qemu-e2k/block/nbd.c

2122 lines
64 KiB
C
Raw Normal View History

/*
* QEMU Block driver for NBD
*
* Copyright (c) 2019 Virtuozzo International GmbH.
* Copyright (C) 2016 Red Hat, Inc.
* Copyright (C) 2008 Bull S.A.S.
* Author: Laurent Vivier <Laurent.Vivier@bull.net>
*
* Some parts:
* Copyright (C) 2007 Anthony Liguori <anthony@codemonkey.ws>
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include "qemu/osdep.h"
#include "trace.h"
#include "qemu/uri.h"
#include "qemu/option.h"
#include "qemu/cutils.h"
#include "qemu/main-loop.h"
#include "qemu/atomic.h"
#include "qapi/qapi-visit-sockets.h"
#include "qapi/qmp/qstring.h"
block/nbd: use non-blocking connect: fix vm hang on connect() This makes nbd's connection_co yield during reconnects, so that reconnect doesn't block the main thread. This is very important in case of an unavailable nbd server host: connect() call may take a long time, blocking the main thread (and due to reconnect, it will hang again and again with small gaps of working time during pauses between connection attempts). Realization notes: - We don't want to implement non-blocking connect() over non-blocking socket, because getaddrinfo() doesn't have portable non-blocking realization anyway, so let's just use a thread for both getaddrinfo() and connect(). - We can't use qio_channel_socket_connect_async (which behaves similarly and starts a thread to execute connect() call), as it's relying on someone iterating main loop (g_main_loop_run() or something like this), which is not always the case. - We can't use thread_pool_submit_co API, as thread pool waits for all threads to finish (but we don't want to wait for blocking reconnect attempt on shutdown. So, we just create the thread by hand. Some additional difficulties are: - We want our connect to avoid blocking drained sections and aio context switches. To achieve this, we make it possible to "cancel" synchronous wait for the connect (which is a coroutine yield actually), still, the thread continues in background, and if successful, its result may be reused on next reconnect attempt. - We don't want to wait for reconnect on shutdown, so there is CONNECT_THREAD_RUNNING_DETACHED thread state, which means that the block layer is no longer interested in a result, and thread should close new connected socket on finish and free the state. How to reproduce the bug, fixed with this commit: 1. Create an image on node1: qemu-img create -f qcow2 xx 100M 2. Start NBD server on node1: qemu-nbd xx 3. Start vm with second nbd disk on node2, like this: ./x86_64-softmmu/qemu-system-x86_64 -nodefaults -drive \ file=/work/images/cent7.qcow2 -drive file=nbd+tcp://192.168.100.2 \ -vnc :0 -qmp stdio -m 2G -enable-kvm -vga std 4. Access the vm through vnc (or some other way?), and check that NBD drive works: dd if=/dev/sdb of=/dev/null bs=1M count=10 - the command should succeed. 5. Now, let's trigger nbd-reconnect loop in Qemu process. For this: 5.1 Kill NBD server on node1 5.2 run "dd if=/dev/sdb of=/dev/null bs=1M count=10" in the guest again. The command should fail and a lot of error messages about failing disk may appear as well. Now NBD client driver in Qemu tries to reconnect. Still, VM works well. 6. Make node1 unavailable on NBD port, so connect() from node2 will last for a long time: On node1 (Note, that 10809 is just a default NBD port): sudo iptables -A INPUT -p tcp --dport 10809 -j DROP After some time the guest hangs, and you may check in gdb that Qemu hangs in connect() call, issued from the main thread. This is the BUG. 7. Don't forget to drop iptables rule from your node1: sudo iptables -D INPUT -p tcp --dport 10809 -j DROP Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20200812145237.4396-1-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: minor wording and formatting tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2020-08-12 16:52:37 +02:00
#include "qapi/clone-visitor.h"
#include "block/qdict.h"
#include "block/nbd.h"
#include "block/block_int.h"
#include "block/coroutines.h"
#include "qemu/yank.h"
#define EN_OPTSTR ":exportname="
#define MAX_NBD_REQUESTS 16
#define HANDLE_TO_INDEX(bs, handle) ((handle) ^ (uint64_t)(intptr_t)(bs))
#define INDEX_TO_HANDLE(bs, index) ((index) ^ (uint64_t)(intptr_t)(bs))
typedef struct {
Coroutine *coroutine;
uint64_t offset; /* original offset of the request */
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
bool receiving; /* sleeping in the yield in nbd_receive_replies */
bool reply_possible; /* reply header not yet received */
} NBDClientRequest;
typedef enum NBDClientState {
NBD_CLIENT_CONNECTING_WAIT,
NBD_CLIENT_CONNECTING_NOWAIT,
NBD_CLIENT_CONNECTED,
NBD_CLIENT_QUIT
} NBDClientState;
typedef struct BDRVNBDState {
QIOChannel *ioc; /* The current I/O channel */
NBDExportInfo info;
CoMutex send_mutex;
CoQueue free_sema;
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
CoMutex receive_mutex;
int in_flight;
NBDClientState state;
block/nbd: fix reconnect-delay reconnect-delay has a design flaw: we handle it in the same loop where we do connection attempt. So, reconnect-delay may be exceeded by unpredictable time of connection attempt. Let's instead use separate timer. How to reproduce the bug: 1. Create an image on node1: qemu-img create -f qcow2 xx 100M 2. Start NBD server on node1: qemu-nbd xx 3. On node2 start qemu-io: ./build/qemu-io --image-opts \ driver=nbd,server.type=inet,server.host=192.168.100.5,server.port=10809,reconnect-delay=15 4. Type 'read 0 512' in qemu-io interface to check that connection works Be careful: you should make steps 5-7 in a short time, less than 15 seconds. 5. Kill nbd server on node1 6. Run 'read 0 512' in qemu-io interface again, to be sure that nbd client goes to reconnect loop. 7. On node1 run the following command sudo iptables -A INPUT -p tcp --dport 10809 -j DROP This will make the connect() call of qemu-io at node2 take a long time. And you'll see that read command in qemu-io will hang for a long time, more than 15 seconds specified by reconnect-delay parameter. It's the bug. 8. Don't forget to drop iptables rule on node1: sudo iptables -D INPUT -p tcp --dport 10809 -j DROP Important note: Step [5] is necessary to reproduce _this_ bug. If we miss step [5], the read command (step 6) will hang for a long time and this commit doesn't help, because there will be not long connect() to unreachable host, but long sendmsg() to unreachable host, which should be fixed by enabling and adjusting keep-alive on the socket, which is a thing for further patch set. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20200903190301.367620-4-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com>
2020-09-03 21:03:00 +02:00
QEMUTimer *reconnect_delay_timer;
QEMUTimer *open_timer;
block/nbd: fix reconnect-delay reconnect-delay has a design flaw: we handle it in the same loop where we do connection attempt. So, reconnect-delay may be exceeded by unpredictable time of connection attempt. Let's instead use separate timer. How to reproduce the bug: 1. Create an image on node1: qemu-img create -f qcow2 xx 100M 2. Start NBD server on node1: qemu-nbd xx 3. On node2 start qemu-io: ./build/qemu-io --image-opts \ driver=nbd,server.type=inet,server.host=192.168.100.5,server.port=10809,reconnect-delay=15 4. Type 'read 0 512' in qemu-io interface to check that connection works Be careful: you should make steps 5-7 in a short time, less than 15 seconds. 5. Kill nbd server on node1 6. Run 'read 0 512' in qemu-io interface again, to be sure that nbd client goes to reconnect loop. 7. On node1 run the following command sudo iptables -A INPUT -p tcp --dport 10809 -j DROP This will make the connect() call of qemu-io at node2 take a long time. And you'll see that read command in qemu-io will hang for a long time, more than 15 seconds specified by reconnect-delay parameter. It's the bug. 8. Don't forget to drop iptables rule on node1: sudo iptables -D INPUT -p tcp --dport 10809 -j DROP Important note: Step [5] is necessary to reproduce _this_ bug. If we miss step [5], the read command (step 6) will hang for a long time and this commit doesn't help, because there will be not long connect() to unreachable host, but long sendmsg() to unreachable host, which should be fixed by enabling and adjusting keep-alive on the socket, which is a thing for further patch set. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20200903190301.367620-4-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com>
2020-09-03 21:03:00 +02:00
NBDClientRequest requests[MAX_NBD_REQUESTS];
NBDReply reply;
BlockDriverState *bs;
/* Connection parameters */
uint32_t reconnect_delay;
uint32_t open_timeout;
SocketAddress *saddr;
char *export, *tlscredsid;
QCryptoTLSCreds *tlscreds;
const char *hostname;
char *x_dirty_bitmap;
bool alloc_depth;
block/nbd: use non-blocking connect: fix vm hang on connect() This makes nbd's connection_co yield during reconnects, so that reconnect doesn't block the main thread. This is very important in case of an unavailable nbd server host: connect() call may take a long time, blocking the main thread (and due to reconnect, it will hang again and again with small gaps of working time during pauses between connection attempts). Realization notes: - We don't want to implement non-blocking connect() over non-blocking socket, because getaddrinfo() doesn't have portable non-blocking realization anyway, so let's just use a thread for both getaddrinfo() and connect(). - We can't use qio_channel_socket_connect_async (which behaves similarly and starts a thread to execute connect() call), as it's relying on someone iterating main loop (g_main_loop_run() or something like this), which is not always the case. - We can't use thread_pool_submit_co API, as thread pool waits for all threads to finish (but we don't want to wait for blocking reconnect attempt on shutdown. So, we just create the thread by hand. Some additional difficulties are: - We want our connect to avoid blocking drained sections and aio context switches. To achieve this, we make it possible to "cancel" synchronous wait for the connect (which is a coroutine yield actually), still, the thread continues in background, and if successful, its result may be reused on next reconnect attempt. - We don't want to wait for reconnect on shutdown, so there is CONNECT_THREAD_RUNNING_DETACHED thread state, which means that the block layer is no longer interested in a result, and thread should close new connected socket on finish and free the state. How to reproduce the bug, fixed with this commit: 1. Create an image on node1: qemu-img create -f qcow2 xx 100M 2. Start NBD server on node1: qemu-nbd xx 3. Start vm with second nbd disk on node2, like this: ./x86_64-softmmu/qemu-system-x86_64 -nodefaults -drive \ file=/work/images/cent7.qcow2 -drive file=nbd+tcp://192.168.100.2 \ -vnc :0 -qmp stdio -m 2G -enable-kvm -vga std 4. Access the vm through vnc (or some other way?), and check that NBD drive works: dd if=/dev/sdb of=/dev/null bs=1M count=10 - the command should succeed. 5. Now, let's trigger nbd-reconnect loop in Qemu process. For this: 5.1 Kill NBD server on node1 5.2 run "dd if=/dev/sdb of=/dev/null bs=1M count=10" in the guest again. The command should fail and a lot of error messages about failing disk may appear as well. Now NBD client driver in Qemu tries to reconnect. Still, VM works well. 6. Make node1 unavailable on NBD port, so connect() from node2 will last for a long time: On node1 (Note, that 10809 is just a default NBD port): sudo iptables -A INPUT -p tcp --dport 10809 -j DROP After some time the guest hangs, and you may check in gdb that Qemu hangs in connect() call, issued from the main thread. This is the BUG. 7. Don't forget to drop iptables rule from your node1: sudo iptables -D INPUT -p tcp --dport 10809 -j DROP Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20200812145237.4396-1-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: minor wording and formatting tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2020-08-12 16:52:37 +02:00
NBDClientConnection *conn;
} BDRVNBDState;
static void nbd_yank(void *opaque);
static void nbd_clear_bdrvstate(BlockDriverState *bs)
{
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
nbd_client_connection_release(s->conn);
s->conn = NULL;
yank_unregister_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name));
object_unref(OBJECT(s->tlscreds));
qapi_free_SocketAddress(s->saddr);
s->saddr = NULL;
g_free(s->export);
s->export = NULL;
g_free(s->tlscredsid);
s->tlscredsid = NULL;
g_free(s->x_dirty_bitmap);
s->x_dirty_bitmap = NULL;
}
static bool nbd_client_connected(BDRVNBDState *s)
{
return qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTED;
}
static bool nbd_recv_coroutine_wake_one(NBDClientRequest *req)
{
if (req->receiving) {
req->receiving = false;
aio_co_wake(req->coroutine);
return true;
}
return false;
}
static void nbd_recv_coroutines_wake(BDRVNBDState *s, bool all)
{
int i;
for (i = 0; i < MAX_NBD_REQUESTS; i++) {
if (nbd_recv_coroutine_wake_one(&s->requests[i]) && !all) {
return;
}
}
}
static void nbd_channel_error(BDRVNBDState *s, int ret)
{
if (nbd_client_connected(s)) {
qio_channel_shutdown(s->ioc, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
}
if (ret == -EIO) {
if (nbd_client_connected(s)) {
s->state = s->reconnect_delay ? NBD_CLIENT_CONNECTING_WAIT :
NBD_CLIENT_CONNECTING_NOWAIT;
}
} else {
s->state = NBD_CLIENT_QUIT;
}
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
nbd_recv_coroutines_wake(s, true);
}
block/nbd: fix reconnect-delay reconnect-delay has a design flaw: we handle it in the same loop where we do connection attempt. So, reconnect-delay may be exceeded by unpredictable time of connection attempt. Let's instead use separate timer. How to reproduce the bug: 1. Create an image on node1: qemu-img create -f qcow2 xx 100M 2. Start NBD server on node1: qemu-nbd xx 3. On node2 start qemu-io: ./build/qemu-io --image-opts \ driver=nbd,server.type=inet,server.host=192.168.100.5,server.port=10809,reconnect-delay=15 4. Type 'read 0 512' in qemu-io interface to check that connection works Be careful: you should make steps 5-7 in a short time, less than 15 seconds. 5. Kill nbd server on node1 6. Run 'read 0 512' in qemu-io interface again, to be sure that nbd client goes to reconnect loop. 7. On node1 run the following command sudo iptables -A INPUT -p tcp --dport 10809 -j DROP This will make the connect() call of qemu-io at node2 take a long time. And you'll see that read command in qemu-io will hang for a long time, more than 15 seconds specified by reconnect-delay parameter. It's the bug. 8. Don't forget to drop iptables rule on node1: sudo iptables -D INPUT -p tcp --dport 10809 -j DROP Important note: Step [5] is necessary to reproduce _this_ bug. If we miss step [5], the read command (step 6) will hang for a long time and this commit doesn't help, because there will be not long connect() to unreachable host, but long sendmsg() to unreachable host, which should be fixed by enabling and adjusting keep-alive on the socket, which is a thing for further patch set. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20200903190301.367620-4-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com>
2020-09-03 21:03:00 +02:00
static void reconnect_delay_timer_del(BDRVNBDState *s)
{
if (s->reconnect_delay_timer) {
timer_free(s->reconnect_delay_timer);
s->reconnect_delay_timer = NULL;
}
}
static void reconnect_delay_timer_cb(void *opaque)
{
BDRVNBDState *s = opaque;
if (qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTING_WAIT) {
block/nbd: fix reconnect-delay reconnect-delay has a design flaw: we handle it in the same loop where we do connection attempt. So, reconnect-delay may be exceeded by unpredictable time of connection attempt. Let's instead use separate timer. How to reproduce the bug: 1. Create an image on node1: qemu-img create -f qcow2 xx 100M 2. Start NBD server on node1: qemu-nbd xx 3. On node2 start qemu-io: ./build/qemu-io --image-opts \ driver=nbd,server.type=inet,server.host=192.168.100.5,server.port=10809,reconnect-delay=15 4. Type 'read 0 512' in qemu-io interface to check that connection works Be careful: you should make steps 5-7 in a short time, less than 15 seconds. 5. Kill nbd server on node1 6. Run 'read 0 512' in qemu-io interface again, to be sure that nbd client goes to reconnect loop. 7. On node1 run the following command sudo iptables -A INPUT -p tcp --dport 10809 -j DROP This will make the connect() call of qemu-io at node2 take a long time. And you'll see that read command in qemu-io will hang for a long time, more than 15 seconds specified by reconnect-delay parameter. It's the bug. 8. Don't forget to drop iptables rule on node1: sudo iptables -D INPUT -p tcp --dport 10809 -j DROP Important note: Step [5] is necessary to reproduce _this_ bug. If we miss step [5], the read command (step 6) will hang for a long time and this commit doesn't help, because there will be not long connect() to unreachable host, but long sendmsg() to unreachable host, which should be fixed by enabling and adjusting keep-alive on the socket, which is a thing for further patch set. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20200903190301.367620-4-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com>
2020-09-03 21:03:00 +02:00
s->state = NBD_CLIENT_CONNECTING_NOWAIT;
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
nbd_co_establish_connection_cancel(s->conn);
block/nbd: fix reconnect-delay reconnect-delay has a design flaw: we handle it in the same loop where we do connection attempt. So, reconnect-delay may be exceeded by unpredictable time of connection attempt. Let's instead use separate timer. How to reproduce the bug: 1. Create an image on node1: qemu-img create -f qcow2 xx 100M 2. Start NBD server on node1: qemu-nbd xx 3. On node2 start qemu-io: ./build/qemu-io --image-opts \ driver=nbd,server.type=inet,server.host=192.168.100.5,server.port=10809,reconnect-delay=15 4. Type 'read 0 512' in qemu-io interface to check that connection works Be careful: you should make steps 5-7 in a short time, less than 15 seconds. 5. Kill nbd server on node1 6. Run 'read 0 512' in qemu-io interface again, to be sure that nbd client goes to reconnect loop. 7. On node1 run the following command sudo iptables -A INPUT -p tcp --dport 10809 -j DROP This will make the connect() call of qemu-io at node2 take a long time. And you'll see that read command in qemu-io will hang for a long time, more than 15 seconds specified by reconnect-delay parameter. It's the bug. 8. Don't forget to drop iptables rule on node1: sudo iptables -D INPUT -p tcp --dport 10809 -j DROP Important note: Step [5] is necessary to reproduce _this_ bug. If we miss step [5], the read command (step 6) will hang for a long time and this commit doesn't help, because there will be not long connect() to unreachable host, but long sendmsg() to unreachable host, which should be fixed by enabling and adjusting keep-alive on the socket, which is a thing for further patch set. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20200903190301.367620-4-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com>
2020-09-03 21:03:00 +02:00
while (qemu_co_enter_next(&s->free_sema, NULL)) {
/* Resume all queued requests */
}
}
reconnect_delay_timer_del(s);
}
static void reconnect_delay_timer_init(BDRVNBDState *s, uint64_t expire_time_ns)
{
if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTING_WAIT) {
block/nbd: fix reconnect-delay reconnect-delay has a design flaw: we handle it in the same loop where we do connection attempt. So, reconnect-delay may be exceeded by unpredictable time of connection attempt. Let's instead use separate timer. How to reproduce the bug: 1. Create an image on node1: qemu-img create -f qcow2 xx 100M 2. Start NBD server on node1: qemu-nbd xx 3. On node2 start qemu-io: ./build/qemu-io --image-opts \ driver=nbd,server.type=inet,server.host=192.168.100.5,server.port=10809,reconnect-delay=15 4. Type 'read 0 512' in qemu-io interface to check that connection works Be careful: you should make steps 5-7 in a short time, less than 15 seconds. 5. Kill nbd server on node1 6. Run 'read 0 512' in qemu-io interface again, to be sure that nbd client goes to reconnect loop. 7. On node1 run the following command sudo iptables -A INPUT -p tcp --dport 10809 -j DROP This will make the connect() call of qemu-io at node2 take a long time. And you'll see that read command in qemu-io will hang for a long time, more than 15 seconds specified by reconnect-delay parameter. It's the bug. 8. Don't forget to drop iptables rule on node1: sudo iptables -D INPUT -p tcp --dport 10809 -j DROP Important note: Step [5] is necessary to reproduce _this_ bug. If we miss step [5], the read command (step 6) will hang for a long time and this commit doesn't help, because there will be not long connect() to unreachable host, but long sendmsg() to unreachable host, which should be fixed by enabling and adjusting keep-alive on the socket, which is a thing for further patch set. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20200903190301.367620-4-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com>
2020-09-03 21:03:00 +02:00
return;
}
assert(!s->reconnect_delay_timer);
s->reconnect_delay_timer = aio_timer_new(bdrv_get_aio_context(s->bs),
QEMU_CLOCK_REALTIME,
SCALE_NS,
reconnect_delay_timer_cb, s);
timer_mod(s->reconnect_delay_timer, expire_time_ns);
}
static void nbd_teardown_connection(BlockDriverState *bs)
{
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
assert(!s->in_flight);
if (s->ioc) {
qio_channel_shutdown(s->ioc, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
yank_unregister_function(BLOCKDEV_YANK_INSTANCE(s->bs->node_name),
nbd_yank, s->bs);
object_unref(OBJECT(s->ioc));
s->ioc = NULL;
}
s->state = NBD_CLIENT_QUIT;
}
static void open_timer_del(BDRVNBDState *s)
{
if (s->open_timer) {
timer_free(s->open_timer);
s->open_timer = NULL;
}
}
static void open_timer_cb(void *opaque)
{
BDRVNBDState *s = opaque;
nbd_co_establish_connection_cancel(s->conn);
open_timer_del(s);
}
static void open_timer_init(BDRVNBDState *s, uint64_t expire_time_ns)
{
assert(!s->open_timer);
s->open_timer = aio_timer_new(bdrv_get_aio_context(s->bs),
QEMU_CLOCK_REALTIME,
SCALE_NS,
open_timer_cb, s);
timer_mod(s->open_timer, expire_time_ns);
}
static bool nbd_client_connecting(BDRVNBDState *s)
{
NBDClientState state = qatomic_load_acquire(&s->state);
return state == NBD_CLIENT_CONNECTING_WAIT ||
state == NBD_CLIENT_CONNECTING_NOWAIT;
}
static bool nbd_client_connecting_wait(BDRVNBDState *s)
{
return qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTING_WAIT;
}
/*
* Update @bs with information learned during a completed negotiation process.
* Return failure if the server's advertised options are incompatible with the
* client's needs.
*/
static int nbd_handle_updated_info(BlockDriverState *bs, Error **errp)
{
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
int ret;
if (s->x_dirty_bitmap) {
if (!s->info.base_allocation) {
error_setg(errp, "requested x-dirty-bitmap %s not found",
s->x_dirty_bitmap);
return -EINVAL;
}
if (strcmp(s->x_dirty_bitmap, "qemu:allocation-depth") == 0) {
s->alloc_depth = true;
}
}
if (s->info.flags & NBD_FLAG_READ_ONLY) {
ret = bdrv_apply_auto_read_only(bs, "NBD export is read-only", errp);
if (ret < 0) {
return ret;
}
}
if (s->info.flags & NBD_FLAG_SEND_FUA) {
bs->supported_write_flags = BDRV_REQ_FUA;
bs->supported_zero_flags |= BDRV_REQ_FUA;
}
if (s->info.flags & NBD_FLAG_SEND_WRITE_ZEROES) {
bs->supported_zero_flags |= BDRV_REQ_MAY_UNMAP;
if (s->info.flags & NBD_FLAG_SEND_FAST_ZERO) {
bs->supported_zero_flags |= BDRV_REQ_NO_FALLBACK;
}
}
trace_nbd_client_handshake_success(s->export);
return 0;
}
int coroutine_fn nbd_co_do_establish_connection(BlockDriverState *bs,
Error **errp)
{
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
int ret;
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
bool blocking = nbd_client_connecting_wait(s);
assert(!s->ioc);
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
s->ioc = nbd_co_establish_connection(s->conn, &s->info, blocking, errp);
if (!s->ioc) {
return -ECONNREFUSED;
}
yank_register_function(BLOCKDEV_YANK_INSTANCE(s->bs->node_name), nbd_yank,
bs);
ret = nbd_handle_updated_info(s->bs, NULL);
if (ret < 0) {
/*
* We have connected, but must fail for other reasons.
* Send NBD_CMD_DISC as a courtesy to the server.
*/
NBDRequest request = { .type = NBD_CMD_DISC };
nbd_send_request(s->ioc, &request);
yank_unregister_function(BLOCKDEV_YANK_INSTANCE(s->bs->node_name),
nbd_yank, bs);
object_unref(OBJECT(s->ioc));
s->ioc = NULL;
return ret;
}
qio_channel_set_blocking(s->ioc, false, NULL);
qio_channel_attach_aio_context(s->ioc, bdrv_get_aio_context(bs));
/* successfully connected */
s->state = NBD_CLIENT_CONNECTED;
qemu_co_queue_restart_all(&s->free_sema);
return 0;
}
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
/* called under s->send_mutex */
static coroutine_fn void nbd_reconnect_attempt(BDRVNBDState *s)
{
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
assert(nbd_client_connecting(s));
assert(s->in_flight == 0);
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
if (nbd_client_connecting_wait(s) && s->reconnect_delay &&
!s->reconnect_delay_timer)
{
/*
* It's first reconnect attempt after switching to
* NBD_CLIENT_CONNECTING_WAIT
*/
reconnect_delay_timer_init(s,
qemu_clock_get_ns(QEMU_CLOCK_REALTIME) +
s->reconnect_delay * NANOSECONDS_PER_SECOND);
}
/*
* Now we are sure that nobody is accessing the channel, and no one will
* try until we set the state to CONNECTED.
*/
/* Finalize previous connection if any */
if (s->ioc) {
qio_channel_detach_aio_context(QIO_CHANNEL(s->ioc));
yank_unregister_function(BLOCKDEV_YANK_INSTANCE(s->bs->node_name),
nbd_yank, s->bs);
object_unref(OBJECT(s->ioc));
s->ioc = NULL;
}
nbd_co_do_establish_connection(s->bs, NULL);
}
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
static coroutine_fn int nbd_receive_replies(BDRVNBDState *s, uint64_t handle)
{
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
int ret;
uint64_t ind = HANDLE_TO_INDEX(s, handle), ind2;
QEMU_LOCK_GUARD(&s->receive_mutex);
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
while (true) {
if (s->reply.handle == handle) {
/* We are done */
return 0;
}
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
if (!nbd_client_connected(s)) {
return -EIO;
}
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
if (s->reply.handle != 0) {
/*
* Some other request is being handled now. It should already be
* woken by whoever set s->reply.handle (or never wait in this
* yield). So, we should not wake it here.
*/
ind2 = HANDLE_TO_INDEX(s, s->reply.handle);
assert(!s->requests[ind2].receiving);
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
s->requests[ind].receiving = true;
qemu_co_mutex_unlock(&s->receive_mutex);
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
qemu_coroutine_yield();
/*
* We may be woken for 3 reasons:
* 1. From this function, executing in parallel coroutine, when our
* handle is received.
* 2. From nbd_channel_error(), when connection is lost.
* 3. From nbd_co_receive_one_chunk(), when previous request is
* finished and s->reply.handle set to 0.
* Anyway, it's OK to lock the mutex and go to the next iteration.
*/
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
qemu_co_mutex_lock(&s->receive_mutex);
assert(!s->requests[ind].receiving);
continue;
}
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
/* We are under mutex and handle is 0. We have to do the dirty work. */
assert(s->reply.handle == 0);
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
ret = nbd_receive_reply(s->bs, s->ioc, &s->reply, NULL);
if (ret <= 0) {
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
ret = ret ? ret : -EIO;
nbd_channel_error(s, ret);
return ret;
}
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
if (nbd_reply_is_structured(&s->reply) && !s->info.structured_reply) {
nbd_channel_error(s, -EINVAL);
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
return -EINVAL;
}
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
if (s->reply.handle == handle) {
/* We are done */
return 0;
}
ind2 = HANDLE_TO_INDEX(s, s->reply.handle);
if (ind2 >= MAX_NBD_REQUESTS || !s->requests[ind2].reply_possible) {
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
nbd_channel_error(s, -EINVAL);
return -EINVAL;
}
nbd_recv_coroutine_wake_one(&s->requests[ind2]);
}
}
static int nbd_co_send_request(BlockDriverState *bs,
NBDRequest *request,
QEMUIOVector *qiov)
{
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
int rc, i = -1;
qemu_co_mutex_lock(&s->send_mutex);
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
while (s->in_flight == MAX_NBD_REQUESTS ||
(!nbd_client_connected(s) && s->in_flight > 0))
{
qemu_co_queue_wait(&s->free_sema, &s->send_mutex);
}
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
if (nbd_client_connecting(s)) {
nbd_reconnect_attempt(s);
}
if (!nbd_client_connected(s)) {
rc = -EIO;
goto err;
}
s->in_flight++;
for (i = 0; i < MAX_NBD_REQUESTS; i++) {
if (s->requests[i].coroutine == NULL) {
break;
}
}
g_assert(qemu_in_coroutine());
assert(i < MAX_NBD_REQUESTS);
s->requests[i].coroutine = qemu_coroutine_self();
s->requests[i].offset = request->from;
s->requests[i].receiving = false;
s->requests[i].reply_possible = true;
request->handle = INDEX_TO_HANDLE(s, i);
assert(s->ioc);
if (qiov) {
qio_channel_set_cork(s->ioc, true);
rc = nbd_send_request(s->ioc, request);
if (nbd_client_connected(s) && rc >= 0) {
if (qio_channel_writev_all(s->ioc, qiov->iov, qiov->niov,
NULL) < 0) {
rc = -EIO;
}
} else if (rc >= 0) {
rc = -EIO;
}
qio_channel_set_cork(s->ioc, false);
} else {
rc = nbd_send_request(s->ioc, request);
}
err:
if (rc < 0) {
nbd_channel_error(s, rc);
if (i != -1) {
s->requests[i].coroutine = NULL;
s->in_flight--;
qemu_co_queue_next(&s->free_sema);
}
}
qemu_co_mutex_unlock(&s->send_mutex);
return rc;
}
static inline uint16_t payload_advance16(uint8_t **payload)
{
*payload += 2;
return lduw_be_p(*payload - 2);
}
static inline uint32_t payload_advance32(uint8_t **payload)
{
*payload += 4;
return ldl_be_p(*payload - 4);
}
static inline uint64_t payload_advance64(uint8_t **payload)
{
*payload += 8;
return ldq_be_p(*payload - 8);
}
static int nbd_parse_offset_hole_payload(BDRVNBDState *s,
NBDStructuredReplyChunk *chunk,
uint8_t *payload, uint64_t orig_offset,
QEMUIOVector *qiov, Error **errp)
{
uint64_t offset;
uint32_t hole_size;
if (chunk->length != sizeof(offset) + sizeof(hole_size)) {
error_setg(errp, "Protocol error: invalid payload for "
"NBD_REPLY_TYPE_OFFSET_HOLE");
return -EINVAL;
}
offset = payload_advance64(&payload);
hole_size = payload_advance32(&payload);
if (!hole_size || offset < orig_offset || hole_size > qiov->size ||
offset > orig_offset + qiov->size - hole_size) {
error_setg(errp, "Protocol error: server sent chunk exceeding requested"
" region");
return -EINVAL;
}
if (s->info.min_block &&
!QEMU_IS_ALIGNED(hole_size, s->info.min_block)) {
trace_nbd_structured_read_compliance("hole");
}
qemu_iovec_memset(qiov, offset - orig_offset, 0, hole_size);
return 0;
}
/*
* nbd_parse_blockstatus_payload
* Based on our request, we expect only one extent in reply, for the
* base:allocation context.
*/
static int nbd_parse_blockstatus_payload(BDRVNBDState *s,
NBDStructuredReplyChunk *chunk,
uint8_t *payload, uint64_t orig_length,
NBDExtent *extent, Error **errp)
{
uint32_t context_id;
/* The server succeeded, so it must have sent [at least] one extent */
if (chunk->length < sizeof(context_id) + sizeof(*extent)) {
error_setg(errp, "Protocol error: invalid payload for "
"NBD_REPLY_TYPE_BLOCK_STATUS");
return -EINVAL;
}
context_id = payload_advance32(&payload);
if (s->info.context_id != context_id) {
error_setg(errp, "Protocol error: unexpected context id %d for "
"NBD_REPLY_TYPE_BLOCK_STATUS, when negotiated context "
"id is %d", context_id,
s->info.context_id);
return -EINVAL;
}
extent->length = payload_advance32(&payload);
extent->flags = payload_advance32(&payload);
if (extent->length == 0) {
error_setg(errp, "Protocol error: server sent status chunk with "
"zero length");
return -EINVAL;
}
/*
* A server sending unaligned block status is in violation of the
* protocol, but as qemu-nbd 3.1 is such a server (at least for
* POSIX files that are not a multiple of 512 bytes, since qemu
* rounds files up to 512-byte multiples but lseek(SEEK_HOLE)
* still sees an implicit hole beyond the real EOF), it's nicer to
* work around the misbehaving server. If the request included
* more than the final unaligned block, truncate it back to an
* aligned result; if the request was only the final block, round
* up to the full block and change the status to fully-allocated
* (always a safe status, even if it loses information).
*/
if (s->info.min_block && !QEMU_IS_ALIGNED(extent->length,
s->info.min_block)) {
trace_nbd_parse_blockstatus_compliance("extent length is unaligned");
if (extent->length > s->info.min_block) {
extent->length = QEMU_ALIGN_DOWN(extent->length,
s->info.min_block);
} else {
extent->length = s->info.min_block;
extent->flags = 0;
}
}
/*
* We used NBD_CMD_FLAG_REQ_ONE, so the server should not have
* sent us any more than one extent, nor should it have included
* status beyond our request in that extent. However, it's easy
* enough to ignore the server's noncompliance without killing the
* connection; just ignore trailing extents, and clamp things to
* the length of our request.
*/
if (chunk->length > sizeof(context_id) + sizeof(*extent)) {
trace_nbd_parse_blockstatus_compliance("more than one extent");
}
if (extent->length > orig_length) {
extent->length = orig_length;
trace_nbd_parse_blockstatus_compliance("extent length too large");
}
/*
* HACK: if we are using x-dirty-bitmaps to access
* qemu:allocation-depth, treat all depths > 2 the same as 2,
* since nbd_client_co_block_status is only expecting the low two
* bits to be set.
*/
if (s->alloc_depth && extent->flags > 2) {
extent->flags = 2;
}
return 0;
}
/*
* nbd_parse_error_payload
* on success @errp contains message describing nbd error reply
*/
static int nbd_parse_error_payload(NBDStructuredReplyChunk *chunk,
uint8_t *payload, int *request_ret,
Error **errp)
{
uint32_t error;
uint16_t message_size;
assert(chunk->type & (1 << 15));
if (chunk->length < sizeof(error) + sizeof(message_size)) {
error_setg(errp,
"Protocol error: invalid payload for structured error");
return -EINVAL;
}
error = nbd_errno_to_system_errno(payload_advance32(&payload));
if (error == 0) {
error_setg(errp, "Protocol error: server sent structured error chunk "
"with error = 0");
return -EINVAL;
}
*request_ret = -error;
message_size = payload_advance16(&payload);
if (message_size > chunk->length - sizeof(error) - sizeof(message_size)) {
error_setg(errp, "Protocol error: server sent structured error chunk "
"with incorrect message size");
return -EINVAL;
}
/* TODO: Add a trace point to mention the server complaint */
/* TODO handle ERROR_OFFSET */
return 0;
}
static int nbd_co_receive_offset_data_payload(BDRVNBDState *s,
uint64_t orig_offset,
QEMUIOVector *qiov, Error **errp)
{
QEMUIOVector sub_qiov;
uint64_t offset;
size_t data_size;
int ret;
NBDStructuredReplyChunk *chunk = &s->reply.structured;
assert(nbd_reply_is_structured(&s->reply));
/* The NBD spec requires at least one byte of payload */
if (chunk->length <= sizeof(offset)) {
error_setg(errp, "Protocol error: invalid payload for "
"NBD_REPLY_TYPE_OFFSET_DATA");
return -EINVAL;
}
if (nbd_read64(s->ioc, &offset, "OFFSET_DATA offset", errp) < 0) {
return -EIO;
}
data_size = chunk->length - sizeof(offset);
assert(data_size);
if (offset < orig_offset || data_size > qiov->size ||
offset > orig_offset + qiov->size - data_size) {
error_setg(errp, "Protocol error: server sent chunk exceeding requested"
" region");
return -EINVAL;
}
if (s->info.min_block && !QEMU_IS_ALIGNED(data_size, s->info.min_block)) {
trace_nbd_structured_read_compliance("data");
}
qemu_iovec_init(&sub_qiov, qiov->niov);
qemu_iovec_concat(&sub_qiov, qiov, offset - orig_offset, data_size);
ret = qio_channel_readv_all(s->ioc, sub_qiov.iov, sub_qiov.niov, errp);
qemu_iovec_destroy(&sub_qiov);
return ret < 0 ? -EIO : 0;
}
#define NBD_MAX_MALLOC_PAYLOAD 1000
static coroutine_fn int nbd_co_receive_structured_payload(
BDRVNBDState *s, void **payload, Error **errp)
{
int ret;
uint32_t len;
assert(nbd_reply_is_structured(&s->reply));
len = s->reply.structured.length;
if (len == 0) {
return 0;
}
if (payload == NULL) {
error_setg(errp, "Unexpected structured payload");
return -EINVAL;
}
if (len > NBD_MAX_MALLOC_PAYLOAD) {
error_setg(errp, "Payload too large");
return -EINVAL;
}
*payload = g_new(char, len);
ret = nbd_read(s->ioc, *payload, len, "structured payload", errp);
if (ret < 0) {
g_free(*payload);
*payload = NULL;
return ret;
}
return 0;
}
/*
* nbd_co_do_receive_one_chunk
* for simple reply:
* set request_ret to received reply error
* if qiov is not NULL: read payload to @qiov
* for structured reply chunk:
* if error chunk: read payload, set @request_ret, do not set @payload
* else if offset_data chunk: read payload data to @qiov, do not set @payload
* else: read payload to @payload
*
* If function fails, @errp contains corresponding error message, and the
* connection with the server is suspect. If it returns 0, then the
* transaction succeeded (although @request_ret may be a negative errno
* corresponding to the server's error reply), and errp is unchanged.
*/
static coroutine_fn int nbd_co_do_receive_one_chunk(
BDRVNBDState *s, uint64_t handle, bool only_structured,
int *request_ret, QEMUIOVector *qiov, void **payload, Error **errp)
{
int ret;
int i = HANDLE_TO_INDEX(s, handle);
void *local_payload = NULL;
NBDStructuredReplyChunk *chunk;
if (payload) {
*payload = NULL;
}
*request_ret = 0;
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
nbd_receive_replies(s, handle);
if (!nbd_client_connected(s)) {
error_setg(errp, "Connection closed");
return -EIO;
}
assert(s->ioc);
assert(s->reply.handle == handle);
if (nbd_reply_is_simple(&s->reply)) {
if (only_structured) {
error_setg(errp, "Protocol error: simple reply when structured "
"reply chunk was expected");
return -EINVAL;
}
*request_ret = -nbd_errno_to_system_errno(s->reply.simple.error);
if (*request_ret < 0 || !qiov) {
return 0;
}
return qio_channel_readv_all(s->ioc, qiov->iov, qiov->niov,
errp) < 0 ? -EIO : 0;
}
/* handle structured reply chunk */
assert(s->info.structured_reply);
chunk = &s->reply.structured;
if (chunk->type == NBD_REPLY_TYPE_NONE) {
if (!(chunk->flags & NBD_REPLY_FLAG_DONE)) {
error_setg(errp, "Protocol error: NBD_REPLY_TYPE_NONE chunk without"
" NBD_REPLY_FLAG_DONE flag set");
return -EINVAL;
}
if (chunk->length) {
error_setg(errp, "Protocol error: NBD_REPLY_TYPE_NONE chunk with"
" nonzero length");
return -EINVAL;
}
return 0;
}
if (chunk->type == NBD_REPLY_TYPE_OFFSET_DATA) {
if (!qiov) {
error_setg(errp, "Unexpected NBD_REPLY_TYPE_OFFSET_DATA chunk");
return -EINVAL;
}
return nbd_co_receive_offset_data_payload(s, s->requests[i].offset,
qiov, errp);
}
if (nbd_reply_type_is_error(chunk->type)) {
payload = &local_payload;
}
ret = nbd_co_receive_structured_payload(s, payload, errp);
if (ret < 0) {
return ret;
}
if (nbd_reply_type_is_error(chunk->type)) {
ret = nbd_parse_error_payload(chunk, local_payload, request_ret, errp);
g_free(local_payload);
return ret;
}
return 0;
}
/*
* nbd_co_receive_one_chunk
* Read reply, wake up connection_co and set s->quit if needed.
* Return value is a fatal error code or normal nbd reply error code
*/
static coroutine_fn int nbd_co_receive_one_chunk(
BDRVNBDState *s, uint64_t handle, bool only_structured,
int *request_ret, QEMUIOVector *qiov, NBDReply *reply, void **payload,
Error **errp)
{
int ret = nbd_co_do_receive_one_chunk(s, handle, only_structured,
request_ret, qiov, payload, errp);
if (ret < 0) {
memset(reply, 0, sizeof(*reply));
nbd_channel_error(s, ret);
} else {
/* For assert at loop start in nbd_connection_entry */
*reply = s->reply;
}
s->reply.handle = 0;
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
nbd_recv_coroutines_wake(s, false);
return ret;
}
typedef struct NBDReplyChunkIter {
int ret;
int request_ret;
Error *err;
bool done, only_structured;
} NBDReplyChunkIter;
static void nbd_iter_channel_error(NBDReplyChunkIter *iter,
int ret, Error **local_err)
{
assert(local_err && *local_err);
assert(ret < 0);
if (!iter->ret) {
iter->ret = ret;
error_propagate(&iter->err, *local_err);
} else {
error_free(*local_err);
}
*local_err = NULL;
}
static void nbd_iter_request_error(NBDReplyChunkIter *iter, int ret)
{
assert(ret < 0);
if (!iter->request_ret) {
iter->request_ret = ret;
}
}
/*
* NBD_FOREACH_REPLY_CHUNK
* The pointer stored in @payload requires g_free() to free it.
*/
#define NBD_FOREACH_REPLY_CHUNK(s, iter, handle, structured, \
qiov, reply, payload) \
for (iter = (NBDReplyChunkIter) { .only_structured = structured }; \
nbd_reply_chunk_iter_receive(s, &iter, handle, qiov, reply, payload);)
/*
* nbd_reply_chunk_iter_receive
* The pointer stored in @payload requires g_free() to free it.
*/
static bool nbd_reply_chunk_iter_receive(BDRVNBDState *s,
NBDReplyChunkIter *iter,
uint64_t handle,
QEMUIOVector *qiov, NBDReply *reply,
void **payload)
{
int ret, request_ret;
NBDReply local_reply;
NBDStructuredReplyChunk *chunk;
Error *local_err = NULL;
if (!nbd_client_connected(s)) {
error_setg(&local_err, "Connection closed");
nbd_iter_channel_error(iter, -EIO, &local_err);
goto break_loop;
}
if (iter->done) {
/* Previous iteration was last. */
goto break_loop;
}
if (reply == NULL) {
reply = &local_reply;
}
ret = nbd_co_receive_one_chunk(s, handle, iter->only_structured,
&request_ret, qiov, reply, payload,
&local_err);
if (ret < 0) {
nbd_iter_channel_error(iter, ret, &local_err);
} else if (request_ret < 0) {
nbd_iter_request_error(iter, request_ret);
}
/* Do not execute the body of NBD_FOREACH_REPLY_CHUNK for simple reply. */
if (nbd_reply_is_simple(reply) || !nbd_client_connected(s)) {
goto break_loop;
}
chunk = &reply->structured;
iter->only_structured = true;
if (chunk->type == NBD_REPLY_TYPE_NONE) {
/* NBD_REPLY_FLAG_DONE is already checked in nbd_co_receive_one_chunk */
assert(chunk->flags & NBD_REPLY_FLAG_DONE);
goto break_loop;
}
if (chunk->flags & NBD_REPLY_FLAG_DONE) {
/* This iteration is last. */
iter->done = true;
}
/* Execute the loop body */
return true;
break_loop:
s->requests[HANDLE_TO_INDEX(s, handle)].coroutine = NULL;
qemu_co_mutex_lock(&s->send_mutex);
s->in_flight--;
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
qemu_co_queue_next(&s->free_sema);
qemu_co_mutex_unlock(&s->send_mutex);
return false;
}
static int nbd_co_receive_return_code(BDRVNBDState *s, uint64_t handle,
int *request_ret, Error **errp)
{
NBDReplyChunkIter iter;
NBD_FOREACH_REPLY_CHUNK(s, iter, handle, false, NULL, NULL, NULL) {
/* nbd_reply_chunk_iter_receive does all the work */
}
error_propagate(errp, iter.err);
*request_ret = iter.request_ret;
return iter.ret;
}
static int nbd_co_receive_cmdread_reply(BDRVNBDState *s, uint64_t handle,
uint64_t offset, QEMUIOVector *qiov,
int *request_ret, Error **errp)
{
NBDReplyChunkIter iter;
NBDReply reply;
void *payload = NULL;
Error *local_err = NULL;
NBD_FOREACH_REPLY_CHUNK(s, iter, handle, s->info.structured_reply,
qiov, &reply, &payload)
{
int ret;
NBDStructuredReplyChunk *chunk = &reply.structured;
assert(nbd_reply_is_structured(&reply));
switch (chunk->type) {
case NBD_REPLY_TYPE_OFFSET_DATA:
/*
* special cased in nbd_co_receive_one_chunk, data is already
* in qiov
*/
break;
case NBD_REPLY_TYPE_OFFSET_HOLE:
ret = nbd_parse_offset_hole_payload(s, &reply.structured, payload,
offset, qiov, &local_err);
if (ret < 0) {
nbd_channel_error(s, ret);
nbd_iter_channel_error(&iter, ret, &local_err);
}
break;
default:
if (!nbd_reply_type_is_error(chunk->type)) {
/* not allowed reply type */
nbd_channel_error(s, -EINVAL);
error_setg(&local_err,
"Unexpected reply type: %d (%s) for CMD_READ",
chunk->type, nbd_reply_type_lookup(chunk->type));
nbd_iter_channel_error(&iter, -EINVAL, &local_err);
}
}
g_free(payload);
payload = NULL;
}
error_propagate(errp, iter.err);
*request_ret = iter.request_ret;
return iter.ret;
}
static int nbd_co_receive_blockstatus_reply(BDRVNBDState *s,
uint64_t handle, uint64_t length,
NBDExtent *extent,
int *request_ret, Error **errp)
{
NBDReplyChunkIter iter;
NBDReply reply;
void *payload = NULL;
Error *local_err = NULL;
bool received = false;
assert(!extent->length);
NBD_FOREACH_REPLY_CHUNK(s, iter, handle, false, NULL, &reply, &payload) {
int ret;
NBDStructuredReplyChunk *chunk = &reply.structured;
assert(nbd_reply_is_structured(&reply));
switch (chunk->type) {
case NBD_REPLY_TYPE_BLOCK_STATUS:
if (received) {
nbd_channel_error(s, -EINVAL);
error_setg(&local_err, "Several BLOCK_STATUS chunks in reply");
nbd_iter_channel_error(&iter, -EINVAL, &local_err);
}
received = true;
ret = nbd_parse_blockstatus_payload(s, &reply.structured,
payload, length, extent,
&local_err);
if (ret < 0) {
nbd_channel_error(s, ret);
nbd_iter_channel_error(&iter, ret, &local_err);
}
break;
default:
if (!nbd_reply_type_is_error(chunk->type)) {
nbd_channel_error(s, -EINVAL);
error_setg(&local_err,
"Unexpected reply type: %d (%s) "
"for CMD_BLOCK_STATUS",
chunk->type, nbd_reply_type_lookup(chunk->type));
nbd_iter_channel_error(&iter, -EINVAL, &local_err);
}
}
g_free(payload);
payload = NULL;
}
if (!extent->length && !iter.request_ret) {
error_setg(&local_err, "Server did not reply with any status extents");
nbd_iter_channel_error(&iter, -EIO, &local_err);
}
error_propagate(errp, iter.err);
*request_ret = iter.request_ret;
return iter.ret;
}
static int nbd_co_request(BlockDriverState *bs, NBDRequest *request,
QEMUIOVector *write_qiov)
{
int ret, request_ret;
Error *local_err = NULL;
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
assert(request->type != NBD_CMD_READ);
if (write_qiov) {
assert(request->type == NBD_CMD_WRITE);
assert(request->len == iov_size(write_qiov->iov, write_qiov->niov));
} else {
assert(request->type != NBD_CMD_WRITE);
}
do {
ret = nbd_co_send_request(bs, request, write_qiov);
if (ret < 0) {
continue;
}
ret = nbd_co_receive_return_code(s, request->handle,
&request_ret, &local_err);
if (local_err) {
trace_nbd_co_request_fail(request->from, request->len,
request->handle, request->flags,
request->type,
nbd_cmd_lookup(request->type),
ret, error_get_pretty(local_err));
error_free(local_err);
local_err = NULL;
}
} while (ret < 0 && nbd_client_connecting_wait(s));
return ret ? ret : request_ret;
}
block: use int64_t instead of uint64_t in driver read handlers We are generally moving to int64_t for both offset and bytes parameters on all io paths. Main motivation is realization of 64-bit write_zeroes operation for fast zeroing large disk chunks, up to the whole disk. We chose signed type, to be consistent with off_t (which is signed) and with possibility for signed return type (where negative value means error). So, convert driver read handlers parameters which are already 64bit to signed type. While being here, convert also flags parameter to be BdrvRequestFlags. Now let's consider all callers. Simple git grep '\->bdrv_\(aio\|co\)_preadv\(_part\)\?' shows that's there three callers of driver function: bdrv_driver_preadv() in block/io.c, passes int64_t, checked by bdrv_check_qiov_request() to be non-negative. qcow2_load_vmstate() does bdrv_check_qiov_request(). do_perform_cow_read() has uint64_t argument. And a lot of things in qcow2 driver are uint64_t, so converting it is big job. But we must not work with requests that don't satisfy bdrv_check_qiov_request(), so let's just assert it here. Still, the functions may be called directly, not only by drv->... Let's check: git grep '\.bdrv_\(aio\|co\)_preadv\(_part\)\?\s*=' | \ awk '{print $4}' | sed 's/,//' | sed 's/&//' | sort | uniq | \ while read func; do git grep "$func(" | \ grep -v "$func(BlockDriverState"; done The only one such caller: QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, &data, 1); ... ret = bdrv_replace_test_co_preadv(bs, 0, 1, &qiov, 0); in tests/unit/test-bdrv-drain.c, and it's OK obviously. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210903102807.27127-4-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: fix typos] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-03 12:27:59 +02:00
static int nbd_client_co_preadv(BlockDriverState *bs, int64_t offset,
int64_t bytes, QEMUIOVector *qiov,
BdrvRequestFlags flags)
{
int ret, request_ret;
Error *local_err = NULL;
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
NBDRequest request = {
.type = NBD_CMD_READ,
.from = offset,
.len = bytes,
};
assert(bytes <= NBD_MAX_BUFFER_SIZE);
assert(!flags);
if (!bytes) {
return 0;
}
/*
* Work around the fact that the block layer doesn't do
* byte-accurate sizing yet - if the read exceeds the server's
* advertised size because the block layer rounded size up, then
* truncate the request to the server and tail-pad with zero.
*/
if (offset >= s->info.size) {
assert(bytes < BDRV_SECTOR_SIZE);
qemu_iovec_memset(qiov, 0, 0, bytes);
return 0;
}
if (offset + bytes > s->info.size) {
uint64_t slop = offset + bytes - s->info.size;
assert(slop < BDRV_SECTOR_SIZE);
qemu_iovec_memset(qiov, bytes - slop, 0, slop);
request.len -= slop;
}
do {
ret = nbd_co_send_request(bs, &request, NULL);
if (ret < 0) {
continue;
}
ret = nbd_co_receive_cmdread_reply(s, request.handle, offset, qiov,
&request_ret, &local_err);
if (local_err) {
trace_nbd_co_request_fail(request.from, request.len, request.handle,
request.flags, request.type,
nbd_cmd_lookup(request.type),
ret, error_get_pretty(local_err));
error_free(local_err);
local_err = NULL;
}
} while (ret < 0 && nbd_client_connecting_wait(s));
return ret ? ret : request_ret;
}
block: use int64_t instead of uint64_t in driver write handlers We are generally moving to int64_t for both offset and bytes parameters on all io paths. Main motivation is realization of 64-bit write_zeroes operation for fast zeroing large disk chunks, up to the whole disk. We chose signed type, to be consistent with off_t (which is signed) and with possibility for signed return type (where negative value means error). So, convert driver write handlers parameters which are already 64bit to signed type. While being here, convert also flags parameter to be BdrvRequestFlags. Now let's consider all callers. Simple git grep '\->bdrv_\(aio\|co\)_pwritev\(_part\)\?' shows that's there three callers of driver function: bdrv_driver_pwritev() and bdrv_driver_pwritev_compressed() in block/io.c, both pass int64_t, checked by bdrv_check_qiov_request() to be non-negative. qcow2_save_vmstate() does bdrv_check_qiov_request(). Still, the functions may be called directly, not only by drv->... Let's check: git grep '\.bdrv_\(aio\|co\)_pwritev\(_part\)\?\s*=' | \ awk '{print $4}' | sed 's/,//' | sed 's/&//' | sort | uniq | \ while read func; do git grep "$func(" | \ grep -v "$func(BlockDriverState"; done shows several callers: qcow2: qcow2_co_truncate() write at most up to @offset, which is checked in generic qcow2_co_truncate() by bdrv_check_request(). qcow2_co_pwritev_compressed_task() pass the request (or part of the request) that already went through normal write path, so it should be OK qcow: qcow_co_pwritev_compressed() pass int64_t, it's updated by this patch quorum: quorum_co_pwrite_zeroes() pass int64_t and int - OK throttle: throttle_co_pwritev_compressed() pass int64_t, it's updated by this patch vmdk: vmdk_co_pwritev_compressed() pass int64_t, it's updated by this patch Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210903102807.27127-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-03 12:28:00 +02:00
static int nbd_client_co_pwritev(BlockDriverState *bs, int64_t offset,
int64_t bytes, QEMUIOVector *qiov,
BdrvRequestFlags flags)
{
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
NBDRequest request = {
.type = NBD_CMD_WRITE,
.from = offset,
.len = bytes,
};
assert(!(s->info.flags & NBD_FLAG_READ_ONLY));
if (flags & BDRV_REQ_FUA) {
assert(s->info.flags & NBD_FLAG_SEND_FUA);
request.flags |= NBD_CMD_FLAG_FUA;
}
assert(bytes <= NBD_MAX_BUFFER_SIZE);
if (!bytes) {
return 0;
}
return nbd_co_request(bs, &request, qiov);
}
static int nbd_client_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset,
block: use int64_t instead of int in driver write_zeroes handlers We are generally moving to int64_t for both offset and bytes parameters on all io paths. Main motivation is realization of 64-bit write_zeroes operation for fast zeroing large disk chunks, up to the whole disk. We chose signed type, to be consistent with off_t (which is signed) and with possibility for signed return type (where negative value means error). So, convert driver write_zeroes handlers bytes parameter to int64_t. The only caller of all updated function is bdrv_co_do_pwrite_zeroes(). bdrv_co_do_pwrite_zeroes() itself is of course OK with widening of callee parameter type. Also, bdrv_co_do_pwrite_zeroes()'s max_write_zeroes is limited to INT_MAX. So, updated functions all are safe, they will not get "bytes" larger than before. Still, let's look through all updated functions, and add assertions to the ones which are actually unprepared to values larger than INT_MAX. For these drivers also set explicit max_pwrite_zeroes limit. Let's go: blkdebug: calculations can't overflow, thanks to bdrv_check_qiov_request() in generic layer. rule_check() and bdrv_co_pwrite_zeroes() both have 64bit argument. blklogwrites: pass to blk_log_writes_co_log() with 64bit argument. blkreplay, copy-on-read, filter-compress: pass to bdrv_co_pwrite_zeroes() which is OK copy-before-write: Calls cbw_do_copy_before_write() and bdrv_co_pwrite_zeroes, both have 64bit argument. file-posix: both handler calls raw_do_pwrite_zeroes, which is updated. In raw_do_pwrite_zeroes() calculations are OK due to bdrv_check_qiov_request(), bytes go to RawPosixAIOData::aio_nbytes which is uint64_t. Check also where that uint64_t gets handed: handle_aiocb_write_zeroes_block() passes a uint64_t[2] to ioctl(BLKZEROOUT), handle_aiocb_write_zeroes() calls do_fallocate() which takes off_t (and we compile to always have 64-bit off_t), as does handle_aiocb_write_zeroes_unmap. All look safe. gluster: bytes go to GlusterAIOCB::size which is int64_t and to glfs_zerofill_async works with off_t. iscsi: Aha, here we deal with iscsi_writesame16_task() that has uint32_t num_blocks argument and iscsi_writesame16_task() has uint16_t argument. Make comments, add assertions and clarify max_pwrite_zeroes calculation. iscsi_allocmap_() functions already has int64_t argument is_byte_request_lun_aligned is simple to update, do it. mirror_top: pass to bdrv_mirror_top_do_write which has uint64_t argument nbd: Aha, here we have protocol limitation, and NBDRequest::len is uint32_t. max_pwrite_zeroes is cleanly set to 32bit value, so we are OK for now. nvme: Again, protocol limitation. And no inherent limit for write-zeroes at all. But from code that calculates cdw12 it's obvious that we do have limit and alignment. Let's clarify it. Also, obviously the code is not prepared to handle bytes=0. Let's handle this case too. trace events already 64bit preallocate: pass to handle_write() and bdrv_co_pwrite_zeroes(), both 64bit. rbd: pass to qemu_rbd_start_co() which is 64bit. qcow2: offset + bytes and alignment still works good (thanks to bdrv_check_qiov_request()), so tail calculation is OK qcow2_subcluster_zeroize() has 64bit argument, should be OK trace events updated qed: qed_co_request wants int nb_sectors. Also in code we have size_t used for request length which may be 32bit. So, let's just keep INT_MAX as a limit (aligning it down to pwrite_zeroes_alignment) and don't care. raw-format: Is OK. raw_adjust_offset and bdrv_co_pwrite_zeroes are both 64bit. throttle: Both throttle_group_co_io_limits_intercept() and bdrv_co_pwrite_zeroes() are 64bit. vmdk: pass to vmdk_pwritev which is 64bit quorum: pass to quorum_co_pwritev() which is 64bit Hooray! At this point all block drivers are prepared to support 64bit write-zero requests, or have explicitly set max_pwrite_zeroes. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210903102807.27127-8-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: use <= rather than < in assertions relying on max_pwrite_zeroes] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-03 12:28:03 +02:00
int64_t bytes, BdrvRequestFlags flags)
{
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
NBDRequest request = {
.type = NBD_CMD_WRITE_ZEROES,
.from = offset,
block: use int64_t instead of int in driver write_zeroes handlers We are generally moving to int64_t for both offset and bytes parameters on all io paths. Main motivation is realization of 64-bit write_zeroes operation for fast zeroing large disk chunks, up to the whole disk. We chose signed type, to be consistent with off_t (which is signed) and with possibility for signed return type (where negative value means error). So, convert driver write_zeroes handlers bytes parameter to int64_t. The only caller of all updated function is bdrv_co_do_pwrite_zeroes(). bdrv_co_do_pwrite_zeroes() itself is of course OK with widening of callee parameter type. Also, bdrv_co_do_pwrite_zeroes()'s max_write_zeroes is limited to INT_MAX. So, updated functions all are safe, they will not get "bytes" larger than before. Still, let's look through all updated functions, and add assertions to the ones which are actually unprepared to values larger than INT_MAX. For these drivers also set explicit max_pwrite_zeroes limit. Let's go: blkdebug: calculations can't overflow, thanks to bdrv_check_qiov_request() in generic layer. rule_check() and bdrv_co_pwrite_zeroes() both have 64bit argument. blklogwrites: pass to blk_log_writes_co_log() with 64bit argument. blkreplay, copy-on-read, filter-compress: pass to bdrv_co_pwrite_zeroes() which is OK copy-before-write: Calls cbw_do_copy_before_write() and bdrv_co_pwrite_zeroes, both have 64bit argument. file-posix: both handler calls raw_do_pwrite_zeroes, which is updated. In raw_do_pwrite_zeroes() calculations are OK due to bdrv_check_qiov_request(), bytes go to RawPosixAIOData::aio_nbytes which is uint64_t. Check also where that uint64_t gets handed: handle_aiocb_write_zeroes_block() passes a uint64_t[2] to ioctl(BLKZEROOUT), handle_aiocb_write_zeroes() calls do_fallocate() which takes off_t (and we compile to always have 64-bit off_t), as does handle_aiocb_write_zeroes_unmap. All look safe. gluster: bytes go to GlusterAIOCB::size which is int64_t and to glfs_zerofill_async works with off_t. iscsi: Aha, here we deal with iscsi_writesame16_task() that has uint32_t num_blocks argument and iscsi_writesame16_task() has uint16_t argument. Make comments, add assertions and clarify max_pwrite_zeroes calculation. iscsi_allocmap_() functions already has int64_t argument is_byte_request_lun_aligned is simple to update, do it. mirror_top: pass to bdrv_mirror_top_do_write which has uint64_t argument nbd: Aha, here we have protocol limitation, and NBDRequest::len is uint32_t. max_pwrite_zeroes is cleanly set to 32bit value, so we are OK for now. nvme: Again, protocol limitation. And no inherent limit for write-zeroes at all. But from code that calculates cdw12 it's obvious that we do have limit and alignment. Let's clarify it. Also, obviously the code is not prepared to handle bytes=0. Let's handle this case too. trace events already 64bit preallocate: pass to handle_write() and bdrv_co_pwrite_zeroes(), both 64bit. rbd: pass to qemu_rbd_start_co() which is 64bit. qcow2: offset + bytes and alignment still works good (thanks to bdrv_check_qiov_request()), so tail calculation is OK qcow2_subcluster_zeroize() has 64bit argument, should be OK trace events updated qed: qed_co_request wants int nb_sectors. Also in code we have size_t used for request length which may be 32bit. So, let's just keep INT_MAX as a limit (aligning it down to pwrite_zeroes_alignment) and don't care. raw-format: Is OK. raw_adjust_offset and bdrv_co_pwrite_zeroes are both 64bit. throttle: Both throttle_group_co_io_limits_intercept() and bdrv_co_pwrite_zeroes() are 64bit. vmdk: pass to vmdk_pwritev which is 64bit quorum: pass to quorum_co_pwritev() which is 64bit Hooray! At this point all block drivers are prepared to support 64bit write-zero requests, or have explicitly set max_pwrite_zeroes. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210903102807.27127-8-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: use <= rather than < in assertions relying on max_pwrite_zeroes] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-03 12:28:03 +02:00
.len = bytes, /* .len is uint32_t actually */
};
block: use int64_t instead of int in driver write_zeroes handlers We are generally moving to int64_t for both offset and bytes parameters on all io paths. Main motivation is realization of 64-bit write_zeroes operation for fast zeroing large disk chunks, up to the whole disk. We chose signed type, to be consistent with off_t (which is signed) and with possibility for signed return type (where negative value means error). So, convert driver write_zeroes handlers bytes parameter to int64_t. The only caller of all updated function is bdrv_co_do_pwrite_zeroes(). bdrv_co_do_pwrite_zeroes() itself is of course OK with widening of callee parameter type. Also, bdrv_co_do_pwrite_zeroes()'s max_write_zeroes is limited to INT_MAX. So, updated functions all are safe, they will not get "bytes" larger than before. Still, let's look through all updated functions, and add assertions to the ones which are actually unprepared to values larger than INT_MAX. For these drivers also set explicit max_pwrite_zeroes limit. Let's go: blkdebug: calculations can't overflow, thanks to bdrv_check_qiov_request() in generic layer. rule_check() and bdrv_co_pwrite_zeroes() both have 64bit argument. blklogwrites: pass to blk_log_writes_co_log() with 64bit argument. blkreplay, copy-on-read, filter-compress: pass to bdrv_co_pwrite_zeroes() which is OK copy-before-write: Calls cbw_do_copy_before_write() and bdrv_co_pwrite_zeroes, both have 64bit argument. file-posix: both handler calls raw_do_pwrite_zeroes, which is updated. In raw_do_pwrite_zeroes() calculations are OK due to bdrv_check_qiov_request(), bytes go to RawPosixAIOData::aio_nbytes which is uint64_t. Check also where that uint64_t gets handed: handle_aiocb_write_zeroes_block() passes a uint64_t[2] to ioctl(BLKZEROOUT), handle_aiocb_write_zeroes() calls do_fallocate() which takes off_t (and we compile to always have 64-bit off_t), as does handle_aiocb_write_zeroes_unmap. All look safe. gluster: bytes go to GlusterAIOCB::size which is int64_t and to glfs_zerofill_async works with off_t. iscsi: Aha, here we deal with iscsi_writesame16_task() that has uint32_t num_blocks argument and iscsi_writesame16_task() has uint16_t argument. Make comments, add assertions and clarify max_pwrite_zeroes calculation. iscsi_allocmap_() functions already has int64_t argument is_byte_request_lun_aligned is simple to update, do it. mirror_top: pass to bdrv_mirror_top_do_write which has uint64_t argument nbd: Aha, here we have protocol limitation, and NBDRequest::len is uint32_t. max_pwrite_zeroes is cleanly set to 32bit value, so we are OK for now. nvme: Again, protocol limitation. And no inherent limit for write-zeroes at all. But from code that calculates cdw12 it's obvious that we do have limit and alignment. Let's clarify it. Also, obviously the code is not prepared to handle bytes=0. Let's handle this case too. trace events already 64bit preallocate: pass to handle_write() and bdrv_co_pwrite_zeroes(), both 64bit. rbd: pass to qemu_rbd_start_co() which is 64bit. qcow2: offset + bytes and alignment still works good (thanks to bdrv_check_qiov_request()), so tail calculation is OK qcow2_subcluster_zeroize() has 64bit argument, should be OK trace events updated qed: qed_co_request wants int nb_sectors. Also in code we have size_t used for request length which may be 32bit. So, let's just keep INT_MAX as a limit (aligning it down to pwrite_zeroes_alignment) and don't care. raw-format: Is OK. raw_adjust_offset and bdrv_co_pwrite_zeroes are both 64bit. throttle: Both throttle_group_co_io_limits_intercept() and bdrv_co_pwrite_zeroes() are 64bit. vmdk: pass to vmdk_pwritev which is 64bit quorum: pass to quorum_co_pwritev() which is 64bit Hooray! At this point all block drivers are prepared to support 64bit write-zero requests, or have explicitly set max_pwrite_zeroes. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210903102807.27127-8-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: use <= rather than < in assertions relying on max_pwrite_zeroes] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-03 12:28:03 +02:00
assert(bytes <= UINT32_MAX); /* rely on max_pwrite_zeroes */
assert(!(s->info.flags & NBD_FLAG_READ_ONLY));
if (!(s->info.flags & NBD_FLAG_SEND_WRITE_ZEROES)) {
return -ENOTSUP;
}
if (flags & BDRV_REQ_FUA) {
assert(s->info.flags & NBD_FLAG_SEND_FUA);
request.flags |= NBD_CMD_FLAG_FUA;
}
if (!(flags & BDRV_REQ_MAY_UNMAP)) {
request.flags |= NBD_CMD_FLAG_NO_HOLE;
}
if (flags & BDRV_REQ_NO_FALLBACK) {
assert(s->info.flags & NBD_FLAG_SEND_FAST_ZERO);
request.flags |= NBD_CMD_FLAG_FAST_ZERO;
}
if (!bytes) {
return 0;
}
return nbd_co_request(bs, &request, NULL);
}
static int nbd_client_co_flush(BlockDriverState *bs)
{
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
NBDRequest request = { .type = NBD_CMD_FLUSH };
if (!(s->info.flags & NBD_FLAG_SEND_FLUSH)) {
return 0;
}
request.from = 0;
request.len = 0;
return nbd_co_request(bs, &request, NULL);
}
static int nbd_client_co_pdiscard(BlockDriverState *bs, int64_t offset,
block: use int64_t instead of int in driver discard handlers We are generally moving to int64_t for both offset and bytes parameters on all io paths. Main motivation is realization of 64-bit write_zeroes operation for fast zeroing large disk chunks, up to the whole disk. We chose signed type, to be consistent with off_t (which is signed) and with possibility for signed return type (where negative value means error). So, convert driver discard handlers bytes parameter to int64_t. The only caller of all updated function is bdrv_co_pdiscard in block/io.c. It is already prepared to work with 64bit requests, but pass at most max(bs->bl.max_pdiscard, INT_MAX) to the driver. Let's look at all updated functions: blkdebug: all calculations are still OK, thanks to bdrv_check_qiov_request(). both rule_check and bdrv_co_pdiscard are 64bit blklogwrites: pass to blk_loc_writes_co_log which is 64bit blkreplay, copy-on-read, filter-compress: pass to bdrv_co_pdiscard, OK copy-before-write: pass to bdrv_co_pdiscard which is 64bit and to cbw_do_copy_before_write which is 64bit file-posix: one handler calls raw_account_discard() is 64bit and both handlers calls raw_do_pdiscard(). Update raw_do_pdiscard, which pass to RawPosixAIOData::aio_nbytes, which is 64bit (and calls raw_account_discard()) gluster: somehow, third argument of glfs_discard_async is size_t. Let's set max_pdiscard accordingly. iscsi: iscsi_allocmap_set_invalid is 64bit, !is_byte_request_lun_aligned is 64bit. list.num is uint32_t. Let's clarify max_pdiscard and pdiscard_alignment. mirror_top: pass to bdrv_mirror_top_do_write() which is 64bit nbd: protocol limitation. max_pdiscard is alredy set strict enough, keep it as is for now. nvme: buf.nlb is uint32_t and we do shift. So, add corresponding limits to nvme_refresh_limits(). preallocate: pass to bdrv_co_pdiscard() which is 64bit. rbd: pass to qemu_rbd_start_co() which is 64bit. qcow2: calculations are still OK, thanks to bdrv_check_qiov_request(), qcow2_cluster_discard() is 64bit. raw-format: raw_adjust_offset() is 64bit, bdrv_co_pdiscard too. throttle: pass to bdrv_co_pdiscard() which is 64bit and to throttle_group_co_io_limits_intercept() which is 64bit as well. test-block-iothread: bytes argument is unused Great! Now all drivers are prepared to handle 64bit discard requests, or else have explicit max_pdiscard limits. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210903102807.27127-11-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-03 12:28:06 +02:00
int64_t bytes)
{
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
NBDRequest request = {
.type = NBD_CMD_TRIM,
.from = offset,
block: use int64_t instead of int in driver discard handlers We are generally moving to int64_t for both offset and bytes parameters on all io paths. Main motivation is realization of 64-bit write_zeroes operation for fast zeroing large disk chunks, up to the whole disk. We chose signed type, to be consistent with off_t (which is signed) and with possibility for signed return type (where negative value means error). So, convert driver discard handlers bytes parameter to int64_t. The only caller of all updated function is bdrv_co_pdiscard in block/io.c. It is already prepared to work with 64bit requests, but pass at most max(bs->bl.max_pdiscard, INT_MAX) to the driver. Let's look at all updated functions: blkdebug: all calculations are still OK, thanks to bdrv_check_qiov_request(). both rule_check and bdrv_co_pdiscard are 64bit blklogwrites: pass to blk_loc_writes_co_log which is 64bit blkreplay, copy-on-read, filter-compress: pass to bdrv_co_pdiscard, OK copy-before-write: pass to bdrv_co_pdiscard which is 64bit and to cbw_do_copy_before_write which is 64bit file-posix: one handler calls raw_account_discard() is 64bit and both handlers calls raw_do_pdiscard(). Update raw_do_pdiscard, which pass to RawPosixAIOData::aio_nbytes, which is 64bit (and calls raw_account_discard()) gluster: somehow, third argument of glfs_discard_async is size_t. Let's set max_pdiscard accordingly. iscsi: iscsi_allocmap_set_invalid is 64bit, !is_byte_request_lun_aligned is 64bit. list.num is uint32_t. Let's clarify max_pdiscard and pdiscard_alignment. mirror_top: pass to bdrv_mirror_top_do_write() which is 64bit nbd: protocol limitation. max_pdiscard is alredy set strict enough, keep it as is for now. nvme: buf.nlb is uint32_t and we do shift. So, add corresponding limits to nvme_refresh_limits(). preallocate: pass to bdrv_co_pdiscard() which is 64bit. rbd: pass to qemu_rbd_start_co() which is 64bit. qcow2: calculations are still OK, thanks to bdrv_check_qiov_request(), qcow2_cluster_discard() is 64bit. raw-format: raw_adjust_offset() is 64bit, bdrv_co_pdiscard too. throttle: pass to bdrv_co_pdiscard() which is 64bit and to throttle_group_co_io_limits_intercept() which is 64bit as well. test-block-iothread: bytes argument is unused Great! Now all drivers are prepared to handle 64bit discard requests, or else have explicit max_pdiscard limits. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210903102807.27127-11-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-03 12:28:06 +02:00
.len = bytes, /* len is uint32_t */
};
block: use int64_t instead of int in driver discard handlers We are generally moving to int64_t for both offset and bytes parameters on all io paths. Main motivation is realization of 64-bit write_zeroes operation for fast zeroing large disk chunks, up to the whole disk. We chose signed type, to be consistent with off_t (which is signed) and with possibility for signed return type (where negative value means error). So, convert driver discard handlers bytes parameter to int64_t. The only caller of all updated function is bdrv_co_pdiscard in block/io.c. It is already prepared to work with 64bit requests, but pass at most max(bs->bl.max_pdiscard, INT_MAX) to the driver. Let's look at all updated functions: blkdebug: all calculations are still OK, thanks to bdrv_check_qiov_request(). both rule_check and bdrv_co_pdiscard are 64bit blklogwrites: pass to blk_loc_writes_co_log which is 64bit blkreplay, copy-on-read, filter-compress: pass to bdrv_co_pdiscard, OK copy-before-write: pass to bdrv_co_pdiscard which is 64bit and to cbw_do_copy_before_write which is 64bit file-posix: one handler calls raw_account_discard() is 64bit and both handlers calls raw_do_pdiscard(). Update raw_do_pdiscard, which pass to RawPosixAIOData::aio_nbytes, which is 64bit (and calls raw_account_discard()) gluster: somehow, third argument of glfs_discard_async is size_t. Let's set max_pdiscard accordingly. iscsi: iscsi_allocmap_set_invalid is 64bit, !is_byte_request_lun_aligned is 64bit. list.num is uint32_t. Let's clarify max_pdiscard and pdiscard_alignment. mirror_top: pass to bdrv_mirror_top_do_write() which is 64bit nbd: protocol limitation. max_pdiscard is alredy set strict enough, keep it as is for now. nvme: buf.nlb is uint32_t and we do shift. So, add corresponding limits to nvme_refresh_limits(). preallocate: pass to bdrv_co_pdiscard() which is 64bit. rbd: pass to qemu_rbd_start_co() which is 64bit. qcow2: calculations are still OK, thanks to bdrv_check_qiov_request(), qcow2_cluster_discard() is 64bit. raw-format: raw_adjust_offset() is 64bit, bdrv_co_pdiscard too. throttle: pass to bdrv_co_pdiscard() which is 64bit and to throttle_group_co_io_limits_intercept() which is 64bit as well. test-block-iothread: bytes argument is unused Great! Now all drivers are prepared to handle 64bit discard requests, or else have explicit max_pdiscard limits. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210903102807.27127-11-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-03 12:28:06 +02:00
assert(bytes <= UINT32_MAX); /* rely on max_pdiscard */
assert(!(s->info.flags & NBD_FLAG_READ_ONLY));
if (!(s->info.flags & NBD_FLAG_SEND_TRIM) || !bytes) {
return 0;
}
return nbd_co_request(bs, &request, NULL);
}
static int coroutine_fn nbd_client_co_block_status(
BlockDriverState *bs, bool want_zero, int64_t offset, int64_t bytes,
int64_t *pnum, int64_t *map, BlockDriverState **file)
{
int ret, request_ret;
NBDExtent extent = { 0 };
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
Error *local_err = NULL;
NBDRequest request = {
.type = NBD_CMD_BLOCK_STATUS,
.from = offset,
.len = MIN(QEMU_ALIGN_DOWN(INT_MAX, bs->bl.request_alignment),
MIN(bytes, s->info.size - offset)),
.flags = NBD_CMD_FLAG_REQ_ONE,
};
if (!s->info.base_allocation) {
*pnum = bytes;
*map = offset;
*file = bs;
return BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
}
/*
* Work around the fact that the block layer doesn't do
* byte-accurate sizing yet - if the status request exceeds the
* server's advertised size because the block layer rounded size
* up, we truncated the request to the server (above), or are
* called on just the hole.
*/
if (offset >= s->info.size) {
*pnum = bytes;
assert(bytes < BDRV_SECTOR_SIZE);
/* Intentionally don't report offset_valid for the hole */
return BDRV_BLOCK_ZERO;
}
if (s->info.min_block) {
assert(QEMU_IS_ALIGNED(request.len, s->info.min_block));
}
do {
ret = nbd_co_send_request(bs, &request, NULL);
if (ret < 0) {
continue;
}
ret = nbd_co_receive_blockstatus_reply(s, request.handle, bytes,
&extent, &request_ret,
&local_err);
if (local_err) {
trace_nbd_co_request_fail(request.from, request.len, request.handle,
request.flags, request.type,
nbd_cmd_lookup(request.type),
ret, error_get_pretty(local_err));
error_free(local_err);
local_err = NULL;
}
} while (ret < 0 && nbd_client_connecting_wait(s));
if (ret < 0 || request_ret < 0) {
return ret ? ret : request_ret;
}
assert(extent.length);
*pnum = extent.length;
*map = offset;
*file = bs;
return (extent.flags & NBD_STATE_HOLE ? 0 : BDRV_BLOCK_DATA) |
(extent.flags & NBD_STATE_ZERO ? BDRV_BLOCK_ZERO : 0) |
BDRV_BLOCK_OFFSET_VALID;
}
static int nbd_client_reopen_prepare(BDRVReopenState *state,
BlockReopenQueue *queue, Error **errp)
{
BDRVNBDState *s = (BDRVNBDState *)state->bs->opaque;
if ((state->flags & BDRV_O_RDWR) && (s->info.flags & NBD_FLAG_READ_ONLY)) {
error_setg(errp, "Can't reopen read-only NBD mount as read/write");
return -EACCES;
}
return 0;
}
static void nbd_yank(void *opaque)
{
BlockDriverState *bs = opaque;
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
qatomic_store_release(&s->state, NBD_CLIENT_QUIT);
qio_channel_shutdown(QIO_CHANNEL(s->ioc), QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
}
static void nbd_client_close(BlockDriverState *bs)
{
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
NBDRequest request = { .type = NBD_CMD_DISC };
if (s->ioc) {
nbd_send_request(s->ioc, &request);
}
nbd_teardown_connection(bs);
}
/*
* Parse nbd_open options
*/
static int nbd_parse_uri(const char *filename, QDict *options)
{
URI *uri;
const char *p;
QueryParams *qp = NULL;
int ret = 0;
bool is_unix;
uri = uri_parse(filename);
if (!uri) {
return -EINVAL;
}
/* transport */
if (!g_strcmp0(uri->scheme, "nbd")) {
is_unix = false;
} else if (!g_strcmp0(uri->scheme, "nbd+tcp")) {
is_unix = false;
} else if (!g_strcmp0(uri->scheme, "nbd+unix")) {
is_unix = true;
} else {
ret = -EINVAL;
goto out;
}
p = uri->path ? uri->path : "";
if (p[0] == '/') {
p++;
}
if (p[0]) {
qdict_put_str(options, "export", p);
}
qp = query_params_parse(uri->query);
if (qp->n > 1 || (is_unix && !qp->n) || (!is_unix && qp->n)) {
ret = -EINVAL;
goto out;
}
if (is_unix) {
/* nbd+unix:///export?socket=path */
if (uri->server || uri->port || strcmp(qp->p[0].name, "socket")) {
ret = -EINVAL;
goto out;
}
qdict_put_str(options, "server.type", "unix");
qdict_put_str(options, "server.path", qp->p[0].value);
} else {
QString *host;
char *port_str;
/* nbd[+tcp]://host[:port]/export */
if (!uri->server) {
ret = -EINVAL;
goto out;
}
/* strip braces from literal IPv6 address */
if (uri->server[0] == '[') {
host = qstring_from_substr(uri->server, 1,
strlen(uri->server) - 1);
} else {
host = qstring_from_str(uri->server);
}
qdict_put_str(options, "server.type", "inet");
nbd: Tidy up blockdev-add interface SocketAddress is a simple union, and simple unions are awkward: they have their variant members wrapped in a "data" object on the wire, and require additional indirections in C. I intend to limit its use to existing external interfaces, and convert all internal interfaces to SocketAddressFlat. BlockdevOptionsNbd is an external interface using SocketAddress. We already use SocketAddressFlat elsewhere in blockdev-add. Replace it by SocketAddressFlat while we can (it's new in 2.9) for simplicity and consistency. For example, { "execute": "blockdev-add", "arguments": { "node-name": "foo", "driver": "nbd", "server": { "type": "inet", "data": { "host": "localhost", "port": "12345" } } } } becomes { "execute": "blockdev-add", "arguments": { "node-name": "foo", "driver": "nbd", "server": { "type": "inet", "host": "localhost", "port": "12345" } } } Since the internal interfaces still take SocketAddress, this requires conversion function socket_address_crumple(). It'll go away when I update the interfaces. Unfortunately, SocketAddress is also visible in -drive since 2.8: -drive if=none,driver=nbd,server.type=inet,server.data.host=127.0.0.1,server.data.port=12345 Nobody should be using it, as it's fairly new and has never been documented, so adding still more compatibility gunk to keep it working isn't worth the trouble. You now have to use -drive if=none,driver=nbd,server.type=inet,server.host=127.0.0.1,server.port=12345 Signed-off-by: Markus Armbruster <armbru@redhat.com> Message-id: 1490895797-29094-9-git-send-email-armbru@redhat.com [mreitz: Change iotest 147 accordingly] Because of this interface change, iotest 147 has to be adapted. Unfortunately, we cannot just flatten all of the addresses because nbd-server-start still takes a plain SocketAddress. Therefore, we need both and this is most easily achieved by writing the SocketAddress into the code and flattening it where necessary. Signed-off-by: Max Reitz <mreitz@redhat.com> Message-id: 20170330221243.17333-1-mreitz@redhat.com Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Max Reitz <mreitz@redhat.com>
2017-03-30 19:43:16 +02:00
qdict_put(options, "server.host", host);
port_str = g_strdup_printf("%d", uri->port ?: NBD_DEFAULT_PORT);
qdict_put_str(options, "server.port", port_str);
g_free(port_str);
}
out:
if (qp) {
query_params_free(qp);
}
uri_free(uri);
return ret;
}
static bool nbd_has_filename_options_conflict(QDict *options, Error **errp)
{
const QDictEntry *e;
for (e = qdict_first(options); e; e = qdict_next(options, e)) {
if (!strcmp(e->key, "host") ||
!strcmp(e->key, "port") ||
!strcmp(e->key, "path") ||
!strcmp(e->key, "export") ||
strstart(e->key, "server.", NULL))
{
error_setg(errp, "Option '%s' cannot be used with a file name",
e->key);
return true;
}
}
return false;
}
static void nbd_parse_filename(const char *filename, QDict *options,
Error **errp)
{
g_autofree char *file = NULL;
char *export_name;
const char *host_spec;
const char *unixpath;
if (nbd_has_filename_options_conflict(options, errp)) {
return;
}
if (strstr(filename, "://")) {
int ret = nbd_parse_uri(filename, options);
if (ret < 0) {
error_setg(errp, "No valid URL specified");
}
return;
}
file = g_strdup(filename);
export_name = strstr(file, EN_OPTSTR);
if (export_name) {
if (export_name[strlen(EN_OPTSTR)] == 0) {
return;
}
export_name[0] = 0; /* truncate 'file' */
export_name += strlen(EN_OPTSTR);
qdict_put_str(options, "export", export_name);
}
/* extract the host_spec - fail if it's not nbd:... */
if (!strstart(file, "nbd:", &host_spec)) {
error_setg(errp, "File name string for NBD must start with 'nbd:'");
return;
}
if (!*host_spec) {
return;
}
/* are we a UNIX or TCP socket? */
if (strstart(host_spec, "unix:", &unixpath)) {
qdict_put_str(options, "server.type", "unix");
qdict_put_str(options, "server.path", unixpath);
} else {
InetSocketAddress *addr = g_new(InetSocketAddress, 1);
if (inet_parse(addr, host_spec, errp)) {
goto out_inet;
}
qdict_put_str(options, "server.type", "inet");
qdict_put_str(options, "server.host", addr->host);
qdict_put_str(options, "server.port", addr->port);
out_inet:
qapi_free_InetSocketAddress(addr);
}
}
static bool nbd_process_legacy_socket_options(QDict *output_options,
QemuOpts *legacy_opts,
Error **errp)
{
const char *path = qemu_opt_get(legacy_opts, "path");
const char *host = qemu_opt_get(legacy_opts, "host");
const char *port = qemu_opt_get(legacy_opts, "port");
const QDictEntry *e;
if (!path && !host && !port) {
return true;
}
for (e = qdict_first(output_options); e; e = qdict_next(output_options, e))
{
if (strstart(e->key, "server.", NULL)) {
error_setg(errp, "Cannot use 'server' and path/host/port at the "
"same time");
return false;
}
}
if (path && host) {
error_setg(errp, "path and host may not be used at the same time");
return false;
} else if (path) {
if (port) {
error_setg(errp, "port may not be used without host");
return false;
}
qdict_put_str(output_options, "server.type", "unix");
qdict_put_str(output_options, "server.path", path);
} else if (host) {
qdict_put_str(output_options, "server.type", "inet");
qdict_put_str(output_options, "server.host", host);
qdict_put_str(output_options, "server.port",
port ?: stringify(NBD_DEFAULT_PORT));
}
return true;
}
static SocketAddress *nbd_config(BDRVNBDState *s, QDict *options,
Error **errp)
{
SocketAddress *saddr = NULL;
QDict *addr = NULL;
Visitor *iv = NULL;
qdict_extract_subqdict(options, &addr, "server.");
if (!qdict_size(addr)) {
error_setg(errp, "NBD server address missing");
goto done;
}
iv = qobject_input_visitor_new_flat_confused(addr, errp);
if (!iv) {
goto done;
}
if (!visit_type_SocketAddress(iv, NULL, &saddr, errp)) {
goto done;
}
if (socket_address_parse_named_fd(saddr, errp) < 0) {
qapi_free_SocketAddress(saddr);
saddr = NULL;
goto done;
}
done:
qobject_unref(addr);
visit_free(iv);
return saddr;
}
static QCryptoTLSCreds *nbd_get_tls_creds(const char *id, Error **errp)
{
Object *obj;
QCryptoTLSCreds *creds;
obj = object_resolve_path_component(
object_get_objects_root(), id);
if (!obj) {
error_setg(errp, "No TLS credentials with id '%s'",
id);
return NULL;
}
creds = (QCryptoTLSCreds *)
object_dynamic_cast(obj, TYPE_QCRYPTO_TLS_CREDS);
if (!creds) {
error_setg(errp, "Object with id '%s' is not TLS credentials",
id);
return NULL;
}
if (!qcrypto_tls_creds_check_endpoint(creds,
QCRYPTO_TLS_CREDS_ENDPOINT_CLIENT,
errp)) {
return NULL;
}
object_ref(obj);
return creds;
}
static QemuOptsList nbd_runtime_opts = {
.name = "nbd",
.head = QTAILQ_HEAD_INITIALIZER(nbd_runtime_opts.head),
.desc = {
{
.name = "host",
.type = QEMU_OPT_STRING,
.help = "TCP host to connect to",
},
{
.name = "port",
.type = QEMU_OPT_STRING,
.help = "TCP port to connect to",
},
{
.name = "path",
.type = QEMU_OPT_STRING,
.help = "Unix socket path to connect to",
},
{
.name = "export",
.type = QEMU_OPT_STRING,
.help = "Name of the NBD export to open",
},
{
.name = "tls-creds",
.type = QEMU_OPT_STRING,
.help = "ID of the TLS credentials to use",
},
{
.name = "x-dirty-bitmap",
.type = QEMU_OPT_STRING,
.help = "experimental: expose named dirty bitmap in place of "
"block status",
},
{
.name = "reconnect-delay",
.type = QEMU_OPT_NUMBER,
.help = "On an unexpected disconnect, the nbd client tries to "
"connect again until succeeding or encountering a serious "
"error. During the first @reconnect-delay seconds, all "
"requests are paused and will be rerun on a successful "
"reconnect. After that time, any delayed requests and all "
"future requests before a successful reconnect will "
"immediately fail. Default 0",
},
{
.name = "open-timeout",
.type = QEMU_OPT_NUMBER,
.help = "In seconds. If zero, the nbd driver tries the connection "
"only once, and fails to open if the connection fails. "
"If non-zero, the nbd driver will repeat connection "
"attempts until successful or until @open-timeout seconds "
"have elapsed. Default 0",
},
block/nbd: fix segmentation fault when .desc is not null-terminated The find_desc_by_name() from util/qemu-option.c relies on the .name not being NULL to call strcmp(). This check becomes unsafe when the list is not NULL-terminated, which is the case of nbd_runtime_opts in block/nbd.c, and can result in segmentation fault when strcmp() tries to access an invalid memory: #0 0x00007fff8c75f7d4 in __strcmp_power9 () from /lib64/libc.so.6 #1 0x00000000102d3ec8 in find_desc_by_name (desc=0x1036d6f0, name=0x28e46670 "server.path") at util/qemu-option.c:166 #2 0x00000000102d93e0 in qemu_opts_absorb_qdict (opts=0x28e47a80, qdict=0x28e469a0, errp=0x7fffec247c98) at util/qemu-option.c:1026 #3 0x000000001012a2e4 in nbd_open (bs=0x28e42290, options=0x28e469a0, flags=24578, errp=0x7fffec247d80) at block/nbd.c:406 #4 0x00000000100144e8 in bdrv_open_driver (bs=0x28e42290, drv=0x1036e070 <bdrv_nbd_unix>, node_name=0x0, options=0x28e469a0, open_flags=24578, errp=0x7fffec247f50) at block.c:1135 #5 0x0000000010015b04 in bdrv_open_common (bs=0x28e42290, file=0x0, options=0x28e469a0, errp=0x7fffec247f50) at block.c:1395 >From gdb, the desc[i].name was not NULL and resulted in strcmp() accessing an invalid memory: >>> p desc[5] $8 = { name = 0x1037f098 "R27A", type = 1561964883, help = 0xc0bbb23e <error: Cannot access memory at address 0xc0bbb23e>, def_value_str = 0x2 <error: Cannot access memory at address 0x2> } >>> p desc[6] $9 = { name = 0x103dac78 <__gcov0.do_qemu_init_bdrv_nbd_init> "\001", type = 272101528, help = 0x29ec0b754403e31f <error: Cannot access memory at address 0x29ec0b754403e31f>, def_value_str = 0x81f343b9 <error: Cannot access memory at address 0x81f343b9> } This patch fixes the segmentation fault in strcmp() by adding a NULL element at the end of nbd_runtime_opts.desc list, which is the common practice to most of other structs like runtime_opts in block/null.c. Thus, the desc[i].name != NULL check becomes safe because it will not evaluate to true when .desc list reached its end. Reported-by: R. Nageswara Sastry <nasastry@in.ibm.com> Buglink: https://bugs.launchpad.net/qemu/+bug/1727259 Signed-off-by: Murilo Opsfelder Araujo <muriloo@linux.vnet.ibm.com> Message-Id: <20180105133241.14141-2-muriloo@linux.vnet.ibm.com> CC: qemu-stable@nongnu.org Fixes: 7ccc44fd7d1dfa62c4d6f3a680df809d6e7068ce Signed-off-by: Eric Blake <eblake@redhat.com>
2018-01-05 14:32:41 +01:00
{ /* end of list */ }
},
};
static int nbd_process_options(BlockDriverState *bs, QDict *options,
Error **errp)
{
BDRVNBDState *s = bs->opaque;
QemuOpts *opts;
int ret = -EINVAL;
opts = qemu_opts_create(&nbd_runtime_opts, NULL, 0, &error_abort);
if (!qemu_opts_absorb_qdict(opts, options, errp)) {
goto error;
}
/* Translate @host, @port, and @path to a SocketAddress */
if (!nbd_process_legacy_socket_options(options, opts, errp)) {
goto error;
}
/* Pop the config into our state object. Exit if invalid. */
s->saddr = nbd_config(s, options, errp);
if (!s->saddr) {
goto error;
}
s->export = g_strdup(qemu_opt_get(opts, "export"));
if (s->export && strlen(s->export) > NBD_MAX_STRING_SIZE) {
error_setg(errp, "export name too long to send to server");
goto error;
}
s->tlscredsid = g_strdup(qemu_opt_get(opts, "tls-creds"));
if (s->tlscredsid) {
s->tlscreds = nbd_get_tls_creds(s->tlscredsid, errp);
if (!s->tlscreds) {
goto error;
}
/* TODO SOCKET_ADDRESS_KIND_FD where fd has AF_INET or AF_INET6 */
if (s->saddr->type != SOCKET_ADDRESS_TYPE_INET) {
error_setg(errp, "TLS only supported over IP sockets");
goto error;
}
s->hostname = s->saddr->u.inet.host;
}
s->x_dirty_bitmap = g_strdup(qemu_opt_get(opts, "x-dirty-bitmap"));
if (s->x_dirty_bitmap && strlen(s->x_dirty_bitmap) > NBD_MAX_STRING_SIZE) {
error_setg(errp, "x-dirty-bitmap query too long to send to server");
goto error;
}
s->reconnect_delay = qemu_opt_get_number(opts, "reconnect-delay", 0);
s->open_timeout = qemu_opt_get_number(opts, "open-timeout", 0);
ret = 0;
error:
qemu_opts_del(opts);
return ret;
}
static int nbd_open(BlockDriverState *bs, QDict *options, int flags,
Error **errp)
{
int ret;
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
s->bs = bs;
qemu_co_mutex_init(&s->send_mutex);
qemu_co_queue_init(&s->free_sema);
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
qemu_co_mutex_init(&s->receive_mutex);
if (!yank_register_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name), errp)) {
return -EEXIST;
}
ret = nbd_process_options(bs, options, errp);
if (ret < 0) {
goto fail;
}
s->conn = nbd_client_connection_new(s->saddr, true, s->export,
s->x_dirty_bitmap, s->tlscreds);
if (s->open_timeout) {
nbd_client_connection_enable_retry(s->conn);
open_timer_init(s, qemu_clock_get_ns(QEMU_CLOCK_REALTIME) +
s->open_timeout * NANOSECONDS_PER_SECOND);
}
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
s->state = NBD_CLIENT_CONNECTING_WAIT;
ret = nbd_do_establish_connection(bs, errp);
if (ret < 0) {
goto fail;
}
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
nbd_client_connection_enable_retry(s->conn);
return 0;
fail:
nbd_clear_bdrvstate(bs);
return ret;
}
static int nbd_co_flush(BlockDriverState *bs)
{
return nbd_client_co_flush(bs);
}
static void nbd_refresh_limits(BlockDriverState *bs, Error **errp)
{
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
nbd: Honor server's advertised minimum block size Commit 79ba8c98 (v2.7) changed the setting of request_alignment to occur only during bdrv_refresh_limits(), rather than at at bdrv_open() time; but at the time, NBD was unaffected, because it still used sector-based callbacks, so the block layer defaulted NBD to use 512 request_alignment. Later, commit 70c4fb26 (also v2.7) changed NBD to use byte-based callbacks, without setting request_alignment. This resulted in NBD using request_alignment of 1, which works great when the server supports it (as is the case for qemu-nbd), but falls apart miserably if the server requires alignment (but only if qemu actually sends a sub-sector request; qemu-io can do it, but most qemu operations still perform on sectors or larger). Even later, the NBD protocol was updated to document that clients should learn the server's minimum alignment during NBD_OPT_GO; and recommended that clients should assume a minimum size of 512 unless the server understands NBD_OPT_GO and replied with a smaller size. Commit 081dd1fe (v2.10) attempted to do that, by assigning request_alignment to whatever was learned from the server; but it has two flaws: the assignment is done during bdrv_open() so it gets unconditionally wiped out back to 1 during any later bdrv_refresh_limits(); and the code is not using a default of 512 when the server did not report a minimum size. Fix these issues by moving the assignment to request_alignment to the right function, and by using a sane default when the server does not advertise a minimum size. CC: qemu-stable@nongnu.org Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20180215032905.27146-1-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy<vsementsov@virtuozzo.com>
2018-02-15 04:29:05 +01:00
uint32_t min = s->info.min_block;
uint32_t max = MIN_NON_ZERO(NBD_MAX_BUFFER_SIZE, s->info.max_block);
nbd/client: Lower min_block for block-status, unaligned size We have a latent bug in our NBD client code, tickled by the brand new nbdkit 1.11.10 block status support: $ nbdkit --filter=log --filter=truncate -U - \ data data="1" size=511 truncate=64K logfile=/dev/stdout \ --run 'qemu-img convert $nbd /var/tmp/out' ... qemu-img: block/io.c:2122: bdrv_co_block_status: Assertion `*pnum && QEMU_IS_ALIGNED(*pnum, align) && align > offset - aligned_offset' failed. The culprit? Our implementation of .bdrv_co_block_status can return unaligned block status for any server that operates with a lower actual alignment than what we tell the block layer in request_alignment, in violation of the block layer's constraints. To date, we've been unable to trip the bug, because qemu as NBD server always advertises block sizing (at which point it is a server bug if the server sends unaligned status - although qemu 3.1 is such a server and I've sent separate patches for 4.0 both to get the server to obey the spec, and to let the client to tolerate server oddities at EOF). But nbdkit does not (yet) advertise block sizing, and therefore is not in violation of the spec for returning block status at whatever boundaries it wants, and those unaligned results can occur anywhere rather than just at EOF. While we are still wise to avoid sending sub-sector read/write requests to a server of unknown origin, we MUST consider that a server telling us block status without an advertised block size is correct. So, we either have to munge unaligned answers from the server into aligned ones that we hand back to the block layer, or we have to tell the block layer about a smaller alignment. Similarly, if the server advertises an image size that is not sector-aligned, we might as well assume that the server intends to let us access those tail bytes, and therefore supports a minimum block size of 1, regardless of whether the server supports block status (although we still need more patches to fix the problem that with an unaligned image, we can send read or block status requests that exceed EOF to the server). Again, qemu as server cannot trip this problem (because it rounds images to sector alignment), but nbdkit advertised unaligned size even before it gained block status support. Solve both alignment problems at once by using better heuristics on what alignment to report to the block layer when the server did not give us something to work with. Note that very few NBD servers implement block status (to date, only qemu and nbdkit are known to do so); and as the NBD spec mentioned block sizing constraints prior to documenting block status, it can be assumed that any future implementations of block status are aware that they must advertise block size if they want a minimum size other than 1. We've had a long history of struggles with picking the right alignment to use in the block layer, as evidenced by the commit message of fd8d372d (v2.12) that introduced the current choice of forced 512-byte alignment. There is no iotest coverage for this fix, because qemu can't provoke it, and I didn't want to make test 241 dependent on nbdkit. Fixes: fd8d372d Reported-by: Richard W.M. Jones <rjones@redhat.com> Signed-off-by: Eric Blake <eblake@redhat.com> Message-Id: <20190329042750.14704-3-eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Tested-by: Richard W.M. Jones <rjones@redhat.com>
2019-03-29 05:27:46 +01:00
/*
* If the server did not advertise an alignment:
* - a size that is not sector-aligned implies that an alignment
* of 1 can be used to access those tail bytes
* - advertisement of block status requires an alignment of 1, so
* that we don't violate block layer constraints that block
* status is always aligned (as we can't control whether the
* server will report sub-sector extents, such as a hole at EOF
* on an unaligned POSIX file)
* - otherwise, assume the server is so old that we are safer avoiding
* sub-sector requests
*/
if (!min) {
min = (!QEMU_IS_ALIGNED(s->info.size, BDRV_SECTOR_SIZE) ||
s->info.base_allocation) ? 1 : BDRV_SECTOR_SIZE;
}
bs->bl.request_alignment = min;
bs->bl.max_pdiscard = QEMU_ALIGN_DOWN(INT_MAX, min);
bs->bl.max_pwrite_zeroes = max;
bs->bl.max_transfer = max;
if (s->info.opt_block &&
s->info.opt_block > bs->bl.opt_transfer) {
bs->bl.opt_transfer = s->info.opt_block;
}
}
static void nbd_close(BlockDriverState *bs)
{
nbd_client_close(bs);
nbd_clear_bdrvstate(bs);
}
block: nbd: Fix convert qcow2 compressed to nbd When converting to qcow2 compressed format, the last step is a special zero length compressed write, ending in a call to bdrv_co_truncate(). This call always fails for the nbd driver since it does not implement bdrv_co_truncate(). For block devices, which have the same limits, the call succeeds since the file driver implements bdrv_co_truncate(). If the caller asked to truncate to the same or smaller size with exact=false, the truncate succeeds. Implement the same logic for nbd. Example failing without this change: In one shell start qemu-nbd: $ truncate -s 1g test.tar $ qemu-nbd --socket=/tmp/nbd.sock --persistent --format=raw --offset 1536 test.tar In another shell convert an image to qcow2 compressed via NBD: $ echo "disk data" > disk.raw $ truncate -s 1g disk.raw $ qemu-img convert -f raw -O qcow2 -c disk1.raw nbd+unix:///?socket=/tmp/nbd.sock; echo $? 1 qemu-img failed, but the conversion was successful: $ qemu-img info nbd+unix:///?socket=/tmp/nbd.sock image: nbd+unix://?socket=/tmp/nbd.sock file format: qcow2 virtual size: 1 GiB (1073741824 bytes) ... $ qemu-img check nbd+unix:///?socket=/tmp/nbd.sock No errors were found on the image. 1/16384 = 0.01% allocated, 100.00% fragmented, 100.00% compressed clusters Image end offset: 393216 $ qemu-img compare disk.raw nbd+unix:///?socket=/tmp/nbd.sock Images are identical. Fixes: https://bugzilla.redhat.com/1860627 Signed-off-by: Nir Soffer <nsoffer@redhat.com> Message-Id: <20200727215846.395443-2-nsoffer@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> [eblake: typo fixes] Signed-off-by: Eric Blake <eblake@redhat.com>
2020-07-27 23:58:43 +02:00
/*
* NBD cannot truncate, but if the caller asks to truncate to the same size, or
* to a smaller size with exact=false, there is no reason to fail the
* operation.
*
* Preallocation mode is ignored since it does not seems useful to fail when
* we never change anything.
*/
static int coroutine_fn nbd_co_truncate(BlockDriverState *bs, int64_t offset,
bool exact, PreallocMode prealloc,
BdrvRequestFlags flags, Error **errp)
{
BDRVNBDState *s = bs->opaque;
if (offset != s->info.size && exact) {
error_setg(errp, "Cannot resize NBD nodes");
return -ENOTSUP;
}
if (offset > s->info.size) {
error_setg(errp, "Cannot grow NBD nodes");
return -EINVAL;
}
return 0;
}
static int64_t nbd_getlength(BlockDriverState *bs)
{
BDRVNBDState *s = bs->opaque;
return s->info.size;
}
static void nbd_refresh_filename(BlockDriverState *bs)
{
BDRVNBDState *s = bs->opaque;
const char *host = NULL, *port = NULL, *path = NULL;
size_t len = 0;
if (s->saddr->type == SOCKET_ADDRESS_TYPE_INET) {
nbd: Tidy up blockdev-add interface SocketAddress is a simple union, and simple unions are awkward: they have their variant members wrapped in a "data" object on the wire, and require additional indirections in C. I intend to limit its use to existing external interfaces, and convert all internal interfaces to SocketAddressFlat. BlockdevOptionsNbd is an external interface using SocketAddress. We already use SocketAddressFlat elsewhere in blockdev-add. Replace it by SocketAddressFlat while we can (it's new in 2.9) for simplicity and consistency. For example, { "execute": "blockdev-add", "arguments": { "node-name": "foo", "driver": "nbd", "server": { "type": "inet", "data": { "host": "localhost", "port": "12345" } } } } becomes { "execute": "blockdev-add", "arguments": { "node-name": "foo", "driver": "nbd", "server": { "type": "inet", "host": "localhost", "port": "12345" } } } Since the internal interfaces still take SocketAddress, this requires conversion function socket_address_crumple(). It'll go away when I update the interfaces. Unfortunately, SocketAddress is also visible in -drive since 2.8: -drive if=none,driver=nbd,server.type=inet,server.data.host=127.0.0.1,server.data.port=12345 Nobody should be using it, as it's fairly new and has never been documented, so adding still more compatibility gunk to keep it working isn't worth the trouble. You now have to use -drive if=none,driver=nbd,server.type=inet,server.host=127.0.0.1,server.port=12345 Signed-off-by: Markus Armbruster <armbru@redhat.com> Message-id: 1490895797-29094-9-git-send-email-armbru@redhat.com [mreitz: Change iotest 147 accordingly] Because of this interface change, iotest 147 has to be adapted. Unfortunately, we cannot just flatten all of the addresses because nbd-server-start still takes a plain SocketAddress. Therefore, we need both and this is most easily achieved by writing the SocketAddress into the code and flattening it where necessary. Signed-off-by: Max Reitz <mreitz@redhat.com> Message-id: 20170330221243.17333-1-mreitz@redhat.com Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Max Reitz <mreitz@redhat.com>
2017-03-30 19:43:16 +02:00
const InetSocketAddress *inet = &s->saddr->u.inet;
if (!inet->has_ipv4 && !inet->has_ipv6 && !inet->has_to) {
host = inet->host;
port = inet->port;
}
} else if (s->saddr->type == SOCKET_ADDRESS_TYPE_UNIX) {
nbd: Tidy up blockdev-add interface SocketAddress is a simple union, and simple unions are awkward: they have their variant members wrapped in a "data" object on the wire, and require additional indirections in C. I intend to limit its use to existing external interfaces, and convert all internal interfaces to SocketAddressFlat. BlockdevOptionsNbd is an external interface using SocketAddress. We already use SocketAddressFlat elsewhere in blockdev-add. Replace it by SocketAddressFlat while we can (it's new in 2.9) for simplicity and consistency. For example, { "execute": "blockdev-add", "arguments": { "node-name": "foo", "driver": "nbd", "server": { "type": "inet", "data": { "host": "localhost", "port": "12345" } } } } becomes { "execute": "blockdev-add", "arguments": { "node-name": "foo", "driver": "nbd", "server": { "type": "inet", "host": "localhost", "port": "12345" } } } Since the internal interfaces still take SocketAddress, this requires conversion function socket_address_crumple(). It'll go away when I update the interfaces. Unfortunately, SocketAddress is also visible in -drive since 2.8: -drive if=none,driver=nbd,server.type=inet,server.data.host=127.0.0.1,server.data.port=12345 Nobody should be using it, as it's fairly new and has never been documented, so adding still more compatibility gunk to keep it working isn't worth the trouble. You now have to use -drive if=none,driver=nbd,server.type=inet,server.host=127.0.0.1,server.port=12345 Signed-off-by: Markus Armbruster <armbru@redhat.com> Message-id: 1490895797-29094-9-git-send-email-armbru@redhat.com [mreitz: Change iotest 147 accordingly] Because of this interface change, iotest 147 has to be adapted. Unfortunately, we cannot just flatten all of the addresses because nbd-server-start still takes a plain SocketAddress. Therefore, we need both and this is most easily achieved by writing the SocketAddress into the code and flattening it where necessary. Signed-off-by: Max Reitz <mreitz@redhat.com> Message-id: 20170330221243.17333-1-mreitz@redhat.com Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Max Reitz <mreitz@redhat.com>
2017-03-30 19:43:16 +02:00
path = s->saddr->u.q_unix.path;
} /* else can't represent as pseudo-filename */
if (path && s->export) {
len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
"nbd+unix:///%s?socket=%s", s->export, path);
} else if (path && !s->export) {
len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
"nbd+unix://?socket=%s", path);
} else if (host && s->export) {
len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
"nbd://%s:%s/%s", host, port, s->export);
} else if (host && !s->export) {
len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
"nbd://%s:%s", host, port);
}
if (len >= sizeof(bs->exact_filename)) {
/* Name is too long to represent exactly, so leave it empty. */
bs->exact_filename[0] = '\0';
}
}
static char *nbd_dirname(BlockDriverState *bs, Error **errp)
{
/* The generic bdrv_dirname() implementation is able to work out some
* directory name for NBD nodes, but that would be wrong. So far there is no
* specification for how "export paths" would work, so NBD does not have
* directory names. */
error_setg(errp, "Cannot generate a base directory for NBD nodes");
return NULL;
}
static const char *const nbd_strong_runtime_opts[] = {
"path",
"host",
"port",
"export",
"tls-creds",
"server.",
NULL
};
static void nbd_cancel_in_flight(BlockDriverState *bs)
{
BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
reconnect_delay_timer_del(s);
if (s->state == NBD_CLIENT_CONNECTING_WAIT) {
s->state = NBD_CLIENT_CONNECTING_NOWAIT;
qemu_co_queue_restart_all(&s->free_sema);
}
block/nbd: drop connection_co OK, that's a big rewrite of the logic. Pre-patch we have an always running coroutine - connection_co. It does reply receiving and reconnecting. And it leads to a lot of difficult and unobvious code around drained sections and context switch. We also abuse bs->in_flight counter which is increased for connection_co and temporary decreased in points where we want to allow drained section to begin. One of these place is in another file: in nbd_read_eof() in nbd/client.c. We also cancel reconnect and requests waiting for reconnect on drained begin which is not correct. And this patch fixes that. Let's finally drop this always running coroutine and go another way: do both reconnect and receiving in request coroutines. The detailed list of changes below (in the sequence of diff hunks). 1. receiving coroutines are woken directly from nbd_channel_error, when we change s->state 2. nbd_co_establish_connection_cancel(): we don't have drain_begin now, and in nbd_teardown_connection() all requests should already be finished (and reconnect is done from request). So nbd_co_establish_connection_cancel() is called from nbd_cancel_in_flight() (to cancel the request that is doing nbd_co_establish_connection()) and from reconnect_delay_timer_cb() (previously we didn't need it, as reconnect delay only should cancel active requests not the reconnection itself). But now reconnection itself is done in the separate thread (we now call nbd_client_connection_enable_retry() in nbd_open()), and we need to cancel the requests that wait in nbd_co_establish_connection() now). 2A. We do receive headers in request coroutine. But we also should dispatch replies for other pending requests. So, nbd_connection_entry() is turned into nbd_receive_replies(), which does reply dispatching while it receives other request headers, and returns when it receives the requested header. 3. All old staff around drained sections and context switch is dropped. In details: - we don't need to move connection_co to new aio context, as we don't have connection_co anymore - we don't have a fake "request" of connection_co (extra increasing in_flight), so don't care with it in drain_begin/end - we don't stop reconnection during drained section anymore. This means that drain_begin may wait for a long time (up to reconnect_delay). But that's an improvement and more correct behavior see below[*] 4. In nbd_teardown_connection() we don't have to wait for connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank is moved here from removed connection_co. 5. In nbd_co_do_establish_connection() we now should handle NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in NBD_CLIENT_CONNECTING_NOWAIT, it still should call nbd_co_establish_connection() (who knows, maybe the connection was already established by another thread in the background). But we shouldn't wait: if nbd_co_establish_connection() can't return new channel immediately the request should fail (we are in NBD_CLIENT_CONNECTING_NOWAIT state). 6. nbd_reconnect_attempt() is simplified: it's now easier to wait for other requests in the caller, so here we just assert that fact. Also delay time is now initialized here: we can easily detect first attempt and start a timer. 7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect retries are fully handle by thread (nbd/client-connection.c), delay timer we initialize in nbd_reconnect_attempt(), we don't have to bother with s->drained and friends. nbd_reconnect_attempt() now called from nbd_co_send_request(). 8. nbd_connection_entry is dropped: reconnect is now handled by nbd_co_send_request(), receiving reply is now handled by nbd_receive_replies(): all handled from request coroutines. 9. So, welcome new nbd_receive_replies() called from request coroutine, that receives reply header instead of nbd_connection_entry(). Like with sending requests, only one coroutine may receive in a moment. So we introduce receive_mutex, which is locked around nbd_receive_reply(). It also protects some related fields. Still, full audit of thread-safety in nbd driver is a separate task. New function waits for a reply with specified handle being received and works rather simple: Under mutex: - if current handle is 0, do receive by hand. If another handle received - switch to other request coroutine, release mutex and yield. Otherwise return success - if current handle == requested handle, we are done - otherwise, release mutex and yield 10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if needed. Also waiting in free_sema queue we now wait for one of two conditions: - connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one) - connectING, in_flight == 0, so we can call nbd_reconnect_attempt() And this logic is protected by s->send_mutex Also, on failure we don't have to care of removed s->connection_co 11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for s->connection_co we just call new nbd_receive_replies(). 12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0, which means that handling of the whole reply is finished. Here we need to wake one of coroutines sleeping in nbd_receive_replies(). If none are sleeping - do nothing. That's another behavior change: we don't have endless recv() in the idle time. It may be considered as a drawback. If so, it may be fixed later. 13. nbd_reply_chunk_iter_receive(): don't care about removed connection_co, just ping in_flight waiters. 14. Don't create connection_co, enable retry in the connection thread (we don't have own reconnect loop anymore) 15. We now need to add a nbd_co_establish_connection_cancel() call in nbd_cancel_in_flight(), to cancel the request that is doing a connection attempt. [*], ok, now we don't cancel reconnect on drain begin. That's correct: reconnect feature leads to possibility of long-running requests (up to reconnect delay). Still, drain begin is not a reason to kill long requests. We should wait for them. This also means, that we can again reproduce a dead-lock, described in 8c517de24a8a1dcbeb54e7e12b5b0fda42a90ace. Why we are OK with it: 1. Now this is not absolutely-dead dead-lock: the vm is unfrozen after reconnect delay. Actually 8c517de24a8a1dc fixed a bug in NBD logic, that was not described in 8c517de24a8a1dc and led to forever dead-lock. The problem was that nobody woke the free_sema queue, but drain_begin can't finish until there is a request in free_sema queue. Now we have a reconnect delay timer that works well. 2. It's not a problem of the NBD driver, but of the ide code, because it does drain_begin under the global mutex; the problem doesn't reproduce when using scsi instead of ide. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com> Reviewed-by: Eric Blake <eblake@redhat.com> [eblake: grammar and comment tweaks] Signed-off-by: Eric Blake <eblake@redhat.com>
2021-09-02 12:38:04 +02:00
nbd_co_establish_connection_cancel(s->conn);
}
static BlockDriver bdrv_nbd = {
.format_name = "nbd",
.protocol_name = "nbd",
.instance_size = sizeof(BDRVNBDState),
.bdrv_parse_filename = nbd_parse_filename,
.bdrv_co_create_opts = bdrv_co_create_opts_simple,
.create_opts = &bdrv_create_opts_simple,
.bdrv_file_open = nbd_open,
.bdrv_reopen_prepare = nbd_client_reopen_prepare,
.bdrv_co_preadv = nbd_client_co_preadv,
.bdrv_co_pwritev = nbd_client_co_pwritev,
.bdrv_co_pwrite_zeroes = nbd_client_co_pwrite_zeroes,
.bdrv_close = nbd_close,
.bdrv_co_flush_to_os = nbd_co_flush,
.bdrv_co_pdiscard = nbd_client_co_pdiscard,
.bdrv_refresh_limits = nbd_refresh_limits,
block: nbd: Fix convert qcow2 compressed to nbd When converting to qcow2 compressed format, the last step is a special zero length compressed write, ending in a call to bdrv_co_truncate(). This call always fails for the nbd driver since it does not implement bdrv_co_truncate(). For block devices, which have the same limits, the call succeeds since the file driver implements bdrv_co_truncate(). If the caller asked to truncate to the same or smaller size with exact=false, the truncate succeeds. Implement the same logic for nbd. Example failing without this change: In one shell start qemu-nbd: $ truncate -s 1g test.tar $ qemu-nbd --socket=/tmp/nbd.sock --persistent --format=raw --offset 1536 test.tar In another shell convert an image to qcow2 compressed via NBD: $ echo "disk data" > disk.raw $ truncate -s 1g disk.raw $ qemu-img convert -f raw -O qcow2 -c disk1.raw nbd+unix:///?socket=/tmp/nbd.sock; echo $? 1 qemu-img failed, but the conversion was successful: $ qemu-img info nbd+unix:///?socket=/tmp/nbd.sock image: nbd+unix://?socket=/tmp/nbd.sock file format: qcow2 virtual size: 1 GiB (1073741824 bytes) ... $ qemu-img check nbd+unix:///?socket=/tmp/nbd.sock No errors were found on the image. 1/16384 = 0.01% allocated, 100.00% fragmented, 100.00% compressed clusters Image end offset: 393216 $ qemu-img compare disk.raw nbd+unix:///?socket=/tmp/nbd.sock Images are identical. Fixes: https://bugzilla.redhat.com/1860627 Signed-off-by: Nir Soffer <nsoffer@redhat.com> Message-Id: <20200727215846.395443-2-nsoffer@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> [eblake: typo fixes] Signed-off-by: Eric Blake <eblake@redhat.com>
2020-07-27 23:58:43 +02:00
.bdrv_co_truncate = nbd_co_truncate,
.bdrv_getlength = nbd_getlength,
.bdrv_refresh_filename = nbd_refresh_filename,
.bdrv_co_block_status = nbd_client_co_block_status,
.bdrv_dirname = nbd_dirname,
.strong_runtime_opts = nbd_strong_runtime_opts,
.bdrv_cancel_in_flight = nbd_cancel_in_flight,
};
static BlockDriver bdrv_nbd_tcp = {
.format_name = "nbd",
.protocol_name = "nbd+tcp",
.instance_size = sizeof(BDRVNBDState),
.bdrv_parse_filename = nbd_parse_filename,
.bdrv_co_create_opts = bdrv_co_create_opts_simple,
.create_opts = &bdrv_create_opts_simple,
.bdrv_file_open = nbd_open,
.bdrv_reopen_prepare = nbd_client_reopen_prepare,
.bdrv_co_preadv = nbd_client_co_preadv,
.bdrv_co_pwritev = nbd_client_co_pwritev,
.bdrv_co_pwrite_zeroes = nbd_client_co_pwrite_zeroes,
.bdrv_close = nbd_close,
.bdrv_co_flush_to_os = nbd_co_flush,
.bdrv_co_pdiscard = nbd_client_co_pdiscard,
.bdrv_refresh_limits = nbd_refresh_limits,
block: nbd: Fix convert qcow2 compressed to nbd When converting to qcow2 compressed format, the last step is a special zero length compressed write, ending in a call to bdrv_co_truncate(). This call always fails for the nbd driver since it does not implement bdrv_co_truncate(). For block devices, which have the same limits, the call succeeds since the file driver implements bdrv_co_truncate(). If the caller asked to truncate to the same or smaller size with exact=false, the truncate succeeds. Implement the same logic for nbd. Example failing without this change: In one shell start qemu-nbd: $ truncate -s 1g test.tar $ qemu-nbd --socket=/tmp/nbd.sock --persistent --format=raw --offset 1536 test.tar In another shell convert an image to qcow2 compressed via NBD: $ echo "disk data" > disk.raw $ truncate -s 1g disk.raw $ qemu-img convert -f raw -O qcow2 -c disk1.raw nbd+unix:///?socket=/tmp/nbd.sock; echo $? 1 qemu-img failed, but the conversion was successful: $ qemu-img info nbd+unix:///?socket=/tmp/nbd.sock image: nbd+unix://?socket=/tmp/nbd.sock file format: qcow2 virtual size: 1 GiB (1073741824 bytes) ... $ qemu-img check nbd+unix:///?socket=/tmp/nbd.sock No errors were found on the image. 1/16384 = 0.01% allocated, 100.00% fragmented, 100.00% compressed clusters Image end offset: 393216 $ qemu-img compare disk.raw nbd+unix:///?socket=/tmp/nbd.sock Images are identical. Fixes: https://bugzilla.redhat.com/1860627 Signed-off-by: Nir Soffer <nsoffer@redhat.com> Message-Id: <20200727215846.395443-2-nsoffer@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> [eblake: typo fixes] Signed-off-by: Eric Blake <eblake@redhat.com>
2020-07-27 23:58:43 +02:00
.bdrv_co_truncate = nbd_co_truncate,
.bdrv_getlength = nbd_getlength,
.bdrv_refresh_filename = nbd_refresh_filename,
.bdrv_co_block_status = nbd_client_co_block_status,
.bdrv_dirname = nbd_dirname,
.strong_runtime_opts = nbd_strong_runtime_opts,
.bdrv_cancel_in_flight = nbd_cancel_in_flight,
};
static BlockDriver bdrv_nbd_unix = {
.format_name = "nbd",
.protocol_name = "nbd+unix",
.instance_size = sizeof(BDRVNBDState),
.bdrv_parse_filename = nbd_parse_filename,
.bdrv_co_create_opts = bdrv_co_create_opts_simple,
.create_opts = &bdrv_create_opts_simple,
.bdrv_file_open = nbd_open,
.bdrv_reopen_prepare = nbd_client_reopen_prepare,
.bdrv_co_preadv = nbd_client_co_preadv,
.bdrv_co_pwritev = nbd_client_co_pwritev,
.bdrv_co_pwrite_zeroes = nbd_client_co_pwrite_zeroes,
.bdrv_close = nbd_close,
.bdrv_co_flush_to_os = nbd_co_flush,
.bdrv_co_pdiscard = nbd_client_co_pdiscard,
.bdrv_refresh_limits = nbd_refresh_limits,
block: nbd: Fix convert qcow2 compressed to nbd When converting to qcow2 compressed format, the last step is a special zero length compressed write, ending in a call to bdrv_co_truncate(). This call always fails for the nbd driver since it does not implement bdrv_co_truncate(). For block devices, which have the same limits, the call succeeds since the file driver implements bdrv_co_truncate(). If the caller asked to truncate to the same or smaller size with exact=false, the truncate succeeds. Implement the same logic for nbd. Example failing without this change: In one shell start qemu-nbd: $ truncate -s 1g test.tar $ qemu-nbd --socket=/tmp/nbd.sock --persistent --format=raw --offset 1536 test.tar In another shell convert an image to qcow2 compressed via NBD: $ echo "disk data" > disk.raw $ truncate -s 1g disk.raw $ qemu-img convert -f raw -O qcow2 -c disk1.raw nbd+unix:///?socket=/tmp/nbd.sock; echo $? 1 qemu-img failed, but the conversion was successful: $ qemu-img info nbd+unix:///?socket=/tmp/nbd.sock image: nbd+unix://?socket=/tmp/nbd.sock file format: qcow2 virtual size: 1 GiB (1073741824 bytes) ... $ qemu-img check nbd+unix:///?socket=/tmp/nbd.sock No errors were found on the image. 1/16384 = 0.01% allocated, 100.00% fragmented, 100.00% compressed clusters Image end offset: 393216 $ qemu-img compare disk.raw nbd+unix:///?socket=/tmp/nbd.sock Images are identical. Fixes: https://bugzilla.redhat.com/1860627 Signed-off-by: Nir Soffer <nsoffer@redhat.com> Message-Id: <20200727215846.395443-2-nsoffer@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> [eblake: typo fixes] Signed-off-by: Eric Blake <eblake@redhat.com>
2020-07-27 23:58:43 +02:00
.bdrv_co_truncate = nbd_co_truncate,
.bdrv_getlength = nbd_getlength,
.bdrv_refresh_filename = nbd_refresh_filename,
.bdrv_co_block_status = nbd_client_co_block_status,
.bdrv_dirname = nbd_dirname,
.strong_runtime_opts = nbd_strong_runtime_opts,
.bdrv_cancel_in_flight = nbd_cancel_in_flight,
};
static void bdrv_nbd_init(void)
{
bdrv_register(&bdrv_nbd);
bdrv_register(&bdrv_nbd_tcp);
bdrv_register(&bdrv_nbd_unix);
}
block_init(bdrv_nbd_init);