qemu-e2k/async.c

418 lines
10 KiB
C
Raw Normal View History

/*
* QEMU System Emulator
*
* Copyright (c) 2003-2008 Fabrice Bellard
*
* 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"
2016-03-14 09:01:28 +01:00
#include "qapi/error.h"
#include "qemu-common.h"
#include "block/aio.h"
#include "block/thread-pool.h"
#include "qemu/main-loop.h"
#include "qemu/atomic.h"
#include "block/raw-aio.h"
/***********************************************************/
/* bottom halves (can be seen as timers which expire ASAP) */
struct QEMUBH {
AioContext *ctx;
QEMUBHFunc *cb;
void *opaque;
QEMUBH *next;
bool scheduled;
bool idle;
bool deleted;
};
void aio_bh_schedule_oneshot(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
{
QEMUBH *bh;
bh = g_new(QEMUBH, 1);
*bh = (QEMUBH){
.ctx = ctx,
.cb = cb,
.opaque = opaque,
};
qemu_mutex_lock(&ctx->bh_lock);
bh->next = ctx->first_bh;
bh->scheduled = 1;
bh->deleted = 1;
/* Make sure that the members are ready before putting bh into list */
smp_wmb();
ctx->first_bh = bh;
qemu_mutex_unlock(&ctx->bh_lock);
aio_notify(ctx);
}
QEMUBH *aio_bh_new(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
{
QEMUBH *bh;
bh = g_new(QEMUBH, 1);
*bh = (QEMUBH){
.ctx = ctx,
.cb = cb,
.opaque = opaque,
};
qemu_mutex_lock(&ctx->bh_lock);
bh->next = ctx->first_bh;
/* Make sure that the members are ready before putting bh into list */
smp_wmb();
ctx->first_bh = bh;
qemu_mutex_unlock(&ctx->bh_lock);
return bh;
}
void aio_bh_call(QEMUBH *bh)
{
bh->cb(bh->opaque);
}
/* Multiple occurrences of aio_bh_poll cannot be called concurrently */
int aio_bh_poll(AioContext *ctx)
{
QEMUBH *bh, **bhp, *next;
int ret;
ctx->walking_bh++;
ret = 0;
for (bh = ctx->first_bh; bh; bh = next) {
/* Make sure that fetching bh happens before accessing its members */
smp_read_barrier_depends();
next = bh->next;
aio: strengthen memory barriers for bottom half scheduling There are two problems with memory barriers in async.c. The fix is to use atomic_xchg in order to achieve sequential consistency between the scheduling of a bottom half and the corresponding execution. First, if bh->scheduled is already 1 in qemu_bh_schedule, QEMU does not execute a memory barrier to order any writes needed by the callback before the read of bh->scheduled. If the other side sees req->state as THREAD_ACTIVE, the callback is not invoked and you get deadlock. Second, the memory barrier in aio_bh_poll is too weak. Without this patch, it is possible that bh->scheduled = 0 is not "published" until after the callback has returned. Another thread wants to schedule the bottom half, but it sees bh->scheduled = 1 and does nothing. This causes a lost wakeup. The memory barrier should have been changed to smp_mb() in commit 924fe12 (aio: fix qemu_bh_schedule() bh->ctx race condition, 2014-06-03) together with qemu_bh_schedule()'s. Guess who reviewed that patch? Both of these involve a store and a load, so they are reproducible on x86_64 as well. It is however much easier on aarch64, where the libguestfs test suite triggers the bug fairly easily. Even there the failure can go away or appear depending on compiler optimization level, tracing options, or even kernel debugging options. Paul Leveille however reported how to trigger the problem within 15 minutes on x86_64 as well. His (untested) recipe, reproduced here for reference, is the following: 1) Qcow2 (or 3) is critical – raw files alone seem to avoid the problem. 2) Use “cache=directsync” rather than the default of “cache=none” to make it happen easier. 3) Use a server with a write-back RAID controller to allow for rapid IO rates. 4) Run a random-access load that (mostly) writes chunks to various files on the virtual block device. a. I use ‘diskload.exe c:25’, a Microsoft HCT load generator, on Windows VMs. b. Iometer can probably be configured to generate a similar load. 5) Run multiple VMs in parallel, against the same storage device, to shake the failure out sooner. 6) IvyBridge and Haswell processors for certain; not sure about others. A similar patch survived over 12 hours of testing, where an unpatched QEMU would fail within 15 minutes. This bug is, most likely, also the cause of failures in the libguestfs testsuite on AArch64. Thanks to Laszlo Ersek for initially reporting this bug, to Stefan Hajnoczi for suggesting closer examination of qemu_bh_schedule, and to Paul for providing test input and a prototype patch. Reported-by: Laszlo Ersek <lersek@redhat.com> Reported-by: Paul Leveille <Paul.Leveille@stratus.com> Reported-by: John Snow <jsnow@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Message-id: 1428419779-26062-1-git-send-email-pbonzini@redhat.com Suggested-by: Paul Leveille <Paul.Leveille@stratus.com> Suggested-by: Stefan Hajnoczi <stefanha@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2015-04-07 17:16:19 +02:00
/* The atomic_xchg is paired with the one in qemu_bh_schedule. The
* implicit memory barrier ensures that the callback sees all writes
* done by the scheduling thread. It also ensures that the scheduling
* thread sees the zero before bh->cb has run, and thus will call
* aio_notify again if necessary.
*/
if (atomic_xchg(&bh->scheduled, 0)) {
/* Idle BHs don't count as progress */
if (!bh->idle) {
ret = 1;
}
bh->idle = 0;
aio_bh_call(bh);
}
}
ctx->walking_bh--;
/* remove deleted bhs */
if (!ctx->walking_bh) {
qemu_mutex_lock(&ctx->bh_lock);
bhp = &ctx->first_bh;
while (*bhp) {
bh = *bhp;
if (bh->deleted && !bh->scheduled) {
*bhp = bh->next;
g_free(bh);
} else {
bhp = &bh->next;
}
}
qemu_mutex_unlock(&ctx->bh_lock);
}
return ret;
}
void qemu_bh_schedule_idle(QEMUBH *bh)
{
bh->idle = 1;
/* Make sure that idle & any writes needed by the callback are done
* before the locations are read in the aio_bh_poll.
*/
aio: strengthen memory barriers for bottom half scheduling There are two problems with memory barriers in async.c. The fix is to use atomic_xchg in order to achieve sequential consistency between the scheduling of a bottom half and the corresponding execution. First, if bh->scheduled is already 1 in qemu_bh_schedule, QEMU does not execute a memory barrier to order any writes needed by the callback before the read of bh->scheduled. If the other side sees req->state as THREAD_ACTIVE, the callback is not invoked and you get deadlock. Second, the memory barrier in aio_bh_poll is too weak. Without this patch, it is possible that bh->scheduled = 0 is not "published" until after the callback has returned. Another thread wants to schedule the bottom half, but it sees bh->scheduled = 1 and does nothing. This causes a lost wakeup. The memory barrier should have been changed to smp_mb() in commit 924fe12 (aio: fix qemu_bh_schedule() bh->ctx race condition, 2014-06-03) together with qemu_bh_schedule()'s. Guess who reviewed that patch? Both of these involve a store and a load, so they are reproducible on x86_64 as well. It is however much easier on aarch64, where the libguestfs test suite triggers the bug fairly easily. Even there the failure can go away or appear depending on compiler optimization level, tracing options, or even kernel debugging options. Paul Leveille however reported how to trigger the problem within 15 minutes on x86_64 as well. His (untested) recipe, reproduced here for reference, is the following: 1) Qcow2 (or 3) is critical – raw files alone seem to avoid the problem. 2) Use “cache=directsync” rather than the default of “cache=none” to make it happen easier. 3) Use a server with a write-back RAID controller to allow for rapid IO rates. 4) Run a random-access load that (mostly) writes chunks to various files on the virtual block device. a. I use ‘diskload.exe c:25’, a Microsoft HCT load generator, on Windows VMs. b. Iometer can probably be configured to generate a similar load. 5) Run multiple VMs in parallel, against the same storage device, to shake the failure out sooner. 6) IvyBridge and Haswell processors for certain; not sure about others. A similar patch survived over 12 hours of testing, where an unpatched QEMU would fail within 15 minutes. This bug is, most likely, also the cause of failures in the libguestfs testsuite on AArch64. Thanks to Laszlo Ersek for initially reporting this bug, to Stefan Hajnoczi for suggesting closer examination of qemu_bh_schedule, and to Paul for providing test input and a prototype patch. Reported-by: Laszlo Ersek <lersek@redhat.com> Reported-by: Paul Leveille <Paul.Leveille@stratus.com> Reported-by: John Snow <jsnow@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Message-id: 1428419779-26062-1-git-send-email-pbonzini@redhat.com Suggested-by: Paul Leveille <Paul.Leveille@stratus.com> Suggested-by: Stefan Hajnoczi <stefanha@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2015-04-07 17:16:19 +02:00
atomic_mb_set(&bh->scheduled, 1);
}
void qemu_bh_schedule(QEMUBH *bh)
{
AioContext *ctx;
ctx = bh->ctx;
bh->idle = 0;
aio: strengthen memory barriers for bottom half scheduling There are two problems with memory barriers in async.c. The fix is to use atomic_xchg in order to achieve sequential consistency between the scheduling of a bottom half and the corresponding execution. First, if bh->scheduled is already 1 in qemu_bh_schedule, QEMU does not execute a memory barrier to order any writes needed by the callback before the read of bh->scheduled. If the other side sees req->state as THREAD_ACTIVE, the callback is not invoked and you get deadlock. Second, the memory barrier in aio_bh_poll is too weak. Without this patch, it is possible that bh->scheduled = 0 is not "published" until after the callback has returned. Another thread wants to schedule the bottom half, but it sees bh->scheduled = 1 and does nothing. This causes a lost wakeup. The memory barrier should have been changed to smp_mb() in commit 924fe12 (aio: fix qemu_bh_schedule() bh->ctx race condition, 2014-06-03) together with qemu_bh_schedule()'s. Guess who reviewed that patch? Both of these involve a store and a load, so they are reproducible on x86_64 as well. It is however much easier on aarch64, where the libguestfs test suite triggers the bug fairly easily. Even there the failure can go away or appear depending on compiler optimization level, tracing options, or even kernel debugging options. Paul Leveille however reported how to trigger the problem within 15 minutes on x86_64 as well. His (untested) recipe, reproduced here for reference, is the following: 1) Qcow2 (or 3) is critical – raw files alone seem to avoid the problem. 2) Use “cache=directsync” rather than the default of “cache=none” to make it happen easier. 3) Use a server with a write-back RAID controller to allow for rapid IO rates. 4) Run a random-access load that (mostly) writes chunks to various files on the virtual block device. a. I use ‘diskload.exe c:25’, a Microsoft HCT load generator, on Windows VMs. b. Iometer can probably be configured to generate a similar load. 5) Run multiple VMs in parallel, against the same storage device, to shake the failure out sooner. 6) IvyBridge and Haswell processors for certain; not sure about others. A similar patch survived over 12 hours of testing, where an unpatched QEMU would fail within 15 minutes. This bug is, most likely, also the cause of failures in the libguestfs testsuite on AArch64. Thanks to Laszlo Ersek for initially reporting this bug, to Stefan Hajnoczi for suggesting closer examination of qemu_bh_schedule, and to Paul for providing test input and a prototype patch. Reported-by: Laszlo Ersek <lersek@redhat.com> Reported-by: Paul Leveille <Paul.Leveille@stratus.com> Reported-by: John Snow <jsnow@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Message-id: 1428419779-26062-1-git-send-email-pbonzini@redhat.com Suggested-by: Paul Leveille <Paul.Leveille@stratus.com> Suggested-by: Stefan Hajnoczi <stefanha@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2015-04-07 17:16:19 +02:00
/* The memory barrier implicit in atomic_xchg makes sure that:
* 1. idle & any writes needed by the callback are done before the
* locations are read in the aio_bh_poll.
* 2. ctx is loaded before scheduled is set and the callback has a chance
* to execute.
*/
aio: strengthen memory barriers for bottom half scheduling There are two problems with memory barriers in async.c. The fix is to use atomic_xchg in order to achieve sequential consistency between the scheduling of a bottom half and the corresponding execution. First, if bh->scheduled is already 1 in qemu_bh_schedule, QEMU does not execute a memory barrier to order any writes needed by the callback before the read of bh->scheduled. If the other side sees req->state as THREAD_ACTIVE, the callback is not invoked and you get deadlock. Second, the memory barrier in aio_bh_poll is too weak. Without this patch, it is possible that bh->scheduled = 0 is not "published" until after the callback has returned. Another thread wants to schedule the bottom half, but it sees bh->scheduled = 1 and does nothing. This causes a lost wakeup. The memory barrier should have been changed to smp_mb() in commit 924fe12 (aio: fix qemu_bh_schedule() bh->ctx race condition, 2014-06-03) together with qemu_bh_schedule()'s. Guess who reviewed that patch? Both of these involve a store and a load, so they are reproducible on x86_64 as well. It is however much easier on aarch64, where the libguestfs test suite triggers the bug fairly easily. Even there the failure can go away or appear depending on compiler optimization level, tracing options, or even kernel debugging options. Paul Leveille however reported how to trigger the problem within 15 minutes on x86_64 as well. His (untested) recipe, reproduced here for reference, is the following: 1) Qcow2 (or 3) is critical – raw files alone seem to avoid the problem. 2) Use “cache=directsync” rather than the default of “cache=none” to make it happen easier. 3) Use a server with a write-back RAID controller to allow for rapid IO rates. 4) Run a random-access load that (mostly) writes chunks to various files on the virtual block device. a. I use ‘diskload.exe c:25’, a Microsoft HCT load generator, on Windows VMs. b. Iometer can probably be configured to generate a similar load. 5) Run multiple VMs in parallel, against the same storage device, to shake the failure out sooner. 6) IvyBridge and Haswell processors for certain; not sure about others. A similar patch survived over 12 hours of testing, where an unpatched QEMU would fail within 15 minutes. This bug is, most likely, also the cause of failures in the libguestfs testsuite on AArch64. Thanks to Laszlo Ersek for initially reporting this bug, to Stefan Hajnoczi for suggesting closer examination of qemu_bh_schedule, and to Paul for providing test input and a prototype patch. Reported-by: Laszlo Ersek <lersek@redhat.com> Reported-by: Paul Leveille <Paul.Leveille@stratus.com> Reported-by: John Snow <jsnow@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Message-id: 1428419779-26062-1-git-send-email-pbonzini@redhat.com Suggested-by: Paul Leveille <Paul.Leveille@stratus.com> Suggested-by: Stefan Hajnoczi <stefanha@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2015-04-07 17:16:19 +02:00
if (atomic_xchg(&bh->scheduled, 1) == 0) {
aio_notify(ctx);
}
}
/* This func is async.
*/
void qemu_bh_cancel(QEMUBH *bh)
{
bh->scheduled = 0;
}
/* This func is async.The bottom half will do the delete action at the finial
* end.
*/
void qemu_bh_delete(QEMUBH *bh)
{
bh->scheduled = 0;
bh->deleted = 1;
}
int64_t
aio_compute_timeout(AioContext *ctx)
{
int64_t deadline;
int timeout = -1;
QEMUBH *bh;
for (bh = ctx->first_bh; bh; bh = bh->next) {
if (bh->scheduled) {
if (bh->idle) {
/* idle bottom halves will be polled at least
* every 10ms */
timeout = 10000000;
} else {
/* non-idle bottom halves will be executed
* immediately */
return 0;
}
}
}
deadline = timerlistgroup_deadline_ns(&ctx->tlg);
if (deadline == 0) {
return 0;
} else {
return qemu_soonest_timeout(timeout, deadline);
}
}
static gboolean
aio_ctx_prepare(GSource *source, gint *timeout)
{
AioContext *ctx = (AioContext *) source;
AioContext: fix broken ctx->dispatching optimization This patch rewrites the ctx->dispatching optimization, which was the cause of some mysterious hangs that could be reproduced on aarch64 KVM only. The hangs were indirectly caused by aio_poll() and in particular by flash memory updates's call to blk_write(), which invokes aio_poll(). Fun stuff: they had an extremely short race window, so much that adding all kind of tracing to either the kernel or QEMU made it go away (a single printf made it half as reproducible). On the plus side, the failure mode (a hang until the next keypress) made it very easy to examine the state of the process with a debugger. And there was a very nice reproducer from Laszlo, which failed pretty often (more than half of the time) on any version of QEMU with a non-debug kernel; it also failed fast, while still in the firmware. So, it could have been worse. For some unknown reason they happened only with virtio-scsi, but that's not important. It's more interesting that they disappeared with io=native, making thread-pool.c a likely suspect for where the bug arose. thread-pool.c is also one of the few places which use bottom halves across threads, by the way. I hope that no other similar bugs exist, but just in case :) I am going to describe how the successful debugging went... Since the likely culprit was the ctx->dispatching optimization, which mostly affects bottom halves, the first observation was that there are two qemu_bh_schedule() invocations in the thread pool: the one in the aio worker and the one in thread_pool_completion_bh. The latter always causes the optimization to trigger, the former may or may not. In order to restrict the possibilities, I introduced new functions qemu_bh_schedule_slow() and qemu_bh_schedule_fast(): /* qemu_bh_schedule_slow: */ ctx = bh->ctx; bh->idle = 0; if (atomic_xchg(&bh->scheduled, 1) == 0) { event_notifier_set(&ctx->notifier); } /* qemu_bh_schedule_fast: */ ctx = bh->ctx; bh->idle = 0; assert(ctx->dispatching); atomic_xchg(&bh->scheduled, 1); Notice how the atomic_xchg is still in qemu_bh_schedule_slow(). This was already debated a few months ago, so I assumed it to be correct. In retrospect this was a very good idea, as you'll see later. Changing thread_pool_completion_bh() to qemu_bh_schedule_fast() didn't trigger the assertion (as expected). Changing the worker's invocation to qemu_bh_schedule_slow() didn't hide the bug (another assumption which luckily held). This already limited heavily the amount of interaction between the threads, hinting that the problematic events must have triggered around thread_pool_completion_bh(). As mentioned early, invoking a debugger to examine the state of a hung process was pretty easy; the iothread was always waiting on a poll(..., -1) system call. Infinite timeouts are much rarer on x86, and this could be the reason why the bug was never observed there. With the buggy sequence more or less resolved to an interaction between thread_pool_completion_bh() and poll(..., -1), my "tracing" strategy was to just add a few qemu_clock_get_ns(QEMU_CLOCK_REALTIME) calls, hoping that the ordering of aio_ctx_prepare(), aio_ctx_dispatch, poll() and qemu_bh_schedule_fast() would provide some hint. The output was: (gdb) p last_prepare $3 = 103885451 (gdb) p last_dispatch $4 = 103876492 (gdb) p last_poll $5 = 115909333 (gdb) p last_schedule $6 = 115925212 Notice how the last call to qemu_poll_ns() came after aio_ctx_dispatch(). This makes little sense unless there is an aio_poll() call involved, and indeed with a slightly different instrumentation you can see that there is one: (gdb) p last_prepare $3 = 107569679 (gdb) p last_dispatch $4 = 107561600 (gdb) p last_aio_poll $5 = 110671400 (gdb) p last_schedule $6 = 110698917 So the scenario becomes clearer: iothread VCPU thread -------------------------------------------------------------------------- aio_ctx_prepare aio_ctx_check qemu_poll_ns(timeout=-1) aio_poll aio_dispatch thread_pool_completion_bh qemu_bh_schedule() At this point bh->scheduled = 1 and the iothread has not been woken up. The solution must be close, but this alone should not be a problem, because the bottom half is only rescheduled to account for rare situations (see commit 3c80ca1, thread-pool: avoid deadlock in nested aio_poll() calls, 2014-07-15). Introducing a third thread---a thread pool worker thread, which also does qemu_bh_schedule()---does bring out the problematic case. The third thread must be awakened *after* the callback is complete and thread_pool_completion_bh has redone the whole loop, explaining the short race window. And then this is what happens: thread pool worker -------------------------------------------------------------------------- <I/O completes> qemu_bh_schedule() Tada, bh->scheduled is already 1, so qemu_bh_schedule() does nothing and the iothread is never woken up. This is where the bh->scheduled optimization comes into play---it is correct, but removing it would have masked the bug. So, what is the bug? Well, the question asked by the ctx->dispatching optimization ("is any active aio_poll dispatching?") was wrong. The right question to ask instead is "is any active aio_poll *not* dispatching", i.e. in the prepare or poll phases? In that case, the aio_poll is sleeping or might go to sleep anytime soon, and the EventNotifier must be invoked to wake it up. In any other case (including if there is *no* active aio_poll at all!) we can just wait for the next prepare phase to pick up the event (e.g. a bottom half); the prepare phase will avoid the blocking and service the bottom half. Expressing the invariant with a logic formula, the broken one looked like: !(exists(thread): in_dispatching(thread)) => !optimize or equivalently: !(exists(thread): in_aio_poll(thread) && in_dispatching(thread)) => !optimize In the correct one, the negation is in a slightly different place: (exists(thread): in_aio_poll(thread) && !in_dispatching(thread)) => !optimize or equivalently: (exists(thread): in_prepare_or_poll(thread)) => !optimize Even if the difference boils down to moving an exclamation mark :) the implementation is quite different. However, I think the new one is simpler to understand. In the old implementation, the "exists" was implemented with a boolean value. This didn't really support well the case of multiple concurrent event loops, but I thought that this was okay: aio_poll holds the AioContext lock so there cannot be concurrent aio_poll invocations, and I was just considering nested event loops. However, aio_poll _could_ indeed be concurrent with the GSource. This is why I came up with the wrong invariant. In the new implementation, "exists" is computed simply by counting how many threads are in the prepare or poll phases. There are some interesting points to consider, but the gist of the idea remains: 1) AioContext can be used through GSource as well; as mentioned in the patch, bit 0 of the counter is reserved for the GSource. 2) the counter need not be updated for a non-blocking aio_poll, because it won't sleep forever anyway. This is just a matter of checking the "blocking" variable. This requires some changes to the win32 implementation, but is otherwise not too complicated. 3) as mentioned above, the new implementation will not call aio_notify when there is *no* active aio_poll at all. The tests have to be adjusted for this change. The calls to aio_notify in async.c are fine; they only want to kick aio_poll out of a blocking wait, but need not do anything if aio_poll is not running. 4) nested aio_poll: these just work with the new implementation; when a nested event loop is invoked, the outer event loop is never in the prepare or poll phases. The outer event loop thus has already decremented the counter. Reported-by: Richard W. M. Jones <rjones@redhat.com> Reported-by: Laszlo Ersek <lersek@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Reviewed-by: Fam Zheng <famz@redhat.com> Tested-by: Richard W.M. Jones <rjones@redhat.com> Message-id: 1437487673-23740-5-git-send-email-pbonzini@redhat.com Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2015-07-21 16:07:51 +02:00
atomic_or(&ctx->notify_me, 1);
/* We assume there is no timeout already supplied */
*timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx));
if (aio_prepare(ctx)) {
*timeout = 0;
}
return *timeout == 0;
}
static gboolean
aio_ctx_check(GSource *source)
{
AioContext *ctx = (AioContext *) source;
QEMUBH *bh;
AioContext: fix broken ctx->dispatching optimization This patch rewrites the ctx->dispatching optimization, which was the cause of some mysterious hangs that could be reproduced on aarch64 KVM only. The hangs were indirectly caused by aio_poll() and in particular by flash memory updates's call to blk_write(), which invokes aio_poll(). Fun stuff: they had an extremely short race window, so much that adding all kind of tracing to either the kernel or QEMU made it go away (a single printf made it half as reproducible). On the plus side, the failure mode (a hang until the next keypress) made it very easy to examine the state of the process with a debugger. And there was a very nice reproducer from Laszlo, which failed pretty often (more than half of the time) on any version of QEMU with a non-debug kernel; it also failed fast, while still in the firmware. So, it could have been worse. For some unknown reason they happened only with virtio-scsi, but that's not important. It's more interesting that they disappeared with io=native, making thread-pool.c a likely suspect for where the bug arose. thread-pool.c is also one of the few places which use bottom halves across threads, by the way. I hope that no other similar bugs exist, but just in case :) I am going to describe how the successful debugging went... Since the likely culprit was the ctx->dispatching optimization, which mostly affects bottom halves, the first observation was that there are two qemu_bh_schedule() invocations in the thread pool: the one in the aio worker and the one in thread_pool_completion_bh. The latter always causes the optimization to trigger, the former may or may not. In order to restrict the possibilities, I introduced new functions qemu_bh_schedule_slow() and qemu_bh_schedule_fast(): /* qemu_bh_schedule_slow: */ ctx = bh->ctx; bh->idle = 0; if (atomic_xchg(&bh->scheduled, 1) == 0) { event_notifier_set(&ctx->notifier); } /* qemu_bh_schedule_fast: */ ctx = bh->ctx; bh->idle = 0; assert(ctx->dispatching); atomic_xchg(&bh->scheduled, 1); Notice how the atomic_xchg is still in qemu_bh_schedule_slow(). This was already debated a few months ago, so I assumed it to be correct. In retrospect this was a very good idea, as you'll see later. Changing thread_pool_completion_bh() to qemu_bh_schedule_fast() didn't trigger the assertion (as expected). Changing the worker's invocation to qemu_bh_schedule_slow() didn't hide the bug (another assumption which luckily held). This already limited heavily the amount of interaction between the threads, hinting that the problematic events must have triggered around thread_pool_completion_bh(). As mentioned early, invoking a debugger to examine the state of a hung process was pretty easy; the iothread was always waiting on a poll(..., -1) system call. Infinite timeouts are much rarer on x86, and this could be the reason why the bug was never observed there. With the buggy sequence more or less resolved to an interaction between thread_pool_completion_bh() and poll(..., -1), my "tracing" strategy was to just add a few qemu_clock_get_ns(QEMU_CLOCK_REALTIME) calls, hoping that the ordering of aio_ctx_prepare(), aio_ctx_dispatch, poll() and qemu_bh_schedule_fast() would provide some hint. The output was: (gdb) p last_prepare $3 = 103885451 (gdb) p last_dispatch $4 = 103876492 (gdb) p last_poll $5 = 115909333 (gdb) p last_schedule $6 = 115925212 Notice how the last call to qemu_poll_ns() came after aio_ctx_dispatch(). This makes little sense unless there is an aio_poll() call involved, and indeed with a slightly different instrumentation you can see that there is one: (gdb) p last_prepare $3 = 107569679 (gdb) p last_dispatch $4 = 107561600 (gdb) p last_aio_poll $5 = 110671400 (gdb) p last_schedule $6 = 110698917 So the scenario becomes clearer: iothread VCPU thread -------------------------------------------------------------------------- aio_ctx_prepare aio_ctx_check qemu_poll_ns(timeout=-1) aio_poll aio_dispatch thread_pool_completion_bh qemu_bh_schedule() At this point bh->scheduled = 1 and the iothread has not been woken up. The solution must be close, but this alone should not be a problem, because the bottom half is only rescheduled to account for rare situations (see commit 3c80ca1, thread-pool: avoid deadlock in nested aio_poll() calls, 2014-07-15). Introducing a third thread---a thread pool worker thread, which also does qemu_bh_schedule()---does bring out the problematic case. The third thread must be awakened *after* the callback is complete and thread_pool_completion_bh has redone the whole loop, explaining the short race window. And then this is what happens: thread pool worker -------------------------------------------------------------------------- <I/O completes> qemu_bh_schedule() Tada, bh->scheduled is already 1, so qemu_bh_schedule() does nothing and the iothread is never woken up. This is where the bh->scheduled optimization comes into play---it is correct, but removing it would have masked the bug. So, what is the bug? Well, the question asked by the ctx->dispatching optimization ("is any active aio_poll dispatching?") was wrong. The right question to ask instead is "is any active aio_poll *not* dispatching", i.e. in the prepare or poll phases? In that case, the aio_poll is sleeping or might go to sleep anytime soon, and the EventNotifier must be invoked to wake it up. In any other case (including if there is *no* active aio_poll at all!) we can just wait for the next prepare phase to pick up the event (e.g. a bottom half); the prepare phase will avoid the blocking and service the bottom half. Expressing the invariant with a logic formula, the broken one looked like: !(exists(thread): in_dispatching(thread)) => !optimize or equivalently: !(exists(thread): in_aio_poll(thread) && in_dispatching(thread)) => !optimize In the correct one, the negation is in a slightly different place: (exists(thread): in_aio_poll(thread) && !in_dispatching(thread)) => !optimize or equivalently: (exists(thread): in_prepare_or_poll(thread)) => !optimize Even if the difference boils down to moving an exclamation mark :) the implementation is quite different. However, I think the new one is simpler to understand. In the old implementation, the "exists" was implemented with a boolean value. This didn't really support well the case of multiple concurrent event loops, but I thought that this was okay: aio_poll holds the AioContext lock so there cannot be concurrent aio_poll invocations, and I was just considering nested event loops. However, aio_poll _could_ indeed be concurrent with the GSource. This is why I came up with the wrong invariant. In the new implementation, "exists" is computed simply by counting how many threads are in the prepare or poll phases. There are some interesting points to consider, but the gist of the idea remains: 1) AioContext can be used through GSource as well; as mentioned in the patch, bit 0 of the counter is reserved for the GSource. 2) the counter need not be updated for a non-blocking aio_poll, because it won't sleep forever anyway. This is just a matter of checking the "blocking" variable. This requires some changes to the win32 implementation, but is otherwise not too complicated. 3) as mentioned above, the new implementation will not call aio_notify when there is *no* active aio_poll at all. The tests have to be adjusted for this change. The calls to aio_notify in async.c are fine; they only want to kick aio_poll out of a blocking wait, but need not do anything if aio_poll is not running. 4) nested aio_poll: these just work with the new implementation; when a nested event loop is invoked, the outer event loop is never in the prepare or poll phases. The outer event loop thus has already decremented the counter. Reported-by: Richard W. M. Jones <rjones@redhat.com> Reported-by: Laszlo Ersek <lersek@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Reviewed-by: Fam Zheng <famz@redhat.com> Tested-by: Richard W.M. Jones <rjones@redhat.com> Message-id: 1437487673-23740-5-git-send-email-pbonzini@redhat.com Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2015-07-21 16:07:51 +02:00
atomic_and(&ctx->notify_me, ~1);
aio_notify_accept(ctx);
AioContext: fix broken placement of event_notifier_test_and_clear event_notifier_test_and_clear must be called before processing events. Otherwise, an aio_poll could "eat" the notification before the main I/O thread invokes ppoll(). The main I/O thread then never wakes up. This is an example of what could happen: i/o thread vcpu thread worker thread --------------------------------------------------------------------- lock_iothread notify_me = 1 ... unlock_iothread bh->scheduled = 1 event_notifier_set lock_iothread notify_me = 3 ppoll notify_me = 1 aio_dispatch aio_bh_poll thread_pool_completion_bh bh->scheduled = 1 event_notifier_set node->io_read(node->opaque) event_notifier_test_and_clear ppoll *** hang *** "Tracing" with qemu_clock_get_ns shows pretty much the same behavior as in the previous bug, so there are no new tricks here---just stare more at the code until it is apparent. One could also use a formal model, of course. The included one shows this with three processes: notifier corresponds to a QEMU thread pool worker, temporary_waiter to a VCPU thread that invokes aio_poll(), waiter to the main I/O thread. I would be happy to say that the formal model found the bug for me, but actually I wrote it after the fact. This patch is a bit of a big hammer. The next one optimizes it, with help (this time for real rather than a posteriori :)) from another, similar formal model. Reported-by: Richard W. M. Jones <rjones@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Reviewed-by: Fam Zheng <famz@redhat.com> Tested-by: Richard W.M. Jones <rjones@redhat.com> Message-id: 1437487673-23740-6-git-send-email-pbonzini@redhat.com Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2015-07-21 16:07:52 +02:00
for (bh = ctx->first_bh; bh; bh = bh->next) {
if (bh->scheduled) {
return true;
}
}
return aio_pending(ctx) || (timerlistgroup_deadline_ns(&ctx->tlg) == 0);
}
static gboolean
aio_ctx_dispatch(GSource *source,
GSourceFunc callback,
gpointer user_data)
{
AioContext *ctx = (AioContext *) source;
assert(callback == NULL);
aio_dispatch(ctx, true);
return true;
}
static void
aio_ctx_finalize(GSource *source)
{
AioContext *ctx = (AioContext *) source;
thread_pool_free(ctx->thread_pool);
#ifdef CONFIG_LINUX_AIO
if (ctx->linux_aio) {
laio_detach_aio_context(ctx->linux_aio, ctx);
laio_cleanup(ctx->linux_aio);
ctx->linux_aio = NULL;
}
#endif
qemu_mutex_lock(&ctx->bh_lock);
while (ctx->first_bh) {
QEMUBH *next = ctx->first_bh->next;
/* qemu_bh_delete() must have been called on BHs in this AioContext */
assert(ctx->first_bh->deleted);
g_free(ctx->first_bh);
ctx->first_bh = next;
}
qemu_mutex_unlock(&ctx->bh_lock);
aio_set_event_notifier(ctx, &ctx->notifier, false, NULL, NULL);
event_notifier_cleanup(&ctx->notifier);
qemu_rec_mutex_destroy(&ctx->lock);
qemu_mutex_destroy(&ctx->bh_lock);
timerlistgroup_deinit(&ctx->tlg);
}
static GSourceFuncs aio_source_funcs = {
aio_ctx_prepare,
aio_ctx_check,
aio_ctx_dispatch,
aio_ctx_finalize
};
GSource *aio_get_g_source(AioContext *ctx)
{
g_source_ref(&ctx->source);
return &ctx->source;
}
ThreadPool *aio_get_thread_pool(AioContext *ctx)
{
if (!ctx->thread_pool) {
ctx->thread_pool = thread_pool_new(ctx);
}
return ctx->thread_pool;
}
#ifdef CONFIG_LINUX_AIO
LinuxAioState *aio_get_linux_aio(AioContext *ctx)
{
if (!ctx->linux_aio) {
ctx->linux_aio = laio_init();
laio_attach_aio_context(ctx->linux_aio, ctx);
}
return ctx->linux_aio;
}
#endif
void aio_notify(AioContext *ctx)
{
AioContext: fix broken ctx->dispatching optimization This patch rewrites the ctx->dispatching optimization, which was the cause of some mysterious hangs that could be reproduced on aarch64 KVM only. The hangs were indirectly caused by aio_poll() and in particular by flash memory updates's call to blk_write(), which invokes aio_poll(). Fun stuff: they had an extremely short race window, so much that adding all kind of tracing to either the kernel or QEMU made it go away (a single printf made it half as reproducible). On the plus side, the failure mode (a hang until the next keypress) made it very easy to examine the state of the process with a debugger. And there was a very nice reproducer from Laszlo, which failed pretty often (more than half of the time) on any version of QEMU with a non-debug kernel; it also failed fast, while still in the firmware. So, it could have been worse. For some unknown reason they happened only with virtio-scsi, but that's not important. It's more interesting that they disappeared with io=native, making thread-pool.c a likely suspect for where the bug arose. thread-pool.c is also one of the few places which use bottom halves across threads, by the way. I hope that no other similar bugs exist, but just in case :) I am going to describe how the successful debugging went... Since the likely culprit was the ctx->dispatching optimization, which mostly affects bottom halves, the first observation was that there are two qemu_bh_schedule() invocations in the thread pool: the one in the aio worker and the one in thread_pool_completion_bh. The latter always causes the optimization to trigger, the former may or may not. In order to restrict the possibilities, I introduced new functions qemu_bh_schedule_slow() and qemu_bh_schedule_fast(): /* qemu_bh_schedule_slow: */ ctx = bh->ctx; bh->idle = 0; if (atomic_xchg(&bh->scheduled, 1) == 0) { event_notifier_set(&ctx->notifier); } /* qemu_bh_schedule_fast: */ ctx = bh->ctx; bh->idle = 0; assert(ctx->dispatching); atomic_xchg(&bh->scheduled, 1); Notice how the atomic_xchg is still in qemu_bh_schedule_slow(). This was already debated a few months ago, so I assumed it to be correct. In retrospect this was a very good idea, as you'll see later. Changing thread_pool_completion_bh() to qemu_bh_schedule_fast() didn't trigger the assertion (as expected). Changing the worker's invocation to qemu_bh_schedule_slow() didn't hide the bug (another assumption which luckily held). This already limited heavily the amount of interaction between the threads, hinting that the problematic events must have triggered around thread_pool_completion_bh(). As mentioned early, invoking a debugger to examine the state of a hung process was pretty easy; the iothread was always waiting on a poll(..., -1) system call. Infinite timeouts are much rarer on x86, and this could be the reason why the bug was never observed there. With the buggy sequence more or less resolved to an interaction between thread_pool_completion_bh() and poll(..., -1), my "tracing" strategy was to just add a few qemu_clock_get_ns(QEMU_CLOCK_REALTIME) calls, hoping that the ordering of aio_ctx_prepare(), aio_ctx_dispatch, poll() and qemu_bh_schedule_fast() would provide some hint. The output was: (gdb) p last_prepare $3 = 103885451 (gdb) p last_dispatch $4 = 103876492 (gdb) p last_poll $5 = 115909333 (gdb) p last_schedule $6 = 115925212 Notice how the last call to qemu_poll_ns() came after aio_ctx_dispatch(). This makes little sense unless there is an aio_poll() call involved, and indeed with a slightly different instrumentation you can see that there is one: (gdb) p last_prepare $3 = 107569679 (gdb) p last_dispatch $4 = 107561600 (gdb) p last_aio_poll $5 = 110671400 (gdb) p last_schedule $6 = 110698917 So the scenario becomes clearer: iothread VCPU thread -------------------------------------------------------------------------- aio_ctx_prepare aio_ctx_check qemu_poll_ns(timeout=-1) aio_poll aio_dispatch thread_pool_completion_bh qemu_bh_schedule() At this point bh->scheduled = 1 and the iothread has not been woken up. The solution must be close, but this alone should not be a problem, because the bottom half is only rescheduled to account for rare situations (see commit 3c80ca1, thread-pool: avoid deadlock in nested aio_poll() calls, 2014-07-15). Introducing a third thread---a thread pool worker thread, which also does qemu_bh_schedule()---does bring out the problematic case. The third thread must be awakened *after* the callback is complete and thread_pool_completion_bh has redone the whole loop, explaining the short race window. And then this is what happens: thread pool worker -------------------------------------------------------------------------- <I/O completes> qemu_bh_schedule() Tada, bh->scheduled is already 1, so qemu_bh_schedule() does nothing and the iothread is never woken up. This is where the bh->scheduled optimization comes into play---it is correct, but removing it would have masked the bug. So, what is the bug? Well, the question asked by the ctx->dispatching optimization ("is any active aio_poll dispatching?") was wrong. The right question to ask instead is "is any active aio_poll *not* dispatching", i.e. in the prepare or poll phases? In that case, the aio_poll is sleeping or might go to sleep anytime soon, and the EventNotifier must be invoked to wake it up. In any other case (including if there is *no* active aio_poll at all!) we can just wait for the next prepare phase to pick up the event (e.g. a bottom half); the prepare phase will avoid the blocking and service the bottom half. Expressing the invariant with a logic formula, the broken one looked like: !(exists(thread): in_dispatching(thread)) => !optimize or equivalently: !(exists(thread): in_aio_poll(thread) && in_dispatching(thread)) => !optimize In the correct one, the negation is in a slightly different place: (exists(thread): in_aio_poll(thread) && !in_dispatching(thread)) => !optimize or equivalently: (exists(thread): in_prepare_or_poll(thread)) => !optimize Even if the difference boils down to moving an exclamation mark :) the implementation is quite different. However, I think the new one is simpler to understand. In the old implementation, the "exists" was implemented with a boolean value. This didn't really support well the case of multiple concurrent event loops, but I thought that this was okay: aio_poll holds the AioContext lock so there cannot be concurrent aio_poll invocations, and I was just considering nested event loops. However, aio_poll _could_ indeed be concurrent with the GSource. This is why I came up with the wrong invariant. In the new implementation, "exists" is computed simply by counting how many threads are in the prepare or poll phases. There are some interesting points to consider, but the gist of the idea remains: 1) AioContext can be used through GSource as well; as mentioned in the patch, bit 0 of the counter is reserved for the GSource. 2) the counter need not be updated for a non-blocking aio_poll, because it won't sleep forever anyway. This is just a matter of checking the "blocking" variable. This requires some changes to the win32 implementation, but is otherwise not too complicated. 3) as mentioned above, the new implementation will not call aio_notify when there is *no* active aio_poll at all. The tests have to be adjusted for this change. The calls to aio_notify in async.c are fine; they only want to kick aio_poll out of a blocking wait, but need not do anything if aio_poll is not running. 4) nested aio_poll: these just work with the new implementation; when a nested event loop is invoked, the outer event loop is never in the prepare or poll phases. The outer event loop thus has already decremented the counter. Reported-by: Richard W. M. Jones <rjones@redhat.com> Reported-by: Laszlo Ersek <lersek@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Reviewed-by: Fam Zheng <famz@redhat.com> Tested-by: Richard W.M. Jones <rjones@redhat.com> Message-id: 1437487673-23740-5-git-send-email-pbonzini@redhat.com Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2015-07-21 16:07:51 +02:00
/* Write e.g. bh->scheduled before reading ctx->notify_me. Pairs
* with atomic_or in aio_ctx_prepare or atomic_add in aio_poll.
*/
smp_mb();
AioContext: fix broken ctx->dispatching optimization This patch rewrites the ctx->dispatching optimization, which was the cause of some mysterious hangs that could be reproduced on aarch64 KVM only. The hangs were indirectly caused by aio_poll() and in particular by flash memory updates's call to blk_write(), which invokes aio_poll(). Fun stuff: they had an extremely short race window, so much that adding all kind of tracing to either the kernel or QEMU made it go away (a single printf made it half as reproducible). On the plus side, the failure mode (a hang until the next keypress) made it very easy to examine the state of the process with a debugger. And there was a very nice reproducer from Laszlo, which failed pretty often (more than half of the time) on any version of QEMU with a non-debug kernel; it also failed fast, while still in the firmware. So, it could have been worse. For some unknown reason they happened only with virtio-scsi, but that's not important. It's more interesting that they disappeared with io=native, making thread-pool.c a likely suspect for where the bug arose. thread-pool.c is also one of the few places which use bottom halves across threads, by the way. I hope that no other similar bugs exist, but just in case :) I am going to describe how the successful debugging went... Since the likely culprit was the ctx->dispatching optimization, which mostly affects bottom halves, the first observation was that there are two qemu_bh_schedule() invocations in the thread pool: the one in the aio worker and the one in thread_pool_completion_bh. The latter always causes the optimization to trigger, the former may or may not. In order to restrict the possibilities, I introduced new functions qemu_bh_schedule_slow() and qemu_bh_schedule_fast(): /* qemu_bh_schedule_slow: */ ctx = bh->ctx; bh->idle = 0; if (atomic_xchg(&bh->scheduled, 1) == 0) { event_notifier_set(&ctx->notifier); } /* qemu_bh_schedule_fast: */ ctx = bh->ctx; bh->idle = 0; assert(ctx->dispatching); atomic_xchg(&bh->scheduled, 1); Notice how the atomic_xchg is still in qemu_bh_schedule_slow(). This was already debated a few months ago, so I assumed it to be correct. In retrospect this was a very good idea, as you'll see later. Changing thread_pool_completion_bh() to qemu_bh_schedule_fast() didn't trigger the assertion (as expected). Changing the worker's invocation to qemu_bh_schedule_slow() didn't hide the bug (another assumption which luckily held). This already limited heavily the amount of interaction between the threads, hinting that the problematic events must have triggered around thread_pool_completion_bh(). As mentioned early, invoking a debugger to examine the state of a hung process was pretty easy; the iothread was always waiting on a poll(..., -1) system call. Infinite timeouts are much rarer on x86, and this could be the reason why the bug was never observed there. With the buggy sequence more or less resolved to an interaction between thread_pool_completion_bh() and poll(..., -1), my "tracing" strategy was to just add a few qemu_clock_get_ns(QEMU_CLOCK_REALTIME) calls, hoping that the ordering of aio_ctx_prepare(), aio_ctx_dispatch, poll() and qemu_bh_schedule_fast() would provide some hint. The output was: (gdb) p last_prepare $3 = 103885451 (gdb) p last_dispatch $4 = 103876492 (gdb) p last_poll $5 = 115909333 (gdb) p last_schedule $6 = 115925212 Notice how the last call to qemu_poll_ns() came after aio_ctx_dispatch(). This makes little sense unless there is an aio_poll() call involved, and indeed with a slightly different instrumentation you can see that there is one: (gdb) p last_prepare $3 = 107569679 (gdb) p last_dispatch $4 = 107561600 (gdb) p last_aio_poll $5 = 110671400 (gdb) p last_schedule $6 = 110698917 So the scenario becomes clearer: iothread VCPU thread -------------------------------------------------------------------------- aio_ctx_prepare aio_ctx_check qemu_poll_ns(timeout=-1) aio_poll aio_dispatch thread_pool_completion_bh qemu_bh_schedule() At this point bh->scheduled = 1 and the iothread has not been woken up. The solution must be close, but this alone should not be a problem, because the bottom half is only rescheduled to account for rare situations (see commit 3c80ca1, thread-pool: avoid deadlock in nested aio_poll() calls, 2014-07-15). Introducing a third thread---a thread pool worker thread, which also does qemu_bh_schedule()---does bring out the problematic case. The third thread must be awakened *after* the callback is complete and thread_pool_completion_bh has redone the whole loop, explaining the short race window. And then this is what happens: thread pool worker -------------------------------------------------------------------------- <I/O completes> qemu_bh_schedule() Tada, bh->scheduled is already 1, so qemu_bh_schedule() does nothing and the iothread is never woken up. This is where the bh->scheduled optimization comes into play---it is correct, but removing it would have masked the bug. So, what is the bug? Well, the question asked by the ctx->dispatching optimization ("is any active aio_poll dispatching?") was wrong. The right question to ask instead is "is any active aio_poll *not* dispatching", i.e. in the prepare or poll phases? In that case, the aio_poll is sleeping or might go to sleep anytime soon, and the EventNotifier must be invoked to wake it up. In any other case (including if there is *no* active aio_poll at all!) we can just wait for the next prepare phase to pick up the event (e.g. a bottom half); the prepare phase will avoid the blocking and service the bottom half. Expressing the invariant with a logic formula, the broken one looked like: !(exists(thread): in_dispatching(thread)) => !optimize or equivalently: !(exists(thread): in_aio_poll(thread) && in_dispatching(thread)) => !optimize In the correct one, the negation is in a slightly different place: (exists(thread): in_aio_poll(thread) && !in_dispatching(thread)) => !optimize or equivalently: (exists(thread): in_prepare_or_poll(thread)) => !optimize Even if the difference boils down to moving an exclamation mark :) the implementation is quite different. However, I think the new one is simpler to understand. In the old implementation, the "exists" was implemented with a boolean value. This didn't really support well the case of multiple concurrent event loops, but I thought that this was okay: aio_poll holds the AioContext lock so there cannot be concurrent aio_poll invocations, and I was just considering nested event loops. However, aio_poll _could_ indeed be concurrent with the GSource. This is why I came up with the wrong invariant. In the new implementation, "exists" is computed simply by counting how many threads are in the prepare or poll phases. There are some interesting points to consider, but the gist of the idea remains: 1) AioContext can be used through GSource as well; as mentioned in the patch, bit 0 of the counter is reserved for the GSource. 2) the counter need not be updated for a non-blocking aio_poll, because it won't sleep forever anyway. This is just a matter of checking the "blocking" variable. This requires some changes to the win32 implementation, but is otherwise not too complicated. 3) as mentioned above, the new implementation will not call aio_notify when there is *no* active aio_poll at all. The tests have to be adjusted for this change. The calls to aio_notify in async.c are fine; they only want to kick aio_poll out of a blocking wait, but need not do anything if aio_poll is not running. 4) nested aio_poll: these just work with the new implementation; when a nested event loop is invoked, the outer event loop is never in the prepare or poll phases. The outer event loop thus has already decremented the counter. Reported-by: Richard W. M. Jones <rjones@redhat.com> Reported-by: Laszlo Ersek <lersek@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Reviewed-by: Fam Zheng <famz@redhat.com> Tested-by: Richard W.M. Jones <rjones@redhat.com> Message-id: 1437487673-23740-5-git-send-email-pbonzini@redhat.com Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2015-07-21 16:07:51 +02:00
if (ctx->notify_me) {
event_notifier_set(&ctx->notifier);
atomic_mb_set(&ctx->notified, true);
}
}
void aio_notify_accept(AioContext *ctx)
{
if (atomic_xchg(&ctx->notified, false)) {
event_notifier_test_and_clear(&ctx->notifier);
}
}
static void aio_timerlist_notify(void *opaque)
{
aio_notify(opaque);
}
AioContext: fix broken placement of event_notifier_test_and_clear event_notifier_test_and_clear must be called before processing events. Otherwise, an aio_poll could "eat" the notification before the main I/O thread invokes ppoll(). The main I/O thread then never wakes up. This is an example of what could happen: i/o thread vcpu thread worker thread --------------------------------------------------------------------- lock_iothread notify_me = 1 ... unlock_iothread bh->scheduled = 1 event_notifier_set lock_iothread notify_me = 3 ppoll notify_me = 1 aio_dispatch aio_bh_poll thread_pool_completion_bh bh->scheduled = 1 event_notifier_set node->io_read(node->opaque) event_notifier_test_and_clear ppoll *** hang *** "Tracing" with qemu_clock_get_ns shows pretty much the same behavior as in the previous bug, so there are no new tricks here---just stare more at the code until it is apparent. One could also use a formal model, of course. The included one shows this with three processes: notifier corresponds to a QEMU thread pool worker, temporary_waiter to a VCPU thread that invokes aio_poll(), waiter to the main I/O thread. I would be happy to say that the formal model found the bug for me, but actually I wrote it after the fact. This patch is a bit of a big hammer. The next one optimizes it, with help (this time for real rather than a posteriori :)) from another, similar formal model. Reported-by: Richard W. M. Jones <rjones@redhat.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Reviewed-by: Fam Zheng <famz@redhat.com> Tested-by: Richard W.M. Jones <rjones@redhat.com> Message-id: 1437487673-23740-6-git-send-email-pbonzini@redhat.com Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2015-07-21 16:07:52 +02:00
static void event_notifier_dummy_cb(EventNotifier *e)
{
}
aio: add polling mode to AioContext The AioContext event loop uses ppoll(2) or epoll_wait(2) to monitor file descriptors or until a timer expires. In cases like virtqueues, Linux AIO, and ThreadPool it is technically possible to wait for events via polling (i.e. continuously checking for events without blocking). Polling can be faster than blocking syscalls because file descriptors, the process scheduler, and system calls are bypassed. The main disadvantage to polling is that it increases CPU utilization. In classic polling configuration a full host CPU thread might run at 100% to respond to events as quickly as possible. This patch implements a timeout so we fall back to blocking syscalls if polling detects no activity. After the timeout no CPU cycles are wasted on polling until the next event loop iteration. The run_poll_handlers_begin() and run_poll_handlers_end() trace events are added to aid performance analysis and troubleshooting. If you need to know whether polling mode is being used, trace these events to find out. Note that the AioContext is now re-acquired before disabling notify_me in the non-polling case. This makes the code cleaner since notify_me was enabled outside the non-polling AioContext release region. This change is correct since it's safe to keep notify_me enabled longer (disabling is an optimization) but potentially causes unnecessary event_notifer_set() calls. I think the chance of performance regression is small here. Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> Message-id: 20161201192652.9509-4-stefanha@redhat.com Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2016-12-01 20:26:42 +01:00
/* Returns true if aio_notify() was called (e.g. a BH was scheduled) */
static bool event_notifier_poll(void *opaque)
{
EventNotifier *e = opaque;
AioContext *ctx = container_of(e, AioContext, notifier);
return atomic_read(&ctx->notified);
}
AioContext *aio_context_new(Error **errp)
{
int ret;
AioContext *ctx;
ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext));
aio_context_setup(ctx);
ret = event_notifier_init(&ctx->notifier, false);
if (ret < 0) {
error_setg_errno(errp, -ret, "Failed to initialize event notifier");
goto fail;
}
g_source_set_can_recurse(&ctx->source, true);
aio_set_event_notifier(ctx, &ctx->notifier,
false,
(EventNotifierHandler *)
event_notifier_dummy_cb,
aio: add polling mode to AioContext The AioContext event loop uses ppoll(2) or epoll_wait(2) to monitor file descriptors or until a timer expires. In cases like virtqueues, Linux AIO, and ThreadPool it is technically possible to wait for events via polling (i.e. continuously checking for events without blocking). Polling can be faster than blocking syscalls because file descriptors, the process scheduler, and system calls are bypassed. The main disadvantage to polling is that it increases CPU utilization. In classic polling configuration a full host CPU thread might run at 100% to respond to events as quickly as possible. This patch implements a timeout so we fall back to blocking syscalls if polling detects no activity. After the timeout no CPU cycles are wasted on polling until the next event loop iteration. The run_poll_handlers_begin() and run_poll_handlers_end() trace events are added to aid performance analysis and troubleshooting. If you need to know whether polling mode is being used, trace these events to find out. Note that the AioContext is now re-acquired before disabling notify_me in the non-polling case. This makes the code cleaner since notify_me was enabled outside the non-polling AioContext release region. This change is correct since it's safe to keep notify_me enabled longer (disabling is an optimization) but potentially causes unnecessary event_notifer_set() calls. I think the chance of performance regression is small here. Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> Message-id: 20161201192652.9509-4-stefanha@redhat.com Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2016-12-01 20:26:42 +01:00
event_notifier_poll);
#ifdef CONFIG_LINUX_AIO
ctx->linux_aio = NULL;
#endif
ctx->thread_pool = NULL;
qemu_mutex_init(&ctx->bh_lock);
qemu_rec_mutex_init(&ctx->lock);
timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx);
ctx->poll_ns = 0;
aio: add polling mode to AioContext The AioContext event loop uses ppoll(2) or epoll_wait(2) to monitor file descriptors or until a timer expires. In cases like virtqueues, Linux AIO, and ThreadPool it is technically possible to wait for events via polling (i.e. continuously checking for events without blocking). Polling can be faster than blocking syscalls because file descriptors, the process scheduler, and system calls are bypassed. The main disadvantage to polling is that it increases CPU utilization. In classic polling configuration a full host CPU thread might run at 100% to respond to events as quickly as possible. This patch implements a timeout so we fall back to blocking syscalls if polling detects no activity. After the timeout no CPU cycles are wasted on polling until the next event loop iteration. The run_poll_handlers_begin() and run_poll_handlers_end() trace events are added to aid performance analysis and troubleshooting. If you need to know whether polling mode is being used, trace these events to find out. Note that the AioContext is now re-acquired before disabling notify_me in the non-polling case. This makes the code cleaner since notify_me was enabled outside the non-polling AioContext release region. This change is correct since it's safe to keep notify_me enabled longer (disabling is an optimization) but potentially causes unnecessary event_notifer_set() calls. I think the chance of performance regression is small here. Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> Message-id: 20161201192652.9509-4-stefanha@redhat.com Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2016-12-01 20:26:42 +01:00
ctx->poll_max_ns = 0;
ctx->poll_grow = 0;
ctx->poll_shrink = 0;
aio: add polling mode to AioContext The AioContext event loop uses ppoll(2) or epoll_wait(2) to monitor file descriptors or until a timer expires. In cases like virtqueues, Linux AIO, and ThreadPool it is technically possible to wait for events via polling (i.e. continuously checking for events without blocking). Polling can be faster than blocking syscalls because file descriptors, the process scheduler, and system calls are bypassed. The main disadvantage to polling is that it increases CPU utilization. In classic polling configuration a full host CPU thread might run at 100% to respond to events as quickly as possible. This patch implements a timeout so we fall back to blocking syscalls if polling detects no activity. After the timeout no CPU cycles are wasted on polling until the next event loop iteration. The run_poll_handlers_begin() and run_poll_handlers_end() trace events are added to aid performance analysis and troubleshooting. If you need to know whether polling mode is being used, trace these events to find out. Note that the AioContext is now re-acquired before disabling notify_me in the non-polling case. This makes the code cleaner since notify_me was enabled outside the non-polling AioContext release region. This change is correct since it's safe to keep notify_me enabled longer (disabling is an optimization) but potentially causes unnecessary event_notifer_set() calls. I think the chance of performance regression is small here. Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> Message-id: 20161201192652.9509-4-stefanha@redhat.com Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2016-12-01 20:26:42 +01:00
return ctx;
fail:
g_source_destroy(&ctx->source);
return NULL;
}
void aio_context_ref(AioContext *ctx)
{
g_source_ref(&ctx->source);
}
void aio_context_unref(AioContext *ctx)
{
g_source_unref(&ctx->source);
}
void aio_context_acquire(AioContext *ctx)
{
qemu_rec_mutex_lock(&ctx->lock);
}
void aio_context_release(AioContext *ctx)
{
qemu_rec_mutex_unlock(&ctx->lock);
}