qemu-e2k/aio-posix.c

495 lines
13 KiB
C
Raw Normal View History

/*
* QEMU aio implementation
*
* Copyright IBM, Corp. 2008
*
* Authors:
* Anthony Liguori <aliguori@us.ibm.com>
*
* This work is licensed under the terms of the GNU GPL, version 2. See
* the COPYING file in the top-level directory.
*
* Contributions after 2012-01-13 are licensed under the terms of the
* GNU GPL, version 2 or (at your option) any later version.
*/
#include "qemu-common.h"
#include "block/block.h"
#include "qemu/queue.h"
#include "qemu/sockets.h"
#ifdef CONFIG_EPOLL
#include <sys/epoll.h>
#endif
struct AioHandler
{
GPollFD pfd;
IOHandler *io_read;
IOHandler *io_write;
int deleted;
void *opaque;
bool is_external;
QLIST_ENTRY(AioHandler) node;
};
#ifdef CONFIG_EPOLL
/* The fd number threashold to switch to epoll */
#define EPOLL_ENABLE_THRESHOLD 64
static void aio_epoll_disable(AioContext *ctx)
{
ctx->epoll_available = false;
if (!ctx->epoll_enabled) {
return;
}
ctx->epoll_enabled = false;
close(ctx->epollfd);
}
static inline int epoll_events_from_pfd(int pfd_events)
{
return (pfd_events & G_IO_IN ? EPOLLIN : 0) |
(pfd_events & G_IO_OUT ? EPOLLOUT : 0) |
(pfd_events & G_IO_HUP ? EPOLLHUP : 0) |
(pfd_events & G_IO_ERR ? EPOLLERR : 0);
}
static bool aio_epoll_try_enable(AioContext *ctx)
{
AioHandler *node;
struct epoll_event event;
QLIST_FOREACH(node, &ctx->aio_handlers, node) {
int r;
if (node->deleted || !node->pfd.events) {
continue;
}
event.events = epoll_events_from_pfd(node->pfd.events);
event.data.ptr = node;
r = epoll_ctl(ctx->epollfd, EPOLL_CTL_ADD, node->pfd.fd, &event);
if (r) {
return false;
}
}
ctx->epoll_enabled = true;
return true;
}
static void aio_epoll_update(AioContext *ctx, AioHandler *node, bool is_new)
{
struct epoll_event event;
int r;
if (!ctx->epoll_enabled) {
return;
}
if (!node->pfd.events) {
r = epoll_ctl(ctx->epollfd, EPOLL_CTL_DEL, node->pfd.fd, &event);
if (r) {
aio_epoll_disable(ctx);
}
} else {
event.data.ptr = node;
event.events = epoll_events_from_pfd(node->pfd.events);
if (is_new) {
r = epoll_ctl(ctx->epollfd, EPOLL_CTL_ADD, node->pfd.fd, &event);
if (r) {
aio_epoll_disable(ctx);
}
} else {
r = epoll_ctl(ctx->epollfd, EPOLL_CTL_MOD, node->pfd.fd, &event);
if (r) {
aio_epoll_disable(ctx);
}
}
}
}
static int aio_epoll(AioContext *ctx, GPollFD *pfds,
unsigned npfd, int64_t timeout)
{
AioHandler *node;
int i, ret = 0;
struct epoll_event events[128];
assert(npfd == 1);
assert(pfds[0].fd == ctx->epollfd);
if (timeout > 0) {
ret = qemu_poll_ns(pfds, npfd, timeout);
}
if (timeout <= 0 || ret > 0) {
ret = epoll_wait(ctx->epollfd, events,
sizeof(events) / sizeof(events[0]),
timeout);
if (ret <= 0) {
goto out;
}
for (i = 0; i < ret; i++) {
int ev = events[i].events;
node = events[i].data.ptr;
node->pfd.revents = (ev & EPOLLIN ? G_IO_IN : 0) |
(ev & EPOLLOUT ? G_IO_OUT : 0) |
(ev & EPOLLHUP ? G_IO_HUP : 0) |
(ev & EPOLLERR ? G_IO_ERR : 0);
}
}
out:
return ret;
}
static bool aio_epoll_enabled(AioContext *ctx)
{
/* Fall back to ppoll when external clients are disabled. */
return !aio_external_disabled(ctx) && ctx->epoll_enabled;
}
static bool aio_epoll_check_poll(AioContext *ctx, GPollFD *pfds,
unsigned npfd, int64_t timeout)
{
if (!ctx->epoll_available) {
return false;
}
if (aio_epoll_enabled(ctx)) {
return true;
}
if (npfd >= EPOLL_ENABLE_THRESHOLD) {
if (aio_epoll_try_enable(ctx)) {
return true;
} else {
aio_epoll_disable(ctx);
}
}
return false;
}
#else
static void aio_epoll_update(AioContext *ctx, AioHandler *node, bool is_new)
{
}
static int aio_epoll(AioContext *ctx, GPollFD *pfds,
unsigned npfd, int64_t timeout)
{
assert(false);
}
static bool aio_epoll_enabled(AioContext *ctx)
{
return false;
}
static bool aio_epoll_check_poll(AioContext *ctx, GPollFD *pfds,
unsigned npfd, int64_t timeout)
{
return false;
}
#endif
static AioHandler *find_aio_handler(AioContext *ctx, int fd)
{
AioHandler *node;
QLIST_FOREACH(node, &ctx->aio_handlers, node) {
if (node->pfd.fd == fd)
if (!node->deleted)
return node;
}
return NULL;
}
void aio_set_fd_handler(AioContext *ctx,
int fd,
bool is_external,
IOHandler *io_read,
IOHandler *io_write,
void *opaque)
{
AioHandler *node;
bool is_new = false;
bool deleted = false;
node = find_aio_handler(ctx, fd);
/* Are we deleting the fd handler? */
if (!io_read && !io_write) {
if (node) {
g_source_remove_poll(&ctx->source, &node->pfd);
/* If the lock is held, just mark the node as deleted */
if (ctx->walking_handlers) {
node->deleted = 1;
node->pfd.revents = 0;
} else {
/* Otherwise, delete it for real. We can't just mark it as
* deleted because deleted nodes are only cleaned up after
* releasing the walking_handlers lock.
*/
QLIST_REMOVE(node, node);
deleted = true;
}
}
} else {
if (node == NULL) {
/* Alloc and insert if it's not already there */
node = g_new0(AioHandler, 1);
node->pfd.fd = fd;
QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node);
g_source_add_poll(&ctx->source, &node->pfd);
is_new = true;
}
/* Update handler with latest information */
node->io_read = io_read;
node->io_write = io_write;
node->opaque = opaque;
node->is_external = is_external;
node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP | G_IO_ERR : 0);
node->pfd.events |= (io_write ? G_IO_OUT | G_IO_ERR : 0);
}
aio_epoll_update(ctx, node, is_new);
aio_notify(ctx);
if (deleted) {
g_free(node);
}
}
void aio_set_event_notifier(AioContext *ctx,
EventNotifier *notifier,
bool is_external,
EventNotifierHandler *io_read)
{
aio_set_fd_handler(ctx, event_notifier_get_fd(notifier),
is_external, (IOHandler *)io_read, NULL, notifier);
}
bool aio_prepare(AioContext *ctx)
{
return false;
}
bool aio_pending(AioContext *ctx)
{
AioHandler *node;
QLIST_FOREACH(node, &ctx->aio_handlers, node) {
int revents;
revents = node->pfd.revents & node->pfd.events;
if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read) {
return true;
}
if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write) {
return true;
}
}
return false;
}
bool aio_dispatch(AioContext *ctx)
{
AioHandler *node;
bool progress = false;
/*
* If there are callbacks left that have been queued, we need to call them.
* Do not call select in this case, because it is possible that the caller
* does not need a complete flush (as is the case for aio_poll loops).
*/
if (aio_bh_poll(ctx)) {
progress = true;
}
/*
* We have to walk very carefully in case aio_set_fd_handler is
* called while we're walking.
*/
node = QLIST_FIRST(&ctx->aio_handlers);
while (node) {
AioHandler *tmp;
int revents;
ctx->walking_handlers++;
revents = node->pfd.revents & node->pfd.events;
node->pfd.revents = 0;
if (!node->deleted &&
(revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) &&
node->io_read) {
node->io_read(node->opaque);
aio: stop using .io_flush() Now that aio_poll() users check their termination condition themselves, it is no longer necessary to call .io_flush() handlers. The behavior of aio_poll() changes as follows: 1. .io_flush() is no longer invoked and file descriptors are *always* monitored. Previously returning 0 from .io_flush() would skip this file descriptor. Due to this change it is essential to check that requests are pending before calling qemu_aio_wait(). Failure to do so means we block, for example, waiting for an idle iSCSI socket to become readable when there are no requests. Currently all qemu_aio_wait()/aio_poll() callers check before calling. 2. aio_poll() now returns true if progress was made (BH or fd handlers executed) and false otherwise. Previously it would return true whenever 'busy', which means that .io_flush() returned true. The 'busy' concept no longer exists so just progress is returned. Due to this change we need to update tests/test-aio.c which asserts aio_poll() return values. Note that QEMU doesn't actually rely on these return values so only tests/test-aio.c cares. Note that ctx->notifier, the EventNotifier fd used for aio_notify(), is now handled as a special case. This is a little ugly but maintains aio_poll() semantics, i.e. aio_notify() does not count as 'progress' and aio_poll() avoids blocking when the user has not set any fd handlers yet. Patches after this remove .io_flush() handler code until we can finally drop the io_flush arguments to aio_set_fd_handler() and friends. Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2013-04-11 16:56:50 +02:00
/* aio_notify() does not count as progress */
if (node->opaque != &ctx->notifier) {
progress = true;
}
}
if (!node->deleted &&
(revents & (G_IO_OUT | G_IO_ERR)) &&
node->io_write) {
node->io_write(node->opaque);
progress = true;
}
tmp = node;
node = QLIST_NEXT(node, node);
ctx->walking_handlers--;
if (!ctx->walking_handlers && tmp->deleted) {
QLIST_REMOVE(tmp, node);
g_free(tmp);
}
}
/* Run our timers */
progress |= timerlistgroup_run_timers(&ctx->tlg);
return progress;
}
/* These thread-local variables are used only in a small part of aio_poll
* around the call to the poll() system call. In particular they are not
* used while aio_poll is performing callbacks, which makes it much easier
* to think about reentrancy!
*
* Stack-allocated arrays would be perfect but they have size limitations;
* heap allocation is expensive enough that we want to reuse arrays across
* calls to aio_poll(). And because poll() has to be called without holding
* any lock, the arrays cannot be stored in AioContext. Thread-local data
* has none of the disadvantages of these three options.
*/
static __thread GPollFD *pollfds;
static __thread AioHandler **nodes;
static __thread unsigned npfd, nalloc;
static __thread Notifier pollfds_cleanup_notifier;
static void pollfds_cleanup(Notifier *n, void *unused)
{
g_assert(npfd == 0);
g_free(pollfds);
g_free(nodes);
nalloc = 0;
}
static void add_pollfd(AioHandler *node)
{
if (npfd == nalloc) {
if (nalloc == 0) {
pollfds_cleanup_notifier.notify = pollfds_cleanup;
qemu_thread_atexit_add(&pollfds_cleanup_notifier);
nalloc = 8;
} else {
g_assert(nalloc <= INT_MAX);
nalloc *= 2;
}
pollfds = g_renew(GPollFD, pollfds, nalloc);
nodes = g_renew(AioHandler *, nodes, nalloc);
}
nodes[npfd] = node;
pollfds[npfd] = (GPollFD) {
.fd = node->pfd.fd,
.events = node->pfd.events,
};
npfd++;
}
bool aio_poll(AioContext *ctx, bool blocking)
{
AioHandler *node;
int i, ret;
aio: stop using .io_flush() Now that aio_poll() users check their termination condition themselves, it is no longer necessary to call .io_flush() handlers. The behavior of aio_poll() changes as follows: 1. .io_flush() is no longer invoked and file descriptors are *always* monitored. Previously returning 0 from .io_flush() would skip this file descriptor. Due to this change it is essential to check that requests are pending before calling qemu_aio_wait(). Failure to do so means we block, for example, waiting for an idle iSCSI socket to become readable when there are no requests. Currently all qemu_aio_wait()/aio_poll() callers check before calling. 2. aio_poll() now returns true if progress was made (BH or fd handlers executed) and false otherwise. Previously it would return true whenever 'busy', which means that .io_flush() returned true. The 'busy' concept no longer exists so just progress is returned. Due to this change we need to update tests/test-aio.c which asserts aio_poll() return values. Note that QEMU doesn't actually rely on these return values so only tests/test-aio.c cares. Note that ctx->notifier, the EventNotifier fd used for aio_notify(), is now handled as a special case. This is a little ugly but maintains aio_poll() semantics, i.e. aio_notify() does not count as 'progress' and aio_poll() avoids blocking when the user has not set any fd handlers yet. Patches after this remove .io_flush() handler code until we can finally drop the io_flush arguments to aio_set_fd_handler() and friends. Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2013-04-11 16:56:50 +02:00
bool progress;
int64_t timeout;
aio_context_acquire(ctx);
progress = false;
/* aio_notify can avoid the expensive event_notifier_set if
* everything (file descriptors, bottom halves, timers) will
* be re-evaluated before the next blocking poll(). This is
* already true when aio_poll is called with blocking == false;
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 blocking == true, it is only true after poll() returns,
* so disable the optimization now.
*/
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 (blocking) {
atomic_add(&ctx->notify_me, 2);
}
ctx->walking_handlers++;
assert(npfd == 0);
/* fill pollfds */
QLIST_FOREACH(node, &ctx->aio_handlers, node) {
if (!node->deleted && node->pfd.events
&& !aio_epoll_enabled(ctx)
&& aio_node_check(ctx, node->is_external)) {
add_pollfd(node);
}
}
timeout = blocking ? aio_compute_timeout(ctx) : 0;
/* wait until next event */
if (timeout) {
aio_context_release(ctx);
}
if (aio_epoll_check_poll(ctx, pollfds, npfd, timeout)) {
AioHandler epoll_handler;
epoll_handler.pfd.fd = ctx->epollfd;
epoll_handler.pfd.events = G_IO_IN | G_IO_OUT | G_IO_HUP | G_IO_ERR;
npfd = 0;
add_pollfd(&epoll_handler);
ret = aio_epoll(ctx, pollfds, npfd, timeout);
} else {
ret = qemu_poll_ns(pollfds, npfd, timeout);
}
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 (blocking) {
atomic_sub(&ctx->notify_me, 2);
}
if (timeout) {
aio_context_acquire(ctx);
}
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
/* if we have any readable fds, dispatch event */
if (ret > 0) {
for (i = 0; i < npfd; i++) {
nodes[i]->pfd.revents = pollfds[i].revents;
}
}
npfd = 0;
ctx->walking_handlers--;
/* Run dispatch even if there were no readable fds to run timers */
if (aio_dispatch(ctx)) {
progress = true;
}
aio_context_release(ctx);
aio: stop using .io_flush() Now that aio_poll() users check their termination condition themselves, it is no longer necessary to call .io_flush() handlers. The behavior of aio_poll() changes as follows: 1. .io_flush() is no longer invoked and file descriptors are *always* monitored. Previously returning 0 from .io_flush() would skip this file descriptor. Due to this change it is essential to check that requests are pending before calling qemu_aio_wait(). Failure to do so means we block, for example, waiting for an idle iSCSI socket to become readable when there are no requests. Currently all qemu_aio_wait()/aio_poll() callers check before calling. 2. aio_poll() now returns true if progress was made (BH or fd handlers executed) and false otherwise. Previously it would return true whenever 'busy', which means that .io_flush() returned true. The 'busy' concept no longer exists so just progress is returned. Due to this change we need to update tests/test-aio.c which asserts aio_poll() return values. Note that QEMU doesn't actually rely on these return values so only tests/test-aio.c cares. Note that ctx->notifier, the EventNotifier fd used for aio_notify(), is now handled as a special case. This is a little ugly but maintains aio_poll() semantics, i.e. aio_notify() does not count as 'progress' and aio_poll() avoids blocking when the user has not set any fd handlers yet. Patches after this remove .io_flush() handler code until we can finally drop the io_flush arguments to aio_set_fd_handler() and friends. Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
2013-04-11 16:56:50 +02:00
return progress;
}
void aio_context_setup(AioContext *ctx, Error **errp)
{
#ifdef CONFIG_EPOLL
assert(!ctx->epollfd);
ctx->epollfd = epoll_create1(EPOLL_CLOEXEC);
if (ctx->epollfd == -1) {
ctx->epoll_available = false;
} else {
ctx->epoll_available = true;
}
#endif
}