From 999006975840f8cdf2038a587d852a6cbfe58e3b Mon Sep 17 00:00:00 2001 From: Fam Zheng Date: Thu, 9 Jul 2015 11:47:58 +0800 Subject: [PATCH 1/7] mirror: Speed up bitmap initial scanning Limiting to sectors_per_chunk for each bdrv_is_allocated_above is slow, because the underlying protocol driver would issue much more queries than necessary. We should coalesce the query. Signed-off-by: Fam Zheng Reviewed-by: Stefan Hajnoczi Message-id: <1436413678-7114-4-git-send-email-famz@redhat.com> Signed-off-by: Stefan Hajnoczi --- block/mirror.c | 14 ++++++-------- 1 file changed, 6 insertions(+), 8 deletions(-) diff --git a/block/mirror.c b/block/mirror.c index 323f747c75..fc4d8f561e 100644 --- a/block/mirror.c +++ b/block/mirror.c @@ -388,7 +388,7 @@ static void coroutine_fn mirror_run(void *opaque) MirrorBlockJob *s = opaque; MirrorExitData *data; BlockDriverState *bs = s->common.bs; - int64_t sector_num, end, sectors_per_chunk, length; + int64_t sector_num, end, length; uint64_t last_pause_ns; BlockDriverInfo bdi; char backing_filename[2]; /* we only need 2 characters because we are only @@ -442,7 +442,6 @@ static void coroutine_fn mirror_run(void *opaque) goto immediate_exit; } - sectors_per_chunk = s->granularity >> BDRV_SECTOR_BITS; mirror_free_init(s); last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); @@ -450,7 +449,9 @@ static void coroutine_fn mirror_run(void *opaque) /* First part, loop on the sectors and initialize the dirty bitmap. */ BlockDriverState *base = s->base; for (sector_num = 0; sector_num < end; ) { - int64_t next = (sector_num | (sectors_per_chunk - 1)) + 1; + /* Just to make sure we are not exceeding int limit. */ + int nb_sectors = MIN(INT_MAX >> BDRV_SECTOR_BITS, + end - sector_num); int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); if (now - last_pause_ns > SLICE_TIME) { @@ -462,8 +463,7 @@ static void coroutine_fn mirror_run(void *opaque) goto immediate_exit; } - ret = bdrv_is_allocated_above(bs, base, - sector_num, next - sector_num, &n); + ret = bdrv_is_allocated_above(bs, base, sector_num, nb_sectors, &n); if (ret < 0) { goto immediate_exit; @@ -472,10 +472,8 @@ static void coroutine_fn mirror_run(void *opaque) assert(n > 0); if (ret == 1) { bdrv_set_dirty_bitmap(s->dirty_bitmap, sector_num, n); - sector_num = next; - } else { - sector_num += n; } + sector_num += n; } } From e4efd8a488d0a68b0af34d8ee245463df7c8bdf4 Mon Sep 17 00:00:00 2001 From: Paolo Bonzini Date: Tue, 21 Jul 2015 16:07:48 +0200 Subject: [PATCH 2/7] qemu-timer: initialize "timers_done_ev" to set The normal value for the event is to be set. If we do not do this, pause_all_vcpus (through qemu_clock_enable) hangs unless timerlist_run_timers has been run at least once for the timerlist. This can happen with the following patches, that make aio_notify do nothing most of the time. Signed-off-by: Paolo Bonzini Reviewed-by: Fam Zheng Tested-by: Richard W.M. Jones Message-id: 1437487673-23740-2-git-send-email-pbonzini@redhat.com Signed-off-by: Stefan Hajnoczi --- qemu-timer.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/qemu-timer.c b/qemu-timer.c index aa6757e359..2463fe6f6a 100644 --- a/qemu-timer.c +++ b/qemu-timer.c @@ -99,7 +99,7 @@ QEMUTimerList *timerlist_new(QEMUClockType type, QEMUClock *clock = qemu_clock_ptr(type); timer_list = g_malloc0(sizeof(QEMUTimerList)); - qemu_event_init(&timer_list->timers_done_ev, false); + qemu_event_init(&timer_list->timers_done_ev, true); timer_list->clock = clock; timer_list->notify_cb = cb; timer_list->notify_opaque = opaque; From 12d69ac03b45156356b240424623719f15d8143e Mon Sep 17 00:00:00 2001 From: Paolo Bonzini Date: Tue, 21 Jul 2015 16:07:49 +0200 Subject: [PATCH 3/7] tests: remove irrelevant assertions from test-aio In these tests, the purpose of the initial calls to aio_poll and g_main_context_iteration is simply to put the AioContext in a known state; the return value of the function does not really matter. The next patch will change those return values; change the assertions to a while loop which expresses the intention better. Signed-off-by: Paolo Bonzini Reviewed-by: Fam Zheng Tested-by: Richard W.M. Jones Message-id: 1437487673-23740-3-git-send-email-pbonzini@redhat.com Signed-off-by: Stefan Hajnoczi --- tests/test-aio.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/tests/test-aio.c b/tests/test-aio.c index a7cb5c9915..e7bbb8345a 100644 --- a/tests/test-aio.c +++ b/tests/test-aio.c @@ -331,7 +331,7 @@ static void test_wait_event_notifier(void) EventNotifierTestData data = { .n = 0, .active = 1 }; event_notifier_init(&data.e, false); aio_set_event_notifier(ctx, &data.e, event_ready_cb); - g_assert(!aio_poll(ctx, false)); + while (aio_poll(ctx, false)); g_assert_cmpint(data.n, ==, 0); g_assert_cmpint(data.active, ==, 1); @@ -356,7 +356,7 @@ static void test_flush_event_notifier(void) EventNotifierTestData data = { .n = 0, .active = 10, .auto_set = true }; event_notifier_init(&data.e, false); aio_set_event_notifier(ctx, &data.e, event_ready_cb); - g_assert(!aio_poll(ctx, false)); + while (aio_poll(ctx, false)); g_assert_cmpint(data.n, ==, 0); g_assert_cmpint(data.active, ==, 10); @@ -669,7 +669,7 @@ static void test_source_wait_event_notifier(void) EventNotifierTestData data = { .n = 0, .active = 1 }; event_notifier_init(&data.e, false); aio_set_event_notifier(ctx, &data.e, event_ready_cb); - g_assert(g_main_context_iteration(NULL, false)); + while (g_main_context_iteration(NULL, false)); g_assert_cmpint(data.n, ==, 0); g_assert_cmpint(data.active, ==, 1); @@ -694,7 +694,7 @@ static void test_source_flush_event_notifier(void) EventNotifierTestData data = { .n = 0, .active = 10, .auto_set = true }; event_notifier_init(&data.e, false); aio_set_event_notifier(ctx, &data.e, event_ready_cb); - g_assert(g_main_context_iteration(NULL, false)); + while (g_main_context_iteration(NULL, false)); g_assert_cmpint(data.n, ==, 0); g_assert_cmpint(data.active, ==, 10); From 6493c975af75be5b8d9ade954239bdf5492b7911 Mon Sep 17 00:00:00 2001 From: Paolo Bonzini Date: Tue, 21 Jul 2015 16:07:50 +0200 Subject: [PATCH 4/7] aio-win32: reorganize polling loop Preparatory bugfixes and tweaks to the loop before the next patch: - disable dispatch optimization during aio_prepare. This fixes a bug. - do not modify "blocking" until after the first WaitForMultipleObjects call. This is needed in the next patch. - change the loop to do...while. This makes it obvious that the loop is always entered at least once. In the next patch this is important because the first iteration undoes the ctx->notify_me increment that happened before entering the loop. Signed-off-by: Paolo Bonzini Reviewed-by: Fam Zheng Tested-by: Richard W.M. Jones Message-id: 1437487673-23740-4-git-send-email-pbonzini@redhat.com Signed-off-by: Stefan Hajnoczi --- aio-win32.c | 21 ++++++++++++--------- 1 file changed, 12 insertions(+), 9 deletions(-) diff --git a/aio-win32.c b/aio-win32.c index 233d8f5d79..9268b5c9fc 100644 --- a/aio-win32.c +++ b/aio-win32.c @@ -284,11 +284,6 @@ bool aio_poll(AioContext *ctx, bool blocking) int timeout; aio_context_acquire(ctx); - have_select_revents = aio_prepare(ctx); - if (have_select_revents) { - blocking = false; - } - was_dispatching = ctx->dispatching; progress = false; @@ -304,6 +299,8 @@ bool aio_poll(AioContext *ctx, bool blocking) */ aio_set_dispatching(ctx, !blocking); + have_select_revents = aio_prepare(ctx); + ctx->walking_handlers++; /* fill fd sets */ @@ -317,12 +314,18 @@ bool aio_poll(AioContext *ctx, bool blocking) ctx->walking_handlers--; first = true; - /* wait until next event */ - while (count > 0) { + /* ctx->notifier is always registered. */ + assert(count > 0); + + /* Multiple iterations, all of them non-blocking except the first, + * may be necessary to process all pending events. After the first + * WaitForMultipleObjects call ctx->notify_me will be decremented. + */ + do { HANDLE event; int ret; - timeout = blocking + timeout = blocking && !have_select_revents ? qemu_timeout_ns_to_ms(aio_compute_timeout(ctx)) : 0; if (timeout) { aio_context_release(ctx); @@ -351,7 +354,7 @@ bool aio_poll(AioContext *ctx, bool blocking) blocking = false; progress |= aio_dispatch_handlers(ctx, event); - } + } while (count > 0); progress |= timerlistgroup_run_timers(&ctx->tlg); From eabc977973103527bbb8fed69c91cfaa6691f8ab Mon Sep 17 00:00:00 2001 From: Paolo Bonzini Date: Tue, 21 Jul 2015 16:07:51 +0200 Subject: [PATCH 5/7] 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 -------------------------------------------------------------------------- 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 Reported-by: Laszlo Ersek Signed-off-by: Paolo Bonzini Reviewed-by: Fam Zheng Tested-by: Richard W.M. Jones Message-id: 1437487673-23740-5-git-send-email-pbonzini@redhat.com Signed-off-by: Stefan Hajnoczi --- aio-posix.c | 18 +++++----- aio-win32.c | 20 +++++------ async.c | 21 ++++------- docs/aio_notify.promela | 77 ++++++++++++++++++----------------------- include/block/aio.h | 29 ++++++++++++---- tests/test-aio.c | 18 ---------- 6 files changed, 81 insertions(+), 102 deletions(-) diff --git a/aio-posix.c b/aio-posix.c index 4abec38866..249889f14c 100644 --- a/aio-posix.c +++ b/aio-posix.c @@ -233,26 +233,23 @@ static void add_pollfd(AioHandler *node) bool aio_poll(AioContext *ctx, bool blocking) { AioHandler *node; - bool was_dispatching; int i, ret; bool progress; int64_t timeout; aio_context_acquire(ctx); - was_dispatching = ctx->dispatching; 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; - * if blocking == true, it is only true after poll() returns. - * - * If we're in a nested event loop, ctx->dispatching might be true. - * In that case we can restore it just before returning, but we - * have to clear it now. + * if blocking == true, it is only true after poll() returns, + * so disable the optimization now. */ - aio_set_dispatching(ctx, !blocking); + if (blocking) { + atomic_add(&ctx->notify_me, 2); + } ctx->walking_handlers++; @@ -272,6 +269,9 @@ bool aio_poll(AioContext *ctx, bool blocking) aio_context_release(ctx); } ret = qemu_poll_ns((GPollFD *)pollfds, npfd, timeout); + if (blocking) { + atomic_sub(&ctx->notify_me, 2); + } if (timeout) { aio_context_acquire(ctx); } @@ -287,12 +287,10 @@ bool aio_poll(AioContext *ctx, bool blocking) ctx->walking_handlers--; /* Run dispatch even if there were no readable fds to run timers */ - aio_set_dispatching(ctx, true); if (aio_dispatch(ctx)) { progress = true; } - aio_set_dispatching(ctx, was_dispatching); aio_context_release(ctx); return progress; diff --git a/aio-win32.c b/aio-win32.c index 9268b5c9fc..ea655b0935 100644 --- a/aio-win32.c +++ b/aio-win32.c @@ -279,25 +279,23 @@ bool aio_poll(AioContext *ctx, bool blocking) { AioHandler *node; HANDLE events[MAXIMUM_WAIT_OBJECTS + 1]; - bool was_dispatching, progress, have_select_revents, first; + bool progress, have_select_revents, first; int count; int timeout; aio_context_acquire(ctx); - was_dispatching = ctx->dispatching; 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; - * if blocking == true, it is only true after poll() returns. - * - * If we're in a nested event loop, ctx->dispatching might be true. - * In that case we can restore it just before returning, but we - * have to clear it now. + * if blocking == true, it is only true after poll() returns, + * so disable the optimization now. */ - aio_set_dispatching(ctx, !blocking); + if (blocking) { + atomic_add(&ctx->notify_me, 2); + } have_select_revents = aio_prepare(ctx); @@ -331,10 +329,13 @@ bool aio_poll(AioContext *ctx, bool blocking) aio_context_release(ctx); } ret = WaitForMultipleObjects(count, events, FALSE, timeout); + if (blocking) { + assert(first); + atomic_sub(&ctx->notify_me, 2); + } if (timeout) { aio_context_acquire(ctx); } - aio_set_dispatching(ctx, true); if (first && aio_bh_poll(ctx)) { progress = true; @@ -358,7 +359,6 @@ bool aio_poll(AioContext *ctx, bool blocking) progress |= timerlistgroup_run_timers(&ctx->tlg); - aio_set_dispatching(ctx, was_dispatching); aio_context_release(ctx); return progress; } diff --git a/async.c b/async.c index 77d080d6f5..a232192148 100644 --- a/async.c +++ b/async.c @@ -184,6 +184,8 @@ aio_ctx_prepare(GSource *source, gint *timeout) { AioContext *ctx = (AioContext *) source; + atomic_or(&ctx->notify_me, 1); + /* We assume there is no timeout already supplied */ *timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx)); @@ -200,6 +202,7 @@ aio_ctx_check(GSource *source) AioContext *ctx = (AioContext *) source; QEMUBH *bh; + atomic_and(&ctx->notify_me, ~1); for (bh = ctx->first_bh; bh; bh = bh->next) { if (!bh->deleted && bh->scheduled) { return true; @@ -254,23 +257,13 @@ ThreadPool *aio_get_thread_pool(AioContext *ctx) return ctx->thread_pool; } -void aio_set_dispatching(AioContext *ctx, bool dispatching) -{ - ctx->dispatching = dispatching; - if (!dispatching) { - /* Write ctx->dispatching before reading e.g. bh->scheduled. - * Optimization: this is only needed when we're entering the "unsafe" - * phase where other threads must call event_notifier_set. - */ - smp_mb(); - } -} - void aio_notify(AioContext *ctx) { - /* Write e.g. bh->scheduled before reading ctx->dispatching. */ + /* 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(); - if (!ctx->dispatching) { + if (ctx->notify_me) { event_notifier_set(&ctx->notifier); } } diff --git a/docs/aio_notify.promela b/docs/aio_notify.promela index ad3f6f08b0..fccc7ee1c3 100644 --- a/docs/aio_notify.promela +++ b/docs/aio_notify.promela @@ -1,5 +1,5 @@ /* - * This model describes the interaction between aio_set_dispatching() + * This model describes the interaction between ctx->notify_me * and aio_notify(). * * Author: Paolo Bonzini @@ -14,57 +14,53 @@ * spin -a docs/aio_notify.promela * gcc -O2 pan.c * ./a.out -a + * + * To verify it (with a bug planted in the model): + * spin -a -DBUG docs/aio_notify.promela + * gcc -O2 pan.c + * ./a.out -a */ #define MAX 4 #define LAST (1 << (MAX - 1)) #define FINAL ((LAST << 1) - 1) -bool dispatching; +bool notify_me; bool event; -int req, done; +int req; +int done; active proctype waiter() { - int fetch, blocking; + int fetch; - do - :: done != FINAL -> { - // Computing "blocking" is separate from execution of the - // "bottom half" - blocking = (req == 0); + do + :: true -> { + notify_me++; + + if +#ifndef BUG + :: (req > 0) -> skip; +#endif + :: else -> + // Wait for a nudge from the other side + do + :: event == 1 -> { event = 0; break; } + od; + fi; + + notify_me--; - // This is our "bottom half" atomic { fetch = req; req = 0; } done = done | fetch; - - // Wait for a nudge from the other side - do - :: event == 1 -> { event = 0; break; } - :: !blocking -> break; - od; - - dispatching = 1; - - // If you are simulating this model, you may want to add - // something like this here: - // - // int foo; foo++; foo++; foo++; - // - // This only wastes some time and makes it more likely - // that the notifier process hits the "fast path". - - dispatching = 0; } - :: else -> break; od } active proctype notifier() { int next = 1; - int sets = 0; do :: next <= LAST -> { @@ -74,8 +70,8 @@ active proctype notifier() // aio_notify if - :: dispatching == 0 -> sets++; event = 1; - :: else -> skip; + :: notify_me == 1 -> event = 1; + :: else -> printf("Skipped event_notifier_set\n"); skip; fi; // Test both synchronous and asynchronous delivery @@ -86,19 +82,12 @@ active proctype notifier() :: 1 -> skip; fi; } - :: else -> break; od; - printf("Skipped %d event_notifier_set\n", MAX - sets); } -#define p (done == FINAL) - -never { - do - :: 1 // after an arbitrarily long prefix - :: p -> break // p becomes true - od; - do - :: !p -> accept: break // it then must remains true forever after - od +never { /* [] done < FINAL */ +accept_init: + do + :: done < FINAL -> skip; + od; } diff --git a/include/block/aio.h b/include/block/aio.h index b46103ece7..be91e3f701 100644 --- a/include/block/aio.h +++ b/include/block/aio.h @@ -63,10 +63,30 @@ struct AioContext { */ int walking_handlers; - /* Used to avoid unnecessary event_notifier_set calls in aio_notify. - * Writes protected by lock or BQL, reads are lockless. + /* Used to avoid unnecessary event_notifier_set calls in aio_notify; + * accessed with atomic primitives. If this field is 0, everything + * (file descriptors, bottom halves, timers) will be re-evaluated + * before the next blocking poll(), thus the event_notifier_set call + * can be skipped. If it is non-zero, you may need to wake up a + * concurrent aio_poll or the glib main event loop, making + * event_notifier_set necessary. + * + * Bit 0 is reserved for GSource usage of the AioContext, and is 1 + * between a call to aio_ctx_check and the next call to aio_ctx_dispatch. + * Bits 1-31 simply count the number of active calls to aio_poll + * that are in the prepare or poll phase. + * + * The GSource and aio_poll must use a different mechanism because + * there is no certainty that a call to GSource's prepare callback + * (via g_main_context_prepare) is indeed followed by check and + * dispatch. It's not clear whether this would be a bug, but let's + * play safe and allow it---it will just cause extra calls to + * event_notifier_set until the next call to dispatch. + * + * Instead, the aio_poll calls include both the prepare and the + * dispatch phase, hence a simple counter is enough for them. */ - bool dispatching; + uint32_t notify_me; /* lock to protect between bh's adders and deleter */ QemuMutex bh_lock; @@ -89,9 +109,6 @@ struct AioContext { QEMUTimerListGroup tlg; }; -/* Used internally to synchronize aio_poll against qemu_bh_schedule. */ -void aio_set_dispatching(AioContext *ctx, bool dispatching); - /** * aio_context_new: Allocate a new AioContext. * diff --git a/tests/test-aio.c b/tests/test-aio.c index e7bbb8345a..217e33772e 100644 --- a/tests/test-aio.c +++ b/tests/test-aio.c @@ -97,14 +97,6 @@ static void event_ready_cb(EventNotifier *e) /* Tests using aio_*. */ -static void test_notify(void) -{ - g_assert(!aio_poll(ctx, false)); - aio_notify(ctx); - g_assert(!aio_poll(ctx, true)); - g_assert(!aio_poll(ctx, false)); -} - typedef struct { QemuMutex start_lock; bool thread_acquired; @@ -494,14 +486,6 @@ static void test_timer_schedule(void) * works well, and that's what I am using. */ -static void test_source_notify(void) -{ - while (g_main_context_iteration(NULL, false)); - aio_notify(ctx); - g_assert(g_main_context_iteration(NULL, true)); - g_assert(!g_main_context_iteration(NULL, false)); -} - static void test_source_flush(void) { g_assert(!g_main_context_iteration(NULL, false)); @@ -830,7 +814,6 @@ int main(int argc, char **argv) while (g_main_context_iteration(NULL, false)); g_test_init(&argc, &argv, NULL); - g_test_add_func("/aio/notify", test_notify); g_test_add_func("/aio/acquire", test_acquire); g_test_add_func("/aio/bh/schedule", test_bh_schedule); g_test_add_func("/aio/bh/schedule10", test_bh_schedule10); @@ -845,7 +828,6 @@ int main(int argc, char **argv) g_test_add_func("/aio/event/flush", test_flush_event_notifier); g_test_add_func("/aio/timer/schedule", test_timer_schedule); - g_test_add_func("/aio-gsource/notify", test_source_notify); g_test_add_func("/aio-gsource/flush", test_source_flush); g_test_add_func("/aio-gsource/bh/schedule", test_source_bh_schedule); g_test_add_func("/aio-gsource/bh/schedule10", test_source_bh_schedule10); From 21a03d17f2edb1e63f7137d97ba355cc6f19d79f Mon Sep 17 00:00:00 2001 From: Paolo Bonzini Date: Tue, 21 Jul 2015 16:07:52 +0200 Subject: [PATCH 6/7] 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 Signed-off-by: Paolo Bonzini Reviewed-by: Fam Zheng Tested-by: Richard W.M. Jones Message-id: 1437487673-23740-6-git-send-email-pbonzini@redhat.com Signed-off-by: Stefan Hajnoczi --- aio-posix.c | 2 + aio-win32.c | 7 +- async.c | 8 ++- docs/aio_notify_bug.promela | 140 ++++++++++++++++++++++++++++++++++++ 4 files changed, 153 insertions(+), 4 deletions(-) create mode 100644 docs/aio_notify_bug.promela diff --git a/aio-posix.c b/aio-posix.c index 249889f14c..5c8b266c72 100644 --- a/aio-posix.c +++ b/aio-posix.c @@ -276,6 +276,8 @@ bool aio_poll(AioContext *ctx, bool blocking) aio_context_acquire(ctx); } + event_notifier_test_and_clear(&ctx->notifier); + /* if we have any readable fds, dispatch event */ if (ret > 0) { for (i = 0; i < npfd; i++) { diff --git a/aio-win32.c b/aio-win32.c index ea655b0935..7afc9992d6 100644 --- a/aio-win32.c +++ b/aio-win32.c @@ -337,10 +337,11 @@ bool aio_poll(AioContext *ctx, bool blocking) aio_context_acquire(ctx); } - if (first && aio_bh_poll(ctx)) { - progress = true; + if (first) { + event_notifier_test_and_clear(&ctx->notifier); + progress |= aio_bh_poll(ctx); + first = false; } - first = false; /* if we have any signaled events, dispatch event */ event = NULL; diff --git a/async.c b/async.c index a232192148..d625e8a803 100644 --- a/async.c +++ b/async.c @@ -203,6 +203,8 @@ aio_ctx_check(GSource *source) QEMUBH *bh; atomic_and(&ctx->notify_me, ~1); + event_notifier_test_and_clear(&ctx->notifier); + for (bh = ctx->first_bh; bh; bh = bh->next) { if (!bh->deleted && bh->scheduled) { return true; @@ -279,6 +281,10 @@ static void aio_rfifolock_cb(void *opaque) aio_notify(opaque); } +static void event_notifier_dummy_cb(EventNotifier *e) +{ +} + AioContext *aio_context_new(Error **errp) { int ret; @@ -293,7 +299,7 @@ AioContext *aio_context_new(Error **errp) g_source_set_can_recurse(&ctx->source, true); aio_set_event_notifier(ctx, &ctx->notifier, (EventNotifierHandler *) - event_notifier_test_and_clear); + event_notifier_dummy_cb); ctx->thread_pool = NULL; qemu_mutex_init(&ctx->bh_lock); rfifolock_init(&ctx->lock, aio_rfifolock_cb, ctx); diff --git a/docs/aio_notify_bug.promela b/docs/aio_notify_bug.promela new file mode 100644 index 0000000000..b3bfca1ca4 --- /dev/null +++ b/docs/aio_notify_bug.promela @@ -0,0 +1,140 @@ +/* + * This model describes a bug in aio_notify. If ctx->notifier is + * cleared too late, a wakeup could be lost. + * + * Author: Paolo Bonzini + * + * This file is in the public domain. If you really want a license, + * the WTFPL will do. + * + * To verify the buggy version: + * spin -a -DBUG docs/aio_notify_bug.promela + * gcc -O2 pan.c + * ./a.out -a -f + * + * To verify the fixed version: + * spin -a docs/aio_notify_bug.promela + * gcc -O2 pan.c + * ./a.out -a -f + * + * Add -DCHECK_REQ to test an alternative invariant and the + * "notify_me" optimization. + */ + +int notify_me; +bool event; +bool req; +bool notifier_done; + +#ifdef CHECK_REQ +#define USE_NOTIFY_ME 1 +#else +#define USE_NOTIFY_ME 0 +#endif + +active proctype notifier() +{ + do + :: true -> { + req = 1; + if + :: !USE_NOTIFY_ME || notify_me -> event = 1; + :: else -> skip; + fi + } + :: true -> break; + od; + notifier_done = 1; +} + +#ifdef BUG +#define AIO_POLL \ + notify_me++; \ + if \ + :: !req -> { \ + if \ + :: event -> skip; \ + fi; \ + } \ + :: else -> skip; \ + fi; \ + notify_me--; \ + \ + req = 0; \ + event = 0; +#else +#define AIO_POLL \ + notify_me++; \ + if \ + :: !req -> { \ + if \ + :: event -> skip; \ + fi; \ + } \ + :: else -> skip; \ + fi; \ + notify_me--; \ + \ + event = 0; \ + req = 0; +#endif + +active proctype waiter() +{ + do + :: true -> AIO_POLL; + od; +} + +/* Same as waiter(), but disappears after a while. */ +active proctype temporary_waiter() +{ + do + :: true -> AIO_POLL; + :: true -> break; + od; +} + +#ifdef CHECK_REQ +never { + do + :: req -> goto accept_if_req_not_eventually_false; + :: true -> skip; + od; + +accept_if_req_not_eventually_false: + if + :: req -> goto accept_if_req_not_eventually_false; + fi; + assert(0); +} + +#else +/* There must be infinitely many transitions of event as long + * as the notifier does not exit. + * + * If event stayed always true, the waiters would be busy looping. + * If event stayed always false, the waiters would be sleeping + * forever. + */ +never { + do + :: !event -> goto accept_if_event_not_eventually_true; + :: event -> goto accept_if_event_not_eventually_false; + :: true -> skip; + od; + +accept_if_event_not_eventually_true: + if + :: !event && notifier_done -> do :: true -> skip; od; + :: !event && !notifier_done -> goto accept_if_event_not_eventually_true; + fi; + assert(0); + +accept_if_event_not_eventually_false: + if + :: event -> goto accept_if_event_not_eventually_false; + fi; + assert(0); +} +#endif From 05e514b1d4d5bd4209e2c8bbc76ff05c85a235f3 Mon Sep 17 00:00:00 2001 From: Paolo Bonzini Date: Tue, 21 Jul 2015 16:07:53 +0200 Subject: [PATCH 7/7] AioContext: optimize clearing the EventNotifier It is pretty rare for aio_notify to actually set the EventNotifier. It can happen with worker threads such as thread-pool.c's, but otherwise it should never be set thanks to the ctx->notify_me optimization. The previous patch, unfortunately, added an unconditional call to event_notifier_test_and_clear; now add a userspace fast path that avoids the call. Note that it is not possible to do the same with event_notifier_set; it would break, as proved (again) by the included formal model. This patch survived over 3000 reboots on aarch64 KVM. Signed-off-by: Paolo Bonzini Reviewed-by: Fam Zheng Tested-by: Richard W.M. Jones Message-id: 1437487673-23740-7-git-send-email-pbonzini@redhat.com Signed-off-by: Stefan Hajnoczi --- aio-posix.c | 2 +- aio-win32.c | 2 +- async.c | 10 ++- docs/aio_notify_accept.promela | 152 +++++++++++++++++++++++++++++++++ include/block/aio.h | 32 ++++++- 5 files changed, 194 insertions(+), 4 deletions(-) create mode 100644 docs/aio_notify_accept.promela diff --git a/aio-posix.c b/aio-posix.c index 5c8b266c72..d4770336c5 100644 --- a/aio-posix.c +++ b/aio-posix.c @@ -276,7 +276,7 @@ bool aio_poll(AioContext *ctx, bool blocking) aio_context_acquire(ctx); } - event_notifier_test_and_clear(&ctx->notifier); + aio_notify_accept(ctx); /* if we have any readable fds, dispatch event */ if (ret > 0) { diff --git a/aio-win32.c b/aio-win32.c index 7afc9992d6..50a6867458 100644 --- a/aio-win32.c +++ b/aio-win32.c @@ -338,7 +338,7 @@ bool aio_poll(AioContext *ctx, bool blocking) } if (first) { - event_notifier_test_and_clear(&ctx->notifier); + aio_notify_accept(ctx); progress |= aio_bh_poll(ctx); first = false; } diff --git a/async.c b/async.c index d625e8a803..9a98a74acb 100644 --- a/async.c +++ b/async.c @@ -203,7 +203,7 @@ aio_ctx_check(GSource *source) QEMUBH *bh; atomic_and(&ctx->notify_me, ~1); - event_notifier_test_and_clear(&ctx->notifier); + aio_notify_accept(ctx); for (bh = ctx->first_bh; bh; bh = bh->next) { if (!bh->deleted && bh->scheduled) { @@ -267,6 +267,14 @@ void aio_notify(AioContext *ctx) smp_mb(); 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); } } diff --git a/docs/aio_notify_accept.promela b/docs/aio_notify_accept.promela new file mode 100644 index 0000000000..9cef2c955d --- /dev/null +++ b/docs/aio_notify_accept.promela @@ -0,0 +1,152 @@ +/* + * This model describes the interaction between ctx->notified + * and ctx->notifier. + * + * Author: Paolo Bonzini + * + * This file is in the public domain. If you really want a license, + * the WTFPL will do. + * + * To verify the buggy version: + * spin -a -DBUG1 docs/aio_notify_bug.promela + * gcc -O2 pan.c + * ./a.out -a -f + * (or -DBUG2) + * + * To verify the fixed version: + * spin -a docs/aio_notify_bug.promela + * gcc -O2 pan.c + * ./a.out -a -f + * + * Add -DCHECK_REQ to test an alternative invariant and the + * "notify_me" optimization. + */ + +int notify_me; +bool notified; +bool event; +bool req; +bool notifier_done; + +#ifdef CHECK_REQ +#define USE_NOTIFY_ME 1 +#else +#define USE_NOTIFY_ME 0 +#endif + +#ifdef BUG +#error Please define BUG1 or BUG2 instead. +#endif + +active proctype notifier() +{ + do + :: true -> { + req = 1; + if + :: !USE_NOTIFY_ME || notify_me -> +#if defined BUG1 + /* CHECK_REQ does not detect this bug! */ + notified = 1; + event = 1; +#elif defined BUG2 + if + :: !notified -> event = 1; + :: else -> skip; + fi; + notified = 1; +#else + event = 1; + notified = 1; +#endif + :: else -> skip; + fi + } + :: true -> break; + od; + notifier_done = 1; +} + +#define AIO_POLL \ + notify_me++; \ + if \ + :: !req -> { \ + if \ + :: event -> skip; \ + fi; \ + } \ + :: else -> skip; \ + fi; \ + notify_me--; \ + \ + atomic { old = notified; notified = 0; } \ + if \ + :: old -> event = 0; \ + :: else -> skip; \ + fi; \ + \ + req = 0; + +active proctype waiter() +{ + bool old; + + do + :: true -> AIO_POLL; + od; +} + +/* Same as waiter(), but disappears after a while. */ +active proctype temporary_waiter() +{ + bool old; + + do + :: true -> AIO_POLL; + :: true -> break; + od; +} + +#ifdef CHECK_REQ +never { + do + :: req -> goto accept_if_req_not_eventually_false; + :: true -> skip; + od; + +accept_if_req_not_eventually_false: + if + :: req -> goto accept_if_req_not_eventually_false; + fi; + assert(0); +} + +#else +/* There must be infinitely many transitions of event as long + * as the notifier does not exit. + * + * If event stayed always true, the waiters would be busy looping. + * If event stayed always false, the waiters would be sleeping + * forever. + */ +never { + do + :: !event -> goto accept_if_event_not_eventually_true; + :: event -> goto accept_if_event_not_eventually_false; + :: true -> skip; + od; + +accept_if_event_not_eventually_true: + if + :: !event && notifier_done -> do :: true -> skip; od; + :: !event && !notifier_done -> goto accept_if_event_not_eventually_true; + fi; + assert(0); + +accept_if_event_not_eventually_false: + if + :: event -> goto accept_if_event_not_eventually_false; + fi; + assert(0); +} +#endif diff --git a/include/block/aio.h b/include/block/aio.h index be91e3f701..9dd32e0f13 100644 --- a/include/block/aio.h +++ b/include/block/aio.h @@ -99,7 +99,19 @@ struct AioContext { */ int walking_bh; - /* Used for aio_notify. */ + /* Used by aio_notify. + * + * "notified" is used to avoid expensive event_notifier_test_and_clear + * calls. When it is clear, the EventNotifier is clear, or one thread + * is going to clear "notified" before processing more events. False + * positives are possible, i.e. "notified" could be set even though the + * EventNotifier is clear. + * + * Note that event_notifier_set *cannot* be optimized the same way. For + * more information on the problem that would result, see "#ifdef BUG2" + * in the docs/aio_notify_accept.promela formal model. + */ + bool notified; EventNotifier notifier; /* Thread pool for performing work and receiving completion callbacks */ @@ -173,6 +185,24 @@ QEMUBH *aio_bh_new(AioContext *ctx, QEMUBHFunc *cb, void *opaque); */ void aio_notify(AioContext *ctx); +/** + * aio_notify_accept: Acknowledge receiving an aio_notify. + * + * aio_notify() uses an EventNotifier in order to wake up a sleeping + * aio_poll() or g_main_context_iteration(). Calls to aio_notify() are + * usually rare, but the AioContext has to clear the EventNotifier on + * every aio_poll() or g_main_context_iteration() in order to avoid + * busy waiting. This event_notifier_test_and_clear() cannot be done + * using the usual aio_context_set_event_notifier(), because it must + * be done before processing all events (file descriptors, bottom halves, + * timers). + * + * aio_notify_accept() is an optimized event_notifier_test_and_clear() + * that is specific to an AioContext's notifier; it is used internally + * to clear the EventNotifier only if aio_notify() had been called. + */ +void aio_notify_accept(AioContext *ctx); + /** * aio_bh_poll: Poll bottom halves for an AioContext. *