aio_epoll_update needs the fields in node, so delay the free.
Reported-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
Message-id: 1447655534-13974-1-git-send-email-famz@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
To minimize code duplication, epoll is hooked into aio-posix's
aio_poll() instead of rolling its own. This approach also has both
compile-time and run-time switchability.
1) When QEMU starts with a small number of fds in the event loop, ppoll
is used.
2) When QEMU starts with a big number of fds, or when more devices are
hot plugged, epoll kicks in when the number of fds hits the threshold.
3) Some fds may not support epoll, such as tty based stdio. In this
case, it falls back to ppoll.
A rough benchmark with scsi-disk on virtio-scsi dataplane (epoll gets
enabled from 64 onward). Numbers are in MB/s.
===============================================
| master | epoll
| |
scsi disks # | read randrw | read randrw
-------------|----------------|----------------
1 | 86 36 | 92 45
8 | 87 43 | 86 41
64 | 71 32 | 70 38
128 | 48 24 | 58 31
256 | 37 19 | 57 28
===============================================
To comply with aio_{disable,enable}_external, we always use ppoll when
aio_external_disabled() is true.
[Removed #ifdef CONFIG_EPOLL around AioContext epollfd field declaration
since the field is also referenced outside CONFIG_EPOLL code.
--Stefan]
Signed-off-by: Fam Zheng <famz@redhat.com>
Message-id: 1446177989-6702-4-git-send-email-famz@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
This is the place to initialize platform specific bits of AioContext.
Signed-off-by: Fam Zheng <famz@redhat.com>
Message-id: 1446177989-6702-3-git-send-email-famz@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
All callers pass in false, and the real external ones will switch to
true in coming patches.
Signed-off-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Jeff Cody <jcody@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
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 <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Tested-by: Richard W.M. Jones <rjones@redhat.com>
Message-id: 1437487673-23740-7-git-send-email-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
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>
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>
This is the first step in pushing down acquire/release, and will let
rfifolock drop the contention callback feature.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1424449612-18215-3-git-send-email-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
By using thread-local storage, aio_poll can stop using global data during
g_poll_ns. This will make it possible to drop callbacks from rfifolock.
[Moved npfd = 0 assignment to end of walking_handlers region as
suggested by Paolo. This resolves the assert(npfd == 0) assertion
failure in pollfds_cleanup().
--Stefan]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 1424449612-18215-2-git-send-email-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
g_new(T, 1) is safer than g_malloc(sizeof(T)), because it returns T *
rather than void *, which lets the compiler catch more type errors.
Missed in commit 02c4f26.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-id: 1417697709-13087-1-git-send-email-armbru@redhat.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This will be used to implement socket polling on Windows.
On Windows, select() and g_poll() are completely different;
sockets are polled with select() before calling g_poll,
and the g_poll must be nonblocking if select() says a
socket is ready.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
So far, aio_poll's scheme was dispatch/poll/dispatch, where
the first dispatch phase was used only in the GSource case in
order to avoid a blocking poll. Earlier patches changed it to
dispatch/prepare/poll/dispatch, where prepare is aio_compute_timeout.
By making aio_dispatch public, we can remove the first dispatch
phase altogether, so that both aio_poll and the GSource use the same
prepare/poll/dispatch scheme.
This patch breaks the invariant that aio_poll(..., true) will not block
the first time it returns false. This used to be fundamental for
qemu_aio_flush's implementation as "while (qemu_aio_wait()) {}" but
no code in QEMU relies on this invariant anymore. The return value
of aio_poll() is now comparable with that of g_main_context_iteration.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Make the dispatching phase the same before blocking and afterwards.
The next patch will make aio_dispatch public and use it directly
for the GSource case, instead of aio_poll. aio_poll can then be
simplified heavily.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Right now, QEMU invokes aio_bh_poll before the "poll" phase
of aio_poll. It is simpler to do it afterwards and skip the
"poll" phase altogether when the OS-dependent parts of AioContext
are invoked from GSource. This way, AioContext behaves more
similarly when used as a GSource vs. when used as stand-alone.
As a start, take bottom halves into account when computing the
poll timeout. If a bottom half is ready, do a non-blocking
poll. As a side effect, this makes idle bottom halves work
with aio_poll; an improvement, but not really an important
one since they are deprecated.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
In many cases, the call to event_notifier_set in aio_notify is unnecessary.
In particular, if we are executing aio_dispatch, or if aio_poll is not
blocking, we know that we will soon get to the next loop iteration (if
necessary); the thread that hosts the AioContext's event loop does not
need any nudging.
The patch includes a Promela formal model that shows that this really
works and does not need any further complication such as generation
counts. It needs a memory barrier though.
The generation counts are not needed because any change to
ctx->dispatching after the memory barrier is okay for aio_notify.
If it changes from zero to one, it is the right thing to skip
event_notifier_set. If it changes from one to zero, the
event_notifier_set is unnecessary but harmless.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The main AioContext should be accessed explicitly via qemu_get_aio_context().
Most of the time, using it is not the right thing to do.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This patch drops a special case where aio_poll(ctx, true) returns false
instead of blocking if no file descriptors are waiting on I/O. Now it
is possible to block in aio_poll() to wait for aio_notify().
This change eliminates busy waiting. bdrv_drain_all() used to rely on
busy waiting to completed throttled I/O requests but this is no longer
required so we can simplify aio_poll().
Note that aio_poll() still returns false when aio_notify() was used. In
other words, stopping a blocking aio_poll() wait is not considered
making progress.
Adjust test-aio /aio/bh/callback-delete/one which assumed aio_poll(ctx,
true) would immediately return false instead of blocking.
Reviewed-by: Alex Bligh <alex@alex.org.uk>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Convert aio_poll to use deadline based on AioContext's timers.
aio_poll has been changed to return accurately whether progress
has occurred. Prior to this commit, aio_poll always returned
true if g_poll was entered, whether or not any progress was
made. This required a change to tests/test-aio.c where an
assert was backwards.
Signed-off-by: Alex Bligh <alex@alex.org.uk>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The .io_flush() handler no longer exists and has no users. Drop the
io_flush argument to aio_set_fd_handler() and related functions.
The AioFlushEventNotifierHandler and AioFlushHandler typedefs are no
longer used and are dropped too.
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
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>
aio-posix.c could not take advantage of G_IO_HUP and G_IO_ERR because
select(2) does not have equivalent events. Now that g_poll(3) is used
we can support G_IO_HUP and G_IO_ERR.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Message-id: 1361356113-11049-11-git-send-email-stefanha@redhat.com
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
AioHandler already has a GPollFD so we can directly use its
events/revents.
Add the int pollfds_idx field to AioContext so we can map g_poll(3)
results back to AioHandlers.
Reuse aio_dispatch() to invoke handlers after g_poll(3).
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Message-id: 1361356113-11049-10-git-send-email-stefanha@redhat.com
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
We will need to loop over AioHandlers calling ->io_read()/->io_write()
when aio_poll() is converted from select(2) to g_poll(2).
Luckily the code for this already exists, extract it into the new
aio_dispatch() function.
Two small changes:
* aio_poll() checks !node->deleted to avoid calling handlers that have
been deleted.
* Fix typo 'then' -> 'them' in aio_poll() comment.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Message-id: 1361356113-11049-9-git-send-email-stefanha@redhat.com
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
aio_poll() must return true if any work is still pending, even if it
didn't make progress, so that bdrv_drain_all() doesn't stop waiting too
early. The possibility of stopping early occasionally lead to a failed
assertion in bdrv_drain_all(), when some in-flight request was missed
and the function didn't really drain all requests.
In order to make that change, the return value as specified in the
function comment must change for blocking = false; fortunately, the
return value of blocking = false callers is only used in test cases, so
this change shouldn't cause any trouble.
Cc: qemu-stable@nongnu.org
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
In the current code, this is done by qemu_set_fd_handler2, which is
called by qemu_aio_set_fd_handler. We need to keep the same behavior
even after removing the call to qemu_set_fd_handler2.
Reviewed-by: Anthony Liguori <aliguori@us.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The Win32 implementation will only accept EventNotifiers, thus a few
drivers are disabled under Windows. EventNotifiers are a good match
for the GSource implementation, too, because the Win32 port of glib
allows to place their HANDLEs in a GPollFD.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Fam Zheng