When we measure FIO read performance (cache=writethrough, bs=4k,
iodepth=64) in VMs, ~80K/s notifications (e.g., EPT_MISCONFIG) are observed
from guest to qemu.
It turns out those frequent notificatons are caused by interference from
worker threads. Worker threads queue bottom halves after completing IO
requests. Pending bottom halves may lead to either aio_compute_timeout()
zeros timeout and pass it to try_poll_mode() or run_poll_handlers() returns
no progress after noticing pending aio_notify() events. Both cause
run_poll_handlers() to call poll_set_started(false) to disable poll mode.
However, for both cases, as timeout is already zeroed, the event loop
(i.e., aio_poll()) just processes bottom halves and then starts the next
event loop iteration. So, disabling poll mode has no value but leads to
unnecessary notifications from guest.
To minimize unnecessary notifications from guest, defer disabling poll
mode to when the event loop is about to be blocked.
With this patch applied, FIO seq-read performance (bs=4k, iodepth=64,
cache=writethrough) in VMs increases from 330K/s to 413K/s IOPS.
Suggested-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Chao Gao <chao.gao@intel.com>
Message-id: 20220710120849.63086-1-chao.gao@intel.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The thread pool regulates itself: when idle, it kills threads until
empty, when in demand, it creates new threads until full. This behaviour
doesn't play well with latency sensitive workloads where the price of
creating a new thread is too high. For example, when paired with qemu's
'-mlock', or using safety features like SafeStack, creating a new thread
has been measured take multiple milliseconds.
In order to mitigate this let's introduce a new 'EventLoopBase'
property to set the thread pool size. The threads will be created during
the pool's initialization or upon updating the property's value, remain
available during its lifetime regardless of demand, and destroyed upon
freeing it. A properly characterized workload will then be able to
configure the pool to avoid any latency spikes.
Signed-off-by: Nicolas Saenz Julienne <nsaenzju@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Message-id: 20220425075723.20019-4-nsaenzju@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
When ->poll() succeeds the AioHandler is placed on the ready list with
revents set to the magic value 0. This magic value causes
aio_dispatch_handler() to invoke ->poll_ready() instead of ->io_read()
for G_IO_IN or ->io_write() for G_IO_OUT.
This magic value 0 hack works for the IOThread where AioHandlers are
placed on ->ready_list and processed by aio_dispatch_ready_handlers().
It does not work for the main loop where all AioHandlers are processed
by aio_dispatch_handlers(), even those that are not ready and have a
revents value of 0.
As a result the main loop invokes ->poll_ready() on AioHandlers that are
not ready. These spurious ->poll_ready() calls waste CPU cycles and
could lead to crashes if the code assumes ->poll() must have succeeded
before ->poll_ready() is called (a reasonable asumption but I haven't
seen it in practice).
Stop using revents to track whether ->poll_ready() will be called on an
AioHandler. Introduce a separate AioHandler->poll_ready field instead.
This eliminates spurious ->poll_ready() calls in the main loop.
Fixes: 826cc32423 ("aio-posix: split poll check from ready handler")
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reported-by: Jason Wang <jasowang@redhat.com>
Tested-by: Jason Wang <jasowang@redhat.com>
Message-id: 20220223155703.136833-1-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Adaptive polling measures the execution time of the polling check plus
handlers called when a polled event becomes ready. Handlers can take a
significant amount of time, making it look like polling was running for
a long time when in fact the event handler was running for a long time.
For example, on Linux the io_submit(2) syscall invoked when a virtio-blk
device's virtqueue becomes ready can take 10s of microseconds. This
can exceed the default polling interval (32 microseconds) and cause
adaptive polling to stop polling.
By excluding the handler's execution time from the polling check we make
the adaptive polling calculation more accurate. As a result, the event
loop now stays in polling mode where previously it would have fallen
back to file descriptor monitoring.
The following data was collected with virtio-blk num-queues=2
event_idx=off using an IOThread. Before:
168k IOPS, IOThread syscalls:
9837.115 ( 0.020 ms): IO iothread1/620155 io_submit(ctx_id: 140512552468480, nr: 16, iocbpp: 0x7fcb9f937db0) = 16
9837.158 ( 0.002 ms): IO iothread1/620155 write(fd: 103, buf: 0x556a2ef71b88, count: 8) = 8
9837.161 ( 0.001 ms): IO iothread1/620155 write(fd: 104, buf: 0x556a2ef71b88, count: 8) = 8
9837.163 ( 0.001 ms): IO iothread1/620155 ppoll(ufds: 0x7fcb90002800, nfds: 4, tsp: 0x7fcb9f1342d0, sigsetsize: 8) = 3
9837.164 ( 0.001 ms): IO iothread1/620155 read(fd: 107, buf: 0x7fcb9f939cc0, count: 512) = 8
9837.174 ( 0.001 ms): IO iothread1/620155 read(fd: 105, buf: 0x7fcb9f939cc0, count: 512) = 8
9837.176 ( 0.001 ms): IO iothread1/620155 read(fd: 106, buf: 0x7fcb9f939cc0, count: 512) = 8
9837.209 ( 0.035 ms): IO iothread1/620155 io_submit(ctx_id: 140512552468480, nr: 32, iocbpp: 0x7fca7d0cebe0) = 32
174k IOPS (+3.6%), IOThread syscalls:
9809.566 ( 0.036 ms): IO iothread1/623061 io_submit(ctx_id: 140539805028352, nr: 32, iocbpp: 0x7fd0cdd62be0) = 32
9809.625 ( 0.001 ms): IO iothread1/623061 write(fd: 103, buf: 0x5647cfba5f58, count: 8) = 8
9809.627 ( 0.002 ms): IO iothread1/623061 write(fd: 104, buf: 0x5647cfba5f58, count: 8) = 8
9809.663 ( 0.036 ms): IO iothread1/623061 io_submit(ctx_id: 140539805028352, nr: 32, iocbpp: 0x7fd0d0388b50) = 32
Notice that ppoll(2) and eventfd read(2) syscalls are eliminated because
the IOThread stays in polling mode instead of falling back to file
descriptor monitoring.
As usual, polling is not implemented on Windows so this patch ignores
the new io_poll_read() callback in aio-win32.c.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
Message-id: 20211207132336.36627-2-stefanha@redhat.com
[Fixed up aio_set_event_notifier() calls in
tests/unit/test-fdmon-epoll.c added after this series was queued.
--Stefan]
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The `aio-max-batch` parameter will be propagated to AIO engines
and it will be used to control the maximum number of queued requests.
When there are in queue a number of requests equal to `aio-max-batch`,
the engine invokes the system call to forward the requests to the kernel.
This parameter allows us to control the maximum batch size to reduce
the latency that requests might accumulate while queued in the AIO
engine queue.
If `aio-max-batch` is equal to 0 (default value), the AIO engine will
use its default maximum batch size value.
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Message-id: 20210721094211.69853-3-sgarzare@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
This moves the QMP dispatcher to a coroutine and runs all QMP command
handlers that declare 'coroutine': true in coroutine context so they
can avoid blocking the main loop while doing I/O or waiting for other
events.
For commands that are not declared safe to run in a coroutine, the
dispatcher drops out of coroutine context by calling the QMP command
handler from a bottom half.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20201005155855.256490-10-kwolf@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
clang's C11 atomic_fetch_*() functions only take a C11 atomic type
pointer argument. QEMU uses direct types (int, etc) and this causes a
compiler error when a QEMU code calls these functions in a source file
that also included <stdatomic.h> via a system header file:
$ CC=clang CXX=clang++ ./configure ... && make
../util/async.c:79:17: error: address argument to atomic operation must be a pointer to _Atomic type ('unsigned int *' invalid)
Avoid using atomic_*() names in QEMU's atomic.h since that namespace is
used by <stdatomic.h>. Prefix QEMU's APIs with 'q' so that atomic.h
and <stdatomic.h> can co-exist. I checked /usr/include on my machine and
searched GitHub for existing "qatomic_" users but there seem to be none.
This patch was generated using:
$ git grep -h -o '\<atomic\(64\)\?_[a-z0-9_]\+' include/qemu/atomic.h | \
sort -u >/tmp/changed_identifiers
$ for identifier in $(</tmp/changed_identifiers); do
sed -i "s%\<$identifier\>%q$identifier%g" \
$(git grep -I -l "\<$identifier\>")
done
I manually fixed line-wrap issues and misaligned rST tables.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Acked-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20200923105646.47864-1-stefanha@redhat.com>
Polling only monitors the ctx->notified field and does not need the
ctx->notifier EventNotifier to be signalled. Keep ctx->aio_notify_me
disabled while polling to avoid unnecessary EventNotifier syscalls.
This optimization improves virtio-blk 4KB random read performance by
18%. The following results are with an IOThread and the null-co block
driver:
Test IOPS Error
Before 244518.62 ± 1.20%
After 290706.11 ± 0.44%
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Message-id: 20200806131802.569478-4-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The glib event loop does not call fdmon_io_uring_wait() so fd handlers
waiting to be submitted build up in the list. There is no benefit is
using io_uring when the glib GSource is being used, so disable it
instead of implementing a more complex fix.
This fixes a memory leak where AioHandlers would build up and increasing
amounts of CPU time were spent iterating them in aio_pending(). The
symptom is that guests become slow when QEMU is built with io_uring
support.
Buglink: https://bugs.launchpad.net/qemu/+bug/1877716
Fixes: 73fd282e7b ("aio-posix: add io_uring fd monitoring implementation")
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Tested-by: Oleksandr Natalenko <oleksandr@redhat.com>
Message-id: 20200511183630.279750-3-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The io_uring file descriptor monitoring implementation has an internal
list of fd handlers that are pending submission to io_uring.
fdmon_io_uring_destroy() deletes all fd handlers on the list.
Don't delete fd handlers directly in fdmon_io_uring_destroy() for two
reasons:
1. This duplicates the aio-posix.c AioHandler deletion code and could
become outdated if the struct changes.
2. Only handlers with the FDMON_IO_URING_REMOVE flag set are safe to
remove. If the flag is not set then something still has a pointer to
the fd handler. Let aio-posix.c and its user worry about that. In
practice this isn't an issue because fdmon_io_uring_destroy() is only
called when shutting down so all users have removed their fd
handlers, but the next patch will need this!
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Tested-by: Oleksandr Natalenko <oleksandr@redhat.com>
Message-id: 20200511183630.279750-2-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
When using C11 atomics, non-seqcst reads and writes do not participate
in the total order of seqcst operations. In util/async.c and util/aio-posix.c,
in particular, the pattern that we use
write ctx->notify_me write bh->scheduled
read bh->scheduled read ctx->notify_me
if !bh->scheduled, sleep if ctx->notify_me, notify
needs to use seqcst operations for both the write and the read. In
general this is something that we do not want, because there can be
many sources that are polled in addition to bottom halves. The
alternative is to place a seqcst memory barrier between the write
and the read. This also comes with a disadvantage, in that the
memory barrier is implicit on strongly-ordered architectures and
it wastes a few dozen clock cycles.
Fortunately, ctx->notify_me is never written concurrently by two
threads, so we can assert that and relax the writes to ctx->notify_me.
The resulting solution works and performs well on both aarch64 and x86.
Note that the atomic_set/atomic_read combination is not an atomic
read-modify-write, and therefore it is even weaker than C11 ATOMIC_RELAXED;
on x86, ATOMIC_RELAXED compiles to a locked operation.
Analyzed-by: Ying Fang <fangying1@huawei.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Tested-by: Ying Fang <fangying1@huawei.com>
Message-Id: <20200407140746.8041-6-pbonzini@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
When there are many poll handlers it's likely that some of them are idle
most of the time. Remove handlers that haven't had activity recently so
that the polling loop scales better for guests with a large number of
devices.
This feature only takes effect for the Linux io_uring fd monitoring
implementation because it is capable of combining fd monitoring with
userspace polling. The other implementations can't do that and risk
starving fds in favor of poll handlers, so don't try this optimization
when they are in use.
IOPS improves from 10k to 105k when the guest has 100
virtio-blk-pci,num-queues=32 devices and 1 virtio-blk-pci,num-queues=1
device for rw=randread,iodepth=1,bs=4k,ioengine=libaio on NVMe.
[Clarified aio_poll_handlers locking discipline explanation in comment
after discussion with Paolo Bonzini <pbonzini@redhat.com>.
--Stefan]
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Link: https://lore.kernel.org/r/20200305170806.1313245-8-stefanha@redhat.com
Message-Id: <20200305170806.1313245-8-stefanha@redhat.com>
Unlike ppoll(2) and epoll(7), Linux io_uring completions can be polled
from userspace. Previously userspace polling was only allowed when all
AioHandler's had an ->io_poll() callback. This prevented starvation of
fds by userspace pollable handlers.
Add the FDMonOps->need_wait() callback that enables userspace polling
even when some AioHandlers lack ->io_poll().
For example, it's now possible to do userspace polling when a TCP/IP
socket is monitored thanks to Linux io_uring.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Link: https://lore.kernel.org/r/20200305170806.1313245-7-stefanha@redhat.com
Message-Id: <20200305170806.1313245-7-stefanha@redhat.com>
The recent Linux io_uring API has several advantages over ppoll(2) and
epoll(2). Details are given in the source code.
Add an io_uring implementation and make it the default on Linux.
Performance is the same as with epoll(7) but later patches add
optimizations that take advantage of io_uring.
It is necessary to change how aio_set_fd_handler() deals with deleting
AioHandlers since removing monitored file descriptors is asynchronous in
io_uring. fdmon_io_uring_remove() marks the AioHandler deleted and
aio_set_fd_handler() will let it handle deletion in that case.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Link: https://lore.kernel.org/r/20200305170806.1313245-6-stefanha@redhat.com
Message-Id: <20200305170806.1313245-6-stefanha@redhat.com>
The AioHandler *node, bool is_new arguments are more complicated to
think about than simply being given AioHandler *old_node, AioHandler
*new_node.
Furthermore, the new Linux io_uring file descriptor monitoring mechanism
added by the new patch requires access to both the old and the new
nodes. Make this change now in preparation.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Link: https://lore.kernel.org/r/20200305170806.1313245-5-stefanha@redhat.com
Message-Id: <20200305170806.1313245-5-stefanha@redhat.com>
The ppoll(2) and epoll(7) file descriptor monitoring implementations are
mixed with the core util/aio-posix.c code. Before adding another
implementation for Linux io_uring, extract out the existing
ones so there is a clear interface and the core code is simpler.
The new interface is AioContext->fdmon_ops, a pointer to a FDMonOps
struct. See the patch for details.
Semantic changes:
1. ppoll(2) now reflects events from pollfds[] back into AioHandlers
while we're still on the clock for adaptive polling. This was
already happening for epoll(7), so if it's really an issue then we'll
need to fix both in the future.
2. epoll(7)'s fallback to ppoll(2) while external events are disabled
was broken when the number of fds exceeded the epoll(7) upgrade
threshold. I guess this code path simply wasn't tested and no one
noticed the bug. I didn't go out of my way to fix it but the correct
code is simpler than preserving the bug.
I also took some liberties in removing the unnecessary
AioContext->epoll_available (just check AioContext->epollfd != -1
instead) and AioContext->epoll_enabled (it's implicit if our
AioContext->fdmon_ops callbacks are being invoked) fields.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Link: https://lore.kernel.org/r/20200305170806.1313245-4-stefanha@redhat.com
Message-Id: <20200305170806.1313245-4-stefanha@redhat.com>
Now that run_poll_handlers_once() is only called by run_poll_handlers()
we can improve the CPU time profile by moving the expensive
RCU_READ_LOCK() out of the polling loop.
This reduces the run_poll_handlers() from 40% CPU to 10% CPU in perf's
sampling profiler output.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Link: https://lore.kernel.org/r/20200305170806.1313245-3-stefanha@redhat.com
Message-Id: <20200305170806.1313245-3-stefanha@redhat.com>
One iteration of polling is always performed even when polling is
disabled. This is done because:
1. Userspace polling is cheaper than making a syscall. We might get
lucky.
2. We must poll once more after polling has stopped in case an event
occurred while stopping polling.
However, there are downsides:
1. Polling becomes a bottleneck when the number of event sources is very
high. It's more efficient to monitor fds in that case.
2. A high-frequency polling event source can starve non-polling event
sources because ppoll(2)/epoll(7) is never invoked.
This patch removes the forced polling iteration so that poll_ns=0 really
means no polling.
IOPS increases from 10k to 60k when the guest has 100
virtio-blk-pci,num-queues=32 devices and 1 virtio-blk-pci,num-queues=1
device because the large number of event sources being polled slows down
the event loop.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Link: https://lore.kernel.org/r/20200305170806.1313245-2-stefanha@redhat.com
Message-Id: <20200305170806.1313245-2-stefanha@redhat.com>
QLIST_SAFE_REMOVE() is confusing here because the node must be on the
list. We actually just wanted to clear the linked list pointers when
removing it from the list. QLIST_REMOVE() now does this, so switch to
it.
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Link: https://lore.kernel.org/r/20200224103406.1894923-3-stefanha@redhat.com
Message-Id: <20200224103406.1894923-3-stefanha@redhat.com>
File descriptor monitoring is O(1) with epoll(7), but
aio_dispatch_handlers() still scans all AioHandlers instead of
dispatching just those that are ready. This makes aio_poll() O(n) with
respect to the total number of registered handlers.
Add a local ready_list to aio_poll() so that each nested aio_poll()
builds a list of handlers ready to be dispatched. Since file descriptor
polling is level-triggered, nested aio_poll() calls also see fds that
were ready in the parent but not yet dispatched. This guarantees that
nested aio_poll() invocations will dispatch all fds, even those that
became ready before the nested invocation.
Since only handlers ready to be dispatched are placed onto the
ready_list, the new aio_dispatch_ready_handlers() function provides O(1)
dispatch.
Note that AioContext polling is still O(n) and currently cannot be fully
disabled. This still needs to be fixed before aio_poll() is fully O(1).
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Sergio Lopez <slp@redhat.com>
Message-id: 20200214171712.541358-6-stefanha@redhat.com
[Fix compilation error on macOS where there is no epoll(87). The
aio_epoll() prototype was out of date and aio_add_ready_list() needed to
be moved outside the ifdef.
--Stefan]
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
It is not necessary to scan all AioHandlers for deletion. Keep a list
of deleted handlers instead of scanning the full list of all handlers.
The AioHandler->deleted field can be dropped. Let's check if the
handler has been inserted into the deleted list instead. Add a new
QLIST_IS_INSERTED() API for this check.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Sergio Lopez <slp@redhat.com>
Message-id: 20200214171712.541358-5-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Don't pass the nanosecond timeout into epoll_wait(), which expects
milliseconds.
The epoll_wait() timeout value does not matter if qemu_poll_ns()
determined that the poll fd is ready, but passing a value in the wrong
units is still ugly. Pass a 0 timeout to epoll_wait() instead.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Sergio Lopez <slp@redhat.com>
Message-id: 20200214171712.541358-3-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
epoll_handler is a stack variable and must not be accessed after it goes
out of scope:
if (aio_epoll_check_poll(ctx, pollfds, npfd, timeout)) {
AioHandler epoll_handler;
...
add_pollfd(&epoll_handler);
ret = aio_epoll(ctx, pollfds, npfd, timeout);
} ...
...
/* if we have any readable fds, dispatch event */
if (ret > 0) {
for (i = 0; i < npfd; i++) {
nodes[i]->pfd.revents = pollfds[i].revents;
}
}
nodes[0] is &epoll_handler, which has already gone out of scope.
There is no need to use pollfds[] for epoll. We don't need an
AioHandler for the epoll fd.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Sergio Lopez <slp@redhat.com>
Message-id: 20200214171712.541358-2-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The first rcu_read_lock/unlock() is expensive. Nested calls are cheap.
This optimization increases IOPS from 73k to 162k with a Linux guest
that has 2 virtio-blk,num-queues=1 and 99 virtio-blk,num-queues=32
devices.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Message-id: 20200218182708.914552-1-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
No header includes qemu-common.h after this commit, as prescribed by
qemu-common.h's file comment.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20190523143508.25387-5-armbru@redhat.com>
[Rebased with conflicts resolved automatically, except for
include/hw/arm/xlnx-zynqmp.h hw/arm/nrf51_soc.c hw/arm/msf2-soc.c
block/qcow2-refcount.c block/qcow2-cluster.c block/qcow2-cache.c
target/arm/cpu.h target/lm32/cpu.h target/m68k/cpu.h target/mips/cpu.h
target/moxie/cpu.h target/nios2/cpu.h target/openrisc/cpu.h
target/riscv/cpu.h target/tilegx/cpu.h target/tricore/cpu.h
target/unicore32/cpu.h target/xtensa/cpu.h; bsd-user/main.c and
net/tap-bsd.c fixed up]
With aio=thread, adaptive polling makes latency worse rather than
better, because it delays the execution of the ThreadPool's
completion bottom half.
event_notifier_poll() does run while polling, detecting that
a bottom half was scheduled by a worker thread, but because
ctx->notifier is explicitly ignored in run_poll_handlers_once(),
scheduling the BH does not count as making progress and
run_poll_handlers() keeps running. Fix this by recomputing
the deadline after *timeout could have changed.
With this change, ThreadPool still cannot participate in polling
but at least it does not suffer from extra latency.
Reported-by: Sergio Lopez <slp@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-id: 20190409122823.12416-1-pbonzini@redhat.com
Cc: Stefan Hajnoczi <stefanha@gmail.com>
Cc: Kevin Wolf <kwolf@redhat.com>
Cc: qemu-block@nongnu.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <1553692145-86728-1-git-send-email-pbonzini@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20190409122823.12416-1-pbonzini@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
aio_poll() has an existing assertion that the function is only called
from the AioContext's home thread if blocking is allowed.
This is not enough, some handlers make assumptions about the thread they
run in. Extend the assertion to non-blocking calls, too.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
It is possible for an io_poll callback to be concurrently executed along
with an aio_set_fd_handlers. This can cause all sorts of problems, like
a NULL callback or a bad opaque pointer.
This changes set_fd_handlers so that it no longer modify existing handlers
entries and instead, always insert those after having proper initialisation.
Tested-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Remy Noel <remy.noel@blade-group.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Message-id: 20181220152030.28035-3-remy.noel@blade-group.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Cleaning the events will cause aio_epoll_update to unregister the fd.
Otherwise, the fd is kept registered until it is destroyed.
Signed-off-by: Remy Noel <remy.noel@blade-group.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Message-id: 20181220152030.28035-2-remy.noel@blade-group.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Commit 70232b5253 ("aio-posix: Don't count ctx->notifier as progress when
2018-08-15), by not reporting progress, causes aio_poll to execute the
system call when polling succeeds because of ctx->notifier. This introduces
latency before the call to aio_bh_poll() and negates the advantages of
polling, unfortunately.
The fix builds on the previous patch, separating the effect of polling on
the timeout from the progress reported to aio_poll(). ctx->notifier
does zero the timeout, causing the caller to skip the system call,
but it does not report progress, so that the bug fix of commit 70232b5253
still stands.
Fixes: 70232b5253
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20180912171040.1732-4-pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
This is a preparation for the next patch, and also a very small
optimization. Compute the timeout only once, before invoking
try_poll_mode, and adjust it in run_poll_handlers. The adjustment
is the polling time when polling fails, or zero (non-blocking) if
polling succeeds.
Fixes: 70232b5253
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20180912171040.1732-3-pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
It is valid for an aio_set_fd_handler to happen concurrently with
aio_poll. In that case, poll_disable_cnt can change under the heels
of aio_poll, and the assertion on poll_disable_cnt can fail in
run_poll_handlers.
Therefore, this patch simply checks the counter on every polling
iteration. There are no particular needs for ordering, since the
polling loop is terminated anyway by aio_notify at the end of
aio_set_fd_handler.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20180912171040.1732-2-pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
The counter is for qemu_lockcnt_inc/dec sections (read side),
qemu_lockcnt_lock/unlock is for the write side.
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
Message-Id: <20180803063917.30292-1-famz@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
An aio_notify() pairs with an aio_notify_accept(). The former should
happen in the main thread or a vCPU thread, and the latter should be
done in the IOThread.
There is one rare case that the main thread or vCPU thread may "steal"
the aio_notify() event just raised by itself, in bdrv_set_aio_context()
[1]. The sequence is like this:
main thread IO Thread
===============================================================
bdrv_drained_begin()
aio_disable_external(ctx)
aio_poll(ctx, true)
ctx->notify_me += 2
...
bdrv_drained_end()
...
aio_notify()
...
bdrv_set_aio_context()
aio_poll(ctx, false)
[1] aio_notify_accept(ctx)
ppoll() /* Hang! */
[1] is problematic. It will clear the ctx->notifier event so that
the blocked ppoll() will not return.
(For the curious, this bug was noticed when booting a number of VMs
simultaneously in RHV. One or two of the VMs will hit this race
condition, making the VIRTIO device unresponsive to I/O commands. When
it hangs, Seabios is busy waiting for a read request to complete (read
MBR), right after initializing the virtio-blk-pci device, using 100%
guest CPU. See also https://bugzilla.redhat.com/show_bug.cgi?id=1562750
for the original bug analysis.)
aio_notify() only injects an event when ctx->notify_me is set,
correspondingly aio_notify_accept() is only useful when ctx->notify_me
_was_ set. Move the call to it into the "blocking" branch. This will
effectively skip [1] and fix the hang.
Furthermore, blocking aio_poll is only allowed on home thread
(in_aio_context_home_thread), because otherwise two blocking
aio_poll()'s can steal each other's ctx->notifier event and cause
hanging just like described above.
Cc: qemu-stable@nongnu.org
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
Message-Id: <20180809132259.18402-3-famz@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
The same logic exists in fd polling. This change is especially important
to avoid busy loop once we limit aio_notify_accept() to blocking
aio_poll().
Cc: qemu-stable@nongnu.org
Signed-off-by: Fam Zheng <famz@redhat.com>
Message-Id: <20180809132259.18402-2-famz@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
When we call addIOThread, the epollfd created in aio_context_setup,
but not close it in the process of delIOThread, so the epollfd will leak.
Reorder the code in aio_epoll_disable and reuse it.
Signed-off-by: Jie Wang <wangjie88@huawei.com>
Message-Id: <1526517763-11108-1-git-send-email-wangjie88@huawei.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
[Mention change to aio_epoll_disable in commit message. - Fam]
Signed-off-by: Fam Zheng <famz@redhat.com>
Applied using the Coccinelle semantic patch scripts/coccinelle/use_osdep.cocci
Signed-off-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Michael Tokarev <mjt@tls.msk.ru>
Reviewed-by: Marc-André Lureau <marcandre.lureau@redhat.com>
This hunk should not have been merged but I forgot to remove it. Let's
remove it before it slips into a QEMU release.
¯\_(ツ)_/¯
Reviewed-by: Thomas Huth <thuth@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 20171103154041.12617-1-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
After iothread is enabled internally inside QEMU with GMainContext, we
may encounter this warning when destroying the iothread:
(qemu-system-x86_64:19925): GLib-CRITICAL **: g_source_remove_poll:
assertion '!SOURCE_DESTROYED (source)' failed
The problem is that g_source_remove_poll() does not allow to remove one
source from array if the source is detached from its owner
context. (peterx: which IMHO does not make much sense)
Fix it on QEMU side by avoid calling g_source_remove_poll() if we know
the object is during destruction, and we won't leak anything after all
since the array will be gone soon cleanly even with that fd.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
Message-id: 20170928025958.1420-6-peterx@redhat.com
[peterx: write the commit message]
Signed-off-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Pull the increment/decrement pair out of aio_bh_poll and into the
callers.
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
Message-id: 20170213135235.12274-18-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
This patch prepares for the removal of unnecessary lockcnt inc/dec pairs.
Extract the dispatching loop for file descriptor handlers into a new
function aio_dispatch_handlers, and then inline aio_dispatch into
aio_poll.
aio_dispatch can now become void.
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
Message-id: 20170213135235.12274-17-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
This covers both file descriptor callbacks and polling callbacks,
since they execute related code.
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
Message-id: 20170213135235.12274-14-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
Message-id: 20170213135235.12274-13-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The AioContext data structures are now protected by list_lock and/or
they are walked with FOREACH_RCU primitives. There is no need anymore
to acquire the AioContext for the entire duration of aio_dispatch.
Instead, just acquire it before and after invoking the callbacks.
The next step is then to push it further down.
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
Message-id: 20170213135235.12274-12-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
AioContext is fairly self contained, the only dependency is QEMUTimer but
that in turn doesn't need anything else. So move them out of block-obj-y
to avoid introducing a dependency from io/ to block-obj-y.
main-loop and its dependency iohandler also need to be moved, because
later in this series io/ will call iohandler_get_aio_context.
[Changed copyright "the QEMU team" to "other QEMU contributors" as
suggested by Daniel Berrange and agreed by Paolo.
--Stefan]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Message-id: 20170213135235.12274-2-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
