AioContext has its own io_uring instance for file descriptor monitoring.
The disk I/O io_uring code was developed separately. Originally I
thought the characteristics of file descriptor monitoring and disk I/O
were too different, requiring separate io_uring instances.
Now it has become clear to me that it's feasible to share a single
io_uring instance for file descriptor monitoring and disk I/O. We're not
using io_uring's IOPOLL feature or anything else that would require a
separate instance.
Unify block/io_uring.c and util/fdmon-io_uring.c using the new
aio_add_sqe() API that allows user-defined io_uring sqe submission. Now
block/io_uring.c just needs to submit readv/writev/fsync and most of the
io_uring-specific logic is handled by fdmon-io_uring.c.
There are two immediate advantages:
1. Fewer system calls. There is no need to monitor the disk I/O io_uring
ring fd from the file descriptor monitoring io_uring instance. Disk
I/O completions are now picked up directly. Also, sqes are
accumulated in the sq ring until the end of the event loop iteration
and there are fewer io_uring_enter(2) syscalls.
2. Less code duplication.
Note that error_setg() messages are not supposed to end with
punctuation, so I removed a '.' for the non-io_uring build error
message.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-ID: <20251104022933.618123-15-stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Introduce the aio_add_sqe() API for submitting io_uring requests in the
current AioContext. This allows other components in QEMU, like the block
layer, to take advantage of io_uring features without creating their own
io_uring context.
This API supports nested event loops just like file descriptor
monitoring and BHs do. This comes at a complexity cost: CQE callbacks
must be placed on a list so that nested event loops can invoke pending
CQE callbacks from parent event loops. If you're wondering why
CqeHandler exists instead of just a callback function pointer, this is
why.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-ID: <20251104022933.618123-14-stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The ppoll and epoll file descriptor monitoring implementations rely on
the event loop's generic file descriptor, timer, and BH dispatch code to
invoke user callbacks.
The io_uring file descriptor monitoring implementation will need
io_uring-specific dispatch logic for CQE handlers for custom SQEs.
Introduce a new FDMonOps ->dispatch() callback that allows file
descriptor monitoring implementations to invoke user callbacks. The next
patch will use this new callback.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-ID: <20251104022933.618123-13-stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
When aio_context_new() -> aio_context_setup() fails at startup it
doesn't really matter whether errors are returned to the caller or the
process terminates immediately.
However, it is not acceptable to terminate when hotplugging --object
iothread at runtime. Refactor aio_context_setup() so that errors can be
propagated. The next commit will set errp when fdmon_io_uring_setup()
fails.
Suggested-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20251104022933.618123-10-stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
g_source_destroy() only removes the GSource from the GMainContext it's
attached to, if any. It does not free it.
Use g_source_unref() instead so that the AioContext (which embeds a
GSource) is freed. There is no need to call g_source_destroy() in
aio_context_new() because the GSource isn't attached to a GMainContext
yet.
aio_ctx_finalize() expects everything to be set up already, so introduce
the new ctx->initialized boolean and do nothing when called with
!initialized. This also requires moving aio_context_setup() down after
event_notifier_init() since aio_ctx_finalize() won't release any
resources that aio_context_setup() acquired.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-ID: <20251104022933.618123-9-stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
There is no need for aio_context_use_g_source() now that epoll(7) and
io_uring(7) file descriptor monitoring works with the glib event loop.
AioContext doesn't need to be notified that GSource is being used.
On hosts with io_uring support this now enables fdmon-io_uring.c by
default, replacing fdmon-poll.c and fdmon-epoll.c. In other words, the
event loop will use io_uring!
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20251104022933.618123-8-stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
AioContext's glib integration only supports ppoll(2) file descriptor
monitoring. epoll(7) and io_uring(7) disable themselves and switch back
to ppoll(2) when the glib event loop is used. The main loop thread
cannot use epoll(7) or io_uring(7) because it always uses the glib event
loop.
Future QEMU features may require io_uring(7). One example is uring_cmd
support in FUSE exports. Each feature could create its own io_uring(7)
context and integrate it into the event loop, but this is inefficient
due to extra syscalls. It would be more efficient to reuse the
AioContext's existing fdmon-io_uring.c io_uring(7) context because
fdmon-io_uring.c will already be active on systems where Linux io_uring
is available.
In order to keep fdmon-io_uring.c's AioContext operational even when the
glib event loop is used, extend FDMonOps with an API similar to
GSourceFuncs so that file descriptor monitoring can integrate into the
glib event loop.
A quick summary of the GSourceFuncs API:
- prepare() is called each event loop iteration before waiting for file
descriptors and timers.
- check() is called to determine whether events are ready to be
dispatched after waiting.
- dispatch() is called to process events.
More details here: https://docs.gtk.org/glib/struct.SourceFuncs.html
Move the ppoll(2)-specific code from aio-posix.c into fdmon-poll.c and
also implement epoll(7)- and io_uring(7)-specific file descriptor
monitoring code for glib event loops.
Note that it's still faster to use aio_poll() rather than the glib event
loop since glib waits for file descriptor activity with ppoll(2) and
does not support adaptive polling. But at least epoll(7) and io_uring(7)
now work in glib event loops.
Splitting this into multiple commits without temporarily breaking
AioContext proved difficult so this commit makes all the changes. The
next commit will remove the aio_context_use_g_source() API because it is
no longer needed.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-ID: <20251104022933.618123-7-stefanha@redhat.com>
[kwolf: Build fixes; fix AioContext.list_lock use after destroy]
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Adaptive polling has a big problem: It doesn't consider that an event
loop can wait for many different events that may have very different
typical latencies.
For example, think of a guest that tends to send a new I/O request soon
after the previous I/O request completes, but the storage on the host is
rather slow. In this case, getting the new request from guest quickly
means that polling is enabled, but the next thing is performing the I/O
request on the backend, which is slow and disables polling again for the
next guest request. This means that in such a scenario, polling could
help for every other event, but is only ever enabled when it can't
succeed.
In order to fix this, keep a separate AioPolledEvent for each
AioHandler. We will then know that the backend file descriptor always
has a high latency and isn't worth polling for, but we also know that
the guest is always fast and we should poll for it. This solves at least
half of the problem, we can now keep polling for those cases where it
makes sense and get the improved performance from it.
Since the event loop doesn't know which event will be next, we still do
some unnecessary polling while we're waiting for the slow disk. I made
some attempts to be more clever than just randomly growing and shrinking
the polling time, and even to let callers be explicit about when they
expect a new event, but so far this hasn't resulted in improved
performance or even caused performance regressions. For now, let's just
fix the part that is easy enough to fix, we can revisit the rest later.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250307221634.71951-6-kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
As a preparation for having multiple adaptive polling states per
AioContext, move the 'ns' field into a separate struct.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20250307221634.71951-4-kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
These names conflict with ones used by future generic thread pool
equivalents.
Generic names should belong to the generic pool type, not specific (AIO)
type.
Acked-by: Fabiano Rosas <farosas@suse.de>
Reviewed-by: Cédric Le Goater <clg@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Link: https://lore.kernel.org/qemu-devel/70f9e0fb4b01042258a1a57996c64d19779dc7f0.1741124640.git.maciej.szmigiero@oracle.com
Signed-off-by: Cédric Le Goater <clg@redhat.com>
Currently the QemuLockCnt data structure and associated functions are
in the include/qemu/thread.h header. Move them to their own
qemu/lockcnt.h. The main reason for doing this is that it means we
can autogenerate the documentation comments into the docs/devel
documentation.
The copyright/author in the new header is drawn from lockcnt.c,
since the header changes were added in the same commit as
lockcnt.c; since neither thread.h nor lockcnt.c state an explicit
license, the standard default of GPL-2-or-later applies.
We include the new header (and the .c file, which was accidentally
omitted previously) in the "RCU" part of MAINTAINERS, since that
is where the lockcnt.rst documentation is categorized.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Acked-by: Paolo Bonzini <pbonzini@redhat.com>
Message-id: 20240816132212.3602106-7-peter.maydell@linaro.org
The main loop has two AioContexts: qemu_aio_context and iohandler_ctx.
The main loop runs them both, but nested aio_poll() calls on
qemu_aio_context exclude iohandler_ctx.
Which one should qemu_get_current_aio_context() return when called from
the main loop? Document that it's always qemu_aio_context.
This has subtle effects on functions that use
qemu_get_current_aio_context(). For example, aio_co_reschedule_self()
does not work when moving from iohandler_ctx to qemu_aio_context because
qemu_get_current_aio_context() does not differentiate these two
AioContexts.
Document this in order to reduce the chance of future bugs.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-ID: <20240506190622.56095-3-stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
During drain, we do not care about virtqueue notifications, which is why
we remove the handlers on it. When removing those handlers, whether vq
notifications are enabled or not depends on whether we were in polling
mode or not; if not, they are enabled (by default); if so, they have
been disabled by the io_poll_start callback.
Because we do not care about those notifications after removing the
handlers, this is fine. However, we have to explicitly ensure they are
enabled when re-attaching the handlers, so we will resume receiving
notifications. We do this in virtio_queue_aio_attach_host_notifier*().
If such a function is called while we are in a polling section,
attaching the notifiers will then invoke the io_poll_start callback,
re-disabling notifications.
Because we will always miss virtqueue updates in the drained section, we
also need to poll the virtqueue once after attaching the notifiers.
Buglink: https://issues.redhat.com/browse/RHEL-3934
Signed-off-by: Hanna Czenczek <hreitz@redhat.com>
Message-ID: <20240202153158.788922-3-hreitz@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The LuringState typedef is defined twice, in include/block/raw-aio.h and
block/io_uring.c. Move it in include/block/aio.h, which is included
everywhere the typedef is needed, since include/block/aio.h already has
to define the forward reference to the struct.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
aio_context_set_aio_params() doesn't use its undocumented
Error** argument. Remove it to simplify.
Note this removes a use of "unchecked Error**" in
iothread_set_aio_context_params().
Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-ID: <20231120171806.19361-1-philmd@linaro.org>
Delete these functions because nothing calls these functions anymore.
I introduced these APIs in commit 98563fc3ec ("aio: add
aio_context_acquire() and aio_context_release()") in 2014. It's with a
sigh of relief that I delete these APIs almost 10 years later.
Thanks to Paolo Bonzini's vision for multi-queue QEMU, we got an
understanding of where the code needed to go in order to remove the
limitations that the original dataplane and the IOThread/AioContext
approach that followed it.
Emanuele Giuseppe Esposito had the splendid determination to convert
large parts of the codebase so that they no longer needed the AioContext
lock. This was a painstaking process, both in the actual code changes
required and the iterations of code review that Emanuele eked out of
Kevin and me over many months.
Kevin Wolf tackled multitudes of graph locking conversions to protect
in-flight I/O from run-time changes to the block graph as well as the
clang Thread Safety Analysis annotations that allow the compiler to
check whether the graph lock is being used correctly.
And me, well, I'm just here to add some pizzazz to the QEMU multi-queue
block layer :). Thank you to everyone who helped with this effort,
including Eric Blake, code reviewer extraordinaire, and others who I've
forgotten to mention.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-ID: <20231205182011.1976568-11-stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
It is forbidden to block on the event loop during a coroutine, as that
can cause deadlocks due to recursive locking.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-ID: <20230908075458.527013-1-pbonzini@redhat.com>
Reviewed-by: Michael Tokarev <mjt@tls.msk.ru>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The synchronous bdrv_aio_cancel() function needs the acb's AioContext so
it can call aio_poll() to wait for cancellation.
It turns out that all users run under the BQL in the main AioContext, so
this callback is not needed.
Remove the callback, mark bdrv_aio_cancel() GLOBAL_STATE_CODE just like
its blk_aio_cancel() caller, and poll the main loop AioContext.
The purpose of this cleanup is to identify bdrv_aio_cancel() as an API
that does not work with the multi-queue block layer.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-ID: <20230912231037.826804-2-stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Klaus Jensen <k.jensen@samsung.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
All callers now pass is_external=false to aio_set_fd_handler() and
aio_set_event_notifier(). The aio_disable_external() API that
temporarily disables fd handlers that were registered is_external=true
is therefore dead code.
Remove aio_disable_external(), aio_enable_external(), and the
is_external arguments to aio_set_fd_handler() and
aio_set_event_notifier().
The entire test-fdmon-epoll test is removed because its sole purpose was
testing aio_disable_external().
Parts of this patch were generated using the following coccinelle
(https://coccinelle.lip6.fr/) semantic patch:
@@
expression ctx, fd, is_external, io_read, io_write, io_poll, io_poll_ready, opaque;
@@
- aio_set_fd_handler(ctx, fd, is_external, io_read, io_write, io_poll, io_poll_ready, opaque)
+ aio_set_fd_handler(ctx, fd, io_read, io_write, io_poll, io_poll_ready, opaque)
@@
expression ctx, notifier, is_external, io_read, io_poll, io_poll_ready;
@@
- aio_set_event_notifier(ctx, notifier, is_external, io_read, io_poll, io_poll_ready)
+ aio_set_event_notifier(ctx, notifier, io_read, io_poll, io_poll_ready)
Reviewed-by: Juan Quintela <quintela@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20230516190238.8401-21-stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Devices can pass their MemoryReentrancyGuard (from their DeviceState),
when creating new BHes. Then, the async API will toggle the guard
before/after calling the BH call-back. This prevents bh->mmio reentrancy
issues.
Signed-off-by: Alexander Bulekov <alxndr@bu.edu>
Reviewed-by: Darren Kenny <darren.kenny@oracle.com>
Message-Id: <20230427211013.2994127-3-alxndr@bu.edu>
[thuth: Fix "line over 90 characters" checkpatch.pl error]
Signed-off-by: Thomas Huth <thuth@redhat.com>
Remove usage of aio_context_acquire by always submitting asynchronous
AIO to the current thread's LuringState.
In order to prevent mistakes from the caller side, avoid passing LuringState
in luring_io_{plug/unplug} and luring_co_submit, and document the functions
to make clear that they work in the current thread's AioContext.
Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Message-Id: <20230203131731.851116-3-eesposit@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Remove usage of aio_context_acquire by always submitting asynchronous
AIO to the current thread's LinuxAioState.
In order to prevent mistakes from the caller side, avoid passing LinuxAioState
in laio_io_{plug/unplug} and laio_co_submit, and document the functions
to make clear that they work in the current thread's AioContext.
Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Message-Id: <20230203131731.851116-2-eesposit@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Message-Id: <20221221131435.3851212-6-armbru@redhat.com>
qemu/coroutine.h and qemu/lockable.h include each other.
They need each other only in macro expansions, so we could simply drop
both inclusions to break the loop, and add suitable includes to files
that expand the macros.
Instead, move a part of qemu/coroutine.h to new qemu/coroutine-core.h
so that qemu/coroutine-core.h doesn't need qemu/lockable.h, and
qemu/lockable.h only needs qemu/coroutine-core.h. Result:
qemu/coroutine.h includes qemu/lockable.h includes
qemu/coroutine-core.h.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20221221131435.3851212-5-armbru@redhat.com>
[Semantic rebase conflict with 7c10cb38cc "accel/tcg: Add debuginfo
support" resolved]
Block layer graph operations are always run under BQL in the main loop.
This is proved by the assertion qemu_in_main_thread() and its wrapper
macro GLOBAL_STATE_CODE.
However, there are also concurrent coroutines running in other iothreads
that always try to traverse the graph. Currently this is protected
(among various other things) by the AioContext lock, but once this is
removed, we need to make sure that reads do not happen while modifying
the graph.
We distinguish between writer (main loop, under BQL) that modifies the
graph, and readers (all other coroutines running in various AioContext),
that go through the graph edges, reading ->parents and->children.
The writer (main loop) has "exclusive" access, so it first waits for any
current read to finish, and then prevents incoming ones from entering
while it has the exclusive access.
The readers (coroutines in multiple AioContext) are free to access the
graph as long the writer is not modifying the graph. In case it is, they
go in a CoQueue and sleep until the writer is done.
If a coroutine changes AioContext, the counter in the original and new
AioContext are left intact, since the writer does not care where the
reader is, but only if there is one.
As a result, some AioContexts might have a negative reader count, to
balance the positive count of the AioContext that took the lock. This
also means that when an AioContext is deleted it may have a nonzero
reader count. In that case we transfer the count to a global shared
counter so that the writer is always aware of all readers.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20221207131838.239125-3-kwolf@redhat.com>
Reviewed-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@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>
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>
It can be difficult to debug issues with BHs in production environments.
Although BHs can usually be identified by looking up their ->cb()
function pointer, this requires debug information for the program. It is
also not possible to print human-readable diagnostics about BHs because
they have no identifier.
This patch adds a name to each BH. The name is not unique per instance
but differentiates between cb() functions, which is usually enough. It's
done by changing aio_bh_new() and friends to macros that stringify cb.
The next patch will use the name field when reporting leaked BHs.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Message-Id: <20210414200247.917496-2-stefanha@redhat.com>
If we want to wake up a coroutine from a worker thread, aio_co_wake()
currently does not work. In that scenario, aio_co_wake() calls
aio_co_enter(), but there is no current AioContext and therefore
qemu_get_current_aio_context() returns the main thread. aio_co_wake()
then attempts to call aio_context_acquire() instead of going through
aio_co_schedule().
The default case of qemu_get_current_aio_context() was added to cover
synchronous I/O started from the vCPU thread, but the main and vCPU
threads are quite different. The main thread is an I/O thread itself,
only running a more complicated event loop; the vCPU thread instead
is essentially a worker thread that occasionally calls
qemu_mutex_lock_iothread(). It is only in those critical sections
that it acts as if it were the home thread of the main AioContext.
Therefore, this patch detaches qemu_get_current_aio_context() from
iothreads, which is a useless complication. The AioContext pointer
is stored directly in the thread-local variable, including for the
main loop. Worker threads (including vCPU threads) optionally behave
as temporary home threads if they have taken the big QEMU lock,
but if that is not the case they will always schedule coroutines
on remote threads via aio_co_schedule().
With this change, the stub qemu_mutex_iothread_locked() must be changed
from true to false. The previous value of true was needed because the
main thread did not have an AioContext in the thread-local variable,
but now it does have one.
Reported-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20210609122234.544153-1-pbonzini@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Tested-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
[eblake: tweak commit message per Vladimir's review]
Signed-off-by: Eric Blake <eblake@redhat.com>
Add a function that can be used to move the currently running coroutine
to a different AioContext (and therefore potentially a different
thread).
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20201005155855.256490-12-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>
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>
Any thread that is not a iothread returns NULL for qemu_get_current_aio_context().
As a result, it would also return true for
in_aio_context_home_thread(qemu_get_aio_context()), causing
AIO_WAIT_WHILE to invoke aio_poll() directly. This is incorrect
if the BQL is not held, because aio_poll() does not expect to
run concurrently from multiple threads, and it can actually
happen when savevm writes to the vmstate file from the
migration thread.
Therefore, restrict in_aio_context_home_thread to return true
for the main AioContext only if the BQL is held.
The function is moved to aio-wait.h because it is mostly used
there and to avoid a circular reference between main-loop.h
and block/aio.h.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20200407140746.8041-5-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>
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>
The ctx->first_bh list contains all created BHs, including those that
are not scheduled. The list is iterated by the event loop and therefore
has O(n) time complexity with respected to the number of created BHs.
Rewrite BHs so that only scheduled or deleted BHs are enqueued.
Only BHs that actually require action will be iterated.
One semantic change is required: qemu_bh_delete() enqueues the BH and
therefore invokes aio_notify(). The
tests/test-aio.c:test_source_bh_delete_from_cb() test case assumed that
g_main_context_iteration(NULL, false) returns false after
qemu_bh_delete() but it now returns true for one iteration. Fix up the
test case.
This patch makes aio_compute_timeout() and aio_bh_poll() drop from a CPU
profile reported by perf-top(1). Previously they combined to 9% CPU
utilization when AioContext polling is commented out and the guest 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: 20200221093951.1414693-1-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Aborts when sqe fails to be set as sqes cannot be returned to the
ring. Adds slow path for short reads for older kernels
Signed-off-by: Aarushi Mehta <mehta.aaru20@gmail.com>
Acked-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 20200120141858.587874-5-stefanha@redhat.com
Message-Id: <20200120141858.587874-5-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]
Attributes are simple flags, associated with individual timers for their
whole lifetime. They intended to be used to mark individual timers for
special handling when they fire.
New/init functions family in timer interface updated and refactored (new
'attribute' argument added, timer_list replaced with timer_list_group+type
combinations, comments improved to avoid info duplication). Also existing
aio interface extended with attribute-enabled variants of functions,
which create/initialize timers.
Signed-off-by: Artem Pisarenko <artem.k.pisarenko@gmail.com>
Message-Id: <f47b81dbce734e9806f9516eba8ca588e6321c2f.1539764043.git.artem.k.pisarenko@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
laio_init() can fail for a couple of reasons, which will lead to a NULL
pointer dereference in laio_attach_aio_context().
To solve this, add a aio_setup_linux_aio() function which is called
early in raw_open_common. If this fails, propagate the error up. The
signature of aio_get_linux_aio() was not modified, because it seems
preferable to return the actual errno from the possible failing
initialization calls.
Additionally, when the AioContext changes, we need to associate a
LinuxAioState with the new AioContext. Use the bdrv_attach_aio_context
callback and call the new aio_setup_linux_aio(), which will allocate a
new AioContext if needed, and return errors on failures. If it fails for
any reason, fallback to threaded AIO with an error message, as the
device is already in-use by the guest.
Add an assert that aio_get_linux_aio() cannot return NULL.
Signed-off-by: Nishanth Aravamudan <naravamudan@digitalocean.com>
Message-id: 20180622193700.6523-1-naravamudan@digitalocean.com
Signed-off-by: Stefan Hajnoczi <stefanha@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>
The name aio_context_in_iothread() is misleading because it also returns
true when called on the main AioContext from the main loop thread, which
is not an IOThread.
This patch renames it to in_aio_context_home_thread() and expands the
doc comment to make the semantics clearer.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The main loop uses aio_disable_external()/aio_enable_external() to
temporarily disable processing of external AioContext clients like
device emulation.
This allows monitor commands to quiesce I/O and prevent the guest from
submitting new requests while a monitor command is in progress.
The aio_enable_external() API is currently broken when an IOThread is in
aio_poll() waiting for fd activity when the main loop re-enables
external clients. Incrementing ctx->external_disable_cnt does not wake
the IOThread from ppoll(2) so fd processing remains suspended and leads
to unresponsive emulated devices.
This patch adds an aio_notify() call to aio_enable_external() so the
IOThread is kicked out of ppoll(2) and will re-arm the file descriptors.
The bug can be reproduced as follows:
$ qemu -M accel=kvm -m 1024 \
-object iothread,id=iothread0 \
-device virtio-scsi-pci,iothread=iothread0,id=virtio-scsi-pci0 \
-drive if=none,id=drive0,aio=native,cache=none,format=raw,file=test.img \
-device scsi-hd,id=scsi-hd0,drive=drive0 \
-qmp tcp::5555,server,nowait
$ scripts/qmp/qmp-shell localhost:5555
(qemu) blockdev-snapshot-sync device=drive0 snapshot-file=sn1.qcow2
mode=absolute-paths format=qcow2
After blockdev-snapshot-sync completes the SCSI disk will be
unresponsive. This leads to request timeouts inside the guest.
Reported-by: Qianqian Zhu <qizhu@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 20170508180705.20609-1-stefanha@redhat.com
Suggested-by: Fam Zheng <famz@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
They start the coroutine on the specified context.
Signed-off-by: Fam Zheng <famz@redhat.com>
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
