bdrv_unref() is called by a lot of places that need to hold the graph
lock (it naturally happens in the context of operations that change the
graph). However, bdrv_unref() takes the graph writer lock internally, so
it can't actually be called while already holding a graph lock without
causing a deadlock.
bdrv_unref() also can't just become GRAPH_WRLOCK because it drains the
node before closing it, and draining requires that the graph is
unlocked.
The solution is to defer deleting the node until we don't hold the lock
any more and draining is possible again.
Note that keeping images open for longer than necessary can create
problems, too: You can't open an image again before it is really closed
(if image locking didn't prevent it, it would cause corruption).
Reopening an image immediately happens at least during bdrv_open() and
bdrv_co_create().
In order to solve this problem, make sure to run the deferred unref in
bdrv_graph_wrunlock(), i.e. the first possible place where we can drain
again. This is also why bdrv_schedule_unref() is marked GRAPH_WRLOCK.
The output of iotest 051 is updated because the additional polling
changes the order of HMP output, resulting in a new "(qemu)" prompt in
the test output that was previously on a separate line and filtered out.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20230911094620.45040-6-kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Now that bdrv_graph_wrlock() temporarily drops the AioContext lock that
its caller holds, it can poll without causing deadlocks. We can now
re-enable graph locking.
This reverts commit ad128dff0bf4b6f971d05eb4335a627883a19c1d.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20230605085711.21261-12-kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
If the caller keeps the AioContext lock for a block node in an iothread,
polling in bdrv_graph_wrlock() deadlocks if the condition isn't
fulfilled immediately.
Now that all callers make sure to actually have the AioContext locked
when they call bdrv_replace_child_noperm() like they should, we can
change bdrv_graph_wrlock() to take a BlockDriverState whose AioContext
lock the caller holds (NULL if it doesn't) and unlock it temporarily
while polling.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-ID: <20230605085711.21261-11-kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
In QEMU 8.0, we've been seeing deadlocks in bdrv_graph_wrlock(). They
come from callers that hold an AioContext lock, which is not allowed
during polling. In theory, we could temporarily release the lock, but
callers are inconsistent about whether they hold a lock, and if they do,
some are also confused about which one they hold. While all of this is
fixable, it's not trivial, and the best course of action for 8.0.1 is
probably just disabling the graph locking code temporarily.
We don't currently rely on graph locking yet. It is supposed to replace
the AioContext lock eventually to enable multiqueue support, but as long
as we still have the AioContext lock, it is sufficient without the graph
lock. Once the AioContext lock goes away, the deadlock doesn't exist any
more either and this commit can be reverted. (Of course, it can also be
reverted while the AioContext lock still exists if the callers have been
fixed.)
Cc: qemu-stable@nongnu.org
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20230517152834.277483-2-kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
There are some conditions under which we don't actually need to do
anything for taking a reader lock: Writing the graph is only possible
from the main context while holding the BQL. So if a reader is running
in the main context under the BQL and knows that it won't be interrupted
until the next writer runs, we don't actually need to do anything.
This is the case if the reader code neither has a nested event loop
(this is forbidden anyway while you hold the lock) nor is a coroutine
(because a writer could run when the coroutine has yielded).
These conditions are exactly what bdrv_graph_rdlock_main_loop() asserts.
They are not fulfilled in bdrv_graph_co_rdlock(), which always runs in a
coroutine.
This deletes the shortcuts in bdrv_graph_co_rdlock() that skip taking
the reader lock in the main thread.
Reported-by: Fiona Ebner <f.ebner@proxmox.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20230510203601.418015-9-kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
reader_count() is a performance bottleneck because the global
aio_context_list_lock mutex causes thread contention. Put this debugging
assertion behind a new ./configure --enable-debug-graph-lock option and
disable it by default.
The --enable-debug-graph-lock option is also enabled by the more general
--enable-debug option.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20230501173443.153062-1-stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The following conversion is safe and does not change behavior:
GLOBAL_STATE_CODE();
...
- AIO_WAIT_WHILE(qemu_get_aio_context(), ...);
+ AIO_WAIT_WHILE_UNLOCKED(NULL, ...);
Since we're in GLOBAL_STATE_CODE(), qemu_get_aio_context() is our home
thread's AioContext. Thus AIO_WAIT_WHILE() does not unlock the
AioContext:
if (ctx_ && in_aio_context_home_thread(ctx_)) { \
while ((cond)) { \
aio_poll(ctx_, true); \
waited_ = true; \
} \
And that means AIO_WAIT_WHILE_UNLOCKED(NULL, ...) can be substituted.
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Tested-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20230309190855.414275-4-stefanha@redhat.com>
Reviewed-by: Wilfred Mallawa <wilfred.mallawa@wdc.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20221207131838.239125-15-kwolf@redhat.com>
Reviewed-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Remove the old assert_bdrv_graph_writable, and replace it with
the new version using graph-lock API.
See the function documentation for more information.
Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20221207131838.239125-14-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>
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>
