Precomputing the hash values allows us to perform more frequent
accesses to the hash table, thereby reaching higher throughputs.
We keep the old behaviour by default, since (1) we might confuse
users if they measured a speedup without changing anything in
the QHT implementation, and (2) benchmarking the hash function
"on line" is also valuable.
Before:
$ taskset -c 0 tests/qht-bench -n 1
Throughput: 38.18 MT/s
After:
$ taskset -c 0 tests/qht-bench -n 1
Throughput: 38.16 MT/s
After (with precomputing):
$ taskset -c 0 tests/qht-bench -n 1 -p
Throughput: 50.87 MT/s
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
These functions do not modify their @ht or @bucket arguments.
Constify those arguments.
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
seqlock_read_begin takes a const param since c04649eeea
("seqlock: constify seqlock_read_begin", 2018-08-23), so
we can constify the entire lookup.
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Accessing the HT from an iterator results almost always
in a deadlock. Given that only one qht-internal function
uses this argument, drop it from the interface.
Suggested-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Perform first the tests that exercise code paths that are
easier to hit at small table sizes, and then resize the table
to speed up subsequent tests. If this resize is not too large,
we can make the test faster with no code coverage loss.
- With gcov enabled:
Before: 20.568s, 90.28% qht.c coverage
After: 5.168s, 93.06% qht.c coverage
The coverage increase is entirely due to calling qht_resize,
which we weren't calling before. Note that the code paths
that remain to be tested are either error handling or
can only occur when several threads are accessing the
hash table concurrently (e.g. seqlock retry, trylock fail).
- Without gcov:
Before: 1.987s
After: 0.528s
The speedup is almost the same as with gcov, although the
"before" run is a lot faster.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This improves coverage by one (!) LoC in qht.c, bringing the
coverage rate up from 90.00% to 90.28%.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This improves qht.c code coverage from 89.44% to 90.00%.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This currently has no users, but the use case is so common that I
think we must support it.
Note that without the appended we cannot safely remove a set of
elements; a 2-step approach (i.e. qht_iter first, keep track of
the to-be-deleted elements, and then a bunch of qht_remove calls)
would be racy, since between the iteration and the removals other
threads might insert additional elements.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
- Drain fixes
- node-name parameters for block-commit
- Refactor block jobs to use transactional callbacks for exiting
-----BEGIN PGP SIGNATURE-----
iQEcBAABAgAGBQJbqlBsAAoJEPQH2wBh1c9ABCQIAJ4adAAlr65kmcktHrOeQc6A
7VwSSCUa9B8BJS+/H3V8XF3eX1fa016cRQCHfH/ua3Wqavw00qcuS8Bz/ggc6qls
S1kNSSGhemvtf6ebTCN++HCxEg8g1RjsdnxaTiNWrYLKadX5kjLGofU1eAM2J/re
k5YsyB01X4RHS2L0eAUbYGgNFs+UJEU4p0aKGBPHsOj1LIYEzFhDTVNJ8OjNUG9R
mcMXFiYmQpJTV0hlIqL+pOtRvlR9YOKgkO8dmFkRe+z82f+GA+EZhLfpACxc7ilQ
HS4V2NMVucZ7G8gGudg9mqvd3u/AV5BiUtIGd0iIQ9pU9fUuPVPl977i5WHewoM=
=3TQt
-----END PGP SIGNATURE-----
Merge remote-tracking branch 'remotes/xanclic/tags/pull-block-2018-09-25' into staging
Block layer patches:
- Drain fixes
- node-name parameters for block-commit
- Refactor block jobs to use transactional callbacks for exiting
# gpg: Signature made Tue 25 Sep 2018 16:12:44 BST
# gpg: using RSA key F407DB0061D5CF40
# gpg: Good signature from "Max Reitz <mreitz@redhat.com>"
# Primary key fingerprint: 91BE B60A 30DB 3E88 57D1 1829 F407 DB00 61D5 CF40
* remotes/xanclic/tags/pull-block-2018-09-25: (42 commits)
test-bdrv-drain: Test draining job source child and parent
block: Use a single global AioWait
test-bdrv-drain: Fix outdated comments
test-bdrv-drain: AIO_WAIT_WHILE() in job .commit/.abort
job: Avoid deadlocks in job_completed_txn_abort()
test-bdrv-drain: Test nested poll in bdrv_drain_poll_top_level()
block: Remove aio_poll() in bdrv_drain_poll variants
blockjob: Lie better in child_job_drained_poll()
block-backend: Decrease in_flight only after callback
block-backend: Fix potential double blk_delete()
block-backend: Add .drained_poll callback
block: Add missing locking in bdrv_co_drain_bh_cb()
test-bdrv-drain: Test AIO_WAIT_WHILE() in completion callback
job: Use AIO_WAIT_WHILE() in job_finish_sync()
test-blockjob: Acquire AioContext around job_cancel_sync()
test-bdrv-drain: Drain with block jobs in an I/O thread
aio-wait: Increase num_waiters even in home thread
blockjob: Wake up BDS when job becomes idle
job: Fix missing locking due to mismerge
job: Fix nested aio_poll() hanging in job_txn_apply
...
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
The ARMv8 architecture defines that an AArch32 CPU starts
in SVC mode, unless EL2 is the highest available EL, in
which case it starts in Hyp mode. (In ARMv7 a CPU with EL2
but not EL3 was not a valid configuration, but we don't
specifically reject this if the user asks for one.)
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Message-id: 20180823135047.16525-1-peter.maydell@linaro.org
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Message-id: 20180921161939.822-6-clg@kaod.org
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
The code looks better, it removes duplicated lines and it will ease
the introduction of common properties for the Aspeed machines.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Message-id: 20180921161939.822-4-clg@kaod.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
The AST2500 evb is shipped with a W25Q256 which has a non volatile bit
to make the chip operate in 4 Byte address mode at power up. This
should be an interesting feature to model as it will exercise a bit
more the SMC controllers and MMIO execution at boot time.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Message-id: 20180921161939.822-3-clg@kaod.org
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
In file included from /home/thuth/devel/qemu/hw/timer/aspeed_timer.c:16:
/home/thuth/devel/qemu/include/hw/misc/aspeed_scu.h:37:3: error:
redefinition of typedef 'AspeedSCUState' is a C11 feature
[-Werror,-Wtypedef-redefinition]
} AspeedSCUState;
^
/home/thuth/devel/qemu/include/hw/timer/aspeed_timer.h:27:31: note:
previous definition is here
typedef struct AspeedSCUState AspeedSCUState;
Reported-by: Thomas Huth <thuth@redhat.com>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Message-id: 20180921161939.822-2-clg@kaod.org
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Add 'break' statements missing from a switch in the APB dual-timer
write function. Spotted by Coverity as CID 1395626 and 1395633.
Reported-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Message-id: 20180924123122.14549-1-peter.maydell@linaro.org
The only difference between our implementation of the pcnet ioport
accessors and the mmio accessors is that the former check BCR_DWIO to
see what access widths are permitted for addresses in the aprom range
(0x0..0xf). In fact our failure to do this in the mmio accessors
is a bug (one which was fixed for the ioport accessors in
commit 7ba7974197 in 2011).
The data sheet for the Am79C970A does not describe the DWIO
bit as only applying for I/O space mapped I/O resources and
not memory mapped I/O resources, and our MMIO accessors already
honour DWIO for accesses in the 0x10..0x1f range (since the
pcnet_ioport_{read,write}{w,l} functions check it).
The data sheet for the later but compatible Am79C976 is clearer:
it states specifically "DWIO mode applies to both I/O- and
memory-mapped acceses." This seems to be reasonable evidence
in favour of interpretating the Am79C970A spec as being the same.
(NB: Linux's pcnet driver only supports I/O accesses, so the
MMIO access part of this device is probably untested anyway.)
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Convert the pcnet-pci device away from using the old_mmio
MemoryRegionOps accessor functions.
This commit is a no-behaviour-change API conversion.
(Since PCNET_PNPMMIO_SIZE is 0x20, the old "addr & 0x10"
check and the new "addr < 0x10" check are exact opposites;
the new code is phrased to be parallel with the
pcnet_io_read/write functions.)
I have left a TODO comment marker because the similarity
between the MMIO and IO accessor behaviour is suspicious
and they could be combined, but this will be left to a
different patch.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
The GIC_BASE_IRQ macro is a leftover from when we shared code
between the GICv2 and the v7M NVIC. Since the NVIC is now
split off, GIC_BASE_IRQ is always 0, and we can just delete it.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Luc Michel <luc.michel@greensocs.com>
Message-id: 20180824161819.11085-1-peter.maydell@linaro.org
The GICv2's QEMU interface (sysbus MMIO regions, IRQs,
etc) is now quite complicated with the addition of the
virtualization extensions. Add a comment in the header
file which documents it.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Luc Michel <luc.michel@greensocs.com>
Message-id: 20180823103818.31189-1-peter.maydell@linaro.org
The event queue management is broken today. Event records
are not properly written as EVT_SET_* macro was not updating
the actual event record. Also the event queue interrupt
is not correctly triggered.
Fixes: bb981004ea ("hw/arm/smmuv3: Event queue recording helper")
Signed-off-by: Eric Auger <eric.auger@redhat.com>
Message-id: 20180921070138.10114-3-eric.auger@redhat.com
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
- Fix some jobs/drain/aio_poll related hangs
- commit: Add top-node/base-node options
- linux-aio: Fix locking for qemu_laio_process_completions()
- Fix use after free error in bdrv_open_inherit
-----BEGIN PGP SIGNATURE-----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=c2Ll
-----END PGP SIGNATURE-----
Merge remote-tracking branch 'kevin/tags/for-upstream' into block
Block layer patches:
- Fix some jobs/drain/aio_poll related hangs
- commit: Add top-node/base-node options
- linux-aio: Fix locking for qemu_laio_process_completions()
- Fix use after free error in bdrv_open_inherit
# gpg: Signature made Tue Sep 25 15:54:01 2018 CEST
# gpg: using RSA key 7F09B272C88F2FD6
# gpg: Good signature from "Kevin Wolf <kwolf@redhat.com>"
# Primary key fingerprint: DC3D EB15 9A9A F95D 3D74 56FE 7F09 B272 C88F 2FD6
* kevin/tags/for-upstream: (26 commits)
test-bdrv-drain: Test draining job source child and parent
block: Use a single global AioWait
test-bdrv-drain: Fix outdated comments
test-bdrv-drain: AIO_WAIT_WHILE() in job .commit/.abort
job: Avoid deadlocks in job_completed_txn_abort()
test-bdrv-drain: Test nested poll in bdrv_drain_poll_top_level()
block: Remove aio_poll() in bdrv_drain_poll variants
blockjob: Lie better in child_job_drained_poll()
block-backend: Decrease in_flight only after callback
block-backend: Fix potential double blk_delete()
block-backend: Add .drained_poll callback
block: Add missing locking in bdrv_co_drain_bh_cb()
test-bdrv-drain: Test AIO_WAIT_WHILE() in completion callback
job: Use AIO_WAIT_WHILE() in job_finish_sync()
test-blockjob: Acquire AioContext around job_cancel_sync()
test-bdrv-drain: Drain with block jobs in an I/O thread
aio-wait: Increase num_waiters even in home thread
blockjob: Wake up BDS when job becomes idle
job: Fix missing locking due to mismerge
job: Fix nested aio_poll() hanging in job_txn_apply
...
Signed-off-by: Max Reitz <mreitz@redhat.com>
For the block job drain test, don't only test draining the source and
the target node, but create a backing chain for the source
(source_backing <- source <- source_overlay) and test draining each of
the nodes in it.
When using iothreads, the source node (and therefore the job) is in a
different AioContext than the drain, which happens from the main
thread. This way, the main thread waits in AIO_WAIT_WHILE() for the
iothread to make process and aio_wait_kick() is required to notify it.
The test validates that calling bdrv_wakeup() for a child or a parent
node will actually notify AIO_WAIT_WHILE() instead of letting it hang.
Increase the sleep time a bit (to 1 ms) because the test case is racy
and with the shorter sleep, it didn't reproduce the bug it is supposed
to test for me under 'rr record -n'.
This was because bdrv_drain_invoke_entry() (in the main thread) was only
called after the job had already reached the pause point, so we got a
bdrv_dec_in_flight() from the main thread and the additional
aio_wait_kick() when the job becomes idle (that we really wanted to test
here) wasn't even necessary any more to make progress.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
When draining a block node, we recurse to its parent and for subtree
drains also to its children. A single AIO_WAIT_WHILE() is then used to
wait for bdrv_drain_poll() to become true, which depends on all of the
nodes we recursed to. However, if the respective child or parent becomes
quiescent and calls bdrv_wakeup(), only the AioWait of the child/parent
is checked, while AIO_WAIT_WHILE() depends on the AioWait of the
original node.
Fix this by using a single AioWait for all callers of AIO_WAIT_WHILE().
This may mean that the draining thread gets a few more unnecessary
wakeups because an unrelated operation got completed, but we already
wake it up when something _could_ have changed rather than only if it
has certainly changed.
Apart from that, drain is a slow path anyway. In theory it would be
possible to use wakeups more selectively and still correctly, but the
gains are likely not worth the additional complexity. In fact, this
patch is a nice simplification for some places in the code.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Commit 89bd030533 changed the test case from using job_sleep_ns() to
using qemu_co_sleep_ns() instead. Also, block_job_sleep_ns() became
job_sleep_ns() in commit 5d43e86e11.
In both cases, some comments in the test case were not updated. Do that
now.
Reported-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
This adds tests for calling AIO_WAIT_WHILE() in the .commit and .abort
callbacks. Both reasons why .abort could be called for a single job are
tested: Either .run or .prepare could return an error.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Amongst others, job_finalize_single() calls the .prepare/.commit/.abort
callbacks of the individual job driver. Recently, their use was adapted
for all block jobs so that they involve code calling AIO_WAIT_WHILE()
now. Such code must be called under the AioContext lock for the
respective job, but without holding any other AioContext lock.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
This is a regression test for a deadlock that could occur in callbacks
called from the aio_poll() in bdrv_drain_poll_top_level(). The
AioContext lock wasn't released and therefore would be taken a second
time in the callback. This would cause a possible AIO_WAIT_WHILE() in
the callback to hang.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
bdrv_drain_poll_top_level() was buggy because it didn't release the
AioContext lock of the node to be drained before calling aio_poll().
This way, callbacks called by aio_poll() would possibly take the lock a
second time and run into a deadlock with a nested AIO_WAIT_WHILE() call.
However, it turns out that the aio_poll() call isn't actually needed any
more. It was introduced in commit 91af091f92, which is effectively
reverted by this patch. The cases it was supposed to fix are now covered
by bdrv_drain_poll(), which waits for block jobs to reach a quiescent
state.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Block jobs claim in .drained_poll() that they are in a quiescent state
as soon as job->deferred_to_main_loop is true. This is obviously wrong,
they still have a completion BH to run. We only get away with this
because commit 91af091f92 added an unconditional aio_poll(false) to the
drain functions, but this is bypassing the regular drain mechanisms.
However, just removing this and telling that the job is still active
doesn't work either: The completion callbacks themselves call drain
functions (directly, or indirectly with bdrv_reopen), so they would
deadlock then.
As a better lie, tell that the job is active as long as the BH is
pending, but falsely call it quiescent from the point in the BH when the
completion callback is called. At this point, nested drain calls won't
deadlock because they ignore the job, and outer drains will wait for the
job to really reach a quiescent state because the callback is already
running.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Request callbacks can do pretty much anything, including operations that
will yield from the coroutine (such as draining the backend). In that
case, a decreased in_flight would be visible to other code and could
lead to a drain completing while the callback hasn't actually completed
yet.
Note that reordering these operations forbids calling drain directly
inside an AIO callback. As Paolo explains, indirectly calling it is
okay:
- Calling it through a coroutine is okay, because then
bdrv_drained_begin() goes through bdrv_co_yield_to_drain() and you
have in_flight=2 when bdrv_co_yield_to_drain() yields, then soon
in_flight=1 when the aio_co_wake() in the AIO callback completes, then
in_flight=0 after the bottom half starts.
- Calling it through a bottom half would be okay too, as long as the AIO
callback remembers to do inc_in_flight/dec_in_flight just like
bdrv_co_yield_to_drain() and bdrv_co_drain_bh_cb() do
A few more important cases that come to mind:
- A coroutine that yields because of I/O is okay, with a sequence
similar to bdrv_co_yield_to_drain().
- A coroutine that yields with no I/O pending will correctly decrease
in_flight to zero before yielding.
- Calling more AIO from the callback won't overflow the counter just
because of mutual recursion, because AIO functions always yield at
least once before invoking the callback.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
blk_unref() first decreases the refcount of the BlockBackend and calls
blk_delete() if the refcount reaches zero. Requests can still be in
flight at this point, they are only drained during blk_delete():
At this point, arbitrary callbacks can run. If any callback takes a
temporary BlockBackend reference, it will first increase the refcount to
1 and then decrease it to 0 again, triggering another blk_delete(). This
will cause a use-after-free crash in the outer blk_delete().
Fix it by draining the BlockBackend before decreasing to refcount to 0.
Assert in blk_ref() that it never takes the first refcount (which would
mean that the BlockBackend is already being deleted).
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
A bdrv_drain operation must ensure that all parents are quiesced, this
includes BlockBackends. Otherwise, callbacks called by requests that are
completed on the BDS layer, but not quite yet on the BlockBackend layer
could still create new requests.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
bdrv_do_drained_begin/end() assume that they are called with the
AioContext lock of bs held. If we call drain functions from a coroutine
with the AioContext lock held, we yield and schedule a BH to move out of
coroutine context. This means that the lock for the home context of the
coroutine is released and must be re-acquired in the bottom half.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
This is a regression test for a deadlock that occurred in block job
completion callbacks (via job_defer_to_main_loop) because the AioContext
lock was taken twice: once in job_finish_sync() and then again in
job_defer_to_main_loop_bh(). This would cause AIO_WAIT_WHILE() to hang.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
job_finish_sync() needs to release the AioContext lock of the job before
calling aio_poll(). Otherwise, callbacks called by aio_poll() would
possibly take the lock a second time and run into a deadlock with a
nested AIO_WAIT_WHILE() call.
Also, job_drain() without aio_poll() isn't necessarily enough to make
progress on a job, it could depend on bottom halves to be executed.
Combine both open-coded while loops into a single AIO_WAIT_WHILE() call
that solves both of these problems.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
All callers in QEMU proper hold the AioContext lock when calling
job_finish_sync(). test-blockjob should do the same when it calls the
function indirectly through job_cancel_sync().
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
This extends the existing drain test with a block job to include
variants where the block job runs in a different AioContext.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Even if AIO_WAIT_WHILE() is called in the home context of the
AioContext, we still want to allow the condition to change depending on
other threads as long as they kick the AioWait. Specfically block jobs
can be running in an I/O thread and should then be able to kick a drain
in the main loop context.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
In the context of draining a BDS, the .drained_poll callback of block
jobs is called. If this returns true (i.e. there is still some activity
pending), the drain operation may call aio_poll() with blocking=true to
wait for completion.
As soon as the pending activity is completed and the job finally arrives
in a quiescent state (i.e. its coroutine either yields with busy=false
or terminates), the block job must notify the aio_poll() loop to wake
up, otherwise we get a deadlock if both are running in different
threads.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
job_completed() had a problem with double locking that was recently
fixed independently by two different commits:
"job: Fix nested aio_poll() hanging in job_txn_apply"
"jobs: add exit shim"
One fix removed the first aio_context_acquire(), the other fix removed
the other one. Now we have a bug again and the code is run without any
locking.
Add it back in one of the places.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
All callers have acquired ctx already. Doing that again results in
aio_poll() hang. This fixes the problem that a BDRV_POLL_WHILE() in the
callback cannot make progress because ctx is recursively locked, for
example, when drive-backup finishes.
There are two callers of job_finalize():
fam@lemon:~/work/qemu [master]$ git grep -w -A1 '^\s*job_finalize'
blockdev.c: job_finalize(&job->job, errp);
blockdev.c- aio_context_release(aio_context);
--
job-qmp.c: job_finalize(job, errp);
job-qmp.c- aio_context_release(aio_context);
--
tests/test-blockjob.c: job_finalize(&job->job, &error_abort);
tests/test-blockjob.c- assert(job->job.status == JOB_STATUS_CONCLUDED);
Ignoring the test, it's easy to see both callers to job_finalize (and
job_do_finalize) have acquired the context.
Cc: qemu-stable@nongnu.org
Reported-by: Gu Nini <ngu@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
AIO Coroutines shouldn't by managed by an AioContext different than the
one assigned when they are created. aio_co_enter avoids entering a
coroutine from a different AioContext, calling aio_co_schedule instead.
Scheduled coroutines are then entered by co_schedule_bh_cb using
qemu_coroutine_enter, which just calls qemu_aio_coroutine_enter with the
current AioContext obtained with qemu_get_current_aio_context.
Eventually, co->ctx will be set to the AioContext passed as an argument
to qemu_aio_coroutine_enter.
This means that, if an IO Thread's AioConext is being processed by the
Main Thread (due to aio_poll being called with a BDS AioContext, as it
happens in AIO_WAIT_WHILE among other places), the AioContext from some
coroutines may be wrongly replaced with the one from the Main Thread.
This is the root cause behind some crashes, mainly triggered by the
drain code at block/io.c. The most common are these abort and failed
assertion:
util/async.c:aio_co_schedule
456 if (scheduled) {
457 fprintf(stderr,
458 "%s: Co-routine was already scheduled in '%s'\n",
459 __func__, scheduled);
460 abort();
461 }
util/qemu-coroutine-lock.c:
286 assert(mutex->holder == self);
But it's also known to cause random errors at different locations, and
even SIGSEGV with broken coroutine backtraces.
By using qemu_aio_coroutine_enter directly in co_schedule_bh_cb, we can
pass the correct AioContext as an argument, making sure co->ctx is not
wrongly altered.
Signed-off-by: Sergio Lopez <slp@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
We just fixed a bug that was causing a use-after-free when QEMU was
unable to create a temporary snapshot. This is a test case for this
scenario.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
When a block device is opened with BDRV_O_SNAPSHOT and the
bdrv_append_temp_snapshot() call fails then the error code path tries
to unref the already destroyed 'options' QDict.
This can be reproduced easily by setting TMPDIR to a location where
the QEMU process can't write:
$ TMPDIR=/nonexistent $QEMU -drive driver=null-co,snapshot=on
Signed-off-by: Alberto Garcia <berto@igalia.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Emilio G. Cota