This is the big patch that removes
aio_context_acquire()/aio_context_release() from the block layer and
affected block layer users.
There isn't a clean way to split this patch and the reviewers are likely
the same group of people, so I decided to do it in one patch.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Paul Durrant <paul@xen.org>
Message-ID: <20231205182011.1976568-7-stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
These public functions are not used anywhere, thus can be dropped.
Also, since this is the final job API that doesn't use AioContext
lock and replaces it with job_lock, adjust all remaining function
documentation to clearly specify if the job lock is taken or not.
Also document the locking requirements for a few functions
where the second version is not removed.
Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20220926093214.506243-22-eesposit@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@yandex-team.ru>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Change the job_{lock/unlock} and macros to use job_mutex.
Now that they are not nop anymore, remove the aiocontext
to avoid deadlocks.
Therefore:
- when possible, remove completely the aiocontext lock/unlock pair
- if it is used by some other function too, reduce the locking
section as much as possible, leaving the job API outside.
- change AIO_WAIT_WHILE in AIO_WAIT_WHILE_UNLOCKED, since we
are not using the aiocontext lock anymore
The only functions that still need the aiocontext lock are:
- the JobDriver callbacks, already documented in job.h
- job_cancel_sync() in replication.c is called with aio_context_lock
taken, but now job is using AIO_WAIT_WHILE_UNLOCKED so we need to
release the lock.
Reduce the locking section to only cover the callback invocation
and document the functions that take the AioContext lock,
to avoid taking it twice.
Also remove real_job_{lock/unlock}, as they are replaced by the
public functions.
Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Message-Id: <20220926093214.506243-19-eesposit@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@yandex-team.ru>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
In order to make it thread safe, implement a "fake rwlock",
where we allow reads under BQL *or* job_mutex held, but
writes only under BQL *and* job_mutex.
The only write we have is in child_job_set_aio_ctx, which always
happens under drain (so the job is paused).
For this reason, introduce job_set_aio_context and make sure that
the context is set under BQL, job_mutex and drain.
Also make sure all other places where the aiocontext is read
are protected.
The reads in commit.c and mirror.c are actually safe, because always
done under BQL.
Note: at this stage, job_{lock/unlock} and job lock guard macros
are *nop*.
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Message-Id: <20220926093214.506243-14-eesposit@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@yandex-team.ru>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
We want to make sure access of job->aio_context is always done
under either BQL or job_mutex. The problem is that using
aio_co_enter(job->aiocontext, job->co) in job_start and job_enter_cond
makes the coroutine immediately resume, so we can't hold the job lock.
And caching it is not safe either, as it might change.
job_start is under BQL, so it can freely read job->aiocontext, but
job_enter_cond is not.
We want to avoid reading job->aio_context in job_enter_cond, therefore:
1) use aio_co_wake(), since it doesn't want an aiocontext as argument
but uses job->co->ctx
2) detect possible discrepancy between job->co->ctx and job->aio_context
by checking right after the coroutine resumes back from yielding if
job->aio_context has changed. If so, reschedule the coroutine to the
new context.
Calling bdrv_try_set_aio_context() will issue the following calls
(simplified):
* in terms of bdrv callbacks:
.drained_begin -> .set_aio_context -> .drained_end
* in terms of child_job functions:
child_job_drained_begin -> child_job_set_aio_context -> child_job_drained_end
* in terms of job functions:
job_pause_locked -> job_set_aio_context -> job_resume_locked
We can see that after setting the new aio_context, job_resume_locked
calls again job_enter_cond, which then invokes aio_co_wake(). But
while job->aiocontext has been set in job_set_aio_context,
job->co->ctx has not changed, so the coroutine would be entering in
the wrong aiocontext.
Using aio_co_schedule in job_resume_locked() might seem as a valid
alternative, but the problem is that the bh resuming the coroutine
is not scheduled immediately, and if in the meanwhile another
bdrv_try_set_aio_context() is run (see test_propagate_mirror() in
test-block-iothread.c), we would have the first schedule in the
wrong aiocontext, and the second set of drains won't even manage
to schedule the coroutine, as job->busy would still be true from
the previous job_resume_locked().
The solution is to stick with aio_co_wake() and detect every time
the coroutine resumes back from yielding if job->aio_context
has changed. If so, we can reschedule it to the new context.
Check for the aiocontext change in job_do_yield_locked because:
1) aio_co_reschedule_self requires to be in the running coroutine
2) since child_job_set_aio_context allows changing the aiocontext only
while the job is paused, this is the exact place where the coroutine
resumes, before running JobDriver's code.
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@yandex-team.ru>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20220926093214.506243-13-eesposit@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This comment applies more on job, it was left in blockjob as in the past
the whole job logic was implemented there.
Note: at this stage, job_{lock/unlock} and job lock guard macros
are *nop*.
No functional change intended.
Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@yandex-team.ru>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20220926093214.506243-7-eesposit@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
With "intact" we mean that all job.h functions implicitly
take the lock. Therefore API callers are unmodified.
This means that:
- many static functions that will be always called with job lock held
become _locked, and call _locked functions
- all public functions take the lock internally if needed, and call _locked
functions
- all public functions called internally by other functions in job.c will have a
_locked counterpart (sometimes public), to avoid deadlocks (job lock already taken).
These functions are not used for now.
- some public functions called only from exernal files (not job.c) do not
have _locked() counterpart and take the lock inside. Others won't need
the lock at all because use fields only set at initialization and
never modified.
job_{lock/unlock} is independent from real_job_{lock/unlock}.
Note: at this stage, job_{lock/unlock} and job lock guard macros
are *nop*
Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@yandex-team.ru>
Message-Id: <20220926093214.506243-6-eesposit@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
job mutex will be used to protect the job struct elements and list,
replacing AioContext locks.
Right now use a shared lock for all jobs, in order to keep things
simple. Once the AioContext lock is gone, we can introduce per-job
locks.
To simplify the switch from aiocontext to job lock, introduce
*nop* lock/unlock functions and macros.
We want to always call job_lock/unlock outside the AioContext locks,
and not vice-versa, otherwise we might get a deadlock. This is not
straightforward to do, and that's why we start with nop functions.
Once everything is protected by job_lock/unlock, we can change the nop into
an actual mutex and remove the aiocontext lock.
Since job_mutex is already being used, add static
real_job_{lock/unlock} for the existing usage.
Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@yandex-team.ru>
Message-Id: <20220926093214.506243-2-eesposit@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Clearing .cancelled before leaving the main loop when the job has been
soft-cancelled is no longer necessary since job_is_cancelled() only
returns true for jobs that have been force-cancelled.
Therefore, this only makes a differences in places that call
job_cancel_requested(). In block/mirror.c, this is done only before
.cancelled was cleared.
In job.c, there are two callers:
- job_completed_txn_abort() asserts that .cancelled is true, so keeping
it true will not affect this place.
- job_complete() refuses to let a job complete that has .cancelled set.
It is correct to refuse to let the user invoke job-complete on mirror
jobs that have already been soft-cancelled.
With this change, there are no places that reset .cancelled to false and
so we can be sure that .force_cancel can only be true if .cancelled is
true as well. Assert this in job_is_cancelled().
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20211006151940.214590-13-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Most callers of job_is_cancelled() actually want to know whether the job
is on its way to immediate termination. For example, we refuse to pause
jobs that are cancelled; but this only makes sense for jobs that are
really actually cancelled.
A mirror job that is cancelled during READY with force=false should
absolutely be allowed to pause. This "cancellation" (which is actually
a kind of completion) may take an indefinite amount of time, and so
should behave like any job during normal operation. For example, with
on-target-error=stop, the job should stop on write errors. (In
contrast, force-cancelled jobs should not get write errors, as they
should just terminate and not do further I/O.)
Therefore, redefine job_is_cancelled() to only return true for jobs that
are force-cancelled (which as of HEAD^ means any job that interprets the
cancellation request as a request for immediate termination), and add
job_cancel_requested() as the general variant, which returns true for
any jobs which have been requested to be cancelled, whether it be
immediately or after an arbitrarily long completion phase.
Finally, here is a justification for how different job_is_cancelled()
invocations are treated by this patch:
- block/mirror.c (mirror_run()):
- The first invocation is a while loop that should loop until the job
has been cancelled or scheduled for completion. What kind of cancel
does not matter, only the fact that the job is supposed to end.
- The second invocation wants to know whether the job has been
soft-cancelled. Calling job_cancel_requested() is a bit too broad,
but if the job were force-cancelled, we should leave the main loop
as soon as possible anyway, so this should not matter here.
- The last two invocations already check force_cancel, so they should
continue to use job_is_cancelled().
- block/backup.c, block/commit.c, block/stream.c, anything in tests/:
These jobs know only force-cancel, so there is no difference between
job_is_cancelled() and job_cancel_requested(). We can continue using
job_is_cancelled().
- job.c:
- job_pause_point(), job_yield(), job_sleep_ns(): Only force-cancelled
jobs should be prevented from being paused. Continue using job_is_cancelled().
- job_update_rc(), job_finalize_single(), job_finish_sync(): These
functions are all called after the job has left its main loop. The
mirror job (the only job that can be soft-cancelled) will clear
.cancelled before leaving the main loop if it has been
soft-cancelled. Therefore, these functions will observe .cancelled
to be true only if the job has been force-cancelled. We can
continue to use job_is_cancelled().
(Furthermore, conceptually, a soft-cancelled mirror job should not
report to have been cancelled. It should report completion (see
also the block-job-cancel QAPI documentation). Therefore, it makes
sense for these functions not to distinguish between a
soft-cancelled mirror job and a job that has completed as normal.)
- job_completed_txn_abort(): All jobs other than @job have been
force-cancelled. job_is_cancelled() must be true for them.
Regarding @job itself: job_completed_txn_abort() is mostly called
when the job's return value is not 0. A soft-cancelled mirror has a
return value of 0, and so will not end up here then.
However, job_cancel() invokes job_completed_txn_abort() if the job
has been deferred to the main loop, which is mostly the case for
completed jobs (which skip the assertion), but not for sure.
To be safe, use job_cancel_requested() in this assertion.
- job_complete(): This is function eventually invoked by the user
(through qmp_block_job_complete() or qmp_job_complete(), or
job_complete_sync(), which comes from qemu-img). The intention here
is to prevent a user from invoking job-complete after the job has
been cancelled. This should also apply to soft cancelling: After a
mirror job has been soft-cancelled, the user should not be able to
decide otherwise and have it complete as normal (i.e. pivoting to
the target).
- job_cancel(): Both functions are equivalent (see comment there), but
we want to use job_is_cancelled(), because this shows that we call
job_completed_txn_abort() only for force-cancelled jobs. (As
explained for job_update_rc(), soft-cancelled jobs should be treated
as if they have completed as normal.)
Buglink: https://gitlab.com/qemu-project/qemu/-/issues/462
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20211006151940.214590-9-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
The only job that supports a soft cancel mode is the mirror job, and in
such a case it resets its .cancelled field before it leaves its .run()
function, so it does not really count as cancelled.
However, it is possible to cancel the job after .run() returns and
before job_exit() (which is run in the main loop) is executed. Then,
.cancelled would still be true and the job would count as cancelled.
This does not seem to be in the interest of the mirror job, so adjust
job_cancel_async() to not set .cancelled in such a case, and
job_cancel() to not invoke job_completed_txn_abort().
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20211006151940.214590-8-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
We largely have two cancel modes for jobs:
First, there is actual cancelling. The job is terminated as soon as
possible, without trying to reach a consistent result.
Second, we have mirror in the READY state. Technically, the job is not
really cancelled, but it just is a different completion mode. The job
can still run for an indefinite amount of time while it tries to reach a
consistent result.
We want to be able to clearly distinguish which cancel mode a job is in
(when it has been cancelled). We can use Job.force_cancel for this, but
right now it only reflects cancel requests from the user with
force=true, but clearly, jobs that do not even distinguish between
force=false and force=true are effectively always force-cancelled.
So this patch has Job.force_cancel signify whether the job will
terminate as soon as possible (force_cancel=true) or whether it will
effectively remain running despite being "cancelled"
(force_cancel=false).
To this end, we let jobs that provide JobDriver.cancel() tell the
generic job code whether they will terminate as soon as possible or not,
and for jobs that do not provide that method we assume they will.
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20211006151940.214590-7-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Callers should be able to specify whether they want job_cancel_sync() to
force-cancel the job or not.
In fact, almost all invocations do not care about consistency of the
result and just want the job to terminate as soon as possible, so they
should pass force=true. The replication block driver is the exception,
specifically the active commit job it runs.
As for job_cancel_sync_all(), all callers want it to force-cancel all
jobs, because that is the point of it: To cancel all remaining jobs as
quickly as possible (generally on process termination). So make it
invoke job_cancel_sync() with force=true.
This changes some iotest outputs, because quitting qemu while a mirror
job is active will now lead to it being cancelled instead of completed,
which is what we want. (Cancelling a READY mirror job with force=false
may take an indefinite amount of time, which we do not want when
quitting. If users want consistent results, they must have all jobs be
done before they quit qemu.)
Buglink: https://gitlab.com/qemu-project/qemu/-/issues/462
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20211006151940.214590-6-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Finalizing the job may cause its AioContext to change. This is noted by
job_exit(), which points at job_txn_apply() to take this fact into
account.
However, job_completed() does not necessarily invoke job_txn_apply()
(through job_completed_txn_success()), but potentially also
job_completed_txn_abort(). The latter stores the context in a local
variable, and so always acquires the same context at its end that it has
released in the beginning -- which may be a different context from the
one that job_exit() releases at its end. If it is different, qemu
aborts ("qemu_mutex_unlock_impl: Operation not permitted").
Drop the local @outer_ctx variable from job_completed_txn_abort(), and
instead re-acquire the actual job's context at the end of the function,
so job_exit() will release the same.
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20211006151940.214590-2-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Progressmeter is protected by the AioContext mutex, which
is taken by the block jobs and their caller (like blockdev).
We would like to remove the dependency of block layer code on the
AioContext mutex, since most drivers and the core I/O code are already
not relying on it.
Create a new C file to implement the ProgressMeter API, but keep the
struct as public, to avoid forcing allocation on the heap.
Also add a mutex to be able to provide an accurate snapshot of the
progress values to the caller.
Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20210614081130.22134-5-eesposit@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
If mirror is READY than cancel operation is not discarding the whole
result of the operation, but instead it's a documented way get a
point-in-time snapshot of source disk.
So, we should not cancel any requests if mirror is READ and
force=false. Let's fix that case.
Note, that bug that we have before this commit is not critical, as the
only .bdrv_cancel_in_flight implementation is nbd_cancel_in_flight()
and it cancels only requests waiting for reconnection, so it should be
rare case.
Fixes: 521ff8b779
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20210421075858.40197-1-vsementsov@virtuozzo.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
