| Message ID | 20220208143513.1077229-19-eesposit@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | job: replace AioContext lock with job_mutex | expand |
On Tue, Feb 08, 2022 at 09:35:11AM -0500, Emanuele Giuseppe Esposito wrote: > static AioContext *child_job_get_parent_aio_context(BdrvChild *c) > { > BlockJob *job = c->opaque; > + assert(qemu_in_main_thread()); > > return job->job.aio_context; > } It's not clear to me that .get_parent_aio_context() should only be called from the main thread. The API is read-only so someone might try to call from I/O code in the future expecting it to work like other read-only graph APIs that are available from I/O code. Currently the assertion is true because the only user is bdrv_attach_child_*() but please document this invariant for bdrv_child_get_parent_aio_context() and the callback. Maybe move the assertion into a higher-level function like bdrv_child_get_parent_aio_context() (if that still covers all cases). > diff --git a/include/qemu/job.h b/include/qemu/job.h > index dfbf2ea501..ca46e46f5b 100644 > --- a/include/qemu/job.h > +++ b/include/qemu/job.h > @@ -75,7 +75,12 @@ typedef struct Job { > ProgressMeter progress; > > > - /** AioContext to run the job coroutine in */ > + /** > + * AioContext to run the job coroutine in. > + * This field can be read when holding either the BQL (so we are in > + * the main loop) or the job_mutex. > + * Instead, it can be only written when we hold *both* BQL and job_mutex. s/Instead,// (It sounds weird because "instead" means "replacement" or "substitution" We're comparing "read" and "write" here, not substituting them. Something like "on the other hand" or "conversely" works.) > + */ > AioContext *aio_context; > > /** Reference count of the block job */ > @@ -706,4 +711,16 @@ void job_dismiss_locked(Job **job, Error **errp); > int job_finish_sync_locked(Job *job, void (*finish)(Job *, Error **errp), > Error **errp); > > +/** > + * Sets the @job->aio_context. > + * Called with job_mutex *not* held. > + * > + * This function must run in the main thread to protect against > + * concurrent read in job_finish_sync_locked(), > + * takes the job_mutex lock to protect against the read in > + * job_do_yield_locked(), and must be called when the coroutine > + * is quiescent. > + */ > +void job_set_aio_context(Job *job, AioContext *ctx); > + > #endif > diff --git a/job.c b/job.c > index f05850a337..7a07d25ec3 100644 > --- a/job.c > +++ b/job.c > @@ -354,6 +354,17 @@ Job *job_get_locked(const char *id) > return NULL; > } > > +void job_set_aio_context(Job *job, AioContext *ctx) > +{ > + /* protect against read in job_finish_sync_locked and job_start */ > + assert(qemu_in_main_thread()); > + /* protect against read in job_do_yield_locked */ > + JOB_LOCK_GUARD(); > + /* ensure the coroutine is quiescent while the AioContext is changed */ > + assert(job->pause_count > 0); > + job->aio_context = ctx; > +} > + > /* Called with job_mutex *not* held. */ > static void job_sleep_timer_cb(void *opaque) > { > @@ -1256,6 +1267,7 @@ int job_finish_sync_locked(Job *job, void (*finish)(Job *, Error **errp), > { > Error *local_err = NULL; > int ret; > + assert(qemu_in_main_thread()); > > job_ref_locked(job); > > -- > 2.31.1 >
Am 08/03/2022 um 14:41 schrieb Stefan Hajnoczi: > It's not clear to me that .get_parent_aio_context() should only be > called from the main thread. The API is read-only so someone might try > to call from I/O code in the future expecting it to work like other > read-only graph APIs that are available from I/O code. > > Currently the assertion is true because the only user is > bdrv_attach_child_*() but please document this invariant for > bdrv_child_get_parent_aio_context() and the callback. Maybe move the > assertion into a higher-level function like > bdrv_child_get_parent_aio_context() (if that still covers all cases). We classified .get_parent_aio_context as GS callback, and indeed bdrv_child_get_parent_aio_context has the macro GLOBAL_STATE_CODE(); So we should be fine. Emanuele
diff --git a/block/replication.c b/block/replication.c index 50ea778937..68018948b9 100644 --- a/block/replication.c +++ b/block/replication.c @@ -148,8 +148,8 @@ static void replication_close(BlockDriverState *bs) } if (s->stage == BLOCK_REPLICATION_FAILOVER) { commit_job = &s->commit_job->job; - assert(commit_job->aio_context == qemu_get_current_aio_context()); WITH_JOB_LOCK_GUARD() { + assert(commit_job->aio_context == qemu_get_current_aio_context()); job_cancel_sync_locked(commit_job, false); } } diff --git a/blockjob.c b/blockjob.c index 04d868f020..afb0e9ad5e 100644 --- a/blockjob.c +++ b/blockjob.c @@ -154,12 +154,13 @@ static void child_job_set_aio_ctx(BdrvChild *c, AioContext *ctx, bdrv_set_aio_context_ignore(sibling->bs, ctx, ignore); } - job->job.aio_context = ctx; + job_set_aio_context(&job->job, ctx); } static AioContext *child_job_get_parent_aio_context(BdrvChild *c) { BlockJob *job = c->opaque; + assert(qemu_in_main_thread()); return job->job.aio_context; } diff --git a/include/qemu/job.h b/include/qemu/job.h index dfbf2ea501..ca46e46f5b 100644 --- a/include/qemu/job.h +++ b/include/qemu/job.h @@ -75,7 +75,12 @@ typedef struct Job { ProgressMeter progress; - /** AioContext to run the job coroutine in */ + /** + * AioContext to run the job coroutine in. + * This field can be read when holding either the BQL (so we are in + * the main loop) or the job_mutex. + * Instead, it can be only written when we hold *both* BQL and job_mutex. + */ AioContext *aio_context; /** Reference count of the block job */ @@ -706,4 +711,16 @@ void job_dismiss_locked(Job **job, Error **errp); int job_finish_sync_locked(Job *job, void (*finish)(Job *, Error **errp), Error **errp); +/** + * Sets the @job->aio_context. + * Called with job_mutex *not* held. + * + * This function must run in the main thread to protect against + * concurrent read in job_finish_sync_locked(), + * takes the job_mutex lock to protect against the read in + * job_do_yield_locked(), and must be called when the coroutine + * is quiescent. + */ +void job_set_aio_context(Job *job, AioContext *ctx); + #endif diff --git a/job.c b/job.c index f05850a337..7a07d25ec3 100644 --- a/job.c +++ b/job.c @@ -354,6 +354,17 @@ Job *job_get_locked(const char *id) return NULL; } +void job_set_aio_context(Job *job, AioContext *ctx) +{ + /* protect against read in job_finish_sync_locked and job_start */ + assert(qemu_in_main_thread()); + /* protect against read in job_do_yield_locked */ + JOB_LOCK_GUARD(); + /* ensure the coroutine is quiescent while the AioContext is changed */ + assert(job->pause_count > 0); + job->aio_context = ctx; +} + /* Called with job_mutex *not* held. */ static void job_sleep_timer_cb(void *opaque) { @@ -1256,6 +1267,7 @@ int job_finish_sync_locked(Job *job, void (*finish)(Job *, Error **errp), { Error *local_err = NULL; int ret; + assert(qemu_in_main_thread()); job_ref_locked(job);
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. Suggested-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com> --- block/replication.c | 2 +- blockjob.c | 3 ++- include/qemu/job.h | 19 ++++++++++++++++++- job.c | 12 ++++++++++++ 4 files changed, 33 insertions(+), 3 deletions(-)