Message ID | 20220105140208.365608-15-eesposit@redhat.com |
---|---|
State | New |
Headers | show |
Series | job: replace AioContext lock with job_mutex | expand |
Getters such as job_get_aio_context are often wrong because the AioContext can change immediately after returning. So, I wonder if job.aio_context should be protected with a kind of "fake rwlock": read under BQL or job_lock, write under BQL+job_lock. For this to work, you can add an assertion for qemu_in_main_thread() to child_job_set_aio_ctx, or even better have the assertion in a wrapper API job_set_aio_context_locked(). And then, we can remove job_get_aio_context(). Let's look at all cases individually: On 1/5/22 15:02, Emanuele Giuseppe Esposito wrote: > diff --git a/block/commit.c b/block/commit.c > index f639eb49c5..961b57edf0 100644 > --- a/block/commit.c > +++ b/block/commit.c > @@ -369,7 +369,7 @@ void commit_start(const char *job_id, BlockDriverState *bs, > goto fail; > } > > - s->base = blk_new(s->common.job.aio_context, > + s->base = blk_new(job_get_aio_context(&s->common.job), > base_perms, > BLK_PERM_CONSISTENT_READ > | BLK_PERM_GRAPH_MOD Here the AioContext is the one of bs. It cannot change because we're under BQL. Replace with bdrv_get_aio_context. > @@ -382,7 +382,7 @@ void commit_start(const char *job_id, BlockDriverState *bs, > s->base_bs = base; > > /* Required permissions are already taken with block_job_add_bdrv() */ > - s->top = blk_new(s->common.job.aio_context, 0, BLK_PERM_ALL); > + s->top = blk_new(job_get_aio_context(&s->common.job), 0, BLK_PERM_ALL); > ret = blk_insert_bs(s->top, top, errp); > if (ret < 0) { > goto fail; Same. > diff --git a/block/mirror.c b/block/mirror.c > index 41450df55c..72b4367b4e 100644 > --- a/block/mirror.c > +++ b/block/mirror.c > @@ -1743,7 +1743,7 @@ static BlockJob *mirror_start_job( > target_perms |= BLK_PERM_GRAPH_MOD; > } > > - s->target = blk_new(s->common.job.aio_context, > + s->target = blk_new(job_get_aio_context(&s->common.job), > target_perms, target_shared_perms); > ret = blk_insert_bs(s->target, target, errp); > if (ret < 0) { Same. > 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); > } > } Ok. > diff --git a/blockjob.c b/blockjob.c > index cf1f49f6c2..468ba735c5 100644 > --- a/blockjob.c > +++ b/blockjob.c > @@ -155,14 +155,16 @@ 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; > + WITH_JOB_LOCK_GUARD() { > + job->job.aio_context = ctx; > + } > } > > static AioContext *child_job_get_parent_aio_context(BdrvChild *c) > { > BlockJob *job = c->opaque; > > - return job->job.aio_context; > + return job_get_aio_context(&job->job); > } > > static const BdrvChildClass child_job = { Both called with BQL held, I think. > @@ -218,19 +220,21 @@ int block_job_add_bdrv(BlockJob *job, const char *name, BlockDriverState *bs, > { > BdrvChild *c; > bool need_context_ops; > + AioContext *job_aiocontext; > assert(qemu_in_main_thread()); > > bdrv_ref(bs); > > - need_context_ops = bdrv_get_aio_context(bs) != job->job.aio_context; > + job_aiocontext = job_get_aio_context(&job->job); > + need_context_ops = bdrv_get_aio_context(bs) != job_aiocontext; > > - if (need_context_ops && job->job.aio_context != qemu_get_aio_context()) { > - aio_context_release(job->job.aio_context); > + if (need_context_ops && job_aiocontext != qemu_get_aio_context()) { > + aio_context_release(job_aiocontext); > } > c = bdrv_root_attach_child(bs, name, &child_job, 0, perm, shared_perm, job, > errp); > - if (need_context_ops && job->job.aio_context != qemu_get_aio_context()) { > - aio_context_acquire(job->job.aio_context); > + if (need_context_ops && job_aiocontext != qemu_get_aio_context()) { > + aio_context_acquire(job_aiocontext); > } > if (c == NULL) { > return -EPERM; BQL held, too. > diff --git a/job.c b/job.c > index f16a4ef542..8a5b710d9b 100644 > --- a/job.c > +++ b/job.c > @@ -566,7 +566,7 @@ void job_enter_cond_locked(Job *job, bool(*fn)(Job *job)) > job->busy = true; > real_job_unlock(); > job_unlock(); > - aio_co_enter(job->aio_context, job->co); > + aio_co_enter(job_get_aio_context(job), job->co); > job_lock(); > } > If you replace aio_co_enter with aio_co_schedule, you can call it without dropping the lock. The difference being that aio_co_schedule will always go through a bottom half. > @@ -1138,7 +1138,6 @@ static void coroutine_fn job_co_entry(void *opaque) > Job *job = opaque; > int ret; > > - assert(job->aio_context == qemu_get_current_aio_context()); > assert(job && job->driver && job->driver->run); > job_pause_point(job); > ret = job->driver->run(job, &job->err); > @@ -1177,7 +1176,7 @@ void job_start(Job *job) > job->paused = false; > job_state_transition_locked(job, JOB_STATUS_RUNNING); > } > - aio_co_enter(job->aio_context, job->co); > + aio_co_enter(job_get_aio_context(job), job->co); Better to use aio_co_schedule here, too, and move it under the previous WITH_JOB_LOCK_GUARD. > } > > void job_cancel_locked(Job *job, bool force) > @@ -1303,7 +1302,8 @@ int job_finish_sync_locked(Job *job, void (*finish)(Job *, Error **errp), > } > > job_unlock(); > - AIO_WAIT_WHILE(job->aio_context, (job_enter(job), !job_is_completed(job))); > + AIO_WAIT_WHILE(job_get_aio_context(job), > + (job_enter(job), !job_is_completed(job))); > job_lock(); Here I think we are also holding the BQL, because this function is "sync", so it's safe to use job->aio_context. Paolo
On 19/01/2022 11:31, Paolo Bonzini wrote: >> diff --git a/blockjob.c b/blockjob.c >> index cf1f49f6c2..468ba735c5 100644 >> --- a/blockjob.c >> +++ b/blockjob.c >> @@ -155,14 +155,16 @@ 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; >> + WITH_JOB_LOCK_GUARD() { >> + job->job.aio_context = ctx; >> + } >> } >> static AioContext *child_job_get_parent_aio_context(BdrvChild *c) >> { >> BlockJob *job = c->opaque; >> - return job->job.aio_context; >> + return job_get_aio_context(&job->job); >> } >> static const BdrvChildClass child_job = { > > Both called with BQL held, I think. Yes, as their callbacks .get_parent_aio_context and .set_aio_context are defined as GS functions in block_int-common.h > >> @@ -218,19 +220,21 @@ int block_job_add_bdrv(BlockJob *job, const char >> *name, BlockDriverState *bs, >> { >> BdrvChild *c; >> bool need_context_ops; >> + AioContext *job_aiocontext; >> assert(qemu_in_main_thread()); >> bdrv_ref(bs); >> - need_context_ops = bdrv_get_aio_context(bs) != job->job.aio_context; >> + job_aiocontext = job_get_aio_context(&job->job); >> + need_context_ops = bdrv_get_aio_context(bs) != job_aiocontext; >> - if (need_context_ops && job->job.aio_context != >> qemu_get_aio_context()) { >> - aio_context_release(job->job.aio_context); >> + if (need_context_ops && job_aiocontext != qemu_get_aio_context()) { >> + aio_context_release(job_aiocontext); >> } >> c = bdrv_root_attach_child(bs, name, &child_job, 0, perm, >> shared_perm, job, >> errp); >> - if (need_context_ops && job->job.aio_context != >> qemu_get_aio_context()) { >> - aio_context_acquire(job->job.aio_context); >> + if (need_context_ops && job_aiocontext != qemu_get_aio_context()) { >> + aio_context_acquire(job_aiocontext); >> } >> if (c == NULL) { >> return -EPERM; > > BQL held, too. Wouldn't it be better to keep job_get_aio_context and implement like this: AioContext *job_get_aio_context(Job *job) { /* * Job AioContext can be written under BQL+job_mutex, * but can be read with just the BQL held. */ assert(qemu_in_main_thread()); return job->aio_context; } and instead job_set_aio_context: void job_set_aio_context(Job *job, AioContext *ctx) { JOB_LOCK_GUARD(); assert(qemu_in_main_thread()); job->aio_context = ctx; } (obviously implement also _locked version, if needed, and probably move the comment in get_aio_context in job.h). Thank you, Emanuele
On 19/01/2022 11:31, Paolo Bonzini wrote: > >> diff --git a/job.c b/job.c >> index f16a4ef542..8a5b710d9b 100644 >> --- a/job.c >> +++ b/job.c >> @@ -566,7 +566,7 @@ void job_enter_cond_locked(Job *job, bool(*fn)(Job >> *job)) >> job->busy = true; >> real_job_unlock(); >> job_unlock(); >> - aio_co_enter(job->aio_context, job->co); >> + aio_co_enter(job_get_aio_context(job), job->co); >> job_lock(); >> } > > If you replace aio_co_enter with aio_co_schedule, you can call it > without dropping the lock. The difference being that aio_co_schedule > will always go through a bottom half. > >> @@ -1138,7 +1138,6 @@ static void coroutine_fn job_co_entry(void *opaque) >> Job *job = opaque; >> int ret; >> - assert(job->aio_context == qemu_get_current_aio_context()); >> assert(job && job->driver && job->driver->run); >> job_pause_point(job); >> ret = job->driver->run(job, &job->err); >> @@ -1177,7 +1176,7 @@ void job_start(Job *job) >> job->paused = false; >> job_state_transition_locked(job, JOB_STATUS_RUNNING); >> } >> - aio_co_enter(job->aio_context, job->co); >> + aio_co_enter(job_get_aio_context(job), job->co); > > Better to use aio_co_schedule here, too, and move it under the previous > WITH_JOB_LOCK_GUARD. Unfortunately this does not work straightforward: aio_co_enter invokes aio_co_schedule only if the context is different from the main loop, otherwise it can directly enter the coroutine with qemu_aio_coroutine_enter. So always replacing it with aio_co_schedule breaks the unit tests assumptions, as they expect that when control is returned the job has already executed. A possible solution is to aio_poll() on the condition we want to assert, waiting for the bh to be scheduled. But I don't know if this is then useful to test something. Thank you, Emanuele
On 21/01/2022 13:33, Emanuele Giuseppe Esposito wrote: > > > On 19/01/2022 11:31, Paolo Bonzini wrote: >>> diff --git a/blockjob.c b/blockjob.c >>> index cf1f49f6c2..468ba735c5 100644 >>> --- a/blockjob.c >>> +++ b/blockjob.c >>> @@ -155,14 +155,16 @@ 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; >>> + WITH_JOB_LOCK_GUARD() { >>> + job->job.aio_context = ctx; >>> + } >>> } >>> static AioContext *child_job_get_parent_aio_context(BdrvChild *c) >>> { >>> BlockJob *job = c->opaque; >>> - return job->job.aio_context; >>> + return job_get_aio_context(&job->job); >>> } >>> static const BdrvChildClass child_job = { >> >> Both called with BQL held, I think. > > Yes, as their callbacks .get_parent_aio_context and .set_aio_context are > defined as GS functions in block_int-common.h >> >>> @@ -218,19 +220,21 @@ int block_job_add_bdrv(BlockJob *job, const >>> char *name, BlockDriverState *bs, >>> { >>> BdrvChild *c; >>> bool need_context_ops; >>> + AioContext *job_aiocontext; >>> assert(qemu_in_main_thread()); >>> bdrv_ref(bs); >>> - need_context_ops = bdrv_get_aio_context(bs) != >>> job->job.aio_context; >>> + job_aiocontext = job_get_aio_context(&job->job); >>> + need_context_ops = bdrv_get_aio_context(bs) != job_aiocontext; >>> - if (need_context_ops && job->job.aio_context != >>> qemu_get_aio_context()) { >>> - aio_context_release(job->job.aio_context); >>> + if (need_context_ops && job_aiocontext != qemu_get_aio_context()) { >>> + aio_context_release(job_aiocontext); >>> } >>> c = bdrv_root_attach_child(bs, name, &child_job, 0, perm, >>> shared_perm, job, >>> errp); >>> - if (need_context_ops && job->job.aio_context != >>> qemu_get_aio_context()) { >>> - aio_context_acquire(job->job.aio_context); >>> + if (need_context_ops && job_aiocontext != qemu_get_aio_context()) { >>> + aio_context_acquire(job_aiocontext); >>> } >>> if (c == NULL) { >>> return -EPERM; >> >> BQL held, too. > > Wouldn't it be better to keep job_get_aio_context and implement like this: > > AioContext *job_get_aio_context(Job *job) > { > /* > * Job AioContext can be written under BQL+job_mutex, > * but can be read with just the BQL held. > */ > assert(qemu_in_main_thread()); > return job->aio_context; > } Uhm ok this one doesn't really work, because it's ok to read it under either BQL or job lock. So I will get rid of job_get_aio_context, but add job_set_aio_context (and use it in child_job_set_aio_ctx). Emanuele > > and instead job_set_aio_context: > > void job_set_aio_context(Job *job, AioContext *ctx) > { > JOB_LOCK_GUARD(); > assert(qemu_in_main_thread()); > job->aio_context = ctx; > } > > (obviously implement also _locked version, if needed, and probably move > the comment in get_aio_context in job.h). > > Thank you, > Emanuele
On 1/21/22 16:18, Emanuele Giuseppe Esposito wrote: >>> >> >> Better to use aio_co_schedule here, too, and move it under the >> previous WITH_JOB_LOCK_GUARD. > > Unfortunately this does not work straightforward: aio_co_enter invokes > aio_co_schedule only if the context is different from the main loop, > otherwise it can directly enter the coroutine with > qemu_aio_coroutine_enter. So always replacing it with aio_co_schedule > breaks the unit tests assumptions, as they expect that when control is > returned the job has already executed. > > A possible solution is to aio_poll() on the condition we want to assert, > waiting for the bh to be scheduled. But I don't know if this is then > useful to test something. I think you sorted that out, based on IRC conversation? Paolo
On 24/01/2022 15:22, Paolo Bonzini wrote: > On 1/21/22 16:18, Emanuele Giuseppe Esposito wrote: >>>> >>> >>> Better to use aio_co_schedule here, too, and move it under the >>> previous WITH_JOB_LOCK_GUARD. >> >> Unfortunately this does not work straightforward: aio_co_enter invokes >> aio_co_schedule only if the context is different from the main loop, >> otherwise it can directly enter the coroutine with >> qemu_aio_coroutine_enter. So always replacing it with aio_co_schedule >> breaks the unit tests assumptions, as they expect that when control is >> returned the job has already executed. >> >> A possible solution is to aio_poll() on the condition we want to >> assert, waiting for the bh to be scheduled. But I don't know if this >> is then useful to test something. > > I think you sorted that out, based on IRC conversation? > Yes. Thank you, Emanuele
diff --git a/block/commit.c b/block/commit.c index f639eb49c5..961b57edf0 100644 --- a/block/commit.c +++ b/block/commit.c @@ -369,7 +369,7 @@ void commit_start(const char *job_id, BlockDriverState *bs, goto fail; } - s->base = blk_new(s->common.job.aio_context, + s->base = blk_new(job_get_aio_context(&s->common.job), base_perms, BLK_PERM_CONSISTENT_READ | BLK_PERM_GRAPH_MOD @@ -382,7 +382,7 @@ void commit_start(const char *job_id, BlockDriverState *bs, s->base_bs = base; /* Required permissions are already taken with block_job_add_bdrv() */ - s->top = blk_new(s->common.job.aio_context, 0, BLK_PERM_ALL); + s->top = blk_new(job_get_aio_context(&s->common.job), 0, BLK_PERM_ALL); ret = blk_insert_bs(s->top, top, errp); if (ret < 0) { goto fail; diff --git a/block/mirror.c b/block/mirror.c index 41450df55c..72b4367b4e 100644 --- a/block/mirror.c +++ b/block/mirror.c @@ -1743,7 +1743,7 @@ static BlockJob *mirror_start_job( target_perms |= BLK_PERM_GRAPH_MOD; } - s->target = blk_new(s->common.job.aio_context, + s->target = blk_new(job_get_aio_context(&s->common.job), target_perms, target_shared_perms); ret = blk_insert_bs(s->target, target, errp); if (ret < 0) { 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 cf1f49f6c2..468ba735c5 100644 --- a/blockjob.c +++ b/blockjob.c @@ -155,14 +155,16 @@ 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; + WITH_JOB_LOCK_GUARD() { + job->job.aio_context = ctx; + } } static AioContext *child_job_get_parent_aio_context(BdrvChild *c) { BlockJob *job = c->opaque; - return job->job.aio_context; + return job_get_aio_context(&job->job); } static const BdrvChildClass child_job = { @@ -218,19 +220,21 @@ int block_job_add_bdrv(BlockJob *job, const char *name, BlockDriverState *bs, { BdrvChild *c; bool need_context_ops; + AioContext *job_aiocontext; assert(qemu_in_main_thread()); bdrv_ref(bs); - need_context_ops = bdrv_get_aio_context(bs) != job->job.aio_context; + job_aiocontext = job_get_aio_context(&job->job); + need_context_ops = bdrv_get_aio_context(bs) != job_aiocontext; - if (need_context_ops && job->job.aio_context != qemu_get_aio_context()) { - aio_context_release(job->job.aio_context); + if (need_context_ops && job_aiocontext != qemu_get_aio_context()) { + aio_context_release(job_aiocontext); } c = bdrv_root_attach_child(bs, name, &child_job, 0, perm, shared_perm, job, errp); - if (need_context_ops && job->job.aio_context != qemu_get_aio_context()) { - aio_context_acquire(job->job.aio_context); + if (need_context_ops && job_aiocontext != qemu_get_aio_context()) { + aio_context_acquire(job_aiocontext); } if (c == NULL) { return -EPERM; diff --git a/job.c b/job.c index f16a4ef542..8a5b710d9b 100644 --- a/job.c +++ b/job.c @@ -566,7 +566,7 @@ void job_enter_cond_locked(Job *job, bool(*fn)(Job *job)) job->busy = true; real_job_unlock(); job_unlock(); - aio_co_enter(job->aio_context, job->co); + aio_co_enter(job_get_aio_context(job), job->co); job_lock(); } @@ -1138,7 +1138,6 @@ static void coroutine_fn job_co_entry(void *opaque) Job *job = opaque; int ret; - assert(job->aio_context == qemu_get_current_aio_context()); assert(job && job->driver && job->driver->run); job_pause_point(job); ret = job->driver->run(job, &job->err); @@ -1177,7 +1176,7 @@ void job_start(Job *job) job->paused = false; job_state_transition_locked(job, JOB_STATUS_RUNNING); } - aio_co_enter(job->aio_context, job->co); + aio_co_enter(job_get_aio_context(job), job->co); } void job_cancel_locked(Job *job, bool force) @@ -1303,7 +1302,8 @@ int job_finish_sync_locked(Job *job, void (*finish)(Job *, Error **errp), } job_unlock(); - AIO_WAIT_WHILE(job->aio_context, (job_enter(job), !job_is_completed(job))); + AIO_WAIT_WHILE(job_get_aio_context(job), + (job_enter(job), !job_is_completed(job))); job_lock(); ret = (job_is_cancelled_locked(job) && job->ret == 0) ? diff --git a/tests/unit/test-block-iothread.c b/tests/unit/test-block-iothread.c index b9309beec2..addcb5846b 100644 --- a/tests/unit/test-block-iothread.c +++ b/tests/unit/test-block-iothread.c @@ -379,7 +379,7 @@ static int coroutine_fn test_job_run(Job *job, Error **errp) job_transition_to_ready(&s->common.job); while (!s->should_complete) { s->n++; - g_assert(qemu_get_current_aio_context() == job->aio_context); + g_assert(qemu_get_current_aio_context() == job_get_aio_context(job)); /* Avoid job_sleep_ns() because it marks the job as !busy. We want to * emulate some actual activity (probably some I/O) here so that the @@ -390,7 +390,7 @@ static int coroutine_fn test_job_run(Job *job, Error **errp) job_pause_point(&s->common.job); } - g_assert(qemu_get_current_aio_context() == job->aio_context); + g_assert(qemu_get_current_aio_context() == job_get_aio_context(job)); return 0; } @@ -642,7 +642,7 @@ static void test_propagate_mirror(void) g_assert(bdrv_get_aio_context(src) == ctx); g_assert(bdrv_get_aio_context(target) == ctx); g_assert(bdrv_get_aio_context(filter) == ctx); - g_assert(job->aio_context == ctx); + g_assert(job_get_aio_context(job) == ctx); /* Change the AioContext of target */ aio_context_acquire(ctx);
If the job->aio_context is accessed under job_mutex, leave as it is. Otherwise use job_get_aio_context(). Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com> --- block/commit.c | 4 ++-- block/mirror.c | 2 +- block/replication.c | 2 +- blockjob.c | 18 +++++++++++------- job.c | 8 ++++---- tests/unit/test-block-iothread.c | 6 +++--- 6 files changed, 22 insertions(+), 18 deletions(-)