[v2,23/23] migration/multifd: Optimize sender side to be lockless

Commit Message

Peter Xu Feb. 2, 2024, 10:28 a.m. UTC

From: Peter Xu <[e-mail]>

When reviewing my attempt to refactor send_prepare(), Fabiano suggested we
try out with dropping the mutex in multifd code [1].

I thought about that before but I never tried to change the code.  Now
maybe it's time to give it a stab.  This only optimizes the sender side.

The trick here is multifd has a clear provider/consumer model, that the
migration main thread publishes requests (either pending_job/pending_sync),
while the multifd sender threads are consumers.  Here we don't have a lot
of comlicated data sharing, and the jobs can logically be submitted lockless.

Arm the code with atomic weapons.  Two things worth mentioning:

  - For multifd_send_pages(): we can use qatomic_load_acquire() when trying
  to find a free channel, but that's expensive if we attach one ACQUIRE per
  channel.  Instead, make it atomic_read() on the pending_job flag, but
  merge the ACQUIRE into one single smp_mb_acquire() later.

  - For pending_sync: it doesn't have any extra data required since now
  p->flags are never touched, it should be safe to not use memory barrier.
  That's different from pending_sync.

Provide rich comments for all the lockless operations to state how they are
paired.  With that, we can remove the mutex.

[1] https://lore.kernel.org/r/[e-mail]

Suggested-by: Fabiano Rosas <[e-mail]>
Signed-off-by: Peter Xu <[e-mail]>
---
 migration/multifd.h |  2 --
 migration/multifd.c | 51 +++++++++++++++++++++++----------------------
 2 files changed, 26 insertions(+), 27 deletions(-)

Comments

Fabiano Rosas Feb. 2, 2024, 9:34 p.m. UTC | #1

[e-mail] writes:

> From: Peter Xu <[e-mail]>
>
> When reviewing my attempt to refactor send_prepare(), Fabiano suggested we
> try out with dropping the mutex in multifd code [1].
>
> I thought about that before but I never tried to change the code.  Now
> maybe it's time to give it a stab.  This only optimizes the sender side.
>
> The trick here is multifd has a clear provider/consumer model, that the
> migration main thread publishes requests (either pending_job/pending_sync),
> while the multifd sender threads are consumers.  Here we don't have a lot
> of comlicated data sharing, and the jobs can logically be submitted lockless.

complicated

>
> Arm the code with atomic weapons.  Two things worth mentioning:
>
>   - For multifd_send_pages(): we can use qatomic_load_acquire() when trying
>   to find a free channel, but that's expensive if we attach one ACQUIRE per
>   channel.  Instead, make it atomic_read() on the pending_job flag, but

s/make it/keep it/

The diff doesn't show the atomic_read already there so it's confusing.

>   merge the ACQUIRE into one single smp_mb_acquire() later.
>
>   - For pending_sync: it doesn't have any extra data required since now
>   p->flags are never touched, it should be safe to not use memory barrier.
>   That's different from pending_sync.

pending_job?

>
> Provide rich comments for all the lockless operations to state how they are
> paired.  With that, we can remove the mutex.
>
> [1] https://lore.kernel.org/r/[e-mail]
>
> Suggested-by: Fabiano Rosas <[e-mail]>
> Signed-off-by: Peter Xu <[e-mail]>
> ---
>  migration/multifd.h |  2 --
>  migration/multifd.c | 51 +++++++++++++++++++++++----------------------
>  2 files changed, 26 insertions(+), 27 deletions(-)
>
> diff --git a/migration/multifd.h b/migration/multifd.h
> index 98876ff94a..78a2317263 100644
> --- a/migration/multifd.h
> +++ b/migration/multifd.h
> @@ -91,8 +91,6 @@ typedef struct {
>      /* syncs main thread and channels */
>      QemuSemaphore sem_sync;
>  
> -    /* this mutex protects the following parameters */
> -    QemuMutex mutex;
>      /* is this channel thread running */
>      bool running;
>      /* multifd flags for each packet */
> diff --git a/migration/multifd.c b/migration/multifd.c
> index b317d57d61..ef13e2e781 100644
> --- a/migration/multifd.c
> +++ b/migration/multifd.c
> @@ -501,19 +501,19 @@ static bool multifd_send_pages(void)
>          }
>      }
>  
> -    qemu_mutex_lock(&p->mutex);
> -    assert(!p->pages->num);
> -    assert(!p->pages->block);
>      /*
> -     * Double check on pending_job==false with the lock.  In the future if
> -     * we can have >1 requester thread, we can replace this with a "goto
> -     * retry", but that is for later.
> +     * Make sure we read p->pending_job before all the rest.  Pairs with
> +     * qatomic_store_release() in multifd_send_thread().
>       */
> -    assert(qatomic_read(&p->pending_job) == false);
> -    qatomic_set(&p->pending_job, true);
> +    smp_mb_acquire();
> +    assert(!p->pages->num);
>      multifd_send_state->pages = p->pages;
>      p->pages = pages;
> -    qemu_mutex_unlock(&p->mutex);
> +    /*
> +     * Making sure p->pages is setup before marking pending_job=true. Pairs
> +     * with the qatomic_load_acquire() in multifd_send_thread().
> +     */
> +    qatomic_store_release(&p->pending_job, true);
>      qemu_sem_post(&p->sem);
>  
>      return true;
> @@ -648,7 +648,6 @@ static bool multifd_send_cleanup_channel(MultiFDSendParams *p, Error **errp)
>      }
>      multifd_send_channel_destroy(p->c);
>      p->c = NULL;
> -    qemu_mutex_destroy(&p->mutex);
>      qemu_sem_destroy(&p->sem);
>      qemu_sem_destroy(&p->sem_sync);
>      g_free(p->name);
> @@ -742,14 +741,12 @@ int multifd_send_sync_main(void)
>  
>          trace_multifd_send_sync_main_signal(p->id);
>  
> -        qemu_mutex_lock(&p->mutex);
>          /*
>           * We should be the only user so far, so not possible to be set by
>           * others concurrently.
>           */
>          assert(qatomic_read(&p->pending_sync) == false);
>          qatomic_set(&p->pending_sync, true);
> -        qemu_mutex_unlock(&p->mutex);
>          qemu_sem_post(&p->sem);
>      }
>      for (i = 0; i < migrate_multifd_channels(); i++) {
> @@ -796,9 +793,12 @@ static void *multifd_send_thread(void *opaque)
>          if (multifd_send_should_exit()) {
>              break;
>          }
> -        qemu_mutex_lock(&p->mutex);
>  
> -        if (qatomic_read(&p->pending_job)) {
> +        /*
> +         * Read pending_job flag before p->pages.  Pairs with the
> +         * qatomic_store_release() in multifd_send_pages().
> +         */
> +        if (qatomic_load_acquire(&p->pending_job)) {
>              MultiFDPages_t *pages = p->pages;
>  
>              p->iovs_num = 0;
> @@ -806,14 +806,12 @@ static void *multifd_send_thread(void *opaque)
>  
>              ret = multifd_send_state->ops->send_prepare(p, &local_err);
>              if (ret != 0) {
> -                qemu_mutex_unlock(&p->mutex);
>                  break;
>              }
>  
>              ret = qio_channel_writev_full_all(p->c, p->iov, p->iovs_num, NULL,
>                                                0, p->write_flags, &local_err);
>              if (ret != 0) {
> -                qemu_mutex_unlock(&p->mutex);
>                  break;
>              }
>  
> @@ -822,24 +820,31 @@ static void *multifd_send_thread(void *opaque)
>  
>              multifd_pages_reset(p->pages);
>              p->next_packet_size = 0;
> -            qatomic_set(&p->pending_job, false);
> -            qemu_mutex_unlock(&p->mutex);
> +
> +            /*
> +             * Making sure p->pages is published before saying "we're
> +             * free".  Pairs with the qatomic_load_acquire() in

smp_mb_acquire()

> +             * multifd_send_pages().
> +             */
> +            qatomic_store_release(&p->pending_job, false);
>          } else {
> -            /* If not a normal job, must be a sync request */
> +            /*
> +             * If not a normal job, must be a sync request.  Note that
> +             * pending_sync is a standalone flag (unlike pending_job), so
> +             * it doesn't require explicit memory barriers.
> +             */
>              assert(qatomic_read(&p->pending_sync));
>              p->flags = MULTIFD_FLAG_SYNC;
>              multifd_send_fill_packet(p);
>              ret = qio_channel_write_all(p->c, (void *)p->packet,
>                                          p->packet_len, &local_err);
>              if (ret != 0) {
> -                qemu_mutex_unlock(&p->mutex);
>                  break;
>              }
>              /* p->next_packet_size will always be zero for a SYNC packet */
>              stat64_add(&mig_stats.multifd_bytes, p->packet_len);
>              p->flags = 0;
>              qatomic_set(&p->pending_sync, false);
> -            qemu_mutex_unlock(&p->mutex);
>              qemu_sem_post(&p->sem_sync);
>          }
>      }
> @@ -853,10 +858,7 @@ out:
>          error_free(local_err);
>      }
>  
> -    qemu_mutex_lock(&p->mutex);
>      p->running = false;
> -    qemu_mutex_unlock(&p->mutex);
> -
>      rcu_unregister_thread();
>      migration_threads_remove(thread);
>      trace_multifd_send_thread_end(p->id, p->packets_sent, p->total_normal_pages);
> @@ -998,7 +1000,6 @@ int multifd_send_setup(Error **errp)
>      for (i = 0; i < thread_count; i++) {
>          MultiFDSendParams *p = &multifd_send_state->params[i];
>  
> -        qemu_mutex_init(&p->mutex);
>          qemu_sem_init(&p->sem, 0);
>          qemu_sem_init(&p->sem_sync, 0);
>          p->id = i;

Peter Xu Feb. 5, 2024, 4:35 a.m. UTC | #2

On Fri, Feb 02, 2024 at 06:34:08PM -0300, Fabiano Rosas wrote:
> [e-mail] writes:
> 
> > From: Peter Xu <[e-mail]>
> >
> > When reviewing my attempt to refactor send_prepare(), Fabiano suggested we
> > try out with dropping the mutex in multifd code [1].
> >
> > I thought about that before but I never tried to change the code.  Now
> > maybe it's time to give it a stab.  This only optimizes the sender side.
> >
> > The trick here is multifd has a clear provider/consumer model, that the
> > migration main thread publishes requests (either pending_job/pending_sync),
> > while the multifd sender threads are consumers.  Here we don't have a lot
> > of comlicated data sharing, and the jobs can logically be submitted lockless.
> 
> complicated
> 
> >
> > Arm the code with atomic weapons.  Two things worth mentioning:
> >
> >   - For multifd_send_pages(): we can use qatomic_load_acquire() when trying
> >   to find a free channel, but that's expensive if we attach one ACQUIRE per
> >   channel.  Instead, make it atomic_read() on the pending_job flag, but
> 
> s/make it/keep it/
> 
> The diff doesn't show the atomic_read already there so it's confusing.

Right.  I also has a trivial typo on s/atomic_read/qatomic_read/..

I tried to rephrase the last sentence:

  - For multifd_send_pages(): we can use qatomic_load_acquire() when trying
  to find a free channel, but that's expensive if we attach one ACQUIRE per
  channel.  Instead, keep the qatomic_read() on reading the pending_job
  flag as we do already, meanwhile use one smp_mb_acquire() after the loop
  to guarantee the memory ordering.

Maybe slightly clearer?

> 
> >   merge the ACQUIRE into one single smp_mb_acquire() later.
> >
> >   - For pending_sync: it doesn't have any extra data required since now
> >   p->flags are never touched, it should be safe to not use memory barrier.
> >   That's different from pending_sync.
> 
> pending_job?

Yep, all the rest fixed.

> 
> >
> > Provide rich comments for all the lockless operations to state how they are
> > paired.  With that, we can remove the mutex.
> >
> > [1] https://lore.kernel.org/r/[e-mail]
> >
> > Suggested-by: Fabiano Rosas <[e-mail]>
> > Signed-off-by: Peter Xu <[e-mail]>
> > ---
> >  migration/multifd.h |  2 --
> >  migration/multifd.c | 51 +++++++++++++++++++++++----------------------
> >  2 files changed, 26 insertions(+), 27 deletions(-)
> >
> > diff --git a/migration/multifd.h b/migration/multifd.h
> > index 98876ff94a..78a2317263 100644
> > --- a/migration/multifd.h
> > +++ b/migration/multifd.h
> > @@ -91,8 +91,6 @@ typedef struct {
> >      /* syncs main thread and channels */
> >      QemuSemaphore sem_sync;
> >  
> > -    /* this mutex protects the following parameters */
> > -    QemuMutex mutex;
> >      /* is this channel thread running */
> >      bool running;
> >      /* multifd flags for each packet */
> > diff --git a/migration/multifd.c b/migration/multifd.c
> > index b317d57d61..ef13e2e781 100644
> > --- a/migration/multifd.c
> > +++ b/migration/multifd.c
> > @@ -501,19 +501,19 @@ static bool multifd_send_pages(void)
> >          }
> >      }
> >  
> > -    qemu_mutex_lock(&p->mutex);
> > -    assert(!p->pages->num);
> > -    assert(!p->pages->block);
> >      /*
> > -     * Double check on pending_job==false with the lock.  In the future if
> > -     * we can have >1 requester thread, we can replace this with a "goto
> > -     * retry", but that is for later.
> > +     * Make sure we read p->pending_job before all the rest.  Pairs with
> > +     * qatomic_store_release() in multifd_send_thread().
> >       */
> > -    assert(qatomic_read(&p->pending_job) == false);
> > -    qatomic_set(&p->pending_job, true);
> > +    smp_mb_acquire();
> > +    assert(!p->pages->num);
> >      multifd_send_state->pages = p->pages;
> >      p->pages = pages;
> > -    qemu_mutex_unlock(&p->mutex);
> > +    /*
> > +     * Making sure p->pages is setup before marking pending_job=true. Pairs
> > +     * with the qatomic_load_acquire() in multifd_send_thread().
> > +     */
> > +    qatomic_store_release(&p->pending_job, true);
> >      qemu_sem_post(&p->sem);
> >  
> >      return true;
> > @@ -648,7 +648,6 @@ static bool multifd_send_cleanup_channel(MultiFDSendParams *p, Error **errp)
> >      }
> >      multifd_send_channel_destroy(p->c);
> >      p->c = NULL;
> > -    qemu_mutex_destroy(&p->mutex);
> >      qemu_sem_destroy(&p->sem);
> >      qemu_sem_destroy(&p->sem_sync);
> >      g_free(p->name);
> > @@ -742,14 +741,12 @@ int multifd_send_sync_main(void)
> >  
> >          trace_multifd_send_sync_main_signal(p->id);
> >  
> > -        qemu_mutex_lock(&p->mutex);
> >          /*
> >           * We should be the only user so far, so not possible to be set by
> >           * others concurrently.
> >           */
> >          assert(qatomic_read(&p->pending_sync) == false);
> >          qatomic_set(&p->pending_sync, true);
> > -        qemu_mutex_unlock(&p->mutex);
> >          qemu_sem_post(&p->sem);
> >      }
> >      for (i = 0; i < migrate_multifd_channels(); i++) {
> > @@ -796,9 +793,12 @@ static void *multifd_send_thread(void *opaque)
> >          if (multifd_send_should_exit()) {
> >              break;
> >          }
> > -        qemu_mutex_lock(&p->mutex);
> >  
> > -        if (qatomic_read(&p->pending_job)) {
> > +        /*
> > +         * Read pending_job flag before p->pages.  Pairs with the
> > +         * qatomic_store_release() in multifd_send_pages().
> > +         */
> > +        if (qatomic_load_acquire(&p->pending_job)) {
> >              MultiFDPages_t *pages = p->pages;
> >  
> >              p->iovs_num = 0;
> > @@ -806,14 +806,12 @@ static void *multifd_send_thread(void *opaque)
> >  
> >              ret = multifd_send_state->ops->send_prepare(p, &local_err);
> >              if (ret != 0) {
> > -                qemu_mutex_unlock(&p->mutex);
> >                  break;
> >              }
> >  
> >              ret = qio_channel_writev_full_all(p->c, p->iov, p->iovs_num, NULL,
> >                                                0, p->write_flags, &local_err);
> >              if (ret != 0) {
> > -                qemu_mutex_unlock(&p->mutex);
> >                  break;
> >              }
> >  
> > @@ -822,24 +820,31 @@ static void *multifd_send_thread(void *opaque)
> >  
> >              multifd_pages_reset(p->pages);
> >              p->next_packet_size = 0;
> > -            qatomic_set(&p->pending_job, false);
> > -            qemu_mutex_unlock(&p->mutex);
> > +
> > +            /*
> > +             * Making sure p->pages is published before saying "we're
> > +             * free".  Pairs with the qatomic_load_acquire() in
> 
> smp_mb_acquire()

Fixed.

Any more comment on the code changes before I repost?

(maybe I can repost this single patch in-place to avoid another round of
 mail bombs..)

Fabiano Rosas Feb. 5, 2024, 2:10 p.m. UTC | #3

Peter Xu <[e-mail]> writes:

> On Fri, Feb 02, 2024 at 06:34:08PM -0300, Fabiano Rosas wrote:
>> [e-mail] writes:
>> 
>> > From: Peter Xu <[e-mail]>
>> >
>> > When reviewing my attempt to refactor send_prepare(), Fabiano suggested we
>> > try out with dropping the mutex in multifd code [1].
>> >
>> > I thought about that before but I never tried to change the code.  Now
>> > maybe it's time to give it a stab.  This only optimizes the sender side.
>> >
>> > The trick here is multifd has a clear provider/consumer model, that the
>> > migration main thread publishes requests (either pending_job/pending_sync),
>> > while the multifd sender threads are consumers.  Here we don't have a lot
>> > of comlicated data sharing, and the jobs can logically be submitted lockless.
>> 
>> complicated
>> 
>> >
>> > Arm the code with atomic weapons.  Two things worth mentioning:
>> >
>> >   - For multifd_send_pages(): we can use qatomic_load_acquire() when trying
>> >   to find a free channel, but that's expensive if we attach one ACQUIRE per
>> >   channel.  Instead, make it atomic_read() on the pending_job flag, but
>> 
>> s/make it/keep it/
>> 
>> The diff doesn't show the atomic_read already there so it's confusing.
>
> Right.  I also has a trivial typo on s/atomic_read/qatomic_read/..
>
> I tried to rephrase the last sentence:
>
>   - For multifd_send_pages(): we can use qatomic_load_acquire() when trying
>   to find a free channel, but that's expensive if we attach one ACQUIRE per
>   channel.  Instead, keep the qatomic_read() on reading the pending_job
>   flag as we do already, meanwhile use one smp_mb_acquire() after the loop
>   to guarantee the memory ordering.
>
> Maybe slightly clearer?
>

Yep, that's better. Thanks.

>> 
>> >   merge the ACQUIRE into one single smp_mb_acquire() later.
>> >
>> >   - For pending_sync: it doesn't have any extra data required since now
>> >   p->flags are never touched, it should be safe to not use memory barrier.
>> >   That's different from pending_sync.
>> 
>> pending_job?
>
> Yep, all the rest fixed.
>
>> 
>> >
>> > Provide rich comments for all the lockless operations to state how they are
>> > paired.  With that, we can remove the mutex.
>> >
>> > [1] https://lore.kernel.org/r/[e-mail]
>> >
>> > Suggested-by: Fabiano Rosas <[e-mail]>
>> > Signed-off-by: Peter Xu <[e-mail]>
>> > ---
>> >  migration/multifd.h |  2 --
>> >  migration/multifd.c | 51 +++++++++++++++++++++++----------------------
>> >  2 files changed, 26 insertions(+), 27 deletions(-)
>> >
>> > diff --git a/migration/multifd.h b/migration/multifd.h
>> > index 98876ff94a..78a2317263 100644
>> > --- a/migration/multifd.h
>> > +++ b/migration/multifd.h
>> > @@ -91,8 +91,6 @@ typedef struct {
>> >      /* syncs main thread and channels */
>> >      QemuSemaphore sem_sync;
>> >  
>> > -    /* this mutex protects the following parameters */
>> > -    QemuMutex mutex;
>> >      /* is this channel thread running */
>> >      bool running;
>> >      /* multifd flags for each packet */
>> > diff --git a/migration/multifd.c b/migration/multifd.c
>> > index b317d57d61..ef13e2e781 100644
>> > --- a/migration/multifd.c
>> > +++ b/migration/multifd.c
>> > @@ -501,19 +501,19 @@ static bool multifd_send_pages(void)
>> >          }
>> >      }
>> >  
>> > -    qemu_mutex_lock(&p->mutex);
>> > -    assert(!p->pages->num);
>> > -    assert(!p->pages->block);
>> >      /*
>> > -     * Double check on pending_job==false with the lock.  In the future if
>> > -     * we can have >1 requester thread, we can replace this with a "goto
>> > -     * retry", but that is for later.
>> > +     * Make sure we read p->pending_job before all the rest.  Pairs with
>> > +     * qatomic_store_release() in multifd_send_thread().
>> >       */
>> > -    assert(qatomic_read(&p->pending_job) == false);
>> > -    qatomic_set(&p->pending_job, true);
>> > +    smp_mb_acquire();
>> > +    assert(!p->pages->num);
>> >      multifd_send_state->pages = p->pages;
>> >      p->pages = pages;
>> > -    qemu_mutex_unlock(&p->mutex);
>> > +    /*
>> > +     * Making sure p->pages is setup before marking pending_job=true. Pairs
>> > +     * with the qatomic_load_acquire() in multifd_send_thread().
>> > +     */
>> > +    qatomic_store_release(&p->pending_job, true);
>> >      qemu_sem_post(&p->sem);
>> >  
>> >      return true;
>> > @@ -648,7 +648,6 @@ static bool multifd_send_cleanup_channel(MultiFDSendParams *p, Error **errp)
>> >      }
>> >      multifd_send_channel_destroy(p->c);
>> >      p->c = NULL;
>> > -    qemu_mutex_destroy(&p->mutex);
>> >      qemu_sem_destroy(&p->sem);
>> >      qemu_sem_destroy(&p->sem_sync);
>> >      g_free(p->name);
>> > @@ -742,14 +741,12 @@ int multifd_send_sync_main(void)
>> >  
>> >          trace_multifd_send_sync_main_signal(p->id);
>> >  
>> > -        qemu_mutex_lock(&p->mutex);
>> >          /*
>> >           * We should be the only user so far, so not possible to be set by
>> >           * others concurrently.
>> >           */
>> >          assert(qatomic_read(&p->pending_sync) == false);
>> >          qatomic_set(&p->pending_sync, true);
>> > -        qemu_mutex_unlock(&p->mutex);
>> >          qemu_sem_post(&p->sem);
>> >      }
>> >      for (i = 0; i < migrate_multifd_channels(); i++) {
>> > @@ -796,9 +793,12 @@ static void *multifd_send_thread(void *opaque)
>> >          if (multifd_send_should_exit()) {
>> >              break;
>> >          }
>> > -        qemu_mutex_lock(&p->mutex);
>> >  
>> > -        if (qatomic_read(&p->pending_job)) {
>> > +        /*
>> > +         * Read pending_job flag before p->pages.  Pairs with the
>> > +         * qatomic_store_release() in multifd_send_pages().
>> > +         */
>> > +        if (qatomic_load_acquire(&p->pending_job)) {
>> >              MultiFDPages_t *pages = p->pages;
>> >  
>> >              p->iovs_num = 0;
>> > @@ -806,14 +806,12 @@ static void *multifd_send_thread(void *opaque)
>> >  
>> >              ret = multifd_send_state->ops->send_prepare(p, &local_err);
>> >              if (ret != 0) {
>> > -                qemu_mutex_unlock(&p->mutex);
>> >                  break;
>> >              }
>> >  
>> >              ret = qio_channel_writev_full_all(p->c, p->iov, p->iovs_num, NULL,
>> >                                                0, p->write_flags, &local_err);
>> >              if (ret != 0) {
>> > -                qemu_mutex_unlock(&p->mutex);
>> >                  break;
>> >              }
>> >  
>> > @@ -822,24 +820,31 @@ static void *multifd_send_thread(void *opaque)
>> >  
>> >              multifd_pages_reset(p->pages);
>> >              p->next_packet_size = 0;
>> > -            qatomic_set(&p->pending_job, false);
>> > -            qemu_mutex_unlock(&p->mutex);
>> > +
>> > +            /*
>> > +             * Making sure p->pages is published before saying "we're
>> > +             * free".  Pairs with the qatomic_load_acquire() in
>> 
>> smp_mb_acquire()
>
> Fixed.
>
> Any more comment on the code changes before I repost?

Nope, that's it.

>
> (maybe I can repost this single patch in-place to avoid another round of
>  mail bombs..)

Sure.

Peter Xu Feb. 5, 2024, 2:24 p.m. UTC | #4

On Mon, Feb 05, 2024 at 11:10:34AM -0300, Fabiano Rosas wrote:
> > (maybe I can repost this single patch in-place to avoid another round of
> >  mail bombs..)
> 
> Sure.

I've got the final version attached here.  Feel free to have a look, thanks.

====
From 6ba337320430feae4ce9d3d906ea19f68430642d Mon Sep 17 00:00:00 2001
From: Peter Xu <[e-mail]>
Date: Fri, 2 Feb 2024 18:28:57 +0800
Subject: [PATCH] migration/multifd: Optimize sender side to be lockless

When reviewing my attempt to refactor send_prepare(), Fabiano suggested we
try out with dropping the mutex in multifd code [1].

I thought about that before but I never tried to change the code.  Now
maybe it's time to give it a stab.  This only optimizes the sender side.

The trick here is multifd has a clear provider/consumer model, that the
migration main thread publishes requests (either pending_job/pending_sync),
while the multifd sender threads are consumers.  Here we don't have a lot
of complicated data sharing, and the jobs can logically be submitted
lockless.

Arm the code with atomic weapons.  Two things worth mentioning:

  - For multifd_send_pages(): we can use qatomic_load_acquire() when trying
  to find a free channel, but that's expensive if we attach one ACQUIRE per
  channel.  Instead, keep the qatomic_read() on reading the pending_job
  flag as we do already, meanwhile use one smp_mb_acquire() after the loop
  to guarantee the memory ordering.

  - For pending_sync: it doesn't have any extra data required since now
  p->flags are never touched, it should be safe to not use memory barrier.
  That's different from pending_job.

Provide rich comments for all the lockless operations to state how they are
paired.  With that, we can remove the mutex.

[1] https://lore.kernel.org/r/[e-mail]

Suggested-by: Fabiano Rosas <[e-mail]>
Link: https://lore.kernel.org/r/[e-mail]
Signed-off-by: Peter Xu <[e-mail]>
---
 migration/multifd.h |  2 --
 migration/multifd.c | 51 +++++++++++++++++++++++----------------------
 2 files changed, 26 insertions(+), 27 deletions(-)

diff --git a/migration/multifd.h b/migration/multifd.h
index 98876ff94a..78a2317263 100644
--- a/migration/multifd.h
+++ b/migration/multifd.h
@@ -91,8 +91,6 @@ typedef struct {
     /* syncs main thread and channels */
     QemuSemaphore sem_sync;
 
-    /* this mutex protects the following parameters */
-    QemuMutex mutex;
     /* is this channel thread running */
     bool running;
     /* multifd flags for each packet */
diff --git a/migration/multifd.c b/migration/multifd.c
index b317d57d61..fbdb129088 100644
--- a/migration/multifd.c
+++ b/migration/multifd.c
@@ -501,19 +501,19 @@ static bool multifd_send_pages(void)
         }
     }
 
-    qemu_mutex_lock(&p->mutex);
-    assert(!p->pages->num);
-    assert(!p->pages->block);
     /*
-     * Double check on pending_job==false with the lock.  In the future if
-     * we can have >1 requester thread, we can replace this with a "goto
-     * retry", but that is for later.
+     * Make sure we read p->pending_job before all the rest.  Pairs with
+     * qatomic_store_release() in multifd_send_thread().
      */
-    assert(qatomic_read(&p->pending_job) == false);
-    qatomic_set(&p->pending_job, true);
+    smp_mb_acquire();
+    assert(!p->pages->num);
     multifd_send_state->pages = p->pages;
     p->pages = pages;
-    qemu_mutex_unlock(&p->mutex);
+    /*
+     * Making sure p->pages is setup before marking pending_job=true. Pairs
+     * with the qatomic_load_acquire() in multifd_send_thread().
+     */
+    qatomic_store_release(&p->pending_job, true);
     qemu_sem_post(&p->sem);
 
     return true;
@@ -648,7 +648,6 @@ static bool multifd_send_cleanup_channel(MultiFDSendParams *p, Error **errp)
     }
     multifd_send_channel_destroy(p->c);
     p->c = NULL;
-    qemu_mutex_destroy(&p->mutex);
     qemu_sem_destroy(&p->sem);
     qemu_sem_destroy(&p->sem_sync);
     g_free(p->name);
@@ -742,14 +741,12 @@ int multifd_send_sync_main(void)
 
         trace_multifd_send_sync_main_signal(p->id);
 
-        qemu_mutex_lock(&p->mutex);
         /*
          * We should be the only user so far, so not possible to be set by
          * others concurrently.
          */
         assert(qatomic_read(&p->pending_sync) == false);
         qatomic_set(&p->pending_sync, true);
-        qemu_mutex_unlock(&p->mutex);
         qemu_sem_post(&p->sem);
     }
     for (i = 0; i < migrate_multifd_channels(); i++) {
@@ -796,9 +793,12 @@ static void *multifd_send_thread(void *opaque)
         if (multifd_send_should_exit()) {
             break;
         }
-        qemu_mutex_lock(&p->mutex);
 
-        if (qatomic_read(&p->pending_job)) {
+        /*
+         * Read pending_job flag before p->pages.  Pairs with the
+         * qatomic_store_release() in multifd_send_pages().
+         */
+        if (qatomic_load_acquire(&p->pending_job)) {
             MultiFDPages_t *pages = p->pages;
 
             p->iovs_num = 0;
@@ -806,14 +806,12 @@ static void *multifd_send_thread(void *opaque)
 
             ret = multifd_send_state->ops->send_prepare(p, &local_err);
             if (ret != 0) {
-                qemu_mutex_unlock(&p->mutex);
                 break;
             }
 
             ret = qio_channel_writev_full_all(p->c, p->iov, p->iovs_num, NULL,
                                               0, p->write_flags, &local_err);
             if (ret != 0) {
-                qemu_mutex_unlock(&p->mutex);
                 break;
             }
 
@@ -822,24 +820,31 @@ static void *multifd_send_thread(void *opaque)
 
             multifd_pages_reset(p->pages);
             p->next_packet_size = 0;
-            qatomic_set(&p->pending_job, false);
-            qemu_mutex_unlock(&p->mutex);
+
+            /*
+             * Making sure p->pages is published before saying "we're
+             * free".  Pairs with the smp_mb_acquire() in
+             * multifd_send_pages().
+             */
+            qatomic_store_release(&p->pending_job, false);
         } else {
-            /* If not a normal job, must be a sync request */
+            /*
+             * If not a normal job, must be a sync request.  Note that
+             * pending_sync is a standalone flag (unlike pending_job), so
+             * it doesn't require explicit memory barriers.
+             */
             assert(qatomic_read(&p->pending_sync));
             p->flags = MULTIFD_FLAG_SYNC;
             multifd_send_fill_packet(p);
             ret = qio_channel_write_all(p->c, (void *)p->packet,
                                         p->packet_len, &local_err);
             if (ret != 0) {
-                qemu_mutex_unlock(&p->mutex);
                 break;
             }
             /* p->next_packet_size will always be zero for a SYNC packet */
             stat64_add(&mig_stats.multifd_bytes, p->packet_len);
             p->flags = 0;
             qatomic_set(&p->pending_sync, false);
-            qemu_mutex_unlock(&p->mutex);
             qemu_sem_post(&p->sem_sync);
         }
     }
@@ -853,10 +858,7 @@ out:
         error_free(local_err);
     }
 
-    qemu_mutex_lock(&p->mutex);
     p->running = false;
-    qemu_mutex_unlock(&p->mutex);
-
     rcu_unregister_thread();
     migration_threads_remove(thread);
     trace_multifd_send_thread_end(p->id, p->packets_sent, p->total_normal_pages);
@@ -998,7 +1000,6 @@ int multifd_send_setup(Error **errp)
     for (i = 0; i < thread_count; i++) {
         MultiFDSendParams *p = &multifd_send_state->params[i];
 
-        qemu_mutex_init(&p->mutex);
         qemu_sem_init(&p->sem, 0);
         qemu_sem_init(&p->sem_sync, 0);
         p->id = i;

Fabiano Rosas Feb. 5, 2024, 5:59 p.m. UTC | #5

Peter Xu <[e-mail]> writes:

> On Mon, Feb 05, 2024 at 11:10:34AM -0300, Fabiano Rosas wrote:
>> > (maybe I can repost this single patch in-place to avoid another round of
>> >  mail bombs..)
>> 
>> Sure.
>
> I've got the final version attached here.  Feel free to have a look, thanks.
>
> ====
> From 6ba337320430feae4ce9d3d906ea19f68430642d Mon Sep 17 00:00:00 2001
> From: Peter Xu <[e-mail]>
> Date: Fri, 2 Feb 2024 18:28:57 +0800
> Subject: [PATCH] migration/multifd: Optimize sender side to be lockless
>
> When reviewing my attempt to refactor send_prepare(), Fabiano suggested we
> try out with dropping the mutex in multifd code [1].
>
> I thought about that before but I never tried to change the code.  Now
> maybe it's time to give it a stab.  This only optimizes the sender side.
>
> The trick here is multifd has a clear provider/consumer model, that the
> migration main thread publishes requests (either pending_job/pending_sync),
> while the multifd sender threads are consumers.  Here we don't have a lot
> of complicated data sharing, and the jobs can logically be submitted
> lockless.
>
> Arm the code with atomic weapons.  Two things worth mentioning:
>
>   - For multifd_send_pages(): we can use qatomic_load_acquire() when trying
>   to find a free channel, but that's expensive if we attach one ACQUIRE per
>   channel.  Instead, keep the qatomic_read() on reading the pending_job
>   flag as we do already, meanwhile use one smp_mb_acquire() after the loop
>   to guarantee the memory ordering.
>
>   - For pending_sync: it doesn't have any extra data required since now
>   p->flags are never touched, it should be safe to not use memory barrier.
>   That's different from pending_job.
>
> Provide rich comments for all the lockless operations to state how they are
> paired.  With that, we can remove the mutex.
>
> [1] https://lore.kernel.org/r/[e-mail]
>
> Suggested-by: Fabiano Rosas <[e-mail]>
> Link: https://lore.kernel.org/r/[e-mail]
> Signed-off-by: Peter Xu <[e-mail]>
> ---
>  migration/multifd.h |  2 --
>  migration/multifd.c | 51 +++++++++++++++++++++++----------------------
>  2 files changed, 26 insertions(+), 27 deletions(-)
>
> diff --git a/migration/multifd.h b/migration/multifd.h
> index 98876ff94a..78a2317263 100644
> --- a/migration/multifd.h
> +++ b/migration/multifd.h
> @@ -91,8 +91,6 @@ typedef struct {
>      /* syncs main thread and channels */
>      QemuSemaphore sem_sync;
>  
> -    /* this mutex protects the following parameters */
> -    QemuMutex mutex;
>      /* is this channel thread running */
>      bool running;
>      /* multifd flags for each packet */
> diff --git a/migration/multifd.c b/migration/multifd.c
> index b317d57d61..fbdb129088 100644
> --- a/migration/multifd.c
> +++ b/migration/multifd.c
> @@ -501,19 +501,19 @@ static bool multifd_send_pages(void)
>          }
>      }
>  
> -    qemu_mutex_lock(&p->mutex);
> -    assert(!p->pages->num);
> -    assert(!p->pages->block);
>      /*
> -     * Double check on pending_job==false with the lock.  In the future if
> -     * we can have >1 requester thread, we can replace this with a "goto
> -     * retry", but that is for later.
> +     * Make sure we read p->pending_job before all the rest.  Pairs with
> +     * qatomic_store_release() in multifd_send_thread().
>       */
> -    assert(qatomic_read(&p->pending_job) == false);
> -    qatomic_set(&p->pending_job, true);
> +    smp_mb_acquire();
> +    assert(!p->pages->num);
>      multifd_send_state->pages = p->pages;
>      p->pages = pages;
> -    qemu_mutex_unlock(&p->mutex);
> +    /*
> +     * Making sure p->pages is setup before marking pending_job=true. Pairs
> +     * with the qatomic_load_acquire() in multifd_send_thread().
> +     */
> +    qatomic_store_release(&p->pending_job, true);
>      qemu_sem_post(&p->sem);
>  
>      return true;
> @@ -648,7 +648,6 @@ static bool multifd_send_cleanup_channel(MultiFDSendParams *p, Error **errp)
>      }
>      multifd_send_channel_destroy(p->c);
>      p->c = NULL;
> -    qemu_mutex_destroy(&p->mutex);
>      qemu_sem_destroy(&p->sem);
>      qemu_sem_destroy(&p->sem_sync);
>      g_free(p->name);
> @@ -742,14 +741,12 @@ int multifd_send_sync_main(void)
>  
>          trace_multifd_send_sync_main_signal(p->id);
>  
> -        qemu_mutex_lock(&p->mutex);
>          /*
>           * We should be the only user so far, so not possible to be set by
>           * others concurrently.
>           */
>          assert(qatomic_read(&p->pending_sync) == false);
>          qatomic_set(&p->pending_sync, true);
> -        qemu_mutex_unlock(&p->mutex);
>          qemu_sem_post(&p->sem);
>      }
>      for (i = 0; i < migrate_multifd_channels(); i++) {
> @@ -796,9 +793,12 @@ static void *multifd_send_thread(void *opaque)
>          if (multifd_send_should_exit()) {
>              break;
>          }
> -        qemu_mutex_lock(&p->mutex);
>  
> -        if (qatomic_read(&p->pending_job)) {
> +        /*
> +         * Read pending_job flag before p->pages.  Pairs with the
> +         * qatomic_store_release() in multifd_send_pages().
> +         */
> +        if (qatomic_load_acquire(&p->pending_job)) {
>              MultiFDPages_t *pages = p->pages;
>  
>              p->iovs_num = 0;
> @@ -806,14 +806,12 @@ static void *multifd_send_thread(void *opaque)
>  
>              ret = multifd_send_state->ops->send_prepare(p, &local_err);
>              if (ret != 0) {
> -                qemu_mutex_unlock(&p->mutex);
>                  break;
>              }
>  
>              ret = qio_channel_writev_full_all(p->c, p->iov, p->iovs_num, NULL,
>                                                0, p->write_flags, &local_err);
>              if (ret != 0) {
> -                qemu_mutex_unlock(&p->mutex);
>                  break;
>              }
>  
> @@ -822,24 +820,31 @@ static void *multifd_send_thread(void *opaque)
>  
>              multifd_pages_reset(p->pages);
>              p->next_packet_size = 0;
> -            qatomic_set(&p->pending_job, false);
> -            qemu_mutex_unlock(&p->mutex);
> +
> +            /*
> +             * Making sure p->pages is published before saying "we're
> +             * free".  Pairs with the smp_mb_acquire() in
> +             * multifd_send_pages().
> +             */
> +            qatomic_store_release(&p->pending_job, false);
>          } else {
> -            /* If not a normal job, must be a sync request */
> +            /*
> +             * If not a normal job, must be a sync request.  Note that
> +             * pending_sync is a standalone flag (unlike pending_job), so
> +             * it doesn't require explicit memory barriers.
> +             */
>              assert(qatomic_read(&p->pending_sync));
>              p->flags = MULTIFD_FLAG_SYNC;
>              multifd_send_fill_packet(p);
>              ret = qio_channel_write_all(p->c, (void *)p->packet,
>                                          p->packet_len, &local_err);
>              if (ret != 0) {
> -                qemu_mutex_unlock(&p->mutex);
>                  break;
>              }
>              /* p->next_packet_size will always be zero for a SYNC packet */
>              stat64_add(&mig_stats.multifd_bytes, p->packet_len);
>              p->flags = 0;
>              qatomic_set(&p->pending_sync, false);
> -            qemu_mutex_unlock(&p->mutex);
>              qemu_sem_post(&p->sem_sync);
>          }
>      }
> @@ -853,10 +858,7 @@ out:
>          error_free(local_err);
>      }
>  
> -    qemu_mutex_lock(&p->mutex);
>      p->running = false;
> -    qemu_mutex_unlock(&p->mutex);
> -
>      rcu_unregister_thread();
>      migration_threads_remove(thread);
>      trace_multifd_send_thread_end(p->id, p->packets_sent, p->total_normal_pages);
> @@ -998,7 +1000,6 @@ int multifd_send_setup(Error **errp)
>      for (i = 0; i < thread_count; i++) {
>          MultiFDSendParams *p = &multifd_send_state->params[i];
>  
> -        qemu_mutex_init(&p->mutex);
>          qemu_sem_init(&p->sem, 0);
>          qemu_sem_init(&p->sem_sync, 0);
>          p->id = i;
> -- 
> 2.43.0

Reviewed-by: Fabiano Rosas <[e-mail]>

Patch

diff --git a/migration/multifd.h b/migration/multifd.h
index 98876ff94a..78a2317263 100644
--- a/migration/multifd.h
+++ b/migration/multifd.h
@@ -91,8 +91,6 @@  typedef struct {
     /* syncs main thread and channels */
     QemuSemaphore sem_sync;
 
-    /* this mutex protects the following parameters */
-    QemuMutex mutex;
     /* is this channel thread running */
     bool running;
     /* multifd flags for each packet */
diff --git a/migration/multifd.c b/migration/multifd.c
index b317d57d61..ef13e2e781 100644
--- a/migration/multifd.c
+++ b/migration/multifd.c
@@ -501,19 +501,19 @@  static bool multifd_send_pages(void)
         }
     }
 
-    qemu_mutex_lock(&p->mutex);
-    assert(!p->pages->num);
-    assert(!p->pages->block);
     /*
-     * Double check on pending_job==false with the lock.  In the future if
-     * we can have >1 requester thread, we can replace this with a "goto
-     * retry", but that is for later.
+     * Make sure we read p->pending_job before all the rest.  Pairs with
+     * qatomic_store_release() in multifd_send_thread().
      */
-    assert(qatomic_read(&p->pending_job) == false);
-    qatomic_set(&p->pending_job, true);
+    smp_mb_acquire();
+    assert(!p->pages->num);
     multifd_send_state->pages = p->pages;
     p->pages = pages;
-    qemu_mutex_unlock(&p->mutex);
+    /*
+     * Making sure p->pages is setup before marking pending_job=true. Pairs
+     * with the qatomic_load_acquire() in multifd_send_thread().
+     */
+    qatomic_store_release(&p->pending_job, true);
     qemu_sem_post(&p->sem);
 
     return true;
@@ -648,7 +648,6 @@  static bool multifd_send_cleanup_channel(MultiFDSendParams *p, Error **errp)
     }
     multifd_send_channel_destroy(p->c);
     p->c = NULL;
-    qemu_mutex_destroy(&p->mutex);
     qemu_sem_destroy(&p->sem);
     qemu_sem_destroy(&p->sem_sync);
     g_free(p->name);
@@ -742,14 +741,12 @@  int multifd_send_sync_main(void)
 
         trace_multifd_send_sync_main_signal(p->id);
 
-        qemu_mutex_lock(&p->mutex);
         /*
          * We should be the only user so far, so not possible to be set by
          * others concurrently.
          */
         assert(qatomic_read(&p->pending_sync) == false);
         qatomic_set(&p->pending_sync, true);
-        qemu_mutex_unlock(&p->mutex);
         qemu_sem_post(&p->sem);
     }
     for (i = 0; i < migrate_multifd_channels(); i++) {
@@ -796,9 +793,12 @@  static void *multifd_send_thread(void *opaque)
         if (multifd_send_should_exit()) {
             break;
         }
-        qemu_mutex_lock(&p->mutex);
 
-        if (qatomic_read(&p->pending_job)) {
+        /*
+         * Read pending_job flag before p->pages.  Pairs with the
+         * qatomic_store_release() in multifd_send_pages().
+         */
+        if (qatomic_load_acquire(&p->pending_job)) {
             MultiFDPages_t *pages = p->pages;
 
             p->iovs_num = 0;
@@ -806,14 +806,12 @@  static void *multifd_send_thread(void *opaque)
 
             ret = multifd_send_state->ops->send_prepare(p, &local_err);
             if (ret != 0) {
-                qemu_mutex_unlock(&p->mutex);
                 break;
             }
 
             ret = qio_channel_writev_full_all(p->c, p->iov, p->iovs_num, NULL,
                                               0, p->write_flags, &local_err);
             if (ret != 0) {
-                qemu_mutex_unlock(&p->mutex);
                 break;
             }
 
@@ -822,24 +820,31 @@  static void *multifd_send_thread(void *opaque)
 
             multifd_pages_reset(p->pages);
             p->next_packet_size = 0;
-            qatomic_set(&p->pending_job, false);
-            qemu_mutex_unlock(&p->mutex);
+
+            /*
+             * Making sure p->pages is published before saying "we're
+             * free".  Pairs with the qatomic_load_acquire() in
+             * multifd_send_pages().
+             */
+            qatomic_store_release(&p->pending_job, false);
         } else {
-            /* If not a normal job, must be a sync request */
+            /*
+             * If not a normal job, must be a sync request.  Note that
+             * pending_sync is a standalone flag (unlike pending_job), so
+             * it doesn't require explicit memory barriers.
+             */
             assert(qatomic_read(&p->pending_sync));
             p->flags = MULTIFD_FLAG_SYNC;
             multifd_send_fill_packet(p);
             ret = qio_channel_write_all(p->c, (void *)p->packet,
                                         p->packet_len, &local_err);
             if (ret != 0) {
-                qemu_mutex_unlock(&p->mutex);
                 break;
             }
             /* p->next_packet_size will always be zero for a SYNC packet */
             stat64_add(&mig_stats.multifd_bytes, p->packet_len);
             p->flags = 0;
             qatomic_set(&p->pending_sync, false);
-            qemu_mutex_unlock(&p->mutex);
             qemu_sem_post(&p->sem_sync);
         }
     }
@@ -853,10 +858,7 @@  out:
         error_free(local_err);
     }
 
-    qemu_mutex_lock(&p->mutex);
     p->running = false;
-    qemu_mutex_unlock(&p->mutex);
-
     rcu_unregister_thread();
     migration_threads_remove(thread);
     trace_multifd_send_thread_end(p->id, p->packets_sent, p->total_normal_pages);
@@ -998,7 +1000,6 @@  int multifd_send_setup(Error **errp)
     for (i = 0; i < thread_count; i++) {
         MultiFDSendParams *p = &multifd_send_state->params[i];
 
-        qemu_mutex_init(&p->mutex);
         qemu_sem_init(&p->sem, 0);
         qemu_sem_init(&p->sem_sync, 0);
         p->id = i;