[ovs-dev,v2,2/5] dpif-netdev: Trigger parallel pmd reloads.

pmd reloads are currently serialised in each steps calling
reload_affected_pmds.
Any pmd processing packets, waiting on a mutex etc... will make other
pmd threads wait for a delay that can be undeterministic when syscalls
adds up.

Switch to a little busy loop on the control thread using the existing
per-pmd reload boolean.

The memory order on this atomic is rel-acq to have an explicit
synchronisation between the pmd threads and the control thread.

Signed-off-by: David Marchand <david.marchand@redhat.com>
---
Changelog since v1:
- removed the introduced reloading_pmds atomic and reuse the existing
  pmd->reload boolean as a single synchronisation point (Ilya)

Changelog since RFC v1:
- added memory ordering on 'reloading_pmds' atomic to serve as a
  synchronisation point between pmd threads and control thread

---
 lib/dpif-netdev.c | 32 ++++++++++++++++++--------------
 1 file changed, 18 insertions(+), 14 deletions(-)

On 26.06.2019 12:08, David Marchand wrote:
> pmd reloads are currently serialised in each steps calling
> reload_affected_pmds.
> Any pmd processing packets, waiting on a mutex etc... will make other
> pmd threads wait for a delay that can be undeterministic when syscalls
> adds up.
> 
> Switch to a little busy loop on the control thread using the existing
> per-pmd reload boolean.
> 
> The memory order on this atomic is rel-acq to have an explicit
> synchronisation between the pmd threads and the control thread.
> 
> Signed-off-by: David Marchand <david.marchand@redhat.com>
> ---
> Changelog since v1:
> - removed the introduced reloading_pmds atomic and reuse the existing
>   pmd->reload boolean as a single synchronisation point (Ilya)
> 
> Changelog since RFC v1:
> - added memory ordering on 'reloading_pmds' atomic to serve as a
>   synchronisation point between pmd threads and control thread
> 
> ---
>  lib/dpif-netdev.c | 32 ++++++++++++++++++--------------
>  1 file changed, 18 insertions(+), 14 deletions(-)
> 
> diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
> index b7f8117..b71fcbe 100644
> --- a/lib/dpif-netdev.c
> +++ b/lib/dpif-netdev.c
> @@ -647,9 +647,6 @@ struct dp_netdev_pmd_thread {
>      struct ovs_refcount ref_cnt;    /* Every reference must be refcount'ed. */
>      struct cmap_node node;          /* In 'dp->poll_threads'. */
>  
> -    pthread_cond_t cond;            /* For synchronizing pmd thread reload. */
> -    struct ovs_mutex cond_mutex;    /* Mutex for condition variable. */
> -
>      /* Per thread exact-match cache.  Note, the instance for cpu core
>       * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
>       * need to be protected by 'non_pmd_mutex'.  Every other instance
> @@ -1754,11 +1751,8 @@ dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
>          return;
>      }
>  
> -    ovs_mutex_lock(&pmd->cond_mutex);
>      seq_change(pmd->reload_seq);
>      atomic_store_explicit(&pmd->reload, true, memory_order_release);
> -    ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
> -    ovs_mutex_unlock(&pmd->cond_mutex);
>  }
>  
>  static uint32_t
> @@ -4643,6 +4637,16 @@ rxq_scheduling(struct dp_netdev *dp, bool pinned) OVS_REQUIRES(dp->port_mutex)
>  }
>  
>  static void
> +wait_reloading_pmd(struct dp_netdev_pmd_thread *pmd)
> +{
> +    bool reload;
> +
> +    do {
> +        atomic_read_explicit(&pmd->reload, &reload, memory_order_acquire);
> +    } while (reload);
> +}
> +
> +static void
>  reload_affected_pmds(struct dp_netdev *dp)
>  {
>      struct dp_netdev_pmd_thread *pmd;
> @@ -4651,6 +4655,12 @@ reload_affected_pmds(struct dp_netdev *dp)
>          if (pmd->need_reload) {
>              flow_mark_flush(pmd);
>              dp_netdev_reload_pmd__(pmd);
> +        }
> +    }
> +
> +    CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
> +        if (pmd->need_reload) {
> +            wait_reloading_pmd(pmd);
>              pmd->need_reload = false;
>          }
>      }
> @@ -5816,11 +5826,8 @@ dpif_netdev_enable_upcall(struct dpif *dpif)
>  static void
>  dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
>  {
> -    ovs_mutex_lock(&pmd->cond_mutex);
> -    atomic_store_relaxed(&pmd->reload, false);
>      pmd->last_reload_seq = seq_read(pmd->reload_seq);
> -    xpthread_cond_signal(&pmd->cond);
> -    ovs_mutex_unlock(&pmd->cond_mutex);
> +    atomic_store_explicit(&pmd->reload, false, memory_order_release);
>  }
>  
>  /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'.  Returns
> @@ -5905,8 +5912,6 @@ dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
>      pmd->reload_seq = seq_create();
>      pmd->last_reload_seq = seq_read(pmd->reload_seq);
>      atomic_init(&pmd->reload, false);
> -    xpthread_cond_init(&pmd->cond, NULL);
> -    ovs_mutex_init(&pmd->cond_mutex);
>      ovs_mutex_init(&pmd->flow_mutex);
>      ovs_mutex_init(&pmd->port_mutex);
>      cmap_init(&pmd->flow_table);
> @@ -5949,8 +5954,6 @@ dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
>      cmap_destroy(&pmd->flow_table);
>      ovs_mutex_destroy(&pmd->flow_mutex);
>      seq_destroy(pmd->reload_seq);
> -    xpthread_cond_destroy(&pmd->cond);
> -    ovs_mutex_destroy(&pmd->cond_mutex);
>      ovs_mutex_destroy(&pmd->port_mutex);
>      free(pmd);
>  }
> @@ -5971,6 +5974,7 @@ dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd)
>      } else {
>          atomic_store_relaxed(&pmd->exit, true);
>          dp_netdev_reload_pmd__(pmd);
> +        wait_reloading_pmd(pmd);

Join will wait for the thread exit. We don't need to wait for reload here.

>          xpthread_join(pmd->thread, NULL);
>      }
>  
>

On Wed, Jun 26, 2019 at 2:43 PM Ilya Maximets <i.maximets@samsung.com>
wrote:

> On 26.06.2019 12:08, David Marchand wrote:
> > pmd reloads are currently serialised in each steps calling
> > reload_affected_pmds.
> > Any pmd processing packets, waiting on a mutex etc... will make other
> > pmd threads wait for a delay that can be undeterministic when syscalls
> > adds up.
> >
> > Switch to a little busy loop on the control thread using the existing
> > per-pmd reload boolean.
> >
> > The memory order on this atomic is rel-acq to have an explicit
> > synchronisation between the pmd threads and the control thread.
> >
> > Signed-off-by: David Marchand <david.marchand@redhat.com>
> > ---
> > Changelog since v1:
> > - removed the introduced reloading_pmds atomic and reuse the existing
> >   pmd->reload boolean as a single synchronisation point (Ilya)
> >
> > Changelog since RFC v1:
> > - added memory ordering on 'reloading_pmds' atomic to serve as a
> >   synchronisation point between pmd threads and control thread
> >
> > ---
> >  lib/dpif-netdev.c | 32 ++++++++++++++++++--------------
> >  1 file changed, 18 insertions(+), 14 deletions(-)
> >
> > diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
> > index b7f8117..b71fcbe 100644
> > --- a/lib/dpif-netdev.c
> > +++ b/lib/dpif-netdev.c
> > @@ -647,9 +647,6 @@ struct dp_netdev_pmd_thread {
> >      struct ovs_refcount ref_cnt;    /* Every reference must be
> refcount'ed. */
> >      struct cmap_node node;          /* In 'dp->poll_threads'. */
> >
> > -    pthread_cond_t cond;            /* For synchronizing pmd thread
> reload. */
> > -    struct ovs_mutex cond_mutex;    /* Mutex for condition variable. */
> > -
> >      /* Per thread exact-match cache.  Note, the instance for cpu core
> >       * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
> >       * need to be protected by 'non_pmd_mutex'.  Every other instance
> > @@ -1754,11 +1751,8 @@ dp_netdev_reload_pmd__(struct
> dp_netdev_pmd_thread *pmd)
> >          return;
> >      }
> >
> > -    ovs_mutex_lock(&pmd->cond_mutex);
> >      seq_change(pmd->reload_seq);
> >      atomic_store_explicit(&pmd->reload, true, memory_order_release);
> > -    ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
> > -    ovs_mutex_unlock(&pmd->cond_mutex);
> >  }
> >
> >  static uint32_t
> > @@ -4643,6 +4637,16 @@ rxq_scheduling(struct dp_netdev *dp, bool pinned)
> OVS_REQUIRES(dp->port_mutex)
> >  }
> >
> >  static void
> > +wait_reloading_pmd(struct dp_netdev_pmd_thread *pmd)
> > +{
> > +    bool reload;
> > +
> > +    do {
> > +        atomic_read_explicit(&pmd->reload, &reload,
> memory_order_acquire);
> > +    } while (reload);
> > +}
> > +
> > +static void
> >  reload_affected_pmds(struct dp_netdev *dp)
> >  {
> >      struct dp_netdev_pmd_thread *pmd;
> > @@ -4651,6 +4655,12 @@ reload_affected_pmds(struct dp_netdev *dp)
> >          if (pmd->need_reload) {
> >              flow_mark_flush(pmd);
> >              dp_netdev_reload_pmd__(pmd);
> > +        }
> > +    }
> > +
> > +    CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
> > +        if (pmd->need_reload) {
> > +            wait_reloading_pmd(pmd);
> >              pmd->need_reload = false;
> >          }
> >      }
> > @@ -5816,11 +5826,8 @@ dpif_netdev_enable_upcall(struct dpif *dpif)
> >  static void
> >  dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
> >  {
> > -    ovs_mutex_lock(&pmd->cond_mutex);
> > -    atomic_store_relaxed(&pmd->reload, false);
> >      pmd->last_reload_seq = seq_read(pmd->reload_seq);
> > -    xpthread_cond_signal(&pmd->cond);
> > -    ovs_mutex_unlock(&pmd->cond_mutex);
> > +    atomic_store_explicit(&pmd->reload, false, memory_order_release);
> >  }
> >
> >  /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'.  Returns
> > @@ -5905,8 +5912,6 @@ dp_netdev_configure_pmd(struct
> dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
> >      pmd->reload_seq = seq_create();
> >      pmd->last_reload_seq = seq_read(pmd->reload_seq);
> >      atomic_init(&pmd->reload, false);
> > -    xpthread_cond_init(&pmd->cond, NULL);
> > -    ovs_mutex_init(&pmd->cond_mutex);
> >      ovs_mutex_init(&pmd->flow_mutex);
> >      ovs_mutex_init(&pmd->port_mutex);
> >      cmap_init(&pmd->flow_table);
> > @@ -5949,8 +5954,6 @@ dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
> *pmd)
> >      cmap_destroy(&pmd->flow_table);
> >      ovs_mutex_destroy(&pmd->flow_mutex);
> >      seq_destroy(pmd->reload_seq);
> > -    xpthread_cond_destroy(&pmd->cond);
> > -    ovs_mutex_destroy(&pmd->cond_mutex);
> >      ovs_mutex_destroy(&pmd->port_mutex);
> >      free(pmd);
> >  }
> > @@ -5971,6 +5974,7 @@ dp_netdev_del_pmd(struct dp_netdev *dp, struct
> dp_netdev_pmd_thread *pmd)
> >      } else {
> >          atomic_store_relaxed(&pmd->exit, true);
> >          dp_netdev_reload_pmd__(pmd);
> > +        wait_reloading_pmd(pmd);
>
> Join will wait for the thread exit. We don't need to wait for reload here.
>

Indeed, and then I can move the wait_reloading_pmd() code directly into
reload_affected_pmds().