Date:   Thu, 8 Aug 2019 09:01:09 -0400 (EDT)
From:   Jason Wang <jasowang@...hat.com>
To:     Jason Gunthorpe <jgg@...pe.ca>
Cc:     kvm@...r.kernel.org, mst@...hat.com, netdev@...r.kernel.org,
        linux-kernel@...r.kernel.org,
        virtualization@...ts.linux-foundation.org, linux-mm@...ck.org
Subject: Re: [PATCH V4 7/9] vhost: do not use RCU to synchronize MMU
 notifier       with worker



----- Original Message -----
> 
> On 2019/8/7 下午10:02, Jason Wang wrote:
> >
> > On 2019/8/7 下午8:07, Jason Gunthorpe wrote:
> >> On Wed, Aug 07, 2019 at 03:06:15AM -0400, Jason Wang wrote:
> >>> We used to use RCU to synchronize MMU notifier with worker. This leads
> >>> calling synchronize_rcu() in invalidate_range_start(). But on a busy
> >>> system, there would be many factors that may slow down the
> >>> synchronize_rcu() which makes it unsuitable to be called in MMU
> >>> notifier.
> >>>
> >>> So this patch switches use seqlock counter to track whether or not the
> >>> map was used. The counter was increased when vq try to start or finish
> >>> uses the map. This means, when it was even, we're sure there's no
> >>> readers and MMU notifier is synchronized. When it was odd, it means
> >>> there's a reader we need to wait it to be even again then we are
> >>> synchronized. Consider the read critical section is pretty small the
> >>> synchronization should be done very fast.
> >>>
> >>> Reported-by: Michael S. Tsirkin <mst@...hat.com>
> >>> Fixes: 7f466032dc9e ("vhost: access vq metadata through kernel
> >>> virtual address")
> >>> Signed-off-by: Jason Wang <jasowang@...hat.com>
> >>>   drivers/vhost/vhost.c | 141
> >>> ++++++++++++++++++++++++++----------------
> >>>   drivers/vhost/vhost.h |   7 ++-
> >>>   2 files changed, 90 insertions(+), 58 deletions(-)
> >>>
> >>> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> >>> index cfc11f9ed9c9..57bfbb60d960 100644
> >>> +++ b/drivers/vhost/vhost.c
> >>> @@ -324,17 +324,16 @@ static void vhost_uninit_vq_maps(struct
> >>> vhost_virtqueue *vq)
> >>>         spin_lock(&vq->mmu_lock);
> >>>       for (i = 0; i < VHOST_NUM_ADDRS; i++) {
> >>> -        map[i] = rcu_dereference_protected(vq->maps[i],
> >>> -                  lockdep_is_held(&vq->mmu_lock));
> >>> +        map[i] = vq->maps[i];
> >>>           if (map[i]) {
> >>>               vhost_set_map_dirty(vq, map[i], i);
> >>> -            rcu_assign_pointer(vq->maps[i], NULL);
> >>> +            vq->maps[i] = NULL;
> >>>           }
> >>>       }
> >>>       spin_unlock(&vq->mmu_lock);
> >>>   -    /* No need for synchronize_rcu() or kfree_rcu() since we are
> >>> -     * serialized with memory accessors (e.g vq mutex held).
> >>> +    /* No need for synchronization since we are serialized with
> >>> +     * memory accessors (e.g vq mutex held).
> >>>        */
> >>>         for (i = 0; i < VHOST_NUM_ADDRS; i++)
> >>> @@ -362,6 +361,40 @@ static bool vhost_map_range_overlap(struct
> >>> vhost_uaddr *uaddr,
> >>>       return !(end < uaddr->uaddr || start > uaddr->uaddr - 1 +
> >>> uaddr->size);
> >>>   }
> >>>   +static void inline vhost_vq_access_map_begin(struct
> >>> vhost_virtqueue *vq)
> >>> +{
> >>> +    write_seqcount_begin(&vq->seq);
> >>> +}
> >>> +
> >>> +static void inline vhost_vq_access_map_end(struct vhost_virtqueue *vq)
> >>> +{
> >>> +    write_seqcount_end(&vq->seq);
> >>> +}
> >> The write side of a seqlock only provides write barriers. Access to
> >>
> >>     map = vq->maps[VHOST_ADDR_USED];
> >>
> >> Still needs a read side barrier, and then I think this will be no
> >> better than a normal spinlock.
> >>
> >> It also doesn't seem like this algorithm even needs a seqlock, as this
> >> is just a one bit flag
> >
> >
> > Right, so then I tend to use spinlock first for correctness.
> >
> >
> >>
> >> atomic_set_bit(using map)
> >> smp_mb__after_atomic()
> >> .. maps [...]
> >> atomic_clear_bit(using map)
> >>
> >>
> >> map = NULL;
> >> smp_mb__before_atomic();
> >> while (atomic_read_bit(using map))
> >>     relax()
> >>
> >> Again, not clear this could be faster than a spinlock when the
> >> barriers are correct...
> >
> 
> I've done some benchmark[1] on x86, and yes it looks even slower. It
> looks to me the atomic stuffs is not necessary, so in order to compare
> it better with spinlock. I tweak it a little bit through
> smp_load_acquire()/store_releaes() + mb() like:
> 

Sorry the format is messed up:

The code should be something like:

static struct vhost_map *vhost_vq_access_map_begin(struct vhost_virtqueue *vq,
                                                   unsigned int type)
{
        ++vq->counter;
        /* Ensure map was read after incresing the counter. Paired
         * with smp_mb() in vhost_vq_sync_access().
         */
        smp_mb();
        return vq->maps[type];
}

static void inline vhost_vq_access_map_end(struct vhost_virtqueue *vq)
{
 	/* Ensure all memory access through map was done before
         * reducing the counter. Paired with smp_load_acquire() in
         * vhost_vq_sync_access() */
        smp_store_release(&vq->counter, --vq->counter);
}

static void inline vhost_vq_sync_access(struct vhost_virtqueue *vq)
{
        /* Ensure new map value is visible before checking counter. */
        smp_mb();
        /* Ensure map was freed after reading counter value, paired
         * with smp_store_release() in vhost_vq_access_map_end().
         */
        while (smp_load_acquire(&vq->counter)) {
                if (need_resched())
                        schedule();
        }
}

And the benchmark result is:

         | base    | direct + atomic bitops | direct + spinlock() | direct + counter + smp_mb() | direct + RCU     |
SMAP on  | 5.0Mpps | 5.0Mpps     (+0%)      | 5.7Mpps  	(+14%)	  | 5.9Mpps  (+18%)	        | 6.2Mpps  (+24%)  |
SMAP off | 7.0Mpps | 7.0Mpps     (+0%)      | 7.0Mpps   (+0%)     | 7.5Mpps  (+7%)	        | 8.2Mpps  (+17%)  |


> 
> 
> base: normal copy_to_user()/copy_from_user() path.
> direct + atomic bitops: using direct mapping but synchronize through
> atomic bitops like you suggested above
> direct + spinlock(): using direct mapping but synchronize through spinlocks
> direct + counter + smp_mb(): using direct mapping but synchronize
> through counter + smp_mb()
> direct + RCU: using direct mapping and synchronize through RCU (buggy
> and need to be addressed by this series)
> 
> 
> So smp_mb() + counter is fastest way. And spinlock can still show some
> improvement (+14%) in the case of SMAP, but no the case when SMAP is off.
> 
> I don't have any objection to convert  to spinlock() but just want to
> know if any case that the above smp_mb() + counter looks good to you?
> 
> Thanks
> 
> 
> >
> > Yes, for next release we may want to use the idea from Michael like to
> > mitigate the impact of mb.
> >
> > https://lwn.net/Articles/775871/
> >
> > Thanks
> >
> >
> >>
> >> Jason
> 
> _______________________________________________
> Virtualization mailing list
> Virtualization@...ts.linux-foundation.org
> https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Powered by blists - more mailing lists