Re: [PATCH net] virtio-net: fix page refcnt leaking when fail to allocate frag skb

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Nov 19, 2013 at 01:36:36PM -0800, Eric Dumazet wrote:
> On Tue, 2013-11-19 at 22:49 +0200, Michael S. Tsirkin wrote:
> > On Tue, Nov 19, 2013 at 06:03:48AM -0800, Eric Dumazet wrote:
> > > On Tue, 2013-11-19 at 16:05 +0800, Jason Wang wrote:
> > > > We need to drop the refcnt of page when we fail to allocate an skb for frag
> > > > list, otherwise it will be leaked. The bug was introduced by commit
> > > > 2613af0ed18a11d5c566a81f9a6510b73180660a ("virtio_net: migrate mergeable rx
> > > > buffers to page frag allocators").
> > > > 
> > > > Cc: Michael Dalton <mwdalton@google.com>
> > > > Cc: Eric Dumazet <edumazet@google.com>
> > > > Cc: Rusty Russell <rusty@rustcorp.com.au>
> > > > Cc: Michael S. Tsirkin <mst@redhat.com>
> > > > Signed-off-by: Jason Wang <jasowang@redhat.com>
> > > > ---
> > > > The patch was needed for 3.12 stable.
> > > 
> > > Good catch, but if we return from receive_mergeable() in the 'middle'
> > > of the frags we would need for the current skb, who will
> > > call the virtqueue_get_buf() to flush the remaining frags ?
> > > 
> > > Don't we also need to call virtqueue_get_buf() like 
> > > 
> > > while (--num_buf) {
> > >     buf = virtqueue_get_buf(rq->vq, &len);
> > >     if (!buf)
> > >         break;
> > >     put_page(virt_to_head_page(buf));
> > > }
> > > 
> > > ?
> > > 
> > > 
> > 
> > 
> > Let me explain what worries me in your suggestion:
> > 
> >                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
> >                         if (unlikely(!nskb)) {
> >                                 head_skb->dev->stats.rx_dropped++;
> >                                 return -ENOMEM;
> >                         }
> > 
> > is this the failure case we are talking about?
> 
> I thought Jason patch was about this, no ?
> 
> > 
> > I think this is a symprom of a larger problem
> > introduced by 2613af0ed18a11d5c566a81f9a6510b73180660a,
> > namely that we now need to allocate memory in the
> > middle of processing a packet.
> > 
> > 
> > I think discarding a completely valid and well-formed
> > packet from the receive queue because we are unable
> > to allocate new memory with GFP_ATOMIC
> > for future packets is not a good idea.
> 
> How is it different with NIC processing in RX path ?


Which NIC? Virtio? Prior to 2613af0ed18a11d5c566a81f9a6510b73180660a
it didn't drop packets received from host as far as I can tell.
virtio is more like a pipe than a real NIC in this respect.

> > 
> > It certainly violates the principle of least surprize:
> > when one sees host pass packet to guest, one expects
> > the packet to get into the networking stack, not get
> > dropped by the driver internally.
> > Guest stack can do with the packet what it sees fit.
> > 
> > We actually wake up a thread if we can't fill up the queue,
> > that will fill it up in GFP_KERNEL context.
> > 
> > So I think we should find a way to pre-allocate if necessary and avoid
> > error paths where allocating new memory is a required to avoid drops.
> > 
> 
> Really, under ATOMIC context, there is no way you can avoid dropping
> packets if you cannot allocate memory. If you cannot allocate sk_buff
> (256 bytes !!), you wont be able to allocate the 1500+ bytes to hold the
> payload of next packets anyway. 

that's why we do:

                if (!try_fill_recv(rq, GFP_ATOMIC))
                        schedule_delayed_work(&vi->refill, 0);


the queues are large enough for a single failure not to be
an immediate problem.


> Same problem on a real NIC.
> 
> Under memory pressure we _do_ packet drops.
> Nobody really complained.
>
> Sure, you can add yet another cache of pre-allocated skbs and pay the
> price of managing yet another cache layer, but still need to trop
> packets under stress.

We don't need a cache even. Just enough to avoid dropping packets
if allocation failed in the middle so we don't dequeue a buffer and then
drop it.

Once we use this reserved skb, we stop processing the queue until
refill gives it back.

> Pre-allocating skb on real NIC has a performance cost, because we clear
> sk_buff way ahead of time. By the time skb is finally received, cpu has
> to bring back into its cache memory cache lines.
> 

Alternatively we can pre-allocate the memory but avoid clearing it maybe?

-- 
MST