On Tue, 2021-01-12 at 21:12 -0800, Eric Dumazet wrote: > From: Eric Dumazet <eduma...@google.com> > > This reverts commit c67f5db82027ba6d2ea4ac9176bc45996a03ae6a. > > While using page fragments instead of a kmalloc backed skb->head might give > a small performance improvement in some cases, there is a huge risk of > memory use under estimation. > > GOOD_COPY_LEN is 128 bytes. This means that we need a small amount > of memory to hold the headers and struct skb_shared_info > > Yet, napi_alloc_skb() might use a whole 32KB page (or 64KB on PowerPc) > for long lived incoming TCP packets. > > We have been tracking OOM issues on GKE hosts hitting tcp_mem limits > but consuming far more memory for TCP buffers than instructed in tcp_mem[2] > > Even if we force napi_alloc_skb() to only use order-0 pages, the issue > would still be there on arches with PAGE_SIZE >= 32768 > > Using alloc_skb() and thus standard kmallloc() for skb->head allocations > will get the benefit of letting other objects in each page being independently > used by other skbs, regardless of the lifetime. > > Note that a similar problem exists for skbs allocated from napi_get_frags(), > this is handled in a separate patch. > > I would like to thank Greg Thelen for his precious help on this matter, > analysing crash dumps is always a time consuming task. > > Signed-off-by: Eric Dumazet <eduma...@google.com> > Cc: Paolo Abeni <pab...@redhat.com> > Cc: Michael S. Tsirkin <m...@redhat.com> > Cc: Greg Thelen <gthe...@google.com> > --- > drivers/net/virtio_net.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c > index > 508408fbe78fbd8658dc226834b5b1b334b8b011..5886504c1acacf3f6148127b5c1cc7f6a906b827 > 100644 > --- a/drivers/net/virtio_net.c > +++ b/drivers/net/virtio_net.c > @@ -386,7 +386,7 @@ static struct sk_buff *page_to_skb(struct virtnet_info > *vi, > p = page_address(page) + offset; > > /* copy small packet so we can reuse these pages for small data */ > - skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN); > + skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN); > if (unlikely(!skb)) > return NULL;
I'm ok with the revert. The gain given by the original change in my tests was measurable, but small. Acked-by: Paolo Abeni <pab...@redhat.com>
