On Tue, Aug 20, 2024 at 12:44:46PM -0700, Si-Wei Liu wrote:
>
>
> On 8/20/2024 12:19 AM, Xuan Zhuo wrote:
> > leads to regression on VM with the sysctl value of:
> >
> > - net.core.high_order_alloc_disable=1
> >
> > which could see reliable crashes or scp failure (scp a file 100M in size
> > to VM):
> >
> > The issue is that the virtnet_rq_dma takes up 16 bytes at the beginning
> > of a new frag. When the frag size is larger than PAGE_SIZE,
> > everything is fine. However, if the frag is only one page and the
> > total size of the buffer and virtnet_rq_dma is larger than one page, an
> > overflow may occur. In this case, if an overflow is possible, I adjust
> > the buffer size. If net.core.high_order_alloc_disable=1, the maximum
> > buffer size is 4096 - 16. If net.core.high_order_alloc_disable=0, only
> > the first buffer of the frag is affected.
> >
> > Fixes: f9dac92ba908 ("virtio_ring: enable premapped mode whatever
> > use_dma_api")
> > Reported-by: "Si-Wei Liu" <si-wei....@oracle.com>
> > Closes:
> > http://lore.kernel.org/all/8b20cc28-45a9-4643-8e87-ba164a540...@oracle.com
> > Signed-off-by: Xuan Zhuo <xuanz...@linux.alibaba.com>
> > ---
> > drivers/net/virtio_net.c | 12 +++++++++---
> > 1 file changed, 9 insertions(+), 3 deletions(-)
> >
> > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> > index c6af18948092..e5286a6da863 100644
> > --- a/drivers/net/virtio_net.c
> > +++ b/drivers/net/virtio_net.c
> > @@ -918,9 +918,6 @@ static void *virtnet_rq_alloc(struct receive_queue *rq,
> > u32 size, gfp_t gfp)
> > void *buf, *head;
> > dma_addr_t addr;
> > - if (unlikely(!skb_page_frag_refill(size, alloc_frag, gfp)))
> > - return NULL;
> > -
> > head = page_address(alloc_frag->page);
> > dma = head;
> > @@ -2421,6 +2418,9 @@ static int add_recvbuf_small(struct virtnet_info *vi,
> > struct receive_queue *rq,
> > len = SKB_DATA_ALIGN(len) +
> > SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
> > + if (unlikely(!skb_page_frag_refill(len, &rq->alloc_frag, gfp)))
> > + return -ENOMEM;
> > +
> Do you want to document the assumption that small packet case won't end up
> crossing the page frag boundary unlike the mergeable case? Add a comment
> block to explain or a WARN_ON() check against potential overflow would work
> with me.
>
> > buf = virtnet_rq_alloc(rq, len, gfp);
> > if (unlikely(!buf))
> > return -ENOMEM;
> > @@ -2521,6 +2521,12 @@ static int add_recvbuf_mergeable(struct virtnet_info
> > *vi,
> > */
> > len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
> > + if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
> > + return -ENOMEM;
> > +
> > + if (!alloc_frag->offset && len + room + sizeof(struct virtnet_rq_dma) >
> > alloc_frag->size)
> > + len -= sizeof(struct virtnet_rq_dma);
> > +
> This could address my previous concern for possibly regressing every buffer
> size for the mergeable case, thanks. Though I still don't get why carving up
> a small chunk from page_frag for storing the virtnet_rq_dma metadata, this
> would cause perf regression on certain MTU size
4Kbyte MTU exactly?
> that happens to end up with
> one more base page (and an extra descriptor as well) to be allocated
> compared to the previous code without the extra virtnet_rq_dma content. How
> hard would it be to allocate a dedicated struct to store the related
> information without affecting the (size of) datapath pages?
>
> FWIW, out of the code review perspective, I've looked up the past
> conversations but didn't see comprehensive benchmark was done before
> removing the old code and making premap the sole default mode. Granted this
> would reduce the footprint of additional code and the associated maintaining
> cost immediately, but I would assume at least there should have been
> thorough performance runs upfront to guarantee no regression is seen with
> every possible use case, or the negative effect is comparatively negligible
> even though there's slight regression in some limited case. If that kind of
> perf measurement hadn't been done before getting accepted/merged, I think at
> least it should allow both modes to coexist for a while such that every user
> could gauge the performance effect.
>
> Thanks,
> -Siwei
>
> > buf = virtnet_rq_alloc(rq, len + room, gfp);
> > if (unlikely(!buf))
> > return -ENOMEM;
