Re: virtio_net: support device stats

[Jason Wang]

On Mon, Oct 14, 2024 at 6:15 PM Colin King (gmail)
<colin.i.king@xxxxxxxxx> wrote:
>
> On 14/10/2024 10:47, Michael S. Tsirkin wrote:
> > On Mon, Oct 14, 2024 at 10:39:26AM +0100, Colin King (gmail) wrote:
> >> Hi,
> >>
> >> Static analysis on Linux-next has detected a potential issue with the
> >> following commit:
> >>
> >> commit 941168f8b40e50518a3bc6ce770a7062a5d99230
> >> Author: Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx>
> >> Date:   Fri Apr 26 11:39:24 2024 +0800
> >>
> >>      virtio_net: support device stats
> >>
> >>
> >> The issue is in function virtnet_stats_ctx_init, in drivers/net/virtio_net.c
> >> as follows:
> >>
> >>          if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_CVQ) {
> >>                  queue_type = VIRTNET_Q_TYPE_CQ;
> >>
> >>                  ctx->bitmap[queue_type]   |= VIRTIO_NET_STATS_TYPE_CVQ;
> >>                  ctx->desc_num[queue_type] +=
> >> ARRAY_SIZE(virtnet_stats_cvq_desc);
> >>                  ctx->size[queue_type]     += sizeof(struct
> >> virtio_net_stats_cvq);
> >>          }
> >>
> >>
> >> ctx->bitmap is declared as a u32 however it is being bit-wise or'd with
> >> VIRTIO_NET_STATS_TYPE_CVQ and this is defined as 1 << 32:
> >>
> >> include/uapi/linux/virtio_net.h:#define VIRTIO_NET_STATS_TYPE_CVQ (1ULL <<
> >> 32)
> >>
> >> ..and hence the bit-wise or operation won't set any bits in ctx->bitmap
> >> because 1ULL < 32 is too wide for a u32.
> >
> > Indeed. Xuan Zhuo how did you test this patch?
> >
> >> I suspect ctx->bitmap should be
> >> declared as u64.
> >>
> >> Colin
> >>
> >>
> >
> > In fact, it is read into a u64:
> >
> >         u64 offset, bitmap;
> > ....
> >          bitmap = ctx->bitmap[queue_type];
> >
> > we'll have to reorder fields to avoid wasting memory.
> > Like this I guess:
> >
> > Signed-off-by: Michael S. Tsirkin <mst@xxxxxxxxxx>
> >
> > Colin, can you confirm pls?
>
> Fix looks sane to be, with u64 bitmap[3] struct size field re-ordering
> does not seem to make any difference on x86-64 (64 bytes) and i586 (56
> bytes) when I compiled with gcc-12, gcc-14 and clang-20.
>
> I can't functionally test this though (not sure how).
>
> Reviewed-by: Colin Ian King <colin.king@xxxxxxxxx>

Maybe it's time to ask for a prototype for each virtio-net feature
before it can be merged into the spec.

Thanks