Re: [PATCH net-next v2 02/13] virtio_ring: split: record extras for indirect buffers
On Wed, Nov 6, 2024 at 3:42 PM Michael S. Tsirkin wrote:
>
> On Wed, Nov 06, 2024 at 09:44:39AM +0800, Jason Wang wrote:
> > > > > while (vq->split.vring.desc[i].flags & nextflag) {
> > > > > - vring_unmap_one_split(vq, i);
> > > > > + vring_unmap_one_split(vq, &extra[i]);
> > > >
> > > > Not sure if I've asked this before. But this part seems to deserve an
> > > > independent fix for -stable.
> > >
> > > What fix?
> >
> > I meant for hardening we need to check the flags stored in the extra
> > instead of the descriptor itself as it could be mangled by the device.
> >
> > Thanks
>
> Good point. Jason, want to cook up a patch?
Will do.
Thanks
>
> --
> MST
>
Re: [PATCH net-next v2 02/13] virtio_ring: split: record extras for indirect buffers
On Wed, Nov 06, 2024 at 09:44:39AM +0800, Jason Wang wrote:
> > > > while (vq->split.vring.desc[i].flags & nextflag) {
> > > > - vring_unmap_one_split(vq, i);
> > > > + vring_unmap_one_split(vq, &extra[i]);
> > >
> > > Not sure if I've asked this before. But this part seems to deserve an
> > > independent fix for -stable.
> >
> > What fix?
>
> I meant for hardening we need to check the flags stored in the extra
> instead of the descriptor itself as it could be mangled by the device.
>
> Thanks
Good point. Jason, want to cook up a patch?
--
MST
Re: [PATCH net-next v2 02/13] virtio_ring: split: record extras for indirect buffers
On Wed, 6 Nov 2024 09:44:39 +0800, Jason Wang wrote:
> On Tue, Nov 5, 2024 at 2:53 PM Xuan Zhuo wrote:
> >
> > On Tue, 5 Nov 2024 11:42:09 +0800, Jason Wang wrote:
> > > On Wed, Oct 30, 2024 at 4:25 PM Xuan Zhuo
> > > wrote:
> > > >
> > > > The subsequent commit needs to know whether every indirect buffer is
> > > > premapped or not. So we need to introduce an extra struct for every
> > > > indirect buffer to record this info.
> > > >
> > > > Signed-off-by: Xuan Zhuo
> > > > ---
> > > > drivers/virtio/virtio_ring.c | 112 ---
> > > > 1 file changed, 52 insertions(+), 60 deletions(-)
> > >
> > > Do we have a performance impact for this patch?
> > >
> > > >
> > > > diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> > > > index 97590c201aa2..dca093744fe1 100644
> > > > --- a/drivers/virtio/virtio_ring.c
> > > > +++ b/drivers/virtio/virtio_ring.c
> > > > @@ -69,7 +69,11 @@
> > > >
> > > > struct vring_desc_state_split {
> > > > void *data; /* Data for callback. */
> > > > - struct vring_desc *indir_desc; /* Indirect descriptor, if any.
> > > > */
> > > > +
> > > > + /* Indirect extra table and desc table, if any. These two will
> > > > be
> > > > +* allocated together. So we won't stress more to the memory
> > > > allocator.
> > > > +*/
> > > > + struct vring_desc *indir_desc;
> > >
> > > So it looks like we put a descriptor table after the extra table. Can
> > > this lead to more crossing page mappings for the indirect descriptors?
> > >
> > > If yes, it seems expensive so we probably need to make the descriptor
> > > table come first.
> >
> > No, the descriptors are before extra table.
>
> Well, you need then tweak the above comment, it said
>
> "Indirect extra table and desc table".
>
> > So, there is not performance impact.
> >
> >
> > >
> > > > };
> > > >
>
> [...]
>
> > > > while (vq->split.vring.desc[i].flags & nextflag) {
> > > > - vring_unmap_one_split(vq, i);
> > > > + vring_unmap_one_split(vq, &extra[i]);
> > >
> > > Not sure if I've asked this before. But this part seems to deserve an
> > > independent fix for -stable.
> >
> > What fix?
>
> I meant for hardening we need to check the flags stored in the extra
> instead of the descriptor itself as it could be mangled by the device.
I see, we can do it in future.
Thanks.
>
> Thanks
>
Re: [PATCH net-next v2 02/13] virtio_ring: split: record extras for indirect buffers
On Tue, Nov 5, 2024 at 2:53 PM Xuan Zhuo wrote:
>
> On Tue, 5 Nov 2024 11:42:09 +0800, Jason Wang wrote:
> > On Wed, Oct 30, 2024 at 4:25 PM Xuan Zhuo
> > wrote:
> > >
> > > The subsequent commit needs to know whether every indirect buffer is
> > > premapped or not. So we need to introduce an extra struct for every
> > > indirect buffer to record this info.
> > >
> > > Signed-off-by: Xuan Zhuo
> > > ---
> > > drivers/virtio/virtio_ring.c | 112 ---
> > > 1 file changed, 52 insertions(+), 60 deletions(-)
> >
> > Do we have a performance impact for this patch?
> >
> > >
> > > diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> > > index 97590c201aa2..dca093744fe1 100644
> > > --- a/drivers/virtio/virtio_ring.c
> > > +++ b/drivers/virtio/virtio_ring.c
> > > @@ -69,7 +69,11 @@
> > >
> > > struct vring_desc_state_split {
> > > void *data; /* Data for callback. */
> > > - struct vring_desc *indir_desc; /* Indirect descriptor, if any. */
> > > +
> > > + /* Indirect extra table and desc table, if any. These two will be
> > > +* allocated together. So we won't stress more to the memory
> > > allocator.
> > > +*/
> > > + struct vring_desc *indir_desc;
> >
> > So it looks like we put a descriptor table after the extra table. Can
> > this lead to more crossing page mappings for the indirect descriptors?
> >
> > If yes, it seems expensive so we probably need to make the descriptor
> > table come first.
>
> No, the descriptors are before extra table.
Well, you need then tweak the above comment, it said
"Indirect extra table and desc table".
> So, there is not performance impact.
>
>
> >
> > > };
> > >
[...]
> > > while (vq->split.vring.desc[i].flags & nextflag) {
> > > - vring_unmap_one_split(vq, i);
> > > + vring_unmap_one_split(vq, &extra[i]);
> >
> > Not sure if I've asked this before. But this part seems to deserve an
> > independent fix for -stable.
>
> What fix?
I meant for hardening we need to check the flags stored in the extra
instead of the descriptor itself as it could be mangled by the device.
Thanks
Re: [PATCH net-next v2 02/13] virtio_ring: split: record extras for indirect buffers
On Tue, 5 Nov 2024 11:42:09 +0800, Jason Wang wrote:
> On Wed, Oct 30, 2024 at 4:25 PM Xuan Zhuo wrote:
> >
> > The subsequent commit needs to know whether every indirect buffer is
> > premapped or not. So we need to introduce an extra struct for every
> > indirect buffer to record this info.
> >
> > Signed-off-by: Xuan Zhuo
> > ---
> > drivers/virtio/virtio_ring.c | 112 ---
> > 1 file changed, 52 insertions(+), 60 deletions(-)
>
> Do we have a performance impact for this patch?
>
> >
> > diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> > index 97590c201aa2..dca093744fe1 100644
> > --- a/drivers/virtio/virtio_ring.c
> > +++ b/drivers/virtio/virtio_ring.c
> > @@ -69,7 +69,11 @@
> >
> > struct vring_desc_state_split {
> > void *data; /* Data for callback. */
> > - struct vring_desc *indir_desc; /* Indirect descriptor, if any. */
> > +
> > + /* Indirect extra table and desc table, if any. These two will be
> > +* allocated together. So we won't stress more to the memory
> > allocator.
> > +*/
> > + struct vring_desc *indir_desc;
>
> So it looks like we put a descriptor table after the extra table. Can
> this lead to more crossing page mappings for the indirect descriptors?
>
> If yes, it seems expensive so we probably need to make the descriptor
> table come first.
No, the descriptors are before extra table.
So, there is not performance impact.
>
> > };
> >
> > struct vring_desc_state_packed {
> > @@ -440,38 +444,20 @@ static void virtqueue_init(struct vring_virtqueue
> > *vq, u32 num)
> > * Split ring specific functions - *_split().
> > */
> >
> > -static void vring_unmap_one_split_indirect(const struct vring_virtqueue
> > *vq,
> > - const struct vring_desc *desc)
> > -{
> > - u16 flags;
> > -
> > - if (!vring_need_unmap_buffer(vq))
> > - return;
> > -
> > - flags = virtio16_to_cpu(vq->vq.vdev, desc->flags);
> > -
> > - dma_unmap_page(vring_dma_dev(vq),
> > - virtio64_to_cpu(vq->vq.vdev, desc->addr),
> > - virtio32_to_cpu(vq->vq.vdev, desc->len),
> > - (flags & VRING_DESC_F_WRITE) ?
> > - DMA_FROM_DEVICE : DMA_TO_DEVICE);
> > -}
> > -
> > static unsigned int vring_unmap_one_split(const struct vring_virtqueue *vq,
> > - unsigned int i)
> > + struct vring_desc_extra *extra)
> > {
> > - struct vring_desc_extra *extra = vq->split.desc_extra;
> > u16 flags;
> >
> > - flags = extra[i].flags;
> > + flags = extra->flags;
> >
> > if (flags & VRING_DESC_F_INDIRECT) {
> > if (!vq->use_dma_api)
> > goto out;
> >
> > dma_unmap_single(vring_dma_dev(vq),
> > -extra[i].addr,
> > -extra[i].len,
> > +extra->addr,
> > +extra->len,
> > (flags & VRING_DESC_F_WRITE) ?
> > DMA_FROM_DEVICE : DMA_TO_DEVICE);
> > } else {
> > @@ -479,22 +465,23 @@ static unsigned int vring_unmap_one_split(const
> > struct vring_virtqueue *vq,
> > goto out;
> >
> > dma_unmap_page(vring_dma_dev(vq),
> > - extra[i].addr,
> > - extra[i].len,
> > + extra->addr,
> > + extra->len,
> >(flags & VRING_DESC_F_WRITE) ?
> >DMA_FROM_DEVICE : DMA_TO_DEVICE);
> > }
> >
> > out:
> > - return extra[i].next;
> > + return extra->next;
> > }
> >
> > static struct vring_desc *alloc_indirect_split(struct virtqueue *_vq,
> >unsigned int total_sg,
> >gfp_t gfp)
> > {
> > + struct vring_desc_extra *extra;
> > struct vring_desc *desc;
> > - unsigned int i;
> > + unsigned int i, size;
> >
> > /*
> > * We require lowmem mappings for the descriptors because
> > @@ -503,40 +490,41 @@ static struct vring_desc *alloc_indirect_split(struct
> > virtqueue *_vq,
> > */
> > gfp &= ~__GFP_HIGHMEM;
> >
> > - desc = kmalloc_array(total_sg, sizeof(struct vring_desc), gfp);
> > + size = sizeof(*desc) * total_sg + sizeof(*extra) * total_sg;
> > +
> > + desc = kmalloc(size, gfp);
> > if (!desc)
> > return NULL;
> >
> > + extra = (struct vring_desc_extra *)&desc[total_sg];
> > +
> > for (i = 0; i < total_sg; i++)
> > - desc[i].next = cpu_to_virtio16(_vq->vdev, i + 1);
>
Re: [PATCH net-next v2 02/13] virtio_ring: split: record extras for indirect buffers
On Wed, Oct 30, 2024 at 4:25 PM Xuan Zhuo wrote:
>
> The subsequent commit needs to know whether every indirect buffer is
> premapped or not. So we need to introduce an extra struct for every
> indirect buffer to record this info.
>
> Signed-off-by: Xuan Zhuo
> ---
> drivers/virtio/virtio_ring.c | 112 ---
> 1 file changed, 52 insertions(+), 60 deletions(-)
Do we have a performance impact for this patch?
>
> diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> index 97590c201aa2..dca093744fe1 100644
> --- a/drivers/virtio/virtio_ring.c
> +++ b/drivers/virtio/virtio_ring.c
> @@ -69,7 +69,11 @@
>
> struct vring_desc_state_split {
> void *data; /* Data for callback. */
> - struct vring_desc *indir_desc; /* Indirect descriptor, if any. */
> +
> + /* Indirect extra table and desc table, if any. These two will be
> +* allocated together. So we won't stress more to the memory
> allocator.
> +*/
> + struct vring_desc *indir_desc;
So it looks like we put a descriptor table after the extra table. Can
this lead to more crossing page mappings for the indirect descriptors?
If yes, it seems expensive so we probably need to make the descriptor
table come first.
> };
>
> struct vring_desc_state_packed {
> @@ -440,38 +444,20 @@ static void virtqueue_init(struct vring_virtqueue *vq,
> u32 num)
> * Split ring specific functions - *_split().
> */
>
> -static void vring_unmap_one_split_indirect(const struct vring_virtqueue *vq,
> - const struct vring_desc *desc)
> -{
> - u16 flags;
> -
> - if (!vring_need_unmap_buffer(vq))
> - return;
> -
> - flags = virtio16_to_cpu(vq->vq.vdev, desc->flags);
> -
> - dma_unmap_page(vring_dma_dev(vq),
> - virtio64_to_cpu(vq->vq.vdev, desc->addr),
> - virtio32_to_cpu(vq->vq.vdev, desc->len),
> - (flags & VRING_DESC_F_WRITE) ?
> - DMA_FROM_DEVICE : DMA_TO_DEVICE);
> -}
> -
> static unsigned int vring_unmap_one_split(const struct vring_virtqueue *vq,
> - unsigned int i)
> + struct vring_desc_extra *extra)
> {
> - struct vring_desc_extra *extra = vq->split.desc_extra;
> u16 flags;
>
> - flags = extra[i].flags;
> + flags = extra->flags;
>
> if (flags & VRING_DESC_F_INDIRECT) {
> if (!vq->use_dma_api)
> goto out;
>
> dma_unmap_single(vring_dma_dev(vq),
> -extra[i].addr,
> -extra[i].len,
> +extra->addr,
> +extra->len,
> (flags & VRING_DESC_F_WRITE) ?
> DMA_FROM_DEVICE : DMA_TO_DEVICE);
> } else {
> @@ -479,22 +465,23 @@ static unsigned int vring_unmap_one_split(const struct
> vring_virtqueue *vq,
> goto out;
>
> dma_unmap_page(vring_dma_dev(vq),
> - extra[i].addr,
> - extra[i].len,
> + extra->addr,
> + extra->len,
>(flags & VRING_DESC_F_WRITE) ?
>DMA_FROM_DEVICE : DMA_TO_DEVICE);
> }
>
> out:
> - return extra[i].next;
> + return extra->next;
> }
>
> static struct vring_desc *alloc_indirect_split(struct virtqueue *_vq,
>unsigned int total_sg,
>gfp_t gfp)
> {
> + struct vring_desc_extra *extra;
> struct vring_desc *desc;
> - unsigned int i;
> + unsigned int i, size;
>
> /*
> * We require lowmem mappings for the descriptors because
> @@ -503,40 +490,41 @@ static struct vring_desc *alloc_indirect_split(struct
> virtqueue *_vq,
> */
> gfp &= ~__GFP_HIGHMEM;
>
> - desc = kmalloc_array(total_sg, sizeof(struct vring_desc), gfp);
> + size = sizeof(*desc) * total_sg + sizeof(*extra) * total_sg;
> +
> + desc = kmalloc(size, gfp);
> if (!desc)
> return NULL;
>
> + extra = (struct vring_desc_extra *)&desc[total_sg];
> +
> for (i = 0; i < total_sg; i++)
> - desc[i].next = cpu_to_virtio16(_vq->vdev, i + 1);
> + extra[i].next = i + 1;
> +
> return desc;
> }
>
> static inline unsigned int virtqueue_add_desc_split(struct virtqueue *vq,
> struct vring_desc *desc,
> + struct vring_desc_extra
> *extra,
> unsigned int i,
>
