Message ID | 20230502174404.668749-1-xiyou.wangcong@gmail.com (mailing list archive) |
---|---|
State | Changes Requested |
Delegated to: | Netdev Maintainers |
Headers | show |
Series | [net] vsock: improve tap delivery accuracy | expand |
On Tue, May 02, 2023 at 04:14:18PM -0400, Stefan Hajnoczi wrote: > On Tue, May 02, 2023 at 10:44:04AM -0700, Cong Wang wrote: > > From: Cong Wang <cong.wang@bytedance.com> > > > > When virtqueue_add_sgs() fails, the skb is put back to send queue, > > we should not deliver the copy to tap device in this case. So we > > need to move virtio_transport_deliver_tap_pkt() down after all > > possible failures. > > > > Fixes: 82dfb540aeb2 ("VSOCK: Add virtio vsock vsockmon hooks") > > Cc: Stefan Hajnoczi <stefanha@redhat.com> > > Cc: Stefano Garzarella <sgarzare@redhat.com> > > Cc: Bobby Eshleman <bobby.eshleman@bytedance.com> > > Signed-off-by: Cong Wang <cong.wang@bytedance.com> > > --- > > net/vmw_vsock/virtio_transport.c | 5 ++--- > > 1 file changed, 2 insertions(+), 3 deletions(-) > > > > diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c > > index e95df847176b..055678628c07 100644 > > --- a/net/vmw_vsock/virtio_transport.c > > +++ b/net/vmw_vsock/virtio_transport.c > > @@ -109,9 +109,6 @@ virtio_transport_send_pkt_work(struct work_struct *work) > > if (!skb) > > break; > > > > - virtio_transport_deliver_tap_pkt(skb); > > - reply = virtio_vsock_skb_reply(skb); > > - > > sg_init_one(&hdr, virtio_vsock_hdr(skb), sizeof(*virtio_vsock_hdr(skb))); > > sgs[out_sg++] = &hdr; > > if (skb->len > 0) { > > @@ -128,6 +125,8 @@ virtio_transport_send_pkt_work(struct work_struct *work) > > break; > > } > > > > + virtio_transport_deliver_tap_pkt(skb); > > + reply = virtio_vsock_skb_reply(skb); > > I don't remember the reason for the ordering, but I'm pretty sure it was > deliberate. Probably because the payload buffers could be freed as soon > as virtqueue_add_sgs() is called. > > If that's no longer true with Bobby's skbuff code, then maybe it's safe > to monitor packets after they have been sent. > > Stefan Hey Stefan, Unfortunately, skbuff doesn't change that behavior. If I understand correctly, the problem flow you are describing would be something like this: Thread 0 Thread 1 guest:virtqueue_add_sgs()[@send_pkt_work] host:vhost_vq_get_desc()[@handle_tx_kick] host:vhost_add_used() host:vhost_signal() guest:virtqueue_get_buf()[@tx_work] guest:consume_skb() guest:deliver_tap_pkt()[@send_pkt_work] ^ use-after-free Which I guess is possible because the receiver can consume the new scatterlist during the processing kicked off for a previous batch? (doesn't have to wait for the subsequent kick) Best, Bobby
On Wed, May 03, 2023 at 09:39:13AM -0400, Stefan Hajnoczi wrote: > On Sun, Apr 16, 2023 at 04:49:00AM +0000, Bobby Eshleman wrote: > > On Tue, May 02, 2023 at 04:14:18PM -0400, Stefan Hajnoczi wrote: > > > On Tue, May 02, 2023 at 10:44:04AM -0700, Cong Wang wrote: > > > > From: Cong Wang <cong.wang@bytedance.com> > > > > > > > > When virtqueue_add_sgs() fails, the skb is put back to send queue, > > > > we should not deliver the copy to tap device in this case. So we > > > > need to move virtio_transport_deliver_tap_pkt() down after all > > > > possible failures. > > > > > > > > Fixes: 82dfb540aeb2 ("VSOCK: Add virtio vsock vsockmon hooks") > > > > Cc: Stefan Hajnoczi <stefanha@redhat.com> > > > > Cc: Stefano Garzarella <sgarzare@redhat.com> > > > > Cc: Bobby Eshleman <bobby.eshleman@bytedance.com> > > > > Signed-off-by: Cong Wang <cong.wang@bytedance.com> > > > > --- > > > > net/vmw_vsock/virtio_transport.c | 5 ++--- > > > > 1 file changed, 2 insertions(+), 3 deletions(-) > > > > > > > > diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c > > > > index e95df847176b..055678628c07 100644 > > > > --- a/net/vmw_vsock/virtio_transport.c > > > > +++ b/net/vmw_vsock/virtio_transport.c > > > > @@ -109,9 +109,6 @@ virtio_transport_send_pkt_work(struct work_struct *work) > > > > if (!skb) > > > > break; > > > > > > > > - virtio_transport_deliver_tap_pkt(skb); > > > > - reply = virtio_vsock_skb_reply(skb); > > > > - > > > > sg_init_one(&hdr, virtio_vsock_hdr(skb), sizeof(*virtio_vsock_hdr(skb))); > > > > sgs[out_sg++] = &hdr; > > > > if (skb->len > 0) { > > > > @@ -128,6 +125,8 @@ virtio_transport_send_pkt_work(struct work_struct *work) > > > > break; > > > > } > > > > > > > > + virtio_transport_deliver_tap_pkt(skb); > > > > + reply = virtio_vsock_skb_reply(skb); > > > > > > I don't remember the reason for the ordering, but I'm pretty sure it was > > > deliberate. Probably because the payload buffers could be freed as soon > > > as virtqueue_add_sgs() is called. > > > > > > If that's no longer true with Bobby's skbuff code, then maybe it's safe > > > to monitor packets after they have been sent. > > > > > > Stefan > > > > Hey Stefan, > > > > Unfortunately, skbuff doesn't change that behavior. > > > > If I understand correctly, the problem flow you are describing > > would be something like this: > > > > Thread 0 Thread 1 > > guest:virtqueue_add_sgs()[@send_pkt_work] > > > > host:vhost_vq_get_desc()[@handle_tx_kick] > > host:vhost_add_used() > > host:vhost_signal() > > guest:virtqueue_get_buf()[@tx_work] > > guest:consume_skb() > > > > guest:deliver_tap_pkt()[@send_pkt_work] > > ^ use-after-free > > > > Which I guess is possible because the receiver can consume the new > > scatterlist during the processing kicked off for a previous batch? > > (doesn't have to wait for the subsequent kick) > > Yes, drivers must assume that the device completes request before > virtqueue_add_sgs() returns. For example, the device is allowed to poll > the virtqueue memory and may see the new descriptors immediately. > > I haven't audited the current vsock code path to determine whether it's > possible to reach consume_skb() before deliver_tap_pkt() returns, so I > can't say whether it's safe or not. > I see, thanks for the clarification. Best, Bobby
On Wed, May 03, 2023 at 09:38:50AM +0200, Stefano Garzarella wrote: > On Sun, Apr 16, 2023 at 04:49:00AM +0000, Bobby Eshleman wrote: > > On Tue, May 02, 2023 at 04:14:18PM -0400, Stefan Hajnoczi wrote: > > > On Tue, May 02, 2023 at 10:44:04AM -0700, Cong Wang wrote: > > > > From: Cong Wang <cong.wang@bytedance.com> > > > > > > > > When virtqueue_add_sgs() fails, the skb is put back to send queue, > > > > we should not deliver the copy to tap device in this case. So we > > > > need to move virtio_transport_deliver_tap_pkt() down after all > > > > possible failures. > > > > > > > > Fixes: 82dfb540aeb2 ("VSOCK: Add virtio vsock vsockmon hooks") > > > > Cc: Stefan Hajnoczi <stefanha@redhat.com> > > > > Cc: Stefano Garzarella <sgarzare@redhat.com> > > > > Cc: Bobby Eshleman <bobby.eshleman@bytedance.com> > > > > Signed-off-by: Cong Wang <cong.wang@bytedance.com> > > > > --- > > > > net/vmw_vsock/virtio_transport.c | 5 ++--- > > > > 1 file changed, 2 insertions(+), 3 deletions(-) > > > > > > > > diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c > > > > index e95df847176b..055678628c07 100644 > > > > --- a/net/vmw_vsock/virtio_transport.c > > > > +++ b/net/vmw_vsock/virtio_transport.c > > > > @@ -109,9 +109,6 @@ virtio_transport_send_pkt_work(struct work_struct *work) > > > > if (!skb) > > > > break; > > > > > > > > - virtio_transport_deliver_tap_pkt(skb); > > > > - reply = virtio_vsock_skb_reply(skb); > > > > - > > > > sg_init_one(&hdr, virtio_vsock_hdr(skb), sizeof(*virtio_vsock_hdr(skb))); > > > > sgs[out_sg++] = &hdr; > > > > if (skb->len > 0) { > > > > @@ -128,6 +125,8 @@ virtio_transport_send_pkt_work(struct work_struct *work) > > > > break; > > > > } > > > > > > > > + virtio_transport_deliver_tap_pkt(skb); > > I would move only the virtio_transport_deliver_tap_pkt(), > virtio_vsock_skb_reply() is not related. > > > > > + reply = virtio_vsock_skb_reply(skb); > > > > > > I don't remember the reason for the ordering, but I'm pretty sure it was > > > deliberate. Probably because the payload buffers could be freed as soon > > > as virtqueue_add_sgs() is called. > > > > > > If that's no longer true with Bobby's skbuff code, then maybe it's safe > > > to monitor packets after they have been sent. > > > > > > Stefan > > > > Hey Stefan, > > > > Unfortunately, skbuff doesn't change that behavior. > > > > If I understand correctly, the problem flow you are describing > > would be something like this: > > > > Thread 0 Thread 1 > > guest:virtqueue_add_sgs()[@send_pkt_work] > > > > host:vhost_vq_get_desc()[@handle_tx_kick] > > host:vhost_add_used() > > host:vhost_signal() > > guest:virtqueue_get_buf()[@tx_work] > > guest:consume_skb() > > > > guest:deliver_tap_pkt()[@send_pkt_work] > > ^ use-after-free > > > > Which I guess is possible because the receiver can consume the new > > scatterlist during the processing kicked off for a previous batch? > > (doesn't have to wait for the subsequent kick) > > This is true, but both `send_pkt_work` and `tx_work` hold `tx_lock`, so can > they really go in parallel? > Oh good point, the tx_lock synchronizes it: Thread 0 Thread 1 guest:virtqueue_add_sgs()[@send_pkt_work] host:vhost_vq_get_desc()[@handle_tx_kick] host:vhost_add_used() host:vhost_signal() guest:mutex_lock()[@tx_work] guest:deliver_tap_pkt()[@send_pkt_work] guest:mutex_unlock() guest:virtqueue_get_buf()[@tx_work] guest:consume_skb() I'm pretty sure this should be safe. Best, Bobby
On Tue, May 02, 2023 at 10:44:04AM -0700, Cong Wang wrote: > From: Cong Wang <cong.wang@bytedance.com> > > When virtqueue_add_sgs() fails, the skb is put back to send queue, > we should not deliver the copy to tap device in this case. So we > need to move virtio_transport_deliver_tap_pkt() down after all > possible failures. > > Fixes: 82dfb540aeb2 ("VSOCK: Add virtio vsock vsockmon hooks") > Cc: Stefan Hajnoczi <stefanha@redhat.com> > Cc: Stefano Garzarella <sgarzare@redhat.com> > Cc: Bobby Eshleman <bobby.eshleman@bytedance.com> > Signed-off-by: Cong Wang <cong.wang@bytedance.com> Reviewed-by: Simon Horman <simon.horman@corigine.com>
On Tue, May 02, 2023 at 10:44:04AM -0700, Cong Wang wrote: > From: Cong Wang <cong.wang@bytedance.com> > > When virtqueue_add_sgs() fails, the skb is put back to send queue, > we should not deliver the copy to tap device in this case. So we > need to move virtio_transport_deliver_tap_pkt() down after all > possible failures. > > Fixes: 82dfb540aeb2 ("VSOCK: Add virtio vsock vsockmon hooks") > Cc: Stefan Hajnoczi <stefanha@redhat.com> > Cc: Stefano Garzarella <sgarzare@redhat.com> > Cc: Bobby Eshleman <bobby.eshleman@bytedance.com> > Signed-off-by: Cong Wang <cong.wang@bytedance.com> > --- > net/vmw_vsock/virtio_transport.c | 5 ++--- > 1 file changed, 2 insertions(+), 3 deletions(-) > > diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c > index e95df847176b..055678628c07 100644 > --- a/net/vmw_vsock/virtio_transport.c > +++ b/net/vmw_vsock/virtio_transport.c > @@ -109,9 +109,6 @@ virtio_transport_send_pkt_work(struct work_struct *work) > if (!skb) > break; > > - virtio_transport_deliver_tap_pkt(skb); > - reply = virtio_vsock_skb_reply(skb); > - > sg_init_one(&hdr, virtio_vsock_hdr(skb), sizeof(*virtio_vsock_hdr(skb))); > sgs[out_sg++] = &hdr; > if (skb->len > 0) { > @@ -128,6 +125,8 @@ virtio_transport_send_pkt_work(struct work_struct *work) > break; > } > > + virtio_transport_deliver_tap_pkt(skb); > + reply = virtio_vsock_skb_reply(skb); I don't remember the reason for the ordering, but I'm pretty sure it was deliberate. Probably because the payload buffers could be freed as soon as virtqueue_add_sgs() is called. If that's no longer true with Bobby's skbuff code, then maybe it's safe to monitor packets after they have been sent. Stefan
On Sun, Apr 16, 2023 at 04:49:00AM +0000, Bobby Eshleman wrote: >On Tue, May 02, 2023 at 04:14:18PM -0400, Stefan Hajnoczi wrote: >> On Tue, May 02, 2023 at 10:44:04AM -0700, Cong Wang wrote: >> > From: Cong Wang <cong.wang@bytedance.com> >> > >> > When virtqueue_add_sgs() fails, the skb is put back to send queue, >> > we should not deliver the copy to tap device in this case. So we >> > need to move virtio_transport_deliver_tap_pkt() down after all >> > possible failures. >> > >> > Fixes: 82dfb540aeb2 ("VSOCK: Add virtio vsock vsockmon hooks") >> > Cc: Stefan Hajnoczi <stefanha@redhat.com> >> > Cc: Stefano Garzarella <sgarzare@redhat.com> >> > Cc: Bobby Eshleman <bobby.eshleman@bytedance.com> >> > Signed-off-by: Cong Wang <cong.wang@bytedance.com> >> > --- >> > net/vmw_vsock/virtio_transport.c | 5 ++--- >> > 1 file changed, 2 insertions(+), 3 deletions(-) >> > >> > diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c >> > index e95df847176b..055678628c07 100644 >> > --- a/net/vmw_vsock/virtio_transport.c >> > +++ b/net/vmw_vsock/virtio_transport.c >> > @@ -109,9 +109,6 @@ virtio_transport_send_pkt_work(struct work_struct *work) >> > if (!skb) >> > break; >> > >> > - virtio_transport_deliver_tap_pkt(skb); >> > - reply = virtio_vsock_skb_reply(skb); >> > - >> > sg_init_one(&hdr, virtio_vsock_hdr(skb), sizeof(*virtio_vsock_hdr(skb))); >> > sgs[out_sg++] = &hdr; >> > if (skb->len > 0) { >> > @@ -128,6 +125,8 @@ virtio_transport_send_pkt_work(struct work_struct *work) >> > break; >> > } >> > >> > + virtio_transport_deliver_tap_pkt(skb); I would move only the virtio_transport_deliver_tap_pkt(), virtio_vsock_skb_reply() is not related. >> > + reply = virtio_vsock_skb_reply(skb); >> >> I don't remember the reason for the ordering, but I'm pretty sure it was >> deliberate. Probably because the payload buffers could be freed as soon >> as virtqueue_add_sgs() is called. >> >> If that's no longer true with Bobby's skbuff code, then maybe it's safe >> to monitor packets after they have been sent. >> >> Stefan > >Hey Stefan, > >Unfortunately, skbuff doesn't change that behavior. > >If I understand correctly, the problem flow you are describing >would be something like this: > >Thread 0 Thread 1 >guest:virtqueue_add_sgs()[@send_pkt_work] > > host:vhost_vq_get_desc()[@handle_tx_kick] > host:vhost_add_used() > host:vhost_signal() > guest:virtqueue_get_buf()[@tx_work] > guest:consume_skb() > >guest:deliver_tap_pkt()[@send_pkt_work] >^ use-after-free > >Which I guess is possible because the receiver can consume the new >scatterlist during the processing kicked off for a previous batch? >(doesn't have to wait for the subsequent kick) This is true, but both `send_pkt_work` and `tx_work` hold `tx_lock`, so can they really go in parallel? Thanks, Stefano
On Sun, Apr 16, 2023 at 04:49:00AM +0000, Bobby Eshleman wrote: > On Tue, May 02, 2023 at 04:14:18PM -0400, Stefan Hajnoczi wrote: > > On Tue, May 02, 2023 at 10:44:04AM -0700, Cong Wang wrote: > > > From: Cong Wang <cong.wang@bytedance.com> > > > > > > When virtqueue_add_sgs() fails, the skb is put back to send queue, > > > we should not deliver the copy to tap device in this case. So we > > > need to move virtio_transport_deliver_tap_pkt() down after all > > > possible failures. > > > > > > Fixes: 82dfb540aeb2 ("VSOCK: Add virtio vsock vsockmon hooks") > > > Cc: Stefan Hajnoczi <stefanha@redhat.com> > > > Cc: Stefano Garzarella <sgarzare@redhat.com> > > > Cc: Bobby Eshleman <bobby.eshleman@bytedance.com> > > > Signed-off-by: Cong Wang <cong.wang@bytedance.com> > > > --- > > > net/vmw_vsock/virtio_transport.c | 5 ++--- > > > 1 file changed, 2 insertions(+), 3 deletions(-) > > > > > > diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c > > > index e95df847176b..055678628c07 100644 > > > --- a/net/vmw_vsock/virtio_transport.c > > > +++ b/net/vmw_vsock/virtio_transport.c > > > @@ -109,9 +109,6 @@ virtio_transport_send_pkt_work(struct work_struct *work) > > > if (!skb) > > > break; > > > > > > - virtio_transport_deliver_tap_pkt(skb); > > > - reply = virtio_vsock_skb_reply(skb); > > > - > > > sg_init_one(&hdr, virtio_vsock_hdr(skb), sizeof(*virtio_vsock_hdr(skb))); > > > sgs[out_sg++] = &hdr; > > > if (skb->len > 0) { > > > @@ -128,6 +125,8 @@ virtio_transport_send_pkt_work(struct work_struct *work) > > > break; > > > } > > > > > > + virtio_transport_deliver_tap_pkt(skb); > > > + reply = virtio_vsock_skb_reply(skb); > > > > I don't remember the reason for the ordering, but I'm pretty sure it was > > deliberate. Probably because the payload buffers could be freed as soon > > as virtqueue_add_sgs() is called. > > > > If that's no longer true with Bobby's skbuff code, then maybe it's safe > > to monitor packets after they have been sent. > > > > Stefan > > Hey Stefan, > > Unfortunately, skbuff doesn't change that behavior. > > If I understand correctly, the problem flow you are describing > would be something like this: > > Thread 0 Thread 1 > guest:virtqueue_add_sgs()[@send_pkt_work] > > host:vhost_vq_get_desc()[@handle_tx_kick] > host:vhost_add_used() > host:vhost_signal() > guest:virtqueue_get_buf()[@tx_work] > guest:consume_skb() > > guest:deliver_tap_pkt()[@send_pkt_work] > ^ use-after-free > > Which I guess is possible because the receiver can consume the new > scatterlist during the processing kicked off for a previous batch? > (doesn't have to wait for the subsequent kick) Yes, drivers must assume that the device completes request before virtqueue_add_sgs() returns. For example, the device is allowed to poll the virtqueue memory and may see the new descriptors immediately. I haven't audited the current vsock code path to determine whether it's possible to reach consume_skb() before deliver_tap_pkt() returns, so I can't say whether it's safe or not. Stefan
diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c index e95df847176b..055678628c07 100644 --- a/net/vmw_vsock/virtio_transport.c +++ b/net/vmw_vsock/virtio_transport.c @@ -109,9 +109,6 @@ virtio_transport_send_pkt_work(struct work_struct *work) if (!skb) break; - virtio_transport_deliver_tap_pkt(skb); - reply = virtio_vsock_skb_reply(skb); - sg_init_one(&hdr, virtio_vsock_hdr(skb), sizeof(*virtio_vsock_hdr(skb))); sgs[out_sg++] = &hdr; if (skb->len > 0) { @@ -128,6 +125,8 @@ virtio_transport_send_pkt_work(struct work_struct *work) break; } + virtio_transport_deliver_tap_pkt(skb); + reply = virtio_vsock_skb_reply(skb); if (reply) { struct virtqueue *rx_vq = vsock->vqs[VSOCK_VQ_RX]; int val;