On Wed, 10 Dec 2025 17:31:25 +0900, Jakub Kicinski wrote:

quoted

In zerocopy_fill_skb_from_iter(), if two copy operations are performed
and the first one succeeds while the second one fails, it returns a
failure but the count in iterator has already been decremented due to
the first successful copy. This ultimately affects the local variable
rest_len in virtio_transport_send_pkt_info(), causing the remaining
count in rest_len to be greater than the actual iterator count. As a
result, packet sending operations continue even when the iterator count
is zero, which further leads to skb->len being 0 and triggers the warning
reported by syzbot [1].

Therefore, if the zerocopy operation fails, we should revert the iterator
to its original state.

The iov_iter_revert() in skb_zerocopy_iter_stream() is no longer needed
and has been removed.

[1]
'send_pkt()' returns 0, but 4096 expected
WARNING: net/vmw_vsock/virtio_transport_common.c:430 at virtio_transport_send_pkt_info+0xd1e/0xef0 net/vmw_vsock/virtio_transport_common.c:428, CPU#1: syz.0.17/5986
Call Trace:
 virtio_transport_stream_enqueue net/vmw_vsock/virtio_transport_common.c:1113 [inline]
 virtio_transport_seqpacket_enqueue+0x143/0x1c0 net/vmw_vsock/virtio_transport_common.c:841
 vsock_connectible_sendmsg+0xabf/0x1040 net/vmw_vsock/af_vsock.c:2158
 sock_sendmsg_nosec net/socket.c:727 [inline]
 __sock_sendmsg+0x21c/0x270 net/socket.c:746

Reported-by: syzbot+28e5f3d207b14bae122a@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=28e5f3d207b14bae122a
Signed-off-by: Edward Adam Davis <redacted>
---
v3:
  - fix test tcp_zerocopy_maxfrags timeout
v2: https://lore.kernel.org/all/tencent_BA768766163C533724966E36344AAE754709@qq.com/ (local)
  - Remove iov_iter_revert() in skb_zerocopy_iter_stream()
v1: https://lore.kernel.org/all/tencent_387517772566B03DBD365896C036264AA809@qq.com/ (local)

Have you investigated the other callers? Given problems with previous
version of this patch I'm worried you have not. If you did please extend
the commit message with the appropriate explanation.

Are you asking if I investigated other zerocopy tests? NO.
The test results [T2] for this version of the patch do not show any
failures related to zerocopy.

[T2] https://patchwork.kernel.org/project/netdevbpf/patch/tencent_3C86DFD37A0374496263BE24483777D76305@qq.com

Alternative would be to add a _full() flavor of this API, but not sure
if other callers actually care.

quoted

diff --git a/net/core/datagram.c b/net/core/datagram.c
index c285c6465923..3a612ebbbe80 100644
--- a/net/core/datagram.c
+++ b/net/core/datagram.c

@@ -748,9 +748,13 @@ int __zerocopy_sg_from_iter(struct msghdr *msg, struct sock *sk,
 			    size_t length,
 			    struct net_devmem_dmabuf_binding *binding)
 {
+	struct iov_iter_state state;

nit: if you respin move this one line down

OK.

quoted

 	unsigned long orig_size = skb->truesize;
 	unsigned long truesize;
-	int ret;
+	int ret, orig_len;
+
+	iov_iter_save_state(from, &state);
+	orig_len = skb->len;

 	if (msg && msg->msg_ubuf && msg->sg_from_iter)
 		ret = msg->sg_from_iter(skb, from, length);

@@ -759,6 +763,9 @@ int __zerocopy_sg_from_iter(struct msghdr *msg, struct sock *sk,
 	else
 		ret = zerocopy_fill_skb_from_iter(skb, from, length);

+	if (ret == -EFAULT || (ret == -EMSGSIZE && skb->len == orig_len))

I'd think that for the purpose of reverting iter the second part of
this condition is completely moot. If skb len didn't change there should
be nothing to revert?

Regarding the second judgment condition, I aligned it with the condition
in skb_zerocopy_iter_stream(). During my testing, I only encountered
-EFAULT failures, not -EMSGSIZE.

quoted

+		iov_iter_restore(from, &state);
+
 	truesize = skb->truesize - orig_size;
 	if (sk && sk->sk_type == SOCK_STREAM) {
 		sk_wmem_queued_add(sk, truesize);

diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index a00808f7be6a..7b8836f668b7 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c

@@ -1908,7 +1908,6 @@ int skb_zerocopy_iter_stream(struct sock *sk, struct sk_buff *skb,
 		struct sock *save_sk = skb->sk;

 		/* Streams do not free skb on error. Reset to prev state. */
-		iov_iter_revert(&msg->msg_iter, skb->len - orig_len);
 		skb->sk = sk;
 		___pskb_trim(skb, orig_len);
 		skb->sk = save_sk;