RE: [REGRESSION 6.12.90 -> 6.12.94] vsock/virtio: large AF_VSOCK transfers reset under backpressure

["Brien Oberstein" <brienpub@gmail.com>]

Hi Stefano,

Thanks, that matches what I'm seeing: large transfers reset mid-stream
instead of the sender being throttled (reliable above ~1.5 MB, fine below
~90 KB).

The bind for me: it's not just this mail bridge -- I use AF_VSOCK for a few
host/guest services, some of which open their own sockets, so the per-socket
buffer workaround can't cover them all. That leaves pinning 6.12.90 (losing
the DoS fix and further kernel updates) as the only blanket option.

A few quick questions:

1. Is a -stable backport of the merging fix likely, and roughly when?
2. Could a smaller interim land in -stable sooner (e.g. more default
   headroom) without reopening the DoS?
3. Will the fix guarantee backpressure for any packet size, or just widen
   the margin?

Happy to test any patch -- I have a solid reproducer and can turn it around
in a day. I'll also file this as a tracked regression so it's not lost.

Thanks again,
Brien

#regzbot introduced: v6.12.90..v6.12.94

-----Original Message-----
From: Stefano Garzarella <sgarzare@redhat.com> 
Sent: Monday, June 22, 2026 6:08 AM
To: Brien Oberstein <brienpub@gmail.com>
Cc: edumazet@google.com
Subject: Re: [REGRESSION 6.12.90 -> 6.12.94] vsock/virtio: large AF_VSOCK transfers reset under backpressure

On Sun, Jun 21, 2026 at 08:42:41AM -0400, Brien Oberstein wrote:
>Hi Stefano, Eric,

Hi Brien,

>
>I'm hitting a regression in the 6.12.y stable series: a bulk transfer 
>over
>AF_VSOCK is torn down mid-stream once the message is large enough to
>exercise receiver-side backpressure. By stable version it lands on
>6.12.94; 6.12.90 is fine.
>
>Setup
>-----
>A host process mails a guest's postfix over an AF_VSOCK bridge:
>
>  host msmtp --(unix sock)--> socat --(AF_VSOCK: host CID 2 ->
>    guest CID 101, port 20025)--> [guest] socat --(TCP 127.0.0.1:25)-->
>    postfix
>
>postfix (TLS-terminating, then writing to its queue) drains the stream
>slower than the host writes it, so the per-socket vsock buffer fills
>during a large message.
>
>Symptom (guest, 6.12.94)
>------------------------
>The guest-side socat exits status=1 mid-transfer and postfix logs:
>
>  postfix/smtpd: NNN: lost connection after DATA (153330 bytes)
>    from localhost[127.0.0.1]
>  postfix/smtpd: disconnect ... data=0/1 commands=5/6
>
>On the host, msmtp reports:
>
>  msmtp: cannot write to TLS connection: The TLS connection was
>    non-properly terminated.        (sendmail exit 74 / EX_TEMPFAIL)
>
>So the AF_VSOCK connection is dropped while data is still flowing, rather
>than the sender being throttled by the credit-based flow control.
>
>Reproduction
>------------
>Send messages of increasing size through the bridge:
>
>  body <= ~88 KB : always succeeds
>  body ~354 KB   : intermittent failure
>  body >= 1.5 MB : fails 12/12
>
>On 6.12.90 the identical test passes 20/20, including 1.5 MB x12,
>2.4 MB x3, 4 MB x3 and 8 MB x2. The only variable is the guest kernel.
>
>Bisection
>---------
>6.12.91, .92 and .93 carry no vsock changes. 6.12.94 pulled in three
>vsock/virtio commits:
>
>  1eca304f  vsock/virtio: fix potential unbounded skb queue
>  f3bf0f3b  vsock/virtio: fix skb overhead accounting to preserve
>            full buf_alloc
>  149205a1  vsock/virtio: fix skb overhead overflow on 32-bit builds
>
>The behaviour (drop/reset under a fast sender + slow receiver instead of
>applying backpressure) makes 1eca304f the prime suspect, but I have only
>A/B tested whole stable releases, not the individual commits.

Yep, I'm working on a followup to improve the status.

Basically, the memory management in AF_VSOCK has always been broken. The 
patches you mentioned are designed to prevent one peer from consuming 
all of the other peer’s memory.
Instead of counting only the payload bytes, we now also take the packet 
metadata into account, using a socket buffer that is double the size set 
(default 256 KB).

So if your system is sending small packets, then this is likely hitting 
this issue.

My advice for now is to increase the socket buffer size. Thanks to 
VMware, AF_VSOCK have specific sockopts :-(:
- SO_VM_SOCKETS_BUFFER_SIZE (0)
- SO_VM_SOCKETS_BUFFER_MAX_SIZE (2)

I suggest to set both to 16 MB (MAX should be set first).
I tried this with socat and seems to work:
   socat VSOCK-LISTEN:4242,setsockopt=40:2:x0000000001000000,setsockopt=40:0:x0000000001000000

Hope this helps.

In the mean time, I'm working on a follow-up for net-next to ensure that 
packets are merged when we exceed a threshold; we might be able to 
backport this to stable, but I'm not sure.

Thanks,
Stefano