The Linux Kernel Mailing List
 help / color / mirror / Atom feed
* [PATCH] vhost/net: bound the count of VHOST_SET_FEATURES_ARRAY
@ 2026-11-22 12:09 Hengyu Liang
  2026-11-22 14:25 ` Michael S. Tsirkin
  0 siblings, 1 reply; 3+ messages in thread
From: Hengyu Liang @ 2026-11-22 12:09 UTC (permalink / raw)
  To: Michael S. Tsirkin
  Cc: Jason Wang, Eugenio Pérez, Paolo Abeni, kvm, virtualization,
	netdev, linux-kernel, stable

Commit 333c515d1896 ("vhost-net: allow configuring extended features")
added VHOST_SET_FEATURES_ARRAY. The ioctl reads the number of feature
words from user space and checks that every word above the supported
ones is zero, one word at a time in a loop.

However, the loop has no upper bound and no scheduling point. As of now,
a caller can pass a count of U64_MAX and a pointer into a large
zero-filled mapping. The ioctl then stays in the kernel for as long as
the mapping is readable, and a fatal signal is not handled until it
ends. On a v6.18 debug kernel booted with preempt=none, two such calls
from a non-root user, each on a 4 GiB mapping, froze a 2-CPU VM for
56 s with soft lockup reports.

Interfaces that take an extensible structure from user space limit it
to one page (openat2, clone3, sched_setattr).

On a KASAN kernel with a 1 GiB mapping: before commit 333c515d1896
("vhost-net: allow configuring extended features"), the ioctl does not
exist and the call fails with EPERM at once. After that commit, it
fails with EFAULT after 26.0 s. With this patch, it fails with E2BIG at
once. A reproducer is available on request.

This patch will reject counts above one page of words (512 on a 4 KiB
page, or 32768 feature bits; 128 are defined today) before the walk
starts.

Fixes: 333c515d1896 ("vhost-net: allow configuring extended features")
Cc: stable@vger.kernel.org # v6.17+
Signed-off-by: Hengyu Liang <hengyul@cs.unc.edu>
---
I found this while comparing the behavior of different kernel versions.
I used an LLM to analyze the problem and to draft the patch and this
message, and I analyzed the result myself.

Tested on the vhost tree at 44293796526f with KASAN and lockdep: counts
of 0, 2, 3 and 512 words still succeed, a nonzero word above the
supported ones still gives EOPNOTSUPP, and VHOST_GET_FEATURES_ARRAY is
unchanged.
 drivers/vhost/net.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
index 631855064b3d..22e06b44d2ed 100644
--- a/drivers/vhost/net.c
+++ b/drivers/vhost/net.c
@@ -1795,6 +1795,8 @@ static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 	case VHOST_SET_FEATURES_ARRAY:
 		if (copy_from_user(&count, featurep, sizeof(count)))
 			return -EFAULT;
+		if (count > PAGE_SIZE / sizeof(u64))
+			return -E2BIG;
 
 		virtio_features_zero(all_features);
 		argp += sizeof(u64);

base-commit: 44293796526f4343639f87b2acc53811c78b5d54
-- 
2.53.0


^ permalink raw reply related	[flat|nested] 3+ messages in thread

* Re: [PATCH] vhost/net: bound the count of VHOST_SET_FEATURES_ARRAY
  2026-11-22 12:09 [PATCH] vhost/net: bound the count of VHOST_SET_FEATURES_ARRAY Hengyu Liang
@ 2026-11-22 14:25 ` Michael S. Tsirkin
  2026-11-22 14:38   ` Hengyu Liang
  0 siblings, 1 reply; 3+ messages in thread
From: Michael S. Tsirkin @ 2026-11-22 14:25 UTC (permalink / raw)
  To: Hengyu Liang
  Cc: Jason Wang, Eugenio Pérez, Paolo Abeni, kvm, virtualization,
	netdev, linux-kernel, stable

On Sun, Nov 22, 2026 at 07:09:19AM -0500, Hengyu Liang wrote:
> Commit 333c515d1896 ("vhost-net: allow configuring extended features")
> added VHOST_SET_FEATURES_ARRAY. The ioctl reads the number of feature
> words from user space and checks that every word above the supported
> ones is zero, one word at a time in a loop.
> 
> However, the loop has no upper bound and no scheduling point. As of now,
> a caller can pass a count of U64_MAX and a pointer into a large
> zero-filled mapping. The ioctl then stays in the kernel for as long as
> the mapping is readable, and a fatal signal is not handled until it
> ends. On a v6.18 debug kernel booted with preempt=none, two such calls
> from a non-root user, each on a 4 GiB mapping, froze a 2-CPU VM for
> 56 s with soft lockup reports.
> 
> Interfaces that take an extensible structure from user space limit it
> to one page (openat2, clone3, sched_setattr).
> 
> On a KASAN kernel with a 1 GiB mapping: before commit 333c515d1896
> ("vhost-net: allow configuring extended features"), the ioctl does not
> exist and the call fails with EPERM at once. After that commit, it
> fails with EFAULT after 26.0 s. With this patch, it fails with E2BIG at
> once. A reproducer is available on request.
> 
> This patch will reject counts above one page of words (512 on a 4 KiB
> page, or 32768 feature bits; 128 are defined today) before the walk
> starts.
> 
> Fixes: 333c515d1896 ("vhost-net: allow configuring extended features")
> Cc: stable@vger.kernel.org # v6.17+
> Signed-off-by: Hengyu Liang <hengyul@cs.unc.edu>
> ---
> I found this while comparing the behavior of different kernel versions.
> I used an LLM to analyze the problem and to draft the patch and this
> message, and I analyzed the result myself.
> 
> Tested on the vhost tree at 44293796526f with KASAN and lockdep: counts
> of 0, 2, 3 and 512 words still succeed, a nonzero word above the
> supported ones still gives EOPNOTSUPP, and VHOST_GET_FEATURES_ARRAY is
> unchanged.
>  drivers/vhost/net.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
> index 631855064b3d..22e06b44d2ed 100644
> --- a/drivers/vhost/net.c
> +++ b/drivers/vhost/net.c
> @@ -1795,6 +1795,8 @@ static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
>  	case VHOST_SET_FEATURES_ARRAY:
>  		if (copy_from_user(&count, featurep, sizeof(count)))
>  			return -EFAULT;
> +		if (count > PAGE_SIZE / sizeof(u64))
> +			return -E2BIG;
>  
>  		virtio_features_zero(all_features);
>  		argp += sizeof(u64);

????
this is followed by:

                virtio_features_zero(all_features);
                argp += sizeof(u64);
                copied = min(count, (u64)VIRTIO_FEATURES_U64S);
                if (copy_from_user(all_features, argp, copied * sizeof(u64)))
                        return -EFAULT;

copies at most 8 bytes.


did you test that this patch actually helps anything?


> base-commit: 44293796526f4343639f87b2acc53811c78b5d54
> -- 
> 2.53.0


^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: [PATCH] vhost/net: bound the count of VHOST_SET_FEATURES_ARRAY
  2026-11-22 14:25 ` Michael S. Tsirkin
@ 2026-11-22 14:38   ` Hengyu Liang
  0 siblings, 0 replies; 3+ messages in thread
From: Hengyu Liang @ 2026-11-22 14:38 UTC (permalink / raw)
  To: Michael S. Tsirkin
  Cc: Jason Wang, Eugenio Pérez, Paolo Abeni, kvm, virtualization,
	netdev, linux-kernel, stable

On Sun, Nov 22, 2026 at 09:25:37AM -0500, Michael S. Tsirkin wrote:
> this is followed by:
>
>                 virtio_features_zero(all_features);
>                 argp += sizeof(u64);
>                 copied = min(count, (u64)VIRTIO_FEATURES_U64S);
>                 if (copy_from_user(all_features, argp, copied * sizeof(u64)))
>                         return -EFAULT;
>
> copies at most 8 bytes.
>
>
> did you test that this patch actually helps anything?

That first copy is bounded, yes. The unbounded part is the loop right
after it, which is outside the diff context:

		for (i = copied; i < count; ++i) {
			if (copy_from_user(&features, featurep + 1 + i,
					   sizeof(features)))
				return -EFAULT;
			if (features)
				return -EOPNOTSUPP;
		}

It reads one u64 per iteration up to the count from user space. With
count = U64_MAX and a large zero-filled mapping, it only stops at the
first unreadable address.

Yes, I tested it, on the vhost tree at 44293796526f (KASAN + lockdep),
count = U64_MAX:

  1 GiB zero-filled mapping:
    without the patch: EFAULT after 27.4 s
    with the patch:    E2BIG at once

  2 GiB mapping, SIGKILL sent 1 s into the ioctl:
    without the patch: the process is reaped 46.7 s after the kill
    with the patch:    the ioctl has already returned

I have a small reproducer (a few lines of C). Since I found this with
the help of an LLM, Documentation/process/security-bugs.rst asks me to
treat it as public and not post the reproducer to a public list, so I
have not included it here. I am happy to send it to you off-list if it
would help.

Thanks,
Hengyu

^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2026-11-22 14:38 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-22 12:09 [PATCH] vhost/net: bound the count of VHOST_SET_FEATURES_ARRAY Hengyu Liang
2026-11-22 14:25 ` Michael S. Tsirkin
2026-11-22 14:38   ` Hengyu Liang

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox