[PATCH] bpf: fix reference count leak in bpf_prog_test_run_xdp()

[PATCH] bpf: fix reference count leak in bpf_prog_test_run_xdp()

[Tetsuo Handa]

syzbot is reporting

  unregister_netdevice: waiting for sit0 to become free. Usage count = 2

problem. A debug printk() patch found that a refcount is obtained at
xdp_convert_md_to_buff() from bpf_prog_test_run_xdp().

According to commit ec94670fcb3b ("bpf: Support specifying ingress via
xdp_md context in BPF_PROG_TEST_RUN"), the refcount obtained by
xdp_convert_md_to_buff() will be released by xdp_convert_buff_to_md().

Therefore, we can consider that the error handling path introduced by
commit 1c1949982524 ("bpf: introduce frags support to
bpf_prog_test_run_xdp()") forgot to call xdp_convert_buff_to_md().

Reported-by: syzbot+881d65229ca4f9ae8c84@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=881d65229ca4f9ae8c84
Fixes: 1c1949982524 ("bpf: introduce frags support to bpf_prog_test_run_xdp()")
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
---
Since syzbot has no reproducer for this problem, I can't test this patch.

 net/bpf/test_run.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/net/bpf/test_run.c b/net/bpf/test_run.c
index 655efac6f133..9a16293ba14b 100644
--- a/net/bpf/test_run.c
+++ b/net/bpf/test_run.c
@@ -1355,13 +1355,13 @@ int bpf_prog_test_run_xdp(struct bpf_prog *prog, const union bpf_attr *kattr,
 
 			if (sinfo->nr_frags == MAX_SKB_FRAGS) {
 				ret = -ENOMEM;
-				goto out;
+				goto out_put_dev;
 			}
 
 			page = alloc_page(GFP_KERNEL);
 			if (!page) {
 				ret = -ENOMEM;
-				goto out;
+				goto out_put_dev;
 			}
 
 			frag = &sinfo->frags[sinfo->nr_frags++];
@@ -1373,7 +1373,7 @@ int bpf_prog_test_run_xdp(struct bpf_prog *prog, const union bpf_attr *kattr,
 			if (copy_from_user(page_address(page), data_in + size,
 					   data_len)) {
 				ret = -EFAULT;
-				goto out;
+				goto out_put_dev;
 			}
 			sinfo->xdp_frags_size += data_len;
 			size += data_len;
@@ -1388,6 +1388,7 @@ int bpf_prog_test_run_xdp(struct bpf_prog *prog, const union bpf_attr *kattr,
 		ret = bpf_test_run_xdp_live(prog, &xdp, repeat, batch_size, &duration);
 	else
 		ret = bpf_test_run(prog, &xdp, repeat, &retval, &duration, true);
+out_put_dev:
 	/* We convert the xdp_buff back to an xdp_md before checking the return
 	 * code so the reference count of any held netdevice will be decremented
 	 * even if the test run failed.
-- 
2.47.3

Re: [PATCH] bpf: fix reference count leak in bpf_prog_test_run_xdp()

[Toke Høiland-Jørgensen]

Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> writes:

> syzbot is reporting
>
>   unregister_netdevice: waiting for sit0 to become free. Usage count = 2
>
> problem. A debug printk() patch found that a refcount is obtained at
> xdp_convert_md_to_buff() from bpf_prog_test_run_xdp().
>
> According to commit ec94670fcb3b ("bpf: Support specifying ingress via
> xdp_md context in BPF_PROG_TEST_RUN"), the refcount obtained by
> xdp_convert_md_to_buff() will be released by xdp_convert_buff_to_md().
>
> Therefore, we can consider that the error handling path introduced by
> commit 1c1949982524 ("bpf: introduce frags support to
> bpf_prog_test_run_xdp()") forgot to call xdp_convert_buff_to_md().
>
> Reported-by: syzbot+881d65229ca4f9ae8c84@syzkaller.appspotmail.com
> Closes: https://syzkaller.appspot.com/bug?extid=881d65229ca4f9ae8c84
> Fixes: 1c1949982524 ("bpf: introduce frags support to bpf_prog_test_run_xdp()")
> Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
> ---
> Since syzbot has no reproducer for this problem, I can't test this patch.
>
>  net/bpf/test_run.c | 7 ++++---
>  1 file changed, 4 insertions(+), 3 deletions(-)
>
> diff --git a/net/bpf/test_run.c b/net/bpf/test_run.c
> index 655efac6f133..9a16293ba14b 100644
> --- a/net/bpf/test_run.c
> +++ b/net/bpf/test_run.c
> @@ -1355,13 +1355,13 @@ int bpf_prog_test_run_xdp(struct bpf_prog *prog, const union bpf_attr *kattr,
>  
>  			if (sinfo->nr_frags == MAX_SKB_FRAGS) {
>  				ret = -ENOMEM;
> -				goto out;
> +				goto out_put_dev;
>  			}
>  
>  			page = alloc_page(GFP_KERNEL);
>  			if (!page) {
>  				ret = -ENOMEM;
> -				goto out;
> +				goto out_put_dev;
>  			}
>  
>  			frag = &sinfo->frags[sinfo->nr_frags++];
> @@ -1373,7 +1373,7 @@ int bpf_prog_test_run_xdp(struct bpf_prog *prog, const union bpf_attr *kattr,
>  			if (copy_from_user(page_address(page), data_in + size,
>  					   data_len)) {
>  				ret = -EFAULT;
> -				goto out;
> +				goto out_put_dev;
>  			}
>  			sinfo->xdp_frags_size += data_len;
>  			size += data_len;
> @@ -1388,6 +1388,7 @@ int bpf_prog_test_run_xdp(struct bpf_prog *prog, const union bpf_attr *kattr,
>  		ret = bpf_test_run_xdp_live(prog, &xdp, repeat, batch_size, &duration);
>  	else
>  		ret = bpf_test_run(prog, &xdp, repeat, &retval, &duration, true);
> +out_put_dev:
>  	/* We convert the xdp_buff back to an xdp_md before checking the return
>  	 * code so the reference count of any held netdevice will be decremented
>  	 * even if the test run failed.

Hmm, this will end up call bpf_ctx_finish() in the error path, which I'm
not sure we want?

Could we just move the xdp_convert_md_to_buff() call to after the frags
have been copied? Not sure there's technically any dependency there,
even though it does look a little off?

-Toke

Re: [PATCH] bpf: fix reference count leak in bpf_prog_test_run_xdp()

[Tetsuo Handa]

On 2026/01/08 23:01, Toke Høiland-Jørgensen wrote:
> Hmm, this will end up call bpf_ctx_finish() in the error path, which I'm
> not sure we want?

Excuse me, but I don't think bpf_ctx_finish() will be called, for

+out_put_dev:
 	/* We convert the xdp_buff back to an xdp_md before checking the return
 	 * code so the reference count of any held netdevice will be decremented
 	 * even if the test run failed.
 	 */
 	xdp_convert_buff_to_md(&xdp, ctx);
 	if (ret) // <== ret was set to non-0 value immediately before the "goto out_put_dev;" line.
 		goto out;
 
 	size = xdp.data_end - xdp.data_meta + sinfo->xdp_frags_size;
 	ret = bpf_test_finish(kattr, uattr, xdp.data_meta, sinfo, size, sinfo->xdp_frags_size,
 			      retval, duration);
 	if (!ret)
 		ret = bpf_ctx_finish(kattr, uattr, ctx,
 				     sizeof(struct xdp_md));

> 
> Could we just move the xdp_convert_md_to_buff() call to after the frags
> have been copied? Not sure there's technically any dependency there,
> even though it does look a little off?

Unless

	xdp_md->data = xdp->data - xdp->data_meta;
	xdp_md->data_end = xdp->data_end - xdp->data_meta;

in xdp_convert_buff_to_md() lines do something bad for the error path,
I think this change will be safe.

Re: [PATCH] bpf: fix reference count leak in bpf_prog_test_run_xdp()

[Toke Høiland-Jørgensen]

Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> writes:

> On 2026/01/08 23:01, Toke Høiland-Jørgensen wrote:
>> Hmm, this will end up call bpf_ctx_finish() in the error path, which I'm
>> not sure we want?
>
> Excuse me, but I don't think bpf_ctx_finish() will be called, for
>
> +out_put_dev:
>  	/* We convert the xdp_buff back to an xdp_md before checking the return
>  	 * code so the reference count of any held netdevice will be decremented
>  	 * even if the test run failed.
>  	 */
>  	xdp_convert_buff_to_md(&xdp, ctx);
>  	if (ret) // <== ret was set to non-0 value immediately before the "goto out_put_dev;" line.
>  		goto out;

Oh, right; I think my brain just pattern matched on "if (ret) right
after a function call" and assumed there was an assignment to ret there
as well :D

Okay, not the clearest code flow, but not sure there's a good way to
make it clearer without quite a bit of refactoring.

>  
>  	size = xdp.data_end - xdp.data_meta + sinfo->xdp_frags_size;
>  	ret = bpf_test_finish(kattr, uattr, xdp.data_meta, sinfo, size, sinfo->xdp_frags_size,
>  			      retval, duration);
>  	if (!ret)
>  		ret = bpf_ctx_finish(kattr, uattr, ctx,
>  				     sizeof(struct xdp_md));
>
>> 
>> Could we just move the xdp_convert_md_to_buff() call to after the frags
>> have been copied? Not sure there's technically any dependency there,
>> even though it does look a little off?
>
> Unless
>
> 	xdp_md->data = xdp->data - xdp->data_meta;
> 	xdp_md->data_end = xdp->data_end - xdp->data_meta;
>
> in xdp_convert_buff_to_md() lines do something bad for the error path,
> I think this change will be safe.

Yeah, sure, this should be fine.

Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>

Re: [PATCH] bpf: fix reference count leak in bpf_prog_test_run_xdp()

[patchwork-bot+netdevbpf]

Hello:

This patch was applied to bpf/bpf.git (master)
by Alexei Starovoitov <ast@kernel.org>:

On Thu, 8 Jan 2026 21:36:48 +0900 you wrote:
> syzbot is reporting
> 
>   unregister_netdevice: waiting for sit0 to become free. Usage count = 2
> 
> problem. A debug printk() patch found that a refcount is obtained at
> xdp_convert_md_to_buff() from bpf_prog_test_run_xdp().
> 
> [...]

Here is the summary with links:
  - bpf: fix reference count leak in bpf_prog_test_run_xdp()
    https://git.kernel.org/bpf/bpf/c/ec69daabe452

You are awesome, thank you!
-- 
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html