[PATCH] media: v4l2: Fix memleak in videobuf_read_one

[PATCH] media: v4l2: Fix memleak in videobuf_read_one

[Dinghao Liu]

When videobuf_waiton() fails, we should execute clean
functions to prevent memleak. It's the same when
__videobuf_copy_to_user() fails.

Fixes: 7a7d9a89d0307 ("V4L/DVB (6251): Replace video-buf to a more generic approach")
Signed-off-by: Dinghao Liu <dinghao.liu@zju.edu.cn>
---
 drivers/media/v4l2-core/videobuf-core.c | 12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)

diff --git a/drivers/media/v4l2-core/videobuf-core.c b/drivers/media/v4l2-core/videobuf-core.c
index 606a271bdd2d..0709b75d11cd 100644
--- a/drivers/media/v4l2-core/videobuf-core.c
+++ b/drivers/media/v4l2-core/videobuf-core.c
@@ -924,8 +924,12 @@ ssize_t videobuf_read_one(struct videobuf_queue *q,
 
 	/* wait until capture is done */
 	retval = videobuf_waiton(q, q->read_buf, nonblocking, 1);
-	if (0 != retval)
+	if (retval != 0) {
+		q->ops->buf_release(q, q->read_buf);
+		kfree(q->read_buf);
+		q->read_buf = NULL;
 		goto done;
+	}
 
 	CALL(q, sync, q, q->read_buf);
 
@@ -940,8 +944,12 @@ ssize_t videobuf_read_one(struct videobuf_queue *q,
 
 	/* Copy to userspace */
 	retval = __videobuf_copy_to_user(q, q->read_buf, data, count, nonblocking);
-	if (retval < 0)
+	if (retval < 0) {
+		q->ops->buf_release(q, q->read_buf);
+		kfree(q->read_buf);
+		q->read_buf = NULL;
 		goto done;
+	}
 
 	q->read_off += retval;
 	if (q->read_off == q->read_buf->size) {
-- 
2.17.1

Re: [PATCH] media: v4l2: Fix memleak in videobuf_read_one

[Hans Verkuil]

On 05/01/2021 08:59, Dinghao Liu wrote:
> When videobuf_waiton() fails, we should execute clean
> functions to prevent memleak. It's the same when
> __videobuf_copy_to_user() fails.
> 
> Fixes: 7a7d9a89d0307 ("V4L/DVB (6251): Replace video-buf to a more generic approach")
> Signed-off-by: Dinghao Liu <dinghao.liu@zju.edu.cn>
> ---
>  drivers/media/v4l2-core/videobuf-core.c | 12 ++++++++++--
>  1 file changed, 10 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/media/v4l2-core/videobuf-core.c b/drivers/media/v4l2-core/videobuf-core.c
> index 606a271bdd2d..0709b75d11cd 100644
> --- a/drivers/media/v4l2-core/videobuf-core.c
> +++ b/drivers/media/v4l2-core/videobuf-core.c
> @@ -924,8 +924,12 @@ ssize_t videobuf_read_one(struct videobuf_queue *q,
>  
>  	/* wait until capture is done */
>  	retval = videobuf_waiton(q, q->read_buf, nonblocking, 1);
> -	if (0 != retval)
> +	if (retval != 0) {
> +		q->ops->buf_release(q, q->read_buf);
> +		kfree(q->read_buf);
> +		q->read_buf = NULL;
>  		goto done;
> +	}

I'm fairly certain that this is wrong: if waiton returns an error, then
that means that the wait is either interrupted or that we are in non-blocking
mode and no buffer has arrived yet. In that case you just go to done since
there is nothing to clean up.

>  
>  	CALL(q, sync, q, q->read_buf);
>  
> @@ -940,8 +944,12 @@ ssize_t videobuf_read_one(struct videobuf_queue *q,
>  
>  	/* Copy to userspace */
>  	retval = __videobuf_copy_to_user(q, q->read_buf, data, count, nonblocking);
> -	if (retval < 0)
> +	if (retval < 0) {
> +		q->ops->buf_release(q, q->read_buf);
> +		kfree(q->read_buf);
> +		q->read_buf = NULL;
>  		goto done;

I'm not sure about this either: if userspace gave a crappy pointer and this
copy_to_user fails, then that doesn't mean you should release the buffer.
The next read() might have a valid pointer or, more likely, the application
exits or crashes and everything is cleaned up when the filehandle is closed.

> +	}
>  
>  	q->read_off += retval;
>  	if (q->read_off == q->read_buf->size) {
> 

Do you have actual proof that this is a memleak? I don't want to mess around
with the old videobuf unless you can show me that there is a real bug.

Regards,

	Hans

Re: Re: [PATCH] media: v4l2: Fix memleak in videobuf_read_one

[dinghao.liu]

> On 05/01/2021 08:59, Dinghao Liu wrote:
> > When videobuf_waiton() fails, we should execute clean
> > functions to prevent memleak. It's the same when
> > __videobuf_copy_to_user() fails.
> > 
> > Fixes: 7a7d9a89d0307 ("V4L/DVB (6251): Replace video-buf to a more generic approach")
> > Signed-off-by: Dinghao Liu <dinghao.liu@zju.edu.cn>
> > ---
> >  drivers/media/v4l2-core/videobuf-core.c | 12 ++++++++++--
> >  1 file changed, 10 insertions(+), 2 deletions(-)
> > 
> > diff --git a/drivers/media/v4l2-core/videobuf-core.c b/drivers/media/v4l2-core/videobuf-core.c
> > index 606a271bdd2d..0709b75d11cd 100644
> > --- a/drivers/media/v4l2-core/videobuf-core.c
> > +++ b/drivers/media/v4l2-core/videobuf-core.c
> > @@ -924,8 +924,12 @@ ssize_t videobuf_read_one(struct videobuf_queue *q,
> >  
> >  	/* wait until capture is done */
> >  	retval = videobuf_waiton(q, q->read_buf, nonblocking, 1);
> > -	if (0 != retval)
> > +	if (retval != 0) {
> > +		q->ops->buf_release(q, q->read_buf);
> > +		kfree(q->read_buf);
> > +		q->read_buf = NULL;
> >  		goto done;
> > +	}
> 
> I'm fairly certain that this is wrong: if waiton returns an error, then
> that means that the wait is either interrupted or that we are in non-blocking
> mode and no buffer has arrived yet. In that case you just go to done since
> there is nothing to clean up.
> 

I found there was a similar error handling in videobuf_read_zerocopy(), where
q->read_buf was freed on failure of videobuf_waiton(), thus I reported this as
a memleak. Do you think the error handling in videobuf_read_zerocopy() is right?

> >  
> >  	CALL(q, sync, q, q->read_buf);
> >  
> > @@ -940,8 +944,12 @@ ssize_t videobuf_read_one(struct videobuf_queue *q,
> >  
> >  	/* Copy to userspace */
> >  	retval = __videobuf_copy_to_user(q, q->read_buf, data, count, nonblocking);
> > -	if (retval < 0)
> > +	if (retval < 0) {
> > +		q->ops->buf_release(q, q->read_buf);
> > +		kfree(q->read_buf);
> > +		q->read_buf = NULL;
> >  		goto done;
> 
> I'm not sure about this either: if userspace gave a crappy pointer and this
> copy_to_user fails, then that doesn't mean you should release the buffer.
> The next read() might have a valid pointer or, more likely, the application
> exits or crashes and everything is cleaned up when the filehandle is closed.
> 

You are right. Let's keep this part as it was for security.

Regards,
Dinghao