Re: [PATCH] media: venus: fix use after free for registeredbufs

[Jeffrey Kardatzke]

On Sat, Mar 7, 2020 at 9:24 AM Stanimir Varbanov
<stanimir.varbanov@xxxxxxxxxx> wrote:
>
> Hi Jeff,
>
> On 3/6/20 10:10 PM, Jeffrey Kardatzke wrote:
> > On Fri, Mar 6, 2020 at 1:03 AM Stanimir Varbanov
> > <stanimir.varbanov@xxxxxxxxxx> wrote:
> >>
> >> Hi Jeff,
> >>
> >> Thanks for the patch!
> >>
> >> On 3/6/20 2:23 AM, Jeffrey Kardatzke wrote:
> >>> In dynamic bufmode we do not manage the buffers in the registeredbufs
> >>> list, so do not add them there when they are initialized. Adding them
> >>> there was causing a use after free of the list_head struct in the buffer
> >>> when new buffers were allocated after existing buffers were freed.
> >>
> >> Is this fixing a real issue? How you come to it?
> >>
> > In our code we were allocating 64x64 capture queue buffers initially,
> > then got a resolution change event for the actual video resolution of
> > 320x256 so we freed all the existing capture buffers and allocated new
> > ones. I had noticed memory poisoning warnings in dmesg and tracked it
> > down to the patch I created here. This is only a problem when the
> > capture queue has its buffers freed and reallocated (which would
> > happen during any resolution change).
>
> Do you call STREAMOFF(CAPTURE) ?
>

Yes, we call STREAMOFF before we destroy the existing buffers and
allocate new ones.

> Better, could you share v4l2 debug logs:
>
> echo 0x3f > /sys/class/video4linux/videoX/dev_debug
>

I'll email you these off list since they are rather large.

> >
> >>>
> >>> Signed-off-by: Jeffrey Kardatzke <jkardatzke@xxxxxxxxxx>
> >>> ---
> >>>  drivers/media/platform/qcom/venus/helpers.c | 4 +++-
> >>>  1 file changed, 3 insertions(+), 1 deletion(-)
> >>>
> >>> diff --git a/drivers/media/platform/qcom/venus/helpers.c b/drivers/media/platform/qcom/venus/helpers.c
> >>> index bcc603804041..688a3593b49b 100644
> >>> --- a/drivers/media/platform/qcom/venus/helpers.c
> >>> +++ b/drivers/media/platform/qcom/venus/helpers.c
> >>> @@ -1054,8 +1054,10 @@ int venus_helper_vb2_buf_init(struct vb2_buffer *vb)
> >>>       buf->size = vb2_plane_size(vb, 0);
> >>>       buf->dma_addr = sg_dma_address(sgt->sgl);
> >>>
> >>> -     if (vb->type == V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE)
> >>> +     if (vb->type == V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE &&
> >>> +         !is_dynamic_bufmode(inst)) {
> >>
> >> If you add !is_dynamic_bufmode here, we will loose the reference frames
> >> mechanism (see venus_helper_release_buf_ref()) which is not good.
> >
> > In my testing, I never see venus_helper_release_buf_ref called.  I
> > think something is wrong with reference frame management. I'm also
>
> The mechanism is valid for Venus v1 and v3, might be you tried on v4
> where we have a set of DPB buffers and use them for reference frames.
>

We are using V4.

> > seeing failure in my tests that very much look like reference frames
> > that were dropped in the decoder (with or without my patch); but they
> > are not consistent.
> >
> >>
> >> Thus, I wonder (depending on when you observe the use-after-free issue)
> >> does this is the correct resolution of the problem.
> >
> > I agree this is likely not the right solution to the problem, there's
> > something deeper that's wrong I think because I never see events
> > coming back from hfi with the release buffer reference event.
> >>
> >>>               list_add_tail(&buf->reg_list, &inst->registeredbufs);
> >>> +     }
> >>>
> >>>       return 0;
> >>>  }
> >>>
> >>
> >> --
> >> regards,
> >> Stan
>
> --
> regards,
> Stan