[FFmpeg-devel] [PATCH] lavu/videotoolbox: add support for memory mapping frames

[Cameron Gutman]

> On Jan 17, 2022, at 3:07 AM, Anton Khirnov <anton at khirnov.net> wrote:
> 
> Quoting Cameron Gutman (2022-01-10 09:17:37)
>> 
>>> On Jan 9, 2022, at 3:24 AM, Anton Khirnov <anton at khirnov.net> wrote:
>>> 
>>> Quoting Cameron Gutman (2022-01-03 01:33:19)
>>>> Signed-off-by: Cameron Gutman <aicommander at gmail.com>
>>>> ---
>>>> libavutil/hwcontext_videotoolbox.c | 25 +++++++++++++++++++++++++
>>>> 1 file changed, 25 insertions(+)
>>>> 
>>>> diff --git a/libavutil/hwcontext_videotoolbox.c b/libavutil/hwcontext_videotoolbox.c
>>>> index 0a8dbe9f33..026127d412 100644
>>>> --- a/libavutil/hwcontext_videotoolbox.c
>>>> +++ b/libavutil/hwcontext_videotoolbox.c
>>>> @@ -711,6 +711,30 @@ fail:
>>>>    return err;
>>>> }
>>>> 
>>>> +static int vt_map_from(AVHWFramesContext *hwfc, AVFrame *dst,
>>>> +                       const AVFrame *src, int flags)
>>>> +{
>>>> +    int err;
>>>> +
>>>> +    if (dst->format == AV_PIX_FMT_NONE)
>>>> +        dst->format = hwfc->sw_format;
>>>> +    else if (dst->format != hwfc->sw_format)
>>>> +        return AVERROR(ENOSYS);
>>>> +
>>>> +    err = vt_map_frame(hwfc, dst, src, flags);
>>>> +    if (err)
>>>> +        return err;
>>>> +
>>>> +    dst->width  = src->width;
>>>> +    dst->height = src->height;
>>>> +
>>>> +    err = av_frame_copy_props(dst, src);
>>>> +    if (err)
>>>> +        return err;
>>> 
>>> Don't you need to unmap the frame in this error path?
>>> 
>> 
>> Maybe? It's complicated...
>> 
>> The mapping is still written to dst and it will be unmapped when
>> av_frame_unref() is called on dst. However, if the caller wants to try again
>> with same dst frame for some reason, then it looks like it will leak the first
>> "failed" mapping. AFAICT, another call to ff_hwframe_map_create() will
>> overwrite dst->buf[0] without unrefing it first, but that makes sense given
>> that the docs say "dst should be blank" (arguably that "should" ought to be a
>> "must" in this case). However, this isn't the full story.
>> 
>> None of the existing map_from() implementations (VAAPI, DRM, DXVA2) actually
>> unmap when av_frame_copy_props() fails. The only ones that don't have this bug
>> are OpenCL and Vulkan, and that's only because they forgot to call
>> av_frame_copy_props() entirely!
>> 
>> Ideally, you'd have nothing modified in dst when av_hwframe_map() fails,
>> but that isn't the case in practice. Much of the mapping process involves
>> irreversible changes to the dst frame, including overwriting dst->format,
>> dst->width, dst->height, dst->data, dst->linesize, even partially copied
>> frame properties prior to av_frame_copy_props() failing.
>> 
>> Given these limitations, it seems like the only sane thing to do with a dst
>> frame after failure of av_hwframe_map() (other than ENOSYS) is to unref/free.
>> The frame is basically in an undefined state after such a failure. If that's
>> the case, then we don't actually have a leak here since av_frame_unref()
>> will do the right thing and free the old mapping.
> 
> Right, but who will call this av_frame_unref(). The doxy for
> av_hwframe_map() does not specify what precisely happens on failure, but
> I would expect it to clean dst.
> 

Whoever allocated the frame will unref/free it when they are done with it.

For example, vf_hwmap.c does this:

        dst = av_frame_alloc();
        if (!dst) {
            av_frame_free(&src);
            return NULL;
        }

        err = av_hwframe_map(dst, src, ctx->mode);
        if (err) {
            av_log(avctx, AV_LOG_ERROR, "Failed to map frame to "
                   "software: %d.\n", err);
            av_frame_free(&src);
            av_frame_free(&dst);
            return NULL;
        }

So the mapping will be freed in the call to av_frame_free(&dst) in this case.

Where we might run into issues with leaks is if someone does:

        /* Try NV12 first */
        dst->format = AV_PIX_FMT_NV12;
        err = av_hwframe_map(dst, src, AV_HWFRAME_MAP_READ);
        if (err) {
            /* That didn't work. Now try YUV420P */
            dst->format = AV_PIX_FMT_YUV420P;
            err = av_hwframe_map(dst, src, AV_HWFRAME_MAP_READ);
        }

Assuming that first av_hwframe_map() fails in av_frame_copy_props(), the second
call (if successful) will cause the first mapping to be leaked. For this case,
it makes sense to clear dst on failure.

The docs say "dst should be blank", so you could also argue that the second
call to av_hwframe_map() is incorrect, since it is not guaranteed that dst is
blank at that point (and today it is guaranteed _not_ to be blank for any of
the map_from() implementations in the case that av_frame_copy_props() fails).

If we do want dst to be cleared upon failure, a good place to do it might be in
av_hwframe_map() itself (since none of the existing map_from() implementations
do it themselves). That would formalize the behavior of cleaning dst on failure
rather than leaving the various hwcontext implementations to handle it how they
choose.

> -- 
> Anton Khirnov
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".