[PATCH v3] media: hantro: Implement support for encoder commands
Chen-Yu Tsai
wenst at chromium.org
Mon Mar 21 21:17:40 PDT 2022
On Mon, Mar 21, 2022 at 8:50 PM Ezequiel Garcia
<ezequiel at vanguardiasur.com.ar> wrote:
>
> Hi Chen-Yu,
>
> On Mon, Mar 21, 2022 at 1:12 AM Chen-Yu Tsai <wenst at chromium.org> wrote:
> >
> > On Sun, Mar 20, 2022 at 8:18 PM Ezequiel Garcia
> > <ezequiel at vanguardiasur.com.ar> wrote:
> > >
> > > Hi Chen-Yu, Benjamin, Nicolas,
> > >
> > > Sorry for the late review.
> > >
> > > On Tue, Mar 01, 2022 at 12:22:25PM +0800, Chen-Yu Tsai wrote:
> > > > The V4L2 stateful encoder uAPI specification requires that drivers
> > > > support the ENCODER_CMD ioctl to allow draining of buffers. This
> > > > however was not implemented, and causes issues for some userspace
> > > > applications.
> > > >
> > > > Implement support for the ENCODER_CMD ioctl using v4l2-mem2mem helpers.
> > > > This is entirely based on existing code found in the vicodec test
> > > > driver.
> > > >
> > > > Fixes: 775fec69008d ("media: add Rockchip VPU JPEG encoder driver")
> > > > Signed-off-by: Chen-Yu Tsai <wenst at chromium.org>
> > > > Reviewed-by: Benjamin Gaignard <benjamin.gaignard at collabora.com>
> > > > ---
> > > >
> > > > Changes since v2:
> > > > - Dropped RFC tag
> > > > - Added Reviewed-by from Benjamin
> > > > - Replace direct access to vb->planes[i].bytesused with
> > > > vb2_set_plane_payload()
> > > >
> > > > Changes since v1:
> > > > - Correctly handle last buffers that are empty
> > > > - Correctly handle last buffers that just got queued
> > > > - Disable (TRY_)ENCODER_CMD ioctls for hantro decoder
> > > >
> > > > This is based on linux-next-20220208, and was tested on RK3399 with
> > > > Gstreamer running the JPEG encoder. It was also tested on ChromeOS
> > > > 5.10 on Kevin with the video encoder used in ChromeOS ARC, which
> > > > requires this. For ChromeOS, both encoder and decoder tests were run
> > > > to check for regressions.
> > > >
> > >
> > > Sounds great, thanks for running those tests!
> > >
> > > > Everything really works OK now, but since I'm not very familiar with
> > > > the mem2mem framework, I might be missing something, causing resource
> > > > leaks. Hence this patch is labeled RFC.
> > > >
> > >
> > > It would be great to apply this to a mainline-ish kernel,
> > > and run some fluster and some stress-tests to ensure this is not
> > > regressing decoding in some way.
> >
> > I can try to get the fluster suite running.
> >
> > Not sure what stress-tests you had in mind though? Upstream only has the
> > JPEG encoder. Maybe multiple gstreamer encoding/decoding instances running
> > in parallel? Or even multiple fluster runs in parallel?
> >
>
> Exactly, I was imagining multiple gstreamer pipelines in parallel,
> focusing on decoding.
>
> > I had just put my RK3399 device away though. Have to dig it out again.
> >
>
> Sorry about that. I will try to shorten my review cadence.
No worries. I just meant that it might take me a few days to get
through testing.
ChenYu
> Thanks,
> Ezequiel
>
> > > > Last, I suppose we could also add support for (TRY_)DECODER_CMD now?
> > > >
> > > > drivers/staging/media/hantro/hantro_drv.c | 17 +++++-
> > > > drivers/staging/media/hantro/hantro_v4l2.c | 68 +++++++++++++++++++++-
> > > > 2 files changed, 81 insertions(+), 4 deletions(-)
> > > >
> > > > diff --git a/drivers/staging/media/hantro/hantro_drv.c b/drivers/staging/media/hantro/hantro_drv.c
> > > > index bc9bcb4eaf46..99bc650a5a93 100644
> > > > --- a/drivers/staging/media/hantro/hantro_drv.c
> > > > +++ b/drivers/staging/media/hantro/hantro_drv.c
> > > > @@ -56,6 +56,10 @@ dma_addr_t hantro_get_ref(struct hantro_ctx *ctx, u64 ts)
> > > > return hantro_get_dec_buf_addr(ctx, buf);
> > > > }
> > > >
> > > > +static const struct v4l2_event hantro_eos_event = {
> > > > + .type = V4L2_EVENT_EOS
> > > > +};
> > > > +
> > > > static void hantro_job_finish_no_pm(struct hantro_dev *vpu,
> > > > struct hantro_ctx *ctx,
> > > > enum vb2_buffer_state result)
> > > > @@ -73,6 +77,12 @@ static void hantro_job_finish_no_pm(struct hantro_dev *vpu,
> > > > src->sequence = ctx->sequence_out++;
> > > > dst->sequence = ctx->sequence_cap++;
> > > >
> > > > + if (v4l2_m2m_is_last_draining_src_buf(ctx->fh.m2m_ctx, src)) {
> > > > + dst->flags |= V4L2_BUF_FLAG_LAST;
> > > > + v4l2_event_queue_fh(&ctx->fh, &hantro_eos_event);
> > > > + v4l2_m2m_mark_stopped(ctx->fh.m2m_ctx);
> > > > + }
> > > > +
> > > > v4l2_m2m_buf_done_and_job_finish(ctx->dev->m2m_dev, ctx->fh.m2m_ctx,
> > > > result);
> > > > }
> > > > @@ -807,10 +817,13 @@ static int hantro_add_func(struct hantro_dev *vpu, unsigned int funcid)
> > > > snprintf(vfd->name, sizeof(vfd->name), "%s-%s", match->compatible,
> > > > funcid == MEDIA_ENT_F_PROC_VIDEO_ENCODER ? "enc" : "dec");
> > > >
> > > > - if (funcid == MEDIA_ENT_F_PROC_VIDEO_ENCODER)
> > > > + if (funcid == MEDIA_ENT_F_PROC_VIDEO_ENCODER) {
> > > > vpu->encoder = func;
> > > > - else
> > > > + } else {
> > > > vpu->decoder = func;
> > > > + v4l2_disable_ioctl(vfd, VIDIOC_TRY_ENCODER_CMD);
> > > > + v4l2_disable_ioctl(vfd, VIDIOC_ENCODER_CMD);
> > > > + }
> > > >
> > > > video_set_drvdata(vfd, vpu);
> > > >
> > > > diff --git a/drivers/staging/media/hantro/hantro_v4l2.c b/drivers/staging/media/hantro/hantro_v4l2.c
> > > > index 67148ba346f5..aa10ecd04c9c 100644
> > > > --- a/drivers/staging/media/hantro/hantro_v4l2.c
> > > > +++ b/drivers/staging/media/hantro/hantro_v4l2.c
> > > > @@ -628,6 +628,39 @@ static int vidioc_s_selection(struct file *file, void *priv,
> > > > return 0;
> > > > }
> > > >
> > > > +static const struct v4l2_event hantro_eos_event = {
> > > > + .type = V4L2_EVENT_EOS
> > > > +};
> > > > +
> > > > +static int vidioc_encoder_cmd(struct file *file, void *priv,
> > > > + struct v4l2_encoder_cmd *ec)
> > > > +{
> > > > + struct hantro_ctx *ctx = fh_to_ctx(priv);
> > > > + int ret;
> > > > +
> > > > + ret = v4l2_m2m_ioctl_try_encoder_cmd(file, priv, ec);
> > > > + if (ret < 0)
> > > > + return ret;
> > > > +
> > > > + if (!vb2_is_streaming(v4l2_m2m_get_src_vq(ctx->fh.m2m_ctx)) ||
> > > > + !vb2_is_streaming(v4l2_m2m_get_dst_vq(ctx->fh.m2m_ctx)))
> > > > + return 0;
> > > > +
> > > > + ret = v4l2_m2m_ioctl_encoder_cmd(file, priv, ec);
> > > > + if (ret < 0)
> > > > + return ret;
> > > > +
> > > > + if (ec->cmd == V4L2_ENC_CMD_STOP &&
> > > > + v4l2_m2m_has_stopped(ctx->fh.m2m_ctx))
> > > > + v4l2_event_queue_fh(&ctx->fh, &hantro_eos_event);
> > > > +
> > > > + if (ec->cmd == V4L2_ENC_CMD_START &&
> > > > + v4l2_m2m_has_stopped(ctx->fh.m2m_ctx))
> > >
> > > This looks odd. The has_stopped flag is cleared by calling
> > > v4l2_m2m_ioctl_encoder_cmd so I can't see how it could be set here.
> > >
> > > This same pattern is in the vicodec driver, the change was introduced
> > > in d4d137de5f31d318ed9acdcdf359b9bd3920808b.
> >
> > Looking at the commit you pointed out, it seems the original intent was to
> > check the prior state. So it looks like d4d137de5f31d318ed9acdcdf359b9bd3920808b
> > introduced a bug.
> >
> > I'll take care of both drivers.
> >
> > > > + vb2_clear_last_buffer_dequeued(&ctx->fh.m2m_ctx->cap_q_ctx.q);
> > > > +
> > > > + return 0;
> > > > +}
> > > > +
> > > > const struct v4l2_ioctl_ops hantro_ioctl_ops = {
> > > > .vidioc_querycap = vidioc_querycap,
> > > > .vidioc_enum_framesizes = vidioc_enum_framesizes,
> > > > @@ -657,6 +690,9 @@ const struct v4l2_ioctl_ops hantro_ioctl_ops = {
> > > >
> > > > .vidioc_g_selection = vidioc_g_selection,
> > > > .vidioc_s_selection = vidioc_s_selection,
> > > > +
> > > > + .vidioc_try_encoder_cmd = v4l2_m2m_ioctl_try_encoder_cmd,
> > > > + .vidioc_encoder_cmd = vidioc_encoder_cmd,
> > > > };
> > > >
> > > > static int
> > > > @@ -733,8 +769,12 @@ static int hantro_buf_prepare(struct vb2_buffer *vb)
> > > > * (for OUTPUT buffers, if userspace passes 0 bytesused, v4l2-core sets
> > > > * it to buffer length).
> > > > */
> > > > - if (V4L2_TYPE_IS_CAPTURE(vq->type))
> > > > - vb2_set_plane_payload(vb, 0, pix_fmt->plane_fmt[0].sizeimage);
> > > > + if (V4L2_TYPE_IS_CAPTURE(vq->type)) {
> > > > + if (ctx->is_encoder)
> > > > + vb2_set_plane_payload(vb, 0, 0);
> > >
> > > This looks like some fix, that could be applied independently of this patch?
> >
> > Yes. I seem to have accidentally squashed that in. Of the top of my head,
> > I think this was clearing the capture buffers to be empty, so unused
> > buffers returned to userspace do look "unused".
> >
> > I'll split this out.
> >
> >
> > Thanks
> > ChenYu
> >
> > > > + else
> > > > + vb2_set_plane_payload(vb, 0, pix_fmt->plane_fmt[0].sizeimage);
> > > > + }
> > > >
> > > > return 0;
> > > > }
> > > > @@ -744,6 +784,22 @@ static void hantro_buf_queue(struct vb2_buffer *vb)
> > > > struct hantro_ctx *ctx = vb2_get_drv_priv(vb->vb2_queue);
> > > > struct vb2_v4l2_buffer *vbuf = to_vb2_v4l2_buffer(vb);
> > > >
> > > > + if (V4L2_TYPE_IS_CAPTURE(vb->vb2_queue->type) &&
> > > > + vb2_is_streaming(vb->vb2_queue) &&
> > > > + v4l2_m2m_dst_buf_is_last(ctx->fh.m2m_ctx)) {
> > > > + unsigned int i;
> > > > +
> > > > + for (i = 0; i < vb->num_planes; i++)
> > > > + vb2_set_plane_payload(vb, i, 0);
> > > > +
> > > > + vbuf->field = V4L2_FIELD_NONE;
> > > > + vbuf->sequence = ctx->sequence_cap++;
> > > > +
> > > > + v4l2_m2m_last_buffer_done(ctx->fh.m2m_ctx, vbuf);
> > > > + v4l2_event_queue_fh(&ctx->fh, &hantro_eos_event);
> > > > + return;
> > > > + }
> > > > +
> > > > v4l2_m2m_buf_queue(ctx->fh.m2m_ctx, vbuf);
> > > > }
> > > >
> > > > @@ -759,6 +815,8 @@ static int hantro_start_streaming(struct vb2_queue *q, unsigned int count)
> > > > struct hantro_ctx *ctx = vb2_get_drv_priv(q);
> > > > int ret = 0;
> > > >
> > > > + v4l2_m2m_update_start_streaming_state(ctx->fh.m2m_ctx, q);
> > > > +
> > > > if (V4L2_TYPE_IS_OUTPUT(q->type))
> > > > ctx->sequence_out = 0;
> > > > else
> > > > @@ -831,6 +889,12 @@ static void hantro_stop_streaming(struct vb2_queue *q)
> > > > hantro_return_bufs(q, v4l2_m2m_src_buf_remove);
> > > > else
> > > > hantro_return_bufs(q, v4l2_m2m_dst_buf_remove);
> > > > +
> > > > + v4l2_m2m_update_stop_streaming_state(ctx->fh.m2m_ctx, q);
> > > > +
> > > > + if (V4L2_TYPE_IS_OUTPUT(q->type) &&
> > > > + v4l2_m2m_has_stopped(ctx->fh.m2m_ctx))
> > > > + v4l2_event_queue_fh(&ctx->fh, &hantro_eos_event);
> > > > }
> > > >
> > > > static void hantro_buf_request_complete(struct vb2_buffer *vb)
> > > > --
> > > > 2.35.1.574.g5d30c73bfb-goog
> > > >
> > >
> > > Thanks,
> > > Ezequiel
More information about the Linux-rockchip
mailing list