Re: [RFC PATCH V2 4/4] platform: mtk-isp: Add Mediatek FD driver

[Tomasz Figa <tfiga-F7+t8E8rja9g9hUCZPvPmw@public.gmane.org>]

On Wed, Sep 4, 2019 at 12:38 PM Jerry-ch Chen
<Jerry-ch.Chen-NuS5LvNUpcJWk0Htik3J/w@public.gmane.org> wrote:
>
> Hi Tomasz,
>
> On Tue, 2019-09-03 at 20:05 +0800, Tomasz Figa wrote:
> > On Tue, Sep 3, 2019 at 8:46 PM Jerry-ch Chen <Jerry-ch.Chen-NuS5LvNUpcJWk0Htik3J/w@public.gmane.org> wrote:
> > >
> > > Hi Tomasz,
> > >
> > > On Tue, 2019-09-03 at 15:04 +0800, Tomasz Figa wrote:
> > > > On Tue, Sep 3, 2019 at 3:44 PM Jerry-ch Chen <Jerry-ch.Chen-NuS5LvNUpcJWk0Htik3J/w@public.gmane.org> wrote:
> > > > >
> > > > > On Tue, 2019-09-03 at 13:19 +0800, Tomasz Figa wrote:
> > > > > > On Mon, Sep 2, 2019 at 8:47 PM Jerry-ch Chen <Jerry-ch.Chen-NuS5LvNUpcJWk0Htik3J/w@public.gmane.org> wrote:
> > > > > > >
> > > > > > > Hi Tomasz,
> > > > > > >
> > > > > > > On Fri, 2019-08-30 at 16:33 +0800, Tomasz Figa wrote:
> > > > > > > > On Wed, Aug 28, 2019 at 11:00 AM Jerry-ch Chen
> > > > > > > > <Jerry-ch.Chen-NuS5LvNUpcJWk0Htik3J/w@public.gmane.org> wrote:
> > > > > > > > >
> > > > > > > > > Hi Tomasz,
> > > > > > > > >
> > > > > > > > > On Mon, 2019-08-26 at 14:36 +0800, Tomasz Figa wrote:
> > > > > > > > > > Hi Jerry,
> > > > > > > > > >
> > > > > > > > > > On Sun, Aug 25, 2019 at 6:18 PM Jerry-ch Chen
> > > > > > > > > > <Jerry-ch.Chen-NuS5LvNUpcJWk0Htik3J/w@public.gmane.org> wrote:
> > > > > > > > > > >
> > > > > > > > > > > Hi Tomasz,
> > > > > > > > > > >
> > > > > > > > > > > On Fri, 2019-08-02 at 16:28 +0800, Tomasz Figa wrote:
> > > > > > > > > > > > Hi Jerry,
> > > > > > > > > > > >
> > > > > > > > > > > > On Tue, Jul 09, 2019 at 04:41:12PM +0800, Jerry-ch Chen wrote:
> [snip]
> > > > > > > > > > [snip]
> > > > > > > > > >
> > > > > > > > > > > > > +static void mtk_fd_vb2_stop_streaming(struct vb2_queue *vq)
> > > > > > > > > > > > > +{
> > > > > > > > > > > > > +   struct mtk_fd_ctx *ctx = vb2_get_drv_priv(vq);
> > > > > > > > > > > > > +   struct vb2_buffer *vb;
> > > > > > > > > > > >
> > > > > > > > > > > > How do we guarantee here that the hardware isn't still accessing the buffers
> > > > > > > > > > > > removed below?
> > > > > > > > > > > >
> > > > > > > > > > > Maybe we can check the driver state flag and aborting the unfinished
> > > > > > > > > > > jobs?
> > > > > > > > > > > (fd_hw->state == FD_ENQ)
> > > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > Yes, we need to either cancel or wait for the currently processing
> > > > > > > > > > job. It depends on hardware capabilities, but cancelling is generally
> > > > > > > > > > preferred for the lower latency.
> > > > > > > > > >
> > > > > > > > > Ok, it the state is ENQ, then we can disable the FD hw by controlling
> > > > > > > > > the registers.
> > > > > > > > >
> > > > > > > > > for example:
> > > > > > > > >         writel(0x0, fd->fd_base + FD_HW_ENABLE);
> > > > > > > > >         writel(0x0, fd->fd_base + FD_INT_EN);
> > > > > > > > >
> > > > > > > >
> > > > > > > > What's exactly the effect of writing 0 to FD_HW_ENABLE?
> > > > > > > >
> > > > > > > Sorry, my last reply didn't solve the question,
> > > > > > > we should implement a mtk_fd_job_abort() for v4l2_m2m_ops().
> > > > > > >
> > > > > > > which is able to readl_poll_timeout_atomic()
> > > > > > > and check the HW busy bits in the register FD_INT_EN;
> > > > > > >
> > > > > > > if they are not cleared until timeout, we could handle the last
> > > > > > > processing job.
> > > > > > > Otherwise, the FD irq handler should have handled the last processing
> > > > > > > job and we could continue the stop_streaming().
> > > > > > >
> > > > > > > For job_abort():
> > > > > > > static void mtk_fd_job_abort(void *priv)
> > > > > > > {
> > > > > > >         struct mtk_fd_ctx *ctx = priv;
> > > > > > >         struct mtk_fd_dev *fd = ctx->fd_dev;
> > > > > > >         u32 val;
> > > > > > >         u32 ret;
> > > > > > >
> > > > > > >         ret = readl_poll_timeout_atomic(fd->fd_base + MTK_FD_REG_OFFSET_INT_EN,
> > > > > > >                                         val,
> > > > > > >                                         (val & MTK_FD_HW_BUSY_MASK) ==
> > > > > > >                                         MTK_FD_HW_STATE_IS_BUSY,
> > > > > > >                                         USEC_PER_MSEC, MTK_FD_STOP_HW_TIMEOUT);
> > > > > >
> > > > > > Hmm, would it be possible to avoid the busy wait by having a
> > > > > > completion that could be signalled from the interrupt handler?
> > > > > >
> > > > > > Best regards,
> > > > > > Tomasz
> > > > >
> > > > > I suppose that would be wakeup a wait queue in the interrupt handler,
> > > > > the the wait_event_interrupt_timeout() will be used in here and system
> > > > > suspend e.g. mtk_fd_suspend().
> > > >
> > > > Yes, that should work.
> > > >
> > > > > Or do you suggest to wait_event_interrupt_timeout() every frame in the
> > > > > mtk_fd_ipi_handler()?
> > > >
> > > > Nope, we shouldn't need that.
> > > >
> > > > > I think maybe the readl_poll_timeout_atomic would be good enough.
> > > > >
> > > >
> > > > Not really. Busy waiting should be avoided as much as possible. What's
> > > > the point of entering suspend if you end up burning the power by
> > > > spinning the CPU for some milliseconds?
> > > >
> > > Ok, I see, busy waiting is not a good idea, I will use the wait queue
> > > instead. the job abort will refine as following:
> > >
> > > static void mtk_fd_job_abort(void *priv)
> > > {
> > >         struct mtk_fd_ctx *ctx = priv;
> > >         struct mtk_fd_dev *fd = ctx->fd_dev;
> > >         u32 ret;
> > >
> > >         ret = wait_event_interruptible_timeout
> > >                 (fd->wq, (fd->fd_irq_result & MTK_FD_HW_IRQ_MASK),
> > >                  usecs_to_jiffies(50000));
> > >         if (ret)
> > >                 mtk_fd_hw_job_finish(fd, VB2_BUF_STATE_ERROR);
> > >         dev_dbg(fd->dev, "%s, ret:%d\n", __func__, ret);
> > >
> > >         fd->fd_irq_result = 0;
> > > }
> > >
> > > static struct v4l2_m2m_ops fd_m2m_ops = {
> > >         .device_run = mtk_fd_device_run,
> > >         .job_abort = mtk_fd_job_abort,
> >
> > I'm not sure we should be using the functon above as the .job_abort
> > callback. It's expected to abort the job, but we're just waiting for
> > it to finish. I think we should just call mtk_fd_job_abort() manually
> > from .stop_streaming.
> >
>
> Ok, I will fix it.
>
> > > };
> > >
> > > and we could use it in suspend.
> > > static int mtk_fd_suspend(struct device *dev)
> > > {
> > >         struct mtk_fd_dev *fd = dev_get_drvdata(dev);
> > >
> > >         if (pm_runtime_suspended(dev))
> > >                 return 0;
> > >
> > >         if (fd->fd_stream_count)
> > >                 mtk_fd_job_abort(fd->ctx);
> > >
> > >         /* suspend FD HW */
> > >         writel(0x0, fd->fd_base + MTK_FD_REG_OFFSET_INT_EN);
> > >         writel(0x0, fd->fd_base + MTK_FD_REG_OFFSET_HW_ENABLE);
> > >         clk_disable_unprepare(fd->fd_clk);
> > >         dev_dbg(dev, "%s:disable clock\n", __func__);
> > >
> > >         return 0;
> > > }
> > >
> > > static irqreturn_t mtk_fd_irq(int irq, void *data)
> > > {
> > >         struct mtk_fd_dev *fd = (struct mtk_fd_dev *)data;
> > >
> > >         fd->fd_irq_result = readl(fd->fd_base + MTK_FD_REG_OFFSET_INT_VAL);
> > >         wake_up_interruptible(&fd->wq);
> >
> > The wake up should be done at the very end of this function. Otherwise
> > we end up with m2m framework racing with the mtk_fd_hw_job_finish()
> > below.
> >
> > Also, I'd use a completion here rather than an open coded wait and
> > wake-up. The driver could reinit_completion() before queuing a job to
> > the hardware and the IRQ handler would complete(). There would be no
> > need to store the IRQ flags in driver data anymore.
> Ok, It will be refined as following:
>
> suspend and stop streaming will call mtk_fd_job_abort()
>
> static void mtk_fd_job_abort(struct mtk_fd_dev *fd)
> {
>         u32 ret;
>
>         ret = wait_for_completion_timeout(&fd->fd_irq_done,
>                                           msecs_to_jiffies(MTK_FD_HW_TIMEOUT));
>         if (ret)

For the _timeout variants, !ret means the timeout and ret > 0 means
that the wait ended successfully before.

Also please make sure that the timeout is big enough not to happen
normally on a heavily loaded system. Something like 1 second should be
good.

>                 mtk_fd_hw_job_finish(fd, VB2_BUF_STATE_ERROR);
> }
>
> complete at irq handler
>
> static irqreturn_t mtk_fd_irq(int irq, void *data)
> {
>         struct mtk_fd_dev *fd = (struct mtk_fd_dev *)data;
>
>         /* must read this register otherwise HW will keep sending irq */
>         readl(fd->fd_base + MTK_FD_REG_OFFSET_INT_VAL);
>         fd->output->number = readl(fd->fd_base + MTK_FD_REG_OFFSET_RESULT);
>         dev_dbg(fd->dev, "mtk_fd_face_num:%d\n", fd->output->number);
>
>         pm_runtime_put((fd->dev));
>         mtk_fd_hw_job_finish(fd, VB2_BUF_STATE_DONE);
>         complete(&fd->fd_irq_done);
>         return IRQ_HANDLED;
> }
>
> and reinit_completion before time before sending a job to HW
>
> static int mtk_fd_hw_job_exec(struct mtk_fd_dev *fd,
>                               struct fd_enq_param *fd_param)
> {
>         struct ipi_message fd_ipi_msg;
>         int ret;
>
>         pm_runtime_get_sync((fd->dev));
>
>         reinit_completion(&fd->fd_irq_done);
>         fd_ipi_msg.cmd_id = MTK_FD_IPI_CMD_ENQUEUE;
>         memcpy(&fd_ipi_msg.fd_enq_param, fd_param, sizeof(struct
> fd_enq_param));
>         ret = scp_ipi_send(fd->scp_pdev, SCP_IPI_FD_CMD, &fd_ipi_msg,
>                            sizeof(fd_ipi_msg), MTK_FD_IPI_SEND_TIMEOUT);
>         if (ret) {
>                 pm_runtime_put((fd->dev));
>                 mtk_fd_hw_job_finish(fd, VB2_BUF_STATE_ERROR);
>                 return ret;
>         }
>         return 0;
> }

Looks good, thanks. Please also don't forget about init_completion()
in driver probe.

Best regards,
Tomasz