Message-ID: <CAAFQd5CeH9GrsN8Z+DfqLGbmHMPbjHca8-sVhcXY17JuL-r6Hg@mail.gmail.com>
Date:   Thu, 6 Aug 2020 14:08:43 +0200
From:   Tomasz Figa <tfiga@...omium.org>
To:     Dafna Hirschfeld <dafna.hirschfeld@...labora.com>
Cc:     Laurent Pinchart <laurent.pinchart@...asonboard.com>,
        Helen Koike <helen.koike@...labora.com>,
        "open list:ARM/Rockchip SoC..." <linux-rockchip@...ts.infradead.org>,
        linux-devicetree <devicetree@...r.kernel.org>,
        Eddie Cai <eddie.cai.linux@...il.com>,
        Mauro Carvalho Chehab <mchehab@...nel.org>,
        Heiko Stübner <heiko@...ech.de>,
        Chen Jacob <jacob2.chen@...k-chips.com>,
        Jeffy <jeffy.chen@...k-chips.com>,
        钟以崇 <zyc@...k-chips.com>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Hans Verkuil <hans.verkuil@...co.com>,
        Sakari Ailus <sakari.ailus@...ux.intel.com>,
        kernel@...labora.com, Ezequiel Garcia <ezequiel@...labora.com>,
        Linux Media Mailing List <linux-media@...r.kernel.org>,
        "list@....net:IOMMU DRIVERS <iommu@...ts.linux-foundation.org>, Joerg
        Roedel <joro@...tes.org>," <linux-arm-kernel@...ts.infradead.org>,
        Shunqian Zheng <zhengsq@...k-chips.com>,
        Jacob Chen <cc@...k-chips.com>,
        Allon Huang <allon.huang@...k-chips.com>
Subject: Re: [PATCH v8 05/14] media: rkisp1: add Rockchip ISP1 subdev driver

On Wed, Aug 5, 2020 at 11:10 PM Dafna Hirschfeld
<dafna.hirschfeld@...labora.com> wrote:
>
> Hi
>
> On 22.07.20 17:24, Tomasz Figa wrote:
> > Hi Dafna,
> >
> > On Sat, Jul 11, 2020 at 01:04:31PM +0200, Dafna Hirschfeld wrote:
> >> Hi Laurent,
> >>
> >> On 16.08.19 02:13, Laurent Pinchart wrote:
> >>> Hello Helen,
> >>>
> >>> Thank you for the patch.
> >>>
> >>> On Tue, Jul 30, 2019 at 03:42:47PM -0300, Helen Koike wrote:
> > [snip]
> >>>> +static void rkisp1_isp_queue_event_sof(struct rkisp1_isp_subdev *isp)
> >>>> +{
> >>>> +  struct v4l2_event event = {
> >>>> +          .type = V4L2_EVENT_FRAME_SYNC,
> >>>> +          .u.frame_sync.frame_sequence =
> >>>> +                  atomic_inc_return(&isp->frm_sync_seq) - 1,
> >>>
> >>> I would move the increment to the caller, hiding it in this function is
> >>> error-prone (and if you look at the caller I'm pointing out one possible
> >>> error :-)).
> >>>
> >>> In general usage of frm_sync_seq through the driver seems to be very
> >>> race-prone. It's read in various IRQ handling functions, all coming from
> >>> the same IRQ, so that part is fine (and wouldn't require an atomic
> >>> variable), but when read from the buffer queue handlers I really get a
> >>> red light flashing in my head. I'll try to investigate more when
> >>> reviewing the next patches.
> >>
> >> I see that the only place were 'frame_sequence' is read outside of the irq
> >> handlers is in the capture in 'rkisp1_vb2_buf_queue':
> >>
> >>      /*
> >>           * If there's no next buffer assigned, queue this buffer directly
> >>           * as the next buffer, and update the memory interface.
> >>           */
> >>          if (cap->is_streaming && !cap->buf.next &&
> >>              atomic_read(&cap->rkisp1->isp.frame_sequence) == -1) {
> >>                  cap->buf.next = ispbuf;
> >>                  rkisp1_set_next_buf(cap);
> >>          } else {
> >>                  list_add_tail(&ispbuf->queue, &cap->buf.queue);
> >>          }
> >> This "if" condition seems very specific, a case where we already stream but v-start was not yet received.
> >> I think it is possible to remove the test 'atomic_read(&cap->rkisp1->isp.frame_sequence) == -1'
> >> from the above condition so that the next buffer is updated in case it is null not just before the first
> >> v-start signal.
> >>
> >
> > We don't have this special case in the Chrome OS code.
> >
> > I suppose it would make it possible to resume the capture 1 frame
> > earlier after a queue underrun, as otherwise the new buffer would be
> > only programmed after the next frame start interrupt and used for the
> > next-next frame.  However, it's racy, because programming of the buffer
> > addresses is not atomic and could end up with the hardware using few
> > plane addresses from the new buffer and few from the dummy buffer.
> >
> > Given that and also the fact that a queue underrun is a very special
> > case, where the system was already having problems catching up, I'd just
> > remove this special case.
> >
> > [snip]
> >>>> +void rkisp1_isp_isr(unsigned int isp_mis, struct rkisp1_device *dev)
> >>>> +{
> >>>> +  void __iomem *base = dev->base_addr;
> >>>> +  unsigned int isp_mis_tmp = 0;
> >>>
> >>> _tmp are never good names :-S
> >>>
> >>>> +  unsigned int isp_err = 0;
> >>>
> >>> Neither of these variable need to be initialised to 0.
> >>>
> >>>> +
> >>>> +  /* start edge of v_sync */
> >>>> +  if (isp_mis & CIF_ISP_V_START) {
> >>>> +          rkisp1_isp_queue_event_sof(&dev->isp_sdev);
> >>>
> >>> This will increment the frame sequence number. What if the interrupt is
> >>> slightly delayed and the next frame starts before we get a change to
> >>> copy the sequence number to the buffers (before they will complete
> >>> below) ?
> >>
> >> Do you mean that we get two sequental v-start signals and then the next
> >> frame-end signal in MI_MIS belongs to the first v-start signal of the two?
> >> How can this be solved? I wonder if any v-start signal has a later signal
> >> that correspond to the same frame so that we can follow it?
> >>
> >> Maybe we should have one counter that is incremented on v-start signal,
> >> and another counter that is incremented uppon some other signal?
> >>
> >
> > We're talking about a hard IRQ. I can't imagine the interrupt handler
> > being delayed for a time close to a full frame interval (~16ms for 60
> > fps) to trigger such scenario.
> >
> >>>
> >>>> +
> >>>> +          writel(CIF_ISP_V_START, base + CIF_ISP_ICR);
> >>>
> >>> Do you need to clear all interrupt bits individually, can't you write
> >>> isp_mis to CIF_ISP_ICR at the beginning of the function to clear them
> >>> all in one go ?
> >>>
> >>>> +          isp_mis_tmp = readl(base + CIF_ISP_MIS);
> >>>> +          if (isp_mis_tmp & CIF_ISP_V_START)
> >>>> +                  v4l2_err(&dev->v4l2_dev, "isp icr v_statr err: 0x%x\n",
> >>>> +                           isp_mis_tmp);
> >>>
> >>> This require some explanation. It looks like a naive way to protect
> >>> against something, but I think it could trigger under normal
> >>> circumstances if IRQ handling is delayed, and wouldn't do much anyway.
> >>> Same for the similar constructs below.
> >>>
> >>>> +  }
> >>>> +
> >>>> +  if ((isp_mis & CIF_ISP_PIC_SIZE_ERROR)) {
> >>>> +          /* Clear pic_size_error */
> >>>> +          writel(CIF_ISP_PIC_SIZE_ERROR, base + CIF_ISP_ICR);
> >>>> +          isp_err = readl(base + CIF_ISP_ERR);
> >>>> +          v4l2_err(&dev->v4l2_dev,
> >>>> +                   "CIF_ISP_PIC_SIZE_ERROR (0x%08x)", isp_err);
> >>>
> >>> What does this mean ?
> >>>
> >>>> +          writel(isp_err, base + CIF_ISP_ERR_CLR);
> >>>> +  } else if ((isp_mis & CIF_ISP_DATA_LOSS)) {
> >>>
> >>> Are CIF_ISP_PIC_SIZE_ERROR and CIF_ISP_DATA_LOSS mutually exclusive ?
> >>>
> >>>> +          /* Clear data_loss */
> >>>> +          writel(CIF_ISP_DATA_LOSS, base + CIF_ISP_ICR);
> >>>> +          v4l2_err(&dev->v4l2_dev, "CIF_ISP_DATA_LOSS\n");
> >>>> +          writel(CIF_ISP_DATA_LOSS, base + CIF_ISP_ICR);
> >>>> +  }
> >>>> +
> >>>> +  /* sampled input frame is complete */
> >>>> +  if (isp_mis & CIF_ISP_FRAME_IN) {
> >>>> +          writel(CIF_ISP_FRAME_IN, base + CIF_ISP_ICR);
> >>>> +          isp_mis_tmp = readl(base + CIF_ISP_MIS);
> >>>> +          if (isp_mis_tmp & CIF_ISP_FRAME_IN)
> >>>> +                  v4l2_err(&dev->v4l2_dev, "isp icr frame_in err: 0x%x\n",
> >>>> +                           isp_mis_tmp);
> >>>> +  }
> >>>> +
> >>>> +  /* frame was completely put out */
> >>>
> >>> "put out" ? :-) What's the difference between ISP_FRAME_IN and ISP_FRAME
> >>> ? The two comments could do with a bit of brush up, and I think the
> >>> ISP_FRAME_IN interrupt could be disabled as it doesn't perform any
> >>> action.
> >>
> >> Those two oneline comments are just copy-paste from the datasheet.
> >>
> >> ""
> >> 5 MIS_FRAME_IN sampled input frame is complete
> >> 1 MIS_FRAME frame was completely put out
> >> ""
> >>
> >> Unfrotunately, the datasheet does not add any further explanation about those signals.
> >>
> >>
> >
> > My loose recollection is that the former is signaled when then frame
> > is fully input to the ISP and the latter when the ISP completes
> > outputting the frame to the next block in the pipeline, but someone
> > would need to verify this, for example by printing timestamps for all
> > the various interrupts.
> >
> >>>
> >>>> +  if (isp_mis & CIF_ISP_FRAME) {
> >>>> +          u32 isp_ris = 0;
> >>>
> >>> No need to initialise this to 0.
> >>>
> >>>> +          /* Clear Frame In (ISP) */
> >>>> +          writel(CIF_ISP_FRAME, base + CIF_ISP_ICR);
> >>>> +          isp_mis_tmp = readl(base + CIF_ISP_MIS);
> >>>> +          if (isp_mis_tmp & CIF_ISP_FRAME)
> >>>> +                  v4l2_err(&dev->v4l2_dev,
> >>>> +                           "isp icr frame end err: 0x%x\n", isp_mis_tmp);
> >>>> +
> >>>> +          isp_ris = readl(base + CIF_ISP_RIS);
> >>>> +          if (isp_ris & (CIF_ISP_AWB_DONE | CIF_ISP_AFM_FIN |
> >>>> +                         CIF_ISP_EXP_END | CIF_ISP_HIST_MEASURE_RDY))
> >>>> +                  rkisp1_stats_isr(&dev->stats_vdev, isp_ris);
> >>>
> >>> Is there a guarantee that the statistics will be fully written out
> >>> before the video frame itself ? And doesn't this test if any of the
> >>> statistics is complete, not all of them ? I think the logic is wrong, it
> >>
> >> The datasheet does not add any explanation of what is expected to come first.
> >> Should we wait until all statistics measurements are done? In the struct
> >> sent to userspace there is a bitmaks for which of the statistics are read.
> >> I think that if only part of the statistics are ready, we can already send the once
> >> that are ready to userspace.
> >>
> >
> > If we look further into the code, rkisp1_stats_isr() checks the
> > interrupt status mask passed to it and reads out only the parameters
> > with indicated completion. The statistics metadata buffer format
> > includes a bit mask which tells the userspace which measurements are
> > available.
> >
> > However, I think I've spotted a bug there. At the beginning of
> > rkisp1_stats_isr(), all the 4 interrupt status bits are cleared,
> > regardless of the mask used later to decide which readouts need to be
> > done. This could mean that with an unfortunate timing, some measurements
> > would be lost. So at least the code should be fixed to only clear the
> > interrupts bits really handled.
>
> I'll fix that
>
> >
> > As for whether to send separate buffers for each measurement, I guess
> > it's not a bad thing to let the userspace access the ones available
> > earlier. Now I only don't recall why we decided to put all the
> > measurements into one metadata structure, rather than splitting the 4
> > into their own structures and buffer queues...
>
> Is it possible to have several queues to the same video node?
>
> >
> >>> seems it should be moved out of the CIF_ISP_FRAME test, to a test of its
> >>> own. It's hard to tell for sure without extra information though (for
> >>> instance why are the stats-related bits read from CIF_ISP_RIS, when
> >>> they seem to be documented as valid in CIF_ISP_ISR), but this should be
> >>> validated, and most probably fixed. Care should be taken to keep
> >>> synchronisation of sequence number between the different queues.
> >>
> >> I see that the capture buffers are done before incrementing the frame_sequence with
> >> the following explanation:
> >>
> >>      /*
> >>           * Call rkisp1_capture_isr() first to handle the frame that
> >>           * potentially completed using the current frame_sequence number before
> >>           * it is potentially incremented by rkisp1_isp_isr() in the vertical
> >>           * sync.
> >>           */
> >>
> >> I think reading the stats/params should also be done before calling rkisp1_capture_isr
> >> for the same reason. (so to match the correct frame_sequence)
> >
> > My recollection of the sequence of interrupts in this hardware is like
> > this:
> >
> > CIF_ISP_V_START (frame 0)
> >    CIF_ISP_FRAME_IN (frame 0)
> >      CIF_ISP_FRAME (frame 0)
> >        CIF_ISP_AWB_DONE
> >        CIF_ISP_AFM_FIN
> >        CIF_ISP_EXP_END
> >        CIF_ISP_HIST_MEASURE_RDY
> >        CIF_MI_FRAME*
> >        CIF_ISP_V_START (frame 1)
> >          CIF_ISP_FRAME_IN (frame 1)
> >            CIF_ISP_FRAME (frame 1)
> >              ...
> >
> > where the interrupts at the same indentation level can happen
> > independently of each other. Again, someone would have to verify this.
>
> I wrote this patch to print the interrupts and the time difference between interrupts:
> https://gitlab.collabora.com/dafna/linux/-/commit/9b9c5ddc2f06a6b87d2c1b210219f69de83296c5
>
> I got this output: http://ix.io/2tl8,
> there is a repeating pattern where only v-start interrupt is sent, indicated by the prints "isp mis 0x00000040" then about 23 milisec later are the other interrupts
> (FRAME_IN, FRAME, MI_FRAME* ) and about 10 milisec the v-start interrupt again.
>
> I am still not sure why the mi_frame interrupt should be handled first. If it happen for example that all the interrupts arrive at once, how can
> we know that the MI_FRAME interrupt relates to the previous v-start interrupt and not the current one?
> I think that for that we need a code that keep track of the previous interrupt.
>

Okay, thanks for double checking this. We are still missing the
CIF_ISP_FRAME_IN interrupt and the interrupts from the CIF_ISP_RIS
register in the picture, though.

Regardless of that, looking at how this hardware seems to work, I'd
imagine that what we need to do is:

1) On an MI interrupt:
  a) Program MI shadow registers for frame N+2
  b) Signal completion of frame N to the user

2) On an RKISP1_CIF_ISP_FRAME interrupt:
  a) Program ISP parameter shadow registers for frame N+2

3) On one of the 4 interrupts from the CIF_ISP_RIS register:
  a) Read out the corresponding ISP statistics
  b.1) Return the partial statistics buffer to the userspace, OR
  b.2) If all the 4 statistics have been read, return the full
statistics buffer to the userspace.

All the 3 cases don't care about each other. Every time each interrupt
is received, the driver does the work for the sequential frame. We
basically assume that no interrupts are ever lost.

If we were to detect interrupt lost cases, we would have to either
look at some hardware frame counter (if one exists) or use some
heuristics based on the system timer and configured frame interval. I
don't think the extra complexity is worth it, because if the system
suffers from interrupt latency high enough to cause lost interrupts
(i.e. close to the shortest frame intervals supported by this ISP),
then it is going to have other problems as well and the underlying
cause of the latency should be fixed instead.

One more note about the current implementation is that it seems to
assume that CIF_ISP_RIS interrupts are already done when the
RKISP1_CIF_ISP_FRAME interrupt arrives. If that's how the hardware
works indeed, then it's a good thing to do, because it decreases the
number of signalled interrupts and thus the total interrupt handling
overhead.

Another thing to note is that when both MP and SP are enabled, I
recall seeing their interrupts sometimes arriving separately.

Now one thing that needs to be done to handle the 3 cases correctly is
to make sure they match the input parameters, output capture buffers
and statistics to the right frame. Since the MIs are asynchronous from
the ISP and each other, I suppose we could make the code less fragile
by making the MIs have their own frame counters, so they don't have to
rely on the order of interrupt handling. Perhaps that would be enough
to address your concern?

Best regards,
Tomasz

Powered by blists - more mailing lists