Re: [PATCH v2] media: imx7-media-csi: Add support for fast-tracking queued buffers

[Laurent Pinchart <laurent.pinchart@ideasonboard.com>]

Hi Paul,

Thank you for the patch.

On Wed, Sep 07, 2022 at 08:47:37PM +0900, Paul Elder wrote:
> The CSI hardware compatible with this driver handles buffers using a
> ping-pong mechanism with two sets of destination addresses. Normally,
> when an interrupt comes in to signal the completion of one buffer, say
> FB0, it assigns the next buffer in the queue to the next FB0, and the
> hardware starts to capture into FB1 in the meantime.

Could you replace FB0 and FB1 with FB1 and FB2 respectively, to match
the naming of the registers ?

> In a buffer underrun situation, in the above example without loss of
> generality, if a new buffer is queued before the interrupt for FB0 comes
> in, we can program the buffer into FB1 (which is programmed with a dummy
> buffer, as there is a buffer underrun).
> 
> This of course races with the interrupt that signals FB0 completion, as
> once that interrupt comes in, we are no longer guaranteed that the
> programming of FB1 was in time and must assume it was too late. This
> race is resolved partly by locking the programming of FB1. If it came
> after the interrupt for FB0, then the variable that is used to determine
> which FB to program would have been swapped by the interrupt handler.
> 
> This alone isn't sufficient, however, because the interrupt could still
> be generated (thus the hardware starts capturing into the other fb)
> while the fast-tracking routine has the irq lock. Thus, after
> programming the fb register to fast-track the buffer, the isr also must
> be checked to confirm that an interrupt didn't come in the meantime. If
> it has, we must assume that programming the register for the
> fast-tracked buffer was not in time, and queue the buffer normally.
> 
> Signed-off-by: Paul Elder <paul.elder@ideasonboard.com>
> Acked-by: Rui Miguel Silva <rmfrfs@gmail.com>
> 
> ---
> Changes in v2:
> - fix the potential race condition where the interrupt comes in while
>   the fast tracking routine has the irqlock
> - change return value from int to bool
> ---
>  drivers/staging/media/imx/imx7-media-csi.c | 63 ++++++++++++++++++++++
>  1 file changed, 63 insertions(+)
> 
> diff --git a/drivers/staging/media/imx/imx7-media-csi.c b/drivers/staging/media/imx/imx7-media-csi.c
> index a0553c24cce4..0ebef44a7627 100644
> --- a/drivers/staging/media/imx/imx7-media-csi.c
> +++ b/drivers/staging/media/imx/imx7-media-csi.c
> @@ -1296,12 +1296,75 @@ static int imx7_csi_video_buf_prepare(struct vb2_buffer *vb)
>  	return 0;
>  }
>  
> +static bool imx7_csi_fast_track_buffer(struct imx7_csi *csi,
> +				       struct imx7_csi_vb2_buffer *buf)
> +{
> +	unsigned long flags;
> +	dma_addr_t phys;
> +	int buf_num;
> +	u32 isr;
> +
> +	if (!csi->is_streaming)
> +		return false;
> +
> +	phys = vb2_dma_contig_plane_dma_addr(&buf->vbuf.vb2_buf, 0);
> +
> +	/*
> +	 * buf_num holds the fb id of the most recently (*not* the next
> +	 * anticipated) triggered interrupt. Without loss of generality, if
> +	 * buf_num is 0 and we get to this section before the irq for fb2, the

s/fb2/FB2/ to match hardware registers and the commit message ?

> +	 * buffer that we are fast-tracking into fb1 should be programmed in
> +	 * time to be captured into. If the irq for fb2 already happened, then
> +	 * buf_num would be 1, and we would fast-track the buffer into fb2
> +	 * instead. This guarantees that we won't try to fast-track into fb1
> +	 * and race against the start-of-capture into fb1.
> +	 *
> +	 * We only fast-track the buffer if the currently programmed buffer is
> +	 * a dummy buffer. We can check the active_vb2_buf instead as it is
> +	 * always modified along with programming the fb[1,2] registers via the
> +	 * lock (besides setup and cleanup).
> +	 */

I think this needs to be updated, it still indicates we handle the race
just by checking buf_num. How about the following ?

	/*
	 * buf_num holds the framebuffer ID of the most recently (*not* the next
	 * anticipated) triggered interrupt. Without loss of generality, if
	 * buf_num is 0, the hardware is capturing to FB2. If FB1 has been
	 * programmed with a dummy buffer (as indicated by active_vb2_buf[0]
	 * being NULL), then we can fast-track the new buffer by programming its
	 * address in FB1 before the hardware completes FB2, instead of adding
	 * it to the buffer queue and incurring a delay of one additional frame.
	 *
	 * The irqlock prevents races with the interrupt handler that queues the
	 * next buffer and updates buf_num, but we can still race with the
	 * hardware if we program the buffer in FB1 just after the hardware
	 * completes FB2 and switches to FB1 and before we notice the buf_num
	 * change. The fast-tracked buffer would then be ignored by the hardware
	 * while the driver would think it has successfully been processed.
	 *
	 * To avoid this problem, if we can't avoid the race, we can detect that
	 * we have lost it by checking, after programming the buffer in FB1, if
	 * the interrupt flag indicated completion of FB2 has been raised. If
	 * that is not the case, fast-tracking succeeded, and we can update
	 * active_vb2_buf[0]. Otherwise, we may or may not have lost the race
	 * (as the interrupt flag may have been raised just after programming
	 * FB1 and before we read the interrupt status register), and we need to
	 * assume the worst case of a race loss and queue the buffer through the
	 * slow path.
	 */

Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>

If you're fine with these changes there's no need to submit a v3, I'll
update the comment and the commit message locally.

> +
> +	spin_lock_irqsave(&csi->irqlock, flags);
> +
> +	buf_num = csi->buf_num;
> +	if (csi->active_vb2_buf[buf_num]) {
> +		spin_unlock_irqrestore(&csi->irqlock, flags);
> +		return false;
> +	}
> +
> +	imx7_csi_update_buf(csi, phys, buf_num);
> +
> +	isr = imx7_csi_reg_read(csi, CSI_CSISR);
> +	/*
> +	 * The interrupt for the /other/ fb just came (the isr hasn't run yet
> +	 * though, because we have the lock here); we can't be sure we've
> +	 * programmed buf_num fb in time, so queue the buffer to the buffer
> +	 * queue normally. No need to undo writing the fb register, since we
> +	 * won't return it as active_vb2_buf is NULL, so it's okay to
> +	 * potentially write it to both fb1 and fb2; only the one where it was
> +	 * queued normally will be returned.
> +	 */
> +	if (isr & (buf_num ? BIT_DMA_TSF_DONE_FB1 : BIT_DMA_TSF_DONE_FB2)) {
> +		spin_unlock_irqrestore(&csi->irqlock, flags);
> +		return false;
> +	}
> +
> +	csi->active_vb2_buf[buf_num] = buf;
> +
> +	spin_unlock_irqrestore(&csi->irqlock, flags);
> +	return true;
> +}
> +
>  static void imx7_csi_video_buf_queue(struct vb2_buffer *vb)
>  {
>  	struct imx7_csi *csi = vb2_get_drv_priv(vb->vb2_queue);
>  	struct imx7_csi_vb2_buffer *buf = to_imx7_csi_vb2_buffer(vb);
>  	unsigned long flags;
>  
> +	if (imx7_csi_fast_track_buffer(csi, buf))
> +		return;
> +
>  	spin_lock_irqsave(&csi->q_lock, flags);
>  
>  	list_add_tail(&buf->list, &csi->ready_q);

-- 
Regards,

Laurent Pinchart