Re: [PATCH 07/55] media: imx8-isi: Stop abusing of min_buffers_needed field

[Laurent Pinchart]

On Wed, Nov 29, 2023 at 05:39:25PM +0900, Tomasz Figa wrote:
> On Wed, Nov 29, 2023 at 5:28 PM Benjamin Gaignard wrote:
> > Le 29/11/2023 à 05:17, Tomasz Figa a écrit :
> > > On Tue, Nov 28, 2023 at 7:26 PM Benjamin Gaignard wrote:
> > >> Le 28/11/2023 à 10:35, Tomasz Figa a écrit :
> > >>> On Tue, Nov 28, 2023 at 6:31 PM Benjamin Gaignard wrote:
> > >>>> Le 27/11/2023 à 18:07, Laurent Pinchart a écrit :
> > >>>>> On Mon, Nov 27, 2023 at 05:54:06PM +0100, Benjamin Gaignard wrote:
> > >>>>>> 'min_buffers_needed' is suppose to be used to indicate the number
> > >>>>>> of buffers needed by DMA engine to start streaming.
> > >>>>>> imx8-isi driver doesn't use DMA engine and just want to specify
> > >>>>>
> > >>>>> What do you mean, "doesn't use DMA engine" ? The ISI surely has DMA
> > >>>>> engines :-)
> > >>>>
> > >>>> I have done assumption on drivers given if they use or dma_* functions.
> > >>>
> > >>> I suspect the use of vb2_dma_sg_plane_desc() and
> > >>> vb2_dma_contig_plane_dma_addr() may be more correlated to whether
> > >>> there is a DMA involved or not. Usually V4L2 drivers don't really have
> > >>> to deal with the DMA API explicitly, because the vb2 framework handles
> > >>> most of the work.
> > >>
> > >> Unfortunately isn't not true either, for example verisilicon driver use
> > >> these function but doesn't need DMA engine.
> > >
> > > That sounds weird. Why would a driver that doesn't have a DMA engine
> > > need to obtain a scatterlist or the DMA address of the buffer?
> >
> > Just because the hardware needs the physical address of the buffer to access it.
> 
> Right, and the part of the hardware that accesses the memory is called
> a DMA engine.
> 
> > >> I haven't found yet a 100% criteria to decide if driver use or not DMA engine
> > >> so I plan to fix case by case given maintainers remarks.
> > >
> > > Yeah, there probably wouldn't be a way that would give one a 100%
> > > certainty, although I'd still insist that the two functions I
> > > mentioned should be close to that. Of course a driver can use those
> > > functions for some queues, while other queues would be pure software
> > > queues, e.g. for some metadata - a simple grep is not enough. Is that
> > > perhaps the case for the verisilicon driver?
> >
> > Verisilicon hardware block doesn't have IOMMU so it needs the physical
> > addresses of all the buffers it use (input buffer, reference frame buffers, etc...).
> > No DMA engine involved here it is just how the hardware is working.
> 
> I think we need to clarify what you mean by DMA engine. If it's
> basically a standalone hardware block that does the DMA for another
> hardware block, i.e. such as the standalone DMA engines under
> drivers/dma, then I'd like to ask what the relation is between using
> an external DMA engine and min_buffers_needed.

Yes, there seems to have been some confusion, DMA engine != dmaengine.h.

> > Expect functions like dma_release_channel() or being in PCI directory,
> > I don't have found any magical way to know if a driver needs a minimum number of buffers before start streaming.
> > I can only read the code and do assumptions for the other cases.
> > I hope maintainers, like Laurent or you, will answer to this question for each driver.
> 
> In theory that could work too, so hopefully we can achieve that. Some
> drivers may not have very active maintainers... And other maintainers
> who never worked with such drivers are as suited to read the code and
> guess the expected state as you. That said, let's make sure that
> everyone involved does their best, without pushing the task around.

We can rely on individual drivers maintainers for review, but the
initial work needs to make a reasonable effort to analyze the drivers
and find the right value for min_buffers_needed and
min_reqbufs_allocation.

> > >>>> I have considers that all PCI drivers are using DMA engine and
> > >>>> I don't know the design for each drivers so I hope to get this information
> > >>>> from maintainers and fix that in v2.
> > >>>> If imx8-isi driver needs a minimum number of buffers before start streaming
> > >>>> I will do a v2 and use min_dma_buffers_needed instead.
> > >>>>
> > >>>>>> the minimum number of buffers to allocate when calling VIDIOC_REQBUFS.
> > >>>>>> That 'min_reqbufs_allocation' field purpose so use it.
> > >>>>>>
> > >>>>>> Signed-off-by: Benjamin Gaignard <benjamin.gaignard@xxxxxxxxxxxxx>
> > >>>>>> CC: Laurent Pinchart <laurent.pinchart@xxxxxxxxxxxxxxxx>
> > >>>>>> CC: Mauro Carvalho Chehab <mchehab@xxxxxxxxxx>
> > >>>>>> CC: Shawn Guo <shawnguo@xxxxxxxxxx>
> > >>>>>> CC: Sascha Hauer <s.hauer@xxxxxxxxxxxxxx>
> > >>>>>> CC: Pengutronix Kernel Team <kernel@xxxxxxxxxxxxxx>
> > >>>>>> CC: Fabio Estevam <festevam@xxxxxxxxx>
> > >>>>>> CC: NXP Linux Team <linux-imx@xxxxxxx>
> > >>>>>> ---
> > >>>>>>     drivers/media/platform/nxp/imx8-isi/imx8-isi-video.c | 2 +-
> > >>>>>>     1 file changed, 1 insertion(+), 1 deletion(-)
> > >>>>>>
> > >>>>>> diff --git a/drivers/media/platform/nxp/imx8-isi/imx8-isi-video.c b/drivers/media/platform/nxp/imx8-isi/imx8-isi-video.c
> > >>>>>> index 49bca2b01cc6..81673ff9084b 100644
> > >>>>>> --- a/drivers/media/platform/nxp/imx8-isi/imx8-isi-video.c
> > >>>>>> +++ b/drivers/media/platform/nxp/imx8-isi/imx8-isi-video.c
> > >>>>>> @@ -1453,7 +1453,7 @@ int mxc_isi_video_register(struct mxc_isi_pipe *pipe,
> > >>>>>>        q->mem_ops = &vb2_dma_contig_memops;
> > >>>>>>        q->buf_struct_size = sizeof(struct mxc_isi_buffer);
> > >>>>>>        q->timestamp_flags = V4L2_BUF_FLAG_TIMESTAMP_MONOTONIC;
> > >>>>>> -    q->min_buffers_needed = 2;
> > >>>>>> +    q->min_reqbufs_allocation = 2;
> > >>>>>>        q->lock = &video->lock;
> > >>>>>>        q->dev = pipe->isi->dev;
> > >>>>>>

-- 
Regards,

Laurent Pinchart