Le vendredi 30 juillet 2021 à 16:11 +0300, Laurent Pinchart a écrit :
> Hi Dafna,
>
> On Fri, Jul 30, 2021 at 02:35:20PM +0200, Dafna Hirschfeld wrote:
> > On 30.07.21 02:19, Laurent Pinchart wrote:
> > > The vimc driver is used for testing purpose, and some test use cases
> > > involve sharing buffers with a consumer device. Consumers often require
> > > DMA contiguous memory, which vimc doesn't currently support. This leads
> > > in the best case to usage of bounce buffers, which is very slow, and in
> > > the worst case in a complete failure.
> > >
> > > Add support for the dma-contig allocator in vimc to support those use
> > > cases properly. The allocator is selected through a new "allocator"
> > > module parameter, which defaults to vmalloc.
> > >
> > > Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@xxxxxxxxxxxxxxxx>
> > > ---
> > > drivers/media/test-drivers/vimc/vimc-capture.c | 9 +++++++--
> > > drivers/media/test-drivers/vimc/vimc-common.h | 2 ++
> > > drivers/media/test-drivers/vimc/vimc-core.c | 10 ++++++++++
> > > 3 files changed, 19 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/media/test-drivers/vimc/vimc-capture.c b/drivers/media/test-drivers/vimc/vimc-capture.c
> > > index 5e9fd902cd37..92b69a6529fb 100644
> > > --- a/drivers/media/test-drivers/vimc/vimc-capture.c
> > > +++ b/drivers/media/test-drivers/vimc/vimc-capture.c
> > > @@ -7,6 +7,7 @@
> > >
> > > #include <media/v4l2-ioctl.h>
> > > #include <media/videobuf2-core.h>
> > > +#include <media/videobuf2-dma-contig.h>
> > > #include <media/videobuf2-vmalloc.h>
> > >
> > > #include "vimc-common.h"
> > > @@ -423,14 +424,18 @@ static struct vimc_ent_device *vimc_cap_add(struct vimc_device *vimc,
> > > /* Initialize the vb2 queue */
> > > q = &vcap->queue;
> > > q->type = V4L2_BUF_TYPE_VIDEO_CAPTURE;
> > > - q->io_modes = VB2_MMAP | VB2_DMABUF | VB2_USERPTR;
> > > + q->io_modes = VB2_MMAP | VB2_DMABUF;
> >
> > maybe to be on the safe side VB2_DMABUF should be set only if vimc_allocator==1 ?
>
> Why so ? vb2-vmalloc can import dma-bufs.
Indeed, should be safe with both allocator, the CMA one will validate that the
pages are contiguous and fail synchronously as expected for CMA. The known
issues are mostly for reading importers (consumers), for writing importer
(producers) it is likely still a bit buggy on cache management.
> > > if (ret) {
> > > diff --git a/drivers/media/test-drivers/vimc/vimc-common.h b/drivers/media/test-drivers/vimc/vimc-common.h
> > > index a289434e75ba..b77939123501 100644
> > > --- a/drivers/media/test-drivers/vimc/vimc-common.h
> > > +++ b/drivers/media/test-drivers/vimc/vimc-common.h
> > > @@ -35,6 +35,8 @@
> > >
> > > #define VIMC_PIX_FMT_MAX_CODES 8
> > >
> > > +extern unsigned int vimc_allocator;
> > > +
> > > /**
> > > * vimc_colorimetry_clamp - Adjust colorimetry parameters
> > > *
> > > diff --git a/drivers/media/test-drivers/vimc/vimc-core.c b/drivers/media/test-drivers/vimc/vimc-core.c
> > > index 4b0ae6f51d76..7badcecb7aed 100644
> > > --- a/drivers/media/test-drivers/vimc/vimc-core.c
> > > +++ b/drivers/media/test-drivers/vimc/vimc-core.c
> > > @@ -5,6 +5,7 @@
> > > * Copyright (C) 2015-2017 Helen Koike <helen.fornazier@xxxxxxxxx>
> > > */
> > >
> > > +#include <linux/dma-mapping.h>
> > > #include <linux/font.h>
> > > #include <linux/init.h>
> > > #include <linux/module.h>
> > > @@ -15,6 +16,12 @@
> > >
> > > #include "vimc-common.h"
> > >
> > > +unsigned int vimc_allocator;
> > > +module_param_named(allocator, vimc_allocator, uint, 0444);
> > > +MODULE_PARM_DESC(allocator, " memory allocator selection, default is 0.\n"
> > > + "\t\t 0 == vmalloc\n"
> > > + "\t\t 1 == dma-contig");
> > > +
> >
> > There is a section 'Module options' in vimc.rst. So a doc should be added there.
>
> OK, I'll update that.
>
> > > #define VIMC_MDEV_MODEL_NAME "VIMC MDEV"
> > >
> > > #define VIMC_ENT_LINK(src, srcpad, sink, sinkpad, link_flags) { \
> > > @@ -278,6 +285,9 @@ static int vimc_probe(struct platform_device *pdev)
> > >
> > > tpg_set_font(font->data);
> > >
> > > + if (vimc_allocator == 1)
> > > + dma_coerce_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
> > > +
> > > vimc = kzalloc(sizeof(*vimc), GFP_KERNEL);
> > > if (!vimc)
> > > return -ENOMEM;
> > >
>
