[PATCH 3/9] dma-heap: Provide accessors so that in-kernel drivers can allocate dmabufs from specific heaps

[Nicolas Dufresne]

Le mardi 12 septembre 2023 à 08:47 +0000, Yong Wu (吴勇) a écrit :
> On Mon, 2023-09-11 at 12:12 -0400, Nicolas Dufresne wrote:
> >  	 
> > External email : Please do not click links or open attachments until
> > you have verified the sender or the content.
> >  Hi,
> > 
> > Le lundi 11 septembre 2023 à 10:30 +0800, Yong Wu a écrit :
> > > From: John Stultz <jstultz at google.com>
> > > 
> > > This allows drivers who don't want to create their own
> > > DMA-BUF exporter to be able to allocate DMA-BUFs directly
> > > from existing DMA-BUF Heaps.
> > > 
> > > There is some concern that the premise of DMA-BUF heaps is
> > > that userland knows better about what type of heap memory
> > > is needed for a pipeline, so it would likely be best for
> > > drivers to import and fill DMA-BUFs allocated by userland
> > > instead of allocating one themselves, but this is still
> > > up for debate.
> > 
> > 
> > Would be nice for the reviewers to provide the information about the
> > user of
> > this new in-kernel API. I noticed it because I was CCed, but
> > strangely it didn't
> > make it to the mailing list yet and its not clear in the cover what
> > this is used
> > with. 
> > 
> > I can explain in my words though, my read is that this is used to
> > allocate both
> > user visible and driver internal memory segments in MTK VCODEC
> > driver.
> > 
> > I'm somewhat concerned that DMABuf objects are used to abstract
> > secure memory
> > allocation from tee. For framebuffers that are going to be exported
> > and shared
> > its probably fair use, but it seems that internal shared memory and
> > codec
> > specific reference buffers also endup with a dmabuf fd (often called
> > a secure fd
> > in the v4l2 patchset) for data that is not being shared, and requires
> > a 1:1
> > mapping to a tee handle anyway. Is that the design we'd like to
> > follow ? 
> 
> Yes. basically this is right.
> 
> > Can't
> > we directly allocate from the tee, adding needed helper to make this
> > as simple
> > as allocating from a HEAP ?
> 
> If this happens, the memory will always be inside TEE. Here we create a
> new _CMA heap, it will cma_alloc/free dynamically. Reserve it before
> SVP start, and release to kernel after SVP done.

Ok, I see the benefit of having a common driver then. It would add to the
complexity, but having a driver for the tee allocator and v4l2/heaps would be
another option?

>   
> Secondly. the v4l2/drm has the mature driver control flow, like
> drm_gem_prime_import_dev that always use dma_buf ops. So we can use the
> current flow as much as possible without having to re-plan a flow in
> the TEE.

>From what I've read from Yunfei series, this is only partially true for V4L2.
The vb2 queue MMAP feature have dmabuf exportation as optional, but its not a
problem to always back it up with a dmabuf object. But for internal SHM buffers
used for firmware communication, I've never seen any driver use a DMABuf.

Same applies for primary decode buffers when frame buffer compression or post-
processing it used (or reconstruction buffer in encoders), these are not user
visible and are usually not DMABuf.

> 
> > 
> > Nicolas
> > 
> > > 
> > > Signed-off-by: John Stultz <jstultz at google.com>
> > > Signed-off-by: T.J. Mercier <tjmercier at google.com>
> > > Signed-off-by: Yong Wu <yong.wu at mediatek.com>
> > > [Yong: Fix the checkpatch alignment warning]
> > > ---
> > >  drivers/dma-buf/dma-heap.c | 60 ++++++++++++++++++++++++++++----
> > ------
> > >  include/linux/dma-heap.h   | 25 ++++++++++++++++
> > >  2 files changed, 69 insertions(+), 16 deletions(-)
> > > 
> [snip]