On 2020-08-19 13:49, Tomasz Figa wrote:
On Wed, Aug 19, 2020 at 1:51 PM Robin Murphy <robin.murphy@xxxxxxx> wrote:
Hi Tomasz,
On 2020-08-19 12:16, Tomasz Figa wrote:
Hi Christoph,
On Wed, Aug 19, 2020 at 8:56 AM Christoph Hellwig <hch@xxxxxx> wrote:
The V4L2-FLAG-MEMORY-NON-CONSISTENT flag is entirely unused,
Could you explain what makes you think it's unused? It's a feature of
the UAPI generally supported by the videobuf2 framework and relied on
by Chromium OS to get any kind of reasonable performance when
accessing V4L2 buffers in the userspace.
and causes
weird gymanstics with the DMA_ATTR_NON_CONSISTENT flag, which is
unimplemented except on PARISC and some MIPS configs, and about to be
removed.
It is implemented by the generic DMA mapping layer [1], which is used
by a number of architectures including ARM64 and supposed to be used
by new architectures going forward.
AFAICS all that V4L2_FLAG_MEMORY_NON_CONSISTENT does is end up
controling whether DMA_ATTR_NON_CONSISTENT is added to vb2_queue::dma_attrs.
Please can you point to where DMA_ATTR_NON_CONSISTENT does anything at
all on arm64?
With the default config it doesn't, but with
CONFIG_DMA_NONCOHERENT_CACHE_SYNC enabled it makes dma_pgprot() keep
the pgprot value as is, without enforcing coherence attributes.
How active are the PA-RISC and MIPS ports of Chromium OS?
Hacking CONFIG_DMA_NONCOHERENT_CACHE_SYNC into an architecture that
doesn't provide dma_cache_sync() is wrong, since at worst it may break
other drivers. If downstream is wildly misusing an API then so be it,
but it's hardly a strong basis for an upstream argument.
Also, I posit that videobuf2 is not actually relying on
DMA_ATTR_NON_CONSISTENT anyway, since it's clearly not using it properly:
"By using this API, you are guaranteeing to the platform
that you have all the correct and necessary sync points for this memory
in the driver should it choose to return non-consistent memory."
$ git grep dma_cache_sync drivers/media
$
AFAIK dma_cache_sync() isn't the only way to perform the cache
synchronization. The earlier patch series that I reviewed relied on
dma_get_sgtable() and then dma_sync_sg_*() (which existed in the
vb2-dc since forever [1]). However, it looks like with the final code
the sgtable isn't acquired and the synchronization isn't happening, so
you have a point.
Using the streaming sync calls on coherent allocations has also always
been wrong per the API, regardless of the bodies of code that have
happened to get away with it for so long.
FWIW, I asked back in time what the plan is for non-coherent
allocations and it seemed like DMA_ATTR_NON_CONSISTENT and
dma_sync_*() was supposed to be the right thing to go with. [2] The
same thread also explains why dma_alloc_pages() isn't suitable for the
users of dma_alloc_attrs() and DMA_ATTR_NON_CONSISTENT.
AFAICS even back then Christoph was implying getting rid of
NON_CONSISTENT and *replacing* it with something streaming-API-based -
i.e. this series - not encouraging mixing the existing APIs. It doesn't
seem impossible to implement a remapping version of this new
dma_alloc_pages() for IOMMU-backed ops if it's really warranted
(although at that point it seems like "non-coherent" vb2-dc starts to
have significant conceptual overlap with vb2-sg).
Robin.
