Hi Inki,
On 2017-11-13 02:24, Inki Dae wrote:
Hi Marek,
2017년 11월 10일 16:35에 Marek Szyprowski 이(가) 쓴 글:
Dear Inki,
On 2017-11-10 04:04, Inki Dae wrote:
2017년 11월 01일 01:28에 Marek Szyprowski 이(가) 쓴 글:
When no IOMMU is available, all GEM buffers allocated by Exynos DRM driver
are contiguous, because of the underlying dma_alloc_attrs() function
provides only such buffers. In such case it makes no sense to keep
What if user disabled CMA support? In this case, it guarantees also to allocate physically contiguous memory?
I think it depends on CMA support so wouldn't be true.
dma_alloc_attrs() always guarantees the contiguity of the allocated memory
in DMA address space. When NOIOMMU is available, this mean that the allocated
buffer is contiguous in the physical memory. When CMA is disabled,
dma_alloc_attrs() uses alloc_pages() allocator. The drawback of alloc_pages()
is that it easily fails if physical memory is fragmented and it cannot
allocate memory larger than MAX_ORDER (4MiB in case of ARM32bit).
You are right. Without IOMMU suppot alloc_pages always tryies to allocate contiguous memory,
order = get_order(size);
page = alloc_pages(..., order);
if (!page)
return NULL;
...
Right
Real problem I think is that user don't know whether the gem buffer allocated with CONTIG or NONCONTIG flag can be used as a SCANOUT buffer.
So user can request a page flip with NONCONTIG buffer to kernel which doesn't support IOMMU.
And another is that user may want to use NONCONTIG buffer for another purpose, not scanout.
So if we enforce on using CONTIG buffer on kernel without IOMMU support, then it wouldn't be really what user intended.
When IOMMU support is disabled, ANY buffer allocated with dma_alloc_attrs()
will be contiguous, so I see no point propagating incorrect flag for it.
The only way to create a NONCONTIG buffer with IOMMU disabled is to import
it from other driver and this path is already handled correctly.
My idea is to provide a new flag - i.e., EXYNOS_BO_SCANOUT - which can allocate a buffer with a different allocation type - CONTIG or NONCONTIG - according to IOMMU support.
And any page flip request with NONCONTIG buffer to kernel without IOMMU support should fail and it has to return a error with a proper error message.
I don't think that we need it. With the proposed patch the same userspace will
work fine in both cases IOMMU enabled and disabled, even if it allocate GEM
with NONCONTIG flag. We can assume that CONTIG is a special case of NONCONTIG
then.
The problem is really not whether user space will work fine or not but is that this may be not what user space wanted.
I.e., user space expects the allocated buffer is NONCONTIG buffer because user space requested NONCONTIG buffer but kernel internally, this is changed to CONTIG buffer.
What's the problem when kernel allocated contiguous buffer but user
requested non-contiguous? Contiguous is a subclass of non-contiguous
in general. There is no driver nor scenario which will not work because
of such change (any code for handing n-segments should be fine with only
1 segment). In some conditions, by luck, kernel might allocate
a contiguous buffer even with IOMMU enabled. When we know that the
buffer is contiguous, then flag it as such.
So you could provide some information - maybe warning message?? - to user space the buffer type is changed to CONTIG buffer due to NO IOMMU support.
I don't think this needs a warning. I just think that when we know that
the allocated buffer IS contiguous there is no point flagging it as
non-contiguous and fail a few calls later although the driver allocated
the contiguous buffer in fact. That's all.
Regarding the buffer type Exynos DRM has, I'm not sure that providing CONTIG and NONCONTIG buffer types to user space is reasonable because I think such buffer types - physically contiguous or non contiguous - should be transparent to user space.
And I looked into other ARM DRM drivers - omap and msm provides SCANOUT buffer type, tegra and rockchip have no such buffer types (contiguous or non-contiguous).
This is also a thing I felt while working on kernel in collaborating with Platform guys - we made user spaces to care about the memory allocation way.
I wonder why you say SCANOUT type is incorrect flag.
I'm not against adding SCANOUT flag. It is a good idea in general,
especially if the driver can simply select which set of flags will
be best for such use case on the given hardware (IOMMU is practically
always used on newer Exynos SoCs, IOMMU has enough TLB cache and
handling of non-contiguous buffers is cheap there).
I would only separate it from fixing the current code.
Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland
_______________________________________________
dri-devel mailing list
dri-devel@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/dri-devel
