P.S. Apologies in advance for the explicit TO list, it seemed adding people who've
touched the dma mapping code (for ARC atleast), would respond sooner ;-)
The question is does dma_map_single() imply a single region (possibly > PAGE_SIZE)
or does it imply PAGE_SIZE. Documentation/DMA-API* is not explicit about one or
the other and the code below seems to point to "a" struct page, although it could
also mean multiple pages, specially if the pages are contiguous, say as those
returned by alloc_pages(with order > 0)
static inline dma_addr_t dma_map_single_attrs(dev, size, dir, attrs)
{
addr = ops->map_page(dev,
virt_to_page(ptr), <--- this is one struct page
offset_in_page(ptr), size, dir, attrs);
}
ARC dma_map_single() currently only handles cache coherency for only one struct
page or PAGE_SIZE worth of memory, and we have a customer who seem to assume that
it handles a region.
Looking at other arches dma mapping backend, it is not clear either what the
semantics are.
Thx,
-Vineet
I would expect that it would support any contiguous range in
the kernel mapping (e.g. no vmalloc and friends). But it's not
documented anywhere, and if no in kernel users makes use of that
fact at the moment it might be better to document a page size
limitation and add asserts to enforce it.
Hi Christoph,
On 07/16/2017 11:42 PM, Christoph Hellwig wrote:
> I would expect that it would support any contiguous range in
> the kernel mapping (e.g. no vmalloc and friends). But it's not
> documented anywhere, and if no in kernel users makes use of that
> fact at the moment it might be better to document a page size
> limitation and add asserts to enforce it.
My first thought was indeed to add a BUG_ON for @size > PAGE_SIZE (also accounting
for offset etc), but I have a feeling this will cause too many breakages. So
perhaps it would be better to add the fact to Documentation that it can handle any
physically contiguous range.
-Vineet
On Mon, 2017-07-17 at 09:06 -0700, Vineet Gupta wrote:
> Hi Christoph,
>
> On 07/16/2017 11:42 PM, Christoph Hellwig wrote:
> >
> > I would expect that it would support any contiguous range in
> > the kernel mapping (e.g. no vmalloc and friends). But it's not
> > documented anywhere, and if no in kernel users makes use of that
> > fact at the moment it might be better to document a page size
> > limitation and add asserts to enforce it.
>
> My first thought was indeed to add a BUG_ON for @size > PAGE_SIZE
> (also accounting for offset etc), but I have a feeling this will
> cause too many breakages. So perhaps it would be better to add the
> fact to Documentation that it can handle any physically contiguous
> range.
Actually, that's not historically right. dma_map_single() was
originally designed to be called on any region that was kmalloc'd
meaning it was capable of mapping physically contiguous > PAGE_SIZE
regions.
For years (decades?) we've been eliminating the specialised
dma_map_single() calls in favour of dma_map_sg, so it's possible there
may not be any large region consumers anymore, so it *may* be safe to
enforce a PAGE_SIZE limit, but not without auditing the remaining
callers.
James