Messages in this thread | | | Date | Fri, 21 Nov 2014 11:36:18 +0000 | From | Catalin Marinas <> | Subject | Re: For the problem when using swiotlb |
| |
On Fri, Nov 21, 2014 at 11:26:45AM +0000, Arnd Bergmann wrote: > On Friday 21 November 2014 11:06:10 Catalin Marinas wrote: > > On Wed, Nov 19, 2014 at 03:56:42PM +0000, Arnd Bergmann wrote: > > > On Wednesday 19 November 2014 15:46:35 Catalin Marinas wrote: > > > > Going back to original topic, the dma_supported() function on arm64 > > > > calls swiotlb_dma_supported() which actually checks whether the swiotlb > > > > bounce buffer is within the dma mask. This transparent bouncing (unlike > > > > arm32 where it needs to be explicit) is not always optimal, though > > > > required for 32-bit only devices on a 64-bit system. The problem is when > > > > the driver is 64-bit capable but forgets to call > > > > dma_set_mask_and_coherent() (that's not the only question I got about > > > > running out of swiotlb buffers). > > > > > > I think it would be nice to warn once per device that starts using the > > > swiotlb. Really all 32-bit DMA masters should have a proper IOMMU > > > attached. > > > > It would be nice to have a dev_warn_once(). > > > > I think it makes sense on arm64 to avoid swiotlb bounce buffers for > > coherent allocations altogether. The __dma_alloc_coherent() function > > already checks coherent_dma_mask and sets GFP_DMA accordingly. If we > > have a device that cannot even cope with a 32-bit ZONE_DMA, we should > > just not support DMA at all on it (without an IOMMU). The arm32 > > __dma_supported() has a similar check. > > If we ever encounter this case, we may have to add a smaller ZONE_DMA > and use ZONE_DMA32 for the normal dma allocations.
Traditionally on x86 I think ZONE_DMA was for ISA and ZONE_DMA32 had to cover the 32-bit physical address space. On arm64 we don't expect ISA, so we only use ZONE_DMA (which is 4G, similar to IA-64, sparc). We had ZONE_DMA32 originally but it broke swiotlb which assumes ZONE_DMA for its bounce buffer.
> > Swiotlb is still required for the streaming DMA since we get bouncing > > for pages allocated outside the driver control (e.g. VFS layer which > > doesn't care about GFP_DMA), hoping a 16M bounce buffer would be enough. > > > > Ding seems to imply that CMA fixes the problem, which means that the > > issue is indeed coherent allocations. > > I wonder what's going on here, since swiotlb_alloc_coherent() actually > tries a regular __get_free_pages(flags, order) first, and when ZONE_DMA > is set here, it just work without using the pool.
As long as coherent_dma_mask is sufficient for ZONE_DMA. I have no idea what this mask is set to in Ding's case (but I've seen the problem previously with an out of tree driver where coherent_dma_mask was some random number; so better reporting here would help).
There could be another case where dma_pfn_offset is required but let's wait for some more info from Ding (ZONE_DMA is 32-bit from the start of RAM which could be 40-bit like on Seattle, so basically such devices would need to set dma_pfn_offset).
-- Catalin
| |