Messages in this thread | | | Date | Mon, 3 Feb 2014 21:14:37 +0900 | Subject | Re: [PATCH v2 1/5] x86: make dma_alloc_coherent() return zeroed memory if CMA is enabled | From | Akinobu Mita <> |
| |
2014-01-29 Akinobu Mita <akinobu.mita@gmail.com>: > 2014-01-28 Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>: >> On Mon, Jan 27, 2014 at 02:54:47PM +0100, Marek Szyprowski wrote: >>> Hello, >>> >>> On 2014-01-14 15:13, Akinobu Mita wrote: >>> >Calling dma_alloc_coherent() with __GFP_ZERO must return zeroed memory. >>> > >>> >But when the contiguous memory allocator (CMA) is enabled on x86 and >>> >the memory region is allocated by dma_alloc_from_contiguous(), it >>> >doesn't return zeroed memory. Because dma_generic_alloc_coherent() >>> >forgot to fill the memory region with zero if it was allocated by >>> >dma_alloc_from_contiguous() >>> >>> I just wonder how it will work with high mem? I've didn't check the x86 >>> dma mapping code yet, but page_address() works only for pages, which comes >>> from low memory. In other patches you have added an option to place CMA >>> area anywhere in the memory. Is the x86 pci dma code ready for the case >>> when cma area is put into high mem and direct mappings are not available? >> >> Yes and no. The swiotbl_bounce does have the code to take that into account. >> But that is it - nothing else does - so I think you would run in the >> possiblity of 'page_address' not providing an correct virtual address. > > Thanks for spotting the issue. I haven't much tested on x86_32. > I'll go through it and try to find the solution.
I have confirmed that locating CMA on highmem range with a 'cma=size@start-end' kernel parameter by this patch set caused the issue on x86_32.
This can be fixed by limiting CMA area upto max_low_pfn to prevent from locating it on highmem at arch/x86/kernel/setup.c:setup_arch()
- dma_contiguous_reserve(0); + dma_contiguous_reserve(max_low_pfn << PAGE_SHIFT);
I'm going to inlcude this change in this patch set.
| |