Thursday 19 August 2010 14:16:21 Marin Mitov napisał(a):
On Thursday, August 19, 2010 02:39:47 pm Guennadi Liakhovetski wrote:
No, I don't think you should go to the next power of 2 - that's too crude. Try rounding your buffer size to the page size, that should suffice.
Guennadi, If you have a look at how a device reserved memory is next allocated to a driver with drivers/base/dma-coherent.c::dma_alloc_from_coherent(), then than you may find my conclusion on a power of 2 as true:
int dma_alloc_from_coherent(struct device *dev, ssize_t size, dma_addr_t *dma_handle, void **ret) { ... int order = get_order(size); ... pageno = bitmap_find_free_region(mem->bitmap, mem->size, order); ... }
Allocated coherent memory is always a power of 2.
Marin, For ARM, this seems true as long as allocated with the above from a device assigned pool, but not true for a (pre)allocation from a generic system RAM. See arch/arm/mm/dma-mapping.c::__dma_alloc_buffer(), where it looks like extra pages are freed:
static struct page *__dma_alloc_buffer(struct device *dev, size_t size, gfp_t gfp) { unsigned long order = get_order(size); ... page = alloc_pages(gfp, order); ... split_page(page, order); for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++) __free_page(p); ... }
Thanks, Janusz