With CONFIG_HIGHMEM enabled, __dma_sync_page() unconditionally goes through __dma_sync_page_highmem(), which disables interrupts and kmaps the buffer one page at a time. This adds needless per-page overhead to every streaming DMA sync on such kernels.
Use the per-page kmap path only when the buffer reaches into highmem and do a single __dma_sync() on the linear mapping otherwise. Signed-off-by: Marek Szyprowski <[email protected]> --- arch/powerpc/mm/dma-noncoherent.c | 12 ++++++++---- 1 file changed, 8 insertions(+), 4 deletions(-) diff --git a/arch/powerpc/mm/dma-noncoherent.c b/arch/powerpc/mm/dma-noncoherent.c index 30260b5d146d..3ce7dfc213d9 100644 --- a/arch/powerpc/mm/dma-noncoherent.c +++ b/arch/powerpc/mm/dma-noncoherent.c @@ -96,12 +96,16 @@ static void __dma_sync_page(phys_addr_t paddr, size_t size, int dir) struct page *page = pfn_to_page(paddr >> PAGE_SHIFT); unsigned offset = paddr & ~PAGE_MASK; + if (!size) + return; + #ifdef CONFIG_HIGHMEM - __dma_sync_page_highmem(page, offset, size, dir); -#else - unsigned long start = (unsigned long)page_address(page) + offset; - __dma_sync((void *)start, size, dir); + if (PageHighMem(pfn_to_page(PHYS_PFN(paddr + size - 1)))) { + __dma_sync_page_highmem(page, offset, size, dir); + return; + } #endif + __dma_sync(page_address(page) + offset, size, dir); } void arch_sync_dma_for_device(phys_addr_t paddr, size_t size, -- 2.43.0
