On 18/09/17 05:22, Huacai Chen wrote: > In non-coherent DMA mode, kernel uses cache flushing operations to > maintain I/O coherency, so the dmapool objects should be aligned to > ARCH_DMA_MINALIGN. Otherwise, it will cause data corruption, at least > on MIPS: > > Step 1, dma_map_single > Step 2, cache_invalidate (no writeback) > Step 3, dma_from_device > Step 4, dma_unmap_single This is a massive red warning flag for the whole series, because DMA pools don't work like that. At best, this will do nothing, and at worst it is papering over egregious bugs elsewhere. Streaming mappings of coherent allocations means completely broken code. > If a DMA buffer and a kernel structure share a same cache line, and if > the kernel structure has dirty data, cache_invalidate (no writeback) > will cause data lost. DMA pools are backed by coherent allocations, and those should already be at *page* granularity, so this doubly cannot happen for correct code. More generally, the whole point of having the DMA APIs is that drivers and subsystems should not have to be aware of details like hardware coherency. Besides, cache line sharing that could pose a correctness issue for non-hardware-coherent systems could still be a performance issue in the presence of hardware coherency (due to unnecessary line migration), so there's still an argument for not treating them differently. Robin. > Cc: stable@xxxxxxxxxxxxxxx > Signed-off-by: Huacai Chen <chenhc@xxxxxxxxxx> > --- > mm/dmapool.c | 3 +++ > 1 file changed, 3 insertions(+) > > diff --git a/mm/dmapool.c b/mm/dmapool.c > index 4d90a64..6263905 100644 > --- a/mm/dmapool.c > +++ b/mm/dmapool.c > @@ -140,6 +140,9 @@ struct dma_pool *dma_pool_create(const char *name, struct device *dev, > else if (align & (align - 1)) > return NULL; > > + if (!device_is_coherent(dev)) > + align = max_t(size_t, align, dma_get_cache_alignment()); > + > if (size == 0) > return NULL; > else if (size < 4) >