The LS1046A features a cache-coherent interconnect and the drivers configure the hardware appropriately, e.g. setting the FMan PRAM_MODE_GLOBAL bit, so the written Ethernet Controllers snoop caches. Yet, we use the standard arm64 cache maintenance routines when the MMU is enabled and thus risk memory corruption if CPU prefetches receive buffers in the time window between dma_map_single() cleaning them to Point-of-Coherency and dma_unmap_single() invalidating them[1]. To properly solve this issue, we need to consult the newly added per-device dma coherent attribute to decide whether to do manual cache maintenance. [1]: https://lore.kernel.org/all/a5d6cc26-cd23-7c31-f56e-f6d535ea39b0@xxxxxxx/ Signed-off-by: Ahmad Fatoum <a.fatoum@xxxxxxxxxxxxxx> --- drivers/dma/map.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/drivers/dma/map.c b/drivers/dma/map.c index 114c0f7db3bd..be0ee258cc59 100644 --- a/drivers/dma/map.c +++ b/drivers/dma/map.c @@ -25,7 +25,8 @@ dma_addr_t dma_map_single(struct device *dev, void *ptr, size_t size, { unsigned long addr = (unsigned long)ptr; - dma_sync_single_for_device(addr, size, dir); + if (dev_is_dma_coherent(dev) <= 0) + dma_sync_single_for_device(addr, size, dir); return cpu_to_dma(dev, ptr); } @@ -35,5 +36,6 @@ void dma_unmap_single(struct device *dev, dma_addr_t dma_addr, size_t size, { unsigned long addr = (unsigned long)dma_to_cpu(dev, dma_addr); - dma_sync_single_for_cpu(addr, size, dir); + if (dev_is_dma_coherent(dev) <= 0) + dma_sync_single_for_cpu(addr, size, dir); } -- 2.30.2
