On Fri, May 5, 2023, at 07:47, Guo Ren wrote: > On Mon, Mar 27, 2023 at 8:15 PM Arnd Bergmann <arnd@xxxxxxxxxx> wrote: >> >> riscv also invalidates the caches before the transfer, which does >> not appear to serve any purpose. > Yes, we can't guarantee the CPU pre-load cache lines randomly during > dma working. > > But I've two purposes to keep invalidates before dma transfer: > - We clearly tell the CPU these cache lines are invalid. The caching > algorithm would use these invalid slots first instead of replacing > valid ones. > - Invalidating is very cheap. Actually, flush and clean have the same > performance in our machine. The main purpose of the series was to get consistent behavior on all machines, so I really don't want a custom optimization on one architecture. You make a good point about cacheline reuse after invalidation, but if we do that, I'd suggest doing this across all architectures. > So, how about: > > diff --git a/arch/riscv/mm/dma-noncoherent.c b/arch/riscv/mm/dma-noncoherent.c > index d919efab6eba..2c52fbc15064 100644 > --- a/arch/riscv/mm/dma-noncoherent.c > +++ b/arch/riscv/mm/dma-noncoherent.c > @@ -22,8 +22,6 @@ void arch_sync_dma_for_device(phys_addr_t paddr, size_t size, > ALT_CMO_OP(clean, vaddr, size, riscv_cbom_block_size); > break; > case DMA_FROM_DEVICE: > - ALT_CMO_OP(clean, vaddr, size, riscv_cbom_block_size); > - break; > case DMA_BIDIRECTIONAL: > ALT_CMO_OP(flush, vaddr, size, riscv_cbom_block_size); > break; This is something we can consider. Unfortunately, this is something that no architecture (except pa-risc, which has other problems) does at the moment, so we'd probably need to have a proper debate about this. We already have two conflicting ways to handle DMA_FROM_DEVICE, either invalidate/invalidate, or clean/invalidate. I can see that flush/invalidate may be a sensible option as well, but I'd want to have that discussion after the series is complete, so we can come to a generic solution that has the same documented behavior across all architectures. In particular, if we end up moving arm64 and riscv back to the traditional invalidate/invalidate for DMA_FROM_DEVICE and document that driver must not rely on buffers getting cleaned before a partial DMA_FROM_DEVICE, the question between clean or flush becomes moot as well. > @@ -42,7 +40,7 @@ void arch_sync_dma_for_cpu(phys_addr_t paddr, size_t size, > break; > case DMA_FROM_DEVICE: > case DMA_BIDIRECTIONAL: > /* I'm not sure all drivers have guaranteed cacheline > alignment. If not, this inval would cause problems */ > - ALT_CMO_OP(flush, vaddr, size, riscv_cbom_block_size); > + ALT_CMO_OP(inval, vaddr, size, riscv_cbom_block_size); > break; This is my original patch, and I would not mix it with the other change. The problem with non-aligned DMA_BIDIRECTIONAL buffers in is that both flush and inval would be wrong if you get simultaneous writes from device and cpu to the same cache line, so there is no way to win this. Using inval instead of flush would at least work if the CPU data in the cacheline is read-only from the CPU, so that seems better than something that is always wrong. The documented API is that sharing the cache line is not allowed at all, so anything that would observe a difference between the two is also a bug. One idea that we have considered already is that we could overwrite the unused bits of the cacheline with poison values and/or mark them as invalid using KASAN for debugging purposes, to find drivers that already violate this. Arnd