On Wed, Nov 13, 2024 at 5:46 AM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote: > > On Mon, Nov 11, 2024 at 04:36:52PM +0800, Jason Wang wrote: > > On Mon, Nov 11, 2024 at 3:30 PM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote: > > > > > > On Mon, Nov 11, 2024 at 10:55:38AM +0800, Jason Wang wrote: > > > > There's no need to sync DMA for CPU on mapping errors. So this patch > > > > skips the CPU sync in the error handling path of DMA mapping. > > > > > > > > Signed-off-by: Jason Wang <jasowang@xxxxxxxxxx> > > > > > > DMA sync is idempotent. > > > Extra work for slow path. Why do we bother? > > > > dma_map_sg() did this, since current virtio hack sg mappings to per > > page mapping, we lose such optimization. > > > > Actually the path is not necessarily slowpath in some setups like > > swiotlb or VDUSE. > > > > Thanks > > I don't get how it's not a slowpath. Example? I'm not sure how to define slowpath but a possible example is when we reach the swiotlb limitation, in this case. vring_map_one_sg() can fail easily. Thanks > > -- > MST > >
