On 12/07/2017 03:07, Fam Zheng wrote: > On Tue, 07/11 12:28, Paolo Bonzini wrote: >> On 11/07/2017 12:05, Stefan Hajnoczi wrote: >>> On Mon, Jul 10, 2017 at 05:08:56PM +0200, Paolo Bonzini wrote: >>>> On 10/07/2017 17:07, Stefan Hajnoczi wrote: >>>>> On Wed, Jul 05, 2017 at 09:36:32PM +0800, Fam Zheng wrote: >>>>>> Allow block driver to map and unmap a buffer for later I/O, as a >>>>>> performance >>>>>> hint. >>>>> The name blk_dma_map() is confusing since other "dma" APIs like >>>>> dma_addr_t and dma_blk_io() deal with guest physical addresses instead >>>>> of host addresses. They are about DMA to/from guest RAM. >>>>> >>>>> Have you considered hiding this cached mapping in block/nvme.c so that >>>>> it isn't exposed? block/nvme.c could keep the last buffer mapped and >>>>> callers would get the performance benefit without a new blk_dma_map() >>>>> API. >>>> >>>> One buffer is enough for qemu-img bench, but not for more complex cases >>>> (e.g. fio). >>> >>> I don't see any other blk_dma_map() callers. >> >> Indeed, the fio plugin is not part of this series, but it also used >> blk_dma_map. Without it, performance is awful. > > How many buffers does fio use, typically? If it's not too many, block/nvme.c > can > cache the last N buffers. I'm with Stefan that hiding the mapping logic from > block layer callers makes a nicer API, especially such that qemu-img is much > easier to maintain good performance across subcommmands.
It depends on the queue depth. I think the API addition is necessary, otherwise we wouldn't have added the RAMBlockNotifier which is a layering violation that does the same thing (create permanent HVA->IOVA mappings). In fact, the RAMBlockNotifier could be moved out of nvme.c and made to use blk_dma_map/unmap, though I'm not proposing to do it now. I don't think qemu-img convert and dd are impacted by IOMMU map/unmap as heavily as bench, because they operate with queue depth 1. But adding map/unmap there would not be hard. Paolo
