> On Feb 7, 2021, at 12:31 AM, Zhou Wang <wangzhou1@xxxxxxxxxxxxx> wrote: > > SVA(share virtual address) offers a way for device to share process virtual > address space safely, which makes more convenient for user space device > driver coding. However, IO page faults may happen when doing DMA > operations. As the latency of IO page fault is relatively big, DMA > performance will be affected severely when there are IO page faults. > From a long term view, DMA performance will be not stable. > > In high-performance I/O cases, accelerators might want to perform > I/O on a memory without IO page faults which can result in dramatically > increased latency. Current memory related APIs could not achieve this > requirement, e.g. mlock can only avoid memory to swap to backup device, > page migration can still trigger IO page fault. > > Various drivers working under traditional non-SVA mode are using > their own specific ioctl to do pin. Such ioctl can be seen in v4l2, > gpu, infiniband, media, vfio, etc. Drivers are usually doing dma > mapping while doing pin. > > But, in SVA mode, pin could be a common need which isn't necessarily > bound with any drivers, and neither is dma mapping needed by drivers > since devices are using the virtual address of CPU. Thus, It is better > to introduce a new common syscall for it. > > This patch leverages the design of userfaultfd and adds mempinfd for pin > to avoid messing up mm_struct. A fd will be got by mempinfd, then user > space can do pin/unpin pages by ioctls of this fd, all pinned pages under > one file will be unpinned in file release process. Like pin page cases in > other places, can_do_mlock is used to check permission and input > parameters. Can you document what the syscall does? Userfaultfd is an fd because one program controls another. Is mempinfd like this?
