> > > In the original process of dma_map/unmap pages for VFIO-devices, to > > make sure the pages are contiguous, we have to check them one by one. > > As a result, dma_map/unmap could spend a long time. > > Using the hugetlb pages, we can avoid this problem. > > All pages in hugetlb pages are contiguous.And the hugetlb page should > > not be split.So we can delete the for loops. > > I know nothing about VFIO, but I'm surprised that you're paying such attention > to PageHuge hugetlbfs pages, rather than to PageCompound > pages: which would also include Transparent Huge Pages, of the traditional > anonymous kind, or the huge tmpfs kind, or the more general (not necessarily > pmd-sized) kind that Matthew Wilcox is currently working on. > > It's true that hugetlbfs is peculiar enough that whatever you write for it may > need some tweaks to cover the THP case too, or vice versa; but wouldn't your > patch be a lot better for covering all cases? > > You mention above that "the hugetlb page should not be split": > perhaps you have been worried that a THP could be split beneath you? > That used to be a possibility some years ago, but nowadays a THP cannot be > split while anyone is pinning it with an extra reference. > > Hugh > Thanks for your suggestions. You mention that a THP cannot be split while anyone is pinning it. Do you mean the check of can_split_huge_page()(in split_huge_page_to_list())? When we want to pin pages, vfio_pin_pages_remote() always gets a normal page first. In this case, a THP cannot be split because of the increased reference. Is this right? Maybe I can optimize for hugetlb pages as a first step, and then cover all compound pages.
