On Mon, Oct 24, 2022 at 8:30 PM Kirill A . Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx> wrote: > > On Fri, Oct 21, 2022 at 04:18:14PM +0000, Sean Christopherson wrote: > > On Fri, Oct 21, 2022, Chao Peng wrote: > > > > > > > > In the context of userspace inaccessible memfd, what would be a > > > > suggested way to enforce NUMA memory policy for physical memory > > > > allocation? mbind[1] won't work here in absence of virtual address > > > > range. > > > > > > How about set_mempolicy(): > > > https://www.man7.org/linux/man-pages/man2/set_mempolicy.2.html > > > > Andy Lutomirski brought this up in an off-list discussion way back when the whole > > private-fd thing was first being proposed. > > > > : The current Linux NUMA APIs (mbind, move_pages) work on virtual addresses. If > > : we want to support them for TDX private memory, we either need TDX private > > : memory to have an HVA or we need file-based equivalents. Arguably we should add > > : fmove_pages and fbind syscalls anyway, since the current API is quite awkward > > : even for tools like numactl. > > Yeah, we definitely have gaps in API wrt NUMA, but I don't think it be > addressed in the initial submission. > > BTW, it is not regression comparing to old KVM slots, if the memory is > backed by memfd or other file: > > MBIND(2) > The specified policy will be ignored for any MAP_SHARED mappings in the > specified memory range. Rather the pages will be allocated according to > the memory policy of the thread that caused the page to be allocated. > Again, this may not be the thread that called mbind(). > > It is not clear how to define fbind(2) semantics, considering that multiple > processes may compete for the same region of page cache. > > Should it be per-inode or per-fd? Or maybe per-range in inode/fd? > David's analysis on mempolicy with shmem seems to be right. set_policy on virtual address range does seem to change the shared policy for the inode irrespective of the mapping type. Maybe having a way to set numa policy per-range in the inode would be at par with what we can do today via mbind on virtual address ranges. > fmove_pages(2) should be relatively straight forward, since it is > best-effort and does not guarantee that the page will note be moved > somewhare else just after return from the syscall. > > -- > Kiryl Shutsemau / Kirill A. Shutemov