On 5/30/19 9:11 AM, Catalin Marinas wrote: > On Wed, May 29, 2019 at 01:16:37PM -0600, Khalid Aziz wrote: >> mmap() can return the same tagged address but I am uneasy about kernel >> pre-coloring the pages. Database can mmap 100's of GB of memory. That is >> lot of work being offloaded to the kernel to pre-color the page even if >> done in batches as pages are faulted in. > > For anonymous mmap() for example, the kernel would have to zero the > faulted in pages anyway. We can handle the colouring at the same time in > clear_user_page() (as I said below, we have to clear the colour anyway > from previous uses, so it's simply extending this to support something > other than tag/colour 0 by default with no additional overhead). > On sparc M7, clear_user_page() ends up in M7clear_user_page defined in arch/sparc/lib/M7memset.S. M7 code use Block Init Store (BIS) to clear the page. BIS on M7 clears the memory tags as well and no separate instructions are needed to clear the tags. As a result when kernel clears a page before returning it to user, the page is not only zeroed out, its tags are also cleared to 0. >>> Since we already need such loop in the kernel, we might as well allow >>> user space to require a certain colour. This comes in handy for large >>> malloc() and another advantage is that the C library won't be stuck >>> trying to paint the whole range (think GB). >> >> If kernel is going to pre-color all pages in a vma, we will need to >> store the default tag in the vma. It will add more time to page fault >> handling code. On sparc M7, kernel will need to execute additional 128 >> stxa instructions to set the tags on a normal page. > > As I said, since the user can retrieve an old colour using ldxa, the > kernel should perform this operation anyway on any newly allocated page > (unless you clear the existing colour on page freeing).> Tags are not cleared on sparc on freeing. They get cleared when the page is allocated again. >>>> We can try to store tags for an entire region in vma but that is >>>> expensive, plus on sparc tags are set in userspace with no >>>> participation from kernel and now we need a way for userspace to >>>> communicate the tags to kernel. >>> >>> We can't support finer granularity through the mmap() syscall and, as >>> you said, the vma is not the right thing to store the individual tags. >>> With the above extension to mmap(), we'd have to store a colour per vma >>> and prevent merging if different colours (we could as well use the >>> pkeys mechanism we already have in the kernel but use a colour per vma >>> instead of a key). >> >> Since tags can change on any part of mmap region on sparc at any time >> without kernel being involved, I am not sure I see much reason for >> kernel to enforce any tag related restrictions. > > It's not enforcing a tag, more like the default colour for a faulted in > page. Anyway, if sparc is going with default 0/untagged, that's fine as > well. We may add this mmap() option to arm64 only. > >>>> From sparc point of view, making kernel responsible for assigning tags >>>> to a page on page fault is full of pitfalls. >>> >>> This could be just some arm64-specific but if you plan to deploy it more >>> generically for sparc (at the C library level), you may find this >>> useful. >> >> Common semantics from app developer point of view will be very useful to >> maintain. If arm64 says mmap with MAP_FIXED and a tagged address will >> return a pre-colored page, I would rather have it be the same on any >> architecture. Is there a use case that justifies kernel doing this extra >> work? > > So if a database program is doing an anonymous mmap(PROT_TBI) of 100GB, > IIUC for sparc the faulted-in pages will have random colours (on 64-byte > granularity). Ignoring the information leak from prior uses of such > pages, it would be the responsibility of the db program to issue the > stxa. On arm64, since we also want to do this via malloc(), any large > allocation would require all pages to be faulted in so that malloc() can > set the write colour before being handed over to the user. That's what > we want to avoid and the user is free to repaint the memory as it likes. > On sparc, any newly allocated page is cleared along with any old tags on it. Since clearing tag happens automatically when page is cleared on sparc, clear_user_page() will need to execute additional stxa instructions to set a new tag. It is doable. In a way it is done already if page is being pre-colored with tag 0 always ;) Where would the pre-defined tag be stored - as part of address stored in vm_start or a new field in vm_area_struct? -- Khalid _______________________________________________ amd-gfx mailing list amd-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/amd-gfx
