On Thu, Sep 24, 2020 at 01:55:31PM -0400, Peter Xu wrote: > On Thu, Sep 24, 2020 at 01:51:52PM -0300, Jason Gunthorpe wrote: > > > Regarding the solution here, I think we can also cover read-only fast-gup too > > > in the future - IIUC what we need to do is to make it pte_protnone() instead of > > > pte_wrprotect(), then in the fault handler we should identify this special > > > pte_protnone() against numa balancing (change_prot_numa()). I think it should > > > work fine too, iiuc, because I don't think we should migrate a page at all if > > > it's pinned for any reason... > > [1] > > > > > With your COW breaking patch the read only fast-gup should break the > > COW because of the write protect, just like for the write side. Not > > seeing why we need to do something more? > > Consider this sequence of a parent process managed to fork() a child: > > buf = malloc(); > // RDONLY gup > pin_user_pages(buf, !WRITE); > // pte of buf duplicated on both sides > fork(); > mprotect(buf, WRITE); > *buf = 1; > // buf page replaced as cow triggered > > Currently when fork() we'll happily share a pinned read-only page with the > child by copying the pte directly. Why? This series prevents that, the page will be maybe_dma_pinned, so fork() will copy it. > As a summary: imho the important thing is we should not allow any kind of > sharing of any dma page, even it's pinned for read. Any sharing that results in COW. MAP_SHARED is fine, for instance My feeling for READ when FOLL_PIN is used GUP_fast will go to the slow path any time it sees a read-only page. The slow path will determine if it is read-only because it could be COW'd or read-only for some other reason Jason
