On Thu, Sep 10, 2020 at 10:35:38AM -0700, Linus Torvalds wrote: > On Thu, Sep 10, 2020 at 2:40 AM Alexander Gordeev > <agordeev@xxxxxxxxxxxxx> wrote: > > > > It is only gup_fast case that exposes the issue. It hits because > > pointers to stack copies are passed to gup_pXd_range iterators, not > > pointers to real page tables itself. > > Can we possibly change fast-gup to not do the stack copies? > > I'd actually rather do something like that, than the "addr_end" thing. > As you say, none of the other page table walking code does what the > GUP code does, and I don't think it's required. As I understand it, the requirement is because fast-gup walks without the page table spinlock, or mmap_sem held so it must READ_ONCE the *pXX. It then checks that it is a valid page table pointer, then calls pXX_offset(). The arch implementation of pXX_offset() derefs again the passed pXX pointer. So it defeats the READ_ONCE and the 2nd load could observe something that is no longer a page table pointer and crash. Passing it the address of the stack value is a way to force pXX_offset() to use the READ_ONCE result which has already been tested to be a page table pointer. Other page walking code that holds the mmap_sem tends to use pmd_trans_unstable() which solves this problem by injecting a barrier. The load hidden in pte_offset() after a pmd_trans_unstable() can't be re-ordered and will only see a page table entry under the mmap_sem. However, I think that logic would have been much clearer following the GUP model of READ_ONCE vs extra reads and a hidden barrier. At least it took me a long time to work it out :( I also think there are real bugs here where places are reading *pXX multiple times without locking the page table. One was found recently in the wild in the huge tlb code IIRC. The mm/pagewalk.c has these missing READ_ONCE bugs too. So.. To change away from the stack option I think we'd have to pass the READ_ONCE value to pXX_offset() as an extra argument instead of it derefing the pointer internally. Jason
