| Message ID | 20220720215319.87839-1-peterx@redhat.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | mm/mprotect: Fix soft-dirty check in can_change_pte_writable() | expand |
On Wed, Jul 20, 2022 at 05:53:19PM -0400, Peter Xu wrote: > void clear_refs_write(void) > { > int fd = open("/proc/self/clear_refs", O_RDWR); > > assert(fd >= 0); > write(fd, "4", 2); > close(fd); > } > Oops, git dropped some of the lines starting with "#".. Let me retry, sorry. > bool dirty = pagemap_read_vaddr(fd, page) & PM_SOFT_DIRTY; \ > if (dirty != expect) { \ > printf("ERROR: %s, soft-dirty=%d (expect: %d)\n", str, dirty, expect); \ > exit(-1); \ > } \ > } while (0)
diff --git a/mm/mprotect.c b/mm/mprotect.c index 0420c3ed936c..804807ab14e6 100644 --- a/mm/mprotect.c +++ b/mm/mprotect.c @@ -48,8 +48,11 @@ static inline bool can_change_pte_writable(struct vm_area_struct *vma, if (pte_protnone(pte) || !pte_dirty(pte)) return false; - /* Do we need write faults for softdirty tracking? */ - if ((vma->vm_flags & VM_SOFTDIRTY) && !pte_soft_dirty(pte)) + /* + * Do we need write faults for softdirty tracking? Note, + * soft-dirty is enabled when !VM_SOFTDIRTY. + */ + if (!(vma->vm_flags & VM_SOFTDIRTY) && !pte_soft_dirty(pte)) return false; /* Do we need write faults for uffd-wp tracking? */
The check wanted to make sure when soft-dirty tracking is enabled we won't grant write bit by accident, as a page fault is needed for dirty tracking. The intention is correct but we didn't check it right because VM_SOFTDIRTY set actually means soft-dirty tracking disabled. Fix it. It wasn't a bug for a long time because we used to only optimize the write bit settings in change_pte_range() for page caches, and since we've got a higher level check in vma_wants_writenotify(), we will never set the bit MM_CP_TRY_CHANGE_WRITABLE for soft-dirty enabled page caches, hence even if we checked with the wrong value of VM_SOFTDIRTY in change_pte_range() it'll just be an no-op. Functionally it was still correct, even if cpu cycles wasted. However after the recent work of anonymous page optimization on exclusive pages we'll start to make it wrong because anonymous page does not require the check in vma_wants_writenotify() hence it'll suffer from the wrong check here in can_change_pte_writable(). We can easily verify this with any exclusive anonymous page, like program below: =======8<====== unsigned int psize; char *page; uint64_t pagemap_read_vaddr(int fd, void *vaddr) { uint64_t value; int ret; ret = pread(fd, &value, sizeof(uint64_t), ((uint64_t)vaddr >> 12) * sizeof(uint64_t)); assert(ret == sizeof(uint64_t)); return value; } void clear_refs_write(void) { int fd = open("/proc/self/clear_refs", O_RDWR); assert(fd >= 0); write(fd, "4", 2); close(fd); } bool dirty = pagemap_read_vaddr(fd, page) & PM_SOFT_DIRTY; \ if (dirty != expect) { \ printf("ERROR: %s, soft-dirty=%d (expect: %d)\n", str, dirty, expect); \ exit(-1); \ } \ } while (0) int main(void) { int fd = open("/proc/self/pagemap", O_RDONLY); assert(fd >= 0); psize = getpagesize(); page = mmap(NULL, psize, PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0); assert(page != MAP_FAILED); *page = 1; check_soft_dirty("Just faulted in page", 1); clear_refs_write(); check_soft_dirty("Clear_refs written", 0); mprotect(page, psize, PROT_READ); check_soft_dirty("Marked RO", 0); mprotect(page, psize, PROT_READ|PROT_WRITE); check_soft_dirty("Marked RW", 0); *page = 2; check_soft_dirty("Wrote page again", 1); munmap(page, psize); close(fd); printf("Test passed.\n"); return 0; } =======8<====== So even if commit 64fe24a3e05e kept the old behavior and didn't attempt to change the behavior here, the bug will only be able to be triggered after commit 64fe24a3e05e because only anonymous page will suffer from it. Fixes: 64fe24a3e05e ("mm/mprotect: try avoiding write faults for exclusive anonymous pages when changing protection") Signed-off-by: Peter Xu <peterx@redhat.com> --- mm/mprotect.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)