| Message ID | 20221003155630.469263-2-peterx@redhat.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | mm/hugetlb: Fix selftest failures with write check | expand |
On 10/03/22 11:56, Peter Xu wrote: > After the recent rework patchset of hugetlb locking on pmd sharing, > kselftest for userfaultfd sometimes fails on hugetlb private tests with > unexpected write fault checks. > > It turns out there's nothing wrong within the locking series regarding this > matter, but it could have changed the timing of threads so it can trigger > an old bug. > > The real bug is when we call hugetlb_no_page() we're not with the pgtable > lock. It means we're reading the pte values lockless. It's perfectly fine > in most cases because before we do normal page allocations we'll take the > lock and check pte_same() again. However before that, there are actually > two paths on userfaultfd missing/minor handling that may directly move on > with the fault process without checking the pte values. > > It means for these two paths we may be generating an uffd message based on > an unstable pte, while an unstable pte can legally be anything as long as > the modifier holds the pgtable lock. > > One example, which is also what happened in the failing kselftest and > caused the test failure, is that for private mappings CoW can happen on one > page. CoW requires pte being cleared before being replaced with a new page > for TLB coherency, but then there can be a race condition: > > thread 1 thread 2 > -------- -------- > > hugetlb_fault hugetlb_fault > private pte RO > hugetlb_wp > pgtable_lock() > huge_ptep_clear_flush > pte=NULL > hugetlb_no_page > generate uffd missing event > even if page existed!! > set_huge_pte_at > pgtable_unlock() Thanks for working on this Peter! I agree with this patch, but I suspect the above race is not possible. Why? In both cases, we take the hugetlb fault mutex when processing a huegtlb page fault. This means only one thread can execute the fault code for a specific mapping/index at a time. This is why I was so confused, and may remain a bit confused about how we end up with userfault processing a write fault. It was part of the reason for my 'unclear' wording about this being more about cpus not seeing updated values. Note that we do drop the fault mutex before calling handle_usefault, but by then we have already made the 'missing' determination. Thoughts? Perhaps, I am still confused.
On Oct 3, 2022, at 8:56 AM, Peter Xu <peterx@redhat.com> wrote: > > + */ > + if (hugetlb_pte_stable(h, mm, ptep, old_pte)) > + ret = hugetlb_handle_userfault( > + vma, mapping, idx, flags, haddr, > + address, VM_UFFD_MISSING); > + else > + /* Retry the fault */ > + ret = 0; Might be unrelated, but at least for the sake of consistency if not potential GUP issues, don’t you want to return VM_FAULT_RETRY ?
On Mon, Oct 03, 2022 at 02:00:36PM -0700, Nadav Amit wrote: > On Oct 3, 2022, at 8:56 AM, Peter Xu <peterx@redhat.com> wrote: > > > > > + */ > > + if (hugetlb_pte_stable(h, mm, ptep, old_pte)) > > + ret = hugetlb_handle_userfault( > > + vma, mapping, idx, flags, haddr, > > + address, VM_UFFD_MISSING); > > + else > > + /* Retry the fault */ > > + ret = 0; > > Might be unrelated, but at least for the sake of consistency if not > potential GUP issues, don’t you want to return VM_FAULT_RETRY ? VM_FAULT_RETRY implies releasing of mmap sem, while we didn't?
On Mon, Oct 03, 2022 at 10:20:29AM -0700, Mike Kravetz wrote: > On 10/03/22 11:56, Peter Xu wrote: > > After the recent rework patchset of hugetlb locking on pmd sharing, > > kselftest for userfaultfd sometimes fails on hugetlb private tests with > > unexpected write fault checks. > > > > It turns out there's nothing wrong within the locking series regarding this > > matter, but it could have changed the timing of threads so it can trigger > > an old bug. > > > > The real bug is when we call hugetlb_no_page() we're not with the pgtable > > lock. It means we're reading the pte values lockless. It's perfectly fine > > in most cases because before we do normal page allocations we'll take the > > lock and check pte_same() again. However before that, there are actually > > two paths on userfaultfd missing/minor handling that may directly move on > > with the fault process without checking the pte values. > > > > It means for these two paths we may be generating an uffd message based on > > an unstable pte, while an unstable pte can legally be anything as long as > > the modifier holds the pgtable lock. > > > > One example, which is also what happened in the failing kselftest and > > caused the test failure, is that for private mappings CoW can happen on one > > page. CoW requires pte being cleared before being replaced with a new page > > for TLB coherency, but then there can be a race condition: > > > > thread 1 thread 2 > > -------- -------- > > > > hugetlb_fault hugetlb_fault > > private pte RO > > hugetlb_wp > > pgtable_lock() > > huge_ptep_clear_flush > > pte=NULL > > hugetlb_no_page > > generate uffd missing event > > even if page existed!! > > set_huge_pte_at > > pgtable_unlock() > > Thanks for working on this Peter! > > I agree with this patch, but I suspect the above race is not possible. Why? > In both cases, we take the hugetlb fault mutex when processing a huegtlb > page fault. This means only one thread can execute the fault code for > a specific mapping/index at a time. This is why I was so confused, and may > remain a bit confused about how we end up with userfault processing a write > fault. It was part of the reason for my 'unclear' wording about this being > more about cpus not seeing updated values. Note that we do drop the fault > mutex before calling handle_usefault, but by then we have already made the > 'missing' determination. > > Thoughts? Perhaps, I am still confused. It's my fault to have the commit message wrong, sorry. :) And thanks for raising this question, I could have overlooked that. It turns out it's not the CoW that's clearing the pte... it's the wr-protect with huge_ptep_modify_prot_start(). So the race is with UFFDIO_WRITEPROTECT, not CoW. Obviously when I was tracking the hpte changes I overlooked huge_ptep_get_and_clear(), only seeing the CoW path and I'm pretty sure that's already a bug which was obvious enough. I didn't prove they happened at the same time during the MISSING event. Then after I further looked at the CoW code I start to question myself on why CoW would trigger at all even with an available fault mutex, since for private mappings mapcount should be 1: if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { if (!PageAnonExclusive(old_page)) page_move_anon_rmap(old_page, vma); if (likely(!unshare)) set_huge_ptep_writable(vma, haddr, ptep); delayacct_wpcopy_end(); return 0; } There could still be something else I didn't catch, even though that may not be relevant to this patchset anymore. I'll wait a few more hours for other reviewers, then prepare a new version with the corrected commit message. Thanks,
On Oct 3, 2022, at 2:16 PM, Peter Xu <peterx@redhat.com> wrote: > On Mon, Oct 03, 2022 at 02:00:36PM -0700, Nadav Amit wrote: >> On Oct 3, 2022, at 8:56 AM, Peter Xu <peterx@redhat.com> wrote: >> >>> + */ >>> + if (hugetlb_pte_stable(h, mm, ptep, old_pte)) >>> + ret = hugetlb_handle_userfault( >>> + vma, mapping, idx, flags, haddr, >>> + address, VM_UFFD_MISSING); >>> + else >>> + /* Retry the fault */ >>> + ret = 0; >> >> Might be unrelated, but at least for the sake of consistency if not >> potential GUP issues, don’t you want to return VM_FAULT_RETRY ? > > VM_FAULT_RETRY implies releasing of mmap sem, while we didn't? Of course. My bad.
On 10/03/22 17:27, Peter Xu wrote: > On Mon, Oct 03, 2022 at 10:20:29AM -0700, Mike Kravetz wrote: > > On 10/03/22 11:56, Peter Xu wrote: > > > After the recent rework patchset of hugetlb locking on pmd sharing, > > > kselftest for userfaultfd sometimes fails on hugetlb private tests with > > > unexpected write fault checks. > > > > > > It turns out there's nothing wrong within the locking series regarding this > > > matter, but it could have changed the timing of threads so it can trigger > > > an old bug. > > > > > > The real bug is when we call hugetlb_no_page() we're not with the pgtable > > > lock. It means we're reading the pte values lockless. It's perfectly fine > > > in most cases because before we do normal page allocations we'll take the > > > lock and check pte_same() again. However before that, there are actually > > > two paths on userfaultfd missing/minor handling that may directly move on > > > with the fault process without checking the pte values. > > > > > > It means for these two paths we may be generating an uffd message based on > > > an unstable pte, while an unstable pte can legally be anything as long as > > > the modifier holds the pgtable lock. > > > > > > One example, which is also what happened in the failing kselftest and > > > caused the test failure, is that for private mappings CoW can happen on one > > > page. CoW requires pte being cleared before being replaced with a new page > > > for TLB coherency, but then there can be a race condition: > > > > > > thread 1 thread 2 > > > -------- -------- > > > > > > hugetlb_fault hugetlb_fault > > > private pte RO > > > hugetlb_wp > > > pgtable_lock() > > > huge_ptep_clear_flush > > > pte=NULL > > > hugetlb_no_page > > > generate uffd missing event > > > even if page existed!! > > > set_huge_pte_at > > > pgtable_unlock() > > > > Thanks for working on this Peter! > > > > I agree with this patch, but I suspect the above race is not possible. Why? > > In both cases, we take the hugetlb fault mutex when processing a huegtlb > > page fault. This means only one thread can execute the fault code for > > a specific mapping/index at a time. This is why I was so confused, and may > > remain a bit confused about how we end up with userfault processing a write > > fault. It was part of the reason for my 'unclear' wording about this being > > more about cpus not seeing updated values. Note that we do drop the fault > > mutex before calling handle_usefault, but by then we have already made the > > 'missing' determination. > > > > Thoughts? Perhaps, I am still confused. > > It's my fault to have the commit message wrong, sorry. :) And thanks for > raising this question, I could have overlooked that. > > It turns out it's not the CoW that's clearing the pte... it's the > wr-protect with huge_ptep_modify_prot_start(). So the race is with > UFFDIO_WRITEPROTECT, not CoW. Thank you! Now I understand. This also explains why the new locking exposes the race. hugetlb_change_protection needs to take the i_mmap_sema in write mode because it could unshare pmds. Previously, hugetlb page faults took i_mmap_sema in read mode so this race could not happen.
On Mon, Oct 03, 2022 at 02:45:47PM -0700, Mike Kravetz wrote: > This also explains why the new locking exposes the race. > hugetlb_change_protection needs to take the i_mmap_sema in write mode because > it could unshare pmds. Previously, hugetlb page faults took i_mmap_sema in > read mode so this race could not happen. Makes sense, thanks for explaining.
diff --git a/mm/hugetlb.c b/mm/hugetlb.c index 9679fe519b90..fa3fcdb0c4b8 100644 --- a/mm/hugetlb.c +++ b/mm/hugetlb.c @@ -5521,6 +5521,23 @@ static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, return ret; } +/* + * Recheck pte with pgtable lock. Returns true if pte didn't change, or + * false if pte changed or is changing. + */ +static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, + pte_t *ptep, pte_t old_pte) +{ + spinlock_t *ptl; + bool same; + + ptl = huge_pte_lock(h, mm, ptep); + same = pte_same(huge_ptep_get(ptep), old_pte); + spin_unlock(ptl); + + return same; +} + static vm_fault_t hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, struct address_space *mapping, pgoff_t idx, @@ -5562,9 +5579,30 @@ static vm_fault_t hugetlb_no_page(struct mm_struct *mm, goto out; /* Check for page in userfault range */ if (userfaultfd_missing(vma)) { - ret = hugetlb_handle_userfault(vma, mapping, idx, - flags, haddr, address, - VM_UFFD_MISSING); + /* + * Since hugetlb_no_page() was examining pte + * without pgtable lock, we need to re-test under + * lock because the pte may not be stable and could + * have changed from under us. Try to detect + * either changed or during-changing ptes and retry + * properly when needed. + * + * Note that userfaultfd is actually fine with + * false positives (e.g. caused by pte changed), + * but not wrong logical events (e.g. caused by + * reading a pte during changing). The latter can + * confuse the userspace, so the strictness is very + * much preferred. E.g., MISSING event should + * never happen on the page after UFFDIO_COPY has + * correctly installed the page and returned. + */ + if (hugetlb_pte_stable(h, mm, ptep, old_pte)) + ret = hugetlb_handle_userfault( + vma, mapping, idx, flags, haddr, + address, VM_UFFD_MISSING); + else + /* Retry the fault */ + ret = 0; goto out; } @@ -5634,9 +5672,14 @@ static vm_fault_t hugetlb_no_page(struct mm_struct *mm, if (userfaultfd_minor(vma)) { unlock_page(page); put_page(page); - ret = hugetlb_handle_userfault(vma, mapping, idx, - flags, haddr, address, - VM_UFFD_MINOR); + /* See comment in userfaultfd_missing() block above */ + if (hugetlb_pte_stable(h, mm, ptep, old_pte)) + ret = hugetlb_handle_userfault( + vma, mapping, idx, flags, haddr, + address, VM_UFFD_MINOR); + else + /* Retry the fault */ + ret = 0; goto out; } }
After the recent rework patchset of hugetlb locking on pmd sharing, kselftest for userfaultfd sometimes fails on hugetlb private tests with unexpected write fault checks. It turns out there's nothing wrong within the locking series regarding this matter, but it could have changed the timing of threads so it can trigger an old bug. The real bug is when we call hugetlb_no_page() we're not with the pgtable lock. It means we're reading the pte values lockless. It's perfectly fine in most cases because before we do normal page allocations we'll take the lock and check pte_same() again. However before that, there are actually two paths on userfaultfd missing/minor handling that may directly move on with the fault process without checking the pte values. It means for these two paths we may be generating an uffd message based on an unstable pte, while an unstable pte can legally be anything as long as the modifier holds the pgtable lock. One example, which is also what happened in the failing kselftest and caused the test failure, is that for private mappings CoW can happen on one page. CoW requires pte being cleared before being replaced with a new page for TLB coherency, but then there can be a race condition: thread 1 thread 2 -------- -------- hugetlb_fault hugetlb_fault private pte RO hugetlb_wp pgtable_lock() huge_ptep_clear_flush pte=NULL hugetlb_no_page generate uffd missing event even if page existed!! set_huge_pte_at pgtable_unlock() Fix this by recheck the pte after pgtable lock for both userfaultfd missing & minor fault paths. This bug should have been around starting from uffd hugetlb introduced, so attaching a Fixes to the commit. Also attach another Fixes to the minor support commit for easier tracking. Note that userfaultfd is actually fine with false positives (e.g. caused by pte changed), but not wrong logical events (e.g. caused by reading a pte during changing). The latter can confuse the userspace, so the strictness is very much preferred. E.g., MISSING event should never happen on the page after UFFDIO_COPY has correctly installed the page and returned. Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: Axel Rasmussen <axelrasmussen@google.com> Cc: Nadav Amit <nadav.amit@gmail.com> Fixes: 1a1aad8a9b7b ("userfaultfd: hugetlbfs: add userfaultfd hugetlb hook") Fixes: 7677f7fd8be7 ("userfaultfd: add minor fault registration mode") Co-developed-by: Mike Kravetz <mike.kravetz@oracle.com> Signed-off-by: Peter Xu <peterx@redhat.com> --- mm/hugetlb.c | 55 ++++++++++++++++++++++++++++++++++++++++++++++------ 1 file changed, 49 insertions(+), 6 deletions(-)