| Message ID | 424274a4-7c13-e14-b380-428fc69a45c5@google.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | mm: allow pte_offset_map[_lock]() to fail | expand |
On Sun, May 21, 2023 at 10:06:32PM -0700, Hugh Dickins wrote: > Instead of worrying whether the pmd is stable, userfaultfd_must_wait() > call pte_offset_map() as before, but go back to try again if that fails. > > Risk of endless loop? It already broke out if pmd_none(), !pmd_present() > or pmd_trans_huge(), and pte_offset_map() would have cleared pmd_bad(): > which leaves pmd_devmap(). Presumably pmd_devmap() is inappropriate in > a vma subject to userfaultfd (it would have been mistreated before), > but add a check just to avoid all possibility of endless loop there. Agreed, afaiu that's for either PFNMAP or MIXEDMAP vmas only. Maybe we can use a WARN_ON_ONCE() for that to be clear, but no strong opinions. > > Signed-off-by: Hugh Dickins <hughd@google.com> Acked-by: Peter Xu <peterx@redhat.com>
diff --git a/fs/userfaultfd.c b/fs/userfaultfd.c index f7a0817b1ec0..ca83423f8d54 100644 --- a/fs/userfaultfd.c +++ b/fs/userfaultfd.c @@ -349,12 +349,13 @@ static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx, if (!pud_present(*pud)) goto out; pmd = pmd_offset(pud, address); +again: _pmd = pmdp_get_lockless(pmd); if (pmd_none(_pmd)) goto out; ret = false; - if (!pmd_present(_pmd)) + if (!pmd_present(_pmd) || pmd_devmap(_pmd)) goto out; if (pmd_trans_huge(_pmd)) { @@ -363,11 +364,11 @@ static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx, goto out; } - /* - * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it - * and use the standard pte_offset_map() instead of parsing _pmd. - */ pte = pte_offset_map(pmd, address); + if (!pte) { + ret = true; + goto again; + } /* * Lockless access: we're in a wait_event so it's ok if it * changes under us. PTE markers should be handled the same as none
Instead of worrying whether the pmd is stable, userfaultfd_must_wait() call pte_offset_map() as before, but go back to try again if that fails. Risk of endless loop? It already broke out if pmd_none(), !pmd_present() or pmd_trans_huge(), and pte_offset_map() would have cleared pmd_bad(): which leaves pmd_devmap(). Presumably pmd_devmap() is inappropriate in a vma subject to userfaultfd (it would have been mistreated before), but add a check just to avoid all possibility of endless loop there. Signed-off-by: Hugh Dickins <hughd@google.com> --- fs/userfaultfd.c | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-)