]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commit
mm: fix memory ordering for mm_lock_seq and vm_lock_seq
authorJann Horn <jannh@google.com>
Fri, 21 Jul 2023 22:51:07 +0000 (00:51 +0200)
committerAndrew Morton <akpm@linux-foundation.org>
Thu, 27 Jul 2023 20:07:04 +0000 (13:07 -0700)
commitb1f02b95758d05b799731d939e76a0bd6da312db
treee463d7987ce0dfb824db51b338d575874e3ed214
parent15571273db93ac2e4415e74280e04566c31d5eb0
mm: fix memory ordering for mm_lock_seq and vm_lock_seq

mm->mm_lock_seq effectively functions as a read/write lock; therefore it
must be used with acquire/release semantics.

A specific example is the interaction between userfaultfd_register() and
lock_vma_under_rcu().

userfaultfd_register() does the following from the point where it changes
a VMA's flags to the point where concurrent readers are permitted again
(in a simple scenario where only a single private VMA is accessed and no
merging/splitting is involved):

userfaultfd_register
  userfaultfd_set_vm_flags
    vm_flags_reset
      vma_start_write
        down_write(&vma->vm_lock->lock)
        vma->vm_lock_seq = mm_lock_seq [marks VMA as busy]
        up_write(&vma->vm_lock->lock)
      vm_flags_init
        [sets VM_UFFD_* in __vm_flags]
  vma->vm_userfaultfd_ctx.ctx = ctx
  mmap_write_unlock
    vma_end_write_all
      WRITE_ONCE(mm->mm_lock_seq, mm->mm_lock_seq + 1) [unlocks VMA]

There are no memory barriers in between the __vm_flags update and the
mm->mm_lock_seq update that unlocks the VMA, so the unlock can be
reordered to above the `vm_flags_init()` call, which means from the
perspective of a concurrent reader, a VMA can be marked as a userfaultfd
VMA while it is not VMA-locked.  That's bad, we definitely need a
store-release for the unlock operation.

The non-atomic write to vma->vm_lock_seq in vma_start_write() is mostly
fine because all accesses to vma->vm_lock_seq that matter are always
protected by the VMA lock.  There is a racy read in vma_start_read()
though that can tolerate false-positives, so we should be using
WRITE_ONCE() to keep things tidy and data-race-free (including for KCSAN).

On the other side, lock_vma_under_rcu() works as follows in the relevant
region for locking and userfaultfd check:

lock_vma_under_rcu
  vma_start_read
    vma->vm_lock_seq == READ_ONCE(vma->vm_mm->mm_lock_seq) [early bailout]
    down_read_trylock(&vma->vm_lock->lock)
    vma->vm_lock_seq == READ_ONCE(vma->vm_mm->mm_lock_seq) [main check]
  userfaultfd_armed
    checks vma->vm_flags & __VM_UFFD_FLAGS

Here, the interesting aspect is how far down the mm->mm_lock_seq read can
be reordered - if this read is reordered down below the vma->vm_flags
access, this could cause lock_vma_under_rcu() to partly operate on
information that was read while the VMA was supposed to be locked.  To
prevent this kind of downwards bleeding of the mm->mm_lock_seq read, we
need to read it with a load-acquire.

Some of the comment wording is based on suggestions by Suren.

BACKPORT WARNING: One of the functions changed by this patch (which I've
written against Linus' tree) is vma_try_start_write(), but this function
no longer exists in mm/mm-everything.  I don't know whether the merged
version of this patch will be ordered before or after the patch that
removes vma_try_start_write().  If you're backporting this patch to a tree
with vma_try_start_write(), make sure this patch changes that function.

Link: https://lkml.kernel.org/r/20230721225107.942336-1-jannh@google.com
Fixes: 5e31275cc997 ("mm: add per-VMA lock and helper functions to control it")
Signed-off-by: Jann Horn <jannh@google.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
include/linux/mm.h
include/linux/mm_types.h
include/linux/mmap_lock.h