On Tue, Dec 17, 2024 at 2:30 AM Peter Zijlstra [off-list ref] wrote:

On Mon, Dec 16, 2024 at 01:44:45PM -0800, Suren Baghdasaryan wrote:

quoted

On Mon, Dec 16, 2024 at 1:38 PM Peter Zijlstra [off-list ref] wrote:

quoted

On Mon, Dec 16, 2024 at 11:24:13AM -0800, Suren Baghdasaryan wrote:

quoted

+static inline void vma_refcount_put(struct vm_area_struct *vma)
+{
+     int refcnt;
+
+     if (!__refcount_dec_and_test(&vma->vm_refcnt, &refcnt)) {
+             rwsem_release(&vma->vmlock_dep_map, _RET_IP_);
+
+             if (refcnt & VMA_STATE_LOCKED)
+                     rcuwait_wake_up(&vma->vm_mm->vma_writer_wait);
+     }
+}
+
 /*
  * Try to read-lock a vma. The function is allowed to occasionally yield false
  * locked result to avoid performance overhead, in which case we fall back to

@@ -710,6 +728,8 @@ static inline void vma_lock_init(struct vm_area_struct *vma)
  */
 static inline bool vma_start_read(struct vm_area_struct *vma)
 {
+     int oldcnt;
+
      /*
       * Check before locking. A race might cause false locked result.
       * We can use READ_ONCE() for the mm_lock_seq here, and don't need

@@ -720,13 +740,20 @@ static inline bool vma_start_read(struct vm_area_struct *vma)
      if (READ_ONCE(vma->vm_lock_seq) == READ_ONCE(vma->vm_mm->mm_lock_seq.sequence))
              return false;

+
+     rwsem_acquire_read(&vma->vmlock_dep_map, 0, 0, _RET_IP_);
+     /* Limit at VMA_STATE_LOCKED - 2 to leave one count for a writer */
+     if (unlikely(!__refcount_inc_not_zero_limited(&vma->vm_refcnt, &oldcnt,
+                                                   VMA_STATE_LOCKED - 2))) {
+             rwsem_release(&vma->vmlock_dep_map, _RET_IP_);
              return false;
+     }
+     lock_acquired(&vma->vmlock_dep_map, _RET_IP_);

      /*
+      * Overflow of vm_lock_seq/mm_lock_seq might produce false locked result.
       * False unlocked result is impossible because we modify and check
+      * vma->vm_lock_seq under vma->vm_refcnt protection and mm->mm_lock_seq
       * modification invalidates all existing locks.
       *
       * We must use ACQUIRE semantics for the mm_lock_seq so that if we are

@@ -734,10 +761,12 @@ static inline bool vma_start_read(struct vm_area_struct *vma)
       * after it has been unlocked.
       * This pairs with RELEASE semantics in vma_end_write_all().
       */
+     if (oldcnt & VMA_STATE_LOCKED ||
+         unlikely(vma->vm_lock_seq == raw_read_seqcount(&vma->vm_mm->mm_lock_seq))) {
+             vma_refcount_put(vma);

Suppose we have detach race with a concurrent RCU lookup like:

                                        vma = mas_lookup();

        vma_start_write();
        mas_detach();
                                        vma_start_read()
                                        rwsem_acquire_read()
                                        inc // success
        vma_mark_detach();
        dec_and_test // assumes 1->0
                     // is actually 2->1

                                        if (vm_lock_seq == vma->vm_mm_mm_lock_seq) // true
                                          vma_refcount_put
                                            dec_and_test() // 1->0
                                              *NO* rwsem_release()

Yes, this is possible. I think that's not a problem until we start
reusing the vmas and I deal with this race later in this patchset.
I think what you described here is the same race I mention in the
description of this patch:
https://lore.kernel.org/all/20241216192419.2970941-14-surenb@google.com/ (local)
I introduce vma_ensure_detached() in that patch to handle this case
and ensure that vmas are detached before they are returned into the
slab cache for reuse. Does that make sense?

So I just replied there, and no, I don't think it makes sense. Just put
the kmem_cache_free() in vma_refcount_put(), to be done on 0.

That's very appealing indeed and makes things much simpler. The
problem I see with that is the case when we detach a vma from the tree
to isolate it, then do some cleanup and only then free it. That's done
in vms_gather_munmap_vmas() here:
https://elixir.bootlin.com/linux/v6.12.5/source/mm/vma.c#L1240 and we
even might reattach detached vmas back:
https://elixir.bootlin.com/linux/v6.12.5/source/mm/vma.c#L1312. IOW,
detached state is not final and we can't destroy the object that
reached this state. We could change states to: 0=unused (we can free
the object), 1=detached, 2=attached, etc. but then vma_start_read()
should do something like refcount_inc_more_than_one() instead of
refcount_inc_not_zero(). Would you be ok with such an approach?

Anyway, my point was more about the weird entanglement of lockdep and
the refcount. Just pull the lockdep annotation out of _put() and put it
explicitly in the vma_start_read() error paths and vma_end_read().

Ok, I think that's easy.

Additionally, having vma_end_write() would allow you to put a lockdep
annotation in vma_{start,end}_write() -- which was I think the original
reason I proposed it a while back, that and having improved clarity when
reading the code, since explicitly marking the end of a section is
helpful.

The vma->vmlock_dep_map is tracking vma->vm_refcnt, not the
vma->vm_lock_seq (similar to how today vma->vm_lock has its lockdep
tracking that rw_semaphore). If I implement vma_end_write() then it
will simply be something like:

void vma_end_write(vma)
{
         vma_assert_write_locked(vma);
         vma->vm_lock_seq = UINT_MAX;
}

so, vmlock_dep_map would not be involved.

If you want to track vma->vm_lock_seq with a separate lockdep, that
would be more complicated. Specifically for vma_end_write_all() that
would require us to call rwsem_release() on all locked vmas, however
we currently do not track individual locked vmas. vma_end_write_all()
allows us not to worry about tracking them, knowing that once we do
mmap_write_unlock() they all will get unlocked with one increment of
mm->mm_lock_seq. If your suggestion is to replace vma_end_write_all()
with vma_end_write() and unlock vmas individually across the mm code,
that would be a sizable effort. If that is indeed your ultimate goal,
I can do that as a separate project: introduce vma_end_write(),
gradually add them in required places (not yet sure how complex that
would be), then retire vma_end_write_all() and add a lockdep for
vma->vm_lock_seq.