There is a race in ceph_put_snap_realm. The change to the nref and the spinlock acquisition are not done atomically, so you could decrement nref, and before you take the spinlock, the nref is incremented again. At that point, you end up putting it on the empty list when it shouldn't be there. Eventually __cleanup_empty_realms runs and frees it when it's still in-use.
Fix this by protecting the 1->0 transition with atomic_dec_and_lock, and just drop the spinlock if we can get the rwsem.
Because these objects can also undergo a 0->1 refcount transition, we must protect that change as well with the spinlock. Increment locklessly unless the value is at 0, in which case we take the spinlock, increment and then take it off the empty list if it did the 0->1 transition.
With these changes, I'm removing the dout() messages from these functions, as well as in __put_snap_realm. They've always been racy, and it's better to not print values that may be misleading.
Cc: stable@vger.kernel.org Cc: Sage Weil sage@redhat.com Reported-by: Mark Nelson mnelson@redhat.com URL: https://tracker.ceph.com/issues/46419 Signed-off-by: Jeff Layton jlayton@kernel.org --- fs/ceph/snap.c | 34 +++++++++++++++++----------------- 1 file changed, 17 insertions(+), 17 deletions(-)
v2: No functional changes, but I cleaned up the comments a bit and added another in __put_snap_realm.
diff --git a/fs/ceph/snap.c b/fs/ceph/snap.c index 9dbc92cfda38..158c11e96fb7 100644 --- a/fs/ceph/snap.c +++ b/fs/ceph/snap.c @@ -67,19 +67,19 @@ void ceph_get_snap_realm(struct ceph_mds_client *mdsc, { lockdep_assert_held(&mdsc->snap_rwsem);
- dout("get_realm %p %d -> %d\n", realm, - atomic_read(&realm->nref), atomic_read(&realm->nref)+1); /* - * since we _only_ increment realm refs or empty the empty - * list with snap_rwsem held, adjusting the empty list here is - * safe. we do need to protect against concurrent empty list - * additions, however. + * The 0->1 and 1->0 transitions must take the snap_empty_lock + * atomically with the refcount change. Go ahead and bump the + * nref here, unless it's 0, in which case we take the spinlock + * and then do the increment and remove it from the list. */ - if (atomic_inc_return(&realm->nref) == 1) { - spin_lock(&mdsc->snap_empty_lock); + if (atomic_add_unless(&realm->nref, 1, 0)) + return; + + spin_lock(&mdsc->snap_empty_lock); + if (atomic_inc_return(&realm->nref) == 1) list_del_init(&realm->empty_item); - spin_unlock(&mdsc->snap_empty_lock); - } + spin_unlock(&mdsc->snap_empty_lock); }
static void __insert_snap_realm(struct rb_root *root, @@ -208,28 +208,28 @@ static void __put_snap_realm(struct ceph_mds_client *mdsc, { lockdep_assert_held_write(&mdsc->snap_rwsem);
- dout("__put_snap_realm %llx %p %d -> %d\n", realm->ino, realm, - atomic_read(&realm->nref), atomic_read(&realm->nref)-1); + /* + * We do not require the snap_empty_lock here, as any caller that + * increments the value must hold the snap_rwsem. + */ if (atomic_dec_and_test(&realm->nref)) __destroy_snap_realm(mdsc, realm); }
/* - * caller needn't hold any locks + * See comments in ceph_get_snap_realm. Caller needn't hold any locks. */ void ceph_put_snap_realm(struct ceph_mds_client *mdsc, struct ceph_snap_realm *realm) { - dout("put_snap_realm %llx %p %d -> %d\n", realm->ino, realm, - atomic_read(&realm->nref), atomic_read(&realm->nref)-1); - if (!atomic_dec_and_test(&realm->nref)) + if (!atomic_dec_and_lock(&realm->nref, &mdsc->snap_empty_lock)) return;
if (down_write_trylock(&mdsc->snap_rwsem)) { + spin_unlock(&mdsc->snap_empty_lock); __destroy_snap_realm(mdsc, realm); up_write(&mdsc->snap_rwsem); } else { - spin_lock(&mdsc->snap_empty_lock); list_add(&realm->empty_item, &mdsc->snap_empty); spin_unlock(&mdsc->snap_empty_lock); }
Jeff Layton jlayton@kernel.org writes:
There is a race in ceph_put_snap_realm. The change to the nref and the spinlock acquisition are not done atomically, so you could decrement nref, and before you take the spinlock, the nref is incremented again. At that point, you end up putting it on the empty list when it shouldn't be there. Eventually __cleanup_empty_realms runs and frees it when it's still in-use.
Fix this by protecting the 1->0 transition with atomic_dec_and_lock, and just drop the spinlock if we can get the rwsem.
Because these objects can also undergo a 0->1 refcount transition, we must protect that change as well with the spinlock. Increment locklessly unless the value is at 0, in which case we take the spinlock, increment and then take it off the empty list if it did the 0->1 transition.
With these changes, I'm removing the dout() messages from these functions, as well as in __put_snap_realm. They've always been racy, and it's better to not print values that may be misleading.
Cc: stable@vger.kernel.org Cc: Sage Weil sage@redhat.com Reported-by: Mark Nelson mnelson@redhat.com URL: https://tracker.ceph.com/issues/46419 Signed-off-by: Jeff Layton jlayton@kernel.org
fs/ceph/snap.c | 34 +++++++++++++++++----------------- 1 file changed, 17 insertions(+), 17 deletions(-)
v2: No functional changes, but I cleaned up the comments a bit and added another in __put_snap_realm.
diff --git a/fs/ceph/snap.c b/fs/ceph/snap.c index 9dbc92cfda38..158c11e96fb7 100644 --- a/fs/ceph/snap.c +++ b/fs/ceph/snap.c @@ -67,19 +67,19 @@ void ceph_get_snap_realm(struct ceph_mds_client *mdsc, { lockdep_assert_held(&mdsc->snap_rwsem);
- dout("get_realm %p %d -> %d\n", realm,
/*atomic_read(&realm->nref), atomic_read(&realm->nref)+1);
* since we _only_ increment realm refs or empty the empty
* list with snap_rwsem held, adjusting the empty list here is
* safe. we do need to protect against concurrent empty list
* additions, however.
* The 0->1 and 1->0 transitions must take the snap_empty_lock
* atomically with the refcount change. Go ahead and bump the
* nref here, unless it's 0, in which case we take the spinlock
*/* and then do the increment and remove it from the list.
- if (atomic_inc_return(&realm->nref) == 1) {
spin_lock(&mdsc->snap_empty_lock);
- if (atomic_add_unless(&realm->nref, 1, 0))
Here you could probably use atomic_inc_not_zero() instead. But other than that it looks good. Thanks a lot for solving yet another locking puzzle!
Reviewed-by: Luis Henriques lhenriques@suse.de
Cheers,
On Wed, 2021-08-04 at 17:26 +0100, Luis Henriques wrote:
Jeff Layton jlayton@kernel.org writes:
There is a race in ceph_put_snap_realm. The change to the nref and the spinlock acquisition are not done atomically, so you could decrement nref, and before you take the spinlock, the nref is incremented again. At that point, you end up putting it on the empty list when it shouldn't be there. Eventually __cleanup_empty_realms runs and frees it when it's still in-use.
Fix this by protecting the 1->0 transition with atomic_dec_and_lock, and just drop the spinlock if we can get the rwsem.
Because these objects can also undergo a 0->1 refcount transition, we must protect that change as well with the spinlock. Increment locklessly unless the value is at 0, in which case we take the spinlock, increment and then take it off the empty list if it did the 0->1 transition.
With these changes, I'm removing the dout() messages from these functions, as well as in __put_snap_realm. They've always been racy, and it's better to not print values that may be misleading.
Cc: stable@vger.kernel.org Cc: Sage Weil sage@redhat.com Reported-by: Mark Nelson mnelson@redhat.com URL: https://tracker.ceph.com/issues/46419 Signed-off-by: Jeff Layton jlayton@kernel.org
fs/ceph/snap.c | 34 +++++++++++++++++----------------- 1 file changed, 17 insertions(+), 17 deletions(-)
v2: No functional changes, but I cleaned up the comments a bit and added another in __put_snap_realm.
diff --git a/fs/ceph/snap.c b/fs/ceph/snap.c index 9dbc92cfda38..158c11e96fb7 100644 --- a/fs/ceph/snap.c +++ b/fs/ceph/snap.c @@ -67,19 +67,19 @@ void ceph_get_snap_realm(struct ceph_mds_client *mdsc, { lockdep_assert_held(&mdsc->snap_rwsem);
- dout("get_realm %p %d -> %d\n", realm,
/*atomic_read(&realm->nref), atomic_read(&realm->nref)+1);
* since we _only_ increment realm refs or empty the empty
* list with snap_rwsem held, adjusting the empty list here is
* safe. we do need to protect against concurrent empty list
* additions, however.
* The 0->1 and 1->0 transitions must take the snap_empty_lock
* atomically with the refcount change. Go ahead and bump the
* nref here, unless it's 0, in which case we take the spinlock
*/* and then do the increment and remove it from the list.
- if (atomic_inc_return(&realm->nref) == 1) {
spin_lock(&mdsc->snap_empty_lock);
- if (atomic_add_unless(&realm->nref, 1, 0))
Here you could probably use atomic_inc_not_zero() instead. But other than that it looks good. Thanks a lot for solving yet another locking puzzle!
Reviewed-by: Luis Henriques lhenriques@suse.de
Cheers,
Good point! That is a little clearer. I'll incorporate that change and merge it.
Thanks,
linux-stable-mirror@lists.linaro.org