cluster/afr: detect in-progress creation in lookup and return ENOENT

if any subvol returned ENOENT while parent entrylk lock was held,
yield and return ENOENT for the entire lookup.

This is how the issue happens:

Multiple clients A, B and C are attempting 'mkdir -p /mnt/a/b/c'

1 Client A is in the middle of mkdir(/a). It has acquired lock.
  It has performed mkdir(/a) on one subvol, and second one is still
  in progress
2 Client B performs a lookup, sees directory /a on one,
  ENOENT on the other, succeeds lookup.
3 Client B performs lookup on /a/b on both subvols, both return ENOENT
  (one subvol because /a/b does not exist, another because /a
  itself does not exist)
4 Client B proceeds to mkdir /a/b. It obtains entrylk on inode=/a with
  basename=b on one subvol, but fails on other subvol as /a is yet to
  be created by Client A.
5 Client A finishes mkdir of /a on other subvol
6 Client C also attempts to create /a/b, lookup returns ENOENT on
  both subvols.
7 Client C tries to obtain entrylk on on inode=/a with basename=b,
  obtains on one subvol (where B had failed), and waits for B to unlock
  on other subvol.
8 Client B finishes mkdir() on one subvol with GFID-1 and completes
  transaction and unlocks
9 Client C gets the lock on the second subvol, At this stage second
  subvol already has /a/b created from Client B, but Client C does not
  check that in the middle of mkdir transaction
10 Client C attempts mkdir /a/b on both subvols. It succeeds on
   ONLY ONE (where Client B could not get lock because of
   missing parent /a dir) with GFID-2, and gets EEXIST from ONE subvol.
This way we have /a/b in GFID mismatch. One subvol got GFID-1 because
Client B performed transaction on only one subvol (because entrylk()
could not be obtained on second subvol because of missing parent dir --
caused by premature/speculative succeeding of lookup() on /a when locks
are detected). Other subvol gets GFID-2 from Client C because while
it was waiting for entrylk() on both subvols, Client B was in the
middle of creating mkdir() on only one subvol, and Client C does not
"expect" this when it is between lock() and pre-op()/op() phase of the
transaction.

Original-author: Anand Avati <avati@redhat.com>
Change-Id: Idca475dbbc2a51e09da6fa0f9e1e37148caef208
BUG: 860210
Signed-off-by: Pranith Kumar K <pkarampu@redhat.com>
Reviewed-on: http://review.gluster.org/4625
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Anand Avati <avati@redhat.com>

1 files changed, 57 insertions, 0 deletions

diff --git a/xlators/cluster/afr/src/afr-common.c b/xlators/cluster/afr/src/afr-common.c
index 0eca21309a8..89d964056a0 100644
--- a/xlators/cluster/afr/src/afr-common.c
+++ b/xlators/cluster/afr/src/afr-common.c
@@ -1920,6 +1920,57 @@ out:
         return;
 }
 
+gf_boolean_t
+afr_is_entry_possibly_under_creation (afr_local_t *local, xlator_t *this)
+{
+        /*
+         * We need to perform this test in lookup done and treat on going
+         * create/DELETE as ENOENT.
+         * Reason:
+        Multiple clients A, B and C are attempting 'mkdir -p /mnt/a/b/c'
+
+        1 Client A is in the middle of mkdir(/a). It has acquired lock.
+          It has performed mkdir(/a) on one subvol, and second one is still
+          in progress
+        2 Client B performs a lookup, sees directory /a on one,
+          ENOENT on the other, succeeds lookup.
+        3 Client B performs lookup on /a/b on both subvols, both return ENOENT
+          (one subvol because /a/b does not exist, another because /a
+          itself does not exist)
+        4 Client B proceeds to mkdir /a/b. It obtains entrylk on inode=/a with
+          basename=b on one subvol, but fails on other subvol as /a is yet to
+          be created by Client A.
+        5 Client A finishes mkdir of /a on other subvol
+        6 Client C also attempts to create /a/b, lookup returns ENOENT on
+          both subvols.
+        7 Client C tries to obtain entrylk on on inode=/a with basename=b,
+          obtains on one subvol (where B had failed), and waits for B to unlock
+          on other subvol.
+        8 Client B finishes mkdir() on one subvol with GFID-1 and completes
+          transaction and unlocks
+        9 Client C gets the lock on the second subvol, At this stage second
+          subvol already has /a/b created from Client B, but Client C does not
+          check that in the middle of mkdir transaction
+        10 Client C attempts mkdir /a/b on both subvols. It succeeds on
+           ONLY ONE (where Client B could not get lock because of
+           missing parent /a dir) with GFID-2, and gets EEXIST from ONE subvol.
+        This way we have /a/b in GFID mismatch. One subvol got GFID-1 because
+        Client B performed transaction on only one subvol (because entrylk()
+        could not be obtained on second subvol because of missing parent dir --
+        caused by premature/speculative succeeding of lookup() on /a when locks
+        are detected). Other subvol gets GFID-2 from Client C because while
+        it was waiting for entrylk() on both subvols, Client B was in the
+        middle of creating mkdir() on only one subvol, and Client C does not
+        "expect" this when it is between lock() and pre-op()/op() phase of the
+        transaction.
+         */
+	if (local->cont.lookup.parent_entrylk && local->enoent_count)
+		return _gf_true;
+
+	return _gf_false;
+}
+
+
 static void
 afr_lookup_done (call_frame_t *frame, xlator_t *this)
 {
@@ -1936,6 +1987,12 @@ afr_lookup_done (call_frame_t *frame, xlator_t *this)
         priv  = this->private;
         local = frame->local;
 
+	if (afr_is_entry_possibly_under_creation (local, this)) {
+		local->op_ret = -1;
+		local->op_errno = ENOENT;
+		goto unwind;
+	}
+
         if (local->op_ret < 0)
                 goto unwind;