From 2ab72d3db1ab8c916bbc33c5833c161ce98eff63 Mon Sep 17 00:00:00 2001

From: Pranith Kumar K <pkarampu@redhat.com>

Date: Tue, 13 Jun 2017 23:35:40 +0530

Subject: [PATCH 514/525] cluster/ec: lk shouldn't be a transaction

Problem:

When application sends a blocking lock, the lk fop actually waits under

inodelk.  This can lead to a dead-lock.

1) Let's say app-1 takes exculsive-fcntl-lock on the file

2) app-2 attempts an exclusive-fcntl-lock on the file which goes to blocking

   stage note: app-2 is blocked inside transaction which holds an inode-lock

3) app-1 tries to perform write which needs inode-lock so it gets blocked on

   app-2 to unlock inodelk and app-2 is blocked on app-1 to unlock fcntl-lock

Fix:

Correct way to fix this issue and make fcntl locks perform well would be to

introduce

2-phase locking for fcntl lock:

1) Implement a try-lock phase where locks xlator will not merge lk call with

   existing calls until a commit-lock phase.

2) If in try-lock phase we get quorum number of success without any EAGAIN

   error, then send a commit-lock which will merge locks.

3) In case there are any errors, unlock should just delete the lock-object

   which was tried earlier and shouldn't touch the committed locks.

Unfortunately this is a sizeable feature and need to be thought through for any

corner cases.  Until then remove transaction from lk call.

 >BUG: 1455049

 >Change-Id: I18a782903ba0eb43f1e6526fb0cf8c626c460159

 >Signed-off-by: Pranith Kumar K <pkarampu@redhat.com>

 >Reviewed-on: https://review.gluster.org/17542

 >Smoke: Gluster Build System <jenkins@build.gluster.org>

 >NetBSD-regression: NetBSD Build System <jenkins@build.gluster.org>

 >CentOS-regression: Gluster Build System <jenkins@build.gluster.org>

 >Reviewed-by: Ashish Pandey <aspandey@redhat.com>

 >Reviewed-by: Xavier Hernandez <xhernandez@datalab.es>

BUG: 1444515

Change-Id: I723e675119a08ec2a9c9073eb5eac73dc455af0d

Signed-off-by: Pranith Kumar K <pkarampu@redhat.com>

Reviewed-on: https://code.engineering.redhat.com/gerrit/109275

---

 xlators/cluster/ec/src/ec-locks.c | 20 +-------------------

 1 file changed, 1 insertion(+), 19 deletions(-)

diff --git a/xlators/cluster/ec/src/ec-locks.c b/xlators/cluster/ec/src/ec-locks.c

index 70b07c3..ff09852 100644

--- a/xlators/cluster/ec/src/ec-locks.c

+++ b/xlators/cluster/ec/src/ec-locks.c

@@ -1020,12 +1020,6 @@ int32_t ec_manager_lk(ec_fop_data_t * fop, int32_t state)

         /* Fall through */

-        case EC_STATE_LOCK:

-            ec_lock_prepare_fd(fop, fop->fd, EC_UPDATE_DATA | EC_QUERY_INFO);

-            ec_lock(fop);

-

-            return EC_STATE_DISPATCH;

-

         case EC_STATE_DISPATCH:

             ec_dispatch_all(fop);

@@ -1080,10 +1074,9 @@ int32_t ec_manager_lk(ec_fop_data_t * fop, int32_t state)

                              cbk->op_errno, &cbk->flock, cbk->xdata);

             }

-            return EC_STATE_LOCK_REUSE;

+            return EC_STATE_END;

         case -EC_STATE_INIT:

-        case -EC_STATE_LOCK:

         case -EC_STATE_DISPATCH:

         case -EC_STATE_REPORT:

             GF_ASSERT(fop->error != 0);

@@ -1094,17 +1087,6 @@ int32_t ec_manager_lk(ec_fop_data_t * fop, int32_t state)

                              NULL, NULL);

             }

-            return EC_STATE_LOCK_REUSE;

-

-        case -EC_STATE_LOCK_REUSE:

-        case EC_STATE_LOCK_REUSE:

-            ec_lock_reuse(fop);

-

-            return EC_STATE_UNLOCK;

-

-        case -EC_STATE_UNLOCK:

-        case EC_STATE_UNLOCK:

-            ec_unlock(fop);

             return EC_STATE_END;

-- 

1.8.3.1