libglusterfs: Implementation of sync lock as recursive lock to avoid crash.

Problem : In glusterd,we are using big lock which is implemented based on sync
task frame work for thread synchronization and rcu lock for data consistency.
sync task frame work swap the threads  if there is no worker poll threads
available,due to this rcu lock and rcu unlock was happening in different threads
(urcu-bp will not allow this),resulting into glusterd crash.

fix : To avoid releasing the sync lock(big lock) in between rcu critical
section,implemented sync lock as recursive lock.

More details:
link : http://www.spinics.net/lists/gluster-devel/msg14632.html

Change-Id: I2b56c1caf3f0470f219b1adcaf62cce29cdc6b88
BUG: 1216942
Signed-off-by: anand <anekkunt@redhat.com>
Reviewed-on: http://review.gluster.org/10285
Reviewed-by: Atin Mukherjee <amukherj@redhat.com>
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Tested-by: NetBSD Build System
Reviewed-by: Vijay Bellur <vbellur@redhat.com>
(cherry picked from commit ada6b3a8800867934af57a57d5312f5a5d8374f0)
Reviewed-on: http://review.gluster.org/10432
Reviewed-by: Krishnan Parthasarathi <kparthas@redhat.com>
Tested-by: Vijay Bellur <vbellur@redhat.com>

2 files changed, 146 insertions, 54 deletions

diff --git a/libglusterfs/src/syncop.c b/libglusterfs/src/syncop.c
index 01ca9153733..c930193cb3c 100644
--- a/libglusterfs/src/syncop.c
+++ b/libglusterfs/src/syncop.c
@@ -827,16 +827,20 @@ syncenv_new (size_t stacksize, int procmin, int procmax)
 
 
 int
-synclock_init (synclock_t *lock)
+synclock_init (synclock_t *lock, lock_attr_t attr)
 {
-	if (!lock)
-		return -1;
+        if (!lock)
+               return -1;
 
-	pthread_cond_init (&lock->cond, 0);
-	lock->lock = 0;
-	INIT_LIST_HEAD (&lock->waitq);
+        pthread_cond_init (&lock->cond, 0);
+        lock->type = LOCK_NULL;
+        lock->owner = NULL;
+        lock->owner_tid = 0;
+        lock->lock = 0;
+        lock->attr = attr;
+        INIT_LIST_HEAD (&lock->waitq);
 
-	return pthread_mutex_init (&lock->guard, 0);
+        return pthread_mutex_init (&lock->guard, 0);
 }
 
 
@@ -854,32 +858,70 @@ synclock_destroy (synclock_t *lock)
 static int
 __synclock_lock (struct synclock *lock)
 {
-	struct synctask *task = NULL;
+        struct synctask *task = NULL;
 
-	if (!lock)
-		return -1;
+        if (!lock)
+                return -1;
 
-	task = synctask_get ();
+        task = synctask_get ();
 
-	while (lock->lock) {
-		if (task) {
-			/* called within a synctask */
-			list_add_tail (&task->waitq, &lock->waitq);
+        if (lock->lock && (lock->attr == SYNC_LOCK_RECURSIVE)) {
+                /*Recursive lock (if same owner requested for lock again then
+                 *increment lock count and return success).
+                 *Note:same number of unlocks required.
+                 */
+                switch (lock->type) {
+                case LOCK_TASK:
+                        if (task == lock->owner) {
+                                lock->lock++;
+                                gf_log ("", GF_LOG_TRACE, "Recursive lock called by "
+                                        "sync task.owner= %p,lock=%d", lock->owner, lock->lock);
+                                return 0;
+                        }
+                        break;
+                case LOCK_THREAD:
+                        if (pthread_self () == lock->owner_tid) {
+                                lock->lock++;
+                                gf_log ("", GF_LOG_TRACE, "Recursive lock called by "
+                                        "thread ,owner=%u lock=%d", (unsigned int) lock->owner_tid,
+                                         lock->lock);
+                                return 0;
+                        }
+                        break;
+                default:
+                        gf_log ("", GF_LOG_CRITICAL, "unknown lock type");
+                        break;
+                }
+        }
+
+
+        while (lock->lock) {
+                if (task) {
+                        /* called within a synctask */
+                        list_add_tail (&task->waitq, &lock->waitq);
                         pthread_mutex_unlock (&lock->guard);
                         synctask_yield (task);
                         /* task is removed from waitq in unlock,
                          * under lock->guard.*/
                         pthread_mutex_lock (&lock->guard);
-		} else {
-			/* called by a non-synctask */
-			pthread_cond_wait (&lock->cond, &lock->guard);
-		}
-	}
+                } else {
+                        /* called by a non-synctask */
+                        pthread_cond_wait (&lock->cond, &lock->guard);
+                }
+        }
 
-	lock->lock = _gf_true;
-	lock->owner = task;
+        if (task) {
+                lock->type = LOCK_TASK;
+                lock->owner = task;    /* for synctask*/
 
-	return 0;
+        } else {
+                lock->type = LOCK_THREAD;
+                lock->owner_tid = pthread_self (); /* for non-synctask */
+
+        }
+        lock->lock = 1;
+
+        return 0;
 }
 
 
@@ -925,37 +967,73 @@ unlock:
 static int
 __synclock_unlock (synclock_t *lock)
 {
-	struct synctask *task = NULL;
-	struct synctask *curr = NULL;
+        struct synctask *task = NULL;
+        struct synctask *curr = NULL;
 
-	if (!lock)
-		return -1;
+        if (!lock)
+               return -1;
+
+        if (lock->lock == 0) {
+                gf_log ("", GF_LOG_CRITICAL, "Unlock called  before lock ");
+                return -1;
+        }
+        curr = synctask_get ();
+        /*unlock should be called by lock owner
+         *i.e this will not allow the lock in nonsync task and unlock
+         * in sync task and vice-versa
+         */
+        switch (lock->type) {
+        case LOCK_TASK:
+                if (curr == lock->owner) {
+                        lock->lock--;
+                        gf_log ("", GF_LOG_TRACE, "Unlock success %p, remaining"
+                                " locks=%d", lock->owner, lock->lock);
+                } else {
+                        gf_log ("", GF_LOG_WARNING, "Unlock called by %p, but"
+                               " lock held by %p", curr, lock->owner);
+                }
 
-	curr = synctask_get ();
+                break;
+        case LOCK_THREAD:
+                if (pthread_self () == lock->owner_tid) {
+                        lock->lock--;
+                        gf_log ("", GF_LOG_TRACE, "Unlock success %u, remaining"
+                                " locks=%d", (unsigned int)lock->owner_tid, lock->lock);
+                } else {
+                        gf_log ("", GF_LOG_WARNING, "Unlock called by %u, but"
+                                " lock held by %u", (unsigned int) pthread_self(),
+                                 (unsigned int) lock->owner_tid);
+                }
 
-	if (lock->owner != curr) {
-		/* warn ? */
-	}
+                break;
+        default:
+                break;
+        }
 
-	lock->lock = _gf_false;
-
-	/* There could be both synctasks and non synctasks
-	   waiting (or none, or either). As a mid-approach
-	   between maintaining too many waiting counters
-	   at one extreme and a thundering herd on unlock
-	   at the other, call a cond_signal (which wakes
-	   one waiter) and first synctask waiter. So at
-	   most we have two threads waking up to grab the
-	   just released lock.
-	*/
-	pthread_cond_signal (&lock->cond);
-	if (!list_empty (&lock->waitq)) {
-		task = list_entry (lock->waitq.next, struct synctask, waitq);
+        if (lock->lock > 0) {
+                 return 0;
+        }
+        lock->type = LOCK_NULL;
+        lock->owner = NULL;
+        lock->owner_tid = 0;
+        lock->lock = 0;
+        /* There could be both synctasks and non synctasks
+           waiting (or none, or either). As a mid-approach
+           between maintaining too many waiting counters
+           at one extreme and a thundering herd on unlock
+           at the other, call a cond_signal (which wakes
+           one waiter) and first synctask waiter. So at
+           most we have two threads waking up to grab the
+           just released lock.
+        */
+        pthread_cond_signal (&lock->cond);
+        if (!list_empty (&lock->waitq)) {
+                task = list_entry (lock->waitq.next, struct synctask, waitq);
                 list_del_init (&task->waitq);
-		synctask_wake (task);
-	}
+                synctask_wake (task);
+        }
 
-	return 0;
+        return 0;
 }
 
 
diff --git a/libglusterfs/src/syncop.h b/libglusterfs/src/syncop.h
index e9d3428c512..4eebd56e5f8 100644
--- a/libglusterfs/src/syncop.h
+++ b/libglusterfs/src/syncop.h
@@ -114,12 +114,26 @@ struct syncenv {
 };
 
 
+typedef enum {
+        LOCK_NULL = 0,
+        LOCK_TASK,
+        LOCK_THREAD
+} lock_type_t;
+
+typedef enum {
+        SYNC_LOCK_DEFAULT = 0,
+        SYNC_LOCK_RECURSIVE,   /*it allows recursive locking*/
+} lock_attr_t;
+
 struct synclock {
-	pthread_mutex_t     guard; /* guard the remaining members, pair @cond */
-	pthread_cond_t      cond;  /* waiting non-synctasks */
-	struct list_head    waitq; /* waiting synctasks */
-	gf_boolean_t        lock;  /* _gf_true or _gf_false, lock status */
-	struct synctask    *owner; /* NULL if current owner is not a synctask */
+        pthread_mutex_t     guard; /* guard the remaining members, pair @cond */
+        pthread_cond_t      cond;  /* waiting non-synctasks */
+        struct list_head    waitq; /* waiting synctasks */
+        volatile int        lock;  /* true(non zero) or false(zero), lock status */
+        lock_attr_t         attr;
+        struct synctask    *owner; /* NULL if current owner is not a synctask */
+        pthread_t           owner_tid;
+        lock_type_t         type;
 };
 typedef struct synclock synclock_t;
 
@@ -336,7 +350,7 @@ syncop_create_frame (xlator_t *this)
 	return frame;
 }
 
-int synclock_init (synclock_t *lock);
+int synclock_init (synclock_t *lock, lock_attr_t attr);
 int synclock_destroy (synclock_t *lock);
 int synclock_lock (synclock_t *lock);
 int synclock_trylock (synclock_t *lock);