branch: HEAD

[fs/lustre-release.git] / lustre / ldlm / ldlm_extent.c

diff --git a/lustre/ldlm/ldlm_extent.c b/lustre/ldlm/ldlm_extent.c
index 5825ca0..03172d6 100644 (file)
--- a/lustre/ldlm/ldlm_extent.c
+++ b/lustre/ldlm/ldlm_extent.c
@@ -49,6 +49,7 @@
 #include <lustre_dlm.h>
 #include <obd_support.h>
 #include <obd.h>
+#include <obd_class.h>
 #include <lustre_lib.h>
 
 #include "ldlm_internal.h"
@@ -64,7 +65,7 @@ static void ldlm_extent_internal_policy_fixup(struct ldlm_lock *req,
         __u64 req_start = req->l_req_extent.start;
         __u64 req_end = req->l_req_extent.end;
         __u64 req_align, mask;
- 
+
         if (conflicting > 32 && (req_mode == LCK_PW || req_mode == LCK_CW)) {
                 if (req_end < req_start + LDLM_MAX_GROWN_EXTENT)
                         new_ex->end = min(req_start + LDLM_MAX_GROWN_EXTENT,
@@ -216,7 +217,7 @@ ldlm_extent_internal_policy_waiting(struct ldlm_lock *req,
                         continue;
 
                 /* We grow extents downwards only as far as they don't overlap
-                 * with already-granted locks, on the assumtion that clients
+                 * with already-granted locks, on the assumption that clients
                  * will be writing beyond the initial requested end and would
                  * then need to enqueue a new lock beyond previous request.
                  * l_req_extent->end strictly < req_start, checked above. */
@@ -284,6 +285,9 @@ static int ldlm_check_contention(struct ldlm_lock *lock, int contended_locks)
         struct ldlm_resource *res = lock->l_resource;
         cfs_time_t now = cfs_time_current();
 
+        if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_SET_CONTENTION))
+                return 1;
+
         CDEBUG(D_DLMTRACE, "contended locks = %d\n", contended_locks);
         if (contended_locks > res->lr_namespace->ns_contended_locks)
                 res->lr_contention_time = now;
@@ -620,9 +624,9 @@ static void discard_bl_list(struct list_head *bl_list)
                 lock->l_flags &= ~LDLM_FL_AST_SENT;
                 LASSERT(lock->l_bl_ast_run == 0);
                 LASSERT(lock->l_blocking_lock);
-                LDLM_LOCK_PUT(lock->l_blocking_lock);
+                LDLM_LOCK_RELEASE(lock->l_blocking_lock);
                 lock->l_blocking_lock = NULL;
-                LDLM_LOCK_PUT(lock);
+                LDLM_LOCK_RELEASE(lock);
         }
         EXIT;
 }
@@ -704,16 +708,31 @@ int ldlm_process_extent_lock(struct ldlm_lock *lock, int *flags, int first_enq,
                         ldlm_resource_add_lock(res, &res->lr_waiting, lock);
                 unlock_res(res);
                 rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_BL_AST);
-                lock_res(res);
 
+                if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
+                    !ns_is_client(res->lr_namespace))
+                        class_fail_export(lock->l_export);
+ 
+                lock_res(res);
                 if (rc == -ERESTART) {
+
+                        /* 15715: The lock was granted and destroyed after
+                         * resource lock was dropped. Interval node was freed
+                         * in ldlm_lock_destroy. Anyway, this always happens
+                         * when a client is being evicted. So it would be
+                         * ok to return an error. -jay */
+                        if (lock->l_destroyed) {
+                                *err = -EAGAIN;
+                                GOTO(out, rc = -EAGAIN);
+                        }
+
                         /* lock was granted while resource was unlocked. */
                         if (lock->l_granted_mode == lock->l_req_mode) {
                                 /* bug 11300: if the lock has been granted,
                                  * break earlier because otherwise, we will go
                                  * to restart and ldlm_resource_unlink will be
                                  * called and it causes the interval node to be
-                                 * freed. Then we will fail at 
+                                 * freed. Then we will fail at
                                  * ldlm_extent_add_lock() */
                                 *flags &= ~(LDLM_FL_BLOCK_GRANTED | LDLM_FL_BLOCK_CONV |
                                             LDLM_FL_BLOCK_WAIT);
@@ -782,7 +801,7 @@ struct ldlm_interval *ldlm_interval_alloc(struct ldlm_lock *lock)
         ENTRY;
 
         LASSERT(lock->l_resource->lr_type == LDLM_EXTENT);
-        OBD_SLAB_ALLOC(node, ldlm_interval_slab, CFS_ALLOC_IO, sizeof(*node));
+        OBD_SLAB_ALLOC_PTR_GFP(node, ldlm_interval_slab, CFS_ALLOC_IO);
         if (node == NULL)
                 RETURN(NULL);
 
@@ -795,6 +814,7 @@ void ldlm_interval_free(struct ldlm_interval *node)
 {
         if (node) {
                 LASSERT(list_empty(&node->li_group));
+                LASSERT(!interval_is_intree(&node->li_node));
                 OBD_SLAB_FREE(node, ldlm_interval_slab, sizeof(*node));
         }
 }
@@ -847,6 +867,7 @@ void ldlm_extent_add_lock(struct ldlm_resource *res,
 
         node = lock->l_tree_node;
         LASSERT(node != NULL);
+        LASSERT(!interval_is_intree(&node->li_node));
 
         idx = lock_mode_to_index(lock->l_granted_mode);
         LASSERT(lock->l_granted_mode == 1 << idx);
@@ -874,14 +895,13 @@ void ldlm_extent_add_lock(struct ldlm_resource *res,
 void ldlm_extent_unlink_lock(struct ldlm_lock *lock)
 {
         struct ldlm_resource *res = lock->l_resource;
-        struct ldlm_interval *node;
+        struct ldlm_interval *node = lock->l_tree_node;
         struct ldlm_interval_tree *tree;
         int idx;
 
-        if (lock->l_granted_mode != lock->l_req_mode)
+        if (!node || !interval_is_intree(&node->li_node)) /* duplicate unlink */
                 return;
 
-        LASSERT(lock->l_tree_node != NULL);
         idx = lock_mode_to_index(lock->l_granted_mode);
         LASSERT(lock->l_granted_mode == 1 << idx);
         tree = &res->lr_itree[idx];