Whamcloud - gitweb
b=13872
[fs/lustre-release.git] / lustre / ldlm / ldlm_flock.c
index 7ef1341..6cb71e7 100644 (file)
 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
  * vim:expandtab:shiftwidth=8:tabstop=8:
  *
- *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
- *   Author: Peter Braam <braam@clusterfs.com>
- *   Author: Phil Schwan <phil@clusterfs.com>
+ *  Copyright (c) 2003 Hewlett-Packard Development Company LP.
+ *   Developed under the sponsorship of the US Government under
+ *   Subcontract No. B514193
  *
- *   This file is part of Lustre, http://www.lustre.org.
+ *   This file is part of the Lustre file system, http://www.lustre.org
+ *   Lustre is a trademark of Cluster File Systems, Inc.
  *
- *   Lustre is free software; you can redistribute it and/or
- *   modify it under the terms of version 2 of the GNU General Public
- *   License as published by the Free Software Foundation.
+ *   You may have signed or agreed to another license before downloading
+ *   this software.  If so, you are bound by the terms and conditions
+ *   of that agreement, and the following does not apply to you.  See the
+ *   LICENSE file included with this distribution for more information.
  *
- *   Lustre is distributed in the hope that it will be useful,
- *   but WITHOUT ANY WARRANTY; without even the implied warranty of
- *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- *   GNU General Public License for more details.
+ *   If you did not agree to a different license, then this copy of Lustre
+ *   is open source software; you can redistribute it and/or modify it
+ *   under the terms of version 2 of the GNU General Public License as
+ *   published by the Free Software Foundation.
  *
- *   You should have received a copy of the GNU General Public License
- *   along with Lustre; if not, write to the Free Software
- *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ *   In either case, Lustre is distributed in the hope that it will be
+ *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
+ *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ *   license text for more details.
  */
 
 #define DEBUG_SUBSYSTEM S_LDLM
 
 #ifdef __KERNEL__
-#include <linux/lustre_dlm.h>
-#include <linux/obd_support.h>
-#include <linux/obd_class.h>
-#include <linux/lustre_lib.h>
+#include <lustre_dlm.h>
+#include <obd_support.h>
+#include <obd_class.h>
+#include <lustre_lib.h>
+#include <libcfs/list.h>
 #else
 #include <liblustre.h>
+#include <obd_class.h>
 #endif
 
+#include "ldlm_internal.h"
+
+#define l_flock_waitq   l_lru
+
+static struct list_head ldlm_flock_waitq = CFS_LIST_HEAD_INIT(ldlm_flock_waitq);
+
+int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
+                            void *data, int flag);
+
+/**
+ * list_for_remaining_safe - iterate over the remaining entries in a list
+ *              and safeguard against removal of a list entry.
+ * @pos:        the &struct list_head to use as a loop counter. pos MUST
+ *              have been initialized prior to using it in this macro.
+ * @n:          another &struct list_head to use as temporary storage
+ * @head:       the head for your list.
+ */
+#define list_for_remaining_safe(pos, n, head) \
+        for (n = pos->next; pos != (head); pos = n, n = pos->next)
+
 static inline int
 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
 {
-        if ((new->l_data.l_flock.pid == lock->l_data.l_flock.pid) &&
-            (new->l_export == lock->l_export))
-                return 1;
-        else
-                return 0;
+        return((new->l_policy_data.l_flock.pid ==
+                lock->l_policy_data.l_flock.pid) &&
+               (new->l_export == lock->l_export));
 }
 
 static inline int
 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
 {
-        if ((new->l_data.l_flock.start <= lock->l_data.l_flock.end) &&
-            (new->l_data.l_flock.end >= lock->l_data.l_flock.start))
-                return 1;
-        else
-                return 0;
+        return((new->l_policy_data.l_flock.start <=
+                lock->l_policy_data.l_flock.end) &&
+               (new->l_policy_data.l_flock.end >=
+                lock->l_policy_data.l_flock.start));
 }
 
 static inline void
-ldlm_flock_destroy(struct ldlm_lock *lock, int flags)
+ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, int flags)
 {
         ENTRY;
 
+        LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: 0x%x)",
+                   mode, flags);
+
+        LASSERT(list_empty(&lock->l_flock_waitq));
+
         list_del_init(&lock->l_res_link);
         if (flags == LDLM_FL_WAIT_NOREPROC) {
-                /* client side */
-                struct lustre_handle lockh;
+                /* client side - set a flag to prevent sending a CANCEL */
+                lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
+                ldlm_lock_decref_internal(lock, mode);
+        }
+
+        ldlm_lock_destroy_nolock(lock);
+        EXIT;
+}
+
+static int
+ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *blocking_lock)
+{
+        struct obd_export *req_export = req->l_export;
+        struct obd_export *blocking_export = blocking_lock->l_export;
+        pid_t req_pid = req->l_policy_data.l_flock.pid;
+        pid_t blocking_pid = blocking_lock->l_policy_data.l_flock.pid;
+        struct ldlm_lock *lock;
+
+restart:
+        list_for_each_entry(lock, &ldlm_flock_waitq, l_flock_waitq) {
+                if ((lock->l_policy_data.l_flock.pid != blocking_pid) ||
+                    (lock->l_export != blocking_export))
+                        continue;
 
-                /* Set a flag to prevent us from sending a CANCEL */
-                lock->l_flags |= LDLM_FL_LOCAL_ONLY;
+                blocking_pid = lock->l_policy_data.l_flock.blocking_pid;
+                blocking_export = (struct obd_export *)(long)
+                        lock->l_policy_data.l_flock.blocking_export;
+                if (blocking_pid == req_pid && blocking_export == req_export)
+                        return 1;
 
-                ldlm_lock2handle(lock, &lockh);
-                ldlm_lock_decref_and_cancel(&lockh, lock->l_granted_mode);
+                goto restart;
         }
 
-        ldlm_lock_destroy(lock);
-        EXIT;
+        return 0;
 }
 
 int
-ldlm_flock_enqueue(struct ldlm_lock **reqp, void *req_cookie, int *flags,
-                   int first_enq, ldlm_error_t *err)
+ldlm_process_flock_lock(struct ldlm_lock *req, int *flags, int first_enq,
+                        ldlm_error_t *err, struct list_head *work_list)
 {
-        struct ldlm_lock *req = *reqp;
-        struct ldlm_lock *new = req;
-        struct ldlm_lock *new2 = NULL;
-        struct ldlm_lock *lock = NULL;
         struct ldlm_resource *res = req->l_resource;
         struct ldlm_namespace *ns = res->lr_namespace;
         struct list_head *tmp;
-        struct list_head *ownlocks;
+        struct list_head *ownlocks = NULL;
+        struct ldlm_lock *lock = NULL;
+        struct ldlm_lock *new = req;
+        struct ldlm_lock *new2 = NULL;
         ldlm_mode_t mode = req->l_req_mode;
-        int added = 0;
+        int local = ns_is_client(ns);
+        int added = (mode == LCK_NL);
         int overlaps = 0;
         ENTRY;
 
-        CDEBUG(D_FLOCK, "flags: 0x%x pid: %d mode: %d start: %llu end: %llu\n",
-               *flags, new->l_data.l_flock.pid, mode,
-               req->l_data.l_flock.start, req->l_data.l_flock.end);
+        CDEBUG(D_DLMTRACE, "flags %#x pid %u mode %u start "LPU64" end "LPU64
+               "\n", *flags, new->l_policy_data.l_flock.pid, mode,
+               req->l_policy_data.l_flock.start,
+               req->l_policy_data.l_flock.end);
 
         *err = ELDLM_OK;
 
-        /* No blocking ASTs are sent for record locks */
-        req->l_blocking_ast = NULL;
+        if (local) {
+                /* No blocking ASTs are sent to the clients for
+                 * Posix file & record locks */
+                req->l_blocking_ast = NULL;
+        } else {
+                /* Called on the server for lock cancels. */
+                req->l_blocking_ast = ldlm_flock_blocking_ast;
+        }
 
-        ownlocks = NULL;
-       if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
+        if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
+                /* This loop determines where this processes locks start
+                 * in the resource lr_granted list. */
                 list_for_each(tmp, &res->lr_granted) {
                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
-
                         if (ldlm_same_flock_owner(lock, req)) {
                                 ownlocks = tmp;
                                 break;
                         }
                 }
         } else {
+                lockmode_verify(mode);
+
+                /* This loop determines if there are existing locks
+                 * that conflict with the new lock request. */
                 list_for_each(tmp, &res->lr_granted) {
                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 
@@ -122,120 +182,143 @@ ldlm_flock_enqueue(struct ldlm_lock **reqp, void *req_cookie, int *flags,
                         /* locks are compatible, overlap doesn't matter */
                         if (lockmode_compat(lock->l_granted_mode, mode))
                                 continue;
-                        
+
                         if (!ldlm_flocks_overlap(lock, req))
                                 continue;
 
+                        if (!first_enq)
+                                RETURN(LDLM_ITER_CONTINUE);
+
                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
-                                ldlm_flock_destroy(req, *flags);
-                                *err = ELDLM_LOCK_ABORTED;
+                                ldlm_flock_destroy(req, mode, *flags);
+                                *err = -EAGAIN;
                                 RETURN(LDLM_ITER_STOP);
                         }
 
                         if (*flags & LDLM_FL_TEST_LOCK) {
-                                req->l_granted_mode = lock->l_granted_mode;
-                                req->l_data.l_flock.pid =
-                                        lock->l_data.l_flock.pid;
-                                req->l_data.l_flock.start =
-                                        lock->l_data.l_flock.start;
-                                req->l_data.l_flock.end =
-                                        lock->l_data.l_flock.end;
-                                ldlm_flock_destroy(req, *flags);
+                                ldlm_flock_destroy(req, mode, *flags);
+                                req->l_req_mode = lock->l_granted_mode;
+                                req->l_policy_data.l_flock.pid =
+                                        lock->l_policy_data.l_flock.pid;
+                                req->l_policy_data.l_flock.start =
+                                        lock->l_policy_data.l_flock.start;
+                                req->l_policy_data.l_flock.end =
+                                        lock->l_policy_data.l_flock.end;
+                                *flags |= LDLM_FL_LOCK_CHANGED;
                                 RETURN(LDLM_ITER_STOP);
                         }
 
-                        if (first_enq) {
-                                /* XXX - add deadlock detection check here */
+                        if (ldlm_flock_deadlock(req, lock)) {
+                                ldlm_flock_destroy(req, mode, *flags);
+                                *err = -EDEADLK;
+                                RETURN(LDLM_ITER_STOP);
                         }
 
+                        req->l_policy_data.l_flock.blocking_pid =
+                                lock->l_policy_data.l_flock.pid;
+                        req->l_policy_data.l_flock.blocking_export =
+                                (long)(void *)lock->l_export;
+
+                        LASSERT(list_empty(&req->l_flock_waitq));
+                        list_add_tail(&req->l_flock_waitq, &ldlm_flock_waitq);
+
+                        ldlm_resource_add_lock(res, &res->lr_waiting, req);
                         *flags |= LDLM_FL_BLOCK_GRANTED;
-                        RETURN(LDLM_ITER_CONTINUE);
+                        RETURN(LDLM_ITER_STOP);
                 }
         }
 
         if (*flags & LDLM_FL_TEST_LOCK) {
-                LASSERT(first_enq);
-                req->l_granted_mode = req->l_req_mode;
+                ldlm_flock_destroy(req, mode, *flags);
+                req->l_req_mode = LCK_NL;
+                *flags |= LDLM_FL_LOCK_CHANGED;
                 RETURN(LDLM_ITER_STOP);
         }
 
-        added = (mode == LCK_NL);
+        /* In case we had slept on this lock request take it off of the
+         * deadlock detection waitq. */
+        list_del_init(&req->l_flock_waitq);
 
-        /* Insert the new lock into the list */
+        /* Scan the locks owned by this process that overlap this request.
+         * We may have to merge or split existing locks. */
 
         if (!ownlocks)
                 ownlocks = &res->lr_granted;
 
-        for (tmp = ownlocks->next; ownlocks != &res->lr_granted;
-             ownlocks = tmp, tmp = ownlocks->next) {
+        list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 
                 if (!ldlm_same_flock_owner(lock, new))
                         break;
 
-               if (lock->l_granted_mode == mode) {
-                       if (lock->l_data.l_flock.end <
-                            (new->l_data.l_flock.start - 1))
-                               continue;
-
-                       if (lock->l_data.l_flock.start >
-                            (new->l_data.l_flock.end + 1))
-                               break;
-
-                       if (lock->l_data.l_flock.start >
-                            new->l_data.l_flock.start)
-                               lock->l_data.l_flock.start =
-                                        new->l_data.l_flock.start;
-                       else
-                               new->l_data.l_flock.start =
-                                        lock->l_data.l_flock.start;
-
-                       if (lock->l_data.l_flock.end <
-                            new->l_data.l_flock.end)
-                               lock->l_data.l_flock.end =
-                                        new->l_data.l_flock.end;
-                       else
-                               new->l_data.l_flock.end =
-                                        lock->l_data.l_flock.end;
-
-                       if (added) {
-                                ldlm_flock_destroy(lock, *flags);
-                       } else {
+                if (lock->l_granted_mode == mode) {
+                        /* If the modes are the same then we need to process
+                         * locks that overlap OR adjoin the new lock. The extra
+                         * logic condition is necessary to deal with arithmetic
+                         * overflow and underflow. */
+                        if ((new->l_policy_data.l_flock.start >
+                             (lock->l_policy_data.l_flock.end + 1))
+                            && (lock->l_policy_data.l_flock.end !=
+                                OBD_OBJECT_EOF))
+                                continue;
+
+                        if ((new->l_policy_data.l_flock.end <
+                             (lock->l_policy_data.l_flock.start - 1))
+                            && (lock->l_policy_data.l_flock.start != 0))
+                                break;
+
+                        if (new->l_policy_data.l_flock.start <
+                            lock->l_policy_data.l_flock.start) {
+                                lock->l_policy_data.l_flock.start =
+                                        new->l_policy_data.l_flock.start;
+                        } else {
+                                new->l_policy_data.l_flock.start =
+                                        lock->l_policy_data.l_flock.start;
+                        }
+
+                        if (new->l_policy_data.l_flock.end >
+                            lock->l_policy_data.l_flock.end) {
+                                lock->l_policy_data.l_flock.end =
+                                        new->l_policy_data.l_flock.end;
+                        } else {
+                                new->l_policy_data.l_flock.end =
+                                        lock->l_policy_data.l_flock.end;
+                        }
+
+                        if (added) {
+                                ldlm_flock_destroy(lock, mode, *flags);
+                        } else {
                                 new = lock;
                                 added = 1;
                         }
                         continue;
-               }
+                }
 
-                if (lock->l_data.l_flock.end < new->l_data.l_flock.start)
+                if (new->l_policy_data.l_flock.start >
+                    lock->l_policy_data.l_flock.end)
                         continue;
-                if (lock->l_data.l_flock.start > new->l_data.l_flock.end)
+
+                if (new->l_policy_data.l_flock.end <
+                    lock->l_policy_data.l_flock.start)
                         break;
 
                 ++overlaps;
 
-                if (new->l_data.l_flock.start <=
-                    lock->l_data.l_flock.start) {
-                        if (new->l_data.l_flock.end <
-                            lock->l_data.l_flock.end) {
-                                lock->l_data.l_flock.start =
-                                        new->l_data.l_flock.end + 1;
+                if (new->l_policy_data.l_flock.start <=
+                    lock->l_policy_data.l_flock.start) {
+                        if (new->l_policy_data.l_flock.end <
+                            lock->l_policy_data.l_flock.end) {
+                                lock->l_policy_data.l_flock.start =
+                                        new->l_policy_data.l_flock.end + 1;
                                 break;
-                        } else if (added) {
-                                ldlm_flock_destroy(lock, *flags);
-                        } else {
-                                lock->l_data.l_flock.start =
-                                        new->l_data.l_flock.start;
-                                lock->l_data.l_flock.end =
-                                        new->l_data.l_flock.end;
-                                new = lock;
-                                added = 1;
                         }
+                        ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
                         continue;
                 }
-                if (new->l_data.l_flock.end >= lock->l_data.l_flock.end) {
-                        lock->l_data.l_flock.end =
-                                new->l_data.l_flock.start - 1;
+                if (new->l_policy_data.l_flock.end >=
+                    lock->l_policy_data.l_flock.end) {
+                        lock->l_policy_data.l_flock.end =
+                                new->l_policy_data.l_flock.start - 1;
                         continue;
                 }
 
@@ -245,86 +328,131 @@ ldlm_flock_enqueue(struct ldlm_lock **reqp, void *req_cookie, int *flags,
                  * allocating a new lock and use the req lock passed in
                  * with the request but this would complicate the reply
                  * processing since updates to req get reflected in the
-                 * reply. The client side must see the original lock data
-                 * so that it can process the unlock properly. */
-
-                /* XXX - if ldlm_lock_new() can sleep we have to
-                 * release the ns_lock, allocate the new lock, and
-                 * restart processing this lock. */
-                new2 = ldlm_lock_create(ns, NULL, res->lr_name, LDLM_FLOCK,
-                                        lock->l_granted_mode, NULL, NULL);
+                 * reply. The client side replays the lock request so
+                 * it must see the original lock data in the reply. */
+
+                /* XXX - if ldlm_lock_new() can sleep we should
+                 * release the ns_lock, allocate the new lock,
+                 * and restart processing this lock. */
+                new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
+                                        lock->l_granted_mode, NULL, NULL, NULL,
+                                        NULL, 0);
                 if (!new2) {
-                        /* LBUG for now */
-                        LASSERT(0);
-                        RETURN(ENOMEM);
+                        ldlm_flock_destroy(req, lock->l_granted_mode, *flags);
+                        *err = -ENOLCK;
+                        RETURN(LDLM_ITER_STOP);
                 }
 
                 new2->l_granted_mode = lock->l_granted_mode;
-                new2->l_data.l_flock.pid = new->l_data.l_flock.pid;
-                new2->l_data.l_flock.start = lock->l_data.l_flock.start;
-                new2->l_data.l_flock.end = new->l_data.l_flock.start - 1;
-                lock->l_data.l_flock.start = new->l_data.l_flock.end + 1;
-                new2->l_connh = lock->l_connh;
-                if ((new2->l_export = lock->l_export) != NULL) {
+                new2->l_policy_data.l_flock.pid =
+                        new->l_policy_data.l_flock.pid;
+                new2->l_policy_data.l_flock.start =
+                        lock->l_policy_data.l_flock.start;
+                new2->l_policy_data.l_flock.end =
+                        new->l_policy_data.l_flock.start - 1;
+                lock->l_policy_data.l_flock.start =
+                        new->l_policy_data.l_flock.end + 1;
+                new2->l_conn_export = lock->l_conn_export;
+                if (lock->l_export != NULL) {
+                        new2->l_export = class_export_get(lock->l_export);
+                        spin_lock(&new2->l_export->exp_ldlm_data.led_lock);
                         list_add(&new2->l_export_chain,
-                                 &new2->l_export->
-                                 exp_ldlm_data.led_held_locks);
+                                 &new2->l_export->exp_ldlm_data.led_held_locks);
+                        spin_unlock(&new2->l_export->exp_ldlm_data.led_lock);
                 }
-                if (*flags == LDLM_FL_WAIT_NOREPROC) {
-                        /* client side */
+                if (*flags == LDLM_FL_WAIT_NOREPROC)
                         ldlm_lock_addref_internal(new2, lock->l_granted_mode);
-                }
 
                 /* insert new2 at lock */
-                list_add_tail(&new2->l_res_link, ownlocks);
+                ldlm_resource_add_lock(res, ownlocks, new2);
                 LDLM_LOCK_PUT(new2);
                 break;
         }
 
-        if (added) {
-                ldlm_flock_destroy(req, *flags);
-        } else {
-                /* insert new at ownlocks */
-                new->l_granted_mode = new->l_req_mode;
-                list_del_init(&new->l_res_link);
-                list_add_tail(&new->l_res_link, ownlocks);
-        }
+        /* At this point we're granting the lock request. */
+        req->l_granted_mode = req->l_req_mode;
 
-       if (*flags != LDLM_FL_WAIT_NOREPROC) {
-                if (req->l_completion_ast)
-                        ldlm_add_ast_work_item(req, NULL, NULL, 0);
-
-                /* The only problem with doing the reprocessing here is that
-                 * the completion ASTs for newly granted locks will be sent
-                 * before the unlock completion is sent. It shouldn't be an
-                 * issue. Also note that ldlm_flock_enqueue() will recurse,
-                 * but only once because there can't be unlock requests on
-                 * the wait queue. */
-                if ((mode == LCK_NL) && overlaps)
-                        ldlm_reprocess_queue(res, &res->lr_waiting);
+        /* Add req to the granted queue before calling ldlm_reprocess_all(). */
+        if (!added) {
+                list_del_init(&req->l_res_link);
+                /* insert new lock before ownlocks in list. */
+                ldlm_resource_add_lock(res, ownlocks, req);
         }
 
-        ldlm_resource_dump(res);
+        if (*flags != LDLM_FL_WAIT_NOREPROC) {
+                if (first_enq) {
+                        /* If this is an unlock, reprocess the waitq and
+                         * send completions ASTs for locks that can now be
+                         * granted. The only problem with doing this
+                         * reprocessing here is that the completion ASTs for
+                         * newly granted locks will be sent before the unlock
+                         * completion is sent. It shouldn't be an issue. Also
+                         * note that ldlm_process_flock_lock() will recurse,
+                         * but only once because first_enq will be false from
+                         * ldlm_reprocess_queue. */
+                        if ((mode == LCK_NL) && overlaps) {
+                                struct list_head rpc_list
+                                                    = CFS_LIST_HEAD_INIT(rpc_list);
+                                int rc;
+restart:
+                                ldlm_reprocess_queue(res, &res->lr_waiting,
+                                                     &rpc_list);
+
+                                unlock_res(res);
+                                rc = ldlm_run_bl_ast_work(&rpc_list);
+                                lock_res(res);
+                                if (rc == -ERESTART)
+                                        GOTO(restart, -ERESTART);
+                       }
+                } else {
+                        LASSERT(req->l_completion_ast);
+                        ldlm_add_ast_work_item(req, NULL, work_list);
+                }
+        }
 
-       RETURN(LDLM_ITER_CONTINUE);
-}
+        /* In case we're reprocessing the requested lock we can't destroy
+         * it until after calling ldlm_ast_work_item() above so that lawi()
+         * can bump the reference count on req. Otherwise req could be freed
+         * before the completion AST can be sent.  */
+        if (added)
+                ldlm_flock_destroy(req, mode, *flags);
 
-static void interrupted_flock_completion_wait(void *data)
-{
+        ldlm_resource_dump(D_OTHER, res);
+        RETURN(LDLM_ITER_CONTINUE);
 }
 
-struct flock_wait_data {
+struct ldlm_flock_wait_data {
         struct ldlm_lock *fwd_lock;
         int               fwd_generation;
 };
 
+static void
+ldlm_flock_interrupted_wait(void *data)
+{
+        struct ldlm_lock *lock;
+        struct lustre_handle lockh;
+        ENTRY;
+
+        lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
+
+        /* take lock off the deadlock detection waitq. */
+        list_del_init(&lock->l_flock_waitq);
+
+        /* client side - set flag to prevent lock from being put on lru list */
+        lock->l_flags |= LDLM_FL_CBPENDING;
+
+        ldlm_lock_decref_internal(lock, lock->l_req_mode);
+        ldlm_lock2handle(lock, &lockh);
+        ldlm_cli_cancel(&lockh);
+        EXIT;
+}
+
 int
 ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 {
         struct ldlm_namespace *ns;
-        struct file_lock *getlk = data;
-        struct flock_wait_data fwd;
-        unsigned long irqflags;
+        cfs_flock_t *getlk = lock->l_ast_data;
+        struct ldlm_flock_wait_data fwd;
         struct obd_device *obd;
         struct obd_import *imp = NULL;
         ldlm_error_t err;
@@ -332,12 +460,10 @@ ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
         struct l_wait_info lwi;
         ENTRY;
 
-        LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
+        CDEBUG(D_DLMTRACE, "flags: 0x%x data: %p getlk: %p\n",
+               flags, data, getlk);
 
-        if (flags == 0) {
-                wake_up(&lock->l_waitq);
-                RETURN(0);
-        }
+        LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 
         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
                        LDLM_FL_BLOCK_CONV)))
@@ -346,73 +472,87 @@ ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
                    "sleeping");
 
-        ldlm_lock_dump(D_OTHER, lock);
-
         fwd.fwd_lock = lock;
-        obd = class_conn2obd(lock->l_connh);
+        obd = class_exp2obd(lock->l_conn_export);
 
         /* if this is a local lock, then there is no import */
         if (obd != NULL)
                 imp = obd->u.cli.cl_import;
 
         if (imp != NULL) {
-                spin_lock_irqsave(&imp->imp_lock, irqflags);
+                spin_lock(&imp->imp_lock);
                 fwd.fwd_generation = imp->imp_generation;
-                spin_unlock_irqrestore(&imp->imp_lock, irqflags);
+                spin_unlock(&imp->imp_lock);
         }
 
-        lwi = LWI_TIMEOUT_INTR(0, NULL, interrupted_flock_completion_wait,
-                               &fwd);
+        lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd);
 
         /* Go to sleep until the lock is granted. */
         rc = l_wait_event(lock->l_waitq,
                           ((lock->l_req_mode == lock->l_granted_mode) ||
                            lock->l_destroyed), &lwi);
 
-        LASSERT(!(lock->l_destroyed));
-
-        if (rc) {
-                LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
-                           rc);
-                RETURN(rc);
-        }
+        LDLM_DEBUG(lock, "client-side enqueue waking up: rc = %d", rc);
+        RETURN(rc);
 
 granted:
 
-        LDLM_DEBUG(lock, "client-side enqueue waking up");
+        LDLM_DEBUG(lock, "client-side enqueue granted");
         ns = lock->l_resource->lr_namespace;
-        l_lock(&ns->ns_lock);
+        lock_res(lock->l_resource);
+
+        /* take lock off the deadlock detection waitq. */
+        list_del_init(&lock->l_flock_waitq);
 
         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
         list_del_init(&lock->l_res_link);
 
-        if (getlk) {
+        if (flags & LDLM_FL_TEST_LOCK) {
                 /* fcntl(F_GETLK) request */
-                if (lock->l_granted_mode == LCK_PR)
-                        getlk->fl_type = F_RDLCK;
-                else if (lock->l_granted_mode == LCK_PW)
-                        getlk->fl_type = F_WRLCK;
-                else
-                        getlk->fl_type = F_UNLCK;
-                getlk->fl_pid = lock->l_data.l_flock.pid;
-                getlk->fl_start = lock->l_data.l_flock.start;
-                getlk->fl_end = lock->l_data.l_flock.end;
-                /* ldlm_flock_destroy(lock); */
+                /* The old mode was saved in getlk->fl_type so that if the mode
+                 * in the lock changes we can decref the approprate refcount. */
+                ldlm_flock_destroy(lock, cfs_flock_type(getlk), LDLM_FL_WAIT_NOREPROC);
+                switch (lock->l_granted_mode) {
+                case LCK_PR:
+                        cfs_flock_set_type(getlk, F_RDLCK);
+                        break;
+                case LCK_PW:
+                        cfs_flock_set_type(getlk, F_WRLCK);
+                        break;
+                default:
+                        cfs_flock_set_type(getlk, F_UNLCK);
+                }
+                cfs_flock_set_pid(getlk, (pid_t)lock->l_policy_data.l_flock.pid);
+                cfs_flock_set_start(getlk, (loff_t)lock->l_policy_data.l_flock.start);
+                cfs_flock_set_end(getlk, (loff_t)lock->l_policy_data.l_flock.end);
         } else {
-                flags = LDLM_FL_WAIT_NOREPROC;
-                /* We need to reprocess the lock to do merges or split */
-                ldlm_flock_enqueue(&lock, NULL, &flags, 1, &err);
+                int noreproc = LDLM_FL_WAIT_NOREPROC;
+
+                /* We need to reprocess the lock to do merges or splits
+                 * with existing locks owned by this process. */
+                ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
+                if (flags == 0)
+                        cfs_waitq_signal(&lock->l_waitq);
         }
-        l_unlock(&ns->ns_lock);
+        unlock_res(lock->l_resource);
         RETURN(0);
 }
+EXPORT_SYMBOL(ldlm_flock_completion_ast);
 
-/* This function is only called on the client when a lock is aborted. */
-int
-ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *ld,
-                        void *data, int flag)
+int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
+                            void *data, int flag)
 {
+        struct ldlm_namespace *ns;
         ENTRY;
-        ldlm_lock_destroy(lock);
+
+        LASSERT(lock);
+        LASSERT(flag == LDLM_CB_CANCELING);
+
+        ns = lock->l_resource->lr_namespace;
+
+        /* take lock off the deadlock detection waitq. */
+        lock_res_and_lock(lock);
+        list_del_init(&lock->l_flock_waitq);
+        unlock_res_and_lock(lock);
         RETURN(0);
 }