Whamcloud - gitweb
b=20748
[fs/lustre-release.git] / lustre / ldlm / ldlm_flock.c
index 273c50c..b3b2cb1 100644 (file)
@@ -1,27 +1,41 @@
 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
  * vim:expandtab:shiftwidth=8:tabstop=8:
  *
- *  Copyright (c) 2003 Hewlett-Packard Development Company LP.
- *   Developed under the sponsorship of the US Government under
- *   Subcontract No. B514193
+ * GPL HEADER START
  *
- *   This file is part of the Lustre file system, http://www.lustre.org
- *   Lustre is a trademark of Cluster File Systems, Inc.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
- *   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.
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
  *
- *   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.
+ * This program 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 version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
  *
- *   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.
+ * You should have received a copy of the GNU General Public License
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
+ *
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ * GPL HEADER END
+ */
+/*
+ * Copyright  2008 Sun Microsystems, Inc. All rights reserved
+ * Use is subject to license terms.
+ *
+ * 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/
+ * Lustre is a trademark of Sun Microsystems, Inc.
  */
 
 #define DEBUG_SUBSYSTEM S_LDLM
 
 #define l_flock_waitq   l_lru
 
+/**
+ * Wait queue for Posix lock deadlock detection, added with
+ * ldlm_lock::l_flock_waitq.
+ */
 static CFS_LIST_HEAD(ldlm_flock_waitq);
+/**
+ * Lock protecting access to ldlm_flock_waitq.
+ */
 spinlock_t ldlm_flock_waitq_lock = SPIN_LOCK_UNLOCKED;
 
 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
@@ -50,10 +71,10 @@ int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 /**
  * 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
+ * \param 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.
+ * \param n     another &struct list_head to use as temporary storage
+ * \param 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)
@@ -87,10 +108,14 @@ ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, int flags)
         LASSERT(list_empty(&lock->l_flock_waitq));
 
         list_del_init(&lock->l_res_link);
-        if (flags == LDLM_FL_WAIT_NOREPROC) {
+        if (flags == LDLM_FL_WAIT_NOREPROC &&
+            !(lock->l_flags & LDLM_FL_FAILED)) {
                 /* 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);
+
+                /* when reaching here, it is under lock_res_and_lock(). Thus,
+                   need call the nolock version of ldlm_lock_decref_internal*/
+                ldlm_lock_decref_internal_nolock(lock, mode);
         }
 
         ldlm_lock_destroy_nolock(lock);
@@ -143,6 +168,8 @@ ldlm_process_flock_lock(struct ldlm_lock *req, int *flags, int first_enq,
         int local = ns_is_client(ns);
         int added = (mode == LCK_NL);
         int overlaps = 0;
+        int splitted = 0;
+        const struct ldlm_callback_suite null_cbs = { NULL };
         ENTRY;
 
         CDEBUG(D_DLMTRACE, "flags %#x pid %u mode %u start "LPU64" end "LPU64
@@ -161,6 +188,7 @@ ldlm_process_flock_lock(struct ldlm_lock *req, int *flags, int first_enq,
                 req->l_blocking_ast = ldlm_flock_blocking_ast;
         }
 
+reprocess:
         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
                 /* This loop determines where this processes locks start
                  * in the resource lr_granted list. */
@@ -344,15 +372,23 @@ ldlm_process_flock_lock(struct ldlm_lock *req, int *flags, int first_enq,
                 /* 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) {
-                        ldlm_flock_destroy(req, lock->l_granted_mode, *flags);
-                        *err = -ENOLCK;
-                        RETURN(LDLM_ITER_STOP);
+                        unlock_res_and_lock(req);
+                         new2 = ldlm_lock_create(ns, &res->lr_name, LDLM_FLOCK,
+                                        lock->l_granted_mode, &null_cbs,
+                                        NULL, 0);
+                        lock_res_and_lock(req);
+                        if (!new2) {
+                                ldlm_flock_destroy(req, lock->l_granted_mode,
+                                                   *flags);
+                                *err = -ENOLCK;
+                                RETURN(LDLM_ITER_STOP);
+                        }
+                        goto reprocess;
                 }
 
+                splitted = 1;
+
                 new2->l_granted_mode = lock->l_granted_mode;
                 new2->l_policy_data.l_flock.pid =
                         new->l_policy_data.l_flock.pid;
@@ -364,21 +400,27 @@ ldlm_process_flock_lock(struct ldlm_lock *req, int *flags, int first_enq,
                         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);
-                        spin_unlock(&new2->l_export->exp_ldlm_data.led_lock);
+                        new2->l_export = class_export_lock_get(lock->l_export);
+                        if (new2->l_export->exp_lock_hash &&
+                            hlist_unhashed(&new2->l_exp_hash))
+                                lustre_hash_add(new2->l_export->exp_lock_hash,
+                                                &new2->l_remote_handle,
+                                                &new2->l_exp_hash);
                 }
                 if (*flags == LDLM_FL_WAIT_NOREPROC)
-                        ldlm_lock_addref_internal(new2, lock->l_granted_mode);
+                        ldlm_lock_addref_internal_nolock(new2,
+                                                         lock->l_granted_mode);
 
                 /* insert new2 at lock */
                 ldlm_resource_add_lock(res, ownlocks, new2);
-                LDLM_LOCK_PUT(new2);
+                LDLM_LOCK_RELEASE(new2);
                 break;
         }
 
+        /* if new2 is created but never used, destroy it*/
+        if (splitted == 0 && new2 != NULL)
+                ldlm_lock_destroy_nolock(new2);
+
         /* At this point we're granting the lock request. */
         req->l_granted_mode = req->l_req_mode;
 
@@ -407,9 +449,10 @@ restart:
                                 ldlm_reprocess_queue(res, &res->lr_waiting,
                                                      &rpc_list);
 
-                                unlock_res(res);
-                                rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_BL_AST);
-                                lock_res(res);
+                                unlock_res_and_lock(req);
+                                rc = ldlm_run_ast_work(&rpc_list,
+                                                       LDLM_WORK_CP_AST);
+                                lock_res_and_lock(req);
                                 if (rc == -ERESTART)
                                         GOTO(restart, -ERESTART);
                        }
@@ -426,7 +469,7 @@ restart:
         if (added)
                 ldlm_flock_destroy(req, mode, *flags);
 
-        ldlm_resource_dump(D_OTHER, res);
+        ldlm_resource_dump(D_INFO, res);
         RETURN(LDLM_ITER_CONTINUE);
 }
 
@@ -462,17 +505,26 @@ ldlm_flock_interrupted_wait(void *data)
         EXIT;
 }
 
+/**
+ * Flock completion calback function.
+ *
+ * \param lock [in,out]: A lock to be handled
+ * \param flags    [in]: flags
+ * \param *data    [in]: ldlm_work_cp_ast_lock() will use ldlm_cb_set_arg
+ *
+ * \retval 0    : success
+ * \retval <0   : failure
+ */
 int
 ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 {
-        struct ldlm_namespace *ns;
-        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;
-        int rc = 0;
-        struct l_wait_info lwi;
+        cfs_flock_t                    *getlk = lock->l_ast_data;
+        struct obd_device              *obd;
+        struct obd_import              *imp = NULL;
+        struct ldlm_flock_wait_data     fwd;
+        struct l_wait_info              lwi;
+        ldlm_error_t                    err;
+        int                             rc = 0;
         ENTRY;
 
         CDEBUG(D_DLMTRACE, "flags: 0x%x data: %p getlk: %p\n",
@@ -483,11 +535,12 @@ ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
          * holding the lock even if app still believes it has it, since
          * server already dropped it anyway. Only for granted locks too. */
         lock_res_and_lock(lock);
-        if ((lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) == 
+        if ((lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) ==
             (LDLM_FL_FAILED|LDLM_FL_LOCAL_ONLY)) {
                 unlock_res_and_lock(lock);
                 if (lock->l_req_mode == lock->l_granted_mode &&
-                    lock->l_granted_mode != LCK_NL)
+                    lock->l_granted_mode != LCK_NL &&
+                    NULL == data)
                         ldlm_lock_decref_internal(lock, lock->l_req_mode);
                 RETURN(0);
         }
@@ -496,20 +549,25 @@ ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 
         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
-                       LDLM_FL_BLOCK_CONV)))
-                goto  granted;
+                       LDLM_FL_BLOCK_CONV))) {
+                if (NULL == data)
+                        /* mds granted the lock in the reply */
+                        goto granted;
+                /* CP AST RPC: lock get granted, wake it up */
+                cfs_waitq_signal(&lock->l_waitq);
+                RETURN(0);
+        }
 
         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
                    "sleeping");
-
         fwd.fwd_lock = lock;
         obd = class_exp2obd(lock->l_conn_export);
 
-        /* if this is a local lock, then there is no import */
-        if (obd != NULL)
+        /* if this is a local lock, there is no import */
+        if (NULL != obd)
                 imp = obd->u.cli.cl_import;
 
-        if (imp != NULL) {
+        if (NULL != imp) {
                 spin_lock(&imp->imp_lock);
                 fwd.fwd_generation = imp->imp_generation;
                 spin_unlock(&imp->imp_lock);
@@ -518,18 +576,31 @@ ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
         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);
+        rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
 
-        LDLM_DEBUG(lock, "client-side enqueue waking up: rc = %d", rc);
-        RETURN(rc);
+        if (rc) {
+                LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
+                           rc);
+                RETURN(rc);
+        }
 
 granted:
+        OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CP_CB_WAIT, 10);
+
+        lock_res_and_lock(lock);
+        if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
+                LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
+                unlock_res(lock->l_resource);
+                RETURN(-EIO);
+        }
+        if (rc) {
+                LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
+                           rc);
+                unlock_res_and_lock(lock);
+                RETURN(rc);
+        }
 
         LDLM_DEBUG(lock, "client-side enqueue granted");
-        ns = lock->l_resource->lr_namespace;
-        lock_res(lock->l_resource);
 
         /* take lock off the deadlock detection waitq. */
         spin_lock(&ldlm_flock_waitq_lock);
@@ -542,8 +613,9 @@ granted:
         if (flags & LDLM_FL_TEST_LOCK) {
                 /* fcntl(F_GETLK) request */
                 /* 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);
+                 * in the lock changes we can decref the appropriate 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);
@@ -554,19 +626,20 @@ granted:
                 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);
+                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 {
                 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);
         }
-        unlock_res(lock->l_resource);
+        unlock_res_and_lock(lock);
         RETURN(0);
 }
 EXPORT_SYMBOL(ldlm_flock_completion_ast);