4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_lockd.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LDLM
44 #include <libcfs/libcfs.h>
45 #include <lustre_dlm.h>
46 #include <obd_class.h>
47 #include <libcfs/list.h>
48 #include "ldlm_internal.h"
50 static int ldlm_num_threads;
51 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
52 "number of DLM service threads to start");
54 static char *ldlm_cpts;
55 CFS_MODULE_PARM(ldlm_cpts, "s", charp, 0444,
56 "CPU partitions ldlm threads should run on");
58 extern struct kmem_cache *ldlm_resource_slab;
59 extern struct kmem_cache *ldlm_lock_slab;
60 static struct mutex ldlm_ref_mutex;
61 static int ldlm_refcount;
63 struct ldlm_cb_async_args {
64 struct ldlm_cb_set_arg *ca_set_arg;
65 struct ldlm_lock *ca_lock;
70 static struct ldlm_state *ldlm_state;
72 inline cfs_time_t round_timeout(cfs_time_t timeout)
74 return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
77 /* timeout for initial callback (AST) reply (bz10399) */
78 static inline unsigned int ldlm_get_rq_timeout(void)
81 unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
83 return timeout < 1 ? 1 : timeout;
88 #define ELT_TERMINATE 2
94 * blp_prio_list is used for callbacks that should be handled
95 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
98 struct list_head blp_prio_list;
101 * blp_list is used for all other callbacks which are likely
102 * to take longer to process.
104 struct list_head blp_list;
106 wait_queue_head_t blp_waitq;
107 struct completion blp_comp;
108 atomic_t blp_num_threads;
109 atomic_t blp_busy_threads;
114 struct ldlm_bl_work_item {
115 struct list_head blwi_entry;
116 struct ldlm_namespace *blwi_ns;
117 struct ldlm_lock_desc blwi_ld;
118 struct ldlm_lock *blwi_lock;
119 struct list_head blwi_head;
121 struct completion blwi_comp;
122 ldlm_cancel_flags_t blwi_flags;
123 int blwi_mem_pressure;
126 #ifdef HAVE_SERVER_SUPPORT
129 * Protects both waiting_locks_list and expired_lock_thread.
131 static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */
134 * List for contended locks.
136 * As soon as a lock is contended, it gets placed on this list and
137 * expected time to get a response is filled in the lock. A special
138 * thread walks the list looking for locks that should be released and
139 * schedules client evictions for those that have not been released in
142 * All access to it should be under waiting_locks_spinlock.
144 static struct list_head waiting_locks_list;
145 static struct timer_list waiting_locks_timer;
147 static struct expired_lock_thread {
148 wait_queue_head_t elt_waitq;
151 struct list_head elt_expired_locks;
152 } expired_lock_thread;
154 static inline int have_expired_locks(void)
159 spin_lock_bh(&waiting_locks_spinlock);
160 need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
161 spin_unlock_bh(&waiting_locks_spinlock);
167 * Check expired lock list for expired locks and time them out.
169 static int expired_lock_main(void *arg)
171 struct list_head *expired = &expired_lock_thread.elt_expired_locks;
172 struct l_wait_info lwi = { 0 };
177 expired_lock_thread.elt_state = ELT_READY;
178 wake_up(&expired_lock_thread.elt_waitq);
181 l_wait_event(expired_lock_thread.elt_waitq,
182 have_expired_locks() ||
183 expired_lock_thread.elt_state == ELT_TERMINATE,
186 spin_lock_bh(&waiting_locks_spinlock);
187 if (expired_lock_thread.elt_dump) {
188 struct libcfs_debug_msg_data msgdata = {
189 .msg_file = __FILE__,
190 .msg_fn = "waiting_locks_callback",
191 .msg_line = expired_lock_thread.elt_dump };
192 spin_unlock_bh(&waiting_locks_spinlock);
194 /* from waiting_locks_callback, but not in timer */
195 libcfs_debug_dumplog();
196 libcfs_run_lbug_upcall(&msgdata);
198 spin_lock_bh(&waiting_locks_spinlock);
199 expired_lock_thread.elt_dump = 0;
204 while (!list_empty(expired)) {
205 struct obd_export *export;
206 struct ldlm_lock *lock;
208 lock = list_entry(expired->next, struct ldlm_lock,
210 if ((void *)lock < LP_POISON + PAGE_CACHE_SIZE &&
211 (void *)lock >= LP_POISON) {
212 spin_unlock_bh(&waiting_locks_spinlock);
213 CERROR("free lock on elt list %p\n", lock);
216 list_del_init(&lock->l_pending_chain);
217 if ((void *)lock->l_export <
218 LP_POISON + PAGE_CACHE_SIZE &&
219 (void *)lock->l_export >= LP_POISON) {
220 CERROR("lock with free export on elt list %p\n",
222 lock->l_export = NULL;
223 LDLM_ERROR(lock, "free export");
224 /* release extra ref grabbed by
225 * ldlm_add_waiting_lock() or
226 * ldlm_failed_ast() */
227 LDLM_LOCK_RELEASE(lock);
231 if (ldlm_is_destroyed(lock)) {
232 /* release the lock refcount where
233 * waiting_locks_callback() founds */
234 LDLM_LOCK_RELEASE(lock);
237 export = class_export_lock_get(lock->l_export, lock);
238 spin_unlock_bh(&waiting_locks_spinlock);
240 spin_lock_bh(&export->exp_bl_list_lock);
241 list_del_init(&lock->l_exp_list);
242 spin_unlock_bh(&export->exp_bl_list_lock);
245 class_fail_export(export);
246 class_export_lock_put(export, lock);
248 /* release extra ref grabbed by ldlm_add_waiting_lock()
249 * or ldlm_failed_ast() */
250 LDLM_LOCK_RELEASE(lock);
252 spin_lock_bh(&waiting_locks_spinlock);
254 spin_unlock_bh(&waiting_locks_spinlock);
256 if (do_dump && obd_dump_on_eviction) {
257 CERROR("dump the log upon eviction\n");
258 libcfs_debug_dumplog();
261 if (expired_lock_thread.elt_state == ELT_TERMINATE)
265 expired_lock_thread.elt_state = ELT_STOPPED;
266 wake_up(&expired_lock_thread.elt_waitq);
270 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
271 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds);
274 * Check if there is a request in the export request list
275 * which prevents the lock canceling.
277 static int ldlm_lock_busy(struct ldlm_lock *lock)
279 struct ptlrpc_request *req;
283 if (lock->l_export == NULL)
286 spin_lock_bh(&lock->l_export->exp_rpc_lock);
287 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
289 if (req->rq_ops->hpreq_lock_match) {
290 match = req->rq_ops->hpreq_lock_match(req, lock);
295 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
299 /* This is called from within a timer interrupt and cannot schedule */
300 static void waiting_locks_callback(unsigned long unused)
302 struct ldlm_lock *lock;
305 spin_lock_bh(&waiting_locks_spinlock);
306 while (!list_empty(&waiting_locks_list)) {
307 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
309 if (cfs_time_after(lock->l_callback_timeout,
310 cfs_time_current()) ||
311 (lock->l_req_mode == LCK_GROUP))
314 /* Check if we need to prolong timeout */
315 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
316 ldlm_lock_busy(lock)) {
319 if (lock->l_pending_chain.next == &waiting_locks_list)
324 spin_unlock_bh(&waiting_locks_spinlock);
325 LDLM_DEBUG(lock, "prolong the busy lock");
326 ldlm_refresh_waiting_lock(lock,
327 ldlm_get_enq_timeout(lock));
328 spin_lock_bh(&waiting_locks_spinlock);
331 LDLM_LOCK_RELEASE(lock);
335 LDLM_LOCK_RELEASE(lock);
338 ldlm_lock_to_ns(lock)->ns_timeouts++;
339 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
340 "evicting client at %s ",
341 cfs_time_current_sec() - lock->l_last_activity,
343 lock->l_export->exp_connection->c_peer.nid));
345 /* no needs to take an extra ref on the lock since it was in
346 * the waiting_locks_list and ldlm_add_waiting_lock()
347 * already grabbed a ref */
348 list_del(&lock->l_pending_chain);
349 list_add(&lock->l_pending_chain,
350 &expired_lock_thread.elt_expired_locks);
354 if (!list_empty(&expired_lock_thread.elt_expired_locks)) {
355 if (obd_dump_on_timeout && need_dump)
356 expired_lock_thread.elt_dump = __LINE__;
358 wake_up(&expired_lock_thread.elt_waitq);
362 * Make sure the timer will fire again if we have any locks
365 if (!list_empty(&waiting_locks_list)) {
366 cfs_time_t timeout_rounded;
367 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
369 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
370 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
372 spin_unlock_bh(&waiting_locks_spinlock);
376 * Add lock to the list of contended locks.
378 * Indicate that we're waiting for a client to call us back cancelling a given
379 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
380 * timer to fire appropriately. (We round up to the next second, to avoid
381 * floods of timer firings during periods of high lock contention and traffic).
382 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
383 * if it has been added to the waiting list (1 is returned).
385 * Called with the namespace lock held.
387 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
390 cfs_time_t timeout_rounded;
392 if (!list_empty(&lock->l_pending_chain))
395 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
396 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
399 timeout = cfs_time_shift(seconds);
400 if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
401 lock->l_callback_timeout = timeout;
403 timeout_rounded = round_timeout(lock->l_callback_timeout);
405 if (cfs_time_before(timeout_rounded,
406 cfs_timer_deadline(&waiting_locks_timer)) ||
407 !cfs_timer_is_armed(&waiting_locks_timer)) {
408 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
410 /* if the new lock has a shorter timeout than something earlier on
411 the list, we'll wait the longer amount of time; no big deal. */
413 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
417 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
420 int timeout = ldlm_get_enq_timeout(lock);
422 /* NB: must be called with hold of lock_res_and_lock() */
423 LASSERT(ldlm_is_res_locked(lock));
424 ldlm_set_waited(lock);
426 LASSERT(!ldlm_is_cancel_on_block(lock));
428 spin_lock_bh(&waiting_locks_spinlock);
429 if (ldlm_is_destroyed(lock)) {
430 static cfs_time_t next;
431 spin_unlock_bh(&waiting_locks_spinlock);
432 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
433 if (cfs_time_after(cfs_time_current(), next)) {
434 next = cfs_time_shift(14400);
435 libcfs_debug_dumpstack(NULL);
440 ret = __ldlm_add_waiting_lock(lock, timeout);
442 /* grab ref on the lock if it has been added to the
446 spin_unlock_bh(&waiting_locks_spinlock);
449 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
450 if (list_empty(&lock->l_exp_list))
451 list_add(&lock->l_exp_list,
452 &lock->l_export->exp_bl_list);
453 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
456 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
457 ret == 0 ? "not re-" : "", timeout,
458 AT_OFF ? "off" : "on");
463 * Remove a lock from the pending list, likely because it had its cancellation
464 * callback arrive without incident. This adjusts the lock-timeout timer if
465 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
466 * As done by ldlm_del_waiting_lock(), the caller must release the lock
467 * reference when the lock is removed from any list (1 is returned).
469 * Called with namespace lock held.
471 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
473 struct list_head *list_next;
475 if (list_empty(&lock->l_pending_chain))
478 list_next = lock->l_pending_chain.next;
479 if (lock->l_pending_chain.prev == &waiting_locks_list) {
480 /* Removing the head of the list, adjust timer. */
481 if (list_next == &waiting_locks_list) {
482 /* No more, just cancel. */
483 cfs_timer_disarm(&waiting_locks_timer);
485 struct ldlm_lock *next;
486 next = list_entry(list_next, struct ldlm_lock,
488 cfs_timer_arm(&waiting_locks_timer,
489 round_timeout(next->l_callback_timeout));
492 list_del_init(&lock->l_pending_chain);
497 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
501 if (lock->l_export == NULL) {
502 /* We don't have a "waiting locks list" on clients. */
503 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
507 spin_lock_bh(&waiting_locks_spinlock);
508 ret = __ldlm_del_waiting_lock(lock);
509 spin_unlock_bh(&waiting_locks_spinlock);
511 /* remove the lock out of export blocking list */
512 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
513 list_del_init(&lock->l_exp_list);
514 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
517 /* release lock ref if it has indeed been removed
519 LDLM_LOCK_RELEASE(lock);
522 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
525 EXPORT_SYMBOL(ldlm_del_waiting_lock);
528 * Prolong the contended lock waiting time.
530 * Called with namespace lock held.
532 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
534 if (lock->l_export == NULL) {
535 /* We don't have a "waiting locks list" on clients. */
536 LDLM_DEBUG(lock, "client lock: no-op");
540 spin_lock_bh(&waiting_locks_spinlock);
542 if (list_empty(&lock->l_pending_chain)) {
543 spin_unlock_bh(&waiting_locks_spinlock);
544 LDLM_DEBUG(lock, "wasn't waiting");
548 /* we remove/add the lock to the waiting list, so no needs to
549 * release/take a lock reference */
550 __ldlm_del_waiting_lock(lock);
551 __ldlm_add_waiting_lock(lock, timeout);
552 spin_unlock_bh(&waiting_locks_spinlock);
554 LDLM_DEBUG(lock, "refreshed");
557 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
559 #else /* HAVE_SERVER_SUPPORT */
561 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
566 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
571 #endif /* !HAVE_SERVER_SUPPORT */
573 #ifdef HAVE_SERVER_SUPPORT
576 * Perform lock cleanup if AST sending failed.
578 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
579 const char *ast_type)
581 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
582 "to a lock %s callback time out: rc %d\n",
583 lock->l_export->exp_obd->obd_name,
584 obd_export_nid2str(lock->l_export), ast_type, rc);
586 if (obd_dump_on_timeout)
587 libcfs_debug_dumplog();
588 spin_lock_bh(&waiting_locks_spinlock);
589 if (__ldlm_del_waiting_lock(lock) == 0)
590 /* the lock was not in any list, grab an extra ref before adding
591 * the lock to the expired list */
593 list_add(&lock->l_pending_chain,
594 &expired_lock_thread.elt_expired_locks);
595 wake_up(&expired_lock_thread.elt_waitq);
596 spin_unlock_bh(&waiting_locks_spinlock);
600 * Perform lock cleanup if AST reply came with error.
602 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
603 struct ptlrpc_request *req, int rc,
604 const char *ast_type)
606 lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
608 if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
609 LASSERT(lock->l_export);
610 if (lock->l_export->exp_libclient) {
611 LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
612 " timeout, just cancelling lock", ast_type,
613 libcfs_nid2str(peer.nid));
614 ldlm_lock_cancel(lock);
616 } else if (ldlm_is_cancel(lock)) {
617 LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
618 "cancel was received (AST reply lost?)",
619 ast_type, libcfs_nid2str(peer.nid));
620 ldlm_lock_cancel(lock);
623 ldlm_del_waiting_lock(lock);
624 ldlm_failed_ast(lock, rc, ast_type);
628 struct ldlm_resource *res = lock->l_resource;
629 LDLM_DEBUG(lock, "client (nid %s) returned %d"
630 " from %s AST - normal race",
631 libcfs_nid2str(peer.nid),
633 lustre_msg_get_status(req->rq_repmsg) : -1,
636 /* update lvbo to return proper attributes.
638 ldlm_resource_getref(res);
639 ldlm_res_lvbo_update(res, NULL, 1);
640 ldlm_resource_putref(res);
644 LDLM_ERROR(lock, "client (nid %s) returned %d: rc=%d "
645 "from %s AST", libcfs_nid2str(peer.nid),
646 (req->rq_repmsg != NULL) ?
647 lustre_msg_get_status(req->rq_repmsg) : 0,
650 ldlm_lock_cancel(lock);
651 /* Server-side AST functions are called from ldlm_reprocess_all,
652 * which needs to be told to please restart its reprocessing. */
659 static int ldlm_cb_interpret(const struct lu_env *env,
660 struct ptlrpc_request *req, void *data, int rc)
662 struct ldlm_cb_async_args *ca = data;
663 struct ldlm_lock *lock = ca->ca_lock;
664 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
667 LASSERT(lock != NULL);
670 case LDLM_GL_CALLBACK:
671 /* Update the LVB from disk if the AST failed
672 * (this is a legal race)
674 * - Glimpse callback of local lock just returns
675 * -ELDLM_NO_LOCK_DATA.
676 * - Glimpse callback of remote lock might return
677 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
679 if (rc == -ELDLM_NO_LOCK_DATA) {
680 LDLM_DEBUG(lock, "lost race - client has a lock but no "
682 ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
683 } else if (rc != 0) {
684 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
686 rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
689 case LDLM_BL_CALLBACK:
691 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
693 case LDLM_CP_CALLBACK:
695 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
698 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
703 /* release extra reference taken in ldlm_ast_fini() */
704 LDLM_LOCK_RELEASE(lock);
707 atomic_inc(&arg->restart);
712 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
714 struct ldlm_cb_async_args *ca = data;
715 struct ldlm_lock *lock = ca->ca_lock;
717 ldlm_refresh_waiting_lock(lock, ldlm_get_enq_timeout(lock));
720 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
721 struct ldlm_cb_set_arg *arg,
722 struct ldlm_lock *lock,
728 if (unlikely(instant_cancel)) {
729 rc = ptl_send_rpc(req, 1);
730 ptlrpc_req_finished(req);
732 atomic_inc(&arg->restart);
735 ptlrpc_set_add_req(arg->set, req);
742 * Check if there are requests in the export request list which prevent
743 * the lock canceling and make these requests high priority ones.
745 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
747 struct ptlrpc_request *req;
750 if (lock->l_export == NULL) {
751 LDLM_DEBUG(lock, "client lock: no-op");
755 spin_lock_bh(&lock->l_export->exp_rpc_lock);
756 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
758 /* Do not process requests that were not yet added to there
759 * incoming queue or were already removed from there for
760 * processing. We evaluate ptlrpc_nrs_req_can_move() without
761 * holding svcpt->scp_req_lock, and then redo the check with
762 * the lock held once we need to obtain a reliable result.
764 if (ptlrpc_nrs_req_can_move(req) &&
765 req->rq_ops->hpreq_lock_match &&
766 req->rq_ops->hpreq_lock_match(req, lock))
767 ptlrpc_nrs_req_hp_move(req);
769 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
774 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
775 * enqueued server lock conflicts with given one.
777 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
778 * to wait for client response.
780 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
781 struct ldlm_lock_desc *desc,
782 void *data, int flag)
784 struct ldlm_cb_async_args *ca;
785 struct ldlm_cb_set_arg *arg = data;
786 struct ldlm_request *body;
787 struct ptlrpc_request *req;
788 int instant_cancel = 0;
792 if (flag == LDLM_CB_CANCELING)
793 /* Don't need to do anything here. */
797 LASSERT(data != NULL);
798 if (lock->l_export->exp_obd->obd_recovering != 0)
799 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
801 ldlm_lock_reorder_req(lock);
803 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
804 &RQF_LDLM_BL_CALLBACK,
805 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
809 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
810 ca = ptlrpc_req_async_args(req);
811 ca->ca_set_arg = arg;
814 req->rq_interpret_reply = ldlm_cb_interpret;
816 lock_res_and_lock(lock);
817 if (lock->l_granted_mode != lock->l_req_mode) {
818 /* this blocking AST will be communicated as part of the
819 * completion AST instead */
820 unlock_res_and_lock(lock);
822 ptlrpc_req_finished(req);
823 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
827 if (ldlm_is_destroyed(lock)) {
828 /* What's the point? */
829 unlock_res_and_lock(lock);
830 ptlrpc_req_finished(req);
834 if (ldlm_is_cancel_on_block(lock))
837 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
838 body->lock_handle[0] = lock->l_remote_handle;
839 body->lock_desc = *desc;
840 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
842 LDLM_DEBUG(lock, "server preparing blocking AST");
844 ptlrpc_request_set_replen(req);
845 if (instant_cancel) {
846 unlock_res_and_lock(lock);
847 ldlm_lock_cancel(lock);
849 req->rq_no_resend = 1;
851 LASSERT(lock->l_granted_mode == lock->l_req_mode);
852 ldlm_add_waiting_lock(lock);
853 unlock_res_and_lock(lock);
855 /* Do not resend after lock callback timeout */
856 req->rq_delay_limit = ldlm_get_enq_timeout(lock);
857 req->rq_resend_cb = ldlm_update_resend;
860 req->rq_send_state = LUSTRE_IMP_FULL;
861 /* ptlrpc_request_alloc_pack already set timeout */
863 req->rq_timeout = ldlm_get_rq_timeout();
865 lock->l_last_activity = cfs_time_current_sec();
867 if (lock->l_export && lock->l_export->exp_nid_stats &&
868 lock->l_export->exp_nid_stats->nid_ldlm_stats)
869 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
870 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
872 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
876 EXPORT_SYMBOL(ldlm_server_blocking_ast);
879 * ->l_completion_ast callback for a remote lock in server namespace.
881 * Sends AST to the client notifying it of lock granting. If initial
882 * lock response was not sent yet, instead of sending another RPC, just
883 * mark the lock as granted and client will understand
885 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
887 struct ldlm_cb_set_arg *arg = data;
888 struct ldlm_request *body;
889 struct ptlrpc_request *req;
890 struct ldlm_cb_async_args *ca;
891 long total_enqueue_wait;
892 int instant_cancel = 0;
897 LASSERT(lock != NULL);
898 LASSERT(data != NULL);
900 total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(),
901 lock->l_last_activity);
903 if (OBD_FAIL_PRECHECK(OBD_FAIL_OST_LDLM_REPLY_NET)) {
904 LDLM_DEBUG(lock, "dropping CP AST");
908 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
909 &RQF_LDLM_CP_CALLBACK);
913 /* server namespace, doesn't need lock */
914 lvb_len = ldlm_lvbo_size(lock);
915 /* LU-3124 & LU-2187: to not return layout in completion AST because
916 * it may deadlock for LU-2187, or client may not have enough space
917 * for large layout. The layout will be returned to client with an
918 * extra RPC to fetch xattr.lov */
919 if (ldlm_has_layout(lock))
922 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
923 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
925 ptlrpc_request_free(req);
929 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
930 ca = ptlrpc_req_async_args(req);
931 ca->ca_set_arg = arg;
934 req->rq_interpret_reply = ldlm_cb_interpret;
935 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
937 body->lock_handle[0] = lock->l_remote_handle;
938 body->lock_flags = ldlm_flags_to_wire(flags);
939 ldlm_lock2desc(lock, &body->lock_desc);
941 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
943 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
945 /* We still need to send the RPC to wake up the blocked
946 * enqueue thread on the client.
948 * Consider old client, there is no better way to notify
949 * the failure, just zero-sized the LVB, then the client
950 * will fail out as "-EPROTO". */
951 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
955 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
960 LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)",
963 lock->l_last_activity = cfs_time_current_sec();
965 /* Server-side enqueue wait time estimate, used in
966 __ldlm_add_waiting_lock to set future enqueue timers */
967 if (total_enqueue_wait < ldlm_get_enq_timeout(lock))
968 at_measured(ldlm_lock_to_ns_at(lock),
971 /* bz18618. Don't add lock enqueue time we spend waiting for a
972 previous callback to fail. Locks waiting legitimately will
973 get extended by ldlm_refresh_waiting_lock regardless of the
974 estimate, so it's okay to underestimate here. */
975 LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. "
976 "It is likely that a previous callback timed out.",
978 at_get(ldlm_lock_to_ns_at(lock)));
980 ptlrpc_request_set_replen(req);
982 req->rq_send_state = LUSTRE_IMP_FULL;
983 /* ptlrpc_request_pack already set timeout */
985 req->rq_timeout = ldlm_get_rq_timeout();
987 /* We only send real blocking ASTs after the lock is granted */
988 lock_res_and_lock(lock);
989 if (ldlm_is_ast_sent(lock)) {
990 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
991 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
992 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
995 /* We might get here prior to ldlm_handle_enqueue setting
996 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
997 * into waiting list, but this is safe and similar code in
998 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
999 * that would not only cancel the lock, but will also remove
1000 * it from waiting list */
1001 if (ldlm_is_cancel_on_block(lock)) {
1002 unlock_res_and_lock(lock);
1003 ldlm_lock_cancel(lock);
1006 req->rq_no_resend = 1;
1008 lock_res_and_lock(lock);
1010 /* start the lock-timeout clock */
1011 ldlm_add_waiting_lock(lock);
1012 /* Do not resend after lock callback timeout */
1013 req->rq_delay_limit = ldlm_get_enq_timeout(lock);
1014 req->rq_resend_cb = ldlm_update_resend;
1017 unlock_res_and_lock(lock);
1019 if (lock->l_export && lock->l_export->exp_nid_stats &&
1020 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1021 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1022 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1024 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1026 RETURN(lvb_len < 0 ? lvb_len : rc);
1028 EXPORT_SYMBOL(ldlm_server_completion_ast);
1031 * Server side ->l_glimpse_ast handler for client locks.
1033 * Sends glimpse AST to the client and waits for reply. Then updates
1034 * lvbo with the result.
1036 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1038 struct ldlm_cb_set_arg *arg = data;
1039 struct ldlm_request *body;
1040 struct ptlrpc_request *req;
1041 struct ldlm_cb_async_args *ca;
1043 struct req_format *req_fmt;
1046 LASSERT(lock != NULL);
1048 if (arg->gl_desc != NULL)
1049 /* There is a glimpse descriptor to pack */
1050 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1052 req_fmt = &RQF_LDLM_GL_CALLBACK;
1054 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1055 req_fmt, LUSTRE_DLM_VERSION,
1061 if (arg->gl_desc != NULL) {
1062 /* copy the GL descriptor */
1063 union ldlm_gl_desc *desc;
1064 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1065 *desc = *arg->gl_desc;
1068 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1069 body->lock_handle[0] = lock->l_remote_handle;
1070 ldlm_lock2desc(lock, &body->lock_desc);
1072 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1073 ca = ptlrpc_req_async_args(req);
1074 ca->ca_set_arg = arg;
1077 /* server namespace, doesn't need lock */
1078 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1079 ldlm_lvbo_size(lock));
1080 ptlrpc_request_set_replen(req);
1082 req->rq_send_state = LUSTRE_IMP_FULL;
1083 /* ptlrpc_request_alloc_pack already set timeout */
1085 req->rq_timeout = ldlm_get_rq_timeout();
1087 lock->l_last_activity = cfs_time_current_sec();
1089 req->rq_interpret_reply = ldlm_cb_interpret;
1091 if (lock->l_export && lock->l_export->exp_nid_stats &&
1092 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1093 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1094 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1096 rc = ldlm_ast_fini(req, arg, lock, 0);
1100 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1102 int ldlm_glimpse_locks(struct ldlm_resource *res,
1103 struct list_head *gl_work_list)
1108 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1110 if (rc == -ERESTART)
1111 ldlm_reprocess_all(res);
1115 EXPORT_SYMBOL(ldlm_glimpse_locks);
1117 /* return LDLM lock associated with a lock callback request */
1118 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1120 struct ldlm_cb_async_args *ca;
1121 struct ldlm_lock *lock;
1124 ca = ptlrpc_req_async_args(req);
1127 RETURN(ERR_PTR(-EFAULT));
1131 EXPORT_SYMBOL(ldlm_request_lock);
1133 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1134 struct lprocfs_stats *srv_stats)
1136 int lock_type = 0, op = 0;
1138 lock_type = dlm_req->lock_desc.l_resource.lr_type;
1140 switch (lock_type) {
1142 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1145 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1146 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1148 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1151 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1154 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1162 lprocfs_counter_incr(srv_stats, op);
1168 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1169 * service threads to carry out client lock enqueueing requests.
1171 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1172 struct ptlrpc_request *req,
1173 const struct ldlm_request *dlm_req,
1174 const struct ldlm_callback_suite *cbs)
1176 struct ldlm_reply *dlm_rep;
1178 ldlm_error_t err = ELDLM_OK;
1179 struct ldlm_lock *lock = NULL;
1180 void *cookie = NULL;
1182 struct ldlm_resource *res = NULL;
1185 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1187 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1188 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1190 LASSERT(req->rq_export);
1192 if (ptlrpc_req2svc(req)->srv_stats != NULL)
1193 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1195 if (req->rq_export && req->rq_export->exp_nid_stats &&
1196 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1197 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1198 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1200 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1201 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1202 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1203 dlm_req->lock_desc.l_resource.lr_type);
1204 GOTO(out, rc = -EFAULT);
1207 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1208 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1209 dlm_req->lock_desc.l_req_mode &
1210 (dlm_req->lock_desc.l_req_mode-1))) {
1211 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1212 dlm_req->lock_desc.l_req_mode);
1213 GOTO(out, rc = -EFAULT);
1216 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) {
1217 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1219 DEBUG_REQ(D_ERROR, req,
1220 "PLAIN lock request from IBITS client?");
1221 GOTO(out, rc = -EPROTO);
1223 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1225 DEBUG_REQ(D_ERROR, req,
1226 "IBITS lock request from unaware client?");
1227 GOTO(out, rc = -EPROTO);
1231 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1232 against server's _CONNECT_SUPPORTED flags? (I don't want to use
1233 ibits for mgc/mgs) */
1235 /* INODEBITS_INTEROP: Perform conversion from plain lock to
1236 * inodebits lock if client does not support them. */
1237 if (!(exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) &&
1238 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1239 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1240 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1241 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1242 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1243 dlm_req->lock_desc.l_req_mode = LCK_CR;
1247 if (unlikely((flags & LDLM_FL_REPLAY) ||
1248 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1249 /* Find an existing lock in the per-export lock hash */
1250 /* In the function below, .hs_keycmp resolves to
1251 * ldlm_export_lock_keycmp() */
1252 /* coverity[overrun-buffer-val] */
1253 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1254 (void *)&dlm_req->lock_handle[0]);
1256 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1257 LPX64, lock->l_handle.h_cookie);
1258 flags |= LDLM_FL_RESENT;
1259 GOTO(existing_lock, rc = 0);
1263 /* The lock's callback data might be set in the policy function */
1264 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1265 dlm_req->lock_desc.l_resource.lr_type,
1266 dlm_req->lock_desc.l_req_mode,
1267 cbs, NULL, 0, LVB_T_NONE);
1274 lock->l_last_activity = cfs_time_current_sec();
1275 lock->l_remote_handle = dlm_req->lock_handle[0];
1276 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1278 /* Initialize resource lvb but not for a lock being replayed since
1279 * Client already got lvb sent in this case.
1280 * This must occur early since some policy methods assume resource
1281 * lvb is available (lr_lvb_data != NULL).
1283 res = lock->l_resource;
1284 if (!(flags & LDLM_FL_REPLAY)) {
1285 /* non-replayed lock, delayed lvb init may need to be done */
1286 rc = ldlm_lvbo_init(res);
1288 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", rc);
1293 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1294 /* Don't enqueue a lock onto the export if it is been disonnected
1295 * due to eviction (bug 3822) or server umount (bug 24324).
1296 * Cancel it now instead. */
1297 if (req->rq_export->exp_disconnected) {
1298 LDLM_ERROR(lock, "lock on disconnected export %p",
1300 GOTO(out, rc = -ENOTCONN);
1303 lock->l_export = class_export_lock_get(req->rq_export, lock);
1304 if (lock->l_export->exp_lock_hash)
1305 cfs_hash_add(lock->l_export->exp_lock_hash,
1306 &lock->l_remote_handle,
1309 /* Inherit the enqueue flags before the operation, because we do not
1310 * keep the res lock on return and next operations (BL AST) may proceed
1312 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1313 LDLM_FL_INHERIT_MASK);
1316 if (flags & LDLM_FL_HAS_INTENT) {
1317 /* In this case, the reply buffer is allocated deep in
1318 * local_lock_enqueue by the policy function. */
1321 /* based on the assumption that lvb size never changes during
1322 * resource life time otherwise it need resource->lr_lock's
1324 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1325 RCL_SERVER, ldlm_lvbo_size(lock));
1327 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1328 GOTO(out, rc = -ENOMEM);
1330 rc = req_capsule_server_pack(&req->rq_pill);
1335 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1336 ldlm_convert_policy_to_local(req->rq_export,
1337 dlm_req->lock_desc.l_resource.lr_type,
1338 &dlm_req->lock_desc.l_policy_data,
1339 &lock->l_policy_data);
1340 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1341 lock->l_req_extent = lock->l_policy_data.l_extent;
1343 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1350 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1352 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1353 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1355 /* We never send a blocking AST until the lock is granted, but
1356 * we can tell it right now */
1357 lock_res_and_lock(lock);
1359 /* Now take into account flags to be inherited from original lock
1360 request both in reply to client and in our own lock flags. */
1361 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1362 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1364 /* Don't move a pending lock onto the export if it has already been
1365 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1366 * Cancel it now instead. */
1367 if (unlikely(req->rq_export->exp_disconnected ||
1368 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1369 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1371 } else if (ldlm_is_ast_sent(lock)) {
1372 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1373 if (lock->l_granted_mode == lock->l_req_mode) {
1375 * Only cancel lock if it was granted, because it would
1376 * be destroyed immediately and would never be granted
1377 * in the future, causing timeouts on client. Not
1378 * granted lock will be cancelled immediately after
1379 * sending completion AST.
1381 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1382 unlock_res_and_lock(lock);
1383 ldlm_lock_cancel(lock);
1384 lock_res_and_lock(lock);
1386 ldlm_add_waiting_lock(lock);
1389 /* Make sure we never ever grant usual metadata locks to liblustre
1391 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1392 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1393 req->rq_export->exp_libclient) {
1394 if (unlikely(!ldlm_is_cancel_on_block(lock) ||
1395 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1396 CERROR("Granting sync lock to libclient. "
1397 "req fl %d, rep fl %d, lock fl "LPX64"\n",
1398 dlm_req->lock_flags, dlm_rep->lock_flags,
1400 LDLM_ERROR(lock, "sync lock");
1401 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1402 struct ldlm_intent *it;
1404 it = req_capsule_client_get(&req->rq_pill,
1407 CERROR("This is intent %s ("LPU64")\n",
1408 ldlm_it2str(it->opc), it->opc);
1414 unlock_res_and_lock(lock);
1418 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1419 if (!req->rq_packed_final) {
1420 err = lustre_pack_reply(req, 1, NULL, NULL);
1425 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1426 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1428 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1429 "(err=%d, rc=%d)", err, rc);
1432 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1434 ldlm_lvbo_size(lock) > 0) {
1438 buf = req_capsule_server_get(&req->rq_pill,
1440 LASSERTF(buf != NULL, "req %p, lock %p\n",
1442 buflen = req_capsule_get_size(&req->rq_pill,
1443 &RMF_DLM_LVB, RCL_SERVER);
1444 /* non-replayed lock, delayed lvb init may
1445 * need to be occur now */
1446 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1447 buflen = ldlm_lvbo_fill(lock, buf,
1453 buflen, RCL_SERVER);
1456 } else if (flags & LDLM_FL_REPLAY) {
1457 /* no LVB resend upon replay */
1472 lock_res_and_lock(lock);
1473 ldlm_resource_unlink_lock(lock);
1474 ldlm_lock_destroy_nolock(lock);
1475 unlock_res_and_lock(lock);
1478 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1479 ldlm_reprocess_all(lock->l_resource);
1481 LDLM_LOCK_RELEASE(lock);
1484 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1489 EXPORT_SYMBOL(ldlm_handle_enqueue0);
1492 * Old-style LDLM main entry point for server code enqueue.
1494 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1495 ldlm_completion_callback completion_callback,
1496 ldlm_blocking_callback blocking_callback,
1497 ldlm_glimpse_callback glimpse_callback)
1499 struct ldlm_request *dlm_req;
1500 struct ldlm_callback_suite cbs = {
1501 .lcs_completion = completion_callback,
1502 .lcs_blocking = blocking_callback,
1503 .lcs_glimpse = glimpse_callback
1507 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1508 if (dlm_req != NULL) {
1509 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1510 req, dlm_req, &cbs);
1516 EXPORT_SYMBOL(ldlm_handle_enqueue);
1519 * Main LDLM entry point for server code to process lock conversion requests.
1521 int ldlm_handle_convert0(struct ptlrpc_request *req,
1522 const struct ldlm_request *dlm_req)
1524 struct ldlm_reply *dlm_rep;
1525 struct ldlm_lock *lock;
1529 if (req->rq_export && req->rq_export->exp_nid_stats &&
1530 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1531 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1532 LDLM_CONVERT - LDLM_FIRST_OPC);
1534 rc = req_capsule_server_pack(&req->rq_pill);
1538 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1539 dlm_rep->lock_flags = dlm_req->lock_flags;
1541 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1543 req->rq_status = LUSTRE_EINVAL;
1547 LDLM_DEBUG(lock, "server-side convert handler START");
1549 lock->l_last_activity = cfs_time_current_sec();
1550 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1551 &dlm_rep->lock_flags);
1553 if (ldlm_del_waiting_lock(lock))
1554 LDLM_DEBUG(lock, "converted waiting lock");
1557 req->rq_status = LUSTRE_EDEADLK;
1562 if (!req->rq_status)
1563 ldlm_reprocess_all(lock->l_resource);
1564 LDLM_DEBUG(lock, "server-side convert handler END");
1565 LDLM_LOCK_PUT(lock);
1567 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1571 EXPORT_SYMBOL(ldlm_handle_convert0);
1574 * Old-style main LDLM entry point for server code to process lock conversion
1577 int ldlm_handle_convert(struct ptlrpc_request *req)
1580 struct ldlm_request *dlm_req;
1582 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1583 if (dlm_req != NULL) {
1584 rc = ldlm_handle_convert0(req, dlm_req);
1586 CERROR ("Can't unpack dlm_req\n");
1591 EXPORT_SYMBOL(ldlm_handle_convert);
1594 * Cancel all the locks whose handles are packed into ldlm_request
1596 * Called by server code expecting such combined cancel activity
1599 int ldlm_request_cancel(struct ptlrpc_request *req,
1600 const struct ldlm_request *dlm_req, int first)
1602 struct ldlm_resource *res, *pres = NULL;
1603 struct ldlm_lock *lock;
1604 int i, count, done = 0;
1607 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1611 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1614 /* There is no lock on the server at the replay time,
1615 * skip lock cancelling to make replay tests to pass. */
1616 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1619 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1620 "starting at %d", count, first);
1622 for (i = first; i < count; i++) {
1623 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1625 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1626 "lock (cookie "LPU64")",
1627 dlm_req->lock_handle[i].cookie);
1631 res = lock->l_resource;
1634 /* This code is an optimization to only attempt lock
1635 * granting on the resource (that could be CPU-expensive)
1636 * after we are done cancelling lock in that resource. */
1639 ldlm_reprocess_all(pres);
1640 LDLM_RESOURCE_DELREF(pres);
1641 ldlm_resource_putref(pres);
1644 ldlm_resource_getref(res);
1645 LDLM_RESOURCE_ADDREF(res);
1646 ldlm_res_lvbo_update(res, NULL, 1);
1650 ldlm_lock_cancel(lock);
1651 LDLM_LOCK_PUT(lock);
1654 ldlm_reprocess_all(pres);
1655 LDLM_RESOURCE_DELREF(pres);
1656 ldlm_resource_putref(pres);
1658 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1661 EXPORT_SYMBOL(ldlm_request_cancel);
1664 * Main LDLM entry point for server code to cancel locks.
1666 * Typically gets called from service handler on LDLM_CANCEL opc.
1668 int ldlm_handle_cancel(struct ptlrpc_request *req)
1670 struct ldlm_request *dlm_req;
1674 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1675 if (dlm_req == NULL) {
1676 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1680 if (req->rq_export && req->rq_export->exp_nid_stats &&
1681 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1682 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1683 LDLM_CANCEL - LDLM_FIRST_OPC);
1685 rc = req_capsule_server_pack(&req->rq_pill);
1689 if (!ldlm_request_cancel(req, dlm_req, 0))
1690 req->rq_status = LUSTRE_ESTALE;
1692 RETURN(ptlrpc_reply(req));
1694 EXPORT_SYMBOL(ldlm_handle_cancel);
1695 #endif /* HAVE_SERVER_SUPPORT */
1698 * Callback handler for receiving incoming blocking ASTs.
1700 * This can only happen on client side.
1702 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1703 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1708 LDLM_DEBUG(lock, "client blocking AST callback handler");
1710 lock_res_and_lock(lock);
1711 ldlm_set_cbpending(lock);
1713 if (ldlm_is_cancel_on_block(lock))
1714 ldlm_set_cancel(lock);
1716 do_ast = (!lock->l_readers && !lock->l_writers);
1717 unlock_res_and_lock(lock);
1720 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1721 lock, lock->l_blocking_ast);
1722 if (lock->l_blocking_ast != NULL)
1723 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1726 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1730 LDLM_DEBUG(lock, "client blocking callback handler END");
1731 LDLM_LOCK_RELEASE(lock);
1736 * Callback handler for receiving incoming completion ASTs.
1738 * This only can happen on client side.
1740 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1741 struct ldlm_namespace *ns,
1742 struct ldlm_request *dlm_req,
1743 struct ldlm_lock *lock)
1745 struct list_head ast_list;
1750 LDLM_DEBUG(lock, "client completion callback handler START");
1752 INIT_LIST_HEAD(&ast_list);
1753 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1754 int to = cfs_time_seconds(1);
1756 schedule_timeout_and_set_state(
1757 TASK_INTERRUPTIBLE, to);
1758 if (lock->l_granted_mode == lock->l_req_mode ||
1759 ldlm_is_destroyed(lock))
1764 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1766 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1767 GOTO(out, rc = lvb_len);
1768 } else if (lvb_len > 0) {
1769 if (lock->l_lvb_len > 0) {
1770 /* for extent lock, lvb contains ost_lvb{}. */
1771 LASSERT(lock->l_lvb_data != NULL);
1773 if (unlikely(lock->l_lvb_len < lvb_len)) {
1774 LDLM_ERROR(lock, "Replied LVB is larger than "
1775 "expectation, expected = %d, "
1777 lock->l_lvb_len, lvb_len);
1778 GOTO(out, rc = -EINVAL);
1780 } else if (ldlm_has_layout(lock)) { /* for layout lock, lvb has
1781 * variable length */
1784 OBD_ALLOC_LARGE(lvb_data, lvb_len);
1785 if (lvb_data == NULL) {
1786 LDLM_ERROR(lock, "No memory: %d.\n", lvb_len);
1787 GOTO(out, rc = -ENOMEM);
1790 lock_res_and_lock(lock);
1791 LASSERT(lock->l_lvb_data == NULL);
1792 lock->l_lvb_type = LVB_T_LAYOUT;
1793 lock->l_lvb_data = lvb_data;
1794 lock->l_lvb_len = lvb_len;
1795 unlock_res_and_lock(lock);
1799 lock_res_and_lock(lock);
1800 if (ldlm_is_destroyed(lock) ||
1801 lock->l_granted_mode == lock->l_req_mode) {
1802 /* bug 11300: the lock has already been granted */
1803 unlock_res_and_lock(lock);
1804 LDLM_DEBUG(lock, "Double grant race happened");
1808 /* If we receive the completion AST before the actual enqueue returned,
1809 * then we might need to switch lock modes, resources, or extents. */
1810 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1811 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1812 LDLM_DEBUG(lock, "completion AST, new lock mode");
1815 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1816 ldlm_convert_policy_to_local(req->rq_export,
1817 dlm_req->lock_desc.l_resource.lr_type,
1818 &dlm_req->lock_desc.l_policy_data,
1819 &lock->l_policy_data);
1820 LDLM_DEBUG(lock, "completion AST, new policy data");
1823 ldlm_resource_unlink_lock(lock);
1824 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1825 &lock->l_resource->lr_name,
1826 sizeof(lock->l_resource->lr_name)) != 0) {
1827 unlock_res_and_lock(lock);
1828 rc = ldlm_lock_change_resource(ns, lock,
1829 &dlm_req->lock_desc.l_resource.lr_name);
1831 LDLM_ERROR(lock, "Failed to allocate resource");
1834 LDLM_DEBUG(lock, "completion AST, new resource");
1835 CERROR("change resource!\n");
1836 lock_res_and_lock(lock);
1839 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1840 /* BL_AST locks are not needed in LRU.
1841 * Let ldlm_cancel_lru() be fast. */
1842 ldlm_lock_remove_from_lru(lock);
1843 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1844 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1847 if (lock->l_lvb_len > 0) {
1848 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1849 lock->l_lvb_data, lvb_len);
1851 unlock_res_and_lock(lock);
1856 ldlm_grant_lock(lock, &ast_list);
1857 unlock_res_and_lock(lock);
1859 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1861 /* Let Enqueue to call osc_lock_upcall() and initialize
1863 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1865 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1867 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1873 lock_res_and_lock(lock);
1874 ldlm_set_failed(lock);
1875 unlock_res_and_lock(lock);
1876 wake_up(&lock->l_waitq);
1878 LDLM_LOCK_RELEASE(lock);
1882 * Callback handler for receiving incoming glimpse ASTs.
1884 * This only can happen on client side. After handling the glimpse AST
1885 * we also consider dropping the lock here if it is unused locally for a
1888 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1889 struct ldlm_namespace *ns,
1890 struct ldlm_request *dlm_req,
1891 struct ldlm_lock *lock)
1896 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1898 if (lock->l_glimpse_ast != NULL)
1899 rc = lock->l_glimpse_ast(lock, req);
1901 if (req->rq_repmsg != NULL) {
1904 req->rq_status = rc;
1908 lock_res_and_lock(lock);
1909 if (lock->l_granted_mode == LCK_PW &&
1910 !lock->l_readers && !lock->l_writers &&
1911 cfs_time_after(cfs_time_current(),
1912 cfs_time_add(lock->l_last_used,
1913 cfs_time_seconds(10)))) {
1914 unlock_res_and_lock(lock);
1915 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1916 ldlm_handle_bl_callback(ns, NULL, lock);
1921 unlock_res_and_lock(lock);
1922 LDLM_LOCK_RELEASE(lock);
1926 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1928 if (req->rq_no_reply)
1931 req->rq_status = rc;
1932 if (!req->rq_packed_final) {
1933 rc = lustre_pack_reply(req, 1, NULL, NULL);
1937 return ptlrpc_reply(req);
1940 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1941 ldlm_cancel_flags_t cancel_flags)
1943 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1946 spin_lock(&blp->blp_lock);
1947 if (blwi->blwi_lock &&
1948 ldlm_is_discard_data(blwi->blwi_lock)) {
1949 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1950 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1952 /* other blocking callbacks are added to the regular list */
1953 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1955 spin_unlock(&blp->blp_lock);
1957 wake_up(&blp->blp_waitq);
1959 /* can not check blwi->blwi_flags as blwi could be already freed in
1961 if (!(cancel_flags & LCF_ASYNC))
1962 wait_for_completion(&blwi->blwi_comp);
1967 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1968 struct ldlm_namespace *ns,
1969 struct ldlm_lock_desc *ld,
1970 struct list_head *cancels, int count,
1971 struct ldlm_lock *lock,
1972 ldlm_cancel_flags_t cancel_flags)
1974 init_completion(&blwi->blwi_comp);
1975 INIT_LIST_HEAD(&blwi->blwi_head);
1977 if (memory_pressure_get())
1978 blwi->blwi_mem_pressure = 1;
1981 blwi->blwi_flags = cancel_flags;
1983 blwi->blwi_ld = *ld;
1985 list_add(&blwi->blwi_head, cancels);
1986 list_del_init(cancels);
1987 blwi->blwi_count = count;
1989 blwi->blwi_lock = lock;
1994 * Queues a list of locks \a cancels containing \a count locks
1995 * for later processing by a blocking thread. If \a count is zero,
1996 * then the lock referenced as \a lock is queued instead.
1998 * The blocking thread would then call ->l_blocking_ast callback in the lock.
1999 * If list addition fails an error is returned and caller is supposed to
2000 * call ->l_blocking_ast itself.
2002 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2003 struct ldlm_lock_desc *ld,
2004 struct ldlm_lock *lock,
2005 struct list_head *cancels, int count,
2006 ldlm_cancel_flags_t cancel_flags)
2010 if (cancels && count == 0)
2013 if (cancel_flags & LCF_ASYNC) {
2014 struct ldlm_bl_work_item *blwi;
2016 OBD_ALLOC(blwi, sizeof(*blwi));
2019 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2021 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2023 /* if it is synchronous call do minimum mem alloc, as it could
2024 * be triggered from kernel shrinker
2026 struct ldlm_bl_work_item blwi;
2028 memset(&blwi, 0, sizeof(blwi));
2029 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2030 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2035 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2036 struct ldlm_lock *lock)
2038 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2041 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2042 struct list_head *cancels, int count,
2043 ldlm_cancel_flags_t cancel_flags)
2045 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2048 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2049 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2051 struct obd_device *obd = req->rq_export->exp_obd;
2058 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2060 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2062 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2064 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2067 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2069 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2071 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2074 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2077 /* We are responsible for swabbing contents of val */
2079 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2080 /* Pass it on to mdc (the "export" in this case) */
2081 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2083 sizeof(KEY_HSM_COPYTOOL_SEND),
2084 KEY_HSM_COPYTOOL_SEND,
2087 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2092 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2093 const char *msg, int rc,
2094 struct lustre_handle *handle)
2096 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2097 "%s: [nid %s] [rc %d] [lock "LPX64"]",
2098 msg, libcfs_id2str(req->rq_peer), rc,
2099 handle ? handle->cookie : 0);
2100 if (req->rq_no_reply)
2101 CWARN("No reply was sent, maybe cause bug 21636.\n");
2103 CWARN("Send reply failed, maybe cause bug 21636.\n");
2106 static int ldlm_handle_qc_callback(struct ptlrpc_request *req)
2108 struct obd_quotactl *oqctl;
2109 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2111 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2112 if (oqctl == NULL) {
2113 CERROR("Can't unpack obd_quotactl\n");
2117 oqctl->qc_stat = ptlrpc_status_ntoh(oqctl->qc_stat);
2119 cli->cl_qchk_stat = oqctl->qc_stat;
2123 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2124 static int ldlm_callback_handler(struct ptlrpc_request *req)
2126 struct ldlm_namespace *ns;
2127 struct ldlm_request *dlm_req;
2128 struct ldlm_lock *lock;
2132 /* Requests arrive in sender's byte order. The ptlrpc service
2133 * handler has already checked and, if necessary, byte-swapped the
2134 * incoming request message body, but I am responsible for the
2135 * message buffers. */
2137 /* do nothing for sec context finalize */
2138 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2141 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2143 if (req->rq_export == NULL) {
2144 rc = ldlm_callback_reply(req, -ENOTCONN);
2145 ldlm_callback_errmsg(req, "Operate on unconnected server",
2150 LASSERT(req->rq_export != NULL);
2151 LASSERT(req->rq_export->exp_obd != NULL);
2153 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2154 case LDLM_BL_CALLBACK:
2155 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET))
2158 case LDLM_CP_CALLBACK:
2159 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2162 case LDLM_GL_CALLBACK:
2163 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2167 rc = ldlm_handle_setinfo(req);
2168 ldlm_callback_reply(req, rc);
2170 case LLOG_ORIGIN_HANDLE_CREATE:
2171 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2172 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2174 rc = llog_origin_handle_open(req);
2175 ldlm_callback_reply(req, rc);
2177 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2178 req_capsule_set(&req->rq_pill,
2179 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2180 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2182 rc = llog_origin_handle_next_block(req);
2183 ldlm_callback_reply(req, rc);
2185 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2186 req_capsule_set(&req->rq_pill,
2187 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2188 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2190 rc = llog_origin_handle_read_header(req);
2191 ldlm_callback_reply(req, rc);
2193 case LLOG_ORIGIN_HANDLE_CLOSE:
2194 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2196 rc = llog_origin_handle_close(req);
2197 ldlm_callback_reply(req, rc);
2199 case OBD_QC_CALLBACK:
2200 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
2201 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
2203 rc = ldlm_handle_qc_callback(req);
2204 ldlm_callback_reply(req, rc);
2207 CERROR("unknown opcode %u\n",
2208 lustre_msg_get_opc(req->rq_reqmsg));
2209 ldlm_callback_reply(req, -EPROTO);
2213 ns = req->rq_export->exp_obd->obd_namespace;
2214 LASSERT(ns != NULL);
2216 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2218 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2219 if (dlm_req == NULL) {
2220 rc = ldlm_callback_reply(req, -EPROTO);
2221 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2226 /* Force a known safe race, send a cancel to the server for a lock
2227 * which the server has already started a blocking callback on. */
2228 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2229 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2230 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2232 CERROR("ldlm_cli_cancel: %d\n", rc);
2235 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2237 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2238 "disappeared\n", dlm_req->lock_handle[0].cookie);
2239 rc = ldlm_callback_reply(req, -EINVAL);
2240 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2241 &dlm_req->lock_handle[0]);
2245 if (ldlm_is_fail_loc(lock) &&
2246 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2247 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2249 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2250 lock_res_and_lock(lock);
2251 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2253 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2254 /* If somebody cancels lock and cache is already dropped,
2255 * or lock is failed before cp_ast received on client,
2256 * we can tell the server we have no lock. Otherwise, we
2257 * should send cancel after dropping the cache. */
2258 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2259 ldlm_is_failed(lock)) {
2260 LDLM_DEBUG(lock, "callback on lock "
2261 LPX64" - lock disappeared\n",
2262 dlm_req->lock_handle[0].cookie);
2263 unlock_res_and_lock(lock);
2264 LDLM_LOCK_RELEASE(lock);
2265 rc = ldlm_callback_reply(req, -EINVAL);
2266 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2267 &dlm_req->lock_handle[0]);
2270 /* BL_AST locks are not needed in LRU.
2271 * Let ldlm_cancel_lru() be fast. */
2272 ldlm_lock_remove_from_lru(lock);
2273 ldlm_set_bl_ast(lock);
2275 unlock_res_and_lock(lock);
2277 /* We want the ost thread to get this reply so that it can respond
2278 * to ost requests (write cache writeback) that might be triggered
2281 * But we'd also like to be able to indicate in the reply that we're
2282 * cancelling right now, because it's unused, or have an intent result
2283 * in the reply, so we might have to push the responsibility for sending
2284 * the reply down into the AST handlers, alas. */
2286 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2287 case LDLM_BL_CALLBACK:
2288 CDEBUG(D_INODE, "blocking ast\n");
2289 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2290 if (!ldlm_is_cancel_on_block(lock)) {
2291 rc = ldlm_callback_reply(req, 0);
2292 if (req->rq_no_reply || rc)
2293 ldlm_callback_errmsg(req, "Normal process", rc,
2294 &dlm_req->lock_handle[0]);
2296 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2297 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2299 case LDLM_CP_CALLBACK:
2300 CDEBUG(D_INODE, "completion ast\n");
2301 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2302 ldlm_callback_reply(req, 0);
2303 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2305 case LDLM_GL_CALLBACK:
2306 CDEBUG(D_INODE, "glimpse ast\n");
2307 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2308 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2311 LBUG(); /* checked above */
2317 #ifdef HAVE_SERVER_SUPPORT
2319 * Main handler for canceld thread.
2321 * Separated into its own thread to avoid deadlocks.
2323 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2328 /* Requests arrive in sender's byte order. The ptlrpc service
2329 * handler has already checked and, if necessary, byte-swapped the
2330 * incoming request message body, but I am responsible for the
2331 * message buffers. */
2333 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2335 if (req->rq_export == NULL) {
2336 struct ldlm_request *dlm_req;
2338 CERROR("%s from %s arrived at %lu with bad export cookie "
2340 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2341 libcfs_nid2str(req->rq_peer.nid),
2342 req->rq_arrival_time.tv_sec,
2343 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2345 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2346 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2347 dlm_req = req_capsule_client_get(&req->rq_pill,
2349 if (dlm_req != NULL)
2350 ldlm_lock_dump_handle(D_ERROR,
2351 &dlm_req->lock_handle[0]);
2353 ldlm_callback_reply(req, -ENOTCONN);
2357 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2359 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2361 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2362 CDEBUG(D_INODE, "cancel\n");
2363 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2364 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))
2366 rc = ldlm_handle_cancel(req);
2371 CERROR("invalid opcode %d\n",
2372 lustre_msg_get_opc(req->rq_reqmsg));
2373 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2374 ldlm_callback_reply(req, -EINVAL);
2380 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2381 struct ldlm_lock *lock)
2383 struct ldlm_request *dlm_req;
2384 struct lustre_handle lockh;
2389 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2390 if (dlm_req == NULL)
2393 ldlm_lock2handle(lock, &lockh);
2394 for (i = 0; i < dlm_req->lock_count; i++) {
2395 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2397 DEBUG_REQ(D_RPCTRACE, req,
2398 "Prio raised by lock "LPX64".", lockh.cookie);
2409 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2411 struct ldlm_request *dlm_req;
2416 /* no prolong in recovery */
2417 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2420 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2421 if (dlm_req == NULL)
2424 for (i = 0; i < dlm_req->lock_count; i++) {
2425 struct ldlm_lock *lock;
2427 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2431 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2433 LDLM_DEBUG(lock, "hpreq cancel lock");
2434 LDLM_LOCK_PUT(lock);
2443 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2444 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2445 .hpreq_check = ldlm_cancel_hpreq_check,
2449 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2453 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2455 if (req->rq_export == NULL)
2458 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2459 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2460 req->rq_ops = &ldlm_cancel_hpreq_ops;
2465 int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2466 struct hlist_node *hnode, void *data)
2469 struct list_head *rpc_list = data;
2470 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2472 lock_res_and_lock(lock);
2474 if (lock->l_req_mode != lock->l_granted_mode) {
2475 unlock_res_and_lock(lock);
2479 LASSERT(lock->l_resource);
2480 if (lock->l_resource->lr_type != LDLM_IBITS &&
2481 lock->l_resource->lr_type != LDLM_PLAIN) {
2482 unlock_res_and_lock(lock);
2486 if (ldlm_is_ast_sent(lock)) {
2487 unlock_res_and_lock(lock);
2491 LASSERT(lock->l_blocking_ast);
2492 LASSERT(!lock->l_blocking_lock);
2494 ldlm_set_ast_sent(lock);
2495 if (lock->l_export && lock->l_export->exp_lock_hash) {
2496 /* NB: it's safe to call cfs_hash_del() even lock isn't
2497 * in exp_lock_hash. */
2498 /* In the function below, .hs_keycmp resolves to
2499 * ldlm_export_lock_keycmp() */
2500 /* coverity[overrun-buffer-val] */
2501 cfs_hash_del(lock->l_export->exp_lock_hash,
2502 &lock->l_remote_handle, &lock->l_exp_hash);
2505 list_add_tail(&lock->l_rk_ast, rpc_list);
2506 LDLM_LOCK_GET(lock);
2508 unlock_res_and_lock(lock);
2512 void ldlm_revoke_export_locks(struct obd_export *exp)
2514 struct list_head rpc_list;
2517 INIT_LIST_HEAD(&rpc_list);
2518 cfs_hash_for_each_empty(exp->exp_lock_hash,
2519 ldlm_revoke_lock_cb, &rpc_list);
2520 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2521 LDLM_WORK_REVOKE_AST);
2525 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2526 #endif /* HAVE_SERVER_SUPPORT */
2528 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2530 struct ldlm_bl_work_item *blwi = NULL;
2531 static unsigned int num_bl = 0;
2533 spin_lock(&blp->blp_lock);
2534 /* process a request from the blp_list at least every blp_num_threads */
2535 if (!list_empty(&blp->blp_list) &&
2536 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2537 blwi = list_entry(blp->blp_list.next,
2538 struct ldlm_bl_work_item, blwi_entry);
2540 if (!list_empty(&blp->blp_prio_list))
2541 blwi = list_entry(blp->blp_prio_list.next,
2542 struct ldlm_bl_work_item,
2546 if (++num_bl >= atomic_read(&blp->blp_num_threads))
2548 list_del(&blwi->blwi_entry);
2550 spin_unlock(&blp->blp_lock);
2555 /* This only contains temporary data until the thread starts */
2556 struct ldlm_bl_thread_data {
2557 char bltd_name[CFS_CURPROC_COMM_MAX];
2558 struct ldlm_bl_pool *bltd_blp;
2559 struct completion bltd_comp;
2563 static int ldlm_bl_thread_main(void *arg);
2565 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2567 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2568 struct task_struct *task;
2570 init_completion(&bltd.bltd_comp);
2571 bltd.bltd_num = atomic_read(&blp->blp_num_threads);
2572 snprintf(bltd.bltd_name, sizeof(bltd.bltd_name) - 1,
2573 "ldlm_bl_%02d", bltd.bltd_num);
2574 task = kthread_run(ldlm_bl_thread_main, &bltd, bltd.bltd_name);
2576 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2577 atomic_read(&blp->blp_num_threads), PTR_ERR(task));
2578 return PTR_ERR(task);
2580 wait_for_completion(&bltd.bltd_comp);
2586 * Main blocking requests processing thread.
2588 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2589 * This thread in the end ends up doing actual call to ->l_blocking_ast
2592 static int ldlm_bl_thread_main(void *arg)
2594 struct ldlm_bl_pool *blp;
2595 struct ldlm_bl_thread_data *bltd = arg;
2598 blp = bltd->bltd_blp;
2600 atomic_inc(&blp->blp_num_threads);
2601 atomic_inc(&blp->blp_busy_threads);
2603 complete(&bltd->bltd_comp);
2604 /* cannot use bltd after this, it is only on caller's stack */
2607 struct l_wait_info lwi = { 0 };
2608 struct ldlm_bl_work_item *blwi = NULL;
2611 blwi = ldlm_bl_get_work(blp);
2614 atomic_dec(&blp->blp_busy_threads);
2615 l_wait_event_exclusive(blp->blp_waitq,
2616 (blwi = ldlm_bl_get_work(blp)) != NULL,
2618 busy = atomic_inc_return(&blp->blp_busy_threads);
2620 busy = atomic_read(&blp->blp_busy_threads);
2623 if (blwi->blwi_ns == NULL)
2624 /* added by ldlm_cleanup() */
2627 /* Not fatal if racy and have a few too many threads */
2628 if (unlikely(busy < blp->blp_max_threads &&
2629 busy >= atomic_read(&blp->blp_num_threads) &&
2630 !blwi->blwi_mem_pressure))
2631 /* discard the return value, we tried */
2632 ldlm_bl_thread_start(blp);
2634 if (blwi->blwi_mem_pressure)
2635 memory_pressure_set();
2637 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2639 if (blwi->blwi_count) {
2641 /* The special case when we cancel locks in LRU
2642 * asynchronously, we pass the list of locks here.
2643 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2644 * canceled locally yet. */
2645 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2648 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2651 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2654 if (blwi->blwi_mem_pressure)
2655 memory_pressure_clr();
2657 if (blwi->blwi_flags & LCF_ASYNC)
2658 OBD_FREE(blwi, sizeof(*blwi));
2660 complete(&blwi->blwi_comp);
2663 atomic_dec(&blp->blp_busy_threads);
2664 atomic_dec(&blp->blp_num_threads);
2665 complete(&blp->blp_comp);
2670 static int ldlm_setup(void);
2671 static int ldlm_cleanup(void);
2673 int ldlm_get_ref(void)
2677 mutex_lock(&ldlm_ref_mutex);
2678 if (++ldlm_refcount == 1) {
2683 mutex_unlock(&ldlm_ref_mutex);
2687 EXPORT_SYMBOL(ldlm_get_ref);
2689 void ldlm_put_ref(void)
2692 mutex_lock(&ldlm_ref_mutex);
2693 if (ldlm_refcount == 1) {
2694 int rc = ldlm_cleanup();
2696 CERROR("ldlm_cleanup failed: %d\n", rc);
2702 mutex_unlock(&ldlm_ref_mutex);
2706 EXPORT_SYMBOL(ldlm_put_ref);
2709 * Export handle<->lock hash operations.
2712 ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
2714 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2718 ldlm_export_lock_key(struct hlist_node *hnode)
2720 struct ldlm_lock *lock;
2722 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2723 return &lock->l_remote_handle;
2727 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2729 struct ldlm_lock *lock;
2731 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2732 lock->l_remote_handle = *(struct lustre_handle *)key;
2736 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2738 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2742 ldlm_export_lock_object(struct hlist_node *hnode)
2744 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2748 ldlm_export_lock_get(cfs_hash_t *hs, struct hlist_node *hnode)
2750 struct ldlm_lock *lock;
2752 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2753 LDLM_LOCK_GET(lock);
2757 ldlm_export_lock_put(cfs_hash_t *hs, struct hlist_node *hnode)
2759 struct ldlm_lock *lock;
2761 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2762 LDLM_LOCK_RELEASE(lock);
2765 static cfs_hash_ops_t ldlm_export_lock_ops = {
2766 .hs_hash = ldlm_export_lock_hash,
2767 .hs_key = ldlm_export_lock_key,
2768 .hs_keycmp = ldlm_export_lock_keycmp,
2769 .hs_keycpy = ldlm_export_lock_keycpy,
2770 .hs_object = ldlm_export_lock_object,
2771 .hs_get = ldlm_export_lock_get,
2772 .hs_put = ldlm_export_lock_put,
2773 .hs_put_locked = ldlm_export_lock_put,
2776 int ldlm_init_export(struct obd_export *exp)
2781 exp->exp_lock_hash =
2782 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2783 HASH_EXP_LOCK_CUR_BITS,
2784 HASH_EXP_LOCK_MAX_BITS,
2785 HASH_EXP_LOCK_BKT_BITS, 0,
2786 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2787 &ldlm_export_lock_ops,
2788 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2789 CFS_HASH_NBLK_CHANGE);
2791 if (!exp->exp_lock_hash)
2794 rc = ldlm_init_flock_export(exp);
2800 ldlm_destroy_export(exp);
2803 EXPORT_SYMBOL(ldlm_init_export);
2805 void ldlm_destroy_export(struct obd_export *exp)
2808 cfs_hash_putref(exp->exp_lock_hash);
2809 exp->exp_lock_hash = NULL;
2811 ldlm_destroy_flock_export(exp);
2814 EXPORT_SYMBOL(ldlm_destroy_export);
2816 static int ldlm_setup(void)
2818 static struct ptlrpc_service_conf conf;
2819 struct ldlm_bl_pool *blp = NULL;
2820 #ifdef HAVE_SERVER_SUPPORT
2821 struct task_struct *task;
2822 #endif /* HAVE_SERVER_SUPPORT */
2828 if (ldlm_state != NULL)
2831 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2832 if (ldlm_state == NULL)
2836 rc = ldlm_proc_setup();
2839 #endif /* LPROCFS */
2841 memset(&conf, 0, sizeof(conf));
2842 conf = (typeof(conf)) {
2843 .psc_name = "ldlm_cbd",
2844 .psc_watchdog_factor = 2,
2846 .bc_nbufs = LDLM_CLIENT_NBUFS,
2847 .bc_buf_size = LDLM_BUFSIZE,
2848 .bc_req_max_size = LDLM_MAXREQSIZE,
2849 .bc_rep_max_size = LDLM_MAXREPSIZE,
2850 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2851 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2854 .tc_thr_name = "ldlm_cb",
2855 .tc_thr_factor = LDLM_THR_FACTOR,
2856 .tc_nthrs_init = LDLM_NTHRS_INIT,
2857 .tc_nthrs_base = LDLM_NTHRS_BASE,
2858 .tc_nthrs_max = LDLM_NTHRS_MAX,
2859 .tc_nthrs_user = ldlm_num_threads,
2860 .tc_cpu_affinity = 1,
2861 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2864 .cc_pattern = ldlm_cpts,
2867 .so_req_handler = ldlm_callback_handler,
2870 ldlm_state->ldlm_cb_service = \
2871 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2872 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2873 CERROR("failed to start service\n");
2874 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2875 ldlm_state->ldlm_cb_service = NULL;
2879 #ifdef HAVE_SERVER_SUPPORT
2880 memset(&conf, 0, sizeof(conf));
2881 conf = (typeof(conf)) {
2882 .psc_name = "ldlm_canceld",
2883 .psc_watchdog_factor = 6,
2885 .bc_nbufs = LDLM_SERVER_NBUFS,
2886 .bc_buf_size = LDLM_BUFSIZE,
2887 .bc_req_max_size = LDLM_MAXREQSIZE,
2888 .bc_rep_max_size = LDLM_MAXREPSIZE,
2889 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
2890 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
2894 .tc_thr_name = "ldlm_cn",
2895 .tc_thr_factor = LDLM_THR_FACTOR,
2896 .tc_nthrs_init = LDLM_NTHRS_INIT,
2897 .tc_nthrs_base = LDLM_NTHRS_BASE,
2898 .tc_nthrs_max = LDLM_NTHRS_MAX,
2899 .tc_nthrs_user = ldlm_num_threads,
2900 .tc_cpu_affinity = 1,
2901 .tc_ctx_tags = LCT_MD_THREAD | \
2906 .cc_pattern = ldlm_cpts,
2909 .so_req_handler = ldlm_cancel_handler,
2910 .so_hpreq_handler = ldlm_hpreq_handler,
2913 ldlm_state->ldlm_cancel_service = \
2914 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2915 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2916 CERROR("failed to start service\n");
2917 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2918 ldlm_state->ldlm_cancel_service = NULL;
2921 #endif /* HAVE_SERVER_SUPPORT */
2923 OBD_ALLOC(blp, sizeof(*blp));
2925 GOTO(out, rc = -ENOMEM);
2926 ldlm_state->ldlm_bl_pool = blp;
2928 spin_lock_init(&blp->blp_lock);
2929 INIT_LIST_HEAD(&blp->blp_list);
2930 INIT_LIST_HEAD(&blp->blp_prio_list);
2931 init_waitqueue_head(&blp->blp_waitq);
2932 atomic_set(&blp->blp_num_threads, 0);
2933 atomic_set(&blp->blp_busy_threads, 0);
2935 if (ldlm_num_threads == 0) {
2936 blp->blp_min_threads = LDLM_NTHRS_INIT;
2937 blp->blp_max_threads = LDLM_NTHRS_MAX;
2939 blp->blp_min_threads = blp->blp_max_threads = \
2940 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
2944 for (i = 0; i < blp->blp_min_threads; i++) {
2945 rc = ldlm_bl_thread_start(blp);
2950 #ifdef HAVE_SERVER_SUPPORT
2951 INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2952 expired_lock_thread.elt_state = ELT_STOPPED;
2953 init_waitqueue_head(&expired_lock_thread.elt_waitq);
2955 INIT_LIST_HEAD(&waiting_locks_list);
2956 spin_lock_init(&waiting_locks_spinlock);
2957 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2959 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
2962 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2966 wait_event(expired_lock_thread.elt_waitq,
2967 expired_lock_thread.elt_state == ELT_READY);
2968 #endif /* HAVE_SERVER_SUPPORT */
2970 rc = ldlm_pools_init();
2972 CERROR("Failed to initialize LDLM pools: %d\n", rc);
2982 static int ldlm_cleanup(void)
2986 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2987 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2988 CERROR("ldlm still has namespaces; clean these up first.\n");
2989 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2990 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2996 if (ldlm_state->ldlm_bl_pool != NULL) {
2997 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2999 while (atomic_read(&blp->blp_num_threads) > 0) {
3000 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3002 init_completion(&blp->blp_comp);
3004 spin_lock(&blp->blp_lock);
3005 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3006 wake_up(&blp->blp_waitq);
3007 spin_unlock(&blp->blp_lock);
3009 wait_for_completion(&blp->blp_comp);
3012 OBD_FREE(blp, sizeof(*blp));
3015 if (ldlm_state->ldlm_cb_service != NULL)
3016 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3017 #ifdef HAVE_SERVER_SUPPORT
3018 if (ldlm_state->ldlm_cancel_service != NULL)
3019 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3022 ldlm_proc_cleanup();
3024 #ifdef HAVE_SERVER_SUPPORT
3025 if (expired_lock_thread.elt_state != ELT_STOPPED) {
3026 expired_lock_thread.elt_state = ELT_TERMINATE;
3027 wake_up(&expired_lock_thread.elt_waitq);
3028 wait_event(expired_lock_thread.elt_waitq,
3029 expired_lock_thread.elt_state == ELT_STOPPED);
3033 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3041 mutex_init(&ldlm_ref_mutex);
3042 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
3043 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
3045 INIT_LIST_HEAD(&ldlm_srv_namespace_list);
3046 INIT_LIST_HEAD(&ldlm_cli_active_namespace_list);
3047 INIT_LIST_HEAD(&ldlm_cli_inactive_namespace_list);
3049 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3050 sizeof(struct ldlm_resource), 0,
3051 SLAB_HWCACHE_ALIGN, NULL);
3052 if (ldlm_resource_slab == NULL)
3055 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3056 sizeof(struct ldlm_lock), 0,
3057 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3058 if (ldlm_lock_slab == NULL) {
3059 kmem_cache_destroy(ldlm_resource_slab);
3063 ldlm_interval_slab = kmem_cache_create("interval_node",
3064 sizeof(struct ldlm_interval),
3065 0, SLAB_HWCACHE_ALIGN, NULL);
3066 if (ldlm_interval_slab == NULL) {
3067 kmem_cache_destroy(ldlm_resource_slab);
3068 kmem_cache_destroy(ldlm_lock_slab);
3071 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3072 class_export_dump_hook = ldlm_dump_export_locks;
3077 void ldlm_exit(void)
3080 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3081 kmem_cache_destroy(ldlm_resource_slab);
3082 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3083 * synchronize_rcu() to wait a grace period elapsed, so that
3084 * ldlm_lock_free() get a chance to be called. */
3086 kmem_cache_destroy(ldlm_lock_slab);
3087 kmem_cache_destroy(ldlm_interval_slab);