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, 2012, 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
45 # include <libcfs/libcfs.h>
47 # include <liblustre.h>
50 #include <lustre_dlm.h>
51 #include <obd_class.h>
52 #include <libcfs/list.h>
53 #include "ldlm_internal.h"
55 static int ldlm_num_threads;
56 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
57 "number of DLM service threads to start");
59 static char *ldlm_cpts;
60 CFS_MODULE_PARM(ldlm_cpts, "s", charp, 0444,
61 "CPU partitions ldlm threads should run on");
63 extern cfs_mem_cache_t *ldlm_resource_slab;
64 extern cfs_mem_cache_t *ldlm_lock_slab;
65 static struct mutex ldlm_ref_mutex;
66 static int ldlm_refcount;
68 struct ldlm_cb_async_args {
69 struct ldlm_cb_set_arg *ca_set_arg;
70 struct ldlm_lock *ca_lock;
75 static struct ldlm_state *ldlm_state;
77 inline cfs_time_t round_timeout(cfs_time_t timeout)
79 return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
82 /* timeout for initial callback (AST) reply (bz10399) */
83 static inline unsigned int ldlm_get_rq_timeout(void)
86 unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
88 return timeout < 1 ? 1 : timeout;
93 #define ELT_TERMINATE 2
99 * blp_prio_list is used for callbacks that should be handled
100 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
103 cfs_list_t blp_prio_list;
106 * blp_list is used for all other callbacks which are likely
107 * to take longer to process.
111 cfs_waitq_t blp_waitq;
112 struct completion blp_comp;
113 cfs_atomic_t blp_num_threads;
114 cfs_atomic_t blp_busy_threads;
119 struct ldlm_bl_work_item {
120 cfs_list_t blwi_entry;
121 struct ldlm_namespace *blwi_ns;
122 struct ldlm_lock_desc blwi_ld;
123 struct ldlm_lock *blwi_lock;
124 cfs_list_t blwi_head;
126 struct completion blwi_comp;
128 int blwi_mem_pressure;
131 #if defined(HAVE_SERVER_SUPPORT) && defined(__KERNEL__)
134 * Protects both waiting_locks_list and expired_lock_thread.
136 static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */
139 * List for contended locks.
141 * As soon as a lock is contended, it gets placed on this list and
142 * expected time to get a response is filled in the lock. A special
143 * thread walks the list looking for locks that should be released and
144 * schedules client evictions for those that have not been released in
147 * All access to it should be under waiting_locks_spinlock.
149 static cfs_list_t waiting_locks_list;
150 static cfs_timer_t waiting_locks_timer;
152 static struct expired_lock_thread {
153 cfs_waitq_t elt_waitq;
156 cfs_list_t elt_expired_locks;
157 } expired_lock_thread;
159 static inline int have_expired_locks(void)
164 spin_lock_bh(&waiting_locks_spinlock);
165 need_to_run = !cfs_list_empty(&expired_lock_thread.elt_expired_locks);
166 spin_unlock_bh(&waiting_locks_spinlock);
172 * Check expired lock list for expired locks and time them out.
174 static int expired_lock_main(void *arg)
176 cfs_list_t *expired = &expired_lock_thread.elt_expired_locks;
177 struct l_wait_info lwi = { 0 };
181 cfs_daemonize("ldlm_elt");
183 expired_lock_thread.elt_state = ELT_READY;
184 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
187 l_wait_event(expired_lock_thread.elt_waitq,
188 have_expired_locks() ||
189 expired_lock_thread.elt_state == ELT_TERMINATE,
192 spin_lock_bh(&waiting_locks_spinlock);
193 if (expired_lock_thread.elt_dump) {
194 struct libcfs_debug_msg_data msgdata = {
195 .msg_file = __FILE__,
196 .msg_fn = "waiting_locks_callback",
197 .msg_line = expired_lock_thread.elt_dump };
198 spin_unlock_bh(&waiting_locks_spinlock);
200 /* from waiting_locks_callback, but not in timer */
201 libcfs_debug_dumplog();
202 libcfs_run_lbug_upcall(&msgdata);
204 spin_lock_bh(&waiting_locks_spinlock);
205 expired_lock_thread.elt_dump = 0;
210 while (!cfs_list_empty(expired)) {
211 struct obd_export *export;
212 struct ldlm_lock *lock;
214 lock = cfs_list_entry(expired->next, struct ldlm_lock,
216 if ((void *)lock < LP_POISON + CFS_PAGE_SIZE &&
217 (void *)lock >= LP_POISON) {
218 spin_unlock_bh(&waiting_locks_spinlock);
219 CERROR("free lock on elt list %p\n", lock);
222 cfs_list_del_init(&lock->l_pending_chain);
223 if ((void *)lock->l_export < LP_POISON + CFS_PAGE_SIZE &&
224 (void *)lock->l_export >= LP_POISON) {
225 CERROR("lock with free export on elt list %p\n",
227 lock->l_export = NULL;
228 LDLM_ERROR(lock, "free export");
229 /* release extra ref grabbed by
230 * ldlm_add_waiting_lock() or
231 * ldlm_failed_ast() */
232 LDLM_LOCK_RELEASE(lock);
236 if (lock->l_destroyed) {
237 /* release the lock refcount where
238 * waiting_locks_callback() founds */
239 LDLM_LOCK_RELEASE(lock);
242 export = class_export_lock_get(lock->l_export, lock);
243 spin_unlock_bh(&waiting_locks_spinlock);
246 class_fail_export(export);
247 class_export_lock_put(export, lock);
249 /* release extra ref grabbed by ldlm_add_waiting_lock()
250 * or ldlm_failed_ast() */
251 LDLM_LOCK_RELEASE(lock);
253 spin_lock_bh(&waiting_locks_spinlock);
255 spin_unlock_bh(&waiting_locks_spinlock);
257 if (do_dump && obd_dump_on_eviction) {
258 CERROR("dump the log upon eviction\n");
259 libcfs_debug_dumplog();
262 if (expired_lock_thread.elt_state == ELT_TERMINATE)
266 expired_lock_thread.elt_state = ELT_STOPPED;
267 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
271 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
272 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds);
275 * Check if there is a request in the export request list
276 * which prevents the lock canceling.
278 static int ldlm_lock_busy(struct ldlm_lock *lock)
280 struct ptlrpc_request *req;
284 if (lock->l_export == NULL)
287 spin_lock_bh(&lock->l_export->exp_rpc_lock);
288 cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
290 if (req->rq_ops->hpreq_lock_match) {
291 match = req->rq_ops->hpreq_lock_match(req, lock);
296 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
300 /* This is called from within a timer interrupt and cannot schedule */
301 static void waiting_locks_callback(unsigned long unused)
303 struct ldlm_lock *lock;
306 spin_lock_bh(&waiting_locks_spinlock);
307 while (!cfs_list_empty(&waiting_locks_list)) {
308 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
310 if (cfs_time_after(lock->l_callback_timeout,
311 cfs_time_current()) ||
312 (lock->l_req_mode == LCK_GROUP))
315 if (ptlrpc_check_suspend()) {
316 /* there is a case when we talk to one mds, holding
317 * lock from another mds. this way we easily can get
318 * here, if second mds is being recovered. so, we
319 * suspend timeouts. bug 6019 */
321 LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
322 lock->l_export->exp_client_uuid.uuid,
323 lock->l_export->exp_connection->c_remote_uuid.uuid,
324 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
326 cfs_list_del_init(&lock->l_pending_chain);
327 if (lock->l_destroyed) {
328 /* relay the lock refcount decrease to
329 * expired lock thread */
330 cfs_list_add(&lock->l_pending_chain,
331 &expired_lock_thread.elt_expired_locks);
333 __ldlm_add_waiting_lock(lock,
334 ldlm_get_enq_timeout(lock));
339 /* if timeout overlaps the activation time of suspended timeouts
340 * then extend it to give a chance for client to reconnect */
341 if (cfs_time_before(cfs_time_sub(lock->l_callback_timeout,
342 cfs_time_seconds(obd_timeout)/2),
343 ptlrpc_suspend_wakeup_time())) {
344 LDLM_ERROR(lock, "extend timeout due to recovery: %s@%s nid %s ",
345 lock->l_export->exp_client_uuid.uuid,
346 lock->l_export->exp_connection->c_remote_uuid.uuid,
347 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
349 cfs_list_del_init(&lock->l_pending_chain);
350 if (lock->l_destroyed) {
351 /* relay the lock refcount decrease to
352 * expired lock thread */
353 cfs_list_add(&lock->l_pending_chain,
354 &expired_lock_thread.elt_expired_locks);
356 __ldlm_add_waiting_lock(lock,
357 ldlm_get_enq_timeout(lock));
362 /* Check if we need to prolong timeout */
363 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
364 ldlm_lock_busy(lock)) {
367 if (lock->l_pending_chain.next == &waiting_locks_list)
372 spin_unlock_bh(&waiting_locks_spinlock);
373 LDLM_DEBUG(lock, "prolong the busy lock");
374 ldlm_refresh_waiting_lock(lock,
375 ldlm_get_enq_timeout(lock));
376 spin_lock_bh(&waiting_locks_spinlock);
379 LDLM_LOCK_RELEASE(lock);
383 LDLM_LOCK_RELEASE(lock);
386 ldlm_lock_to_ns(lock)->ns_timeouts++;
387 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
388 "evicting client at %s ",
389 cfs_time_current_sec()- lock->l_last_activity,
391 lock->l_export->exp_connection->c_peer.nid));
393 /* no needs to take an extra ref on the lock since it was in
394 * the waiting_locks_list and ldlm_add_waiting_lock()
395 * already grabbed a ref */
396 cfs_list_del(&lock->l_pending_chain);
397 cfs_list_add(&lock->l_pending_chain,
398 &expired_lock_thread.elt_expired_locks);
402 if (!cfs_list_empty(&expired_lock_thread.elt_expired_locks)) {
403 if (obd_dump_on_timeout && need_dump)
404 expired_lock_thread.elt_dump = __LINE__;
406 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
410 * Make sure the timer will fire again if we have any locks
413 if (!cfs_list_empty(&waiting_locks_list)) {
414 cfs_time_t timeout_rounded;
415 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
417 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
418 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
420 spin_unlock_bh(&waiting_locks_spinlock);
424 * Add lock to the list of contended locks.
426 * Indicate that we're waiting for a client to call us back cancelling a given
427 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
428 * timer to fire appropriately. (We round up to the next second, to avoid
429 * floods of timer firings during periods of high lock contention and traffic).
430 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
431 * if it has been added to the waiting list (1 is returned).
433 * Called with the namespace lock held.
435 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
438 cfs_time_t timeout_rounded;
440 if (!cfs_list_empty(&lock->l_pending_chain))
443 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
444 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
447 timeout = cfs_time_shift(seconds);
448 if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
449 lock->l_callback_timeout = timeout;
451 timeout_rounded = round_timeout(lock->l_callback_timeout);
453 if (cfs_time_before(timeout_rounded,
454 cfs_timer_deadline(&waiting_locks_timer)) ||
455 !cfs_timer_is_armed(&waiting_locks_timer)) {
456 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
458 /* if the new lock has a shorter timeout than something earlier on
459 the list, we'll wait the longer amount of time; no big deal. */
461 cfs_list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
465 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
468 int timeout = ldlm_get_enq_timeout(lock);
470 /* NB: must be called with hold of lock_res_and_lock() */
471 LASSERT(lock->l_res_locked);
474 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
476 spin_lock_bh(&waiting_locks_spinlock);
477 if (lock->l_destroyed) {
478 static cfs_time_t next;
479 spin_unlock_bh(&waiting_locks_spinlock);
480 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
481 if (cfs_time_after(cfs_time_current(), next)) {
482 next = cfs_time_shift(14400);
483 libcfs_debug_dumpstack(NULL);
488 ret = __ldlm_add_waiting_lock(lock, timeout);
490 /* grab ref on the lock if it has been added to the
494 spin_unlock_bh(&waiting_locks_spinlock);
497 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
498 if (cfs_list_empty(&lock->l_exp_list))
499 cfs_list_add(&lock->l_exp_list,
500 &lock->l_export->exp_bl_list);
501 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
504 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
505 ret == 0 ? "not re-" : "", timeout,
506 AT_OFF ? "off" : "on");
511 * Remove a lock from the pending list, likely because it had its cancellation
512 * callback arrive without incident. This adjusts the lock-timeout timer if
513 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
514 * As done by ldlm_del_waiting_lock(), the caller must release the lock
515 * reference when the lock is removed from any list (1 is returned).
517 * Called with namespace lock held.
519 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
521 cfs_list_t *list_next;
523 if (cfs_list_empty(&lock->l_pending_chain))
526 list_next = lock->l_pending_chain.next;
527 if (lock->l_pending_chain.prev == &waiting_locks_list) {
528 /* Removing the head of the list, adjust timer. */
529 if (list_next == &waiting_locks_list) {
530 /* No more, just cancel. */
531 cfs_timer_disarm(&waiting_locks_timer);
533 struct ldlm_lock *next;
534 next = cfs_list_entry(list_next, struct ldlm_lock,
536 cfs_timer_arm(&waiting_locks_timer,
537 round_timeout(next->l_callback_timeout));
540 cfs_list_del_init(&lock->l_pending_chain);
545 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
549 if (lock->l_export == NULL) {
550 /* We don't have a "waiting locks list" on clients. */
551 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
555 spin_lock_bh(&waiting_locks_spinlock);
556 ret = __ldlm_del_waiting_lock(lock);
557 spin_unlock_bh(&waiting_locks_spinlock);
559 /* remove the lock out of export blocking list */
560 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
561 cfs_list_del_init(&lock->l_exp_list);
562 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
565 /* release lock ref if it has indeed been removed
567 LDLM_LOCK_RELEASE(lock);
570 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
573 EXPORT_SYMBOL(ldlm_del_waiting_lock);
576 * Prolong the contended lock waiting time.
578 * Called with namespace lock held.
580 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
582 if (lock->l_export == NULL) {
583 /* We don't have a "waiting locks list" on clients. */
584 LDLM_DEBUG(lock, "client lock: no-op");
588 spin_lock_bh(&waiting_locks_spinlock);
590 if (cfs_list_empty(&lock->l_pending_chain)) {
591 spin_unlock_bh(&waiting_locks_spinlock);
592 LDLM_DEBUG(lock, "wasn't waiting");
596 /* we remove/add the lock to the waiting list, so no needs to
597 * release/take a lock reference */
598 __ldlm_del_waiting_lock(lock);
599 __ldlm_add_waiting_lock(lock, timeout);
600 spin_unlock_bh(&waiting_locks_spinlock);
602 LDLM_DEBUG(lock, "refreshed");
605 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
607 #else /* !HAVE_SERVER_SUPPORT || !__KERNEL__ */
609 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
614 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
619 # ifdef HAVE_SERVER_SUPPORT
620 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
622 LASSERT(lock->l_res_locked);
623 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
628 #endif /* HAVE_SERVER_SUPPORT && __KERNEL__ */
630 #ifdef HAVE_SERVER_SUPPORT
633 * Perform lock cleanup if AST sending failed.
635 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
636 const char *ast_type)
638 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
639 "to a lock %s callback time out: rc %d\n",
640 lock->l_export->exp_obd->obd_name,
641 obd_export_nid2str(lock->l_export), ast_type, rc);
643 if (obd_dump_on_timeout)
644 libcfs_debug_dumplog();
646 spin_lock_bh(&waiting_locks_spinlock);
647 if (__ldlm_del_waiting_lock(lock) == 0)
648 /* the lock was not in any list, grab an extra ref before adding
649 * the lock to the expired list */
651 cfs_list_add(&lock->l_pending_chain,
652 &expired_lock_thread.elt_expired_locks);
653 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
654 spin_unlock_bh(&waiting_locks_spinlock);
656 class_fail_export(lock->l_export);
661 * Perform lock cleanup if AST reply came with error.
663 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
664 struct ptlrpc_request *req, int rc,
665 const char *ast_type)
667 lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
669 if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
670 LASSERT(lock->l_export);
671 if (lock->l_export->exp_libclient) {
672 LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
673 " timeout, just cancelling lock", ast_type,
674 libcfs_nid2str(peer.nid));
675 ldlm_lock_cancel(lock);
677 } else if (lock->l_flags & LDLM_FL_CANCEL) {
678 LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
679 "cancel was received (AST reply lost?)",
680 ast_type, libcfs_nid2str(peer.nid));
681 ldlm_lock_cancel(lock);
684 ldlm_del_waiting_lock(lock);
685 ldlm_failed_ast(lock, rc, ast_type);
689 struct ldlm_resource *res = lock->l_resource;
690 LDLM_DEBUG(lock, "client (nid %s) returned %d"
691 " from %s AST - normal race",
692 libcfs_nid2str(peer.nid),
694 lustre_msg_get_status(req->rq_repmsg) : -1,
697 /* update lvbo to return proper attributes.
699 ldlm_resource_getref(res);
700 ldlm_res_lvbo_update(res, NULL, 1);
701 ldlm_resource_putref(res);
705 LDLM_ERROR(lock, "client (nid %s) returned %d "
706 "from %s AST", libcfs_nid2str(peer.nid),
707 (req->rq_repmsg != NULL) ?
708 lustre_msg_get_status(req->rq_repmsg) : 0,
711 ldlm_lock_cancel(lock);
712 /* Server-side AST functions are called from ldlm_reprocess_all,
713 * which needs to be told to please restart its reprocessing. */
720 static int ldlm_cb_interpret(const struct lu_env *env,
721 struct ptlrpc_request *req, void *data, int rc)
723 struct ldlm_cb_async_args *ca = data;
724 struct ldlm_lock *lock = ca->ca_lock;
725 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
728 LASSERT(lock != NULL);
731 case LDLM_GL_CALLBACK:
732 /* Update the LVB from disk if the AST failed
733 * (this is a legal race)
735 * - Glimpse callback of local lock just returns
736 * -ELDLM_NO_LOCK_DATA.
737 * - Glimpse callback of remote lock might return
738 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
740 if (rc == -ELDLM_NO_LOCK_DATA) {
741 LDLM_DEBUG(lock, "lost race - client has a lock but no "
743 ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
744 } else if (rc != 0) {
745 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
747 rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
750 case LDLM_BL_CALLBACK:
752 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
754 case LDLM_CP_CALLBACK:
756 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
759 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
764 /* release extra reference taken in ldlm_ast_fini() */
765 LDLM_LOCK_RELEASE(lock);
768 cfs_atomic_inc(&arg->restart);
773 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
774 struct ldlm_cb_set_arg *arg,
775 struct ldlm_lock *lock,
781 if (unlikely(instant_cancel)) {
782 rc = ptl_send_rpc(req, 1);
783 ptlrpc_req_finished(req);
785 cfs_atomic_inc(&arg->restart);
788 ptlrpc_set_add_req(arg->set, req);
795 * Check if there are requests in the export request list which prevent
796 * the lock canceling and make these requests high priority ones.
798 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
800 struct ptlrpc_request *req;
803 if (lock->l_export == NULL) {
804 LDLM_DEBUG(lock, "client lock: no-op");
808 spin_lock_bh(&lock->l_export->exp_rpc_lock);
809 cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
811 /* Do not process requests that were not yet added to there
812 * incoming queue or were already removed from there for
814 if (!req->rq_hp && !cfs_list_empty(&req->rq_list) &&
815 req->rq_ops->hpreq_lock_match &&
816 req->rq_ops->hpreq_lock_match(req, lock))
817 ptlrpc_hpreq_reorder(req);
819 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
824 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
825 * enqueued server lock conflicts with given one.
827 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
828 * to wait for client response.
830 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
831 struct ldlm_lock_desc *desc,
832 void *data, int flag)
834 struct ldlm_cb_async_args *ca;
835 struct ldlm_cb_set_arg *arg = data;
836 struct ldlm_request *body;
837 struct ptlrpc_request *req;
838 int instant_cancel = 0;
842 if (flag == LDLM_CB_CANCELING)
843 /* Don't need to do anything here. */
847 LASSERT(data != NULL);
848 if (lock->l_export->exp_obd->obd_recovering != 0)
849 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
851 ldlm_lock_reorder_req(lock);
853 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
854 &RQF_LDLM_BL_CALLBACK,
855 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
859 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
860 ca = ptlrpc_req_async_args(req);
861 ca->ca_set_arg = arg;
864 req->rq_interpret_reply = ldlm_cb_interpret;
865 req->rq_no_resend = 1;
867 lock_res_and_lock(lock);
868 if (lock->l_granted_mode != lock->l_req_mode) {
869 /* this blocking AST will be communicated as part of the
870 * completion AST instead */
871 unlock_res_and_lock(lock);
873 ptlrpc_req_finished(req);
874 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
878 if (lock->l_destroyed) {
879 /* What's the point? */
880 unlock_res_and_lock(lock);
881 ptlrpc_req_finished(req);
885 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
888 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
889 body->lock_handle[0] = lock->l_remote_handle;
890 body->lock_desc = *desc;
891 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_AST_FLAGS);
893 LDLM_DEBUG(lock, "server preparing blocking AST");
895 ptlrpc_request_set_replen(req);
896 if (instant_cancel) {
897 unlock_res_and_lock(lock);
898 ldlm_lock_cancel(lock);
900 LASSERT(lock->l_granted_mode == lock->l_req_mode);
901 ldlm_add_waiting_lock(lock);
902 unlock_res_and_lock(lock);
905 req->rq_send_state = LUSTRE_IMP_FULL;
906 /* ptlrpc_request_alloc_pack already set timeout */
908 req->rq_timeout = ldlm_get_rq_timeout();
910 if (lock->l_export && lock->l_export->exp_nid_stats &&
911 lock->l_export->exp_nid_stats->nid_ldlm_stats)
912 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
913 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
915 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
919 EXPORT_SYMBOL(ldlm_server_blocking_ast);
922 * ->l_completion_ast callback for a remote lock in server namespace.
924 * Sends AST to the client notifying it of lock granting. If initial
925 * lock response was not sent yet, instead of sending another RPC, just
926 * mark the lock as granted and client will understand
928 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
930 struct ldlm_cb_set_arg *arg = data;
931 struct ldlm_request *body;
932 struct ptlrpc_request *req;
933 struct ldlm_cb_async_args *ca;
934 long total_enqueue_wait;
935 int instant_cancel = 0;
940 LASSERT(lock != NULL);
941 LASSERT(data != NULL);
943 total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(),
944 lock->l_last_activity);
946 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
947 &RQF_LDLM_CP_CALLBACK);
951 /* server namespace, doesn't need lock */
952 lvb_len = ldlm_lvbo_size(lock);
953 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
954 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
956 ptlrpc_request_free(req);
960 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
961 ca = ptlrpc_req_async_args(req);
962 ca->ca_set_arg = arg;
965 req->rq_interpret_reply = ldlm_cb_interpret;
966 req->rq_no_resend = 1;
967 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
969 body->lock_handle[0] = lock->l_remote_handle;
970 body->lock_flags = ldlm_flags_to_wire(flags);
971 ldlm_lock2desc(lock, &body->lock_desc);
973 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
975 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
976 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
977 lvb_len, RCL_CLIENT);
980 LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)",
983 /* Server-side enqueue wait time estimate, used in
984 __ldlm_add_waiting_lock to set future enqueue timers */
985 if (total_enqueue_wait < ldlm_get_enq_timeout(lock))
986 at_measured(ldlm_lock_to_ns_at(lock),
989 /* bz18618. Don't add lock enqueue time we spend waiting for a
990 previous callback to fail. Locks waiting legitimately will
991 get extended by ldlm_refresh_waiting_lock regardless of the
992 estimate, so it's okay to underestimate here. */
993 LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. "
994 "It is likely that a previous callback timed out.",
996 at_get(ldlm_lock_to_ns_at(lock)));
998 ptlrpc_request_set_replen(req);
1000 req->rq_send_state = LUSTRE_IMP_FULL;
1001 /* ptlrpc_request_pack already set timeout */
1003 req->rq_timeout = ldlm_get_rq_timeout();
1005 /* We only send real blocking ASTs after the lock is granted */
1006 lock_res_and_lock(lock);
1007 if (lock->l_flags & LDLM_FL_AST_SENT) {
1008 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1009 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1010 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1013 /* We might get here prior to ldlm_handle_enqueue setting
1014 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1015 * into waiting list, but this is safe and similar code in
1016 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1017 * that would not only cancel the lock, but will also remove
1018 * it from waiting list */
1019 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1020 unlock_res_and_lock(lock);
1021 ldlm_lock_cancel(lock);
1023 lock_res_and_lock(lock);
1025 /* start the lock-timeout clock */
1026 ldlm_add_waiting_lock(lock);
1029 unlock_res_and_lock(lock);
1031 if (lock->l_export && lock->l_export->exp_nid_stats &&
1032 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1033 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1034 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1036 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1040 EXPORT_SYMBOL(ldlm_server_completion_ast);
1043 * Server side ->l_glimpse_ast handler for client locks.
1045 * Sends glimpse AST to the client and waits for reply. Then updates
1046 * lvbo with the result.
1048 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1050 struct ldlm_cb_set_arg *arg = data;
1051 struct ldlm_request *body;
1052 struct ptlrpc_request *req;
1053 struct ldlm_cb_async_args *ca;
1055 struct req_format *req_fmt;
1058 LASSERT(lock != NULL);
1060 if (arg->gl_desc != NULL)
1061 /* There is a glimpse descriptor to pack */
1062 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1064 req_fmt = &RQF_LDLM_GL_CALLBACK;
1066 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1067 req_fmt, LUSTRE_DLM_VERSION,
1073 if (arg->gl_desc != NULL) {
1074 /* copy the GL descriptor */
1075 union ldlm_gl_desc *desc;
1076 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1077 *desc = *arg->gl_desc;
1080 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1081 body->lock_handle[0] = lock->l_remote_handle;
1082 ldlm_lock2desc(lock, &body->lock_desc);
1084 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1085 ca = ptlrpc_req_async_args(req);
1086 ca->ca_set_arg = arg;
1089 /* server namespace, doesn't need lock */
1090 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1091 ldlm_lvbo_size(lock));
1092 ptlrpc_request_set_replen(req);
1094 req->rq_send_state = LUSTRE_IMP_FULL;
1095 /* ptlrpc_request_alloc_pack already set timeout */
1097 req->rq_timeout = ldlm_get_rq_timeout();
1099 req->rq_interpret_reply = ldlm_cb_interpret;
1101 if (lock->l_export && lock->l_export->exp_nid_stats &&
1102 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1103 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1104 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1106 rc = ldlm_ast_fini(req, arg, lock, 0);
1110 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1112 int ldlm_glimpse_locks(struct ldlm_resource *res, cfs_list_t *gl_work_list)
1117 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1119 if (rc == -ERESTART)
1120 ldlm_reprocess_all(res);
1124 EXPORT_SYMBOL(ldlm_glimpse_locks);
1126 /* return LDLM lock associated with a lock callback request */
1127 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1129 struct ldlm_cb_async_args *ca;
1130 struct ldlm_lock *lock;
1133 ca = ptlrpc_req_async_args(req);
1136 RETURN(ERR_PTR(-EFAULT));
1140 EXPORT_SYMBOL(ldlm_request_lock);
1142 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1143 struct lprocfs_stats *srv_stats)
1145 int lock_type = 0, op = 0;
1147 lock_type = dlm_req->lock_desc.l_resource.lr_type;
1149 switch (lock_type) {
1151 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1154 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1155 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1157 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1160 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1163 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1171 lprocfs_counter_incr(srv_stats, op);
1177 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1178 * service threads to carry out client lock enqueueing requests.
1180 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1181 struct ptlrpc_request *req,
1182 const struct ldlm_request *dlm_req,
1183 const struct ldlm_callback_suite *cbs)
1185 struct ldlm_reply *dlm_rep;
1187 ldlm_error_t err = ELDLM_OK;
1188 struct ldlm_lock *lock = NULL;
1189 void *cookie = NULL;
1193 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1195 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1196 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1198 LASSERT(req->rq_export);
1200 if (ptlrpc_req2svc(req)->srv_stats != NULL)
1201 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1203 if (req->rq_export && req->rq_export->exp_nid_stats &&
1204 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1205 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1206 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1208 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1209 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1210 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1211 dlm_req->lock_desc.l_resource.lr_type);
1212 GOTO(out, rc = -EFAULT);
1215 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1216 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1217 dlm_req->lock_desc.l_req_mode &
1218 (dlm_req->lock_desc.l_req_mode-1))) {
1219 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1220 dlm_req->lock_desc.l_req_mode);
1221 GOTO(out, rc = -EFAULT);
1224 if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
1225 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1227 DEBUG_REQ(D_ERROR, req,
1228 "PLAIN lock request from IBITS client?");
1229 GOTO(out, rc = -EPROTO);
1231 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1233 DEBUG_REQ(D_ERROR, req,
1234 "IBITS lock request from unaware client?");
1235 GOTO(out, rc = -EPROTO);
1239 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1240 against server's _CONNECT_SUPPORTED flags? (I don't want to use
1241 ibits for mgc/mgs) */
1243 /* INODEBITS_INTEROP: Perform conversion from plain lock to
1244 * inodebits lock if client does not support them. */
1245 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
1246 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1247 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1248 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1249 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1250 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1251 dlm_req->lock_desc.l_req_mode = LCK_CR;
1255 if (unlikely(flags & LDLM_FL_REPLAY)) {
1256 /* Find an existing lock in the per-export lock hash */
1257 /* In the function below, .hs_keycmp resolves to
1258 * ldlm_export_lock_keycmp() */
1259 /* coverity[overrun-buffer-val] */
1260 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1261 (void *)&dlm_req->lock_handle[0]);
1263 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1264 LPX64, lock->l_handle.h_cookie);
1265 GOTO(existing_lock, rc = 0);
1269 /* The lock's callback data might be set in the policy function */
1270 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1271 dlm_req->lock_desc.l_resource.lr_type,
1272 dlm_req->lock_desc.l_req_mode,
1273 cbs, NULL, 0, LVB_T_NONE);
1275 GOTO(out, rc = -ENOMEM);
1277 lock->l_last_activity = cfs_time_current_sec();
1278 lock->l_remote_handle = dlm_req->lock_handle[0];
1279 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1281 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1282 /* Don't enqueue a lock onto the export if it is been disonnected
1283 * due to eviction (bug 3822) or server umount (bug 24324).
1284 * Cancel it now instead. */
1285 if (req->rq_export->exp_disconnected) {
1286 LDLM_ERROR(lock, "lock on disconnected export %p",
1288 GOTO(out, rc = -ENOTCONN);
1291 lock->l_export = class_export_lock_get(req->rq_export, lock);
1292 if (lock->l_export->exp_lock_hash)
1293 cfs_hash_add(lock->l_export->exp_lock_hash,
1294 &lock->l_remote_handle,
1299 if (flags & LDLM_FL_HAS_INTENT) {
1300 /* In this case, the reply buffer is allocated deep in
1301 * local_lock_enqueue by the policy function. */
1304 /* based on the assumption that lvb size never changes during
1305 * resource life time otherwise it need resource->lr_lock's
1307 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1308 RCL_SERVER, ldlm_lvbo_size(lock));
1310 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1311 GOTO(out, rc = -ENOMEM);
1313 rc = req_capsule_server_pack(&req->rq_pill);
1318 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1319 ldlm_convert_policy_to_local(req->rq_export,
1320 dlm_req->lock_desc.l_resource.lr_type,
1321 &dlm_req->lock_desc.l_policy_data,
1322 &lock->l_policy_data);
1323 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1324 lock->l_req_extent = lock->l_policy_data.l_extent;
1326 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1330 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1331 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1333 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1334 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1336 /* We never send a blocking AST until the lock is granted, but
1337 * we can tell it right now */
1338 lock_res_and_lock(lock);
1340 /* Now take into account flags to be inherited from original lock
1341 request both in reply to client and in our own lock flags. */
1342 dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1343 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1344 LDLM_INHERIT_FLAGS);
1346 /* Don't move a pending lock onto the export if it has already been
1347 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1348 * Cancel it now instead. */
1349 if (unlikely(req->rq_export->exp_disconnected ||
1350 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1351 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1353 } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1354 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1355 if (lock->l_granted_mode == lock->l_req_mode) {
1357 * Only cancel lock if it was granted, because it would
1358 * be destroyed immediately and would never be granted
1359 * in the future, causing timeouts on client. Not
1360 * granted lock will be cancelled immediately after
1361 * sending completion AST.
1363 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1364 unlock_res_and_lock(lock);
1365 ldlm_lock_cancel(lock);
1366 lock_res_and_lock(lock);
1368 ldlm_add_waiting_lock(lock);
1371 /* Make sure we never ever grant usual metadata locks to liblustre
1373 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1374 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1375 req->rq_export->exp_libclient) {
1376 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1377 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1378 CERROR("Granting sync lock to libclient. "
1379 "req fl %d, rep fl %d, lock fl "LPX64"\n",
1380 dlm_req->lock_flags, dlm_rep->lock_flags,
1382 LDLM_ERROR(lock, "sync lock");
1383 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1384 struct ldlm_intent *it;
1386 it = req_capsule_client_get(&req->rq_pill,
1389 CERROR("This is intent %s ("LPU64")\n",
1390 ldlm_it2str(it->opc), it->opc);
1396 unlock_res_and_lock(lock);
1400 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1401 if (!req->rq_packed_final) {
1402 err = lustre_pack_reply(req, 1, NULL, NULL);
1407 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1408 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1410 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1411 "(err=%d, rc=%d)", err, rc);
1414 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1416 ldlm_lvbo_size(lock) > 0) {
1420 buf = req_capsule_server_get(&req->rq_pill,
1422 LASSERTF(buf != NULL, "req %p, lock %p\n",
1424 buflen = req_capsule_get_size(&req->rq_pill,
1425 &RMF_DLM_LVB, RCL_SERVER);
1426 buflen = ldlm_lvbo_fill(lock, buf, buflen);
1427 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
1428 buflen, RCL_SERVER);
1431 lock_res_and_lock(lock);
1432 ldlm_resource_unlink_lock(lock);
1433 ldlm_lock_destroy_nolock(lock);
1434 unlock_res_and_lock(lock);
1437 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1438 ldlm_reprocess_all(lock->l_resource);
1440 LDLM_LOCK_RELEASE(lock);
1443 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1448 EXPORT_SYMBOL(ldlm_handle_enqueue0);
1451 * Old-style LDLM main entry point for server code enqueue.
1453 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1454 ldlm_completion_callback completion_callback,
1455 ldlm_blocking_callback blocking_callback,
1456 ldlm_glimpse_callback glimpse_callback)
1458 struct ldlm_request *dlm_req;
1459 struct ldlm_callback_suite cbs = {
1460 .lcs_completion = completion_callback,
1461 .lcs_blocking = blocking_callback,
1462 .lcs_glimpse = glimpse_callback
1466 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1467 if (dlm_req != NULL) {
1468 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1469 req, dlm_req, &cbs);
1475 EXPORT_SYMBOL(ldlm_handle_enqueue);
1478 * Main LDLM entry point for server code to process lock conversion requests.
1480 int ldlm_handle_convert0(struct ptlrpc_request *req,
1481 const struct ldlm_request *dlm_req)
1483 struct ldlm_reply *dlm_rep;
1484 struct ldlm_lock *lock;
1488 if (req->rq_export && req->rq_export->exp_nid_stats &&
1489 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1490 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1491 LDLM_CONVERT - LDLM_FIRST_OPC);
1493 rc = req_capsule_server_pack(&req->rq_pill);
1497 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1498 dlm_rep->lock_flags = dlm_req->lock_flags;
1500 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1502 req->rq_status = EINVAL;
1506 LDLM_DEBUG(lock, "server-side convert handler START");
1508 lock->l_last_activity = cfs_time_current_sec();
1509 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1510 &dlm_rep->lock_flags);
1512 if (ldlm_del_waiting_lock(lock))
1513 LDLM_DEBUG(lock, "converted waiting lock");
1516 req->rq_status = EDEADLOCK;
1521 if (!req->rq_status)
1522 ldlm_reprocess_all(lock->l_resource);
1523 LDLM_DEBUG(lock, "server-side convert handler END");
1524 LDLM_LOCK_PUT(lock);
1526 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1530 EXPORT_SYMBOL(ldlm_handle_convert0);
1533 * Old-style main LDLM entry point for server code to process lock conversion
1536 int ldlm_handle_convert(struct ptlrpc_request *req)
1539 struct ldlm_request *dlm_req;
1541 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1542 if (dlm_req != NULL) {
1543 rc = ldlm_handle_convert0(req, dlm_req);
1545 CERROR ("Can't unpack dlm_req\n");
1550 EXPORT_SYMBOL(ldlm_handle_convert);
1553 * Cancel all the locks whose handles are packed into ldlm_request
1555 * Called by server code expecting such combined cancel activity
1558 int ldlm_request_cancel(struct ptlrpc_request *req,
1559 const struct ldlm_request *dlm_req, int first)
1561 struct ldlm_resource *res, *pres = NULL;
1562 struct ldlm_lock *lock;
1563 int i, count, done = 0;
1566 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1570 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1573 /* There is no lock on the server at the replay time,
1574 * skip lock cancelling to make replay tests to pass. */
1575 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1578 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1579 "starting at %d", count, first);
1581 for (i = first; i < count; i++) {
1582 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1584 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1585 "lock (cookie "LPU64")",
1586 dlm_req->lock_handle[i].cookie);
1590 res = lock->l_resource;
1593 /* This code is an optimization to only attempt lock
1594 * granting on the resource (that could be CPU-expensive)
1595 * after we are done cancelling lock in that resource. */
1598 ldlm_reprocess_all(pres);
1599 LDLM_RESOURCE_DELREF(pres);
1600 ldlm_resource_putref(pres);
1603 ldlm_resource_getref(res);
1604 LDLM_RESOURCE_ADDREF(res);
1605 ldlm_res_lvbo_update(res, NULL, 1);
1609 ldlm_lock_cancel(lock);
1610 LDLM_LOCK_PUT(lock);
1613 ldlm_reprocess_all(pres);
1614 LDLM_RESOURCE_DELREF(pres);
1615 ldlm_resource_putref(pres);
1617 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1620 EXPORT_SYMBOL(ldlm_request_cancel);
1623 * Main LDLM entry point for server code to cancel locks.
1625 * Typically gets called from service handler on LDLM_CANCEL opc.
1627 int ldlm_handle_cancel(struct ptlrpc_request *req)
1629 struct ldlm_request *dlm_req;
1633 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1634 if (dlm_req == NULL) {
1635 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1639 if (req->rq_export && req->rq_export->exp_nid_stats &&
1640 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1641 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1642 LDLM_CANCEL - LDLM_FIRST_OPC);
1644 rc = req_capsule_server_pack(&req->rq_pill);
1648 if (!ldlm_request_cancel(req, dlm_req, 0))
1649 req->rq_status = ESTALE;
1651 RETURN(ptlrpc_reply(req));
1653 EXPORT_SYMBOL(ldlm_handle_cancel);
1654 #endif /* HAVE_SERVER_SUPPORT */
1657 * Callback handler for receiving incoming blocking ASTs.
1659 * This can only happen on client side.
1661 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1662 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1667 LDLM_DEBUG(lock, "client blocking AST callback handler");
1669 lock_res_and_lock(lock);
1670 lock->l_flags |= LDLM_FL_CBPENDING;
1672 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1673 lock->l_flags |= LDLM_FL_CANCEL;
1675 do_ast = (!lock->l_readers && !lock->l_writers);
1676 unlock_res_and_lock(lock);
1679 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1680 lock, lock->l_blocking_ast);
1681 if (lock->l_blocking_ast != NULL)
1682 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1685 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1689 LDLM_DEBUG(lock, "client blocking callback handler END");
1690 LDLM_LOCK_RELEASE(lock);
1695 * Callback handler for receiving incoming completion ASTs.
1697 * This only can happen on client side.
1699 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1700 struct ldlm_namespace *ns,
1701 struct ldlm_request *dlm_req,
1702 struct ldlm_lock *lock)
1705 CFS_LIST_HEAD(ast_list);
1709 LDLM_DEBUG(lock, "client completion callback handler START");
1711 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1712 int to = cfs_time_seconds(1);
1714 cfs_schedule_timeout_and_set_state(
1715 CFS_TASK_INTERRUPTIBLE, to);
1716 if (lock->l_granted_mode == lock->l_req_mode ||
1722 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1724 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1725 GOTO(out, rc = lvb_len);
1726 } else if (lvb_len > 0) {
1727 if (lock->l_lvb_len > 0) {
1728 /* for extent lock, lvb contains ost_lvb{}. */
1729 LASSERT(lock->l_lvb_data != NULL);
1731 if (unlikely(lock->l_lvb_len < lvb_len)) {
1732 LDLM_ERROR(lock, "Replied LVB is larger than "
1733 "expectation, expected = %d, "
1735 lock->l_lvb_len, lvb_len);
1736 GOTO(out, rc = -EINVAL);
1738 } else if (ldlm_has_layout(lock)) { /* for layout lock, lvb has
1739 * variable length */
1742 OBD_ALLOC(lvb_data, lvb_len);
1743 if (lvb_data == NULL) {
1744 LDLM_ERROR(lock, "No memory: %d.\n", lvb_len);
1745 GOTO(out, rc = -ENOMEM);
1748 lock_res_and_lock(lock);
1749 LASSERT(lock->l_lvb_data == NULL);
1750 lock->l_lvb_data = lvb_data;
1751 lock->l_lvb_len = lvb_len;
1752 unlock_res_and_lock(lock);
1756 lock_res_and_lock(lock);
1757 if (lock->l_destroyed ||
1758 lock->l_granted_mode == lock->l_req_mode) {
1759 /* bug 11300: the lock has already been granted */
1760 unlock_res_and_lock(lock);
1761 LDLM_DEBUG(lock, "Double grant race happened");
1765 /* If we receive the completion AST before the actual enqueue returned,
1766 * then we might need to switch lock modes, resources, or extents. */
1767 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1768 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1769 LDLM_DEBUG(lock, "completion AST, new lock mode");
1772 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1773 ldlm_convert_policy_to_local(req->rq_export,
1774 dlm_req->lock_desc.l_resource.lr_type,
1775 &dlm_req->lock_desc.l_policy_data,
1776 &lock->l_policy_data);
1777 LDLM_DEBUG(lock, "completion AST, new policy data");
1780 ldlm_resource_unlink_lock(lock);
1781 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1782 &lock->l_resource->lr_name,
1783 sizeof(lock->l_resource->lr_name)) != 0) {
1784 unlock_res_and_lock(lock);
1785 rc = ldlm_lock_change_resource(ns, lock,
1786 &dlm_req->lock_desc.l_resource.lr_name);
1788 LDLM_ERROR(lock, "Failed to allocate resource");
1791 LDLM_DEBUG(lock, "completion AST, new resource");
1792 CERROR("change resource!\n");
1793 lock_res_and_lock(lock);
1796 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1797 /* BL_AST locks are not needed in LRU.
1798 * Let ldlm_cancel_lru() be fast. */
1799 ldlm_lock_remove_from_lru(lock);
1800 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1801 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1804 if (lock->l_lvb_len > 0) {
1805 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1806 lock->l_lvb_data, lvb_len);
1808 unlock_res_and_lock(lock);
1813 ldlm_grant_lock(lock, &ast_list);
1814 unlock_res_and_lock(lock);
1816 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1818 /* Let Enqueue to call osc_lock_upcall() and initialize
1820 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1822 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1824 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1830 lock_res_and_lock(lock);
1831 lock->l_flags |= LDLM_FL_FAILED;
1832 unlock_res_and_lock(lock);
1833 cfs_waitq_signal(&lock->l_waitq);
1835 LDLM_LOCK_RELEASE(lock);
1839 * Callback handler for receiving incoming glimpse ASTs.
1841 * This only can happen on client side. After handling the glimpse AST
1842 * we also consider dropping the lock here if it is unused locally for a
1845 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1846 struct ldlm_namespace *ns,
1847 struct ldlm_request *dlm_req,
1848 struct ldlm_lock *lock)
1853 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1855 if (lock->l_glimpse_ast != NULL)
1856 rc = lock->l_glimpse_ast(lock, req);
1858 if (req->rq_repmsg != NULL) {
1861 req->rq_status = rc;
1865 lock_res_and_lock(lock);
1866 if (lock->l_granted_mode == LCK_PW &&
1867 !lock->l_readers && !lock->l_writers &&
1868 cfs_time_after(cfs_time_current(),
1869 cfs_time_add(lock->l_last_used,
1870 cfs_time_seconds(10)))) {
1871 unlock_res_and_lock(lock);
1872 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1873 ldlm_handle_bl_callback(ns, NULL, lock);
1878 unlock_res_and_lock(lock);
1879 LDLM_LOCK_RELEASE(lock);
1883 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1885 if (req->rq_no_reply)
1888 req->rq_status = rc;
1889 if (!req->rq_packed_final) {
1890 rc = lustre_pack_reply(req, 1, NULL, NULL);
1894 return ptlrpc_reply(req);
1898 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi, int mode)
1900 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1903 spin_lock(&blp->blp_lock);
1904 if (blwi->blwi_lock &&
1905 blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) {
1906 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1907 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1909 /* other blocking callbacks are added to the regular list */
1910 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1912 spin_unlock(&blp->blp_lock);
1914 cfs_waitq_signal(&blp->blp_waitq);
1916 /* can not use blwi->blwi_mode as blwi could be already freed in
1918 if (mode == LDLM_SYNC)
1919 wait_for_completion(&blwi->blwi_comp);
1924 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1925 struct ldlm_namespace *ns,
1926 struct ldlm_lock_desc *ld,
1927 cfs_list_t *cancels, int count,
1928 struct ldlm_lock *lock,
1931 init_completion(&blwi->blwi_comp);
1932 CFS_INIT_LIST_HEAD(&blwi->blwi_head);
1934 if (cfs_memory_pressure_get())
1935 blwi->blwi_mem_pressure = 1;
1938 blwi->blwi_mode = mode;
1940 blwi->blwi_ld = *ld;
1942 cfs_list_add(&blwi->blwi_head, cancels);
1943 cfs_list_del_init(cancels);
1944 blwi->blwi_count = count;
1946 blwi->blwi_lock = lock;
1951 * Queues a list of locks \a cancels containing \a count locks
1952 * for later processing by a blocking thread. If \a count is zero,
1953 * then the lock referenced as \a lock is queued instead.
1955 * The blocking thread would then call ->l_blocking_ast callback in the lock.
1956 * If list addition fails an error is returned and caller is supposed to
1957 * call ->l_blocking_ast itself.
1959 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1960 struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
1961 cfs_list_t *cancels, int count, int mode)
1965 if (cancels && count == 0)
1968 if (mode == LDLM_SYNC) {
1969 /* if it is synchronous call do minimum mem alloc, as it could
1970 * be triggered from kernel shrinker
1972 struct ldlm_bl_work_item blwi;
1973 memset(&blwi, 0, sizeof(blwi));
1974 init_blwi(&blwi, ns, ld, cancels, count, lock, LDLM_SYNC);
1975 RETURN(__ldlm_bl_to_thread(&blwi, LDLM_SYNC));
1977 struct ldlm_bl_work_item *blwi;
1978 OBD_ALLOC(blwi, sizeof(*blwi));
1981 init_blwi(blwi, ns, ld, cancels, count, lock, LDLM_ASYNC);
1983 RETURN(__ldlm_bl_to_thread(blwi, LDLM_ASYNC));
1989 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1990 struct ldlm_lock *lock)
1993 RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LDLM_ASYNC));
1999 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2000 cfs_list_t *cancels, int count, int mode)
2003 RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count, mode));
2009 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2010 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2012 struct obd_device *obd = req->rq_export->exp_obd;
2019 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2021 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2023 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2025 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2028 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2030 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2032 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2035 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2038 /* We are responsible for swabbing contents of val */
2040 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2041 /* Pass it on to mdc (the "export" in this case) */
2042 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2044 sizeof(KEY_HSM_COPYTOOL_SEND),
2045 KEY_HSM_COPYTOOL_SEND,
2048 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2053 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2054 const char *msg, int rc,
2055 struct lustre_handle *handle)
2057 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2058 "%s: [nid %s] [rc %d] [lock "LPX64"]",
2059 msg, libcfs_id2str(req->rq_peer), rc,
2060 handle ? handle->cookie : 0);
2061 if (req->rq_no_reply)
2062 CWARN("No reply was sent, maybe cause bug 21636.\n");
2064 CWARN("Send reply failed, maybe cause bug 21636.\n");
2067 static int ldlm_handle_qc_callback(struct ptlrpc_request *req)
2069 struct obd_quotactl *oqctl;
2070 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2072 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2073 if (oqctl == NULL) {
2074 CERROR("Can't unpack obd_quotactl\n");
2078 cli->cl_qchk_stat = oqctl->qc_stat;
2082 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2083 static int ldlm_callback_handler(struct ptlrpc_request *req)
2085 struct ldlm_namespace *ns;
2086 struct ldlm_request *dlm_req;
2087 struct ldlm_lock *lock;
2091 /* Requests arrive in sender's byte order. The ptlrpc service
2092 * handler has already checked and, if necessary, byte-swapped the
2093 * incoming request message body, but I am responsible for the
2094 * message buffers. */
2096 /* do nothing for sec context finalize */
2097 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2100 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2102 if (req->rq_export == NULL) {
2103 rc = ldlm_callback_reply(req, -ENOTCONN);
2104 ldlm_callback_errmsg(req, "Operate on unconnected server",
2109 LASSERT(req->rq_export != NULL);
2110 LASSERT(req->rq_export->exp_obd != NULL);
2112 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2113 case LDLM_BL_CALLBACK:
2114 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET))
2117 case LDLM_CP_CALLBACK:
2118 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2121 case LDLM_GL_CALLBACK:
2122 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2126 rc = ldlm_handle_setinfo(req);
2127 ldlm_callback_reply(req, rc);
2129 case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
2130 CERROR("shouldn't be handling OBD_LOG_CANCEL on DLM thread\n");
2131 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2132 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2134 rc = llog_origin_handle_cancel(req);
2135 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2137 ldlm_callback_reply(req, rc);
2139 case LLOG_ORIGIN_HANDLE_CREATE:
2140 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2141 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2143 rc = llog_origin_handle_open(req);
2144 ldlm_callback_reply(req, rc);
2146 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2147 req_capsule_set(&req->rq_pill,
2148 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2149 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2151 rc = llog_origin_handle_next_block(req);
2152 ldlm_callback_reply(req, rc);
2154 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2155 req_capsule_set(&req->rq_pill,
2156 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2157 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2159 rc = llog_origin_handle_read_header(req);
2160 ldlm_callback_reply(req, rc);
2162 case LLOG_ORIGIN_HANDLE_CLOSE:
2163 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2165 rc = llog_origin_handle_close(req);
2166 ldlm_callback_reply(req, rc);
2168 case OBD_QC_CALLBACK:
2169 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
2170 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
2172 rc = ldlm_handle_qc_callback(req);
2173 ldlm_callback_reply(req, rc);
2176 CERROR("unknown opcode %u\n",
2177 lustre_msg_get_opc(req->rq_reqmsg));
2178 ldlm_callback_reply(req, -EPROTO);
2182 ns = req->rq_export->exp_obd->obd_namespace;
2183 LASSERT(ns != NULL);
2185 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2187 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2188 if (dlm_req == NULL) {
2189 rc = ldlm_callback_reply(req, -EPROTO);
2190 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2195 /* Force a known safe race, send a cancel to the server for a lock
2196 * which the server has already started a blocking callback on. */
2197 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2198 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2199 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]);
2201 CERROR("ldlm_cli_cancel: %d\n", rc);
2204 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2206 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2207 "disappeared\n", dlm_req->lock_handle[0].cookie);
2208 rc = ldlm_callback_reply(req, -EINVAL);
2209 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2210 &dlm_req->lock_handle[0]);
2214 if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
2215 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2216 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2218 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2219 lock_res_and_lock(lock);
2220 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2222 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2223 /* If somebody cancels lock and cache is already dropped,
2224 * or lock is failed before cp_ast received on client,
2225 * we can tell the server we have no lock. Otherwise, we
2226 * should send cancel after dropping the cache. */
2227 if (((lock->l_flags & LDLM_FL_CANCELING) &&
2228 (lock->l_flags & LDLM_FL_BL_DONE)) ||
2229 (lock->l_flags & LDLM_FL_FAILED)) {
2230 LDLM_DEBUG(lock, "callback on lock "
2231 LPX64" - lock disappeared\n",
2232 dlm_req->lock_handle[0].cookie);
2233 unlock_res_and_lock(lock);
2234 LDLM_LOCK_RELEASE(lock);
2235 rc = ldlm_callback_reply(req, -EINVAL);
2236 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2237 &dlm_req->lock_handle[0]);
2240 /* BL_AST locks are not needed in LRU.
2241 * Let ldlm_cancel_lru() be fast. */
2242 ldlm_lock_remove_from_lru(lock);
2243 lock->l_flags |= LDLM_FL_BL_AST;
2245 unlock_res_and_lock(lock);
2247 /* We want the ost thread to get this reply so that it can respond
2248 * to ost requests (write cache writeback) that might be triggered
2251 * But we'd also like to be able to indicate in the reply that we're
2252 * cancelling right now, because it's unused, or have an intent result
2253 * in the reply, so we might have to push the responsibility for sending
2254 * the reply down into the AST handlers, alas. */
2256 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2257 case LDLM_BL_CALLBACK:
2258 CDEBUG(D_INODE, "blocking ast\n");
2259 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2260 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) {
2261 rc = ldlm_callback_reply(req, 0);
2262 if (req->rq_no_reply || rc)
2263 ldlm_callback_errmsg(req, "Normal process", rc,
2264 &dlm_req->lock_handle[0]);
2266 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2267 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2269 case LDLM_CP_CALLBACK:
2270 CDEBUG(D_INODE, "completion ast\n");
2271 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2272 ldlm_callback_reply(req, 0);
2273 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2275 case LDLM_GL_CALLBACK:
2276 CDEBUG(D_INODE, "glimpse ast\n");
2277 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2278 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2281 LBUG(); /* checked above */
2287 #ifdef HAVE_SERVER_SUPPORT
2289 * Main handler for canceld thread.
2291 * Separated into its own thread to avoid deadlocks.
2293 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2298 /* Requests arrive in sender's byte order. The ptlrpc service
2299 * handler has already checked and, if necessary, byte-swapped the
2300 * incoming request message body, but I am responsible for the
2301 * message buffers. */
2303 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2305 if (req->rq_export == NULL) {
2306 struct ldlm_request *dlm_req;
2308 CERROR("%s from %s arrived at %lu with bad export cookie "
2310 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2311 libcfs_nid2str(req->rq_peer.nid),
2312 req->rq_arrival_time.tv_sec,
2313 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2315 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2316 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2317 dlm_req = req_capsule_client_get(&req->rq_pill,
2319 if (dlm_req != NULL)
2320 ldlm_lock_dump_handle(D_ERROR,
2321 &dlm_req->lock_handle[0]);
2323 ldlm_callback_reply(req, -ENOTCONN);
2327 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2329 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2331 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2332 CDEBUG(D_INODE, "cancel\n");
2333 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2334 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))
2336 rc = ldlm_handle_cancel(req);
2340 case OBD_LOG_CANCEL:
2341 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2342 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2344 rc = llog_origin_handle_cancel(req);
2345 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2347 ldlm_callback_reply(req, rc);
2350 CERROR("invalid opcode %d\n",
2351 lustre_msg_get_opc(req->rq_reqmsg));
2352 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2353 ldlm_callback_reply(req, -EINVAL);
2359 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2360 struct ldlm_lock *lock)
2362 struct ldlm_request *dlm_req;
2363 struct lustre_handle lockh;
2368 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2369 if (dlm_req == NULL)
2372 ldlm_lock2handle(lock, &lockh);
2373 for (i = 0; i < dlm_req->lock_count; i++) {
2374 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2376 DEBUG_REQ(D_RPCTRACE, req,
2377 "Prio raised by lock "LPX64".", lockh.cookie);
2388 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2390 struct ldlm_request *dlm_req;
2395 /* no prolong in recovery */
2396 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2399 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2400 if (dlm_req == NULL)
2403 for (i = 0; i < dlm_req->lock_count; i++) {
2404 struct ldlm_lock *lock;
2406 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2410 rc = !!(lock->l_flags & LDLM_FL_AST_SENT);
2412 LDLM_DEBUG(lock, "hpreq cancel lock");
2413 LDLM_LOCK_PUT(lock);
2422 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2423 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2424 .hpreq_check = ldlm_cancel_hpreq_check,
2428 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2432 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2434 if (req->rq_export == NULL)
2437 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2438 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2439 req->rq_ops = &ldlm_cancel_hpreq_ops;
2444 int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2445 cfs_hlist_node_t *hnode, void *data)
2448 cfs_list_t *rpc_list = data;
2449 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2451 lock_res_and_lock(lock);
2453 if (lock->l_req_mode != lock->l_granted_mode) {
2454 unlock_res_and_lock(lock);
2458 LASSERT(lock->l_resource);
2459 if (lock->l_resource->lr_type != LDLM_IBITS &&
2460 lock->l_resource->lr_type != LDLM_PLAIN) {
2461 unlock_res_and_lock(lock);
2465 if (lock->l_flags & LDLM_FL_AST_SENT) {
2466 unlock_res_and_lock(lock);
2470 LASSERT(lock->l_blocking_ast);
2471 LASSERT(!lock->l_blocking_lock);
2473 lock->l_flags |= LDLM_FL_AST_SENT;
2474 if (lock->l_export && lock->l_export->exp_lock_hash) {
2475 /* NB: it's safe to call cfs_hash_del() even lock isn't
2476 * in exp_lock_hash. */
2477 /* In the function below, .hs_keycmp resolves to
2478 * ldlm_export_lock_keycmp() */
2479 /* coverity[overrun-buffer-val] */
2480 cfs_hash_del(lock->l_export->exp_lock_hash,
2481 &lock->l_remote_handle, &lock->l_exp_hash);
2484 cfs_list_add_tail(&lock->l_rk_ast, rpc_list);
2485 LDLM_LOCK_GET(lock);
2487 unlock_res_and_lock(lock);
2491 void ldlm_revoke_export_locks(struct obd_export *exp)
2493 cfs_list_t rpc_list;
2496 CFS_INIT_LIST_HEAD(&rpc_list);
2497 cfs_hash_for_each_empty(exp->exp_lock_hash,
2498 ldlm_revoke_lock_cb, &rpc_list);
2499 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2500 LDLM_WORK_REVOKE_AST);
2504 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2505 #endif /* HAVE_SERVER_SUPPORT */
2508 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2510 struct ldlm_bl_work_item *blwi = NULL;
2511 static unsigned int num_bl = 0;
2513 spin_lock(&blp->blp_lock);
2514 /* process a request from the blp_list at least every blp_num_threads */
2515 if (!cfs_list_empty(&blp->blp_list) &&
2516 (cfs_list_empty(&blp->blp_prio_list) || num_bl == 0))
2517 blwi = cfs_list_entry(blp->blp_list.next,
2518 struct ldlm_bl_work_item, blwi_entry);
2520 if (!cfs_list_empty(&blp->blp_prio_list))
2521 blwi = cfs_list_entry(blp->blp_prio_list.next,
2522 struct ldlm_bl_work_item,
2526 if (++num_bl >= cfs_atomic_read(&blp->blp_num_threads))
2528 cfs_list_del(&blwi->blwi_entry);
2530 spin_unlock(&blp->blp_lock);
2535 /* This only contains temporary data until the thread starts */
2536 struct ldlm_bl_thread_data {
2537 char bltd_name[CFS_CURPROC_COMM_MAX];
2538 struct ldlm_bl_pool *bltd_blp;
2539 struct completion bltd_comp;
2543 static int ldlm_bl_thread_main(void *arg);
2545 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2547 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2550 init_completion(&bltd.bltd_comp);
2551 rc = cfs_create_thread(ldlm_bl_thread_main, &bltd, 0);
2553 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
2554 cfs_atomic_read(&blp->blp_num_threads), rc);
2557 wait_for_completion(&bltd.bltd_comp);
2563 * Main blocking requests processing thread.
2565 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2566 * This thread in the end ends up doing actual call to ->l_blocking_ast
2569 static int ldlm_bl_thread_main(void *arg)
2571 struct ldlm_bl_pool *blp;
2575 struct ldlm_bl_thread_data *bltd = arg;
2577 blp = bltd->bltd_blp;
2580 cfs_atomic_inc_return(&blp->blp_num_threads) - 1;
2581 cfs_atomic_inc(&blp->blp_busy_threads);
2583 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
2584 "ldlm_bl_%02d", bltd->bltd_num);
2585 cfs_daemonize(bltd->bltd_name);
2587 complete(&bltd->bltd_comp);
2588 /* cannot use bltd after this, it is only on caller's stack */
2592 struct l_wait_info lwi = { 0 };
2593 struct ldlm_bl_work_item *blwi = NULL;
2596 blwi = ldlm_bl_get_work(blp);
2599 cfs_atomic_dec(&blp->blp_busy_threads);
2600 l_wait_event_exclusive(blp->blp_waitq,
2601 (blwi = ldlm_bl_get_work(blp)) != NULL,
2603 busy = cfs_atomic_inc_return(&blp->blp_busy_threads);
2605 busy = cfs_atomic_read(&blp->blp_busy_threads);
2608 if (blwi->blwi_ns == NULL)
2609 /* added by ldlm_cleanup() */
2612 /* Not fatal if racy and have a few too many threads */
2613 if (unlikely(busy < blp->blp_max_threads &&
2614 busy >= cfs_atomic_read(&blp->blp_num_threads) &&
2615 !blwi->blwi_mem_pressure))
2616 /* discard the return value, we tried */
2617 ldlm_bl_thread_start(blp);
2619 if (blwi->blwi_mem_pressure)
2620 cfs_memory_pressure_set();
2622 if (blwi->blwi_count) {
2624 /* The special case when we cancel locks in LRU
2625 * asynchronously, we pass the list of locks here.
2626 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2627 * canceled locally yet. */
2628 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2631 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL, 0);
2633 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2636 if (blwi->blwi_mem_pressure)
2637 cfs_memory_pressure_clr();
2639 if (blwi->blwi_mode == LDLM_ASYNC)
2640 OBD_FREE(blwi, sizeof(*blwi));
2642 complete(&blwi->blwi_comp);
2645 cfs_atomic_dec(&blp->blp_busy_threads);
2646 cfs_atomic_dec(&blp->blp_num_threads);
2647 complete(&blp->blp_comp);
2653 static int ldlm_setup(void);
2654 static int ldlm_cleanup(void);
2656 int ldlm_get_ref(void)
2660 mutex_lock(&ldlm_ref_mutex);
2661 if (++ldlm_refcount == 1) {
2666 mutex_unlock(&ldlm_ref_mutex);
2670 EXPORT_SYMBOL(ldlm_get_ref);
2672 void ldlm_put_ref(void)
2675 mutex_lock(&ldlm_ref_mutex);
2676 if (ldlm_refcount == 1) {
2677 int rc = ldlm_cleanup();
2679 CERROR("ldlm_cleanup failed: %d\n", rc);
2685 mutex_unlock(&ldlm_ref_mutex);
2689 EXPORT_SYMBOL(ldlm_put_ref);
2692 * Export handle<->lock hash operations.
2695 ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
2697 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2701 ldlm_export_lock_key(cfs_hlist_node_t *hnode)
2703 struct ldlm_lock *lock;
2705 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2706 return &lock->l_remote_handle;
2710 ldlm_export_lock_keycpy(cfs_hlist_node_t *hnode, void *key)
2712 struct ldlm_lock *lock;
2714 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2715 lock->l_remote_handle = *(struct lustre_handle *)key;
2719 ldlm_export_lock_keycmp(const void *key, cfs_hlist_node_t *hnode)
2721 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2725 ldlm_export_lock_object(cfs_hlist_node_t *hnode)
2727 return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2731 ldlm_export_lock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2733 struct ldlm_lock *lock;
2735 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2736 LDLM_LOCK_GET(lock);
2740 ldlm_export_lock_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2742 struct ldlm_lock *lock;
2744 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2745 LDLM_LOCK_RELEASE(lock);
2748 static cfs_hash_ops_t ldlm_export_lock_ops = {
2749 .hs_hash = ldlm_export_lock_hash,
2750 .hs_key = ldlm_export_lock_key,
2751 .hs_keycmp = ldlm_export_lock_keycmp,
2752 .hs_keycpy = ldlm_export_lock_keycpy,
2753 .hs_object = ldlm_export_lock_object,
2754 .hs_get = ldlm_export_lock_get,
2755 .hs_put = ldlm_export_lock_put,
2756 .hs_put_locked = ldlm_export_lock_put,
2759 int ldlm_init_export(struct obd_export *exp)
2763 exp->exp_lock_hash =
2764 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2765 HASH_EXP_LOCK_CUR_BITS,
2766 HASH_EXP_LOCK_MAX_BITS,
2767 HASH_EXP_LOCK_BKT_BITS, 0,
2768 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2769 &ldlm_export_lock_ops,
2770 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2771 CFS_HASH_NBLK_CHANGE);
2773 if (!exp->exp_lock_hash)
2778 EXPORT_SYMBOL(ldlm_init_export);
2780 void ldlm_destroy_export(struct obd_export *exp)
2783 cfs_hash_putref(exp->exp_lock_hash);
2784 exp->exp_lock_hash = NULL;
2786 ldlm_destroy_flock_export(exp);
2789 EXPORT_SYMBOL(ldlm_destroy_export);
2791 static int ldlm_setup(void)
2793 static struct ptlrpc_service_conf conf;
2794 struct ldlm_bl_pool *blp = NULL;
2801 if (ldlm_state != NULL)
2804 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2805 if (ldlm_state == NULL)
2809 rc = ldlm_proc_setup();
2814 memset(&conf, 0, sizeof(conf));
2815 conf = (typeof(conf)) {
2816 .psc_name = "ldlm_cbd",
2817 .psc_watchdog_factor = 2,
2819 .bc_nbufs = LDLM_NBUFS,
2820 .bc_buf_size = LDLM_BUFSIZE,
2821 .bc_req_max_size = LDLM_MAXREQSIZE,
2822 .bc_rep_max_size = LDLM_MAXREPSIZE,
2823 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2824 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2827 .tc_thr_name = "ldlm_cb",
2828 .tc_thr_factor = LDLM_THR_FACTOR,
2829 .tc_nthrs_init = LDLM_NTHRS_INIT,
2830 .tc_nthrs_base = LDLM_NTHRS_BASE,
2831 .tc_nthrs_max = LDLM_NTHRS_MAX,
2832 .tc_nthrs_user = ldlm_num_threads,
2833 .tc_cpu_affinity = 1,
2834 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2837 .cc_pattern = ldlm_cpts,
2840 .so_req_handler = ldlm_callback_handler,
2843 ldlm_state->ldlm_cb_service = \
2844 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2845 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2846 CERROR("failed to start service\n");
2847 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2848 ldlm_state->ldlm_cb_service = NULL;
2852 #ifdef HAVE_SERVER_SUPPORT
2853 memset(&conf, 0, sizeof(conf));
2854 conf = (typeof(conf)) {
2855 .psc_name = "ldlm_canceld",
2856 .psc_watchdog_factor = 6,
2858 .bc_nbufs = LDLM_NBUFS,
2859 .bc_buf_size = LDLM_BUFSIZE,
2860 .bc_req_max_size = LDLM_MAXREQSIZE,
2861 .bc_rep_max_size = LDLM_MAXREPSIZE,
2862 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
2863 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
2867 .tc_thr_name = "ldlm_cn",
2868 .tc_thr_factor = LDLM_THR_FACTOR,
2869 .tc_nthrs_init = LDLM_NTHRS_INIT,
2870 .tc_nthrs_base = LDLM_NTHRS_BASE,
2871 .tc_nthrs_max = LDLM_NTHRS_MAX,
2872 .tc_nthrs_user = ldlm_num_threads,
2873 .tc_cpu_affinity = 1,
2874 .tc_ctx_tags = LCT_MD_THREAD | \
2879 .cc_pattern = ldlm_cpts,
2882 .so_req_handler = ldlm_cancel_handler,
2883 .so_hpreq_handler = ldlm_hpreq_handler,
2886 ldlm_state->ldlm_cancel_service = \
2887 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2888 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2889 CERROR("failed to start service\n");
2890 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2891 ldlm_state->ldlm_cancel_service = NULL;
2896 OBD_ALLOC(blp, sizeof(*blp));
2898 GOTO(out, rc = -ENOMEM);
2899 ldlm_state->ldlm_bl_pool = blp;
2901 spin_lock_init(&blp->blp_lock);
2902 CFS_INIT_LIST_HEAD(&blp->blp_list);
2903 CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2904 cfs_waitq_init(&blp->blp_waitq);
2905 cfs_atomic_set(&blp->blp_num_threads, 0);
2906 cfs_atomic_set(&blp->blp_busy_threads, 0);
2909 if (ldlm_num_threads == 0) {
2910 blp->blp_min_threads = LDLM_NTHRS_INIT;
2911 blp->blp_max_threads = LDLM_NTHRS_MAX;
2913 blp->blp_min_threads = blp->blp_max_threads = \
2914 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
2918 for (i = 0; i < blp->blp_min_threads; i++) {
2919 rc = ldlm_bl_thread_start(blp);
2924 # ifdef HAVE_SERVER_SUPPORT
2925 CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2926 expired_lock_thread.elt_state = ELT_STOPPED;
2927 cfs_waitq_init(&expired_lock_thread.elt_waitq);
2929 CFS_INIT_LIST_HEAD(&waiting_locks_list);
2930 spin_lock_init(&waiting_locks_spinlock);
2931 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2933 rc = cfs_create_thread(expired_lock_main, NULL, CFS_DAEMON_FLAGS);
2935 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2939 cfs_wait_event(expired_lock_thread.elt_waitq,
2940 expired_lock_thread.elt_state == ELT_READY);
2941 # endif /* HAVE_SERVER_SUPPORT */
2943 rc = ldlm_pools_init();
2945 CERROR("Failed to initialize LDLM pools: %d\n", rc);
2956 static int ldlm_cleanup(void)
2960 if (!cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2961 !cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2962 CERROR("ldlm still has namespaces; clean these up first.\n");
2963 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2964 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2971 if (ldlm_state->ldlm_bl_pool != NULL) {
2972 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2974 while (cfs_atomic_read(&blp->blp_num_threads) > 0) {
2975 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2977 init_completion(&blp->blp_comp);
2979 spin_lock(&blp->blp_lock);
2980 cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2981 cfs_waitq_signal(&blp->blp_waitq);
2982 spin_unlock(&blp->blp_lock);
2984 wait_for_completion(&blp->blp_comp);
2987 OBD_FREE(blp, sizeof(*blp));
2989 #endif /* __KERNEL__ */
2991 if (ldlm_state->ldlm_cb_service != NULL)
2992 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2993 # ifdef HAVE_SERVER_SUPPORT
2994 if (ldlm_state->ldlm_cancel_service != NULL)
2995 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2999 ldlm_proc_cleanup();
3001 # ifdef HAVE_SERVER_SUPPORT
3002 if (expired_lock_thread.elt_state != ELT_STOPPED) {
3003 expired_lock_thread.elt_state = ELT_TERMINATE;
3004 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
3005 cfs_wait_event(expired_lock_thread.elt_waitq,
3006 expired_lock_thread.elt_state == ELT_STOPPED);
3009 #endif /* __KERNEL__ */
3011 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3019 mutex_init(&ldlm_ref_mutex);
3020 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
3021 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
3022 ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
3023 sizeof(struct ldlm_resource), 0,
3024 CFS_SLAB_HWCACHE_ALIGN);
3025 if (ldlm_resource_slab == NULL)
3028 ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
3029 sizeof(struct ldlm_lock), 0,
3030 CFS_SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU);
3031 if (ldlm_lock_slab == NULL) {
3032 cfs_mem_cache_destroy(ldlm_resource_slab);
3036 ldlm_interval_slab = cfs_mem_cache_create("interval_node",
3037 sizeof(struct ldlm_interval),
3038 0, CFS_SLAB_HWCACHE_ALIGN);
3039 if (ldlm_interval_slab == NULL) {
3040 cfs_mem_cache_destroy(ldlm_resource_slab);
3041 cfs_mem_cache_destroy(ldlm_lock_slab);
3044 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3045 class_export_dump_hook = ldlm_dump_export_locks;
3050 void ldlm_exit(void)
3054 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3055 rc = cfs_mem_cache_destroy(ldlm_resource_slab);
3056 LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
3058 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3059 * synchronize_rcu() to wait a grace period elapsed, so that
3060 * ldlm_lock_free() get a chance to be called. */
3063 rc = cfs_mem_cache_destroy(ldlm_lock_slab);
3064 LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
3065 rc = cfs_mem_cache_destroy(ldlm_interval_slab);
3066 LASSERTF(rc == 0, "couldn't free interval node slab\n");