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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lockd.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <linux/kthread.h>
41 #include <linux/list.h>
42 #include <libcfs/libcfs.h>
43 #include <lustre_errno.h>
44 #include <lustre_dlm.h>
45 #include <obd_class.h>
46 #include "ldlm_internal.h"
48 static int ldlm_num_threads;
49 module_param(ldlm_num_threads, int, 0444);
50 MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start");
52 static char *ldlm_cpts;
53 module_param(ldlm_cpts, charp, 0444);
54 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
56 static DEFINE_MUTEX(ldlm_ref_mutex);
57 static int ldlm_refcount;
59 struct kobject *ldlm_kobj;
60 struct kset *ldlm_ns_kset;
61 struct kset *ldlm_svc_kset;
65 static struct ldlm_state *ldlm_state;
67 /* timeout for initial callback (AST) reply (bz10399)
68 * Due to having to send a 32 bit time value over the
69 * wire return it as time_t instead of time64_t
71 static inline time_t ldlm_get_rq_timeout(void)
74 time_t timeout = min(ldlm_timeout, obd_timeout / 3);
76 return timeout < 1 ? 1 : timeout;
83 * blp_prio_list is used for callbacks that should be handled
84 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
87 struct list_head blp_prio_list;
90 * blp_list is used for all other callbacks which are likely
91 * to take longer to process.
93 struct list_head blp_list;
95 wait_queue_head_t blp_waitq;
96 struct completion blp_comp;
97 atomic_t blp_num_threads;
98 atomic_t blp_busy_threads;
103 struct ldlm_bl_work_item {
104 struct list_head blwi_entry;
105 struct ldlm_namespace *blwi_ns;
106 struct ldlm_lock_desc blwi_ld;
107 struct ldlm_lock *blwi_lock;
108 struct list_head blwi_head;
110 struct completion blwi_comp;
111 enum ldlm_cancel_flags blwi_flags;
112 int blwi_mem_pressure;
115 #ifdef HAVE_SERVER_SUPPORT
118 * Protects both waiting_locks_list and expired_lock_thread.
120 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
123 * List for contended locks.
125 * As soon as a lock is contended, it gets placed on this list and
126 * expected time to get a response is filled in the lock. A special
127 * thread walks the list looking for locks that should be released and
128 * schedules client evictions for those that have not been released in
131 * All access to it should be under waiting_locks_spinlock.
133 static LIST_HEAD(waiting_locks_list);
134 static void waiting_locks_callback(unsigned long unused);
135 static DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
143 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
144 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
145 static int expired_lock_dump;
146 static LIST_HEAD(expired_lock_list);
148 static int ldlm_lock_busy(struct ldlm_lock *lock);
149 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout);
150 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds);
152 static inline int have_expired_locks(void)
157 spin_lock_bh(&waiting_locks_spinlock);
158 need_to_run = !list_empty(&expired_lock_list);
159 spin_unlock_bh(&waiting_locks_spinlock);
165 * Check expired lock list for expired locks and time them out.
167 static int expired_lock_main(void *arg)
169 struct list_head *expired = &expired_lock_list;
170 struct l_wait_info lwi = { 0 };
175 expired_lock_thread_state = ELT_READY;
176 wake_up(&expired_lock_wait_queue);
179 l_wait_event(expired_lock_wait_queue,
180 have_expired_locks() ||
181 expired_lock_thread_state == ELT_TERMINATE,
184 spin_lock_bh(&waiting_locks_spinlock);
185 if (expired_lock_dump) {
186 spin_unlock_bh(&waiting_locks_spinlock);
188 /* from waiting_locks_callback, but not in timer */
189 libcfs_debug_dumplog();
191 spin_lock_bh(&waiting_locks_spinlock);
192 expired_lock_dump = 0;
197 while (!list_empty(expired)) {
198 struct obd_export *export;
199 struct ldlm_lock *lock;
201 lock = list_entry(expired->next, struct ldlm_lock,
203 if ((void *)lock < LP_POISON + PAGE_SIZE &&
204 (void *)lock >= LP_POISON) {
205 spin_unlock_bh(&waiting_locks_spinlock);
206 CERROR("free lock on elt list %p\n", lock);
209 list_del_init(&lock->l_pending_chain);
210 if ((void *)lock->l_export <
211 LP_POISON + PAGE_SIZE &&
212 (void *)lock->l_export >= LP_POISON) {
213 CERROR("lock with free export on elt list %p\n",
215 lock->l_export = NULL;
216 LDLM_ERROR(lock, "free export");
217 /* release extra ref grabbed by
218 * ldlm_add_waiting_lock() or
219 * ldlm_failed_ast() */
220 LDLM_LOCK_RELEASE(lock);
224 if (ldlm_is_destroyed(lock)) {
225 /* release the lock refcount where
226 * waiting_locks_callback() founds */
227 LDLM_LOCK_RELEASE(lock);
230 export = class_export_lock_get(lock->l_export, lock);
231 spin_unlock_bh(&waiting_locks_spinlock);
233 /* Check if we need to prolong timeout */
234 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
235 lock->l_callback_timeout != 0 && /* not AST error */
236 ldlm_lock_busy(lock)) {
237 LDLM_DEBUG(lock, "prolong the busy lock");
238 lock_res_and_lock(lock);
239 ldlm_add_waiting_lock(lock,
240 ldlm_bl_timeout(lock) >> 1);
241 unlock_res_and_lock(lock);
243 spin_lock_bh(&export->exp_bl_list_lock);
244 list_del_init(&lock->l_exp_list);
245 spin_unlock_bh(&export->exp_bl_list_lock);
248 "lock callback timer expired after "
249 "%llds: evicting client at %s ",
250 cfs_time_current_sec() -
251 lock->l_last_activity,
252 obd_export_nid2str(export));
253 ldlm_lock_to_ns(lock)->ns_timeouts++;
255 class_fail_export(export);
257 class_export_lock_put(export, lock);
258 /* release extra ref grabbed by ldlm_add_waiting_lock()
259 * or ldlm_failed_ast() */
260 LDLM_LOCK_RELEASE(lock);
262 spin_lock_bh(&waiting_locks_spinlock);
264 spin_unlock_bh(&waiting_locks_spinlock);
266 if (do_dump && obd_dump_on_eviction) {
267 CERROR("dump the log upon eviction\n");
268 libcfs_debug_dumplog();
271 if (expired_lock_thread_state == ELT_TERMINATE)
275 expired_lock_thread_state = ELT_STOPPED;
276 wake_up(&expired_lock_wait_queue);
281 * Check if there is a request in the export request list
282 * which prevents the lock canceling.
284 static int ldlm_lock_busy(struct ldlm_lock *lock)
286 struct ptlrpc_request *req;
290 if (lock->l_export == NULL)
293 spin_lock(&lock->l_export->exp_rpc_lock);
294 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
296 if (req->rq_ops->hpreq_lock_match) {
297 match = req->rq_ops->hpreq_lock_match(req, lock);
302 spin_unlock(&lock->l_export->exp_rpc_lock);
306 /* This is called from within a timer interrupt and cannot schedule */
307 static void waiting_locks_callback(unsigned long unused)
309 struct ldlm_lock *lock;
312 spin_lock_bh(&waiting_locks_spinlock);
313 while (!list_empty(&waiting_locks_list)) {
314 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
316 if (lock->l_callback_timeout > ktime_get_seconds() ||
317 lock->l_req_mode == LCK_GROUP)
320 /* no needs to take an extra ref on the lock since it was in
321 * the waiting_locks_list and ldlm_add_waiting_lock()
322 * already grabbed a ref */
323 list_del(&lock->l_pending_chain);
324 list_add(&lock->l_pending_chain, &expired_lock_list);
328 if (!list_empty(&expired_lock_list)) {
329 if (obd_dump_on_timeout && need_dump)
330 expired_lock_dump = __LINE__;
332 wake_up(&expired_lock_wait_queue);
336 * Make sure the timer will fire again if we have any locks
339 if (!list_empty(&waiting_locks_list)) {
340 unsigned long timeout_jiffies;
342 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
344 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
345 mod_timer(&waiting_locks_timer, timeout_jiffies);
347 spin_unlock_bh(&waiting_locks_spinlock);
351 * Add lock to the list of contended locks.
353 * Indicate that we're waiting for a client to call us back cancelling a given
354 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
355 * timer to fire appropriately. (We round up to the next second, to avoid
356 * floods of timer firings during periods of high lock contention and traffic).
357 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
358 * if it has been added to the waiting list (1 is returned).
360 * Called with the namespace lock held.
362 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds)
364 unsigned long timeout_jiffies;
367 if (!list_empty(&lock->l_pending_chain))
370 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
371 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
374 timeout = ktime_get_seconds() + seconds;
375 if (likely(timeout > lock->l_callback_timeout))
376 lock->l_callback_timeout = timeout;
378 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
380 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
381 !timer_pending(&waiting_locks_timer))
382 mod_timer(&waiting_locks_timer, timeout_jiffies);
384 /* if the new lock has a shorter timeout than something earlier on
385 * the list, we'll wait the longer amount of time; no big deal.
388 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
392 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
394 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
395 if (list_empty(&lock->l_exp_list)) {
396 if (lock->l_granted_mode != lock->l_req_mode)
397 list_add_tail(&lock->l_exp_list,
398 &lock->l_export->exp_bl_list);
400 list_add(&lock->l_exp_list,
401 &lock->l_export->exp_bl_list);
403 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
405 /* A blocked lock is added. Adjust the position in
406 * the stale list if the export is in the list.
407 * If export is stale and not in the list - it is being
408 * processed and will be placed on the right position
409 * on obd_stale_export_put(). */
410 if (!list_empty(&lock->l_export->exp_stale_list))
411 obd_stale_export_adjust(lock->l_export);
414 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
418 /* NB: must be called with hold of lock_res_and_lock() */
419 LASSERT(ldlm_is_res_locked(lock));
420 LASSERT(!ldlm_is_cancel_on_block(lock));
422 /* Do not put cross-MDT lock in the waiting list, since we
423 * will not evict it due to timeout for now */
424 if (lock->l_export != NULL &&
425 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
428 spin_lock_bh(&waiting_locks_spinlock);
429 if (ldlm_is_cancel(lock)) {
430 spin_unlock_bh(&waiting_locks_spinlock);
434 if (ldlm_is_destroyed(lock)) {
435 static time64_t next;
437 spin_unlock_bh(&waiting_locks_spinlock);
438 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
439 if (ktime_get_seconds() > next) {
440 next = ktime_get_seconds() + 14400;
441 libcfs_debug_dumpstack(NULL);
446 ldlm_set_waited(lock);
447 lock->l_last_activity = ktime_get_real_seconds();
448 ret = __ldlm_add_waiting_lock(lock, timeout);
450 /* grab ref on the lock if it has been added to the
454 spin_unlock_bh(&waiting_locks_spinlock);
457 ldlm_add_blocked_lock(lock);
459 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)",
460 ret == 0 ? "not re-" : "", timeout,
461 AT_OFF ? "off" : "on");
466 * Remove a lock from the pending list, likely because it had its cancellation
467 * callback arrive without incident. This adjusts the lock-timeout timer if
468 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
469 * As done by ldlm_del_waiting_lock(), the caller must release the lock
470 * reference when the lock is removed from any list (1 is returned).
472 * Called with namespace lock held.
474 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
476 struct list_head *list_next;
478 if (list_empty(&lock->l_pending_chain))
481 list_next = lock->l_pending_chain.next;
482 if (lock->l_pending_chain.prev == &waiting_locks_list) {
483 /* Removing the head of the list, adjust timer. */
484 if (list_next == &waiting_locks_list) {
485 /* No more, just cancel. */
486 del_timer(&waiting_locks_timer);
488 struct ldlm_lock *next;
490 next = list_entry(list_next, struct ldlm_lock,
492 mod_timer(&waiting_locks_timer,
493 cfs_time_seconds(next->l_callback_timeout));
496 list_del_init(&lock->l_pending_chain);
501 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
505 if (lock->l_export == NULL) {
506 /* We don't have a "waiting locks list" on clients. */
507 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
511 spin_lock_bh(&waiting_locks_spinlock);
512 ret = __ldlm_del_waiting_lock(lock);
513 ldlm_clear_waited(lock);
514 spin_unlock_bh(&waiting_locks_spinlock);
516 /* remove the lock out of export blocking list */
517 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
518 list_del_init(&lock->l_exp_list);
519 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
522 /* release lock ref if it has indeed been removed
524 LDLM_LOCK_RELEASE(lock);
527 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
532 * Prolong the contended lock waiting time.
534 * Called with namespace lock held.
536 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
538 if (lock->l_export == NULL) {
539 /* We don't have a "waiting locks list" on clients. */
540 LDLM_DEBUG(lock, "client lock: no-op");
544 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
545 /* We don't have a "waiting locks list" on OSP. */
546 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
550 spin_lock_bh(&waiting_locks_spinlock);
552 if (list_empty(&lock->l_pending_chain)) {
553 spin_unlock_bh(&waiting_locks_spinlock);
554 LDLM_DEBUG(lock, "wasn't waiting");
558 /* we remove/add the lock to the waiting list, so no needs to
559 * release/take a lock reference */
560 __ldlm_del_waiting_lock(lock);
561 __ldlm_add_waiting_lock(lock, timeout);
562 spin_unlock_bh(&waiting_locks_spinlock);
564 LDLM_DEBUG(lock, "refreshed");
567 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
569 #else /* HAVE_SERVER_SUPPORT */
571 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
576 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
581 #endif /* !HAVE_SERVER_SUPPORT */
583 #ifdef HAVE_SERVER_SUPPORT
586 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
587 * lock cancel, and their replies). Used for lock callback timeout and AST
590 * \param[in] lock lock which is getting the blocking callback
592 * \retval timeout in seconds to wait for the client reply
594 time64_t ldlm_bl_timeout(struct ldlm_lock *lock)
599 return obd_timeout / 2;
601 /* Since these are non-updating timeouts, we should be conservative.
602 * Take more than usually, 150%
603 * It would be nice to have some kind of "early reply" mechanism for
604 * lock callbacks too... */
605 timeout = at_get(&lock->l_export->exp_bl_lock_at);
606 return max(timeout + (timeout >> 1), (time64_t)ldlm_enqueue_min);
608 EXPORT_SYMBOL(ldlm_bl_timeout);
611 * Perform lock cleanup if AST sending failed.
613 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
614 const char *ast_type)
616 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
617 "to a lock %s callback time out: rc %d\n",
618 lock->l_export->exp_obd->obd_name,
619 obd_export_nid2str(lock->l_export), ast_type, rc);
621 if (obd_dump_on_timeout)
622 libcfs_debug_dumplog();
623 spin_lock_bh(&waiting_locks_spinlock);
624 if (__ldlm_del_waiting_lock(lock) == 0)
625 /* the lock was not in any list, grab an extra ref before adding
626 * the lock to the expired list */
628 lock->l_callback_timeout = 0; /* differentiate it from expired locks */
629 list_add(&lock->l_pending_chain, &expired_lock_list);
630 wake_up(&expired_lock_wait_queue);
631 spin_unlock_bh(&waiting_locks_spinlock);
635 * Perform lock cleanup if AST reply came with error.
637 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
638 struct ptlrpc_request *req, int rc,
639 const char *ast_type)
641 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
643 if (!req->rq_replied || (rc && rc != -EINVAL)) {
644 if (ldlm_is_cancel(lock)) {
646 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
647 ast_type, req, req->rq_xid,
648 libcfs_nid2str(peer.nid));
649 ldlm_lock_cancel(lock);
651 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
653 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
654 /* Upon umount process the AST fails because cannot be
655 * sent. This shouldn't lead to the client eviction.
656 * -ENODEV error is returned by ptl_send_rpc() for
657 * new request in such import.
658 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
659 * if imp_invalid is set or obd_no_recov.
660 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
661 * in ptlrpc_import_delay_req() as well with -EIO code.
662 * In all such cases errors are ignored.
664 LDLM_DEBUG(lock, "%s AST can't be sent due to a server"
665 " %s failure or umount process: rc = %d\n",
667 req->rq_import->imp_obd->obd_name, rc);
670 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
671 libcfs_nid2str(peer.nid),
672 req->rq_replied ? "returned error from" :
673 "failed to reply to",
674 ast_type, req, req->rq_xid,
675 (req->rq_repmsg != NULL) ?
676 lustre_msg_get_status(req->rq_repmsg) : 0,
678 ldlm_failed_ast(lock, rc, ast_type);
684 struct ldlm_resource *res = lock->l_resource;
687 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
688 libcfs_nid2str(peer.nid),
690 lustre_msg_get_status(req->rq_repmsg) : -1,
691 ast_type, req, req->rq_xid);
693 /* update lvbo to return proper attributes.
695 ldlm_resource_getref(res);
696 ldlm_lvbo_update(res, lock, NULL, 1);
697 ldlm_resource_putref(res);
699 ldlm_lock_cancel(lock);
706 static int ldlm_cb_interpret(const struct lu_env *env,
707 struct ptlrpc_request *req, void *data, int rc)
709 struct ldlm_cb_async_args *ca = data;
710 struct ldlm_lock *lock = ca->ca_lock;
711 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
714 LASSERT(lock != NULL);
717 case LDLM_GL_CALLBACK:
718 /* Update the LVB from disk if the AST failed
719 * (this is a legal race)
721 * - Glimpse callback of local lock just returns
722 * -ELDLM_NO_LOCK_DATA.
723 * - Glimpse callback of remote lock might return
724 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
726 if (unlikely(arg->gl_interpret_reply)) {
727 rc = arg->gl_interpret_reply(env, req, data, rc);
728 } else if (rc == -ELDLM_NO_LOCK_DATA) {
729 LDLM_DEBUG(lock, "lost race - client has a lock but no "
731 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
732 } else if (rc != 0) {
733 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
735 rc = ldlm_lvbo_update(lock->l_resource, lock, req, 1);
738 case LDLM_BL_CALLBACK:
740 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
742 case LDLM_CP_CALLBACK:
744 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
747 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
752 /* release extra reference taken in ldlm_ast_fini() */
753 LDLM_LOCK_RELEASE(lock);
756 atomic_inc(&arg->restart);
761 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
763 struct ldlm_cb_async_args *ca = data;
764 struct ldlm_lock *lock = ca->ca_lock;
766 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
769 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
770 struct ldlm_cb_set_arg *arg,
771 struct ldlm_lock *lock,
777 if (unlikely(instant_cancel)) {
778 rc = ptl_send_rpc(req, 1);
779 ptlrpc_req_finished(req);
781 atomic_inc(&arg->restart);
784 ptlrpc_set_add_req(arg->set, req);
791 * Check if there are requests in the export request list which prevent
792 * the lock canceling and make these requests high priority ones.
794 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
796 struct ptlrpc_request *req;
799 if (lock->l_export == NULL) {
800 LDLM_DEBUG(lock, "client lock: no-op");
804 spin_lock(&lock->l_export->exp_rpc_lock);
805 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
807 /* Do not process requests that were not yet added to there
808 * incoming queue or were already removed from there for
809 * processing. We evaluate ptlrpc_nrs_req_can_move() without
810 * holding svcpt->scp_req_lock, and then redo the check with
811 * the lock held once we need to obtain a reliable result.
813 if (ptlrpc_nrs_req_can_move(req) &&
814 req->rq_ops->hpreq_lock_match &&
815 req->rq_ops->hpreq_lock_match(req, lock))
816 ptlrpc_nrs_req_hp_move(req);
818 spin_unlock(&lock->l_export->exp_rpc_lock);
823 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
824 * enqueued server lock conflicts with given one.
826 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
827 * to wait for client response.
829 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
830 struct ldlm_lock_desc *desc,
831 void *data, int flag)
833 struct ldlm_cb_async_args *ca;
834 struct ldlm_cb_set_arg *arg = data;
835 struct ldlm_request *body;
836 struct ptlrpc_request *req;
837 int instant_cancel = 0;
841 if (flag == LDLM_CB_CANCELING)
842 /* Don't need to do anything here. */
845 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
846 LDLM_DEBUG(lock, "dropping BL AST");
851 LASSERT(data != NULL);
852 if (lock->l_export->exp_obd->obd_recovering != 0)
853 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
855 ldlm_lock_reorder_req(lock);
857 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
858 &RQF_LDLM_BL_CALLBACK,
859 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
863 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
864 ca = ptlrpc_req_async_args(req);
865 ca->ca_set_arg = arg;
868 req->rq_interpret_reply = ldlm_cb_interpret;
870 lock_res_and_lock(lock);
871 if (ldlm_is_destroyed(lock)) {
872 /* What's the point? */
873 unlock_res_and_lock(lock);
874 ptlrpc_req_finished(req);
878 if (lock->l_granted_mode != lock->l_req_mode) {
879 /* this blocking AST will be communicated as part of the
880 * completion AST instead */
881 ldlm_add_blocked_lock(lock);
882 ldlm_set_waited(lock);
883 unlock_res_and_lock(lock);
885 ptlrpc_req_finished(req);
886 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
890 if (ldlm_is_cancel_on_block(lock))
893 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
894 body->lock_handle[0] = lock->l_remote_handle;
895 body->lock_desc = *desc;
896 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
898 LDLM_DEBUG(lock, "server preparing blocking AST");
900 ptlrpc_request_set_replen(req);
901 ldlm_set_cbpending(lock);
902 if (instant_cancel) {
903 unlock_res_and_lock(lock);
904 ldlm_lock_cancel(lock);
906 req->rq_no_resend = 1;
908 LASSERT(lock->l_granted_mode == lock->l_req_mode);
909 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
910 unlock_res_and_lock(lock);
912 /* Do not resend after lock callback timeout */
913 req->rq_delay_limit = ldlm_bl_timeout(lock);
914 req->rq_resend_cb = ldlm_update_resend;
917 req->rq_send_state = LUSTRE_IMP_FULL;
918 /* ptlrpc_request_alloc_pack already set timeout */
920 req->rq_timeout = ldlm_get_rq_timeout();
922 lock->l_last_activity = ktime_get_real_seconds();
924 if (lock->l_export && lock->l_export->exp_nid_stats &&
925 lock->l_export->exp_nid_stats->nid_ldlm_stats)
926 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
927 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
929 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
935 * ->l_completion_ast callback for a remote lock in server namespace.
937 * Sends AST to the client notifying it of lock granting. If initial
938 * lock response was not sent yet, instead of sending another RPC, just
939 * mark the lock as granted and client will understand
941 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
943 struct ldlm_cb_set_arg *arg = data;
944 struct ldlm_request *body;
945 struct ptlrpc_request *req;
946 struct ldlm_cb_async_args *ca;
947 int instant_cancel = 0;
952 LASSERT(lock != NULL);
953 LASSERT(data != NULL);
955 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
956 LDLM_DEBUG(lock, "dropping CP AST");
960 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
961 &RQF_LDLM_CP_CALLBACK);
965 /* server namespace, doesn't need lock */
966 lvb_len = ldlm_lvbo_size(lock);
967 /* LU-3124 & LU-2187: to not return layout in completion AST because
968 * it may deadlock for LU-2187, or client may not have enough space
969 * for large layout. The layout will be returned to client with an
970 * extra RPC to fetch xattr.lov */
971 if (ldlm_has_layout(lock))
974 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
975 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
977 ptlrpc_request_free(req);
981 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
982 ca = ptlrpc_req_async_args(req);
983 ca->ca_set_arg = arg;
986 req->rq_interpret_reply = ldlm_cb_interpret;
987 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
989 body->lock_handle[0] = lock->l_remote_handle;
990 body->lock_flags = ldlm_flags_to_wire(flags);
991 ldlm_lock2desc(lock, &body->lock_desc);
993 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
995 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
997 /* We still need to send the RPC to wake up the blocked
998 * enqueue thread on the client.
1000 * Consider old client, there is no better way to notify
1001 * the failure, just zero-sized the LVB, then the client
1002 * will fail out as "-EPROTO". */
1003 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1007 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1012 lock->l_last_activity = ktime_get_real_seconds();
1014 LDLM_DEBUG(lock, "server preparing completion AST");
1016 ptlrpc_request_set_replen(req);
1018 req->rq_send_state = LUSTRE_IMP_FULL;
1019 /* ptlrpc_request_pack already set timeout */
1021 req->rq_timeout = ldlm_get_rq_timeout();
1023 /* We only send real blocking ASTs after the lock is granted */
1024 lock_res_and_lock(lock);
1025 if (ldlm_is_ast_sent(lock)) {
1026 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1027 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1028 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1031 /* We might get here prior to ldlm_handle_enqueue setting
1032 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1033 * into waiting list, but this is safe and similar code in
1034 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1035 * that would not only cancel the lock, but will also remove
1036 * it from waiting list */
1037 if (ldlm_is_cancel_on_block(lock)) {
1038 unlock_res_and_lock(lock);
1039 ldlm_lock_cancel(lock);
1042 req->rq_no_resend = 1;
1044 lock_res_and_lock(lock);
1046 /* start the lock-timeout clock */
1047 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1048 /* Do not resend after lock callback timeout */
1049 req->rq_delay_limit = ldlm_bl_timeout(lock);
1050 req->rq_resend_cb = ldlm_update_resend;
1053 unlock_res_and_lock(lock);
1055 if (lock->l_export && lock->l_export->exp_nid_stats &&
1056 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1057 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1058 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1060 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1062 RETURN(lvb_len < 0 ? lvb_len : rc);
1066 * Server side ->l_glimpse_ast handler for client locks.
1068 * Sends glimpse AST to the client and waits for reply. Then updates
1069 * lvbo with the result.
1071 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1073 struct ldlm_cb_set_arg *arg = data;
1074 struct ldlm_request *body;
1075 struct ptlrpc_request *req;
1076 struct ldlm_cb_async_args *ca;
1078 struct req_format *req_fmt;
1081 LASSERT(lock != NULL);
1083 if (arg->gl_desc != NULL)
1084 /* There is a glimpse descriptor to pack */
1085 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1087 req_fmt = &RQF_LDLM_GL_CALLBACK;
1089 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1090 req_fmt, LUSTRE_DLM_VERSION,
1096 if (arg->gl_desc != NULL) {
1097 /* copy the GL descriptor */
1098 union ldlm_gl_desc *desc;
1099 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1100 *desc = *arg->gl_desc;
1103 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1104 body->lock_handle[0] = lock->l_remote_handle;
1105 ldlm_lock2desc(lock, &body->lock_desc);
1107 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1108 ca = ptlrpc_req_async_args(req);
1109 ca->ca_set_arg = arg;
1112 /* server namespace, doesn't need lock */
1113 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1114 ldlm_lvbo_size(lock));
1115 ptlrpc_request_set_replen(req);
1117 req->rq_send_state = LUSTRE_IMP_FULL;
1118 /* ptlrpc_request_alloc_pack already set timeout */
1120 req->rq_timeout = ldlm_get_rq_timeout();
1122 lock->l_last_activity = ktime_get_real_seconds();
1124 req->rq_interpret_reply = ldlm_cb_interpret;
1126 if (lock->l_export && lock->l_export->exp_nid_stats &&
1127 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1128 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1129 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1131 rc = ldlm_ast_fini(req, arg, lock, 0);
1135 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1137 int ldlm_glimpse_locks(struct ldlm_resource *res,
1138 struct list_head *gl_work_list)
1143 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1145 if (rc == -ERESTART)
1146 ldlm_reprocess_all(res);
1150 EXPORT_SYMBOL(ldlm_glimpse_locks);
1152 /* return LDLM lock associated with a lock callback request */
1153 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1155 struct ldlm_cb_async_args *ca;
1156 struct ldlm_lock *lock;
1159 ca = ptlrpc_req_async_args(req);
1162 RETURN(ERR_PTR(-EFAULT));
1166 EXPORT_SYMBOL(ldlm_request_lock);
1169 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1170 * service threads to carry out client lock enqueueing requests.
1172 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1173 struct ptlrpc_request *req,
1174 const struct ldlm_request *dlm_req,
1175 const struct ldlm_callback_suite *cbs)
1177 struct ldlm_reply *dlm_rep;
1179 enum ldlm_error err = ELDLM_OK;
1180 struct ldlm_lock *lock = NULL;
1181 void *cookie = NULL;
1183 struct ldlm_resource *res = NULL;
1186 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1188 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1189 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1191 LASSERT(req->rq_export);
1193 /* for intent enqueue the stat will be updated inside intent policy */
1194 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1195 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1196 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1198 if (req->rq_export && req->rq_export->exp_nid_stats &&
1199 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1200 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1201 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1203 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1204 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1205 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1206 dlm_req->lock_desc.l_resource.lr_type);
1207 GOTO(out, rc = -EFAULT);
1210 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1211 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1212 dlm_req->lock_desc.l_req_mode &
1213 (dlm_req->lock_desc.l_req_mode-1))) {
1214 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1215 dlm_req->lock_desc.l_req_mode);
1216 GOTO(out, rc = -EFAULT);
1219 if (unlikely((flags & LDLM_FL_REPLAY) ||
1220 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1221 /* Find an existing lock in the per-export lock hash */
1222 /* In the function below, .hs_keycmp resolves to
1223 * ldlm_export_lock_keycmp() */
1224 /* coverity[overrun-buffer-val] */
1225 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1226 (void *)&dlm_req->lock_handle[0]);
1228 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie %#llx",
1229 lock->l_handle.h_cookie);
1230 flags |= LDLM_FL_RESENT;
1231 GOTO(existing_lock, rc = 0);
1234 if (ldlm_reclaim_full()) {
1235 DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1236 "reject current enqueue request and let the "
1237 "client retry later.\n");
1238 GOTO(out, rc = -EINPROGRESS);
1242 /* The lock's callback data might be set in the policy function */
1243 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1244 dlm_req->lock_desc.l_resource.lr_type,
1245 dlm_req->lock_desc.l_req_mode,
1246 cbs, NULL, 0, LVB_T_NONE);
1253 lock->l_remote_handle = dlm_req->lock_handle[0];
1254 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1256 /* Initialize resource lvb but not for a lock being replayed since
1257 * Client already got lvb sent in this case.
1258 * This must occur early since some policy methods assume resource
1259 * lvb is available (lr_lvb_data != NULL).
1261 res = lock->l_resource;
1262 if (!(flags & LDLM_FL_REPLAY)) {
1263 /* non-replayed lock, delayed lvb init may need to be done */
1264 rc = ldlm_lvbo_init(res);
1266 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1271 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1272 /* Don't enqueue a lock onto the export if it is been disonnected
1273 * due to eviction (bug 3822) or server umount (bug 24324).
1274 * Cancel it now instead. */
1275 if (req->rq_export->exp_disconnected) {
1276 LDLM_ERROR(lock, "lock on disconnected export %p",
1278 GOTO(out, rc = -ENOTCONN);
1281 lock->l_export = class_export_lock_get(req->rq_export, lock);
1282 if (lock->l_export->exp_lock_hash)
1283 cfs_hash_add(lock->l_export->exp_lock_hash,
1284 &lock->l_remote_handle,
1287 /* Inherit the enqueue flags before the operation, because we do not
1288 * keep the res lock on return and next operations (BL AST) may proceed
1290 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1291 LDLM_FL_INHERIT_MASK);
1293 ldlm_convert_policy_to_local(req->rq_export,
1294 dlm_req->lock_desc.l_resource.lr_type,
1295 &dlm_req->lock_desc.l_policy_data,
1296 &lock->l_policy_data);
1297 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1298 lock->l_req_extent = lock->l_policy_data.l_extent;
1301 if (flags & LDLM_FL_HAS_INTENT) {
1302 /* In this case, the reply buffer is allocated deep in
1303 * local_lock_enqueue by the policy function. */
1306 /* based on the assumption that lvb size never changes during
1307 * resource life time otherwise it need resource->lr_lock's
1309 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1310 RCL_SERVER, ldlm_lvbo_size(lock));
1312 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1313 GOTO(out, rc = -ENOMEM);
1315 rc = req_capsule_server_pack(&req->rq_pill);
1320 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1327 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1329 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1330 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1332 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1333 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1335 /* We never send a blocking AST until the lock is granted, but
1336 * we can tell it right now */
1337 lock_res_and_lock(lock);
1339 /* Now take into account flags to be inherited from original lock
1340 request both in reply to client and in our own lock flags. */
1341 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1342 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1344 /* Don't move a pending lock onto the export if it has already been
1345 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1346 * Cancel it now instead. */
1347 if (unlikely(req->rq_export->exp_disconnected ||
1348 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1349 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1351 } else if (ldlm_is_ast_sent(lock)) {
1352 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1353 if (lock->l_granted_mode == lock->l_req_mode) {
1355 * Only cancel lock if it was granted, because it would
1356 * be destroyed immediately and would never be granted
1357 * in the future, causing timeouts on client. Not
1358 * granted lock will be cancelled immediately after
1359 * sending completion AST.
1361 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1362 unlock_res_and_lock(lock);
1363 ldlm_lock_cancel(lock);
1364 lock_res_and_lock(lock);
1366 ldlm_add_waiting_lock(lock,
1367 ldlm_bl_timeout(lock));
1371 unlock_res_and_lock(lock);
1375 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1376 if (!req->rq_packed_final) {
1377 err = lustre_pack_reply(req, 1, NULL, NULL);
1382 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1383 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1385 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1386 "(err=%d, rc=%d)", err, rc);
1389 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1391 ldlm_lvbo_size(lock) > 0) {
1395 buf = req_capsule_server_get(&req->rq_pill,
1397 LASSERTF(buf != NULL, "req %p, lock %p\n",
1399 buflen = req_capsule_get_size(&req->rq_pill,
1400 &RMF_DLM_LVB, RCL_SERVER);
1401 /* non-replayed lock, delayed lvb init may
1402 * need to be occur now */
1403 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1404 buflen = ldlm_lvbo_fill(lock, buf,
1410 buflen, RCL_SERVER);
1413 } else if (flags & LDLM_FL_REPLAY) {
1414 /* no LVB resend upon replay */
1428 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1429 if (lock->l_export) {
1430 ldlm_lock_cancel(lock);
1432 lock_res_and_lock(lock);
1433 ldlm_resource_unlink_lock(lock);
1434 ldlm_lock_destroy_nolock(lock);
1435 unlock_res_and_lock(lock);
1440 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1441 ldlm_reprocess_all(lock->l_resource);
1443 LDLM_LOCK_RELEASE(lock);
1446 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1453 * Main LDLM entry point for server code to process lock conversion requests.
1455 int ldlm_handle_convert0(struct ptlrpc_request *req,
1456 const struct ldlm_request *dlm_req)
1462 * Cancel all the locks whose handles are packed into ldlm_request
1464 * Called by server code expecting such combined cancel activity
1467 int ldlm_request_cancel(struct ptlrpc_request *req,
1468 const struct ldlm_request *dlm_req,
1469 int first, enum lustre_at_flags flags)
1471 struct ldlm_resource *res, *pres = NULL;
1472 struct ldlm_lock *lock;
1473 int i, count, done = 0;
1476 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1480 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1483 /* There is no lock on the server at the replay time,
1484 * skip lock cancelling to make replay tests to pass. */
1485 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1488 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1489 "starting at %d", count, first);
1491 for (i = first; i < count; i++) {
1492 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1494 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1495 "lock (cookie %llu)",
1496 dlm_req->lock_handle[i].cookie);
1500 res = lock->l_resource;
1503 /* This code is an optimization to only attempt lock
1504 * granting on the resource (that could be CPU-expensive)
1505 * after we are done cancelling lock in that resource. */
1508 ldlm_reprocess_all(pres);
1509 LDLM_RESOURCE_DELREF(pres);
1510 ldlm_resource_putref(pres);
1513 ldlm_resource_getref(res);
1514 LDLM_RESOURCE_ADDREF(res);
1516 if (!ldlm_is_discard_data(lock))
1517 ldlm_lvbo_update(res, lock, NULL, 1);
1522 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock)) {
1523 time64_t delay = ktime_get_real_seconds() -
1524 lock->l_last_activity;
1525 LDLM_DEBUG(lock, "server cancels blocked lock after %llds",
1527 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1529 ldlm_lock_cancel(lock);
1530 LDLM_LOCK_PUT(lock);
1533 ldlm_reprocess_all(pres);
1534 LDLM_RESOURCE_DELREF(pres);
1535 ldlm_resource_putref(pres);
1537 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1540 EXPORT_SYMBOL(ldlm_request_cancel);
1543 * Main LDLM entry point for server code to cancel locks.
1545 * Typically gets called from service handler on LDLM_CANCEL opc.
1547 int ldlm_handle_cancel(struct ptlrpc_request *req)
1549 struct ldlm_request *dlm_req;
1553 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1554 if (dlm_req == NULL) {
1555 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1559 if (req->rq_export && req->rq_export->exp_nid_stats &&
1560 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1561 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1562 LDLM_CANCEL - LDLM_FIRST_OPC);
1564 rc = req_capsule_server_pack(&req->rq_pill);
1568 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1569 req->rq_status = LUSTRE_ESTALE;
1571 RETURN(ptlrpc_reply(req));
1573 #endif /* HAVE_SERVER_SUPPORT */
1576 * Callback handler for receiving incoming blocking ASTs.
1578 * This can only happen on client side.
1580 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1581 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1586 LDLM_DEBUG(lock, "client blocking AST callback handler");
1588 lock_res_and_lock(lock);
1589 ldlm_set_cbpending(lock);
1591 if (ldlm_is_cancel_on_block(lock))
1592 ldlm_set_cancel(lock);
1594 do_ast = (!lock->l_readers && !lock->l_writers);
1595 unlock_res_and_lock(lock);
1598 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1599 lock, lock->l_blocking_ast);
1600 if (lock->l_blocking_ast != NULL)
1601 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1604 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1608 LDLM_DEBUG(lock, "client blocking callback handler END");
1609 LDLM_LOCK_RELEASE(lock);
1614 * Callback handler for receiving incoming completion ASTs.
1616 * This only can happen on client side.
1618 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1619 struct ldlm_namespace *ns,
1620 struct ldlm_request *dlm_req,
1621 struct ldlm_lock *lock)
1623 struct list_head ast_list;
1628 LDLM_DEBUG(lock, "client completion callback handler START");
1630 INIT_LIST_HEAD(&ast_list);
1631 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1632 long to = cfs_time_seconds(1);
1635 set_current_state(TASK_INTERRUPTIBLE);
1636 schedule_timeout(to);
1637 if (lock->l_granted_mode == lock->l_req_mode ||
1638 ldlm_is_destroyed(lock))
1643 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1645 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1646 GOTO(out, rc = lvb_len);
1647 } else if (lvb_len > 0) {
1648 if (lock->l_lvb_len > 0) {
1649 /* for extent lock, lvb contains ost_lvb{}. */
1650 LASSERT(lock->l_lvb_data != NULL);
1652 if (unlikely(lock->l_lvb_len < lvb_len)) {
1653 LDLM_ERROR(lock, "Replied LVB is larger than "
1654 "expectation, expected = %d, "
1656 lock->l_lvb_len, lvb_len);
1657 GOTO(out, rc = -EINVAL);
1662 lock_res_and_lock(lock);
1663 if (ldlm_is_destroyed(lock) ||
1664 lock->l_granted_mode == lock->l_req_mode) {
1665 /* bug 11300: the lock has already been granted */
1666 unlock_res_and_lock(lock);
1667 LDLM_DEBUG(lock, "Double grant race happened");
1671 /* If we receive the completion AST before the actual enqueue returned,
1672 * then we might need to switch lock modes, resources, or extents. */
1673 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1674 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1675 LDLM_DEBUG(lock, "completion AST, new lock mode");
1678 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1679 ldlm_convert_policy_to_local(req->rq_export,
1680 dlm_req->lock_desc.l_resource.lr_type,
1681 &dlm_req->lock_desc.l_policy_data,
1682 &lock->l_policy_data);
1683 LDLM_DEBUG(lock, "completion AST, new policy data");
1686 ldlm_resource_unlink_lock(lock);
1687 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1688 &lock->l_resource->lr_name,
1689 sizeof(lock->l_resource->lr_name)) != 0) {
1690 unlock_res_and_lock(lock);
1691 rc = ldlm_lock_change_resource(ns, lock,
1692 &dlm_req->lock_desc.l_resource.lr_name);
1694 LDLM_ERROR(lock, "Failed to allocate resource");
1697 LDLM_DEBUG(lock, "completion AST, new resource");
1698 CERROR("change resource!\n");
1699 lock_res_and_lock(lock);
1702 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1703 /* BL_AST locks are not needed in LRU.
1704 * Let ldlm_cancel_lru() be fast. */
1705 ldlm_lock_remove_from_lru(lock);
1706 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1707 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1710 if (lock->l_lvb_len > 0) {
1711 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1712 lock->l_lvb_data, lvb_len);
1714 unlock_res_and_lock(lock);
1719 ldlm_grant_lock(lock, &ast_list);
1720 unlock_res_and_lock(lock);
1722 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1724 /* Let Enqueue to call osc_lock_upcall() and initialize
1726 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1728 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1730 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1736 lock_res_and_lock(lock);
1737 ldlm_set_failed(lock);
1738 unlock_res_and_lock(lock);
1739 wake_up(&lock->l_waitq);
1741 LDLM_LOCK_RELEASE(lock);
1745 * Callback handler for receiving incoming glimpse ASTs.
1747 * This only can happen on client side. After handling the glimpse AST
1748 * we also consider dropping the lock here if it is unused locally for a
1751 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1752 struct ldlm_namespace *ns,
1753 struct ldlm_request *dlm_req,
1754 struct ldlm_lock *lock)
1759 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1761 if (lock->l_glimpse_ast != NULL)
1762 rc = lock->l_glimpse_ast(lock, req);
1764 if (req->rq_repmsg != NULL) {
1767 req->rq_status = rc;
1771 lock_res_and_lock(lock);
1772 if (lock->l_granted_mode == LCK_PW &&
1773 !lock->l_readers && !lock->l_writers &&
1774 ktime_after(ktime_get(),
1775 ktime_add(lock->l_last_used,
1776 ktime_set(10, 0)))) {
1777 unlock_res_and_lock(lock);
1778 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1779 ldlm_handle_bl_callback(ns, NULL, lock);
1784 unlock_res_and_lock(lock);
1785 LDLM_LOCK_RELEASE(lock);
1789 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1791 if (req->rq_no_reply)
1794 req->rq_status = rc;
1795 if (!req->rq_packed_final) {
1796 rc = lustre_pack_reply(req, 1, NULL, NULL);
1800 return ptlrpc_reply(req);
1803 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1804 enum ldlm_cancel_flags cancel_flags)
1806 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1809 spin_lock(&blp->blp_lock);
1810 if (blwi->blwi_lock &&
1811 ldlm_is_discard_data(blwi->blwi_lock)) {
1812 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1813 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1815 /* other blocking callbacks are added to the regular list */
1816 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1818 spin_unlock(&blp->blp_lock);
1820 wake_up(&blp->blp_waitq);
1822 /* can not check blwi->blwi_flags as blwi could be already freed in
1824 if (!(cancel_flags & LCF_ASYNC))
1825 wait_for_completion(&blwi->blwi_comp);
1830 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1831 struct ldlm_namespace *ns,
1832 struct ldlm_lock_desc *ld,
1833 struct list_head *cancels, int count,
1834 struct ldlm_lock *lock,
1835 enum ldlm_cancel_flags cancel_flags)
1837 init_completion(&blwi->blwi_comp);
1838 INIT_LIST_HEAD(&blwi->blwi_head);
1840 if (memory_pressure_get())
1841 blwi->blwi_mem_pressure = 1;
1844 blwi->blwi_flags = cancel_flags;
1846 blwi->blwi_ld = *ld;
1848 list_add(&blwi->blwi_head, cancels);
1849 list_del_init(cancels);
1850 blwi->blwi_count = count;
1852 blwi->blwi_lock = lock;
1857 * Queues a list of locks \a cancels containing \a count locks
1858 * for later processing by a blocking thread. If \a count is zero,
1859 * then the lock referenced as \a lock is queued instead.
1861 * The blocking thread would then call ->l_blocking_ast callback in the lock.
1862 * If list addition fails an error is returned and caller is supposed to
1863 * call ->l_blocking_ast itself.
1865 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1866 struct ldlm_lock_desc *ld,
1867 struct ldlm_lock *lock,
1868 struct list_head *cancels, int count,
1869 enum ldlm_cancel_flags cancel_flags)
1873 if (cancels && count == 0)
1876 if (cancel_flags & LCF_ASYNC) {
1877 struct ldlm_bl_work_item *blwi;
1879 OBD_ALLOC(blwi, sizeof(*blwi));
1882 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
1884 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
1886 /* if it is synchronous call do minimum mem alloc, as it could
1887 * be triggered from kernel shrinker
1889 struct ldlm_bl_work_item blwi;
1891 memset(&blwi, 0, sizeof(blwi));
1892 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
1893 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
1898 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1899 struct ldlm_lock *lock)
1901 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
1904 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1905 struct list_head *cancels, int count,
1906 enum ldlm_cancel_flags cancel_flags)
1908 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
1911 int ldlm_bl_thread_wakeup(void)
1913 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
1917 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
1918 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
1920 struct obd_device *obd = req->rq_export->exp_obd;
1927 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
1929 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
1931 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1933 DEBUG_REQ(D_IOCTL, req, "no set_info key");
1936 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
1938 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1940 DEBUG_REQ(D_IOCTL, req, "no set_info val");
1943 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
1946 /* We are responsible for swabbing contents of val */
1948 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
1949 /* Pass it on to mdc (the "export" in this case) */
1950 rc = obd_set_info_async(req->rq_svc_thread->t_env,
1952 sizeof(KEY_HSM_COPYTOOL_SEND),
1953 KEY_HSM_COPYTOOL_SEND,
1956 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
1961 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
1962 const char *msg, int rc,
1963 const struct lustre_handle *handle)
1965 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
1966 "%s: [nid %s] [rc %d] [lock %#llx]",
1967 msg, libcfs_id2str(req->rq_peer), rc,
1968 handle ? handle->cookie : 0);
1969 if (req->rq_no_reply)
1970 CWARN("No reply was sent, maybe cause bug 21636.\n");
1972 CWARN("Send reply failed, maybe cause bug 21636.\n");
1975 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1976 static int ldlm_callback_handler(struct ptlrpc_request *req)
1978 struct ldlm_namespace *ns;
1979 struct ldlm_request *dlm_req;
1980 struct ldlm_lock *lock;
1984 /* Requests arrive in sender's byte order. The ptlrpc service
1985 * handler has already checked and, if necessary, byte-swapped the
1986 * incoming request message body, but I am responsible for the
1987 * message buffers. */
1989 /* do nothing for sec context finalize */
1990 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
1993 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1995 if (req->rq_export == NULL) {
1996 rc = ldlm_callback_reply(req, -ENOTCONN);
1997 ldlm_callback_errmsg(req, "Operate on unconnected server",
2002 LASSERT(req->rq_export != NULL);
2003 LASSERT(req->rq_export->exp_obd != NULL);
2005 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2006 case LDLM_BL_CALLBACK:
2007 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2009 ldlm_callback_reply(req, -(int)cfs_fail_err);
2013 case LDLM_CP_CALLBACK:
2014 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2017 case LDLM_GL_CALLBACK:
2018 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2022 rc = ldlm_handle_setinfo(req);
2023 ldlm_callback_reply(req, rc);
2025 case LLOG_ORIGIN_HANDLE_CREATE:
2026 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2027 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2029 rc = llog_origin_handle_open(req);
2030 ldlm_callback_reply(req, rc);
2032 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2033 req_capsule_set(&req->rq_pill,
2034 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2035 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2037 rc = llog_origin_handle_next_block(req);
2038 ldlm_callback_reply(req, rc);
2040 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2041 req_capsule_set(&req->rq_pill,
2042 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2043 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2045 rc = llog_origin_handle_read_header(req);
2046 ldlm_callback_reply(req, rc);
2048 case LLOG_ORIGIN_HANDLE_CLOSE:
2049 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2051 rc = llog_origin_handle_close(req);
2052 ldlm_callback_reply(req, rc);
2055 CERROR("unknown opcode %u\n",
2056 lustre_msg_get_opc(req->rq_reqmsg));
2057 ldlm_callback_reply(req, -EPROTO);
2061 ns = req->rq_export->exp_obd->obd_namespace;
2062 LASSERT(ns != NULL);
2064 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2066 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2067 if (dlm_req == NULL) {
2068 rc = ldlm_callback_reply(req, -EPROTO);
2069 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2074 /* Force a known safe race, send a cancel to the server for a lock
2075 * which the server has already started a blocking callback on. */
2076 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2077 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2078 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2080 CERROR("ldlm_cli_cancel: %d\n", rc);
2083 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2085 CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock "
2086 "disappeared\n", dlm_req->lock_handle[0].cookie);
2087 rc = ldlm_callback_reply(req, -EINVAL);
2088 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2089 &dlm_req->lock_handle[0]);
2093 if (ldlm_is_fail_loc(lock) &&
2094 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2095 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2097 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2098 lock_res_and_lock(lock);
2099 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2101 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2102 /* If somebody cancels lock and cache is already dropped,
2103 * or lock is failed before cp_ast received on client,
2104 * we can tell the server we have no lock. Otherwise, we
2105 * should send cancel after dropping the cache. */
2106 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2107 ldlm_is_failed(lock)) {
2108 LDLM_DEBUG(lock, "callback on lock %llx - lock disappeared",
2109 dlm_req->lock_handle[0].cookie);
2110 unlock_res_and_lock(lock);
2111 LDLM_LOCK_RELEASE(lock);
2112 rc = ldlm_callback_reply(req, -EINVAL);
2113 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2114 &dlm_req->lock_handle[0]);
2117 /* BL_AST locks are not needed in LRU.
2118 * Let ldlm_cancel_lru() be fast. */
2119 ldlm_lock_remove_from_lru(lock);
2120 ldlm_set_bl_ast(lock);
2122 unlock_res_and_lock(lock);
2124 /* We want the ost thread to get this reply so that it can respond
2125 * to ost requests (write cache writeback) that might be triggered
2128 * But we'd also like to be able to indicate in the reply that we're
2129 * cancelling right now, because it's unused, or have an intent result
2130 * in the reply, so we might have to push the responsibility for sending
2131 * the reply down into the AST handlers, alas. */
2133 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2134 case LDLM_BL_CALLBACK:
2135 CDEBUG(D_INODE, "blocking ast\n");
2136 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2137 if (!ldlm_is_cancel_on_block(lock)) {
2138 rc = ldlm_callback_reply(req, 0);
2139 if (req->rq_no_reply || rc)
2140 ldlm_callback_errmsg(req, "Normal process", rc,
2141 &dlm_req->lock_handle[0]);
2143 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2144 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2146 case LDLM_CP_CALLBACK:
2147 CDEBUG(D_INODE, "completion ast\n");
2148 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2149 ldlm_callback_reply(req, 0);
2150 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2152 case LDLM_GL_CALLBACK:
2153 CDEBUG(D_INODE, "glimpse ast\n");
2154 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2155 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2158 LBUG(); /* checked above */
2164 #ifdef HAVE_SERVER_SUPPORT
2166 * Main handler for canceld thread.
2168 * Separated into its own thread to avoid deadlocks.
2170 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2175 /* Requests arrive in sender's byte order. The ptlrpc service
2176 * handler has already checked and, if necessary, byte-swapped the
2177 * incoming request message body, but I am responsible for the
2178 * message buffers. */
2180 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2182 if (req->rq_export == NULL) {
2183 struct ldlm_request *dlm_req;
2185 CERROR("%s from %s arrived at %lu with bad export cookie "
2187 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2188 libcfs_nid2str(req->rq_peer.nid),
2189 req->rq_arrival_time.tv_sec,
2190 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2192 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2193 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2194 dlm_req = req_capsule_client_get(&req->rq_pill,
2196 if (dlm_req != NULL)
2197 ldlm_lock_dump_handle(D_ERROR,
2198 &dlm_req->lock_handle[0]);
2200 ldlm_callback_reply(req, -ENOTCONN);
2204 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2206 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2208 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2209 CDEBUG(D_INODE, "cancel\n");
2210 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2211 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2212 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2214 rc = ldlm_handle_cancel(req);
2219 CERROR("invalid opcode %d\n",
2220 lustre_msg_get_opc(req->rq_reqmsg));
2221 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2222 ldlm_callback_reply(req, -EINVAL);
2228 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2229 struct ldlm_lock *lock)
2231 struct ldlm_request *dlm_req;
2232 struct lustre_handle lockh;
2237 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2238 if (dlm_req == NULL)
2241 ldlm_lock2handle(lock, &lockh);
2242 for (i = 0; i < dlm_req->lock_count; i++) {
2243 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2245 DEBUG_REQ(D_RPCTRACE, req,
2246 "Prio raised by lock %#llx.", lockh.cookie);
2257 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2259 struct ldlm_request *dlm_req;
2264 /* no prolong in recovery */
2265 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2268 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2269 if (dlm_req == NULL)
2272 for (i = 0; i < dlm_req->lock_count; i++) {
2273 struct ldlm_lock *lock;
2275 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2279 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2281 LDLM_DEBUG(lock, "hpreq cancel lock");
2282 LDLM_LOCK_PUT(lock);
2291 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2292 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2293 .hpreq_check = ldlm_cancel_hpreq_check,
2297 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2301 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2303 if (req->rq_export == NULL)
2306 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2307 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2308 req->rq_ops = &ldlm_cancel_hpreq_ops;
2313 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2314 struct hlist_node *hnode, void *data)
2317 struct list_head *rpc_list = data;
2318 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2320 lock_res_and_lock(lock);
2322 if (lock->l_req_mode != lock->l_granted_mode) {
2323 unlock_res_and_lock(lock);
2327 LASSERT(lock->l_resource);
2328 if (lock->l_resource->lr_type != LDLM_IBITS &&
2329 lock->l_resource->lr_type != LDLM_PLAIN) {
2330 unlock_res_and_lock(lock);
2334 if (ldlm_is_ast_sent(lock)) {
2335 unlock_res_and_lock(lock);
2339 LASSERT(lock->l_blocking_ast);
2340 LASSERT(!lock->l_blocking_lock);
2342 ldlm_set_ast_sent(lock);
2343 if (lock->l_export && lock->l_export->exp_lock_hash) {
2344 /* NB: it's safe to call cfs_hash_del() even lock isn't
2345 * in exp_lock_hash. */
2346 /* In the function below, .hs_keycmp resolves to
2347 * ldlm_export_lock_keycmp() */
2348 /* coverity[overrun-buffer-val] */
2349 cfs_hash_del(lock->l_export->exp_lock_hash,
2350 &lock->l_remote_handle, &lock->l_exp_hash);
2353 list_add_tail(&lock->l_rk_ast, rpc_list);
2354 LDLM_LOCK_GET(lock);
2356 unlock_res_and_lock(lock);
2360 void ldlm_revoke_export_locks(struct obd_export *exp)
2362 struct list_head rpc_list;
2365 INIT_LIST_HEAD(&rpc_list);
2366 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2367 ldlm_revoke_lock_cb, &rpc_list, 0);
2368 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2369 LDLM_WORK_REVOKE_AST);
2373 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2374 #endif /* HAVE_SERVER_SUPPORT */
2376 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2377 struct ldlm_bl_work_item **p_blwi,
2378 struct obd_export **p_exp)
2380 struct ldlm_bl_work_item *blwi = NULL;
2381 static unsigned int num_bl = 0;
2382 static unsigned int num_stale;
2383 int num_th = atomic_read(&blp->blp_num_threads);
2385 *p_exp = obd_stale_export_get();
2387 spin_lock(&blp->blp_lock);
2388 if (*p_exp != NULL) {
2389 if (num_th == 1 || ++num_stale < num_th) {
2390 spin_unlock(&blp->blp_lock);
2397 /* process a request from the blp_list at least every blp_num_threads */
2398 if (!list_empty(&blp->blp_list) &&
2399 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2400 blwi = list_entry(blp->blp_list.next,
2401 struct ldlm_bl_work_item, blwi_entry);
2403 if (!list_empty(&blp->blp_prio_list))
2404 blwi = list_entry(blp->blp_prio_list.next,
2405 struct ldlm_bl_work_item,
2409 if (++num_bl >= num_th)
2411 list_del(&blwi->blwi_entry);
2413 spin_unlock(&blp->blp_lock);
2416 if (*p_exp != NULL && *p_blwi != NULL) {
2417 obd_stale_export_put(*p_exp);
2421 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2424 /* This only contains temporary data until the thread starts */
2425 struct ldlm_bl_thread_data {
2426 struct ldlm_bl_pool *bltd_blp;
2427 struct completion bltd_comp;
2431 static int ldlm_bl_thread_main(void *arg);
2433 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2435 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2436 struct task_struct *task;
2438 init_completion(&bltd.bltd_comp);
2440 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2441 if (bltd.bltd_num >= blp->blp_max_threads) {
2442 atomic_dec(&blp->blp_num_threads);
2446 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2448 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2449 atomic_dec(&blp->blp_num_threads);
2453 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2456 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2457 bltd.bltd_num, PTR_ERR(task));
2458 atomic_dec(&blp->blp_num_threads);
2459 return PTR_ERR(task);
2461 wait_for_completion(&bltd.bltd_comp);
2466 /* Not fatal if racy and have a few too many threads */
2467 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2468 struct ldlm_bl_work_item *blwi)
2470 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2473 if (atomic_read(&blp->blp_busy_threads) <
2474 atomic_read(&blp->blp_num_threads))
2477 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2478 blwi->blwi_mem_pressure))
2484 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2485 struct ldlm_bl_work_item *blwi)
2489 if (blwi->blwi_ns == NULL)
2490 /* added by ldlm_cleanup() */
2491 RETURN(LDLM_ITER_STOP);
2493 if (blwi->blwi_mem_pressure)
2494 memory_pressure_set();
2496 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2498 if (blwi->blwi_count) {
2500 /* The special case when we cancel locks in lru
2501 * asynchronously, we pass the list of locks here.
2502 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2503 * canceled locally yet. */
2504 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2507 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2510 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2513 if (blwi->blwi_mem_pressure)
2514 memory_pressure_clr();
2516 if (blwi->blwi_flags & LCF_ASYNC)
2517 OBD_FREE(blwi, sizeof(*blwi));
2519 complete(&blwi->blwi_comp);
2525 * Cancel stale locks on export. Cancel blocked locks first.
2526 * If the given export has blocked locks, the next in the list may have
2527 * them too, thus cancel not blocked locks only if the current export has
2530 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2531 struct obd_export *exp)
2536 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2538 num = ldlm_export_cancel_blocked_locks(exp);
2540 ldlm_export_cancel_locks(exp);
2542 obd_stale_export_put(exp);
2549 * Main blocking requests processing thread.
2551 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2552 * This thread in the end ends up doing actual call to ->l_blocking_ast
2555 static int ldlm_bl_thread_main(void *arg)
2557 struct ldlm_bl_pool *blp;
2558 struct ldlm_bl_thread_data *bltd = arg;
2561 blp = bltd->bltd_blp;
2563 complete(&bltd->bltd_comp);
2564 /* cannot use bltd after this, it is only on caller's stack */
2567 struct l_wait_info lwi = { 0 };
2568 struct ldlm_bl_work_item *blwi = NULL;
2569 struct obd_export *exp = NULL;
2572 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2575 l_wait_event_exclusive(blp->blp_waitq,
2576 ldlm_bl_get_work(blp, &blwi,
2579 atomic_inc(&blp->blp_busy_threads);
2581 if (ldlm_bl_thread_need_create(blp, blwi))
2582 /* discard the return value, we tried */
2583 ldlm_bl_thread_start(blp, true);
2586 rc = ldlm_bl_thread_exports(blp, exp);
2588 rc = ldlm_bl_thread_blwi(blp, blwi);
2590 atomic_dec(&blp->blp_busy_threads);
2592 if (rc == LDLM_ITER_STOP)
2595 /* If there are many namespaces, we will not sleep waiting for
2596 * work, and must do a cond_resched to avoid holding the CPU
2601 atomic_dec(&blp->blp_num_threads);
2602 complete(&blp->blp_comp);
2607 static int ldlm_setup(void);
2608 static int ldlm_cleanup(void);
2610 int ldlm_get_ref(void)
2614 mutex_lock(&ldlm_ref_mutex);
2615 if (++ldlm_refcount == 1) {
2620 mutex_unlock(&ldlm_ref_mutex);
2625 void ldlm_put_ref(void)
2628 mutex_lock(&ldlm_ref_mutex);
2629 if (ldlm_refcount == 1) {
2630 int rc = ldlm_cleanup();
2632 CERROR("ldlm_cleanup failed: %d\n", rc);
2638 mutex_unlock(&ldlm_ref_mutex);
2644 * Export handle<->lock hash operations.
2647 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2649 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2653 ldlm_export_lock_key(struct hlist_node *hnode)
2655 struct ldlm_lock *lock;
2657 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2658 return &lock->l_remote_handle;
2662 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2664 struct ldlm_lock *lock;
2666 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2667 lock->l_remote_handle = *(struct lustre_handle *)key;
2671 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2673 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2677 ldlm_export_lock_object(struct hlist_node *hnode)
2679 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2683 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2685 struct ldlm_lock *lock;
2687 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2688 LDLM_LOCK_GET(lock);
2692 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2694 struct ldlm_lock *lock;
2696 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2697 LDLM_LOCK_RELEASE(lock);
2700 static struct cfs_hash_ops ldlm_export_lock_ops = {
2701 .hs_hash = ldlm_export_lock_hash,
2702 .hs_key = ldlm_export_lock_key,
2703 .hs_keycmp = ldlm_export_lock_keycmp,
2704 .hs_keycpy = ldlm_export_lock_keycpy,
2705 .hs_object = ldlm_export_lock_object,
2706 .hs_get = ldlm_export_lock_get,
2707 .hs_put = ldlm_export_lock_put,
2708 .hs_put_locked = ldlm_export_lock_put,
2711 int ldlm_init_export(struct obd_export *exp)
2716 exp->exp_lock_hash =
2717 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2718 HASH_EXP_LOCK_CUR_BITS,
2719 HASH_EXP_LOCK_MAX_BITS,
2720 HASH_EXP_LOCK_BKT_BITS, 0,
2721 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2722 &ldlm_export_lock_ops,
2723 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2724 CFS_HASH_NBLK_CHANGE);
2726 if (!exp->exp_lock_hash)
2729 rc = ldlm_init_flock_export(exp);
2735 ldlm_destroy_export(exp);
2738 EXPORT_SYMBOL(ldlm_init_export);
2740 void ldlm_destroy_export(struct obd_export *exp)
2743 cfs_hash_putref(exp->exp_lock_hash);
2744 exp->exp_lock_hash = NULL;
2746 ldlm_destroy_flock_export(exp);
2749 EXPORT_SYMBOL(ldlm_destroy_export);
2751 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
2752 struct attribute *attr,
2755 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
2758 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
2759 struct attribute *attr,
2766 rc = kstrtoul(buffer, 10, &val);
2770 ldlm_cancel_unused_locks_before_replay = val;
2774 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
2776 static struct attribute *ldlm_attrs[] = {
2777 &lustre_attr_cancel_unused_locks_before_replay.attr,
2781 static struct attribute_group ldlm_attr_group = {
2782 .attrs = ldlm_attrs,
2785 static int ldlm_setup(void)
2787 static struct ptlrpc_service_conf conf;
2788 struct ldlm_bl_pool *blp = NULL;
2789 #ifdef HAVE_SERVER_SUPPORT
2790 struct task_struct *task;
2791 #endif /* HAVE_SERVER_SUPPORT */
2797 if (ldlm_state != NULL)
2800 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2801 if (ldlm_state == NULL)
2804 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
2808 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
2812 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
2816 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
2820 rc = ldlm_debugfs_setup();
2824 memset(&conf, 0, sizeof(conf));
2825 conf = (typeof(conf)) {
2826 .psc_name = "ldlm_cbd",
2827 .psc_watchdog_factor = 2,
2829 .bc_nbufs = LDLM_CLIENT_NBUFS,
2830 .bc_buf_size = LDLM_BUFSIZE,
2831 .bc_req_max_size = LDLM_MAXREQSIZE,
2832 .bc_rep_max_size = LDLM_MAXREPSIZE,
2833 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2834 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2837 .tc_thr_name = "ldlm_cb",
2838 .tc_thr_factor = LDLM_THR_FACTOR,
2839 .tc_nthrs_init = LDLM_NTHRS_INIT,
2840 .tc_nthrs_base = LDLM_NTHRS_BASE,
2841 .tc_nthrs_max = LDLM_NTHRS_MAX,
2842 .tc_nthrs_user = ldlm_num_threads,
2843 .tc_cpu_affinity = 1,
2844 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2847 .cc_pattern = ldlm_cpts,
2850 .so_req_handler = ldlm_callback_handler,
2853 ldlm_state->ldlm_cb_service = \
2854 ptlrpc_register_service(&conf, ldlm_svc_kset,
2855 ldlm_svc_debugfs_dir);
2856 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2857 CERROR("failed to start service\n");
2858 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2859 ldlm_state->ldlm_cb_service = NULL;
2863 #ifdef HAVE_SERVER_SUPPORT
2864 memset(&conf, 0, sizeof(conf));
2865 conf = (typeof(conf)) {
2866 .psc_name = "ldlm_canceld",
2867 .psc_watchdog_factor = 6,
2869 .bc_nbufs = LDLM_SERVER_NBUFS,
2870 .bc_buf_size = LDLM_BUFSIZE,
2871 .bc_req_max_size = LDLM_MAXREQSIZE,
2872 .bc_rep_max_size = LDLM_MAXREPSIZE,
2873 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
2874 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
2878 .tc_thr_name = "ldlm_cn",
2879 .tc_thr_factor = LDLM_THR_FACTOR,
2880 .tc_nthrs_init = LDLM_NTHRS_INIT,
2881 .tc_nthrs_base = LDLM_NTHRS_BASE,
2882 .tc_nthrs_max = LDLM_NTHRS_MAX,
2883 .tc_nthrs_user = ldlm_num_threads,
2884 .tc_cpu_affinity = 1,
2885 .tc_ctx_tags = LCT_MD_THREAD | \
2890 .cc_pattern = ldlm_cpts,
2893 .so_req_handler = ldlm_cancel_handler,
2894 .so_hpreq_handler = ldlm_hpreq_handler,
2897 ldlm_state->ldlm_cancel_service = \
2898 ptlrpc_register_service(&conf, ldlm_svc_kset,
2899 ldlm_svc_debugfs_dir);
2900 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2901 CERROR("failed to start service\n");
2902 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2903 ldlm_state->ldlm_cancel_service = NULL;
2906 #endif /* HAVE_SERVER_SUPPORT */
2908 OBD_ALLOC(blp, sizeof(*blp));
2910 GOTO(out, rc = -ENOMEM);
2911 ldlm_state->ldlm_bl_pool = blp;
2913 spin_lock_init(&blp->blp_lock);
2914 INIT_LIST_HEAD(&blp->blp_list);
2915 INIT_LIST_HEAD(&blp->blp_prio_list);
2916 init_waitqueue_head(&blp->blp_waitq);
2917 atomic_set(&blp->blp_num_threads, 0);
2918 atomic_set(&blp->blp_busy_threads, 0);
2920 if (ldlm_num_threads == 0) {
2921 blp->blp_min_threads = LDLM_NTHRS_INIT;
2922 blp->blp_max_threads = LDLM_NTHRS_MAX;
2924 blp->blp_min_threads = blp->blp_max_threads = \
2925 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
2929 for (i = 0; i < blp->blp_min_threads; i++) {
2930 rc = ldlm_bl_thread_start(blp, false);
2935 #ifdef HAVE_SERVER_SUPPORT
2936 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
2939 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2943 wait_event(expired_lock_wait_queue,
2944 expired_lock_thread_state == ELT_READY);
2945 #endif /* HAVE_SERVER_SUPPORT */
2947 rc = ldlm_pools_init();
2949 CERROR("Failed to initialize LDLM pools: %d\n", rc);
2953 rc = ldlm_reclaim_setup();
2955 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
2965 static int ldlm_cleanup(void)
2969 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2970 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2971 CERROR("ldlm still has namespaces; clean these up first.\n");
2972 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2973 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2977 ldlm_reclaim_cleanup();
2980 if (ldlm_state->ldlm_bl_pool != NULL) {
2981 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2983 while (atomic_read(&blp->blp_num_threads) > 0) {
2984 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2986 init_completion(&blp->blp_comp);
2988 spin_lock(&blp->blp_lock);
2989 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2990 wake_up(&blp->blp_waitq);
2991 spin_unlock(&blp->blp_lock);
2993 wait_for_completion(&blp->blp_comp);
2996 OBD_FREE(blp, sizeof(*blp));
2999 if (ldlm_state->ldlm_cb_service != NULL)
3000 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3001 #ifdef HAVE_SERVER_SUPPORT
3002 if (ldlm_state->ldlm_cancel_service != NULL)
3003 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3007 kset_unregister(ldlm_ns_kset);
3009 kset_unregister(ldlm_svc_kset);
3011 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3012 kobject_put(ldlm_kobj);
3015 ldlm_debugfs_cleanup();
3017 #ifdef HAVE_SERVER_SUPPORT
3018 if (expired_lock_thread_state != ELT_STOPPED) {
3019 expired_lock_thread_state = ELT_TERMINATE;
3020 wake_up(&expired_lock_wait_queue);
3021 wait_event(expired_lock_wait_queue,
3022 expired_lock_thread_state == ELT_STOPPED);
3026 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3034 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3035 sizeof(struct ldlm_resource), 0,
3036 SLAB_HWCACHE_ALIGN, NULL);
3037 if (ldlm_resource_slab == NULL)
3040 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3041 sizeof(struct ldlm_lock), 0,
3042 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3043 if (ldlm_lock_slab == NULL)
3046 ldlm_interval_slab = kmem_cache_create("interval_node",
3047 sizeof(struct ldlm_interval),
3048 0, SLAB_HWCACHE_ALIGN, NULL);
3049 if (ldlm_interval_slab == NULL)
3052 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3053 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3054 0, SLAB_HWCACHE_ALIGN, NULL);
3055 if (ldlm_interval_tree_slab == NULL)
3058 #ifdef HAVE_SERVER_SUPPORT
3059 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3060 sizeof(struct ldlm_glimpse_work),
3062 if (ldlm_glimpse_work_kmem == NULL)
3063 goto out_interval_tree;
3066 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3067 class_export_dump_hook = ldlm_dump_export_locks;
3070 #ifdef HAVE_SERVER_SUPPORT
3072 kmem_cache_destroy(ldlm_interval_tree_slab);
3075 kmem_cache_destroy(ldlm_interval_slab);
3077 kmem_cache_destroy(ldlm_lock_slab);
3079 kmem_cache_destroy(ldlm_resource_slab);
3084 void ldlm_exit(void)
3087 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3088 kmem_cache_destroy(ldlm_resource_slab);
3089 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3090 * synchronize_rcu() to wait a grace period elapsed, so that
3091 * ldlm_lock_free() get a chance to be called. */
3093 kmem_cache_destroy(ldlm_lock_slab);
3094 kmem_cache_destroy(ldlm_interval_slab);
3095 kmem_cache_destroy(ldlm_interval_tree_slab);
3096 #ifdef HAVE_SERVER_SUPPORT
3097 kmem_cache_destroy(ldlm_glimpse_work_kmem);