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 unsigned int ldlm_cpu_bind = 1;
53 module_param(ldlm_cpu_bind, uint, 0444);
54 MODULE_PARM_DESC(ldlm_cpu_bind,
55 "bind DLM service threads to particular CPU partitions");
57 static char *ldlm_cpts;
58 module_param(ldlm_cpts, charp, 0444);
59 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
61 static DEFINE_MUTEX(ldlm_ref_mutex);
62 static int ldlm_refcount;
64 struct kobject *ldlm_kobj;
65 struct kset *ldlm_ns_kset;
66 struct kset *ldlm_svc_kset;
70 static struct ldlm_state *ldlm_state;
73 * timeout for initial callback (AST) reply (bz10399)
74 * Due to having to send a 32 bit time value over the
75 * wire return it as time_t instead of time64_t
77 static inline time_t ldlm_get_rq_timeout(void)
80 time_t timeout = min(ldlm_timeout, obd_timeout / 3);
82 return timeout < 1 ? 1 : timeout;
89 * blp_prio_list is used for callbacks that should be handled
90 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
93 struct list_head blp_prio_list;
96 * blp_list is used for all other callbacks which are likely
97 * to take longer to process.
99 struct list_head blp_list;
101 wait_queue_head_t blp_waitq;
102 struct completion blp_comp;
103 atomic_t blp_num_threads;
104 atomic_t blp_busy_threads;
109 struct ldlm_bl_work_item {
110 struct list_head blwi_entry;
111 struct ldlm_namespace *blwi_ns;
112 struct ldlm_lock_desc blwi_ld;
113 struct ldlm_lock *blwi_lock;
114 struct list_head blwi_head;
116 struct completion blwi_comp;
117 enum ldlm_cancel_flags blwi_flags;
118 int blwi_mem_pressure;
121 #ifdef HAVE_SERVER_SUPPORT
124 * Protects both waiting_locks_list and expired_lock_thread.
126 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
129 * List for contended locks.
131 * As soon as a lock is contended, it gets placed on this list and
132 * expected time to get a response is filled in the lock. A special
133 * thread walks the list looking for locks that should be released and
134 * schedules client evictions for those that have not been released in
137 * All access to it should be under waiting_locks_spinlock.
139 static LIST_HEAD(waiting_locks_list);
140 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
141 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
149 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
150 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
151 static int expired_lock_dump;
152 static LIST_HEAD(expired_lock_list);
154 static int ldlm_lock_busy(struct ldlm_lock *lock);
155 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout);
156 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds);
158 static inline int have_expired_locks(void)
163 spin_lock_bh(&waiting_locks_spinlock);
164 need_to_run = !list_empty(&expired_lock_list);
165 spin_unlock_bh(&waiting_locks_spinlock);
171 * Check expired lock list for expired locks and time them out.
173 static int expired_lock_main(void *arg)
175 struct list_head *expired = &expired_lock_list;
176 struct l_wait_info lwi = { 0 };
181 expired_lock_thread_state = ELT_READY;
182 wake_up(&expired_lock_wait_queue);
185 l_wait_event(expired_lock_wait_queue,
186 have_expired_locks() ||
187 expired_lock_thread_state == ELT_TERMINATE,
190 spin_lock_bh(&waiting_locks_spinlock);
191 if (expired_lock_dump) {
192 spin_unlock_bh(&waiting_locks_spinlock);
194 /* from waiting_locks_callback, but not in timer */
195 libcfs_debug_dumplog();
197 spin_lock_bh(&waiting_locks_spinlock);
198 expired_lock_dump = 0;
203 while (!list_empty(expired)) {
204 struct obd_export *export;
205 struct ldlm_lock *lock;
207 lock = list_entry(expired->next, struct ldlm_lock,
209 if ((void *)lock < LP_POISON + PAGE_SIZE &&
210 (void *)lock >= LP_POISON) {
211 spin_unlock_bh(&waiting_locks_spinlock);
212 CERROR("free lock on elt list %p\n", lock);
215 list_del_init(&lock->l_pending_chain);
216 if ((void *)lock->l_export <
217 LP_POISON + PAGE_SIZE &&
218 (void *)lock->l_export >= LP_POISON) {
219 CERROR("lock with free export on elt list %p\n",
221 lock->l_export = NULL;
222 LDLM_ERROR(lock, "free export");
224 * release extra ref grabbed by
225 * ldlm_add_waiting_lock() or
228 LDLM_LOCK_RELEASE(lock);
232 if (ldlm_is_destroyed(lock)) {
234 * release the lock refcount where
235 * waiting_locks_callback() founds
237 LDLM_LOCK_RELEASE(lock);
240 export = class_export_lock_get(lock->l_export, lock);
241 spin_unlock_bh(&waiting_locks_spinlock);
243 /* Check if we need to prolong timeout */
244 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
245 lock->l_callback_timeout != 0 && /* not AST error */
246 ldlm_lock_busy(lock)) {
247 LDLM_DEBUG(lock, "prolong the busy lock");
248 lock_res_and_lock(lock);
249 ldlm_add_waiting_lock(lock,
250 ldlm_bl_timeout(lock) >> 1);
251 unlock_res_and_lock(lock);
253 spin_lock_bh(&export->exp_bl_list_lock);
254 list_del_init(&lock->l_exp_list);
255 spin_unlock_bh(&export->exp_bl_list_lock);
258 "lock callback timer expired after %llds: evicting client at %s ",
259 ktime_get_real_seconds() -
261 obd_export_nid2str(export));
262 ldlm_lock_to_ns(lock)->ns_timeouts++;
264 class_fail_export(export);
266 class_export_lock_put(export, lock);
268 * release extra ref grabbed by ldlm_add_waiting_lock()
269 * or ldlm_failed_ast()
271 LDLM_LOCK_RELEASE(lock);
273 spin_lock_bh(&waiting_locks_spinlock);
275 spin_unlock_bh(&waiting_locks_spinlock);
277 if (do_dump && obd_dump_on_eviction) {
278 CERROR("dump the log upon eviction\n");
279 libcfs_debug_dumplog();
282 if (expired_lock_thread_state == ELT_TERMINATE)
286 expired_lock_thread_state = ELT_STOPPED;
287 wake_up(&expired_lock_wait_queue);
292 * Check if there is a request in the export request list
293 * which prevents the lock canceling.
295 static int ldlm_lock_busy(struct ldlm_lock *lock)
297 struct ptlrpc_request *req;
302 if (lock->l_export == NULL)
305 spin_lock(&lock->l_export->exp_rpc_lock);
306 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
308 if (req->rq_ops->hpreq_lock_match) {
309 match = req->rq_ops->hpreq_lock_match(req, lock);
314 spin_unlock(&lock->l_export->exp_rpc_lock);
318 /* This is called from within a timer interrupt and cannot schedule */
319 static void waiting_locks_callback(TIMER_DATA_TYPE unused)
321 struct ldlm_lock *lock;
324 spin_lock_bh(&waiting_locks_spinlock);
325 while (!list_empty(&waiting_locks_list)) {
326 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
328 if (lock->l_callback_timeout > ktime_get_seconds() ||
329 lock->l_req_mode == LCK_GROUP)
333 * no needs to take an extra ref on the lock since it was in
334 * the waiting_locks_list and ldlm_add_waiting_lock()
335 * already grabbed a ref
337 list_del(&lock->l_pending_chain);
338 list_add(&lock->l_pending_chain, &expired_lock_list);
342 if (!list_empty(&expired_lock_list)) {
343 if (obd_dump_on_timeout && need_dump)
344 expired_lock_dump = __LINE__;
346 wake_up(&expired_lock_wait_queue);
350 * Make sure the timer will fire again if we have any locks
353 if (!list_empty(&waiting_locks_list)) {
354 unsigned long timeout_jiffies;
356 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
358 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
359 mod_timer(&waiting_locks_timer, timeout_jiffies);
361 spin_unlock_bh(&waiting_locks_spinlock);
365 * Add lock to the list of contended locks.
367 * Indicate that we're waiting for a client to call us back cancelling a given
368 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
369 * timer to fire appropriately. (We round up to the next second, to avoid
370 * floods of timer firings during periods of high lock contention and traffic).
371 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
372 * if it has been added to the waiting list (1 is returned).
374 * Called with the namespace lock held.
376 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds)
378 unsigned long timeout_jiffies;
381 if (!list_empty(&lock->l_pending_chain))
384 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
385 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
388 timeout = ktime_get_seconds() + seconds;
389 if (likely(timeout > lock->l_callback_timeout))
390 lock->l_callback_timeout = timeout;
392 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
394 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
395 !timer_pending(&waiting_locks_timer))
396 mod_timer(&waiting_locks_timer, timeout_jiffies);
399 * if the new lock has a shorter timeout than something earlier on
400 * the list, we'll wait the longer amount of time; no big deal.
403 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
407 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
409 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
410 if (list_empty(&lock->l_exp_list)) {
411 if (!ldlm_is_granted(lock))
412 list_add_tail(&lock->l_exp_list,
413 &lock->l_export->exp_bl_list);
415 list_add(&lock->l_exp_list,
416 &lock->l_export->exp_bl_list);
418 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
421 * A blocked lock is added. Adjust the position in
422 * the stale list if the export is in the list.
423 * If export is stale and not in the list - it is being
424 * processed and will be placed on the right position
425 * on obd_stale_export_put().
427 if (!list_empty(&lock->l_export->exp_stale_list))
428 obd_stale_export_adjust(lock->l_export);
431 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
435 /* NB: must be called with hold of lock_res_and_lock() */
436 LASSERT(ldlm_is_res_locked(lock));
437 LASSERT(!ldlm_is_cancel_on_block(lock));
440 * Do not put cross-MDT lock in the waiting list, since we
441 * will not evict it due to timeout for now
443 if (lock->l_export != NULL &&
444 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
447 spin_lock_bh(&waiting_locks_spinlock);
448 if (ldlm_is_cancel(lock)) {
449 spin_unlock_bh(&waiting_locks_spinlock);
453 if (ldlm_is_destroyed(lock)) {
454 static time64_t next;
456 spin_unlock_bh(&waiting_locks_spinlock);
457 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
458 if (ktime_get_seconds() > next) {
459 next = ktime_get_seconds() + 14400;
460 libcfs_debug_dumpstack(NULL);
465 ldlm_set_waited(lock);
466 lock->l_blast_sent = ktime_get_real_seconds();
467 ret = __ldlm_add_waiting_lock(lock, timeout);
470 * grab ref on the lock if it has been added to the
475 spin_unlock_bh(&waiting_locks_spinlock);
478 ldlm_add_blocked_lock(lock);
480 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)",
481 ret == 0 ? "not re-" : "", timeout,
482 AT_OFF ? "off" : "on");
487 * Remove a lock from the pending list, likely because it had its cancellation
488 * callback arrive without incident. This adjusts the lock-timeout timer if
489 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
490 * As done by ldlm_del_waiting_lock(), the caller must release the lock
491 * reference when the lock is removed from any list (1 is returned).
493 * Called with namespace lock held.
495 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
497 struct list_head *list_next;
499 if (list_empty(&lock->l_pending_chain))
502 list_next = lock->l_pending_chain.next;
503 if (lock->l_pending_chain.prev == &waiting_locks_list) {
504 /* Removing the head of the list, adjust timer. */
505 if (list_next == &waiting_locks_list) {
506 /* No more, just cancel. */
507 del_timer(&waiting_locks_timer);
509 struct ldlm_lock *next;
511 next = list_entry(list_next, struct ldlm_lock,
513 mod_timer(&waiting_locks_timer,
514 cfs_time_seconds(next->l_callback_timeout));
517 list_del_init(&lock->l_pending_chain);
522 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
526 if (lock->l_export == NULL) {
527 /* We don't have a "waiting locks list" on clients. */
528 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
532 spin_lock_bh(&waiting_locks_spinlock);
533 ret = __ldlm_del_waiting_lock(lock);
534 ldlm_clear_waited(lock);
535 spin_unlock_bh(&waiting_locks_spinlock);
537 /* remove the lock out of export blocking list */
538 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
539 list_del_init(&lock->l_exp_list);
540 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
544 * release lock ref if it has indeed been removed
547 LDLM_LOCK_RELEASE(lock);
550 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
555 * Prolong the contended lock waiting time.
557 * Called with namespace lock held.
559 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
561 if (lock->l_export == NULL) {
562 /* We don't have a "waiting locks list" on clients. */
563 LDLM_DEBUG(lock, "client lock: no-op");
567 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
568 /* We don't have a "waiting locks list" on OSP. */
569 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
573 spin_lock_bh(&waiting_locks_spinlock);
575 if (list_empty(&lock->l_pending_chain)) {
576 spin_unlock_bh(&waiting_locks_spinlock);
577 LDLM_DEBUG(lock, "wasn't waiting");
582 * we remove/add the lock to the waiting list, so no needs to
583 * release/take a lock reference
585 __ldlm_del_waiting_lock(lock);
586 __ldlm_add_waiting_lock(lock, timeout);
587 spin_unlock_bh(&waiting_locks_spinlock);
589 LDLM_DEBUG(lock, "refreshed");
592 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
594 #else /* HAVE_SERVER_SUPPORT */
596 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
601 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
606 #endif /* !HAVE_SERVER_SUPPORT */
608 #ifdef HAVE_SERVER_SUPPORT
611 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
612 * lock cancel, and their replies). Used for lock callback timeout and AST
615 * \param[in] lock lock which is getting the blocking callback
617 * \retval timeout in seconds to wait for the client reply
619 time64_t ldlm_bl_timeout(struct ldlm_lock *lock)
624 return obd_timeout / 2;
627 * Since these are non-updating timeouts, we should be conservative.
628 * Take more than usually, 150%
629 * It would be nice to have some kind of "early reply" mechanism for
630 * lock callbacks too...
632 timeout = at_get(&lock->l_export->exp_bl_lock_at);
633 return max(timeout + (timeout >> 1), (time64_t)ldlm_enqueue_min);
635 EXPORT_SYMBOL(ldlm_bl_timeout);
638 * Perform lock cleanup if AST sending failed.
640 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
641 const char *ast_type)
643 LCONSOLE_ERROR_MSG(0x138,
644 "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
645 lock->l_export->exp_obd->obd_name,
646 obd_export_nid2str(lock->l_export), ast_type, rc);
648 if (obd_dump_on_timeout)
649 libcfs_debug_dumplog();
650 spin_lock_bh(&waiting_locks_spinlock);
651 if (__ldlm_del_waiting_lock(lock) == 0)
653 * the lock was not in any list, grab an extra ref before adding
654 * the lock to the expired list
657 lock->l_callback_timeout = 0; /* differentiate it from expired locks */
658 list_add(&lock->l_pending_chain, &expired_lock_list);
659 wake_up(&expired_lock_wait_queue);
660 spin_unlock_bh(&waiting_locks_spinlock);
664 * Perform lock cleanup if AST reply came with error.
666 static int ldlm_handle_ast_error(const struct lu_env *env,
667 struct ldlm_lock *lock,
668 struct ptlrpc_request *req, int rc,
669 const char *ast_type)
671 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
673 if (!req->rq_replied || (rc && rc != -EINVAL)) {
674 if (ldlm_is_cancel(lock)) {
676 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
677 ast_type, req, req->rq_xid,
678 libcfs_nid2str(peer.nid));
679 ldlm_lock_cancel(lock);
681 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
683 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
685 * Upon umount process the AST fails because cannot be
686 * sent. This shouldn't lead to the client eviction.
687 * -ENODEV error is returned by ptl_send_rpc() for
688 * new request in such import.
689 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
690 * if imp_invalid is set or obd_no_recov.
691 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
692 * in ptlrpc_import_delay_req() as well with -EIO code.
693 * In all such cases errors are ignored.
696 "%s AST can't be sent due to a server %s failure or umount process: rc = %d\n",
698 req->rq_import->imp_obd->obd_name, rc);
701 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
702 libcfs_nid2str(peer.nid),
703 req->rq_replied ? "returned error from" :
704 "failed to reply to",
705 ast_type, req, req->rq_xid,
706 (req->rq_repmsg != NULL) ?
707 lustre_msg_get_status(req->rq_repmsg) : 0,
709 ldlm_failed_ast(lock, rc, ast_type);
715 struct ldlm_resource *res = lock->l_resource;
718 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
719 libcfs_nid2str(peer.nid),
721 lustre_msg_get_status(req->rq_repmsg) : -1,
722 ast_type, req, req->rq_xid);
725 * update lvbo to return proper attributes.
728 ldlm_resource_getref(res);
729 ldlm_lvbo_update(env, res, lock, NULL, 1);
730 ldlm_resource_putref(res);
732 ldlm_lock_cancel(lock);
739 static int ldlm_cb_interpret(const struct lu_env *env,
740 struct ptlrpc_request *req, void *args, int rc)
742 struct ldlm_cb_async_args *ca = args;
743 struct ldlm_lock *lock = ca->ca_lock;
744 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
748 LASSERT(lock != NULL);
751 case LDLM_GL_CALLBACK:
753 * Update the LVB from disk if the AST failed
754 * (this is a legal race)
756 * - Glimpse callback of local lock just returns
757 * -ELDLM_NO_LOCK_DATA.
758 * - Glimpse callback of remote lock might return
759 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
761 if (unlikely(arg->gl_interpret_reply)) {
762 rc = arg->gl_interpret_reply(env, req, args, rc);
763 } else if (rc == -ELDLM_NO_LOCK_DATA) {
765 "lost race - client has a lock but no inode");
766 ldlm_lvbo_update(env, lock->l_resource, lock, NULL, 1);
767 } else if (rc != 0) {
768 rc = ldlm_handle_ast_error(env, lock, req,
771 rc = ldlm_lvbo_update(env, lock->l_resource,
775 case LDLM_BL_CALLBACK:
777 rc = ldlm_handle_ast_error(env, lock, req,
780 case LDLM_CP_CALLBACK:
782 rc = ldlm_handle_ast_error(env, lock, req,
786 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
791 /* release extra reference taken in ldlm_ast_fini() */
792 LDLM_LOCK_RELEASE(lock);
795 atomic_inc(&arg->restart);
800 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
802 struct ldlm_cb_async_args *ca = data;
803 struct ldlm_lock *lock = ca->ca_lock;
805 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
808 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
809 struct ldlm_cb_set_arg *arg,
810 struct ldlm_lock *lock,
817 if (unlikely(instant_cancel)) {
818 rc = ptl_send_rpc(req, 1);
819 ptlrpc_req_finished(req);
821 atomic_inc(&arg->restart);
824 ptlrpc_set_add_req(arg->set, req);
831 * Check if there are requests in the export request list which prevent
832 * the lock canceling and make these requests high priority ones.
834 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
836 struct ptlrpc_request *req;
840 if (lock->l_export == NULL) {
841 LDLM_DEBUG(lock, "client lock: no-op");
845 spin_lock(&lock->l_export->exp_rpc_lock);
846 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
849 * Do not process requests that were not yet added to there
850 * incoming queue or were already removed from there for
851 * processing. We evaluate ptlrpc_nrs_req_can_move() without
852 * holding svcpt->scp_req_lock, and then redo the check with
853 * the lock held once we need to obtain a reliable result.
855 if (ptlrpc_nrs_req_can_move(req) &&
856 req->rq_ops->hpreq_lock_match &&
857 req->rq_ops->hpreq_lock_match(req, lock))
858 ptlrpc_nrs_req_hp_move(req);
860 spin_unlock(&lock->l_export->exp_rpc_lock);
865 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
866 * enqueued server lock conflicts with given one.
868 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
869 * to wait for client response.
871 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
872 struct ldlm_lock_desc *desc,
873 void *data, int flag)
875 struct ldlm_cb_async_args *ca;
876 struct ldlm_cb_set_arg *arg = data;
877 struct ldlm_request *body;
878 struct ptlrpc_request *req;
879 int instant_cancel = 0;
884 if (flag == LDLM_CB_CANCELING)
885 /* Don't need to do anything here. */
888 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
889 LDLM_DEBUG(lock, "dropping BL AST");
894 LASSERT(data != NULL);
895 if (lock->l_export->exp_obd->obd_recovering != 0)
896 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
898 ldlm_lock_reorder_req(lock);
900 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
901 &RQF_LDLM_BL_CALLBACK,
902 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
906 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
907 ca = ptlrpc_req_async_args(req);
908 ca->ca_set_arg = arg;
911 req->rq_interpret_reply = ldlm_cb_interpret;
913 lock_res_and_lock(lock);
914 if (ldlm_is_destroyed(lock)) {
915 /* What's the point? */
916 unlock_res_and_lock(lock);
917 ptlrpc_req_finished(req);
921 if (!ldlm_is_granted(lock)) {
923 * this blocking AST will be communicated as part of the
924 * completion AST instead
926 ldlm_add_blocked_lock(lock);
927 ldlm_set_waited(lock);
928 unlock_res_and_lock(lock);
930 ptlrpc_req_finished(req);
931 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
935 if (ldlm_is_cancel_on_block(lock))
938 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
939 body->lock_handle[0] = lock->l_remote_handle;
940 body->lock_desc = *desc;
941 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
943 LDLM_DEBUG(lock, "server preparing blocking AST");
945 ptlrpc_request_set_replen(req);
946 ldlm_set_cbpending(lock);
947 if (instant_cancel) {
948 unlock_res_and_lock(lock);
949 ldlm_lock_cancel(lock);
951 req->rq_no_resend = 1;
953 LASSERT(ldlm_is_granted(lock));
954 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
955 unlock_res_and_lock(lock);
957 /* Do not resend after lock callback timeout */
958 req->rq_delay_limit = ldlm_bl_timeout(lock);
959 req->rq_resend_cb = ldlm_update_resend;
962 req->rq_send_state = LUSTRE_IMP_FULL;
963 /* ptlrpc_request_alloc_pack already set timeout */
965 req->rq_timeout = ldlm_get_rq_timeout();
967 if (lock->l_export && lock->l_export->exp_nid_stats &&
968 lock->l_export->exp_nid_stats->nid_ldlm_stats)
969 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
970 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
972 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
978 * ->l_completion_ast callback for a remote lock in server namespace.
980 * Sends AST to the client notifying it of lock granting. If initial
981 * lock response was not sent yet, instead of sending another RPC, just
982 * mark the lock as granted and client will understand
984 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
986 struct ldlm_cb_set_arg *arg = data;
987 struct ldlm_request *body;
988 struct ptlrpc_request *req;
989 struct ldlm_cb_async_args *ca;
990 int instant_cancel = 0;
996 LASSERT(lock != NULL);
997 LASSERT(data != NULL);
999 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1000 LDLM_DEBUG(lock, "dropping CP AST");
1004 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1005 &RQF_LDLM_CP_CALLBACK);
1009 /* server namespace, doesn't need lock */
1010 lvb_len = ldlm_lvbo_size(lock);
1012 * LU-3124 & LU-2187: to not return layout in completion AST because
1013 * it may deadlock for LU-2187, or client may not have enough space
1014 * for large layout. The layout will be returned to client with an
1015 * extra RPC to fetch xattr.lov
1017 if (ldlm_has_layout(lock))
1020 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1021 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1023 ptlrpc_request_free(req);
1027 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1028 ca = ptlrpc_req_async_args(req);
1029 ca->ca_set_arg = arg;
1032 req->rq_interpret_reply = ldlm_cb_interpret;
1033 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1035 body->lock_handle[0] = lock->l_remote_handle;
1036 body->lock_flags = ldlm_flags_to_wire(flags);
1037 ldlm_lock2desc(lock, &body->lock_desc);
1039 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1040 const struct lu_env *env = NULL;
1042 if (req->rq_svc_thread)
1043 env = req->rq_svc_thread->t_env;
1045 lvb_len = ldlm_lvbo_fill(env, lock, lvb, &lvb_len);
1048 * We still need to send the RPC to wake up the blocked
1049 * enqueue thread on the client.
1051 * Consider old client, there is no better way to notify
1052 * the failure, just zero-sized the LVB, then the client
1053 * will fail out as "-EPROTO".
1055 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1059 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1064 LDLM_DEBUG(lock, "server preparing completion AST");
1066 ptlrpc_request_set_replen(req);
1068 req->rq_send_state = LUSTRE_IMP_FULL;
1069 /* ptlrpc_request_pack already set timeout */
1071 req->rq_timeout = ldlm_get_rq_timeout();
1073 /* We only send real blocking ASTs after the lock is granted */
1074 lock_res_and_lock(lock);
1075 if (ldlm_is_ast_sent(lock)) {
1076 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1077 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1078 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1082 * We might get here prior to ldlm_handle_enqueue setting
1083 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1084 * into waiting list, but this is safe and similar code in
1085 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1086 * that would not only cancel the lock, but will also remove
1087 * it from waiting list
1089 if (ldlm_is_cancel_on_block(lock)) {
1090 unlock_res_and_lock(lock);
1091 ldlm_lock_cancel(lock);
1094 req->rq_no_resend = 1;
1096 lock_res_and_lock(lock);
1098 /* start the lock-timeout clock */
1099 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1100 /* Do not resend after lock callback timeout */
1101 req->rq_delay_limit = ldlm_bl_timeout(lock);
1102 req->rq_resend_cb = ldlm_update_resend;
1105 unlock_res_and_lock(lock);
1107 if (lock->l_export && lock->l_export->exp_nid_stats &&
1108 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1109 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1110 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1112 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1114 RETURN(lvb_len < 0 ? lvb_len : rc);
1118 * Server side ->l_glimpse_ast handler for client locks.
1120 * Sends glimpse AST to the client and waits for reply. Then updates
1121 * lvbo with the result.
1123 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1125 struct ldlm_cb_set_arg *arg = data;
1126 struct ldlm_request *body;
1127 struct ptlrpc_request *req;
1128 struct ldlm_cb_async_args *ca;
1130 struct req_format *req_fmt;
1134 LASSERT(lock != NULL);
1136 if (arg->gl_desc != NULL)
1137 /* There is a glimpse descriptor to pack */
1138 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1140 req_fmt = &RQF_LDLM_GL_CALLBACK;
1142 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1143 req_fmt, LUSTRE_DLM_VERSION,
1149 if (arg->gl_desc != NULL) {
1150 /* copy the GL descriptor */
1151 union ldlm_gl_desc *desc;
1153 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1154 *desc = *arg->gl_desc;
1157 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1158 body->lock_handle[0] = lock->l_remote_handle;
1159 ldlm_lock2desc(lock, &body->lock_desc);
1161 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1162 ca = ptlrpc_req_async_args(req);
1163 ca->ca_set_arg = arg;
1166 /* server namespace, doesn't need lock */
1167 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1168 ldlm_lvbo_size(lock));
1169 ptlrpc_request_set_replen(req);
1171 req->rq_send_state = LUSTRE_IMP_FULL;
1172 /* ptlrpc_request_alloc_pack already set timeout */
1174 req->rq_timeout = ldlm_get_rq_timeout();
1176 req->rq_interpret_reply = ldlm_cb_interpret;
1178 if (lock->l_export && lock->l_export->exp_nid_stats) {
1179 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1181 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1182 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1185 rc = ldlm_ast_fini(req, arg, lock, 0);
1189 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1191 int ldlm_glimpse_locks(struct ldlm_resource *res,
1192 struct list_head *gl_work_list)
1198 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1200 if (rc == -ERESTART)
1201 ldlm_reprocess_all(res);
1205 EXPORT_SYMBOL(ldlm_glimpse_locks);
1207 /* return LDLM lock associated with a lock callback request */
1208 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1210 struct ldlm_cb_async_args *ca;
1211 struct ldlm_lock *lock;
1215 ca = ptlrpc_req_async_args(req);
1218 RETURN(ERR_PTR(-EFAULT));
1222 EXPORT_SYMBOL(ldlm_request_lock);
1225 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1226 * service threads to carry out client lock enqueueing requests.
1228 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1229 struct ptlrpc_request *req,
1230 const struct ldlm_request *dlm_req,
1231 const struct ldlm_callback_suite *cbs)
1233 struct ldlm_reply *dlm_rep;
1235 enum ldlm_error err = ELDLM_OK;
1236 struct ldlm_lock *lock = NULL;
1237 void *cookie = NULL;
1239 struct ldlm_resource *res = NULL;
1240 const struct lu_env *env = req->rq_svc_thread->t_env;
1244 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1246 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1247 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1249 LASSERT(req->rq_export);
1251 /* for intent enqueue the stat will be updated inside intent policy */
1252 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1253 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1254 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1256 if (req->rq_export && req->rq_export->exp_nid_stats &&
1257 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1258 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1259 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1261 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1262 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1263 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1264 dlm_req->lock_desc.l_resource.lr_type);
1265 GOTO(out, rc = -EFAULT);
1268 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1269 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1270 dlm_req->lock_desc.l_req_mode &
1271 (dlm_req->lock_desc.l_req_mode-1))) {
1272 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1273 dlm_req->lock_desc.l_req_mode);
1274 GOTO(out, rc = -EFAULT);
1277 if (unlikely((flags & LDLM_FL_REPLAY) ||
1278 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1279 /* Find an existing lock in the per-export lock hash */
1281 * In the function below, .hs_keycmp resolves to
1282 * ldlm_export_lock_keycmp()
1284 /* coverity[overrun-buffer-val] */
1285 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1286 (void *)&dlm_req->lock_handle[0]);
1288 DEBUG_REQ(D_DLMTRACE, req,
1289 "found existing lock cookie %#llx",
1290 lock->l_handle.h_cookie);
1291 flags |= LDLM_FL_RESENT;
1292 GOTO(existing_lock, rc = 0);
1295 if (ldlm_reclaim_full()) {
1296 DEBUG_REQ(D_DLMTRACE, req,
1297 "Too many granted locks, reject current enqueue request and let the client retry later.\n");
1298 GOTO(out, rc = -EINPROGRESS);
1302 /* The lock's callback data might be set in the policy function */
1303 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1304 dlm_req->lock_desc.l_resource.lr_type,
1305 dlm_req->lock_desc.l_req_mode,
1306 cbs, NULL, 0, LVB_T_NONE);
1313 lock->l_remote_handle = dlm_req->lock_handle[0];
1314 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1317 * Initialize resource lvb but not for a lock being replayed since
1318 * Client already got lvb sent in this case.
1319 * This must occur early since some policy methods assume resource
1320 * lvb is available (lr_lvb_data != NULL).
1322 res = lock->l_resource;
1323 if (!(flags & LDLM_FL_REPLAY)) {
1324 /* non-replayed lock, delayed lvb init may need to be done */
1325 rc = ldlm_lvbo_init(env, res);
1327 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1332 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1334 * Don't enqueue a lock onto the export if it is been disonnected
1335 * due to eviction (b=3822) or server umount (b=24324).
1336 * Cancel it now instead.
1338 if (req->rq_export->exp_disconnected) {
1339 LDLM_ERROR(lock, "lock on disconnected export %p",
1341 GOTO(out, rc = -ENOTCONN);
1344 lock->l_export = class_export_lock_get(req->rq_export, lock);
1345 if (lock->l_export->exp_lock_hash)
1346 cfs_hash_add(lock->l_export->exp_lock_hash,
1347 &lock->l_remote_handle,
1351 * Inherit the enqueue flags before the operation, because we do not
1352 * keep the res lock on return and next operations (BL AST) may proceed
1355 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1356 LDLM_FL_INHERIT_MASK);
1358 ldlm_convert_policy_to_local(req->rq_export,
1359 dlm_req->lock_desc.l_resource.lr_type,
1360 &dlm_req->lock_desc.l_policy_data,
1361 &lock->l_policy_data);
1362 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1363 lock->l_req_extent = lock->l_policy_data.l_extent;
1366 if (flags & LDLM_FL_HAS_INTENT) {
1368 * In this case, the reply buffer is allocated deep in
1369 * local_lock_enqueue by the policy function.
1374 * based on the assumption that lvb size never changes during
1375 * resource life time otherwise it need resource->lr_lock's
1378 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1379 RCL_SERVER, ldlm_lvbo_size(lock));
1381 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1382 GOTO(out, rc = -ENOMEM);
1384 rc = req_capsule_server_pack(&req->rq_pill);
1389 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1396 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1398 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1399 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1401 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1402 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1405 * We never send a blocking AST until the lock is granted, but
1406 * we can tell it right now
1408 lock_res_and_lock(lock);
1411 * Now take into account flags to be inherited from original lock
1412 * request both in reply to client and in our own lock flags.
1414 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1415 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1418 * Don't move a pending lock onto the export if it has already been
1419 * disconnected due to eviction (b=5683) or server umount (b=24324).
1420 * Cancel it now instead.
1422 if (unlikely(req->rq_export->exp_disconnected ||
1423 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1424 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1426 } else if (ldlm_is_ast_sent(lock)) {
1427 /* fill lock desc for possible lock convert */
1428 if (lock->l_blocking_lock &&
1429 lock->l_resource->lr_type == LDLM_IBITS) {
1430 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1431 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1434 "save blocking bits %llx in granted lock",
1435 bl_lock->l_policy_data.l_inodebits.bits);
1437 * If lock is blocked then save blocking ibits
1438 * in returned lock policy for the possible lock
1439 * convert on a client.
1441 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1442 bl_lock->l_policy_data.l_inodebits.bits;
1444 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1445 if (ldlm_is_granted(lock)) {
1447 * Only cancel lock if it was granted, because it would
1448 * be destroyed immediately and would never be granted
1449 * in the future, causing timeouts on client. Not
1450 * granted lock will be cancelled immediately after
1451 * sending completion AST.
1453 if (ldlm_is_cancel_on_block(lock)) {
1454 unlock_res_and_lock(lock);
1455 ldlm_lock_cancel(lock);
1456 lock_res_and_lock(lock);
1458 ldlm_add_waiting_lock(lock,
1459 ldlm_bl_timeout(lock));
1463 unlock_res_and_lock(lock);
1467 req->rq_status = rc ?: err; /* return either error - b=11190 */
1468 if (!req->rq_packed_final) {
1469 err = lustre_pack_reply(req, 1, NULL, NULL);
1475 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1476 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1480 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1484 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1486 ldlm_lvbo_size(lock) > 0) {
1491 buf = req_capsule_server_get(&req->rq_pill,
1493 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1494 buflen = req_capsule_get_size(&req->rq_pill,
1495 &RMF_DLM_LVB, RCL_SERVER);
1497 * non-replayed lock, delayed lvb init may
1498 * need to be occur now
1500 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1503 rc2 = ldlm_lvbo_fill(env, lock, buf, &buflen);
1505 req_capsule_shrink(&req->rq_pill,
1508 } else if (rc2 == -ERANGE) {
1509 rc2 = req_capsule_server_grow(
1511 &RMF_DLM_LVB, buflen);
1516 * if we can't grow the buffer,
1517 * it's ok to return empty lvb
1528 } else if (flags & LDLM_FL_REPLAY) {
1529 /* no LVB resend upon replay */
1531 req_capsule_shrink(&req->rq_pill,
1541 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1542 if (lock->l_export) {
1543 ldlm_lock_cancel(lock);
1545 lock_res_and_lock(lock);
1546 ldlm_resource_unlink_lock(lock);
1547 ldlm_lock_destroy_nolock(lock);
1548 unlock_res_and_lock(lock);
1553 if (!err && !ldlm_is_cbpending(lock) &&
1554 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1555 ldlm_reprocess_all(lock->l_resource);
1557 LDLM_LOCK_RELEASE(lock);
1560 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1567 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1568 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1570 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1572 if (lock->l_blocking_lock) {
1573 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1574 lock->l_blocking_lock = NULL;
1578 /* A lock can be converted to new ibits or mode and should be considered
1579 * as new lock. Clear all states related to a previous blocking AST
1580 * processing so new conflicts will cause new blocking ASTs.
1582 * This is used during lock convert below and lock downgrade to COS mode in
1583 * ldlm_lock_mode_downgrade().
1585 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1587 ldlm_clear_ast_sent(lock);
1588 lock->l_bl_ast_run = 0;
1589 ldlm_clear_blocking_lock(lock);
1593 * Main LDLM entry point for server code to process lock conversion requests.
1595 int ldlm_handle_convert0(struct ptlrpc_request *req,
1596 const struct ldlm_request *dlm_req)
1598 struct obd_export *exp = req->rq_export;
1599 struct ldlm_reply *dlm_rep;
1600 struct ldlm_lock *lock;
1605 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1606 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1607 LDLM_CONVERT - LDLM_FIRST_OPC);
1609 rc = req_capsule_server_pack(&req->rq_pill);
1613 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1614 dlm_rep->lock_flags = dlm_req->lock_flags;
1616 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1621 bits = lock->l_policy_data.l_inodebits.bits;
1622 new = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1623 LDLM_DEBUG(lock, "server-side convert handler START");
1625 if (ldlm_is_cancel(lock)) {
1626 LDLM_ERROR(lock, "convert on canceled lock!");
1627 rc = ELDLM_NO_LOCK_DATA;
1628 } else if (dlm_req->lock_desc.l_req_mode !=
1629 lock->l_granted_mode) {
1630 LDLM_ERROR(lock, "lock mode differs!");
1631 rc = ELDLM_NO_LOCK_DATA;
1632 } else if (bits == new) {
1634 * This can be valid situation if CONVERT RPCs are
1635 * re-ordered. Just finish silently
1637 LDLM_DEBUG(lock, "lock is converted already!");
1640 lock_res_and_lock(lock);
1641 if (ldlm_is_waited(lock))
1642 ldlm_del_waiting_lock(lock);
1644 ldlm_clear_cbpending(lock);
1645 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1646 ldlm_inodebits_drop(lock, bits & ~new);
1648 ldlm_clear_blocking_data(lock);
1649 unlock_res_and_lock(lock);
1651 ldlm_reprocess_all(lock->l_resource);
1655 if (rc == ELDLM_OK) {
1656 dlm_rep->lock_handle = lock->l_remote_handle;
1657 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1658 &dlm_rep->lock_desc.l_policy_data);
1661 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d",
1663 LDLM_LOCK_PUT(lock);
1665 rc = ELDLM_NO_LOCK_DATA;
1666 LDLM_DEBUG_NOLOCK("server-side convert handler END, rc = %d",
1670 req->rq_status = rc;
1676 * Cancel all the locks whose handles are packed into ldlm_request
1678 * Called by server code expecting such combined cancel activity
1681 int ldlm_request_cancel(struct ptlrpc_request *req,
1682 const struct ldlm_request *dlm_req,
1683 int first, enum lustre_at_flags flags)
1685 const struct lu_env *env = req->rq_svc_thread->t_env;
1686 struct ldlm_resource *res, *pres = NULL;
1687 struct ldlm_lock *lock;
1688 int i, count, done = 0;
1692 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1696 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1700 * There is no lock on the server at the replay time,
1701 * skip lock cancelling to make replay tests to pass.
1703 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1706 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1709 for (i = first; i < count; i++) {
1710 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1712 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1713 dlm_req->lock_handle[i].cookie);
1717 res = lock->l_resource;
1721 * This code is an optimization to only attempt lock
1722 * granting on the resource (that could be CPU-expensive)
1723 * after we are done cancelling lock in that resource.
1727 ldlm_reprocess_all(pres);
1728 LDLM_RESOURCE_DELREF(pres);
1729 ldlm_resource_putref(pres);
1732 ldlm_resource_getref(res);
1733 LDLM_RESOURCE_ADDREF(res);
1735 if (!ldlm_is_discard_data(lock))
1736 ldlm_lvbo_update(env, res, lock,
1742 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1743 lock->l_blast_sent != 0) {
1744 time64_t delay = ktime_get_real_seconds() -
1747 "server cancels blocked lock after %llds",
1749 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1751 ldlm_lock_cancel(lock);
1752 LDLM_LOCK_PUT(lock);
1755 ldlm_reprocess_all(pres);
1756 LDLM_RESOURCE_DELREF(pres);
1757 ldlm_resource_putref(pres);
1759 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1762 EXPORT_SYMBOL(ldlm_request_cancel);
1765 * Main LDLM entry point for server code to cancel locks.
1767 * Typically gets called from service handler on LDLM_CANCEL opc.
1769 int ldlm_handle_cancel(struct ptlrpc_request *req)
1771 struct ldlm_request *dlm_req;
1776 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1777 if (dlm_req == NULL) {
1778 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1782 if (req->rq_export && req->rq_export->exp_nid_stats &&
1783 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1784 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1785 LDLM_CANCEL - LDLM_FIRST_OPC);
1787 rc = req_capsule_server_pack(&req->rq_pill);
1791 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1792 req->rq_status = LUSTRE_ESTALE;
1794 RETURN(ptlrpc_reply(req));
1796 #endif /* HAVE_SERVER_SUPPORT */
1799 * Callback handler for receiving incoming blocking ASTs.
1801 * This can only happen on client side.
1803 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1804 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1810 LDLM_DEBUG(lock, "client blocking AST callback handler");
1812 lock_res_and_lock(lock);
1814 /* set bits to cancel for this lock for possible lock convert */
1815 if (ns_is_client(ns) && (lock->l_resource->lr_type == LDLM_IBITS)) {
1817 * Lock description contains policy of blocking lock,
1818 * and its cancel_bits is used to pass conflicting bits.
1819 * NOTE: ld can be NULL or can be not NULL but zeroed if
1820 * passed from ldlm_bl_thread_blwi(), check below used bits
1821 * in ld to make sure it is valid description.
1823 * If server may replace lock resource keeping the same cookie,
1824 * never use cancel bits from different resource, full cancel
1827 if (ld && ld->l_policy_data.l_inodebits.bits &&
1828 ldlm_res_eq(&ld->l_resource.lr_name,
1829 &lock->l_resource->lr_name))
1830 lock->l_policy_data.l_inodebits.cancel_bits =
1831 ld->l_policy_data.l_inodebits.cancel_bits;
1833 * if there is no valid ld and lock is cbpending already
1834 * then cancel_bits should be kept, otherwise it is zeroed.
1836 else if (!ldlm_is_cbpending(lock))
1837 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1839 ldlm_set_cbpending(lock);
1841 do_ast = (!lock->l_readers && !lock->l_writers);
1842 unlock_res_and_lock(lock);
1846 "Lock %p already unused, calling callback (%p)\n",
1847 lock, lock->l_blocking_ast);
1848 if (lock->l_blocking_ast != NULL)
1849 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1853 "Lock %p is referenced, will be cancelled later\n",
1857 LDLM_DEBUG(lock, "client blocking callback handler END");
1858 LDLM_LOCK_RELEASE(lock);
1863 * Callback handler for receiving incoming completion ASTs.
1865 * This only can happen on client side.
1867 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1868 struct ldlm_namespace *ns,
1869 struct ldlm_request *dlm_req,
1870 struct ldlm_lock *lock)
1872 struct list_head ast_list;
1878 LDLM_DEBUG(lock, "client completion callback handler START");
1880 INIT_LIST_HEAD(&ast_list);
1881 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1882 long to = cfs_time_seconds(1);
1885 set_current_state(TASK_INTERRUPTIBLE);
1886 schedule_timeout(to);
1887 if (ldlm_is_granted(lock) ||
1888 ldlm_is_destroyed(lock))
1893 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1895 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1896 GOTO(out, rc = lvb_len);
1897 } else if (lvb_len > 0) {
1898 if (lock->l_lvb_len > 0) {
1899 /* for extent lock, lvb contains ost_lvb{}. */
1900 LASSERT(lock->l_lvb_data != NULL);
1902 if (unlikely(lock->l_lvb_len < lvb_len)) {
1904 "Replied LVB is larger than expectation, expected = %d, replied = %d",
1905 lock->l_lvb_len, lvb_len);
1906 GOTO(out, rc = -EINVAL);
1911 lock_res_and_lock(lock);
1913 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1914 &lock->l_resource->lr_name)) {
1915 ldlm_resource_unlink_lock(lock);
1916 unlock_res_and_lock(lock);
1917 rc = ldlm_lock_change_resource(ns, lock,
1918 &dlm_req->lock_desc.l_resource.lr_name);
1920 LDLM_ERROR(lock, "Failed to allocate resource");
1923 LDLM_DEBUG(lock, "completion AST, new resource");
1924 lock_res_and_lock(lock);
1927 if (ldlm_is_destroyed(lock) ||
1928 ldlm_is_granted(lock)) {
1929 /* b=11300: the lock has already been granted */
1930 unlock_res_and_lock(lock);
1931 LDLM_DEBUG(lock, "Double grant race happened");
1936 * If we receive the completion AST before the actual enqueue returned,
1937 * then we might need to switch lock modes, resources, or extents.
1939 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1940 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1941 LDLM_DEBUG(lock, "completion AST, new lock mode");
1944 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1945 ldlm_convert_policy_to_local(req->rq_export,
1946 dlm_req->lock_desc.l_resource.lr_type,
1947 &dlm_req->lock_desc.l_policy_data,
1948 &lock->l_policy_data);
1949 LDLM_DEBUG(lock, "completion AST, new policy data");
1952 ldlm_resource_unlink_lock(lock);
1954 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1956 * BL_AST locks are not needed in LRU.
1957 * Let ldlm_cancel_lru() be fast.
1959 ldlm_lock_remove_from_lru(lock);
1960 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1961 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1964 if (lock->l_lvb_len > 0) {
1965 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1966 lock->l_lvb_data, lvb_len);
1968 unlock_res_and_lock(lock);
1973 ldlm_grant_lock(lock, &ast_list);
1974 unlock_res_and_lock(lock);
1976 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1979 * Let Enqueue to call osc_lock_upcall() and initialize
1982 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1984 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1986 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1992 lock_res_and_lock(lock);
1993 ldlm_set_failed(lock);
1994 unlock_res_and_lock(lock);
1995 wake_up(&lock->l_waitq);
1997 LDLM_LOCK_RELEASE(lock);
2001 * Callback handler for receiving incoming glimpse ASTs.
2003 * This only can happen on client side. After handling the glimpse AST
2004 * we also consider dropping the lock here if it is unused locally for a
2007 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2008 struct ldlm_namespace *ns,
2009 struct ldlm_request *dlm_req,
2010 struct ldlm_lock *lock)
2016 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2018 if (lock->l_glimpse_ast != NULL)
2019 rc = lock->l_glimpse_ast(lock, req);
2021 if (req->rq_repmsg != NULL) {
2024 req->rq_status = rc;
2028 lock_res_and_lock(lock);
2029 if (lock->l_granted_mode == LCK_PW &&
2030 !lock->l_readers && !lock->l_writers &&
2031 ktime_after(ktime_get(),
2032 ktime_add(lock->l_last_used,
2033 ktime_set(ns->ns_dirty_age_limit, 0)))) {
2034 unlock_res_and_lock(lock);
2035 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
2036 ldlm_handle_bl_callback(ns, NULL, lock);
2041 unlock_res_and_lock(lock);
2042 LDLM_LOCK_RELEASE(lock);
2046 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
2048 if (req->rq_no_reply)
2051 req->rq_status = rc;
2052 if (!req->rq_packed_final) {
2053 rc = lustre_pack_reply(req, 1, NULL, NULL);
2057 return ptlrpc_reply(req);
2060 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2061 enum ldlm_cancel_flags cancel_flags)
2063 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2067 spin_lock(&blp->blp_lock);
2068 if (blwi->blwi_lock &&
2069 ldlm_is_discard_data(blwi->blwi_lock)) {
2070 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2071 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2073 /* other blocking callbacks are added to the regular list */
2074 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2076 spin_unlock(&blp->blp_lock);
2078 wake_up(&blp->blp_waitq);
2081 * can not check blwi->blwi_flags as blwi could be already freed in
2084 if (!(cancel_flags & LCF_ASYNC))
2085 wait_for_completion(&blwi->blwi_comp);
2090 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2091 struct ldlm_namespace *ns,
2092 struct ldlm_lock_desc *ld,
2093 struct list_head *cancels, int count,
2094 struct ldlm_lock *lock,
2095 enum ldlm_cancel_flags cancel_flags)
2097 init_completion(&blwi->blwi_comp);
2098 INIT_LIST_HEAD(&blwi->blwi_head);
2100 if (memory_pressure_get())
2101 blwi->blwi_mem_pressure = 1;
2104 blwi->blwi_flags = cancel_flags;
2106 blwi->blwi_ld = *ld;
2108 list_add(&blwi->blwi_head, cancels);
2109 list_del_init(cancels);
2110 blwi->blwi_count = count;
2112 blwi->blwi_lock = lock;
2117 * Queues a list of locks \a cancels containing \a count locks
2118 * for later processing by a blocking thread. If \a count is zero,
2119 * then the lock referenced as \a lock is queued instead.
2121 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2122 * If list addition fails an error is returned and caller is supposed to
2123 * call ->l_blocking_ast itself.
2125 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2126 struct ldlm_lock_desc *ld,
2127 struct ldlm_lock *lock,
2128 struct list_head *cancels, int count,
2129 enum ldlm_cancel_flags cancel_flags)
2133 if (cancels && count == 0)
2136 if (cancel_flags & LCF_ASYNC) {
2137 struct ldlm_bl_work_item *blwi;
2139 OBD_ALLOC(blwi, sizeof(*blwi));
2142 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2144 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2147 * if it is synchronous call do minimum mem alloc, as it could
2148 * be triggered from kernel shrinker
2150 struct ldlm_bl_work_item blwi;
2152 memset(&blwi, 0, sizeof(blwi));
2153 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2154 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2159 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2160 struct ldlm_lock *lock)
2162 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2165 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2166 struct list_head *cancels, int count,
2167 enum ldlm_cancel_flags cancel_flags)
2169 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2172 int ldlm_bl_thread_wakeup(void)
2174 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2178 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2179 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2181 struct obd_device *obd = req->rq_export->exp_obd;
2189 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2191 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2193 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2195 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2198 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2200 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2202 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2205 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2208 /* We are responsible for swabbing contents of val */
2210 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2211 /* Pass it on to mdc (the "export" in this case) */
2212 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2214 sizeof(KEY_HSM_COPYTOOL_SEND),
2215 KEY_HSM_COPYTOOL_SEND,
2218 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2223 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2224 const char *msg, int rc,
2225 const struct lustre_handle *handle)
2227 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2228 "%s: [nid %s] [rc %d] [lock %#llx]",
2229 msg, libcfs_id2str(req->rq_peer), rc,
2230 handle ? handle->cookie : 0);
2231 if (req->rq_no_reply)
2232 CWARN("No reply was sent, maybe cause b=21636.\n");
2234 CWARN("Send reply failed, maybe cause b=21636.\n");
2237 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2238 static int ldlm_callback_handler(struct ptlrpc_request *req)
2240 struct ldlm_namespace *ns;
2241 struct ldlm_request *dlm_req;
2242 struct ldlm_lock *lock;
2248 * Requests arrive in sender's byte order. The ptlrpc service
2249 * handler has already checked and, if necessary, byte-swapped the
2250 * incoming request message body, but I am responsible for the
2254 /* do nothing for sec context finalize */
2255 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2258 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2260 if (req->rq_export == NULL) {
2261 rc = ldlm_callback_reply(req, -ENOTCONN);
2262 ldlm_callback_errmsg(req, "Operate on unconnected server",
2267 LASSERT(req->rq_export != NULL);
2268 LASSERT(req->rq_export->exp_obd != NULL);
2270 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2271 case LDLM_BL_CALLBACK:
2272 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2274 ldlm_callback_reply(req, -(int)cfs_fail_err);
2278 case LDLM_CP_CALLBACK:
2279 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2282 case LDLM_GL_CALLBACK:
2283 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2287 rc = ldlm_handle_setinfo(req);
2288 ldlm_callback_reply(req, rc);
2291 CERROR("unknown opcode %u\n",
2292 lustre_msg_get_opc(req->rq_reqmsg));
2293 ldlm_callback_reply(req, -EPROTO);
2297 ns = req->rq_export->exp_obd->obd_namespace;
2298 LASSERT(ns != NULL);
2300 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2302 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2303 if (dlm_req == NULL) {
2304 rc = ldlm_callback_reply(req, -EPROTO);
2305 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2311 * Force a known safe race, send a cancel to the server for a lock
2312 * which the server has already started a blocking callback on.
2314 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2315 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2316 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2318 CERROR("ldlm_cli_cancel: %d\n", rc);
2321 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2324 "callback on lock %#llx - lock disappeared\n",
2325 dlm_req->lock_handle[0].cookie);
2326 rc = ldlm_callback_reply(req, -EINVAL);
2327 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2328 &dlm_req->lock_handle[0]);
2332 if (ldlm_is_fail_loc(lock) &&
2333 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2334 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2336 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2337 lock_res_and_lock(lock);
2338 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2340 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2342 * If somebody cancels lock and cache is already dropped,
2343 * or lock is failed before cp_ast received on client,
2344 * we can tell the server we have no lock. Otherwise, we
2345 * should send cancel after dropping the cache.
2347 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2348 ldlm_is_failed(lock)) {
2350 "callback on lock %llx - lock disappeared",
2351 dlm_req->lock_handle[0].cookie);
2352 unlock_res_and_lock(lock);
2353 LDLM_LOCK_RELEASE(lock);
2354 rc = ldlm_callback_reply(req, -EINVAL);
2355 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2356 &dlm_req->lock_handle[0]);
2360 * BL_AST locks are not needed in LRU.
2361 * Let ldlm_cancel_lru() be fast.
2363 ldlm_lock_remove_from_lru(lock);
2364 ldlm_set_bl_ast(lock);
2366 unlock_res_and_lock(lock);
2369 * We want the ost thread to get this reply so that it can respond
2370 * to ost requests (write cache writeback) that might be triggered
2373 * But we'd also like to be able to indicate in the reply that we're
2374 * cancelling right now, because it's unused, or have an intent result
2375 * in the reply, so we might have to push the responsibility for sending
2376 * the reply down into the AST handlers, alas.
2379 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2380 case LDLM_BL_CALLBACK:
2381 CDEBUG(D_INODE, "blocking ast\n");
2382 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2383 if (!ldlm_is_cancel_on_block(lock)) {
2384 rc = ldlm_callback_reply(req, 0);
2385 if (req->rq_no_reply || rc)
2386 ldlm_callback_errmsg(req, "Normal process", rc,
2387 &dlm_req->lock_handle[0]);
2389 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2390 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2392 case LDLM_CP_CALLBACK:
2393 CDEBUG(D_INODE, "completion ast\n");
2394 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2395 ldlm_callback_reply(req, 0);
2396 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2398 case LDLM_GL_CALLBACK:
2399 CDEBUG(D_INODE, "glimpse ast\n");
2400 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2401 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2404 LBUG(); /* checked above */
2410 #ifdef HAVE_SERVER_SUPPORT
2412 * Main handler for canceld thread.
2414 * Separated into its own thread to avoid deadlocks.
2416 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2423 * Requests arrive in sender's byte order. The ptlrpc service
2424 * handler has already checked and, if necessary, byte-swapped the
2425 * incoming request message body, but I am responsible for the
2429 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2431 if (req->rq_export == NULL) {
2432 struct ldlm_request *dlm_req;
2434 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2435 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2436 libcfs_nid2str(req->rq_peer.nid),
2437 (unsigned long long)req->rq_arrival_time.tv_sec,
2438 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2440 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2441 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2442 dlm_req = req_capsule_client_get(&req->rq_pill,
2444 if (dlm_req != NULL)
2445 ldlm_lock_dump_handle(D_ERROR,
2446 &dlm_req->lock_handle[0]);
2448 ldlm_callback_reply(req, -ENOTCONN);
2452 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2453 /* XXX FIXME move this back to mds/handler.c, b=249 */
2455 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2456 CDEBUG(D_INODE, "cancel\n");
2457 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2458 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2459 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2461 rc = ldlm_handle_cancel(req);
2465 struct ldlm_request *dlm_req;
2467 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2468 CDEBUG(D_INODE, "convert\n");
2470 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2471 if (dlm_req == NULL) {
2472 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2473 rc = ldlm_callback_reply(req, -EPROTO);
2475 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2476 rc = ptlrpc_reply(req);
2481 CERROR("invalid opcode %d\n",
2482 lustre_msg_get_opc(req->rq_reqmsg));
2483 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2484 rc = ldlm_callback_reply(req, -EINVAL);
2490 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2491 struct ldlm_lock *lock)
2493 struct ldlm_request *dlm_req;
2494 struct lustre_handle lockh;
2500 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2501 if (dlm_req == NULL)
2504 ldlm_lock2handle(lock, &lockh);
2505 for (i = 0; i < dlm_req->lock_count; i++) {
2506 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2508 DEBUG_REQ(D_RPCTRACE, req,
2509 "Prio raised by lock %#llx.", lockh.cookie);
2518 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2520 struct ldlm_request *dlm_req;
2526 /* no prolong in recovery */
2527 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2530 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2531 if (dlm_req == NULL)
2534 for (i = 0; i < dlm_req->lock_count; i++) {
2535 struct ldlm_lock *lock;
2537 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2541 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2543 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2544 LDLM_LOCK_PUT(lock);
2553 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2554 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2555 .hpreq_check = ldlm_cancel_hpreq_check,
2559 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2563 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2565 if (req->rq_export == NULL)
2568 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2569 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2570 req->rq_ops = &ldlm_cancel_hpreq_ops;
2571 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2572 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2573 req->rq_ops = &ldlm_cancel_hpreq_ops;
2578 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2579 struct hlist_node *hnode, void *data)
2582 struct list_head *rpc_list = data;
2583 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2585 lock_res_and_lock(lock);
2587 if (!ldlm_is_granted(lock)) {
2588 unlock_res_and_lock(lock);
2592 LASSERT(lock->l_resource);
2593 if (lock->l_resource->lr_type != LDLM_IBITS &&
2594 lock->l_resource->lr_type != LDLM_PLAIN) {
2595 unlock_res_and_lock(lock);
2599 if (ldlm_is_ast_sent(lock)) {
2600 unlock_res_and_lock(lock);
2604 LASSERT(lock->l_blocking_ast);
2605 LASSERT(!lock->l_blocking_lock);
2607 ldlm_set_ast_sent(lock);
2608 if (lock->l_export && lock->l_export->exp_lock_hash) {
2610 * NB: it's safe to call cfs_hash_del() even lock isn't
2614 * In the function below, .hs_keycmp resolves to
2615 * ldlm_export_lock_keycmp()
2617 /* coverity[overrun-buffer-val] */
2618 cfs_hash_del(lock->l_export->exp_lock_hash,
2619 &lock->l_remote_handle, &lock->l_exp_hash);
2622 list_add_tail(&lock->l_rk_ast, rpc_list);
2623 LDLM_LOCK_GET(lock);
2625 unlock_res_and_lock(lock);
2629 void ldlm_revoke_export_locks(struct obd_export *exp)
2631 struct list_head rpc_list;
2635 INIT_LIST_HEAD(&rpc_list);
2636 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2637 ldlm_revoke_lock_cb, &rpc_list, 0);
2638 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2639 LDLM_WORK_REVOKE_AST);
2643 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2644 #endif /* HAVE_SERVER_SUPPORT */
2646 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2647 struct ldlm_bl_work_item **p_blwi,
2648 struct obd_export **p_exp)
2650 struct ldlm_bl_work_item *blwi = NULL;
2651 static unsigned int num_bl;
2652 static unsigned int num_stale;
2653 int num_th = atomic_read(&blp->blp_num_threads);
2655 *p_exp = obd_stale_export_get();
2657 spin_lock(&blp->blp_lock);
2658 if (*p_exp != NULL) {
2659 if (num_th == 1 || ++num_stale < num_th) {
2660 spin_unlock(&blp->blp_lock);
2666 /* process a request from the blp_list at least every blp_num_threads */
2667 if (!list_empty(&blp->blp_list) &&
2668 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2669 blwi = list_entry(blp->blp_list.next,
2670 struct ldlm_bl_work_item, blwi_entry);
2672 if (!list_empty(&blp->blp_prio_list))
2673 blwi = list_entry(blp->blp_prio_list.next,
2674 struct ldlm_bl_work_item,
2678 if (++num_bl >= num_th)
2680 list_del(&blwi->blwi_entry);
2682 spin_unlock(&blp->blp_lock);
2685 if (*p_exp != NULL && *p_blwi != NULL) {
2686 obd_stale_export_put(*p_exp);
2690 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2693 /* This only contains temporary data until the thread starts */
2694 struct ldlm_bl_thread_data {
2695 struct ldlm_bl_pool *bltd_blp;
2696 struct completion bltd_comp;
2700 static int ldlm_bl_thread_main(void *arg);
2702 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2704 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2705 struct task_struct *task;
2707 init_completion(&bltd.bltd_comp);
2709 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2710 if (bltd.bltd_num >= blp->blp_max_threads) {
2711 atomic_dec(&blp->blp_num_threads);
2715 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2717 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2718 atomic_dec(&blp->blp_num_threads);
2722 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2725 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2726 bltd.bltd_num, PTR_ERR(task));
2727 atomic_dec(&blp->blp_num_threads);
2728 return PTR_ERR(task);
2730 wait_for_completion(&bltd.bltd_comp);
2735 /* Not fatal if racy and have a few too many threads */
2736 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2737 struct ldlm_bl_work_item *blwi)
2739 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2742 if (atomic_read(&blp->blp_busy_threads) <
2743 atomic_read(&blp->blp_num_threads))
2746 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2747 blwi->blwi_mem_pressure))
2753 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2754 struct ldlm_bl_work_item *blwi)
2758 if (blwi->blwi_ns == NULL)
2759 /* added by ldlm_cleanup() */
2760 RETURN(LDLM_ITER_STOP);
2762 if (blwi->blwi_mem_pressure)
2763 memory_pressure_set();
2765 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2767 if (blwi->blwi_count) {
2770 * The special case when we cancel locks in lru
2771 * asynchronously, we pass the list of locks here.
2772 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2773 * canceled locally yet.
2775 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2778 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2781 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2784 if (blwi->blwi_mem_pressure)
2785 memory_pressure_clr();
2787 if (blwi->blwi_flags & LCF_ASYNC)
2788 OBD_FREE(blwi, sizeof(*blwi));
2790 complete(&blwi->blwi_comp);
2796 * Cancel stale locks on export. Cancel blocked locks first.
2797 * If the given export has blocked locks, the next in the list may have
2798 * them too, thus cancel not blocked locks only if the current export has
2801 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2802 struct obd_export *exp)
2808 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2810 num = ldlm_export_cancel_blocked_locks(exp);
2812 ldlm_export_cancel_locks(exp);
2814 obd_stale_export_put(exp);
2821 * Main blocking requests processing thread.
2823 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2824 * This thread in the end ends up doing actual call to ->l_blocking_ast
2827 static int ldlm_bl_thread_main(void *arg)
2829 struct ldlm_bl_pool *blp;
2830 struct ldlm_bl_thread_data *bltd = arg;
2834 blp = bltd->bltd_blp;
2836 complete(&bltd->bltd_comp);
2837 /* cannot use bltd after this, it is only on caller's stack */
2840 struct l_wait_info lwi = { 0 };
2841 struct ldlm_bl_work_item *blwi = NULL;
2842 struct obd_export *exp = NULL;
2845 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2848 l_wait_event_exclusive(blp->blp_waitq,
2849 ldlm_bl_get_work(blp, &blwi,
2852 atomic_inc(&blp->blp_busy_threads);
2854 if (ldlm_bl_thread_need_create(blp, blwi))
2855 /* discard the return value, we tried */
2856 ldlm_bl_thread_start(blp, true);
2859 rc = ldlm_bl_thread_exports(blp, exp);
2861 rc = ldlm_bl_thread_blwi(blp, blwi);
2863 atomic_dec(&blp->blp_busy_threads);
2865 if (rc == LDLM_ITER_STOP)
2869 * If there are many namespaces, we will not sleep waiting for
2870 * work, and must do a cond_resched to avoid holding the CPU
2876 atomic_dec(&blp->blp_num_threads);
2877 complete(&blp->blp_comp);
2882 static int ldlm_setup(void);
2883 static int ldlm_cleanup(void);
2885 int ldlm_get_ref(void)
2890 mutex_lock(&ldlm_ref_mutex);
2891 if (++ldlm_refcount == 1) {
2896 mutex_unlock(&ldlm_ref_mutex);
2901 void ldlm_put_ref(void)
2904 mutex_lock(&ldlm_ref_mutex);
2905 if (ldlm_refcount == 1) {
2906 int rc = ldlm_cleanup();
2909 CERROR("ldlm_cleanup failed: %d\n", rc);
2915 mutex_unlock(&ldlm_ref_mutex);
2921 * Export handle<->lock hash operations.
2924 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
2926 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2930 ldlm_export_lock_key(struct hlist_node *hnode)
2932 struct ldlm_lock *lock;
2934 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2935 return &lock->l_remote_handle;
2939 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2941 struct ldlm_lock *lock;
2943 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2944 lock->l_remote_handle = *(struct lustre_handle *)key;
2948 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2950 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2954 ldlm_export_lock_object(struct hlist_node *hnode)
2956 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2960 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2962 struct ldlm_lock *lock;
2964 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2965 LDLM_LOCK_GET(lock);
2969 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2971 struct ldlm_lock *lock;
2973 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2974 LDLM_LOCK_RELEASE(lock);
2977 static struct cfs_hash_ops ldlm_export_lock_ops = {
2978 .hs_hash = ldlm_export_lock_hash,
2979 .hs_key = ldlm_export_lock_key,
2980 .hs_keycmp = ldlm_export_lock_keycmp,
2981 .hs_keycpy = ldlm_export_lock_keycpy,
2982 .hs_object = ldlm_export_lock_object,
2983 .hs_get = ldlm_export_lock_get,
2984 .hs_put = ldlm_export_lock_put,
2985 .hs_put_locked = ldlm_export_lock_put,
2988 int ldlm_init_export(struct obd_export *exp)
2994 exp->exp_lock_hash =
2995 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2996 HASH_EXP_LOCK_CUR_BITS,
2997 HASH_EXP_LOCK_MAX_BITS,
2998 HASH_EXP_LOCK_BKT_BITS, 0,
2999 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3000 &ldlm_export_lock_ops,
3001 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3002 CFS_HASH_NBLK_CHANGE);
3004 if (!exp->exp_lock_hash)
3007 rc = ldlm_init_flock_export(exp);
3013 ldlm_destroy_export(exp);
3016 EXPORT_SYMBOL(ldlm_init_export);
3018 void ldlm_destroy_export(struct obd_export *exp)
3021 cfs_hash_putref(exp->exp_lock_hash);
3022 exp->exp_lock_hash = NULL;
3024 ldlm_destroy_flock_export(exp);
3027 EXPORT_SYMBOL(ldlm_destroy_export);
3029 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3030 struct attribute *attr,
3033 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3036 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3037 struct attribute *attr,
3044 rc = kstrtoul(buffer, 10, &val);
3048 ldlm_cancel_unused_locks_before_replay = val;
3052 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3054 static struct attribute *ldlm_attrs[] = {
3055 &lustre_attr_cancel_unused_locks_before_replay.attr,
3059 static struct attribute_group ldlm_attr_group = {
3060 .attrs = ldlm_attrs,
3063 static int ldlm_setup(void)
3065 static struct ptlrpc_service_conf conf;
3066 struct ldlm_bl_pool *blp = NULL;
3067 #ifdef HAVE_SERVER_SUPPORT
3068 struct task_struct *task;
3069 #endif /* HAVE_SERVER_SUPPORT */
3075 if (ldlm_state != NULL)
3078 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3079 if (ldlm_state == NULL)
3082 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3086 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3090 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3094 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3098 rc = ldlm_debugfs_setup();
3102 memset(&conf, 0, sizeof(conf));
3103 conf = (typeof(conf)) {
3104 .psc_name = "ldlm_cbd",
3105 .psc_watchdog_factor = 2,
3107 .bc_nbufs = LDLM_CLIENT_NBUFS,
3108 .bc_buf_size = LDLM_BUFSIZE,
3109 .bc_req_max_size = LDLM_MAXREQSIZE,
3110 .bc_rep_max_size = LDLM_MAXREPSIZE,
3111 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3112 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3115 .tc_thr_name = "ldlm_cb",
3116 .tc_thr_factor = LDLM_THR_FACTOR,
3117 .tc_nthrs_init = LDLM_NTHRS_INIT,
3118 .tc_nthrs_base = LDLM_NTHRS_BASE,
3119 .tc_nthrs_max = LDLM_NTHRS_MAX,
3120 .tc_nthrs_user = ldlm_num_threads,
3121 .tc_cpu_bind = ldlm_cpu_bind,
3122 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3125 .cc_pattern = ldlm_cpts,
3126 .cc_affinity = true,
3129 .so_req_handler = ldlm_callback_handler,
3132 ldlm_state->ldlm_cb_service = \
3133 ptlrpc_register_service(&conf, ldlm_svc_kset,
3134 ldlm_svc_debugfs_dir);
3135 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3136 CERROR("failed to start service\n");
3137 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3138 ldlm_state->ldlm_cb_service = NULL;
3142 #ifdef HAVE_SERVER_SUPPORT
3143 memset(&conf, 0, sizeof(conf));
3144 conf = (typeof(conf)) {
3145 .psc_name = "ldlm_canceld",
3146 .psc_watchdog_factor = 6,
3148 .bc_nbufs = LDLM_SERVER_NBUFS,
3149 .bc_buf_size = LDLM_BUFSIZE,
3150 .bc_req_max_size = LDLM_MAXREQSIZE,
3151 .bc_rep_max_size = LDLM_MAXREPSIZE,
3152 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3153 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3157 .tc_thr_name = "ldlm_cn",
3158 .tc_thr_factor = LDLM_THR_FACTOR,
3159 .tc_nthrs_init = LDLM_NTHRS_INIT,
3160 .tc_nthrs_base = LDLM_NTHRS_BASE,
3161 .tc_nthrs_max = LDLM_NTHRS_MAX,
3162 .tc_nthrs_user = ldlm_num_threads,
3163 .tc_cpu_bind = ldlm_cpu_bind,
3164 .tc_ctx_tags = LCT_MD_THREAD | \
3169 .cc_pattern = ldlm_cpts,
3170 .cc_affinity = true,
3173 .so_req_handler = ldlm_cancel_handler,
3174 .so_hpreq_handler = ldlm_hpreq_handler,
3177 ldlm_state->ldlm_cancel_service = \
3178 ptlrpc_register_service(&conf, ldlm_svc_kset,
3179 ldlm_svc_debugfs_dir);
3180 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3181 CERROR("failed to start service\n");
3182 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3183 ldlm_state->ldlm_cancel_service = NULL;
3186 #endif /* HAVE_SERVER_SUPPORT */
3188 OBD_ALLOC(blp, sizeof(*blp));
3190 GOTO(out, rc = -ENOMEM);
3191 ldlm_state->ldlm_bl_pool = blp;
3193 spin_lock_init(&blp->blp_lock);
3194 INIT_LIST_HEAD(&blp->blp_list);
3195 INIT_LIST_HEAD(&blp->blp_prio_list);
3196 init_waitqueue_head(&blp->blp_waitq);
3197 atomic_set(&blp->blp_num_threads, 0);
3198 atomic_set(&blp->blp_busy_threads, 0);
3200 if (ldlm_num_threads == 0) {
3201 blp->blp_min_threads = LDLM_NTHRS_INIT;
3202 blp->blp_max_threads = LDLM_NTHRS_MAX;
3204 blp->blp_min_threads = blp->blp_max_threads = \
3205 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3209 for (i = 0; i < blp->blp_min_threads; i++) {
3210 rc = ldlm_bl_thread_start(blp, false);
3215 #ifdef HAVE_SERVER_SUPPORT
3216 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3219 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3223 wait_event(expired_lock_wait_queue,
3224 expired_lock_thread_state == ELT_READY);
3225 #endif /* HAVE_SERVER_SUPPORT */
3227 rc = ldlm_pools_init();
3229 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3233 rc = ldlm_reclaim_setup();
3235 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3245 static int ldlm_cleanup(void)
3249 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3250 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3251 CERROR("ldlm still has namespaces; clean these up first.\n");
3252 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3253 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3257 ldlm_reclaim_cleanup();
3260 if (ldlm_state->ldlm_bl_pool != NULL) {
3261 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3263 while (atomic_read(&blp->blp_num_threads) > 0) {
3264 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3266 init_completion(&blp->blp_comp);
3268 spin_lock(&blp->blp_lock);
3269 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3270 wake_up(&blp->blp_waitq);
3271 spin_unlock(&blp->blp_lock);
3273 wait_for_completion(&blp->blp_comp);
3276 OBD_FREE(blp, sizeof(*blp));
3279 if (ldlm_state->ldlm_cb_service != NULL)
3280 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3281 #ifdef HAVE_SERVER_SUPPORT
3282 if (ldlm_state->ldlm_cancel_service != NULL)
3283 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3287 kset_unregister(ldlm_ns_kset);
3289 kset_unregister(ldlm_svc_kset);
3291 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3292 kobject_put(ldlm_kobj);
3295 ldlm_debugfs_cleanup();
3297 #ifdef HAVE_SERVER_SUPPORT
3298 if (expired_lock_thread_state != ELT_STOPPED) {
3299 expired_lock_thread_state = ELT_TERMINATE;
3300 wake_up(&expired_lock_wait_queue);
3301 wait_event(expired_lock_wait_queue,
3302 expired_lock_thread_state == ELT_STOPPED);
3306 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3314 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3315 sizeof(struct ldlm_resource), 0,
3316 SLAB_HWCACHE_ALIGN, NULL);
3317 if (ldlm_resource_slab == NULL)
3320 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3321 sizeof(struct ldlm_lock), 0,
3322 SLAB_HWCACHE_ALIGN, NULL);
3323 if (ldlm_lock_slab == NULL)
3326 ldlm_interval_slab = kmem_cache_create("interval_node",
3327 sizeof(struct ldlm_interval),
3328 0, SLAB_HWCACHE_ALIGN, NULL);
3329 if (ldlm_interval_slab == NULL)
3332 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3333 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3334 0, SLAB_HWCACHE_ALIGN, NULL);
3335 if (ldlm_interval_tree_slab == NULL)
3338 #ifdef HAVE_SERVER_SUPPORT
3339 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3340 sizeof(struct ldlm_glimpse_work),
3342 if (ldlm_glimpse_work_kmem == NULL)
3343 goto out_interval_tree;
3346 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3347 class_export_dump_hook = ldlm_dump_export_locks;
3350 #ifdef HAVE_SERVER_SUPPORT
3352 kmem_cache_destroy(ldlm_interval_tree_slab);
3355 kmem_cache_destroy(ldlm_interval_slab);
3357 kmem_cache_destroy(ldlm_lock_slab);
3359 kmem_cache_destroy(ldlm_resource_slab);
3364 void ldlm_exit(void)
3367 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3368 kmem_cache_destroy(ldlm_resource_slab);
3370 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3371 * synchronize_rcu() to wait a grace period elapsed, so that
3372 * ldlm_lock_free() get a chance to be called.
3375 kmem_cache_destroy(ldlm_lock_slab);
3376 kmem_cache_destroy(ldlm_interval_slab);
3377 kmem_cache_destroy(ldlm_interval_tree_slab);
3378 #ifdef HAVE_SERVER_SUPPORT
3379 kmem_cache_destroy(ldlm_glimpse_work_kmem);