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(struct ldlm_lock *lock,
667 struct ptlrpc_request *req, int rc,
668 const char *ast_type)
670 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
672 if (!req->rq_replied || (rc && rc != -EINVAL)) {
673 if (ldlm_is_cancel(lock)) {
675 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
676 ast_type, req, req->rq_xid,
677 libcfs_nid2str(peer.nid));
678 ldlm_lock_cancel(lock);
680 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
682 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
684 * Upon umount process the AST fails because cannot be
685 * sent. This shouldn't lead to the client eviction.
686 * -ENODEV error is returned by ptl_send_rpc() for
687 * new request in such import.
688 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
689 * if imp_invalid is set or obd_no_recov.
690 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
691 * in ptlrpc_import_delay_req() as well with -EIO code.
692 * In all such cases errors are ignored.
695 "%s AST can't be sent due to a server %s failure or umount process: rc = %d\n",
697 req->rq_import->imp_obd->obd_name, rc);
700 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
701 libcfs_nid2str(peer.nid),
702 req->rq_replied ? "returned error from" :
703 "failed to reply to",
704 ast_type, req, req->rq_xid,
705 (req->rq_repmsg != NULL) ?
706 lustre_msg_get_status(req->rq_repmsg) : 0,
708 ldlm_failed_ast(lock, rc, ast_type);
714 struct ldlm_resource *res = lock->l_resource;
717 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
718 libcfs_nid2str(peer.nid),
720 lustre_msg_get_status(req->rq_repmsg) : -1,
721 ast_type, req, req->rq_xid);
724 * update lvbo to return proper attributes.
727 ldlm_resource_getref(res);
728 ldlm_lvbo_update(res, lock, NULL, 1);
729 ldlm_resource_putref(res);
731 ldlm_lock_cancel(lock);
738 static int ldlm_cb_interpret(const struct lu_env *env,
739 struct ptlrpc_request *req, void *args, int rc)
741 struct ldlm_cb_async_args *ca = args;
742 struct ldlm_lock *lock = ca->ca_lock;
743 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
747 LASSERT(lock != NULL);
750 case LDLM_GL_CALLBACK:
752 * Update the LVB from disk if the AST failed
753 * (this is a legal race)
755 * - Glimpse callback of local lock just returns
756 * -ELDLM_NO_LOCK_DATA.
757 * - Glimpse callback of remote lock might return
758 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
760 if (unlikely(arg->gl_interpret_reply)) {
761 rc = arg->gl_interpret_reply(NULL, req, args, rc);
762 } else if (rc == -ELDLM_NO_LOCK_DATA) {
764 "lost race - client has a lock but no inode");
765 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
766 } else if (rc != 0) {
767 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
769 rc = ldlm_lvbo_update(lock->l_resource,
773 case LDLM_BL_CALLBACK:
775 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
777 case LDLM_CP_CALLBACK:
779 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
782 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
787 /* release extra reference taken in ldlm_ast_fini() */
788 LDLM_LOCK_RELEASE(lock);
791 atomic_inc(&arg->restart);
796 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
798 struct ldlm_cb_async_args *ca = data;
799 struct ldlm_lock *lock = ca->ca_lock;
801 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
804 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
805 struct ldlm_cb_set_arg *arg,
806 struct ldlm_lock *lock,
813 if (unlikely(instant_cancel)) {
814 rc = ptl_send_rpc(req, 1);
815 ptlrpc_req_finished(req);
817 atomic_inc(&arg->restart);
820 ptlrpc_set_add_req(arg->set, req);
827 * Check if there are requests in the export request list which prevent
828 * the lock canceling and make these requests high priority ones.
830 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
832 struct ptlrpc_request *req;
836 if (lock->l_export == NULL) {
837 LDLM_DEBUG(lock, "client lock: no-op");
841 spin_lock(&lock->l_export->exp_rpc_lock);
842 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
845 * Do not process requests that were not yet added to there
846 * incoming queue or were already removed from there for
847 * processing. We evaluate ptlrpc_nrs_req_can_move() without
848 * holding svcpt->scp_req_lock, and then redo the check with
849 * the lock held once we need to obtain a reliable result.
851 if (ptlrpc_nrs_req_can_move(req) &&
852 req->rq_ops->hpreq_lock_match &&
853 req->rq_ops->hpreq_lock_match(req, lock))
854 ptlrpc_nrs_req_hp_move(req);
856 spin_unlock(&lock->l_export->exp_rpc_lock);
861 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
862 * enqueued server lock conflicts with given one.
864 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
865 * to wait for client response.
867 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
868 struct ldlm_lock_desc *desc,
869 void *data, int flag)
871 struct ldlm_cb_async_args *ca;
872 struct ldlm_cb_set_arg *arg = data;
873 struct ldlm_request *body;
874 struct ptlrpc_request *req;
875 int instant_cancel = 0;
880 if (flag == LDLM_CB_CANCELING)
881 /* Don't need to do anything here. */
884 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
885 LDLM_DEBUG(lock, "dropping BL AST");
890 LASSERT(data != NULL);
891 if (lock->l_export->exp_obd->obd_recovering != 0)
892 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
894 ldlm_lock_reorder_req(lock);
896 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
897 &RQF_LDLM_BL_CALLBACK,
898 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
902 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
903 ca = ptlrpc_req_async_args(req);
904 ca->ca_set_arg = arg;
907 req->rq_interpret_reply = ldlm_cb_interpret;
909 lock_res_and_lock(lock);
910 if (ldlm_is_destroyed(lock)) {
911 /* What's the point? */
912 unlock_res_and_lock(lock);
913 ptlrpc_req_finished(req);
917 if (!ldlm_is_granted(lock)) {
919 * this blocking AST will be communicated as part of the
920 * completion AST instead
922 ldlm_add_blocked_lock(lock);
923 ldlm_set_waited(lock);
924 unlock_res_and_lock(lock);
926 ptlrpc_req_finished(req);
927 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
931 if (ldlm_is_cancel_on_block(lock))
934 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
935 body->lock_handle[0] = lock->l_remote_handle;
936 body->lock_desc = *desc;
937 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
939 LDLM_DEBUG(lock, "server preparing blocking AST");
941 ptlrpc_request_set_replen(req);
942 ldlm_set_cbpending(lock);
943 if (instant_cancel) {
944 unlock_res_and_lock(lock);
945 ldlm_lock_cancel(lock);
947 req->rq_no_resend = 1;
949 LASSERT(ldlm_is_granted(lock));
950 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
951 unlock_res_and_lock(lock);
953 /* Do not resend after lock callback timeout */
954 req->rq_delay_limit = ldlm_bl_timeout(lock);
955 req->rq_resend_cb = ldlm_update_resend;
958 req->rq_send_state = LUSTRE_IMP_FULL;
959 /* ptlrpc_request_alloc_pack already set timeout */
961 req->rq_timeout = ldlm_get_rq_timeout();
963 if (lock->l_export && lock->l_export->exp_nid_stats &&
964 lock->l_export->exp_nid_stats->nid_ldlm_stats)
965 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
966 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
968 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
974 * ->l_completion_ast callback for a remote lock in server namespace.
976 * Sends AST to the client notifying it of lock granting. If initial
977 * lock response was not sent yet, instead of sending another RPC, just
978 * mark the lock as granted and client will understand
980 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
982 struct ldlm_cb_set_arg *arg = data;
983 struct ldlm_request *body;
984 struct ptlrpc_request *req;
985 struct ldlm_cb_async_args *ca;
986 int instant_cancel = 0;
992 LASSERT(lock != NULL);
993 LASSERT(data != NULL);
995 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
996 LDLM_DEBUG(lock, "dropping CP AST");
1000 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1001 &RQF_LDLM_CP_CALLBACK);
1005 /* server namespace, doesn't need lock */
1006 lvb_len = ldlm_lvbo_size(lock);
1008 * LU-3124 & LU-2187: to not return layout in completion AST because
1009 * it may deadlock for LU-2187, or client may not have enough space
1010 * for large layout. The layout will be returned to client with an
1011 * extra RPC to fetch xattr.lov
1013 if (ldlm_has_layout(lock))
1016 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1017 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1019 ptlrpc_request_free(req);
1023 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1024 ca = ptlrpc_req_async_args(req);
1025 ca->ca_set_arg = arg;
1028 req->rq_interpret_reply = ldlm_cb_interpret;
1029 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1031 body->lock_handle[0] = lock->l_remote_handle;
1032 body->lock_flags = ldlm_flags_to_wire(flags);
1033 ldlm_lock2desc(lock, &body->lock_desc);
1035 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1036 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1039 * We still need to send the RPC to wake up the blocked
1040 * enqueue thread on the client.
1042 * Consider old client, there is no better way to notify
1043 * the failure, just zero-sized the LVB, then the client
1044 * will fail out as "-EPROTO".
1046 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1050 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1055 LDLM_DEBUG(lock, "server preparing completion AST");
1057 ptlrpc_request_set_replen(req);
1059 req->rq_send_state = LUSTRE_IMP_FULL;
1060 /* ptlrpc_request_pack already set timeout */
1062 req->rq_timeout = ldlm_get_rq_timeout();
1064 /* We only send real blocking ASTs after the lock is granted */
1065 lock_res_and_lock(lock);
1066 if (ldlm_is_ast_sent(lock)) {
1067 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1068 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1069 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1073 * We might get here prior to ldlm_handle_enqueue setting
1074 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1075 * into waiting list, but this is safe and similar code in
1076 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1077 * that would not only cancel the lock, but will also remove
1078 * it from waiting list
1080 if (ldlm_is_cancel_on_block(lock)) {
1081 unlock_res_and_lock(lock);
1082 ldlm_lock_cancel(lock);
1085 req->rq_no_resend = 1;
1087 lock_res_and_lock(lock);
1089 /* start the lock-timeout clock */
1090 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1091 /* Do not resend after lock callback timeout */
1092 req->rq_delay_limit = ldlm_bl_timeout(lock);
1093 req->rq_resend_cb = ldlm_update_resend;
1096 unlock_res_and_lock(lock);
1098 if (lock->l_export && lock->l_export->exp_nid_stats &&
1099 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1100 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1101 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1103 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1105 RETURN(lvb_len < 0 ? lvb_len : rc);
1109 * Server side ->l_glimpse_ast handler for client locks.
1111 * Sends glimpse AST to the client and waits for reply. Then updates
1112 * lvbo with the result.
1114 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1116 struct ldlm_cb_set_arg *arg = data;
1117 struct ldlm_request *body;
1118 struct ptlrpc_request *req;
1119 struct ldlm_cb_async_args *ca;
1121 struct req_format *req_fmt;
1125 LASSERT(lock != NULL);
1127 if (arg->gl_desc != NULL)
1128 /* There is a glimpse descriptor to pack */
1129 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1131 req_fmt = &RQF_LDLM_GL_CALLBACK;
1133 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1134 req_fmt, LUSTRE_DLM_VERSION,
1140 if (arg->gl_desc != NULL) {
1141 /* copy the GL descriptor */
1142 union ldlm_gl_desc *desc;
1144 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1145 *desc = *arg->gl_desc;
1148 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1149 body->lock_handle[0] = lock->l_remote_handle;
1150 ldlm_lock2desc(lock, &body->lock_desc);
1152 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1153 ca = ptlrpc_req_async_args(req);
1154 ca->ca_set_arg = arg;
1157 /* server namespace, doesn't need lock */
1158 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1159 ldlm_lvbo_size(lock));
1160 ptlrpc_request_set_replen(req);
1162 req->rq_send_state = LUSTRE_IMP_FULL;
1163 /* ptlrpc_request_alloc_pack already set timeout */
1165 req->rq_timeout = ldlm_get_rq_timeout();
1167 req->rq_interpret_reply = ldlm_cb_interpret;
1169 if (lock->l_export && lock->l_export->exp_nid_stats) {
1170 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1172 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1173 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1176 rc = ldlm_ast_fini(req, arg, lock, 0);
1180 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1182 int ldlm_glimpse_locks(struct ldlm_resource *res,
1183 struct list_head *gl_work_list)
1189 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1191 if (rc == -ERESTART)
1192 ldlm_reprocess_all(res);
1196 EXPORT_SYMBOL(ldlm_glimpse_locks);
1198 /* return LDLM lock associated with a lock callback request */
1199 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1201 struct ldlm_cb_async_args *ca;
1202 struct ldlm_lock *lock;
1206 ca = ptlrpc_req_async_args(req);
1209 RETURN(ERR_PTR(-EFAULT));
1213 EXPORT_SYMBOL(ldlm_request_lock);
1216 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1217 * service threads to carry out client lock enqueueing requests.
1219 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1220 struct ptlrpc_request *req,
1221 const struct ldlm_request *dlm_req,
1222 const struct ldlm_callback_suite *cbs)
1224 struct ldlm_reply *dlm_rep;
1226 enum ldlm_error err = ELDLM_OK;
1227 struct ldlm_lock *lock = NULL;
1228 void *cookie = NULL;
1230 struct ldlm_resource *res = NULL;
1231 const struct lu_env *env = req->rq_svc_thread->t_env;
1235 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1237 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1238 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1240 LASSERT(req->rq_export);
1242 /* for intent enqueue the stat will be updated inside intent policy */
1243 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1244 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1245 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1247 if (req->rq_export && req->rq_export->exp_nid_stats &&
1248 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1249 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1250 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1252 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1253 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1254 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1255 dlm_req->lock_desc.l_resource.lr_type);
1256 GOTO(out, rc = -EFAULT);
1259 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1260 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1261 dlm_req->lock_desc.l_req_mode &
1262 (dlm_req->lock_desc.l_req_mode-1))) {
1263 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1264 dlm_req->lock_desc.l_req_mode);
1265 GOTO(out, rc = -EFAULT);
1268 if (unlikely((flags & LDLM_FL_REPLAY) ||
1269 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1270 /* Find an existing lock in the per-export lock hash */
1272 * In the function below, .hs_keycmp resolves to
1273 * ldlm_export_lock_keycmp()
1275 /* coverity[overrun-buffer-val] */
1276 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1277 (void *)&dlm_req->lock_handle[0]);
1279 DEBUG_REQ(D_DLMTRACE, req,
1280 "found existing lock cookie %#llx",
1281 lock->l_handle.h_cookie);
1282 flags |= LDLM_FL_RESENT;
1283 GOTO(existing_lock, rc = 0);
1286 if (ldlm_reclaim_full()) {
1287 DEBUG_REQ(D_DLMTRACE, req,
1288 "Too many granted locks, reject current enqueue request and let the client retry later.\n");
1289 GOTO(out, rc = -EINPROGRESS);
1293 /* The lock's callback data might be set in the policy function */
1294 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1295 dlm_req->lock_desc.l_resource.lr_type,
1296 dlm_req->lock_desc.l_req_mode,
1297 cbs, NULL, 0, LVB_T_NONE);
1304 lock->l_remote_handle = dlm_req->lock_handle[0];
1305 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1308 * Initialize resource lvb but not for a lock being replayed since
1309 * Client already got lvb sent in this case.
1310 * This must occur early since some policy methods assume resource
1311 * lvb is available (lr_lvb_data != NULL).
1313 res = lock->l_resource;
1314 if (!(flags & LDLM_FL_REPLAY)) {
1315 /* non-replayed lock, delayed lvb init may need to be done */
1316 rc = ldlm_lvbo_init(res);
1318 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1323 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1325 * Don't enqueue a lock onto the export if it is been disonnected
1326 * due to eviction (b=3822) or server umount (b=24324).
1327 * Cancel it now instead.
1329 if (req->rq_export->exp_disconnected) {
1330 LDLM_ERROR(lock, "lock on disconnected export %p",
1332 GOTO(out, rc = -ENOTCONN);
1335 lock->l_export = class_export_lock_get(req->rq_export, lock);
1336 if (lock->l_export->exp_lock_hash)
1337 cfs_hash_add(lock->l_export->exp_lock_hash,
1338 &lock->l_remote_handle,
1342 * Inherit the enqueue flags before the operation, because we do not
1343 * keep the res lock on return and next operations (BL AST) may proceed
1346 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1347 LDLM_FL_INHERIT_MASK);
1349 ldlm_convert_policy_to_local(req->rq_export,
1350 dlm_req->lock_desc.l_resource.lr_type,
1351 &dlm_req->lock_desc.l_policy_data,
1352 &lock->l_policy_data);
1353 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1354 lock->l_req_extent = lock->l_policy_data.l_extent;
1357 if (flags & LDLM_FL_HAS_INTENT) {
1359 * In this case, the reply buffer is allocated deep in
1360 * local_lock_enqueue by the policy function.
1365 * based on the assumption that lvb size never changes during
1366 * resource life time otherwise it need resource->lr_lock's
1369 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1370 RCL_SERVER, ldlm_lvbo_size(lock));
1372 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1373 GOTO(out, rc = -ENOMEM);
1375 rc = req_capsule_server_pack(&req->rq_pill);
1380 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1387 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1389 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1390 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1392 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1393 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1396 * We never send a blocking AST until the lock is granted, but
1397 * we can tell it right now
1399 lock_res_and_lock(lock);
1402 * Now take into account flags to be inherited from original lock
1403 * request both in reply to client and in our own lock flags.
1405 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1406 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1409 * Don't move a pending lock onto the export if it has already been
1410 * disconnected due to eviction (b=5683) or server umount (b=24324).
1411 * Cancel it now instead.
1413 if (unlikely(req->rq_export->exp_disconnected ||
1414 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1415 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1417 } else if (ldlm_is_ast_sent(lock)) {
1418 /* fill lock desc for possible lock convert */
1419 if (lock->l_blocking_lock &&
1420 lock->l_resource->lr_type == LDLM_IBITS) {
1421 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1422 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1425 "save blocking bits %llx in granted lock",
1426 bl_lock->l_policy_data.l_inodebits.bits);
1428 * If lock is blocked then save blocking ibits
1429 * in returned lock policy for the possible lock
1430 * convert on a client.
1432 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1433 bl_lock->l_policy_data.l_inodebits.bits;
1435 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1436 if (ldlm_is_granted(lock)) {
1438 * Only cancel lock if it was granted, because it would
1439 * be destroyed immediately and would never be granted
1440 * in the future, causing timeouts on client. Not
1441 * granted lock will be cancelled immediately after
1442 * sending completion AST.
1444 if (ldlm_is_cancel_on_block(lock)) {
1445 unlock_res_and_lock(lock);
1446 ldlm_lock_cancel(lock);
1447 lock_res_and_lock(lock);
1449 ldlm_add_waiting_lock(lock,
1450 ldlm_bl_timeout(lock));
1454 unlock_res_and_lock(lock);
1458 req->rq_status = rc ?: err; /* return either error - b=11190 */
1459 if (!req->rq_packed_final) {
1460 err = lustre_pack_reply(req, 1, NULL, NULL);
1466 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1467 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1471 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1475 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1477 ldlm_lvbo_size(lock) > 0) {
1482 buf = req_capsule_server_get(&req->rq_pill,
1484 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1485 buflen = req_capsule_get_size(&req->rq_pill,
1486 &RMF_DLM_LVB, RCL_SERVER);
1488 * non-replayed lock, delayed lvb init may
1489 * need to be occur now
1491 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1494 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1496 req_capsule_shrink(&req->rq_pill,
1499 } else if (rc2 == -ERANGE) {
1500 rc2 = req_capsule_server_grow(
1502 &RMF_DLM_LVB, buflen);
1507 * if we can't grow the buffer,
1508 * it's ok to return empty lvb
1519 } else if (flags & LDLM_FL_REPLAY) {
1520 /* no LVB resend upon replay */
1522 req_capsule_shrink(&req->rq_pill,
1532 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1533 if (lock->l_export) {
1534 ldlm_lock_cancel(lock);
1536 lock_res_and_lock(lock);
1537 ldlm_resource_unlink_lock(lock);
1538 ldlm_lock_destroy_nolock(lock);
1539 unlock_res_and_lock(lock);
1544 if (!err && !ldlm_is_cbpending(lock) &&
1545 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1546 ldlm_reprocess_all(lock->l_resource);
1548 LDLM_LOCK_RELEASE(lock);
1551 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1558 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1559 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1561 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1563 if (lock->l_blocking_lock) {
1564 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1565 lock->l_blocking_lock = NULL;
1569 /* A lock can be converted to new ibits or mode and should be considered
1570 * as new lock. Clear all states related to a previous blocking AST
1571 * processing so new conflicts will cause new blocking ASTs.
1573 * This is used during lock convert below and lock downgrade to COS mode in
1574 * ldlm_lock_mode_downgrade().
1576 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1578 ldlm_clear_ast_sent(lock);
1579 lock->l_bl_ast_run = 0;
1580 ldlm_clear_blocking_lock(lock);
1584 * Main LDLM entry point for server code to process lock conversion requests.
1586 int ldlm_handle_convert0(struct ptlrpc_request *req,
1587 const struct ldlm_request *dlm_req)
1589 struct obd_export *exp = req->rq_export;
1590 struct ldlm_reply *dlm_rep;
1591 struct ldlm_lock *lock;
1596 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1597 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1598 LDLM_CONVERT - LDLM_FIRST_OPC);
1600 rc = req_capsule_server_pack(&req->rq_pill);
1604 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1605 dlm_rep->lock_flags = dlm_req->lock_flags;
1607 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1612 bits = lock->l_policy_data.l_inodebits.bits;
1613 new = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1614 LDLM_DEBUG(lock, "server-side convert handler START");
1616 if (ldlm_is_cancel(lock)) {
1617 LDLM_ERROR(lock, "convert on canceled lock!");
1618 rc = ELDLM_NO_LOCK_DATA;
1619 } else if (dlm_req->lock_desc.l_req_mode !=
1620 lock->l_granted_mode) {
1621 LDLM_ERROR(lock, "lock mode differs!");
1622 rc = ELDLM_NO_LOCK_DATA;
1623 } else if (bits == new) {
1625 * This can be valid situation if CONVERT RPCs are
1626 * re-ordered. Just finish silently
1628 LDLM_DEBUG(lock, "lock is converted already!");
1631 lock_res_and_lock(lock);
1632 if (ldlm_is_waited(lock))
1633 ldlm_del_waiting_lock(lock);
1635 ldlm_clear_cbpending(lock);
1636 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1637 ldlm_inodebits_drop(lock, bits & ~new);
1639 ldlm_clear_blocking_data(lock);
1640 unlock_res_and_lock(lock);
1642 ldlm_reprocess_all(lock->l_resource);
1646 if (rc == ELDLM_OK) {
1647 dlm_rep->lock_handle = lock->l_remote_handle;
1648 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1649 &dlm_rep->lock_desc.l_policy_data);
1652 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d",
1654 LDLM_LOCK_PUT(lock);
1656 rc = ELDLM_NO_LOCK_DATA;
1657 LDLM_DEBUG_NOLOCK("server-side convert handler END, rc = %d",
1661 req->rq_status = rc;
1667 * Cancel all the locks whose handles are packed into ldlm_request
1669 * Called by server code expecting such combined cancel activity
1672 int ldlm_request_cancel(struct ptlrpc_request *req,
1673 const struct ldlm_request *dlm_req,
1674 int first, enum lustre_at_flags flags)
1676 struct ldlm_resource *res, *pres = NULL;
1677 struct ldlm_lock *lock;
1678 int i, count, done = 0;
1682 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1686 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1690 * There is no lock on the server at the replay time,
1691 * skip lock cancelling to make replay tests to pass.
1693 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1696 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1699 for (i = first; i < count; i++) {
1700 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1702 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1703 dlm_req->lock_handle[i].cookie);
1707 res = lock->l_resource;
1711 * This code is an optimization to only attempt lock
1712 * granting on the resource (that could be CPU-expensive)
1713 * after we are done cancelling lock in that resource.
1717 ldlm_reprocess_all(pres);
1718 LDLM_RESOURCE_DELREF(pres);
1719 ldlm_resource_putref(pres);
1722 ldlm_resource_getref(res);
1723 LDLM_RESOURCE_ADDREF(res);
1725 if (!ldlm_is_discard_data(lock))
1726 ldlm_lvbo_update(res, lock,
1732 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1733 lock->l_blast_sent != 0) {
1734 time64_t delay = ktime_get_real_seconds() -
1737 "server cancels blocked lock after %llds",
1739 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1741 ldlm_lock_cancel(lock);
1742 LDLM_LOCK_PUT(lock);
1745 ldlm_reprocess_all(pres);
1746 LDLM_RESOURCE_DELREF(pres);
1747 ldlm_resource_putref(pres);
1749 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1752 EXPORT_SYMBOL(ldlm_request_cancel);
1755 * Main LDLM entry point for server code to cancel locks.
1757 * Typically gets called from service handler on LDLM_CANCEL opc.
1759 int ldlm_handle_cancel(struct ptlrpc_request *req)
1761 struct ldlm_request *dlm_req;
1766 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1767 if (dlm_req == NULL) {
1768 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1772 if (req->rq_export && req->rq_export->exp_nid_stats &&
1773 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1774 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1775 LDLM_CANCEL - LDLM_FIRST_OPC);
1777 rc = req_capsule_server_pack(&req->rq_pill);
1781 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1782 req->rq_status = LUSTRE_ESTALE;
1784 RETURN(ptlrpc_reply(req));
1786 #endif /* HAVE_SERVER_SUPPORT */
1789 * Callback handler for receiving incoming blocking ASTs.
1791 * This can only happen on client side.
1793 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1794 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1800 LDLM_DEBUG(lock, "client blocking AST callback handler");
1802 lock_res_and_lock(lock);
1804 /* set bits to cancel for this lock for possible lock convert */
1805 if (ns_is_client(ns) && (lock->l_resource->lr_type == LDLM_IBITS)) {
1807 * Lock description contains policy of blocking lock,
1808 * and its cancel_bits is used to pass conflicting bits.
1809 * NOTE: ld can be NULL or can be not NULL but zeroed if
1810 * passed from ldlm_bl_thread_blwi(), check below used bits
1811 * in ld to make sure it is valid description.
1813 * If server may replace lock resource keeping the same cookie,
1814 * never use cancel bits from different resource, full cancel
1817 if (ld && ld->l_policy_data.l_inodebits.bits &&
1818 ldlm_res_eq(&ld->l_resource.lr_name,
1819 &lock->l_resource->lr_name))
1820 lock->l_policy_data.l_inodebits.cancel_bits =
1821 ld->l_policy_data.l_inodebits.cancel_bits;
1823 * if there is no valid ld and lock is cbpending already
1824 * then cancel_bits should be kept, otherwise it is zeroed.
1826 else if (!ldlm_is_cbpending(lock))
1827 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1829 ldlm_set_cbpending(lock);
1831 do_ast = (!lock->l_readers && !lock->l_writers);
1832 unlock_res_and_lock(lock);
1836 "Lock %p already unused, calling callback (%p)\n",
1837 lock, lock->l_blocking_ast);
1838 if (lock->l_blocking_ast != NULL)
1839 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1843 "Lock %p is referenced, will be cancelled later\n",
1847 LDLM_DEBUG(lock, "client blocking callback handler END");
1848 LDLM_LOCK_RELEASE(lock);
1853 * Callback handler for receiving incoming completion ASTs.
1855 * This only can happen on client side.
1857 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1858 struct ldlm_namespace *ns,
1859 struct ldlm_request *dlm_req,
1860 struct ldlm_lock *lock)
1862 struct list_head ast_list;
1868 LDLM_DEBUG(lock, "client completion callback handler START");
1870 INIT_LIST_HEAD(&ast_list);
1871 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1872 long to = cfs_time_seconds(1);
1875 set_current_state(TASK_INTERRUPTIBLE);
1876 schedule_timeout(to);
1877 if (ldlm_is_granted(lock) ||
1878 ldlm_is_destroyed(lock))
1883 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1885 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1886 GOTO(out, rc = lvb_len);
1887 } else if (lvb_len > 0) {
1888 if (lock->l_lvb_len > 0) {
1889 /* for extent lock, lvb contains ost_lvb{}. */
1890 LASSERT(lock->l_lvb_data != NULL);
1892 if (unlikely(lock->l_lvb_len < lvb_len)) {
1894 "Replied LVB is larger than expectation, expected = %d, replied = %d",
1895 lock->l_lvb_len, lvb_len);
1896 GOTO(out, rc = -EINVAL);
1901 lock_res_and_lock(lock);
1903 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1904 &lock->l_resource->lr_name)) {
1905 ldlm_resource_unlink_lock(lock);
1906 unlock_res_and_lock(lock);
1907 rc = ldlm_lock_change_resource(ns, lock,
1908 &dlm_req->lock_desc.l_resource.lr_name);
1910 LDLM_ERROR(lock, "Failed to allocate resource");
1913 LDLM_DEBUG(lock, "completion AST, new resource");
1914 lock_res_and_lock(lock);
1917 if (ldlm_is_destroyed(lock) ||
1918 ldlm_is_granted(lock)) {
1919 /* b=11300: the lock has already been granted */
1920 unlock_res_and_lock(lock);
1921 LDLM_DEBUG(lock, "Double grant race happened");
1926 * If we receive the completion AST before the actual enqueue returned,
1927 * then we might need to switch lock modes, resources, or extents.
1929 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1930 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1931 LDLM_DEBUG(lock, "completion AST, new lock mode");
1934 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1935 ldlm_convert_policy_to_local(req->rq_export,
1936 dlm_req->lock_desc.l_resource.lr_type,
1937 &dlm_req->lock_desc.l_policy_data,
1938 &lock->l_policy_data);
1939 LDLM_DEBUG(lock, "completion AST, new policy data");
1942 ldlm_resource_unlink_lock(lock);
1944 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1946 * BL_AST locks are not needed in LRU.
1947 * Let ldlm_cancel_lru() be fast.
1949 ldlm_lock_remove_from_lru(lock);
1950 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1951 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1954 if (lock->l_lvb_len > 0) {
1955 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1956 lock->l_lvb_data, lvb_len);
1958 unlock_res_and_lock(lock);
1963 ldlm_grant_lock(lock, &ast_list);
1964 unlock_res_and_lock(lock);
1966 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1969 * Let Enqueue to call osc_lock_upcall() and initialize
1972 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1974 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1976 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1982 lock_res_and_lock(lock);
1983 ldlm_set_failed(lock);
1984 unlock_res_and_lock(lock);
1985 wake_up(&lock->l_waitq);
1987 LDLM_LOCK_RELEASE(lock);
1991 * Callback handler for receiving incoming glimpse ASTs.
1993 * This only can happen on client side. After handling the glimpse AST
1994 * we also consider dropping the lock here if it is unused locally for a
1997 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1998 struct ldlm_namespace *ns,
1999 struct ldlm_request *dlm_req,
2000 struct ldlm_lock *lock)
2006 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2008 if (lock->l_glimpse_ast != NULL)
2009 rc = lock->l_glimpse_ast(lock, req);
2011 if (req->rq_repmsg != NULL) {
2014 req->rq_status = rc;
2018 lock_res_and_lock(lock);
2019 if (lock->l_granted_mode == LCK_PW &&
2020 !lock->l_readers && !lock->l_writers &&
2021 ktime_after(ktime_get(),
2022 ktime_add(lock->l_last_used,
2023 ktime_set(ns->ns_dirty_age_limit, 0)))) {
2024 unlock_res_and_lock(lock);
2025 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
2026 ldlm_handle_bl_callback(ns, NULL, lock);
2031 unlock_res_and_lock(lock);
2032 LDLM_LOCK_RELEASE(lock);
2036 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
2038 if (req->rq_no_reply)
2041 req->rq_status = rc;
2042 if (!req->rq_packed_final) {
2043 rc = lustre_pack_reply(req, 1, NULL, NULL);
2047 return ptlrpc_reply(req);
2050 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2051 enum ldlm_cancel_flags cancel_flags)
2053 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2057 spin_lock(&blp->blp_lock);
2058 if (blwi->blwi_lock &&
2059 ldlm_is_discard_data(blwi->blwi_lock)) {
2060 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2061 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2063 /* other blocking callbacks are added to the regular list */
2064 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2066 spin_unlock(&blp->blp_lock);
2068 wake_up(&blp->blp_waitq);
2071 * can not check blwi->blwi_flags as blwi could be already freed in
2074 if (!(cancel_flags & LCF_ASYNC))
2075 wait_for_completion(&blwi->blwi_comp);
2080 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2081 struct ldlm_namespace *ns,
2082 struct ldlm_lock_desc *ld,
2083 struct list_head *cancels, int count,
2084 struct ldlm_lock *lock,
2085 enum ldlm_cancel_flags cancel_flags)
2087 init_completion(&blwi->blwi_comp);
2088 INIT_LIST_HEAD(&blwi->blwi_head);
2090 if (memory_pressure_get())
2091 blwi->blwi_mem_pressure = 1;
2094 blwi->blwi_flags = cancel_flags;
2096 blwi->blwi_ld = *ld;
2098 list_add(&blwi->blwi_head, cancels);
2099 list_del_init(cancels);
2100 blwi->blwi_count = count;
2102 blwi->blwi_lock = lock;
2107 * Queues a list of locks \a cancels containing \a count locks
2108 * for later processing by a blocking thread. If \a count is zero,
2109 * then the lock referenced as \a lock is queued instead.
2111 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2112 * If list addition fails an error is returned and caller is supposed to
2113 * call ->l_blocking_ast itself.
2115 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2116 struct ldlm_lock_desc *ld,
2117 struct ldlm_lock *lock,
2118 struct list_head *cancels, int count,
2119 enum ldlm_cancel_flags cancel_flags)
2123 if (cancels && count == 0)
2126 if (cancel_flags & LCF_ASYNC) {
2127 struct ldlm_bl_work_item *blwi;
2129 OBD_ALLOC(blwi, sizeof(*blwi));
2132 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2134 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2137 * if it is synchronous call do minimum mem alloc, as it could
2138 * be triggered from kernel shrinker
2140 struct ldlm_bl_work_item blwi;
2142 memset(&blwi, 0, sizeof(blwi));
2143 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2144 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2149 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2150 struct ldlm_lock *lock)
2152 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2155 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2156 struct list_head *cancels, int count,
2157 enum ldlm_cancel_flags cancel_flags)
2159 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2162 int ldlm_bl_thread_wakeup(void)
2164 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2168 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2169 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2171 struct obd_device *obd = req->rq_export->exp_obd;
2179 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2181 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2183 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2185 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2188 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2190 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2192 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2195 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2198 /* We are responsible for swabbing contents of val */
2200 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2201 /* Pass it on to mdc (the "export" in this case) */
2202 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2204 sizeof(KEY_HSM_COPYTOOL_SEND),
2205 KEY_HSM_COPYTOOL_SEND,
2208 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2213 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2214 const char *msg, int rc,
2215 const struct lustre_handle *handle)
2217 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2218 "%s: [nid %s] [rc %d] [lock %#llx]",
2219 msg, libcfs_id2str(req->rq_peer), rc,
2220 handle ? handle->cookie : 0);
2221 if (req->rq_no_reply)
2222 CWARN("No reply was sent, maybe cause b=21636.\n");
2224 CWARN("Send reply failed, maybe cause b=21636.\n");
2227 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2228 static int ldlm_callback_handler(struct ptlrpc_request *req)
2230 struct ldlm_namespace *ns;
2231 struct ldlm_request *dlm_req;
2232 struct ldlm_lock *lock;
2238 * Requests arrive in sender's byte order. The ptlrpc service
2239 * handler has already checked and, if necessary, byte-swapped the
2240 * incoming request message body, but I am responsible for the
2244 /* do nothing for sec context finalize */
2245 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2248 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2250 if (req->rq_export == NULL) {
2251 rc = ldlm_callback_reply(req, -ENOTCONN);
2252 ldlm_callback_errmsg(req, "Operate on unconnected server",
2257 LASSERT(req->rq_export != NULL);
2258 LASSERT(req->rq_export->exp_obd != NULL);
2260 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2261 case LDLM_BL_CALLBACK:
2262 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2264 ldlm_callback_reply(req, -(int)cfs_fail_err);
2268 case LDLM_CP_CALLBACK:
2269 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2272 case LDLM_GL_CALLBACK:
2273 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2277 rc = ldlm_handle_setinfo(req);
2278 ldlm_callback_reply(req, rc);
2281 CERROR("unknown opcode %u\n",
2282 lustre_msg_get_opc(req->rq_reqmsg));
2283 ldlm_callback_reply(req, -EPROTO);
2287 ns = req->rq_export->exp_obd->obd_namespace;
2288 LASSERT(ns != NULL);
2290 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2292 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2293 if (dlm_req == NULL) {
2294 rc = ldlm_callback_reply(req, -EPROTO);
2295 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2301 * Force a known safe race, send a cancel to the server for a lock
2302 * which the server has already started a blocking callback on.
2304 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2305 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2306 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2308 CERROR("ldlm_cli_cancel: %d\n", rc);
2311 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2314 "callback on lock %#llx - lock disappeared\n",
2315 dlm_req->lock_handle[0].cookie);
2316 rc = ldlm_callback_reply(req, -EINVAL);
2317 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2318 &dlm_req->lock_handle[0]);
2322 if (ldlm_is_fail_loc(lock) &&
2323 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2324 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2326 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2327 lock_res_and_lock(lock);
2328 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2330 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2332 * If somebody cancels lock and cache is already dropped,
2333 * or lock is failed before cp_ast received on client,
2334 * we can tell the server we have no lock. Otherwise, we
2335 * should send cancel after dropping the cache.
2337 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2338 ldlm_is_failed(lock)) {
2340 "callback on lock %llx - lock disappeared",
2341 dlm_req->lock_handle[0].cookie);
2342 unlock_res_and_lock(lock);
2343 LDLM_LOCK_RELEASE(lock);
2344 rc = ldlm_callback_reply(req, -EINVAL);
2345 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2346 &dlm_req->lock_handle[0]);
2350 * BL_AST locks are not needed in LRU.
2351 * Let ldlm_cancel_lru() be fast.
2353 ldlm_lock_remove_from_lru(lock);
2354 ldlm_set_bl_ast(lock);
2356 unlock_res_and_lock(lock);
2359 * We want the ost thread to get this reply so that it can respond
2360 * to ost requests (write cache writeback) that might be triggered
2363 * But we'd also like to be able to indicate in the reply that we're
2364 * cancelling right now, because it's unused, or have an intent result
2365 * in the reply, so we might have to push the responsibility for sending
2366 * the reply down into the AST handlers, alas.
2369 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2370 case LDLM_BL_CALLBACK:
2371 CDEBUG(D_INODE, "blocking ast\n");
2372 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2373 if (!ldlm_is_cancel_on_block(lock)) {
2374 rc = ldlm_callback_reply(req, 0);
2375 if (req->rq_no_reply || rc)
2376 ldlm_callback_errmsg(req, "Normal process", rc,
2377 &dlm_req->lock_handle[0]);
2379 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2380 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2382 case LDLM_CP_CALLBACK:
2383 CDEBUG(D_INODE, "completion ast\n");
2384 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2385 ldlm_callback_reply(req, 0);
2386 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2388 case LDLM_GL_CALLBACK:
2389 CDEBUG(D_INODE, "glimpse ast\n");
2390 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2391 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2394 LBUG(); /* checked above */
2400 #ifdef HAVE_SERVER_SUPPORT
2402 * Main handler for canceld thread.
2404 * Separated into its own thread to avoid deadlocks.
2406 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2413 * Requests arrive in sender's byte order. The ptlrpc service
2414 * handler has already checked and, if necessary, byte-swapped the
2415 * incoming request message body, but I am responsible for the
2419 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2421 if (req->rq_export == NULL) {
2422 struct ldlm_request *dlm_req;
2424 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2425 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2426 libcfs_nid2str(req->rq_peer.nid),
2427 (unsigned long long)req->rq_arrival_time.tv_sec,
2428 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2430 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2431 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2432 dlm_req = req_capsule_client_get(&req->rq_pill,
2434 if (dlm_req != NULL)
2435 ldlm_lock_dump_handle(D_ERROR,
2436 &dlm_req->lock_handle[0]);
2438 ldlm_callback_reply(req, -ENOTCONN);
2442 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2443 /* XXX FIXME move this back to mds/handler.c, b=249 */
2445 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2446 CDEBUG(D_INODE, "cancel\n");
2447 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2448 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2449 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2451 rc = ldlm_handle_cancel(req);
2455 struct ldlm_request *dlm_req;
2457 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2458 CDEBUG(D_INODE, "convert\n");
2460 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2461 if (dlm_req == NULL) {
2462 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2463 rc = ldlm_callback_reply(req, -EPROTO);
2465 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2466 rc = ptlrpc_reply(req);
2471 CERROR("invalid opcode %d\n",
2472 lustre_msg_get_opc(req->rq_reqmsg));
2473 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2474 rc = ldlm_callback_reply(req, -EINVAL);
2480 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2481 struct ldlm_lock *lock)
2483 struct ldlm_request *dlm_req;
2484 struct lustre_handle lockh;
2490 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2491 if (dlm_req == NULL)
2494 ldlm_lock2handle(lock, &lockh);
2495 for (i = 0; i < dlm_req->lock_count; i++) {
2496 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2498 DEBUG_REQ(D_RPCTRACE, req,
2499 "Prio raised by lock %#llx.", lockh.cookie);
2508 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2510 struct ldlm_request *dlm_req;
2516 /* no prolong in recovery */
2517 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2520 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2521 if (dlm_req == NULL)
2524 for (i = 0; i < dlm_req->lock_count; i++) {
2525 struct ldlm_lock *lock;
2527 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2531 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2533 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2534 LDLM_LOCK_PUT(lock);
2543 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2544 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2545 .hpreq_check = ldlm_cancel_hpreq_check,
2549 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2553 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2555 if (req->rq_export == NULL)
2558 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2559 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2560 req->rq_ops = &ldlm_cancel_hpreq_ops;
2561 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2562 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2563 req->rq_ops = &ldlm_cancel_hpreq_ops;
2568 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2569 struct hlist_node *hnode, void *data)
2572 struct list_head *rpc_list = data;
2573 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2575 lock_res_and_lock(lock);
2577 if (!ldlm_is_granted(lock)) {
2578 unlock_res_and_lock(lock);
2582 LASSERT(lock->l_resource);
2583 if (lock->l_resource->lr_type != LDLM_IBITS &&
2584 lock->l_resource->lr_type != LDLM_PLAIN) {
2585 unlock_res_and_lock(lock);
2589 if (ldlm_is_ast_sent(lock)) {
2590 unlock_res_and_lock(lock);
2594 LASSERT(lock->l_blocking_ast);
2595 LASSERT(!lock->l_blocking_lock);
2597 ldlm_set_ast_sent(lock);
2598 if (lock->l_export && lock->l_export->exp_lock_hash) {
2600 * NB: it's safe to call cfs_hash_del() even lock isn't
2604 * In the function below, .hs_keycmp resolves to
2605 * ldlm_export_lock_keycmp()
2607 /* coverity[overrun-buffer-val] */
2608 cfs_hash_del(lock->l_export->exp_lock_hash,
2609 &lock->l_remote_handle, &lock->l_exp_hash);
2612 list_add_tail(&lock->l_rk_ast, rpc_list);
2613 LDLM_LOCK_GET(lock);
2615 unlock_res_and_lock(lock);
2619 void ldlm_revoke_export_locks(struct obd_export *exp)
2621 struct list_head rpc_list;
2625 INIT_LIST_HEAD(&rpc_list);
2626 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2627 ldlm_revoke_lock_cb, &rpc_list, 0);
2628 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2629 LDLM_WORK_REVOKE_AST);
2633 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2634 #endif /* HAVE_SERVER_SUPPORT */
2636 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2637 struct ldlm_bl_work_item **p_blwi,
2638 struct obd_export **p_exp)
2640 struct ldlm_bl_work_item *blwi = NULL;
2641 static unsigned int num_bl;
2642 static unsigned int num_stale;
2643 int num_th = atomic_read(&blp->blp_num_threads);
2645 *p_exp = obd_stale_export_get();
2647 spin_lock(&blp->blp_lock);
2648 if (*p_exp != NULL) {
2649 if (num_th == 1 || ++num_stale < num_th) {
2650 spin_unlock(&blp->blp_lock);
2656 /* process a request from the blp_list at least every blp_num_threads */
2657 if (!list_empty(&blp->blp_list) &&
2658 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2659 blwi = list_entry(blp->blp_list.next,
2660 struct ldlm_bl_work_item, blwi_entry);
2662 if (!list_empty(&blp->blp_prio_list))
2663 blwi = list_entry(blp->blp_prio_list.next,
2664 struct ldlm_bl_work_item,
2668 if (++num_bl >= num_th)
2670 list_del(&blwi->blwi_entry);
2672 spin_unlock(&blp->blp_lock);
2675 if (*p_exp != NULL && *p_blwi != NULL) {
2676 obd_stale_export_put(*p_exp);
2680 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2683 /* This only contains temporary data until the thread starts */
2684 struct ldlm_bl_thread_data {
2685 struct ldlm_bl_pool *bltd_blp;
2686 struct completion bltd_comp;
2690 static int ldlm_bl_thread_main(void *arg);
2692 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2694 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2695 struct task_struct *task;
2697 init_completion(&bltd.bltd_comp);
2699 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2700 if (bltd.bltd_num >= blp->blp_max_threads) {
2701 atomic_dec(&blp->blp_num_threads);
2705 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2707 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2708 atomic_dec(&blp->blp_num_threads);
2712 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2715 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2716 bltd.bltd_num, PTR_ERR(task));
2717 atomic_dec(&blp->blp_num_threads);
2718 return PTR_ERR(task);
2720 wait_for_completion(&bltd.bltd_comp);
2725 /* Not fatal if racy and have a few too many threads */
2726 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2727 struct ldlm_bl_work_item *blwi)
2729 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2732 if (atomic_read(&blp->blp_busy_threads) <
2733 atomic_read(&blp->blp_num_threads))
2736 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2737 blwi->blwi_mem_pressure))
2743 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2744 struct ldlm_bl_work_item *blwi)
2748 if (blwi->blwi_ns == NULL)
2749 /* added by ldlm_cleanup() */
2750 RETURN(LDLM_ITER_STOP);
2752 if (blwi->blwi_mem_pressure)
2753 memory_pressure_set();
2755 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2757 if (blwi->blwi_count) {
2760 * The special case when we cancel locks in lru
2761 * asynchronously, we pass the list of locks here.
2762 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2763 * canceled locally yet.
2765 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2768 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2771 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2774 if (blwi->blwi_mem_pressure)
2775 memory_pressure_clr();
2777 if (blwi->blwi_flags & LCF_ASYNC)
2778 OBD_FREE(blwi, sizeof(*blwi));
2780 complete(&blwi->blwi_comp);
2786 * Cancel stale locks on export. Cancel blocked locks first.
2787 * If the given export has blocked locks, the next in the list may have
2788 * them too, thus cancel not blocked locks only if the current export has
2791 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2792 struct obd_export *exp)
2798 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2800 num = ldlm_export_cancel_blocked_locks(exp);
2802 ldlm_export_cancel_locks(exp);
2804 obd_stale_export_put(exp);
2811 * Main blocking requests processing thread.
2813 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2814 * This thread in the end ends up doing actual call to ->l_blocking_ast
2817 static int ldlm_bl_thread_main(void *arg)
2820 struct ldlm_bl_pool *blp;
2821 struct ldlm_bl_thread_data *bltd = arg;
2829 rc = lu_env_init(env, LCT_DT_THREAD);
2832 rc = lu_env_add(env);
2834 GOTO(out_env_fini, rc);
2836 blp = bltd->bltd_blp;
2838 complete(&bltd->bltd_comp);
2839 /* cannot use bltd after this, it is only on caller's stack */
2842 struct l_wait_info lwi = { 0 };
2843 struct ldlm_bl_work_item *blwi = NULL;
2844 struct obd_export *exp = NULL;
2847 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2850 l_wait_event_exclusive(blp->blp_waitq,
2851 ldlm_bl_get_work(blp, &blwi,
2854 atomic_inc(&blp->blp_busy_threads);
2856 if (ldlm_bl_thread_need_create(blp, blwi))
2857 /* discard the return value, we tried */
2858 ldlm_bl_thread_start(blp, true);
2861 rc = ldlm_bl_thread_exports(blp, exp);
2863 rc = ldlm_bl_thread_blwi(blp, blwi);
2865 atomic_dec(&blp->blp_busy_threads);
2867 if (rc == LDLM_ITER_STOP)
2871 * If there are many namespaces, we will not sleep waiting for
2872 * work, and must do a cond_resched to avoid holding the CPU
2878 atomic_dec(&blp->blp_num_threads);
2879 complete(&blp->blp_comp);
2890 static int ldlm_setup(void);
2891 static int ldlm_cleanup(void);
2893 int ldlm_get_ref(void)
2898 mutex_lock(&ldlm_ref_mutex);
2899 if (++ldlm_refcount == 1) {
2904 mutex_unlock(&ldlm_ref_mutex);
2909 void ldlm_put_ref(void)
2912 mutex_lock(&ldlm_ref_mutex);
2913 if (ldlm_refcount == 1) {
2914 int rc = ldlm_cleanup();
2917 CERROR("ldlm_cleanup failed: %d\n", rc);
2923 mutex_unlock(&ldlm_ref_mutex);
2929 * Export handle<->lock hash operations.
2932 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
2934 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2938 ldlm_export_lock_key(struct hlist_node *hnode)
2940 struct ldlm_lock *lock;
2942 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2943 return &lock->l_remote_handle;
2947 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2949 struct ldlm_lock *lock;
2951 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2952 lock->l_remote_handle = *(struct lustre_handle *)key;
2956 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2958 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2962 ldlm_export_lock_object(struct hlist_node *hnode)
2964 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2968 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2970 struct ldlm_lock *lock;
2972 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2973 LDLM_LOCK_GET(lock);
2977 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2979 struct ldlm_lock *lock;
2981 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2982 LDLM_LOCK_RELEASE(lock);
2985 static struct cfs_hash_ops ldlm_export_lock_ops = {
2986 .hs_hash = ldlm_export_lock_hash,
2987 .hs_key = ldlm_export_lock_key,
2988 .hs_keycmp = ldlm_export_lock_keycmp,
2989 .hs_keycpy = ldlm_export_lock_keycpy,
2990 .hs_object = ldlm_export_lock_object,
2991 .hs_get = ldlm_export_lock_get,
2992 .hs_put = ldlm_export_lock_put,
2993 .hs_put_locked = ldlm_export_lock_put,
2996 int ldlm_init_export(struct obd_export *exp)
3002 exp->exp_lock_hash =
3003 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3004 HASH_EXP_LOCK_CUR_BITS,
3005 HASH_EXP_LOCK_MAX_BITS,
3006 HASH_EXP_LOCK_BKT_BITS, 0,
3007 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3008 &ldlm_export_lock_ops,
3009 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3010 CFS_HASH_NBLK_CHANGE);
3012 if (!exp->exp_lock_hash)
3015 rc = ldlm_init_flock_export(exp);
3021 ldlm_destroy_export(exp);
3024 EXPORT_SYMBOL(ldlm_init_export);
3026 void ldlm_destroy_export(struct obd_export *exp)
3029 cfs_hash_putref(exp->exp_lock_hash);
3030 exp->exp_lock_hash = NULL;
3032 ldlm_destroy_flock_export(exp);
3035 EXPORT_SYMBOL(ldlm_destroy_export);
3037 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3038 struct attribute *attr,
3041 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3044 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3045 struct attribute *attr,
3052 rc = kstrtoul(buffer, 10, &val);
3056 ldlm_cancel_unused_locks_before_replay = val;
3060 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3062 static struct attribute *ldlm_attrs[] = {
3063 &lustre_attr_cancel_unused_locks_before_replay.attr,
3067 static struct attribute_group ldlm_attr_group = {
3068 .attrs = ldlm_attrs,
3071 static int ldlm_setup(void)
3073 static struct ptlrpc_service_conf conf;
3074 struct ldlm_bl_pool *blp = NULL;
3075 #ifdef HAVE_SERVER_SUPPORT
3076 struct task_struct *task;
3077 #endif /* HAVE_SERVER_SUPPORT */
3083 if (ldlm_state != NULL)
3086 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3087 if (ldlm_state == NULL)
3090 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3094 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3098 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3102 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3106 rc = ldlm_debugfs_setup();
3110 memset(&conf, 0, sizeof(conf));
3111 conf = (typeof(conf)) {
3112 .psc_name = "ldlm_cbd",
3113 .psc_watchdog_factor = 2,
3115 .bc_nbufs = LDLM_CLIENT_NBUFS,
3116 .bc_buf_size = LDLM_BUFSIZE,
3117 .bc_req_max_size = LDLM_MAXREQSIZE,
3118 .bc_rep_max_size = LDLM_MAXREPSIZE,
3119 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3120 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3123 .tc_thr_name = "ldlm_cb",
3124 .tc_thr_factor = LDLM_THR_FACTOR,
3125 .tc_nthrs_init = LDLM_NTHRS_INIT,
3126 .tc_nthrs_base = LDLM_NTHRS_BASE,
3127 .tc_nthrs_max = LDLM_NTHRS_MAX,
3128 .tc_nthrs_user = ldlm_num_threads,
3129 .tc_cpu_bind = ldlm_cpu_bind,
3130 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3133 .cc_pattern = ldlm_cpts,
3134 .cc_affinity = true,
3137 .so_req_handler = ldlm_callback_handler,
3140 ldlm_state->ldlm_cb_service = \
3141 ptlrpc_register_service(&conf, ldlm_svc_kset,
3142 ldlm_svc_debugfs_dir);
3143 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3144 CERROR("failed to start service\n");
3145 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3146 ldlm_state->ldlm_cb_service = NULL;
3150 #ifdef HAVE_SERVER_SUPPORT
3151 memset(&conf, 0, sizeof(conf));
3152 conf = (typeof(conf)) {
3153 .psc_name = "ldlm_canceld",
3154 .psc_watchdog_factor = 6,
3156 .bc_nbufs = LDLM_SERVER_NBUFS,
3157 .bc_buf_size = LDLM_BUFSIZE,
3158 .bc_req_max_size = LDLM_MAXREQSIZE,
3159 .bc_rep_max_size = LDLM_MAXREPSIZE,
3160 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3161 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3165 .tc_thr_name = "ldlm_cn",
3166 .tc_thr_factor = LDLM_THR_FACTOR,
3167 .tc_nthrs_init = LDLM_NTHRS_INIT,
3168 .tc_nthrs_base = LDLM_NTHRS_BASE,
3169 .tc_nthrs_max = LDLM_NTHRS_MAX,
3170 .tc_nthrs_user = ldlm_num_threads,
3171 .tc_cpu_bind = ldlm_cpu_bind,
3172 .tc_ctx_tags = LCT_MD_THREAD | \
3177 .cc_pattern = ldlm_cpts,
3178 .cc_affinity = true,
3181 .so_req_handler = ldlm_cancel_handler,
3182 .so_hpreq_handler = ldlm_hpreq_handler,
3185 ldlm_state->ldlm_cancel_service = \
3186 ptlrpc_register_service(&conf, ldlm_svc_kset,
3187 ldlm_svc_debugfs_dir);
3188 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3189 CERROR("failed to start service\n");
3190 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3191 ldlm_state->ldlm_cancel_service = NULL;
3194 #endif /* HAVE_SERVER_SUPPORT */
3196 OBD_ALLOC(blp, sizeof(*blp));
3198 GOTO(out, rc = -ENOMEM);
3199 ldlm_state->ldlm_bl_pool = blp;
3201 spin_lock_init(&blp->blp_lock);
3202 INIT_LIST_HEAD(&blp->blp_list);
3203 INIT_LIST_HEAD(&blp->blp_prio_list);
3204 init_waitqueue_head(&blp->blp_waitq);
3205 atomic_set(&blp->blp_num_threads, 0);
3206 atomic_set(&blp->blp_busy_threads, 0);
3208 if (ldlm_num_threads == 0) {
3209 blp->blp_min_threads = LDLM_NTHRS_INIT;
3210 blp->blp_max_threads = LDLM_NTHRS_MAX;
3212 blp->blp_min_threads = blp->blp_max_threads = \
3213 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3217 for (i = 0; i < blp->blp_min_threads; i++) {
3218 rc = ldlm_bl_thread_start(blp, false);
3223 #ifdef HAVE_SERVER_SUPPORT
3224 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3227 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3231 wait_event(expired_lock_wait_queue,
3232 expired_lock_thread_state == ELT_READY);
3233 #endif /* HAVE_SERVER_SUPPORT */
3235 rc = ldlm_pools_init();
3237 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3241 rc = ldlm_reclaim_setup();
3243 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3253 static int ldlm_cleanup(void)
3257 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3258 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3259 CERROR("ldlm still has namespaces; clean these up first.\n");
3260 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3261 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3265 ldlm_reclaim_cleanup();
3268 if (ldlm_state->ldlm_bl_pool != NULL) {
3269 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3271 while (atomic_read(&blp->blp_num_threads) > 0) {
3272 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3274 init_completion(&blp->blp_comp);
3276 spin_lock(&blp->blp_lock);
3277 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3278 wake_up(&blp->blp_waitq);
3279 spin_unlock(&blp->blp_lock);
3281 wait_for_completion(&blp->blp_comp);
3284 OBD_FREE(blp, sizeof(*blp));
3287 if (ldlm_state->ldlm_cb_service != NULL)
3288 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3289 #ifdef HAVE_SERVER_SUPPORT
3290 if (ldlm_state->ldlm_cancel_service != NULL)
3291 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3295 kset_unregister(ldlm_ns_kset);
3297 kset_unregister(ldlm_svc_kset);
3299 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3300 kobject_put(ldlm_kobj);
3303 ldlm_debugfs_cleanup();
3305 #ifdef HAVE_SERVER_SUPPORT
3306 if (expired_lock_thread_state != ELT_STOPPED) {
3307 expired_lock_thread_state = ELT_TERMINATE;
3308 wake_up(&expired_lock_wait_queue);
3309 wait_event(expired_lock_wait_queue,
3310 expired_lock_thread_state == ELT_STOPPED);
3314 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3322 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3323 sizeof(struct ldlm_resource), 0,
3324 SLAB_HWCACHE_ALIGN, NULL);
3325 if (ldlm_resource_slab == NULL)
3328 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3329 sizeof(struct ldlm_lock), 0,
3330 SLAB_HWCACHE_ALIGN, NULL);
3331 if (ldlm_lock_slab == NULL)
3334 ldlm_interval_slab = kmem_cache_create("interval_node",
3335 sizeof(struct ldlm_interval),
3336 0, SLAB_HWCACHE_ALIGN, NULL);
3337 if (ldlm_interval_slab == NULL)
3340 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3341 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3342 0, SLAB_HWCACHE_ALIGN, NULL);
3343 if (ldlm_interval_tree_slab == NULL)
3346 #ifdef HAVE_SERVER_SUPPORT
3347 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3348 sizeof(struct ldlm_glimpse_work),
3350 if (ldlm_glimpse_work_kmem == NULL)
3351 goto out_interval_tree;
3354 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3355 class_export_dump_hook = ldlm_dump_export_locks;
3358 #ifdef HAVE_SERVER_SUPPORT
3360 kmem_cache_destroy(ldlm_interval_tree_slab);
3363 kmem_cache_destroy(ldlm_interval_slab);
3365 kmem_cache_destroy(ldlm_lock_slab);
3367 kmem_cache_destroy(ldlm_resource_slab);
3372 void ldlm_exit(void)
3375 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3376 kmem_cache_destroy(ldlm_resource_slab);
3378 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3379 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3380 * so that ldlm_lock_free() get a chance to be called.
3383 kmem_cache_destroy(ldlm_lock_slab);
3384 kmem_cache_destroy(ldlm_interval_slab);
3385 kmem_cache_destroy(ldlm_interval_tree_slab);
3386 #ifdef HAVE_SERVER_SUPPORT
3387 kmem_cache_destroy(ldlm_glimpse_work_kmem);