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;
72 /* timeout for initial callback (AST) reply (bz10399)
73 * Due to having to send a 32 bit time value over the
74 * wire return it as time_t instead of time64_t
76 static inline time_t ldlm_get_rq_timeout(void)
79 time_t timeout = min(ldlm_timeout, obd_timeout / 3);
81 return timeout < 1 ? 1 : timeout;
88 * blp_prio_list is used for callbacks that should be handled
89 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
92 struct list_head blp_prio_list;
95 * blp_list is used for all other callbacks which are likely
96 * to take longer to process.
98 struct list_head blp_list;
100 wait_queue_head_t blp_waitq;
101 struct completion blp_comp;
102 atomic_t blp_num_threads;
103 atomic_t blp_busy_threads;
108 struct ldlm_bl_work_item {
109 struct list_head blwi_entry;
110 struct ldlm_namespace *blwi_ns;
111 struct ldlm_lock_desc blwi_ld;
112 struct ldlm_lock *blwi_lock;
113 struct list_head blwi_head;
115 struct completion blwi_comp;
116 enum ldlm_cancel_flags blwi_flags;
117 int blwi_mem_pressure;
120 #ifdef HAVE_SERVER_SUPPORT
123 * Protects both waiting_locks_list and expired_lock_thread.
125 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
128 * List for contended locks.
130 * As soon as a lock is contended, it gets placed on this list and
131 * expected time to get a response is filled in the lock. A special
132 * thread walks the list looking for locks that should be released and
133 * schedules client evictions for those that have not been released in
136 * All access to it should be under waiting_locks_spinlock.
138 static LIST_HEAD(waiting_locks_list);
139 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
140 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
148 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
149 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
150 static int expired_lock_dump;
151 static LIST_HEAD(expired_lock_list);
153 static int ldlm_lock_busy(struct ldlm_lock *lock);
154 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout);
155 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds);
157 static inline int have_expired_locks(void)
162 spin_lock_bh(&waiting_locks_spinlock);
163 need_to_run = !list_empty(&expired_lock_list);
164 spin_unlock_bh(&waiting_locks_spinlock);
170 * Check expired lock list for expired locks and time them out.
172 static int expired_lock_main(void *arg)
174 struct list_head *expired = &expired_lock_list;
175 struct l_wait_info lwi = { 0 };
180 expired_lock_thread_state = ELT_READY;
181 wake_up(&expired_lock_wait_queue);
184 l_wait_event(expired_lock_wait_queue,
185 have_expired_locks() ||
186 expired_lock_thread_state == ELT_TERMINATE,
189 spin_lock_bh(&waiting_locks_spinlock);
190 if (expired_lock_dump) {
191 spin_unlock_bh(&waiting_locks_spinlock);
193 /* from waiting_locks_callback, but not in timer */
194 libcfs_debug_dumplog();
196 spin_lock_bh(&waiting_locks_spinlock);
197 expired_lock_dump = 0;
202 while (!list_empty(expired)) {
203 struct obd_export *export;
204 struct ldlm_lock *lock;
206 lock = list_entry(expired->next, struct ldlm_lock,
208 if ((void *)lock < LP_POISON + PAGE_SIZE &&
209 (void *)lock >= LP_POISON) {
210 spin_unlock_bh(&waiting_locks_spinlock);
211 CERROR("free lock on elt list %p\n", lock);
214 list_del_init(&lock->l_pending_chain);
215 if ((void *)lock->l_export <
216 LP_POISON + PAGE_SIZE &&
217 (void *)lock->l_export >= LP_POISON) {
218 CERROR("lock with free export on elt list %p\n",
220 lock->l_export = NULL;
221 LDLM_ERROR(lock, "free export");
222 /* release extra ref grabbed by
223 * ldlm_add_waiting_lock() or
224 * ldlm_failed_ast() */
225 LDLM_LOCK_RELEASE(lock);
229 if (ldlm_is_destroyed(lock)) {
230 /* release the lock refcount where
231 * waiting_locks_callback() founds */
232 LDLM_LOCK_RELEASE(lock);
235 export = class_export_lock_get(lock->l_export, lock);
236 spin_unlock_bh(&waiting_locks_spinlock);
238 /* Check if we need to prolong timeout */
239 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
240 lock->l_callback_timeout != 0 && /* not AST error */
241 ldlm_lock_busy(lock)) {
242 LDLM_DEBUG(lock, "prolong the busy lock");
243 lock_res_and_lock(lock);
244 ldlm_add_waiting_lock(lock,
245 ldlm_bl_timeout(lock) >> 1);
246 unlock_res_and_lock(lock);
248 spin_lock_bh(&export->exp_bl_list_lock);
249 list_del_init(&lock->l_exp_list);
250 spin_unlock_bh(&export->exp_bl_list_lock);
253 "lock callback timer expired after %llds: evicting client at %s ",
254 ktime_get_real_seconds() -
256 obd_export_nid2str(export));
257 ldlm_lock_to_ns(lock)->ns_timeouts++;
259 class_fail_export(export);
261 class_export_lock_put(export, lock);
262 /* release extra ref grabbed by ldlm_add_waiting_lock()
263 * or ldlm_failed_ast() */
264 LDLM_LOCK_RELEASE(lock);
266 spin_lock_bh(&waiting_locks_spinlock);
268 spin_unlock_bh(&waiting_locks_spinlock);
270 if (do_dump && obd_dump_on_eviction) {
271 CERROR("dump the log upon eviction\n");
272 libcfs_debug_dumplog();
275 if (expired_lock_thread_state == ELT_TERMINATE)
279 expired_lock_thread_state = ELT_STOPPED;
280 wake_up(&expired_lock_wait_queue);
285 * Check if there is a request in the export request list
286 * which prevents the lock canceling.
288 static int ldlm_lock_busy(struct ldlm_lock *lock)
290 struct ptlrpc_request *req;
294 if (lock->l_export == NULL)
297 spin_lock(&lock->l_export->exp_rpc_lock);
298 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
300 if (req->rq_ops->hpreq_lock_match) {
301 match = req->rq_ops->hpreq_lock_match(req, lock);
306 spin_unlock(&lock->l_export->exp_rpc_lock);
310 /* This is called from within a timer interrupt and cannot schedule */
311 static void waiting_locks_callback(TIMER_DATA_TYPE unused)
313 struct ldlm_lock *lock;
316 spin_lock_bh(&waiting_locks_spinlock);
317 while (!list_empty(&waiting_locks_list)) {
318 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
320 if (lock->l_callback_timeout > ktime_get_seconds() ||
321 lock->l_req_mode == LCK_GROUP)
324 /* no needs to take an extra ref on the lock since it was in
325 * the waiting_locks_list and ldlm_add_waiting_lock()
326 * already grabbed a ref */
327 list_del(&lock->l_pending_chain);
328 list_add(&lock->l_pending_chain, &expired_lock_list);
332 if (!list_empty(&expired_lock_list)) {
333 if (obd_dump_on_timeout && need_dump)
334 expired_lock_dump = __LINE__;
336 wake_up(&expired_lock_wait_queue);
340 * Make sure the timer will fire again if we have any locks
343 if (!list_empty(&waiting_locks_list)) {
344 unsigned long timeout_jiffies;
346 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
348 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
349 mod_timer(&waiting_locks_timer, timeout_jiffies);
351 spin_unlock_bh(&waiting_locks_spinlock);
355 * Add lock to the list of contended locks.
357 * Indicate that we're waiting for a client to call us back cancelling a given
358 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
359 * timer to fire appropriately. (We round up to the next second, to avoid
360 * floods of timer firings during periods of high lock contention and traffic).
361 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
362 * if it has been added to the waiting list (1 is returned).
364 * Called with the namespace lock held.
366 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds)
368 unsigned long timeout_jiffies;
371 if (!list_empty(&lock->l_pending_chain))
374 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
375 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
378 timeout = ktime_get_seconds() + seconds;
379 if (likely(timeout > lock->l_callback_timeout))
380 lock->l_callback_timeout = timeout;
382 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
384 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
385 !timer_pending(&waiting_locks_timer))
386 mod_timer(&waiting_locks_timer, timeout_jiffies);
388 /* if the new lock has a shorter timeout than something earlier on
389 * the list, we'll wait the longer amount of time; no big deal.
392 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
396 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
398 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
399 if (list_empty(&lock->l_exp_list)) {
400 if (lock->l_granted_mode != lock->l_req_mode)
401 list_add_tail(&lock->l_exp_list,
402 &lock->l_export->exp_bl_list);
404 list_add(&lock->l_exp_list,
405 &lock->l_export->exp_bl_list);
407 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
409 /* A blocked lock is added. Adjust the position in
410 * the stale list if the export is in the list.
411 * If export is stale and not in the list - it is being
412 * processed and will be placed on the right position
413 * on obd_stale_export_put(). */
414 if (!list_empty(&lock->l_export->exp_stale_list))
415 obd_stale_export_adjust(lock->l_export);
418 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
422 /* NB: must be called with hold of lock_res_and_lock() */
423 LASSERT(ldlm_is_res_locked(lock));
424 LASSERT(!ldlm_is_cancel_on_block(lock));
426 /* Do not put cross-MDT lock in the waiting list, since we
427 * will not evict it due to timeout for now */
428 if (lock->l_export != NULL &&
429 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
432 spin_lock_bh(&waiting_locks_spinlock);
433 if (ldlm_is_cancel(lock)) {
434 spin_unlock_bh(&waiting_locks_spinlock);
438 if (ldlm_is_destroyed(lock)) {
439 static time64_t next;
441 spin_unlock_bh(&waiting_locks_spinlock);
442 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
443 if (ktime_get_seconds() > next) {
444 next = ktime_get_seconds() + 14400;
445 libcfs_debug_dumpstack(NULL);
450 ldlm_set_waited(lock);
451 lock->l_blast_sent = ktime_get_real_seconds();
452 ret = __ldlm_add_waiting_lock(lock, timeout);
454 /* grab ref on the lock if it has been added to the
458 spin_unlock_bh(&waiting_locks_spinlock);
461 ldlm_add_blocked_lock(lock);
463 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)",
464 ret == 0 ? "not re-" : "", timeout,
465 AT_OFF ? "off" : "on");
470 * Remove a lock from the pending list, likely because it had its cancellation
471 * callback arrive without incident. This adjusts the lock-timeout timer if
472 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
473 * As done by ldlm_del_waiting_lock(), the caller must release the lock
474 * reference when the lock is removed from any list (1 is returned).
476 * Called with namespace lock held.
478 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
480 struct list_head *list_next;
482 if (list_empty(&lock->l_pending_chain))
485 list_next = lock->l_pending_chain.next;
486 if (lock->l_pending_chain.prev == &waiting_locks_list) {
487 /* Removing the head of the list, adjust timer. */
488 if (list_next == &waiting_locks_list) {
489 /* No more, just cancel. */
490 del_timer(&waiting_locks_timer);
492 struct ldlm_lock *next;
494 next = list_entry(list_next, struct ldlm_lock,
496 mod_timer(&waiting_locks_timer,
497 cfs_time_seconds(next->l_callback_timeout));
500 list_del_init(&lock->l_pending_chain);
505 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
509 if (lock->l_export == NULL) {
510 /* We don't have a "waiting locks list" on clients. */
511 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
515 spin_lock_bh(&waiting_locks_spinlock);
516 ret = __ldlm_del_waiting_lock(lock);
517 ldlm_clear_waited(lock);
518 spin_unlock_bh(&waiting_locks_spinlock);
520 /* remove the lock out of export blocking list */
521 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
522 list_del_init(&lock->l_exp_list);
523 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
526 /* release lock ref if it has indeed been removed
528 LDLM_LOCK_RELEASE(lock);
531 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
536 * Prolong the contended lock waiting time.
538 * Called with namespace lock held.
540 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
542 if (lock->l_export == NULL) {
543 /* We don't have a "waiting locks list" on clients. */
544 LDLM_DEBUG(lock, "client lock: no-op");
548 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
549 /* We don't have a "waiting locks list" on OSP. */
550 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
554 spin_lock_bh(&waiting_locks_spinlock);
556 if (list_empty(&lock->l_pending_chain)) {
557 spin_unlock_bh(&waiting_locks_spinlock);
558 LDLM_DEBUG(lock, "wasn't waiting");
562 /* we remove/add the lock to the waiting list, so no needs to
563 * release/take a lock reference */
564 __ldlm_del_waiting_lock(lock);
565 __ldlm_add_waiting_lock(lock, timeout);
566 spin_unlock_bh(&waiting_locks_spinlock);
568 LDLM_DEBUG(lock, "refreshed");
571 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
573 #else /* HAVE_SERVER_SUPPORT */
575 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
580 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
585 #endif /* !HAVE_SERVER_SUPPORT */
587 #ifdef HAVE_SERVER_SUPPORT
590 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
591 * lock cancel, and their replies). Used for lock callback timeout and AST
594 * \param[in] lock lock which is getting the blocking callback
596 * \retval timeout in seconds to wait for the client reply
598 time64_t ldlm_bl_timeout(struct ldlm_lock *lock)
603 return obd_timeout / 2;
605 /* Since these are non-updating timeouts, we should be conservative.
606 * Take more than usually, 150%
607 * It would be nice to have some kind of "early reply" mechanism for
608 * lock callbacks too... */
609 timeout = at_get(&lock->l_export->exp_bl_lock_at);
610 return max(timeout + (timeout >> 1), (time64_t)ldlm_enqueue_min);
612 EXPORT_SYMBOL(ldlm_bl_timeout);
615 * Perform lock cleanup if AST sending failed.
617 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
618 const char *ast_type)
620 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
621 "to a lock %s callback time out: rc %d\n",
622 lock->l_export->exp_obd->obd_name,
623 obd_export_nid2str(lock->l_export), ast_type, rc);
625 if (obd_dump_on_timeout)
626 libcfs_debug_dumplog();
627 spin_lock_bh(&waiting_locks_spinlock);
628 if (__ldlm_del_waiting_lock(lock) == 0)
629 /* the lock was not in any list, grab an extra ref before adding
630 * the lock to the expired list */
632 lock->l_callback_timeout = 0; /* differentiate it from expired locks */
633 list_add(&lock->l_pending_chain, &expired_lock_list);
634 wake_up(&expired_lock_wait_queue);
635 spin_unlock_bh(&waiting_locks_spinlock);
639 * Perform lock cleanup if AST reply came with error.
641 static int ldlm_handle_ast_error(const struct lu_env *env,
642 struct ldlm_lock *lock,
643 struct ptlrpc_request *req, int rc,
644 const char *ast_type)
646 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
648 if (!req->rq_replied || (rc && rc != -EINVAL)) {
649 if (ldlm_is_cancel(lock)) {
651 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
652 ast_type, req, req->rq_xid,
653 libcfs_nid2str(peer.nid));
654 ldlm_lock_cancel(lock);
656 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
658 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
659 /* Upon umount process the AST fails because cannot be
660 * sent. This shouldn't lead to the client eviction.
661 * -ENODEV error is returned by ptl_send_rpc() for
662 * new request in such import.
663 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
664 * if imp_invalid is set or obd_no_recov.
665 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
666 * in ptlrpc_import_delay_req() as well with -EIO code.
667 * In all such cases errors are ignored.
669 LDLM_DEBUG(lock, "%s AST can't be sent due to a server"
670 " %s failure or umount process: rc = %d\n",
672 req->rq_import->imp_obd->obd_name, rc);
675 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
676 libcfs_nid2str(peer.nid),
677 req->rq_replied ? "returned error from" :
678 "failed to reply to",
679 ast_type, req, req->rq_xid,
680 (req->rq_repmsg != NULL) ?
681 lustre_msg_get_status(req->rq_repmsg) : 0,
683 ldlm_failed_ast(lock, rc, ast_type);
689 struct ldlm_resource *res = lock->l_resource;
692 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
693 libcfs_nid2str(peer.nid),
695 lustre_msg_get_status(req->rq_repmsg) : -1,
696 ast_type, req, req->rq_xid);
698 /* update lvbo to return proper attributes.
700 ldlm_resource_getref(res);
701 ldlm_lvbo_update(env, res, lock, NULL, 1);
702 ldlm_resource_putref(res);
704 ldlm_lock_cancel(lock);
711 static int ldlm_cb_interpret(const struct lu_env *env,
712 struct ptlrpc_request *req, void *args, int rc)
714 struct ldlm_cb_async_args *ca = args;
715 struct ldlm_lock *lock = ca->ca_lock;
716 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
720 LASSERT(lock != NULL);
723 case LDLM_GL_CALLBACK:
724 /* Update the LVB from disk if the AST failed
725 * (this is a legal race)
727 * - Glimpse callback of local lock just returns
728 * -ELDLM_NO_LOCK_DATA.
729 * - Glimpse callback of remote lock might return
730 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
732 if (unlikely(arg->gl_interpret_reply)) {
733 rc = arg->gl_interpret_reply(env, req, args, rc);
734 } else if (rc == -ELDLM_NO_LOCK_DATA) {
735 LDLM_DEBUG(lock, "lost race - client has a lock but no "
737 ldlm_lvbo_update(env, lock->l_resource, lock, NULL, 1);
738 } else if (rc != 0) {
739 rc = ldlm_handle_ast_error(env, lock, req,
742 rc = ldlm_lvbo_update(env, lock->l_resource,
746 case LDLM_BL_CALLBACK:
748 rc = ldlm_handle_ast_error(env, lock, req,
751 case LDLM_CP_CALLBACK:
753 rc = ldlm_handle_ast_error(env, lock, req,
757 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
762 /* release extra reference taken in ldlm_ast_fini() */
763 LDLM_LOCK_RELEASE(lock);
766 atomic_inc(&arg->restart);
771 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
773 struct ldlm_cb_async_args *ca = data;
774 struct ldlm_lock *lock = ca->ca_lock;
776 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
779 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
780 struct ldlm_cb_set_arg *arg,
781 struct ldlm_lock *lock,
787 if (unlikely(instant_cancel)) {
788 rc = ptl_send_rpc(req, 1);
789 ptlrpc_req_finished(req);
791 atomic_inc(&arg->restart);
794 ptlrpc_set_add_req(arg->set, req);
801 * Check if there are requests in the export request list which prevent
802 * the lock canceling and make these requests high priority ones.
804 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
806 struct ptlrpc_request *req;
809 if (lock->l_export == NULL) {
810 LDLM_DEBUG(lock, "client lock: no-op");
814 spin_lock(&lock->l_export->exp_rpc_lock);
815 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
817 /* Do not process requests that were not yet added to there
818 * incoming queue or were already removed from there for
819 * processing. We evaluate ptlrpc_nrs_req_can_move() without
820 * holding svcpt->scp_req_lock, and then redo the check with
821 * the lock held once we need to obtain a reliable result.
823 if (ptlrpc_nrs_req_can_move(req) &&
824 req->rq_ops->hpreq_lock_match &&
825 req->rq_ops->hpreq_lock_match(req, lock))
826 ptlrpc_nrs_req_hp_move(req);
828 spin_unlock(&lock->l_export->exp_rpc_lock);
833 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
834 * enqueued server lock conflicts with given one.
836 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
837 * to wait for client response.
839 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
840 struct ldlm_lock_desc *desc,
841 void *data, int flag)
843 struct ldlm_cb_async_args *ca;
844 struct ldlm_cb_set_arg *arg = data;
845 struct ldlm_request *body;
846 struct ptlrpc_request *req;
847 int instant_cancel = 0;
851 if (flag == LDLM_CB_CANCELING)
852 /* Don't need to do anything here. */
855 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
856 LDLM_DEBUG(lock, "dropping BL AST");
861 LASSERT(data != NULL);
862 if (lock->l_export->exp_obd->obd_recovering != 0)
863 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
865 ldlm_lock_reorder_req(lock);
867 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
868 &RQF_LDLM_BL_CALLBACK,
869 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
873 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
874 ca = ptlrpc_req_async_args(req);
875 ca->ca_set_arg = arg;
878 req->rq_interpret_reply = ldlm_cb_interpret;
880 lock_res_and_lock(lock);
881 if (ldlm_is_destroyed(lock)) {
882 /* What's the point? */
883 unlock_res_and_lock(lock);
884 ptlrpc_req_finished(req);
888 if (lock->l_granted_mode != lock->l_req_mode) {
889 /* this blocking AST will be communicated as part of the
890 * completion AST instead */
891 ldlm_add_blocked_lock(lock);
892 ldlm_set_waited(lock);
893 unlock_res_and_lock(lock);
895 ptlrpc_req_finished(req);
896 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
900 if (ldlm_is_cancel_on_block(lock))
903 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
904 body->lock_handle[0] = lock->l_remote_handle;
905 body->lock_desc = *desc;
906 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
908 LDLM_DEBUG(lock, "server preparing blocking AST");
910 ptlrpc_request_set_replen(req);
911 ldlm_set_cbpending(lock);
912 if (instant_cancel) {
913 unlock_res_and_lock(lock);
914 ldlm_lock_cancel(lock);
916 req->rq_no_resend = 1;
918 LASSERT(lock->l_granted_mode == lock->l_req_mode);
919 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
920 unlock_res_and_lock(lock);
922 /* Do not resend after lock callback timeout */
923 req->rq_delay_limit = ldlm_bl_timeout(lock);
924 req->rq_resend_cb = ldlm_update_resend;
927 req->rq_send_state = LUSTRE_IMP_FULL;
928 /* ptlrpc_request_alloc_pack already set timeout */
930 req->rq_timeout = ldlm_get_rq_timeout();
932 if (lock->l_export && lock->l_export->exp_nid_stats &&
933 lock->l_export->exp_nid_stats->nid_ldlm_stats)
934 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
935 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
937 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
943 * ->l_completion_ast callback for a remote lock in server namespace.
945 * Sends AST to the client notifying it of lock granting. If initial
946 * lock response was not sent yet, instead of sending another RPC, just
947 * mark the lock as granted and client will understand
949 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
951 struct ldlm_cb_set_arg *arg = data;
952 struct ldlm_request *body;
953 struct ptlrpc_request *req;
954 struct ldlm_cb_async_args *ca;
955 int instant_cancel = 0;
960 LASSERT(lock != NULL);
961 LASSERT(data != NULL);
963 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
964 LDLM_DEBUG(lock, "dropping CP AST");
968 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
969 &RQF_LDLM_CP_CALLBACK);
973 /* server namespace, doesn't need lock */
974 lvb_len = ldlm_lvbo_size(lock);
975 /* LU-3124 & LU-2187: to not return layout in completion AST because
976 * it may deadlock for LU-2187, or client may not have enough space
977 * for large layout. The layout will be returned to client with an
978 * extra RPC to fetch xattr.lov */
979 if (ldlm_has_layout(lock))
982 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
983 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
985 ptlrpc_request_free(req);
989 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
990 ca = ptlrpc_req_async_args(req);
991 ca->ca_set_arg = arg;
994 req->rq_interpret_reply = ldlm_cb_interpret;
995 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
997 body->lock_handle[0] = lock->l_remote_handle;
998 body->lock_flags = ldlm_flags_to_wire(flags);
999 ldlm_lock2desc(lock, &body->lock_desc);
1001 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1002 const struct lu_env *env = NULL;
1004 if (req->rq_svc_thread)
1005 env = req->rq_svc_thread->t_env;
1007 lvb_len = ldlm_lvbo_fill(env, lock, lvb, &lvb_len);
1009 /* We still need to send the RPC to wake up the blocked
1010 * enqueue thread on the client.
1012 * Consider old client, there is no better way to notify
1013 * the failure, just zero-sized the LVB, then the client
1014 * will fail out as "-EPROTO". */
1015 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1019 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1024 LDLM_DEBUG(lock, "server preparing completion AST");
1026 ptlrpc_request_set_replen(req);
1028 req->rq_send_state = LUSTRE_IMP_FULL;
1029 /* ptlrpc_request_pack already set timeout */
1031 req->rq_timeout = ldlm_get_rq_timeout();
1033 /* We only send real blocking ASTs after the lock is granted */
1034 lock_res_and_lock(lock);
1035 if (ldlm_is_ast_sent(lock)) {
1036 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1037 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1038 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1041 /* We might get here prior to ldlm_handle_enqueue setting
1042 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1043 * into waiting list, but this is safe and similar code in
1044 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1045 * that would not only cancel the lock, but will also remove
1046 * it from waiting list */
1047 if (ldlm_is_cancel_on_block(lock)) {
1048 unlock_res_and_lock(lock);
1049 ldlm_lock_cancel(lock);
1052 req->rq_no_resend = 1;
1054 lock_res_and_lock(lock);
1056 /* start the lock-timeout clock */
1057 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1058 /* Do not resend after lock callback timeout */
1059 req->rq_delay_limit = ldlm_bl_timeout(lock);
1060 req->rq_resend_cb = ldlm_update_resend;
1063 unlock_res_and_lock(lock);
1065 if (lock->l_export && lock->l_export->exp_nid_stats &&
1066 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1067 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1068 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1070 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1072 RETURN(lvb_len < 0 ? lvb_len : rc);
1076 * Server side ->l_glimpse_ast handler for client locks.
1078 * Sends glimpse AST to the client and waits for reply. Then updates
1079 * lvbo with the result.
1081 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1083 struct ldlm_cb_set_arg *arg = data;
1084 struct ldlm_request *body;
1085 struct ptlrpc_request *req;
1086 struct ldlm_cb_async_args *ca;
1088 struct req_format *req_fmt;
1091 LASSERT(lock != NULL);
1093 if (arg->gl_desc != NULL)
1094 /* There is a glimpse descriptor to pack */
1095 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1097 req_fmt = &RQF_LDLM_GL_CALLBACK;
1099 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1100 req_fmt, LUSTRE_DLM_VERSION,
1106 if (arg->gl_desc != NULL) {
1107 /* copy the GL descriptor */
1108 union ldlm_gl_desc *desc;
1109 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1110 *desc = *arg->gl_desc;
1113 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1114 body->lock_handle[0] = lock->l_remote_handle;
1115 ldlm_lock2desc(lock, &body->lock_desc);
1117 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1118 ca = ptlrpc_req_async_args(req);
1119 ca->ca_set_arg = arg;
1122 /* server namespace, doesn't need lock */
1123 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1124 ldlm_lvbo_size(lock));
1125 ptlrpc_request_set_replen(req);
1127 req->rq_send_state = LUSTRE_IMP_FULL;
1128 /* ptlrpc_request_alloc_pack already set timeout */
1130 req->rq_timeout = ldlm_get_rq_timeout();
1132 req->rq_interpret_reply = ldlm_cb_interpret;
1134 if (lock->l_export && lock->l_export->exp_nid_stats) {
1135 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1137 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1138 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1141 rc = ldlm_ast_fini(req, arg, lock, 0);
1145 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1147 int ldlm_glimpse_locks(struct ldlm_resource *res,
1148 struct list_head *gl_work_list)
1153 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1155 if (rc == -ERESTART)
1156 ldlm_reprocess_all(res);
1160 EXPORT_SYMBOL(ldlm_glimpse_locks);
1162 /* return LDLM lock associated with a lock callback request */
1163 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1165 struct ldlm_cb_async_args *ca;
1166 struct ldlm_lock *lock;
1169 ca = ptlrpc_req_async_args(req);
1172 RETURN(ERR_PTR(-EFAULT));
1176 EXPORT_SYMBOL(ldlm_request_lock);
1179 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1180 * service threads to carry out client lock enqueueing requests.
1182 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1183 struct ptlrpc_request *req,
1184 const struct ldlm_request *dlm_req,
1185 const struct ldlm_callback_suite *cbs)
1187 struct ldlm_reply *dlm_rep;
1189 enum ldlm_error err = ELDLM_OK;
1190 struct ldlm_lock *lock = NULL;
1191 void *cookie = NULL;
1193 struct ldlm_resource *res = NULL;
1194 const struct lu_env *env = req->rq_svc_thread->t_env;
1197 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1199 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1200 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1202 LASSERT(req->rq_export);
1204 /* for intent enqueue the stat will be updated inside intent policy */
1205 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1206 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1207 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1209 if (req->rq_export && req->rq_export->exp_nid_stats &&
1210 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1211 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1212 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1214 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1215 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1216 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1217 dlm_req->lock_desc.l_resource.lr_type);
1218 GOTO(out, rc = -EFAULT);
1221 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1222 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1223 dlm_req->lock_desc.l_req_mode &
1224 (dlm_req->lock_desc.l_req_mode-1))) {
1225 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1226 dlm_req->lock_desc.l_req_mode);
1227 GOTO(out, rc = -EFAULT);
1230 if (unlikely((flags & LDLM_FL_REPLAY) ||
1231 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1232 /* Find an existing lock in the per-export lock hash */
1233 /* In the function below, .hs_keycmp resolves to
1234 * ldlm_export_lock_keycmp() */
1235 /* coverity[overrun-buffer-val] */
1236 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1237 (void *)&dlm_req->lock_handle[0]);
1239 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie %#llx",
1240 lock->l_handle.h_cookie);
1241 flags |= LDLM_FL_RESENT;
1242 GOTO(existing_lock, rc = 0);
1245 if (ldlm_reclaim_full()) {
1246 DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1247 "reject current enqueue request and let the "
1248 "client retry later.\n");
1249 GOTO(out, rc = -EINPROGRESS);
1253 /* The lock's callback data might be set in the policy function */
1254 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1255 dlm_req->lock_desc.l_resource.lr_type,
1256 dlm_req->lock_desc.l_req_mode,
1257 cbs, NULL, 0, LVB_T_NONE);
1264 lock->l_remote_handle = dlm_req->lock_handle[0];
1265 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1267 /* Initialize resource lvb but not for a lock being replayed since
1268 * Client already got lvb sent in this case.
1269 * This must occur early since some policy methods assume resource
1270 * lvb is available (lr_lvb_data != NULL).
1272 res = lock->l_resource;
1273 if (!(flags & LDLM_FL_REPLAY)) {
1274 /* non-replayed lock, delayed lvb init may need to be done */
1275 rc = ldlm_lvbo_init(env, res);
1277 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1282 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1283 /* Don't enqueue a lock onto the export if it is been disonnected
1284 * due to eviction (bug 3822) or server umount (bug 24324).
1285 * Cancel it now instead. */
1286 if (req->rq_export->exp_disconnected) {
1287 LDLM_ERROR(lock, "lock on disconnected export %p",
1289 GOTO(out, rc = -ENOTCONN);
1292 lock->l_export = class_export_lock_get(req->rq_export, lock);
1293 if (lock->l_export->exp_lock_hash)
1294 cfs_hash_add(lock->l_export->exp_lock_hash,
1295 &lock->l_remote_handle,
1298 /* Inherit the enqueue flags before the operation, because we do not
1299 * keep the res lock on return and next operations (BL AST) may proceed
1301 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1302 LDLM_FL_INHERIT_MASK);
1304 ldlm_convert_policy_to_local(req->rq_export,
1305 dlm_req->lock_desc.l_resource.lr_type,
1306 &dlm_req->lock_desc.l_policy_data,
1307 &lock->l_policy_data);
1308 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1309 lock->l_req_extent = lock->l_policy_data.l_extent;
1312 if (flags & LDLM_FL_HAS_INTENT) {
1313 /* In this case, the reply buffer is allocated deep in
1314 * local_lock_enqueue by the policy function. */
1317 /* based on the assumption that lvb size never changes during
1318 * resource life time otherwise it need resource->lr_lock's
1320 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1321 RCL_SERVER, ldlm_lvbo_size(lock));
1323 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1324 GOTO(out, rc = -ENOMEM);
1326 rc = req_capsule_server_pack(&req->rq_pill);
1331 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1338 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1340 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1341 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1343 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1344 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1346 /* We never send a blocking AST until the lock is granted, but
1347 * we can tell it right now */
1348 lock_res_and_lock(lock);
1350 /* Now take into account flags to be inherited from original lock
1351 request both in reply to client and in our own lock flags. */
1352 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1353 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1355 /* Don't move a pending lock onto the export if it has already been
1356 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1357 * Cancel it now instead. */
1358 if (unlikely(req->rq_export->exp_disconnected ||
1359 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1360 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1362 } else if (ldlm_is_ast_sent(lock)) {
1363 /* fill lock desc for possible lock convert */
1364 if (lock->l_blocking_lock &&
1365 lock->l_resource->lr_type == LDLM_IBITS) {
1366 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1367 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1370 "save blocking bits %llx in granted lock",
1371 bl_lock->l_policy_data.l_inodebits.bits);
1372 /* If lock is blocked then save blocking ibits
1373 * in returned lock policy for the possible lock
1374 * convert on a client.
1376 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1377 bl_lock->l_policy_data.l_inodebits.bits;
1379 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1380 if (lock->l_granted_mode == lock->l_req_mode) {
1382 * Only cancel lock if it was granted, because it would
1383 * be destroyed immediately and would never be granted
1384 * in the future, causing timeouts on client. Not
1385 * granted lock will be cancelled immediately after
1386 * sending completion AST.
1388 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1389 unlock_res_and_lock(lock);
1390 ldlm_lock_cancel(lock);
1391 lock_res_and_lock(lock);
1393 ldlm_add_waiting_lock(lock,
1394 ldlm_bl_timeout(lock));
1398 unlock_res_and_lock(lock);
1402 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1403 if (!req->rq_packed_final) {
1404 err = lustre_pack_reply(req, 1, NULL, NULL);
1409 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1410 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1412 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1413 "(err=%d, rc=%d)", err, rc);
1416 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1418 ldlm_lvbo_size(lock) > 0) {
1423 buf = req_capsule_server_get(&req->rq_pill,
1425 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1426 buflen = req_capsule_get_size(&req->rq_pill,
1427 &RMF_DLM_LVB, RCL_SERVER);
1428 /* non-replayed lock, delayed lvb init may
1429 * need to be occur now
1431 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1434 rc2 = ldlm_lvbo_fill(env, lock, buf, &buflen);
1436 req_capsule_shrink(&req->rq_pill,
1439 } else if (rc2 == -ERANGE) {
1440 rc2 = req_capsule_server_grow(
1442 &RMF_DLM_LVB, buflen);
1446 /* if we can't grow the buffer,
1447 * it's ok to return empty lvb
1458 } else if (flags & LDLM_FL_REPLAY) {
1459 /* no LVB resend upon replay */
1461 req_capsule_shrink(&req->rq_pill,
1471 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1472 if (lock->l_export) {
1473 ldlm_lock_cancel(lock);
1475 lock_res_and_lock(lock);
1476 ldlm_resource_unlink_lock(lock);
1477 ldlm_lock_destroy_nolock(lock);
1478 unlock_res_and_lock(lock);
1483 if (!err && !ldlm_is_cbpending(lock) &&
1484 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1485 ldlm_reprocess_all(lock->l_resource);
1487 LDLM_LOCK_RELEASE(lock);
1490 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1496 /* Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1497 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1499 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1501 if (lock->l_blocking_lock) {
1502 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1503 lock->l_blocking_lock = NULL;
1507 /* A lock can be converted to new ibits or mode and should be considered
1508 * as new lock. Clear all states related to a previous blocking AST
1509 * processing so new conflicts will cause new blocking ASTs.
1511 * This is used during lock convert below and lock downgrade to COS mode in
1512 * ldlm_lock_mode_downgrade().
1514 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1516 ldlm_clear_ast_sent(lock);
1517 lock->l_bl_ast_run = 0;
1518 ldlm_clear_blocking_lock(lock);
1522 * Main LDLM entry point for server code to process lock conversion requests.
1524 int ldlm_handle_convert0(struct ptlrpc_request *req,
1525 const struct ldlm_request *dlm_req)
1527 struct obd_export *exp = req->rq_export;
1528 struct ldlm_reply *dlm_rep;
1529 struct ldlm_lock *lock;
1534 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1535 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1536 LDLM_CONVERT - LDLM_FIRST_OPC);
1538 rc = req_capsule_server_pack(&req->rq_pill);
1542 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1543 dlm_rep->lock_flags = dlm_req->lock_flags;
1545 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1550 bits = lock->l_policy_data.l_inodebits.bits;
1551 new = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1552 LDLM_DEBUG(lock, "server-side convert handler START");
1554 if (ldlm_is_cancel(lock)) {
1555 LDLM_ERROR(lock, "convert on canceled lock!");
1556 rc = ELDLM_NO_LOCK_DATA;
1557 } else if (dlm_req->lock_desc.l_req_mode !=
1558 lock->l_granted_mode) {
1559 LDLM_ERROR(lock, "lock mode differs!");
1560 rc = ELDLM_NO_LOCK_DATA;
1561 } else if (bits == new) {
1562 /* This can be valid situation if CONVERT RPCs are
1563 * re-ordered. Just finish silently */
1564 LDLM_DEBUG(lock, "lock is converted already!");
1567 lock_res_and_lock(lock);
1568 if (ldlm_is_waited(lock))
1569 ldlm_del_waiting_lock(lock);
1571 ldlm_clear_cbpending(lock);
1572 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1573 ldlm_inodebits_drop(lock, bits & ~new);
1575 ldlm_clear_blocking_data(lock);
1576 unlock_res_and_lock(lock);
1578 ldlm_reprocess_all(lock->l_resource);
1582 if (rc == ELDLM_OK) {
1583 dlm_rep->lock_handle = lock->l_remote_handle;
1584 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1585 &dlm_rep->lock_desc.l_policy_data);
1588 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d",
1590 LDLM_LOCK_PUT(lock);
1592 rc = ELDLM_NO_LOCK_DATA;
1593 LDLM_DEBUG_NOLOCK("server-side convert handler END, rc = %d",
1597 req->rq_status = rc;
1603 * Cancel all the locks whose handles are packed into ldlm_request
1605 * Called by server code expecting such combined cancel activity
1608 int ldlm_request_cancel(struct ptlrpc_request *req,
1609 const struct ldlm_request *dlm_req,
1610 int first, enum lustre_at_flags flags)
1612 const struct lu_env *env = req->rq_svc_thread->t_env;
1613 struct ldlm_resource *res, *pres = NULL;
1614 struct ldlm_lock *lock;
1615 int i, count, done = 0;
1618 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1622 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1625 /* There is no lock on the server at the replay time,
1626 * skip lock cancelling to make replay tests to pass. */
1627 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1630 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1631 "starting at %d", count, first);
1633 for (i = first; i < count; i++) {
1634 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1636 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1637 "lock (cookie %llu)",
1638 dlm_req->lock_handle[i].cookie);
1642 res = lock->l_resource;
1645 /* This code is an optimization to only attempt lock
1646 * granting on the resource (that could be CPU-expensive)
1647 * after we are done cancelling lock in that resource. */
1650 ldlm_reprocess_all(pres);
1651 LDLM_RESOURCE_DELREF(pres);
1652 ldlm_resource_putref(pres);
1655 ldlm_resource_getref(res);
1656 LDLM_RESOURCE_ADDREF(res);
1658 if (!ldlm_is_discard_data(lock))
1659 ldlm_lvbo_update(env, res, lock,
1665 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1666 lock->l_blast_sent != 0) {
1667 time64_t delay = ktime_get_real_seconds() -
1669 LDLM_DEBUG(lock, "server cancels blocked lock after %llds",
1671 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1673 ldlm_lock_cancel(lock);
1674 LDLM_LOCK_PUT(lock);
1677 ldlm_reprocess_all(pres);
1678 LDLM_RESOURCE_DELREF(pres);
1679 ldlm_resource_putref(pres);
1681 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1684 EXPORT_SYMBOL(ldlm_request_cancel);
1687 * Main LDLM entry point for server code to cancel locks.
1689 * Typically gets called from service handler on LDLM_CANCEL opc.
1691 int ldlm_handle_cancel(struct ptlrpc_request *req)
1693 struct ldlm_request *dlm_req;
1697 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1698 if (dlm_req == NULL) {
1699 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1703 if (req->rq_export && req->rq_export->exp_nid_stats &&
1704 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1705 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1706 LDLM_CANCEL - LDLM_FIRST_OPC);
1708 rc = req_capsule_server_pack(&req->rq_pill);
1712 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1713 req->rq_status = LUSTRE_ESTALE;
1715 RETURN(ptlrpc_reply(req));
1717 #endif /* HAVE_SERVER_SUPPORT */
1720 * Callback handler for receiving incoming blocking ASTs.
1722 * This can only happen on client side.
1724 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1725 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1730 LDLM_DEBUG(lock, "client blocking AST callback handler");
1732 lock_res_and_lock(lock);
1734 /* set bits to cancel for this lock for possible lock convert */
1735 if (ns_is_client(ns) && (lock->l_resource->lr_type == LDLM_IBITS)) {
1736 /* Lock description contains policy of blocking lock,
1737 * and its cancel_bits is used to pass conflicting bits.
1738 * NOTE: ld can be NULL or can be not NULL but zeroed if
1739 * passed from ldlm_bl_thread_blwi(), check below used bits
1740 * in ld to make sure it is valid description.
1742 if (ld && ld->l_policy_data.l_inodebits.bits)
1743 lock->l_policy_data.l_inodebits.cancel_bits =
1744 ld->l_policy_data.l_inodebits.cancel_bits;
1745 /* if there is no valid ld and lock is cbpending already
1746 * then cancel_bits should be kept, otherwise it is zeroed.
1748 else if (!ldlm_is_cbpending(lock))
1749 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1751 ldlm_set_cbpending(lock);
1753 if (ldlm_is_cancel_on_block(lock))
1754 ldlm_set_cancel(lock);
1756 do_ast = (!lock->l_readers && !lock->l_writers);
1757 unlock_res_and_lock(lock);
1760 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1761 lock, lock->l_blocking_ast);
1762 if (lock->l_blocking_ast != NULL)
1763 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1766 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1770 LDLM_DEBUG(lock, "client blocking callback handler END");
1771 LDLM_LOCK_RELEASE(lock);
1776 * Callback handler for receiving incoming completion ASTs.
1778 * This only can happen on client side.
1780 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1781 struct ldlm_namespace *ns,
1782 struct ldlm_request *dlm_req,
1783 struct ldlm_lock *lock)
1785 struct list_head ast_list;
1790 LDLM_DEBUG(lock, "client completion callback handler START");
1792 INIT_LIST_HEAD(&ast_list);
1793 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1794 long to = cfs_time_seconds(1);
1797 set_current_state(TASK_INTERRUPTIBLE);
1798 schedule_timeout(to);
1799 if (lock->l_granted_mode == lock->l_req_mode ||
1800 ldlm_is_destroyed(lock))
1805 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1807 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1808 GOTO(out, rc = lvb_len);
1809 } else if (lvb_len > 0) {
1810 if (lock->l_lvb_len > 0) {
1811 /* for extent lock, lvb contains ost_lvb{}. */
1812 LASSERT(lock->l_lvb_data != NULL);
1814 if (unlikely(lock->l_lvb_len < lvb_len)) {
1815 LDLM_ERROR(lock, "Replied LVB is larger than "
1816 "expectation, expected = %d, "
1818 lock->l_lvb_len, lvb_len);
1819 GOTO(out, rc = -EINVAL);
1824 lock_res_and_lock(lock);
1826 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1827 &lock->l_resource->lr_name)) {
1828 ldlm_resource_unlink_lock(lock);
1829 unlock_res_and_lock(lock);
1830 rc = ldlm_lock_change_resource(ns, lock,
1831 &dlm_req->lock_desc.l_resource.lr_name);
1833 LDLM_ERROR(lock, "Failed to allocate resource");
1836 LDLM_DEBUG(lock, "completion AST, new resource");
1837 lock_res_and_lock(lock);
1840 if (ldlm_is_destroyed(lock) ||
1841 lock->l_granted_mode == lock->l_req_mode) {
1842 /* bug 11300: the lock has already been granted */
1843 unlock_res_and_lock(lock);
1844 LDLM_DEBUG(lock, "Double grant race happened");
1848 /* If we receive the completion AST before the actual enqueue returned,
1849 * then we might need to switch lock modes, resources, or extents. */
1850 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1851 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1852 LDLM_DEBUG(lock, "completion AST, new lock mode");
1855 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1856 ldlm_convert_policy_to_local(req->rq_export,
1857 dlm_req->lock_desc.l_resource.lr_type,
1858 &dlm_req->lock_desc.l_policy_data,
1859 &lock->l_policy_data);
1860 LDLM_DEBUG(lock, "completion AST, new policy data");
1863 ldlm_resource_unlink_lock(lock);
1865 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1866 /* BL_AST locks are not needed in LRU.
1867 * Let ldlm_cancel_lru() be fast. */
1868 ldlm_lock_remove_from_lru(lock);
1869 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1870 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1873 if (lock->l_lvb_len > 0) {
1874 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1875 lock->l_lvb_data, lvb_len);
1877 unlock_res_and_lock(lock);
1882 ldlm_grant_lock(lock, &ast_list);
1883 unlock_res_and_lock(lock);
1885 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1887 /* Let Enqueue to call osc_lock_upcall() and initialize
1889 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1891 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1893 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1899 lock_res_and_lock(lock);
1900 ldlm_set_failed(lock);
1901 unlock_res_and_lock(lock);
1902 wake_up(&lock->l_waitq);
1904 LDLM_LOCK_RELEASE(lock);
1908 * Callback handler for receiving incoming glimpse ASTs.
1910 * This only can happen on client side. After handling the glimpse AST
1911 * we also consider dropping the lock here if it is unused locally for a
1914 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1915 struct ldlm_namespace *ns,
1916 struct ldlm_request *dlm_req,
1917 struct ldlm_lock *lock)
1922 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1924 if (lock->l_glimpse_ast != NULL)
1925 rc = lock->l_glimpse_ast(lock, req);
1927 if (req->rq_repmsg != NULL) {
1930 req->rq_status = rc;
1934 lock_res_and_lock(lock);
1935 if (lock->l_granted_mode == LCK_PW &&
1936 !lock->l_readers && !lock->l_writers &&
1937 ktime_after(ktime_get(),
1938 ktime_add(lock->l_last_used,
1939 ktime_set(ns->ns_dirty_age_limit, 0)))) {
1940 unlock_res_and_lock(lock);
1941 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1942 ldlm_handle_bl_callback(ns, NULL, lock);
1947 unlock_res_and_lock(lock);
1948 LDLM_LOCK_RELEASE(lock);
1952 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1954 if (req->rq_no_reply)
1957 req->rq_status = rc;
1958 if (!req->rq_packed_final) {
1959 rc = lustre_pack_reply(req, 1, NULL, NULL);
1963 return ptlrpc_reply(req);
1966 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1967 enum ldlm_cancel_flags cancel_flags)
1969 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1972 spin_lock(&blp->blp_lock);
1973 if (blwi->blwi_lock &&
1974 ldlm_is_discard_data(blwi->blwi_lock)) {
1975 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1976 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1978 /* other blocking callbacks are added to the regular list */
1979 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1981 spin_unlock(&blp->blp_lock);
1983 wake_up(&blp->blp_waitq);
1985 /* can not check blwi->blwi_flags as blwi could be already freed in
1987 if (!(cancel_flags & LCF_ASYNC))
1988 wait_for_completion(&blwi->blwi_comp);
1993 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1994 struct ldlm_namespace *ns,
1995 struct ldlm_lock_desc *ld,
1996 struct list_head *cancels, int count,
1997 struct ldlm_lock *lock,
1998 enum ldlm_cancel_flags cancel_flags)
2000 init_completion(&blwi->blwi_comp);
2001 INIT_LIST_HEAD(&blwi->blwi_head);
2003 if (memory_pressure_get())
2004 blwi->blwi_mem_pressure = 1;
2007 blwi->blwi_flags = cancel_flags;
2009 blwi->blwi_ld = *ld;
2011 list_add(&blwi->blwi_head, cancels);
2012 list_del_init(cancels);
2013 blwi->blwi_count = count;
2015 blwi->blwi_lock = lock;
2020 * Queues a list of locks \a cancels containing \a count locks
2021 * for later processing by a blocking thread. If \a count is zero,
2022 * then the lock referenced as \a lock is queued instead.
2024 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2025 * If list addition fails an error is returned and caller is supposed to
2026 * call ->l_blocking_ast itself.
2028 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2029 struct ldlm_lock_desc *ld,
2030 struct ldlm_lock *lock,
2031 struct list_head *cancels, int count,
2032 enum ldlm_cancel_flags cancel_flags)
2036 if (cancels && count == 0)
2039 if (cancel_flags & LCF_ASYNC) {
2040 struct ldlm_bl_work_item *blwi;
2042 OBD_ALLOC(blwi, sizeof(*blwi));
2045 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2047 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2049 /* if it is synchronous call do minimum mem alloc, as it could
2050 * be triggered from kernel shrinker
2052 struct ldlm_bl_work_item blwi;
2054 memset(&blwi, 0, sizeof(blwi));
2055 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2056 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2061 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2062 struct ldlm_lock *lock)
2064 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2067 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2068 struct list_head *cancels, int count,
2069 enum ldlm_cancel_flags cancel_flags)
2071 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2074 int ldlm_bl_thread_wakeup(void)
2076 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2080 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2081 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2083 struct obd_device *obd = req->rq_export->exp_obd;
2090 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2092 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2094 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2096 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2099 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2101 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2103 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2106 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2109 /* We are responsible for swabbing contents of val */
2111 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2112 /* Pass it on to mdc (the "export" in this case) */
2113 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2115 sizeof(KEY_HSM_COPYTOOL_SEND),
2116 KEY_HSM_COPYTOOL_SEND,
2119 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2124 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2125 const char *msg, int rc,
2126 const struct lustre_handle *handle)
2128 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2129 "%s: [nid %s] [rc %d] [lock %#llx]",
2130 msg, libcfs_id2str(req->rq_peer), rc,
2131 handle ? handle->cookie : 0);
2132 if (req->rq_no_reply)
2133 CWARN("No reply was sent, maybe cause bug 21636.\n");
2135 CWARN("Send reply failed, maybe cause bug 21636.\n");
2138 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2139 static int ldlm_callback_handler(struct ptlrpc_request *req)
2141 struct ldlm_namespace *ns;
2142 struct ldlm_request *dlm_req;
2143 struct ldlm_lock *lock;
2147 /* Requests arrive in sender's byte order. The ptlrpc service
2148 * handler has already checked and, if necessary, byte-swapped the
2149 * incoming request message body, but I am responsible for the
2150 * message buffers. */
2152 /* do nothing for sec context finalize */
2153 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2156 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2158 if (req->rq_export == NULL) {
2159 rc = ldlm_callback_reply(req, -ENOTCONN);
2160 ldlm_callback_errmsg(req, "Operate on unconnected server",
2165 LASSERT(req->rq_export != NULL);
2166 LASSERT(req->rq_export->exp_obd != NULL);
2168 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2169 case LDLM_BL_CALLBACK:
2170 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2172 ldlm_callback_reply(req, -(int)cfs_fail_err);
2176 case LDLM_CP_CALLBACK:
2177 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2180 case LDLM_GL_CALLBACK:
2181 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2185 rc = ldlm_handle_setinfo(req);
2186 ldlm_callback_reply(req, rc);
2189 CERROR("unknown opcode %u\n",
2190 lustre_msg_get_opc(req->rq_reqmsg));
2191 ldlm_callback_reply(req, -EPROTO);
2195 ns = req->rq_export->exp_obd->obd_namespace;
2196 LASSERT(ns != NULL);
2198 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2200 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2201 if (dlm_req == NULL) {
2202 rc = ldlm_callback_reply(req, -EPROTO);
2203 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2208 /* Force a known safe race, send a cancel to the server for a lock
2209 * which the server has already started a blocking callback on. */
2210 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2211 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2212 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2214 CERROR("ldlm_cli_cancel: %d\n", rc);
2217 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2219 CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock "
2220 "disappeared\n", dlm_req->lock_handle[0].cookie);
2221 rc = ldlm_callback_reply(req, -EINVAL);
2222 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2223 &dlm_req->lock_handle[0]);
2227 if (ldlm_is_fail_loc(lock) &&
2228 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2229 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2231 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2232 lock_res_and_lock(lock);
2233 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2235 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2236 /* If somebody cancels lock and cache is already dropped,
2237 * or lock is failed before cp_ast received on client,
2238 * we can tell the server we have no lock. Otherwise, we
2239 * should send cancel after dropping the cache. */
2240 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2241 ldlm_is_failed(lock)) {
2242 LDLM_DEBUG(lock, "callback on lock %llx - lock disappeared",
2243 dlm_req->lock_handle[0].cookie);
2244 unlock_res_and_lock(lock);
2245 LDLM_LOCK_RELEASE(lock);
2246 rc = ldlm_callback_reply(req, -EINVAL);
2247 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2248 &dlm_req->lock_handle[0]);
2251 /* BL_AST locks are not needed in LRU.
2252 * Let ldlm_cancel_lru() be fast. */
2253 ldlm_lock_remove_from_lru(lock);
2254 ldlm_set_bl_ast(lock);
2256 unlock_res_and_lock(lock);
2258 /* We want the ost thread to get this reply so that it can respond
2259 * to ost requests (write cache writeback) that might be triggered
2262 * But we'd also like to be able to indicate in the reply that we're
2263 * cancelling right now, because it's unused, or have an intent result
2264 * in the reply, so we might have to push the responsibility for sending
2265 * the reply down into the AST handlers, alas. */
2267 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2268 case LDLM_BL_CALLBACK:
2269 CDEBUG(D_INODE, "blocking ast\n");
2270 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2271 if (!ldlm_is_cancel_on_block(lock)) {
2272 rc = ldlm_callback_reply(req, 0);
2273 if (req->rq_no_reply || rc)
2274 ldlm_callback_errmsg(req, "Normal process", rc,
2275 &dlm_req->lock_handle[0]);
2277 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2278 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2280 case LDLM_CP_CALLBACK:
2281 CDEBUG(D_INODE, "completion ast\n");
2282 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2283 ldlm_callback_reply(req, 0);
2284 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2286 case LDLM_GL_CALLBACK:
2287 CDEBUG(D_INODE, "glimpse ast\n");
2288 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2289 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2292 LBUG(); /* checked above */
2298 #ifdef HAVE_SERVER_SUPPORT
2300 * Main handler for canceld thread.
2302 * Separated into its own thread to avoid deadlocks.
2304 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2310 /* Requests arrive in sender's byte order. The ptlrpc service
2311 * handler has already checked and, if necessary, byte-swapped the
2312 * incoming request message body, but I am responsible for the
2313 * message buffers. */
2315 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2317 if (req->rq_export == NULL) {
2318 struct ldlm_request *dlm_req;
2320 CERROR("%s from %s arrived at %lu with bad export cookie "
2322 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2323 libcfs_nid2str(req->rq_peer.nid),
2324 req->rq_arrival_time.tv_sec,
2325 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2327 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2328 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2329 dlm_req = req_capsule_client_get(&req->rq_pill,
2331 if (dlm_req != NULL)
2332 ldlm_lock_dump_handle(D_ERROR,
2333 &dlm_req->lock_handle[0]);
2335 ldlm_callback_reply(req, -ENOTCONN);
2339 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2340 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2342 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2343 CDEBUG(D_INODE, "cancel\n");
2344 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2345 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2346 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2348 rc = ldlm_handle_cancel(req);
2352 struct ldlm_request *dlm_req;
2354 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2355 CDEBUG(D_INODE, "convert\n");
2357 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2358 if (dlm_req == NULL) {
2359 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2360 rc = ldlm_callback_reply(req, -EPROTO);
2362 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2363 rc = ptlrpc_reply(req);
2368 CERROR("invalid opcode %d\n",
2369 lustre_msg_get_opc(req->rq_reqmsg));
2370 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2371 rc = ldlm_callback_reply(req, -EINVAL);
2377 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2378 struct ldlm_lock *lock)
2380 struct ldlm_request *dlm_req;
2381 struct lustre_handle lockh;
2387 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2388 if (dlm_req == NULL)
2391 ldlm_lock2handle(lock, &lockh);
2392 for (i = 0; i < dlm_req->lock_count; i++) {
2393 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2395 DEBUG_REQ(D_RPCTRACE, req,
2396 "Prio raised by lock %#llx.", lockh.cookie);
2405 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2407 struct ldlm_request *dlm_req;
2413 /* no prolong in recovery */
2414 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2417 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2418 if (dlm_req == NULL)
2421 for (i = 0; i < dlm_req->lock_count; i++) {
2422 struct ldlm_lock *lock;
2424 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2428 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2430 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2431 LDLM_LOCK_PUT(lock);
2440 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2441 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2442 .hpreq_check = ldlm_cancel_hpreq_check,
2446 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2450 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2452 if (req->rq_export == NULL)
2455 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2456 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2457 req->rq_ops = &ldlm_cancel_hpreq_ops;
2458 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2459 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2460 req->rq_ops = &ldlm_cancel_hpreq_ops;
2465 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2466 struct hlist_node *hnode, void *data)
2469 struct list_head *rpc_list = data;
2470 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2472 lock_res_and_lock(lock);
2474 if (lock->l_req_mode != lock->l_granted_mode) {
2475 unlock_res_and_lock(lock);
2479 LASSERT(lock->l_resource);
2480 if (lock->l_resource->lr_type != LDLM_IBITS &&
2481 lock->l_resource->lr_type != LDLM_PLAIN) {
2482 unlock_res_and_lock(lock);
2486 if (ldlm_is_ast_sent(lock)) {
2487 unlock_res_and_lock(lock);
2491 LASSERT(lock->l_blocking_ast);
2492 LASSERT(!lock->l_blocking_lock);
2494 ldlm_set_ast_sent(lock);
2495 if (lock->l_export && lock->l_export->exp_lock_hash) {
2496 /* NB: it's safe to call cfs_hash_del() even lock isn't
2497 * in exp_lock_hash. */
2498 /* In the function below, .hs_keycmp resolves to
2499 * ldlm_export_lock_keycmp() */
2500 /* coverity[overrun-buffer-val] */
2501 cfs_hash_del(lock->l_export->exp_lock_hash,
2502 &lock->l_remote_handle, &lock->l_exp_hash);
2505 list_add_tail(&lock->l_rk_ast, rpc_list);
2506 LDLM_LOCK_GET(lock);
2508 unlock_res_and_lock(lock);
2512 void ldlm_revoke_export_locks(struct obd_export *exp)
2514 struct list_head rpc_list;
2517 INIT_LIST_HEAD(&rpc_list);
2518 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2519 ldlm_revoke_lock_cb, &rpc_list, 0);
2520 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2521 LDLM_WORK_REVOKE_AST);
2525 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2526 #endif /* HAVE_SERVER_SUPPORT */
2528 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2529 struct ldlm_bl_work_item **p_blwi,
2530 struct obd_export **p_exp)
2532 struct ldlm_bl_work_item *blwi = NULL;
2533 static unsigned int num_bl = 0;
2534 static unsigned int num_stale;
2535 int num_th = atomic_read(&blp->blp_num_threads);
2537 *p_exp = obd_stale_export_get();
2539 spin_lock(&blp->blp_lock);
2540 if (*p_exp != NULL) {
2541 if (num_th == 1 || ++num_stale < num_th) {
2542 spin_unlock(&blp->blp_lock);
2549 /* process a request from the blp_list at least every blp_num_threads */
2550 if (!list_empty(&blp->blp_list) &&
2551 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2552 blwi = list_entry(blp->blp_list.next,
2553 struct ldlm_bl_work_item, blwi_entry);
2555 if (!list_empty(&blp->blp_prio_list))
2556 blwi = list_entry(blp->blp_prio_list.next,
2557 struct ldlm_bl_work_item,
2561 if (++num_bl >= num_th)
2563 list_del(&blwi->blwi_entry);
2565 spin_unlock(&blp->blp_lock);
2568 if (*p_exp != NULL && *p_blwi != NULL) {
2569 obd_stale_export_put(*p_exp);
2573 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2576 /* This only contains temporary data until the thread starts */
2577 struct ldlm_bl_thread_data {
2578 struct ldlm_bl_pool *bltd_blp;
2579 struct completion bltd_comp;
2583 static int ldlm_bl_thread_main(void *arg);
2585 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2587 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2588 struct task_struct *task;
2590 init_completion(&bltd.bltd_comp);
2592 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2593 if (bltd.bltd_num >= blp->blp_max_threads) {
2594 atomic_dec(&blp->blp_num_threads);
2598 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2600 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2601 atomic_dec(&blp->blp_num_threads);
2605 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2608 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2609 bltd.bltd_num, PTR_ERR(task));
2610 atomic_dec(&blp->blp_num_threads);
2611 return PTR_ERR(task);
2613 wait_for_completion(&bltd.bltd_comp);
2618 /* Not fatal if racy and have a few too many threads */
2619 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2620 struct ldlm_bl_work_item *blwi)
2622 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2625 if (atomic_read(&blp->blp_busy_threads) <
2626 atomic_read(&blp->blp_num_threads))
2629 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2630 blwi->blwi_mem_pressure))
2636 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2637 struct ldlm_bl_work_item *blwi)
2641 if (blwi->blwi_ns == NULL)
2642 /* added by ldlm_cleanup() */
2643 RETURN(LDLM_ITER_STOP);
2645 if (blwi->blwi_mem_pressure)
2646 memory_pressure_set();
2648 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2650 if (blwi->blwi_count) {
2652 /* The special case when we cancel locks in lru
2653 * asynchronously, we pass the list of locks here.
2654 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2655 * canceled locally yet. */
2656 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2659 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2662 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2665 if (blwi->blwi_mem_pressure)
2666 memory_pressure_clr();
2668 if (blwi->blwi_flags & LCF_ASYNC)
2669 OBD_FREE(blwi, sizeof(*blwi));
2671 complete(&blwi->blwi_comp);
2677 * Cancel stale locks on export. Cancel blocked locks first.
2678 * If the given export has blocked locks, the next in the list may have
2679 * them too, thus cancel not blocked locks only if the current export has
2682 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2683 struct obd_export *exp)
2688 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2690 num = ldlm_export_cancel_blocked_locks(exp);
2692 ldlm_export_cancel_locks(exp);
2694 obd_stale_export_put(exp);
2701 * Main blocking requests processing thread.
2703 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2704 * This thread in the end ends up doing actual call to ->l_blocking_ast
2707 static int ldlm_bl_thread_main(void *arg)
2709 struct ldlm_bl_pool *blp;
2710 struct ldlm_bl_thread_data *bltd = arg;
2713 blp = bltd->bltd_blp;
2715 complete(&bltd->bltd_comp);
2716 /* cannot use bltd after this, it is only on caller's stack */
2719 struct l_wait_info lwi = { 0 };
2720 struct ldlm_bl_work_item *blwi = NULL;
2721 struct obd_export *exp = NULL;
2724 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2727 l_wait_event_exclusive(blp->blp_waitq,
2728 ldlm_bl_get_work(blp, &blwi,
2731 atomic_inc(&blp->blp_busy_threads);
2733 if (ldlm_bl_thread_need_create(blp, blwi))
2734 /* discard the return value, we tried */
2735 ldlm_bl_thread_start(blp, true);
2738 rc = ldlm_bl_thread_exports(blp, exp);
2740 rc = ldlm_bl_thread_blwi(blp, blwi);
2742 atomic_dec(&blp->blp_busy_threads);
2744 if (rc == LDLM_ITER_STOP)
2747 /* If there are many namespaces, we will not sleep waiting for
2748 * work, and must do a cond_resched to avoid holding the CPU
2753 atomic_dec(&blp->blp_num_threads);
2754 complete(&blp->blp_comp);
2759 static int ldlm_setup(void);
2760 static int ldlm_cleanup(void);
2762 int ldlm_get_ref(void)
2766 mutex_lock(&ldlm_ref_mutex);
2767 if (++ldlm_refcount == 1) {
2772 mutex_unlock(&ldlm_ref_mutex);
2777 void ldlm_put_ref(void)
2780 mutex_lock(&ldlm_ref_mutex);
2781 if (ldlm_refcount == 1) {
2782 int rc = ldlm_cleanup();
2784 CERROR("ldlm_cleanup failed: %d\n", rc);
2790 mutex_unlock(&ldlm_ref_mutex);
2796 * Export handle<->lock hash operations.
2799 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2801 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2805 ldlm_export_lock_key(struct hlist_node *hnode)
2807 struct ldlm_lock *lock;
2809 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2810 return &lock->l_remote_handle;
2814 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2816 struct ldlm_lock *lock;
2818 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2819 lock->l_remote_handle = *(struct lustre_handle *)key;
2823 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2825 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2829 ldlm_export_lock_object(struct hlist_node *hnode)
2831 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2835 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2837 struct ldlm_lock *lock;
2839 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2840 LDLM_LOCK_GET(lock);
2844 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2846 struct ldlm_lock *lock;
2848 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2849 LDLM_LOCK_RELEASE(lock);
2852 static struct cfs_hash_ops ldlm_export_lock_ops = {
2853 .hs_hash = ldlm_export_lock_hash,
2854 .hs_key = ldlm_export_lock_key,
2855 .hs_keycmp = ldlm_export_lock_keycmp,
2856 .hs_keycpy = ldlm_export_lock_keycpy,
2857 .hs_object = ldlm_export_lock_object,
2858 .hs_get = ldlm_export_lock_get,
2859 .hs_put = ldlm_export_lock_put,
2860 .hs_put_locked = ldlm_export_lock_put,
2863 int ldlm_init_export(struct obd_export *exp)
2868 exp->exp_lock_hash =
2869 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2870 HASH_EXP_LOCK_CUR_BITS,
2871 HASH_EXP_LOCK_MAX_BITS,
2872 HASH_EXP_LOCK_BKT_BITS, 0,
2873 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2874 &ldlm_export_lock_ops,
2875 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2876 CFS_HASH_NBLK_CHANGE);
2878 if (!exp->exp_lock_hash)
2881 rc = ldlm_init_flock_export(exp);
2887 ldlm_destroy_export(exp);
2890 EXPORT_SYMBOL(ldlm_init_export);
2892 void ldlm_destroy_export(struct obd_export *exp)
2895 cfs_hash_putref(exp->exp_lock_hash);
2896 exp->exp_lock_hash = NULL;
2898 ldlm_destroy_flock_export(exp);
2901 EXPORT_SYMBOL(ldlm_destroy_export);
2903 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
2904 struct attribute *attr,
2907 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
2910 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
2911 struct attribute *attr,
2918 rc = kstrtoul(buffer, 10, &val);
2922 ldlm_cancel_unused_locks_before_replay = val;
2926 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
2928 static struct attribute *ldlm_attrs[] = {
2929 &lustre_attr_cancel_unused_locks_before_replay.attr,
2933 static struct attribute_group ldlm_attr_group = {
2934 .attrs = ldlm_attrs,
2937 static int ldlm_setup(void)
2939 static struct ptlrpc_service_conf conf;
2940 struct ldlm_bl_pool *blp = NULL;
2941 #ifdef HAVE_SERVER_SUPPORT
2942 struct task_struct *task;
2943 #endif /* HAVE_SERVER_SUPPORT */
2949 if (ldlm_state != NULL)
2952 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2953 if (ldlm_state == NULL)
2956 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
2960 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
2964 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
2968 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
2972 rc = ldlm_debugfs_setup();
2976 memset(&conf, 0, sizeof(conf));
2977 conf = (typeof(conf)) {
2978 .psc_name = "ldlm_cbd",
2979 .psc_watchdog_factor = 2,
2981 .bc_nbufs = LDLM_CLIENT_NBUFS,
2982 .bc_buf_size = LDLM_BUFSIZE,
2983 .bc_req_max_size = LDLM_MAXREQSIZE,
2984 .bc_rep_max_size = LDLM_MAXREPSIZE,
2985 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2986 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2989 .tc_thr_name = "ldlm_cb",
2990 .tc_thr_factor = LDLM_THR_FACTOR,
2991 .tc_nthrs_init = LDLM_NTHRS_INIT,
2992 .tc_nthrs_base = LDLM_NTHRS_BASE,
2993 .tc_nthrs_max = LDLM_NTHRS_MAX,
2994 .tc_nthrs_user = ldlm_num_threads,
2995 .tc_cpu_bind = ldlm_cpu_bind,
2996 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2999 .cc_pattern = ldlm_cpts,
3000 .cc_affinity = true,
3003 .so_req_handler = ldlm_callback_handler,
3006 ldlm_state->ldlm_cb_service = \
3007 ptlrpc_register_service(&conf, ldlm_svc_kset,
3008 ldlm_svc_debugfs_dir);
3009 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3010 CERROR("failed to start service\n");
3011 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3012 ldlm_state->ldlm_cb_service = NULL;
3016 #ifdef HAVE_SERVER_SUPPORT
3017 memset(&conf, 0, sizeof(conf));
3018 conf = (typeof(conf)) {
3019 .psc_name = "ldlm_canceld",
3020 .psc_watchdog_factor = 6,
3022 .bc_nbufs = LDLM_SERVER_NBUFS,
3023 .bc_buf_size = LDLM_BUFSIZE,
3024 .bc_req_max_size = LDLM_MAXREQSIZE,
3025 .bc_rep_max_size = LDLM_MAXREPSIZE,
3026 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3027 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3031 .tc_thr_name = "ldlm_cn",
3032 .tc_thr_factor = LDLM_THR_FACTOR,
3033 .tc_nthrs_init = LDLM_NTHRS_INIT,
3034 .tc_nthrs_base = LDLM_NTHRS_BASE,
3035 .tc_nthrs_max = LDLM_NTHRS_MAX,
3036 .tc_nthrs_user = ldlm_num_threads,
3037 .tc_cpu_bind = ldlm_cpu_bind,
3038 .tc_ctx_tags = LCT_MD_THREAD | \
3043 .cc_pattern = ldlm_cpts,
3044 .cc_affinity = true,
3047 .so_req_handler = ldlm_cancel_handler,
3048 .so_hpreq_handler = ldlm_hpreq_handler,
3051 ldlm_state->ldlm_cancel_service = \
3052 ptlrpc_register_service(&conf, ldlm_svc_kset,
3053 ldlm_svc_debugfs_dir);
3054 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3055 CERROR("failed to start service\n");
3056 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3057 ldlm_state->ldlm_cancel_service = NULL;
3060 #endif /* HAVE_SERVER_SUPPORT */
3062 OBD_ALLOC(blp, sizeof(*blp));
3064 GOTO(out, rc = -ENOMEM);
3065 ldlm_state->ldlm_bl_pool = blp;
3067 spin_lock_init(&blp->blp_lock);
3068 INIT_LIST_HEAD(&blp->blp_list);
3069 INIT_LIST_HEAD(&blp->blp_prio_list);
3070 init_waitqueue_head(&blp->blp_waitq);
3071 atomic_set(&blp->blp_num_threads, 0);
3072 atomic_set(&blp->blp_busy_threads, 0);
3074 if (ldlm_num_threads == 0) {
3075 blp->blp_min_threads = LDLM_NTHRS_INIT;
3076 blp->blp_max_threads = LDLM_NTHRS_MAX;
3078 blp->blp_min_threads = blp->blp_max_threads = \
3079 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3083 for (i = 0; i < blp->blp_min_threads; i++) {
3084 rc = ldlm_bl_thread_start(blp, false);
3089 #ifdef HAVE_SERVER_SUPPORT
3090 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3093 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3097 wait_event(expired_lock_wait_queue,
3098 expired_lock_thread_state == ELT_READY);
3099 #endif /* HAVE_SERVER_SUPPORT */
3101 rc = ldlm_pools_init();
3103 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3107 rc = ldlm_reclaim_setup();
3109 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3119 static int ldlm_cleanup(void)
3123 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3124 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3125 CERROR("ldlm still has namespaces; clean these up first.\n");
3126 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3127 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3131 ldlm_reclaim_cleanup();
3134 if (ldlm_state->ldlm_bl_pool != NULL) {
3135 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3137 while (atomic_read(&blp->blp_num_threads) > 0) {
3138 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3140 init_completion(&blp->blp_comp);
3142 spin_lock(&blp->blp_lock);
3143 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3144 wake_up(&blp->blp_waitq);
3145 spin_unlock(&blp->blp_lock);
3147 wait_for_completion(&blp->blp_comp);
3150 OBD_FREE(blp, sizeof(*blp));
3153 if (ldlm_state->ldlm_cb_service != NULL)
3154 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3155 #ifdef HAVE_SERVER_SUPPORT
3156 if (ldlm_state->ldlm_cancel_service != NULL)
3157 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3161 kset_unregister(ldlm_ns_kset);
3163 kset_unregister(ldlm_svc_kset);
3165 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3166 kobject_put(ldlm_kobj);
3169 ldlm_debugfs_cleanup();
3171 #ifdef HAVE_SERVER_SUPPORT
3172 if (expired_lock_thread_state != ELT_STOPPED) {
3173 expired_lock_thread_state = ELT_TERMINATE;
3174 wake_up(&expired_lock_wait_queue);
3175 wait_event(expired_lock_wait_queue,
3176 expired_lock_thread_state == ELT_STOPPED);
3180 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3188 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3189 sizeof(struct ldlm_resource), 0,
3190 SLAB_HWCACHE_ALIGN, NULL);
3191 if (ldlm_resource_slab == NULL)
3194 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3195 sizeof(struct ldlm_lock), 0,
3196 SLAB_HWCACHE_ALIGN, NULL);
3197 if (ldlm_lock_slab == NULL)
3200 ldlm_interval_slab = kmem_cache_create("interval_node",
3201 sizeof(struct ldlm_interval),
3202 0, SLAB_HWCACHE_ALIGN, NULL);
3203 if (ldlm_interval_slab == NULL)
3206 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3207 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3208 0, SLAB_HWCACHE_ALIGN, NULL);
3209 if (ldlm_interval_tree_slab == NULL)
3212 #ifdef HAVE_SERVER_SUPPORT
3213 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3214 sizeof(struct ldlm_glimpse_work),
3216 if (ldlm_glimpse_work_kmem == NULL)
3217 goto out_interval_tree;
3220 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3221 class_export_dump_hook = ldlm_dump_export_locks;
3224 #ifdef HAVE_SERVER_SUPPORT
3226 kmem_cache_destroy(ldlm_interval_tree_slab);
3229 kmem_cache_destroy(ldlm_interval_slab);
3231 kmem_cache_destroy(ldlm_lock_slab);
3233 kmem_cache_destroy(ldlm_resource_slab);
3238 void ldlm_exit(void)
3241 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3242 kmem_cache_destroy(ldlm_resource_slab);
3243 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3244 * synchronize_rcu() to wait a grace period elapsed, so that
3245 * ldlm_lock_free() get a chance to be called. */
3247 kmem_cache_destroy(ldlm_lock_slab);
3248 kmem_cache_destroy(ldlm_interval_slab);
3249 kmem_cache_destroy(ldlm_interval_tree_slab);
3250 #ifdef HAVE_SERVER_SUPPORT
3251 kmem_cache_destroy(ldlm_glimpse_work_kmem);