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 *data, int rc)
714 struct ldlm_cb_async_args *ca = data;
715 struct ldlm_lock *lock = ca->ca_lock;
716 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
719 LASSERT(lock != NULL);
722 case LDLM_GL_CALLBACK:
723 /* Update the LVB from disk if the AST failed
724 * (this is a legal race)
726 * - Glimpse callback of local lock just returns
727 * -ELDLM_NO_LOCK_DATA.
728 * - Glimpse callback of remote lock might return
729 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
731 if (unlikely(arg->gl_interpret_reply)) {
732 rc = arg->gl_interpret_reply(env, req, data, rc);
733 } else if (rc == -ELDLM_NO_LOCK_DATA) {
734 LDLM_DEBUG(lock, "lost race - client has a lock but no "
736 ldlm_lvbo_update(env, lock->l_resource, lock, NULL, 1);
737 } else if (rc != 0) {
738 rc = ldlm_handle_ast_error(env, lock, req,
741 rc = ldlm_lvbo_update(env, lock->l_resource,
745 case LDLM_BL_CALLBACK:
747 rc = ldlm_handle_ast_error(env, lock, req,
750 case LDLM_CP_CALLBACK:
752 rc = ldlm_handle_ast_error(env, lock, req,
756 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
761 /* release extra reference taken in ldlm_ast_fini() */
762 LDLM_LOCK_RELEASE(lock);
765 atomic_inc(&arg->restart);
770 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
772 struct ldlm_cb_async_args *ca = data;
773 struct ldlm_lock *lock = ca->ca_lock;
775 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
778 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
779 struct ldlm_cb_set_arg *arg,
780 struct ldlm_lock *lock,
786 if (unlikely(instant_cancel)) {
787 rc = ptl_send_rpc(req, 1);
788 ptlrpc_req_finished(req);
790 atomic_inc(&arg->restart);
793 ptlrpc_set_add_req(arg->set, req);
800 * Check if there are requests in the export request list which prevent
801 * the lock canceling and make these requests high priority ones.
803 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
805 struct ptlrpc_request *req;
808 if (lock->l_export == NULL) {
809 LDLM_DEBUG(lock, "client lock: no-op");
813 spin_lock(&lock->l_export->exp_rpc_lock);
814 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
816 /* Do not process requests that were not yet added to there
817 * incoming queue or were already removed from there for
818 * processing. We evaluate ptlrpc_nrs_req_can_move() without
819 * holding svcpt->scp_req_lock, and then redo the check with
820 * the lock held once we need to obtain a reliable result.
822 if (ptlrpc_nrs_req_can_move(req) &&
823 req->rq_ops->hpreq_lock_match &&
824 req->rq_ops->hpreq_lock_match(req, lock))
825 ptlrpc_nrs_req_hp_move(req);
827 spin_unlock(&lock->l_export->exp_rpc_lock);
832 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
833 * enqueued server lock conflicts with given one.
835 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
836 * to wait for client response.
838 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
839 struct ldlm_lock_desc *desc,
840 void *data, int flag)
842 struct ldlm_cb_async_args *ca;
843 struct ldlm_cb_set_arg *arg = data;
844 struct ldlm_request *body;
845 struct ptlrpc_request *req;
846 int instant_cancel = 0;
850 if (flag == LDLM_CB_CANCELING)
851 /* Don't need to do anything here. */
854 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
855 LDLM_DEBUG(lock, "dropping BL AST");
860 LASSERT(data != NULL);
861 if (lock->l_export->exp_obd->obd_recovering != 0)
862 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
864 ldlm_lock_reorder_req(lock);
866 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
867 &RQF_LDLM_BL_CALLBACK,
868 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
872 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
873 ca = ptlrpc_req_async_args(req);
874 ca->ca_set_arg = arg;
877 req->rq_interpret_reply = ldlm_cb_interpret;
879 lock_res_and_lock(lock);
880 if (ldlm_is_destroyed(lock)) {
881 /* What's the point? */
882 unlock_res_and_lock(lock);
883 ptlrpc_req_finished(req);
887 if (lock->l_granted_mode != lock->l_req_mode) {
888 /* this blocking AST will be communicated as part of the
889 * completion AST instead */
890 ldlm_add_blocked_lock(lock);
891 ldlm_set_waited(lock);
892 unlock_res_and_lock(lock);
894 ptlrpc_req_finished(req);
895 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
899 if (ldlm_is_cancel_on_block(lock))
902 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
903 body->lock_handle[0] = lock->l_remote_handle;
904 body->lock_desc = *desc;
905 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
907 LDLM_DEBUG(lock, "server preparing blocking AST");
909 ptlrpc_request_set_replen(req);
910 ldlm_set_cbpending(lock);
911 if (instant_cancel) {
912 unlock_res_and_lock(lock);
913 ldlm_lock_cancel(lock);
915 req->rq_no_resend = 1;
917 LASSERT(lock->l_granted_mode == lock->l_req_mode);
918 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
919 unlock_res_and_lock(lock);
921 /* Do not resend after lock callback timeout */
922 req->rq_delay_limit = ldlm_bl_timeout(lock);
923 req->rq_resend_cb = ldlm_update_resend;
926 req->rq_send_state = LUSTRE_IMP_FULL;
927 /* ptlrpc_request_alloc_pack already set timeout */
929 req->rq_timeout = ldlm_get_rq_timeout();
931 if (lock->l_export && lock->l_export->exp_nid_stats &&
932 lock->l_export->exp_nid_stats->nid_ldlm_stats)
933 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
934 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
936 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
942 * ->l_completion_ast callback for a remote lock in server namespace.
944 * Sends AST to the client notifying it of lock granting. If initial
945 * lock response was not sent yet, instead of sending another RPC, just
946 * mark the lock as granted and client will understand
948 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
950 struct ldlm_cb_set_arg *arg = data;
951 struct ldlm_request *body;
952 struct ptlrpc_request *req;
953 struct ldlm_cb_async_args *ca;
954 int instant_cancel = 0;
959 LASSERT(lock != NULL);
960 LASSERT(data != NULL);
962 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
963 LDLM_DEBUG(lock, "dropping CP AST");
967 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
968 &RQF_LDLM_CP_CALLBACK);
972 /* server namespace, doesn't need lock */
973 lvb_len = ldlm_lvbo_size(lock);
974 /* LU-3124 & LU-2187: to not return layout in completion AST because
975 * it may deadlock for LU-2187, or client may not have enough space
976 * for large layout. The layout will be returned to client with an
977 * extra RPC to fetch xattr.lov */
978 if (ldlm_has_layout(lock))
981 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
982 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
984 ptlrpc_request_free(req);
988 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
989 ca = ptlrpc_req_async_args(req);
990 ca->ca_set_arg = arg;
993 req->rq_interpret_reply = ldlm_cb_interpret;
994 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
996 body->lock_handle[0] = lock->l_remote_handle;
997 body->lock_flags = ldlm_flags_to_wire(flags);
998 ldlm_lock2desc(lock, &body->lock_desc);
1000 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1001 const struct lu_env *env = NULL;
1003 if (req->rq_svc_thread)
1004 env = req->rq_svc_thread->t_env;
1006 lvb_len = ldlm_lvbo_fill(env, lock, lvb, &lvb_len);
1008 /* We still need to send the RPC to wake up the blocked
1009 * enqueue thread on the client.
1011 * Consider old client, there is no better way to notify
1012 * the failure, just zero-sized the LVB, then the client
1013 * will fail out as "-EPROTO". */
1014 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1018 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1023 LDLM_DEBUG(lock, "server preparing completion AST");
1025 ptlrpc_request_set_replen(req);
1027 req->rq_send_state = LUSTRE_IMP_FULL;
1028 /* ptlrpc_request_pack already set timeout */
1030 req->rq_timeout = ldlm_get_rq_timeout();
1032 /* We only send real blocking ASTs after the lock is granted */
1033 lock_res_and_lock(lock);
1034 if (ldlm_is_ast_sent(lock)) {
1035 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1036 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1037 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1040 /* We might get here prior to ldlm_handle_enqueue setting
1041 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1042 * into waiting list, but this is safe and similar code in
1043 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1044 * that would not only cancel the lock, but will also remove
1045 * it from waiting list */
1046 if (ldlm_is_cancel_on_block(lock)) {
1047 unlock_res_and_lock(lock);
1048 ldlm_lock_cancel(lock);
1051 req->rq_no_resend = 1;
1053 lock_res_and_lock(lock);
1055 /* start the lock-timeout clock */
1056 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1057 /* Do not resend after lock callback timeout */
1058 req->rq_delay_limit = ldlm_bl_timeout(lock);
1059 req->rq_resend_cb = ldlm_update_resend;
1062 unlock_res_and_lock(lock);
1064 if (lock->l_export && lock->l_export->exp_nid_stats &&
1065 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1066 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1067 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1069 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1071 RETURN(lvb_len < 0 ? lvb_len : rc);
1075 * Server side ->l_glimpse_ast handler for client locks.
1077 * Sends glimpse AST to the client and waits for reply. Then updates
1078 * lvbo with the result.
1080 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1082 struct ldlm_cb_set_arg *arg = data;
1083 struct ldlm_request *body;
1084 struct ptlrpc_request *req;
1085 struct ldlm_cb_async_args *ca;
1087 struct req_format *req_fmt;
1090 LASSERT(lock != NULL);
1092 if (arg->gl_desc != NULL)
1093 /* There is a glimpse descriptor to pack */
1094 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1096 req_fmt = &RQF_LDLM_GL_CALLBACK;
1098 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1099 req_fmt, LUSTRE_DLM_VERSION,
1105 if (arg->gl_desc != NULL) {
1106 /* copy the GL descriptor */
1107 union ldlm_gl_desc *desc;
1108 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1109 *desc = *arg->gl_desc;
1112 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1113 body->lock_handle[0] = lock->l_remote_handle;
1114 ldlm_lock2desc(lock, &body->lock_desc);
1116 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1117 ca = ptlrpc_req_async_args(req);
1118 ca->ca_set_arg = arg;
1121 /* server namespace, doesn't need lock */
1122 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1123 ldlm_lvbo_size(lock));
1124 ptlrpc_request_set_replen(req);
1126 req->rq_send_state = LUSTRE_IMP_FULL;
1127 /* ptlrpc_request_alloc_pack already set timeout */
1129 req->rq_timeout = ldlm_get_rq_timeout();
1131 req->rq_interpret_reply = ldlm_cb_interpret;
1133 if (lock->l_export && lock->l_export->exp_nid_stats &&
1134 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1135 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1136 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1138 rc = ldlm_ast_fini(req, arg, lock, 0);
1142 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1144 int ldlm_glimpse_locks(struct ldlm_resource *res,
1145 struct list_head *gl_work_list)
1150 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1152 if (rc == -ERESTART)
1153 ldlm_reprocess_all(res);
1157 EXPORT_SYMBOL(ldlm_glimpse_locks);
1159 /* return LDLM lock associated with a lock callback request */
1160 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1162 struct ldlm_cb_async_args *ca;
1163 struct ldlm_lock *lock;
1166 ca = ptlrpc_req_async_args(req);
1169 RETURN(ERR_PTR(-EFAULT));
1173 EXPORT_SYMBOL(ldlm_request_lock);
1176 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1177 * service threads to carry out client lock enqueueing requests.
1179 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1180 struct ptlrpc_request *req,
1181 const struct ldlm_request *dlm_req,
1182 const struct ldlm_callback_suite *cbs)
1184 struct ldlm_reply *dlm_rep;
1186 enum ldlm_error err = ELDLM_OK;
1187 struct ldlm_lock *lock = NULL;
1188 void *cookie = NULL;
1190 struct ldlm_resource *res = NULL;
1191 const struct lu_env *env = req->rq_svc_thread->t_env;
1194 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1196 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1197 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1199 LASSERT(req->rq_export);
1201 /* for intent enqueue the stat will be updated inside intent policy */
1202 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1203 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1204 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1206 if (req->rq_export && req->rq_export->exp_nid_stats &&
1207 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1208 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1209 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1211 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1212 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1213 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1214 dlm_req->lock_desc.l_resource.lr_type);
1215 GOTO(out, rc = -EFAULT);
1218 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1219 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1220 dlm_req->lock_desc.l_req_mode &
1221 (dlm_req->lock_desc.l_req_mode-1))) {
1222 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1223 dlm_req->lock_desc.l_req_mode);
1224 GOTO(out, rc = -EFAULT);
1227 if (unlikely((flags & LDLM_FL_REPLAY) ||
1228 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1229 /* Find an existing lock in the per-export lock hash */
1230 /* In the function below, .hs_keycmp resolves to
1231 * ldlm_export_lock_keycmp() */
1232 /* coverity[overrun-buffer-val] */
1233 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1234 (void *)&dlm_req->lock_handle[0]);
1236 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie %#llx",
1237 lock->l_handle.h_cookie);
1238 flags |= LDLM_FL_RESENT;
1239 GOTO(existing_lock, rc = 0);
1242 if (ldlm_reclaim_full()) {
1243 DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1244 "reject current enqueue request and let the "
1245 "client retry later.\n");
1246 GOTO(out, rc = -EINPROGRESS);
1250 /* The lock's callback data might be set in the policy function */
1251 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1252 dlm_req->lock_desc.l_resource.lr_type,
1253 dlm_req->lock_desc.l_req_mode,
1254 cbs, NULL, 0, LVB_T_NONE);
1261 lock->l_remote_handle = dlm_req->lock_handle[0];
1262 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1264 /* Initialize resource lvb but not for a lock being replayed since
1265 * Client already got lvb sent in this case.
1266 * This must occur early since some policy methods assume resource
1267 * lvb is available (lr_lvb_data != NULL).
1269 res = lock->l_resource;
1270 if (!(flags & LDLM_FL_REPLAY)) {
1271 /* non-replayed lock, delayed lvb init may need to be done */
1272 rc = ldlm_lvbo_init(env, res);
1274 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1279 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1280 /* Don't enqueue a lock onto the export if it is been disonnected
1281 * due to eviction (bug 3822) or server umount (bug 24324).
1282 * Cancel it now instead. */
1283 if (req->rq_export->exp_disconnected) {
1284 LDLM_ERROR(lock, "lock on disconnected export %p",
1286 GOTO(out, rc = -ENOTCONN);
1289 lock->l_export = class_export_lock_get(req->rq_export, lock);
1290 if (lock->l_export->exp_lock_hash)
1291 cfs_hash_add(lock->l_export->exp_lock_hash,
1292 &lock->l_remote_handle,
1295 /* Inherit the enqueue flags before the operation, because we do not
1296 * keep the res lock on return and next operations (BL AST) may proceed
1298 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1299 LDLM_FL_INHERIT_MASK);
1301 ldlm_convert_policy_to_local(req->rq_export,
1302 dlm_req->lock_desc.l_resource.lr_type,
1303 &dlm_req->lock_desc.l_policy_data,
1304 &lock->l_policy_data);
1305 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1306 lock->l_req_extent = lock->l_policy_data.l_extent;
1309 if (flags & LDLM_FL_HAS_INTENT) {
1310 /* In this case, the reply buffer is allocated deep in
1311 * local_lock_enqueue by the policy function. */
1314 /* based on the assumption that lvb size never changes during
1315 * resource life time otherwise it need resource->lr_lock's
1317 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1318 RCL_SERVER, ldlm_lvbo_size(lock));
1320 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1321 GOTO(out, rc = -ENOMEM);
1323 rc = req_capsule_server_pack(&req->rq_pill);
1328 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1335 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1337 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1338 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1340 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1341 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1343 /* We never send a blocking AST until the lock is granted, but
1344 * we can tell it right now */
1345 lock_res_and_lock(lock);
1347 /* Now take into account flags to be inherited from original lock
1348 request both in reply to client and in our own lock flags. */
1349 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1350 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1352 /* Don't move a pending lock onto the export if it has already been
1353 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1354 * Cancel it now instead. */
1355 if (unlikely(req->rq_export->exp_disconnected ||
1356 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1357 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1359 } else if (ldlm_is_ast_sent(lock)) {
1360 /* fill lock desc for possible lock convert */
1361 if (lock->l_blocking_lock &&
1362 lock->l_resource->lr_type == LDLM_IBITS) {
1363 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1364 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1367 "save blocking bits %llx in granted lock",
1368 bl_lock->l_policy_data.l_inodebits.bits);
1369 /* If lock is blocked then save blocking ibits
1370 * in returned lock policy for the possible lock
1371 * convert on a client.
1373 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1374 bl_lock->l_policy_data.l_inodebits.bits;
1376 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1377 if (lock->l_granted_mode == lock->l_req_mode) {
1379 * Only cancel lock if it was granted, because it would
1380 * be destroyed immediately and would never be granted
1381 * in the future, causing timeouts on client. Not
1382 * granted lock will be cancelled immediately after
1383 * sending completion AST.
1385 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1386 unlock_res_and_lock(lock);
1387 ldlm_lock_cancel(lock);
1388 lock_res_and_lock(lock);
1390 ldlm_add_waiting_lock(lock,
1391 ldlm_bl_timeout(lock));
1395 unlock_res_and_lock(lock);
1399 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1400 if (!req->rq_packed_final) {
1401 err = lustre_pack_reply(req, 1, NULL, NULL);
1406 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1407 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1409 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1410 "(err=%d, rc=%d)", err, rc);
1413 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1415 ldlm_lvbo_size(lock) > 0) {
1420 buf = req_capsule_server_get(&req->rq_pill,
1422 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1423 buflen = req_capsule_get_size(&req->rq_pill,
1424 &RMF_DLM_LVB, RCL_SERVER);
1425 /* non-replayed lock, delayed lvb init may
1426 * need to be occur now
1428 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1431 rc2 = ldlm_lvbo_fill(env, lock, buf, &buflen);
1433 req_capsule_shrink(&req->rq_pill,
1436 } else if (rc2 == -ERANGE) {
1437 rc2 = req_capsule_server_grow(
1439 &RMF_DLM_LVB, buflen);
1443 /* if we can't grow the buffer,
1444 * it's ok to return empty lvb
1455 } else if (flags & LDLM_FL_REPLAY) {
1456 /* no LVB resend upon replay */
1458 req_capsule_shrink(&req->rq_pill,
1468 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1469 if (lock->l_export) {
1470 ldlm_lock_cancel(lock);
1472 lock_res_and_lock(lock);
1473 ldlm_resource_unlink_lock(lock);
1474 ldlm_lock_destroy_nolock(lock);
1475 unlock_res_and_lock(lock);
1480 if (!err && !ldlm_is_cbpending(lock) &&
1481 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1482 ldlm_reprocess_all(lock->l_resource);
1484 LDLM_LOCK_RELEASE(lock);
1487 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1493 /* Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1494 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1496 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1498 if (lock->l_blocking_lock) {
1499 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1500 lock->l_blocking_lock = NULL;
1504 /* A lock can be converted to new ibits or mode and should be considered
1505 * as new lock. Clear all states related to a previous blocking AST
1506 * processing so new conflicts will cause new blocking ASTs.
1508 * This is used during lock convert below and lock downgrade to COS mode in
1509 * ldlm_lock_mode_downgrade().
1511 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1513 ldlm_clear_ast_sent(lock);
1514 lock->l_bl_ast_run = 0;
1515 ldlm_clear_blocking_lock(lock);
1519 * Main LDLM entry point for server code to process lock conversion requests.
1521 int ldlm_handle_convert0(struct ptlrpc_request *req,
1522 const struct ldlm_request *dlm_req)
1524 struct obd_export *exp = req->rq_export;
1525 struct ldlm_reply *dlm_rep;
1526 struct ldlm_lock *lock;
1531 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1532 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1533 LDLM_CONVERT - LDLM_FIRST_OPC);
1535 rc = req_capsule_server_pack(&req->rq_pill);
1539 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1540 dlm_rep->lock_flags = dlm_req->lock_flags;
1542 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1547 bits = lock->l_policy_data.l_inodebits.bits;
1548 new = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1549 LDLM_DEBUG(lock, "server-side convert handler START");
1551 if (ldlm_is_cancel(lock)) {
1552 LDLM_ERROR(lock, "convert on canceled lock!");
1553 rc = ELDLM_NO_LOCK_DATA;
1554 } else if (dlm_req->lock_desc.l_req_mode !=
1555 lock->l_granted_mode) {
1556 LDLM_ERROR(lock, "lock mode differs!");
1557 rc = ELDLM_NO_LOCK_DATA;
1558 } else if (bits == new) {
1559 /* This can be valid situation if CONVERT RPCs are
1560 * re-ordered. Just finish silently */
1561 LDLM_DEBUG(lock, "lock is converted already!");
1564 lock_res_and_lock(lock);
1565 if (ldlm_is_waited(lock))
1566 ldlm_del_waiting_lock(lock);
1568 ldlm_clear_cbpending(lock);
1569 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1570 ldlm_inodebits_drop(lock, bits & ~new);
1572 ldlm_clear_blocking_data(lock);
1573 unlock_res_and_lock(lock);
1575 ldlm_reprocess_all(lock->l_resource);
1579 if (rc == ELDLM_OK) {
1580 dlm_rep->lock_handle = lock->l_remote_handle;
1581 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1582 &dlm_rep->lock_desc.l_policy_data);
1585 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d",
1587 LDLM_LOCK_PUT(lock);
1589 rc = ELDLM_NO_LOCK_DATA;
1590 LDLM_DEBUG_NOLOCK("server-side convert handler END, rc = %d",
1594 req->rq_status = rc;
1600 * Cancel all the locks whose handles are packed into ldlm_request
1602 * Called by server code expecting such combined cancel activity
1605 int ldlm_request_cancel(struct ptlrpc_request *req,
1606 const struct ldlm_request *dlm_req,
1607 int first, enum lustre_at_flags flags)
1609 const struct lu_env *env = req->rq_svc_thread->t_env;
1610 struct ldlm_resource *res, *pres = NULL;
1611 struct ldlm_lock *lock;
1612 int i, count, done = 0;
1615 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1619 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1622 /* There is no lock on the server at the replay time,
1623 * skip lock cancelling to make replay tests to pass. */
1624 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1627 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1628 "starting at %d", count, first);
1630 for (i = first; i < count; i++) {
1631 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1633 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1634 "lock (cookie %llu)",
1635 dlm_req->lock_handle[i].cookie);
1639 res = lock->l_resource;
1642 /* This code is an optimization to only attempt lock
1643 * granting on the resource (that could be CPU-expensive)
1644 * after we are done cancelling lock in that resource. */
1647 ldlm_reprocess_all(pres);
1648 LDLM_RESOURCE_DELREF(pres);
1649 ldlm_resource_putref(pres);
1652 ldlm_resource_getref(res);
1653 LDLM_RESOURCE_ADDREF(res);
1655 if (!ldlm_is_discard_data(lock))
1656 ldlm_lvbo_update(env, res, lock,
1662 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1663 lock->l_blast_sent != 0) {
1664 time64_t delay = ktime_get_real_seconds() -
1666 LDLM_DEBUG(lock, "server cancels blocked lock after %llds",
1668 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1670 ldlm_lock_cancel(lock);
1671 LDLM_LOCK_PUT(lock);
1674 ldlm_reprocess_all(pres);
1675 LDLM_RESOURCE_DELREF(pres);
1676 ldlm_resource_putref(pres);
1678 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1681 EXPORT_SYMBOL(ldlm_request_cancel);
1684 * Main LDLM entry point for server code to cancel locks.
1686 * Typically gets called from service handler on LDLM_CANCEL opc.
1688 int ldlm_handle_cancel(struct ptlrpc_request *req)
1690 struct ldlm_request *dlm_req;
1694 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1695 if (dlm_req == NULL) {
1696 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1700 if (req->rq_export && req->rq_export->exp_nid_stats &&
1701 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1702 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1703 LDLM_CANCEL - LDLM_FIRST_OPC);
1705 rc = req_capsule_server_pack(&req->rq_pill);
1709 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1710 req->rq_status = LUSTRE_ESTALE;
1712 RETURN(ptlrpc_reply(req));
1714 #endif /* HAVE_SERVER_SUPPORT */
1717 * Callback handler for receiving incoming blocking ASTs.
1719 * This can only happen on client side.
1721 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1722 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1727 LDLM_DEBUG(lock, "client blocking AST callback handler");
1729 lock_res_and_lock(lock);
1731 /* set bits to cancel for this lock for possible lock convert */
1732 if (ns_is_client(ns) && (lock->l_resource->lr_type == LDLM_IBITS)) {
1733 /* Lock description contains policy of blocking lock,
1734 * and its cancel_bits is used to pass conflicting bits.
1735 * NOTE: ld can be NULL or can be not NULL but zeroed if
1736 * passed from ldlm_bl_thread_blwi(), check below used bits
1737 * in ld to make sure it is valid description.
1739 if (ld && ld->l_policy_data.l_inodebits.bits)
1740 lock->l_policy_data.l_inodebits.cancel_bits =
1741 ld->l_policy_data.l_inodebits.cancel_bits;
1742 /* if there is no valid ld and lock is cbpending already
1743 * then cancel_bits should be kept, otherwise it is zeroed.
1745 else if (!ldlm_is_cbpending(lock))
1746 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1748 ldlm_set_cbpending(lock);
1750 if (ldlm_is_cancel_on_block(lock))
1751 ldlm_set_cancel(lock);
1753 do_ast = (!lock->l_readers && !lock->l_writers);
1754 unlock_res_and_lock(lock);
1757 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1758 lock, lock->l_blocking_ast);
1759 if (lock->l_blocking_ast != NULL)
1760 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1763 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1767 LDLM_DEBUG(lock, "client blocking callback handler END");
1768 LDLM_LOCK_RELEASE(lock);
1773 * Callback handler for receiving incoming completion ASTs.
1775 * This only can happen on client side.
1777 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1778 struct ldlm_namespace *ns,
1779 struct ldlm_request *dlm_req,
1780 struct ldlm_lock *lock)
1782 struct list_head ast_list;
1787 LDLM_DEBUG(lock, "client completion callback handler START");
1789 INIT_LIST_HEAD(&ast_list);
1790 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1791 long to = cfs_time_seconds(1);
1794 set_current_state(TASK_INTERRUPTIBLE);
1795 schedule_timeout(to);
1796 if (lock->l_granted_mode == lock->l_req_mode ||
1797 ldlm_is_destroyed(lock))
1802 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1804 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1805 GOTO(out, rc = lvb_len);
1806 } else if (lvb_len > 0) {
1807 if (lock->l_lvb_len > 0) {
1808 /* for extent lock, lvb contains ost_lvb{}. */
1809 LASSERT(lock->l_lvb_data != NULL);
1811 if (unlikely(lock->l_lvb_len < lvb_len)) {
1812 LDLM_ERROR(lock, "Replied LVB is larger than "
1813 "expectation, expected = %d, "
1815 lock->l_lvb_len, lvb_len);
1816 GOTO(out, rc = -EINVAL);
1821 lock_res_and_lock(lock);
1823 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1824 &lock->l_resource->lr_name)) {
1825 ldlm_resource_unlink_lock(lock);
1826 unlock_res_and_lock(lock);
1827 rc = ldlm_lock_change_resource(ns, lock,
1828 &dlm_req->lock_desc.l_resource.lr_name);
1830 LDLM_ERROR(lock, "Failed to allocate resource");
1833 LDLM_DEBUG(lock, "completion AST, new resource");
1834 lock_res_and_lock(lock);
1837 if (ldlm_is_destroyed(lock) ||
1838 lock->l_granted_mode == lock->l_req_mode) {
1839 /* bug 11300: the lock has already been granted */
1840 unlock_res_and_lock(lock);
1841 LDLM_DEBUG(lock, "Double grant race happened");
1845 /* If we receive the completion AST before the actual enqueue returned,
1846 * then we might need to switch lock modes, resources, or extents. */
1847 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1848 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1849 LDLM_DEBUG(lock, "completion AST, new lock mode");
1852 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1853 ldlm_convert_policy_to_local(req->rq_export,
1854 dlm_req->lock_desc.l_resource.lr_type,
1855 &dlm_req->lock_desc.l_policy_data,
1856 &lock->l_policy_data);
1857 LDLM_DEBUG(lock, "completion AST, new policy data");
1860 ldlm_resource_unlink_lock(lock);
1862 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1863 /* BL_AST locks are not needed in LRU.
1864 * Let ldlm_cancel_lru() be fast. */
1865 ldlm_lock_remove_from_lru(lock);
1866 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1867 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1870 if (lock->l_lvb_len > 0) {
1871 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1872 lock->l_lvb_data, lvb_len);
1874 unlock_res_and_lock(lock);
1879 ldlm_grant_lock(lock, &ast_list);
1880 unlock_res_and_lock(lock);
1882 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1884 /* Let Enqueue to call osc_lock_upcall() and initialize
1886 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1888 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1890 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1896 lock_res_and_lock(lock);
1897 ldlm_set_failed(lock);
1898 unlock_res_and_lock(lock);
1899 wake_up(&lock->l_waitq);
1901 LDLM_LOCK_RELEASE(lock);
1905 * Callback handler for receiving incoming glimpse ASTs.
1907 * This only can happen on client side. After handling the glimpse AST
1908 * we also consider dropping the lock here if it is unused locally for a
1911 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1912 struct ldlm_namespace *ns,
1913 struct ldlm_request *dlm_req,
1914 struct ldlm_lock *lock)
1919 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1921 if (lock->l_glimpse_ast != NULL)
1922 rc = lock->l_glimpse_ast(lock, req);
1924 if (req->rq_repmsg != NULL) {
1927 req->rq_status = rc;
1931 lock_res_and_lock(lock);
1932 if (lock->l_granted_mode == LCK_PW &&
1933 !lock->l_readers && !lock->l_writers &&
1934 ktime_after(ktime_get(),
1935 ktime_add(lock->l_last_used,
1936 ktime_set(ns->ns_dirty_age_limit, 0)))) {
1937 unlock_res_and_lock(lock);
1938 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1939 ldlm_handle_bl_callback(ns, NULL, lock);
1944 unlock_res_and_lock(lock);
1945 LDLM_LOCK_RELEASE(lock);
1949 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1951 if (req->rq_no_reply)
1954 req->rq_status = rc;
1955 if (!req->rq_packed_final) {
1956 rc = lustre_pack_reply(req, 1, NULL, NULL);
1960 return ptlrpc_reply(req);
1963 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1964 enum ldlm_cancel_flags cancel_flags)
1966 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1969 spin_lock(&blp->blp_lock);
1970 if (blwi->blwi_lock &&
1971 ldlm_is_discard_data(blwi->blwi_lock)) {
1972 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1973 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1975 /* other blocking callbacks are added to the regular list */
1976 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1978 spin_unlock(&blp->blp_lock);
1980 wake_up(&blp->blp_waitq);
1982 /* can not check blwi->blwi_flags as blwi could be already freed in
1984 if (!(cancel_flags & LCF_ASYNC))
1985 wait_for_completion(&blwi->blwi_comp);
1990 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1991 struct ldlm_namespace *ns,
1992 struct ldlm_lock_desc *ld,
1993 struct list_head *cancels, int count,
1994 struct ldlm_lock *lock,
1995 enum ldlm_cancel_flags cancel_flags)
1997 init_completion(&blwi->blwi_comp);
1998 INIT_LIST_HEAD(&blwi->blwi_head);
2000 if (memory_pressure_get())
2001 blwi->blwi_mem_pressure = 1;
2004 blwi->blwi_flags = cancel_flags;
2006 blwi->blwi_ld = *ld;
2008 list_add(&blwi->blwi_head, cancels);
2009 list_del_init(cancels);
2010 blwi->blwi_count = count;
2012 blwi->blwi_lock = lock;
2017 * Queues a list of locks \a cancels containing \a count locks
2018 * for later processing by a blocking thread. If \a count is zero,
2019 * then the lock referenced as \a lock is queued instead.
2021 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2022 * If list addition fails an error is returned and caller is supposed to
2023 * call ->l_blocking_ast itself.
2025 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2026 struct ldlm_lock_desc *ld,
2027 struct ldlm_lock *lock,
2028 struct list_head *cancels, int count,
2029 enum ldlm_cancel_flags cancel_flags)
2033 if (cancels && count == 0)
2036 if (cancel_flags & LCF_ASYNC) {
2037 struct ldlm_bl_work_item *blwi;
2039 OBD_ALLOC(blwi, sizeof(*blwi));
2042 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2044 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2046 /* if it is synchronous call do minimum mem alloc, as it could
2047 * be triggered from kernel shrinker
2049 struct ldlm_bl_work_item blwi;
2051 memset(&blwi, 0, sizeof(blwi));
2052 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2053 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2058 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2059 struct ldlm_lock *lock)
2061 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2064 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2065 struct list_head *cancels, int count,
2066 enum ldlm_cancel_flags cancel_flags)
2068 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2071 int ldlm_bl_thread_wakeup(void)
2073 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2077 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2078 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2080 struct obd_device *obd = req->rq_export->exp_obd;
2087 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2089 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2091 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2093 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2096 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2098 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2100 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2103 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2106 /* We are responsible for swabbing contents of val */
2108 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2109 /* Pass it on to mdc (the "export" in this case) */
2110 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2112 sizeof(KEY_HSM_COPYTOOL_SEND),
2113 KEY_HSM_COPYTOOL_SEND,
2116 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2121 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2122 const char *msg, int rc,
2123 const struct lustre_handle *handle)
2125 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2126 "%s: [nid %s] [rc %d] [lock %#llx]",
2127 msg, libcfs_id2str(req->rq_peer), rc,
2128 handle ? handle->cookie : 0);
2129 if (req->rq_no_reply)
2130 CWARN("No reply was sent, maybe cause bug 21636.\n");
2132 CWARN("Send reply failed, maybe cause bug 21636.\n");
2135 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2136 static int ldlm_callback_handler(struct ptlrpc_request *req)
2138 struct ldlm_namespace *ns;
2139 struct ldlm_request *dlm_req;
2140 struct ldlm_lock *lock;
2144 /* Requests arrive in sender's byte order. The ptlrpc service
2145 * handler has already checked and, if necessary, byte-swapped the
2146 * incoming request message body, but I am responsible for the
2147 * message buffers. */
2149 /* do nothing for sec context finalize */
2150 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2153 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2155 if (req->rq_export == NULL) {
2156 rc = ldlm_callback_reply(req, -ENOTCONN);
2157 ldlm_callback_errmsg(req, "Operate on unconnected server",
2162 LASSERT(req->rq_export != NULL);
2163 LASSERT(req->rq_export->exp_obd != NULL);
2165 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2166 case LDLM_BL_CALLBACK:
2167 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2169 ldlm_callback_reply(req, -(int)cfs_fail_err);
2173 case LDLM_CP_CALLBACK:
2174 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2177 case LDLM_GL_CALLBACK:
2178 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2182 rc = ldlm_handle_setinfo(req);
2183 ldlm_callback_reply(req, rc);
2186 CERROR("unknown opcode %u\n",
2187 lustre_msg_get_opc(req->rq_reqmsg));
2188 ldlm_callback_reply(req, -EPROTO);
2192 ns = req->rq_export->exp_obd->obd_namespace;
2193 LASSERT(ns != NULL);
2195 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2197 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2198 if (dlm_req == NULL) {
2199 rc = ldlm_callback_reply(req, -EPROTO);
2200 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2205 /* Force a known safe race, send a cancel to the server for a lock
2206 * which the server has already started a blocking callback on. */
2207 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2208 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2209 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2211 CERROR("ldlm_cli_cancel: %d\n", rc);
2214 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2216 CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock "
2217 "disappeared\n", dlm_req->lock_handle[0].cookie);
2218 rc = ldlm_callback_reply(req, -EINVAL);
2219 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2220 &dlm_req->lock_handle[0]);
2224 if (ldlm_is_fail_loc(lock) &&
2225 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2226 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2228 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2229 lock_res_and_lock(lock);
2230 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2232 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2233 /* If somebody cancels lock and cache is already dropped,
2234 * or lock is failed before cp_ast received on client,
2235 * we can tell the server we have no lock. Otherwise, we
2236 * should send cancel after dropping the cache. */
2237 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2238 ldlm_is_failed(lock)) {
2239 LDLM_DEBUG(lock, "callback on lock %llx - lock disappeared",
2240 dlm_req->lock_handle[0].cookie);
2241 unlock_res_and_lock(lock);
2242 LDLM_LOCK_RELEASE(lock);
2243 rc = ldlm_callback_reply(req, -EINVAL);
2244 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2245 &dlm_req->lock_handle[0]);
2248 /* BL_AST locks are not needed in LRU.
2249 * Let ldlm_cancel_lru() be fast. */
2250 ldlm_lock_remove_from_lru(lock);
2251 ldlm_set_bl_ast(lock);
2253 unlock_res_and_lock(lock);
2255 /* We want the ost thread to get this reply so that it can respond
2256 * to ost requests (write cache writeback) that might be triggered
2259 * But we'd also like to be able to indicate in the reply that we're
2260 * cancelling right now, because it's unused, or have an intent result
2261 * in the reply, so we might have to push the responsibility for sending
2262 * the reply down into the AST handlers, alas. */
2264 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2265 case LDLM_BL_CALLBACK:
2266 CDEBUG(D_INODE, "blocking ast\n");
2267 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2268 if (!ldlm_is_cancel_on_block(lock)) {
2269 rc = ldlm_callback_reply(req, 0);
2270 if (req->rq_no_reply || rc)
2271 ldlm_callback_errmsg(req, "Normal process", rc,
2272 &dlm_req->lock_handle[0]);
2274 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2275 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2277 case LDLM_CP_CALLBACK:
2278 CDEBUG(D_INODE, "completion ast\n");
2279 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2280 ldlm_callback_reply(req, 0);
2281 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2283 case LDLM_GL_CALLBACK:
2284 CDEBUG(D_INODE, "glimpse ast\n");
2285 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2286 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2289 LBUG(); /* checked above */
2295 #ifdef HAVE_SERVER_SUPPORT
2297 * Main handler for canceld thread.
2299 * Separated into its own thread to avoid deadlocks.
2301 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2307 /* Requests arrive in sender's byte order. The ptlrpc service
2308 * handler has already checked and, if necessary, byte-swapped the
2309 * incoming request message body, but I am responsible for the
2310 * message buffers. */
2312 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2314 if (req->rq_export == NULL) {
2315 struct ldlm_request *dlm_req;
2317 CERROR("%s from %s arrived at %lu with bad export cookie "
2319 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2320 libcfs_nid2str(req->rq_peer.nid),
2321 req->rq_arrival_time.tv_sec,
2322 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2324 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2325 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2326 dlm_req = req_capsule_client_get(&req->rq_pill,
2328 if (dlm_req != NULL)
2329 ldlm_lock_dump_handle(D_ERROR,
2330 &dlm_req->lock_handle[0]);
2332 ldlm_callback_reply(req, -ENOTCONN);
2336 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2337 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2339 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2340 CDEBUG(D_INODE, "cancel\n");
2341 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2342 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2343 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2345 rc = ldlm_handle_cancel(req);
2349 struct ldlm_request *dlm_req;
2351 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2352 CDEBUG(D_INODE, "convert\n");
2354 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2355 if (dlm_req == NULL) {
2356 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2357 rc = ldlm_callback_reply(req, -EPROTO);
2359 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2360 rc = ptlrpc_reply(req);
2365 CERROR("invalid opcode %d\n",
2366 lustre_msg_get_opc(req->rq_reqmsg));
2367 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2368 rc = ldlm_callback_reply(req, -EINVAL);
2374 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2375 struct ldlm_lock *lock)
2377 struct ldlm_request *dlm_req;
2378 struct lustre_handle lockh;
2384 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2385 if (dlm_req == NULL)
2388 ldlm_lock2handle(lock, &lockh);
2389 for (i = 0; i < dlm_req->lock_count; i++) {
2390 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2392 DEBUG_REQ(D_RPCTRACE, req,
2393 "Prio raised by lock %#llx.", lockh.cookie);
2402 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2404 struct ldlm_request *dlm_req;
2410 /* no prolong in recovery */
2411 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2414 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2415 if (dlm_req == NULL)
2418 for (i = 0; i < dlm_req->lock_count; i++) {
2419 struct ldlm_lock *lock;
2421 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2425 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2427 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2428 LDLM_LOCK_PUT(lock);
2437 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2438 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2439 .hpreq_check = ldlm_cancel_hpreq_check,
2443 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2447 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2449 if (req->rq_export == NULL)
2452 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2453 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2454 req->rq_ops = &ldlm_cancel_hpreq_ops;
2455 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2456 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2457 req->rq_ops = &ldlm_cancel_hpreq_ops;
2462 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2463 struct hlist_node *hnode, void *data)
2466 struct list_head *rpc_list = data;
2467 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2469 lock_res_and_lock(lock);
2471 if (lock->l_req_mode != lock->l_granted_mode) {
2472 unlock_res_and_lock(lock);
2476 LASSERT(lock->l_resource);
2477 if (lock->l_resource->lr_type != LDLM_IBITS &&
2478 lock->l_resource->lr_type != LDLM_PLAIN) {
2479 unlock_res_and_lock(lock);
2483 if (ldlm_is_ast_sent(lock)) {
2484 unlock_res_and_lock(lock);
2488 LASSERT(lock->l_blocking_ast);
2489 LASSERT(!lock->l_blocking_lock);
2491 ldlm_set_ast_sent(lock);
2492 if (lock->l_export && lock->l_export->exp_lock_hash) {
2493 /* NB: it's safe to call cfs_hash_del() even lock isn't
2494 * in exp_lock_hash. */
2495 /* In the function below, .hs_keycmp resolves to
2496 * ldlm_export_lock_keycmp() */
2497 /* coverity[overrun-buffer-val] */
2498 cfs_hash_del(lock->l_export->exp_lock_hash,
2499 &lock->l_remote_handle, &lock->l_exp_hash);
2502 list_add_tail(&lock->l_rk_ast, rpc_list);
2503 LDLM_LOCK_GET(lock);
2505 unlock_res_and_lock(lock);
2509 void ldlm_revoke_export_locks(struct obd_export *exp)
2511 struct list_head rpc_list;
2514 INIT_LIST_HEAD(&rpc_list);
2515 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2516 ldlm_revoke_lock_cb, &rpc_list, 0);
2517 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2518 LDLM_WORK_REVOKE_AST);
2522 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2523 #endif /* HAVE_SERVER_SUPPORT */
2525 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2526 struct ldlm_bl_work_item **p_blwi,
2527 struct obd_export **p_exp)
2529 struct ldlm_bl_work_item *blwi = NULL;
2530 static unsigned int num_bl = 0;
2531 static unsigned int num_stale;
2532 int num_th = atomic_read(&blp->blp_num_threads);
2534 *p_exp = obd_stale_export_get();
2536 spin_lock(&blp->blp_lock);
2537 if (*p_exp != NULL) {
2538 if (num_th == 1 || ++num_stale < num_th) {
2539 spin_unlock(&blp->blp_lock);
2546 /* process a request from the blp_list at least every blp_num_threads */
2547 if (!list_empty(&blp->blp_list) &&
2548 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2549 blwi = list_entry(blp->blp_list.next,
2550 struct ldlm_bl_work_item, blwi_entry);
2552 if (!list_empty(&blp->blp_prio_list))
2553 blwi = list_entry(blp->blp_prio_list.next,
2554 struct ldlm_bl_work_item,
2558 if (++num_bl >= num_th)
2560 list_del(&blwi->blwi_entry);
2562 spin_unlock(&blp->blp_lock);
2565 if (*p_exp != NULL && *p_blwi != NULL) {
2566 obd_stale_export_put(*p_exp);
2570 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2573 /* This only contains temporary data until the thread starts */
2574 struct ldlm_bl_thread_data {
2575 struct ldlm_bl_pool *bltd_blp;
2576 struct completion bltd_comp;
2580 static int ldlm_bl_thread_main(void *arg);
2582 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2584 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2585 struct task_struct *task;
2587 init_completion(&bltd.bltd_comp);
2589 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2590 if (bltd.bltd_num >= blp->blp_max_threads) {
2591 atomic_dec(&blp->blp_num_threads);
2595 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2597 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2598 atomic_dec(&blp->blp_num_threads);
2602 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2605 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2606 bltd.bltd_num, PTR_ERR(task));
2607 atomic_dec(&blp->blp_num_threads);
2608 return PTR_ERR(task);
2610 wait_for_completion(&bltd.bltd_comp);
2615 /* Not fatal if racy and have a few too many threads */
2616 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2617 struct ldlm_bl_work_item *blwi)
2619 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2622 if (atomic_read(&blp->blp_busy_threads) <
2623 atomic_read(&blp->blp_num_threads))
2626 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2627 blwi->blwi_mem_pressure))
2633 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2634 struct ldlm_bl_work_item *blwi)
2638 if (blwi->blwi_ns == NULL)
2639 /* added by ldlm_cleanup() */
2640 RETURN(LDLM_ITER_STOP);
2642 if (blwi->blwi_mem_pressure)
2643 memory_pressure_set();
2645 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2647 if (blwi->blwi_count) {
2649 /* The special case when we cancel locks in lru
2650 * asynchronously, we pass the list of locks here.
2651 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2652 * canceled locally yet. */
2653 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2656 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2659 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2662 if (blwi->blwi_mem_pressure)
2663 memory_pressure_clr();
2665 if (blwi->blwi_flags & LCF_ASYNC)
2666 OBD_FREE(blwi, sizeof(*blwi));
2668 complete(&blwi->blwi_comp);
2674 * Cancel stale locks on export. Cancel blocked locks first.
2675 * If the given export has blocked locks, the next in the list may have
2676 * them too, thus cancel not blocked locks only if the current export has
2679 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2680 struct obd_export *exp)
2685 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2687 num = ldlm_export_cancel_blocked_locks(exp);
2689 ldlm_export_cancel_locks(exp);
2691 obd_stale_export_put(exp);
2698 * Main blocking requests processing thread.
2700 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2701 * This thread in the end ends up doing actual call to ->l_blocking_ast
2704 static int ldlm_bl_thread_main(void *arg)
2706 struct ldlm_bl_pool *blp;
2707 struct ldlm_bl_thread_data *bltd = arg;
2710 blp = bltd->bltd_blp;
2712 complete(&bltd->bltd_comp);
2713 /* cannot use bltd after this, it is only on caller's stack */
2716 struct l_wait_info lwi = { 0 };
2717 struct ldlm_bl_work_item *blwi = NULL;
2718 struct obd_export *exp = NULL;
2721 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2724 l_wait_event_exclusive(blp->blp_waitq,
2725 ldlm_bl_get_work(blp, &blwi,
2728 atomic_inc(&blp->blp_busy_threads);
2730 if (ldlm_bl_thread_need_create(blp, blwi))
2731 /* discard the return value, we tried */
2732 ldlm_bl_thread_start(blp, true);
2735 rc = ldlm_bl_thread_exports(blp, exp);
2737 rc = ldlm_bl_thread_blwi(blp, blwi);
2739 atomic_dec(&blp->blp_busy_threads);
2741 if (rc == LDLM_ITER_STOP)
2744 /* If there are many namespaces, we will not sleep waiting for
2745 * work, and must do a cond_resched to avoid holding the CPU
2750 atomic_dec(&blp->blp_num_threads);
2751 complete(&blp->blp_comp);
2756 static int ldlm_setup(void);
2757 static int ldlm_cleanup(void);
2759 int ldlm_get_ref(void)
2763 mutex_lock(&ldlm_ref_mutex);
2764 if (++ldlm_refcount == 1) {
2769 mutex_unlock(&ldlm_ref_mutex);
2774 void ldlm_put_ref(void)
2777 mutex_lock(&ldlm_ref_mutex);
2778 if (ldlm_refcount == 1) {
2779 int rc = ldlm_cleanup();
2781 CERROR("ldlm_cleanup failed: %d\n", rc);
2787 mutex_unlock(&ldlm_ref_mutex);
2793 * Export handle<->lock hash operations.
2796 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2798 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2802 ldlm_export_lock_key(struct hlist_node *hnode)
2804 struct ldlm_lock *lock;
2806 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2807 return &lock->l_remote_handle;
2811 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2813 struct ldlm_lock *lock;
2815 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2816 lock->l_remote_handle = *(struct lustre_handle *)key;
2820 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2822 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2826 ldlm_export_lock_object(struct hlist_node *hnode)
2828 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2832 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2834 struct ldlm_lock *lock;
2836 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2837 LDLM_LOCK_GET(lock);
2841 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2843 struct ldlm_lock *lock;
2845 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2846 LDLM_LOCK_RELEASE(lock);
2849 static struct cfs_hash_ops ldlm_export_lock_ops = {
2850 .hs_hash = ldlm_export_lock_hash,
2851 .hs_key = ldlm_export_lock_key,
2852 .hs_keycmp = ldlm_export_lock_keycmp,
2853 .hs_keycpy = ldlm_export_lock_keycpy,
2854 .hs_object = ldlm_export_lock_object,
2855 .hs_get = ldlm_export_lock_get,
2856 .hs_put = ldlm_export_lock_put,
2857 .hs_put_locked = ldlm_export_lock_put,
2860 int ldlm_init_export(struct obd_export *exp)
2865 exp->exp_lock_hash =
2866 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2867 HASH_EXP_LOCK_CUR_BITS,
2868 HASH_EXP_LOCK_MAX_BITS,
2869 HASH_EXP_LOCK_BKT_BITS, 0,
2870 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2871 &ldlm_export_lock_ops,
2872 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2873 CFS_HASH_NBLK_CHANGE);
2875 if (!exp->exp_lock_hash)
2878 rc = ldlm_init_flock_export(exp);
2884 ldlm_destroy_export(exp);
2887 EXPORT_SYMBOL(ldlm_init_export);
2889 void ldlm_destroy_export(struct obd_export *exp)
2892 cfs_hash_putref(exp->exp_lock_hash);
2893 exp->exp_lock_hash = NULL;
2895 ldlm_destroy_flock_export(exp);
2898 EXPORT_SYMBOL(ldlm_destroy_export);
2900 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
2901 struct attribute *attr,
2904 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
2907 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
2908 struct attribute *attr,
2915 rc = kstrtoul(buffer, 10, &val);
2919 ldlm_cancel_unused_locks_before_replay = val;
2923 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
2925 static struct attribute *ldlm_attrs[] = {
2926 &lustre_attr_cancel_unused_locks_before_replay.attr,
2930 static struct attribute_group ldlm_attr_group = {
2931 .attrs = ldlm_attrs,
2934 static int ldlm_setup(void)
2936 static struct ptlrpc_service_conf conf;
2937 struct ldlm_bl_pool *blp = NULL;
2938 #ifdef HAVE_SERVER_SUPPORT
2939 struct task_struct *task;
2940 #endif /* HAVE_SERVER_SUPPORT */
2946 if (ldlm_state != NULL)
2949 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2950 if (ldlm_state == NULL)
2953 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
2957 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
2961 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
2965 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
2969 rc = ldlm_debugfs_setup();
2973 memset(&conf, 0, sizeof(conf));
2974 conf = (typeof(conf)) {
2975 .psc_name = "ldlm_cbd",
2976 .psc_watchdog_factor = 2,
2978 .bc_nbufs = LDLM_CLIENT_NBUFS,
2979 .bc_buf_size = LDLM_BUFSIZE,
2980 .bc_req_max_size = LDLM_MAXREQSIZE,
2981 .bc_rep_max_size = LDLM_MAXREPSIZE,
2982 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2983 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2986 .tc_thr_name = "ldlm_cb",
2987 .tc_thr_factor = LDLM_THR_FACTOR,
2988 .tc_nthrs_init = LDLM_NTHRS_INIT,
2989 .tc_nthrs_base = LDLM_NTHRS_BASE,
2990 .tc_nthrs_max = LDLM_NTHRS_MAX,
2991 .tc_nthrs_user = ldlm_num_threads,
2992 .tc_cpu_bind = ldlm_cpu_bind,
2993 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2996 .cc_pattern = ldlm_cpts,
2997 .cc_affinity = true,
3000 .so_req_handler = ldlm_callback_handler,
3003 ldlm_state->ldlm_cb_service = \
3004 ptlrpc_register_service(&conf, ldlm_svc_kset,
3005 ldlm_svc_debugfs_dir);
3006 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3007 CERROR("failed to start service\n");
3008 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3009 ldlm_state->ldlm_cb_service = NULL;
3013 #ifdef HAVE_SERVER_SUPPORT
3014 memset(&conf, 0, sizeof(conf));
3015 conf = (typeof(conf)) {
3016 .psc_name = "ldlm_canceld",
3017 .psc_watchdog_factor = 6,
3019 .bc_nbufs = LDLM_SERVER_NBUFS,
3020 .bc_buf_size = LDLM_BUFSIZE,
3021 .bc_req_max_size = LDLM_MAXREQSIZE,
3022 .bc_rep_max_size = LDLM_MAXREPSIZE,
3023 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3024 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3028 .tc_thr_name = "ldlm_cn",
3029 .tc_thr_factor = LDLM_THR_FACTOR,
3030 .tc_nthrs_init = LDLM_NTHRS_INIT,
3031 .tc_nthrs_base = LDLM_NTHRS_BASE,
3032 .tc_nthrs_max = LDLM_NTHRS_MAX,
3033 .tc_nthrs_user = ldlm_num_threads,
3034 .tc_cpu_bind = ldlm_cpu_bind,
3035 .tc_ctx_tags = LCT_MD_THREAD | \
3040 .cc_pattern = ldlm_cpts,
3041 .cc_affinity = true,
3044 .so_req_handler = ldlm_cancel_handler,
3045 .so_hpreq_handler = ldlm_hpreq_handler,
3048 ldlm_state->ldlm_cancel_service = \
3049 ptlrpc_register_service(&conf, ldlm_svc_kset,
3050 ldlm_svc_debugfs_dir);
3051 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3052 CERROR("failed to start service\n");
3053 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3054 ldlm_state->ldlm_cancel_service = NULL;
3057 #endif /* HAVE_SERVER_SUPPORT */
3059 OBD_ALLOC(blp, sizeof(*blp));
3061 GOTO(out, rc = -ENOMEM);
3062 ldlm_state->ldlm_bl_pool = blp;
3064 spin_lock_init(&blp->blp_lock);
3065 INIT_LIST_HEAD(&blp->blp_list);
3066 INIT_LIST_HEAD(&blp->blp_prio_list);
3067 init_waitqueue_head(&blp->blp_waitq);
3068 atomic_set(&blp->blp_num_threads, 0);
3069 atomic_set(&blp->blp_busy_threads, 0);
3071 if (ldlm_num_threads == 0) {
3072 blp->blp_min_threads = LDLM_NTHRS_INIT;
3073 blp->blp_max_threads = LDLM_NTHRS_MAX;
3075 blp->blp_min_threads = blp->blp_max_threads = \
3076 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3080 for (i = 0; i < blp->blp_min_threads; i++) {
3081 rc = ldlm_bl_thread_start(blp, false);
3086 #ifdef HAVE_SERVER_SUPPORT
3087 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3090 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3094 wait_event(expired_lock_wait_queue,
3095 expired_lock_thread_state == ELT_READY);
3096 #endif /* HAVE_SERVER_SUPPORT */
3098 rc = ldlm_pools_init();
3100 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3104 rc = ldlm_reclaim_setup();
3106 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3116 static int ldlm_cleanup(void)
3120 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3121 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3122 CERROR("ldlm still has namespaces; clean these up first.\n");
3123 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3124 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3128 ldlm_reclaim_cleanup();
3131 if (ldlm_state->ldlm_bl_pool != NULL) {
3132 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3134 while (atomic_read(&blp->blp_num_threads) > 0) {
3135 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3137 init_completion(&blp->blp_comp);
3139 spin_lock(&blp->blp_lock);
3140 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3141 wake_up(&blp->blp_waitq);
3142 spin_unlock(&blp->blp_lock);
3144 wait_for_completion(&blp->blp_comp);
3147 OBD_FREE(blp, sizeof(*blp));
3150 if (ldlm_state->ldlm_cb_service != NULL)
3151 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3152 #ifdef HAVE_SERVER_SUPPORT
3153 if (ldlm_state->ldlm_cancel_service != NULL)
3154 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3158 kset_unregister(ldlm_ns_kset);
3160 kset_unregister(ldlm_svc_kset);
3162 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3163 kobject_put(ldlm_kobj);
3166 ldlm_debugfs_cleanup();
3168 #ifdef HAVE_SERVER_SUPPORT
3169 if (expired_lock_thread_state != ELT_STOPPED) {
3170 expired_lock_thread_state = ELT_TERMINATE;
3171 wake_up(&expired_lock_wait_queue);
3172 wait_event(expired_lock_wait_queue,
3173 expired_lock_thread_state == ELT_STOPPED);
3177 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3185 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3186 sizeof(struct ldlm_resource), 0,
3187 SLAB_HWCACHE_ALIGN, NULL);
3188 if (ldlm_resource_slab == NULL)
3191 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3192 sizeof(struct ldlm_lock), 0,
3193 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3194 if (ldlm_lock_slab == NULL)
3197 ldlm_interval_slab = kmem_cache_create("interval_node",
3198 sizeof(struct ldlm_interval),
3199 0, SLAB_HWCACHE_ALIGN, NULL);
3200 if (ldlm_interval_slab == NULL)
3203 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3204 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3205 0, SLAB_HWCACHE_ALIGN, NULL);
3206 if (ldlm_interval_tree_slab == NULL)
3209 #ifdef HAVE_SERVER_SUPPORT
3210 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3211 sizeof(struct ldlm_glimpse_work),
3213 if (ldlm_glimpse_work_kmem == NULL)
3214 goto out_interval_tree;
3217 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3218 class_export_dump_hook = ldlm_dump_export_locks;
3221 #ifdef HAVE_SERVER_SUPPORT
3223 kmem_cache_destroy(ldlm_interval_tree_slab);
3226 kmem_cache_destroy(ldlm_interval_slab);
3228 kmem_cache_destroy(ldlm_lock_slab);
3230 kmem_cache_destroy(ldlm_resource_slab);
3235 void ldlm_exit(void)
3238 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3239 kmem_cache_destroy(ldlm_resource_slab);
3240 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3241 * synchronize_rcu() to wait a grace period elapsed, so that
3242 * ldlm_lock_free() get a chance to be called. */
3244 kmem_cache_destroy(ldlm_lock_slab);
3245 kmem_cache_destroy(ldlm_interval_slab);
3246 kmem_cache_destroy(ldlm_interval_tree_slab);
3247 #ifdef HAVE_SERVER_SUPPORT
3248 kmem_cache_destroy(ldlm_glimpse_work_kmem);