4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lockd.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <linux/kthread.h>
41 #include <linux/list.h>
42 #include <libcfs/libcfs.h>
43 #include <lustre_errno.h>
44 #include <lustre_dlm.h>
45 #include <obd_class.h>
46 #include "ldlm_internal.h"
48 static int ldlm_num_threads;
49 module_param(ldlm_num_threads, int, 0444);
50 MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start");
52 static unsigned int ldlm_cpu_bind = 1;
53 module_param(ldlm_cpu_bind, uint, 0444);
54 MODULE_PARM_DESC(ldlm_cpu_bind,
55 "bind DLM service threads to particular CPU partitions");
57 static char *ldlm_cpts;
58 module_param(ldlm_cpts, charp, 0444);
59 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
61 static DEFINE_MUTEX(ldlm_ref_mutex);
62 static int ldlm_refcount;
64 struct kobject *ldlm_kobj;
65 struct kset *ldlm_ns_kset;
66 struct kset *ldlm_svc_kset;
70 static struct ldlm_state *ldlm_state;
73 * timeout for initial callback (AST) reply (bz10399)
74 * Due to having to send a 32 bit time value over the
75 * wire return it as timeout_t instead of time64_t
77 static inline timeout_t ldlm_get_rq_timeout(void)
80 timeout_t timeout = min(ldlm_timeout, obd_timeout / 3);
82 return timeout < 1 ? 1 : timeout;
89 * blp_prio_list is used for callbacks that should be handled
90 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
93 struct list_head blp_prio_list;
96 * blp_list is used for all other callbacks which are likely
97 * to take longer to process.
99 struct list_head blp_list;
101 wait_queue_head_t blp_waitq;
102 struct completion blp_comp;
103 atomic_t blp_num_threads;
104 atomic_t blp_busy_threads;
109 struct ldlm_bl_work_item {
110 struct list_head blwi_entry;
111 struct ldlm_namespace *blwi_ns;
112 struct ldlm_lock_desc blwi_ld;
113 struct ldlm_lock *blwi_lock;
114 struct list_head blwi_head;
116 struct completion blwi_comp;
117 enum ldlm_cancel_flags blwi_flags;
118 int blwi_mem_pressure;
121 #ifdef HAVE_SERVER_SUPPORT
124 * Protects both waiting_locks_list and expired_lock_thread.
126 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
129 * List for contended locks.
131 * As soon as a lock is contended, it gets placed on this list and
132 * expected time to get a response is filled in the lock. A special
133 * thread walks the list looking for locks that should be released and
134 * schedules client evictions for those that have not been released in
137 * All access to it should be under waiting_locks_spinlock.
139 static LIST_HEAD(waiting_locks_list);
140 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
141 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
149 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
150 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
151 static int expired_lock_dump;
152 static LIST_HEAD(expired_lock_list);
154 static int ldlm_lock_busy(struct ldlm_lock *lock);
155 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout);
156 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout);
158 static inline int have_expired_locks(void)
163 spin_lock_bh(&waiting_locks_spinlock);
164 need_to_run = !list_empty(&expired_lock_list);
165 spin_unlock_bh(&waiting_locks_spinlock);
171 * Check expired lock list for expired locks and time them out.
173 static int expired_lock_main(void *arg)
175 struct list_head *expired = &expired_lock_list;
180 expired_lock_thread_state = ELT_READY;
181 wake_up(&expired_lock_wait_queue);
184 wait_event_idle(expired_lock_wait_queue,
185 have_expired_locks() ||
186 expired_lock_thread_state == ELT_TERMINATE);
188 spin_lock_bh(&waiting_locks_spinlock);
189 if (expired_lock_dump) {
190 spin_unlock_bh(&waiting_locks_spinlock);
192 /* from waiting_locks_callback, but not in timer */
193 libcfs_debug_dumplog();
195 spin_lock_bh(&waiting_locks_spinlock);
196 expired_lock_dump = 0;
201 while (!list_empty(expired)) {
202 struct obd_export *export;
203 struct ldlm_lock *lock;
205 lock = list_entry(expired->next, struct ldlm_lock,
207 if ((void *)lock < LP_POISON + PAGE_SIZE &&
208 (void *)lock >= LP_POISON) {
209 spin_unlock_bh(&waiting_locks_spinlock);
210 CERROR("free lock on elt list %p\n", lock);
213 list_del_init(&lock->l_pending_chain);
214 if ((void *)lock->l_export <
215 LP_POISON + PAGE_SIZE &&
216 (void *)lock->l_export >= LP_POISON) {
217 CERROR("lock with free export on elt list %p\n",
219 lock->l_export = NULL;
220 LDLM_ERROR(lock, "free export");
222 * release extra ref grabbed by
223 * ldlm_add_waiting_lock() or
226 LDLM_LOCK_RELEASE(lock);
230 if (ldlm_is_destroyed(lock)) {
232 * release the lock refcount where
233 * waiting_locks_callback() founds
235 LDLM_LOCK_RELEASE(lock);
238 export = class_export_lock_get(lock->l_export, lock);
239 spin_unlock_bh(&waiting_locks_spinlock);
241 /* Check if we need to prolong timeout */
242 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
243 lock->l_callback_timestamp != 0 && /* not AST error */
244 ldlm_lock_busy(lock)) {
245 LDLM_DEBUG(lock, "prolong the busy lock");
246 lock_res_and_lock(lock);
247 ldlm_add_waiting_lock(lock,
248 ldlm_bl_timeout(lock) >> 1);
249 unlock_res_and_lock(lock);
251 spin_lock_bh(&export->exp_bl_list_lock);
252 list_del_init(&lock->l_exp_list);
253 spin_unlock_bh(&export->exp_bl_list_lock);
256 "lock callback timer expired after %llds: evicting client at %s ",
257 ktime_get_real_seconds() -
259 obd_export_nid2str(export));
260 ldlm_lock_to_ns(lock)->ns_timeouts++;
262 class_fail_export(export);
264 class_export_lock_put(export, lock);
266 * release extra ref grabbed by ldlm_add_waiting_lock()
267 * or ldlm_failed_ast()
269 LDLM_LOCK_RELEASE(lock);
271 spin_lock_bh(&waiting_locks_spinlock);
273 spin_unlock_bh(&waiting_locks_spinlock);
275 if (do_dump && obd_dump_on_eviction) {
276 CERROR("dump the log upon eviction\n");
277 libcfs_debug_dumplog();
280 if (expired_lock_thread_state == ELT_TERMINATE)
284 expired_lock_thread_state = ELT_STOPPED;
285 wake_up(&expired_lock_wait_queue);
290 * Check if there is a request in the export request list
291 * which prevents the lock canceling.
293 static int ldlm_lock_busy(struct ldlm_lock *lock)
295 struct ptlrpc_request *req;
300 if (lock->l_export == NULL)
303 spin_lock(&lock->l_export->exp_rpc_lock);
304 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
306 if (req->rq_ops->hpreq_lock_match) {
307 match = req->rq_ops->hpreq_lock_match(req, lock);
312 spin_unlock(&lock->l_export->exp_rpc_lock);
316 /* This is called from within a timer interrupt and cannot schedule */
317 static void waiting_locks_callback(TIMER_DATA_TYPE unused)
319 struct ldlm_lock *lock;
322 spin_lock_bh(&waiting_locks_spinlock);
323 while (!list_empty(&waiting_locks_list)) {
324 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
326 if (lock->l_callback_timestamp > ktime_get_seconds() ||
327 lock->l_req_mode == LCK_GROUP)
331 * no needs to take an extra ref on the lock since it was in
332 * the waiting_locks_list and ldlm_add_waiting_lock()
333 * already grabbed a ref
335 list_move(&lock->l_pending_chain, &expired_lock_list);
339 if (!list_empty(&expired_lock_list)) {
340 if (obd_dump_on_timeout && need_dump)
341 expired_lock_dump = __LINE__;
343 wake_up(&expired_lock_wait_queue);
347 * Make sure the timer will fire again if we have any locks
350 if (!list_empty(&waiting_locks_list)) {
351 time64_t now = ktime_get_seconds();
354 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
356 if (lock->l_callback_timestamp - now > 0)
357 delta = lock->l_callback_timestamp - now;
358 mod_timer(&waiting_locks_timer,
359 jiffies + cfs_time_seconds(delta));
361 spin_unlock_bh(&waiting_locks_spinlock);
365 * Add lock to the list of contended locks.
367 * Indicate that we're waiting for a client to call us back cancelling a given
368 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
369 * timer to fire appropriately. (We round up to the next second, to avoid
370 * floods of timer firings during periods of high lock contention and traffic).
371 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
372 * if it has been added to the waiting list (1 is returned).
374 * Called with the namespace lock held.
376 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t delay)
378 unsigned long timeout_jiffies = jiffies;
379 time64_t now = ktime_get_seconds();
383 if (!list_empty(&lock->l_pending_chain))
386 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
387 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
390 deadline = now + delay;
391 if (likely(deadline > lock->l_callback_timestamp))
392 lock->l_callback_timestamp = deadline;
394 timeout = clamp_t(timeout_t, lock->l_callback_timestamp - now,
396 timeout_jiffies += cfs_time_seconds(timeout);
398 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
399 !timer_pending(&waiting_locks_timer))
400 mod_timer(&waiting_locks_timer, timeout_jiffies);
403 * if the new lock has a shorter timeout than something earlier on
404 * the list, we'll wait the longer amount of time; no big deal.
407 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
411 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
413 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
414 if (list_empty(&lock->l_exp_list)) {
415 if (!ldlm_is_granted(lock))
416 list_add_tail(&lock->l_exp_list,
417 &lock->l_export->exp_bl_list);
419 list_add(&lock->l_exp_list,
420 &lock->l_export->exp_bl_list);
422 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
425 * A blocked lock is added. Adjust the position in
426 * the stale list if the export is in the list.
427 * If export is stale and not in the list - it is being
428 * processed and will be placed on the right position
429 * on obd_stale_export_put().
431 if (!list_empty(&lock->l_export->exp_stale_list))
432 obd_stale_export_adjust(lock->l_export);
435 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
439 /* NB: must be called with hold of lock_res_and_lock() */
440 LASSERT(ldlm_is_res_locked(lock));
441 LASSERT(!ldlm_is_cancel_on_block(lock));
444 * Do not put cross-MDT lock in the waiting list, since we
445 * will not evict it due to timeout for now
447 if (lock->l_export != NULL &&
448 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
451 spin_lock_bh(&waiting_locks_spinlock);
452 if (ldlm_is_cancel(lock)) {
453 spin_unlock_bh(&waiting_locks_spinlock);
457 if (ldlm_is_destroyed(lock)) {
458 static time64_t next;
460 spin_unlock_bh(&waiting_locks_spinlock);
461 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
462 if (ktime_get_seconds() > next) {
463 next = ktime_get_seconds() + 14400;
464 libcfs_debug_dumpstack(NULL);
469 ldlm_set_waited(lock);
470 lock->l_blast_sent = ktime_get_real_seconds();
471 ret = __ldlm_add_waiting_lock(lock, timeout);
474 * grab ref on the lock if it has been added to the
479 spin_unlock_bh(&waiting_locks_spinlock);
482 ldlm_add_blocked_lock(lock);
484 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
485 ret == 0 ? "not re-" : "", timeout,
486 AT_OFF ? "off" : "on");
491 * Remove a lock from the pending list, likely because it had its cancellation
492 * callback arrive without incident. This adjusts the lock-timeout timer if
493 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
494 * As done by ldlm_del_waiting_lock(), the caller must release the lock
495 * reference when the lock is removed from any list (1 is returned).
497 * Called with namespace lock held.
499 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
501 struct list_head *list_next;
503 if (list_empty(&lock->l_pending_chain))
506 list_next = lock->l_pending_chain.next;
507 if (lock->l_pending_chain.prev == &waiting_locks_list) {
508 /* Removing the head of the list, adjust timer. */
509 if (list_next == &waiting_locks_list) {
510 /* No more, just cancel. */
511 del_timer(&waiting_locks_timer);
513 time64_t now = ktime_get_seconds();
514 struct ldlm_lock *next;
517 next = list_entry(list_next, struct ldlm_lock,
519 if (next->l_callback_timestamp - now > 0)
520 delta = lock->l_callback_timestamp - now;
522 mod_timer(&waiting_locks_timer,
523 jiffies + cfs_time_seconds(delta));
526 list_del_init(&lock->l_pending_chain);
531 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
535 if (lock->l_export == NULL) {
536 /* We don't have a "waiting locks list" on clients. */
537 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
541 spin_lock_bh(&waiting_locks_spinlock);
542 ret = __ldlm_del_waiting_lock(lock);
543 ldlm_clear_waited(lock);
544 spin_unlock_bh(&waiting_locks_spinlock);
546 /* remove the lock out of export blocking list */
547 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
548 list_del_init(&lock->l_exp_list);
549 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
553 * release lock ref if it has indeed been removed
556 LDLM_LOCK_RELEASE(lock);
559 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
564 * Prolong the contended lock waiting time.
566 * Called with namespace lock held.
568 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
570 if (lock->l_export == NULL) {
571 /* We don't have a "waiting locks list" on clients. */
572 LDLM_DEBUG(lock, "client lock: no-op");
576 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
577 /* We don't have a "waiting locks list" on OSP. */
578 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
582 spin_lock_bh(&waiting_locks_spinlock);
584 if (list_empty(&lock->l_pending_chain)) {
585 spin_unlock_bh(&waiting_locks_spinlock);
586 LDLM_DEBUG(lock, "wasn't waiting");
591 * we remove/add the lock to the waiting list, so no needs to
592 * release/take a lock reference
594 __ldlm_del_waiting_lock(lock);
595 __ldlm_add_waiting_lock(lock, timeout);
596 spin_unlock_bh(&waiting_locks_spinlock);
598 LDLM_DEBUG(lock, "refreshed");
601 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
603 #else /* HAVE_SERVER_SUPPORT */
605 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
610 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
615 #endif /* !HAVE_SERVER_SUPPORT */
617 #ifdef HAVE_SERVER_SUPPORT
620 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
621 * lock cancel, and their replies). Used for lock callback timeout and AST
624 * \param[in] lock lock which is getting the blocking callback
626 * \retval timeout in seconds to wait for the client reply
628 timeout_t ldlm_bl_timeout(struct ldlm_lock *lock)
633 return obd_timeout / 2;
636 * Since these are non-updating timeouts, we should be conservative.
637 * Take more than usually, 150%
638 * It would be nice to have some kind of "early reply" mechanism for
639 * lock callbacks too...
641 timeout = at_get(&lock->l_export->exp_bl_lock_at);
642 return max_t(timeout_t, timeout + (timeout >> 1),
643 (timeout_t)ldlm_enqueue_min);
645 EXPORT_SYMBOL(ldlm_bl_timeout);
648 * Perform lock cleanup if AST sending failed.
650 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
651 const char *ast_type)
653 LCONSOLE_ERROR_MSG(0x138,
654 "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
655 lock->l_export->exp_obd->obd_name,
656 obd_export_nid2str(lock->l_export), ast_type, rc);
658 if (obd_dump_on_timeout)
659 libcfs_debug_dumplog();
660 spin_lock_bh(&waiting_locks_spinlock);
661 if (__ldlm_del_waiting_lock(lock) == 0)
663 * the lock was not in any list, grab an extra ref before adding
664 * the lock to the expired list
667 /* differentiate it from expired locks */
668 lock->l_callback_timestamp = 0;
669 list_add(&lock->l_pending_chain, &expired_lock_list);
670 wake_up(&expired_lock_wait_queue);
671 spin_unlock_bh(&waiting_locks_spinlock);
675 * Perform lock cleanup if AST reply came with error.
677 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
678 struct ptlrpc_request *req, int rc,
679 const char *ast_type)
681 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
683 if (!req->rq_replied || (rc && rc != -EINVAL)) {
684 if (ldlm_is_cancel(lock)) {
686 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
687 ast_type, req, req->rq_xid,
688 libcfs_nid2str(peer.nid));
689 ldlm_lock_cancel(lock);
691 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
693 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
695 * Upon umount process the AST fails because cannot be
696 * sent. This shouldn't lead to the client eviction.
697 * -ENODEV error is returned by ptl_send_rpc() for
698 * new request in such import.
699 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
700 * if imp_invalid is set or obd_no_recov.
701 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
702 * in ptlrpc_import_delay_req() as well with -EIO code.
703 * In all such cases errors are ignored.
706 "%s AST can't be sent due to a server %s failure or umount process: rc = %d\n",
708 req->rq_import->imp_obd->obd_name, rc);
711 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
712 libcfs_nid2str(peer.nid),
713 req->rq_replied ? "returned error from" :
714 "failed to reply to",
715 ast_type, req, req->rq_xid,
716 (req->rq_repmsg != NULL) ?
717 lustre_msg_get_status(req->rq_repmsg) : 0,
719 ldlm_failed_ast(lock, rc, ast_type);
725 struct ldlm_resource *res = lock->l_resource;
728 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
729 libcfs_nid2str(peer.nid),
731 lustre_msg_get_status(req->rq_repmsg) : -1,
732 ast_type, req, req->rq_xid);
735 * update lvbo to return proper attributes.
738 ldlm_resource_getref(res);
739 ldlm_lvbo_update(res, lock, NULL, 1);
740 ldlm_resource_putref(res);
742 ldlm_lock_cancel(lock);
749 static int ldlm_cb_interpret(const struct lu_env *env,
750 struct ptlrpc_request *req, void *args, int rc)
752 struct ldlm_cb_async_args *ca = args;
753 struct ldlm_lock *lock = ca->ca_lock;
754 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
758 LASSERT(lock != NULL);
761 case LDLM_GL_CALLBACK:
763 * Update the LVB from disk if the AST failed
764 * (this is a legal race)
766 * - Glimpse callback of local lock just returns
767 * -ELDLM_NO_LOCK_DATA.
768 * - Glimpse callback of remote lock might return
769 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
771 if (unlikely(arg->gl_interpret_reply)) {
772 rc = arg->gl_interpret_reply(NULL, req, args, rc);
773 } else if (rc == -ELDLM_NO_LOCK_DATA) {
775 "lost race - client has a lock but no inode");
776 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
777 } else if (rc != 0) {
778 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
780 rc = ldlm_lvbo_update(lock->l_resource,
784 case LDLM_BL_CALLBACK:
786 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
788 case LDLM_CP_CALLBACK:
790 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
793 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
798 /* release extra reference taken in ldlm_ast_fini() */
799 LDLM_LOCK_RELEASE(lock);
802 atomic_inc(&arg->restart);
807 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
809 struct ldlm_cb_async_args *ca = data;
810 struct ldlm_lock *lock = ca->ca_lock;
812 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
815 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
816 struct ldlm_cb_set_arg *arg,
817 struct ldlm_lock *lock,
824 if (unlikely(instant_cancel)) {
825 rc = ptl_send_rpc(req, 1);
826 ptlrpc_req_finished(req);
828 atomic_inc(&arg->restart);
831 ptlrpc_set_add_req(arg->set, req);
838 * Check if there are requests in the export request list which prevent
839 * the lock canceling and make these requests high priority ones.
841 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
843 struct ptlrpc_request *req;
847 if (lock->l_export == NULL) {
848 LDLM_DEBUG(lock, "client lock: no-op");
852 spin_lock(&lock->l_export->exp_rpc_lock);
853 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
856 * Do not process requests that were not yet added to there
857 * incoming queue or were already removed from there for
858 * processing. We evaluate ptlrpc_nrs_req_can_move() without
859 * holding svcpt->scp_req_lock, and then redo the check with
860 * the lock held once we need to obtain a reliable result.
862 if (ptlrpc_nrs_req_can_move(req) &&
863 req->rq_ops->hpreq_lock_match &&
864 req->rq_ops->hpreq_lock_match(req, lock))
865 ptlrpc_nrs_req_hp_move(req);
867 spin_unlock(&lock->l_export->exp_rpc_lock);
872 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
873 * enqueued server lock conflicts with given one.
875 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
876 * to wait for client response.
878 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
879 struct ldlm_lock_desc *desc,
880 void *data, int flag)
882 struct ldlm_cb_async_args *ca;
883 struct ldlm_cb_set_arg *arg = data;
884 struct ldlm_request *body;
885 struct ptlrpc_request *req;
886 int instant_cancel = 0;
891 if (flag == LDLM_CB_CANCELING)
892 /* Don't need to do anything here. */
895 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
896 LDLM_DEBUG(lock, "dropping BL AST");
901 LASSERT(data != NULL);
902 if (lock->l_export->exp_obd->obd_recovering != 0)
903 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
905 ldlm_lock_reorder_req(lock);
907 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
908 &RQF_LDLM_BL_CALLBACK,
909 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
913 ca = ptlrpc_req_async_args(ca, req);
914 ca->ca_set_arg = arg;
917 req->rq_interpret_reply = ldlm_cb_interpret;
919 lock_res_and_lock(lock);
920 if (ldlm_is_destroyed(lock)) {
921 /* What's the point? */
922 unlock_res_and_lock(lock);
923 ptlrpc_req_finished(req);
927 if (!ldlm_is_granted(lock)) {
929 * this blocking AST will be communicated as part of the
930 * completion AST instead
932 ldlm_add_blocked_lock(lock);
933 ldlm_set_waited(lock);
934 unlock_res_and_lock(lock);
936 ptlrpc_req_finished(req);
937 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
941 if (ldlm_is_cancel_on_block(lock))
944 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
945 body->lock_handle[0] = lock->l_remote_handle;
946 body->lock_desc = *desc;
947 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
949 LDLM_DEBUG(lock, "server preparing blocking AST");
951 ptlrpc_request_set_replen(req);
952 ldlm_set_cbpending(lock);
953 if (instant_cancel) {
954 unlock_res_and_lock(lock);
955 ldlm_lock_cancel(lock);
957 req->rq_no_resend = 1;
959 LASSERT(ldlm_is_granted(lock));
960 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
961 unlock_res_and_lock(lock);
963 /* Do not resend after lock callback timeout */
964 req->rq_delay_limit = ldlm_bl_timeout(lock);
965 req->rq_resend_cb = ldlm_update_resend;
968 req->rq_send_state = LUSTRE_IMP_FULL;
969 /* ptlrpc_request_alloc_pack already set timeout */
971 req->rq_timeout = ldlm_get_rq_timeout();
973 if (lock->l_export && lock->l_export->exp_nid_stats &&
974 lock->l_export->exp_nid_stats->nid_ldlm_stats)
975 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
976 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
978 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
984 * ->l_completion_ast callback for a remote lock in server namespace.
986 * Sends AST to the client notifying it of lock granting. If initial
987 * lock response was not sent yet, instead of sending another RPC, just
988 * mark the lock as granted and client will understand
990 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
992 struct ldlm_cb_set_arg *arg = data;
993 struct ldlm_request *body;
994 struct ptlrpc_request *req;
995 struct ldlm_cb_async_args *ca;
996 int instant_cancel = 0;
1002 LASSERT(lock != NULL);
1003 LASSERT(data != NULL);
1005 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1006 LDLM_DEBUG(lock, "dropping CP AST");
1010 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1011 &RQF_LDLM_CP_CALLBACK);
1015 /* server namespace, doesn't need lock */
1016 lvb_len = ldlm_lvbo_size(lock);
1018 * LU-3124 & LU-2187: to not return layout in completion AST because
1019 * it may deadlock for LU-2187, or client may not have enough space
1020 * for large layout. The layout will be returned to client with an
1021 * extra RPC to fetch xattr.lov
1023 if (ldlm_has_layout(lock))
1026 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1027 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1029 ptlrpc_request_free(req);
1033 ca = ptlrpc_req_async_args(ca, req);
1034 ca->ca_set_arg = arg;
1037 req->rq_interpret_reply = ldlm_cb_interpret;
1038 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1040 body->lock_handle[0] = lock->l_remote_handle;
1041 body->lock_flags = ldlm_flags_to_wire(flags);
1042 ldlm_lock2desc(lock, &body->lock_desc);
1044 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1045 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1048 * We still need to send the RPC to wake up the blocked
1049 * enqueue thread on the client.
1051 * Consider old client, there is no better way to notify
1052 * the failure, just zero-sized the LVB, then the client
1053 * will fail out as "-EPROTO".
1055 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1059 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1064 LDLM_DEBUG(lock, "server preparing completion AST");
1066 ptlrpc_request_set_replen(req);
1068 req->rq_send_state = LUSTRE_IMP_FULL;
1069 /* ptlrpc_request_pack already set timeout */
1071 req->rq_timeout = ldlm_get_rq_timeout();
1073 /* We only send real blocking ASTs after the lock is granted */
1074 lock_res_and_lock(lock);
1075 if (ldlm_is_ast_sent(lock)) {
1076 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1077 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1078 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1082 * We might get here prior to ldlm_handle_enqueue setting
1083 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1084 * into waiting list, but this is safe and similar code in
1085 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1086 * that would not only cancel the lock, but will also remove
1087 * it from waiting list
1089 if (ldlm_is_cancel_on_block(lock)) {
1090 unlock_res_and_lock(lock);
1091 ldlm_lock_cancel(lock);
1094 req->rq_no_resend = 1;
1096 lock_res_and_lock(lock);
1098 /* start the lock-timeout clock */
1099 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1100 /* Do not resend after lock callback timeout */
1101 req->rq_delay_limit = ldlm_bl_timeout(lock);
1102 req->rq_resend_cb = ldlm_update_resend;
1105 unlock_res_and_lock(lock);
1107 if (lock->l_export && lock->l_export->exp_nid_stats &&
1108 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1109 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1110 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1112 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1114 RETURN(lvb_len < 0 ? lvb_len : rc);
1118 * Server side ->l_glimpse_ast handler for client locks.
1120 * Sends glimpse AST to the client and waits for reply. Then updates
1121 * lvbo with the result.
1123 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1125 struct ldlm_cb_set_arg *arg = data;
1126 struct ldlm_request *body;
1127 struct ptlrpc_request *req;
1128 struct ldlm_cb_async_args *ca;
1130 struct req_format *req_fmt;
1134 LASSERT(lock != NULL);
1136 if (arg->gl_desc != NULL)
1137 /* There is a glimpse descriptor to pack */
1138 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1140 req_fmt = &RQF_LDLM_GL_CALLBACK;
1142 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1143 req_fmt, LUSTRE_DLM_VERSION,
1149 if (arg->gl_desc != NULL) {
1150 /* copy the GL descriptor */
1151 union ldlm_gl_desc *desc;
1153 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1154 *desc = *arg->gl_desc;
1157 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1158 body->lock_handle[0] = lock->l_remote_handle;
1159 ldlm_lock2desc(lock, &body->lock_desc);
1161 ca = ptlrpc_req_async_args(ca, req);
1162 ca->ca_set_arg = arg;
1165 /* server namespace, doesn't need lock */
1166 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1167 ldlm_lvbo_size(lock));
1168 ptlrpc_request_set_replen(req);
1170 req->rq_send_state = LUSTRE_IMP_FULL;
1171 /* ptlrpc_request_alloc_pack already set timeout */
1173 req->rq_timeout = ldlm_get_rq_timeout();
1175 req->rq_interpret_reply = ldlm_cb_interpret;
1177 if (lock->l_export && lock->l_export->exp_nid_stats) {
1178 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1180 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1181 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1184 rc = ldlm_ast_fini(req, arg, lock, 0);
1188 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1190 int ldlm_glimpse_locks(struct ldlm_resource *res,
1191 struct list_head *gl_work_list)
1197 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1199 if (rc == -ERESTART)
1200 ldlm_reprocess_all(res, NULL);
1204 EXPORT_SYMBOL(ldlm_glimpse_locks);
1206 /* return LDLM lock associated with a lock callback request */
1207 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1209 struct ldlm_cb_async_args *ca;
1210 struct ldlm_lock *lock;
1214 ca = ptlrpc_req_async_args(ca, req);
1217 RETURN(ERR_PTR(-EFAULT));
1221 EXPORT_SYMBOL(ldlm_request_lock);
1224 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1225 * service threads to carry out client lock enqueueing requests.
1227 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1228 struct ptlrpc_request *req,
1229 const struct ldlm_request *dlm_req,
1230 const struct ldlm_callback_suite *cbs)
1232 struct ldlm_reply *dlm_rep;
1234 enum ldlm_error err = ELDLM_OK;
1235 struct ldlm_lock *lock = NULL;
1236 void *cookie = NULL;
1238 struct ldlm_resource *res = NULL;
1239 const struct lu_env *env = req->rq_svc_thread->t_env;
1243 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1245 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1246 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1248 LASSERT(req->rq_export);
1250 /* for intent enqueue the stat will be updated inside intent policy */
1251 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1252 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1253 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1255 if (req->rq_export && req->rq_export->exp_nid_stats &&
1256 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1257 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1258 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1260 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1261 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1262 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1263 dlm_req->lock_desc.l_resource.lr_type);
1264 GOTO(out, rc = -EFAULT);
1267 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1268 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1269 dlm_req->lock_desc.l_req_mode &
1270 (dlm_req->lock_desc.l_req_mode-1))) {
1271 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1272 dlm_req->lock_desc.l_req_mode);
1273 GOTO(out, rc = -EFAULT);
1276 if (unlikely((flags & LDLM_FL_REPLAY) ||
1277 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1278 /* Find an existing lock in the per-export lock hash */
1280 * In the function below, .hs_keycmp resolves to
1281 * ldlm_export_lock_keycmp()
1283 /* coverity[overrun-buffer-val] */
1284 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1285 (void *)&dlm_req->lock_handle[0]);
1287 DEBUG_REQ(D_DLMTRACE, req,
1288 "found existing lock cookie %#llx",
1289 lock->l_handle.h_cookie);
1290 flags |= LDLM_FL_RESENT;
1291 GOTO(existing_lock, rc = 0);
1294 if (ldlm_reclaim_full()) {
1295 DEBUG_REQ(D_DLMTRACE, req,
1296 "Too many granted locks, reject current enqueue request and let the client retry later");
1297 GOTO(out, rc = -EINPROGRESS);
1301 /* The lock's callback data might be set in the policy function */
1302 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1303 dlm_req->lock_desc.l_resource.lr_type,
1304 dlm_req->lock_desc.l_req_mode,
1305 cbs, NULL, 0, LVB_T_NONE);
1312 lock->l_remote_handle = dlm_req->lock_handle[0];
1313 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1316 * Initialize resource lvb but not for a lock being replayed since
1317 * Client already got lvb sent in this case.
1318 * This must occur early since some policy methods assume resource
1319 * lvb is available (lr_lvb_data != NULL).
1321 res = lock->l_resource;
1322 if (!(flags & LDLM_FL_REPLAY)) {
1323 /* non-replayed lock, delayed lvb init may need to be done */
1324 rc = ldlm_lvbo_init(res);
1326 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1331 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1333 * Don't enqueue a lock onto the export if it is been disonnected
1334 * due to eviction (b=3822) or server umount (b=24324).
1335 * Cancel it now instead.
1337 if (req->rq_export->exp_disconnected) {
1338 LDLM_ERROR(lock, "lock on disconnected export %p",
1340 GOTO(out, rc = -ENOTCONN);
1343 lock->l_export = class_export_lock_get(req->rq_export, lock);
1344 if (lock->l_export->exp_lock_hash)
1345 cfs_hash_add(lock->l_export->exp_lock_hash,
1346 &lock->l_remote_handle,
1350 * Inherit the enqueue flags before the operation, because we do not
1351 * keep the res lock on return and next operations (BL AST) may proceed
1354 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1355 LDLM_FL_INHERIT_MASK);
1357 ldlm_convert_policy_to_local(req->rq_export,
1358 dlm_req->lock_desc.l_resource.lr_type,
1359 &dlm_req->lock_desc.l_policy_data,
1360 &lock->l_policy_data);
1361 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1362 lock->l_req_extent = lock->l_policy_data.l_extent;
1363 else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS)
1364 lock->l_policy_data.l_inodebits.try_bits =
1365 dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1369 if (!(flags & LDLM_FL_HAS_INTENT)) {
1370 /* based on the assumption that lvb size never changes during
1371 * resource life time otherwise it need resource->lr_lock's
1373 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1374 RCL_SERVER, ldlm_lvbo_size(lock));
1376 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1377 GOTO(out, rc = -ENOMEM);
1379 rc = req_capsule_server_pack(&req->rq_pill);
1384 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1391 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1393 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1394 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1396 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1397 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1400 * We never send a blocking AST until the lock is granted, but
1401 * we can tell it right now
1403 lock_res_and_lock(lock);
1406 * Now take into account flags to be inherited from original lock
1407 * request both in reply to client and in our own lock flags.
1409 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1410 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1413 * Don't move a pending lock onto the export if it has already been
1414 * disconnected due to eviction (b=5683) or server umount (b=24324).
1415 * Cancel it now instead.
1417 if (unlikely(req->rq_export->exp_disconnected ||
1418 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1419 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1421 } else if (ldlm_is_ast_sent(lock)) {
1422 /* fill lock desc for possible lock convert */
1423 if (lock->l_blocking_lock &&
1424 lock->l_resource->lr_type == LDLM_IBITS) {
1425 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1426 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1429 "save blocking bits %llx in granted lock",
1430 bl_lock->l_policy_data.l_inodebits.bits);
1432 * If lock is blocked then save blocking ibits
1433 * in returned lock policy for the possible lock
1434 * convert on a client.
1436 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1437 bl_lock->l_policy_data.l_inodebits.bits;
1439 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1440 if (ldlm_is_granted(lock)) {
1442 * Only cancel lock if it was granted, because it would
1443 * be destroyed immediately and would never be granted
1444 * in the future, causing timeouts on client. Not
1445 * granted lock will be cancelled immediately after
1446 * sending completion AST.
1448 if (ldlm_is_cancel_on_block(lock)) {
1449 unlock_res_and_lock(lock);
1450 ldlm_lock_cancel(lock);
1451 lock_res_and_lock(lock);
1453 ldlm_add_waiting_lock(lock,
1454 ldlm_bl_timeout(lock));
1458 unlock_res_and_lock(lock);
1462 req->rq_status = rc ?: err; /* return either error - b=11190 */
1463 if (!req->rq_packed_final) {
1464 err = lustre_pack_reply(req, 1, NULL, NULL);
1470 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1471 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1475 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1479 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1481 ldlm_lvbo_size(lock) > 0) {
1486 buf = req_capsule_server_get(&req->rq_pill,
1488 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1489 buflen = req_capsule_get_size(&req->rq_pill,
1490 &RMF_DLM_LVB, RCL_SERVER);
1492 * non-replayed lock, delayed lvb init may
1493 * need to be occur now
1495 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1498 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1500 req_capsule_shrink(&req->rq_pill,
1503 } else if (rc2 == -ERANGE) {
1504 rc2 = req_capsule_server_grow(
1506 &RMF_DLM_LVB, buflen);
1511 * if we can't grow the buffer,
1512 * it's ok to return empty lvb
1523 } else if (flags & LDLM_FL_REPLAY) {
1524 /* no LVB resend upon replay */
1526 req_capsule_shrink(&req->rq_pill,
1536 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1537 if (lock->l_export) {
1538 ldlm_lock_cancel(lock);
1540 lock_res_and_lock(lock);
1541 ldlm_resource_unlink_lock(lock);
1542 ldlm_lock_destroy_nolock(lock);
1543 unlock_res_and_lock(lock);
1548 if (!err && !ldlm_is_cbpending(lock) &&
1549 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1550 ldlm_reprocess_all(lock->l_resource, lock);
1552 LDLM_LOCK_RELEASE(lock);
1555 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1562 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1563 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1565 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1567 if (lock->l_blocking_lock) {
1568 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1569 lock->l_blocking_lock = NULL;
1573 /* A lock can be converted to new ibits or mode and should be considered
1574 * as new lock. Clear all states related to a previous blocking AST
1575 * processing so new conflicts will cause new blocking ASTs.
1577 * This is used during lock convert below and lock downgrade to COS mode in
1578 * ldlm_lock_mode_downgrade().
1580 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1582 ldlm_clear_ast_sent(lock);
1583 lock->l_bl_ast_run = 0;
1584 ldlm_clear_blocking_lock(lock);
1588 * Main LDLM entry point for server code to process lock conversion requests.
1590 int ldlm_handle_convert0(struct ptlrpc_request *req,
1591 const struct ldlm_request *dlm_req)
1593 struct obd_export *exp = req->rq_export;
1594 struct ldlm_reply *dlm_rep;
1595 struct ldlm_lock *lock;
1602 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1603 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1604 LDLM_CONVERT - LDLM_FIRST_OPC);
1606 rc = req_capsule_server_pack(&req->rq_pill);
1610 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1611 dlm_rep->lock_flags = dlm_req->lock_flags;
1613 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1615 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1616 req->rq_status = ELDLM_NO_LOCK_DATA;
1620 LDLM_DEBUG(lock, "server-side convert handler START");
1622 lock_res_and_lock(lock);
1623 bits = lock->l_policy_data.l_inodebits.bits;
1624 new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1626 if (ldlm_is_cancel(lock)) {
1627 LDLM_DEBUG(lock, "convert on canceled lock!");
1628 unlock_res_and_lock(lock);
1629 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1632 if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1633 LDLM_ERROR(lock, "lock mode differs!");
1634 unlock_res_and_lock(lock);
1635 GOTO(out_put, rc = -EPROTO);
1638 if (bits == new_bits) {
1640 * This can be valid situation if CONVERT RPCs are
1641 * re-ordered. Just finish silently
1643 LDLM_DEBUG(lock, "lock is converted already!");
1644 unlock_res_and_lock(lock);
1646 if (ldlm_is_waited(lock))
1647 ldlm_del_waiting_lock(lock);
1649 ldlm_clear_cbpending(lock);
1650 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1651 ldlm_inodebits_drop(lock, bits & ~new_bits);
1653 ldlm_clear_blocking_data(lock);
1654 unlock_res_and_lock(lock);
1656 ldlm_reprocess_all(lock->l_resource, NULL);
1659 dlm_rep->lock_handle = lock->l_remote_handle;
1660 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1661 &dlm_rep->lock_desc.l_policy_data);
1665 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1666 LDLM_LOCK_PUT(lock);
1667 req->rq_status = rc;
1672 * Cancel all the locks whose handles are packed into ldlm_request
1674 * Called by server code expecting such combined cancel activity
1677 int ldlm_request_cancel(struct ptlrpc_request *req,
1678 const struct ldlm_request *dlm_req,
1679 int first, enum lustre_at_flags flags)
1681 struct ldlm_resource *res, *pres = NULL;
1682 struct ldlm_lock *lock;
1683 int i, count, done = 0;
1688 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1689 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1690 (size - offsetof(struct ldlm_request, lock_handle)) /
1691 sizeof(struct lustre_handle) < dlm_req->lock_count)
1694 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1698 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1702 * There is no lock on the server at the replay time,
1703 * skip lock cancelling to make replay tests to pass.
1705 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1708 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1711 for (i = first; i < count; i++) {
1712 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1714 /* below message checked in replay-single.sh test_36 */
1715 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1716 dlm_req->lock_handle[i].cookie);
1720 res = lock->l_resource;
1724 * This code is an optimization to only attempt lock
1725 * granting on the resource (that could be CPU-expensive)
1726 * after we are done cancelling lock in that resource.
1730 ldlm_reprocess_all(pres, NULL);
1731 LDLM_RESOURCE_DELREF(pres);
1732 ldlm_resource_putref(pres);
1735 ldlm_resource_getref(res);
1736 LDLM_RESOURCE_ADDREF(res);
1738 if (!ldlm_is_discard_data(lock))
1739 ldlm_lvbo_update(res, lock,
1745 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1746 lock->l_blast_sent != 0) {
1747 timeout_t delay = 0;
1749 if (ktime_get_real_seconds() > lock->l_blast_sent)
1750 delay = ktime_get_real_seconds() -
1753 "server cancels blocked lock after %ds",
1755 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1757 ldlm_lock_cancel(lock);
1758 LDLM_LOCK_PUT(lock);
1761 ldlm_reprocess_all(pres, NULL);
1762 LDLM_RESOURCE_DELREF(pres);
1763 ldlm_resource_putref(pres);
1765 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1768 EXPORT_SYMBOL(ldlm_request_cancel);
1771 * Main LDLM entry point for server code to cancel locks.
1773 * Typically gets called from service handler on LDLM_CANCEL opc.
1775 int ldlm_handle_cancel(struct ptlrpc_request *req)
1777 struct ldlm_request *dlm_req;
1782 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1783 if (dlm_req == NULL) {
1784 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1788 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1789 offsetof(struct ldlm_request, lock_handle[1]))
1792 if (req->rq_export && req->rq_export->exp_nid_stats &&
1793 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1794 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1795 LDLM_CANCEL - LDLM_FIRST_OPC);
1797 rc = req_capsule_server_pack(&req->rq_pill);
1801 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1802 req->rq_status = LUSTRE_ESTALE;
1804 RETURN(ptlrpc_reply(req));
1806 #endif /* HAVE_SERVER_SUPPORT */
1809 * Server may pass additional information about blocking lock.
1810 * For IBITS locks it is conflicting bits which can be used for
1811 * lock convert instead of cancel.
1813 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1815 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1817 check_res_locked(lock->l_resource);
1818 if (ns_is_client(ns) && ld &&
1819 (lock->l_resource->lr_type == LDLM_IBITS)) {
1821 * Lock description contains policy of blocking lock,
1822 * and its cancel_bits is used to pass conflicting bits.
1823 * NOTE: ld can be NULL or can be not NULL but zeroed if
1824 * passed from ldlm_bl_thread_blwi(), check below used bits
1825 * in ld to make sure it is valid description.
1827 * If server may replace lock resource keeping the same cookie,
1828 * never use cancel bits from different resource, full cancel
1831 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1832 ldlm_res_eq(&ld->l_resource.lr_name,
1833 &lock->l_resource->lr_name) &&
1834 !(ldlm_is_cbpending(lock) &&
1835 lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1836 /* always combine conflicting ibits */
1837 lock->l_policy_data.l_inodebits.cancel_bits |=
1838 ld->l_policy_data.l_inodebits.cancel_bits;
1840 /* If cancel_bits are not obtained or
1841 * if the lock is already CBPENDING and
1842 * has no cancel_bits set
1843 * - the full lock is to be cancelled
1845 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1851 * Callback handler for receiving incoming blocking ASTs.
1853 * This can only happen on client side.
1855 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1856 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1862 LDLM_DEBUG(lock, "client blocking AST callback handler");
1864 lock_res_and_lock(lock);
1866 /* get extra information from desc if any */
1867 ldlm_bl_desc2lock(ld, lock);
1868 ldlm_set_cbpending(lock);
1870 do_ast = (!lock->l_readers && !lock->l_writers);
1871 unlock_res_and_lock(lock);
1875 "Lock %p already unused, calling callback (%p)\n",
1876 lock, lock->l_blocking_ast);
1877 if (lock->l_blocking_ast != NULL)
1878 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1882 "Lock %p is referenced, will be cancelled later\n",
1886 LDLM_DEBUG(lock, "client blocking callback handler END");
1887 LDLM_LOCK_RELEASE(lock);
1891 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1893 if (req->rq_no_reply)
1896 req->rq_status = rc;
1897 if (!req->rq_packed_final) {
1898 rc = lustre_pack_reply(req, 1, NULL, NULL);
1902 return ptlrpc_reply(req);
1906 * Callback handler for receiving incoming completion ASTs.
1908 * This only can happen on client side.
1910 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1911 struct ldlm_namespace *ns,
1912 struct ldlm_request *dlm_req,
1913 struct ldlm_lock *lock)
1915 LIST_HEAD(ast_list);
1921 LDLM_DEBUG(lock, "client completion callback handler START");
1923 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1924 long to = cfs_time_seconds(1);
1926 ldlm_callback_reply(req, 0);
1929 schedule_timeout_interruptible(to);
1930 if (ldlm_is_granted(lock) ||
1931 ldlm_is_destroyed(lock))
1936 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1938 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1939 GOTO(out, rc = lvb_len);
1940 } else if (lvb_len > 0) {
1941 if (lock->l_lvb_len > 0) {
1942 /* for extent lock, lvb contains ost_lvb{}. */
1943 LASSERT(lock->l_lvb_data != NULL);
1945 if (unlikely(lock->l_lvb_len < lvb_len)) {
1947 "Replied LVB is larger than expectation, expected = %d, replied = %d",
1948 lock->l_lvb_len, lvb_len);
1949 GOTO(out, rc = -EINVAL);
1954 lock_res_and_lock(lock);
1956 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1957 &lock->l_resource->lr_name)) {
1958 ldlm_resource_unlink_lock(lock);
1959 unlock_res_and_lock(lock);
1960 rc = ldlm_lock_change_resource(ns, lock,
1961 &dlm_req->lock_desc.l_resource.lr_name);
1963 LDLM_ERROR(lock, "Failed to allocate resource");
1966 LDLM_DEBUG(lock, "completion AST, new resource");
1967 lock_res_and_lock(lock);
1970 if (ldlm_is_failed(lock)) {
1971 unlock_res_and_lock(lock);
1972 LDLM_LOCK_RELEASE(lock);
1976 if (ldlm_is_destroyed(lock) ||
1977 ldlm_is_granted(lock)) {
1978 /* b=11300: the lock has already been granted */
1979 unlock_res_and_lock(lock);
1980 LDLM_DEBUG(lock, "Double grant race happened");
1985 * If we receive the completion AST before the actual enqueue returned,
1986 * then we might need to switch lock modes, resources, or extents.
1988 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1989 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1990 LDLM_DEBUG(lock, "completion AST, new lock mode");
1993 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1994 ldlm_convert_policy_to_local(req->rq_export,
1995 dlm_req->lock_desc.l_resource.lr_type,
1996 &dlm_req->lock_desc.l_policy_data,
1997 &lock->l_policy_data);
1998 LDLM_DEBUG(lock, "completion AST, new policy data");
2001 ldlm_resource_unlink_lock(lock);
2003 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2005 * BL_AST locks are not needed in LRU.
2006 * Let ldlm_cancel_lru() be fast.
2008 ldlm_lock_remove_from_lru(lock);
2009 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2010 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2011 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2014 if (lock->l_lvb_len > 0) {
2015 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2016 lock->l_lvb_data, lvb_len);
2018 unlock_res_and_lock(lock);
2023 ldlm_grant_lock(lock, &ast_list);
2024 unlock_res_and_lock(lock);
2026 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2029 * Let Enqueue to call osc_lock_upcall() and initialize
2032 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2034 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2036 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2042 lock_res_and_lock(lock);
2043 ldlm_set_failed(lock);
2044 unlock_res_and_lock(lock);
2045 wake_up(&lock->l_waitq);
2047 LDLM_LOCK_RELEASE(lock);
2053 * Callback handler for receiving incoming glimpse ASTs.
2055 * This only can happen on client side. After handling the glimpse AST
2056 * we also consider dropping the lock here if it is unused locally for a
2059 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2060 struct ldlm_namespace *ns,
2061 struct ldlm_request *dlm_req,
2062 struct ldlm_lock *lock)
2064 struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2069 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2071 if (lock->l_glimpse_ast != NULL)
2072 rc = lock->l_glimpse_ast(lock, req);
2074 if (req->rq_repmsg != NULL) {
2077 req->rq_status = rc;
2081 lock_res_and_lock(lock);
2082 if (lock->l_granted_mode == LCK_PW &&
2083 !lock->l_readers && !lock->l_writers &&
2084 ktime_after(ktime_get(),
2085 ktime_add(lock->l_last_used,
2086 ktime_set(ns->ns_dirty_age_limit, 0)))) {
2087 unlock_res_and_lock(lock);
2089 /* For MDS glimpse it is always DOM lock, set corresponding
2090 * cancel_bits to perform lock convert if needed
2092 if (lock->l_resource->lr_type == LDLM_IBITS)
2093 ld->l_policy_data.l_inodebits.cancel_bits =
2095 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2096 ldlm_handle_bl_callback(ns, ld, lock);
2101 unlock_res_and_lock(lock);
2102 LDLM_LOCK_RELEASE(lock);
2106 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2107 enum ldlm_cancel_flags cancel_flags)
2109 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2113 spin_lock(&blp->blp_lock);
2114 if (blwi->blwi_lock &&
2115 ldlm_is_discard_data(blwi->blwi_lock)) {
2116 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2117 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2119 /* other blocking callbacks are added to the regular list */
2120 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2122 spin_unlock(&blp->blp_lock);
2124 wake_up(&blp->blp_waitq);
2127 * can not check blwi->blwi_flags as blwi could be already freed in
2130 if (!(cancel_flags & LCF_ASYNC))
2131 wait_for_completion(&blwi->blwi_comp);
2136 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2137 struct ldlm_namespace *ns,
2138 struct ldlm_lock_desc *ld,
2139 struct list_head *cancels, int count,
2140 struct ldlm_lock *lock,
2141 enum ldlm_cancel_flags cancel_flags)
2143 init_completion(&blwi->blwi_comp);
2144 INIT_LIST_HEAD(&blwi->blwi_head);
2146 if (memory_pressure_get())
2147 blwi->blwi_mem_pressure = 1;
2150 blwi->blwi_flags = cancel_flags;
2152 blwi->blwi_ld = *ld;
2154 list_splice_init(cancels, &blwi->blwi_head);
2155 blwi->blwi_count = count;
2157 blwi->blwi_lock = lock;
2162 * Queues a list of locks \a cancels containing \a count locks
2163 * for later processing by a blocking thread. If \a count is zero,
2164 * then the lock referenced as \a lock is queued instead.
2166 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2167 * If list addition fails an error is returned and caller is supposed to
2168 * call ->l_blocking_ast itself.
2170 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2171 struct ldlm_lock_desc *ld,
2172 struct ldlm_lock *lock,
2173 struct list_head *cancels, int count,
2174 enum ldlm_cancel_flags cancel_flags)
2178 if (cancels && count == 0)
2181 if (cancel_flags & LCF_ASYNC) {
2182 struct ldlm_bl_work_item *blwi;
2184 OBD_ALLOC(blwi, sizeof(*blwi));
2187 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2189 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2192 * if it is synchronous call do minimum mem alloc, as it could
2193 * be triggered from kernel shrinker
2195 struct ldlm_bl_work_item blwi;
2197 memset(&blwi, 0, sizeof(blwi));
2198 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2199 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2204 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2205 struct ldlm_lock *lock)
2207 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2210 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2211 struct list_head *cancels, int count,
2212 enum ldlm_cancel_flags cancel_flags)
2214 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2217 int ldlm_bl_thread_wakeup(void)
2219 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2223 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2224 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2226 struct obd_device *obd = req->rq_export->exp_obd;
2234 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2236 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2238 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2240 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2243 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2245 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2247 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2250 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2253 /* We are responsible for swabbing contents of val */
2255 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2256 /* Pass it on to mdc (the "export" in this case) */
2257 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2259 sizeof(KEY_HSM_COPYTOOL_SEND),
2260 KEY_HSM_COPYTOOL_SEND,
2263 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2268 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2269 const char *msg, int rc,
2270 const struct lustre_handle *handle)
2272 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2273 "%s, NID=%s lock=%#llx: rc = %d",
2274 msg, libcfs_id2str(req->rq_peer),
2275 handle ? handle->cookie : 0, rc);
2276 if (req->rq_no_reply)
2277 CWARN("No reply was sent, maybe cause b=21636.\n");
2279 CWARN("Send reply failed, maybe cause b=21636.\n");
2282 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2283 static int ldlm_callback_handler(struct ptlrpc_request *req)
2285 struct ldlm_namespace *ns;
2286 struct ldlm_request *dlm_req;
2287 struct ldlm_lock *lock;
2293 * Requests arrive in sender's byte order. The ptlrpc service
2294 * handler has already checked and, if necessary, byte-swapped the
2295 * incoming request message body, but I am responsible for the
2299 /* do nothing for sec context finalize */
2300 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2303 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2305 if (req->rq_export == NULL) {
2306 rc = ldlm_callback_reply(req, -ENOTCONN);
2307 ldlm_callback_errmsg(req, "Operate on unconnected server",
2312 LASSERT(req->rq_export != NULL);
2313 LASSERT(req->rq_export->exp_obd != NULL);
2315 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2316 case LDLM_BL_CALLBACK:
2317 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2319 ldlm_callback_reply(req, -(int)cfs_fail_err);
2323 case LDLM_CP_CALLBACK:
2324 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2327 case LDLM_GL_CALLBACK:
2328 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2332 rc = ldlm_handle_setinfo(req);
2333 ldlm_callback_reply(req, rc);
2336 CERROR("unknown opcode %u\n",
2337 lustre_msg_get_opc(req->rq_reqmsg));
2338 ldlm_callback_reply(req, -EPROTO);
2342 ns = req->rq_export->exp_obd->obd_namespace;
2343 LASSERT(ns != NULL);
2345 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2347 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2348 if (dlm_req == NULL) {
2349 rc = ldlm_callback_reply(req, -EPROTO);
2350 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2356 * Force a known safe race, send a cancel to the server for a lock
2357 * which the server has already started a blocking callback on.
2359 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2360 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2361 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2363 CERROR("ldlm_cli_cancel: %d\n", rc);
2366 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2369 "callback on lock %#llx - lock disappeared\n",
2370 dlm_req->lock_handle[0].cookie);
2371 rc = ldlm_callback_reply(req, -EINVAL);
2372 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2373 &dlm_req->lock_handle[0]);
2377 if (ldlm_is_fail_loc(lock) &&
2378 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2379 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2381 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2382 lock_res_and_lock(lock);
2383 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2385 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2387 * If somebody cancels lock and cache is already dropped,
2388 * or lock is failed before cp_ast received on client,
2389 * we can tell the server we have no lock. Otherwise, we
2390 * should send cancel after dropping the cache.
2392 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2393 ldlm_is_failed(lock)) {
2395 "callback on lock %llx - lock disappeared",
2396 dlm_req->lock_handle[0].cookie);
2397 unlock_res_and_lock(lock);
2398 LDLM_LOCK_RELEASE(lock);
2399 rc = ldlm_callback_reply(req, -EINVAL);
2400 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2401 &dlm_req->lock_handle[0]);
2405 * BL_AST locks are not needed in LRU.
2406 * Let ldlm_cancel_lru() be fast.
2408 ldlm_lock_remove_from_lru(lock);
2409 ldlm_set_bl_ast(lock);
2411 unlock_res_and_lock(lock);
2414 * We want the ost thread to get this reply so that it can respond
2415 * to ost requests (write cache writeback) that might be triggered
2418 * But we'd also like to be able to indicate in the reply that we're
2419 * cancelling right now, because it's unused, or have an intent result
2420 * in the reply, so we might have to push the responsibility for sending
2421 * the reply down into the AST handlers, alas.
2424 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2425 case LDLM_BL_CALLBACK:
2426 CDEBUG(D_INODE, "blocking ast\n");
2427 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2428 if (!ldlm_is_cancel_on_block(lock)) {
2429 rc = ldlm_callback_reply(req, 0);
2430 if (req->rq_no_reply || rc)
2431 ldlm_callback_errmsg(req, "Normal process", rc,
2432 &dlm_req->lock_handle[0]);
2434 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2435 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2437 case LDLM_CP_CALLBACK:
2438 CDEBUG(D_INODE, "completion ast\n");
2439 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2440 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2441 if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2442 ldlm_callback_reply(req, rc);
2444 case LDLM_GL_CALLBACK:
2445 CDEBUG(D_INODE, "glimpse ast\n");
2446 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2447 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2450 LBUG(); /* checked above */
2456 #ifdef HAVE_SERVER_SUPPORT
2458 * Main handler for canceld thread.
2460 * Separated into its own thread to avoid deadlocks.
2462 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2469 * Requests arrive in sender's byte order. The ptlrpc service
2470 * handler has already checked and, if necessary, byte-swapped the
2471 * incoming request message body, but I am responsible for the
2475 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2477 if (req->rq_export == NULL) {
2478 struct ldlm_request *dlm_req;
2480 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2481 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2482 libcfs_nid2str(req->rq_peer.nid),
2483 (unsigned long long)req->rq_arrival_time.tv_sec,
2484 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2486 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2487 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2488 dlm_req = req_capsule_client_get(&req->rq_pill,
2490 if (dlm_req != NULL)
2491 ldlm_lock_dump_handle(D_ERROR,
2492 &dlm_req->lock_handle[0]);
2494 ldlm_callback_reply(req, -ENOTCONN);
2498 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2499 /* XXX FIXME move this back to mds/handler.c, b=249 */
2501 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2502 CDEBUG(D_INODE, "cancel\n");
2503 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2504 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2505 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2507 rc = ldlm_handle_cancel(req);
2511 struct ldlm_request *dlm_req;
2513 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2514 CDEBUG(D_INODE, "convert\n");
2516 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2517 if (dlm_req == NULL) {
2518 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2519 rc = ldlm_callback_reply(req, -EPROTO);
2521 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2522 rc = ptlrpc_reply(req);
2527 CERROR("invalid opcode %d\n",
2528 lustre_msg_get_opc(req->rq_reqmsg));
2529 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2530 rc = ldlm_callback_reply(req, -EINVAL);
2536 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2537 struct ldlm_lock *lock)
2539 struct ldlm_request *dlm_req;
2540 struct lustre_handle lockh;
2546 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2547 if (dlm_req == NULL)
2550 ldlm_lock2handle(lock, &lockh);
2551 for (i = 0; i < dlm_req->lock_count; i++) {
2552 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2554 DEBUG_REQ(D_RPCTRACE, req,
2555 "Prio raised by lock %#llx", lockh.cookie);
2564 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2566 struct ldlm_request *dlm_req;
2573 /* no prolong in recovery */
2574 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2577 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2578 if (dlm_req == NULL)
2581 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2582 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2583 (size - offsetof(struct ldlm_request, lock_handle)) /
2584 sizeof(struct lustre_handle) < dlm_req->lock_count)
2587 for (i = 0; i < dlm_req->lock_count; i++) {
2588 struct ldlm_lock *lock;
2590 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2594 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2596 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2597 LDLM_LOCK_PUT(lock);
2606 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2607 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2608 .hpreq_check = ldlm_cancel_hpreq_check,
2612 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2616 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2618 if (req->rq_export == NULL)
2621 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2622 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2623 req->rq_ops = &ldlm_cancel_hpreq_ops;
2624 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2625 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2626 req->rq_ops = &ldlm_cancel_hpreq_ops;
2631 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2632 struct hlist_node *hnode, void *data)
2635 struct list_head *rpc_list = data;
2636 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2638 lock_res_and_lock(lock);
2640 if (!ldlm_is_granted(lock)) {
2641 unlock_res_and_lock(lock);
2645 LASSERT(lock->l_resource);
2646 if (lock->l_resource->lr_type != LDLM_IBITS &&
2647 lock->l_resource->lr_type != LDLM_PLAIN) {
2648 unlock_res_and_lock(lock);
2652 if (ldlm_is_ast_sent(lock)) {
2653 unlock_res_and_lock(lock);
2657 LASSERT(lock->l_blocking_ast);
2658 LASSERT(!lock->l_blocking_lock);
2660 ldlm_set_ast_sent(lock);
2661 if (lock->l_export && lock->l_export->exp_lock_hash) {
2663 * NB: it's safe to call cfs_hash_del() even lock isn't
2667 * In the function below, .hs_keycmp resolves to
2668 * ldlm_export_lock_keycmp()
2670 /* coverity[overrun-buffer-val] */
2671 cfs_hash_del(lock->l_export->exp_lock_hash,
2672 &lock->l_remote_handle, &lock->l_exp_hash);
2675 list_add_tail(&lock->l_rk_ast, rpc_list);
2676 LDLM_LOCK_GET(lock);
2678 unlock_res_and_lock(lock);
2682 void ldlm_revoke_export_locks(struct obd_export *exp)
2684 LIST_HEAD(rpc_list);
2688 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2689 ldlm_revoke_lock_cb, &rpc_list, 0);
2690 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2691 LDLM_WORK_REVOKE_AST);
2695 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2696 #endif /* HAVE_SERVER_SUPPORT */
2698 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2699 struct ldlm_bl_work_item **p_blwi,
2700 struct obd_export **p_exp)
2702 struct ldlm_bl_work_item *blwi = NULL;
2703 static unsigned int num_bl;
2704 static unsigned int num_stale;
2705 int num_th = atomic_read(&blp->blp_num_threads);
2707 *p_exp = obd_stale_export_get();
2709 spin_lock(&blp->blp_lock);
2710 if (*p_exp != NULL) {
2711 if (num_th == 1 || ++num_stale < num_th) {
2712 spin_unlock(&blp->blp_lock);
2718 /* process a request from the blp_list at least every blp_num_threads */
2719 if (!list_empty(&blp->blp_list) &&
2720 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2721 blwi = list_entry(blp->blp_list.next,
2722 struct ldlm_bl_work_item, blwi_entry);
2724 if (!list_empty(&blp->blp_prio_list))
2725 blwi = list_entry(blp->blp_prio_list.next,
2726 struct ldlm_bl_work_item,
2730 if (++num_bl >= num_th)
2732 list_del(&blwi->blwi_entry);
2734 spin_unlock(&blp->blp_lock);
2737 if (*p_exp != NULL && *p_blwi != NULL) {
2738 obd_stale_export_put(*p_exp);
2742 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2745 /* This only contains temporary data until the thread starts */
2746 struct ldlm_bl_thread_data {
2747 struct ldlm_bl_pool *bltd_blp;
2748 struct completion bltd_comp;
2752 static int ldlm_bl_thread_main(void *arg);
2754 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2756 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2757 struct task_struct *task;
2759 init_completion(&bltd.bltd_comp);
2761 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2762 if (bltd.bltd_num >= blp->blp_max_threads) {
2763 atomic_dec(&blp->blp_num_threads);
2767 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2769 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2770 atomic_dec(&blp->blp_num_threads);
2774 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2777 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2778 bltd.bltd_num, PTR_ERR(task));
2779 atomic_dec(&blp->blp_num_threads);
2780 return PTR_ERR(task);
2782 wait_for_completion(&bltd.bltd_comp);
2787 /* Not fatal if racy and have a few too many threads */
2788 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2789 struct ldlm_bl_work_item *blwi)
2791 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2794 if (atomic_read(&blp->blp_busy_threads) <
2795 atomic_read(&blp->blp_num_threads))
2798 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2799 blwi->blwi_mem_pressure))
2805 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2806 struct ldlm_bl_work_item *blwi)
2810 if (blwi->blwi_ns == NULL)
2811 /* added by ldlm_cleanup() */
2812 RETURN(LDLM_ITER_STOP);
2814 if (blwi->blwi_mem_pressure)
2815 memory_pressure_set();
2817 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2819 if (blwi->blwi_count) {
2822 * The special case when we cancel locks in lru
2823 * asynchronously, we pass the list of locks here.
2824 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2825 * canceled locally yet.
2827 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2830 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2833 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2836 if (blwi->blwi_mem_pressure)
2837 memory_pressure_clr();
2839 if (blwi->blwi_flags & LCF_ASYNC)
2840 OBD_FREE(blwi, sizeof(*blwi));
2842 complete(&blwi->blwi_comp);
2848 * Cancel stale locks on export. Cancel blocked locks first.
2849 * If the given export has blocked locks, the next in the list may have
2850 * them too, thus cancel not blocked locks only if the current export has
2853 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2854 struct obd_export *exp)
2860 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2862 num = ldlm_export_cancel_blocked_locks(exp);
2864 ldlm_export_cancel_locks(exp);
2866 obd_stale_export_put(exp);
2873 * Main blocking requests processing thread.
2875 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2876 * This thread in the end ends up doing actual call to ->l_blocking_ast
2879 static int ldlm_bl_thread_main(void *arg)
2882 struct ldlm_bl_pool *blp;
2883 struct ldlm_bl_thread_data *bltd = arg;
2891 rc = lu_env_init(env, LCT_DT_THREAD);
2894 rc = lu_env_add(env);
2896 GOTO(out_env_fini, rc);
2898 blp = bltd->bltd_blp;
2900 complete(&bltd->bltd_comp);
2901 /* cannot use bltd after this, it is only on caller's stack */
2904 struct ldlm_bl_work_item *blwi = NULL;
2905 struct obd_export *exp = NULL;
2908 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2911 wait_event_idle_exclusive(blp->blp_waitq,
2912 ldlm_bl_get_work(blp, &blwi,
2914 atomic_inc(&blp->blp_busy_threads);
2916 if (ldlm_bl_thread_need_create(blp, blwi))
2917 /* discard the return value, we tried */
2918 ldlm_bl_thread_start(blp, true);
2921 rc = ldlm_bl_thread_exports(blp, exp);
2923 rc = ldlm_bl_thread_blwi(blp, blwi);
2925 atomic_dec(&blp->blp_busy_threads);
2927 if (rc == LDLM_ITER_STOP)
2931 * If there are many namespaces, we will not sleep waiting for
2932 * work, and must do a cond_resched to avoid holding the CPU
2938 atomic_dec(&blp->blp_num_threads);
2939 complete(&blp->blp_comp);
2950 static int ldlm_setup(void);
2951 static int ldlm_cleanup(void);
2953 int ldlm_get_ref(void)
2958 mutex_lock(&ldlm_ref_mutex);
2959 if (++ldlm_refcount == 1) {
2964 mutex_unlock(&ldlm_ref_mutex);
2969 void ldlm_put_ref(void)
2972 mutex_lock(&ldlm_ref_mutex);
2973 if (ldlm_refcount == 1) {
2974 int rc = ldlm_cleanup();
2977 CERROR("ldlm_cleanup failed: %d\n", rc);
2983 mutex_unlock(&ldlm_ref_mutex);
2989 * Export handle<->lock hash operations.
2992 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
2994 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2998 ldlm_export_lock_key(struct hlist_node *hnode)
3000 struct ldlm_lock *lock;
3002 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3003 return &lock->l_remote_handle;
3007 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3009 struct ldlm_lock *lock;
3011 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3012 lock->l_remote_handle = *(struct lustre_handle *)key;
3016 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3018 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3022 ldlm_export_lock_object(struct hlist_node *hnode)
3024 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3028 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3030 struct ldlm_lock *lock;
3032 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3033 LDLM_LOCK_GET(lock);
3037 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3039 struct ldlm_lock *lock;
3041 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3042 LDLM_LOCK_RELEASE(lock);
3045 static struct cfs_hash_ops ldlm_export_lock_ops = {
3046 .hs_hash = ldlm_export_lock_hash,
3047 .hs_key = ldlm_export_lock_key,
3048 .hs_keycmp = ldlm_export_lock_keycmp,
3049 .hs_keycpy = ldlm_export_lock_keycpy,
3050 .hs_object = ldlm_export_lock_object,
3051 .hs_get = ldlm_export_lock_get,
3052 .hs_put = ldlm_export_lock_put,
3053 .hs_put_locked = ldlm_export_lock_put,
3056 int ldlm_init_export(struct obd_export *exp)
3062 exp->exp_lock_hash =
3063 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3064 HASH_EXP_LOCK_CUR_BITS,
3065 HASH_EXP_LOCK_MAX_BITS,
3066 HASH_EXP_LOCK_BKT_BITS, 0,
3067 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3068 &ldlm_export_lock_ops,
3069 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3070 CFS_HASH_NBLK_CHANGE);
3072 if (!exp->exp_lock_hash)
3075 rc = ldlm_init_flock_export(exp);
3081 ldlm_destroy_export(exp);
3084 EXPORT_SYMBOL(ldlm_init_export);
3086 void ldlm_destroy_export(struct obd_export *exp)
3089 cfs_hash_putref(exp->exp_lock_hash);
3090 exp->exp_lock_hash = NULL;
3092 ldlm_destroy_flock_export(exp);
3095 EXPORT_SYMBOL(ldlm_destroy_export);
3097 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3098 struct attribute *attr,
3101 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3104 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3105 struct attribute *attr,
3112 rc = kstrtoul(buffer, 10, &val);
3116 ldlm_cancel_unused_locks_before_replay = val;
3120 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3122 static struct attribute *ldlm_attrs[] = {
3123 &lustre_attr_cancel_unused_locks_before_replay.attr,
3127 static struct attribute_group ldlm_attr_group = {
3128 .attrs = ldlm_attrs,
3131 static int ldlm_setup(void)
3133 static struct ptlrpc_service_conf conf;
3134 struct ldlm_bl_pool *blp = NULL;
3135 #ifdef HAVE_SERVER_SUPPORT
3136 struct task_struct *task;
3137 #endif /* HAVE_SERVER_SUPPORT */
3143 if (ldlm_state != NULL)
3146 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3147 if (ldlm_state == NULL)
3150 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3154 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3158 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3162 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3166 rc = ldlm_debugfs_setup();
3170 memset(&conf, 0, sizeof(conf));
3171 conf = (typeof(conf)) {
3172 .psc_name = "ldlm_cbd",
3173 .psc_watchdog_factor = 2,
3175 .bc_nbufs = LDLM_CLIENT_NBUFS,
3176 .bc_buf_size = LDLM_BUFSIZE,
3177 .bc_req_max_size = LDLM_MAXREQSIZE,
3178 .bc_rep_max_size = LDLM_MAXREPSIZE,
3179 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3180 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3183 .tc_thr_name = "ldlm_cb",
3184 .tc_thr_factor = LDLM_THR_FACTOR,
3185 .tc_nthrs_init = LDLM_NTHRS_INIT,
3186 .tc_nthrs_base = LDLM_NTHRS_BASE,
3187 .tc_nthrs_max = LDLM_NTHRS_MAX,
3188 .tc_nthrs_user = ldlm_num_threads,
3189 .tc_cpu_bind = ldlm_cpu_bind,
3190 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3193 .cc_pattern = ldlm_cpts,
3194 .cc_affinity = true,
3197 .so_req_handler = ldlm_callback_handler,
3200 ldlm_state->ldlm_cb_service = \
3201 ptlrpc_register_service(&conf, ldlm_svc_kset,
3202 ldlm_svc_debugfs_dir);
3203 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3204 CERROR("failed to start service\n");
3205 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3206 ldlm_state->ldlm_cb_service = NULL;
3210 #ifdef HAVE_SERVER_SUPPORT
3211 memset(&conf, 0, sizeof(conf));
3212 conf = (typeof(conf)) {
3213 .psc_name = "ldlm_canceld",
3214 .psc_watchdog_factor = 6,
3216 .bc_nbufs = LDLM_SERVER_NBUFS,
3217 .bc_buf_size = LDLM_BUFSIZE,
3218 .bc_req_max_size = LDLM_MAXREQSIZE,
3219 .bc_rep_max_size = LDLM_MAXREPSIZE,
3220 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3221 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3225 .tc_thr_name = "ldlm_cn",
3226 .tc_thr_factor = LDLM_THR_FACTOR,
3227 .tc_nthrs_init = LDLM_NTHRS_INIT,
3228 .tc_nthrs_base = LDLM_NTHRS_BASE,
3229 .tc_nthrs_max = LDLM_NTHRS_MAX,
3230 .tc_nthrs_user = ldlm_num_threads,
3231 .tc_cpu_bind = ldlm_cpu_bind,
3232 .tc_ctx_tags = LCT_MD_THREAD | \
3237 .cc_pattern = ldlm_cpts,
3238 .cc_affinity = true,
3241 .so_req_handler = ldlm_cancel_handler,
3242 .so_hpreq_handler = ldlm_hpreq_handler,
3245 ldlm_state->ldlm_cancel_service = \
3246 ptlrpc_register_service(&conf, ldlm_svc_kset,
3247 ldlm_svc_debugfs_dir);
3248 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3249 CERROR("failed to start service\n");
3250 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3251 ldlm_state->ldlm_cancel_service = NULL;
3254 #endif /* HAVE_SERVER_SUPPORT */
3256 OBD_ALLOC(blp, sizeof(*blp));
3258 GOTO(out, rc = -ENOMEM);
3259 ldlm_state->ldlm_bl_pool = blp;
3261 spin_lock_init(&blp->blp_lock);
3262 INIT_LIST_HEAD(&blp->blp_list);
3263 INIT_LIST_HEAD(&blp->blp_prio_list);
3264 init_waitqueue_head(&blp->blp_waitq);
3265 atomic_set(&blp->blp_num_threads, 0);
3266 atomic_set(&blp->blp_busy_threads, 0);
3268 if (ldlm_num_threads == 0) {
3269 blp->blp_min_threads = LDLM_NTHRS_INIT;
3270 blp->blp_max_threads = LDLM_NTHRS_MAX;
3272 blp->blp_min_threads = blp->blp_max_threads = \
3273 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3277 for (i = 0; i < blp->blp_min_threads; i++) {
3278 rc = ldlm_bl_thread_start(blp, false);
3283 #ifdef HAVE_SERVER_SUPPORT
3284 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3287 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3291 wait_event(expired_lock_wait_queue,
3292 expired_lock_thread_state == ELT_READY);
3293 #endif /* HAVE_SERVER_SUPPORT */
3295 rc = ldlm_pools_init();
3297 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3301 rc = ldlm_reclaim_setup();
3303 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3313 static int ldlm_cleanup(void)
3317 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3318 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3319 CERROR("ldlm still has namespaces; clean these up first.\n");
3320 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3321 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3325 ldlm_reclaim_cleanup();
3328 if (ldlm_state->ldlm_bl_pool != NULL) {
3329 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3331 while (atomic_read(&blp->blp_num_threads) > 0) {
3332 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3334 init_completion(&blp->blp_comp);
3336 spin_lock(&blp->blp_lock);
3337 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3338 wake_up(&blp->blp_waitq);
3339 spin_unlock(&blp->blp_lock);
3341 wait_for_completion(&blp->blp_comp);
3344 OBD_FREE(blp, sizeof(*blp));
3347 if (ldlm_state->ldlm_cb_service != NULL)
3348 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3349 #ifdef HAVE_SERVER_SUPPORT
3350 if (ldlm_state->ldlm_cancel_service != NULL)
3351 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3355 kset_unregister(ldlm_ns_kset);
3357 kset_unregister(ldlm_svc_kset);
3359 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3360 kobject_put(ldlm_kobj);
3363 ldlm_debugfs_cleanup();
3365 #ifdef HAVE_SERVER_SUPPORT
3366 if (expired_lock_thread_state != ELT_STOPPED) {
3367 expired_lock_thread_state = ELT_TERMINATE;
3368 wake_up(&expired_lock_wait_queue);
3369 wait_event(expired_lock_wait_queue,
3370 expired_lock_thread_state == ELT_STOPPED);
3374 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3382 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3383 sizeof(struct ldlm_resource), 0,
3384 SLAB_HWCACHE_ALIGN, NULL);
3385 if (ldlm_resource_slab == NULL)
3388 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3389 sizeof(struct ldlm_lock), 0,
3390 SLAB_HWCACHE_ALIGN, NULL);
3391 if (ldlm_lock_slab == NULL)
3394 ldlm_interval_slab = kmem_cache_create("interval_node",
3395 sizeof(struct ldlm_interval),
3396 0, SLAB_HWCACHE_ALIGN, NULL);
3397 if (ldlm_interval_slab == NULL)
3400 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3401 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3402 0, SLAB_HWCACHE_ALIGN, NULL);
3403 if (ldlm_interval_tree_slab == NULL)
3406 #ifdef HAVE_SERVER_SUPPORT
3407 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3408 sizeof(struct ldlm_ibits_node),
3409 0, SLAB_HWCACHE_ALIGN, NULL);
3410 if (ldlm_inodebits_slab == NULL)
3411 goto out_interval_tree;
3413 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3414 sizeof(struct ldlm_glimpse_work),
3416 if (ldlm_glimpse_work_kmem == NULL)
3420 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3421 class_export_dump_hook = ldlm_dump_export_locks;
3424 #ifdef HAVE_SERVER_SUPPORT
3426 kmem_cache_destroy(ldlm_inodebits_slab);
3428 kmem_cache_destroy(ldlm_interval_tree_slab);
3431 kmem_cache_destroy(ldlm_interval_slab);
3433 kmem_cache_destroy(ldlm_lock_slab);
3435 kmem_cache_destroy(ldlm_resource_slab);
3440 void ldlm_exit(void)
3443 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3444 kmem_cache_destroy(ldlm_resource_slab);
3446 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3447 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3448 * so that ldlm_lock_free() get a chance to be called.
3451 kmem_cache_destroy(ldlm_lock_slab);
3452 kmem_cache_destroy(ldlm_interval_slab);
3453 kmem_cache_destroy(ldlm_interval_tree_slab);
3454 #ifdef HAVE_SERVER_SUPPORT
3455 kmem_cache_destroy(ldlm_inodebits_slab);
3456 kmem_cache_destroy(ldlm_glimpse_work_kmem);