4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lockd.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <linux/kthread.h>
41 #include <linux/list.h>
42 #include <libcfs/libcfs.h>
43 #include <lustre_errno.h>
44 #include <lustre_dlm.h>
45 #include <obd_class.h>
46 #include "ldlm_internal.h"
48 static int ldlm_num_threads;
49 module_param(ldlm_num_threads, int, 0444);
50 MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start");
52 static unsigned int ldlm_cpu_bind = 1;
53 module_param(ldlm_cpu_bind, uint, 0444);
54 MODULE_PARM_DESC(ldlm_cpu_bind,
55 "bind DLM service threads to particular CPU partitions");
57 static char *ldlm_cpts;
58 module_param(ldlm_cpts, charp, 0444);
59 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
61 static DEFINE_MUTEX(ldlm_ref_mutex);
62 static int ldlm_refcount;
64 struct kobject *ldlm_kobj;
65 struct kset *ldlm_ns_kset;
66 struct kset *ldlm_svc_kset;
70 static struct ldlm_state *ldlm_state;
73 * timeout for initial callback (AST) reply (bz10399)
74 * Due to having to send a 32 bit time value over the
75 * wire return it as time_t instead of time64_t
77 static inline time_t ldlm_get_rq_timeout(void)
80 time_t timeout = min(ldlm_timeout, obd_timeout / 3);
82 return timeout < 1 ? 1 : timeout;
89 * blp_prio_list is used for callbacks that should be handled
90 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
93 struct list_head blp_prio_list;
96 * blp_list is used for all other callbacks which are likely
97 * to take longer to process.
99 struct list_head blp_list;
101 wait_queue_head_t blp_waitq;
102 struct completion blp_comp;
103 atomic_t blp_num_threads;
104 atomic_t blp_busy_threads;
109 struct ldlm_bl_work_item {
110 struct list_head blwi_entry;
111 struct ldlm_namespace *blwi_ns;
112 struct ldlm_lock_desc blwi_ld;
113 struct ldlm_lock *blwi_lock;
114 struct list_head blwi_head;
116 struct completion blwi_comp;
117 enum ldlm_cancel_flags blwi_flags;
118 int blwi_mem_pressure;
121 #ifdef HAVE_SERVER_SUPPORT
124 * Protects both waiting_locks_list and expired_lock_thread.
126 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
129 * List for contended locks.
131 * As soon as a lock is contended, it gets placed on this list and
132 * expected time to get a response is filled in the lock. A special
133 * thread walks the list looking for locks that should be released and
134 * schedules client evictions for those that have not been released in
137 * All access to it should be under waiting_locks_spinlock.
139 static LIST_HEAD(waiting_locks_list);
140 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
141 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
149 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
150 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
151 static int expired_lock_dump;
152 static LIST_HEAD(expired_lock_list);
154 static int ldlm_lock_busy(struct ldlm_lock *lock);
155 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout);
156 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds);
158 static inline int have_expired_locks(void)
163 spin_lock_bh(&waiting_locks_spinlock);
164 need_to_run = !list_empty(&expired_lock_list);
165 spin_unlock_bh(&waiting_locks_spinlock);
171 * Check expired lock list for expired locks and time them out.
173 static int expired_lock_main(void *arg)
175 struct list_head *expired = &expired_lock_list;
176 struct l_wait_info lwi = { 0 };
181 expired_lock_thread_state = ELT_READY;
182 wake_up(&expired_lock_wait_queue);
185 l_wait_event(expired_lock_wait_queue,
186 have_expired_locks() ||
187 expired_lock_thread_state == ELT_TERMINATE,
190 spin_lock_bh(&waiting_locks_spinlock);
191 if (expired_lock_dump) {
192 spin_unlock_bh(&waiting_locks_spinlock);
194 /* from waiting_locks_callback, but not in timer */
195 libcfs_debug_dumplog();
197 spin_lock_bh(&waiting_locks_spinlock);
198 expired_lock_dump = 0;
203 while (!list_empty(expired)) {
204 struct obd_export *export;
205 struct ldlm_lock *lock;
207 lock = list_entry(expired->next, struct ldlm_lock,
209 if ((void *)lock < LP_POISON + PAGE_SIZE &&
210 (void *)lock >= LP_POISON) {
211 spin_unlock_bh(&waiting_locks_spinlock);
212 CERROR("free lock on elt list %p\n", lock);
215 list_del_init(&lock->l_pending_chain);
216 if ((void *)lock->l_export <
217 LP_POISON + PAGE_SIZE &&
218 (void *)lock->l_export >= LP_POISON) {
219 CERROR("lock with free export on elt list %p\n",
221 lock->l_export = NULL;
222 LDLM_ERROR(lock, "free export");
224 * release extra ref grabbed by
225 * ldlm_add_waiting_lock() or
228 LDLM_LOCK_RELEASE(lock);
232 if (ldlm_is_destroyed(lock)) {
234 * release the lock refcount where
235 * waiting_locks_callback() founds
237 LDLM_LOCK_RELEASE(lock);
240 export = class_export_lock_get(lock->l_export, lock);
241 spin_unlock_bh(&waiting_locks_spinlock);
243 /* Check if we need to prolong timeout */
244 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
245 lock->l_callback_timeout != 0 && /* not AST error */
246 ldlm_lock_busy(lock)) {
247 LDLM_DEBUG(lock, "prolong the busy lock");
248 lock_res_and_lock(lock);
249 ldlm_add_waiting_lock(lock,
250 ldlm_bl_timeout(lock) >> 1);
251 unlock_res_and_lock(lock);
253 spin_lock_bh(&export->exp_bl_list_lock);
254 list_del_init(&lock->l_exp_list);
255 spin_unlock_bh(&export->exp_bl_list_lock);
258 "lock callback timer expired after %llds: evicting client at %s ",
259 ktime_get_real_seconds() -
261 obd_export_nid2str(export));
262 ldlm_lock_to_ns(lock)->ns_timeouts++;
264 class_fail_export(export);
266 class_export_lock_put(export, lock);
268 * release extra ref grabbed by ldlm_add_waiting_lock()
269 * or ldlm_failed_ast()
271 LDLM_LOCK_RELEASE(lock);
273 spin_lock_bh(&waiting_locks_spinlock);
275 spin_unlock_bh(&waiting_locks_spinlock);
277 if (do_dump && obd_dump_on_eviction) {
278 CERROR("dump the log upon eviction\n");
279 libcfs_debug_dumplog();
282 if (expired_lock_thread_state == ELT_TERMINATE)
286 expired_lock_thread_state = ELT_STOPPED;
287 wake_up(&expired_lock_wait_queue);
292 * Check if there is a request in the export request list
293 * which prevents the lock canceling.
295 static int ldlm_lock_busy(struct ldlm_lock *lock)
297 struct ptlrpc_request *req;
302 if (lock->l_export == NULL)
305 spin_lock(&lock->l_export->exp_rpc_lock);
306 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
308 if (req->rq_ops->hpreq_lock_match) {
309 match = req->rq_ops->hpreq_lock_match(req, lock);
314 spin_unlock(&lock->l_export->exp_rpc_lock);
318 /* This is called from within a timer interrupt and cannot schedule */
319 static void waiting_locks_callback(TIMER_DATA_TYPE unused)
321 struct ldlm_lock *lock;
324 spin_lock_bh(&waiting_locks_spinlock);
325 while (!list_empty(&waiting_locks_list)) {
326 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
328 if (lock->l_callback_timeout > ktime_get_seconds() ||
329 lock->l_req_mode == LCK_GROUP)
333 * no needs to take an extra ref on the lock since it was in
334 * the waiting_locks_list and ldlm_add_waiting_lock()
335 * already grabbed a ref
337 list_del(&lock->l_pending_chain);
338 list_add(&lock->l_pending_chain, &expired_lock_list);
342 if (!list_empty(&expired_lock_list)) {
343 if (obd_dump_on_timeout && need_dump)
344 expired_lock_dump = __LINE__;
346 wake_up(&expired_lock_wait_queue);
350 * Make sure the timer will fire again if we have any locks
353 if (!list_empty(&waiting_locks_list)) {
354 unsigned long timeout_jiffies;
356 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
358 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
359 mod_timer(&waiting_locks_timer, timeout_jiffies);
361 spin_unlock_bh(&waiting_locks_spinlock);
365 * Add lock to the list of contended locks.
367 * Indicate that we're waiting for a client to call us back cancelling a given
368 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
369 * timer to fire appropriately. (We round up to the next second, to avoid
370 * floods of timer firings during periods of high lock contention and traffic).
371 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
372 * if it has been added to the waiting list (1 is returned).
374 * Called with the namespace lock held.
376 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds)
378 unsigned long timeout_jiffies;
381 if (!list_empty(&lock->l_pending_chain))
384 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
385 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
388 timeout = ktime_get_seconds() + seconds;
389 if (likely(timeout > lock->l_callback_timeout))
390 lock->l_callback_timeout = timeout;
392 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
394 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
395 !timer_pending(&waiting_locks_timer))
396 mod_timer(&waiting_locks_timer, timeout_jiffies);
399 * if the new lock has a shorter timeout than something earlier on
400 * the list, we'll wait the longer amount of time; no big deal.
403 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
407 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
409 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
410 if (list_empty(&lock->l_exp_list)) {
411 if (!ldlm_is_granted(lock))
412 list_add_tail(&lock->l_exp_list,
413 &lock->l_export->exp_bl_list);
415 list_add(&lock->l_exp_list,
416 &lock->l_export->exp_bl_list);
418 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
421 * A blocked lock is added. Adjust the position in
422 * the stale list if the export is in the list.
423 * If export is stale and not in the list - it is being
424 * processed and will be placed on the right position
425 * on obd_stale_export_put().
427 if (!list_empty(&lock->l_export->exp_stale_list))
428 obd_stale_export_adjust(lock->l_export);
431 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
435 /* NB: must be called with hold of lock_res_and_lock() */
436 LASSERT(ldlm_is_res_locked(lock));
437 LASSERT(!ldlm_is_cancel_on_block(lock));
440 * Do not put cross-MDT lock in the waiting list, since we
441 * will not evict it due to timeout for now
443 if (lock->l_export != NULL &&
444 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
447 spin_lock_bh(&waiting_locks_spinlock);
448 if (ldlm_is_cancel(lock)) {
449 spin_unlock_bh(&waiting_locks_spinlock);
453 if (ldlm_is_destroyed(lock)) {
454 static time64_t next;
456 spin_unlock_bh(&waiting_locks_spinlock);
457 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
458 if (ktime_get_seconds() > next) {
459 next = ktime_get_seconds() + 14400;
460 libcfs_debug_dumpstack(NULL);
465 ldlm_set_waited(lock);
466 lock->l_blast_sent = ktime_get_real_seconds();
467 ret = __ldlm_add_waiting_lock(lock, timeout);
470 * grab ref on the lock if it has been added to the
475 spin_unlock_bh(&waiting_locks_spinlock);
478 ldlm_add_blocked_lock(lock);
480 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)",
481 ret == 0 ? "not re-" : "", timeout,
482 AT_OFF ? "off" : "on");
487 * Remove a lock from the pending list, likely because it had its cancellation
488 * callback arrive without incident. This adjusts the lock-timeout timer if
489 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
490 * As done by ldlm_del_waiting_lock(), the caller must release the lock
491 * reference when the lock is removed from any list (1 is returned).
493 * Called with namespace lock held.
495 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
497 struct list_head *list_next;
499 if (list_empty(&lock->l_pending_chain))
502 list_next = lock->l_pending_chain.next;
503 if (lock->l_pending_chain.prev == &waiting_locks_list) {
504 /* Removing the head of the list, adjust timer. */
505 if (list_next == &waiting_locks_list) {
506 /* No more, just cancel. */
507 del_timer(&waiting_locks_timer);
509 struct ldlm_lock *next;
511 next = list_entry(list_next, struct ldlm_lock,
513 mod_timer(&waiting_locks_timer,
514 cfs_time_seconds(next->l_callback_timeout));
517 list_del_init(&lock->l_pending_chain);
522 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
526 if (lock->l_export == NULL) {
527 /* We don't have a "waiting locks list" on clients. */
528 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
532 spin_lock_bh(&waiting_locks_spinlock);
533 ret = __ldlm_del_waiting_lock(lock);
534 ldlm_clear_waited(lock);
535 spin_unlock_bh(&waiting_locks_spinlock);
537 /* remove the lock out of export blocking list */
538 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
539 list_del_init(&lock->l_exp_list);
540 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
544 * release lock ref if it has indeed been removed
547 LDLM_LOCK_RELEASE(lock);
550 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
555 * Prolong the contended lock waiting time.
557 * Called with namespace lock held.
559 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
561 if (lock->l_export == NULL) {
562 /* We don't have a "waiting locks list" on clients. */
563 LDLM_DEBUG(lock, "client lock: no-op");
567 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
568 /* We don't have a "waiting locks list" on OSP. */
569 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
573 spin_lock_bh(&waiting_locks_spinlock);
575 if (list_empty(&lock->l_pending_chain)) {
576 spin_unlock_bh(&waiting_locks_spinlock);
577 LDLM_DEBUG(lock, "wasn't waiting");
582 * we remove/add the lock to the waiting list, so no needs to
583 * release/take a lock reference
585 __ldlm_del_waiting_lock(lock);
586 __ldlm_add_waiting_lock(lock, timeout);
587 spin_unlock_bh(&waiting_locks_spinlock);
589 LDLM_DEBUG(lock, "refreshed");
592 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
594 #else /* HAVE_SERVER_SUPPORT */
596 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
601 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
606 #endif /* !HAVE_SERVER_SUPPORT */
608 #ifdef HAVE_SERVER_SUPPORT
611 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
612 * lock cancel, and their replies). Used for lock callback timeout and AST
615 * \param[in] lock lock which is getting the blocking callback
617 * \retval timeout in seconds to wait for the client reply
619 time64_t ldlm_bl_timeout(struct ldlm_lock *lock)
624 return obd_timeout / 2;
627 * Since these are non-updating timeouts, we should be conservative.
628 * Take more than usually, 150%
629 * It would be nice to have some kind of "early reply" mechanism for
630 * lock callbacks too...
632 timeout = at_get(&lock->l_export->exp_bl_lock_at);
633 return max(timeout + (timeout >> 1), (time64_t)ldlm_enqueue_min);
635 EXPORT_SYMBOL(ldlm_bl_timeout);
638 * Perform lock cleanup if AST sending failed.
640 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
641 const char *ast_type)
643 LCONSOLE_ERROR_MSG(0x138,
644 "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
645 lock->l_export->exp_obd->obd_name,
646 obd_export_nid2str(lock->l_export), ast_type, rc);
648 if (obd_dump_on_timeout)
649 libcfs_debug_dumplog();
650 spin_lock_bh(&waiting_locks_spinlock);
651 if (__ldlm_del_waiting_lock(lock) == 0)
653 * the lock was not in any list, grab an extra ref before adding
654 * the lock to the expired list
657 lock->l_callback_timeout = 0; /* differentiate it from expired locks */
658 list_add(&lock->l_pending_chain, &expired_lock_list);
659 wake_up(&expired_lock_wait_queue);
660 spin_unlock_bh(&waiting_locks_spinlock);
664 * Perform lock cleanup if AST reply came with error.
666 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
667 struct ptlrpc_request *req, int rc,
668 const char *ast_type)
670 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
672 if (!req->rq_replied || (rc && rc != -EINVAL)) {
673 if (ldlm_is_cancel(lock)) {
675 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
676 ast_type, req, req->rq_xid,
677 libcfs_nid2str(peer.nid));
678 ldlm_lock_cancel(lock);
680 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
682 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
684 * Upon umount process the AST fails because cannot be
685 * sent. This shouldn't lead to the client eviction.
686 * -ENODEV error is returned by ptl_send_rpc() for
687 * new request in such import.
688 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
689 * if imp_invalid is set or obd_no_recov.
690 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
691 * in ptlrpc_import_delay_req() as well with -EIO code.
692 * In all such cases errors are ignored.
695 "%s AST can't be sent due to a server %s failure or umount process: rc = %d\n",
697 req->rq_import->imp_obd->obd_name, rc);
700 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
701 libcfs_nid2str(peer.nid),
702 req->rq_replied ? "returned error from" :
703 "failed to reply to",
704 ast_type, req, req->rq_xid,
705 (req->rq_repmsg != NULL) ?
706 lustre_msg_get_status(req->rq_repmsg) : 0,
708 ldlm_failed_ast(lock, rc, ast_type);
714 struct ldlm_resource *res = lock->l_resource;
717 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
718 libcfs_nid2str(peer.nid),
720 lustre_msg_get_status(req->rq_repmsg) : -1,
721 ast_type, req, req->rq_xid);
724 * update lvbo to return proper attributes.
727 ldlm_resource_getref(res);
728 ldlm_lvbo_update(res, lock, NULL, 1);
729 ldlm_resource_putref(res);
731 ldlm_lock_cancel(lock);
738 static int ldlm_cb_interpret(const struct lu_env *env,
739 struct ptlrpc_request *req, void *args, int rc)
741 struct ldlm_cb_async_args *ca = args;
742 struct ldlm_lock *lock = ca->ca_lock;
743 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
747 LASSERT(lock != NULL);
750 case LDLM_GL_CALLBACK:
752 * Update the LVB from disk if the AST failed
753 * (this is a legal race)
755 * - Glimpse callback of local lock just returns
756 * -ELDLM_NO_LOCK_DATA.
757 * - Glimpse callback of remote lock might return
758 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
760 if (unlikely(arg->gl_interpret_reply)) {
761 rc = arg->gl_interpret_reply(NULL, req, args, rc);
762 } else if (rc == -ELDLM_NO_LOCK_DATA) {
764 "lost race - client has a lock but no inode");
765 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
766 } else if (rc != 0) {
767 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
769 rc = ldlm_lvbo_update(lock->l_resource,
773 case LDLM_BL_CALLBACK:
775 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
777 case LDLM_CP_CALLBACK:
779 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
782 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
787 /* release extra reference taken in ldlm_ast_fini() */
788 LDLM_LOCK_RELEASE(lock);
791 atomic_inc(&arg->restart);
796 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
798 struct ldlm_cb_async_args *ca = data;
799 struct ldlm_lock *lock = ca->ca_lock;
801 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
804 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
805 struct ldlm_cb_set_arg *arg,
806 struct ldlm_lock *lock,
813 if (unlikely(instant_cancel)) {
814 rc = ptl_send_rpc(req, 1);
815 ptlrpc_req_finished(req);
817 atomic_inc(&arg->restart);
820 ptlrpc_set_add_req(arg->set, req);
827 * Check if there are requests in the export request list which prevent
828 * the lock canceling and make these requests high priority ones.
830 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
832 struct ptlrpc_request *req;
836 if (lock->l_export == NULL) {
837 LDLM_DEBUG(lock, "client lock: no-op");
841 spin_lock(&lock->l_export->exp_rpc_lock);
842 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
845 * Do not process requests that were not yet added to there
846 * incoming queue or were already removed from there for
847 * processing. We evaluate ptlrpc_nrs_req_can_move() without
848 * holding svcpt->scp_req_lock, and then redo the check with
849 * the lock held once we need to obtain a reliable result.
851 if (ptlrpc_nrs_req_can_move(req) &&
852 req->rq_ops->hpreq_lock_match &&
853 req->rq_ops->hpreq_lock_match(req, lock))
854 ptlrpc_nrs_req_hp_move(req);
856 spin_unlock(&lock->l_export->exp_rpc_lock);
861 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
862 * enqueued server lock conflicts with given one.
864 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
865 * to wait for client response.
867 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
868 struct ldlm_lock_desc *desc,
869 void *data, int flag)
871 struct ldlm_cb_async_args *ca;
872 struct ldlm_cb_set_arg *arg = data;
873 struct ldlm_request *body;
874 struct ptlrpc_request *req;
875 int instant_cancel = 0;
880 if (flag == LDLM_CB_CANCELING)
881 /* Don't need to do anything here. */
884 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
885 LDLM_DEBUG(lock, "dropping BL AST");
890 LASSERT(data != NULL);
891 if (lock->l_export->exp_obd->obd_recovering != 0)
892 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
894 ldlm_lock_reorder_req(lock);
896 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
897 &RQF_LDLM_BL_CALLBACK,
898 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
902 ca = ptlrpc_req_async_args(ca, req);
903 ca->ca_set_arg = arg;
906 req->rq_interpret_reply = ldlm_cb_interpret;
908 lock_res_and_lock(lock);
909 if (ldlm_is_destroyed(lock)) {
910 /* What's the point? */
911 unlock_res_and_lock(lock);
912 ptlrpc_req_finished(req);
916 if (!ldlm_is_granted(lock)) {
918 * this blocking AST will be communicated as part of the
919 * completion AST instead
921 ldlm_add_blocked_lock(lock);
922 ldlm_set_waited(lock);
923 unlock_res_and_lock(lock);
925 ptlrpc_req_finished(req);
926 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
930 if (ldlm_is_cancel_on_block(lock))
933 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
934 body->lock_handle[0] = lock->l_remote_handle;
935 body->lock_desc = *desc;
936 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
938 LDLM_DEBUG(lock, "server preparing blocking AST");
940 ptlrpc_request_set_replen(req);
941 ldlm_set_cbpending(lock);
942 if (instant_cancel) {
943 unlock_res_and_lock(lock);
944 ldlm_lock_cancel(lock);
946 req->rq_no_resend = 1;
948 LASSERT(ldlm_is_granted(lock));
949 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
950 unlock_res_and_lock(lock);
952 /* Do not resend after lock callback timeout */
953 req->rq_delay_limit = ldlm_bl_timeout(lock);
954 req->rq_resend_cb = ldlm_update_resend;
957 req->rq_send_state = LUSTRE_IMP_FULL;
958 /* ptlrpc_request_alloc_pack already set timeout */
960 req->rq_timeout = ldlm_get_rq_timeout();
962 if (lock->l_export && lock->l_export->exp_nid_stats &&
963 lock->l_export->exp_nid_stats->nid_ldlm_stats)
964 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
965 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
967 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
973 * ->l_completion_ast callback for a remote lock in server namespace.
975 * Sends AST to the client notifying it of lock granting. If initial
976 * lock response was not sent yet, instead of sending another RPC, just
977 * mark the lock as granted and client will understand
979 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
981 struct ldlm_cb_set_arg *arg = data;
982 struct ldlm_request *body;
983 struct ptlrpc_request *req;
984 struct ldlm_cb_async_args *ca;
985 int instant_cancel = 0;
991 LASSERT(lock != NULL);
992 LASSERT(data != NULL);
994 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
995 LDLM_DEBUG(lock, "dropping CP AST");
999 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1000 &RQF_LDLM_CP_CALLBACK);
1004 /* server namespace, doesn't need lock */
1005 lvb_len = ldlm_lvbo_size(lock);
1007 * LU-3124 & LU-2187: to not return layout in completion AST because
1008 * it may deadlock for LU-2187, or client may not have enough space
1009 * for large layout. The layout will be returned to client with an
1010 * extra RPC to fetch xattr.lov
1012 if (ldlm_has_layout(lock))
1015 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1016 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1018 ptlrpc_request_free(req);
1022 ca = ptlrpc_req_async_args(ca, req);
1023 ca->ca_set_arg = arg;
1026 req->rq_interpret_reply = ldlm_cb_interpret;
1027 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1029 body->lock_handle[0] = lock->l_remote_handle;
1030 body->lock_flags = ldlm_flags_to_wire(flags);
1031 ldlm_lock2desc(lock, &body->lock_desc);
1033 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1034 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1037 * We still need to send the RPC to wake up the blocked
1038 * enqueue thread on the client.
1040 * Consider old client, there is no better way to notify
1041 * the failure, just zero-sized the LVB, then the client
1042 * will fail out as "-EPROTO".
1044 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1048 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1053 LDLM_DEBUG(lock, "server preparing completion AST");
1055 ptlrpc_request_set_replen(req);
1057 req->rq_send_state = LUSTRE_IMP_FULL;
1058 /* ptlrpc_request_pack already set timeout */
1060 req->rq_timeout = ldlm_get_rq_timeout();
1062 /* We only send real blocking ASTs after the lock is granted */
1063 lock_res_and_lock(lock);
1064 if (ldlm_is_ast_sent(lock)) {
1065 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1066 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1067 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1071 * We might get here prior to ldlm_handle_enqueue setting
1072 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1073 * into waiting list, but this is safe and similar code in
1074 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1075 * that would not only cancel the lock, but will also remove
1076 * it from waiting list
1078 if (ldlm_is_cancel_on_block(lock)) {
1079 unlock_res_and_lock(lock);
1080 ldlm_lock_cancel(lock);
1083 req->rq_no_resend = 1;
1085 lock_res_and_lock(lock);
1087 /* start the lock-timeout clock */
1088 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1089 /* Do not resend after lock callback timeout */
1090 req->rq_delay_limit = ldlm_bl_timeout(lock);
1091 req->rq_resend_cb = ldlm_update_resend;
1094 unlock_res_and_lock(lock);
1096 if (lock->l_export && lock->l_export->exp_nid_stats &&
1097 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1098 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1099 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1101 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1103 RETURN(lvb_len < 0 ? lvb_len : rc);
1107 * Server side ->l_glimpse_ast handler for client locks.
1109 * Sends glimpse AST to the client and waits for reply. Then updates
1110 * lvbo with the result.
1112 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1114 struct ldlm_cb_set_arg *arg = data;
1115 struct ldlm_request *body;
1116 struct ptlrpc_request *req;
1117 struct ldlm_cb_async_args *ca;
1119 struct req_format *req_fmt;
1123 LASSERT(lock != NULL);
1125 if (arg->gl_desc != NULL)
1126 /* There is a glimpse descriptor to pack */
1127 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1129 req_fmt = &RQF_LDLM_GL_CALLBACK;
1131 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1132 req_fmt, LUSTRE_DLM_VERSION,
1138 if (arg->gl_desc != NULL) {
1139 /* copy the GL descriptor */
1140 union ldlm_gl_desc *desc;
1142 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1143 *desc = *arg->gl_desc;
1146 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1147 body->lock_handle[0] = lock->l_remote_handle;
1148 ldlm_lock2desc(lock, &body->lock_desc);
1150 ca = ptlrpc_req_async_args(ca, req);
1151 ca->ca_set_arg = arg;
1154 /* server namespace, doesn't need lock */
1155 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1156 ldlm_lvbo_size(lock));
1157 ptlrpc_request_set_replen(req);
1159 req->rq_send_state = LUSTRE_IMP_FULL;
1160 /* ptlrpc_request_alloc_pack already set timeout */
1162 req->rq_timeout = ldlm_get_rq_timeout();
1164 req->rq_interpret_reply = ldlm_cb_interpret;
1166 if (lock->l_export && lock->l_export->exp_nid_stats) {
1167 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1169 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1170 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1173 rc = ldlm_ast_fini(req, arg, lock, 0);
1177 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1179 int ldlm_glimpse_locks(struct ldlm_resource *res,
1180 struct list_head *gl_work_list)
1186 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1188 if (rc == -ERESTART)
1189 ldlm_reprocess_all(res, NULL);
1193 EXPORT_SYMBOL(ldlm_glimpse_locks);
1195 /* return LDLM lock associated with a lock callback request */
1196 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1198 struct ldlm_cb_async_args *ca;
1199 struct ldlm_lock *lock;
1203 ca = ptlrpc_req_async_args(ca, req);
1206 RETURN(ERR_PTR(-EFAULT));
1210 EXPORT_SYMBOL(ldlm_request_lock);
1213 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1214 * service threads to carry out client lock enqueueing requests.
1216 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1217 struct ptlrpc_request *req,
1218 const struct ldlm_request *dlm_req,
1219 const struct ldlm_callback_suite *cbs)
1221 struct ldlm_reply *dlm_rep;
1223 enum ldlm_error err = ELDLM_OK;
1224 struct ldlm_lock *lock = NULL;
1225 void *cookie = NULL;
1227 struct ldlm_resource *res = NULL;
1228 const struct lu_env *env = req->rq_svc_thread->t_env;
1232 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1234 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1235 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1237 LASSERT(req->rq_export);
1239 /* for intent enqueue the stat will be updated inside intent policy */
1240 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1241 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1242 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1244 if (req->rq_export && req->rq_export->exp_nid_stats &&
1245 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1246 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1247 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1249 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1250 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1251 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1252 dlm_req->lock_desc.l_resource.lr_type);
1253 GOTO(out, rc = -EFAULT);
1256 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1257 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1258 dlm_req->lock_desc.l_req_mode &
1259 (dlm_req->lock_desc.l_req_mode-1))) {
1260 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1261 dlm_req->lock_desc.l_req_mode);
1262 GOTO(out, rc = -EFAULT);
1265 if (unlikely((flags & LDLM_FL_REPLAY) ||
1266 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1267 /* Find an existing lock in the per-export lock hash */
1269 * In the function below, .hs_keycmp resolves to
1270 * ldlm_export_lock_keycmp()
1272 /* coverity[overrun-buffer-val] */
1273 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1274 (void *)&dlm_req->lock_handle[0]);
1276 DEBUG_REQ(D_DLMTRACE, req,
1277 "found existing lock cookie %#llx",
1278 lock->l_handle.h_cookie);
1279 flags |= LDLM_FL_RESENT;
1280 GOTO(existing_lock, rc = 0);
1283 if (ldlm_reclaim_full()) {
1284 DEBUG_REQ(D_DLMTRACE, req,
1285 "Too many granted locks, reject current enqueue request and let the client retry later");
1286 GOTO(out, rc = -EINPROGRESS);
1290 /* The lock's callback data might be set in the policy function */
1291 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1292 dlm_req->lock_desc.l_resource.lr_type,
1293 dlm_req->lock_desc.l_req_mode,
1294 cbs, NULL, 0, LVB_T_NONE);
1301 lock->l_remote_handle = dlm_req->lock_handle[0];
1302 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1305 * Initialize resource lvb but not for a lock being replayed since
1306 * Client already got lvb sent in this case.
1307 * This must occur early since some policy methods assume resource
1308 * lvb is available (lr_lvb_data != NULL).
1310 res = lock->l_resource;
1311 if (!(flags & LDLM_FL_REPLAY)) {
1312 /* non-replayed lock, delayed lvb init may need to be done */
1313 rc = ldlm_lvbo_init(res);
1315 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1320 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1322 * Don't enqueue a lock onto the export if it is been disonnected
1323 * due to eviction (b=3822) or server umount (b=24324).
1324 * Cancel it now instead.
1326 if (req->rq_export->exp_disconnected) {
1327 LDLM_ERROR(lock, "lock on disconnected export %p",
1329 GOTO(out, rc = -ENOTCONN);
1332 lock->l_export = class_export_lock_get(req->rq_export, lock);
1333 if (lock->l_export->exp_lock_hash)
1334 cfs_hash_add(lock->l_export->exp_lock_hash,
1335 &lock->l_remote_handle,
1339 * Inherit the enqueue flags before the operation, because we do not
1340 * keep the res lock on return and next operations (BL AST) may proceed
1343 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1344 LDLM_FL_INHERIT_MASK);
1346 ldlm_convert_policy_to_local(req->rq_export,
1347 dlm_req->lock_desc.l_resource.lr_type,
1348 &dlm_req->lock_desc.l_policy_data,
1349 &lock->l_policy_data);
1350 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1351 lock->l_req_extent = lock->l_policy_data.l_extent;
1354 if (flags & LDLM_FL_HAS_INTENT) {
1356 * In this case, the reply buffer is allocated deep in
1357 * local_lock_enqueue by the policy function.
1362 * based on the assumption that lvb size never changes during
1363 * resource life time otherwise it need resource->lr_lock's
1366 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1367 RCL_SERVER, ldlm_lvbo_size(lock));
1369 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1370 GOTO(out, rc = -ENOMEM);
1372 rc = req_capsule_server_pack(&req->rq_pill);
1377 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1384 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1386 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1387 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1389 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1390 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1393 * We never send a blocking AST until the lock is granted, but
1394 * we can tell it right now
1396 lock_res_and_lock(lock);
1399 * Now take into account flags to be inherited from original lock
1400 * request both in reply to client and in our own lock flags.
1402 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1403 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1406 * Don't move a pending lock onto the export if it has already been
1407 * disconnected due to eviction (b=5683) or server umount (b=24324).
1408 * Cancel it now instead.
1410 if (unlikely(req->rq_export->exp_disconnected ||
1411 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1412 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1414 } else if (ldlm_is_ast_sent(lock)) {
1415 /* fill lock desc for possible lock convert */
1416 if (lock->l_blocking_lock &&
1417 lock->l_resource->lr_type == LDLM_IBITS) {
1418 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1419 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1422 "save blocking bits %llx in granted lock",
1423 bl_lock->l_policy_data.l_inodebits.bits);
1425 * If lock is blocked then save blocking ibits
1426 * in returned lock policy for the possible lock
1427 * convert on a client.
1429 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1430 bl_lock->l_policy_data.l_inodebits.bits;
1432 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1433 if (ldlm_is_granted(lock)) {
1435 * Only cancel lock if it was granted, because it would
1436 * be destroyed immediately and would never be granted
1437 * in the future, causing timeouts on client. Not
1438 * granted lock will be cancelled immediately after
1439 * sending completion AST.
1441 if (ldlm_is_cancel_on_block(lock)) {
1442 unlock_res_and_lock(lock);
1443 ldlm_lock_cancel(lock);
1444 lock_res_and_lock(lock);
1446 ldlm_add_waiting_lock(lock,
1447 ldlm_bl_timeout(lock));
1451 unlock_res_and_lock(lock);
1455 req->rq_status = rc ?: err; /* return either error - b=11190 */
1456 if (!req->rq_packed_final) {
1457 err = lustre_pack_reply(req, 1, NULL, NULL);
1463 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1464 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1468 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1472 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1474 ldlm_lvbo_size(lock) > 0) {
1479 buf = req_capsule_server_get(&req->rq_pill,
1481 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1482 buflen = req_capsule_get_size(&req->rq_pill,
1483 &RMF_DLM_LVB, RCL_SERVER);
1485 * non-replayed lock, delayed lvb init may
1486 * need to be occur now
1488 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1491 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1493 req_capsule_shrink(&req->rq_pill,
1496 } else if (rc2 == -ERANGE) {
1497 rc2 = req_capsule_server_grow(
1499 &RMF_DLM_LVB, buflen);
1504 * if we can't grow the buffer,
1505 * it's ok to return empty lvb
1516 } else if (flags & LDLM_FL_REPLAY) {
1517 /* no LVB resend upon replay */
1519 req_capsule_shrink(&req->rq_pill,
1529 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1530 if (lock->l_export) {
1531 ldlm_lock_cancel(lock);
1533 lock_res_and_lock(lock);
1534 ldlm_resource_unlink_lock(lock);
1535 ldlm_lock_destroy_nolock(lock);
1536 unlock_res_and_lock(lock);
1541 if (!err && !ldlm_is_cbpending(lock) &&
1542 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1543 ldlm_reprocess_all(lock->l_resource, lock);
1545 LDLM_LOCK_RELEASE(lock);
1548 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1555 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1556 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1558 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1560 if (lock->l_blocking_lock) {
1561 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1562 lock->l_blocking_lock = NULL;
1566 /* A lock can be converted to new ibits or mode and should be considered
1567 * as new lock. Clear all states related to a previous blocking AST
1568 * processing so new conflicts will cause new blocking ASTs.
1570 * This is used during lock convert below and lock downgrade to COS mode in
1571 * ldlm_lock_mode_downgrade().
1573 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1575 ldlm_clear_ast_sent(lock);
1576 lock->l_bl_ast_run = 0;
1577 ldlm_clear_blocking_lock(lock);
1581 * Main LDLM entry point for server code to process lock conversion requests.
1583 int ldlm_handle_convert0(struct ptlrpc_request *req,
1584 const struct ldlm_request *dlm_req)
1586 struct obd_export *exp = req->rq_export;
1587 struct ldlm_reply *dlm_rep;
1588 struct ldlm_lock *lock;
1593 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1594 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1595 LDLM_CONVERT - LDLM_FIRST_OPC);
1597 rc = req_capsule_server_pack(&req->rq_pill);
1601 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1602 dlm_rep->lock_flags = dlm_req->lock_flags;
1604 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1609 bits = lock->l_policy_data.l_inodebits.bits;
1610 new = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1611 LDLM_DEBUG(lock, "server-side convert handler START");
1613 if (ldlm_is_cancel(lock)) {
1614 LDLM_ERROR(lock, "convert on canceled lock!");
1615 rc = ELDLM_NO_LOCK_DATA;
1616 } else if (dlm_req->lock_desc.l_req_mode !=
1617 lock->l_granted_mode) {
1618 LDLM_ERROR(lock, "lock mode differs!");
1619 rc = ELDLM_NO_LOCK_DATA;
1620 } else if (bits == new) {
1622 * This can be valid situation if CONVERT RPCs are
1623 * re-ordered. Just finish silently
1625 LDLM_DEBUG(lock, "lock is converted already!");
1628 lock_res_and_lock(lock);
1629 if (ldlm_is_waited(lock))
1630 ldlm_del_waiting_lock(lock);
1632 ldlm_clear_cbpending(lock);
1633 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1634 ldlm_inodebits_drop(lock, bits & ~new);
1636 ldlm_clear_blocking_data(lock);
1637 unlock_res_and_lock(lock);
1639 ldlm_reprocess_all(lock->l_resource, NULL);
1643 if (rc == ELDLM_OK) {
1644 dlm_rep->lock_handle = lock->l_remote_handle;
1645 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1646 &dlm_rep->lock_desc.l_policy_data);
1649 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d",
1651 LDLM_LOCK_PUT(lock);
1653 rc = ELDLM_NO_LOCK_DATA;
1654 LDLM_DEBUG_NOLOCK("server-side convert handler END, rc = %d",
1658 req->rq_status = rc;
1664 * Cancel all the locks whose handles are packed into ldlm_request
1666 * Called by server code expecting such combined cancel activity
1669 int ldlm_request_cancel(struct ptlrpc_request *req,
1670 const struct ldlm_request *dlm_req,
1671 int first, enum lustre_at_flags flags)
1673 struct ldlm_resource *res, *pres = NULL;
1674 struct ldlm_lock *lock;
1675 int i, count, done = 0;
1680 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1681 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1682 (size - offsetof(struct ldlm_request, lock_handle)) /
1683 sizeof(struct lustre_handle) < dlm_req->lock_count)
1686 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1690 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1694 * There is no lock on the server at the replay time,
1695 * skip lock cancelling to make replay tests to pass.
1697 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1700 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1703 for (i = first; i < count; i++) {
1704 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1706 /* below message checked in replay-single.sh test_36 */
1707 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1708 dlm_req->lock_handle[i].cookie);
1712 res = lock->l_resource;
1716 * This code is an optimization to only attempt lock
1717 * granting on the resource (that could be CPU-expensive)
1718 * after we are done cancelling lock in that resource.
1722 ldlm_reprocess_all(pres, NULL);
1723 LDLM_RESOURCE_DELREF(pres);
1724 ldlm_resource_putref(pres);
1727 ldlm_resource_getref(res);
1728 LDLM_RESOURCE_ADDREF(res);
1730 if (!ldlm_is_discard_data(lock))
1731 ldlm_lvbo_update(res, lock,
1737 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1738 lock->l_blast_sent != 0) {
1739 time64_t delay = ktime_get_real_seconds() -
1742 "server cancels blocked lock after %llds",
1744 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1746 ldlm_lock_cancel(lock);
1747 LDLM_LOCK_PUT(lock);
1750 ldlm_reprocess_all(pres, NULL);
1751 LDLM_RESOURCE_DELREF(pres);
1752 ldlm_resource_putref(pres);
1754 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1757 EXPORT_SYMBOL(ldlm_request_cancel);
1760 * Main LDLM entry point for server code to cancel locks.
1762 * Typically gets called from service handler on LDLM_CANCEL opc.
1764 int ldlm_handle_cancel(struct ptlrpc_request *req)
1766 struct ldlm_request *dlm_req;
1771 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1772 if (dlm_req == NULL) {
1773 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1777 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1778 offsetof(struct ldlm_request, lock_handle[1]))
1781 if (req->rq_export && req->rq_export->exp_nid_stats &&
1782 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1783 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1784 LDLM_CANCEL - LDLM_FIRST_OPC);
1786 rc = req_capsule_server_pack(&req->rq_pill);
1790 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1791 req->rq_status = LUSTRE_ESTALE;
1793 RETURN(ptlrpc_reply(req));
1795 #endif /* HAVE_SERVER_SUPPORT */
1798 * Callback handler for receiving incoming blocking ASTs.
1800 * This can only happen on client side.
1802 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1803 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1809 LDLM_DEBUG(lock, "client blocking AST callback handler");
1811 lock_res_and_lock(lock);
1813 /* set bits to cancel for this lock for possible lock convert */
1814 if (ns_is_client(ns) && (lock->l_resource->lr_type == LDLM_IBITS)) {
1816 * Lock description contains policy of blocking lock,
1817 * and its cancel_bits is used to pass conflicting bits.
1818 * NOTE: ld can be NULL or can be not NULL but zeroed if
1819 * passed from ldlm_bl_thread_blwi(), check below used bits
1820 * in ld to make sure it is valid description.
1822 * If server may replace lock resource keeping the same cookie,
1823 * never use cancel bits from different resource, full cancel
1826 if (ld && ld->l_policy_data.l_inodebits.bits &&
1827 ldlm_res_eq(&ld->l_resource.lr_name,
1828 &lock->l_resource->lr_name))
1829 lock->l_policy_data.l_inodebits.cancel_bits =
1830 ld->l_policy_data.l_inodebits.cancel_bits;
1832 * if there is no valid ld and lock is cbpending already
1833 * then cancel_bits should be kept, otherwise it is zeroed.
1835 else if (!ldlm_is_cbpending(lock))
1836 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1838 ldlm_set_cbpending(lock);
1840 do_ast = (!lock->l_readers && !lock->l_writers);
1841 unlock_res_and_lock(lock);
1845 "Lock %p already unused, calling callback (%p)\n",
1846 lock, lock->l_blocking_ast);
1847 if (lock->l_blocking_ast != NULL)
1848 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1852 "Lock %p is referenced, will be cancelled later\n",
1856 LDLM_DEBUG(lock, "client blocking callback handler END");
1857 LDLM_LOCK_RELEASE(lock);
1862 * Callback handler for receiving incoming completion ASTs.
1864 * This only can happen on client side.
1866 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1867 struct ldlm_namespace *ns,
1868 struct ldlm_request *dlm_req,
1869 struct ldlm_lock *lock)
1871 struct list_head ast_list;
1877 LDLM_DEBUG(lock, "client completion callback handler START");
1879 INIT_LIST_HEAD(&ast_list);
1880 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1881 long to = cfs_time_seconds(1);
1884 set_current_state(TASK_INTERRUPTIBLE);
1885 schedule_timeout(to);
1886 if (ldlm_is_granted(lock) ||
1887 ldlm_is_destroyed(lock))
1892 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1894 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1895 GOTO(out, rc = lvb_len);
1896 } else if (lvb_len > 0) {
1897 if (lock->l_lvb_len > 0) {
1898 /* for extent lock, lvb contains ost_lvb{}. */
1899 LASSERT(lock->l_lvb_data != NULL);
1901 if (unlikely(lock->l_lvb_len < lvb_len)) {
1903 "Replied LVB is larger than expectation, expected = %d, replied = %d",
1904 lock->l_lvb_len, lvb_len);
1905 GOTO(out, rc = -EINVAL);
1910 lock_res_and_lock(lock);
1912 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1913 &lock->l_resource->lr_name)) {
1914 ldlm_resource_unlink_lock(lock);
1915 unlock_res_and_lock(lock);
1916 rc = ldlm_lock_change_resource(ns, lock,
1917 &dlm_req->lock_desc.l_resource.lr_name);
1919 LDLM_ERROR(lock, "Failed to allocate resource");
1922 LDLM_DEBUG(lock, "completion AST, new resource");
1923 lock_res_and_lock(lock);
1926 if (ldlm_is_destroyed(lock) ||
1927 ldlm_is_granted(lock)) {
1928 /* b=11300: the lock has already been granted */
1929 unlock_res_and_lock(lock);
1930 LDLM_DEBUG(lock, "Double grant race happened");
1935 * If we receive the completion AST before the actual enqueue returned,
1936 * then we might need to switch lock modes, resources, or extents.
1938 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1939 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1940 LDLM_DEBUG(lock, "completion AST, new lock mode");
1943 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1944 ldlm_convert_policy_to_local(req->rq_export,
1945 dlm_req->lock_desc.l_resource.lr_type,
1946 &dlm_req->lock_desc.l_policy_data,
1947 &lock->l_policy_data);
1948 LDLM_DEBUG(lock, "completion AST, new policy data");
1951 ldlm_resource_unlink_lock(lock);
1953 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1955 * BL_AST locks are not needed in LRU.
1956 * Let ldlm_cancel_lru() be fast.
1958 ldlm_lock_remove_from_lru(lock);
1959 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1960 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1963 if (lock->l_lvb_len > 0) {
1964 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1965 lock->l_lvb_data, lvb_len);
1967 unlock_res_and_lock(lock);
1972 ldlm_grant_lock(lock, &ast_list);
1973 unlock_res_and_lock(lock);
1975 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1978 * Let Enqueue to call osc_lock_upcall() and initialize
1981 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1983 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1985 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1991 lock_res_and_lock(lock);
1992 ldlm_set_failed(lock);
1993 unlock_res_and_lock(lock);
1994 wake_up(&lock->l_waitq);
1996 LDLM_LOCK_RELEASE(lock);
2000 * Callback handler for receiving incoming glimpse ASTs.
2002 * This only can happen on client side. After handling the glimpse AST
2003 * we also consider dropping the lock here if it is unused locally for a
2006 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2007 struct ldlm_namespace *ns,
2008 struct ldlm_request *dlm_req,
2009 struct ldlm_lock *lock)
2015 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2017 if (lock->l_glimpse_ast != NULL)
2018 rc = lock->l_glimpse_ast(lock, req);
2020 if (req->rq_repmsg != NULL) {
2023 req->rq_status = rc;
2027 lock_res_and_lock(lock);
2028 if (lock->l_granted_mode == LCK_PW &&
2029 !lock->l_readers && !lock->l_writers &&
2030 ktime_after(ktime_get(),
2031 ktime_add(lock->l_last_used,
2032 ktime_set(ns->ns_dirty_age_limit, 0)))) {
2033 unlock_res_and_lock(lock);
2034 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
2035 ldlm_handle_bl_callback(ns, NULL, lock);
2040 unlock_res_and_lock(lock);
2041 LDLM_LOCK_RELEASE(lock);
2045 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
2047 if (req->rq_no_reply)
2050 req->rq_status = rc;
2051 if (!req->rq_packed_final) {
2052 rc = lustre_pack_reply(req, 1, NULL, NULL);
2056 return ptlrpc_reply(req);
2059 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2060 enum ldlm_cancel_flags cancel_flags)
2062 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2066 spin_lock(&blp->blp_lock);
2067 if (blwi->blwi_lock &&
2068 ldlm_is_discard_data(blwi->blwi_lock)) {
2069 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2070 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2072 /* other blocking callbacks are added to the regular list */
2073 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2075 spin_unlock(&blp->blp_lock);
2077 wake_up(&blp->blp_waitq);
2080 * can not check blwi->blwi_flags as blwi could be already freed in
2083 if (!(cancel_flags & LCF_ASYNC))
2084 wait_for_completion(&blwi->blwi_comp);
2089 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2090 struct ldlm_namespace *ns,
2091 struct ldlm_lock_desc *ld,
2092 struct list_head *cancels, int count,
2093 struct ldlm_lock *lock,
2094 enum ldlm_cancel_flags cancel_flags)
2096 init_completion(&blwi->blwi_comp);
2097 INIT_LIST_HEAD(&blwi->blwi_head);
2099 if (memory_pressure_get())
2100 blwi->blwi_mem_pressure = 1;
2103 blwi->blwi_flags = cancel_flags;
2105 blwi->blwi_ld = *ld;
2107 list_add(&blwi->blwi_head, cancels);
2108 list_del_init(cancels);
2109 blwi->blwi_count = count;
2111 blwi->blwi_lock = lock;
2116 * Queues a list of locks \a cancels containing \a count locks
2117 * for later processing by a blocking thread. If \a count is zero,
2118 * then the lock referenced as \a lock is queued instead.
2120 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2121 * If list addition fails an error is returned and caller is supposed to
2122 * call ->l_blocking_ast itself.
2124 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2125 struct ldlm_lock_desc *ld,
2126 struct ldlm_lock *lock,
2127 struct list_head *cancels, int count,
2128 enum ldlm_cancel_flags cancel_flags)
2132 if (cancels && count == 0)
2135 if (cancel_flags & LCF_ASYNC) {
2136 struct ldlm_bl_work_item *blwi;
2138 OBD_ALLOC(blwi, sizeof(*blwi));
2141 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2143 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2146 * if it is synchronous call do minimum mem alloc, as it could
2147 * be triggered from kernel shrinker
2149 struct ldlm_bl_work_item blwi;
2151 memset(&blwi, 0, sizeof(blwi));
2152 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2153 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2158 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2159 struct ldlm_lock *lock)
2161 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2164 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2165 struct list_head *cancels, int count,
2166 enum ldlm_cancel_flags cancel_flags)
2168 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2171 int ldlm_bl_thread_wakeup(void)
2173 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2177 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2178 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2180 struct obd_device *obd = req->rq_export->exp_obd;
2188 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2190 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2192 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2194 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2197 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2199 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2201 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2204 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2207 /* We are responsible for swabbing contents of val */
2209 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2210 /* Pass it on to mdc (the "export" in this case) */
2211 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2213 sizeof(KEY_HSM_COPYTOOL_SEND),
2214 KEY_HSM_COPYTOOL_SEND,
2217 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2222 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2223 const char *msg, int rc,
2224 const struct lustre_handle *handle)
2226 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2227 "%s, NID=%s lock=%#llx: rc = %d",
2228 msg, libcfs_id2str(req->rq_peer),
2229 handle ? handle->cookie : 0, rc);
2230 if (req->rq_no_reply)
2231 CWARN("No reply was sent, maybe cause b=21636.\n");
2233 CWARN("Send reply failed, maybe cause b=21636.\n");
2236 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2237 static int ldlm_callback_handler(struct ptlrpc_request *req)
2239 struct ldlm_namespace *ns;
2240 struct ldlm_request *dlm_req;
2241 struct ldlm_lock *lock;
2247 * Requests arrive in sender's byte order. The ptlrpc service
2248 * handler has already checked and, if necessary, byte-swapped the
2249 * incoming request message body, but I am responsible for the
2253 /* do nothing for sec context finalize */
2254 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2257 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2259 if (req->rq_export == NULL) {
2260 rc = ldlm_callback_reply(req, -ENOTCONN);
2261 ldlm_callback_errmsg(req, "Operate on unconnected server",
2266 LASSERT(req->rq_export != NULL);
2267 LASSERT(req->rq_export->exp_obd != NULL);
2269 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2270 case LDLM_BL_CALLBACK:
2271 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2273 ldlm_callback_reply(req, -(int)cfs_fail_err);
2277 case LDLM_CP_CALLBACK:
2278 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2281 case LDLM_GL_CALLBACK:
2282 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2286 rc = ldlm_handle_setinfo(req);
2287 ldlm_callback_reply(req, rc);
2290 CERROR("unknown opcode %u\n",
2291 lustre_msg_get_opc(req->rq_reqmsg));
2292 ldlm_callback_reply(req, -EPROTO);
2296 ns = req->rq_export->exp_obd->obd_namespace;
2297 LASSERT(ns != NULL);
2299 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2301 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2302 if (dlm_req == NULL) {
2303 rc = ldlm_callback_reply(req, -EPROTO);
2304 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2310 * Force a known safe race, send a cancel to the server for a lock
2311 * which the server has already started a blocking callback on.
2313 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2314 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2315 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2317 CERROR("ldlm_cli_cancel: %d\n", rc);
2320 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2323 "callback on lock %#llx - lock disappeared\n",
2324 dlm_req->lock_handle[0].cookie);
2325 rc = ldlm_callback_reply(req, -EINVAL);
2326 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2327 &dlm_req->lock_handle[0]);
2331 if (ldlm_is_fail_loc(lock) &&
2332 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2333 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2335 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2336 lock_res_and_lock(lock);
2337 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2339 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2341 * If somebody cancels lock and cache is already dropped,
2342 * or lock is failed before cp_ast received on client,
2343 * we can tell the server we have no lock. Otherwise, we
2344 * should send cancel after dropping the cache.
2346 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2347 ldlm_is_failed(lock)) {
2349 "callback on lock %llx - lock disappeared",
2350 dlm_req->lock_handle[0].cookie);
2351 unlock_res_and_lock(lock);
2352 LDLM_LOCK_RELEASE(lock);
2353 rc = ldlm_callback_reply(req, -EINVAL);
2354 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2355 &dlm_req->lock_handle[0]);
2359 * BL_AST locks are not needed in LRU.
2360 * Let ldlm_cancel_lru() be fast.
2362 ldlm_lock_remove_from_lru(lock);
2363 ldlm_set_bl_ast(lock);
2365 unlock_res_and_lock(lock);
2368 * We want the ost thread to get this reply so that it can respond
2369 * to ost requests (write cache writeback) that might be triggered
2372 * But we'd also like to be able to indicate in the reply that we're
2373 * cancelling right now, because it's unused, or have an intent result
2374 * in the reply, so we might have to push the responsibility for sending
2375 * the reply down into the AST handlers, alas.
2378 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2379 case LDLM_BL_CALLBACK:
2380 CDEBUG(D_INODE, "blocking ast\n");
2381 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2382 if (!ldlm_is_cancel_on_block(lock)) {
2383 rc = ldlm_callback_reply(req, 0);
2384 if (req->rq_no_reply || rc)
2385 ldlm_callback_errmsg(req, "Normal process", rc,
2386 &dlm_req->lock_handle[0]);
2388 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2389 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2391 case LDLM_CP_CALLBACK:
2392 CDEBUG(D_INODE, "completion ast\n");
2393 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2394 ldlm_callback_reply(req, 0);
2395 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2397 case LDLM_GL_CALLBACK:
2398 CDEBUG(D_INODE, "glimpse ast\n");
2399 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2400 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2403 LBUG(); /* checked above */
2409 #ifdef HAVE_SERVER_SUPPORT
2411 * Main handler for canceld thread.
2413 * Separated into its own thread to avoid deadlocks.
2415 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2422 * Requests arrive in sender's byte order. The ptlrpc service
2423 * handler has already checked and, if necessary, byte-swapped the
2424 * incoming request message body, but I am responsible for the
2428 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2430 if (req->rq_export == NULL) {
2431 struct ldlm_request *dlm_req;
2433 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2434 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2435 libcfs_nid2str(req->rq_peer.nid),
2436 (unsigned long long)req->rq_arrival_time.tv_sec,
2437 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2439 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2440 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2441 dlm_req = req_capsule_client_get(&req->rq_pill,
2443 if (dlm_req != NULL)
2444 ldlm_lock_dump_handle(D_ERROR,
2445 &dlm_req->lock_handle[0]);
2447 ldlm_callback_reply(req, -ENOTCONN);
2451 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2452 /* XXX FIXME move this back to mds/handler.c, b=249 */
2454 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2455 CDEBUG(D_INODE, "cancel\n");
2456 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2457 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2458 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2460 rc = ldlm_handle_cancel(req);
2464 struct ldlm_request *dlm_req;
2466 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2467 CDEBUG(D_INODE, "convert\n");
2469 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2470 if (dlm_req == NULL) {
2471 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2472 rc = ldlm_callback_reply(req, -EPROTO);
2474 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2475 rc = ptlrpc_reply(req);
2480 CERROR("invalid opcode %d\n",
2481 lustre_msg_get_opc(req->rq_reqmsg));
2482 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2483 rc = ldlm_callback_reply(req, -EINVAL);
2489 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2490 struct ldlm_lock *lock)
2492 struct ldlm_request *dlm_req;
2493 struct lustre_handle lockh;
2499 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2500 if (dlm_req == NULL)
2503 ldlm_lock2handle(lock, &lockh);
2504 for (i = 0; i < dlm_req->lock_count; i++) {
2505 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2507 DEBUG_REQ(D_RPCTRACE, req,
2508 "Prio raised by lock %#llx", lockh.cookie);
2517 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2519 struct ldlm_request *dlm_req;
2526 /* no prolong in recovery */
2527 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2530 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2531 if (dlm_req == NULL)
2534 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2535 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2536 (size - offsetof(struct ldlm_request, lock_handle)) /
2537 sizeof(struct lustre_handle) < dlm_req->lock_count)
2540 for (i = 0; i < dlm_req->lock_count; i++) {
2541 struct ldlm_lock *lock;
2543 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2547 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2549 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2550 LDLM_LOCK_PUT(lock);
2559 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2560 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2561 .hpreq_check = ldlm_cancel_hpreq_check,
2565 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2569 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2571 if (req->rq_export == NULL)
2574 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2575 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2576 req->rq_ops = &ldlm_cancel_hpreq_ops;
2577 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2578 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2579 req->rq_ops = &ldlm_cancel_hpreq_ops;
2584 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2585 struct hlist_node *hnode, void *data)
2588 struct list_head *rpc_list = data;
2589 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2591 lock_res_and_lock(lock);
2593 if (!ldlm_is_granted(lock)) {
2594 unlock_res_and_lock(lock);
2598 LASSERT(lock->l_resource);
2599 if (lock->l_resource->lr_type != LDLM_IBITS &&
2600 lock->l_resource->lr_type != LDLM_PLAIN) {
2601 unlock_res_and_lock(lock);
2605 if (ldlm_is_ast_sent(lock)) {
2606 unlock_res_and_lock(lock);
2610 LASSERT(lock->l_blocking_ast);
2611 LASSERT(!lock->l_blocking_lock);
2613 ldlm_set_ast_sent(lock);
2614 if (lock->l_export && lock->l_export->exp_lock_hash) {
2616 * NB: it's safe to call cfs_hash_del() even lock isn't
2620 * In the function below, .hs_keycmp resolves to
2621 * ldlm_export_lock_keycmp()
2623 /* coverity[overrun-buffer-val] */
2624 cfs_hash_del(lock->l_export->exp_lock_hash,
2625 &lock->l_remote_handle, &lock->l_exp_hash);
2628 list_add_tail(&lock->l_rk_ast, rpc_list);
2629 LDLM_LOCK_GET(lock);
2631 unlock_res_and_lock(lock);
2635 void ldlm_revoke_export_locks(struct obd_export *exp)
2637 struct list_head rpc_list;
2641 INIT_LIST_HEAD(&rpc_list);
2642 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2643 ldlm_revoke_lock_cb, &rpc_list, 0);
2644 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2645 LDLM_WORK_REVOKE_AST);
2649 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2650 #endif /* HAVE_SERVER_SUPPORT */
2652 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2653 struct ldlm_bl_work_item **p_blwi,
2654 struct obd_export **p_exp)
2656 struct ldlm_bl_work_item *blwi = NULL;
2657 static unsigned int num_bl;
2658 static unsigned int num_stale;
2659 int num_th = atomic_read(&blp->blp_num_threads);
2661 *p_exp = obd_stale_export_get();
2663 spin_lock(&blp->blp_lock);
2664 if (*p_exp != NULL) {
2665 if (num_th == 1 || ++num_stale < num_th) {
2666 spin_unlock(&blp->blp_lock);
2672 /* process a request from the blp_list at least every blp_num_threads */
2673 if (!list_empty(&blp->blp_list) &&
2674 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2675 blwi = list_entry(blp->blp_list.next,
2676 struct ldlm_bl_work_item, blwi_entry);
2678 if (!list_empty(&blp->blp_prio_list))
2679 blwi = list_entry(blp->blp_prio_list.next,
2680 struct ldlm_bl_work_item,
2684 if (++num_bl >= num_th)
2686 list_del(&blwi->blwi_entry);
2688 spin_unlock(&blp->blp_lock);
2691 if (*p_exp != NULL && *p_blwi != NULL) {
2692 obd_stale_export_put(*p_exp);
2696 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2699 /* This only contains temporary data until the thread starts */
2700 struct ldlm_bl_thread_data {
2701 struct ldlm_bl_pool *bltd_blp;
2702 struct completion bltd_comp;
2706 static int ldlm_bl_thread_main(void *arg);
2708 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2710 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2711 struct task_struct *task;
2713 init_completion(&bltd.bltd_comp);
2715 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2716 if (bltd.bltd_num >= blp->blp_max_threads) {
2717 atomic_dec(&blp->blp_num_threads);
2721 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2723 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2724 atomic_dec(&blp->blp_num_threads);
2728 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2731 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2732 bltd.bltd_num, PTR_ERR(task));
2733 atomic_dec(&blp->blp_num_threads);
2734 return PTR_ERR(task);
2736 wait_for_completion(&bltd.bltd_comp);
2741 /* Not fatal if racy and have a few too many threads */
2742 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2743 struct ldlm_bl_work_item *blwi)
2745 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2748 if (atomic_read(&blp->blp_busy_threads) <
2749 atomic_read(&blp->blp_num_threads))
2752 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2753 blwi->blwi_mem_pressure))
2759 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2760 struct ldlm_bl_work_item *blwi)
2764 if (blwi->blwi_ns == NULL)
2765 /* added by ldlm_cleanup() */
2766 RETURN(LDLM_ITER_STOP);
2768 if (blwi->blwi_mem_pressure)
2769 memory_pressure_set();
2771 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2773 if (blwi->blwi_count) {
2776 * The special case when we cancel locks in lru
2777 * asynchronously, we pass the list of locks here.
2778 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2779 * canceled locally yet.
2781 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2784 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2787 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2790 if (blwi->blwi_mem_pressure)
2791 memory_pressure_clr();
2793 if (blwi->blwi_flags & LCF_ASYNC)
2794 OBD_FREE(blwi, sizeof(*blwi));
2796 complete(&blwi->blwi_comp);
2802 * Cancel stale locks on export. Cancel blocked locks first.
2803 * If the given export has blocked locks, the next in the list may have
2804 * them too, thus cancel not blocked locks only if the current export has
2807 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2808 struct obd_export *exp)
2814 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2816 num = ldlm_export_cancel_blocked_locks(exp);
2818 ldlm_export_cancel_locks(exp);
2820 obd_stale_export_put(exp);
2827 * Main blocking requests processing thread.
2829 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2830 * This thread in the end ends up doing actual call to ->l_blocking_ast
2833 static int ldlm_bl_thread_main(void *arg)
2836 struct ldlm_bl_pool *blp;
2837 struct ldlm_bl_thread_data *bltd = arg;
2845 rc = lu_env_init(env, LCT_DT_THREAD);
2848 rc = lu_env_add(env);
2850 GOTO(out_env_fini, rc);
2852 blp = bltd->bltd_blp;
2854 complete(&bltd->bltd_comp);
2855 /* cannot use bltd after this, it is only on caller's stack */
2858 struct l_wait_info lwi = { 0 };
2859 struct ldlm_bl_work_item *blwi = NULL;
2860 struct obd_export *exp = NULL;
2863 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2866 l_wait_event_exclusive(blp->blp_waitq,
2867 ldlm_bl_get_work(blp, &blwi,
2870 atomic_inc(&blp->blp_busy_threads);
2872 if (ldlm_bl_thread_need_create(blp, blwi))
2873 /* discard the return value, we tried */
2874 ldlm_bl_thread_start(blp, true);
2877 rc = ldlm_bl_thread_exports(blp, exp);
2879 rc = ldlm_bl_thread_blwi(blp, blwi);
2881 atomic_dec(&blp->blp_busy_threads);
2883 if (rc == LDLM_ITER_STOP)
2887 * If there are many namespaces, we will not sleep waiting for
2888 * work, and must do a cond_resched to avoid holding the CPU
2894 atomic_dec(&blp->blp_num_threads);
2895 complete(&blp->blp_comp);
2906 static int ldlm_setup(void);
2907 static int ldlm_cleanup(void);
2909 int ldlm_get_ref(void)
2914 mutex_lock(&ldlm_ref_mutex);
2915 if (++ldlm_refcount == 1) {
2920 mutex_unlock(&ldlm_ref_mutex);
2925 void ldlm_put_ref(void)
2928 mutex_lock(&ldlm_ref_mutex);
2929 if (ldlm_refcount == 1) {
2930 int rc = ldlm_cleanup();
2933 CERROR("ldlm_cleanup failed: %d\n", rc);
2939 mutex_unlock(&ldlm_ref_mutex);
2945 * Export handle<->lock hash operations.
2948 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
2950 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2954 ldlm_export_lock_key(struct hlist_node *hnode)
2956 struct ldlm_lock *lock;
2958 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2959 return &lock->l_remote_handle;
2963 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2965 struct ldlm_lock *lock;
2967 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2968 lock->l_remote_handle = *(struct lustre_handle *)key;
2972 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2974 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2978 ldlm_export_lock_object(struct hlist_node *hnode)
2980 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2984 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2986 struct ldlm_lock *lock;
2988 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2989 LDLM_LOCK_GET(lock);
2993 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2995 struct ldlm_lock *lock;
2997 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2998 LDLM_LOCK_RELEASE(lock);
3001 static struct cfs_hash_ops ldlm_export_lock_ops = {
3002 .hs_hash = ldlm_export_lock_hash,
3003 .hs_key = ldlm_export_lock_key,
3004 .hs_keycmp = ldlm_export_lock_keycmp,
3005 .hs_keycpy = ldlm_export_lock_keycpy,
3006 .hs_object = ldlm_export_lock_object,
3007 .hs_get = ldlm_export_lock_get,
3008 .hs_put = ldlm_export_lock_put,
3009 .hs_put_locked = ldlm_export_lock_put,
3012 int ldlm_init_export(struct obd_export *exp)
3018 exp->exp_lock_hash =
3019 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3020 HASH_EXP_LOCK_CUR_BITS,
3021 HASH_EXP_LOCK_MAX_BITS,
3022 HASH_EXP_LOCK_BKT_BITS, 0,
3023 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3024 &ldlm_export_lock_ops,
3025 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3026 CFS_HASH_NBLK_CHANGE);
3028 if (!exp->exp_lock_hash)
3031 rc = ldlm_init_flock_export(exp);
3037 ldlm_destroy_export(exp);
3040 EXPORT_SYMBOL(ldlm_init_export);
3042 void ldlm_destroy_export(struct obd_export *exp)
3045 cfs_hash_putref(exp->exp_lock_hash);
3046 exp->exp_lock_hash = NULL;
3048 ldlm_destroy_flock_export(exp);
3051 EXPORT_SYMBOL(ldlm_destroy_export);
3053 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3054 struct attribute *attr,
3057 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3060 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3061 struct attribute *attr,
3068 rc = kstrtoul(buffer, 10, &val);
3072 ldlm_cancel_unused_locks_before_replay = val;
3076 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3078 static struct attribute *ldlm_attrs[] = {
3079 &lustre_attr_cancel_unused_locks_before_replay.attr,
3083 static struct attribute_group ldlm_attr_group = {
3084 .attrs = ldlm_attrs,
3087 static int ldlm_setup(void)
3089 static struct ptlrpc_service_conf conf;
3090 struct ldlm_bl_pool *blp = NULL;
3091 #ifdef HAVE_SERVER_SUPPORT
3092 struct task_struct *task;
3093 #endif /* HAVE_SERVER_SUPPORT */
3099 if (ldlm_state != NULL)
3102 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3103 if (ldlm_state == NULL)
3106 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3110 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3114 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3118 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3122 rc = ldlm_debugfs_setup();
3126 memset(&conf, 0, sizeof(conf));
3127 conf = (typeof(conf)) {
3128 .psc_name = "ldlm_cbd",
3129 .psc_watchdog_factor = 2,
3131 .bc_nbufs = LDLM_CLIENT_NBUFS,
3132 .bc_buf_size = LDLM_BUFSIZE,
3133 .bc_req_max_size = LDLM_MAXREQSIZE,
3134 .bc_rep_max_size = LDLM_MAXREPSIZE,
3135 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3136 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3139 .tc_thr_name = "ldlm_cb",
3140 .tc_thr_factor = LDLM_THR_FACTOR,
3141 .tc_nthrs_init = LDLM_NTHRS_INIT,
3142 .tc_nthrs_base = LDLM_NTHRS_BASE,
3143 .tc_nthrs_max = LDLM_NTHRS_MAX,
3144 .tc_nthrs_user = ldlm_num_threads,
3145 .tc_cpu_bind = ldlm_cpu_bind,
3146 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3149 .cc_pattern = ldlm_cpts,
3150 .cc_affinity = true,
3153 .so_req_handler = ldlm_callback_handler,
3156 ldlm_state->ldlm_cb_service = \
3157 ptlrpc_register_service(&conf, ldlm_svc_kset,
3158 ldlm_svc_debugfs_dir);
3159 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3160 CERROR("failed to start service\n");
3161 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3162 ldlm_state->ldlm_cb_service = NULL;
3166 #ifdef HAVE_SERVER_SUPPORT
3167 memset(&conf, 0, sizeof(conf));
3168 conf = (typeof(conf)) {
3169 .psc_name = "ldlm_canceld",
3170 .psc_watchdog_factor = 6,
3172 .bc_nbufs = LDLM_SERVER_NBUFS,
3173 .bc_buf_size = LDLM_BUFSIZE,
3174 .bc_req_max_size = LDLM_MAXREQSIZE,
3175 .bc_rep_max_size = LDLM_MAXREPSIZE,
3176 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3177 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3181 .tc_thr_name = "ldlm_cn",
3182 .tc_thr_factor = LDLM_THR_FACTOR,
3183 .tc_nthrs_init = LDLM_NTHRS_INIT,
3184 .tc_nthrs_base = LDLM_NTHRS_BASE,
3185 .tc_nthrs_max = LDLM_NTHRS_MAX,
3186 .tc_nthrs_user = ldlm_num_threads,
3187 .tc_cpu_bind = ldlm_cpu_bind,
3188 .tc_ctx_tags = LCT_MD_THREAD | \
3193 .cc_pattern = ldlm_cpts,
3194 .cc_affinity = true,
3197 .so_req_handler = ldlm_cancel_handler,
3198 .so_hpreq_handler = ldlm_hpreq_handler,
3201 ldlm_state->ldlm_cancel_service = \
3202 ptlrpc_register_service(&conf, ldlm_svc_kset,
3203 ldlm_svc_debugfs_dir);
3204 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3205 CERROR("failed to start service\n");
3206 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3207 ldlm_state->ldlm_cancel_service = NULL;
3210 #endif /* HAVE_SERVER_SUPPORT */
3212 OBD_ALLOC(blp, sizeof(*blp));
3214 GOTO(out, rc = -ENOMEM);
3215 ldlm_state->ldlm_bl_pool = blp;
3217 spin_lock_init(&blp->blp_lock);
3218 INIT_LIST_HEAD(&blp->blp_list);
3219 INIT_LIST_HEAD(&blp->blp_prio_list);
3220 init_waitqueue_head(&blp->blp_waitq);
3221 atomic_set(&blp->blp_num_threads, 0);
3222 atomic_set(&blp->blp_busy_threads, 0);
3224 if (ldlm_num_threads == 0) {
3225 blp->blp_min_threads = LDLM_NTHRS_INIT;
3226 blp->blp_max_threads = LDLM_NTHRS_MAX;
3228 blp->blp_min_threads = blp->blp_max_threads = \
3229 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3233 for (i = 0; i < blp->blp_min_threads; i++) {
3234 rc = ldlm_bl_thread_start(blp, false);
3239 #ifdef HAVE_SERVER_SUPPORT
3240 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3243 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3247 wait_event(expired_lock_wait_queue,
3248 expired_lock_thread_state == ELT_READY);
3249 #endif /* HAVE_SERVER_SUPPORT */
3251 rc = ldlm_pools_init();
3253 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3257 rc = ldlm_reclaim_setup();
3259 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3269 static int ldlm_cleanup(void)
3273 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3274 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3275 CERROR("ldlm still has namespaces; clean these up first.\n");
3276 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3277 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3281 ldlm_reclaim_cleanup();
3284 if (ldlm_state->ldlm_bl_pool != NULL) {
3285 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3287 while (atomic_read(&blp->blp_num_threads) > 0) {
3288 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3290 init_completion(&blp->blp_comp);
3292 spin_lock(&blp->blp_lock);
3293 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3294 wake_up(&blp->blp_waitq);
3295 spin_unlock(&blp->blp_lock);
3297 wait_for_completion(&blp->blp_comp);
3300 OBD_FREE(blp, sizeof(*blp));
3303 if (ldlm_state->ldlm_cb_service != NULL)
3304 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3305 #ifdef HAVE_SERVER_SUPPORT
3306 if (ldlm_state->ldlm_cancel_service != NULL)
3307 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3311 kset_unregister(ldlm_ns_kset);
3313 kset_unregister(ldlm_svc_kset);
3315 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3316 kobject_put(ldlm_kobj);
3319 ldlm_debugfs_cleanup();
3321 #ifdef HAVE_SERVER_SUPPORT
3322 if (expired_lock_thread_state != ELT_STOPPED) {
3323 expired_lock_thread_state = ELT_TERMINATE;
3324 wake_up(&expired_lock_wait_queue);
3325 wait_event(expired_lock_wait_queue,
3326 expired_lock_thread_state == ELT_STOPPED);
3330 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3338 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3339 sizeof(struct ldlm_resource), 0,
3340 SLAB_HWCACHE_ALIGN, NULL);
3341 if (ldlm_resource_slab == NULL)
3344 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3345 sizeof(struct ldlm_lock), 0,
3346 SLAB_HWCACHE_ALIGN, NULL);
3347 if (ldlm_lock_slab == NULL)
3350 ldlm_interval_slab = kmem_cache_create("interval_node",
3351 sizeof(struct ldlm_interval),
3352 0, SLAB_HWCACHE_ALIGN, NULL);
3353 if (ldlm_interval_slab == NULL)
3356 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3357 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3358 0, SLAB_HWCACHE_ALIGN, NULL);
3359 if (ldlm_interval_tree_slab == NULL)
3362 #ifdef HAVE_SERVER_SUPPORT
3363 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3364 sizeof(struct ldlm_ibits_node),
3365 0, SLAB_HWCACHE_ALIGN, NULL);
3366 if (ldlm_inodebits_slab == NULL)
3367 goto out_interval_tree;
3369 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3370 sizeof(struct ldlm_glimpse_work),
3372 if (ldlm_glimpse_work_kmem == NULL)
3376 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3377 class_export_dump_hook = ldlm_dump_export_locks;
3380 #ifdef HAVE_SERVER_SUPPORT
3382 kmem_cache_destroy(ldlm_inodebits_slab);
3384 kmem_cache_destroy(ldlm_interval_tree_slab);
3387 kmem_cache_destroy(ldlm_interval_slab);
3389 kmem_cache_destroy(ldlm_lock_slab);
3391 kmem_cache_destroy(ldlm_resource_slab);
3396 void ldlm_exit(void)
3399 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3400 kmem_cache_destroy(ldlm_resource_slab);
3402 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3403 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3404 * so that ldlm_lock_free() get a chance to be called.
3407 kmem_cache_destroy(ldlm_lock_slab);
3408 kmem_cache_destroy(ldlm_interval_slab);
3409 kmem_cache_destroy(ldlm_interval_tree_slab);
3410 #ifdef HAVE_SERVER_SUPPORT
3411 kmem_cache_destroy(ldlm_inodebits_slab);
3412 kmem_cache_destroy(ldlm_glimpse_work_kmem);