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;
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_timeout != 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_timeout > 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_del(&lock->l_pending_chain);
336 list_add(&lock->l_pending_chain, &expired_lock_list);
340 if (!list_empty(&expired_lock_list)) {
341 if (obd_dump_on_timeout && need_dump)
342 expired_lock_dump = __LINE__;
344 wake_up(&expired_lock_wait_queue);
348 * Make sure the timer will fire again if we have any locks
351 if (!list_empty(&waiting_locks_list)) {
352 unsigned long timeout_jiffies;
354 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
356 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
357 mod_timer(&waiting_locks_timer, timeout_jiffies);
359 spin_unlock_bh(&waiting_locks_spinlock);
363 * Add lock to the list of contended locks.
365 * Indicate that we're waiting for a client to call us back cancelling a given
366 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
367 * timer to fire appropriately. (We round up to the next second, to avoid
368 * floods of timer firings during periods of high lock contention and traffic).
369 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
370 * if it has been added to the waiting list (1 is returned).
372 * Called with the namespace lock held.
374 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds)
376 unsigned long timeout_jiffies;
379 if (!list_empty(&lock->l_pending_chain))
382 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
383 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
386 timeout = ktime_get_seconds() + seconds;
387 if (likely(timeout > lock->l_callback_timeout))
388 lock->l_callback_timeout = timeout;
390 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
392 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
393 !timer_pending(&waiting_locks_timer))
394 mod_timer(&waiting_locks_timer, timeout_jiffies);
397 * if the new lock has a shorter timeout than something earlier on
398 * the list, we'll wait the longer amount of time; no big deal.
401 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
405 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
407 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
408 if (list_empty(&lock->l_exp_list)) {
409 if (!ldlm_is_granted(lock))
410 list_add_tail(&lock->l_exp_list,
411 &lock->l_export->exp_bl_list);
413 list_add(&lock->l_exp_list,
414 &lock->l_export->exp_bl_list);
416 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
419 * A blocked lock is added. Adjust the position in
420 * the stale list if the export is in the list.
421 * If export is stale and not in the list - it is being
422 * processed and will be placed on the right position
423 * on obd_stale_export_put().
425 if (!list_empty(&lock->l_export->exp_stale_list))
426 obd_stale_export_adjust(lock->l_export);
429 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
433 /* NB: must be called with hold of lock_res_and_lock() */
434 LASSERT(ldlm_is_res_locked(lock));
435 LASSERT(!ldlm_is_cancel_on_block(lock));
438 * Do not put cross-MDT lock in the waiting list, since we
439 * will not evict it due to timeout for now
441 if (lock->l_export != NULL &&
442 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
445 spin_lock_bh(&waiting_locks_spinlock);
446 if (ldlm_is_cancel(lock)) {
447 spin_unlock_bh(&waiting_locks_spinlock);
451 if (ldlm_is_destroyed(lock)) {
452 static time64_t next;
454 spin_unlock_bh(&waiting_locks_spinlock);
455 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
456 if (ktime_get_seconds() > next) {
457 next = ktime_get_seconds() + 14400;
458 libcfs_debug_dumpstack(NULL);
463 ldlm_set_waited(lock);
464 lock->l_blast_sent = ktime_get_real_seconds();
465 ret = __ldlm_add_waiting_lock(lock, timeout);
468 * grab ref on the lock if it has been added to the
473 spin_unlock_bh(&waiting_locks_spinlock);
476 ldlm_add_blocked_lock(lock);
478 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)",
479 ret == 0 ? "not re-" : "", timeout,
480 AT_OFF ? "off" : "on");
485 * Remove a lock from the pending list, likely because it had its cancellation
486 * callback arrive without incident. This adjusts the lock-timeout timer if
487 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
488 * As done by ldlm_del_waiting_lock(), the caller must release the lock
489 * reference when the lock is removed from any list (1 is returned).
491 * Called with namespace lock held.
493 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
495 struct list_head *list_next;
497 if (list_empty(&lock->l_pending_chain))
500 list_next = lock->l_pending_chain.next;
501 if (lock->l_pending_chain.prev == &waiting_locks_list) {
502 /* Removing the head of the list, adjust timer. */
503 if (list_next == &waiting_locks_list) {
504 /* No more, just cancel. */
505 del_timer(&waiting_locks_timer);
507 struct ldlm_lock *next;
509 next = list_entry(list_next, struct ldlm_lock,
511 mod_timer(&waiting_locks_timer,
512 cfs_time_seconds(next->l_callback_timeout));
515 list_del_init(&lock->l_pending_chain);
520 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
524 if (lock->l_export == NULL) {
525 /* We don't have a "waiting locks list" on clients. */
526 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
530 spin_lock_bh(&waiting_locks_spinlock);
531 ret = __ldlm_del_waiting_lock(lock);
532 ldlm_clear_waited(lock);
533 spin_unlock_bh(&waiting_locks_spinlock);
535 /* remove the lock out of export blocking list */
536 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
537 list_del_init(&lock->l_exp_list);
538 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
542 * release lock ref if it has indeed been removed
545 LDLM_LOCK_RELEASE(lock);
548 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
553 * Prolong the contended lock waiting time.
555 * Called with namespace lock held.
557 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
559 if (lock->l_export == NULL) {
560 /* We don't have a "waiting locks list" on clients. */
561 LDLM_DEBUG(lock, "client lock: no-op");
565 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
566 /* We don't have a "waiting locks list" on OSP. */
567 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
571 spin_lock_bh(&waiting_locks_spinlock);
573 if (list_empty(&lock->l_pending_chain)) {
574 spin_unlock_bh(&waiting_locks_spinlock);
575 LDLM_DEBUG(lock, "wasn't waiting");
580 * we remove/add the lock to the waiting list, so no needs to
581 * release/take a lock reference
583 __ldlm_del_waiting_lock(lock);
584 __ldlm_add_waiting_lock(lock, timeout);
585 spin_unlock_bh(&waiting_locks_spinlock);
587 LDLM_DEBUG(lock, "refreshed");
590 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
592 #else /* HAVE_SERVER_SUPPORT */
594 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
599 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
604 #endif /* !HAVE_SERVER_SUPPORT */
606 #ifdef HAVE_SERVER_SUPPORT
609 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
610 * lock cancel, and their replies). Used for lock callback timeout and AST
613 * \param[in] lock lock which is getting the blocking callback
615 * \retval timeout in seconds to wait for the client reply
617 time64_t ldlm_bl_timeout(struct ldlm_lock *lock)
622 return obd_timeout / 2;
625 * Since these are non-updating timeouts, we should be conservative.
626 * Take more than usually, 150%
627 * It would be nice to have some kind of "early reply" mechanism for
628 * lock callbacks too...
630 timeout = at_get(&lock->l_export->exp_bl_lock_at);
631 return max(timeout + (timeout >> 1), (time64_t)ldlm_enqueue_min);
633 EXPORT_SYMBOL(ldlm_bl_timeout);
636 * Perform lock cleanup if AST sending failed.
638 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
639 const char *ast_type)
641 LCONSOLE_ERROR_MSG(0x138,
642 "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
643 lock->l_export->exp_obd->obd_name,
644 obd_export_nid2str(lock->l_export), ast_type, rc);
646 if (obd_dump_on_timeout)
647 libcfs_debug_dumplog();
648 spin_lock_bh(&waiting_locks_spinlock);
649 if (__ldlm_del_waiting_lock(lock) == 0)
651 * the lock was not in any list, grab an extra ref before adding
652 * the lock to the expired list
655 lock->l_callback_timeout = 0; /* differentiate it from expired locks */
656 list_add(&lock->l_pending_chain, &expired_lock_list);
657 wake_up(&expired_lock_wait_queue);
658 spin_unlock_bh(&waiting_locks_spinlock);
662 * Perform lock cleanup if AST reply came with error.
664 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
665 struct ptlrpc_request *req, int rc,
666 const char *ast_type)
668 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
670 if (!req->rq_replied || (rc && rc != -EINVAL)) {
671 if (ldlm_is_cancel(lock)) {
673 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
674 ast_type, req, req->rq_xid,
675 libcfs_nid2str(peer.nid));
676 ldlm_lock_cancel(lock);
678 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
680 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
682 * Upon umount process the AST fails because cannot be
683 * sent. This shouldn't lead to the client eviction.
684 * -ENODEV error is returned by ptl_send_rpc() for
685 * new request in such import.
686 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
687 * if imp_invalid is set or obd_no_recov.
688 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
689 * in ptlrpc_import_delay_req() as well with -EIO code.
690 * In all such cases errors are ignored.
693 "%s AST can't be sent due to a server %s failure or umount process: rc = %d\n",
695 req->rq_import->imp_obd->obd_name, rc);
698 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
699 libcfs_nid2str(peer.nid),
700 req->rq_replied ? "returned error from" :
701 "failed to reply to",
702 ast_type, req, req->rq_xid,
703 (req->rq_repmsg != NULL) ?
704 lustre_msg_get_status(req->rq_repmsg) : 0,
706 ldlm_failed_ast(lock, rc, ast_type);
712 struct ldlm_resource *res = lock->l_resource;
715 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
716 libcfs_nid2str(peer.nid),
718 lustre_msg_get_status(req->rq_repmsg) : -1,
719 ast_type, req, req->rq_xid);
722 * update lvbo to return proper attributes.
725 ldlm_resource_getref(res);
726 ldlm_lvbo_update(res, lock, NULL, 1);
727 ldlm_resource_putref(res);
729 ldlm_lock_cancel(lock);
736 static int ldlm_cb_interpret(const struct lu_env *env,
737 struct ptlrpc_request *req, void *args, int rc)
739 struct ldlm_cb_async_args *ca = args;
740 struct ldlm_lock *lock = ca->ca_lock;
741 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
745 LASSERT(lock != NULL);
748 case LDLM_GL_CALLBACK:
750 * Update the LVB from disk if the AST failed
751 * (this is a legal race)
753 * - Glimpse callback of local lock just returns
754 * -ELDLM_NO_LOCK_DATA.
755 * - Glimpse callback of remote lock might return
756 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
758 if (unlikely(arg->gl_interpret_reply)) {
759 rc = arg->gl_interpret_reply(NULL, req, args, rc);
760 } else if (rc == -ELDLM_NO_LOCK_DATA) {
762 "lost race - client has a lock but no inode");
763 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
764 } else if (rc != 0) {
765 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
767 rc = ldlm_lvbo_update(lock->l_resource,
771 case LDLM_BL_CALLBACK:
773 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
775 case LDLM_CP_CALLBACK:
777 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
780 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
785 /* release extra reference taken in ldlm_ast_fini() */
786 LDLM_LOCK_RELEASE(lock);
789 atomic_inc(&arg->restart);
794 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
796 struct ldlm_cb_async_args *ca = data;
797 struct ldlm_lock *lock = ca->ca_lock;
799 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
802 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
803 struct ldlm_cb_set_arg *arg,
804 struct ldlm_lock *lock,
811 if (unlikely(instant_cancel)) {
812 rc = ptl_send_rpc(req, 1);
813 ptlrpc_req_finished(req);
815 atomic_inc(&arg->restart);
818 ptlrpc_set_add_req(arg->set, req);
825 * Check if there are requests in the export request list which prevent
826 * the lock canceling and make these requests high priority ones.
828 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
830 struct ptlrpc_request *req;
834 if (lock->l_export == NULL) {
835 LDLM_DEBUG(lock, "client lock: no-op");
839 spin_lock(&lock->l_export->exp_rpc_lock);
840 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
843 * Do not process requests that were not yet added to there
844 * incoming queue or were already removed from there for
845 * processing. We evaluate ptlrpc_nrs_req_can_move() without
846 * holding svcpt->scp_req_lock, and then redo the check with
847 * the lock held once we need to obtain a reliable result.
849 if (ptlrpc_nrs_req_can_move(req) &&
850 req->rq_ops->hpreq_lock_match &&
851 req->rq_ops->hpreq_lock_match(req, lock))
852 ptlrpc_nrs_req_hp_move(req);
854 spin_unlock(&lock->l_export->exp_rpc_lock);
859 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
860 * enqueued server lock conflicts with given one.
862 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
863 * to wait for client response.
865 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
866 struct ldlm_lock_desc *desc,
867 void *data, int flag)
869 struct ldlm_cb_async_args *ca;
870 struct ldlm_cb_set_arg *arg = data;
871 struct ldlm_request *body;
872 struct ptlrpc_request *req;
873 int instant_cancel = 0;
878 if (flag == LDLM_CB_CANCELING)
879 /* Don't need to do anything here. */
882 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
883 LDLM_DEBUG(lock, "dropping BL AST");
888 LASSERT(data != NULL);
889 if (lock->l_export->exp_obd->obd_recovering != 0)
890 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
892 ldlm_lock_reorder_req(lock);
894 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
895 &RQF_LDLM_BL_CALLBACK,
896 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
900 ca = ptlrpc_req_async_args(ca, req);
901 ca->ca_set_arg = arg;
904 req->rq_interpret_reply = ldlm_cb_interpret;
906 lock_res_and_lock(lock);
907 if (ldlm_is_destroyed(lock)) {
908 /* What's the point? */
909 unlock_res_and_lock(lock);
910 ptlrpc_req_finished(req);
914 if (!ldlm_is_granted(lock)) {
916 * this blocking AST will be communicated as part of the
917 * completion AST instead
919 ldlm_add_blocked_lock(lock);
920 ldlm_set_waited(lock);
921 unlock_res_and_lock(lock);
923 ptlrpc_req_finished(req);
924 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
928 if (ldlm_is_cancel_on_block(lock))
931 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
932 body->lock_handle[0] = lock->l_remote_handle;
933 body->lock_desc = *desc;
934 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
936 LDLM_DEBUG(lock, "server preparing blocking AST");
938 ptlrpc_request_set_replen(req);
939 ldlm_set_cbpending(lock);
940 if (instant_cancel) {
941 unlock_res_and_lock(lock);
942 ldlm_lock_cancel(lock);
944 req->rq_no_resend = 1;
946 LASSERT(ldlm_is_granted(lock));
947 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
948 unlock_res_and_lock(lock);
950 /* Do not resend after lock callback timeout */
951 req->rq_delay_limit = ldlm_bl_timeout(lock);
952 req->rq_resend_cb = ldlm_update_resend;
955 req->rq_send_state = LUSTRE_IMP_FULL;
956 /* ptlrpc_request_alloc_pack already set timeout */
958 req->rq_timeout = ldlm_get_rq_timeout();
960 if (lock->l_export && lock->l_export->exp_nid_stats &&
961 lock->l_export->exp_nid_stats->nid_ldlm_stats)
962 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
963 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
965 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
971 * ->l_completion_ast callback for a remote lock in server namespace.
973 * Sends AST to the client notifying it of lock granting. If initial
974 * lock response was not sent yet, instead of sending another RPC, just
975 * mark the lock as granted and client will understand
977 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
979 struct ldlm_cb_set_arg *arg = data;
980 struct ldlm_request *body;
981 struct ptlrpc_request *req;
982 struct ldlm_cb_async_args *ca;
983 int instant_cancel = 0;
989 LASSERT(lock != NULL);
990 LASSERT(data != NULL);
992 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
993 LDLM_DEBUG(lock, "dropping CP AST");
997 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
998 &RQF_LDLM_CP_CALLBACK);
1002 /* server namespace, doesn't need lock */
1003 lvb_len = ldlm_lvbo_size(lock);
1005 * LU-3124 & LU-2187: to not return layout in completion AST because
1006 * it may deadlock for LU-2187, or client may not have enough space
1007 * for large layout. The layout will be returned to client with an
1008 * extra RPC to fetch xattr.lov
1010 if (ldlm_has_layout(lock))
1013 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1014 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1016 ptlrpc_request_free(req);
1020 ca = ptlrpc_req_async_args(ca, req);
1021 ca->ca_set_arg = arg;
1024 req->rq_interpret_reply = ldlm_cb_interpret;
1025 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1027 body->lock_handle[0] = lock->l_remote_handle;
1028 body->lock_flags = ldlm_flags_to_wire(flags);
1029 ldlm_lock2desc(lock, &body->lock_desc);
1031 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1032 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1035 * We still need to send the RPC to wake up the blocked
1036 * enqueue thread on the client.
1038 * Consider old client, there is no better way to notify
1039 * the failure, just zero-sized the LVB, then the client
1040 * will fail out as "-EPROTO".
1042 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1046 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1051 LDLM_DEBUG(lock, "server preparing completion AST");
1053 ptlrpc_request_set_replen(req);
1055 req->rq_send_state = LUSTRE_IMP_FULL;
1056 /* ptlrpc_request_pack already set timeout */
1058 req->rq_timeout = ldlm_get_rq_timeout();
1060 /* We only send real blocking ASTs after the lock is granted */
1061 lock_res_and_lock(lock);
1062 if (ldlm_is_ast_sent(lock)) {
1063 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1064 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1065 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1069 * We might get here prior to ldlm_handle_enqueue setting
1070 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1071 * into waiting list, but this is safe and similar code in
1072 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1073 * that would not only cancel the lock, but will also remove
1074 * it from waiting list
1076 if (ldlm_is_cancel_on_block(lock)) {
1077 unlock_res_and_lock(lock);
1078 ldlm_lock_cancel(lock);
1081 req->rq_no_resend = 1;
1083 lock_res_and_lock(lock);
1085 /* start the lock-timeout clock */
1086 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1087 /* Do not resend after lock callback timeout */
1088 req->rq_delay_limit = ldlm_bl_timeout(lock);
1089 req->rq_resend_cb = ldlm_update_resend;
1092 unlock_res_and_lock(lock);
1094 if (lock->l_export && lock->l_export->exp_nid_stats &&
1095 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1096 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1097 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1099 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1101 RETURN(lvb_len < 0 ? lvb_len : rc);
1105 * Server side ->l_glimpse_ast handler for client locks.
1107 * Sends glimpse AST to the client and waits for reply. Then updates
1108 * lvbo with the result.
1110 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1112 struct ldlm_cb_set_arg *arg = data;
1113 struct ldlm_request *body;
1114 struct ptlrpc_request *req;
1115 struct ldlm_cb_async_args *ca;
1117 struct req_format *req_fmt;
1121 LASSERT(lock != NULL);
1123 if (arg->gl_desc != NULL)
1124 /* There is a glimpse descriptor to pack */
1125 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1127 req_fmt = &RQF_LDLM_GL_CALLBACK;
1129 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1130 req_fmt, LUSTRE_DLM_VERSION,
1136 if (arg->gl_desc != NULL) {
1137 /* copy the GL descriptor */
1138 union ldlm_gl_desc *desc;
1140 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1141 *desc = *arg->gl_desc;
1144 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1145 body->lock_handle[0] = lock->l_remote_handle;
1146 ldlm_lock2desc(lock, &body->lock_desc);
1148 ca = ptlrpc_req_async_args(ca, req);
1149 ca->ca_set_arg = arg;
1152 /* server namespace, doesn't need lock */
1153 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1154 ldlm_lvbo_size(lock));
1155 ptlrpc_request_set_replen(req);
1157 req->rq_send_state = LUSTRE_IMP_FULL;
1158 /* ptlrpc_request_alloc_pack already set timeout */
1160 req->rq_timeout = ldlm_get_rq_timeout();
1162 req->rq_interpret_reply = ldlm_cb_interpret;
1164 if (lock->l_export && lock->l_export->exp_nid_stats) {
1165 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1167 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1168 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1171 rc = ldlm_ast_fini(req, arg, lock, 0);
1175 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1177 int ldlm_glimpse_locks(struct ldlm_resource *res,
1178 struct list_head *gl_work_list)
1184 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1186 if (rc == -ERESTART)
1187 ldlm_reprocess_all(res, NULL);
1191 EXPORT_SYMBOL(ldlm_glimpse_locks);
1193 /* return LDLM lock associated with a lock callback request */
1194 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1196 struct ldlm_cb_async_args *ca;
1197 struct ldlm_lock *lock;
1201 ca = ptlrpc_req_async_args(ca, req);
1204 RETURN(ERR_PTR(-EFAULT));
1208 EXPORT_SYMBOL(ldlm_request_lock);
1211 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1212 * service threads to carry out client lock enqueueing requests.
1214 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1215 struct ptlrpc_request *req,
1216 const struct ldlm_request *dlm_req,
1217 const struct ldlm_callback_suite *cbs)
1219 struct ldlm_reply *dlm_rep;
1221 enum ldlm_error err = ELDLM_OK;
1222 struct ldlm_lock *lock = NULL;
1223 void *cookie = NULL;
1225 struct ldlm_resource *res = NULL;
1226 const struct lu_env *env = req->rq_svc_thread->t_env;
1230 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1232 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1233 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1235 LASSERT(req->rq_export);
1237 /* for intent enqueue the stat will be updated inside intent policy */
1238 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1239 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1240 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1242 if (req->rq_export && req->rq_export->exp_nid_stats &&
1243 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1244 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1245 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1247 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1248 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1249 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1250 dlm_req->lock_desc.l_resource.lr_type);
1251 GOTO(out, rc = -EFAULT);
1254 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1255 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1256 dlm_req->lock_desc.l_req_mode &
1257 (dlm_req->lock_desc.l_req_mode-1))) {
1258 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1259 dlm_req->lock_desc.l_req_mode);
1260 GOTO(out, rc = -EFAULT);
1263 if (unlikely((flags & LDLM_FL_REPLAY) ||
1264 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1265 /* Find an existing lock in the per-export lock hash */
1267 * In the function below, .hs_keycmp resolves to
1268 * ldlm_export_lock_keycmp()
1270 /* coverity[overrun-buffer-val] */
1271 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1272 (void *)&dlm_req->lock_handle[0]);
1274 DEBUG_REQ(D_DLMTRACE, req,
1275 "found existing lock cookie %#llx",
1276 lock->l_handle.h_cookie);
1277 flags |= LDLM_FL_RESENT;
1278 GOTO(existing_lock, rc = 0);
1281 if (ldlm_reclaim_full()) {
1282 DEBUG_REQ(D_DLMTRACE, req,
1283 "Too many granted locks, reject current enqueue request and let the client retry later");
1284 GOTO(out, rc = -EINPROGRESS);
1288 /* The lock's callback data might be set in the policy function */
1289 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1290 dlm_req->lock_desc.l_resource.lr_type,
1291 dlm_req->lock_desc.l_req_mode,
1292 cbs, NULL, 0, LVB_T_NONE);
1299 lock->l_remote_handle = dlm_req->lock_handle[0];
1300 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1303 * Initialize resource lvb but not for a lock being replayed since
1304 * Client already got lvb sent in this case.
1305 * This must occur early since some policy methods assume resource
1306 * lvb is available (lr_lvb_data != NULL).
1308 res = lock->l_resource;
1309 if (!(flags & LDLM_FL_REPLAY)) {
1310 /* non-replayed lock, delayed lvb init may need to be done */
1311 rc = ldlm_lvbo_init(res);
1313 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1318 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1320 * Don't enqueue a lock onto the export if it is been disonnected
1321 * due to eviction (b=3822) or server umount (b=24324).
1322 * Cancel it now instead.
1324 if (req->rq_export->exp_disconnected) {
1325 LDLM_ERROR(lock, "lock on disconnected export %p",
1327 GOTO(out, rc = -ENOTCONN);
1330 lock->l_export = class_export_lock_get(req->rq_export, lock);
1331 if (lock->l_export->exp_lock_hash)
1332 cfs_hash_add(lock->l_export->exp_lock_hash,
1333 &lock->l_remote_handle,
1337 * Inherit the enqueue flags before the operation, because we do not
1338 * keep the res lock on return and next operations (BL AST) may proceed
1341 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1342 LDLM_FL_INHERIT_MASK);
1344 ldlm_convert_policy_to_local(req->rq_export,
1345 dlm_req->lock_desc.l_resource.lr_type,
1346 &dlm_req->lock_desc.l_policy_data,
1347 &lock->l_policy_data);
1348 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1349 lock->l_req_extent = lock->l_policy_data.l_extent;
1352 if (flags & LDLM_FL_HAS_INTENT) {
1354 * In this case, the reply buffer is allocated deep in
1355 * local_lock_enqueue by the policy function.
1360 * based on the assumption that lvb size never changes during
1361 * resource life time otherwise it need resource->lr_lock's
1364 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1365 RCL_SERVER, ldlm_lvbo_size(lock));
1367 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1368 GOTO(out, rc = -ENOMEM);
1370 rc = req_capsule_server_pack(&req->rq_pill);
1375 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1382 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1384 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1385 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1387 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1388 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1391 * We never send a blocking AST until the lock is granted, but
1392 * we can tell it right now
1394 lock_res_and_lock(lock);
1397 * Now take into account flags to be inherited from original lock
1398 * request both in reply to client and in our own lock flags.
1400 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1401 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1404 * Don't move a pending lock onto the export if it has already been
1405 * disconnected due to eviction (b=5683) or server umount (b=24324).
1406 * Cancel it now instead.
1408 if (unlikely(req->rq_export->exp_disconnected ||
1409 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1410 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1412 } else if (ldlm_is_ast_sent(lock)) {
1413 /* fill lock desc for possible lock convert */
1414 if (lock->l_blocking_lock &&
1415 lock->l_resource->lr_type == LDLM_IBITS) {
1416 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1417 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1420 "save blocking bits %llx in granted lock",
1421 bl_lock->l_policy_data.l_inodebits.bits);
1423 * If lock is blocked then save blocking ibits
1424 * in returned lock policy for the possible lock
1425 * convert on a client.
1427 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1428 bl_lock->l_policy_data.l_inodebits.bits;
1430 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1431 if (ldlm_is_granted(lock)) {
1433 * Only cancel lock if it was granted, because it would
1434 * be destroyed immediately and would never be granted
1435 * in the future, causing timeouts on client. Not
1436 * granted lock will be cancelled immediately after
1437 * sending completion AST.
1439 if (ldlm_is_cancel_on_block(lock)) {
1440 unlock_res_and_lock(lock);
1441 ldlm_lock_cancel(lock);
1442 lock_res_and_lock(lock);
1444 ldlm_add_waiting_lock(lock,
1445 ldlm_bl_timeout(lock));
1449 unlock_res_and_lock(lock);
1453 req->rq_status = rc ?: err; /* return either error - b=11190 */
1454 if (!req->rq_packed_final) {
1455 err = lustre_pack_reply(req, 1, NULL, NULL);
1461 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1462 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1466 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1470 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1472 ldlm_lvbo_size(lock) > 0) {
1477 buf = req_capsule_server_get(&req->rq_pill,
1479 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1480 buflen = req_capsule_get_size(&req->rq_pill,
1481 &RMF_DLM_LVB, RCL_SERVER);
1483 * non-replayed lock, delayed lvb init may
1484 * need to be occur now
1486 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1489 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1491 req_capsule_shrink(&req->rq_pill,
1494 } else if (rc2 == -ERANGE) {
1495 rc2 = req_capsule_server_grow(
1497 &RMF_DLM_LVB, buflen);
1502 * if we can't grow the buffer,
1503 * it's ok to return empty lvb
1514 } else if (flags & LDLM_FL_REPLAY) {
1515 /* no LVB resend upon replay */
1517 req_capsule_shrink(&req->rq_pill,
1527 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1528 if (lock->l_export) {
1529 ldlm_lock_cancel(lock);
1531 lock_res_and_lock(lock);
1532 ldlm_resource_unlink_lock(lock);
1533 ldlm_lock_destroy_nolock(lock);
1534 unlock_res_and_lock(lock);
1539 if (!err && !ldlm_is_cbpending(lock) &&
1540 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1541 ldlm_reprocess_all(lock->l_resource, lock);
1543 LDLM_LOCK_RELEASE(lock);
1546 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1553 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1554 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1556 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1558 if (lock->l_blocking_lock) {
1559 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1560 lock->l_blocking_lock = NULL;
1564 /* A lock can be converted to new ibits or mode and should be considered
1565 * as new lock. Clear all states related to a previous blocking AST
1566 * processing so new conflicts will cause new blocking ASTs.
1568 * This is used during lock convert below and lock downgrade to COS mode in
1569 * ldlm_lock_mode_downgrade().
1571 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1573 ldlm_clear_ast_sent(lock);
1574 lock->l_bl_ast_run = 0;
1575 ldlm_clear_blocking_lock(lock);
1579 * Main LDLM entry point for server code to process lock conversion requests.
1581 int ldlm_handle_convert0(struct ptlrpc_request *req,
1582 const struct ldlm_request *dlm_req)
1584 struct obd_export *exp = req->rq_export;
1585 struct ldlm_reply *dlm_rep;
1586 struct ldlm_lock *lock;
1591 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1592 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1593 LDLM_CONVERT - LDLM_FIRST_OPC);
1595 rc = req_capsule_server_pack(&req->rq_pill);
1599 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1600 dlm_rep->lock_flags = dlm_req->lock_flags;
1602 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1607 bits = lock->l_policy_data.l_inodebits.bits;
1608 new = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1609 LDLM_DEBUG(lock, "server-side convert handler START");
1611 if (ldlm_is_cancel(lock)) {
1612 LDLM_ERROR(lock, "convert on canceled lock!");
1613 rc = ELDLM_NO_LOCK_DATA;
1614 } else if (dlm_req->lock_desc.l_req_mode !=
1615 lock->l_granted_mode) {
1616 LDLM_ERROR(lock, "lock mode differs!");
1617 rc = ELDLM_NO_LOCK_DATA;
1618 } else if (bits == new) {
1620 * This can be valid situation if CONVERT RPCs are
1621 * re-ordered. Just finish silently
1623 LDLM_DEBUG(lock, "lock is converted already!");
1626 lock_res_and_lock(lock);
1627 if (ldlm_is_waited(lock))
1628 ldlm_del_waiting_lock(lock);
1630 ldlm_clear_cbpending(lock);
1631 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1632 ldlm_inodebits_drop(lock, bits & ~new);
1634 ldlm_clear_blocking_data(lock);
1635 unlock_res_and_lock(lock);
1637 ldlm_reprocess_all(lock->l_resource, NULL);
1641 if (rc == ELDLM_OK) {
1642 dlm_rep->lock_handle = lock->l_remote_handle;
1643 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1644 &dlm_rep->lock_desc.l_policy_data);
1647 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d",
1649 LDLM_LOCK_PUT(lock);
1651 rc = ELDLM_NO_LOCK_DATA;
1652 LDLM_DEBUG_NOLOCK("server-side convert handler END, rc = %d",
1656 req->rq_status = rc;
1662 * Cancel all the locks whose handles are packed into ldlm_request
1664 * Called by server code expecting such combined cancel activity
1667 int ldlm_request_cancel(struct ptlrpc_request *req,
1668 const struct ldlm_request *dlm_req,
1669 int first, enum lustre_at_flags flags)
1671 struct ldlm_resource *res, *pres = NULL;
1672 struct ldlm_lock *lock;
1673 int i, count, done = 0;
1678 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1679 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1680 (size - offsetof(struct ldlm_request, lock_handle)) /
1681 sizeof(struct lustre_handle) < dlm_req->lock_count)
1684 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1688 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1692 * There is no lock on the server at the replay time,
1693 * skip lock cancelling to make replay tests to pass.
1695 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1698 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1701 for (i = first; i < count; i++) {
1702 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1704 /* below message checked in replay-single.sh test_36 */
1705 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1706 dlm_req->lock_handle[i].cookie);
1710 res = lock->l_resource;
1714 * This code is an optimization to only attempt lock
1715 * granting on the resource (that could be CPU-expensive)
1716 * after we are done cancelling lock in that resource.
1720 ldlm_reprocess_all(pres, NULL);
1721 LDLM_RESOURCE_DELREF(pres);
1722 ldlm_resource_putref(pres);
1725 ldlm_resource_getref(res);
1726 LDLM_RESOURCE_ADDREF(res);
1728 if (!ldlm_is_discard_data(lock))
1729 ldlm_lvbo_update(res, lock,
1735 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1736 lock->l_blast_sent != 0) {
1737 time64_t delay = ktime_get_real_seconds() -
1740 "server cancels blocked lock after %llds",
1742 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1744 ldlm_lock_cancel(lock);
1745 LDLM_LOCK_PUT(lock);
1748 ldlm_reprocess_all(pres, NULL);
1749 LDLM_RESOURCE_DELREF(pres);
1750 ldlm_resource_putref(pres);
1752 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1755 EXPORT_SYMBOL(ldlm_request_cancel);
1758 * Main LDLM entry point for server code to cancel locks.
1760 * Typically gets called from service handler on LDLM_CANCEL opc.
1762 int ldlm_handle_cancel(struct ptlrpc_request *req)
1764 struct ldlm_request *dlm_req;
1769 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1770 if (dlm_req == NULL) {
1771 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1775 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1776 offsetof(struct ldlm_request, lock_handle[1]))
1779 if (req->rq_export && req->rq_export->exp_nid_stats &&
1780 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1781 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1782 LDLM_CANCEL - LDLM_FIRST_OPC);
1784 rc = req_capsule_server_pack(&req->rq_pill);
1788 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1789 req->rq_status = LUSTRE_ESTALE;
1791 RETURN(ptlrpc_reply(req));
1793 #endif /* HAVE_SERVER_SUPPORT */
1796 * Callback handler for receiving incoming blocking ASTs.
1798 * This can only happen on client side.
1800 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1801 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1807 LDLM_DEBUG(lock, "client blocking AST callback handler");
1809 lock_res_and_lock(lock);
1811 /* set bits to cancel for this lock for possible lock convert */
1812 if (ns_is_client(ns) && (lock->l_resource->lr_type == LDLM_IBITS)) {
1814 * Lock description contains policy of blocking lock,
1815 * and its cancel_bits is used to pass conflicting bits.
1816 * NOTE: ld can be NULL or can be not NULL but zeroed if
1817 * passed from ldlm_bl_thread_blwi(), check below used bits
1818 * in ld to make sure it is valid description.
1820 * If server may replace lock resource keeping the same cookie,
1821 * never use cancel bits from different resource, full cancel
1824 if (ld && ld->l_policy_data.l_inodebits.bits &&
1825 ldlm_res_eq(&ld->l_resource.lr_name,
1826 &lock->l_resource->lr_name))
1827 lock->l_policy_data.l_inodebits.cancel_bits =
1828 ld->l_policy_data.l_inodebits.cancel_bits;
1830 * if there is no valid ld and lock is cbpending already
1831 * then cancel_bits should be kept, otherwise it is zeroed.
1833 else if (!ldlm_is_cbpending(lock))
1834 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1836 ldlm_set_cbpending(lock);
1838 do_ast = (!lock->l_readers && !lock->l_writers);
1839 unlock_res_and_lock(lock);
1843 "Lock %p already unused, calling callback (%p)\n",
1844 lock, lock->l_blocking_ast);
1845 if (lock->l_blocking_ast != NULL)
1846 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1850 "Lock %p is referenced, will be cancelled later\n",
1854 LDLM_DEBUG(lock, "client blocking callback handler END");
1855 LDLM_LOCK_RELEASE(lock);
1860 * Callback handler for receiving incoming completion ASTs.
1862 * This only can happen on client side.
1864 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1865 struct ldlm_namespace *ns,
1866 struct ldlm_request *dlm_req,
1867 struct ldlm_lock *lock)
1869 struct list_head ast_list;
1875 LDLM_DEBUG(lock, "client completion callback handler START");
1877 INIT_LIST_HEAD(&ast_list);
1878 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1879 long to = cfs_time_seconds(1);
1882 set_current_state(TASK_INTERRUPTIBLE);
1883 schedule_timeout(to);
1884 if (ldlm_is_granted(lock) ||
1885 ldlm_is_destroyed(lock))
1890 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1892 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1893 GOTO(out, rc = lvb_len);
1894 } else if (lvb_len > 0) {
1895 if (lock->l_lvb_len > 0) {
1896 /* for extent lock, lvb contains ost_lvb{}. */
1897 LASSERT(lock->l_lvb_data != NULL);
1899 if (unlikely(lock->l_lvb_len < lvb_len)) {
1901 "Replied LVB is larger than expectation, expected = %d, replied = %d",
1902 lock->l_lvb_len, lvb_len);
1903 GOTO(out, rc = -EINVAL);
1908 lock_res_and_lock(lock);
1910 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1911 &lock->l_resource->lr_name)) {
1912 ldlm_resource_unlink_lock(lock);
1913 unlock_res_and_lock(lock);
1914 rc = ldlm_lock_change_resource(ns, lock,
1915 &dlm_req->lock_desc.l_resource.lr_name);
1917 LDLM_ERROR(lock, "Failed to allocate resource");
1920 LDLM_DEBUG(lock, "completion AST, new resource");
1921 lock_res_and_lock(lock);
1924 if (ldlm_is_destroyed(lock) ||
1925 ldlm_is_granted(lock)) {
1926 /* b=11300: the lock has already been granted */
1927 unlock_res_and_lock(lock);
1928 LDLM_DEBUG(lock, "Double grant race happened");
1933 * If we receive the completion AST before the actual enqueue returned,
1934 * then we might need to switch lock modes, resources, or extents.
1936 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1937 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1938 LDLM_DEBUG(lock, "completion AST, new lock mode");
1941 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1942 ldlm_convert_policy_to_local(req->rq_export,
1943 dlm_req->lock_desc.l_resource.lr_type,
1944 &dlm_req->lock_desc.l_policy_data,
1945 &lock->l_policy_data);
1946 LDLM_DEBUG(lock, "completion AST, new policy data");
1949 ldlm_resource_unlink_lock(lock);
1951 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1953 * BL_AST locks are not needed in LRU.
1954 * Let ldlm_cancel_lru() be fast.
1956 ldlm_lock_remove_from_lru(lock);
1957 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1958 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1961 if (lock->l_lvb_len > 0) {
1962 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1963 lock->l_lvb_data, lvb_len);
1965 unlock_res_and_lock(lock);
1970 ldlm_grant_lock(lock, &ast_list);
1971 unlock_res_and_lock(lock);
1973 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1976 * Let Enqueue to call osc_lock_upcall() and initialize
1979 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1981 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1983 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1989 lock_res_and_lock(lock);
1990 ldlm_set_failed(lock);
1991 unlock_res_and_lock(lock);
1992 wake_up(&lock->l_waitq);
1994 LDLM_LOCK_RELEASE(lock);
1998 * Callback handler for receiving incoming glimpse ASTs.
2000 * This only can happen on client side. After handling the glimpse AST
2001 * we also consider dropping the lock here if it is unused locally for a
2004 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2005 struct ldlm_namespace *ns,
2006 struct ldlm_request *dlm_req,
2007 struct ldlm_lock *lock)
2013 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2015 if (lock->l_glimpse_ast != NULL)
2016 rc = lock->l_glimpse_ast(lock, req);
2018 if (req->rq_repmsg != NULL) {
2021 req->rq_status = rc;
2025 lock_res_and_lock(lock);
2026 if (lock->l_granted_mode == LCK_PW &&
2027 !lock->l_readers && !lock->l_writers &&
2028 ktime_after(ktime_get(),
2029 ktime_add(lock->l_last_used,
2030 ktime_set(ns->ns_dirty_age_limit, 0)))) {
2031 unlock_res_and_lock(lock);
2032 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
2033 ldlm_handle_bl_callback(ns, NULL, lock);
2038 unlock_res_and_lock(lock);
2039 LDLM_LOCK_RELEASE(lock);
2043 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
2045 if (req->rq_no_reply)
2048 req->rq_status = rc;
2049 if (!req->rq_packed_final) {
2050 rc = lustre_pack_reply(req, 1, NULL, NULL);
2054 return ptlrpc_reply(req);
2057 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2058 enum ldlm_cancel_flags cancel_flags)
2060 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2064 spin_lock(&blp->blp_lock);
2065 if (blwi->blwi_lock &&
2066 ldlm_is_discard_data(blwi->blwi_lock)) {
2067 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2068 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2070 /* other blocking callbacks are added to the regular list */
2071 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2073 spin_unlock(&blp->blp_lock);
2075 wake_up(&blp->blp_waitq);
2078 * can not check blwi->blwi_flags as blwi could be already freed in
2081 if (!(cancel_flags & LCF_ASYNC))
2082 wait_for_completion(&blwi->blwi_comp);
2087 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2088 struct ldlm_namespace *ns,
2089 struct ldlm_lock_desc *ld,
2090 struct list_head *cancels, int count,
2091 struct ldlm_lock *lock,
2092 enum ldlm_cancel_flags cancel_flags)
2094 init_completion(&blwi->blwi_comp);
2095 INIT_LIST_HEAD(&blwi->blwi_head);
2097 if (memory_pressure_get())
2098 blwi->blwi_mem_pressure = 1;
2101 blwi->blwi_flags = cancel_flags;
2103 blwi->blwi_ld = *ld;
2105 list_add(&blwi->blwi_head, cancels);
2106 list_del_init(cancels);
2107 blwi->blwi_count = count;
2109 blwi->blwi_lock = lock;
2114 * Queues a list of locks \a cancels containing \a count locks
2115 * for later processing by a blocking thread. If \a count is zero,
2116 * then the lock referenced as \a lock is queued instead.
2118 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2119 * If list addition fails an error is returned and caller is supposed to
2120 * call ->l_blocking_ast itself.
2122 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2123 struct ldlm_lock_desc *ld,
2124 struct ldlm_lock *lock,
2125 struct list_head *cancels, int count,
2126 enum ldlm_cancel_flags cancel_flags)
2130 if (cancels && count == 0)
2133 if (cancel_flags & LCF_ASYNC) {
2134 struct ldlm_bl_work_item *blwi;
2136 OBD_ALLOC(blwi, sizeof(*blwi));
2139 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2141 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2144 * if it is synchronous call do minimum mem alloc, as it could
2145 * be triggered from kernel shrinker
2147 struct ldlm_bl_work_item blwi;
2149 memset(&blwi, 0, sizeof(blwi));
2150 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2151 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2156 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2157 struct ldlm_lock *lock)
2159 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2162 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2163 struct list_head *cancels, int count,
2164 enum ldlm_cancel_flags cancel_flags)
2166 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2169 int ldlm_bl_thread_wakeup(void)
2171 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2175 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2176 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2178 struct obd_device *obd = req->rq_export->exp_obd;
2186 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2188 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2190 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2192 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2195 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2197 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2199 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2202 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2205 /* We are responsible for swabbing contents of val */
2207 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2208 /* Pass it on to mdc (the "export" in this case) */
2209 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2211 sizeof(KEY_HSM_COPYTOOL_SEND),
2212 KEY_HSM_COPYTOOL_SEND,
2215 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2220 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2221 const char *msg, int rc,
2222 const struct lustre_handle *handle)
2224 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2225 "%s, NID=%s lock=%#llx: rc = %d",
2226 msg, libcfs_id2str(req->rq_peer),
2227 handle ? handle->cookie : 0, rc);
2228 if (req->rq_no_reply)
2229 CWARN("No reply was sent, maybe cause b=21636.\n");
2231 CWARN("Send reply failed, maybe cause b=21636.\n");
2234 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2235 static int ldlm_callback_handler(struct ptlrpc_request *req)
2237 struct ldlm_namespace *ns;
2238 struct ldlm_request *dlm_req;
2239 struct ldlm_lock *lock;
2245 * Requests arrive in sender's byte order. The ptlrpc service
2246 * handler has already checked and, if necessary, byte-swapped the
2247 * incoming request message body, but I am responsible for the
2251 /* do nothing for sec context finalize */
2252 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2255 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2257 if (req->rq_export == NULL) {
2258 rc = ldlm_callback_reply(req, -ENOTCONN);
2259 ldlm_callback_errmsg(req, "Operate on unconnected server",
2264 LASSERT(req->rq_export != NULL);
2265 LASSERT(req->rq_export->exp_obd != NULL);
2267 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2268 case LDLM_BL_CALLBACK:
2269 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2271 ldlm_callback_reply(req, -(int)cfs_fail_err);
2275 case LDLM_CP_CALLBACK:
2276 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2279 case LDLM_GL_CALLBACK:
2280 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2284 rc = ldlm_handle_setinfo(req);
2285 ldlm_callback_reply(req, rc);
2288 CERROR("unknown opcode %u\n",
2289 lustre_msg_get_opc(req->rq_reqmsg));
2290 ldlm_callback_reply(req, -EPROTO);
2294 ns = req->rq_export->exp_obd->obd_namespace;
2295 LASSERT(ns != NULL);
2297 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2299 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2300 if (dlm_req == NULL) {
2301 rc = ldlm_callback_reply(req, -EPROTO);
2302 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2308 * Force a known safe race, send a cancel to the server for a lock
2309 * which the server has already started a blocking callback on.
2311 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2312 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2313 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2315 CERROR("ldlm_cli_cancel: %d\n", rc);
2318 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2321 "callback on lock %#llx - lock disappeared\n",
2322 dlm_req->lock_handle[0].cookie);
2323 rc = ldlm_callback_reply(req, -EINVAL);
2324 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2325 &dlm_req->lock_handle[0]);
2329 if (ldlm_is_fail_loc(lock) &&
2330 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2331 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2333 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2334 lock_res_and_lock(lock);
2335 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2337 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2339 * If somebody cancels lock and cache is already dropped,
2340 * or lock is failed before cp_ast received on client,
2341 * we can tell the server we have no lock. Otherwise, we
2342 * should send cancel after dropping the cache.
2344 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2345 ldlm_is_failed(lock)) {
2347 "callback on lock %llx - lock disappeared",
2348 dlm_req->lock_handle[0].cookie);
2349 unlock_res_and_lock(lock);
2350 LDLM_LOCK_RELEASE(lock);
2351 rc = ldlm_callback_reply(req, -EINVAL);
2352 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2353 &dlm_req->lock_handle[0]);
2357 * BL_AST locks are not needed in LRU.
2358 * Let ldlm_cancel_lru() be fast.
2360 ldlm_lock_remove_from_lru(lock);
2361 ldlm_set_bl_ast(lock);
2363 unlock_res_and_lock(lock);
2366 * We want the ost thread to get this reply so that it can respond
2367 * to ost requests (write cache writeback) that might be triggered
2370 * But we'd also like to be able to indicate in the reply that we're
2371 * cancelling right now, because it's unused, or have an intent result
2372 * in the reply, so we might have to push the responsibility for sending
2373 * the reply down into the AST handlers, alas.
2376 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2377 case LDLM_BL_CALLBACK:
2378 CDEBUG(D_INODE, "blocking ast\n");
2379 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2380 if (!ldlm_is_cancel_on_block(lock)) {
2381 rc = ldlm_callback_reply(req, 0);
2382 if (req->rq_no_reply || rc)
2383 ldlm_callback_errmsg(req, "Normal process", rc,
2384 &dlm_req->lock_handle[0]);
2386 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2387 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2389 case LDLM_CP_CALLBACK:
2390 CDEBUG(D_INODE, "completion ast\n");
2391 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2392 ldlm_callback_reply(req, 0);
2393 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2395 case LDLM_GL_CALLBACK:
2396 CDEBUG(D_INODE, "glimpse ast\n");
2397 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2398 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2401 LBUG(); /* checked above */
2407 #ifdef HAVE_SERVER_SUPPORT
2409 * Main handler for canceld thread.
2411 * Separated into its own thread to avoid deadlocks.
2413 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2420 * Requests arrive in sender's byte order. The ptlrpc service
2421 * handler has already checked and, if necessary, byte-swapped the
2422 * incoming request message body, but I am responsible for the
2426 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2428 if (req->rq_export == NULL) {
2429 struct ldlm_request *dlm_req;
2431 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2432 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2433 libcfs_nid2str(req->rq_peer.nid),
2434 (unsigned long long)req->rq_arrival_time.tv_sec,
2435 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2437 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2438 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2439 dlm_req = req_capsule_client_get(&req->rq_pill,
2441 if (dlm_req != NULL)
2442 ldlm_lock_dump_handle(D_ERROR,
2443 &dlm_req->lock_handle[0]);
2445 ldlm_callback_reply(req, -ENOTCONN);
2449 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2450 /* XXX FIXME move this back to mds/handler.c, b=249 */
2452 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2453 CDEBUG(D_INODE, "cancel\n");
2454 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2455 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2456 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2458 rc = ldlm_handle_cancel(req);
2462 struct ldlm_request *dlm_req;
2464 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2465 CDEBUG(D_INODE, "convert\n");
2467 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2468 if (dlm_req == NULL) {
2469 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2470 rc = ldlm_callback_reply(req, -EPROTO);
2472 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2473 rc = ptlrpc_reply(req);
2478 CERROR("invalid opcode %d\n",
2479 lustre_msg_get_opc(req->rq_reqmsg));
2480 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2481 rc = ldlm_callback_reply(req, -EINVAL);
2487 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2488 struct ldlm_lock *lock)
2490 struct ldlm_request *dlm_req;
2491 struct lustre_handle lockh;
2497 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2498 if (dlm_req == NULL)
2501 ldlm_lock2handle(lock, &lockh);
2502 for (i = 0; i < dlm_req->lock_count; i++) {
2503 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2505 DEBUG_REQ(D_RPCTRACE, req,
2506 "Prio raised by lock %#llx", lockh.cookie);
2515 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2517 struct ldlm_request *dlm_req;
2524 /* no prolong in recovery */
2525 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2528 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2529 if (dlm_req == NULL)
2532 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2533 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2534 (size - offsetof(struct ldlm_request, lock_handle)) /
2535 sizeof(struct lustre_handle) < dlm_req->lock_count)
2538 for (i = 0; i < dlm_req->lock_count; i++) {
2539 struct ldlm_lock *lock;
2541 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2545 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2547 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2548 LDLM_LOCK_PUT(lock);
2557 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2558 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2559 .hpreq_check = ldlm_cancel_hpreq_check,
2563 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2567 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2569 if (req->rq_export == NULL)
2572 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2573 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2574 req->rq_ops = &ldlm_cancel_hpreq_ops;
2575 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2576 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2577 req->rq_ops = &ldlm_cancel_hpreq_ops;
2582 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2583 struct hlist_node *hnode, void *data)
2586 struct list_head *rpc_list = data;
2587 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2589 lock_res_and_lock(lock);
2591 if (!ldlm_is_granted(lock)) {
2592 unlock_res_and_lock(lock);
2596 LASSERT(lock->l_resource);
2597 if (lock->l_resource->lr_type != LDLM_IBITS &&
2598 lock->l_resource->lr_type != LDLM_PLAIN) {
2599 unlock_res_and_lock(lock);
2603 if (ldlm_is_ast_sent(lock)) {
2604 unlock_res_and_lock(lock);
2608 LASSERT(lock->l_blocking_ast);
2609 LASSERT(!lock->l_blocking_lock);
2611 ldlm_set_ast_sent(lock);
2612 if (lock->l_export && lock->l_export->exp_lock_hash) {
2614 * NB: it's safe to call cfs_hash_del() even lock isn't
2618 * In the function below, .hs_keycmp resolves to
2619 * ldlm_export_lock_keycmp()
2621 /* coverity[overrun-buffer-val] */
2622 cfs_hash_del(lock->l_export->exp_lock_hash,
2623 &lock->l_remote_handle, &lock->l_exp_hash);
2626 list_add_tail(&lock->l_rk_ast, rpc_list);
2627 LDLM_LOCK_GET(lock);
2629 unlock_res_and_lock(lock);
2633 void ldlm_revoke_export_locks(struct obd_export *exp)
2635 struct list_head rpc_list;
2639 INIT_LIST_HEAD(&rpc_list);
2640 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2641 ldlm_revoke_lock_cb, &rpc_list, 0);
2642 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2643 LDLM_WORK_REVOKE_AST);
2647 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2648 #endif /* HAVE_SERVER_SUPPORT */
2650 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2651 struct ldlm_bl_work_item **p_blwi,
2652 struct obd_export **p_exp)
2654 struct ldlm_bl_work_item *blwi = NULL;
2655 static unsigned int num_bl;
2656 static unsigned int num_stale;
2657 int num_th = atomic_read(&blp->blp_num_threads);
2659 *p_exp = obd_stale_export_get();
2661 spin_lock(&blp->blp_lock);
2662 if (*p_exp != NULL) {
2663 if (num_th == 1 || ++num_stale < num_th) {
2664 spin_unlock(&blp->blp_lock);
2670 /* process a request from the blp_list at least every blp_num_threads */
2671 if (!list_empty(&blp->blp_list) &&
2672 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2673 blwi = list_entry(blp->blp_list.next,
2674 struct ldlm_bl_work_item, blwi_entry);
2676 if (!list_empty(&blp->blp_prio_list))
2677 blwi = list_entry(blp->blp_prio_list.next,
2678 struct ldlm_bl_work_item,
2682 if (++num_bl >= num_th)
2684 list_del(&blwi->blwi_entry);
2686 spin_unlock(&blp->blp_lock);
2689 if (*p_exp != NULL && *p_blwi != NULL) {
2690 obd_stale_export_put(*p_exp);
2694 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2697 /* This only contains temporary data until the thread starts */
2698 struct ldlm_bl_thread_data {
2699 struct ldlm_bl_pool *bltd_blp;
2700 struct completion bltd_comp;
2704 static int ldlm_bl_thread_main(void *arg);
2706 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2708 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2709 struct task_struct *task;
2711 init_completion(&bltd.bltd_comp);
2713 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2714 if (bltd.bltd_num >= blp->blp_max_threads) {
2715 atomic_dec(&blp->blp_num_threads);
2719 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2721 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2722 atomic_dec(&blp->blp_num_threads);
2726 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2729 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2730 bltd.bltd_num, PTR_ERR(task));
2731 atomic_dec(&blp->blp_num_threads);
2732 return PTR_ERR(task);
2734 wait_for_completion(&bltd.bltd_comp);
2739 /* Not fatal if racy and have a few too many threads */
2740 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2741 struct ldlm_bl_work_item *blwi)
2743 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2746 if (atomic_read(&blp->blp_busy_threads) <
2747 atomic_read(&blp->blp_num_threads))
2750 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2751 blwi->blwi_mem_pressure))
2757 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2758 struct ldlm_bl_work_item *blwi)
2762 if (blwi->blwi_ns == NULL)
2763 /* added by ldlm_cleanup() */
2764 RETURN(LDLM_ITER_STOP);
2766 if (blwi->blwi_mem_pressure)
2767 memory_pressure_set();
2769 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2771 if (blwi->blwi_count) {
2774 * The special case when we cancel locks in lru
2775 * asynchronously, we pass the list of locks here.
2776 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2777 * canceled locally yet.
2779 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2782 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2785 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2788 if (blwi->blwi_mem_pressure)
2789 memory_pressure_clr();
2791 if (blwi->blwi_flags & LCF_ASYNC)
2792 OBD_FREE(blwi, sizeof(*blwi));
2794 complete(&blwi->blwi_comp);
2800 * Cancel stale locks on export. Cancel blocked locks first.
2801 * If the given export has blocked locks, the next in the list may have
2802 * them too, thus cancel not blocked locks only if the current export has
2805 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2806 struct obd_export *exp)
2812 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2814 num = ldlm_export_cancel_blocked_locks(exp);
2816 ldlm_export_cancel_locks(exp);
2818 obd_stale_export_put(exp);
2825 * Main blocking requests processing thread.
2827 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2828 * This thread in the end ends up doing actual call to ->l_blocking_ast
2831 static int ldlm_bl_thread_main(void *arg)
2834 struct ldlm_bl_pool *blp;
2835 struct ldlm_bl_thread_data *bltd = arg;
2843 rc = lu_env_init(env, LCT_DT_THREAD);
2846 rc = lu_env_add(env);
2848 GOTO(out_env_fini, rc);
2850 blp = bltd->bltd_blp;
2852 complete(&bltd->bltd_comp);
2853 /* cannot use bltd after this, it is only on caller's stack */
2856 struct ldlm_bl_work_item *blwi = NULL;
2857 struct obd_export *exp = NULL;
2860 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2863 wait_event_idle_exclusive(blp->blp_waitq,
2864 ldlm_bl_get_work(blp, &blwi,
2866 atomic_inc(&blp->blp_busy_threads);
2868 if (ldlm_bl_thread_need_create(blp, blwi))
2869 /* discard the return value, we tried */
2870 ldlm_bl_thread_start(blp, true);
2873 rc = ldlm_bl_thread_exports(blp, exp);
2875 rc = ldlm_bl_thread_blwi(blp, blwi);
2877 atomic_dec(&blp->blp_busy_threads);
2879 if (rc == LDLM_ITER_STOP)
2883 * If there are many namespaces, we will not sleep waiting for
2884 * work, and must do a cond_resched to avoid holding the CPU
2890 atomic_dec(&blp->blp_num_threads);
2891 complete(&blp->blp_comp);
2902 static int ldlm_setup(void);
2903 static int ldlm_cleanup(void);
2905 int ldlm_get_ref(void)
2910 mutex_lock(&ldlm_ref_mutex);
2911 if (++ldlm_refcount == 1) {
2916 mutex_unlock(&ldlm_ref_mutex);
2921 void ldlm_put_ref(void)
2924 mutex_lock(&ldlm_ref_mutex);
2925 if (ldlm_refcount == 1) {
2926 int rc = ldlm_cleanup();
2929 CERROR("ldlm_cleanup failed: %d\n", rc);
2935 mutex_unlock(&ldlm_ref_mutex);
2941 * Export handle<->lock hash operations.
2944 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
2946 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2950 ldlm_export_lock_key(struct hlist_node *hnode)
2952 struct ldlm_lock *lock;
2954 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2955 return &lock->l_remote_handle;
2959 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2961 struct ldlm_lock *lock;
2963 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2964 lock->l_remote_handle = *(struct lustre_handle *)key;
2968 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2970 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2974 ldlm_export_lock_object(struct hlist_node *hnode)
2976 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2980 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2982 struct ldlm_lock *lock;
2984 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2985 LDLM_LOCK_GET(lock);
2989 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2991 struct ldlm_lock *lock;
2993 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2994 LDLM_LOCK_RELEASE(lock);
2997 static struct cfs_hash_ops ldlm_export_lock_ops = {
2998 .hs_hash = ldlm_export_lock_hash,
2999 .hs_key = ldlm_export_lock_key,
3000 .hs_keycmp = ldlm_export_lock_keycmp,
3001 .hs_keycpy = ldlm_export_lock_keycpy,
3002 .hs_object = ldlm_export_lock_object,
3003 .hs_get = ldlm_export_lock_get,
3004 .hs_put = ldlm_export_lock_put,
3005 .hs_put_locked = ldlm_export_lock_put,
3008 int ldlm_init_export(struct obd_export *exp)
3014 exp->exp_lock_hash =
3015 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3016 HASH_EXP_LOCK_CUR_BITS,
3017 HASH_EXP_LOCK_MAX_BITS,
3018 HASH_EXP_LOCK_BKT_BITS, 0,
3019 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3020 &ldlm_export_lock_ops,
3021 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3022 CFS_HASH_NBLK_CHANGE);
3024 if (!exp->exp_lock_hash)
3027 rc = ldlm_init_flock_export(exp);
3033 ldlm_destroy_export(exp);
3036 EXPORT_SYMBOL(ldlm_init_export);
3038 void ldlm_destroy_export(struct obd_export *exp)
3041 cfs_hash_putref(exp->exp_lock_hash);
3042 exp->exp_lock_hash = NULL;
3044 ldlm_destroy_flock_export(exp);
3047 EXPORT_SYMBOL(ldlm_destroy_export);
3049 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3050 struct attribute *attr,
3053 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3056 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3057 struct attribute *attr,
3064 rc = kstrtoul(buffer, 10, &val);
3068 ldlm_cancel_unused_locks_before_replay = val;
3072 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3074 static struct attribute *ldlm_attrs[] = {
3075 &lustre_attr_cancel_unused_locks_before_replay.attr,
3079 static struct attribute_group ldlm_attr_group = {
3080 .attrs = ldlm_attrs,
3083 static int ldlm_setup(void)
3085 static struct ptlrpc_service_conf conf;
3086 struct ldlm_bl_pool *blp = NULL;
3087 #ifdef HAVE_SERVER_SUPPORT
3088 struct task_struct *task;
3089 #endif /* HAVE_SERVER_SUPPORT */
3095 if (ldlm_state != NULL)
3098 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3099 if (ldlm_state == NULL)
3102 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3106 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3110 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3114 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3118 rc = ldlm_debugfs_setup();
3122 memset(&conf, 0, sizeof(conf));
3123 conf = (typeof(conf)) {
3124 .psc_name = "ldlm_cbd",
3125 .psc_watchdog_factor = 2,
3127 .bc_nbufs = LDLM_CLIENT_NBUFS,
3128 .bc_buf_size = LDLM_BUFSIZE,
3129 .bc_req_max_size = LDLM_MAXREQSIZE,
3130 .bc_rep_max_size = LDLM_MAXREPSIZE,
3131 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3132 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3135 .tc_thr_name = "ldlm_cb",
3136 .tc_thr_factor = LDLM_THR_FACTOR,
3137 .tc_nthrs_init = LDLM_NTHRS_INIT,
3138 .tc_nthrs_base = LDLM_NTHRS_BASE,
3139 .tc_nthrs_max = LDLM_NTHRS_MAX,
3140 .tc_nthrs_user = ldlm_num_threads,
3141 .tc_cpu_bind = ldlm_cpu_bind,
3142 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3145 .cc_pattern = ldlm_cpts,
3146 .cc_affinity = true,
3149 .so_req_handler = ldlm_callback_handler,
3152 ldlm_state->ldlm_cb_service = \
3153 ptlrpc_register_service(&conf, ldlm_svc_kset,
3154 ldlm_svc_debugfs_dir);
3155 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3156 CERROR("failed to start service\n");
3157 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3158 ldlm_state->ldlm_cb_service = NULL;
3162 #ifdef HAVE_SERVER_SUPPORT
3163 memset(&conf, 0, sizeof(conf));
3164 conf = (typeof(conf)) {
3165 .psc_name = "ldlm_canceld",
3166 .psc_watchdog_factor = 6,
3168 .bc_nbufs = LDLM_SERVER_NBUFS,
3169 .bc_buf_size = LDLM_BUFSIZE,
3170 .bc_req_max_size = LDLM_MAXREQSIZE,
3171 .bc_rep_max_size = LDLM_MAXREPSIZE,
3172 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3173 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3177 .tc_thr_name = "ldlm_cn",
3178 .tc_thr_factor = LDLM_THR_FACTOR,
3179 .tc_nthrs_init = LDLM_NTHRS_INIT,
3180 .tc_nthrs_base = LDLM_NTHRS_BASE,
3181 .tc_nthrs_max = LDLM_NTHRS_MAX,
3182 .tc_nthrs_user = ldlm_num_threads,
3183 .tc_cpu_bind = ldlm_cpu_bind,
3184 .tc_ctx_tags = LCT_MD_THREAD | \
3189 .cc_pattern = ldlm_cpts,
3190 .cc_affinity = true,
3193 .so_req_handler = ldlm_cancel_handler,
3194 .so_hpreq_handler = ldlm_hpreq_handler,
3197 ldlm_state->ldlm_cancel_service = \
3198 ptlrpc_register_service(&conf, ldlm_svc_kset,
3199 ldlm_svc_debugfs_dir);
3200 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3201 CERROR("failed to start service\n");
3202 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3203 ldlm_state->ldlm_cancel_service = NULL;
3206 #endif /* HAVE_SERVER_SUPPORT */
3208 OBD_ALLOC(blp, sizeof(*blp));
3210 GOTO(out, rc = -ENOMEM);
3211 ldlm_state->ldlm_bl_pool = blp;
3213 spin_lock_init(&blp->blp_lock);
3214 INIT_LIST_HEAD(&blp->blp_list);
3215 INIT_LIST_HEAD(&blp->blp_prio_list);
3216 init_waitqueue_head(&blp->blp_waitq);
3217 atomic_set(&blp->blp_num_threads, 0);
3218 atomic_set(&blp->blp_busy_threads, 0);
3220 if (ldlm_num_threads == 0) {
3221 blp->blp_min_threads = LDLM_NTHRS_INIT;
3222 blp->blp_max_threads = LDLM_NTHRS_MAX;
3224 blp->blp_min_threads = blp->blp_max_threads = \
3225 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3229 for (i = 0; i < blp->blp_min_threads; i++) {
3230 rc = ldlm_bl_thread_start(blp, false);
3235 #ifdef HAVE_SERVER_SUPPORT
3236 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3239 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3243 wait_event(expired_lock_wait_queue,
3244 expired_lock_thread_state == ELT_READY);
3245 #endif /* HAVE_SERVER_SUPPORT */
3247 rc = ldlm_pools_init();
3249 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3253 rc = ldlm_reclaim_setup();
3255 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3265 static int ldlm_cleanup(void)
3269 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3270 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3271 CERROR("ldlm still has namespaces; clean these up first.\n");
3272 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3273 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3277 ldlm_reclaim_cleanup();
3280 if (ldlm_state->ldlm_bl_pool != NULL) {
3281 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3283 while (atomic_read(&blp->blp_num_threads) > 0) {
3284 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3286 init_completion(&blp->blp_comp);
3288 spin_lock(&blp->blp_lock);
3289 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3290 wake_up(&blp->blp_waitq);
3291 spin_unlock(&blp->blp_lock);
3293 wait_for_completion(&blp->blp_comp);
3296 OBD_FREE(blp, sizeof(*blp));
3299 if (ldlm_state->ldlm_cb_service != NULL)
3300 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3301 #ifdef HAVE_SERVER_SUPPORT
3302 if (ldlm_state->ldlm_cancel_service != NULL)
3303 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3307 kset_unregister(ldlm_ns_kset);
3309 kset_unregister(ldlm_svc_kset);
3311 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3312 kobject_put(ldlm_kobj);
3315 ldlm_debugfs_cleanup();
3317 #ifdef HAVE_SERVER_SUPPORT
3318 if (expired_lock_thread_state != ELT_STOPPED) {
3319 expired_lock_thread_state = ELT_TERMINATE;
3320 wake_up(&expired_lock_wait_queue);
3321 wait_event(expired_lock_wait_queue,
3322 expired_lock_thread_state == ELT_STOPPED);
3326 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3334 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3335 sizeof(struct ldlm_resource), 0,
3336 SLAB_HWCACHE_ALIGN, NULL);
3337 if (ldlm_resource_slab == NULL)
3340 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3341 sizeof(struct ldlm_lock), 0,
3342 SLAB_HWCACHE_ALIGN, NULL);
3343 if (ldlm_lock_slab == NULL)
3346 ldlm_interval_slab = kmem_cache_create("interval_node",
3347 sizeof(struct ldlm_interval),
3348 0, SLAB_HWCACHE_ALIGN, NULL);
3349 if (ldlm_interval_slab == NULL)
3352 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3353 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3354 0, SLAB_HWCACHE_ALIGN, NULL);
3355 if (ldlm_interval_tree_slab == NULL)
3358 #ifdef HAVE_SERVER_SUPPORT
3359 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3360 sizeof(struct ldlm_ibits_node),
3361 0, SLAB_HWCACHE_ALIGN, NULL);
3362 if (ldlm_inodebits_slab == NULL)
3363 goto out_interval_tree;
3365 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3366 sizeof(struct ldlm_glimpse_work),
3368 if (ldlm_glimpse_work_kmem == NULL)
3372 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3373 class_export_dump_hook = ldlm_dump_export_locks;
3376 #ifdef HAVE_SERVER_SUPPORT
3378 kmem_cache_destroy(ldlm_inodebits_slab);
3380 kmem_cache_destroy(ldlm_interval_tree_slab);
3383 kmem_cache_destroy(ldlm_interval_slab);
3385 kmem_cache_destroy(ldlm_lock_slab);
3387 kmem_cache_destroy(ldlm_resource_slab);
3392 void ldlm_exit(void)
3395 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3396 kmem_cache_destroy(ldlm_resource_slab);
3398 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3399 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3400 * so that ldlm_lock_free() get a chance to be called.
3403 kmem_cache_destroy(ldlm_lock_slab);
3404 kmem_cache_destroy(ldlm_interval_slab);
3405 kmem_cache_destroy(ldlm_interval_tree_slab);
3406 #ifdef HAVE_SERVER_SUPPORT
3407 kmem_cache_destroy(ldlm_inodebits_slab);
3408 kmem_cache_destroy(ldlm_glimpse_work_kmem);