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_move(&lock->l_pending_chain, &expired_lock_list);
339 if (!list_empty(&expired_lock_list)) {
340 if (obd_dump_on_timeout && need_dump)
341 expired_lock_dump = __LINE__;
343 wake_up(&expired_lock_wait_queue);
347 * Make sure the timer will fire again if we have any locks
350 if (!list_empty(&waiting_locks_list)) {
351 time64_t now = ktime_get_seconds();
354 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
356 if (lock->l_callback_timeout - now > 0)
357 delta = lock->l_callback_timeout - now;
358 mod_timer(&waiting_locks_timer,
359 jiffies + cfs_time_seconds(delta));
361 spin_unlock_bh(&waiting_locks_spinlock);
365 * Add lock to the list of contended locks.
367 * Indicate that we're waiting for a client to call us back cancelling a given
368 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
369 * timer to fire appropriately. (We round up to the next second, to avoid
370 * floods of timer firings during periods of high lock contention and traffic).
371 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
372 * if it has been added to the waiting list (1 is returned).
374 * Called with the namespace lock held.
376 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds)
378 unsigned long timeout_jiffies = jiffies;
379 time64_t now = ktime_get_seconds();
383 if (!list_empty(&lock->l_pending_chain))
386 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
387 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
390 deadline = now + seconds;
391 if (likely(deadline > lock->l_callback_timeout))
392 lock->l_callback_timeout = deadline;
394 timeout = clamp_t(time_t, lock->l_callback_timeout - now,
396 timeout_jiffies += cfs_time_seconds(timeout);
398 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
399 !timer_pending(&waiting_locks_timer))
400 mod_timer(&waiting_locks_timer, timeout_jiffies);
403 * if the new lock has a shorter timeout than something earlier on
404 * the list, we'll wait the longer amount of time; no big deal.
407 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
411 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
413 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
414 if (list_empty(&lock->l_exp_list)) {
415 if (!ldlm_is_granted(lock))
416 list_add_tail(&lock->l_exp_list,
417 &lock->l_export->exp_bl_list);
419 list_add(&lock->l_exp_list,
420 &lock->l_export->exp_bl_list);
422 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
425 * A blocked lock is added. Adjust the position in
426 * the stale list if the export is in the list.
427 * If export is stale and not in the list - it is being
428 * processed and will be placed on the right position
429 * on obd_stale_export_put().
431 if (!list_empty(&lock->l_export->exp_stale_list))
432 obd_stale_export_adjust(lock->l_export);
435 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
439 /* NB: must be called with hold of lock_res_and_lock() */
440 LASSERT(ldlm_is_res_locked(lock));
441 LASSERT(!ldlm_is_cancel_on_block(lock));
444 * Do not put cross-MDT lock in the waiting list, since we
445 * will not evict it due to timeout for now
447 if (lock->l_export != NULL &&
448 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
451 spin_lock_bh(&waiting_locks_spinlock);
452 if (ldlm_is_cancel(lock)) {
453 spin_unlock_bh(&waiting_locks_spinlock);
457 if (ldlm_is_destroyed(lock)) {
458 static time64_t next;
460 spin_unlock_bh(&waiting_locks_spinlock);
461 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
462 if (ktime_get_seconds() > next) {
463 next = ktime_get_seconds() + 14400;
464 libcfs_debug_dumpstack(NULL);
469 ldlm_set_waited(lock);
470 lock->l_blast_sent = ktime_get_real_seconds();
471 ret = __ldlm_add_waiting_lock(lock, timeout);
474 * grab ref on the lock if it has been added to the
479 spin_unlock_bh(&waiting_locks_spinlock);
482 ldlm_add_blocked_lock(lock);
484 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)",
485 ret == 0 ? "not re-" : "", timeout,
486 AT_OFF ? "off" : "on");
491 * Remove a lock from the pending list, likely because it had its cancellation
492 * callback arrive without incident. This adjusts the lock-timeout timer if
493 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
494 * As done by ldlm_del_waiting_lock(), the caller must release the lock
495 * reference when the lock is removed from any list (1 is returned).
497 * Called with namespace lock held.
499 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
501 struct list_head *list_next;
503 if (list_empty(&lock->l_pending_chain))
506 list_next = lock->l_pending_chain.next;
507 if (lock->l_pending_chain.prev == &waiting_locks_list) {
508 /* Removing the head of the list, adjust timer. */
509 if (list_next == &waiting_locks_list) {
510 /* No more, just cancel. */
511 del_timer(&waiting_locks_timer);
513 time64_t now = ktime_get_seconds();
514 struct ldlm_lock *next;
517 next = list_entry(list_next, struct ldlm_lock,
519 if (next->l_callback_timeout - now > 0)
520 delta = lock->l_callback_timeout - now;
522 mod_timer(&waiting_locks_timer,
523 jiffies + cfs_time_seconds(delta));
526 list_del_init(&lock->l_pending_chain);
531 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
535 if (lock->l_export == NULL) {
536 /* We don't have a "waiting locks list" on clients. */
537 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
541 spin_lock_bh(&waiting_locks_spinlock);
542 ret = __ldlm_del_waiting_lock(lock);
543 ldlm_clear_waited(lock);
544 spin_unlock_bh(&waiting_locks_spinlock);
546 /* remove the lock out of export blocking list */
547 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
548 list_del_init(&lock->l_exp_list);
549 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
553 * release lock ref if it has indeed been removed
556 LDLM_LOCK_RELEASE(lock);
559 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
564 * Prolong the contended lock waiting time.
566 * Called with namespace lock held.
568 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
570 if (lock->l_export == NULL) {
571 /* We don't have a "waiting locks list" on clients. */
572 LDLM_DEBUG(lock, "client lock: no-op");
576 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
577 /* We don't have a "waiting locks list" on OSP. */
578 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
582 spin_lock_bh(&waiting_locks_spinlock);
584 if (list_empty(&lock->l_pending_chain)) {
585 spin_unlock_bh(&waiting_locks_spinlock);
586 LDLM_DEBUG(lock, "wasn't waiting");
591 * we remove/add the lock to the waiting list, so no needs to
592 * release/take a lock reference
594 __ldlm_del_waiting_lock(lock);
595 __ldlm_add_waiting_lock(lock, timeout);
596 spin_unlock_bh(&waiting_locks_spinlock);
598 LDLM_DEBUG(lock, "refreshed");
601 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
603 #else /* HAVE_SERVER_SUPPORT */
605 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
610 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
615 #endif /* !HAVE_SERVER_SUPPORT */
617 #ifdef HAVE_SERVER_SUPPORT
620 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
621 * lock cancel, and their replies). Used for lock callback timeout and AST
624 * \param[in] lock lock which is getting the blocking callback
626 * \retval timeout in seconds to wait for the client reply
628 time64_t ldlm_bl_timeout(struct ldlm_lock *lock)
633 return obd_timeout / 2;
636 * Since these are non-updating timeouts, we should be conservative.
637 * Take more than usually, 150%
638 * It would be nice to have some kind of "early reply" mechanism for
639 * lock callbacks too...
641 timeout = at_get(&lock->l_export->exp_bl_lock_at);
642 return max(timeout + (timeout >> 1), (time64_t)ldlm_enqueue_min);
644 EXPORT_SYMBOL(ldlm_bl_timeout);
647 * Perform lock cleanup if AST sending failed.
649 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
650 const char *ast_type)
652 LCONSOLE_ERROR_MSG(0x138,
653 "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
654 lock->l_export->exp_obd->obd_name,
655 obd_export_nid2str(lock->l_export), ast_type, rc);
657 if (obd_dump_on_timeout)
658 libcfs_debug_dumplog();
659 spin_lock_bh(&waiting_locks_spinlock);
660 if (__ldlm_del_waiting_lock(lock) == 0)
662 * the lock was not in any list, grab an extra ref before adding
663 * the lock to the expired list
666 lock->l_callback_timeout = 0; /* differentiate it from expired locks */
667 list_add(&lock->l_pending_chain, &expired_lock_list);
668 wake_up(&expired_lock_wait_queue);
669 spin_unlock_bh(&waiting_locks_spinlock);
673 * Perform lock cleanup if AST reply came with error.
675 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
676 struct ptlrpc_request *req, int rc,
677 const char *ast_type)
679 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
681 if (!req->rq_replied || (rc && rc != -EINVAL)) {
682 if (ldlm_is_cancel(lock)) {
684 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
685 ast_type, req, req->rq_xid,
686 libcfs_nid2str(peer.nid));
687 ldlm_lock_cancel(lock);
689 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
691 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
693 * Upon umount process the AST fails because cannot be
694 * sent. This shouldn't lead to the client eviction.
695 * -ENODEV error is returned by ptl_send_rpc() for
696 * new request in such import.
697 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
698 * if imp_invalid is set or obd_no_recov.
699 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
700 * in ptlrpc_import_delay_req() as well with -EIO code.
701 * In all such cases errors are ignored.
704 "%s AST can't be sent due to a server %s failure or umount process: rc = %d\n",
706 req->rq_import->imp_obd->obd_name, rc);
709 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
710 libcfs_nid2str(peer.nid),
711 req->rq_replied ? "returned error from" :
712 "failed to reply to",
713 ast_type, req, req->rq_xid,
714 (req->rq_repmsg != NULL) ?
715 lustre_msg_get_status(req->rq_repmsg) : 0,
717 ldlm_failed_ast(lock, rc, ast_type);
723 struct ldlm_resource *res = lock->l_resource;
726 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
727 libcfs_nid2str(peer.nid),
729 lustre_msg_get_status(req->rq_repmsg) : -1,
730 ast_type, req, req->rq_xid);
733 * update lvbo to return proper attributes.
736 ldlm_resource_getref(res);
737 ldlm_lvbo_update(res, lock, NULL, 1);
738 ldlm_resource_putref(res);
740 ldlm_lock_cancel(lock);
747 static int ldlm_cb_interpret(const struct lu_env *env,
748 struct ptlrpc_request *req, void *args, int rc)
750 struct ldlm_cb_async_args *ca = args;
751 struct ldlm_lock *lock = ca->ca_lock;
752 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
756 LASSERT(lock != NULL);
759 case LDLM_GL_CALLBACK:
761 * Update the LVB from disk if the AST failed
762 * (this is a legal race)
764 * - Glimpse callback of local lock just returns
765 * -ELDLM_NO_LOCK_DATA.
766 * - Glimpse callback of remote lock might return
767 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
769 if (unlikely(arg->gl_interpret_reply)) {
770 rc = arg->gl_interpret_reply(NULL, req, args, rc);
771 } else if (rc == -ELDLM_NO_LOCK_DATA) {
773 "lost race - client has a lock but no inode");
774 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
775 } else if (rc != 0) {
776 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
778 rc = ldlm_lvbo_update(lock->l_resource,
782 case LDLM_BL_CALLBACK:
784 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
786 case LDLM_CP_CALLBACK:
788 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
791 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
796 /* release extra reference taken in ldlm_ast_fini() */
797 LDLM_LOCK_RELEASE(lock);
800 atomic_inc(&arg->restart);
805 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
807 struct ldlm_cb_async_args *ca = data;
808 struct ldlm_lock *lock = ca->ca_lock;
810 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
813 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
814 struct ldlm_cb_set_arg *arg,
815 struct ldlm_lock *lock,
822 if (unlikely(instant_cancel)) {
823 rc = ptl_send_rpc(req, 1);
824 ptlrpc_req_finished(req);
826 atomic_inc(&arg->restart);
829 ptlrpc_set_add_req(arg->set, req);
836 * Check if there are requests in the export request list which prevent
837 * the lock canceling and make these requests high priority ones.
839 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
841 struct ptlrpc_request *req;
845 if (lock->l_export == NULL) {
846 LDLM_DEBUG(lock, "client lock: no-op");
850 spin_lock(&lock->l_export->exp_rpc_lock);
851 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
854 * Do not process requests that were not yet added to there
855 * incoming queue or were already removed from there for
856 * processing. We evaluate ptlrpc_nrs_req_can_move() without
857 * holding svcpt->scp_req_lock, and then redo the check with
858 * the lock held once we need to obtain a reliable result.
860 if (ptlrpc_nrs_req_can_move(req) &&
861 req->rq_ops->hpreq_lock_match &&
862 req->rq_ops->hpreq_lock_match(req, lock))
863 ptlrpc_nrs_req_hp_move(req);
865 spin_unlock(&lock->l_export->exp_rpc_lock);
870 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
871 * enqueued server lock conflicts with given one.
873 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
874 * to wait for client response.
876 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
877 struct ldlm_lock_desc *desc,
878 void *data, int flag)
880 struct ldlm_cb_async_args *ca;
881 struct ldlm_cb_set_arg *arg = data;
882 struct ldlm_request *body;
883 struct ptlrpc_request *req;
884 int instant_cancel = 0;
889 if (flag == LDLM_CB_CANCELING)
890 /* Don't need to do anything here. */
893 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
894 LDLM_DEBUG(lock, "dropping BL AST");
899 LASSERT(data != NULL);
900 if (lock->l_export->exp_obd->obd_recovering != 0)
901 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
903 ldlm_lock_reorder_req(lock);
905 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
906 &RQF_LDLM_BL_CALLBACK,
907 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
911 ca = ptlrpc_req_async_args(ca, req);
912 ca->ca_set_arg = arg;
915 req->rq_interpret_reply = ldlm_cb_interpret;
917 lock_res_and_lock(lock);
918 if (ldlm_is_destroyed(lock)) {
919 /* What's the point? */
920 unlock_res_and_lock(lock);
921 ptlrpc_req_finished(req);
925 if (!ldlm_is_granted(lock)) {
927 * this blocking AST will be communicated as part of the
928 * completion AST instead
930 ldlm_add_blocked_lock(lock);
931 ldlm_set_waited(lock);
932 unlock_res_and_lock(lock);
934 ptlrpc_req_finished(req);
935 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
939 if (ldlm_is_cancel_on_block(lock))
942 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
943 body->lock_handle[0] = lock->l_remote_handle;
944 body->lock_desc = *desc;
945 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
947 LDLM_DEBUG(lock, "server preparing blocking AST");
949 ptlrpc_request_set_replen(req);
950 ldlm_set_cbpending(lock);
951 if (instant_cancel) {
952 unlock_res_and_lock(lock);
953 ldlm_lock_cancel(lock);
955 req->rq_no_resend = 1;
957 LASSERT(ldlm_is_granted(lock));
958 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
959 unlock_res_and_lock(lock);
961 /* Do not resend after lock callback timeout */
962 req->rq_delay_limit = ldlm_bl_timeout(lock);
963 req->rq_resend_cb = ldlm_update_resend;
966 req->rq_send_state = LUSTRE_IMP_FULL;
967 /* ptlrpc_request_alloc_pack already set timeout */
969 req->rq_timeout = ldlm_get_rq_timeout();
971 if (lock->l_export && lock->l_export->exp_nid_stats &&
972 lock->l_export->exp_nid_stats->nid_ldlm_stats)
973 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
974 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
976 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
982 * ->l_completion_ast callback for a remote lock in server namespace.
984 * Sends AST to the client notifying it of lock granting. If initial
985 * lock response was not sent yet, instead of sending another RPC, just
986 * mark the lock as granted and client will understand
988 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
990 struct ldlm_cb_set_arg *arg = data;
991 struct ldlm_request *body;
992 struct ptlrpc_request *req;
993 struct ldlm_cb_async_args *ca;
994 int instant_cancel = 0;
1000 LASSERT(lock != NULL);
1001 LASSERT(data != NULL);
1003 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1004 LDLM_DEBUG(lock, "dropping CP AST");
1008 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1009 &RQF_LDLM_CP_CALLBACK);
1013 /* server namespace, doesn't need lock */
1014 lvb_len = ldlm_lvbo_size(lock);
1016 * LU-3124 & LU-2187: to not return layout in completion AST because
1017 * it may deadlock for LU-2187, or client may not have enough space
1018 * for large layout. The layout will be returned to client with an
1019 * extra RPC to fetch xattr.lov
1021 if (ldlm_has_layout(lock))
1024 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1025 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1027 ptlrpc_request_free(req);
1031 ca = ptlrpc_req_async_args(ca, req);
1032 ca->ca_set_arg = arg;
1035 req->rq_interpret_reply = ldlm_cb_interpret;
1036 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1038 body->lock_handle[0] = lock->l_remote_handle;
1039 body->lock_flags = ldlm_flags_to_wire(flags);
1040 ldlm_lock2desc(lock, &body->lock_desc);
1042 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1043 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1046 * We still need to send the RPC to wake up the blocked
1047 * enqueue thread on the client.
1049 * Consider old client, there is no better way to notify
1050 * the failure, just zero-sized the LVB, then the client
1051 * will fail out as "-EPROTO".
1053 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1057 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1062 LDLM_DEBUG(lock, "server preparing completion AST");
1064 ptlrpc_request_set_replen(req);
1066 req->rq_send_state = LUSTRE_IMP_FULL;
1067 /* ptlrpc_request_pack already set timeout */
1069 req->rq_timeout = ldlm_get_rq_timeout();
1071 /* We only send real blocking ASTs after the lock is granted */
1072 lock_res_and_lock(lock);
1073 if (ldlm_is_ast_sent(lock)) {
1074 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1075 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1076 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1080 * We might get here prior to ldlm_handle_enqueue setting
1081 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1082 * into waiting list, but this is safe and similar code in
1083 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1084 * that would not only cancel the lock, but will also remove
1085 * it from waiting list
1087 if (ldlm_is_cancel_on_block(lock)) {
1088 unlock_res_and_lock(lock);
1089 ldlm_lock_cancel(lock);
1092 req->rq_no_resend = 1;
1094 lock_res_and_lock(lock);
1096 /* start the lock-timeout clock */
1097 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1098 /* Do not resend after lock callback timeout */
1099 req->rq_delay_limit = ldlm_bl_timeout(lock);
1100 req->rq_resend_cb = ldlm_update_resend;
1103 unlock_res_and_lock(lock);
1105 if (lock->l_export && lock->l_export->exp_nid_stats &&
1106 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1107 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1108 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1110 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1112 RETURN(lvb_len < 0 ? lvb_len : rc);
1116 * Server side ->l_glimpse_ast handler for client locks.
1118 * Sends glimpse AST to the client and waits for reply. Then updates
1119 * lvbo with the result.
1121 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1123 struct ldlm_cb_set_arg *arg = data;
1124 struct ldlm_request *body;
1125 struct ptlrpc_request *req;
1126 struct ldlm_cb_async_args *ca;
1128 struct req_format *req_fmt;
1132 LASSERT(lock != NULL);
1134 if (arg->gl_desc != NULL)
1135 /* There is a glimpse descriptor to pack */
1136 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1138 req_fmt = &RQF_LDLM_GL_CALLBACK;
1140 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1141 req_fmt, LUSTRE_DLM_VERSION,
1147 if (arg->gl_desc != NULL) {
1148 /* copy the GL descriptor */
1149 union ldlm_gl_desc *desc;
1151 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1152 *desc = *arg->gl_desc;
1155 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1156 body->lock_handle[0] = lock->l_remote_handle;
1157 ldlm_lock2desc(lock, &body->lock_desc);
1159 ca = ptlrpc_req_async_args(ca, req);
1160 ca->ca_set_arg = arg;
1163 /* server namespace, doesn't need lock */
1164 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1165 ldlm_lvbo_size(lock));
1166 ptlrpc_request_set_replen(req);
1168 req->rq_send_state = LUSTRE_IMP_FULL;
1169 /* ptlrpc_request_alloc_pack already set timeout */
1171 req->rq_timeout = ldlm_get_rq_timeout();
1173 req->rq_interpret_reply = ldlm_cb_interpret;
1175 if (lock->l_export && lock->l_export->exp_nid_stats) {
1176 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1178 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1179 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1182 rc = ldlm_ast_fini(req, arg, lock, 0);
1186 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1188 int ldlm_glimpse_locks(struct ldlm_resource *res,
1189 struct list_head *gl_work_list)
1195 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1197 if (rc == -ERESTART)
1198 ldlm_reprocess_all(res, NULL);
1202 EXPORT_SYMBOL(ldlm_glimpse_locks);
1204 /* return LDLM lock associated with a lock callback request */
1205 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1207 struct ldlm_cb_async_args *ca;
1208 struct ldlm_lock *lock;
1212 ca = ptlrpc_req_async_args(ca, req);
1215 RETURN(ERR_PTR(-EFAULT));
1219 EXPORT_SYMBOL(ldlm_request_lock);
1222 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1223 * service threads to carry out client lock enqueueing requests.
1225 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1226 struct ptlrpc_request *req,
1227 const struct ldlm_request *dlm_req,
1228 const struct ldlm_callback_suite *cbs)
1230 struct ldlm_reply *dlm_rep;
1232 enum ldlm_error err = ELDLM_OK;
1233 struct ldlm_lock *lock = NULL;
1234 void *cookie = NULL;
1236 struct ldlm_resource *res = NULL;
1237 const struct lu_env *env = req->rq_svc_thread->t_env;
1241 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1243 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1244 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1246 LASSERT(req->rq_export);
1248 /* for intent enqueue the stat will be updated inside intent policy */
1249 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1250 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1251 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1253 if (req->rq_export && req->rq_export->exp_nid_stats &&
1254 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1255 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1256 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1258 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1259 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1260 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1261 dlm_req->lock_desc.l_resource.lr_type);
1262 GOTO(out, rc = -EFAULT);
1265 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1266 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1267 dlm_req->lock_desc.l_req_mode &
1268 (dlm_req->lock_desc.l_req_mode-1))) {
1269 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1270 dlm_req->lock_desc.l_req_mode);
1271 GOTO(out, rc = -EFAULT);
1274 if (unlikely((flags & LDLM_FL_REPLAY) ||
1275 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1276 /* Find an existing lock in the per-export lock hash */
1278 * In the function below, .hs_keycmp resolves to
1279 * ldlm_export_lock_keycmp()
1281 /* coverity[overrun-buffer-val] */
1282 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1283 (void *)&dlm_req->lock_handle[0]);
1285 DEBUG_REQ(D_DLMTRACE, req,
1286 "found existing lock cookie %#llx",
1287 lock->l_handle.h_cookie);
1288 flags |= LDLM_FL_RESENT;
1289 GOTO(existing_lock, rc = 0);
1292 if (ldlm_reclaim_full()) {
1293 DEBUG_REQ(D_DLMTRACE, req,
1294 "Too many granted locks, reject current enqueue request and let the client retry later");
1295 GOTO(out, rc = -EINPROGRESS);
1299 /* The lock's callback data might be set in the policy function */
1300 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1301 dlm_req->lock_desc.l_resource.lr_type,
1302 dlm_req->lock_desc.l_req_mode,
1303 cbs, NULL, 0, LVB_T_NONE);
1310 lock->l_remote_handle = dlm_req->lock_handle[0];
1311 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1314 * Initialize resource lvb but not for a lock being replayed since
1315 * Client already got lvb sent in this case.
1316 * This must occur early since some policy methods assume resource
1317 * lvb is available (lr_lvb_data != NULL).
1319 res = lock->l_resource;
1320 if (!(flags & LDLM_FL_REPLAY)) {
1321 /* non-replayed lock, delayed lvb init may need to be done */
1322 rc = ldlm_lvbo_init(res);
1324 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1329 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1331 * Don't enqueue a lock onto the export if it is been disonnected
1332 * due to eviction (b=3822) or server umount (b=24324).
1333 * Cancel it now instead.
1335 if (req->rq_export->exp_disconnected) {
1336 LDLM_ERROR(lock, "lock on disconnected export %p",
1338 GOTO(out, rc = -ENOTCONN);
1341 lock->l_export = class_export_lock_get(req->rq_export, lock);
1342 if (lock->l_export->exp_lock_hash)
1343 cfs_hash_add(lock->l_export->exp_lock_hash,
1344 &lock->l_remote_handle,
1348 * Inherit the enqueue flags before the operation, because we do not
1349 * keep the res lock on return and next operations (BL AST) may proceed
1352 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1353 LDLM_FL_INHERIT_MASK);
1355 ldlm_convert_policy_to_local(req->rq_export,
1356 dlm_req->lock_desc.l_resource.lr_type,
1357 &dlm_req->lock_desc.l_policy_data,
1358 &lock->l_policy_data);
1359 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1360 lock->l_req_extent = lock->l_policy_data.l_extent;
1361 else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS)
1362 lock->l_policy_data.l_inodebits.try_bits =
1363 dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1367 if (!(flags & LDLM_FL_HAS_INTENT)) {
1368 /* based on the assumption that lvb size never changes during
1369 * resource life time otherwise it need resource->lr_lock's
1371 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1372 RCL_SERVER, ldlm_lvbo_size(lock));
1374 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1375 GOTO(out, rc = -ENOMEM);
1377 rc = req_capsule_server_pack(&req->rq_pill);
1382 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1389 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1391 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1392 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1394 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1395 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1398 * We never send a blocking AST until the lock is granted, but
1399 * we can tell it right now
1401 lock_res_and_lock(lock);
1404 * Now take into account flags to be inherited from original lock
1405 * request both in reply to client and in our own lock flags.
1407 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1408 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1411 * Don't move a pending lock onto the export if it has already been
1412 * disconnected due to eviction (b=5683) or server umount (b=24324).
1413 * Cancel it now instead.
1415 if (unlikely(req->rq_export->exp_disconnected ||
1416 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1417 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1419 } else if (ldlm_is_ast_sent(lock)) {
1420 /* fill lock desc for possible lock convert */
1421 if (lock->l_blocking_lock &&
1422 lock->l_resource->lr_type == LDLM_IBITS) {
1423 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1424 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1427 "save blocking bits %llx in granted lock",
1428 bl_lock->l_policy_data.l_inodebits.bits);
1430 * If lock is blocked then save blocking ibits
1431 * in returned lock policy for the possible lock
1432 * convert on a client.
1434 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1435 bl_lock->l_policy_data.l_inodebits.bits;
1437 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1438 if (ldlm_is_granted(lock)) {
1440 * Only cancel lock if it was granted, because it would
1441 * be destroyed immediately and would never be granted
1442 * in the future, causing timeouts on client. Not
1443 * granted lock will be cancelled immediately after
1444 * sending completion AST.
1446 if (ldlm_is_cancel_on_block(lock)) {
1447 unlock_res_and_lock(lock);
1448 ldlm_lock_cancel(lock);
1449 lock_res_and_lock(lock);
1451 ldlm_add_waiting_lock(lock,
1452 ldlm_bl_timeout(lock));
1456 unlock_res_and_lock(lock);
1460 req->rq_status = rc ?: err; /* return either error - b=11190 */
1461 if (!req->rq_packed_final) {
1462 err = lustre_pack_reply(req, 1, NULL, NULL);
1468 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1469 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1473 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1477 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1479 ldlm_lvbo_size(lock) > 0) {
1484 buf = req_capsule_server_get(&req->rq_pill,
1486 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1487 buflen = req_capsule_get_size(&req->rq_pill,
1488 &RMF_DLM_LVB, RCL_SERVER);
1490 * non-replayed lock, delayed lvb init may
1491 * need to be occur now
1493 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1496 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1498 req_capsule_shrink(&req->rq_pill,
1501 } else if (rc2 == -ERANGE) {
1502 rc2 = req_capsule_server_grow(
1504 &RMF_DLM_LVB, buflen);
1509 * if we can't grow the buffer,
1510 * it's ok to return empty lvb
1521 } else if (flags & LDLM_FL_REPLAY) {
1522 /* no LVB resend upon replay */
1524 req_capsule_shrink(&req->rq_pill,
1534 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1535 if (lock->l_export) {
1536 ldlm_lock_cancel(lock);
1538 lock_res_and_lock(lock);
1539 ldlm_resource_unlink_lock(lock);
1540 ldlm_lock_destroy_nolock(lock);
1541 unlock_res_and_lock(lock);
1546 if (!err && !ldlm_is_cbpending(lock) &&
1547 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1548 ldlm_reprocess_all(lock->l_resource, lock);
1550 LDLM_LOCK_RELEASE(lock);
1553 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1560 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1561 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1563 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1565 if (lock->l_blocking_lock) {
1566 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1567 lock->l_blocking_lock = NULL;
1571 /* A lock can be converted to new ibits or mode and should be considered
1572 * as new lock. Clear all states related to a previous blocking AST
1573 * processing so new conflicts will cause new blocking ASTs.
1575 * This is used during lock convert below and lock downgrade to COS mode in
1576 * ldlm_lock_mode_downgrade().
1578 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1580 ldlm_clear_ast_sent(lock);
1581 lock->l_bl_ast_run = 0;
1582 ldlm_clear_blocking_lock(lock);
1586 * Main LDLM entry point for server code to process lock conversion requests.
1588 int ldlm_handle_convert0(struct ptlrpc_request *req,
1589 const struct ldlm_request *dlm_req)
1591 struct obd_export *exp = req->rq_export;
1592 struct ldlm_reply *dlm_rep;
1593 struct ldlm_lock *lock;
1600 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1601 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1602 LDLM_CONVERT - LDLM_FIRST_OPC);
1604 rc = req_capsule_server_pack(&req->rq_pill);
1608 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1609 dlm_rep->lock_flags = dlm_req->lock_flags;
1611 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1613 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1614 req->rq_status = ELDLM_NO_LOCK_DATA;
1618 LDLM_DEBUG(lock, "server-side convert handler START");
1620 lock_res_and_lock(lock);
1621 bits = lock->l_policy_data.l_inodebits.bits;
1622 new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1624 if (ldlm_is_cancel(lock)) {
1625 LDLM_DEBUG(lock, "convert on canceled lock!");
1626 unlock_res_and_lock(lock);
1627 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1630 if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1631 LDLM_ERROR(lock, "lock mode differs!");
1632 unlock_res_and_lock(lock);
1633 GOTO(out_put, rc = -EPROTO);
1636 if (bits == new_bits) {
1638 * This can be valid situation if CONVERT RPCs are
1639 * re-ordered. Just finish silently
1641 LDLM_DEBUG(lock, "lock is converted already!");
1642 unlock_res_and_lock(lock);
1644 if (ldlm_is_waited(lock))
1645 ldlm_del_waiting_lock(lock);
1647 ldlm_clear_cbpending(lock);
1648 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1649 ldlm_inodebits_drop(lock, bits & ~new_bits);
1651 ldlm_clear_blocking_data(lock);
1652 unlock_res_and_lock(lock);
1654 ldlm_reprocess_all(lock->l_resource, NULL);
1657 dlm_rep->lock_handle = lock->l_remote_handle;
1658 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1659 &dlm_rep->lock_desc.l_policy_data);
1663 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1664 LDLM_LOCK_PUT(lock);
1665 req->rq_status = rc;
1670 * Cancel all the locks whose handles are packed into ldlm_request
1672 * Called by server code expecting such combined cancel activity
1675 int ldlm_request_cancel(struct ptlrpc_request *req,
1676 const struct ldlm_request *dlm_req,
1677 int first, enum lustre_at_flags flags)
1679 struct ldlm_resource *res, *pres = NULL;
1680 struct ldlm_lock *lock;
1681 int i, count, done = 0;
1686 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1687 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1688 (size - offsetof(struct ldlm_request, lock_handle)) /
1689 sizeof(struct lustre_handle) < dlm_req->lock_count)
1692 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1696 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1700 * There is no lock on the server at the replay time,
1701 * skip lock cancelling to make replay tests to pass.
1703 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1706 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1709 for (i = first; i < count; i++) {
1710 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1712 /* below message checked in replay-single.sh test_36 */
1713 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1714 dlm_req->lock_handle[i].cookie);
1718 res = lock->l_resource;
1722 * This code is an optimization to only attempt lock
1723 * granting on the resource (that could be CPU-expensive)
1724 * after we are done cancelling lock in that resource.
1728 ldlm_reprocess_all(pres, NULL);
1729 LDLM_RESOURCE_DELREF(pres);
1730 ldlm_resource_putref(pres);
1733 ldlm_resource_getref(res);
1734 LDLM_RESOURCE_ADDREF(res);
1736 if (!ldlm_is_discard_data(lock))
1737 ldlm_lvbo_update(res, lock,
1743 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1744 lock->l_blast_sent != 0) {
1745 time64_t delay = ktime_get_real_seconds() -
1748 "server cancels blocked lock after %llds",
1750 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1752 ldlm_lock_cancel(lock);
1753 LDLM_LOCK_PUT(lock);
1756 ldlm_reprocess_all(pres, NULL);
1757 LDLM_RESOURCE_DELREF(pres);
1758 ldlm_resource_putref(pres);
1760 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1763 EXPORT_SYMBOL(ldlm_request_cancel);
1766 * Main LDLM entry point for server code to cancel locks.
1768 * Typically gets called from service handler on LDLM_CANCEL opc.
1770 int ldlm_handle_cancel(struct ptlrpc_request *req)
1772 struct ldlm_request *dlm_req;
1777 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1778 if (dlm_req == NULL) {
1779 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1783 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1784 offsetof(struct ldlm_request, lock_handle[1]))
1787 if (req->rq_export && req->rq_export->exp_nid_stats &&
1788 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1789 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1790 LDLM_CANCEL - LDLM_FIRST_OPC);
1792 rc = req_capsule_server_pack(&req->rq_pill);
1796 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1797 req->rq_status = LUSTRE_ESTALE;
1799 RETURN(ptlrpc_reply(req));
1801 #endif /* HAVE_SERVER_SUPPORT */
1804 * Server may pass additional information about blocking lock.
1805 * For IBITS locks it is conflicting bits which can be used for
1806 * lock convert instead of cancel.
1808 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1810 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1812 check_res_locked(lock->l_resource);
1813 if (ns_is_client(ns) && ld &&
1814 (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->l_policy_data.l_inodebits.cancel_bits &&
1827 ldlm_res_eq(&ld->l_resource.lr_name,
1828 &lock->l_resource->lr_name) &&
1829 !(ldlm_is_cbpending(lock) &&
1830 lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1831 /* always combine conflicting ibits */
1832 lock->l_policy_data.l_inodebits.cancel_bits |=
1833 ld->l_policy_data.l_inodebits.cancel_bits;
1835 /* If cancel_bits are not obtained or
1836 * if the lock is already CBPENDING and
1837 * has no cancel_bits set
1838 * - the full lock is to be cancelled
1840 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1846 * Callback handler for receiving incoming blocking ASTs.
1848 * This can only happen on client side.
1850 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1851 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1857 LDLM_DEBUG(lock, "client blocking AST callback handler");
1859 lock_res_and_lock(lock);
1861 /* get extra information from desc if any */
1862 ldlm_bl_desc2lock(ld, lock);
1863 ldlm_set_cbpending(lock);
1865 do_ast = (!lock->l_readers && !lock->l_writers);
1866 unlock_res_and_lock(lock);
1870 "Lock %p already unused, calling callback (%p)\n",
1871 lock, lock->l_blocking_ast);
1872 if (lock->l_blocking_ast != NULL)
1873 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1877 "Lock %p is referenced, will be cancelled later\n",
1881 LDLM_DEBUG(lock, "client blocking callback handler END");
1882 LDLM_LOCK_RELEASE(lock);
1886 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1888 if (req->rq_no_reply)
1891 req->rq_status = rc;
1892 if (!req->rq_packed_final) {
1893 rc = lustre_pack_reply(req, 1, NULL, NULL);
1897 return ptlrpc_reply(req);
1901 * Callback handler for receiving incoming completion ASTs.
1903 * This only can happen on client side.
1905 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1906 struct ldlm_namespace *ns,
1907 struct ldlm_request *dlm_req,
1908 struct ldlm_lock *lock)
1910 LIST_HEAD(ast_list);
1916 LDLM_DEBUG(lock, "client completion callback handler START");
1918 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1919 long to = cfs_time_seconds(1);
1921 ldlm_callback_reply(req, 0);
1924 schedule_timeout_interruptible(to);
1925 if (ldlm_is_granted(lock) ||
1926 ldlm_is_destroyed(lock))
1931 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1933 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1934 GOTO(out, rc = lvb_len);
1935 } else if (lvb_len > 0) {
1936 if (lock->l_lvb_len > 0) {
1937 /* for extent lock, lvb contains ost_lvb{}. */
1938 LASSERT(lock->l_lvb_data != NULL);
1940 if (unlikely(lock->l_lvb_len < lvb_len)) {
1942 "Replied LVB is larger than expectation, expected = %d, replied = %d",
1943 lock->l_lvb_len, lvb_len);
1944 GOTO(out, rc = -EINVAL);
1949 lock_res_and_lock(lock);
1951 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1952 &lock->l_resource->lr_name)) {
1953 ldlm_resource_unlink_lock(lock);
1954 unlock_res_and_lock(lock);
1955 rc = ldlm_lock_change_resource(ns, lock,
1956 &dlm_req->lock_desc.l_resource.lr_name);
1958 LDLM_ERROR(lock, "Failed to allocate resource");
1961 LDLM_DEBUG(lock, "completion AST, new resource");
1962 lock_res_and_lock(lock);
1965 if (ldlm_is_failed(lock)) {
1966 unlock_res_and_lock(lock);
1967 LDLM_LOCK_RELEASE(lock);
1971 if (ldlm_is_destroyed(lock) ||
1972 ldlm_is_granted(lock)) {
1973 /* b=11300: the lock has already been granted */
1974 unlock_res_and_lock(lock);
1975 LDLM_DEBUG(lock, "Double grant race happened");
1980 * If we receive the completion AST before the actual enqueue returned,
1981 * then we might need to switch lock modes, resources, or extents.
1983 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1984 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1985 LDLM_DEBUG(lock, "completion AST, new lock mode");
1988 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1989 ldlm_convert_policy_to_local(req->rq_export,
1990 dlm_req->lock_desc.l_resource.lr_type,
1991 &dlm_req->lock_desc.l_policy_data,
1992 &lock->l_policy_data);
1993 LDLM_DEBUG(lock, "completion AST, new policy data");
1996 ldlm_resource_unlink_lock(lock);
1998 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2000 * BL_AST locks are not needed in LRU.
2001 * Let ldlm_cancel_lru() be fast.
2003 ldlm_lock_remove_from_lru(lock);
2004 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2005 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2006 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2009 if (lock->l_lvb_len > 0) {
2010 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2011 lock->l_lvb_data, lvb_len);
2013 unlock_res_and_lock(lock);
2018 ldlm_grant_lock(lock, &ast_list);
2019 unlock_res_and_lock(lock);
2021 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2024 * Let Enqueue to call osc_lock_upcall() and initialize
2027 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2029 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2031 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2037 lock_res_and_lock(lock);
2038 ldlm_set_failed(lock);
2039 unlock_res_and_lock(lock);
2040 wake_up(&lock->l_waitq);
2042 LDLM_LOCK_RELEASE(lock);
2048 * Callback handler for receiving incoming glimpse ASTs.
2050 * This only can happen on client side. After handling the glimpse AST
2051 * we also consider dropping the lock here if it is unused locally for a
2054 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2055 struct ldlm_namespace *ns,
2056 struct ldlm_request *dlm_req,
2057 struct ldlm_lock *lock)
2059 struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2064 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2066 if (lock->l_glimpse_ast != NULL)
2067 rc = lock->l_glimpse_ast(lock, req);
2069 if (req->rq_repmsg != NULL) {
2072 req->rq_status = rc;
2076 lock_res_and_lock(lock);
2077 if (lock->l_granted_mode == LCK_PW &&
2078 !lock->l_readers && !lock->l_writers &&
2079 ktime_after(ktime_get(),
2080 ktime_add(lock->l_last_used,
2081 ktime_set(ns->ns_dirty_age_limit, 0)))) {
2082 unlock_res_and_lock(lock);
2084 /* For MDS glimpse it is always DOM lock, set corresponding
2085 * cancel_bits to perform lock convert if needed
2087 if (lock->l_resource->lr_type == LDLM_IBITS)
2088 ld->l_policy_data.l_inodebits.cancel_bits =
2090 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2091 ldlm_handle_bl_callback(ns, ld, lock);
2096 unlock_res_and_lock(lock);
2097 LDLM_LOCK_RELEASE(lock);
2101 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2102 enum ldlm_cancel_flags cancel_flags)
2104 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2108 spin_lock(&blp->blp_lock);
2109 if (blwi->blwi_lock &&
2110 ldlm_is_discard_data(blwi->blwi_lock)) {
2111 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2112 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2114 /* other blocking callbacks are added to the regular list */
2115 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2117 spin_unlock(&blp->blp_lock);
2119 wake_up(&blp->blp_waitq);
2122 * can not check blwi->blwi_flags as blwi could be already freed in
2125 if (!(cancel_flags & LCF_ASYNC))
2126 wait_for_completion(&blwi->blwi_comp);
2131 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2132 struct ldlm_namespace *ns,
2133 struct ldlm_lock_desc *ld,
2134 struct list_head *cancels, int count,
2135 struct ldlm_lock *lock,
2136 enum ldlm_cancel_flags cancel_flags)
2138 init_completion(&blwi->blwi_comp);
2139 INIT_LIST_HEAD(&blwi->blwi_head);
2141 if (memory_pressure_get())
2142 blwi->blwi_mem_pressure = 1;
2145 blwi->blwi_flags = cancel_flags;
2147 blwi->blwi_ld = *ld;
2149 list_splice_init(cancels, &blwi->blwi_head);
2150 blwi->blwi_count = count;
2152 blwi->blwi_lock = lock;
2157 * Queues a list of locks \a cancels containing \a count locks
2158 * for later processing by a blocking thread. If \a count is zero,
2159 * then the lock referenced as \a lock is queued instead.
2161 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2162 * If list addition fails an error is returned and caller is supposed to
2163 * call ->l_blocking_ast itself.
2165 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2166 struct ldlm_lock_desc *ld,
2167 struct ldlm_lock *lock,
2168 struct list_head *cancels, int count,
2169 enum ldlm_cancel_flags cancel_flags)
2173 if (cancels && count == 0)
2176 if (cancel_flags & LCF_ASYNC) {
2177 struct ldlm_bl_work_item *blwi;
2179 OBD_ALLOC(blwi, sizeof(*blwi));
2182 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2184 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2187 * if it is synchronous call do minimum mem alloc, as it could
2188 * be triggered from kernel shrinker
2190 struct ldlm_bl_work_item blwi;
2192 memset(&blwi, 0, sizeof(blwi));
2193 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2194 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2199 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2200 struct ldlm_lock *lock)
2202 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2205 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2206 struct list_head *cancels, int count,
2207 enum ldlm_cancel_flags cancel_flags)
2209 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2212 int ldlm_bl_thread_wakeup(void)
2214 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2218 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2219 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2221 struct obd_device *obd = req->rq_export->exp_obd;
2229 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2231 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2233 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2235 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2238 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2240 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2242 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2245 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2248 /* We are responsible for swabbing contents of val */
2250 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2251 /* Pass it on to mdc (the "export" in this case) */
2252 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2254 sizeof(KEY_HSM_COPYTOOL_SEND),
2255 KEY_HSM_COPYTOOL_SEND,
2258 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2263 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2264 const char *msg, int rc,
2265 const struct lustre_handle *handle)
2267 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2268 "%s, NID=%s lock=%#llx: rc = %d",
2269 msg, libcfs_id2str(req->rq_peer),
2270 handle ? handle->cookie : 0, rc);
2271 if (req->rq_no_reply)
2272 CWARN("No reply was sent, maybe cause b=21636.\n");
2274 CWARN("Send reply failed, maybe cause b=21636.\n");
2277 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2278 static int ldlm_callback_handler(struct ptlrpc_request *req)
2280 struct ldlm_namespace *ns;
2281 struct ldlm_request *dlm_req;
2282 struct ldlm_lock *lock;
2288 * Requests arrive in sender's byte order. The ptlrpc service
2289 * handler has already checked and, if necessary, byte-swapped the
2290 * incoming request message body, but I am responsible for the
2294 /* do nothing for sec context finalize */
2295 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2298 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2300 if (req->rq_export == NULL) {
2301 rc = ldlm_callback_reply(req, -ENOTCONN);
2302 ldlm_callback_errmsg(req, "Operate on unconnected server",
2307 LASSERT(req->rq_export != NULL);
2308 LASSERT(req->rq_export->exp_obd != NULL);
2310 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2311 case LDLM_BL_CALLBACK:
2312 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2314 ldlm_callback_reply(req, -(int)cfs_fail_err);
2318 case LDLM_CP_CALLBACK:
2319 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2322 case LDLM_GL_CALLBACK:
2323 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2327 rc = ldlm_handle_setinfo(req);
2328 ldlm_callback_reply(req, rc);
2331 CERROR("unknown opcode %u\n",
2332 lustre_msg_get_opc(req->rq_reqmsg));
2333 ldlm_callback_reply(req, -EPROTO);
2337 ns = req->rq_export->exp_obd->obd_namespace;
2338 LASSERT(ns != NULL);
2340 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2342 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2343 if (dlm_req == NULL) {
2344 rc = ldlm_callback_reply(req, -EPROTO);
2345 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2351 * Force a known safe race, send a cancel to the server for a lock
2352 * which the server has already started a blocking callback on.
2354 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2355 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2356 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2358 CERROR("ldlm_cli_cancel: %d\n", rc);
2361 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2364 "callback on lock %#llx - lock disappeared\n",
2365 dlm_req->lock_handle[0].cookie);
2366 rc = ldlm_callback_reply(req, -EINVAL);
2367 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2368 &dlm_req->lock_handle[0]);
2372 if (ldlm_is_fail_loc(lock) &&
2373 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2374 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2376 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2377 lock_res_and_lock(lock);
2378 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2380 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2382 * If somebody cancels lock and cache is already dropped,
2383 * or lock is failed before cp_ast received on client,
2384 * we can tell the server we have no lock. Otherwise, we
2385 * should send cancel after dropping the cache.
2387 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2388 ldlm_is_failed(lock)) {
2390 "callback on lock %llx - lock disappeared",
2391 dlm_req->lock_handle[0].cookie);
2392 unlock_res_and_lock(lock);
2393 LDLM_LOCK_RELEASE(lock);
2394 rc = ldlm_callback_reply(req, -EINVAL);
2395 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2396 &dlm_req->lock_handle[0]);
2400 * BL_AST locks are not needed in LRU.
2401 * Let ldlm_cancel_lru() be fast.
2403 ldlm_lock_remove_from_lru(lock);
2404 ldlm_set_bl_ast(lock);
2406 unlock_res_and_lock(lock);
2409 * We want the ost thread to get this reply so that it can respond
2410 * to ost requests (write cache writeback) that might be triggered
2413 * But we'd also like to be able to indicate in the reply that we're
2414 * cancelling right now, because it's unused, or have an intent result
2415 * in the reply, so we might have to push the responsibility for sending
2416 * the reply down into the AST handlers, alas.
2419 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2420 case LDLM_BL_CALLBACK:
2421 CDEBUG(D_INODE, "blocking ast\n");
2422 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2423 if (!ldlm_is_cancel_on_block(lock)) {
2424 rc = ldlm_callback_reply(req, 0);
2425 if (req->rq_no_reply || rc)
2426 ldlm_callback_errmsg(req, "Normal process", rc,
2427 &dlm_req->lock_handle[0]);
2429 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2430 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2432 case LDLM_CP_CALLBACK:
2433 CDEBUG(D_INODE, "completion ast\n");
2434 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2435 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2436 if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2437 ldlm_callback_reply(req, rc);
2439 case LDLM_GL_CALLBACK:
2440 CDEBUG(D_INODE, "glimpse ast\n");
2441 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2442 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2445 LBUG(); /* checked above */
2451 #ifdef HAVE_SERVER_SUPPORT
2453 * Main handler for canceld thread.
2455 * Separated into its own thread to avoid deadlocks.
2457 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2464 * Requests arrive in sender's byte order. The ptlrpc service
2465 * handler has already checked and, if necessary, byte-swapped the
2466 * incoming request message body, but I am responsible for the
2470 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2472 if (req->rq_export == NULL) {
2473 struct ldlm_request *dlm_req;
2475 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2476 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2477 libcfs_nid2str(req->rq_peer.nid),
2478 (unsigned long long)req->rq_arrival_time.tv_sec,
2479 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2481 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2482 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2483 dlm_req = req_capsule_client_get(&req->rq_pill,
2485 if (dlm_req != NULL)
2486 ldlm_lock_dump_handle(D_ERROR,
2487 &dlm_req->lock_handle[0]);
2489 ldlm_callback_reply(req, -ENOTCONN);
2493 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2494 /* XXX FIXME move this back to mds/handler.c, b=249 */
2496 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2497 CDEBUG(D_INODE, "cancel\n");
2498 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2499 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2500 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2502 rc = ldlm_handle_cancel(req);
2506 struct ldlm_request *dlm_req;
2508 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2509 CDEBUG(D_INODE, "convert\n");
2511 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2512 if (dlm_req == NULL) {
2513 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2514 rc = ldlm_callback_reply(req, -EPROTO);
2516 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2517 rc = ptlrpc_reply(req);
2522 CERROR("invalid opcode %d\n",
2523 lustre_msg_get_opc(req->rq_reqmsg));
2524 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2525 rc = ldlm_callback_reply(req, -EINVAL);
2531 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2532 struct ldlm_lock *lock)
2534 struct ldlm_request *dlm_req;
2535 struct lustre_handle lockh;
2541 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2542 if (dlm_req == NULL)
2545 ldlm_lock2handle(lock, &lockh);
2546 for (i = 0; i < dlm_req->lock_count; i++) {
2547 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2549 DEBUG_REQ(D_RPCTRACE, req,
2550 "Prio raised by lock %#llx", lockh.cookie);
2559 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2561 struct ldlm_request *dlm_req;
2568 /* no prolong in recovery */
2569 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2572 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2573 if (dlm_req == NULL)
2576 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2577 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2578 (size - offsetof(struct ldlm_request, lock_handle)) /
2579 sizeof(struct lustre_handle) < dlm_req->lock_count)
2582 for (i = 0; i < dlm_req->lock_count; i++) {
2583 struct ldlm_lock *lock;
2585 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2589 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2591 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2592 LDLM_LOCK_PUT(lock);
2601 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2602 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2603 .hpreq_check = ldlm_cancel_hpreq_check,
2607 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2611 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2613 if (req->rq_export == NULL)
2616 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2617 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2618 req->rq_ops = &ldlm_cancel_hpreq_ops;
2619 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2620 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2621 req->rq_ops = &ldlm_cancel_hpreq_ops;
2626 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2627 struct hlist_node *hnode, void *data)
2630 struct list_head *rpc_list = data;
2631 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2633 lock_res_and_lock(lock);
2635 if (!ldlm_is_granted(lock)) {
2636 unlock_res_and_lock(lock);
2640 LASSERT(lock->l_resource);
2641 if (lock->l_resource->lr_type != LDLM_IBITS &&
2642 lock->l_resource->lr_type != LDLM_PLAIN) {
2643 unlock_res_and_lock(lock);
2647 if (ldlm_is_ast_sent(lock)) {
2648 unlock_res_and_lock(lock);
2652 LASSERT(lock->l_blocking_ast);
2653 LASSERT(!lock->l_blocking_lock);
2655 ldlm_set_ast_sent(lock);
2656 if (lock->l_export && lock->l_export->exp_lock_hash) {
2658 * NB: it's safe to call cfs_hash_del() even lock isn't
2662 * In the function below, .hs_keycmp resolves to
2663 * ldlm_export_lock_keycmp()
2665 /* coverity[overrun-buffer-val] */
2666 cfs_hash_del(lock->l_export->exp_lock_hash,
2667 &lock->l_remote_handle, &lock->l_exp_hash);
2670 list_add_tail(&lock->l_rk_ast, rpc_list);
2671 LDLM_LOCK_GET(lock);
2673 unlock_res_and_lock(lock);
2677 void ldlm_revoke_export_locks(struct obd_export *exp)
2679 LIST_HEAD(rpc_list);
2683 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2684 ldlm_revoke_lock_cb, &rpc_list, 0);
2685 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2686 LDLM_WORK_REVOKE_AST);
2690 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2691 #endif /* HAVE_SERVER_SUPPORT */
2693 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2694 struct ldlm_bl_work_item **p_blwi,
2695 struct obd_export **p_exp)
2697 struct ldlm_bl_work_item *blwi = NULL;
2698 static unsigned int num_bl;
2699 static unsigned int num_stale;
2700 int num_th = atomic_read(&blp->blp_num_threads);
2702 *p_exp = obd_stale_export_get();
2704 spin_lock(&blp->blp_lock);
2705 if (*p_exp != NULL) {
2706 if (num_th == 1 || ++num_stale < num_th) {
2707 spin_unlock(&blp->blp_lock);
2713 /* process a request from the blp_list at least every blp_num_threads */
2714 if (!list_empty(&blp->blp_list) &&
2715 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2716 blwi = list_entry(blp->blp_list.next,
2717 struct ldlm_bl_work_item, blwi_entry);
2719 if (!list_empty(&blp->blp_prio_list))
2720 blwi = list_entry(blp->blp_prio_list.next,
2721 struct ldlm_bl_work_item,
2725 if (++num_bl >= num_th)
2727 list_del(&blwi->blwi_entry);
2729 spin_unlock(&blp->blp_lock);
2732 if (*p_exp != NULL && *p_blwi != NULL) {
2733 obd_stale_export_put(*p_exp);
2737 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2740 /* This only contains temporary data until the thread starts */
2741 struct ldlm_bl_thread_data {
2742 struct ldlm_bl_pool *bltd_blp;
2743 struct completion bltd_comp;
2747 static int ldlm_bl_thread_main(void *arg);
2749 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2751 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2752 struct task_struct *task;
2754 init_completion(&bltd.bltd_comp);
2756 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2757 if (bltd.bltd_num >= blp->blp_max_threads) {
2758 atomic_dec(&blp->blp_num_threads);
2762 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2764 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2765 atomic_dec(&blp->blp_num_threads);
2769 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2772 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2773 bltd.bltd_num, PTR_ERR(task));
2774 atomic_dec(&blp->blp_num_threads);
2775 return PTR_ERR(task);
2777 wait_for_completion(&bltd.bltd_comp);
2782 /* Not fatal if racy and have a few too many threads */
2783 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2784 struct ldlm_bl_work_item *blwi)
2786 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2789 if (atomic_read(&blp->blp_busy_threads) <
2790 atomic_read(&blp->blp_num_threads))
2793 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2794 blwi->blwi_mem_pressure))
2800 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2801 struct ldlm_bl_work_item *blwi)
2805 if (blwi->blwi_ns == NULL)
2806 /* added by ldlm_cleanup() */
2807 RETURN(LDLM_ITER_STOP);
2809 if (blwi->blwi_mem_pressure)
2810 memory_pressure_set();
2812 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2814 if (blwi->blwi_count) {
2817 * The special case when we cancel locks in lru
2818 * asynchronously, we pass the list of locks here.
2819 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2820 * canceled locally yet.
2822 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2825 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2828 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2831 if (blwi->blwi_mem_pressure)
2832 memory_pressure_clr();
2834 if (blwi->blwi_flags & LCF_ASYNC)
2835 OBD_FREE(blwi, sizeof(*blwi));
2837 complete(&blwi->blwi_comp);
2843 * Cancel stale locks on export. Cancel blocked locks first.
2844 * If the given export has blocked locks, the next in the list may have
2845 * them too, thus cancel not blocked locks only if the current export has
2848 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2849 struct obd_export *exp)
2855 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2857 num = ldlm_export_cancel_blocked_locks(exp);
2859 ldlm_export_cancel_locks(exp);
2861 obd_stale_export_put(exp);
2868 * Main blocking requests processing thread.
2870 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2871 * This thread in the end ends up doing actual call to ->l_blocking_ast
2874 static int ldlm_bl_thread_main(void *arg)
2877 struct ldlm_bl_pool *blp;
2878 struct ldlm_bl_thread_data *bltd = arg;
2886 rc = lu_env_init(env, LCT_DT_THREAD);
2889 rc = lu_env_add(env);
2891 GOTO(out_env_fini, rc);
2893 blp = bltd->bltd_blp;
2895 complete(&bltd->bltd_comp);
2896 /* cannot use bltd after this, it is only on caller's stack */
2899 struct ldlm_bl_work_item *blwi = NULL;
2900 struct obd_export *exp = NULL;
2903 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2906 wait_event_idle_exclusive(blp->blp_waitq,
2907 ldlm_bl_get_work(blp, &blwi,
2909 atomic_inc(&blp->blp_busy_threads);
2911 if (ldlm_bl_thread_need_create(blp, blwi))
2912 /* discard the return value, we tried */
2913 ldlm_bl_thread_start(blp, true);
2916 rc = ldlm_bl_thread_exports(blp, exp);
2918 rc = ldlm_bl_thread_blwi(blp, blwi);
2920 atomic_dec(&blp->blp_busy_threads);
2922 if (rc == LDLM_ITER_STOP)
2926 * If there are many namespaces, we will not sleep waiting for
2927 * work, and must do a cond_resched to avoid holding the CPU
2933 atomic_dec(&blp->blp_num_threads);
2934 complete(&blp->blp_comp);
2945 static int ldlm_setup(void);
2946 static int ldlm_cleanup(void);
2948 int ldlm_get_ref(void)
2953 mutex_lock(&ldlm_ref_mutex);
2954 if (++ldlm_refcount == 1) {
2959 mutex_unlock(&ldlm_ref_mutex);
2964 void ldlm_put_ref(void)
2967 mutex_lock(&ldlm_ref_mutex);
2968 if (ldlm_refcount == 1) {
2969 int rc = ldlm_cleanup();
2972 CERROR("ldlm_cleanup failed: %d\n", rc);
2978 mutex_unlock(&ldlm_ref_mutex);
2984 * Export handle<->lock hash operations.
2987 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
2989 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2993 ldlm_export_lock_key(struct hlist_node *hnode)
2995 struct ldlm_lock *lock;
2997 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2998 return &lock->l_remote_handle;
3002 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3004 struct ldlm_lock *lock;
3006 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3007 lock->l_remote_handle = *(struct lustre_handle *)key;
3011 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3013 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3017 ldlm_export_lock_object(struct hlist_node *hnode)
3019 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3023 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3025 struct ldlm_lock *lock;
3027 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3028 LDLM_LOCK_GET(lock);
3032 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3034 struct ldlm_lock *lock;
3036 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3037 LDLM_LOCK_RELEASE(lock);
3040 static struct cfs_hash_ops ldlm_export_lock_ops = {
3041 .hs_hash = ldlm_export_lock_hash,
3042 .hs_key = ldlm_export_lock_key,
3043 .hs_keycmp = ldlm_export_lock_keycmp,
3044 .hs_keycpy = ldlm_export_lock_keycpy,
3045 .hs_object = ldlm_export_lock_object,
3046 .hs_get = ldlm_export_lock_get,
3047 .hs_put = ldlm_export_lock_put,
3048 .hs_put_locked = ldlm_export_lock_put,
3051 int ldlm_init_export(struct obd_export *exp)
3057 exp->exp_lock_hash =
3058 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3059 HASH_EXP_LOCK_CUR_BITS,
3060 HASH_EXP_LOCK_MAX_BITS,
3061 HASH_EXP_LOCK_BKT_BITS, 0,
3062 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3063 &ldlm_export_lock_ops,
3064 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3065 CFS_HASH_NBLK_CHANGE);
3067 if (!exp->exp_lock_hash)
3070 rc = ldlm_init_flock_export(exp);
3076 ldlm_destroy_export(exp);
3079 EXPORT_SYMBOL(ldlm_init_export);
3081 void ldlm_destroy_export(struct obd_export *exp)
3084 cfs_hash_putref(exp->exp_lock_hash);
3085 exp->exp_lock_hash = NULL;
3087 ldlm_destroy_flock_export(exp);
3090 EXPORT_SYMBOL(ldlm_destroy_export);
3092 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3093 struct attribute *attr,
3096 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3099 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3100 struct attribute *attr,
3107 rc = kstrtoul(buffer, 10, &val);
3111 ldlm_cancel_unused_locks_before_replay = val;
3115 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3117 static struct attribute *ldlm_attrs[] = {
3118 &lustre_attr_cancel_unused_locks_before_replay.attr,
3122 static struct attribute_group ldlm_attr_group = {
3123 .attrs = ldlm_attrs,
3126 static int ldlm_setup(void)
3128 static struct ptlrpc_service_conf conf;
3129 struct ldlm_bl_pool *blp = NULL;
3130 #ifdef HAVE_SERVER_SUPPORT
3131 struct task_struct *task;
3132 #endif /* HAVE_SERVER_SUPPORT */
3138 if (ldlm_state != NULL)
3141 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3142 if (ldlm_state == NULL)
3145 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3149 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3153 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3157 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3161 rc = ldlm_debugfs_setup();
3165 memset(&conf, 0, sizeof(conf));
3166 conf = (typeof(conf)) {
3167 .psc_name = "ldlm_cbd",
3168 .psc_watchdog_factor = 2,
3170 .bc_nbufs = LDLM_CLIENT_NBUFS,
3171 .bc_buf_size = LDLM_BUFSIZE,
3172 .bc_req_max_size = LDLM_MAXREQSIZE,
3173 .bc_rep_max_size = LDLM_MAXREPSIZE,
3174 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3175 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3178 .tc_thr_name = "ldlm_cb",
3179 .tc_thr_factor = LDLM_THR_FACTOR,
3180 .tc_nthrs_init = LDLM_NTHRS_INIT,
3181 .tc_nthrs_base = LDLM_NTHRS_BASE,
3182 .tc_nthrs_max = LDLM_NTHRS_MAX,
3183 .tc_nthrs_user = ldlm_num_threads,
3184 .tc_cpu_bind = ldlm_cpu_bind,
3185 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3188 .cc_pattern = ldlm_cpts,
3189 .cc_affinity = true,
3192 .so_req_handler = ldlm_callback_handler,
3195 ldlm_state->ldlm_cb_service = \
3196 ptlrpc_register_service(&conf, ldlm_svc_kset,
3197 ldlm_svc_debugfs_dir);
3198 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3199 CERROR("failed to start service\n");
3200 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3201 ldlm_state->ldlm_cb_service = NULL;
3205 #ifdef HAVE_SERVER_SUPPORT
3206 memset(&conf, 0, sizeof(conf));
3207 conf = (typeof(conf)) {
3208 .psc_name = "ldlm_canceld",
3209 .psc_watchdog_factor = 6,
3211 .bc_nbufs = LDLM_SERVER_NBUFS,
3212 .bc_buf_size = LDLM_BUFSIZE,
3213 .bc_req_max_size = LDLM_MAXREQSIZE,
3214 .bc_rep_max_size = LDLM_MAXREPSIZE,
3215 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3216 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3220 .tc_thr_name = "ldlm_cn",
3221 .tc_thr_factor = LDLM_THR_FACTOR,
3222 .tc_nthrs_init = LDLM_NTHRS_INIT,
3223 .tc_nthrs_base = LDLM_NTHRS_BASE,
3224 .tc_nthrs_max = LDLM_NTHRS_MAX,
3225 .tc_nthrs_user = ldlm_num_threads,
3226 .tc_cpu_bind = ldlm_cpu_bind,
3227 .tc_ctx_tags = LCT_MD_THREAD | \
3232 .cc_pattern = ldlm_cpts,
3233 .cc_affinity = true,
3236 .so_req_handler = ldlm_cancel_handler,
3237 .so_hpreq_handler = ldlm_hpreq_handler,
3240 ldlm_state->ldlm_cancel_service = \
3241 ptlrpc_register_service(&conf, ldlm_svc_kset,
3242 ldlm_svc_debugfs_dir);
3243 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3244 CERROR("failed to start service\n");
3245 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3246 ldlm_state->ldlm_cancel_service = NULL;
3249 #endif /* HAVE_SERVER_SUPPORT */
3251 OBD_ALLOC(blp, sizeof(*blp));
3253 GOTO(out, rc = -ENOMEM);
3254 ldlm_state->ldlm_bl_pool = blp;
3256 spin_lock_init(&blp->blp_lock);
3257 INIT_LIST_HEAD(&blp->blp_list);
3258 INIT_LIST_HEAD(&blp->blp_prio_list);
3259 init_waitqueue_head(&blp->blp_waitq);
3260 atomic_set(&blp->blp_num_threads, 0);
3261 atomic_set(&blp->blp_busy_threads, 0);
3263 if (ldlm_num_threads == 0) {
3264 blp->blp_min_threads = LDLM_NTHRS_INIT;
3265 blp->blp_max_threads = LDLM_NTHRS_MAX;
3267 blp->blp_min_threads = blp->blp_max_threads = \
3268 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3272 for (i = 0; i < blp->blp_min_threads; i++) {
3273 rc = ldlm_bl_thread_start(blp, false);
3278 #ifdef HAVE_SERVER_SUPPORT
3279 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3282 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3286 wait_event(expired_lock_wait_queue,
3287 expired_lock_thread_state == ELT_READY);
3288 #endif /* HAVE_SERVER_SUPPORT */
3290 rc = ldlm_pools_init();
3292 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3296 rc = ldlm_reclaim_setup();
3298 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3308 static int ldlm_cleanup(void)
3312 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3313 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3314 CERROR("ldlm still has namespaces; clean these up first.\n");
3315 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3316 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3320 ldlm_reclaim_cleanup();
3323 if (ldlm_state->ldlm_bl_pool != NULL) {
3324 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3326 while (atomic_read(&blp->blp_num_threads) > 0) {
3327 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3329 init_completion(&blp->blp_comp);
3331 spin_lock(&blp->blp_lock);
3332 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3333 wake_up(&blp->blp_waitq);
3334 spin_unlock(&blp->blp_lock);
3336 wait_for_completion(&blp->blp_comp);
3339 OBD_FREE(blp, sizeof(*blp));
3342 if (ldlm_state->ldlm_cb_service != NULL)
3343 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3344 #ifdef HAVE_SERVER_SUPPORT
3345 if (ldlm_state->ldlm_cancel_service != NULL)
3346 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3350 kset_unregister(ldlm_ns_kset);
3352 kset_unregister(ldlm_svc_kset);
3354 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3355 kobject_put(ldlm_kobj);
3358 ldlm_debugfs_cleanup();
3360 #ifdef HAVE_SERVER_SUPPORT
3361 if (expired_lock_thread_state != ELT_STOPPED) {
3362 expired_lock_thread_state = ELT_TERMINATE;
3363 wake_up(&expired_lock_wait_queue);
3364 wait_event(expired_lock_wait_queue,
3365 expired_lock_thread_state == ELT_STOPPED);
3369 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3377 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3378 sizeof(struct ldlm_resource), 0,
3379 SLAB_HWCACHE_ALIGN, NULL);
3380 if (ldlm_resource_slab == NULL)
3383 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3384 sizeof(struct ldlm_lock), 0,
3385 SLAB_HWCACHE_ALIGN, NULL);
3386 if (ldlm_lock_slab == NULL)
3389 ldlm_interval_slab = kmem_cache_create("interval_node",
3390 sizeof(struct ldlm_interval),
3391 0, SLAB_HWCACHE_ALIGN, NULL);
3392 if (ldlm_interval_slab == NULL)
3395 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3396 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3397 0, SLAB_HWCACHE_ALIGN, NULL);
3398 if (ldlm_interval_tree_slab == NULL)
3401 #ifdef HAVE_SERVER_SUPPORT
3402 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3403 sizeof(struct ldlm_ibits_node),
3404 0, SLAB_HWCACHE_ALIGN, NULL);
3405 if (ldlm_inodebits_slab == NULL)
3406 goto out_interval_tree;
3408 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3409 sizeof(struct ldlm_glimpse_work),
3411 if (ldlm_glimpse_work_kmem == NULL)
3415 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3416 class_export_dump_hook = ldlm_dump_export_locks;
3419 #ifdef HAVE_SERVER_SUPPORT
3421 kmem_cache_destroy(ldlm_inodebits_slab);
3423 kmem_cache_destroy(ldlm_interval_tree_slab);
3426 kmem_cache_destroy(ldlm_interval_slab);
3428 kmem_cache_destroy(ldlm_lock_slab);
3430 kmem_cache_destroy(ldlm_resource_slab);
3435 void ldlm_exit(void)
3438 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3439 kmem_cache_destroy(ldlm_resource_slab);
3441 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3442 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3443 * so that ldlm_lock_free() get a chance to be called.
3446 kmem_cache_destroy(ldlm_lock_slab);
3447 kmem_cache_destroy(ldlm_interval_slab);
3448 kmem_cache_destroy(ldlm_interval_tree_slab);
3449 #ifdef HAVE_SERVER_SUPPORT
3450 kmem_cache_destroy(ldlm_inodebits_slab);
3451 kmem_cache_destroy(ldlm_glimpse_work_kmem);