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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_lockd.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LDLM
44 #include <linux/kthread.h>
45 #include <libcfs/libcfs.h>
46 #include <lustre_dlm.h>
47 #include <obd_class.h>
48 #include <libcfs/list.h>
49 #include "ldlm_internal.h"
51 static int ldlm_num_threads;
52 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
53 "number of DLM service threads to start");
55 static char *ldlm_cpts;
56 CFS_MODULE_PARM(ldlm_cpts, "s", charp, 0444,
57 "CPU partitions ldlm threads should run on");
59 static struct mutex ldlm_ref_mutex;
60 static int ldlm_refcount;
62 struct ldlm_cb_async_args {
63 struct ldlm_cb_set_arg *ca_set_arg;
64 struct ldlm_lock *ca_lock;
69 static struct ldlm_state *ldlm_state;
71 inline cfs_time_t round_timeout(cfs_time_t timeout)
73 return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
76 /* timeout for initial callback (AST) reply (bz10399) */
77 static inline unsigned int ldlm_get_rq_timeout(void)
80 unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
82 return timeout < 1 ? 1 : timeout;
87 #define ELT_TERMINATE 2
93 * blp_prio_list is used for callbacks that should be handled
94 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
97 struct list_head blp_prio_list;
100 * blp_list is used for all other callbacks which are likely
101 * to take longer to process.
103 struct list_head blp_list;
105 wait_queue_head_t blp_waitq;
106 struct completion blp_comp;
107 atomic_t blp_num_threads;
108 atomic_t blp_busy_threads;
113 struct ldlm_bl_work_item {
114 struct list_head blwi_entry;
115 struct ldlm_namespace *blwi_ns;
116 struct ldlm_lock_desc blwi_ld;
117 struct ldlm_lock *blwi_lock;
118 struct list_head blwi_head;
120 struct completion blwi_comp;
121 ldlm_cancel_flags_t blwi_flags;
122 int blwi_mem_pressure;
125 #ifdef HAVE_SERVER_SUPPORT
128 * Protects both waiting_locks_list and expired_lock_thread.
130 static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */
133 * List for contended locks.
135 * As soon as a lock is contended, it gets placed on this list and
136 * expected time to get a response is filled in the lock. A special
137 * thread walks the list looking for locks that should be released and
138 * schedules client evictions for those that have not been released in
141 * All access to it should be under waiting_locks_spinlock.
143 static struct list_head waiting_locks_list;
144 static struct timer_list waiting_locks_timer;
146 static struct expired_lock_thread {
147 wait_queue_head_t elt_waitq;
150 struct list_head elt_expired_locks;
151 } expired_lock_thread;
153 static inline int have_expired_locks(void)
158 spin_lock_bh(&waiting_locks_spinlock);
159 need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
160 spin_unlock_bh(&waiting_locks_spinlock);
166 * Check expired lock list for expired locks and time them out.
168 static int expired_lock_main(void *arg)
170 struct list_head *expired = &expired_lock_thread.elt_expired_locks;
171 struct l_wait_info lwi = { 0 };
176 expired_lock_thread.elt_state = ELT_READY;
177 wake_up(&expired_lock_thread.elt_waitq);
180 l_wait_event(expired_lock_thread.elt_waitq,
181 have_expired_locks() ||
182 expired_lock_thread.elt_state == ELT_TERMINATE,
185 spin_lock_bh(&waiting_locks_spinlock);
186 if (expired_lock_thread.elt_dump) {
187 struct libcfs_debug_msg_data msgdata = {
188 .msg_file = __FILE__,
189 .msg_fn = "waiting_locks_callback",
190 .msg_line = expired_lock_thread.elt_dump };
191 spin_unlock_bh(&waiting_locks_spinlock);
193 /* from waiting_locks_callback, but not in timer */
194 libcfs_debug_dumplog();
195 libcfs_run_lbug_upcall(&msgdata);
197 spin_lock_bh(&waiting_locks_spinlock);
198 expired_lock_thread.elt_dump = 0;
203 while (!list_empty(expired)) {
204 struct obd_export *export;
205 struct ldlm_lock *lock;
207 lock = list_entry(expired->next, struct ldlm_lock,
209 if ((void *)lock < LP_POISON + PAGE_CACHE_SIZE &&
210 (void *)lock >= LP_POISON) {
211 spin_unlock_bh(&waiting_locks_spinlock);
212 CERROR("free lock on elt list %p\n", lock);
215 list_del_init(&lock->l_pending_chain);
216 if ((void *)lock->l_export <
217 LP_POISON + PAGE_CACHE_SIZE &&
218 (void *)lock->l_export >= LP_POISON) {
219 CERROR("lock with free export on elt list %p\n",
221 lock->l_export = NULL;
222 LDLM_ERROR(lock, "free export");
223 /* release extra ref grabbed by
224 * ldlm_add_waiting_lock() or
225 * ldlm_failed_ast() */
226 LDLM_LOCK_RELEASE(lock);
230 if (ldlm_is_destroyed(lock)) {
231 /* release the lock refcount where
232 * waiting_locks_callback() founds */
233 LDLM_LOCK_RELEASE(lock);
236 export = class_export_lock_get(lock->l_export, lock);
237 spin_unlock_bh(&waiting_locks_spinlock);
239 spin_lock_bh(&export->exp_bl_list_lock);
240 list_del_init(&lock->l_exp_list);
241 spin_unlock_bh(&export->exp_bl_list_lock);
244 class_fail_export(export);
245 class_export_lock_put(export, lock);
247 /* release extra ref grabbed by ldlm_add_waiting_lock()
248 * or ldlm_failed_ast() */
249 LDLM_LOCK_RELEASE(lock);
251 spin_lock_bh(&waiting_locks_spinlock);
253 spin_unlock_bh(&waiting_locks_spinlock);
255 if (do_dump && obd_dump_on_eviction) {
256 CERROR("dump the log upon eviction\n");
257 libcfs_debug_dumplog();
260 if (expired_lock_thread.elt_state == ELT_TERMINATE)
264 expired_lock_thread.elt_state = ELT_STOPPED;
265 wake_up(&expired_lock_thread.elt_waitq);
269 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
270 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds);
273 * Check if there is a request in the export request list
274 * which prevents the lock canceling.
276 static int ldlm_lock_busy(struct ldlm_lock *lock)
278 struct ptlrpc_request *req;
282 if (lock->l_export == NULL)
285 spin_lock_bh(&lock->l_export->exp_rpc_lock);
286 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
288 if (req->rq_ops->hpreq_lock_match) {
289 match = req->rq_ops->hpreq_lock_match(req, lock);
294 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
298 /* This is called from within a timer interrupt and cannot schedule */
299 static void waiting_locks_callback(unsigned long unused)
301 struct ldlm_lock *lock;
304 spin_lock_bh(&waiting_locks_spinlock);
305 while (!list_empty(&waiting_locks_list)) {
306 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
308 if (cfs_time_after(lock->l_callback_timeout,
309 cfs_time_current()) ||
310 (lock->l_req_mode == LCK_GROUP))
313 /* Check if we need to prolong timeout */
314 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
315 ldlm_lock_busy(lock)) {
318 if (lock->l_pending_chain.next == &waiting_locks_list)
323 spin_unlock_bh(&waiting_locks_spinlock);
324 LDLM_DEBUG(lock, "prolong the busy lock");
325 ldlm_refresh_waiting_lock(lock,
326 ldlm_bl_timeout(lock) >> 1);
327 spin_lock_bh(&waiting_locks_spinlock);
330 LDLM_LOCK_RELEASE(lock);
334 LDLM_LOCK_RELEASE(lock);
337 ldlm_lock_to_ns(lock)->ns_timeouts++;
338 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
339 "evicting client at %s ",
340 cfs_time_current_sec() - lock->l_last_activity,
342 lock->l_export->exp_connection->c_peer.nid));
344 /* no needs to take an extra ref on the lock since it was in
345 * the waiting_locks_list and ldlm_add_waiting_lock()
346 * already grabbed a ref */
347 list_del(&lock->l_pending_chain);
348 list_add(&lock->l_pending_chain,
349 &expired_lock_thread.elt_expired_locks);
353 if (!list_empty(&expired_lock_thread.elt_expired_locks)) {
354 if (obd_dump_on_timeout && need_dump)
355 expired_lock_thread.elt_dump = __LINE__;
357 wake_up(&expired_lock_thread.elt_waitq);
361 * Make sure the timer will fire again if we have any locks
364 if (!list_empty(&waiting_locks_list)) {
365 cfs_time_t timeout_rounded;
366 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
368 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
369 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
371 spin_unlock_bh(&waiting_locks_spinlock);
375 * Add lock to the list of contended locks.
377 * Indicate that we're waiting for a client to call us back cancelling a given
378 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
379 * timer to fire appropriately. (We round up to the next second, to avoid
380 * floods of timer firings during periods of high lock contention and traffic).
381 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
382 * if it has been added to the waiting list (1 is returned).
384 * Called with the namespace lock held.
386 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
389 cfs_time_t timeout_rounded;
391 if (!list_empty(&lock->l_pending_chain))
394 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
395 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
398 timeout = cfs_time_shift(seconds);
399 if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
400 lock->l_callback_timeout = timeout;
402 timeout_rounded = round_timeout(lock->l_callback_timeout);
404 if (cfs_time_before(timeout_rounded,
405 cfs_timer_deadline(&waiting_locks_timer)) ||
406 !cfs_timer_is_armed(&waiting_locks_timer)) {
407 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
409 /* if the new lock has a shorter timeout than something earlier on
410 the list, we'll wait the longer amount of time; no big deal. */
412 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
416 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
419 int timeout = ldlm_bl_timeout(lock);
421 /* NB: must be called with hold of lock_res_and_lock() */
422 LASSERT(ldlm_is_res_locked(lock));
423 LASSERT(!ldlm_is_cancel_on_block(lock));
425 spin_lock_bh(&waiting_locks_spinlock);
426 if (ldlm_is_cancel(lock)) {
427 spin_unlock_bh(&waiting_locks_spinlock);
431 if (ldlm_is_destroyed(lock)) {
432 static cfs_time_t next;
434 spin_unlock_bh(&waiting_locks_spinlock);
435 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
436 if (cfs_time_after(cfs_time_current(), next)) {
437 next = cfs_time_shift(14400);
438 libcfs_debug_dumpstack(NULL);
443 ldlm_set_waited(lock);
444 lock->l_last_activity = cfs_time_current_sec();
445 ret = __ldlm_add_waiting_lock(lock, timeout);
447 /* grab ref on the lock if it has been added to the
451 spin_unlock_bh(&waiting_locks_spinlock);
454 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
455 if (list_empty(&lock->l_exp_list))
456 list_add(&lock->l_exp_list,
457 &lock->l_export->exp_bl_list);
458 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
461 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
462 ret == 0 ? "not re-" : "", timeout,
463 AT_OFF ? "off" : "on");
468 * Remove a lock from the pending list, likely because it had its cancellation
469 * callback arrive without incident. This adjusts the lock-timeout timer if
470 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
471 * As done by ldlm_del_waiting_lock(), the caller must release the lock
472 * reference when the lock is removed from any list (1 is returned).
474 * Called with namespace lock held.
476 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
478 struct list_head *list_next;
480 if (list_empty(&lock->l_pending_chain))
483 list_next = lock->l_pending_chain.next;
484 if (lock->l_pending_chain.prev == &waiting_locks_list) {
485 /* Removing the head of the list, adjust timer. */
486 if (list_next == &waiting_locks_list) {
487 /* No more, just cancel. */
488 cfs_timer_disarm(&waiting_locks_timer);
490 struct ldlm_lock *next;
491 next = list_entry(list_next, struct ldlm_lock,
493 cfs_timer_arm(&waiting_locks_timer,
494 round_timeout(next->l_callback_timeout));
497 list_del_init(&lock->l_pending_chain);
502 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
506 if (lock->l_export == NULL) {
507 /* We don't have a "waiting locks list" on clients. */
508 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
512 spin_lock_bh(&waiting_locks_spinlock);
513 ret = __ldlm_del_waiting_lock(lock);
514 ldlm_clear_waited(lock);
515 spin_unlock_bh(&waiting_locks_spinlock);
517 /* remove the lock out of export blocking list */
518 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
519 list_del_init(&lock->l_exp_list);
520 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
523 /* release lock ref if it has indeed been removed
525 LDLM_LOCK_RELEASE(lock);
528 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
533 * Prolong the contended lock waiting time.
535 * Called with namespace lock held.
537 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
539 if (lock->l_export == NULL) {
540 /* We don't have a "waiting locks list" on clients. */
541 LDLM_DEBUG(lock, "client lock: no-op");
545 spin_lock_bh(&waiting_locks_spinlock);
547 if (list_empty(&lock->l_pending_chain)) {
548 spin_unlock_bh(&waiting_locks_spinlock);
549 LDLM_DEBUG(lock, "wasn't waiting");
553 /* we remove/add the lock to the waiting list, so no needs to
554 * release/take a lock reference */
555 __ldlm_del_waiting_lock(lock);
556 __ldlm_add_waiting_lock(lock, timeout);
557 spin_unlock_bh(&waiting_locks_spinlock);
559 LDLM_DEBUG(lock, "refreshed");
562 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
564 #else /* HAVE_SERVER_SUPPORT */
566 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
571 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
576 #endif /* !HAVE_SERVER_SUPPORT */
578 #ifdef HAVE_SERVER_SUPPORT
581 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
582 * lock cancel, and their replies). Used for lock callback timeout and AST
585 * \param[in] lock lock which is getting the blocking callback
587 * \retval timeout in seconds to wait for the client reply
589 unsigned int ldlm_bl_timeout(struct ldlm_lock *lock)
591 unsigned int timeout;
594 return obd_timeout / 2;
596 /* Since these are non-updating timeouts, we should be conservative.
597 * Take more than usually, 150%
598 * It would be nice to have some kind of "early reply" mechanism for
599 * lock callbacks too... */
600 timeout = at_get(&lock->l_export->exp_bl_lock_at);
601 return max(timeout + (timeout >> 1), ldlm_enqueue_min);
603 EXPORT_SYMBOL(ldlm_bl_timeout);
606 * Perform lock cleanup if AST sending failed.
608 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
609 const char *ast_type)
611 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
612 "to a lock %s callback time out: rc %d\n",
613 lock->l_export->exp_obd->obd_name,
614 obd_export_nid2str(lock->l_export), ast_type, rc);
616 if (obd_dump_on_timeout)
617 libcfs_debug_dumplog();
618 spin_lock_bh(&waiting_locks_spinlock);
619 if (__ldlm_del_waiting_lock(lock) == 0)
620 /* the lock was not in any list, grab an extra ref before adding
621 * the lock to the expired list */
623 list_add(&lock->l_pending_chain,
624 &expired_lock_thread.elt_expired_locks);
625 wake_up(&expired_lock_thread.elt_waitq);
626 spin_unlock_bh(&waiting_locks_spinlock);
630 * Perform lock cleanup if AST reply came with error.
632 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
633 struct ptlrpc_request *req, int rc,
634 const char *ast_type)
636 lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
638 if (!req->rq_replied || (rc && rc != -EINVAL)) {
639 if (lock->l_export && lock->l_export->exp_libclient) {
640 LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
641 " timeout, just cancelling lock", ast_type,
642 libcfs_nid2str(peer.nid));
643 ldlm_lock_cancel(lock);
645 } else if (ldlm_is_cancel(lock)) {
646 LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
647 "cancel was received (AST reply lost?)",
648 ast_type, libcfs_nid2str(peer.nid));
649 ldlm_lock_cancel(lock);
652 LDLM_ERROR(lock, "client (nid %s) %s %s AST "
653 "(req status %d rc %d), evict it",
654 libcfs_nid2str(peer.nid),
655 req->rq_replied ? "returned error from" :
656 "failed to reply to",
658 (req->rq_repmsg != NULL) ?
659 lustre_msg_get_status(req->rq_repmsg) : 0,
661 ldlm_failed_ast(lock, rc, ast_type);
667 struct ldlm_resource *res = lock->l_resource;
669 LDLM_DEBUG(lock, "client (nid %s) returned %d"
670 " from %s AST - normal race",
671 libcfs_nid2str(peer.nid),
673 lustre_msg_get_status(req->rq_repmsg) : -1,
676 /* update lvbo to return proper attributes.
678 ldlm_resource_getref(res);
679 ldlm_res_lvbo_update(res, NULL, 1);
680 ldlm_resource_putref(res);
682 ldlm_lock_cancel(lock);
689 static int ldlm_cb_interpret(const struct lu_env *env,
690 struct ptlrpc_request *req, void *data, int rc)
692 struct ldlm_cb_async_args *ca = data;
693 struct ldlm_lock *lock = ca->ca_lock;
694 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
697 LASSERT(lock != NULL);
700 case LDLM_GL_CALLBACK:
701 /* Update the LVB from disk if the AST failed
702 * (this is a legal race)
704 * - Glimpse callback of local lock just returns
705 * -ELDLM_NO_LOCK_DATA.
706 * - Glimpse callback of remote lock might return
707 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
709 if (rc == -ELDLM_NO_LOCK_DATA) {
710 LDLM_DEBUG(lock, "lost race - client has a lock but no "
712 ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
713 } else if (rc != 0) {
714 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
716 rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
719 case LDLM_BL_CALLBACK:
721 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
723 case LDLM_CP_CALLBACK:
725 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
728 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
733 /* release extra reference taken in ldlm_ast_fini() */
734 LDLM_LOCK_RELEASE(lock);
737 atomic_inc(&arg->restart);
742 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
744 struct ldlm_cb_async_args *ca = data;
745 struct ldlm_lock *lock = ca->ca_lock;
747 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
750 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
751 struct ldlm_cb_set_arg *arg,
752 struct ldlm_lock *lock,
758 if (unlikely(instant_cancel)) {
759 rc = ptl_send_rpc(req, 1);
760 ptlrpc_req_finished(req);
762 atomic_inc(&arg->restart);
765 ptlrpc_set_add_req(arg->set, req);
772 * Check if there are requests in the export request list which prevent
773 * the lock canceling and make these requests high priority ones.
775 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
777 struct ptlrpc_request *req;
780 if (lock->l_export == NULL) {
781 LDLM_DEBUG(lock, "client lock: no-op");
785 spin_lock_bh(&lock->l_export->exp_rpc_lock);
786 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
788 /* Do not process requests that were not yet added to there
789 * incoming queue or were already removed from there for
790 * processing. We evaluate ptlrpc_nrs_req_can_move() without
791 * holding svcpt->scp_req_lock, and then redo the check with
792 * the lock held once we need to obtain a reliable result.
794 if (ptlrpc_nrs_req_can_move(req) &&
795 req->rq_ops->hpreq_lock_match &&
796 req->rq_ops->hpreq_lock_match(req, lock))
797 ptlrpc_nrs_req_hp_move(req);
799 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
804 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
805 * enqueued server lock conflicts with given one.
807 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
808 * to wait for client response.
810 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
811 struct ldlm_lock_desc *desc,
812 void *data, int flag)
814 struct ldlm_cb_async_args *ca;
815 struct ldlm_cb_set_arg *arg = data;
816 struct ldlm_request *body;
817 struct ptlrpc_request *req;
818 int instant_cancel = 0;
822 if (flag == LDLM_CB_CANCELING)
823 /* Don't need to do anything here. */
827 LASSERT(data != NULL);
828 if (lock->l_export->exp_obd->obd_recovering != 0)
829 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
831 ldlm_lock_reorder_req(lock);
833 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
834 &RQF_LDLM_BL_CALLBACK,
835 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
839 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
840 ca = ptlrpc_req_async_args(req);
841 ca->ca_set_arg = arg;
844 req->rq_interpret_reply = ldlm_cb_interpret;
846 lock_res_and_lock(lock);
847 if (lock->l_granted_mode != lock->l_req_mode) {
848 /* this blocking AST will be communicated as part of the
849 * completion AST instead */
850 unlock_res_and_lock(lock);
852 ptlrpc_req_finished(req);
853 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
857 if (ldlm_is_destroyed(lock)) {
858 /* What's the point? */
859 unlock_res_and_lock(lock);
860 ptlrpc_req_finished(req);
864 if (ldlm_is_cancel_on_block(lock))
867 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
868 body->lock_handle[0] = lock->l_remote_handle;
869 body->lock_desc = *desc;
870 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
872 LDLM_DEBUG(lock, "server preparing blocking AST");
874 ptlrpc_request_set_replen(req);
875 if (instant_cancel) {
876 unlock_res_and_lock(lock);
877 ldlm_lock_cancel(lock);
879 req->rq_no_resend = 1;
881 LASSERT(lock->l_granted_mode == lock->l_req_mode);
882 ldlm_add_waiting_lock(lock);
883 unlock_res_and_lock(lock);
885 /* Do not resend after lock callback timeout */
886 req->rq_delay_limit = ldlm_bl_timeout(lock);
887 req->rq_resend_cb = ldlm_update_resend;
890 req->rq_send_state = LUSTRE_IMP_FULL;
891 /* ptlrpc_request_alloc_pack already set timeout */
893 req->rq_timeout = ldlm_get_rq_timeout();
895 lock->l_last_activity = cfs_time_current_sec();
897 if (lock->l_export && lock->l_export->exp_nid_stats &&
898 lock->l_export->exp_nid_stats->nid_ldlm_stats)
899 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
900 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
902 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
908 * ->l_completion_ast callback for a remote lock in server namespace.
910 * Sends AST to the client notifying it of lock granting. If initial
911 * lock response was not sent yet, instead of sending another RPC, just
912 * mark the lock as granted and client will understand
914 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
916 struct ldlm_cb_set_arg *arg = data;
917 struct ldlm_request *body;
918 struct ptlrpc_request *req;
919 struct ldlm_cb_async_args *ca;
920 int instant_cancel = 0;
925 LASSERT(lock != NULL);
926 LASSERT(data != NULL);
928 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
929 LDLM_DEBUG(lock, "dropping CP AST");
933 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
934 &RQF_LDLM_CP_CALLBACK);
938 /* server namespace, doesn't need lock */
939 lvb_len = ldlm_lvbo_size(lock);
940 /* LU-3124 & LU-2187: to not return layout in completion AST because
941 * it may deadlock for LU-2187, or client may not have enough space
942 * for large layout. The layout will be returned to client with an
943 * extra RPC to fetch xattr.lov */
944 if (ldlm_has_layout(lock))
947 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
948 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
950 ptlrpc_request_free(req);
954 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
955 ca = ptlrpc_req_async_args(req);
956 ca->ca_set_arg = arg;
959 req->rq_interpret_reply = ldlm_cb_interpret;
960 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
962 body->lock_handle[0] = lock->l_remote_handle;
963 body->lock_flags = ldlm_flags_to_wire(flags);
964 ldlm_lock2desc(lock, &body->lock_desc);
966 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
968 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
970 /* We still need to send the RPC to wake up the blocked
971 * enqueue thread on the client.
973 * Consider old client, there is no better way to notify
974 * the failure, just zero-sized the LVB, then the client
975 * will fail out as "-EPROTO". */
976 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
980 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
985 lock->l_last_activity = cfs_time_current_sec();
987 LDLM_DEBUG(lock, "server preparing completion AST");
989 ptlrpc_request_set_replen(req);
991 req->rq_send_state = LUSTRE_IMP_FULL;
992 /* ptlrpc_request_pack already set timeout */
994 req->rq_timeout = ldlm_get_rq_timeout();
996 /* We only send real blocking ASTs after the lock is granted */
997 lock_res_and_lock(lock);
998 if (ldlm_is_ast_sent(lock)) {
999 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1000 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1001 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1004 /* We might get here prior to ldlm_handle_enqueue setting
1005 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1006 * into waiting list, but this is safe and similar code in
1007 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1008 * that would not only cancel the lock, but will also remove
1009 * it from waiting list */
1010 if (ldlm_is_cancel_on_block(lock)) {
1011 unlock_res_and_lock(lock);
1012 ldlm_lock_cancel(lock);
1015 req->rq_no_resend = 1;
1017 lock_res_and_lock(lock);
1019 /* start the lock-timeout clock */
1020 ldlm_add_waiting_lock(lock);
1021 /* Do not resend after lock callback timeout */
1022 req->rq_delay_limit = ldlm_bl_timeout(lock);
1023 req->rq_resend_cb = ldlm_update_resend;
1026 unlock_res_and_lock(lock);
1028 if (lock->l_export && lock->l_export->exp_nid_stats &&
1029 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1030 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1031 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1033 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1035 RETURN(lvb_len < 0 ? lvb_len : rc);
1039 * Server side ->l_glimpse_ast handler for client locks.
1041 * Sends glimpse AST to the client and waits for reply. Then updates
1042 * lvbo with the result.
1044 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1046 struct ldlm_cb_set_arg *arg = data;
1047 struct ldlm_request *body;
1048 struct ptlrpc_request *req;
1049 struct ldlm_cb_async_args *ca;
1051 struct req_format *req_fmt;
1054 LASSERT(lock != NULL);
1056 if (arg->gl_desc != NULL)
1057 /* There is a glimpse descriptor to pack */
1058 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1060 req_fmt = &RQF_LDLM_GL_CALLBACK;
1062 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1063 req_fmt, LUSTRE_DLM_VERSION,
1069 if (arg->gl_desc != NULL) {
1070 /* copy the GL descriptor */
1071 union ldlm_gl_desc *desc;
1072 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1073 *desc = *arg->gl_desc;
1076 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1077 body->lock_handle[0] = lock->l_remote_handle;
1078 ldlm_lock2desc(lock, &body->lock_desc);
1080 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1081 ca = ptlrpc_req_async_args(req);
1082 ca->ca_set_arg = arg;
1085 /* server namespace, doesn't need lock */
1086 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1087 ldlm_lvbo_size(lock));
1088 ptlrpc_request_set_replen(req);
1090 req->rq_send_state = LUSTRE_IMP_FULL;
1091 /* ptlrpc_request_alloc_pack already set timeout */
1093 req->rq_timeout = ldlm_get_rq_timeout();
1095 lock->l_last_activity = cfs_time_current_sec();
1097 req->rq_interpret_reply = ldlm_cb_interpret;
1099 if (lock->l_export && lock->l_export->exp_nid_stats &&
1100 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1101 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1102 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1104 rc = ldlm_ast_fini(req, arg, lock, 0);
1109 int ldlm_glimpse_locks(struct ldlm_resource *res,
1110 struct list_head *gl_work_list)
1115 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1117 if (rc == -ERESTART)
1118 ldlm_reprocess_all(res);
1122 EXPORT_SYMBOL(ldlm_glimpse_locks);
1124 /* return LDLM lock associated with a lock callback request */
1125 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1127 struct ldlm_cb_async_args *ca;
1128 struct ldlm_lock *lock;
1131 ca = ptlrpc_req_async_args(req);
1134 RETURN(ERR_PTR(-EFAULT));
1138 EXPORT_SYMBOL(ldlm_request_lock);
1140 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1141 struct lprocfs_stats *srv_stats)
1143 int lock_type = 0, op = 0;
1145 lock_type = dlm_req->lock_desc.l_resource.lr_type;
1147 switch (lock_type) {
1149 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1152 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1153 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1155 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1158 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1161 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1169 lprocfs_counter_incr(srv_stats, op);
1175 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1176 * service threads to carry out client lock enqueueing requests.
1178 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1179 struct ptlrpc_request *req,
1180 const struct ldlm_request *dlm_req,
1181 const struct ldlm_callback_suite *cbs)
1183 struct ldlm_reply *dlm_rep;
1185 ldlm_error_t err = ELDLM_OK;
1186 struct ldlm_lock *lock = NULL;
1187 void *cookie = NULL;
1189 struct ldlm_resource *res = NULL;
1192 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1194 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1195 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1197 LASSERT(req->rq_export);
1199 if (ptlrpc_req2svc(req)->srv_stats != NULL)
1200 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1202 if (req->rq_export && req->rq_export->exp_nid_stats &&
1203 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1204 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1205 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1207 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1208 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1209 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1210 dlm_req->lock_desc.l_resource.lr_type);
1211 GOTO(out, rc = -EFAULT);
1214 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1215 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1216 dlm_req->lock_desc.l_req_mode &
1217 (dlm_req->lock_desc.l_req_mode-1))) {
1218 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1219 dlm_req->lock_desc.l_req_mode);
1220 GOTO(out, rc = -EFAULT);
1223 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) {
1224 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1226 DEBUG_REQ(D_ERROR, req,
1227 "PLAIN lock request from IBITS client?");
1228 GOTO(out, rc = -EPROTO);
1230 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1232 DEBUG_REQ(D_ERROR, req,
1233 "IBITS lock request from unaware client?");
1234 GOTO(out, rc = -EPROTO);
1238 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1239 against server's _CONNECT_SUPPORTED flags? (I don't want to use
1240 ibits for mgc/mgs) */
1242 /* INODEBITS_INTEROP: Perform conversion from plain lock to
1243 * inodebits lock if client does not support them. */
1244 if (!(exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) &&
1245 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1246 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1247 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1248 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1249 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1250 dlm_req->lock_desc.l_req_mode = LCK_CR;
1254 if (unlikely((flags & LDLM_FL_REPLAY) ||
1255 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1256 /* Find an existing lock in the per-export lock hash */
1257 /* In the function below, .hs_keycmp resolves to
1258 * ldlm_export_lock_keycmp() */
1259 /* coverity[overrun-buffer-val] */
1260 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1261 (void *)&dlm_req->lock_handle[0]);
1263 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1264 LPX64, lock->l_handle.h_cookie);
1265 flags |= LDLM_FL_RESENT;
1266 GOTO(existing_lock, rc = 0);
1269 if (ldlm_reclaim_full()) {
1270 DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1271 "reject current enqueue request and let the "
1272 "client retry later.\n");
1273 GOTO(out, rc = -EINPROGRESS);
1277 /* The lock's callback data might be set in the policy function */
1278 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1279 dlm_req->lock_desc.l_resource.lr_type,
1280 dlm_req->lock_desc.l_req_mode,
1281 cbs, NULL, 0, LVB_T_NONE);
1288 lock->l_remote_handle = dlm_req->lock_handle[0];
1289 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1291 /* Initialize resource lvb but not for a lock being replayed since
1292 * Client already got lvb sent in this case.
1293 * This must occur early since some policy methods assume resource
1294 * lvb is available (lr_lvb_data != NULL).
1296 res = lock->l_resource;
1297 if (!(flags & LDLM_FL_REPLAY)) {
1298 /* non-replayed lock, delayed lvb init may need to be done */
1299 rc = ldlm_lvbo_init(res);
1301 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1306 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1307 /* Don't enqueue a lock onto the export if it is been disonnected
1308 * due to eviction (bug 3822) or server umount (bug 24324).
1309 * Cancel it now instead. */
1310 if (req->rq_export->exp_disconnected) {
1311 LDLM_ERROR(lock, "lock on disconnected export %p",
1313 GOTO(out, rc = -ENOTCONN);
1316 lock->l_export = class_export_lock_get(req->rq_export, lock);
1317 if (lock->l_export->exp_lock_hash)
1318 cfs_hash_add(lock->l_export->exp_lock_hash,
1319 &lock->l_remote_handle,
1322 /* Inherit the enqueue flags before the operation, because we do not
1323 * keep the res lock on return and next operations (BL AST) may proceed
1325 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1326 LDLM_FL_INHERIT_MASK);
1329 if (flags & LDLM_FL_HAS_INTENT) {
1330 /* In this case, the reply buffer is allocated deep in
1331 * local_lock_enqueue by the policy function. */
1334 /* based on the assumption that lvb size never changes during
1335 * resource life time otherwise it need resource->lr_lock's
1337 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1338 RCL_SERVER, ldlm_lvbo_size(lock));
1340 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1341 GOTO(out, rc = -ENOMEM);
1343 rc = req_capsule_server_pack(&req->rq_pill);
1348 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1349 ldlm_convert_policy_to_local(req->rq_export,
1350 dlm_req->lock_desc.l_resource.lr_type,
1351 &dlm_req->lock_desc.l_policy_data,
1352 &lock->l_policy_data);
1353 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1354 lock->l_req_extent = lock->l_policy_data.l_extent;
1356 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1363 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1365 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1366 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1368 /* We never send a blocking AST until the lock is granted, but
1369 * we can tell it right now */
1370 lock_res_and_lock(lock);
1372 /* Now take into account flags to be inherited from original lock
1373 request both in reply to client and in our own lock flags. */
1374 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1375 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1377 /* Don't move a pending lock onto the export if it has already been
1378 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1379 * Cancel it now instead. */
1380 if (unlikely(req->rq_export->exp_disconnected ||
1381 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1382 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1384 } else if (ldlm_is_ast_sent(lock)) {
1385 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1386 if (lock->l_granted_mode == lock->l_req_mode) {
1388 * Only cancel lock if it was granted, because it would
1389 * be destroyed immediately and would never be granted
1390 * in the future, causing timeouts on client. Not
1391 * granted lock will be cancelled immediately after
1392 * sending completion AST.
1394 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1395 unlock_res_and_lock(lock);
1396 ldlm_lock_cancel(lock);
1397 lock_res_and_lock(lock);
1399 ldlm_add_waiting_lock(lock);
1402 /* Make sure we never ever grant usual metadata locks to liblustre
1404 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1405 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1406 req->rq_export->exp_libclient) {
1407 if (unlikely(!ldlm_is_cancel_on_block(lock) ||
1408 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1409 CERROR("Granting sync lock to libclient. "
1410 "req fl %d, rep fl %d, lock fl "LPX64"\n",
1411 dlm_req->lock_flags, dlm_rep->lock_flags,
1413 LDLM_ERROR(lock, "sync lock");
1414 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1415 struct ldlm_intent *it;
1417 it = req_capsule_client_get(&req->rq_pill,
1420 CERROR("This is intent %s ("LPU64")\n",
1421 ldlm_it2str(it->opc), it->opc);
1427 unlock_res_and_lock(lock);
1431 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1432 if (!req->rq_packed_final) {
1433 err = lustre_pack_reply(req, 1, NULL, NULL);
1438 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1439 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1441 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1442 "(err=%d, rc=%d)", err, rc);
1445 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1447 ldlm_lvbo_size(lock) > 0) {
1451 buf = req_capsule_server_get(&req->rq_pill,
1453 LASSERTF(buf != NULL, "req %p, lock %p\n",
1455 buflen = req_capsule_get_size(&req->rq_pill,
1456 &RMF_DLM_LVB, RCL_SERVER);
1457 /* non-replayed lock, delayed lvb init may
1458 * need to be occur now */
1459 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1460 buflen = ldlm_lvbo_fill(lock, buf,
1466 buflen, RCL_SERVER);
1469 } else if (flags & LDLM_FL_REPLAY) {
1470 /* no LVB resend upon replay */
1485 lock_res_and_lock(lock);
1486 ldlm_resource_unlink_lock(lock);
1487 ldlm_lock_destroy_nolock(lock);
1488 unlock_res_and_lock(lock);
1491 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1492 ldlm_reprocess_all(lock->l_resource);
1494 LDLM_LOCK_RELEASE(lock);
1497 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1504 * Old-style LDLM main entry point for server code enqueue.
1506 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1507 ldlm_completion_callback completion_callback,
1508 ldlm_blocking_callback blocking_callback,
1509 ldlm_glimpse_callback glimpse_callback)
1511 struct ldlm_request *dlm_req;
1512 struct ldlm_callback_suite cbs = {
1513 .lcs_completion = completion_callback,
1514 .lcs_blocking = blocking_callback,
1515 .lcs_glimpse = glimpse_callback
1519 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1520 if (dlm_req != NULL) {
1521 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1522 req, dlm_req, &cbs);
1530 * Main LDLM entry point for server code to process lock conversion requests.
1532 int ldlm_handle_convert0(struct ptlrpc_request *req,
1533 const struct ldlm_request *dlm_req)
1535 struct ldlm_reply *dlm_rep;
1536 struct ldlm_lock *lock;
1540 if (req->rq_export && req->rq_export->exp_nid_stats &&
1541 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1542 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1543 LDLM_CONVERT - LDLM_FIRST_OPC);
1545 rc = req_capsule_server_pack(&req->rq_pill);
1549 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1550 dlm_rep->lock_flags = dlm_req->lock_flags;
1552 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1554 req->rq_status = LUSTRE_EINVAL;
1558 LDLM_DEBUG(lock, "server-side convert handler START");
1560 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1561 &dlm_rep->lock_flags);
1563 if (ldlm_del_waiting_lock(lock))
1564 LDLM_DEBUG(lock, "converted waiting lock");
1567 req->rq_status = LUSTRE_EDEADLK;
1572 if (!req->rq_status)
1573 ldlm_reprocess_all(lock->l_resource);
1574 LDLM_DEBUG(lock, "server-side convert handler END");
1575 LDLM_LOCK_PUT(lock);
1577 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1583 * Old-style main LDLM entry point for server code to process lock conversion
1586 int ldlm_handle_convert(struct ptlrpc_request *req)
1589 struct ldlm_request *dlm_req;
1591 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1592 if (dlm_req != NULL) {
1593 rc = ldlm_handle_convert0(req, dlm_req);
1595 CERROR ("Can't unpack dlm_req\n");
1602 * Cancel all the locks whose handles are packed into ldlm_request
1604 * Called by server code expecting such combined cancel activity
1607 int ldlm_request_cancel(struct ptlrpc_request *req,
1608 const struct ldlm_request *dlm_req,
1609 int first, enum lustre_at_flags flags)
1611 struct ldlm_resource *res, *pres = NULL;
1612 struct ldlm_lock *lock;
1613 int i, count, done = 0;
1616 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1620 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1623 /* There is no lock on the server at the replay time,
1624 * skip lock cancelling to make replay tests to pass. */
1625 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1628 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1629 "starting at %d", count, first);
1631 for (i = first; i < count; i++) {
1632 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1634 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1635 "lock (cookie "LPU64")",
1636 dlm_req->lock_handle[i].cookie);
1640 res = lock->l_resource;
1643 /* This code is an optimization to only attempt lock
1644 * granting on the resource (that could be CPU-expensive)
1645 * after we are done cancelling lock in that resource. */
1648 ldlm_reprocess_all(pres);
1649 LDLM_RESOURCE_DELREF(pres);
1650 ldlm_resource_putref(pres);
1653 ldlm_resource_getref(res);
1654 LDLM_RESOURCE_ADDREF(res);
1655 ldlm_res_lvbo_update(res, NULL, 1);
1660 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock)) {
1661 long delay = cfs_time_sub(cfs_time_current_sec(),
1662 lock->l_last_activity);
1663 LDLM_DEBUG(lock, "server cancels blocked lock after "
1664 CFS_DURATION_T"s", delay);
1665 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1667 ldlm_lock_cancel(lock);
1668 LDLM_LOCK_PUT(lock);
1671 ldlm_reprocess_all(pres);
1672 LDLM_RESOURCE_DELREF(pres);
1673 ldlm_resource_putref(pres);
1675 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1678 EXPORT_SYMBOL(ldlm_request_cancel);
1681 * Main LDLM entry point for server code to cancel locks.
1683 * Typically gets called from service handler on LDLM_CANCEL opc.
1685 int ldlm_handle_cancel(struct ptlrpc_request *req)
1687 struct ldlm_request *dlm_req;
1691 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1692 if (dlm_req == NULL) {
1693 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1697 if (req->rq_export && req->rq_export->exp_nid_stats &&
1698 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1699 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1700 LDLM_CANCEL - LDLM_FIRST_OPC);
1702 rc = req_capsule_server_pack(&req->rq_pill);
1706 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1707 req->rq_status = LUSTRE_ESTALE;
1709 RETURN(ptlrpc_reply(req));
1711 #endif /* HAVE_SERVER_SUPPORT */
1714 * Callback handler for receiving incoming blocking ASTs.
1716 * This can only happen on client side.
1718 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1719 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1724 LDLM_DEBUG(lock, "client blocking AST callback handler");
1726 lock_res_and_lock(lock);
1727 ldlm_set_cbpending(lock);
1729 if (ldlm_is_cancel_on_block(lock))
1730 ldlm_set_cancel(lock);
1732 do_ast = (!lock->l_readers && !lock->l_writers);
1733 unlock_res_and_lock(lock);
1736 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1737 lock, lock->l_blocking_ast);
1738 if (lock->l_blocking_ast != NULL)
1739 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1742 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1746 LDLM_DEBUG(lock, "client blocking callback handler END");
1747 LDLM_LOCK_RELEASE(lock);
1752 * Callback handler for receiving incoming completion ASTs.
1754 * This only can happen on client side.
1756 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1757 struct ldlm_namespace *ns,
1758 struct ldlm_request *dlm_req,
1759 struct ldlm_lock *lock)
1761 struct list_head ast_list;
1766 LDLM_DEBUG(lock, "client completion callback handler START");
1768 INIT_LIST_HEAD(&ast_list);
1769 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1770 int to = cfs_time_seconds(1);
1772 set_current_state(TASK_INTERRUPTIBLE);
1773 schedule_timeout(to);
1774 if (lock->l_granted_mode == lock->l_req_mode ||
1775 ldlm_is_destroyed(lock))
1780 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1782 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1783 GOTO(out, rc = lvb_len);
1784 } else if (lvb_len > 0) {
1785 if (lock->l_lvb_len > 0) {
1786 /* for extent lock, lvb contains ost_lvb{}. */
1787 LASSERT(lock->l_lvb_data != NULL);
1789 if (unlikely(lock->l_lvb_len < lvb_len)) {
1790 LDLM_ERROR(lock, "Replied LVB is larger than "
1791 "expectation, expected = %d, "
1793 lock->l_lvb_len, lvb_len);
1794 GOTO(out, rc = -EINVAL);
1799 lock_res_and_lock(lock);
1800 if (ldlm_is_destroyed(lock) ||
1801 lock->l_granted_mode == lock->l_req_mode) {
1802 /* bug 11300: the lock has already been granted */
1803 unlock_res_and_lock(lock);
1804 LDLM_DEBUG(lock, "Double grant race happened");
1808 /* If we receive the completion AST before the actual enqueue returned,
1809 * then we might need to switch lock modes, resources, or extents. */
1810 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1811 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1812 LDLM_DEBUG(lock, "completion AST, new lock mode");
1815 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1816 ldlm_convert_policy_to_local(req->rq_export,
1817 dlm_req->lock_desc.l_resource.lr_type,
1818 &dlm_req->lock_desc.l_policy_data,
1819 &lock->l_policy_data);
1820 LDLM_DEBUG(lock, "completion AST, new policy data");
1823 ldlm_resource_unlink_lock(lock);
1824 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1825 &lock->l_resource->lr_name,
1826 sizeof(lock->l_resource->lr_name)) != 0) {
1827 unlock_res_and_lock(lock);
1828 rc = ldlm_lock_change_resource(ns, lock,
1829 &dlm_req->lock_desc.l_resource.lr_name);
1831 LDLM_ERROR(lock, "Failed to allocate resource");
1834 LDLM_DEBUG(lock, "completion AST, new resource");
1835 CERROR("change resource!\n");
1836 lock_res_and_lock(lock);
1839 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1840 /* BL_AST locks are not needed in LRU.
1841 * Let ldlm_cancel_lru() be fast. */
1842 ldlm_lock_remove_from_lru(lock);
1843 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1844 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1847 if (lock->l_lvb_len > 0) {
1848 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1849 lock->l_lvb_data, lvb_len);
1851 unlock_res_and_lock(lock);
1856 ldlm_grant_lock(lock, &ast_list);
1857 unlock_res_and_lock(lock);
1859 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1861 /* Let Enqueue to call osc_lock_upcall() and initialize
1863 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1865 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1867 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1873 lock_res_and_lock(lock);
1874 ldlm_set_failed(lock);
1875 unlock_res_and_lock(lock);
1876 wake_up(&lock->l_waitq);
1878 LDLM_LOCK_RELEASE(lock);
1882 * Callback handler for receiving incoming glimpse ASTs.
1884 * This only can happen on client side. After handling the glimpse AST
1885 * we also consider dropping the lock here if it is unused locally for a
1888 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1889 struct ldlm_namespace *ns,
1890 struct ldlm_request *dlm_req,
1891 struct ldlm_lock *lock)
1896 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1898 if (lock->l_glimpse_ast != NULL)
1899 rc = lock->l_glimpse_ast(lock, req);
1901 if (req->rq_repmsg != NULL) {
1904 req->rq_status = rc;
1908 lock_res_and_lock(lock);
1909 if (lock->l_granted_mode == LCK_PW &&
1910 !lock->l_readers && !lock->l_writers &&
1911 cfs_time_after(cfs_time_current(),
1912 cfs_time_add(lock->l_last_used,
1913 cfs_time_seconds(10)))) {
1914 unlock_res_and_lock(lock);
1915 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1916 ldlm_handle_bl_callback(ns, NULL, lock);
1921 unlock_res_and_lock(lock);
1922 LDLM_LOCK_RELEASE(lock);
1926 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1928 if (req->rq_no_reply)
1931 req->rq_status = rc;
1932 if (!req->rq_packed_final) {
1933 rc = lustre_pack_reply(req, 1, NULL, NULL);
1937 return ptlrpc_reply(req);
1940 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1941 ldlm_cancel_flags_t cancel_flags)
1943 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1946 spin_lock(&blp->blp_lock);
1947 if (blwi->blwi_lock &&
1948 ldlm_is_discard_data(blwi->blwi_lock)) {
1949 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1950 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1952 /* other blocking callbacks are added to the regular list */
1953 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1955 spin_unlock(&blp->blp_lock);
1957 wake_up(&blp->blp_waitq);
1959 /* can not check blwi->blwi_flags as blwi could be already freed in
1961 if (!(cancel_flags & LCF_ASYNC))
1962 wait_for_completion(&blwi->blwi_comp);
1967 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1968 struct ldlm_namespace *ns,
1969 struct ldlm_lock_desc *ld,
1970 struct list_head *cancels, int count,
1971 struct ldlm_lock *lock,
1972 ldlm_cancel_flags_t cancel_flags)
1974 init_completion(&blwi->blwi_comp);
1975 INIT_LIST_HEAD(&blwi->blwi_head);
1977 if (memory_pressure_get())
1978 blwi->blwi_mem_pressure = 1;
1981 blwi->blwi_flags = cancel_flags;
1983 blwi->blwi_ld = *ld;
1985 list_add(&blwi->blwi_head, cancels);
1986 list_del_init(cancels);
1987 blwi->blwi_count = count;
1989 blwi->blwi_lock = lock;
1994 * Queues a list of locks \a cancels containing \a count locks
1995 * for later processing by a blocking thread. If \a count is zero,
1996 * then the lock referenced as \a lock is queued instead.
1998 * The blocking thread would then call ->l_blocking_ast callback in the lock.
1999 * If list addition fails an error is returned and caller is supposed to
2000 * call ->l_blocking_ast itself.
2002 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2003 struct ldlm_lock_desc *ld,
2004 struct ldlm_lock *lock,
2005 struct list_head *cancels, int count,
2006 ldlm_cancel_flags_t cancel_flags)
2010 if (cancels && count == 0)
2013 if (cancel_flags & LCF_ASYNC) {
2014 struct ldlm_bl_work_item *blwi;
2016 OBD_ALLOC(blwi, sizeof(*blwi));
2019 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2021 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2023 /* if it is synchronous call do minimum mem alloc, as it could
2024 * be triggered from kernel shrinker
2026 struct ldlm_bl_work_item blwi;
2028 memset(&blwi, 0, sizeof(blwi));
2029 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2030 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2035 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2036 struct ldlm_lock *lock)
2038 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2041 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2042 struct list_head *cancels, int count,
2043 ldlm_cancel_flags_t cancel_flags)
2045 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2048 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2049 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2051 struct obd_device *obd = req->rq_export->exp_obd;
2058 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2060 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2062 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2064 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2067 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2069 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2071 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2074 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2077 /* We are responsible for swabbing contents of val */
2079 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2080 /* Pass it on to mdc (the "export" in this case) */
2081 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2083 sizeof(KEY_HSM_COPYTOOL_SEND),
2084 KEY_HSM_COPYTOOL_SEND,
2087 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2092 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2093 const char *msg, int rc,
2094 struct lustre_handle *handle)
2096 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2097 "%s: [nid %s] [rc %d] [lock "LPX64"]",
2098 msg, libcfs_id2str(req->rq_peer), rc,
2099 handle ? handle->cookie : 0);
2100 if (req->rq_no_reply)
2101 CWARN("No reply was sent, maybe cause bug 21636.\n");
2103 CWARN("Send reply failed, maybe cause bug 21636.\n");
2106 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2107 static int ldlm_callback_handler(struct ptlrpc_request *req)
2109 struct ldlm_namespace *ns;
2110 struct ldlm_request *dlm_req;
2111 struct ldlm_lock *lock;
2115 /* Requests arrive in sender's byte order. The ptlrpc service
2116 * handler has already checked and, if necessary, byte-swapped the
2117 * incoming request message body, but I am responsible for the
2118 * message buffers. */
2120 /* do nothing for sec context finalize */
2121 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2124 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2126 if (req->rq_export == NULL) {
2127 rc = ldlm_callback_reply(req, -ENOTCONN);
2128 ldlm_callback_errmsg(req, "Operate on unconnected server",
2133 LASSERT(req->rq_export != NULL);
2134 LASSERT(req->rq_export->exp_obd != NULL);
2136 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2137 case LDLM_BL_CALLBACK:
2138 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2140 ldlm_callback_reply(req, -(int)cfs_fail_err);
2144 case LDLM_CP_CALLBACK:
2145 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2148 case LDLM_GL_CALLBACK:
2149 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2153 rc = ldlm_handle_setinfo(req);
2154 ldlm_callback_reply(req, rc);
2156 case LLOG_ORIGIN_HANDLE_CREATE:
2157 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2158 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2160 rc = llog_origin_handle_open(req);
2161 ldlm_callback_reply(req, rc);
2163 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2164 req_capsule_set(&req->rq_pill,
2165 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2166 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2168 rc = llog_origin_handle_next_block(req);
2169 ldlm_callback_reply(req, rc);
2171 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2172 req_capsule_set(&req->rq_pill,
2173 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2174 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2176 rc = llog_origin_handle_read_header(req);
2177 ldlm_callback_reply(req, rc);
2179 case LLOG_ORIGIN_HANDLE_CLOSE:
2180 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2182 rc = llog_origin_handle_close(req);
2183 ldlm_callback_reply(req, rc);
2186 CERROR("unknown opcode %u\n",
2187 lustre_msg_get_opc(req->rq_reqmsg));
2188 ldlm_callback_reply(req, -EPROTO);
2192 ns = req->rq_export->exp_obd->obd_namespace;
2193 LASSERT(ns != NULL);
2195 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2197 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2198 if (dlm_req == NULL) {
2199 rc = ldlm_callback_reply(req, -EPROTO);
2200 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2205 /* Force a known safe race, send a cancel to the server for a lock
2206 * which the server has already started a blocking callback on. */
2207 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2208 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2209 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2211 CERROR("ldlm_cli_cancel: %d\n", rc);
2214 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2216 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2217 "disappeared\n", dlm_req->lock_handle[0].cookie);
2218 rc = ldlm_callback_reply(req, -EINVAL);
2219 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2220 &dlm_req->lock_handle[0]);
2224 if (ldlm_is_fail_loc(lock) &&
2225 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2226 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2228 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2229 lock_res_and_lock(lock);
2230 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2232 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2233 /* If somebody cancels lock and cache is already dropped,
2234 * or lock is failed before cp_ast received on client,
2235 * we can tell the server we have no lock. Otherwise, we
2236 * should send cancel after dropping the cache. */
2237 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2238 ldlm_is_failed(lock)) {
2239 LDLM_DEBUG(lock, "callback on lock "
2240 LPX64" - lock disappeared\n",
2241 dlm_req->lock_handle[0].cookie);
2242 unlock_res_and_lock(lock);
2243 LDLM_LOCK_RELEASE(lock);
2244 rc = ldlm_callback_reply(req, -EINVAL);
2245 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2246 &dlm_req->lock_handle[0]);
2249 /* BL_AST locks are not needed in LRU.
2250 * Let ldlm_cancel_lru() be fast. */
2251 ldlm_lock_remove_from_lru(lock);
2252 ldlm_set_bl_ast(lock);
2254 unlock_res_and_lock(lock);
2256 /* We want the ost thread to get this reply so that it can respond
2257 * to ost requests (write cache writeback) that might be triggered
2260 * But we'd also like to be able to indicate in the reply that we're
2261 * cancelling right now, because it's unused, or have an intent result
2262 * in the reply, so we might have to push the responsibility for sending
2263 * the reply down into the AST handlers, alas. */
2265 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2266 case LDLM_BL_CALLBACK:
2267 CDEBUG(D_INODE, "blocking ast\n");
2268 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2269 if (!ldlm_is_cancel_on_block(lock)) {
2270 rc = ldlm_callback_reply(req, 0);
2271 if (req->rq_no_reply || rc)
2272 ldlm_callback_errmsg(req, "Normal process", rc,
2273 &dlm_req->lock_handle[0]);
2275 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2276 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2278 case LDLM_CP_CALLBACK:
2279 CDEBUG(D_INODE, "completion ast\n");
2280 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2281 ldlm_callback_reply(req, 0);
2282 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2284 case LDLM_GL_CALLBACK:
2285 CDEBUG(D_INODE, "glimpse ast\n");
2286 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2287 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2290 LBUG(); /* checked above */
2296 #ifdef HAVE_SERVER_SUPPORT
2298 * Main handler for canceld thread.
2300 * Separated into its own thread to avoid deadlocks.
2302 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2307 /* Requests arrive in sender's byte order. The ptlrpc service
2308 * handler has already checked and, if necessary, byte-swapped the
2309 * incoming request message body, but I am responsible for the
2310 * message buffers. */
2312 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2314 if (req->rq_export == NULL) {
2315 struct ldlm_request *dlm_req;
2317 CERROR("%s from %s arrived at %lu with bad export cookie "
2319 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2320 libcfs_nid2str(req->rq_peer.nid),
2321 req->rq_arrival_time.tv_sec,
2322 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2324 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2325 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2326 dlm_req = req_capsule_client_get(&req->rq_pill,
2328 if (dlm_req != NULL)
2329 ldlm_lock_dump_handle(D_ERROR,
2330 &dlm_req->lock_handle[0]);
2332 ldlm_callback_reply(req, -ENOTCONN);
2336 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2338 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2340 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2341 CDEBUG(D_INODE, "cancel\n");
2342 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2343 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))
2345 rc = ldlm_handle_cancel(req);
2350 CERROR("invalid opcode %d\n",
2351 lustre_msg_get_opc(req->rq_reqmsg));
2352 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2353 ldlm_callback_reply(req, -EINVAL);
2359 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2360 struct ldlm_lock *lock)
2362 struct ldlm_request *dlm_req;
2363 struct lustre_handle lockh;
2368 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2369 if (dlm_req == NULL)
2372 ldlm_lock2handle(lock, &lockh);
2373 for (i = 0; i < dlm_req->lock_count; i++) {
2374 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2376 DEBUG_REQ(D_RPCTRACE, req,
2377 "Prio raised by lock "LPX64".", lockh.cookie);
2388 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2390 struct ldlm_request *dlm_req;
2395 /* no prolong in recovery */
2396 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2399 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2400 if (dlm_req == NULL)
2403 for (i = 0; i < dlm_req->lock_count; i++) {
2404 struct ldlm_lock *lock;
2406 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2410 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2412 LDLM_DEBUG(lock, "hpreq cancel lock");
2413 LDLM_LOCK_PUT(lock);
2422 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2423 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2424 .hpreq_check = ldlm_cancel_hpreq_check,
2428 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2432 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2434 if (req->rq_export == NULL)
2437 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2438 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2439 req->rq_ops = &ldlm_cancel_hpreq_ops;
2444 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2445 struct hlist_node *hnode, void *data)
2448 struct list_head *rpc_list = data;
2449 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2451 lock_res_and_lock(lock);
2453 if (lock->l_req_mode != lock->l_granted_mode) {
2454 unlock_res_and_lock(lock);
2458 LASSERT(lock->l_resource);
2459 if (lock->l_resource->lr_type != LDLM_IBITS &&
2460 lock->l_resource->lr_type != LDLM_PLAIN) {
2461 unlock_res_and_lock(lock);
2465 if (ldlm_is_ast_sent(lock)) {
2466 unlock_res_and_lock(lock);
2470 LASSERT(lock->l_blocking_ast);
2471 LASSERT(!lock->l_blocking_lock);
2473 ldlm_set_ast_sent(lock);
2474 if (lock->l_export && lock->l_export->exp_lock_hash) {
2475 /* NB: it's safe to call cfs_hash_del() even lock isn't
2476 * in exp_lock_hash. */
2477 /* In the function below, .hs_keycmp resolves to
2478 * ldlm_export_lock_keycmp() */
2479 /* coverity[overrun-buffer-val] */
2480 cfs_hash_del(lock->l_export->exp_lock_hash,
2481 &lock->l_remote_handle, &lock->l_exp_hash);
2484 list_add_tail(&lock->l_rk_ast, rpc_list);
2485 LDLM_LOCK_GET(lock);
2487 unlock_res_and_lock(lock);
2491 void ldlm_revoke_export_locks(struct obd_export *exp)
2493 struct list_head rpc_list;
2496 INIT_LIST_HEAD(&rpc_list);
2497 cfs_hash_for_each_empty(exp->exp_lock_hash,
2498 ldlm_revoke_lock_cb, &rpc_list);
2499 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2500 LDLM_WORK_REVOKE_AST);
2504 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2505 #endif /* HAVE_SERVER_SUPPORT */
2507 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2509 struct ldlm_bl_work_item *blwi = NULL;
2510 static unsigned int num_bl = 0;
2512 spin_lock(&blp->blp_lock);
2513 /* process a request from the blp_list at least every blp_num_threads */
2514 if (!list_empty(&blp->blp_list) &&
2515 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2516 blwi = list_entry(blp->blp_list.next,
2517 struct ldlm_bl_work_item, blwi_entry);
2519 if (!list_empty(&blp->blp_prio_list))
2520 blwi = list_entry(blp->blp_prio_list.next,
2521 struct ldlm_bl_work_item,
2525 if (++num_bl >= atomic_read(&blp->blp_num_threads))
2527 list_del(&blwi->blwi_entry);
2529 spin_unlock(&blp->blp_lock);
2534 /* This only contains temporary data until the thread starts */
2535 struct ldlm_bl_thread_data {
2536 char bltd_name[CFS_CURPROC_COMM_MAX];
2537 struct ldlm_bl_pool *bltd_blp;
2538 struct completion bltd_comp;
2542 static int ldlm_bl_thread_main(void *arg);
2544 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2546 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2547 struct task_struct *task;
2549 init_completion(&bltd.bltd_comp);
2550 bltd.bltd_num = atomic_read(&blp->blp_num_threads);
2551 snprintf(bltd.bltd_name, sizeof(bltd.bltd_name) - 1,
2552 "ldlm_bl_%02d", bltd.bltd_num);
2553 task = kthread_run(ldlm_bl_thread_main, &bltd, bltd.bltd_name);
2555 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2556 atomic_read(&blp->blp_num_threads), PTR_ERR(task));
2557 return PTR_ERR(task);
2559 wait_for_completion(&bltd.bltd_comp);
2565 * Main blocking requests processing thread.
2567 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2568 * This thread in the end ends up doing actual call to ->l_blocking_ast
2571 static int ldlm_bl_thread_main(void *arg)
2573 struct ldlm_bl_pool *blp;
2574 struct ldlm_bl_thread_data *bltd = arg;
2577 blp = bltd->bltd_blp;
2579 atomic_inc(&blp->blp_num_threads);
2580 atomic_inc(&blp->blp_busy_threads);
2582 complete(&bltd->bltd_comp);
2583 /* cannot use bltd after this, it is only on caller's stack */
2586 struct l_wait_info lwi = { 0 };
2587 struct ldlm_bl_work_item *blwi = NULL;
2590 blwi = ldlm_bl_get_work(blp);
2593 atomic_dec(&blp->blp_busy_threads);
2594 l_wait_event_exclusive(blp->blp_waitq,
2595 (blwi = ldlm_bl_get_work(blp)) != NULL,
2597 busy = atomic_inc_return(&blp->blp_busy_threads);
2599 busy = atomic_read(&blp->blp_busy_threads);
2602 if (blwi->blwi_ns == NULL)
2603 /* added by ldlm_cleanup() */
2606 /* Not fatal if racy and have a few too many threads */
2607 if (unlikely(busy < blp->blp_max_threads &&
2608 busy >= atomic_read(&blp->blp_num_threads) &&
2609 !blwi->blwi_mem_pressure))
2610 /* discard the return value, we tried */
2611 ldlm_bl_thread_start(blp);
2613 if (blwi->blwi_mem_pressure)
2614 memory_pressure_set();
2616 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2618 if (blwi->blwi_count) {
2620 /* The special case when we cancel locks in LRU
2621 * asynchronously, we pass the list of locks here.
2622 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2623 * canceled locally yet. */
2624 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2627 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2630 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2633 if (blwi->blwi_mem_pressure)
2634 memory_pressure_clr();
2636 if (blwi->blwi_flags & LCF_ASYNC)
2637 OBD_FREE(blwi, sizeof(*blwi));
2639 complete(&blwi->blwi_comp);
2642 atomic_dec(&blp->blp_busy_threads);
2643 atomic_dec(&blp->blp_num_threads);
2644 complete(&blp->blp_comp);
2649 static int ldlm_setup(void);
2650 static int ldlm_cleanup(void);
2652 int ldlm_get_ref(void)
2656 mutex_lock(&ldlm_ref_mutex);
2657 if (++ldlm_refcount == 1) {
2662 mutex_unlock(&ldlm_ref_mutex);
2667 void ldlm_put_ref(void)
2670 mutex_lock(&ldlm_ref_mutex);
2671 if (ldlm_refcount == 1) {
2672 int rc = ldlm_cleanup();
2674 CERROR("ldlm_cleanup failed: %d\n", rc);
2680 mutex_unlock(&ldlm_ref_mutex);
2686 * Export handle<->lock hash operations.
2689 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2691 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2695 ldlm_export_lock_key(struct hlist_node *hnode)
2697 struct ldlm_lock *lock;
2699 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2700 return &lock->l_remote_handle;
2704 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2706 struct ldlm_lock *lock;
2708 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2709 lock->l_remote_handle = *(struct lustre_handle *)key;
2713 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2715 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2719 ldlm_export_lock_object(struct hlist_node *hnode)
2721 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2725 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2727 struct ldlm_lock *lock;
2729 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2730 LDLM_LOCK_GET(lock);
2734 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2736 struct ldlm_lock *lock;
2738 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2739 LDLM_LOCK_RELEASE(lock);
2742 static struct cfs_hash_ops ldlm_export_lock_ops = {
2743 .hs_hash = ldlm_export_lock_hash,
2744 .hs_key = ldlm_export_lock_key,
2745 .hs_keycmp = ldlm_export_lock_keycmp,
2746 .hs_keycpy = ldlm_export_lock_keycpy,
2747 .hs_object = ldlm_export_lock_object,
2748 .hs_get = ldlm_export_lock_get,
2749 .hs_put = ldlm_export_lock_put,
2750 .hs_put_locked = ldlm_export_lock_put,
2753 int ldlm_init_export(struct obd_export *exp)
2758 exp->exp_lock_hash =
2759 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2760 HASH_EXP_LOCK_CUR_BITS,
2761 HASH_EXP_LOCK_MAX_BITS,
2762 HASH_EXP_LOCK_BKT_BITS, 0,
2763 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2764 &ldlm_export_lock_ops,
2765 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2766 CFS_HASH_NBLK_CHANGE);
2768 if (!exp->exp_lock_hash)
2771 rc = ldlm_init_flock_export(exp);
2777 ldlm_destroy_export(exp);
2780 EXPORT_SYMBOL(ldlm_init_export);
2782 void ldlm_destroy_export(struct obd_export *exp)
2785 cfs_hash_putref(exp->exp_lock_hash);
2786 exp->exp_lock_hash = NULL;
2788 ldlm_destroy_flock_export(exp);
2791 EXPORT_SYMBOL(ldlm_destroy_export);
2793 static int ldlm_setup(void)
2795 static struct ptlrpc_service_conf conf;
2796 struct ldlm_bl_pool *blp = NULL;
2797 #ifdef HAVE_SERVER_SUPPORT
2798 struct task_struct *task;
2799 #endif /* HAVE_SERVER_SUPPORT */
2805 if (ldlm_state != NULL)
2808 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2809 if (ldlm_state == NULL)
2812 #ifdef CONFIG_PROC_FS
2813 rc = ldlm_proc_setup();
2816 #endif /* CONFIG_PROC_FS */
2818 memset(&conf, 0, sizeof(conf));
2819 conf = (typeof(conf)) {
2820 .psc_name = "ldlm_cbd",
2821 .psc_watchdog_factor = 2,
2823 .bc_nbufs = LDLM_CLIENT_NBUFS,
2824 .bc_buf_size = LDLM_BUFSIZE,
2825 .bc_req_max_size = LDLM_MAXREQSIZE,
2826 .bc_rep_max_size = LDLM_MAXREPSIZE,
2827 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2828 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2831 .tc_thr_name = "ldlm_cb",
2832 .tc_thr_factor = LDLM_THR_FACTOR,
2833 .tc_nthrs_init = LDLM_NTHRS_INIT,
2834 .tc_nthrs_base = LDLM_NTHRS_BASE,
2835 .tc_nthrs_max = LDLM_NTHRS_MAX,
2836 .tc_nthrs_user = ldlm_num_threads,
2837 .tc_cpu_affinity = 1,
2838 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2841 .cc_pattern = ldlm_cpts,
2844 .so_req_handler = ldlm_callback_handler,
2847 ldlm_state->ldlm_cb_service = \
2848 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2849 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2850 CERROR("failed to start service\n");
2851 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2852 ldlm_state->ldlm_cb_service = NULL;
2856 #ifdef HAVE_SERVER_SUPPORT
2857 memset(&conf, 0, sizeof(conf));
2858 conf = (typeof(conf)) {
2859 .psc_name = "ldlm_canceld",
2860 .psc_watchdog_factor = 6,
2862 .bc_nbufs = LDLM_SERVER_NBUFS,
2863 .bc_buf_size = LDLM_BUFSIZE,
2864 .bc_req_max_size = LDLM_MAXREQSIZE,
2865 .bc_rep_max_size = LDLM_MAXREPSIZE,
2866 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
2867 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
2871 .tc_thr_name = "ldlm_cn",
2872 .tc_thr_factor = LDLM_THR_FACTOR,
2873 .tc_nthrs_init = LDLM_NTHRS_INIT,
2874 .tc_nthrs_base = LDLM_NTHRS_BASE,
2875 .tc_nthrs_max = LDLM_NTHRS_MAX,
2876 .tc_nthrs_user = ldlm_num_threads,
2877 .tc_cpu_affinity = 1,
2878 .tc_ctx_tags = LCT_MD_THREAD | \
2883 .cc_pattern = ldlm_cpts,
2886 .so_req_handler = ldlm_cancel_handler,
2887 .so_hpreq_handler = ldlm_hpreq_handler,
2890 ldlm_state->ldlm_cancel_service = \
2891 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2892 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2893 CERROR("failed to start service\n");
2894 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2895 ldlm_state->ldlm_cancel_service = NULL;
2898 #endif /* HAVE_SERVER_SUPPORT */
2900 OBD_ALLOC(blp, sizeof(*blp));
2902 GOTO(out, rc = -ENOMEM);
2903 ldlm_state->ldlm_bl_pool = blp;
2905 spin_lock_init(&blp->blp_lock);
2906 INIT_LIST_HEAD(&blp->blp_list);
2907 INIT_LIST_HEAD(&blp->blp_prio_list);
2908 init_waitqueue_head(&blp->blp_waitq);
2909 atomic_set(&blp->blp_num_threads, 0);
2910 atomic_set(&blp->blp_busy_threads, 0);
2912 if (ldlm_num_threads == 0) {
2913 blp->blp_min_threads = LDLM_NTHRS_INIT;
2914 blp->blp_max_threads = LDLM_NTHRS_MAX;
2916 blp->blp_min_threads = blp->blp_max_threads = \
2917 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
2921 for (i = 0; i < blp->blp_min_threads; i++) {
2922 rc = ldlm_bl_thread_start(blp);
2927 #ifdef HAVE_SERVER_SUPPORT
2928 INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2929 expired_lock_thread.elt_state = ELT_STOPPED;
2930 init_waitqueue_head(&expired_lock_thread.elt_waitq);
2932 INIT_LIST_HEAD(&waiting_locks_list);
2933 spin_lock_init(&waiting_locks_spinlock);
2934 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, NULL);
2936 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
2939 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2943 wait_event(expired_lock_thread.elt_waitq,
2944 expired_lock_thread.elt_state == ELT_READY);
2945 #endif /* HAVE_SERVER_SUPPORT */
2947 rc = ldlm_pools_init();
2949 CERROR("Failed to initialize LDLM pools: %d\n", rc);
2953 rc = ldlm_reclaim_setup();
2955 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
2965 static int ldlm_cleanup(void)
2969 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2970 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2971 CERROR("ldlm still has namespaces; clean these up first.\n");
2972 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2973 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2977 ldlm_reclaim_cleanup();
2980 if (ldlm_state->ldlm_bl_pool != NULL) {
2981 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2983 while (atomic_read(&blp->blp_num_threads) > 0) {
2984 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2986 init_completion(&blp->blp_comp);
2988 spin_lock(&blp->blp_lock);
2989 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2990 wake_up(&blp->blp_waitq);
2991 spin_unlock(&blp->blp_lock);
2993 wait_for_completion(&blp->blp_comp);
2996 OBD_FREE(blp, sizeof(*blp));
2999 if (ldlm_state->ldlm_cb_service != NULL)
3000 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3001 #ifdef HAVE_SERVER_SUPPORT
3002 if (ldlm_state->ldlm_cancel_service != NULL)
3003 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3006 ldlm_proc_cleanup();
3008 #ifdef HAVE_SERVER_SUPPORT
3009 if (expired_lock_thread.elt_state != ELT_STOPPED) {
3010 expired_lock_thread.elt_state = ELT_TERMINATE;
3011 wake_up(&expired_lock_thread.elt_waitq);
3012 wait_event(expired_lock_thread.elt_waitq,
3013 expired_lock_thread.elt_state == ELT_STOPPED);
3017 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3025 mutex_init(&ldlm_ref_mutex);
3026 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
3027 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
3029 INIT_LIST_HEAD(&ldlm_srv_namespace_list);
3030 INIT_LIST_HEAD(&ldlm_cli_active_namespace_list);
3031 INIT_LIST_HEAD(&ldlm_cli_inactive_namespace_list);
3033 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3034 sizeof(struct ldlm_resource), 0,
3035 SLAB_HWCACHE_ALIGN, NULL);
3036 if (ldlm_resource_slab == NULL)
3039 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3040 sizeof(struct ldlm_lock), 0,
3041 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3042 if (ldlm_lock_slab == NULL)
3045 ldlm_interval_slab = kmem_cache_create("interval_node",
3046 sizeof(struct ldlm_interval),
3047 0, SLAB_HWCACHE_ALIGN, NULL);
3048 if (ldlm_interval_slab == NULL)
3051 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3052 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3053 0, SLAB_HWCACHE_ALIGN, NULL);
3054 if (ldlm_interval_tree_slab == NULL)
3057 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3058 class_export_dump_hook = ldlm_dump_export_locks;
3063 kmem_cache_destroy(ldlm_interval_slab);
3065 kmem_cache_destroy(ldlm_lock_slab);
3067 kmem_cache_destroy(ldlm_resource_slab);
3072 void ldlm_exit(void)
3075 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3076 kmem_cache_destroy(ldlm_resource_slab);
3077 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3078 * synchronize_rcu() to wait a grace period elapsed, so that
3079 * ldlm_lock_free() get a chance to be called. */
3081 kmem_cache_destroy(ldlm_lock_slab);
3082 kmem_cache_destroy(ldlm_interval_slab);
3083 kmem_cache_destroy(ldlm_interval_tree_slab);