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);
1270 /* The lock's callback data might be set in the policy function */
1271 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1272 dlm_req->lock_desc.l_resource.lr_type,
1273 dlm_req->lock_desc.l_req_mode,
1274 cbs, NULL, 0, LVB_T_NONE);
1281 lock->l_remote_handle = dlm_req->lock_handle[0];
1282 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1284 /* Initialize resource lvb but not for a lock being replayed since
1285 * Client already got lvb sent in this case.
1286 * This must occur early since some policy methods assume resource
1287 * lvb is available (lr_lvb_data != NULL).
1289 res = lock->l_resource;
1290 if (!(flags & LDLM_FL_REPLAY)) {
1291 /* non-replayed lock, delayed lvb init may need to be done */
1292 rc = ldlm_lvbo_init(res);
1294 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1299 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1300 /* Don't enqueue a lock onto the export if it is been disonnected
1301 * due to eviction (bug 3822) or server umount (bug 24324).
1302 * Cancel it now instead. */
1303 if (req->rq_export->exp_disconnected) {
1304 LDLM_ERROR(lock, "lock on disconnected export %p",
1306 GOTO(out, rc = -ENOTCONN);
1309 lock->l_export = class_export_lock_get(req->rq_export, lock);
1310 if (lock->l_export->exp_lock_hash)
1311 cfs_hash_add(lock->l_export->exp_lock_hash,
1312 &lock->l_remote_handle,
1315 /* Inherit the enqueue flags before the operation, because we do not
1316 * keep the res lock on return and next operations (BL AST) may proceed
1318 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1319 LDLM_FL_INHERIT_MASK);
1322 if (flags & LDLM_FL_HAS_INTENT) {
1323 /* In this case, the reply buffer is allocated deep in
1324 * local_lock_enqueue by the policy function. */
1327 /* based on the assumption that lvb size never changes during
1328 * resource life time otherwise it need resource->lr_lock's
1330 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1331 RCL_SERVER, ldlm_lvbo_size(lock));
1333 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1334 GOTO(out, rc = -ENOMEM);
1336 rc = req_capsule_server_pack(&req->rq_pill);
1341 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1342 ldlm_convert_policy_to_local(req->rq_export,
1343 dlm_req->lock_desc.l_resource.lr_type,
1344 &dlm_req->lock_desc.l_policy_data,
1345 &lock->l_policy_data);
1346 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1347 lock->l_req_extent = lock->l_policy_data.l_extent;
1349 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1356 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1358 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1359 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1361 /* We never send a blocking AST until the lock is granted, but
1362 * we can tell it right now */
1363 lock_res_and_lock(lock);
1365 /* Now take into account flags to be inherited from original lock
1366 request both in reply to client and in our own lock flags. */
1367 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1368 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1370 /* Don't move a pending lock onto the export if it has already been
1371 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1372 * Cancel it now instead. */
1373 if (unlikely(req->rq_export->exp_disconnected ||
1374 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1375 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1377 } else if (ldlm_is_ast_sent(lock)) {
1378 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1379 if (lock->l_granted_mode == lock->l_req_mode) {
1381 * Only cancel lock if it was granted, because it would
1382 * be destroyed immediately and would never be granted
1383 * in the future, causing timeouts on client. Not
1384 * granted lock will be cancelled immediately after
1385 * sending completion AST.
1387 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1388 unlock_res_and_lock(lock);
1389 ldlm_lock_cancel(lock);
1390 lock_res_and_lock(lock);
1392 ldlm_add_waiting_lock(lock);
1395 /* Make sure we never ever grant usual metadata locks to liblustre
1397 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1398 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1399 req->rq_export->exp_libclient) {
1400 if (unlikely(!ldlm_is_cancel_on_block(lock) ||
1401 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1402 CERROR("Granting sync lock to libclient. "
1403 "req fl %d, rep fl %d, lock fl "LPX64"\n",
1404 dlm_req->lock_flags, dlm_rep->lock_flags,
1406 LDLM_ERROR(lock, "sync lock");
1407 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1408 struct ldlm_intent *it;
1410 it = req_capsule_client_get(&req->rq_pill,
1413 CERROR("This is intent %s ("LPU64")\n",
1414 ldlm_it2str(it->opc), it->opc);
1420 unlock_res_and_lock(lock);
1424 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1425 if (!req->rq_packed_final) {
1426 err = lustre_pack_reply(req, 1, NULL, NULL);
1431 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1432 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1434 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1435 "(err=%d, rc=%d)", err, rc);
1438 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1440 ldlm_lvbo_size(lock) > 0) {
1444 buf = req_capsule_server_get(&req->rq_pill,
1446 LASSERTF(buf != NULL, "req %p, lock %p\n",
1448 buflen = req_capsule_get_size(&req->rq_pill,
1449 &RMF_DLM_LVB, RCL_SERVER);
1450 /* non-replayed lock, delayed lvb init may
1451 * need to be occur now */
1452 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1453 buflen = ldlm_lvbo_fill(lock, buf,
1459 buflen, RCL_SERVER);
1462 } else if (flags & LDLM_FL_REPLAY) {
1463 /* no LVB resend upon replay */
1478 lock_res_and_lock(lock);
1479 ldlm_resource_unlink_lock(lock);
1480 ldlm_lock_destroy_nolock(lock);
1481 unlock_res_and_lock(lock);
1484 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1485 ldlm_reprocess_all(lock->l_resource);
1487 LDLM_LOCK_RELEASE(lock);
1490 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1497 * Old-style LDLM main entry point for server code enqueue.
1499 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1500 ldlm_completion_callback completion_callback,
1501 ldlm_blocking_callback blocking_callback,
1502 ldlm_glimpse_callback glimpse_callback)
1504 struct ldlm_request *dlm_req;
1505 struct ldlm_callback_suite cbs = {
1506 .lcs_completion = completion_callback,
1507 .lcs_blocking = blocking_callback,
1508 .lcs_glimpse = glimpse_callback
1512 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1513 if (dlm_req != NULL) {
1514 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1515 req, dlm_req, &cbs);
1523 * Main LDLM entry point for server code to process lock conversion requests.
1525 int ldlm_handle_convert0(struct ptlrpc_request *req,
1526 const struct ldlm_request *dlm_req)
1528 struct ldlm_reply *dlm_rep;
1529 struct ldlm_lock *lock;
1533 if (req->rq_export && req->rq_export->exp_nid_stats &&
1534 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1535 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1536 LDLM_CONVERT - LDLM_FIRST_OPC);
1538 rc = req_capsule_server_pack(&req->rq_pill);
1542 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1543 dlm_rep->lock_flags = dlm_req->lock_flags;
1545 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1547 req->rq_status = LUSTRE_EINVAL;
1551 LDLM_DEBUG(lock, "server-side convert handler START");
1553 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1554 &dlm_rep->lock_flags);
1556 if (ldlm_del_waiting_lock(lock))
1557 LDLM_DEBUG(lock, "converted waiting lock");
1560 req->rq_status = LUSTRE_EDEADLK;
1565 if (!req->rq_status)
1566 ldlm_reprocess_all(lock->l_resource);
1567 LDLM_DEBUG(lock, "server-side convert handler END");
1568 LDLM_LOCK_PUT(lock);
1570 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1576 * Old-style main LDLM entry point for server code to process lock conversion
1579 int ldlm_handle_convert(struct ptlrpc_request *req)
1582 struct ldlm_request *dlm_req;
1584 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1585 if (dlm_req != NULL) {
1586 rc = ldlm_handle_convert0(req, dlm_req);
1588 CERROR ("Can't unpack dlm_req\n");
1595 * Cancel all the locks whose handles are packed into ldlm_request
1597 * Called by server code expecting such combined cancel activity
1600 int ldlm_request_cancel(struct ptlrpc_request *req,
1601 const struct ldlm_request *dlm_req,
1602 int first, enum lustre_at_flags flags)
1604 struct ldlm_resource *res, *pres = NULL;
1605 struct ldlm_lock *lock;
1606 int i, count, done = 0;
1609 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1613 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1616 /* There is no lock on the server at the replay time,
1617 * skip lock cancelling to make replay tests to pass. */
1618 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1621 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1622 "starting at %d", count, first);
1624 for (i = first; i < count; i++) {
1625 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1627 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1628 "lock (cookie "LPU64")",
1629 dlm_req->lock_handle[i].cookie);
1633 res = lock->l_resource;
1636 /* This code is an optimization to only attempt lock
1637 * granting on the resource (that could be CPU-expensive)
1638 * after we are done cancelling lock in that resource. */
1641 ldlm_reprocess_all(pres);
1642 LDLM_RESOURCE_DELREF(pres);
1643 ldlm_resource_putref(pres);
1646 ldlm_resource_getref(res);
1647 LDLM_RESOURCE_ADDREF(res);
1648 ldlm_res_lvbo_update(res, NULL, 1);
1653 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock)) {
1654 long delay = cfs_time_sub(cfs_time_current_sec(),
1655 lock->l_last_activity);
1656 LDLM_DEBUG(lock, "server cancels blocked lock after "
1657 CFS_DURATION_T"s", delay);
1658 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1660 ldlm_lock_cancel(lock);
1661 LDLM_LOCK_PUT(lock);
1664 ldlm_reprocess_all(pres);
1665 LDLM_RESOURCE_DELREF(pres);
1666 ldlm_resource_putref(pres);
1668 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1671 EXPORT_SYMBOL(ldlm_request_cancel);
1674 * Main LDLM entry point for server code to cancel locks.
1676 * Typically gets called from service handler on LDLM_CANCEL opc.
1678 int ldlm_handle_cancel(struct ptlrpc_request *req)
1680 struct ldlm_request *dlm_req;
1684 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1685 if (dlm_req == NULL) {
1686 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1690 if (req->rq_export && req->rq_export->exp_nid_stats &&
1691 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1692 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1693 LDLM_CANCEL - LDLM_FIRST_OPC);
1695 rc = req_capsule_server_pack(&req->rq_pill);
1699 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1700 req->rq_status = LUSTRE_ESTALE;
1702 RETURN(ptlrpc_reply(req));
1704 #endif /* HAVE_SERVER_SUPPORT */
1707 * Callback handler for receiving incoming blocking ASTs.
1709 * This can only happen on client side.
1711 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1712 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1717 LDLM_DEBUG(lock, "client blocking AST callback handler");
1719 lock_res_and_lock(lock);
1720 ldlm_set_cbpending(lock);
1722 if (ldlm_is_cancel_on_block(lock))
1723 ldlm_set_cancel(lock);
1725 do_ast = (!lock->l_readers && !lock->l_writers);
1726 unlock_res_and_lock(lock);
1729 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1730 lock, lock->l_blocking_ast);
1731 if (lock->l_blocking_ast != NULL)
1732 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1735 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1739 LDLM_DEBUG(lock, "client blocking callback handler END");
1740 LDLM_LOCK_RELEASE(lock);
1745 * Callback handler for receiving incoming completion ASTs.
1747 * This only can happen on client side.
1749 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1750 struct ldlm_namespace *ns,
1751 struct ldlm_request *dlm_req,
1752 struct ldlm_lock *lock)
1754 struct list_head ast_list;
1759 LDLM_DEBUG(lock, "client completion callback handler START");
1761 INIT_LIST_HEAD(&ast_list);
1762 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1763 int to = cfs_time_seconds(1);
1765 set_current_state(TASK_INTERRUPTIBLE);
1766 schedule_timeout(to);
1767 if (lock->l_granted_mode == lock->l_req_mode ||
1768 ldlm_is_destroyed(lock))
1773 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1775 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1776 GOTO(out, rc = lvb_len);
1777 } else if (lvb_len > 0) {
1778 if (lock->l_lvb_len > 0) {
1779 /* for extent lock, lvb contains ost_lvb{}. */
1780 LASSERT(lock->l_lvb_data != NULL);
1782 if (unlikely(lock->l_lvb_len < lvb_len)) {
1783 LDLM_ERROR(lock, "Replied LVB is larger than "
1784 "expectation, expected = %d, "
1786 lock->l_lvb_len, lvb_len);
1787 GOTO(out, rc = -EINVAL);
1789 } else if (ldlm_has_layout(lock)) { /* for layout lock, lvb has
1790 * variable length */
1793 OBD_ALLOC_LARGE(lvb_data, lvb_len);
1794 if (lvb_data == NULL) {
1795 LDLM_ERROR(lock, "No memory: %d.\n", lvb_len);
1796 GOTO(out, rc = -ENOMEM);
1799 lock_res_and_lock(lock);
1800 LASSERT(lock->l_lvb_data == NULL);
1801 lock->l_lvb_type = LVB_T_LAYOUT;
1802 lock->l_lvb_data = lvb_data;
1803 lock->l_lvb_len = lvb_len;
1804 unlock_res_and_lock(lock);
1808 lock_res_and_lock(lock);
1809 if (ldlm_is_destroyed(lock) ||
1810 lock->l_granted_mode == lock->l_req_mode) {
1811 /* bug 11300: the lock has already been granted */
1812 unlock_res_and_lock(lock);
1813 LDLM_DEBUG(lock, "Double grant race happened");
1817 /* If we receive the completion AST before the actual enqueue returned,
1818 * then we might need to switch lock modes, resources, or extents. */
1819 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1820 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1821 LDLM_DEBUG(lock, "completion AST, new lock mode");
1824 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1825 ldlm_convert_policy_to_local(req->rq_export,
1826 dlm_req->lock_desc.l_resource.lr_type,
1827 &dlm_req->lock_desc.l_policy_data,
1828 &lock->l_policy_data);
1829 LDLM_DEBUG(lock, "completion AST, new policy data");
1832 ldlm_resource_unlink_lock(lock);
1833 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1834 &lock->l_resource->lr_name,
1835 sizeof(lock->l_resource->lr_name)) != 0) {
1836 unlock_res_and_lock(lock);
1837 rc = ldlm_lock_change_resource(ns, lock,
1838 &dlm_req->lock_desc.l_resource.lr_name);
1840 LDLM_ERROR(lock, "Failed to allocate resource");
1843 LDLM_DEBUG(lock, "completion AST, new resource");
1844 CERROR("change resource!\n");
1845 lock_res_and_lock(lock);
1848 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1849 /* BL_AST locks are not needed in LRU.
1850 * Let ldlm_cancel_lru() be fast. */
1851 ldlm_lock_remove_from_lru(lock);
1852 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1853 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1856 if (lock->l_lvb_len > 0) {
1857 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1858 lock->l_lvb_data, lvb_len);
1860 unlock_res_and_lock(lock);
1865 ldlm_grant_lock(lock, &ast_list);
1866 unlock_res_and_lock(lock);
1868 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1870 /* Let Enqueue to call osc_lock_upcall() and initialize
1872 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1874 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1876 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1882 lock_res_and_lock(lock);
1883 ldlm_set_failed(lock);
1884 unlock_res_and_lock(lock);
1885 wake_up(&lock->l_waitq);
1887 LDLM_LOCK_RELEASE(lock);
1891 * Callback handler for receiving incoming glimpse ASTs.
1893 * This only can happen on client side. After handling the glimpse AST
1894 * we also consider dropping the lock here if it is unused locally for a
1897 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1898 struct ldlm_namespace *ns,
1899 struct ldlm_request *dlm_req,
1900 struct ldlm_lock *lock)
1905 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1907 if (lock->l_glimpse_ast != NULL)
1908 rc = lock->l_glimpse_ast(lock, req);
1910 if (req->rq_repmsg != NULL) {
1913 req->rq_status = rc;
1917 lock_res_and_lock(lock);
1918 if (lock->l_granted_mode == LCK_PW &&
1919 !lock->l_readers && !lock->l_writers &&
1920 cfs_time_after(cfs_time_current(),
1921 cfs_time_add(lock->l_last_used,
1922 cfs_time_seconds(10)))) {
1923 unlock_res_and_lock(lock);
1924 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1925 ldlm_handle_bl_callback(ns, NULL, lock);
1930 unlock_res_and_lock(lock);
1931 LDLM_LOCK_RELEASE(lock);
1935 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1937 if (req->rq_no_reply)
1940 req->rq_status = rc;
1941 if (!req->rq_packed_final) {
1942 rc = lustre_pack_reply(req, 1, NULL, NULL);
1946 return ptlrpc_reply(req);
1949 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1950 ldlm_cancel_flags_t cancel_flags)
1952 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1955 spin_lock(&blp->blp_lock);
1956 if (blwi->blwi_lock &&
1957 ldlm_is_discard_data(blwi->blwi_lock)) {
1958 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1959 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1961 /* other blocking callbacks are added to the regular list */
1962 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1964 spin_unlock(&blp->blp_lock);
1966 wake_up(&blp->blp_waitq);
1968 /* can not check blwi->blwi_flags as blwi could be already freed in
1970 if (!(cancel_flags & LCF_ASYNC))
1971 wait_for_completion(&blwi->blwi_comp);
1976 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1977 struct ldlm_namespace *ns,
1978 struct ldlm_lock_desc *ld,
1979 struct list_head *cancels, int count,
1980 struct ldlm_lock *lock,
1981 ldlm_cancel_flags_t cancel_flags)
1983 init_completion(&blwi->blwi_comp);
1984 INIT_LIST_HEAD(&blwi->blwi_head);
1986 if (memory_pressure_get())
1987 blwi->blwi_mem_pressure = 1;
1990 blwi->blwi_flags = cancel_flags;
1992 blwi->blwi_ld = *ld;
1994 list_add(&blwi->blwi_head, cancels);
1995 list_del_init(cancels);
1996 blwi->blwi_count = count;
1998 blwi->blwi_lock = lock;
2003 * Queues a list of locks \a cancels containing \a count locks
2004 * for later processing by a blocking thread. If \a count is zero,
2005 * then the lock referenced as \a lock is queued instead.
2007 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2008 * If list addition fails an error is returned and caller is supposed to
2009 * call ->l_blocking_ast itself.
2011 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2012 struct ldlm_lock_desc *ld,
2013 struct ldlm_lock *lock,
2014 struct list_head *cancels, int count,
2015 ldlm_cancel_flags_t cancel_flags)
2019 if (cancels && count == 0)
2022 if (cancel_flags & LCF_ASYNC) {
2023 struct ldlm_bl_work_item *blwi;
2025 OBD_ALLOC(blwi, sizeof(*blwi));
2028 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2030 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2032 /* if it is synchronous call do minimum mem alloc, as it could
2033 * be triggered from kernel shrinker
2035 struct ldlm_bl_work_item blwi;
2037 memset(&blwi, 0, sizeof(blwi));
2038 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2039 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2044 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2045 struct ldlm_lock *lock)
2047 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2050 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2051 struct list_head *cancels, int count,
2052 ldlm_cancel_flags_t cancel_flags)
2054 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2057 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2058 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2060 struct obd_device *obd = req->rq_export->exp_obd;
2067 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2069 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2071 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2073 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2076 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2078 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2080 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2083 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2086 /* We are responsible for swabbing contents of val */
2088 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2089 /* Pass it on to mdc (the "export" in this case) */
2090 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2092 sizeof(KEY_HSM_COPYTOOL_SEND),
2093 KEY_HSM_COPYTOOL_SEND,
2096 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2101 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2102 const char *msg, int rc,
2103 struct lustre_handle *handle)
2105 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2106 "%s: [nid %s] [rc %d] [lock "LPX64"]",
2107 msg, libcfs_id2str(req->rq_peer), rc,
2108 handle ? handle->cookie : 0);
2109 if (req->rq_no_reply)
2110 CWARN("No reply was sent, maybe cause bug 21636.\n");
2112 CWARN("Send reply failed, maybe cause bug 21636.\n");
2115 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2116 static int ldlm_callback_handler(struct ptlrpc_request *req)
2118 struct ldlm_namespace *ns;
2119 struct ldlm_request *dlm_req;
2120 struct ldlm_lock *lock;
2124 /* Requests arrive in sender's byte order. The ptlrpc service
2125 * handler has already checked and, if necessary, byte-swapped the
2126 * incoming request message body, but I am responsible for the
2127 * message buffers. */
2129 /* do nothing for sec context finalize */
2130 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2133 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2135 if (req->rq_export == NULL) {
2136 rc = ldlm_callback_reply(req, -ENOTCONN);
2137 ldlm_callback_errmsg(req, "Operate on unconnected server",
2142 LASSERT(req->rq_export != NULL);
2143 LASSERT(req->rq_export->exp_obd != NULL);
2145 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2146 case LDLM_BL_CALLBACK:
2147 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2149 ldlm_callback_reply(req, -(int)cfs_fail_err);
2153 case LDLM_CP_CALLBACK:
2154 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2157 case LDLM_GL_CALLBACK:
2158 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2162 rc = ldlm_handle_setinfo(req);
2163 ldlm_callback_reply(req, rc);
2165 case LLOG_ORIGIN_HANDLE_CREATE:
2166 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2167 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2169 rc = llog_origin_handle_open(req);
2170 ldlm_callback_reply(req, rc);
2172 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2173 req_capsule_set(&req->rq_pill,
2174 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2175 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2177 rc = llog_origin_handle_next_block(req);
2178 ldlm_callback_reply(req, rc);
2180 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2181 req_capsule_set(&req->rq_pill,
2182 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2183 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2185 rc = llog_origin_handle_read_header(req);
2186 ldlm_callback_reply(req, rc);
2188 case LLOG_ORIGIN_HANDLE_CLOSE:
2189 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2191 rc = llog_origin_handle_close(req);
2192 ldlm_callback_reply(req, rc);
2195 CERROR("unknown opcode %u\n",
2196 lustre_msg_get_opc(req->rq_reqmsg));
2197 ldlm_callback_reply(req, -EPROTO);
2201 ns = req->rq_export->exp_obd->obd_namespace;
2202 LASSERT(ns != NULL);
2204 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2206 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2207 if (dlm_req == NULL) {
2208 rc = ldlm_callback_reply(req, -EPROTO);
2209 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2214 /* Force a known safe race, send a cancel to the server for a lock
2215 * which the server has already started a blocking callback on. */
2216 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2217 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2218 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2220 CERROR("ldlm_cli_cancel: %d\n", rc);
2223 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2225 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2226 "disappeared\n", dlm_req->lock_handle[0].cookie);
2227 rc = ldlm_callback_reply(req, -EINVAL);
2228 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2229 &dlm_req->lock_handle[0]);
2233 if (ldlm_is_fail_loc(lock) &&
2234 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2235 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2237 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2238 lock_res_and_lock(lock);
2239 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2241 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2242 /* If somebody cancels lock and cache is already dropped,
2243 * or lock is failed before cp_ast received on client,
2244 * we can tell the server we have no lock. Otherwise, we
2245 * should send cancel after dropping the cache. */
2246 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2247 ldlm_is_failed(lock)) {
2248 LDLM_DEBUG(lock, "callback on lock "
2249 LPX64" - lock disappeared\n",
2250 dlm_req->lock_handle[0].cookie);
2251 unlock_res_and_lock(lock);
2252 LDLM_LOCK_RELEASE(lock);
2253 rc = ldlm_callback_reply(req, -EINVAL);
2254 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2255 &dlm_req->lock_handle[0]);
2258 /* BL_AST locks are not needed in LRU.
2259 * Let ldlm_cancel_lru() be fast. */
2260 ldlm_lock_remove_from_lru(lock);
2261 ldlm_set_bl_ast(lock);
2263 unlock_res_and_lock(lock);
2265 /* We want the ost thread to get this reply so that it can respond
2266 * to ost requests (write cache writeback) that might be triggered
2269 * But we'd also like to be able to indicate in the reply that we're
2270 * cancelling right now, because it's unused, or have an intent result
2271 * in the reply, so we might have to push the responsibility for sending
2272 * the reply down into the AST handlers, alas. */
2274 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2275 case LDLM_BL_CALLBACK:
2276 CDEBUG(D_INODE, "blocking ast\n");
2277 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2278 if (!ldlm_is_cancel_on_block(lock)) {
2279 rc = ldlm_callback_reply(req, 0);
2280 if (req->rq_no_reply || rc)
2281 ldlm_callback_errmsg(req, "Normal process", rc,
2282 &dlm_req->lock_handle[0]);
2284 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2285 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2287 case LDLM_CP_CALLBACK:
2288 CDEBUG(D_INODE, "completion ast\n");
2289 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2290 ldlm_callback_reply(req, 0);
2291 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2293 case LDLM_GL_CALLBACK:
2294 CDEBUG(D_INODE, "glimpse ast\n");
2295 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2296 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2299 LBUG(); /* checked above */
2305 #ifdef HAVE_SERVER_SUPPORT
2307 * Main handler for canceld thread.
2309 * Separated into its own thread to avoid deadlocks.
2311 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2316 /* Requests arrive in sender's byte order. The ptlrpc service
2317 * handler has already checked and, if necessary, byte-swapped the
2318 * incoming request message body, but I am responsible for the
2319 * message buffers. */
2321 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2323 if (req->rq_export == NULL) {
2324 struct ldlm_request *dlm_req;
2326 CERROR("%s from %s arrived at %lu with bad export cookie "
2328 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2329 libcfs_nid2str(req->rq_peer.nid),
2330 req->rq_arrival_time.tv_sec,
2331 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2333 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2334 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2335 dlm_req = req_capsule_client_get(&req->rq_pill,
2337 if (dlm_req != NULL)
2338 ldlm_lock_dump_handle(D_ERROR,
2339 &dlm_req->lock_handle[0]);
2341 ldlm_callback_reply(req, -ENOTCONN);
2345 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2347 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2349 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2350 CDEBUG(D_INODE, "cancel\n");
2351 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2352 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))
2354 rc = ldlm_handle_cancel(req);
2359 CERROR("invalid opcode %d\n",
2360 lustre_msg_get_opc(req->rq_reqmsg));
2361 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2362 ldlm_callback_reply(req, -EINVAL);
2368 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2369 struct ldlm_lock *lock)
2371 struct ldlm_request *dlm_req;
2372 struct lustre_handle lockh;
2377 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2378 if (dlm_req == NULL)
2381 ldlm_lock2handle(lock, &lockh);
2382 for (i = 0; i < dlm_req->lock_count; i++) {
2383 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2385 DEBUG_REQ(D_RPCTRACE, req,
2386 "Prio raised by lock "LPX64".", lockh.cookie);
2397 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2399 struct ldlm_request *dlm_req;
2404 /* no prolong in recovery */
2405 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2408 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2409 if (dlm_req == NULL)
2412 for (i = 0; i < dlm_req->lock_count; i++) {
2413 struct ldlm_lock *lock;
2415 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2419 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2421 LDLM_DEBUG(lock, "hpreq cancel lock");
2422 LDLM_LOCK_PUT(lock);
2431 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2432 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2433 .hpreq_check = ldlm_cancel_hpreq_check,
2437 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2441 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2443 if (req->rq_export == NULL)
2446 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2447 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2448 req->rq_ops = &ldlm_cancel_hpreq_ops;
2453 static int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2454 struct hlist_node *hnode, void *data)
2457 struct list_head *rpc_list = data;
2458 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2460 lock_res_and_lock(lock);
2462 if (lock->l_req_mode != lock->l_granted_mode) {
2463 unlock_res_and_lock(lock);
2467 LASSERT(lock->l_resource);
2468 if (lock->l_resource->lr_type != LDLM_IBITS &&
2469 lock->l_resource->lr_type != LDLM_PLAIN) {
2470 unlock_res_and_lock(lock);
2474 if (ldlm_is_ast_sent(lock)) {
2475 unlock_res_and_lock(lock);
2479 LASSERT(lock->l_blocking_ast);
2480 LASSERT(!lock->l_blocking_lock);
2482 ldlm_set_ast_sent(lock);
2483 if (lock->l_export && lock->l_export->exp_lock_hash) {
2484 /* NB: it's safe to call cfs_hash_del() even lock isn't
2485 * in exp_lock_hash. */
2486 /* In the function below, .hs_keycmp resolves to
2487 * ldlm_export_lock_keycmp() */
2488 /* coverity[overrun-buffer-val] */
2489 cfs_hash_del(lock->l_export->exp_lock_hash,
2490 &lock->l_remote_handle, &lock->l_exp_hash);
2493 list_add_tail(&lock->l_rk_ast, rpc_list);
2494 LDLM_LOCK_GET(lock);
2496 unlock_res_and_lock(lock);
2500 void ldlm_revoke_export_locks(struct obd_export *exp)
2502 struct list_head rpc_list;
2505 INIT_LIST_HEAD(&rpc_list);
2506 cfs_hash_for_each_empty(exp->exp_lock_hash,
2507 ldlm_revoke_lock_cb, &rpc_list);
2508 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2509 LDLM_WORK_REVOKE_AST);
2513 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2514 #endif /* HAVE_SERVER_SUPPORT */
2516 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2518 struct ldlm_bl_work_item *blwi = NULL;
2519 static unsigned int num_bl = 0;
2521 spin_lock(&blp->blp_lock);
2522 /* process a request from the blp_list at least every blp_num_threads */
2523 if (!list_empty(&blp->blp_list) &&
2524 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2525 blwi = list_entry(blp->blp_list.next,
2526 struct ldlm_bl_work_item, blwi_entry);
2528 if (!list_empty(&blp->blp_prio_list))
2529 blwi = list_entry(blp->blp_prio_list.next,
2530 struct ldlm_bl_work_item,
2534 if (++num_bl >= atomic_read(&blp->blp_num_threads))
2536 list_del(&blwi->blwi_entry);
2538 spin_unlock(&blp->blp_lock);
2543 /* This only contains temporary data until the thread starts */
2544 struct ldlm_bl_thread_data {
2545 char bltd_name[CFS_CURPROC_COMM_MAX];
2546 struct ldlm_bl_pool *bltd_blp;
2547 struct completion bltd_comp;
2551 static int ldlm_bl_thread_main(void *arg);
2553 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2555 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2556 struct task_struct *task;
2558 init_completion(&bltd.bltd_comp);
2559 bltd.bltd_num = atomic_read(&blp->blp_num_threads);
2560 snprintf(bltd.bltd_name, sizeof(bltd.bltd_name) - 1,
2561 "ldlm_bl_%02d", bltd.bltd_num);
2562 task = kthread_run(ldlm_bl_thread_main, &bltd, bltd.bltd_name);
2564 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2565 atomic_read(&blp->blp_num_threads), PTR_ERR(task));
2566 return PTR_ERR(task);
2568 wait_for_completion(&bltd.bltd_comp);
2574 * Main blocking requests processing thread.
2576 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2577 * This thread in the end ends up doing actual call to ->l_blocking_ast
2580 static int ldlm_bl_thread_main(void *arg)
2582 struct ldlm_bl_pool *blp;
2583 struct ldlm_bl_thread_data *bltd = arg;
2586 blp = bltd->bltd_blp;
2588 atomic_inc(&blp->blp_num_threads);
2589 atomic_inc(&blp->blp_busy_threads);
2591 complete(&bltd->bltd_comp);
2592 /* cannot use bltd after this, it is only on caller's stack */
2595 struct l_wait_info lwi = { 0 };
2596 struct ldlm_bl_work_item *blwi = NULL;
2599 blwi = ldlm_bl_get_work(blp);
2602 atomic_dec(&blp->blp_busy_threads);
2603 l_wait_event_exclusive(blp->blp_waitq,
2604 (blwi = ldlm_bl_get_work(blp)) != NULL,
2606 busy = atomic_inc_return(&blp->blp_busy_threads);
2608 busy = atomic_read(&blp->blp_busy_threads);
2611 if (blwi->blwi_ns == NULL)
2612 /* added by ldlm_cleanup() */
2615 /* Not fatal if racy and have a few too many threads */
2616 if (unlikely(busy < blp->blp_max_threads &&
2617 busy >= atomic_read(&blp->blp_num_threads) &&
2618 !blwi->blwi_mem_pressure))
2619 /* discard the return value, we tried */
2620 ldlm_bl_thread_start(blp);
2622 if (blwi->blwi_mem_pressure)
2623 memory_pressure_set();
2625 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2627 if (blwi->blwi_count) {
2629 /* The special case when we cancel locks in LRU
2630 * asynchronously, we pass the list of locks here.
2631 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2632 * canceled locally yet. */
2633 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2636 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2639 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2642 if (blwi->blwi_mem_pressure)
2643 memory_pressure_clr();
2645 if (blwi->blwi_flags & LCF_ASYNC)
2646 OBD_FREE(blwi, sizeof(*blwi));
2648 complete(&blwi->blwi_comp);
2651 atomic_dec(&blp->blp_busy_threads);
2652 atomic_dec(&blp->blp_num_threads);
2653 complete(&blp->blp_comp);
2658 static int ldlm_setup(void);
2659 static int ldlm_cleanup(void);
2661 int ldlm_get_ref(void)
2665 mutex_lock(&ldlm_ref_mutex);
2666 if (++ldlm_refcount == 1) {
2671 mutex_unlock(&ldlm_ref_mutex);
2676 void ldlm_put_ref(void)
2679 mutex_lock(&ldlm_ref_mutex);
2680 if (ldlm_refcount == 1) {
2681 int rc = ldlm_cleanup();
2683 CERROR("ldlm_cleanup failed: %d\n", rc);
2689 mutex_unlock(&ldlm_ref_mutex);
2695 * Export handle<->lock hash operations.
2698 ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
2700 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2704 ldlm_export_lock_key(struct hlist_node *hnode)
2706 struct ldlm_lock *lock;
2708 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2709 return &lock->l_remote_handle;
2713 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2715 struct ldlm_lock *lock;
2717 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2718 lock->l_remote_handle = *(struct lustre_handle *)key;
2722 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2724 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2728 ldlm_export_lock_object(struct hlist_node *hnode)
2730 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2734 ldlm_export_lock_get(cfs_hash_t *hs, struct hlist_node *hnode)
2736 struct ldlm_lock *lock;
2738 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2739 LDLM_LOCK_GET(lock);
2743 ldlm_export_lock_put(cfs_hash_t *hs, struct hlist_node *hnode)
2745 struct ldlm_lock *lock;
2747 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2748 LDLM_LOCK_RELEASE(lock);
2751 static cfs_hash_ops_t ldlm_export_lock_ops = {
2752 .hs_hash = ldlm_export_lock_hash,
2753 .hs_key = ldlm_export_lock_key,
2754 .hs_keycmp = ldlm_export_lock_keycmp,
2755 .hs_keycpy = ldlm_export_lock_keycpy,
2756 .hs_object = ldlm_export_lock_object,
2757 .hs_get = ldlm_export_lock_get,
2758 .hs_put = ldlm_export_lock_put,
2759 .hs_put_locked = ldlm_export_lock_put,
2762 int ldlm_init_export(struct obd_export *exp)
2767 exp->exp_lock_hash =
2768 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2769 HASH_EXP_LOCK_CUR_BITS,
2770 HASH_EXP_LOCK_MAX_BITS,
2771 HASH_EXP_LOCK_BKT_BITS, 0,
2772 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2773 &ldlm_export_lock_ops,
2774 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2775 CFS_HASH_NBLK_CHANGE);
2777 if (!exp->exp_lock_hash)
2780 rc = ldlm_init_flock_export(exp);
2786 ldlm_destroy_export(exp);
2789 EXPORT_SYMBOL(ldlm_init_export);
2791 void ldlm_destroy_export(struct obd_export *exp)
2794 cfs_hash_putref(exp->exp_lock_hash);
2795 exp->exp_lock_hash = NULL;
2797 ldlm_destroy_flock_export(exp);
2800 EXPORT_SYMBOL(ldlm_destroy_export);
2802 static int ldlm_setup(void)
2804 static struct ptlrpc_service_conf conf;
2805 struct ldlm_bl_pool *blp = NULL;
2806 #ifdef HAVE_SERVER_SUPPORT
2807 struct task_struct *task;
2808 #endif /* HAVE_SERVER_SUPPORT */
2814 if (ldlm_state != NULL)
2817 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2818 if (ldlm_state == NULL)
2821 #ifdef CONFIG_PROC_FS
2822 rc = ldlm_proc_setup();
2825 #endif /* CONFIG_PROC_FS */
2827 memset(&conf, 0, sizeof(conf));
2828 conf = (typeof(conf)) {
2829 .psc_name = "ldlm_cbd",
2830 .psc_watchdog_factor = 2,
2832 .bc_nbufs = LDLM_CLIENT_NBUFS,
2833 .bc_buf_size = LDLM_BUFSIZE,
2834 .bc_req_max_size = LDLM_MAXREQSIZE,
2835 .bc_rep_max_size = LDLM_MAXREPSIZE,
2836 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2837 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2840 .tc_thr_name = "ldlm_cb",
2841 .tc_thr_factor = LDLM_THR_FACTOR,
2842 .tc_nthrs_init = LDLM_NTHRS_INIT,
2843 .tc_nthrs_base = LDLM_NTHRS_BASE,
2844 .tc_nthrs_max = LDLM_NTHRS_MAX,
2845 .tc_nthrs_user = ldlm_num_threads,
2846 .tc_cpu_affinity = 1,
2847 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2850 .cc_pattern = ldlm_cpts,
2853 .so_req_handler = ldlm_callback_handler,
2856 ldlm_state->ldlm_cb_service = \
2857 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2858 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2859 CERROR("failed to start service\n");
2860 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2861 ldlm_state->ldlm_cb_service = NULL;
2865 #ifdef HAVE_SERVER_SUPPORT
2866 memset(&conf, 0, sizeof(conf));
2867 conf = (typeof(conf)) {
2868 .psc_name = "ldlm_canceld",
2869 .psc_watchdog_factor = 6,
2871 .bc_nbufs = LDLM_SERVER_NBUFS,
2872 .bc_buf_size = LDLM_BUFSIZE,
2873 .bc_req_max_size = LDLM_MAXREQSIZE,
2874 .bc_rep_max_size = LDLM_MAXREPSIZE,
2875 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
2876 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
2880 .tc_thr_name = "ldlm_cn",
2881 .tc_thr_factor = LDLM_THR_FACTOR,
2882 .tc_nthrs_init = LDLM_NTHRS_INIT,
2883 .tc_nthrs_base = LDLM_NTHRS_BASE,
2884 .tc_nthrs_max = LDLM_NTHRS_MAX,
2885 .tc_nthrs_user = ldlm_num_threads,
2886 .tc_cpu_affinity = 1,
2887 .tc_ctx_tags = LCT_MD_THREAD | \
2892 .cc_pattern = ldlm_cpts,
2895 .so_req_handler = ldlm_cancel_handler,
2896 .so_hpreq_handler = ldlm_hpreq_handler,
2899 ldlm_state->ldlm_cancel_service = \
2900 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2901 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2902 CERROR("failed to start service\n");
2903 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2904 ldlm_state->ldlm_cancel_service = NULL;
2907 #endif /* HAVE_SERVER_SUPPORT */
2909 OBD_ALLOC(blp, sizeof(*blp));
2911 GOTO(out, rc = -ENOMEM);
2912 ldlm_state->ldlm_bl_pool = blp;
2914 spin_lock_init(&blp->blp_lock);
2915 INIT_LIST_HEAD(&blp->blp_list);
2916 INIT_LIST_HEAD(&blp->blp_prio_list);
2917 init_waitqueue_head(&blp->blp_waitq);
2918 atomic_set(&blp->blp_num_threads, 0);
2919 atomic_set(&blp->blp_busy_threads, 0);
2921 if (ldlm_num_threads == 0) {
2922 blp->blp_min_threads = LDLM_NTHRS_INIT;
2923 blp->blp_max_threads = LDLM_NTHRS_MAX;
2925 blp->blp_min_threads = blp->blp_max_threads = \
2926 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
2930 for (i = 0; i < blp->blp_min_threads; i++) {
2931 rc = ldlm_bl_thread_start(blp);
2936 #ifdef HAVE_SERVER_SUPPORT
2937 INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2938 expired_lock_thread.elt_state = ELT_STOPPED;
2939 init_waitqueue_head(&expired_lock_thread.elt_waitq);
2941 INIT_LIST_HEAD(&waiting_locks_list);
2942 spin_lock_init(&waiting_locks_spinlock);
2943 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, NULL);
2945 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
2948 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2952 wait_event(expired_lock_thread.elt_waitq,
2953 expired_lock_thread.elt_state == ELT_READY);
2954 #endif /* HAVE_SERVER_SUPPORT */
2956 rc = ldlm_pools_init();
2958 CERROR("Failed to initialize LDLM pools: %d\n", rc);
2968 static int ldlm_cleanup(void)
2972 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2973 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2974 CERROR("ldlm still has namespaces; clean these up first.\n");
2975 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2976 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2982 if (ldlm_state->ldlm_bl_pool != NULL) {
2983 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2985 while (atomic_read(&blp->blp_num_threads) > 0) {
2986 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2988 init_completion(&blp->blp_comp);
2990 spin_lock(&blp->blp_lock);
2991 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2992 wake_up(&blp->blp_waitq);
2993 spin_unlock(&blp->blp_lock);
2995 wait_for_completion(&blp->blp_comp);
2998 OBD_FREE(blp, sizeof(*blp));
3001 if (ldlm_state->ldlm_cb_service != NULL)
3002 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3003 #ifdef HAVE_SERVER_SUPPORT
3004 if (ldlm_state->ldlm_cancel_service != NULL)
3005 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3008 ldlm_proc_cleanup();
3010 #ifdef HAVE_SERVER_SUPPORT
3011 if (expired_lock_thread.elt_state != ELT_STOPPED) {
3012 expired_lock_thread.elt_state = ELT_TERMINATE;
3013 wake_up(&expired_lock_thread.elt_waitq);
3014 wait_event(expired_lock_thread.elt_waitq,
3015 expired_lock_thread.elt_state == ELT_STOPPED);
3019 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3027 mutex_init(&ldlm_ref_mutex);
3028 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
3029 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
3031 INIT_LIST_HEAD(&ldlm_srv_namespace_list);
3032 INIT_LIST_HEAD(&ldlm_cli_active_namespace_list);
3033 INIT_LIST_HEAD(&ldlm_cli_inactive_namespace_list);
3035 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3036 sizeof(struct ldlm_resource), 0,
3037 SLAB_HWCACHE_ALIGN, NULL);
3038 if (ldlm_resource_slab == NULL)
3041 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3042 sizeof(struct ldlm_lock), 0,
3043 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3044 if (ldlm_lock_slab == NULL) {
3045 kmem_cache_destroy(ldlm_resource_slab);
3049 ldlm_interval_slab = kmem_cache_create("interval_node",
3050 sizeof(struct ldlm_interval),
3051 0, SLAB_HWCACHE_ALIGN, NULL);
3052 if (ldlm_interval_slab == NULL) {
3053 kmem_cache_destroy(ldlm_resource_slab);
3054 kmem_cache_destroy(ldlm_lock_slab);
3057 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3058 class_export_dump_hook = ldlm_dump_export_locks;
3063 void ldlm_exit(void)
3066 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3067 kmem_cache_destroy(ldlm_resource_slab);
3068 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3069 * synchronize_rcu() to wait a grace period elapsed, so that
3070 * ldlm_lock_free() get a chance to be called. */
3072 kmem_cache_destroy(ldlm_lock_slab);
3073 kmem_cache_destroy(ldlm_interval_slab);