1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
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>
43 # define EXPORT_SYMTAB
45 #define DEBUG_SUBSYSTEM S_LDLM
48 # include <libcfs/libcfs.h>
50 # include <liblustre.h>
53 #include <lustre_dlm.h>
54 #include <obd_class.h>
55 #include <libcfs/list.h>
56 #include "ldlm_internal.h"
59 static int ldlm_num_threads;
60 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
61 "number of DLM service threads to start");
64 extern cfs_mem_cache_t *ldlm_resource_slab;
65 extern cfs_mem_cache_t *ldlm_lock_slab;
66 static struct semaphore ldlm_ref_sem;
67 static int ldlm_refcount;
71 static struct ldlm_state *ldlm_state;
73 inline cfs_time_t round_timeout(cfs_time_t timeout)
75 return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
78 /* timeout for initial callback (AST) reply (bz10399) */
79 static inline unsigned int ldlm_get_rq_timeout(void)
82 unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
84 return timeout < 1 ? 1 : timeout;
88 /* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */
89 static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */
90 static struct list_head waiting_locks_list;
91 static cfs_timer_t waiting_locks_timer;
93 static struct expired_lock_thread {
94 cfs_waitq_t elt_waitq;
97 struct list_head elt_expired_locks;
98 } expired_lock_thread;
101 #define ELT_STOPPED 0
103 #define ELT_TERMINATE 2
105 struct ldlm_bl_pool {
109 * blp_prio_list is used for callbacks that should be handled
110 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
113 struct list_head blp_prio_list;
116 * blp_list is used for all other callbacks which are likely
117 * to take longer to process.
119 struct list_head blp_list;
121 cfs_waitq_t blp_waitq;
122 struct completion blp_comp;
123 atomic_t blp_num_threads;
124 atomic_t blp_busy_threads;
129 struct ldlm_bl_work_item {
130 struct list_head blwi_entry;
131 struct ldlm_namespace *blwi_ns;
132 struct ldlm_lock_desc blwi_ld;
133 struct ldlm_lock *blwi_lock;
134 struct list_head blwi_head;
140 static inline int have_expired_locks(void)
145 spin_lock_bh(&waiting_locks_spinlock);
146 need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
147 spin_unlock_bh(&waiting_locks_spinlock);
152 static int expired_lock_main(void *arg)
154 struct list_head *expired = &expired_lock_thread.elt_expired_locks;
155 struct l_wait_info lwi = { 0 };
159 cfs_daemonize("ldlm_elt");
161 expired_lock_thread.elt_state = ELT_READY;
162 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
165 l_wait_event(expired_lock_thread.elt_waitq,
166 have_expired_locks() ||
167 expired_lock_thread.elt_state == ELT_TERMINATE,
170 spin_lock_bh(&waiting_locks_spinlock);
171 if (expired_lock_thread.elt_dump) {
172 spin_unlock_bh(&waiting_locks_spinlock);
174 /* from waiting_locks_callback, but not in timer */
175 libcfs_debug_dumplog();
176 libcfs_run_lbug_upcall(__FILE__,
177 "waiting_locks_callback",
178 expired_lock_thread.elt_dump);
180 spin_lock_bh(&waiting_locks_spinlock);
181 expired_lock_thread.elt_dump = 0;
186 while (!list_empty(expired)) {
187 struct obd_export *export;
188 struct ldlm_lock *lock;
190 lock = list_entry(expired->next, struct ldlm_lock,
192 if ((void *)lock < LP_POISON + CFS_PAGE_SIZE &&
193 (void *)lock >= LP_POISON) {
194 spin_unlock_bh(&waiting_locks_spinlock);
195 CERROR("free lock on elt list %p\n", lock);
198 list_del_init(&lock->l_pending_chain);
199 if ((void *)lock->l_export < LP_POISON + CFS_PAGE_SIZE &&
200 (void *)lock->l_export >= LP_POISON) {
201 CERROR("lock with free export on elt list %p\n",
203 lock->l_export = NULL;
204 LDLM_ERROR(lock, "free export");
207 export = class_export_get(lock->l_export);
208 spin_unlock_bh(&waiting_locks_spinlock);
211 class_fail_export(export);
212 class_export_put(export);
213 spin_lock_bh(&waiting_locks_spinlock);
215 spin_unlock_bh(&waiting_locks_spinlock);
217 if (do_dump && obd_dump_on_eviction) {
218 CERROR("dump the log upon eviction\n");
219 libcfs_debug_dumplog();
222 if (expired_lock_thread.elt_state == ELT_TERMINATE)
226 expired_lock_thread.elt_state = ELT_STOPPED;
227 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
231 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
233 /* This is called from within a timer interrupt and cannot schedule */
234 static void waiting_locks_callback(unsigned long unused)
236 struct ldlm_lock *lock, *last = NULL;
239 spin_lock_bh(&waiting_locks_spinlock);
240 while (!list_empty(&waiting_locks_list)) {
241 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
244 if (cfs_time_after(lock->l_callback_timeout, cfs_time_current()) ||
245 (lock->l_req_mode == LCK_GROUP))
248 if (ptlrpc_check_suspend()) {
249 /* there is a case when we talk to one mds, holding
250 * lock from another mds. this way we easily can get
251 * here, if second mds is being recovered. so, we
252 * suspend timeouts. bug 6019 */
254 LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
255 lock->l_export->exp_client_uuid.uuid,
256 lock->l_export->exp_connection->c_remote_uuid.uuid,
257 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
259 list_del_init(&lock->l_pending_chain);
260 spin_unlock_bh(&waiting_locks_spinlock);
261 ldlm_add_waiting_lock(lock);
265 /* if timeout overlaps the activation time of suspended timeouts
266 * then extend it to give a chance for client to reconnect */
267 if (cfs_time_before(cfs_time_sub(lock->l_callback_timeout,
268 cfs_time_seconds(obd_timeout)/2),
269 ptlrpc_suspend_wakeup_time())) {
270 LDLM_ERROR(lock, "extend timeout due to recovery: %s@%s nid %s ",
271 lock->l_export->exp_client_uuid.uuid,
272 lock->l_export->exp_connection->c_remote_uuid.uuid,
273 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
275 list_del_init(&lock->l_pending_chain);
276 spin_unlock_bh(&waiting_locks_spinlock);
277 ldlm_add_waiting_lock(lock);
281 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
282 "evicting client at %s ",
283 cfs_time_current_sec()- lock->l_enqueued_time.tv_sec,
285 lock->l_export->exp_connection->c_peer.nid));
289 list_del(&lock->l_pending_chain);
290 list_add(&lock->l_pending_chain,
291 &expired_lock_thread.elt_expired_locks);
294 if (!list_empty(&expired_lock_thread.elt_expired_locks)) {
295 if (obd_dump_on_timeout)
296 expired_lock_thread.elt_dump = __LINE__;
298 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
302 * Make sure the timer will fire again if we have any locks
305 if (!list_empty(&waiting_locks_list)) {
306 cfs_time_t timeout_rounded;
307 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
309 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
310 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
312 spin_unlock_bh(&waiting_locks_spinlock);
316 * Indicate that we're waiting for a client to call us back cancelling a given
317 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
318 * timer to fire appropriately. (We round up to the next second, to avoid
319 * floods of timer firings during periods of high lock contention and traffic).
321 * Called with the namespace lock held.
323 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock)
326 cfs_time_t timeout_rounded;
328 if (!list_empty(&lock->l_pending_chain))
331 timeout = ldlm_get_enq_timeout(lock);
333 lock->l_callback_timeout = cfs_time_shift(timeout);
335 timeout_rounded = round_timeout(lock->l_callback_timeout);
337 if (cfs_time_before(timeout_rounded,
338 cfs_timer_deadline(&waiting_locks_timer)) ||
339 !cfs_timer_is_armed(&waiting_locks_timer)) {
340 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
342 /* if the new lock has a shorter timeout than something earlier on
343 the list, we'll wait the longer amount of time; no big deal. */
344 list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
348 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
352 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
354 spin_lock_bh(&waiting_locks_spinlock);
355 if (lock->l_destroyed) {
356 static cfs_time_t next;
357 spin_unlock_bh(&waiting_locks_spinlock);
358 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
359 if (cfs_time_after(cfs_time_current(), next)) {
360 next = cfs_time_shift(14400);
361 libcfs_debug_dumpstack(NULL);
366 ret = __ldlm_add_waiting_lock(lock);
367 spin_unlock_bh(&waiting_locks_spinlock);
369 LDLM_DEBUG(lock, "%sadding to wait list",
370 ret == 0 ? "not re-" : "");
375 * Remove a lock from the pending list, likely because it had its cancellation
376 * callback arrive without incident. This adjusts the lock-timeout timer if
377 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
379 * Called with namespace lock held.
381 int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
383 struct list_head *list_next;
385 if (list_empty(&lock->l_pending_chain))
388 list_next = lock->l_pending_chain.next;
389 if (lock->l_pending_chain.prev == &waiting_locks_list) {
390 /* Removing the head of the list, adjust timer. */
391 if (list_next == &waiting_locks_list) {
392 /* No more, just cancel. */
393 cfs_timer_disarm(&waiting_locks_timer);
395 struct ldlm_lock *next;
396 next = list_entry(list_next, struct ldlm_lock,
398 cfs_timer_arm(&waiting_locks_timer,
399 round_timeout(next->l_callback_timeout));
402 list_del_init(&lock->l_pending_chain);
407 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
411 if (lock->l_export == NULL) {
412 /* We don't have a "waiting locks list" on clients. */
413 LDLM_DEBUG(lock, "client lock: no-op");
417 spin_lock_bh(&waiting_locks_spinlock);
418 ret = __ldlm_del_waiting_lock(lock);
419 spin_unlock_bh(&waiting_locks_spinlock);
421 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
428 * Called with namespace lock held.
430 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
432 if (lock->l_export == NULL) {
433 /* We don't have a "waiting locks list" on clients. */
434 LDLM_DEBUG(lock, "client lock: no-op");
438 spin_lock_bh(&waiting_locks_spinlock);
440 if (list_empty(&lock->l_pending_chain)) {
441 spin_unlock_bh(&waiting_locks_spinlock);
442 LDLM_DEBUG(lock, "wasn't waiting");
446 __ldlm_del_waiting_lock(lock);
447 __ldlm_add_waiting_lock(lock);
448 spin_unlock_bh(&waiting_locks_spinlock);
450 LDLM_DEBUG(lock, "refreshed");
454 #else /* !__KERNEL__ */
456 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
458 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
462 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
467 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
471 #endif /* __KERNEL__ */
473 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
474 const char *ast_type)
476 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
477 "to a lock %s callback time out: rc %d\n",
478 lock->l_export->exp_obd->obd_name,
479 obd_export_nid2str(lock->l_export), ast_type, rc);
481 if (obd_dump_on_timeout)
482 libcfs_debug_dumplog();
484 spin_lock_bh(&waiting_locks_spinlock);
485 list_add(&lock->l_pending_chain, &expired_lock_thread.elt_expired_locks);
486 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
487 spin_unlock_bh(&waiting_locks_spinlock);
489 class_fail_export(lock->l_export);
493 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
494 struct ptlrpc_request *req, int rc,
495 const char *ast_type)
497 lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
499 if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
500 LASSERT(lock->l_export);
501 if (lock->l_export->exp_libclient) {
502 LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
503 " timeout, just cancelling lock", ast_type,
504 libcfs_nid2str(peer.nid));
505 ldlm_lock_cancel(lock);
507 } else if (lock->l_flags & LDLM_FL_CANCEL) {
508 LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
509 "cancel was received (AST reply lost?)",
510 ast_type, libcfs_nid2str(peer.nid));
511 ldlm_lock_cancel(lock);
514 ldlm_del_waiting_lock(lock);
515 ldlm_failed_ast(lock, rc, ast_type);
519 LDLM_DEBUG(lock, "client (nid %s) returned %d"
520 " from %s AST - normal race",
521 libcfs_nid2str(peer.nid),
523 lustre_msg_get_status(req->rq_repmsg) : -1,
526 LDLM_ERROR(lock, "client (nid %s) returned %d "
527 "from %s AST", libcfs_nid2str(peer.nid),
528 (req->rq_repmsg != NULL) ?
529 lustre_msg_get_status(req->rq_repmsg) : 0,
531 ldlm_lock_cancel(lock);
532 /* Server-side AST functions are called from ldlm_reprocess_all,
533 * which needs to be told to please restart its reprocessing. */
540 static int ldlm_cb_interpret(const struct lu_env *env,
541 struct ptlrpc_request *req, void *data, int rc)
543 struct ldlm_cb_set_arg *arg;
544 struct ldlm_lock *lock;
547 LASSERT(data != NULL);
549 arg = req->rq_async_args.pointer_arg[0];
550 lock = req->rq_async_args.pointer_arg[1];
551 LASSERT(lock != NULL);
553 /* If client canceled the lock but the cancel has not
554 * been recieved yet, we need to update lvbo to have the
555 * proper attributes cached. */
556 if (rc == -EINVAL && arg->type == LDLM_BL_CALLBACK)
557 ldlm_res_lvbo_update(lock->l_resource, NULL,
559 rc = ldlm_handle_ast_error(lock, req, rc,
560 arg->type == LDLM_BL_CALLBACK
561 ? "blocking" : "completion");
564 LDLM_LOCK_RELEASE(lock);
567 atomic_set(&arg->restart, 1);
572 static inline int ldlm_bl_and_cp_ast_fini(struct ptlrpc_request *req,
573 struct ldlm_cb_set_arg *arg,
574 struct ldlm_lock *lock,
580 if (unlikely(instant_cancel)) {
581 rc = ptl_send_rpc(req, 1);
582 ptlrpc_req_finished(req);
584 /* If we cancelled the lock, we need to restart
585 * ldlm_reprocess_queue */
586 atomic_set(&arg->restart, 1);
589 ptlrpc_set_add_req(arg->set, req);
596 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
597 * enqueued server lock conflicts with given one.
599 * Sends blocking ast rpc to the client owning that lock; arms timeout timer
600 * to wait for client response.
602 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
603 struct ldlm_lock_desc *desc,
604 void *data, int flag)
606 struct ldlm_cb_set_arg *arg = data;
607 struct ldlm_request *body;
608 struct ptlrpc_request *req;
609 int instant_cancel = 0;
613 if (flag == LDLM_CB_CANCELING)
614 /* Don't need to do anything here. */
618 LASSERT(data != NULL);
619 if (lock->l_export->exp_obd->obd_recovering != 0) {
620 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
621 ldlm_lock_dump(D_ERROR, lock, 0);
624 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
625 &RQF_LDLM_BL_CALLBACK,
626 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
630 req->rq_async_args.pointer_arg[0] = arg;
631 req->rq_async_args.pointer_arg[1] = lock;
632 req->rq_interpret_reply = ldlm_cb_interpret;
633 req->rq_no_resend = 1;
635 lock_res(lock->l_resource);
636 if (lock->l_granted_mode != lock->l_req_mode) {
637 /* this blocking AST will be communicated as part of the
638 * completion AST instead */
639 unlock_res(lock->l_resource);
640 ptlrpc_req_finished(req);
641 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
645 if (lock->l_destroyed) {
646 /* What's the point? */
647 unlock_res(lock->l_resource);
648 ptlrpc_req_finished(req);
652 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
655 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
656 body->lock_handle[0] = lock->l_remote_handle;
657 body->lock_desc = *desc;
658 body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
660 LDLM_DEBUG(lock, "server preparing blocking AST");
662 ptlrpc_request_set_replen(req);
663 if (instant_cancel) {
664 unlock_res(lock->l_resource);
665 ldlm_lock_cancel(lock);
667 LASSERT(lock->l_granted_mode == lock->l_req_mode);
668 ldlm_add_waiting_lock(lock);
669 unlock_res(lock->l_resource);
672 req->rq_send_state = LUSTRE_IMP_FULL;
673 /* ptlrpc_prep_req already set timeout */
675 req->rq_timeout = ldlm_get_rq_timeout();
677 if (lock->l_export && lock->l_export->exp_nid_stats &&
678 lock->l_export->exp_nid_stats->nid_ldlm_stats)
679 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
680 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
682 rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
687 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
689 struct ldlm_cb_set_arg *arg = data;
690 struct ldlm_request *body;
691 struct ptlrpc_request *req;
692 struct timeval granted_time;
693 long total_enqueue_wait;
694 int instant_cancel = 0;
698 LASSERT(lock != NULL);
699 LASSERT(data != NULL);
701 do_gettimeofday(&granted_time);
702 total_enqueue_wait = cfs_timeval_sub(&granted_time,
703 &lock->l_enqueued_time, NULL);
705 if (total_enqueue_wait / ONE_MILLION > obd_timeout)
706 /* non-fatal with AT - change to LDLM_DEBUG? */
707 LDLM_ERROR(lock, "enqueue wait took %luus from "CFS_TIME_T,
708 total_enqueue_wait, lock->l_enqueued_time.tv_sec);
710 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
711 &RQF_LDLM_CP_CALLBACK);
715 lock_res_and_lock(lock);
716 if (lock->l_resource->lr_lvb_len)
717 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT,
718 lock->l_resource->lr_lvb_len);
719 unlock_res_and_lock(lock);
721 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
723 ptlrpc_request_free(req);
727 req->rq_async_args.pointer_arg[0] = arg;
728 req->rq_async_args.pointer_arg[1] = lock;
729 req->rq_interpret_reply = ldlm_cb_interpret;
730 req->rq_no_resend = 1;
731 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
733 body->lock_handle[0] = lock->l_remote_handle;
734 body->lock_flags = flags;
735 ldlm_lock2desc(lock, &body->lock_desc);
736 if (lock->l_resource->lr_lvb_len) {
737 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
739 lock_res_and_lock(lock);
740 memcpy(lvb, lock->l_resource->lr_lvb_data,
741 lock->l_resource->lr_lvb_len);
742 unlock_res_and_lock(lock);
745 LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
748 /* Server-side enqueue wait time estimate, used in
749 __ldlm_add_waiting_lock to set future enqueue timers */
750 at_add(&lock->l_resource->lr_namespace->ns_at_estimate,
751 total_enqueue_wait / ONE_MILLION);
753 ptlrpc_request_set_replen(req);
755 req->rq_send_state = LUSTRE_IMP_FULL;
756 /* ptlrpc_prep_req already set timeout */
758 req->rq_timeout = ldlm_get_rq_timeout();
760 /* We only send real blocking ASTs after the lock is granted */
761 lock_res_and_lock(lock);
762 if (lock->l_flags & LDLM_FL_AST_SENT) {
763 body->lock_flags |= LDLM_FL_AST_SENT;
765 /* We might get here prior to ldlm_handle_enqueue setting
766 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
767 * into waiting list, but this is safe and similar code in
768 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
769 * that would not only cancel the lock, but will also remove
770 * it from waiting list */
771 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
772 unlock_res_and_lock(lock);
773 ldlm_lock_cancel(lock);
775 lock_res_and_lock(lock);
777 /* start the lock-timeout clock */
778 ldlm_add_waiting_lock(lock);
781 unlock_res_and_lock(lock);
783 if (lock->l_export && lock->l_export->exp_nid_stats &&
784 lock->l_export->exp_nid_stats->nid_ldlm_stats)
785 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
786 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
788 rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
793 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
795 struct ldlm_resource *res = lock->l_resource;
796 struct ldlm_request *body;
797 struct ptlrpc_request *req;
801 LASSERT(lock != NULL);
803 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
804 &RQF_LDLM_GL_CALLBACK,
805 LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK);
810 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
811 body->lock_handle[0] = lock->l_remote_handle;
812 ldlm_lock2desc(lock, &body->lock_desc);
814 lock_res_and_lock(lock);
815 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
816 lock->l_resource->lr_lvb_len);
817 unlock_res_and_lock(lock);
818 res = lock->l_resource;
819 ptlrpc_request_set_replen(req);
822 req->rq_send_state = LUSTRE_IMP_FULL;
823 /* ptlrpc_prep_req already set timeout */
825 req->rq_timeout = ldlm_get_rq_timeout();
827 if (lock->l_export && lock->l_export->exp_nid_stats &&
828 lock->l_export->exp_nid_stats->nid_ldlm_stats)
829 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
830 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
832 rc = ptlrpc_queue_wait(req);
833 if (rc == -ELDLM_NO_LOCK_DATA)
834 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
836 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
838 rc = ldlm_res_lvbo_update(res, req->rq_repmsg,
840 ptlrpc_req_finished(req);
842 ldlm_reprocess_all(res);
848 extern unsigned long long lu_time_stamp_get(void);
850 #define lu_time_stamp_get() time(NULL)
853 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
854 struct lprocfs_stats *srv_stats)
856 int lock_type = 0, op = 0;
858 lock_type = dlm_req->lock_desc.l_resource.lr_type;
862 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
865 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
866 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
868 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
871 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
874 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
882 lprocfs_counter_incr(srv_stats, op);
888 * Main server-side entry point into LDLM. This is called by ptlrpc service
889 * threads to carry out client lock enqueueing requests.
891 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
892 struct ptlrpc_request *req,
893 const struct ldlm_request *dlm_req,
894 const struct ldlm_callback_suite *cbs)
896 struct ldlm_reply *dlm_rep;
898 ldlm_error_t err = ELDLM_OK;
899 struct ldlm_lock *lock = NULL;
904 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
906 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
907 flags = dlm_req->lock_flags;
909 LASSERT(req->rq_export);
911 if (req->rq_rqbd->rqbd_service->srv_stats)
912 ldlm_svc_get_eopc(dlm_req,
913 req->rq_rqbd->rqbd_service->srv_stats);
915 if (req->rq_export && req->rq_export->exp_nid_stats &&
916 req->rq_export->exp_nid_stats->nid_ldlm_stats)
917 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
918 LDLM_ENQUEUE - LDLM_FIRST_OPC);
920 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
921 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
922 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
923 dlm_req->lock_desc.l_resource.lr_type);
924 GOTO(out, rc = -EFAULT);
927 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
928 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
929 dlm_req->lock_desc.l_req_mode &
930 (dlm_req->lock_desc.l_req_mode-1))) {
931 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
932 dlm_req->lock_desc.l_req_mode);
933 GOTO(out, rc = -EFAULT);
936 if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
937 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
939 DEBUG_REQ(D_ERROR, req,
940 "PLAIN lock request from IBITS client?");
941 GOTO(out, rc = -EPROTO);
943 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
945 DEBUG_REQ(D_ERROR, req,
946 "IBITS lock request from unaware client?");
947 GOTO(out, rc = -EPROTO);
951 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
952 against server's _CONNECT_SUPPORTED flags? (I don't want to use
953 ibits for mgc/mgs) */
955 /* INODEBITS_INTEROP: Perform conversion from plain lock to
956 * inodebits lock if client does not support them. */
957 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
958 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
959 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
960 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
961 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
962 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
963 dlm_req->lock_desc.l_req_mode = LCK_CR;
967 if (unlikely(flags & LDLM_FL_REPLAY)) {
968 /* Find an existing lock in the per-export lock hash */
969 lock = lustre_hash_lookup(req->rq_export->exp_lock_hash,
970 (void *)&dlm_req->lock_handle[0]);
972 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
973 LPX64, lock->l_handle.h_cookie);
974 GOTO(existing_lock, rc = 0);
978 /* The lock's callback data might be set in the policy function */
979 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
980 dlm_req->lock_desc.l_resource.lr_type,
981 dlm_req->lock_desc.l_req_mode,
985 GOTO(out, rc = -ENOMEM);
987 do_gettimeofday(&lock->l_enqueued_time);
988 lock->l_remote_handle = dlm_req->lock_handle[0];
989 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
991 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
992 /* Don't enqueue a lock onto the export if it has already
993 * been evicted. Cancel it now instead. (bug 3822) */
994 if (req->rq_export->exp_failed) {
995 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
996 GOTO(out, rc = -ENOTCONN);
998 lock->l_export = class_export_get(req->rq_export);
1000 if (lock->l_export->exp_lock_hash)
1001 lustre_hash_add(lock->l_export->exp_lock_hash,
1002 &lock->l_remote_handle,
1007 if (flags & LDLM_FL_HAS_INTENT) {
1008 /* In this case, the reply buffer is allocated deep in
1009 * local_lock_enqueue by the policy function. */
1012 lock_res_and_lock(lock);
1013 if (lock->l_resource->lr_lvb_len) {
1014 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1016 lock->l_resource->lr_lvb_len);
1018 unlock_res_and_lock(lock);
1020 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1021 GOTO(out, rc = -ENOMEM);
1023 rc = req_capsule_server_pack(&req->rq_pill);
1028 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1029 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1030 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1031 lock->l_req_extent = lock->l_policy_data.l_extent;
1033 err = ldlm_lock_enqueue(ns, &lock, cookie, (int *)&flags);
1037 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1038 dlm_rep->lock_flags = flags;
1040 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1041 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1043 /* We never send a blocking AST until the lock is granted, but
1044 * we can tell it right now */
1045 lock_res_and_lock(lock);
1047 /* Now take into account flags to be inherited from original lock
1048 request both in reply to client and in our own lock flags. */
1049 dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1050 lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1052 /* Don't move a pending lock onto the export if it has already
1053 * been evicted. Cancel it now instead. (bug 5683) */
1054 if (unlikely(req->rq_export->exp_failed ||
1055 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1056 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1058 } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1059 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
1060 if (lock->l_granted_mode == lock->l_req_mode) {
1062 * Only cancel lock if it was granted, because it would
1063 * be destroyed immediatelly and would never be granted
1064 * in the future, causing timeouts on client. Not
1065 * granted lock will be cancelled immediatelly after
1066 * sending completion AST.
1068 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1069 unlock_res_and_lock(lock);
1070 ldlm_lock_cancel(lock);
1071 lock_res_and_lock(lock);
1073 ldlm_add_waiting_lock(lock);
1076 /* Make sure we never ever grant usual metadata locks to liblustre
1078 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1079 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1080 req->rq_export->exp_libclient) {
1081 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1082 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1083 CERROR("Granting sync lock to libclient. "
1084 "req fl %d, rep fl %d, lock fl %d\n",
1085 dlm_req->lock_flags, dlm_rep->lock_flags,
1087 LDLM_ERROR(lock, "sync lock");
1088 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1089 struct ldlm_intent *it;
1091 it = req_capsule_client_get(&req->rq_pill,
1094 CERROR("This is intent %s ("LPU64")\n",
1095 ldlm_it2str(it->opc), it->opc);
1101 unlock_res_and_lock(lock);
1105 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1106 if (!req->rq_packed_final) {
1107 err = lustre_pack_reply(req, 1, NULL, NULL);
1112 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1113 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1115 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1116 "(err=%d, rc=%d)", err, rc);
1118 lock_res_and_lock(lock);
1120 if (lock->l_resource->lr_lvb_len > 0) {
1123 lvb = req_capsule_server_get(&req->rq_pill,
1125 LASSERTF(lvb != NULL, "req %p, lock %p\n",
1128 memcpy(lvb, lock->l_resource->lr_lvb_data,
1129 lock->l_resource->lr_lvb_len);
1132 ldlm_resource_unlink_lock(lock);
1133 ldlm_lock_destroy_nolock(lock);
1135 unlock_res_and_lock(lock);
1137 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1138 ldlm_reprocess_all(lock->l_resource);
1140 LDLM_LOCK_RELEASE(lock);
1143 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1149 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1150 ldlm_completion_callback completion_callback,
1151 ldlm_blocking_callback blocking_callback,
1152 ldlm_glimpse_callback glimpse_callback)
1154 struct ldlm_request *dlm_req;
1155 struct ldlm_callback_suite cbs = {
1156 .lcs_completion = completion_callback,
1157 .lcs_blocking = blocking_callback,
1158 .lcs_glimpse = glimpse_callback
1162 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1163 if (dlm_req != NULL) {
1164 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1165 req, dlm_req, &cbs);
1172 int ldlm_handle_convert0(struct ptlrpc_request *req,
1173 const struct ldlm_request *dlm_req)
1175 struct ldlm_reply *dlm_rep;
1176 struct ldlm_lock *lock;
1180 if (req->rq_export && req->rq_export->exp_nid_stats &&
1181 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1182 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1183 LDLM_CONVERT - LDLM_FIRST_OPC);
1185 rc = req_capsule_server_pack(&req->rq_pill);
1189 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1190 dlm_rep->lock_flags = dlm_req->lock_flags;
1192 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1194 req->rq_status = EINVAL;
1198 LDLM_DEBUG(lock, "server-side convert handler START");
1200 do_gettimeofday(&lock->l_enqueued_time);
1201 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1202 &dlm_rep->lock_flags);
1204 if (ldlm_del_waiting_lock(lock))
1205 LDLM_DEBUG(lock, "converted waiting lock");
1208 req->rq_status = EDEADLOCK;
1213 if (!req->rq_status)
1214 ldlm_reprocess_all(lock->l_resource);
1215 LDLM_DEBUG(lock, "server-side convert handler END");
1216 LDLM_LOCK_PUT(lock);
1218 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1223 int ldlm_handle_convert(struct ptlrpc_request *req)
1226 struct ldlm_request *dlm_req;
1228 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1229 if (dlm_req != NULL) {
1230 rc = ldlm_handle_convert0(req, dlm_req);
1232 CERROR ("Can't unpack dlm_req\n");
1238 /* Cancel all the locks whos handles are packed into ldlm_request */
1239 int ldlm_request_cancel(struct ptlrpc_request *req,
1240 const struct ldlm_request *dlm_req, int first)
1242 struct ldlm_resource *res, *pres = NULL;
1243 struct ldlm_lock *lock;
1244 int i, count, done = 0;
1247 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1251 /* There is no lock on the server at the replay time,
1252 * skip lock cancelling to make replay tests to pass. */
1253 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1256 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1257 "starting at %d", count, first);
1259 for (i = first; i < count; i++) {
1260 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1262 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1263 "lock (cookie "LPU64")",
1264 dlm_req->lock_handle[i].cookie);
1268 res = lock->l_resource;
1273 ldlm_reprocess_all(pres);
1274 LDLM_RESOURCE_DELREF(pres);
1275 ldlm_resource_putref(pres);
1278 ldlm_resource_getref(res);
1279 LDLM_RESOURCE_ADDREF(res);
1280 ldlm_res_lvbo_update(res, NULL, 0, 1);
1284 ldlm_lock_cancel(lock);
1285 LDLM_LOCK_PUT(lock);
1288 ldlm_reprocess_all(pres);
1289 LDLM_RESOURCE_DELREF(pres);
1290 ldlm_resource_putref(pres);
1292 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1296 int ldlm_handle_cancel(struct ptlrpc_request *req)
1298 struct ldlm_request *dlm_req;
1302 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1303 if (dlm_req == NULL) {
1304 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1308 if (req->rq_export && req->rq_export->exp_nid_stats &&
1309 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1310 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1311 LDLM_CANCEL - LDLM_FIRST_OPC);
1313 rc = req_capsule_server_pack(&req->rq_pill);
1317 if (!ldlm_request_cancel(req, dlm_req, 0))
1318 req->rq_status = ESTALE;
1320 if (ptlrpc_reply(req) != 0)
1326 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1327 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1332 LDLM_DEBUG(lock, "client blocking AST callback handler START");
1334 lock_res_and_lock(lock);
1335 lock->l_flags |= LDLM_FL_CBPENDING;
1337 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1338 lock->l_flags |= LDLM_FL_CANCEL;
1340 do_ast = (!lock->l_readers && !lock->l_writers);
1341 unlock_res_and_lock(lock);
1344 LDLM_DEBUG(lock, "already unused, calling "
1345 "callback (%p)", lock->l_blocking_ast);
1346 if (lock->l_blocking_ast != NULL)
1347 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1350 LDLM_DEBUG(lock, "Lock still has references, will be"
1351 " cancelled later");
1354 LDLM_DEBUG(lock, "client blocking callback handler END");
1355 LDLM_LOCK_RELEASE(lock);
1359 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1360 struct ldlm_namespace *ns,
1361 struct ldlm_request *dlm_req,
1362 struct ldlm_lock *lock)
1364 CFS_LIST_HEAD(ast_list);
1367 LDLM_DEBUG(lock, "client completion callback handler START");
1369 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1370 int to = cfs_time_seconds(1);
1372 cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE, to);
1373 if (lock->l_granted_mode == lock->l_req_mode ||
1379 lock_res_and_lock(lock);
1380 if (lock->l_destroyed ||
1381 lock->l_granted_mode == lock->l_req_mode) {
1382 /* bug 11300: the lock has already been granted */
1383 unlock_res_and_lock(lock);
1384 LDLM_DEBUG(lock, "Double grant race happened");
1385 LDLM_LOCK_RELEASE(lock);
1390 /* If we receive the completion AST before the actual enqueue returned,
1391 * then we might need to switch lock modes, resources, or extents. */
1392 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1393 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1394 LDLM_DEBUG(lock, "completion AST, new lock mode");
1397 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1398 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1399 LDLM_DEBUG(lock, "completion AST, new policy data");
1402 ldlm_resource_unlink_lock(lock);
1403 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1404 &lock->l_resource->lr_name,
1405 sizeof(lock->l_resource->lr_name)) != 0) {
1406 unlock_res_and_lock(lock);
1407 if (ldlm_lock_change_resource(ns, lock,
1408 &dlm_req->lock_desc.l_resource.lr_name) != 0) {
1409 LDLM_ERROR(lock, "Failed to allocate resource");
1410 LDLM_LOCK_RELEASE(lock);
1414 LDLM_DEBUG(lock, "completion AST, new resource");
1415 CERROR("change resource!\n");
1416 lock_res_and_lock(lock);
1419 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1420 /* BL_AST locks are not needed in lru.
1421 * let ldlm_cancel_lru() be fast. */
1422 ldlm_lock_remove_from_lru(lock);
1423 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1424 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1427 if (lock->l_lvb_len) {
1428 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
1429 RCL_CLIENT) < lock->l_lvb_len) {
1430 LDLM_ERROR(lock, "completion AST did not contain "
1433 void *lvb = req_capsule_client_swab_get(&req->rq_pill,
1435 (void *)lock->l_lvb_swabber);
1436 memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1440 ldlm_grant_lock(lock, &ast_list);
1441 unlock_res_and_lock(lock);
1443 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1445 ldlm_run_ast_work(&ast_list, LDLM_WORK_CP_AST);
1447 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1449 LDLM_LOCK_RELEASE(lock);
1453 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1454 struct ldlm_namespace *ns,
1455 struct ldlm_request *dlm_req,
1456 struct ldlm_lock *lock)
1461 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1463 if (lock->l_glimpse_ast != NULL)
1464 rc = lock->l_glimpse_ast(lock, req);
1466 if (req->rq_repmsg != NULL) {
1469 req->rq_status = rc;
1473 lock_res_and_lock(lock);
1474 if (lock->l_granted_mode == LCK_PW &&
1475 !lock->l_readers && !lock->l_writers &&
1476 cfs_time_after(cfs_time_current(),
1477 cfs_time_add(lock->l_last_used,
1478 cfs_time_seconds(10)))) {
1479 unlock_res_and_lock(lock);
1480 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1481 ldlm_handle_bl_callback(ns, NULL, lock);
1486 unlock_res_and_lock(lock);
1487 LDLM_LOCK_RELEASE(lock);
1491 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1493 if (req->rq_no_reply)
1496 req->rq_status = rc;
1497 if (!req->rq_packed_final) {
1498 rc = lustre_pack_reply(req, 1, NULL, NULL);
1502 return ptlrpc_reply(req);
1506 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1507 struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
1508 struct list_head *cancels, int count)
1510 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1511 struct ldlm_bl_work_item *blwi;
1514 if (cancels && count == 0)
1517 OBD_ALLOC(blwi, sizeof(*blwi));
1523 blwi->blwi_ld = *ld;
1525 list_add(&blwi->blwi_head, cancels);
1526 list_del_init(cancels);
1527 blwi->blwi_count = count;
1529 blwi->blwi_lock = lock;
1531 spin_lock(&blp->blp_lock);
1532 if (lock && lock->l_flags & LDLM_FL_DISCARD_DATA) {
1533 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1534 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1536 /* other blocking callbacks are added to the regular list */
1537 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1539 cfs_waitq_signal(&blp->blp_waitq);
1540 spin_unlock(&blp->blp_lock);
1546 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1547 struct ldlm_lock *lock)
1550 RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0));
1556 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1557 struct list_head *cancels, int count)
1560 RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count));
1566 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1567 static int ldlm_callback_handler(struct ptlrpc_request *req)
1569 struct ldlm_namespace *ns;
1570 struct ldlm_request *dlm_req;
1571 struct ldlm_lock *lock;
1575 /* Requests arrive in sender's byte order. The ptlrpc service
1576 * handler has already checked and, if necessary, byte-swapped the
1577 * incoming request message body, but I am responsible for the
1578 * message buffers. */
1580 /* do nothing for sec context finalize */
1581 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
1584 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1586 if (req->rq_export == NULL) {
1587 struct ldlm_request *dlm_req;
1589 CDEBUG(D_RPCTRACE, "operation %d from %s with bad "
1590 "export cookie "LPX64"; this is "
1591 "normal if this node rebooted with a lock held\n",
1592 lustre_msg_get_opc(req->rq_reqmsg),
1593 libcfs_id2str(req->rq_peer),
1594 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
1596 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1597 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1598 if (dlm_req != NULL)
1599 CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1600 dlm_req->lock_handle[0].cookie);
1602 ldlm_callback_reply(req, -ENOTCONN);
1606 LASSERT(req->rq_export != NULL);
1607 LASSERT(req->rq_export->exp_obd != NULL);
1609 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1610 case LDLM_BL_CALLBACK:
1611 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK))
1614 case LDLM_CP_CALLBACK:
1615 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK))
1618 case LDLM_GL_CALLBACK:
1619 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK))
1622 case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
1623 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1624 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1626 rc = llog_origin_handle_cancel(req);
1627 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
1629 ldlm_callback_reply(req, rc);
1631 case OBD_QC_CALLBACK:
1632 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
1633 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
1635 rc = target_handle_qc_callback(req);
1636 ldlm_callback_reply(req, rc);
1640 /* reply in handler */
1641 req_capsule_set(&req->rq_pill, &RQF_MDS_QUOTA_DQACQ);
1642 rc = target_handle_dqacq_callback(req);
1644 case LLOG_ORIGIN_HANDLE_CREATE:
1645 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
1646 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1648 rc = llog_origin_handle_create(req);
1649 ldlm_callback_reply(req, rc);
1651 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1652 req_capsule_set(&req->rq_pill,
1653 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
1654 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1656 rc = llog_origin_handle_next_block(req);
1657 ldlm_callback_reply(req, rc);
1659 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1660 req_capsule_set(&req->rq_pill,
1661 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
1662 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1664 rc = llog_origin_handle_read_header(req);
1665 ldlm_callback_reply(req, rc);
1667 case LLOG_ORIGIN_HANDLE_CLOSE:
1668 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1670 rc = llog_origin_handle_close(req);
1671 ldlm_callback_reply(req, rc);
1674 CERROR("unknown opcode %u\n",
1675 lustre_msg_get_opc(req->rq_reqmsg));
1676 ldlm_callback_reply(req, -EPROTO);
1680 ns = req->rq_export->exp_obd->obd_namespace;
1681 LASSERT(ns != NULL);
1683 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1685 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1686 if (dlm_req == NULL) {
1687 ldlm_callback_reply(req, -EPROTO);
1691 /* Force a known safe race, send a cancel to the server for a lock
1692 * which the server has already started a blocking callback on. */
1693 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
1694 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1695 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]);
1697 CERROR("ldlm_cli_cancel: %d\n", rc);
1700 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
1702 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
1703 "disappeared\n", dlm_req->lock_handle[0].cookie);
1704 ldlm_callback_reply(req, -EINVAL);
1708 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1709 lock_res_and_lock(lock);
1710 lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1711 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1712 /* If somebody cancels locks and cache is already droped,
1713 * we can tell the server we have no lock. Otherwise, we
1714 * should send cancel after dropping the cache. */
1715 if ((lock->l_flags & LDLM_FL_CANCELING) &&
1716 (lock->l_flags & LDLM_FL_BL_DONE)) {
1717 LDLM_DEBUG(lock, "callback on lock "
1718 LPX64" - lock disappeared\n",
1719 dlm_req->lock_handle[0].cookie);
1720 unlock_res_and_lock(lock);
1721 LDLM_LOCK_RELEASE(lock);
1722 ldlm_callback_reply(req, -EINVAL);
1725 /* BL_AST locks are not needed in lru.
1726 * let ldlm_cancel_lru() be fast. */
1727 ldlm_lock_remove_from_lru(lock);
1728 lock->l_flags |= LDLM_FL_BL_AST;
1730 unlock_res_and_lock(lock);
1732 /* We want the ost thread to get this reply so that it can respond
1733 * to ost requests (write cache writeback) that might be triggered
1736 * But we'd also like to be able to indicate in the reply that we're
1737 * cancelling right now, because it's unused, or have an intent result
1738 * in the reply, so we might have to push the responsibility for sending
1739 * the reply down into the AST handlers, alas. */
1741 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1742 case LDLM_BL_CALLBACK:
1743 CDEBUG(D_INODE, "blocking ast\n");
1744 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
1745 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK))
1746 ldlm_callback_reply(req, 0);
1747 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
1748 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1750 case LDLM_CP_CALLBACK:
1751 CDEBUG(D_INODE, "completion ast\n");
1752 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
1753 ldlm_callback_reply(req, 0);
1754 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1756 case LDLM_GL_CALLBACK:
1757 CDEBUG(D_INODE, "glimpse ast\n");
1758 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
1759 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1762 LBUG(); /* checked above */
1768 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1773 /* Requests arrive in sender's byte order. The ptlrpc service
1774 * handler has already checked and, if necessary, byte-swapped the
1775 * incoming request message body, but I am responsible for the
1776 * message buffers. */
1778 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1780 if (req->rq_export == NULL) {
1781 struct ldlm_request *dlm_req;
1783 CERROR("operation %d from %s with bad export cookie "LPU64"\n",
1784 lustre_msg_get_opc(req->rq_reqmsg),
1785 libcfs_id2str(req->rq_peer),
1786 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
1788 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1789 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1790 if (dlm_req != NULL)
1791 ldlm_lock_dump_handle(D_ERROR,
1792 &dlm_req->lock_handle[0]);
1793 ldlm_callback_reply(req, -ENOTCONN);
1797 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1799 /* XXX FIXME move this back to mds/handler.c, bug 249 */
1801 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
1802 CDEBUG(D_INODE, "cancel\n");
1803 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
1805 rc = ldlm_handle_cancel(req);
1809 case OBD_LOG_CANCEL:
1810 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1811 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1813 rc = llog_origin_handle_cancel(req);
1814 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
1816 ldlm_callback_reply(req, rc);
1819 CERROR("invalid opcode %d\n",
1820 lustre_msg_get_opc(req->rq_reqmsg));
1821 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1822 ldlm_callback_reply(req, -EINVAL);
1828 void ldlm_revoke_lock_cb(void *obj, void *data)
1830 struct list_head *rpc_list = data;
1831 struct ldlm_lock *lock = obj;
1833 lock_res_and_lock(lock);
1835 if (lock->l_req_mode != lock->l_granted_mode) {
1836 unlock_res_and_lock(lock);
1840 LASSERT(lock->l_resource);
1841 if (lock->l_resource->lr_type != LDLM_IBITS &&
1842 lock->l_resource->lr_type != LDLM_PLAIN) {
1843 unlock_res_and_lock(lock);
1847 if (lock->l_flags & LDLM_FL_AST_SENT) {
1848 unlock_res_and_lock(lock);
1852 LASSERT(lock->l_blocking_ast);
1853 LASSERT(!lock->l_blocking_lock);
1855 lock->l_flags |= LDLM_FL_AST_SENT;
1856 if (lock->l_export && lock->l_export->exp_lock_hash)
1857 lustre_hash_del(lock->l_export->exp_lock_hash,
1858 &lock->l_remote_handle, &lock->l_exp_hash);
1859 list_add_tail(&lock->l_rk_ast, rpc_list);
1860 LDLM_LOCK_GET(lock);
1862 unlock_res_and_lock(lock);
1865 void ldlm_revoke_export_locks(struct obd_export *exp)
1867 struct list_head rpc_list;
1870 CFS_INIT_LIST_HEAD(&rpc_list);
1871 lustre_hash_for_each_empty(exp->exp_lock_hash,
1872 ldlm_revoke_lock_cb, &rpc_list);
1873 ldlm_run_ast_work(&rpc_list, LDLM_WORK_REVOKE_AST);
1879 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1881 struct ldlm_bl_work_item *blwi = NULL;
1882 static unsigned int num_bl = 0;
1884 spin_lock(&blp->blp_lock);
1885 /* process a request from the blp_list at least every blp_num_threads */
1886 if (!list_empty(&blp->blp_list) &&
1887 (list_empty(&blp->blp_prio_list) || num_bl == 0))
1888 blwi = list_entry(blp->blp_list.next,
1889 struct ldlm_bl_work_item, blwi_entry);
1891 if (!list_empty(&blp->blp_prio_list))
1892 blwi = list_entry(blp->blp_prio_list.next,
1893 struct ldlm_bl_work_item, blwi_entry);
1896 if (++num_bl >= atomic_read(&blp->blp_num_threads))
1898 list_del(&blwi->blwi_entry);
1900 spin_unlock(&blp->blp_lock);
1905 /* This only contains temporary data until the thread starts */
1906 struct ldlm_bl_thread_data {
1907 char bltd_name[CFS_CURPROC_COMM_MAX];
1908 struct ldlm_bl_pool *bltd_blp;
1909 struct completion bltd_comp;
1913 static int ldlm_bl_thread_main(void *arg);
1915 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
1917 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
1920 init_completion(&bltd.bltd_comp);
1921 rc = cfs_kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1923 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
1924 atomic_read(&blp->blp_num_threads), rc);
1927 wait_for_completion(&bltd.bltd_comp);
1932 static int ldlm_bl_thread_main(void *arg)
1934 struct ldlm_bl_pool *blp;
1938 struct ldlm_bl_thread_data *bltd = arg;
1940 blp = bltd->bltd_blp;
1942 bltd->bltd_num = atomic_inc_return(&blp->blp_num_threads) - 1;
1943 atomic_inc(&blp->blp_busy_threads);
1945 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
1946 "ldlm_bl_%02d", bltd->bltd_num);
1947 cfs_daemonize(bltd->bltd_name);
1949 complete(&bltd->bltd_comp);
1950 /* cannot use bltd after this, it is only on caller's stack */
1954 struct l_wait_info lwi = { 0 };
1955 struct ldlm_bl_work_item *blwi = NULL;
1957 blwi = ldlm_bl_get_work(blp);
1962 atomic_dec(&blp->blp_busy_threads);
1963 l_wait_event_exclusive(blp->blp_waitq,
1964 (blwi = ldlm_bl_get_work(blp)) != NULL,
1966 busy = atomic_inc_return(&blp->blp_busy_threads);
1968 if (blwi->blwi_ns == NULL)
1969 /* added by ldlm_cleanup() */
1972 /* Not fatal if racy and have a few too many threads */
1973 if (unlikely(busy < blp->blp_max_threads &&
1974 busy >= atomic_read(&blp->blp_num_threads)))
1975 /* discard the return value, we tried */
1976 ldlm_bl_thread_start(blp);
1978 if (blwi->blwi_ns == NULL)
1979 /* added by ldlm_cleanup() */
1983 if (blwi->blwi_count) {
1984 /* The special case when we cancel locks in lru
1985 * asynchronously, we pass the list of locks here.
1986 * Thus lock is marked LDLM_FL_CANCELING, and already
1987 * canceled locally. */
1988 ldlm_cli_cancel_list(&blwi->blwi_head,
1989 blwi->blwi_count, NULL, 0);
1991 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1994 OBD_FREE(blwi, sizeof(*blwi));
1997 atomic_dec(&blp->blp_busy_threads);
1998 atomic_dec(&blp->blp_num_threads);
1999 complete(&blp->blp_comp);
2005 static int ldlm_setup(void);
2006 static int ldlm_cleanup(void);
2008 int ldlm_get_ref(void)
2012 mutex_down(&ldlm_ref_sem);
2013 if (++ldlm_refcount == 1) {
2018 mutex_up(&ldlm_ref_sem);
2023 void ldlm_put_ref(void)
2026 mutex_down(&ldlm_ref_sem);
2027 if (ldlm_refcount == 1) {
2028 int rc = ldlm_cleanup();
2030 CERROR("ldlm_cleanup failed: %d\n", rc);
2036 mutex_up(&ldlm_ref_sem);
2042 * Export handle<->lock hash operations.
2045 ldlm_export_lock_hash(lustre_hash_t *lh, void *key, unsigned mask)
2047 return lh_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2051 ldlm_export_lock_key(struct hlist_node *hnode)
2053 struct ldlm_lock *lock;
2056 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2057 RETURN(&lock->l_remote_handle);
2061 ldlm_export_lock_compare(void *key, struct hlist_node *hnode)
2064 RETURN(lustre_handle_equal(ldlm_export_lock_key(hnode), key));
2068 ldlm_export_lock_get(struct hlist_node *hnode)
2070 struct ldlm_lock *lock;
2073 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2074 LDLM_LOCK_GET(lock);
2080 ldlm_export_lock_put(struct hlist_node *hnode)
2082 struct ldlm_lock *lock;
2085 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2086 LDLM_LOCK_RELEASE(lock);
2091 static lustre_hash_ops_t ldlm_export_lock_ops = {
2092 .lh_hash = ldlm_export_lock_hash,
2093 .lh_key = ldlm_export_lock_key,
2094 .lh_compare = ldlm_export_lock_compare,
2095 .lh_get = ldlm_export_lock_get,
2096 .lh_put = ldlm_export_lock_put
2099 int ldlm_init_export(struct obd_export *exp)
2103 exp->exp_lock_hash =
2104 lustre_hash_init(obd_uuid2str(&exp->exp_client_uuid),
2105 128, 65536, &ldlm_export_lock_ops, LH_REHASH);
2107 if (!exp->exp_lock_hash)
2112 EXPORT_SYMBOL(ldlm_init_export);
2114 void ldlm_destroy_export(struct obd_export *exp)
2117 lustre_hash_exit(exp->exp_lock_hash);
2118 exp->exp_lock_hash = NULL;
2121 EXPORT_SYMBOL(ldlm_destroy_export);
2123 static int ldlm_setup(void)
2125 struct ldlm_bl_pool *blp;
2127 int ldlm_min_threads = LDLM_THREADS_AUTO_MIN;
2128 int ldlm_max_threads = LDLM_THREADS_AUTO_MAX;
2134 if (ldlm_state != NULL)
2137 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2138 if (ldlm_state == NULL)
2142 rc = ldlm_proc_setup();
2148 if (ldlm_num_threads) {
2149 /* If ldlm_num_threads is set, it is the min and the max. */
2150 if (ldlm_num_threads > LDLM_THREADS_AUTO_MAX)
2151 ldlm_num_threads = LDLM_THREADS_AUTO_MAX;
2152 if (ldlm_num_threads < LDLM_THREADS_AUTO_MIN)
2153 ldlm_num_threads = LDLM_THREADS_AUTO_MIN;
2154 ldlm_min_threads = ldlm_max_threads = ldlm_num_threads;
2158 ldlm_state->ldlm_cb_service =
2159 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2160 LDLM_MAXREPSIZE, LDLM_CB_REQUEST_PORTAL,
2161 LDLM_CB_REPLY_PORTAL, 1800,
2162 ldlm_callback_handler, "ldlm_cbd",
2163 ldlm_svc_proc_dir, NULL,
2164 ldlm_min_threads, ldlm_max_threads,
2166 LCT_MD_THREAD|LCT_DT_THREAD);
2168 if (!ldlm_state->ldlm_cb_service) {
2169 CERROR("failed to start service\n");
2170 GOTO(out_proc, rc = -ENOMEM);
2173 ldlm_state->ldlm_cancel_service =
2174 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2175 LDLM_MAXREPSIZE, LDLM_CANCEL_REQUEST_PORTAL,
2176 LDLM_CANCEL_REPLY_PORTAL, 6000,
2177 ldlm_cancel_handler, "ldlm_canceld",
2178 ldlm_svc_proc_dir, NULL,
2179 ldlm_min_threads, ldlm_max_threads,
2181 LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD);
2183 if (!ldlm_state->ldlm_cancel_service) {
2184 CERROR("failed to start service\n");
2185 GOTO(out_proc, rc = -ENOMEM);
2188 OBD_ALLOC(blp, sizeof(*blp));
2190 GOTO(out_proc, rc = -ENOMEM);
2191 ldlm_state->ldlm_bl_pool = blp;
2193 spin_lock_init(&blp->blp_lock);
2194 CFS_INIT_LIST_HEAD(&blp->blp_list);
2195 CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2196 cfs_waitq_init(&blp->blp_waitq);
2197 atomic_set(&blp->blp_num_threads, 0);
2198 atomic_set(&blp->blp_busy_threads, 0);
2199 blp->blp_min_threads = ldlm_min_threads;
2200 blp->blp_max_threads = ldlm_max_threads;
2203 for (i = 0; i < blp->blp_min_threads; i++) {
2204 rc = ldlm_bl_thread_start(blp);
2206 GOTO(out_thread, rc);
2209 rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cancel_service);
2211 GOTO(out_thread, rc);
2213 rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cb_service);
2215 GOTO(out_thread, rc);
2217 CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2218 expired_lock_thread.elt_state = ELT_STOPPED;
2219 cfs_waitq_init(&expired_lock_thread.elt_waitq);
2221 CFS_INIT_LIST_HEAD(&waiting_locks_list);
2222 spin_lock_init(&waiting_locks_spinlock);
2223 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2225 rc = cfs_kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FILES);
2227 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2228 GOTO(out_thread, rc);
2231 wait_event(expired_lock_thread.elt_waitq,
2232 expired_lock_thread.elt_state == ELT_READY);
2236 rc = ldlm_pools_init();
2238 GOTO(out_thread, rc);
2244 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2245 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2250 ldlm_proc_cleanup();
2253 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2258 static int ldlm_cleanup(void)
2261 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2265 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2266 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2267 CERROR("ldlm still has namespaces; clean these up first.\n");
2268 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2269 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2278 while (atomic_read(&blp->blp_num_threads) > 0) {
2279 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2281 init_completion(&blp->blp_comp);
2283 spin_lock(&blp->blp_lock);
2284 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2285 cfs_waitq_signal(&blp->blp_waitq);
2286 spin_unlock(&blp->blp_lock);
2288 wait_for_completion(&blp->blp_comp);
2290 OBD_FREE(blp, sizeof(*blp));
2292 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2293 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2294 ldlm_proc_cleanup();
2296 expired_lock_thread.elt_state = ELT_TERMINATE;
2297 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
2298 wait_event(expired_lock_thread.elt_waitq,
2299 expired_lock_thread.elt_state == ELT_STOPPED);
2301 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2302 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2305 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2311 int __init ldlm_init(void)
2313 init_mutex(&ldlm_ref_sem);
2314 init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
2315 init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
2316 ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
2317 sizeof(struct ldlm_resource), 0,
2318 SLAB_HWCACHE_ALIGN);
2319 if (ldlm_resource_slab == NULL)
2322 ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
2323 sizeof(struct ldlm_lock), 0,
2324 SLAB_HWCACHE_ALIGN);
2325 if (ldlm_lock_slab == NULL) {
2326 cfs_mem_cache_destroy(ldlm_resource_slab);
2330 ldlm_interval_slab = cfs_mem_cache_create("interval_node",
2331 sizeof(struct ldlm_interval),
2332 0, SLAB_HWCACHE_ALIGN);
2333 if (ldlm_interval_slab == NULL) {
2334 cfs_mem_cache_destroy(ldlm_resource_slab);
2335 cfs_mem_cache_destroy(ldlm_lock_slab);
2342 void __exit ldlm_exit(void)
2346 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
2347 rc = cfs_mem_cache_destroy(ldlm_resource_slab);
2348 LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
2349 rc = cfs_mem_cache_destroy(ldlm_lock_slab);
2350 LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
2351 rc = cfs_mem_cache_destroy(ldlm_interval_slab);
2352 LASSERTF(rc == 0, "couldn't free interval node slab\n");
2356 EXPORT_SYMBOL(ldlm_extent_shift_kms);
2359 EXPORT_SYMBOL(ldlm_get_processing_policy);
2360 EXPORT_SYMBOL(ldlm_lock2desc);
2361 EXPORT_SYMBOL(ldlm_register_intent);
2362 EXPORT_SYMBOL(ldlm_lockname);
2363 EXPORT_SYMBOL(ldlm_typename);
2364 EXPORT_SYMBOL(ldlm_lock2handle);
2365 EXPORT_SYMBOL(__ldlm_handle2lock);
2366 EXPORT_SYMBOL(ldlm_lock_get);
2367 EXPORT_SYMBOL(ldlm_lock_put);
2368 EXPORT_SYMBOL(ldlm_lock_fast_match);
2369 EXPORT_SYMBOL(ldlm_lock_fast_release);
2370 EXPORT_SYMBOL(ldlm_lock_match);
2371 EXPORT_SYMBOL(ldlm_lock_cancel);
2372 EXPORT_SYMBOL(ldlm_lock_addref);
2373 EXPORT_SYMBOL(ldlm_lock_addref_try);
2374 EXPORT_SYMBOL(ldlm_lock_decref);
2375 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
2376 EXPORT_SYMBOL(ldlm_lock_change_resource);
2377 EXPORT_SYMBOL(ldlm_lock_set_data);
2378 EXPORT_SYMBOL(ldlm_it2str);
2379 EXPORT_SYMBOL(ldlm_lock_dump);
2380 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2381 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
2382 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
2383 EXPORT_SYMBOL(ldlm_lock_allow_match);
2385 /* ldlm_request.c */
2386 EXPORT_SYMBOL(ldlm_completion_ast_async);
2387 EXPORT_SYMBOL(ldlm_completion_ast);
2388 EXPORT_SYMBOL(ldlm_blocking_ast);
2389 EXPORT_SYMBOL(ldlm_glimpse_ast);
2390 EXPORT_SYMBOL(ldlm_expired_completion_wait);
2391 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
2392 EXPORT_SYMBOL(ldlm_prep_elc_req);
2393 EXPORT_SYMBOL(ldlm_cli_convert);
2394 EXPORT_SYMBOL(ldlm_cli_enqueue);
2395 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
2396 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
2397 EXPORT_SYMBOL(ldlm_cli_cancel);
2398 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2399 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2400 EXPORT_SYMBOL(ldlm_cli_cancel_req);
2401 EXPORT_SYMBOL(ldlm_replay_locks);
2402 EXPORT_SYMBOL(ldlm_resource_foreach);
2403 EXPORT_SYMBOL(ldlm_namespace_foreach);
2404 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
2405 EXPORT_SYMBOL(ldlm_resource_iterate);
2406 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2407 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2410 EXPORT_SYMBOL(ldlm_server_blocking_ast);
2411 EXPORT_SYMBOL(ldlm_server_completion_ast);
2412 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
2413 EXPORT_SYMBOL(ldlm_handle_enqueue);
2414 EXPORT_SYMBOL(ldlm_handle_enqueue0);
2415 EXPORT_SYMBOL(ldlm_handle_cancel);
2416 EXPORT_SYMBOL(ldlm_request_cancel);
2417 EXPORT_SYMBOL(ldlm_handle_convert);
2418 EXPORT_SYMBOL(ldlm_handle_convert0);
2419 EXPORT_SYMBOL(ldlm_del_waiting_lock);
2420 EXPORT_SYMBOL(ldlm_get_ref);
2421 EXPORT_SYMBOL(ldlm_put_ref);
2422 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
2423 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2425 /* ldlm_resource.c */
2426 EXPORT_SYMBOL(ldlm_namespace_new);
2427 EXPORT_SYMBOL(ldlm_namespace_cleanup);
2428 EXPORT_SYMBOL(ldlm_namespace_free);
2429 EXPORT_SYMBOL(ldlm_namespace_dump);
2430 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
2431 EXPORT_SYMBOL(ldlm_resource_get);
2432 EXPORT_SYMBOL(ldlm_resource_putref);
2433 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
2436 EXPORT_SYMBOL(client_import_add_conn);
2437 EXPORT_SYMBOL(client_import_del_conn);
2438 EXPORT_SYMBOL(client_obd_setup);
2439 EXPORT_SYMBOL(client_obd_cleanup);
2440 EXPORT_SYMBOL(client_connect_import);
2441 EXPORT_SYMBOL(client_disconnect_export);
2442 EXPORT_SYMBOL(target_start_recovery_thread);
2443 EXPORT_SYMBOL(target_stop_recovery_thread);
2444 EXPORT_SYMBOL(target_handle_connect);
2445 EXPORT_SYMBOL(target_cleanup_recovery);
2446 EXPORT_SYMBOL(target_destroy_export);
2447 EXPORT_SYMBOL(target_cancel_recovery_timer);
2448 EXPORT_SYMBOL(target_send_reply);
2449 EXPORT_SYMBOL(target_queue_recovery_request);
2450 EXPORT_SYMBOL(target_handle_ping);
2451 EXPORT_SYMBOL(target_pack_pool_reply);
2452 EXPORT_SYMBOL(target_handle_disconnect);
2455 EXPORT_SYMBOL(lock_res_and_lock);
2456 EXPORT_SYMBOL(unlock_res_and_lock);