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 struct ptlrpc_connection *conn = lock->l_export->exp_connection;
477 char *str = libcfs_nid2str(conn->c_peer.nid);
479 LCONSOLE_ERROR_MSG(0x138, "A client on nid %s was evicted from "
480 "service %s.\n", str,
481 lock->l_export->exp_obd->obd_name);
483 LCONSOLE_ERROR_MSG(0x012, "Lock %s callback to %s timed out for "
484 "resource %d\n", ast_type,
485 obd_export_nid2str(lock->l_export), rc);
487 if (obd_dump_on_timeout)
488 libcfs_debug_dumplog();
489 class_fail_export(lock->l_export);
492 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
493 struct ptlrpc_request *req, int rc,
494 const char *ast_type)
496 lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
498 if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
499 LASSERT(lock->l_export);
500 if (lock->l_export->exp_libclient) {
501 LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
502 " timeout, just cancelling lock", ast_type,
503 libcfs_nid2str(peer.nid));
504 ldlm_lock_cancel(lock);
506 } else if (lock->l_flags & LDLM_FL_CANCEL) {
507 LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
508 "cancel was received (AST reply lost?)",
509 ast_type, libcfs_nid2str(peer.nid));
510 ldlm_lock_cancel(lock);
513 ldlm_del_waiting_lock(lock);
514 ldlm_failed_ast(lock, rc, ast_type);
518 LDLM_DEBUG(lock, "client (nid %s) returned %d"
519 " from %s AST - normal race",
520 libcfs_nid2str(peer.nid),
522 lustre_msg_get_status(req->rq_repmsg) : -1,
525 LDLM_ERROR(lock, "client (nid %s) returned %d "
526 "from %s AST", libcfs_nid2str(peer.nid),
527 (req->rq_repmsg != NULL) ?
528 lustre_msg_get_status(req->rq_repmsg) : 0,
530 ldlm_lock_cancel(lock);
531 /* Server-side AST functions are called from ldlm_reprocess_all,
532 * which needs to be told to please restart its reprocessing. */
539 static int ldlm_cb_interpret(struct ptlrpc_request *req, void *data, int rc)
541 struct ldlm_cb_set_arg *arg;
542 struct ldlm_lock *lock;
545 LASSERT(data != NULL);
547 arg = req->rq_async_args.pointer_arg[0];
548 lock = req->rq_async_args.pointer_arg[1];
549 LASSERT(lock != NULL);
551 /* If client canceled the lock but the cancel has not
552 * been recieved yet, we need to update lvbo to have the
553 * proper attributes cached. */
554 if (rc == -EINVAL && arg->type == LDLM_BL_CALLBACK)
555 ldlm_res_lvbo_update(lock->l_resource, NULL,
557 rc = ldlm_handle_ast_error(lock, req, rc,
558 arg->type == LDLM_BL_CALLBACK
559 ? "blocking" : "completion");
565 atomic_set(&arg->restart, 1);
570 static inline int ldlm_bl_and_cp_ast_fini(struct ptlrpc_request *req,
571 struct ldlm_cb_set_arg *arg,
572 struct ldlm_lock *lock,
578 if (unlikely(instant_cancel)) {
579 rc = ptl_send_rpc(req, 1);
580 ptlrpc_req_finished(req);
582 /* If we cancelled the lock, we need to restart
583 * ldlm_reprocess_queue */
584 atomic_set(&arg->restart, 1);
587 ptlrpc_set_add_req(arg->set, req);
594 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
595 * enqueued server lock conflicts with given one.
597 * Sends blocking ast rpc to the client owning that lock; arms timeout timer
598 * to wait for client response.
600 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
601 struct ldlm_lock_desc *desc,
602 void *data, int flag)
604 struct ldlm_cb_set_arg *arg = data;
605 struct ldlm_request *body;
606 struct ptlrpc_request *req;
607 int instant_cancel = 0;
611 if (flag == LDLM_CB_CANCELING)
612 /* Don't need to do anything here. */
616 LASSERT(data != NULL);
617 if (lock->l_export->exp_obd->obd_recovering != 0) {
618 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
619 ldlm_lock_dump(D_ERROR, lock, 0);
622 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
623 &RQF_LDLM_BL_CALLBACK,
624 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
628 req->rq_async_args.pointer_arg[0] = arg;
629 req->rq_async_args.pointer_arg[1] = lock;
630 req->rq_interpret_reply = ldlm_cb_interpret;
631 req->rq_no_resend = 1;
633 lock_res(lock->l_resource);
634 if (lock->l_granted_mode != lock->l_req_mode) {
635 /* this blocking AST will be communicated as part of the
636 * completion AST instead */
637 unlock_res(lock->l_resource);
638 ptlrpc_req_finished(req);
639 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
643 if (lock->l_destroyed) {
644 /* What's the point? */
645 unlock_res(lock->l_resource);
646 ptlrpc_req_finished(req);
650 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
653 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
654 body->lock_handle[0] = lock->l_remote_handle;
655 body->lock_desc = *desc;
656 body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
658 LDLM_DEBUG(lock, "server preparing blocking AST");
660 ptlrpc_request_set_replen(req);
661 if (instant_cancel) {
662 unlock_res(lock->l_resource);
663 ldlm_lock_cancel(lock);
665 LASSERT(lock->l_granted_mode == lock->l_req_mode);
666 ldlm_add_waiting_lock(lock);
667 unlock_res(lock->l_resource);
670 req->rq_send_state = LUSTRE_IMP_FULL;
671 /* ptlrpc_prep_req already set timeout */
673 req->rq_timeout = ldlm_get_rq_timeout();
675 if (lock->l_export && lock->l_export->exp_nid_stats &&
676 lock->l_export->exp_nid_stats->nid_ldlm_stats)
677 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
678 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
680 rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
685 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
687 struct ldlm_cb_set_arg *arg = data;
688 struct ldlm_request *body;
689 struct ptlrpc_request *req;
690 struct timeval granted_time;
691 long total_enqueue_wait;
692 int instant_cancel = 0;
696 LASSERT(lock != NULL);
697 LASSERT(data != NULL);
699 do_gettimeofday(&granted_time);
700 total_enqueue_wait = cfs_timeval_sub(&granted_time,
701 &lock->l_enqueued_time, NULL);
703 if (total_enqueue_wait / ONE_MILLION > obd_timeout)
704 /* non-fatal with AT - change to LDLM_DEBUG? */
705 LDLM_ERROR(lock, "enqueue wait took %luus from "CFS_TIME_T,
706 total_enqueue_wait, lock->l_enqueued_time.tv_sec);
708 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
709 &RQF_LDLM_CP_CALLBACK);
713 lock_res_and_lock(lock);
714 if (lock->l_resource->lr_lvb_len)
715 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT,
716 lock->l_resource->lr_lvb_len);
717 unlock_res_and_lock(lock);
719 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
721 ptlrpc_request_free(req);
725 req->rq_async_args.pointer_arg[0] = arg;
726 req->rq_async_args.pointer_arg[1] = lock;
727 req->rq_interpret_reply = ldlm_cb_interpret;
728 req->rq_no_resend = 1;
729 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
731 body->lock_handle[0] = lock->l_remote_handle;
732 body->lock_flags = flags;
733 ldlm_lock2desc(lock, &body->lock_desc);
734 if (lock->l_resource->lr_lvb_len) {
735 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
737 lock_res_and_lock(lock);
738 memcpy(lvb, lock->l_resource->lr_lvb_data,
739 lock->l_resource->lr_lvb_len);
740 unlock_res_and_lock(lock);
743 LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
746 /* Server-side enqueue wait time estimate, used in
747 __ldlm_add_waiting_lock to set future enqueue timers */
748 at_add(&lock->l_resource->lr_namespace->ns_at_estimate,
749 total_enqueue_wait / ONE_MILLION);
751 ptlrpc_request_set_replen(req);
753 req->rq_send_state = LUSTRE_IMP_FULL;
754 /* ptlrpc_prep_req already set timeout */
756 req->rq_timeout = ldlm_get_rq_timeout();
758 /* We only send real blocking ASTs after the lock is granted */
759 lock_res_and_lock(lock);
760 if (lock->l_flags & LDLM_FL_AST_SENT) {
761 body->lock_flags |= LDLM_FL_AST_SENT;
763 /* We might get here prior to ldlm_handle_enqueue setting
764 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
765 * into waiting list, but this is safe and similar code in
766 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
767 * that would not only cancel the lock, but will also remove
768 * it from waiting list */
769 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
770 unlock_res_and_lock(lock);
771 ldlm_lock_cancel(lock);
773 lock_res_and_lock(lock);
775 /* start the lock-timeout clock */
776 ldlm_add_waiting_lock(lock);
779 unlock_res_and_lock(lock);
781 if (lock->l_export && lock->l_export->exp_nid_stats &&
782 lock->l_export->exp_nid_stats->nid_ldlm_stats)
783 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
784 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
786 rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
791 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
793 struct ldlm_resource *res = lock->l_resource;
794 struct ldlm_request *body;
795 struct ptlrpc_request *req;
799 LASSERT(lock != NULL);
801 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
802 &RQF_LDLM_GL_CALLBACK,
803 LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK);
808 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
809 body->lock_handle[0] = lock->l_remote_handle;
810 ldlm_lock2desc(lock, &body->lock_desc);
812 lock_res_and_lock(lock);
813 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
814 lock->l_resource->lr_lvb_len);
815 unlock_res_and_lock(lock);
816 res = lock->l_resource;
817 ptlrpc_request_set_replen(req);
820 req->rq_send_state = LUSTRE_IMP_FULL;
821 /* ptlrpc_prep_req already set timeout */
823 req->rq_timeout = ldlm_get_rq_timeout();
825 if (lock->l_export && lock->l_export->exp_nid_stats &&
826 lock->l_export->exp_nid_stats->nid_ldlm_stats)
827 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
828 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
830 rc = ptlrpc_queue_wait(req);
831 if (rc == -ELDLM_NO_LOCK_DATA)
832 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
834 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
836 rc = ldlm_res_lvbo_update(res, req->rq_repmsg,
838 ptlrpc_req_finished(req);
840 ldlm_reprocess_all(res);
846 extern unsigned long long lu_time_stamp_get(void);
848 #define lu_time_stamp_get() time(NULL)
851 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
852 struct lprocfs_stats *srv_stats)
854 int lock_type = 0, op = 0;
856 lock_type = dlm_req->lock_desc.l_resource.lr_type;
860 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
863 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
864 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
866 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
869 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
872 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
880 lprocfs_counter_incr(srv_stats, op);
886 * Main server-side entry point into LDLM. This is called by ptlrpc service
887 * threads to carry out client lock enqueueing requests.
889 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
890 struct ptlrpc_request *req,
891 const struct ldlm_request *dlm_req,
892 const struct ldlm_callback_suite *cbs)
894 struct ldlm_reply *dlm_rep;
896 ldlm_error_t err = ELDLM_OK;
897 struct ldlm_lock *lock = NULL;
902 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
904 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
905 flags = dlm_req->lock_flags;
907 LASSERT(req->rq_export);
909 if (req->rq_rqbd->rqbd_service->srv_stats)
910 ldlm_svc_get_eopc(dlm_req,
911 req->rq_rqbd->rqbd_service->srv_stats);
913 if (req->rq_export && req->rq_export->exp_nid_stats &&
914 req->rq_export->exp_nid_stats->nid_ldlm_stats)
915 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
916 LDLM_ENQUEUE - LDLM_FIRST_OPC);
918 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
919 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
920 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
921 dlm_req->lock_desc.l_resource.lr_type);
922 GOTO(out, rc = -EFAULT);
925 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
926 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
927 dlm_req->lock_desc.l_req_mode &
928 (dlm_req->lock_desc.l_req_mode-1))) {
929 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
930 dlm_req->lock_desc.l_req_mode);
931 GOTO(out, rc = -EFAULT);
934 if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
935 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
937 DEBUG_REQ(D_ERROR, req,
938 "PLAIN lock request from IBITS client?");
939 GOTO(out, rc = -EPROTO);
941 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
943 DEBUG_REQ(D_ERROR, req,
944 "IBITS lock request from unaware client?");
945 GOTO(out, rc = -EPROTO);
949 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
950 against server's _CONNECT_SUPPORTED flags? (I don't want to use
951 ibits for mgc/mgs) */
953 /* INODEBITS_INTEROP: Perform conversion from plain lock to
954 * inodebits lock if client does not support them. */
955 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
956 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
957 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
958 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
959 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
960 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
961 dlm_req->lock_desc.l_req_mode = LCK_CR;
965 if (unlikely(flags & LDLM_FL_REPLAY)) {
966 /* Find an existing lock in the per-export lock hash */
967 lock = lustre_hash_lookup(req->rq_export->exp_lock_hash,
968 (void *)&dlm_req->lock_handle[0]);
970 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
971 LPX64, lock->l_handle.h_cookie);
972 GOTO(existing_lock, rc = 0);
976 /* The lock's callback data might be set in the policy function */
977 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
978 dlm_req->lock_desc.l_resource.lr_type,
979 dlm_req->lock_desc.l_req_mode,
980 cbs->lcs_blocking, cbs->lcs_completion,
981 cbs->lcs_glimpse, NULL, 0);
984 GOTO(out, rc = -ENOMEM);
986 do_gettimeofday(&lock->l_enqueued_time);
987 lock->l_remote_handle = dlm_req->lock_handle[0];
988 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
990 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
991 /* Don't enqueue a lock onto the export if it has already
992 * been evicted. Cancel it now instead. (bug 3822) */
993 if (req->rq_export->exp_failed) {
994 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
995 GOTO(out, rc = -ENOTCONN);
997 lock->l_export = class_export_get(req->rq_export);
999 if (lock->l_export->exp_lock_hash)
1000 lustre_hash_add(lock->l_export->exp_lock_hash,
1001 &lock->l_remote_handle,
1006 if (flags & LDLM_FL_HAS_INTENT) {
1007 /* In this case, the reply buffer is allocated deep in
1008 * local_lock_enqueue by the policy function. */
1011 lock_res_and_lock(lock);
1012 if (lock->l_resource->lr_lvb_len) {
1013 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1015 lock->l_resource->lr_lvb_len);
1017 unlock_res_and_lock(lock);
1019 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1020 GOTO(out, rc = -ENOMEM);
1022 rc = req_capsule_server_pack(&req->rq_pill);
1027 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1028 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1029 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1030 lock->l_req_extent = lock->l_policy_data.l_extent;
1032 err = ldlm_lock_enqueue(ns, &lock, cookie, (int *)&flags);
1036 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1037 dlm_rep->lock_flags = flags;
1039 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1040 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1042 /* We never send a blocking AST until the lock is granted, but
1043 * we can tell it right now */
1044 lock_res_and_lock(lock);
1046 /* Now take into account flags to be inherited from original lock
1047 request both in reply to client and in our own lock flags. */
1048 dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1049 lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1051 /* Don't move a pending lock onto the export if it has already
1052 * been evicted. Cancel it now instead. (bug 5683) */
1053 if (unlikely(req->rq_export->exp_failed ||
1054 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1055 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1057 } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1058 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
1059 if (lock->l_granted_mode == lock->l_req_mode) {
1061 * Only cancel lock if it was granted, because it would
1062 * be destroyed immediatelly and would never be granted
1063 * in the future, causing timeouts on client. Not
1064 * granted lock will be cancelled immediatelly after
1065 * sending completion AST.
1067 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1068 unlock_res_and_lock(lock);
1069 ldlm_lock_cancel(lock);
1070 lock_res_and_lock(lock);
1072 ldlm_add_waiting_lock(lock);
1075 /* Make sure we never ever grant usual metadata locks to liblustre
1077 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1078 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1079 req->rq_export->exp_libclient) {
1080 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1081 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1082 CERROR("Granting sync lock to libclient. "
1083 "req fl %d, rep fl %d, lock fl %d\n",
1084 dlm_req->lock_flags, dlm_rep->lock_flags,
1086 LDLM_ERROR(lock, "sync lock");
1087 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1088 struct ldlm_intent *it;
1090 it = req_capsule_client_get(&req->rq_pill,
1093 CERROR("This is intent %s ("LPU64")\n",
1094 ldlm_it2str(it->opc), it->opc);
1100 unlock_res_and_lock(lock);
1104 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1105 if (!req->rq_packed_final) {
1106 err = lustre_pack_reply(req, 1, NULL, NULL);
1111 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1112 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1114 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1115 "(err=%d, rc=%d)", err, rc);
1117 lock_res_and_lock(lock);
1119 if (lock->l_resource->lr_lvb_len > 0) {
1122 lvb = req_capsule_server_get(&req->rq_pill,
1124 LASSERTF(lvb != NULL, "req %p, lock %p\n",
1127 memcpy(lvb, lock->l_resource->lr_lvb_data,
1128 lock->l_resource->lr_lvb_len);
1131 ldlm_resource_unlink_lock(lock);
1132 ldlm_lock_destroy_nolock(lock);
1134 unlock_res_and_lock(lock);
1136 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1137 ldlm_reprocess_all(lock->l_resource);
1139 LDLM_LOCK_PUT(lock);
1142 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1148 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1149 ldlm_completion_callback completion_callback,
1150 ldlm_blocking_callback blocking_callback,
1151 ldlm_glimpse_callback glimpse_callback)
1153 struct ldlm_request *dlm_req;
1154 struct ldlm_callback_suite cbs = {
1155 .lcs_completion = completion_callback,
1156 .lcs_blocking = blocking_callback,
1157 .lcs_glimpse = glimpse_callback
1161 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1162 if (dlm_req != NULL) {
1163 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1164 req, dlm_req, &cbs);
1171 int ldlm_handle_convert0(struct ptlrpc_request *req,
1172 const struct ldlm_request *dlm_req)
1174 struct ldlm_reply *dlm_rep;
1175 struct ldlm_lock *lock;
1179 if (req->rq_export && req->rq_export->exp_nid_stats &&
1180 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1181 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1182 LDLM_CONVERT - LDLM_FIRST_OPC);
1184 rc = req_capsule_server_pack(&req->rq_pill);
1188 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1189 dlm_rep->lock_flags = dlm_req->lock_flags;
1191 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1193 req->rq_status = EINVAL;
1197 LDLM_DEBUG(lock, "server-side convert handler START");
1199 do_gettimeofday(&lock->l_enqueued_time);
1200 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1201 &dlm_rep->lock_flags);
1203 if (ldlm_del_waiting_lock(lock))
1204 LDLM_DEBUG(lock, "converted waiting lock");
1207 req->rq_status = EDEADLOCK;
1212 if (!req->rq_status)
1213 ldlm_reprocess_all(lock->l_resource);
1214 LDLM_DEBUG(lock, "server-side convert handler END");
1215 LDLM_LOCK_PUT(lock);
1217 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1222 int ldlm_handle_convert(struct ptlrpc_request *req)
1225 struct ldlm_request *dlm_req;
1227 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1228 if (dlm_req != NULL) {
1229 rc = ldlm_handle_convert0(req, dlm_req);
1231 CERROR ("Can't unpack dlm_req\n");
1237 /* Cancel all the locks whos handles are packed into ldlm_request */
1238 int ldlm_request_cancel(struct ptlrpc_request *req,
1239 const struct ldlm_request *dlm_req, int first)
1241 struct ldlm_resource *res, *pres = NULL;
1242 struct ldlm_lock *lock;
1243 int i, count, done = 0;
1246 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1250 /* There is no lock on the server at the replay time,
1251 * skip lock cancelling to make replay tests to pass. */
1252 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1255 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1256 "starting at %d", count, first);
1258 for (i = first; i < count; i++) {
1259 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1261 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1262 "lock (cookie "LPU64")",
1263 dlm_req->lock_handle[i].cookie);
1267 res = lock->l_resource;
1272 ldlm_reprocess_all(pres);
1273 ldlm_resource_putref(pres);
1276 ldlm_resource_getref(res);
1277 ldlm_res_lvbo_update(res, NULL, 0, 1);
1281 ldlm_lock_cancel(lock);
1282 LDLM_LOCK_PUT(lock);
1285 ldlm_reprocess_all(pres);
1286 ldlm_resource_putref(pres);
1288 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1292 int ldlm_handle_cancel(struct ptlrpc_request *req)
1294 struct ldlm_request *dlm_req;
1298 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1299 if (dlm_req == NULL) {
1300 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1304 if (req->rq_export && req->rq_export->exp_nid_stats &&
1305 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1306 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1307 LDLM_CANCEL - LDLM_FIRST_OPC);
1309 rc = req_capsule_server_pack(&req->rq_pill);
1313 if (!ldlm_request_cancel(req, dlm_req, 0))
1314 req->rq_status = ESTALE;
1316 if (ptlrpc_reply(req) != 0)
1322 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1323 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1328 LDLM_DEBUG(lock, "client blocking AST callback handler START");
1330 lock_res_and_lock(lock);
1331 lock->l_flags |= LDLM_FL_CBPENDING;
1333 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1334 lock->l_flags |= LDLM_FL_CANCEL;
1336 do_ast = (!lock->l_readers && !lock->l_writers);
1337 unlock_res_and_lock(lock);
1340 LDLM_DEBUG(lock, "already unused, calling "
1341 "callback (%p)", lock->l_blocking_ast);
1342 if (lock->l_blocking_ast != NULL)
1343 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1346 LDLM_DEBUG(lock, "Lock still has references, will be"
1347 " cancelled later");
1350 LDLM_DEBUG(lock, "client blocking callback handler END");
1351 LDLM_LOCK_PUT(lock);
1355 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1356 struct ldlm_namespace *ns,
1357 struct ldlm_request *dlm_req,
1358 struct ldlm_lock *lock)
1360 CFS_LIST_HEAD(ast_list);
1363 LDLM_DEBUG(lock, "client completion callback handler START");
1365 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1366 int to = cfs_time_seconds(1);
1368 cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE, to);
1369 if (lock->l_granted_mode == lock->l_req_mode ||
1375 lock_res_and_lock(lock);
1376 if (lock->l_destroyed ||
1377 lock->l_granted_mode == lock->l_req_mode) {
1378 /* bug 11300: the lock has already been granted */
1379 unlock_res_and_lock(lock);
1380 LDLM_DEBUG(lock, "Double grant race happened");
1381 LDLM_LOCK_PUT(lock);
1386 /* If we receive the completion AST before the actual enqueue returned,
1387 * then we might need to switch lock modes, resources, or extents. */
1388 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1389 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1390 LDLM_DEBUG(lock, "completion AST, new lock mode");
1393 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1394 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1395 LDLM_DEBUG(lock, "completion AST, new policy data");
1398 ldlm_resource_unlink_lock(lock);
1399 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1400 &lock->l_resource->lr_name,
1401 sizeof(lock->l_resource->lr_name)) != 0) {
1402 unlock_res_and_lock(lock);
1403 if (ldlm_lock_change_resource(ns, lock,
1404 &dlm_req->lock_desc.l_resource.lr_name) != 0) {
1405 LDLM_ERROR(lock, "Failed to allocate resource");
1406 LDLM_LOCK_PUT(lock);
1410 LDLM_DEBUG(lock, "completion AST, new resource");
1411 CERROR("change resource!\n");
1412 lock_res_and_lock(lock);
1415 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1416 /* BL_AST locks are not needed in lru.
1417 * let ldlm_cancel_lru() be fast. */
1418 ldlm_lock_remove_from_lru(lock);
1419 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1420 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1423 if (lock->l_lvb_len) {
1424 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
1425 RCL_CLIENT) < lock->l_lvb_len) {
1426 LDLM_ERROR(lock, "completion AST did not contain "
1429 void *lvb = req_capsule_client_swab_get(&req->rq_pill,
1431 (void *)lock->l_lvb_swabber);
1432 memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1436 ldlm_grant_lock(lock, &ast_list);
1437 unlock_res_and_lock(lock);
1439 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1441 ldlm_run_ast_work(&ast_list, LDLM_WORK_CP_AST);
1443 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1445 LDLM_LOCK_PUT(lock);
1449 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1450 struct ldlm_namespace *ns,
1451 struct ldlm_request *dlm_req,
1452 struct ldlm_lock *lock)
1457 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1459 if (lock->l_glimpse_ast != NULL)
1460 rc = lock->l_glimpse_ast(lock, req);
1462 if (req->rq_repmsg != NULL) {
1465 req->rq_status = rc;
1469 lock_res_and_lock(lock);
1470 if (lock->l_granted_mode == LCK_PW &&
1471 !lock->l_readers && !lock->l_writers &&
1472 cfs_time_after(cfs_time_current(),
1473 cfs_time_add(lock->l_last_used,
1474 cfs_time_seconds(10)))) {
1475 unlock_res_and_lock(lock);
1476 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1477 ldlm_handle_bl_callback(ns, NULL, lock);
1482 unlock_res_and_lock(lock);
1483 LDLM_LOCK_PUT(lock);
1487 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1489 if (req->rq_no_reply)
1492 req->rq_status = rc;
1493 if (!req->rq_packed_final) {
1494 rc = lustre_pack_reply(req, 1, NULL, NULL);
1498 return ptlrpc_reply(req);
1502 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1503 struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
1504 struct list_head *cancels, int count)
1506 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1507 struct ldlm_bl_work_item *blwi;
1510 if (cancels && count == 0)
1513 OBD_ALLOC(blwi, sizeof(*blwi));
1519 blwi->blwi_ld = *ld;
1521 list_add(&blwi->blwi_head, cancels);
1522 list_del_init(cancels);
1523 blwi->blwi_count = count;
1525 blwi->blwi_lock = lock;
1527 spin_lock(&blp->blp_lock);
1528 if (lock && lock->l_flags & LDLM_FL_DISCARD_DATA) {
1529 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1530 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1532 /* other blocking callbacks are added to the regular list */
1533 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1535 cfs_waitq_signal(&blp->blp_waitq);
1536 spin_unlock(&blp->blp_lock);
1542 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1543 struct ldlm_lock *lock)
1546 RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0));
1552 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1553 struct list_head *cancels, int count)
1556 RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count));
1562 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1563 static int ldlm_callback_handler(struct ptlrpc_request *req)
1565 struct ldlm_namespace *ns;
1566 struct ldlm_request *dlm_req;
1567 struct ldlm_lock *lock;
1571 /* Requests arrive in sender's byte order. The ptlrpc service
1572 * handler has already checked and, if necessary, byte-swapped the
1573 * incoming request message body, but I am responsible for the
1574 * message buffers. */
1576 /* do nothing for sec context finalize */
1577 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
1580 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1582 if (req->rq_export == NULL) {
1583 struct ldlm_request *dlm_req;
1585 CDEBUG(D_RPCTRACE, "operation %d from %s with bad "
1586 "export cookie "LPX64"; this is "
1587 "normal if this node rebooted with a lock held\n",
1588 lustre_msg_get_opc(req->rq_reqmsg),
1589 libcfs_id2str(req->rq_peer),
1590 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
1592 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1593 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1594 if (dlm_req != NULL)
1595 CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1596 dlm_req->lock_handle[0].cookie);
1598 ldlm_callback_reply(req, -ENOTCONN);
1602 LASSERT(req->rq_export != NULL);
1603 LASSERT(req->rq_export->exp_obd != NULL);
1605 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1606 case LDLM_BL_CALLBACK:
1607 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK))
1610 case LDLM_CP_CALLBACK:
1611 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK))
1614 case LDLM_GL_CALLBACK:
1615 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK))
1618 case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
1619 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1620 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1622 rc = llog_origin_handle_cancel(req);
1623 ldlm_callback_reply(req, rc);
1625 case OBD_QC_CALLBACK:
1626 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
1627 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
1629 rc = target_handle_qc_callback(req);
1630 ldlm_callback_reply(req, rc);
1634 /* reply in handler */
1635 req_capsule_set(&req->rq_pill, &RQF_MDS_QUOTA_DQACQ);
1636 rc = target_handle_dqacq_callback(req);
1638 case LLOG_ORIGIN_HANDLE_CREATE:
1639 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
1640 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1642 rc = llog_origin_handle_create(req);
1643 ldlm_callback_reply(req, rc);
1645 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1646 req_capsule_set(&req->rq_pill,
1647 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
1648 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1650 rc = llog_origin_handle_next_block(req);
1651 ldlm_callback_reply(req, rc);
1653 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1654 req_capsule_set(&req->rq_pill,
1655 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
1656 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1658 rc = llog_origin_handle_read_header(req);
1659 ldlm_callback_reply(req, rc);
1661 case LLOG_ORIGIN_HANDLE_CLOSE:
1662 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1664 rc = llog_origin_handle_close(req);
1665 ldlm_callback_reply(req, rc);
1668 CERROR("unknown opcode %u\n",
1669 lustre_msg_get_opc(req->rq_reqmsg));
1670 ldlm_callback_reply(req, -EPROTO);
1674 ns = req->rq_export->exp_obd->obd_namespace;
1675 LASSERT(ns != NULL);
1677 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1679 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1680 if (dlm_req == NULL) {
1681 ldlm_callback_reply(req, -EPROTO);
1685 /* Force a known safe race, send a cancel to the server for a lock
1686 * which the server has already started a blocking callback on. */
1687 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
1688 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1689 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]);
1691 CERROR("ldlm_cli_cancel: %d\n", rc);
1694 lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle[0]);
1696 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
1697 "disappeared\n", dlm_req->lock_handle[0].cookie);
1698 ldlm_callback_reply(req, -EINVAL);
1702 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1703 lock_res_and_lock(lock);
1704 lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1705 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1706 /* If somebody cancels locks and cache is already droped,
1707 * we can tell the server we have no lock. Otherwise, we
1708 * should send cancel after dropping the cache. */
1709 if ((lock->l_flags & LDLM_FL_CANCELING) &&
1710 (lock->l_flags & LDLM_FL_BL_DONE)) {
1711 LDLM_DEBUG(lock, "callback on lock "
1712 LPX64" - lock disappeared\n",
1713 dlm_req->lock_handle[0].cookie);
1714 unlock_res_and_lock(lock);
1715 LDLM_LOCK_PUT(lock);
1716 ldlm_callback_reply(req, -EINVAL);
1719 /* BL_AST locks are not needed in lru.
1720 * let ldlm_cancel_lru() be fast. */
1721 ldlm_lock_remove_from_lru(lock);
1722 lock->l_flags |= LDLM_FL_BL_AST;
1724 unlock_res_and_lock(lock);
1726 /* We want the ost thread to get this reply so that it can respond
1727 * to ost requests (write cache writeback) that might be triggered
1730 * But we'd also like to be able to indicate in the reply that we're
1731 * cancelling right now, because it's unused, or have an intent result
1732 * in the reply, so we might have to push the responsibility for sending
1733 * the reply down into the AST handlers, alas. */
1735 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1736 case LDLM_BL_CALLBACK:
1737 CDEBUG(D_INODE, "blocking ast\n");
1738 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
1739 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK))
1740 ldlm_callback_reply(req, 0);
1741 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
1742 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1744 case LDLM_CP_CALLBACK:
1745 CDEBUG(D_INODE, "completion ast\n");
1746 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
1747 ldlm_callback_reply(req, 0);
1748 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1750 case LDLM_GL_CALLBACK:
1751 CDEBUG(D_INODE, "glimpse ast\n");
1752 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
1753 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1756 LBUG(); /* checked above */
1762 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1767 /* Requests arrive in sender's byte order. The ptlrpc service
1768 * handler has already checked and, if necessary, byte-swapped the
1769 * incoming request message body, but I am responsible for the
1770 * message buffers. */
1772 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1774 if (req->rq_export == NULL) {
1775 struct ldlm_request *dlm_req;
1777 CERROR("operation %d from %s with bad export cookie "LPU64"\n",
1778 lustre_msg_get_opc(req->rq_reqmsg),
1779 libcfs_id2str(req->rq_peer),
1780 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
1782 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1783 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1784 if (dlm_req != NULL)
1785 ldlm_lock_dump_handle(D_ERROR,
1786 &dlm_req->lock_handle[0]);
1787 ldlm_callback_reply(req, -ENOTCONN);
1791 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1793 /* XXX FIXME move this back to mds/handler.c, bug 249 */
1795 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
1796 CDEBUG(D_INODE, "cancel\n");
1797 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
1799 rc = ldlm_handle_cancel(req);
1803 case OBD_LOG_CANCEL:
1804 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1805 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1807 rc = llog_origin_handle_cancel(req);
1808 ldlm_callback_reply(req, rc);
1811 CERROR("invalid opcode %d\n",
1812 lustre_msg_get_opc(req->rq_reqmsg));
1813 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1814 ldlm_callback_reply(req, -EINVAL);
1820 void ldlm_revoke_lock_cb(void *obj, void *data)
1822 struct list_head *rpc_list = data;
1823 struct ldlm_lock *lock = obj;
1825 lock_res_and_lock(lock);
1827 if (lock->l_req_mode != lock->l_granted_mode) {
1828 unlock_res_and_lock(lock);
1832 LASSERT(lock->l_resource);
1833 if (lock->l_resource->lr_type != LDLM_IBITS &&
1834 lock->l_resource->lr_type != LDLM_PLAIN) {
1835 unlock_res_and_lock(lock);
1839 if (lock->l_flags & LDLM_FL_AST_SENT) {
1840 unlock_res_and_lock(lock);
1844 LASSERT(lock->l_blocking_ast);
1845 LASSERT(!lock->l_blocking_lock);
1847 lock->l_flags |= LDLM_FL_AST_SENT;
1848 if (lock->l_export && lock->l_export->exp_lock_hash)
1849 lustre_hash_del(lock->l_export->exp_lock_hash,
1850 &lock->l_remote_handle, &lock->l_exp_hash);
1851 list_add_tail(&lock->l_rk_ast, rpc_list);
1852 LDLM_LOCK_GET(lock);
1854 unlock_res_and_lock(lock);
1857 void ldlm_revoke_export_locks(struct obd_export *exp)
1859 struct list_head rpc_list;
1862 CFS_INIT_LIST_HEAD(&rpc_list);
1863 lustre_hash_for_each_empty(exp->exp_lock_hash,
1864 ldlm_revoke_lock_cb, &rpc_list);
1865 ldlm_run_ast_work(&rpc_list, LDLM_WORK_REVOKE_AST);
1871 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1873 struct ldlm_bl_work_item *blwi = NULL;
1874 static unsigned int num_bl = 0;
1876 spin_lock(&blp->blp_lock);
1877 /* process a request from the blp_list at least every blp_num_threads */
1878 if (!list_empty(&blp->blp_list) &&
1879 (list_empty(&blp->blp_prio_list) || num_bl == 0))
1880 blwi = list_entry(blp->blp_list.next,
1881 struct ldlm_bl_work_item, blwi_entry);
1883 if (!list_empty(&blp->blp_prio_list))
1884 blwi = list_entry(blp->blp_prio_list.next,
1885 struct ldlm_bl_work_item, blwi_entry);
1888 if (++num_bl >= atomic_read(&blp->blp_num_threads))
1890 list_del(&blwi->blwi_entry);
1892 spin_unlock(&blp->blp_lock);
1897 /* This only contains temporary data until the thread starts */
1898 struct ldlm_bl_thread_data {
1899 char bltd_name[CFS_CURPROC_COMM_MAX];
1900 struct ldlm_bl_pool *bltd_blp;
1901 struct completion bltd_comp;
1905 static int ldlm_bl_thread_main(void *arg);
1907 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
1909 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
1912 init_completion(&bltd.bltd_comp);
1913 rc = cfs_kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1915 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
1916 atomic_read(&blp->blp_num_threads), rc);
1919 wait_for_completion(&bltd.bltd_comp);
1924 static int ldlm_bl_thread_main(void *arg)
1926 struct ldlm_bl_pool *blp;
1930 struct ldlm_bl_thread_data *bltd = arg;
1932 blp = bltd->bltd_blp;
1934 bltd->bltd_num = atomic_inc_return(&blp->blp_num_threads) - 1;
1935 atomic_inc(&blp->blp_busy_threads);
1937 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
1938 "ldlm_bl_%02d", bltd->bltd_num);
1939 cfs_daemonize(bltd->bltd_name);
1941 complete(&bltd->bltd_comp);
1942 /* cannot use bltd after this, it is only on caller's stack */
1946 struct l_wait_info lwi = { 0 };
1947 struct ldlm_bl_work_item *blwi = NULL;
1949 blwi = ldlm_bl_get_work(blp);
1954 atomic_dec(&blp->blp_busy_threads);
1955 l_wait_event_exclusive(blp->blp_waitq,
1956 (blwi = ldlm_bl_get_work(blp)) != NULL,
1958 busy = atomic_inc_return(&blp->blp_busy_threads);
1960 if (blwi->blwi_ns == NULL)
1961 /* added by ldlm_cleanup() */
1964 /* Not fatal if racy and have a few too many threads */
1965 if (unlikely(busy < blp->blp_max_threads &&
1966 busy >= atomic_read(&blp->blp_num_threads)))
1967 /* discard the return value, we tried */
1968 ldlm_bl_thread_start(blp);
1970 if (blwi->blwi_ns == NULL)
1971 /* added by ldlm_cleanup() */
1975 if (blwi->blwi_count) {
1976 /* The special case when we cancel locks in lru
1977 * asynchronously, we pass the list of locks here.
1978 * Thus lock is marked LDLM_FL_CANCELING, and already
1979 * canceled locally. */
1980 ldlm_cli_cancel_list(&blwi->blwi_head,
1981 blwi->blwi_count, NULL, 0);
1983 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1986 OBD_FREE(blwi, sizeof(*blwi));
1989 atomic_dec(&blp->blp_busy_threads);
1990 atomic_dec(&blp->blp_num_threads);
1991 complete(&blp->blp_comp);
1997 static int ldlm_setup(void);
1998 static int ldlm_cleanup(void);
2000 int ldlm_get_ref(void)
2004 mutex_down(&ldlm_ref_sem);
2005 if (++ldlm_refcount == 1) {
2010 mutex_up(&ldlm_ref_sem);
2015 void ldlm_put_ref(void)
2018 mutex_down(&ldlm_ref_sem);
2019 if (ldlm_refcount == 1) {
2020 int rc = ldlm_cleanup();
2022 CERROR("ldlm_cleanup failed: %d\n", rc);
2028 mutex_up(&ldlm_ref_sem);
2034 * Export handle<->lock hash operations.
2037 ldlm_export_lock_hash(lustre_hash_t *lh, void *key, unsigned mask)
2039 return lh_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2043 ldlm_export_lock_key(struct hlist_node *hnode)
2045 struct ldlm_lock *lock;
2048 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2049 RETURN(&lock->l_remote_handle);
2053 ldlm_export_lock_compare(void *key, struct hlist_node *hnode)
2056 RETURN(lustre_handle_equal(ldlm_export_lock_key(hnode), key));
2060 ldlm_export_lock_get(struct hlist_node *hnode)
2062 struct ldlm_lock *lock;
2065 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2066 LDLM_LOCK_GET(lock);
2072 ldlm_export_lock_put(struct hlist_node *hnode)
2074 struct ldlm_lock *lock;
2077 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2078 LDLM_LOCK_PUT(lock);
2083 static lustre_hash_ops_t ldlm_export_lock_ops = {
2084 .lh_hash = ldlm_export_lock_hash,
2085 .lh_key = ldlm_export_lock_key,
2086 .lh_compare = ldlm_export_lock_compare,
2087 .lh_get = ldlm_export_lock_get,
2088 .lh_put = ldlm_export_lock_put
2091 int ldlm_init_export(struct obd_export *exp)
2095 exp->exp_lock_hash =
2096 lustre_hash_init(obd_uuid2str(&exp->exp_client_uuid),
2097 128, 65536, &ldlm_export_lock_ops, LH_REHASH);
2099 if (!exp->exp_lock_hash)
2104 EXPORT_SYMBOL(ldlm_init_export);
2106 void ldlm_destroy_export(struct obd_export *exp)
2109 lustre_hash_exit(exp->exp_lock_hash);
2110 exp->exp_lock_hash = NULL;
2113 EXPORT_SYMBOL(ldlm_destroy_export);
2115 static int ldlm_setup(void)
2117 struct ldlm_bl_pool *blp;
2119 int ldlm_min_threads = LDLM_THREADS_AUTO_MIN;
2120 int ldlm_max_threads = LDLM_THREADS_AUTO_MAX;
2126 if (ldlm_state != NULL)
2129 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2130 if (ldlm_state == NULL)
2134 rc = ldlm_proc_setup();
2140 if (ldlm_num_threads) {
2141 /* If ldlm_num_threads is set, it is the min and the max. */
2142 if (ldlm_num_threads > LDLM_THREADS_AUTO_MAX)
2143 ldlm_num_threads = LDLM_THREADS_AUTO_MAX;
2144 if (ldlm_num_threads < LDLM_THREADS_AUTO_MIN)
2145 ldlm_num_threads = LDLM_THREADS_AUTO_MIN;
2146 ldlm_min_threads = ldlm_max_threads = ldlm_num_threads;
2150 ldlm_state->ldlm_cb_service =
2151 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2152 LDLM_MAXREPSIZE, LDLM_CB_REQUEST_PORTAL,
2153 LDLM_CB_REPLY_PORTAL, 1800,
2154 ldlm_callback_handler, "ldlm_cbd",
2155 ldlm_svc_proc_dir, NULL,
2156 ldlm_min_threads, ldlm_max_threads,
2158 LCT_MD_THREAD|LCT_DT_THREAD);
2160 if (!ldlm_state->ldlm_cb_service) {
2161 CERROR("failed to start service\n");
2162 GOTO(out_proc, rc = -ENOMEM);
2165 ldlm_state->ldlm_cancel_service =
2166 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2167 LDLM_MAXREPSIZE, LDLM_CANCEL_REQUEST_PORTAL,
2168 LDLM_CANCEL_REPLY_PORTAL, 6000,
2169 ldlm_cancel_handler, "ldlm_canceld",
2170 ldlm_svc_proc_dir, NULL,
2171 ldlm_min_threads, ldlm_max_threads,
2173 LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD);
2175 if (!ldlm_state->ldlm_cancel_service) {
2176 CERROR("failed to start service\n");
2177 GOTO(out_proc, rc = -ENOMEM);
2180 OBD_ALLOC(blp, sizeof(*blp));
2182 GOTO(out_proc, rc = -ENOMEM);
2183 ldlm_state->ldlm_bl_pool = blp;
2185 spin_lock_init(&blp->blp_lock);
2186 CFS_INIT_LIST_HEAD(&blp->blp_list);
2187 CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2188 cfs_waitq_init(&blp->blp_waitq);
2189 atomic_set(&blp->blp_num_threads, 0);
2190 atomic_set(&blp->blp_busy_threads, 0);
2191 blp->blp_min_threads = ldlm_min_threads;
2192 blp->blp_max_threads = ldlm_max_threads;
2195 for (i = 0; i < blp->blp_min_threads; i++) {
2196 rc = ldlm_bl_thread_start(blp);
2198 GOTO(out_thread, rc);
2201 rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cancel_service);
2203 GOTO(out_thread, rc);
2205 rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cb_service);
2207 GOTO(out_thread, rc);
2209 CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2210 expired_lock_thread.elt_state = ELT_STOPPED;
2211 cfs_waitq_init(&expired_lock_thread.elt_waitq);
2213 CFS_INIT_LIST_HEAD(&waiting_locks_list);
2214 spin_lock_init(&waiting_locks_spinlock);
2215 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2217 rc = cfs_kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FILES);
2219 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2220 GOTO(out_thread, rc);
2223 wait_event(expired_lock_thread.elt_waitq,
2224 expired_lock_thread.elt_state == ELT_READY);
2228 rc = ldlm_pools_init();
2230 GOTO(out_thread, rc);
2236 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2237 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2242 ldlm_proc_cleanup();
2245 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2250 static int ldlm_cleanup(void)
2253 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2257 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2258 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2259 CERROR("ldlm still has namespaces; clean these up first.\n");
2260 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2261 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2270 while (atomic_read(&blp->blp_num_threads) > 0) {
2271 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2273 init_completion(&blp->blp_comp);
2275 spin_lock(&blp->blp_lock);
2276 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2277 cfs_waitq_signal(&blp->blp_waitq);
2278 spin_unlock(&blp->blp_lock);
2280 wait_for_completion(&blp->blp_comp);
2282 OBD_FREE(blp, sizeof(*blp));
2284 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2285 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2286 ldlm_proc_cleanup();
2288 expired_lock_thread.elt_state = ELT_TERMINATE;
2289 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
2290 wait_event(expired_lock_thread.elt_waitq,
2291 expired_lock_thread.elt_state == ELT_STOPPED);
2293 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2294 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2297 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2303 int __init ldlm_init(void)
2305 init_mutex(&ldlm_ref_sem);
2306 init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
2307 init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
2308 ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
2309 sizeof(struct ldlm_resource), 0,
2310 SLAB_HWCACHE_ALIGN);
2311 if (ldlm_resource_slab == NULL)
2314 ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
2315 sizeof(struct ldlm_lock), 0,
2316 SLAB_HWCACHE_ALIGN);
2317 if (ldlm_lock_slab == NULL) {
2318 cfs_mem_cache_destroy(ldlm_resource_slab);
2322 ldlm_interval_slab = cfs_mem_cache_create("interval_node",
2323 sizeof(struct ldlm_interval),
2324 0, SLAB_HWCACHE_ALIGN);
2325 if (ldlm_interval_slab == NULL) {
2326 cfs_mem_cache_destroy(ldlm_resource_slab);
2327 cfs_mem_cache_destroy(ldlm_lock_slab);
2334 void __exit ldlm_exit(void)
2338 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
2339 rc = cfs_mem_cache_destroy(ldlm_resource_slab);
2340 LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
2341 rc = cfs_mem_cache_destroy(ldlm_lock_slab);
2342 LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
2343 rc = cfs_mem_cache_destroy(ldlm_interval_slab);
2344 LASSERTF(rc == 0, "couldn't free interval node slab\n");
2348 EXPORT_SYMBOL(ldlm_extent_shift_kms);
2351 EXPORT_SYMBOL(ldlm_get_processing_policy);
2352 EXPORT_SYMBOL(ldlm_lock2desc);
2353 EXPORT_SYMBOL(ldlm_register_intent);
2354 EXPORT_SYMBOL(ldlm_lockname);
2355 EXPORT_SYMBOL(ldlm_typename);
2356 EXPORT_SYMBOL(ldlm_lock2handle);
2357 EXPORT_SYMBOL(__ldlm_handle2lock);
2358 EXPORT_SYMBOL(ldlm_lock_get);
2359 EXPORT_SYMBOL(ldlm_lock_put);
2360 EXPORT_SYMBOL(ldlm_lock_fast_match);
2361 EXPORT_SYMBOL(ldlm_lock_fast_release);
2362 EXPORT_SYMBOL(ldlm_lock_match);
2363 EXPORT_SYMBOL(ldlm_lock_cancel);
2364 EXPORT_SYMBOL(ldlm_lock_addref);
2365 EXPORT_SYMBOL(ldlm_lock_decref);
2366 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
2367 EXPORT_SYMBOL(ldlm_lock_change_resource);
2368 EXPORT_SYMBOL(ldlm_lock_set_data);
2369 EXPORT_SYMBOL(ldlm_it2str);
2370 EXPORT_SYMBOL(ldlm_lock_dump);
2371 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2372 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
2373 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
2374 EXPORT_SYMBOL(ldlm_lock_allow_match);
2376 /* ldlm_request.c */
2377 EXPORT_SYMBOL(ldlm_completion_ast);
2378 EXPORT_SYMBOL(ldlm_blocking_ast);
2379 EXPORT_SYMBOL(ldlm_glimpse_ast);
2380 EXPORT_SYMBOL(ldlm_expired_completion_wait);
2381 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
2382 EXPORT_SYMBOL(ldlm_prep_elc_req);
2383 EXPORT_SYMBOL(ldlm_cli_convert);
2384 EXPORT_SYMBOL(ldlm_cli_enqueue);
2385 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
2386 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
2387 EXPORT_SYMBOL(ldlm_cli_cancel);
2388 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2389 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2390 EXPORT_SYMBOL(ldlm_cli_cancel_req);
2391 EXPORT_SYMBOL(ldlm_cli_join_lru);
2392 EXPORT_SYMBOL(ldlm_replay_locks);
2393 EXPORT_SYMBOL(ldlm_resource_foreach);
2394 EXPORT_SYMBOL(ldlm_namespace_foreach);
2395 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
2396 EXPORT_SYMBOL(ldlm_resource_iterate);
2397 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2398 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2401 EXPORT_SYMBOL(ldlm_server_blocking_ast);
2402 EXPORT_SYMBOL(ldlm_server_completion_ast);
2403 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
2404 EXPORT_SYMBOL(ldlm_handle_enqueue);
2405 EXPORT_SYMBOL(ldlm_handle_enqueue0);
2406 EXPORT_SYMBOL(ldlm_handle_cancel);
2407 EXPORT_SYMBOL(ldlm_request_cancel);
2408 EXPORT_SYMBOL(ldlm_handle_convert);
2409 EXPORT_SYMBOL(ldlm_handle_convert0);
2410 EXPORT_SYMBOL(ldlm_del_waiting_lock);
2411 EXPORT_SYMBOL(ldlm_get_ref);
2412 EXPORT_SYMBOL(ldlm_put_ref);
2413 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
2414 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2416 /* ldlm_resource.c */
2417 EXPORT_SYMBOL(ldlm_namespace_new);
2418 EXPORT_SYMBOL(ldlm_namespace_cleanup);
2419 EXPORT_SYMBOL(ldlm_namespace_free);
2420 EXPORT_SYMBOL(ldlm_namespace_dump);
2421 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
2422 EXPORT_SYMBOL(ldlm_resource_get);
2423 EXPORT_SYMBOL(ldlm_resource_putref);
2424 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
2427 EXPORT_SYMBOL(client_import_add_conn);
2428 EXPORT_SYMBOL(client_import_del_conn);
2429 EXPORT_SYMBOL(client_obd_setup);
2430 EXPORT_SYMBOL(client_obd_cleanup);
2431 EXPORT_SYMBOL(client_connect_import);
2432 EXPORT_SYMBOL(client_disconnect_export);
2433 EXPORT_SYMBOL(target_start_recovery_thread);
2434 EXPORT_SYMBOL(target_stop_recovery_thread);
2435 EXPORT_SYMBOL(target_handle_connect);
2436 EXPORT_SYMBOL(target_cleanup_recovery);
2437 EXPORT_SYMBOL(target_destroy_export);
2438 EXPORT_SYMBOL(target_cancel_recovery_timer);
2439 EXPORT_SYMBOL(target_send_reply);
2440 EXPORT_SYMBOL(target_queue_recovery_request);
2441 EXPORT_SYMBOL(target_handle_ping);
2442 EXPORT_SYMBOL(target_pack_pool_reply);
2443 EXPORT_SYMBOL(target_handle_disconnect);
2446 EXPORT_SYMBOL(lock_res_and_lock);
2447 EXPORT_SYMBOL(unlock_res_and_lock);