1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002-2004 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Phil Schwan <phil@clusterfs.com>
8 * This file is part of the Lustre file system, http://www.lustre.org
9 * Lustre is a trademark of Cluster File Systems, Inc.
11 * You may have signed or agreed to another license before downloading
12 * this software. If so, you are bound by the terms and conditions
13 * of that agreement, and the following does not apply to you. See the
14 * LICENSE file included with this distribution for more information.
16 * If you did not agree to a different license, then this copy of Lustre
17 * is open source software; you can redistribute it and/or modify it
18 * under the terms of version 2 of the GNU General Public License as
19 * published by the Free Software Foundation.
21 * In either case, Lustre is distributed in the hope that it will be
22 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
23 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
24 * license text for more details.
28 # define EXPORT_SYMTAB
30 #define DEBUG_SUBSYSTEM S_LDLM
33 # include <libcfs/libcfs.h>
35 # include <liblustre.h>
38 #include <lustre_dlm.h>
39 #include <obd_class.h>
40 #include <libcfs/list.h>
41 #include "ldlm_internal.h"
44 static int ldlm_num_threads;
45 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
46 "number of DLM service threads to start");
49 extern cfs_mem_cache_t *ldlm_resource_slab;
50 extern cfs_mem_cache_t *ldlm_lock_slab;
51 static struct semaphore ldlm_ref_sem;
52 static int ldlm_refcount;
56 static struct ldlm_state *ldlm_state;
58 inline cfs_time_t round_timeout(cfs_time_t timeout)
60 return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
63 /* timeout for initial callback (AST) reply (bz10399) */
64 static inline unsigned int ldlm_get_rq_timeout(void)
67 unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
69 return timeout < 1 ? 1 : timeout;
73 /* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */
74 static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */
75 static struct list_head waiting_locks_list;
76 static cfs_timer_t waiting_locks_timer;
78 static struct expired_lock_thread {
79 cfs_waitq_t elt_waitq;
82 struct list_head elt_expired_locks;
83 } expired_lock_thread;
88 #define ELT_TERMINATE 2
94 * blp_prio_list is used for callbacks that should be handled
95 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
98 struct list_head blp_prio_list;
101 * blp_list is used for all other callbacks which are likely
102 * to take longer to process.
104 struct list_head blp_list;
106 cfs_waitq_t blp_waitq;
107 struct completion blp_comp;
108 atomic_t blp_num_threads;
109 atomic_t blp_busy_threads;
114 struct ldlm_bl_work_item {
115 struct list_head blwi_entry;
116 struct ldlm_namespace *blwi_ns;
117 struct ldlm_lock_desc blwi_ld;
118 struct ldlm_lock *blwi_lock;
119 struct list_head blwi_head;
125 static inline int have_expired_locks(void)
130 spin_lock_bh(&waiting_locks_spinlock);
131 need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
132 spin_unlock_bh(&waiting_locks_spinlock);
137 static int expired_lock_main(void *arg)
139 struct list_head *expired = &expired_lock_thread.elt_expired_locks;
140 struct l_wait_info lwi = { 0 };
144 cfs_daemonize("ldlm_elt");
146 expired_lock_thread.elt_state = ELT_READY;
147 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
150 l_wait_event(expired_lock_thread.elt_waitq,
151 have_expired_locks() ||
152 expired_lock_thread.elt_state == ELT_TERMINATE,
155 spin_lock_bh(&waiting_locks_spinlock);
156 if (expired_lock_thread.elt_dump) {
157 spin_unlock_bh(&waiting_locks_spinlock);
159 /* from waiting_locks_callback, but not in timer */
160 libcfs_debug_dumplog();
161 libcfs_run_lbug_upcall(__FILE__,
162 "waiting_locks_callback",
163 expired_lock_thread.elt_dump);
165 spin_lock_bh(&waiting_locks_spinlock);
166 expired_lock_thread.elt_dump = 0;
171 while (!list_empty(expired)) {
172 struct obd_export *export;
173 struct ldlm_lock *lock;
175 lock = list_entry(expired->next, struct ldlm_lock,
177 if ((void *)lock < LP_POISON + CFS_PAGE_SIZE &&
178 (void *)lock >= LP_POISON) {
179 spin_unlock_bh(&waiting_locks_spinlock);
180 CERROR("free lock on elt list %p\n", lock);
183 list_del_init(&lock->l_pending_chain);
184 if ((void *)lock->l_export < LP_POISON + CFS_PAGE_SIZE &&
185 (void *)lock->l_export >= LP_POISON) {
186 CERROR("lock with free export on elt list %p\n",
188 lock->l_export = NULL;
189 LDLM_ERROR(lock, "free export");
192 export = class_export_get(lock->l_export);
193 spin_unlock_bh(&waiting_locks_spinlock);
196 class_fail_export(export);
197 class_export_put(export);
198 spin_lock_bh(&waiting_locks_spinlock);
200 spin_unlock_bh(&waiting_locks_spinlock);
202 if (do_dump && obd_dump_on_eviction) {
203 CERROR("dump the log upon eviction\n");
204 libcfs_debug_dumplog();
207 if (expired_lock_thread.elt_state == ELT_TERMINATE)
211 expired_lock_thread.elt_state = ELT_STOPPED;
212 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
216 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
218 /* This is called from within a timer interrupt and cannot schedule */
219 static void waiting_locks_callback(unsigned long unused)
221 struct ldlm_lock *lock, *last = NULL;
224 spin_lock_bh(&waiting_locks_spinlock);
225 while (!list_empty(&waiting_locks_list)) {
226 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
229 if (cfs_time_after(lock->l_callback_timeout, cfs_time_current()) ||
230 (lock->l_req_mode == LCK_GROUP))
233 if (ptlrpc_check_suspend()) {
234 /* there is a case when we talk to one mds, holding
235 * lock from another mds. this way we easily can get
236 * here, if second mds is being recovered. so, we
237 * suspend timeouts. bug 6019 */
239 LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
240 lock->l_export->exp_client_uuid.uuid,
241 lock->l_export->exp_connection->c_remote_uuid.uuid,
242 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
244 list_del_init(&lock->l_pending_chain);
245 spin_unlock_bh(&waiting_locks_spinlock);
246 ldlm_add_waiting_lock(lock);
250 /* if timeout overlaps the activation time of suspended timeouts
251 * then extend it to give a chance for client to reconnect */
252 if (cfs_time_before(cfs_time_sub(lock->l_callback_timeout,
253 cfs_time_seconds(obd_timeout)/2),
254 ptlrpc_suspend_wakeup_time())) {
255 LDLM_ERROR(lock, "extend timeout due to recovery: %s@%s nid %s ",
256 lock->l_export->exp_client_uuid.uuid,
257 lock->l_export->exp_connection->c_remote_uuid.uuid,
258 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
260 list_del_init(&lock->l_pending_chain);
261 spin_unlock_bh(&waiting_locks_spinlock);
262 ldlm_add_waiting_lock(lock);
266 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
267 "evicting client at %s ",
268 cfs_time_current_sec()- lock->l_enqueued_time.tv_sec,
270 lock->l_export->exp_connection->c_peer.nid));
274 list_del(&lock->l_pending_chain);
275 list_add(&lock->l_pending_chain,
276 &expired_lock_thread.elt_expired_locks);
279 if (!list_empty(&expired_lock_thread.elt_expired_locks)) {
280 if (obd_dump_on_timeout)
281 expired_lock_thread.elt_dump = __LINE__;
283 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
287 * Make sure the timer will fire again if we have any locks
290 if (!list_empty(&waiting_locks_list)) {
291 cfs_time_t timeout_rounded;
292 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
294 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
295 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
297 spin_unlock_bh(&waiting_locks_spinlock);
301 * Indicate that we're waiting for a client to call us back cancelling a given
302 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
303 * timer to fire appropriately. (We round up to the next second, to avoid
304 * floods of timer firings during periods of high lock contention and traffic).
306 * Called with the namespace lock held.
308 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock)
311 cfs_time_t timeout_rounded;
313 if (!list_empty(&lock->l_pending_chain))
316 timeout = ldlm_get_enq_timeout(lock);
318 lock->l_callback_timeout = cfs_time_shift(timeout);
320 timeout_rounded = round_timeout(lock->l_callback_timeout);
322 if (cfs_time_before(timeout_rounded,
323 cfs_timer_deadline(&waiting_locks_timer)) ||
324 !cfs_timer_is_armed(&waiting_locks_timer)) {
325 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
327 /* if the new lock has a shorter timeout than something earlier on
328 the list, we'll wait the longer amount of time; no big deal. */
329 list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
333 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
337 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
339 spin_lock_bh(&waiting_locks_spinlock);
340 if (lock->l_destroyed) {
341 static cfs_time_t next;
342 spin_unlock_bh(&waiting_locks_spinlock);
343 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
344 if (cfs_time_after(cfs_time_current(), next)) {
345 next = cfs_time_shift(14400);
346 libcfs_debug_dumpstack(NULL);
351 ret = __ldlm_add_waiting_lock(lock);
352 spin_unlock_bh(&waiting_locks_spinlock);
354 LDLM_DEBUG(lock, "%sadding to wait list",
355 ret == 0 ? "not re-" : "");
360 * Remove a lock from the pending list, likely because it had its cancellation
361 * callback arrive without incident. This adjusts the lock-timeout timer if
362 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
364 * Called with namespace lock held.
366 int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
368 struct list_head *list_next;
370 if (list_empty(&lock->l_pending_chain))
373 list_next = lock->l_pending_chain.next;
374 if (lock->l_pending_chain.prev == &waiting_locks_list) {
375 /* Removing the head of the list, adjust timer. */
376 if (list_next == &waiting_locks_list) {
377 /* No more, just cancel. */
378 cfs_timer_disarm(&waiting_locks_timer);
380 struct ldlm_lock *next;
381 next = list_entry(list_next, struct ldlm_lock,
383 cfs_timer_arm(&waiting_locks_timer,
384 round_timeout(next->l_callback_timeout));
387 list_del_init(&lock->l_pending_chain);
392 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
396 if (lock->l_export == NULL) {
397 /* We don't have a "waiting locks list" on clients. */
398 LDLM_DEBUG(lock, "client lock: no-op");
402 spin_lock_bh(&waiting_locks_spinlock);
403 ret = __ldlm_del_waiting_lock(lock);
404 spin_unlock_bh(&waiting_locks_spinlock);
406 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
413 * Called with namespace lock held.
415 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
417 if (lock->l_export == NULL) {
418 /* We don't have a "waiting locks list" on clients. */
419 LDLM_DEBUG(lock, "client lock: no-op");
423 spin_lock_bh(&waiting_locks_spinlock);
425 if (list_empty(&lock->l_pending_chain)) {
426 spin_unlock_bh(&waiting_locks_spinlock);
427 LDLM_DEBUG(lock, "wasn't waiting");
431 __ldlm_del_waiting_lock(lock);
432 __ldlm_add_waiting_lock(lock);
433 spin_unlock_bh(&waiting_locks_spinlock);
435 LDLM_DEBUG(lock, "refreshed");
439 #else /* !__KERNEL__ */
441 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
443 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
447 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
452 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
456 #endif /* __KERNEL__ */
458 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
459 const char *ast_type)
461 struct ptlrpc_connection *conn = lock->l_export->exp_connection;
462 char *str = libcfs_nid2str(conn->c_peer.nid);
464 LCONSOLE_ERROR_MSG(0x138, "A client on nid %s was evicted from "
465 "service %s.\n", str,
466 lock->l_export->exp_obd->obd_name);
468 LCONSOLE_ERROR_MSG(0x012, "Lock %s callback to %s timed out for "
469 "resource %d\n", ast_type,
470 obd_export_nid2str(lock->l_export), rc);
472 if (obd_dump_on_timeout)
473 libcfs_debug_dumplog();
474 class_fail_export(lock->l_export);
477 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
478 struct ptlrpc_request *req, int rc,
479 const char *ast_type)
481 lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
483 if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
484 LASSERT(lock->l_export);
485 if (lock->l_export->exp_libclient) {
486 LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
487 " timeout, just cancelling lock", ast_type,
488 libcfs_nid2str(peer.nid));
489 ldlm_lock_cancel(lock);
491 } else if (lock->l_flags & LDLM_FL_CANCEL) {
492 LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
493 "cancel was received (AST reply lost?)",
494 ast_type, libcfs_nid2str(peer.nid));
495 ldlm_lock_cancel(lock);
498 ldlm_del_waiting_lock(lock);
499 ldlm_failed_ast(lock, rc, ast_type);
503 LDLM_DEBUG(lock, "client (nid %s) returned %d"
504 " from %s AST - normal race",
505 libcfs_nid2str(peer.nid),
507 lustre_msg_get_status(req->rq_repmsg) : -1,
510 LDLM_ERROR(lock, "client (nid %s) returned %d "
511 "from %s AST", libcfs_nid2str(peer.nid),
512 (req->rq_repmsg != NULL) ?
513 lustre_msg_get_status(req->rq_repmsg) : 0,
515 ldlm_lock_cancel(lock);
516 /* Server-side AST functions are called from ldlm_reprocess_all,
517 * which needs to be told to please restart its reprocessing. */
524 static int ldlm_cb_interpret(struct ptlrpc_request *req, void *data, int rc)
526 struct ldlm_cb_set_arg *arg;
527 struct ldlm_lock *lock;
530 LASSERT(data != NULL);
532 arg = req->rq_async_args.pointer_arg[0];
533 lock = req->rq_async_args.pointer_arg[1];
534 LASSERT(lock != NULL);
536 /* If client canceled the lock but the cancel has not
537 * been recieved yet, we need to update lvbo to have the
538 * proper attributes cached. */
539 if (rc == -EINVAL && arg->type == LDLM_BL_CALLBACK)
540 ldlm_res_lvbo_update(lock->l_resource, NULL,
542 rc = ldlm_handle_ast_error(lock, req, rc,
543 arg->type == LDLM_BL_CALLBACK
544 ? "blocking" : "completion");
550 atomic_set(&arg->restart, 1);
555 static inline int ldlm_bl_and_cp_ast_fini(struct ptlrpc_request *req,
556 struct ldlm_cb_set_arg *arg,
557 struct ldlm_lock *lock,
563 if (unlikely(instant_cancel)) {
564 rc = ptl_send_rpc(req, 1);
565 ptlrpc_req_finished(req);
567 /* If we cancelled the lock, we need to restart
568 * ldlm_reprocess_queue */
569 atomic_set(&arg->restart, 1);
572 ptlrpc_set_add_req(arg->set, req);
579 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
580 * enqueued server lock conflicts with given one.
582 * Sends blocking ast rpc to the client owning that lock; arms timeout timer
583 * to wait for client response.
585 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
586 struct ldlm_lock_desc *desc,
587 void *data, int flag)
589 struct ldlm_cb_set_arg *arg = data;
590 struct ldlm_request *body;
591 struct ptlrpc_request *req;
592 int instant_cancel = 0;
596 if (flag == LDLM_CB_CANCELING)
597 /* Don't need to do anything here. */
601 LASSERT(data != NULL);
602 if (lock->l_export->exp_obd->obd_recovering != 0) {
603 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
604 ldlm_lock_dump(D_ERROR, lock, 0);
607 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
608 &RQF_LDLM_BL_CALLBACK,
609 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
613 req->rq_async_args.pointer_arg[0] = arg;
614 req->rq_async_args.pointer_arg[1] = lock;
615 req->rq_interpret_reply = ldlm_cb_interpret;
616 req->rq_no_resend = 1;
618 lock_res(lock->l_resource);
619 if (lock->l_granted_mode != lock->l_req_mode) {
620 /* this blocking AST will be communicated as part of the
621 * completion AST instead */
622 unlock_res(lock->l_resource);
623 ptlrpc_req_finished(req);
624 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
628 if (lock->l_destroyed) {
629 /* What's the point? */
630 unlock_res(lock->l_resource);
631 ptlrpc_req_finished(req);
635 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
638 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
639 body->lock_handle[0] = lock->l_remote_handle;
640 body->lock_desc = *desc;
641 body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
643 LDLM_DEBUG(lock, "server preparing blocking AST");
645 ptlrpc_request_set_replen(req);
646 if (instant_cancel) {
647 unlock_res(lock->l_resource);
648 ldlm_lock_cancel(lock);
650 LASSERT(lock->l_granted_mode == lock->l_req_mode);
651 ldlm_add_waiting_lock(lock);
652 unlock_res(lock->l_resource);
655 req->rq_send_state = LUSTRE_IMP_FULL;
656 /* ptlrpc_prep_req already set timeout */
658 req->rq_timeout = ldlm_get_rq_timeout();
660 if (lock->l_export && lock->l_export->exp_ldlm_stats)
661 lprocfs_counter_incr(lock->l_export->exp_ldlm_stats,
662 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
664 rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
669 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
671 struct ldlm_cb_set_arg *arg = data;
672 struct ldlm_request *body;
673 struct ptlrpc_request *req;
674 struct timeval granted_time;
675 long total_enqueue_wait;
676 int instant_cancel = 0;
680 LASSERT(lock != NULL);
681 LASSERT(data != NULL);
683 do_gettimeofday(&granted_time);
684 total_enqueue_wait = cfs_timeval_sub(&granted_time,
685 &lock->l_enqueued_time, NULL);
687 if (total_enqueue_wait / ONE_MILLION > obd_timeout)
688 /* non-fatal with AT - change to LDLM_DEBUG? */
689 LDLM_ERROR(lock, "enqueue wait took %luus from "CFS_TIME_T,
690 total_enqueue_wait, lock->l_enqueued_time.tv_sec);
692 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
693 &RQF_LDLM_CP_CALLBACK);
697 lock_res_and_lock(lock);
698 if (lock->l_resource->lr_lvb_len)
699 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT,
700 lock->l_resource->lr_lvb_len);
701 unlock_res_and_lock(lock);
703 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
705 ptlrpc_request_free(req);
709 req->rq_async_args.pointer_arg[0] = arg;
710 req->rq_async_args.pointer_arg[1] = lock;
711 req->rq_interpret_reply = ldlm_cb_interpret;
712 req->rq_no_resend = 1;
713 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
715 body->lock_handle[0] = lock->l_remote_handle;
716 body->lock_flags = flags;
717 ldlm_lock2desc(lock, &body->lock_desc);
718 if (lock->l_resource->lr_lvb_len) {
719 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
721 lock_res_and_lock(lock);
722 memcpy(lvb, lock->l_resource->lr_lvb_data,
723 lock->l_resource->lr_lvb_len);
724 unlock_res_and_lock(lock);
727 LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
730 /* Server-side enqueue wait time estimate, used in
731 __ldlm_add_waiting_lock to set future enqueue timers */
732 at_add(&lock->l_resource->lr_namespace->ns_at_estimate,
733 total_enqueue_wait / ONE_MILLION);
735 ptlrpc_request_set_replen(req);
737 req->rq_send_state = LUSTRE_IMP_FULL;
738 /* ptlrpc_prep_req already set timeout */
740 req->rq_timeout = ldlm_get_rq_timeout();
742 /* We only send real blocking ASTs after the lock is granted */
743 lock_res_and_lock(lock);
744 if (lock->l_flags & LDLM_FL_AST_SENT) {
745 body->lock_flags |= LDLM_FL_AST_SENT;
747 /* We might get here prior to ldlm_handle_enqueue setting
748 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
749 * into waiting list, but this is safe and similar code in
750 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
751 * that would not only cancel the lock, but will also remove
752 * it from waiting list */
753 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
754 unlock_res_and_lock(lock);
755 ldlm_lock_cancel(lock);
757 lock_res_and_lock(lock);
759 /* start the lock-timeout clock */
760 ldlm_add_waiting_lock(lock);
763 unlock_res_and_lock(lock);
765 if (lock->l_export && lock->l_export->exp_ldlm_stats)
766 lprocfs_counter_incr(lock->l_export->exp_ldlm_stats,
767 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
769 rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
774 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
776 struct ldlm_resource *res = lock->l_resource;
777 struct ldlm_request *body;
778 struct ptlrpc_request *req;
782 LASSERT(lock != NULL);
784 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
785 &RQF_LDLM_GL_CALLBACK,
786 LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK);
791 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
792 body->lock_handle[0] = lock->l_remote_handle;
793 ldlm_lock2desc(lock, &body->lock_desc);
795 lock_res_and_lock(lock);
796 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
797 lock->l_resource->lr_lvb_len);
798 unlock_res_and_lock(lock);
799 res = lock->l_resource;
800 ptlrpc_request_set_replen(req);
803 req->rq_send_state = LUSTRE_IMP_FULL;
804 /* ptlrpc_prep_req already set timeout */
806 req->rq_timeout = ldlm_get_rq_timeout();
808 if (lock->l_export && lock->l_export->exp_ldlm_stats)
809 lprocfs_counter_incr(lock->l_export->exp_ldlm_stats,
810 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
812 rc = ptlrpc_queue_wait(req);
813 if (rc == -ELDLM_NO_LOCK_DATA)
814 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
816 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
818 rc = ldlm_res_lvbo_update(res, req->rq_repmsg,
820 ptlrpc_req_finished(req);
822 ldlm_reprocess_all(res);
827 static struct ldlm_lock *
828 find_existing_lock(struct obd_export *exp,
829 const struct lustre_handle *remote_hdl)
831 struct list_head *iter;
833 spin_lock(&exp->exp_ldlm_data.led_lock);
834 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
835 struct ldlm_lock *lock;
836 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
837 if (lock->l_remote_handle.cookie == remote_hdl->cookie) {
839 spin_unlock(&exp->exp_ldlm_data.led_lock);
843 spin_unlock(&exp->exp_ldlm_data.led_lock);
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->exp_ldlm_stats)
916 lprocfs_counter_incr(req->rq_export->exp_ldlm_stats,
917 LDLM_ENQUEUE - LDLM_FIRST_OPC);
919 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
920 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
921 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
922 dlm_req->lock_desc.l_resource.lr_type);
923 GOTO(out, rc = -EFAULT);
926 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
927 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
928 dlm_req->lock_desc.l_req_mode &
929 (dlm_req->lock_desc.l_req_mode-1))) {
930 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
931 dlm_req->lock_desc.l_req_mode);
932 GOTO(out, rc = -EFAULT);
935 if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
936 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
938 DEBUG_REQ(D_ERROR, req,
939 "PLAIN lock request from IBITS client?");
940 GOTO(out, rc = -EPROTO);
942 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
944 DEBUG_REQ(D_ERROR, req,
945 "IBITS lock request from unaware client?");
946 GOTO(out, rc = -EPROTO);
950 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
951 against server's _CONNECT_SUPPORTED flags? (I don't want to use
952 ibits for mgc/mgs) */
954 /* INODEBITS_INTEROP: Perform conversion from plain lock to
955 * inodebits lock if client does not support them. */
956 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
957 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
958 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
959 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
960 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
961 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
962 dlm_req->lock_desc.l_req_mode = LCK_CR;
966 if (unlikely(flags & LDLM_FL_REPLAY)) {
967 lock = find_existing_lock(req->rq_export,
968 &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);
998 spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
999 list_add(&lock->l_export_chain,
1000 &lock->l_export->exp_ldlm_data.led_held_locks);
1001 spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
1005 if (flags & LDLM_FL_HAS_INTENT) {
1006 /* In this case, the reply buffer is allocated deep in
1007 * local_lock_enqueue by the policy function. */
1010 lock_res_and_lock(lock);
1011 if (lock->l_resource->lr_lvb_len) {
1012 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1014 lock->l_resource->lr_lvb_len);
1016 unlock_res_and_lock(lock);
1018 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1019 GOTO(out, rc = -ENOMEM);
1021 rc = req_capsule_server_pack(&req->rq_pill);
1026 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1027 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1028 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1029 lock->l_req_extent = lock->l_policy_data.l_extent;
1031 err = ldlm_lock_enqueue(ns, &lock, cookie, (int *)&flags);
1035 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1036 dlm_rep->lock_flags = flags;
1038 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1039 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1041 /* We never send a blocking AST until the lock is granted, but
1042 * we can tell it right now */
1043 lock_res_and_lock(lock);
1045 /* Now take into account flags to be inherited from original lock
1046 request both in reply to client and in our own lock flags. */
1047 dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1048 lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1050 /* Don't move a pending lock onto the export if it has already
1051 * been evicted. Cancel it now instead. (bug 5683) */
1052 if (unlikely(req->rq_export->exp_failed ||
1053 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1054 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1056 } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1057 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
1058 if (lock->l_granted_mode == lock->l_req_mode) {
1060 * Only cancel lock if it was granted, because it would
1061 * be destroyed immediatelly and would never be granted
1062 * in the future, causing timeouts on client. Not
1063 * granted lock will be cancelled immediatelly after
1064 * sending completion AST.
1066 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1067 unlock_res_and_lock(lock);
1068 ldlm_lock_cancel(lock);
1069 lock_res_and_lock(lock);
1071 ldlm_add_waiting_lock(lock);
1074 /* Make sure we never ever grant usual metadata locks to liblustre
1076 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1077 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1078 req->rq_export->exp_libclient) {
1079 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1080 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1081 CERROR("Granting sync lock to libclient. "
1082 "req fl %d, rep fl %d, lock fl %d\n",
1083 dlm_req->lock_flags, dlm_rep->lock_flags,
1085 LDLM_ERROR(lock, "sync lock");
1086 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1087 struct ldlm_intent *it;
1089 it = req_capsule_client_get(&req->rq_pill,
1092 CERROR("This is intent %s ("LPU64")\n",
1093 ldlm_it2str(it->opc), it->opc);
1099 unlock_res_and_lock(lock);
1103 req->rq_status = err;
1104 if (!req->rq_packed_final) {
1105 err = lustre_pack_reply(req, 1, NULL, NULL);
1108 req->rq_status = rc;
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_ldlm_stats)
1180 lprocfs_counter_incr(req->rq_export->exp_ldlm_stats,
1181 LDLM_CONVERT - LDLM_FIRST_OPC);
1183 rc = req_capsule_server_pack(&req->rq_pill);
1187 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1188 dlm_rep->lock_flags = dlm_req->lock_flags;
1190 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1192 req->rq_status = EINVAL;
1196 LDLM_DEBUG(lock, "server-side convert handler START");
1198 do_gettimeofday(&lock->l_enqueued_time);
1199 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1200 &dlm_rep->lock_flags);
1202 if (ldlm_del_waiting_lock(lock))
1203 LDLM_DEBUG(lock, "converted waiting lock");
1206 req->rq_status = EDEADLOCK;
1211 if (!req->rq_status)
1212 ldlm_reprocess_all(lock->l_resource);
1213 LDLM_DEBUG(lock, "server-side convert handler END");
1214 LDLM_LOCK_PUT(lock);
1216 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1221 int ldlm_handle_convert(struct ptlrpc_request *req)
1224 struct ldlm_request *dlm_req;
1226 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1227 if (dlm_req != NULL) {
1228 rc = ldlm_handle_convert0(req, dlm_req);
1230 CERROR ("Can't unpack dlm_req\n");
1236 /* Cancel all the locks whos handles are packed into ldlm_request */
1237 int ldlm_request_cancel(struct ptlrpc_request *req,
1238 const struct ldlm_request *dlm_req, int first)
1240 struct ldlm_resource *res, *pres = NULL;
1241 struct ldlm_lock *lock;
1242 int i, count, done = 0;
1245 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1249 /* There is no lock on the server at the replay time,
1250 * skip lock cancelling to make replay tests to pass. */
1251 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1254 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1255 "starting at %d", count, first);
1257 for (i = first; i < count; i++) {
1258 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1260 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1261 "lock (cookie "LPU64")",
1262 dlm_req->lock_handle[i].cookie);
1266 res = lock->l_resource;
1271 ldlm_reprocess_all(pres);
1272 ldlm_resource_putref(pres);
1275 ldlm_resource_getref(res);
1276 ldlm_res_lvbo_update(res, NULL, 0, 1);
1280 ldlm_lock_cancel(lock);
1281 LDLM_LOCK_PUT(lock);
1284 ldlm_reprocess_all(pres);
1285 ldlm_resource_putref(pres);
1287 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1291 int ldlm_handle_cancel(struct ptlrpc_request *req)
1293 struct ldlm_request *dlm_req;
1297 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1298 if (dlm_req == NULL) {
1299 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1303 if (req->rq_export && req->rq_export->exp_ldlm_stats)
1304 lprocfs_counter_incr(req->rq_export->exp_ldlm_stats,
1305 LDLM_CANCEL - LDLM_FIRST_OPC);
1307 rc = req_capsule_server_pack(&req->rq_pill);
1311 if (!ldlm_request_cancel(req, dlm_req, 0))
1312 req->rq_status = ESTALE;
1314 if (ptlrpc_reply(req) != 0)
1320 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1321 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1326 LDLM_DEBUG(lock, "client blocking AST callback handler START");
1328 lock_res_and_lock(lock);
1329 lock->l_flags |= LDLM_FL_CBPENDING;
1331 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1332 lock->l_flags |= LDLM_FL_CANCEL;
1334 do_ast = (!lock->l_readers && !lock->l_writers);
1335 unlock_res_and_lock(lock);
1338 LDLM_DEBUG(lock, "already unused, calling "
1339 "callback (%p)", lock->l_blocking_ast);
1340 if (lock->l_blocking_ast != NULL)
1341 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1344 LDLM_DEBUG(lock, "Lock still has references, will be"
1345 " cancelled later");
1348 LDLM_DEBUG(lock, "client blocking callback handler END");
1349 LDLM_LOCK_PUT(lock);
1353 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1354 struct ldlm_namespace *ns,
1355 struct ldlm_request *dlm_req,
1356 struct ldlm_lock *lock)
1358 CFS_LIST_HEAD(ast_list);
1361 LDLM_DEBUG(lock, "client completion callback handler START");
1363 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1364 int to = cfs_time_seconds(1);
1366 to = schedule_timeout(to);
1367 if (lock->l_granted_mode == lock->l_req_mode ||
1373 lock_res_and_lock(lock);
1374 if (lock->l_destroyed ||
1375 lock->l_granted_mode == lock->l_req_mode) {
1376 /* bug 11300: the lock has already been granted */
1377 unlock_res_and_lock(lock);
1378 LDLM_DEBUG(lock, "Double grant race happened");
1379 LDLM_LOCK_PUT(lock);
1384 /* If we receive the completion AST before the actual enqueue returned,
1385 * then we might need to switch lock modes, resources, or extents. */
1386 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1387 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1388 LDLM_DEBUG(lock, "completion AST, new lock mode");
1391 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1392 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1393 LDLM_DEBUG(lock, "completion AST, new policy data");
1396 ldlm_resource_unlink_lock(lock);
1397 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1398 &lock->l_resource->lr_name,
1399 sizeof(lock->l_resource->lr_name)) != 0) {
1400 unlock_res_and_lock(lock);
1401 if (ldlm_lock_change_resource(ns, lock,
1402 &dlm_req->lock_desc.l_resource.lr_name) != 0) {
1403 LDLM_ERROR(lock, "Failed to allocate resource");
1404 LDLM_LOCK_PUT(lock);
1408 LDLM_DEBUG(lock, "completion AST, new resource");
1409 CERROR("change resource!\n");
1410 lock_res_and_lock(lock);
1413 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1414 /* BL_AST locks are not needed in lru.
1415 * let ldlm_cancel_lru() be fast. */
1416 ldlm_lock_remove_from_lru(lock);
1417 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1418 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1421 if (lock->l_lvb_len) {
1422 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
1423 RCL_CLIENT) < lock->l_lvb_len) {
1424 LDLM_ERROR(lock, "completion AST did not contain "
1427 void *lvb = req_capsule_client_swab_get(&req->rq_pill,
1429 (void *)lock->l_lvb_swabber);
1430 memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1434 ldlm_grant_lock(lock, &ast_list);
1435 unlock_res_and_lock(lock);
1437 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1439 ldlm_run_ast_work(&ast_list, LDLM_WORK_CP_AST);
1441 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1443 LDLM_LOCK_PUT(lock);
1447 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1448 struct ldlm_namespace *ns,
1449 struct ldlm_request *dlm_req,
1450 struct ldlm_lock *lock)
1455 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1457 if (lock->l_glimpse_ast != NULL)
1458 rc = lock->l_glimpse_ast(lock, req);
1460 if (req->rq_repmsg != NULL) {
1463 req->rq_status = rc;
1467 lock_res_and_lock(lock);
1468 if (lock->l_granted_mode == LCK_PW &&
1469 !lock->l_readers && !lock->l_writers &&
1470 cfs_time_after(cfs_time_current(),
1471 cfs_time_add(lock->l_last_used,
1472 cfs_time_seconds(10)))) {
1473 unlock_res_and_lock(lock);
1474 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1475 ldlm_handle_bl_callback(ns, NULL, lock);
1480 unlock_res_and_lock(lock);
1481 LDLM_LOCK_PUT(lock);
1485 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1487 if (req->rq_no_reply)
1490 req->rq_status = rc;
1491 if (!req->rq_packed_final) {
1492 rc = lustre_pack_reply(req, 1, NULL, NULL);
1496 return ptlrpc_reply(req);
1500 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1501 struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
1502 struct list_head *cancels, int count)
1504 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1505 struct ldlm_bl_work_item *blwi;
1508 if (cancels && count == 0)
1511 OBD_ALLOC(blwi, sizeof(*blwi));
1517 blwi->blwi_ld = *ld;
1519 list_add(&blwi->blwi_head, cancels);
1520 list_del_init(cancels);
1521 blwi->blwi_count = count;
1523 blwi->blwi_lock = lock;
1525 spin_lock(&blp->blp_lock);
1526 if (lock && lock->l_flags & LDLM_FL_DISCARD_DATA) {
1527 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1528 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1530 /* other blocking callbacks are added to the regular list */
1531 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1533 cfs_waitq_signal(&blp->blp_waitq);
1534 spin_unlock(&blp->blp_lock);
1540 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1541 struct ldlm_lock *lock)
1544 RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0));
1550 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1551 struct list_head *cancels, int count)
1554 RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count));
1560 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1561 static int ldlm_callback_handler(struct ptlrpc_request *req)
1563 struct ldlm_namespace *ns;
1564 struct ldlm_request *dlm_req;
1565 struct ldlm_lock *lock;
1569 /* Requests arrive in sender's byte order. The ptlrpc service
1570 * handler has already checked and, if necessary, byte-swapped the
1571 * incoming request message body, but I am responsible for the
1572 * message buffers. */
1574 /* do nothing for sec context finalize */
1575 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
1578 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1580 if (req->rq_export == NULL) {
1581 struct ldlm_request *dlm_req;
1583 CDEBUG(D_RPCTRACE, "operation %d from %s with bad "
1584 "export cookie "LPX64"; this is "
1585 "normal if this node rebooted with a lock held\n",
1586 lustre_msg_get_opc(req->rq_reqmsg),
1587 libcfs_id2str(req->rq_peer),
1588 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
1590 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1591 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1592 if (dlm_req != NULL)
1593 CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1594 dlm_req->lock_handle[0].cookie);
1596 ldlm_callback_reply(req, -ENOTCONN);
1600 LASSERT(req->rq_export != NULL);
1601 LASSERT(req->rq_export->exp_obd != NULL);
1603 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1604 case LDLM_BL_CALLBACK:
1605 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK))
1608 case LDLM_CP_CALLBACK:
1609 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK))
1612 case LDLM_GL_CALLBACK:
1613 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK))
1616 case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
1617 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1618 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1620 rc = llog_origin_handle_cancel(req);
1621 ldlm_callback_reply(req, rc);
1623 case OBD_QC_CALLBACK:
1624 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
1625 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
1627 rc = target_handle_qc_callback(req);
1628 ldlm_callback_reply(req, rc);
1632 /* reply in handler */
1633 req_capsule_set(&req->rq_pill, &RQF_MDS_QUOTA_DQACQ);
1634 rc = target_handle_dqacq_callback(req);
1636 case LLOG_ORIGIN_HANDLE_CREATE:
1637 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
1638 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1640 rc = llog_origin_handle_create(req);
1641 ldlm_callback_reply(req, rc);
1643 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1644 req_capsule_set(&req->rq_pill,
1645 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
1646 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1648 rc = llog_origin_handle_next_block(req);
1649 ldlm_callback_reply(req, rc);
1651 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1652 req_capsule_set(&req->rq_pill,
1653 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
1654 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1656 rc = llog_origin_handle_read_header(req);
1657 ldlm_callback_reply(req, rc);
1659 case LLOG_ORIGIN_HANDLE_CLOSE:
1660 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1662 rc = llog_origin_handle_close(req);
1663 ldlm_callback_reply(req, rc);
1666 CERROR("unknown opcode %u\n",
1667 lustre_msg_get_opc(req->rq_reqmsg));
1668 ldlm_callback_reply(req, -EPROTO);
1672 ns = req->rq_export->exp_obd->obd_namespace;
1673 LASSERT(ns != NULL);
1675 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1677 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1678 if (dlm_req == NULL) {
1679 ldlm_callback_reply(req, -EPROTO);
1683 /* Force a known safe race, send a cancel to the server for a lock
1684 * which the server has already started a blocking callback on. */
1685 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
1686 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1687 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]);
1689 CERROR("ldlm_cli_cancel: %d\n", rc);
1692 lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle[0]);
1694 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
1695 "disappeared\n", dlm_req->lock_handle[0].cookie);
1696 ldlm_callback_reply(req, -EINVAL);
1700 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1701 lock_res_and_lock(lock);
1702 lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1703 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1704 /* If somebody cancels locks and cache is already droped,
1705 * we can tell the server we have no lock. Otherwise, we
1706 * should send cancel after dropping the cache. */
1707 if ((lock->l_flags & LDLM_FL_CANCELING) &&
1708 (lock->l_flags & LDLM_FL_BL_DONE)) {
1709 LDLM_DEBUG(lock, "callback on lock "
1710 LPX64" - lock disappeared\n",
1711 dlm_req->lock_handle[0].cookie);
1712 unlock_res_and_lock(lock);
1713 LDLM_LOCK_PUT(lock);
1714 ldlm_callback_reply(req, -EINVAL);
1717 /* BL_AST locks are not needed in lru.
1718 * let ldlm_cancel_lru() be fast. */
1719 ldlm_lock_remove_from_lru(lock);
1720 lock->l_flags |= LDLM_FL_BL_AST;
1722 unlock_res_and_lock(lock);
1724 /* We want the ost thread to get this reply so that it can respond
1725 * to ost requests (write cache writeback) that might be triggered
1728 * But we'd also like to be able to indicate in the reply that we're
1729 * cancelling right now, because it's unused, or have an intent result
1730 * in the reply, so we might have to push the responsibility for sending
1731 * the reply down into the AST handlers, alas. */
1733 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1734 case LDLM_BL_CALLBACK:
1735 CDEBUG(D_INODE, "blocking ast\n");
1736 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
1737 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK))
1738 ldlm_callback_reply(req, 0);
1739 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
1740 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1742 case LDLM_CP_CALLBACK:
1743 CDEBUG(D_INODE, "completion ast\n");
1744 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
1745 ldlm_callback_reply(req, 0);
1746 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1748 case LDLM_GL_CALLBACK:
1749 CDEBUG(D_INODE, "glimpse ast\n");
1750 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
1751 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1754 LBUG(); /* checked above */
1760 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1765 /* Requests arrive in sender's byte order. The ptlrpc service
1766 * handler has already checked and, if necessary, byte-swapped the
1767 * incoming request message body, but I am responsible for the
1768 * message buffers. */
1770 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1772 if (req->rq_export == NULL) {
1773 struct ldlm_request *dlm_req;
1775 CERROR("operation %d from %s with bad export cookie "LPU64"\n",
1776 lustre_msg_get_opc(req->rq_reqmsg),
1777 libcfs_id2str(req->rq_peer),
1778 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
1780 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1781 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1782 if (dlm_req != NULL)
1783 ldlm_lock_dump_handle(D_ERROR,
1784 &dlm_req->lock_handle[0]);
1785 ldlm_callback_reply(req, -ENOTCONN);
1789 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1791 /* XXX FIXME move this back to mds/handler.c, bug 249 */
1793 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
1794 CDEBUG(D_INODE, "cancel\n");
1795 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
1797 rc = ldlm_handle_cancel(req);
1801 case OBD_LOG_CANCEL:
1802 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1803 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1805 rc = llog_origin_handle_cancel(req);
1806 ldlm_callback_reply(req, rc);
1809 CERROR("invalid opcode %d\n",
1810 lustre_msg_get_opc(req->rq_reqmsg));
1811 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1812 ldlm_callback_reply(req, -EINVAL);
1818 void ldlm_revoke_export_locks(struct obd_export *exp)
1820 struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
1821 struct list_head rpc_list;
1822 struct ldlm_lock *lock, *next;
1825 CFS_INIT_LIST_HEAD(&rpc_list);
1827 spin_lock(&exp->exp_ldlm_data.led_lock);
1828 list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
1829 lock_res_and_lock(lock);
1831 if (lock->l_req_mode != lock->l_granted_mode) {
1832 unlock_res_and_lock(lock);
1836 LASSERT(lock->l_resource);
1837 if (lock->l_resource->lr_type != LDLM_IBITS &&
1838 lock->l_resource->lr_type != LDLM_PLAIN) {
1839 unlock_res_and_lock(lock);
1843 if (lock->l_flags & LDLM_FL_AST_SENT) {
1844 unlock_res_and_lock(lock);
1848 LASSERT(lock->l_blocking_ast);
1849 LASSERT(!lock->l_blocking_lock);
1851 lock->l_flags |= LDLM_FL_AST_SENT;
1852 list_move(&lock->l_export_chain, &rpc_list);
1853 LDLM_LOCK_GET(lock);
1855 unlock_res_and_lock(lock);
1857 spin_unlock(&exp->exp_ldlm_data.led_lock);
1859 ldlm_run_ast_work(&rpc_list, LDLM_WORK_REVOKE_AST);
1865 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1867 struct ldlm_bl_work_item *blwi = NULL;
1868 static unsigned int num_bl = 0;
1870 spin_lock(&blp->blp_lock);
1871 /* process a request from the blp_list at least every blp_num_threads */
1872 if (!list_empty(&blp->blp_list) &&
1873 (list_empty(&blp->blp_prio_list) || num_bl == 0))
1874 blwi = list_entry(blp->blp_list.next,
1875 struct ldlm_bl_work_item, blwi_entry);
1877 if (!list_empty(&blp->blp_prio_list))
1878 blwi = list_entry(blp->blp_prio_list.next,
1879 struct ldlm_bl_work_item, blwi_entry);
1882 if (++num_bl >= atomic_read(&blp->blp_num_threads))
1884 list_del(&blwi->blwi_entry);
1886 spin_unlock(&blp->blp_lock);
1891 /* This only contains temporary data until the thread starts */
1892 struct ldlm_bl_thread_data {
1893 char bltd_name[CFS_CURPROC_COMM_MAX];
1894 struct ldlm_bl_pool *bltd_blp;
1895 struct completion bltd_comp;
1899 static int ldlm_bl_thread_main(void *arg);
1901 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
1903 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
1906 init_completion(&bltd.bltd_comp);
1907 rc = cfs_kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1909 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
1910 atomic_read(&blp->blp_num_threads), rc);
1913 wait_for_completion(&bltd.bltd_comp);
1918 static int ldlm_bl_thread_main(void *arg)
1920 struct ldlm_bl_pool *blp;
1924 struct ldlm_bl_thread_data *bltd = arg;
1926 blp = bltd->bltd_blp;
1928 bltd->bltd_num = atomic_inc_return(&blp->blp_num_threads) - 1;
1929 atomic_inc(&blp->blp_busy_threads);
1931 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
1932 "ldlm_bl_%02d", bltd->bltd_num);
1933 cfs_daemonize(bltd->bltd_name);
1935 complete(&bltd->bltd_comp);
1936 /* cannot use bltd after this, it is only on caller's stack */
1940 struct l_wait_info lwi = { 0 };
1941 struct ldlm_bl_work_item *blwi = NULL;
1943 blwi = ldlm_bl_get_work(blp);
1948 atomic_dec(&blp->blp_busy_threads);
1949 l_wait_event_exclusive(blp->blp_waitq,
1950 (blwi = ldlm_bl_get_work(blp)) != NULL,
1952 busy = atomic_inc_return(&blp->blp_busy_threads);
1954 if (blwi->blwi_ns == NULL)
1955 /* added by ldlm_cleanup() */
1958 /* Not fatal if racy and have a few too many threads */
1959 if (unlikely(busy < blp->blp_max_threads &&
1960 busy >= atomic_read(&blp->blp_num_threads)))
1961 /* discard the return value, we tried */
1962 ldlm_bl_thread_start(blp);
1964 if (blwi->blwi_ns == NULL)
1965 /* added by ldlm_cleanup() */
1969 if (blwi->blwi_count) {
1970 /* The special case when we cancel locks in lru
1971 * asynchronously, we pass the list of locks here.
1972 * Thus lock is marked LDLM_FL_CANCELING, and already
1973 * canceled locally. */
1974 ldlm_cli_cancel_list(&blwi->blwi_head,
1975 blwi->blwi_count, NULL, 0);
1977 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1980 OBD_FREE(blwi, sizeof(*blwi));
1983 atomic_dec(&blp->blp_busy_threads);
1984 atomic_dec(&blp->blp_num_threads);
1985 complete(&blp->blp_comp);
1991 static int ldlm_setup(void);
1992 static int ldlm_cleanup(void);
1994 int ldlm_get_ref(void)
1998 mutex_down(&ldlm_ref_sem);
1999 if (++ldlm_refcount == 1) {
2004 mutex_up(&ldlm_ref_sem);
2009 void ldlm_put_ref(void)
2012 mutex_down(&ldlm_ref_sem);
2013 if (ldlm_refcount == 1) {
2014 int rc = ldlm_cleanup();
2016 CERROR("ldlm_cleanup failed: %d\n", rc);
2022 mutex_up(&ldlm_ref_sem);
2027 static int ldlm_setup(void)
2029 struct ldlm_bl_pool *blp;
2031 int ldlm_min_threads = LDLM_THREADS_AUTO_MIN;
2032 int ldlm_max_threads = LDLM_THREADS_AUTO_MAX;
2038 if (ldlm_state != NULL)
2041 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2042 if (ldlm_state == NULL)
2046 rc = ldlm_proc_setup();
2052 if (ldlm_num_threads) {
2053 /* If ldlm_num_threads is set, it is the min and the max. */
2054 if (ldlm_num_threads > LDLM_THREADS_AUTO_MAX)
2055 ldlm_num_threads = LDLM_THREADS_AUTO_MAX;
2056 if (ldlm_num_threads < LDLM_THREADS_AUTO_MIN)
2057 ldlm_num_threads = LDLM_THREADS_AUTO_MIN;
2058 ldlm_min_threads = ldlm_max_threads = ldlm_num_threads;
2062 ldlm_state->ldlm_cb_service =
2063 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2064 LDLM_MAXREPSIZE, LDLM_CB_REQUEST_PORTAL,
2065 LDLM_CB_REPLY_PORTAL, 1800,
2066 ldlm_callback_handler, "ldlm_cbd",
2067 ldlm_svc_proc_dir, NULL,
2068 ldlm_min_threads, ldlm_max_threads,
2070 LCT_MD_THREAD|LCT_DT_THREAD);
2072 if (!ldlm_state->ldlm_cb_service) {
2073 CERROR("failed to start service\n");
2074 GOTO(out_proc, rc = -ENOMEM);
2077 ldlm_state->ldlm_cancel_service =
2078 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2079 LDLM_MAXREPSIZE, LDLM_CANCEL_REQUEST_PORTAL,
2080 LDLM_CANCEL_REPLY_PORTAL, 6000,
2081 ldlm_cancel_handler, "ldlm_canceld",
2082 ldlm_svc_proc_dir, NULL,
2083 ldlm_min_threads, ldlm_max_threads,
2085 LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD);
2087 if (!ldlm_state->ldlm_cancel_service) {
2088 CERROR("failed to start service\n");
2089 GOTO(out_proc, rc = -ENOMEM);
2092 OBD_ALLOC(blp, sizeof(*blp));
2094 GOTO(out_proc, rc = -ENOMEM);
2095 ldlm_state->ldlm_bl_pool = blp;
2097 spin_lock_init(&blp->blp_lock);
2098 CFS_INIT_LIST_HEAD(&blp->blp_list);
2099 CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2100 cfs_waitq_init(&blp->blp_waitq);
2101 atomic_set(&blp->blp_num_threads, 0);
2102 atomic_set(&blp->blp_busy_threads, 0);
2103 blp->blp_min_threads = ldlm_min_threads;
2104 blp->blp_max_threads = ldlm_max_threads;
2107 for (i = 0; i < blp->blp_min_threads; i++) {
2108 rc = ldlm_bl_thread_start(blp);
2110 GOTO(out_thread, rc);
2113 rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cancel_service);
2115 GOTO(out_thread, rc);
2117 rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cb_service);
2119 GOTO(out_thread, rc);
2121 CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2122 expired_lock_thread.elt_state = ELT_STOPPED;
2123 cfs_waitq_init(&expired_lock_thread.elt_waitq);
2125 CFS_INIT_LIST_HEAD(&waiting_locks_list);
2126 spin_lock_init(&waiting_locks_spinlock);
2127 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2129 rc = cfs_kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FILES);
2131 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2132 GOTO(out_thread, rc);
2135 wait_event(expired_lock_thread.elt_waitq,
2136 expired_lock_thread.elt_state == ELT_READY);
2140 rc = ldlm_pools_init();
2142 GOTO(out_thread, rc);
2148 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2149 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2154 ldlm_proc_cleanup();
2157 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2162 static int ldlm_cleanup(void)
2165 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2169 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2170 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2171 CERROR("ldlm still has namespaces; clean these up first.\n");
2172 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2173 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2182 while (atomic_read(&blp->blp_num_threads) > 0) {
2183 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2185 init_completion(&blp->blp_comp);
2187 spin_lock(&blp->blp_lock);
2188 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2189 cfs_waitq_signal(&blp->blp_waitq);
2190 spin_unlock(&blp->blp_lock);
2192 wait_for_completion(&blp->blp_comp);
2194 OBD_FREE(blp, sizeof(*blp));
2196 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2197 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2198 ldlm_proc_cleanup();
2200 expired_lock_thread.elt_state = ELT_TERMINATE;
2201 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
2202 wait_event(expired_lock_thread.elt_waitq,
2203 expired_lock_thread.elt_state == ELT_STOPPED);
2205 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2206 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2209 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2215 int __init ldlm_init(void)
2217 init_mutex(&ldlm_ref_sem);
2218 init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
2219 init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
2220 ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
2221 sizeof(struct ldlm_resource), 0,
2222 SLAB_HWCACHE_ALIGN);
2223 if (ldlm_resource_slab == NULL)
2226 ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
2227 sizeof(struct ldlm_lock), 0,
2228 SLAB_HWCACHE_ALIGN);
2229 if (ldlm_lock_slab == NULL) {
2230 cfs_mem_cache_destroy(ldlm_resource_slab);
2234 ldlm_interval_slab = cfs_mem_cache_create("interval_node",
2235 sizeof(struct ldlm_interval),
2236 0, SLAB_HWCACHE_ALIGN);
2237 if (ldlm_interval_slab == NULL) {
2238 cfs_mem_cache_destroy(ldlm_resource_slab);
2239 cfs_mem_cache_destroy(ldlm_lock_slab);
2246 void __exit ldlm_exit(void)
2250 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
2251 rc = cfs_mem_cache_destroy(ldlm_resource_slab);
2252 LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
2253 rc = cfs_mem_cache_destroy(ldlm_lock_slab);
2254 LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
2255 rc = cfs_mem_cache_destroy(ldlm_interval_slab);
2256 LASSERTF(rc == 0, "couldn't free interval node slab\n");
2260 EXPORT_SYMBOL(ldlm_extent_shift_kms);
2263 EXPORT_SYMBOL(ldlm_get_processing_policy);
2264 EXPORT_SYMBOL(ldlm_lock2desc);
2265 EXPORT_SYMBOL(ldlm_register_intent);
2266 EXPORT_SYMBOL(ldlm_lockname);
2267 EXPORT_SYMBOL(ldlm_typename);
2268 EXPORT_SYMBOL(ldlm_lock2handle);
2269 EXPORT_SYMBOL(__ldlm_handle2lock);
2270 EXPORT_SYMBOL(ldlm_lock_get);
2271 EXPORT_SYMBOL(ldlm_lock_put);
2272 EXPORT_SYMBOL(ldlm_lock_fast_match);
2273 EXPORT_SYMBOL(ldlm_lock_fast_release);
2274 EXPORT_SYMBOL(ldlm_lock_match);
2275 EXPORT_SYMBOL(ldlm_lock_cancel);
2276 EXPORT_SYMBOL(ldlm_lock_addref);
2277 EXPORT_SYMBOL(ldlm_lock_decref);
2278 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
2279 EXPORT_SYMBOL(ldlm_lock_change_resource);
2280 EXPORT_SYMBOL(ldlm_lock_set_data);
2281 EXPORT_SYMBOL(ldlm_it2str);
2282 EXPORT_SYMBOL(ldlm_lock_dump);
2283 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2284 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
2285 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
2286 EXPORT_SYMBOL(ldlm_lock_allow_match);
2288 /* ldlm_request.c */
2289 EXPORT_SYMBOL(ldlm_completion_ast);
2290 EXPORT_SYMBOL(ldlm_blocking_ast);
2291 EXPORT_SYMBOL(ldlm_glimpse_ast);
2292 EXPORT_SYMBOL(ldlm_expired_completion_wait);
2293 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
2294 EXPORT_SYMBOL(ldlm_prep_elc_req);
2295 EXPORT_SYMBOL(ldlm_cli_convert);
2296 EXPORT_SYMBOL(ldlm_cli_enqueue);
2297 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
2298 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
2299 EXPORT_SYMBOL(ldlm_cli_cancel);
2300 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2301 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2302 EXPORT_SYMBOL(ldlm_cli_cancel_req);
2303 EXPORT_SYMBOL(ldlm_cli_join_lru);
2304 EXPORT_SYMBOL(ldlm_replay_locks);
2305 EXPORT_SYMBOL(ldlm_resource_foreach);
2306 EXPORT_SYMBOL(ldlm_namespace_foreach);
2307 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
2308 EXPORT_SYMBOL(ldlm_resource_iterate);
2309 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2310 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2313 EXPORT_SYMBOL(ldlm_server_blocking_ast);
2314 EXPORT_SYMBOL(ldlm_server_completion_ast);
2315 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
2316 EXPORT_SYMBOL(ldlm_handle_enqueue);
2317 EXPORT_SYMBOL(ldlm_handle_enqueue0);
2318 EXPORT_SYMBOL(ldlm_handle_cancel);
2319 EXPORT_SYMBOL(ldlm_request_cancel);
2320 EXPORT_SYMBOL(ldlm_handle_convert);
2321 EXPORT_SYMBOL(ldlm_handle_convert0);
2322 EXPORT_SYMBOL(ldlm_del_waiting_lock);
2323 EXPORT_SYMBOL(ldlm_get_ref);
2324 EXPORT_SYMBOL(ldlm_put_ref);
2325 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
2326 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2328 /* ldlm_resource.c */
2329 EXPORT_SYMBOL(ldlm_namespace_new);
2330 EXPORT_SYMBOL(ldlm_namespace_cleanup);
2331 EXPORT_SYMBOL(ldlm_namespace_free);
2332 EXPORT_SYMBOL(ldlm_namespace_dump);
2333 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
2334 EXPORT_SYMBOL(ldlm_resource_get);
2335 EXPORT_SYMBOL(ldlm_resource_putref);
2336 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
2339 EXPORT_SYMBOL(client_import_add_conn);
2340 EXPORT_SYMBOL(client_import_del_conn);
2341 EXPORT_SYMBOL(client_obd_setup);
2342 EXPORT_SYMBOL(client_obd_cleanup);
2343 EXPORT_SYMBOL(client_connect_import);
2344 EXPORT_SYMBOL(client_disconnect_export);
2345 EXPORT_SYMBOL(target_start_recovery_thread);
2346 EXPORT_SYMBOL(target_stop_recovery_thread);
2347 EXPORT_SYMBOL(target_handle_connect);
2348 EXPORT_SYMBOL(target_cleanup_recovery);
2349 EXPORT_SYMBOL(target_destroy_export);
2350 EXPORT_SYMBOL(target_cancel_recovery_timer);
2351 EXPORT_SYMBOL(target_send_reply);
2352 EXPORT_SYMBOL(target_queue_recovery_request);
2353 EXPORT_SYMBOL(target_handle_ping);
2354 EXPORT_SYMBOL(target_pack_pool_reply);
2355 EXPORT_SYMBOL(target_handle_disconnect);
2358 EXPORT_SYMBOL(lock_res_and_lock);
2359 EXPORT_SYMBOL(unlock_res_and_lock);