4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_lockd.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LDLM
45 # include <libcfs/libcfs.h>
47 # include <liblustre.h>
50 #include <lustre_dlm.h>
51 #include <obd_class.h>
52 #include <libcfs/list.h>
53 #include "ldlm_internal.h"
55 static int ldlm_num_threads;
56 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
57 "number of DLM service threads to start");
59 static char *ldlm_cpts;
60 CFS_MODULE_PARM(ldlm_cpts, "s", charp, 0444,
61 "CPU partitions ldlm threads should run on");
63 extern cfs_mem_cache_t *ldlm_resource_slab;
64 extern cfs_mem_cache_t *ldlm_lock_slab;
65 static cfs_mutex_t ldlm_ref_mutex;
66 static int ldlm_refcount;
68 struct ldlm_cb_async_args {
69 struct ldlm_cb_set_arg *ca_set_arg;
70 struct ldlm_lock *ca_lock;
75 static struct ldlm_state *ldlm_state;
77 inline cfs_time_t round_timeout(cfs_time_t timeout)
79 return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
82 /* timeout for initial callback (AST) reply (bz10399) */
83 static inline unsigned int ldlm_get_rq_timeout(void)
86 unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
88 return timeout < 1 ? 1 : timeout;
93 #define ELT_TERMINATE 2
96 cfs_spinlock_t blp_lock;
99 * blp_prio_list is used for callbacks that should be handled
100 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
103 cfs_list_t blp_prio_list;
106 * blp_list is used for all other callbacks which are likely
107 * to take longer to process.
111 cfs_waitq_t blp_waitq;
112 cfs_completion_t blp_comp;
113 cfs_atomic_t blp_num_threads;
114 cfs_atomic_t blp_busy_threads;
119 struct ldlm_bl_work_item {
120 cfs_list_t blwi_entry;
121 struct ldlm_namespace *blwi_ns;
122 struct ldlm_lock_desc blwi_ld;
123 struct ldlm_lock *blwi_lock;
124 cfs_list_t blwi_head;
126 cfs_completion_t blwi_comp;
128 int blwi_mem_pressure;
131 #if defined(HAVE_SERVER_SUPPORT) && defined(__KERNEL__)
133 /* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */
134 static cfs_spinlock_t waiting_locks_spinlock; /* BH lock (timer) */
135 static cfs_list_t waiting_locks_list;
136 static cfs_timer_t waiting_locks_timer;
138 static struct expired_lock_thread {
139 cfs_waitq_t elt_waitq;
142 cfs_list_t elt_expired_locks;
143 } expired_lock_thread;
145 static inline int have_expired_locks(void)
150 cfs_spin_lock_bh(&waiting_locks_spinlock);
151 need_to_run = !cfs_list_empty(&expired_lock_thread.elt_expired_locks);
152 cfs_spin_unlock_bh(&waiting_locks_spinlock);
157 static int expired_lock_main(void *arg)
159 cfs_list_t *expired = &expired_lock_thread.elt_expired_locks;
160 struct l_wait_info lwi = { 0 };
164 cfs_daemonize("ldlm_elt");
166 expired_lock_thread.elt_state = ELT_READY;
167 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
170 l_wait_event(expired_lock_thread.elt_waitq,
171 have_expired_locks() ||
172 expired_lock_thread.elt_state == ELT_TERMINATE,
175 cfs_spin_lock_bh(&waiting_locks_spinlock);
176 if (expired_lock_thread.elt_dump) {
177 struct libcfs_debug_msg_data msgdata = {
178 .msg_file = __FILE__,
179 .msg_fn = "waiting_locks_callback",
180 .msg_line = expired_lock_thread.elt_dump };
181 cfs_spin_unlock_bh(&waiting_locks_spinlock);
183 /* from waiting_locks_callback, but not in timer */
184 libcfs_debug_dumplog();
185 libcfs_run_lbug_upcall(&msgdata);
187 cfs_spin_lock_bh(&waiting_locks_spinlock);
188 expired_lock_thread.elt_dump = 0;
193 while (!cfs_list_empty(expired)) {
194 struct obd_export *export;
195 struct ldlm_lock *lock;
197 lock = cfs_list_entry(expired->next, struct ldlm_lock,
199 if ((void *)lock < LP_POISON + CFS_PAGE_SIZE &&
200 (void *)lock >= LP_POISON) {
201 cfs_spin_unlock_bh(&waiting_locks_spinlock);
202 CERROR("free lock on elt list %p\n", lock);
205 cfs_list_del_init(&lock->l_pending_chain);
206 if ((void *)lock->l_export < LP_POISON + CFS_PAGE_SIZE &&
207 (void *)lock->l_export >= LP_POISON) {
208 CERROR("lock with free export on elt list %p\n",
210 lock->l_export = NULL;
211 LDLM_ERROR(lock, "free export");
212 /* release extra ref grabbed by
213 * ldlm_add_waiting_lock() or
214 * ldlm_failed_ast() */
215 LDLM_LOCK_RELEASE(lock);
219 if (lock->l_destroyed) {
220 /* release the lock refcount where
221 * waiting_locks_callback() founds */
222 LDLM_LOCK_RELEASE(lock);
225 export = class_export_lock_get(lock->l_export, lock);
226 cfs_spin_unlock_bh(&waiting_locks_spinlock);
229 class_fail_export(export);
230 class_export_lock_put(export, lock);
232 /* release extra ref grabbed by ldlm_add_waiting_lock()
233 * or ldlm_failed_ast() */
234 LDLM_LOCK_RELEASE(lock);
236 cfs_spin_lock_bh(&waiting_locks_spinlock);
238 cfs_spin_unlock_bh(&waiting_locks_spinlock);
240 if (do_dump && obd_dump_on_eviction) {
241 CERROR("dump the log upon eviction\n");
242 libcfs_debug_dumplog();
245 if (expired_lock_thread.elt_state == ELT_TERMINATE)
249 expired_lock_thread.elt_state = ELT_STOPPED;
250 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
254 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
255 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds);
258 * Check if there is a request in the export request list
259 * which prevents the lock canceling.
261 static int ldlm_lock_busy(struct ldlm_lock *lock)
263 struct ptlrpc_request *req;
267 if (lock->l_export == NULL)
270 cfs_spin_lock_bh(&lock->l_export->exp_rpc_lock);
271 cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
273 if (req->rq_ops->hpreq_lock_match) {
274 match = req->rq_ops->hpreq_lock_match(req, lock);
279 cfs_spin_unlock_bh(&lock->l_export->exp_rpc_lock);
283 /* This is called from within a timer interrupt and cannot schedule */
284 static void waiting_locks_callback(unsigned long unused)
286 struct ldlm_lock *lock;
289 cfs_spin_lock_bh(&waiting_locks_spinlock);
290 while (!cfs_list_empty(&waiting_locks_list)) {
291 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
293 if (cfs_time_after(lock->l_callback_timeout,
294 cfs_time_current()) ||
295 (lock->l_req_mode == LCK_GROUP))
298 if (ptlrpc_check_suspend()) {
299 /* there is a case when we talk to one mds, holding
300 * lock from another mds. this way we easily can get
301 * here, if second mds is being recovered. so, we
302 * suspend timeouts. bug 6019 */
304 LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
305 lock->l_export->exp_client_uuid.uuid,
306 lock->l_export->exp_connection->c_remote_uuid.uuid,
307 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
309 cfs_list_del_init(&lock->l_pending_chain);
310 if (lock->l_destroyed) {
311 /* relay the lock refcount decrease to
312 * expired lock thread */
313 cfs_list_add(&lock->l_pending_chain,
314 &expired_lock_thread.elt_expired_locks);
316 __ldlm_add_waiting_lock(lock,
317 ldlm_get_enq_timeout(lock));
322 /* if timeout overlaps the activation time of suspended timeouts
323 * then extend it to give a chance for client to reconnect */
324 if (cfs_time_before(cfs_time_sub(lock->l_callback_timeout,
325 cfs_time_seconds(obd_timeout)/2),
326 ptlrpc_suspend_wakeup_time())) {
327 LDLM_ERROR(lock, "extend timeout due to recovery: %s@%s nid %s ",
328 lock->l_export->exp_client_uuid.uuid,
329 lock->l_export->exp_connection->c_remote_uuid.uuid,
330 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
332 cfs_list_del_init(&lock->l_pending_chain);
333 if (lock->l_destroyed) {
334 /* relay the lock refcount decrease to
335 * expired lock thread */
336 cfs_list_add(&lock->l_pending_chain,
337 &expired_lock_thread.elt_expired_locks);
339 __ldlm_add_waiting_lock(lock,
340 ldlm_get_enq_timeout(lock));
345 /* Check if we need to prolong timeout */
346 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
347 ldlm_lock_busy(lock)) {
350 if (lock->l_pending_chain.next == &waiting_locks_list)
355 cfs_spin_unlock_bh(&waiting_locks_spinlock);
356 LDLM_DEBUG(lock, "prolong the busy lock");
357 ldlm_refresh_waiting_lock(lock,
358 ldlm_get_enq_timeout(lock));
359 cfs_spin_lock_bh(&waiting_locks_spinlock);
362 LDLM_LOCK_RELEASE(lock);
366 LDLM_LOCK_RELEASE(lock);
369 ldlm_lock_to_ns(lock)->ns_timeouts++;
370 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
371 "evicting client at %s ",
372 cfs_time_current_sec()- lock->l_last_activity,
374 lock->l_export->exp_connection->c_peer.nid));
376 /* no needs to take an extra ref on the lock since it was in
377 * the waiting_locks_list and ldlm_add_waiting_lock()
378 * already grabbed a ref */
379 cfs_list_del(&lock->l_pending_chain);
380 cfs_list_add(&lock->l_pending_chain,
381 &expired_lock_thread.elt_expired_locks);
385 if (!cfs_list_empty(&expired_lock_thread.elt_expired_locks)) {
386 if (obd_dump_on_timeout && need_dump)
387 expired_lock_thread.elt_dump = __LINE__;
389 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
393 * Make sure the timer will fire again if we have any locks
396 if (!cfs_list_empty(&waiting_locks_list)) {
397 cfs_time_t timeout_rounded;
398 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
400 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
401 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
403 cfs_spin_unlock_bh(&waiting_locks_spinlock);
407 * Indicate that we're waiting for a client to call us back cancelling a given
408 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
409 * timer to fire appropriately. (We round up to the next second, to avoid
410 * floods of timer firings during periods of high lock contention and traffic).
411 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
412 * if it has been added to the waiting list (1 is returned).
414 * Called with the namespace lock held.
416 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
419 cfs_time_t timeout_rounded;
421 if (!cfs_list_empty(&lock->l_pending_chain))
424 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
425 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
428 timeout = cfs_time_shift(seconds);
429 if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
430 lock->l_callback_timeout = timeout;
432 timeout_rounded = round_timeout(lock->l_callback_timeout);
434 if (cfs_time_before(timeout_rounded,
435 cfs_timer_deadline(&waiting_locks_timer)) ||
436 !cfs_timer_is_armed(&waiting_locks_timer)) {
437 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
439 /* if the new lock has a shorter timeout than something earlier on
440 the list, we'll wait the longer amount of time; no big deal. */
442 cfs_list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
446 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
449 int timeout = ldlm_get_enq_timeout(lock);
451 /* NB: must be called with hold of lock_res_and_lock() */
452 LASSERT(lock->l_res_locked);
455 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
457 cfs_spin_lock_bh(&waiting_locks_spinlock);
458 if (lock->l_destroyed) {
459 static cfs_time_t next;
460 cfs_spin_unlock_bh(&waiting_locks_spinlock);
461 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
462 if (cfs_time_after(cfs_time_current(), next)) {
463 next = cfs_time_shift(14400);
464 libcfs_debug_dumpstack(NULL);
469 ret = __ldlm_add_waiting_lock(lock, timeout);
471 /* grab ref on the lock if it has been added to the
475 cfs_spin_unlock_bh(&waiting_locks_spinlock);
478 cfs_spin_lock_bh(&lock->l_export->exp_bl_list_lock);
479 if (cfs_list_empty(&lock->l_exp_list))
480 cfs_list_add(&lock->l_exp_list,
481 &lock->l_export->exp_bl_list);
482 cfs_spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
485 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
486 ret == 0 ? "not re-" : "", timeout,
487 AT_OFF ? "off" : "on");
492 * Remove a lock from the pending list, likely because it had its cancellation
493 * callback arrive without incident. This adjusts the lock-timeout timer if
494 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
495 * As done by ldlm_del_waiting_lock(), the caller must release the lock
496 * reference when the lock is removed from any list (1 is returned).
498 * Called with namespace lock held.
500 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
502 cfs_list_t *list_next;
504 if (cfs_list_empty(&lock->l_pending_chain))
507 list_next = lock->l_pending_chain.next;
508 if (lock->l_pending_chain.prev == &waiting_locks_list) {
509 /* Removing the head of the list, adjust timer. */
510 if (list_next == &waiting_locks_list) {
511 /* No more, just cancel. */
512 cfs_timer_disarm(&waiting_locks_timer);
514 struct ldlm_lock *next;
515 next = cfs_list_entry(list_next, struct ldlm_lock,
517 cfs_timer_arm(&waiting_locks_timer,
518 round_timeout(next->l_callback_timeout));
521 cfs_list_del_init(&lock->l_pending_chain);
526 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
530 if (lock->l_export == NULL) {
531 /* We don't have a "waiting locks list" on clients. */
532 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
536 cfs_spin_lock_bh(&waiting_locks_spinlock);
537 ret = __ldlm_del_waiting_lock(lock);
538 cfs_spin_unlock_bh(&waiting_locks_spinlock);
540 /* remove the lock out of export blocking list */
541 cfs_spin_lock_bh(&lock->l_export->exp_bl_list_lock);
542 cfs_list_del_init(&lock->l_exp_list);
543 cfs_spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
546 /* release lock ref if it has indeed been removed
548 LDLM_LOCK_RELEASE(lock);
551 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
558 * Called with namespace lock held.
560 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
562 if (lock->l_export == NULL) {
563 /* We don't have a "waiting locks list" on clients. */
564 LDLM_DEBUG(lock, "client lock: no-op");
568 cfs_spin_lock_bh(&waiting_locks_spinlock);
570 if (cfs_list_empty(&lock->l_pending_chain)) {
571 cfs_spin_unlock_bh(&waiting_locks_spinlock);
572 LDLM_DEBUG(lock, "wasn't waiting");
576 /* we remove/add the lock to the waiting list, so no needs to
577 * release/take a lock reference */
578 __ldlm_del_waiting_lock(lock);
579 __ldlm_add_waiting_lock(lock, timeout);
580 cfs_spin_unlock_bh(&waiting_locks_spinlock);
582 LDLM_DEBUG(lock, "refreshed");
586 #else /* !HAVE_SERVER_SUPPORT || !__KERNEL__ */
588 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
593 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
598 # ifdef HAVE_SERVER_SUPPORT
599 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
601 LASSERT(lock->l_res_locked);
602 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
607 #endif /* HAVE_SERVER_SUPPORT && __KERNEL__ */
609 #ifdef HAVE_SERVER_SUPPORT
611 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
612 const char *ast_type)
614 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
615 "to a lock %s callback time out: rc %d\n",
616 lock->l_export->exp_obd->obd_name,
617 obd_export_nid2str(lock->l_export), ast_type, rc);
619 if (obd_dump_on_timeout)
620 libcfs_debug_dumplog();
622 cfs_spin_lock_bh(&waiting_locks_spinlock);
623 if (__ldlm_del_waiting_lock(lock) == 0)
624 /* the lock was not in any list, grab an extra ref before adding
625 * the lock to the expired list */
627 cfs_list_add(&lock->l_pending_chain,
628 &expired_lock_thread.elt_expired_locks);
629 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
630 cfs_spin_unlock_bh(&waiting_locks_spinlock);
632 class_fail_export(lock->l_export);
636 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
637 struct ptlrpc_request *req, int rc,
638 const char *ast_type)
640 lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
642 if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
643 LASSERT(lock->l_export);
644 if (lock->l_export->exp_libclient) {
645 LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
646 " timeout, just cancelling lock", ast_type,
647 libcfs_nid2str(peer.nid));
648 ldlm_lock_cancel(lock);
650 } else if (lock->l_flags & LDLM_FL_CANCEL) {
651 LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
652 "cancel was received (AST reply lost?)",
653 ast_type, libcfs_nid2str(peer.nid));
654 ldlm_lock_cancel(lock);
657 ldlm_del_waiting_lock(lock);
658 ldlm_failed_ast(lock, rc, ast_type);
662 struct ldlm_resource *res = lock->l_resource;
663 LDLM_DEBUG(lock, "client (nid %s) returned %d"
664 " from %s AST - normal race",
665 libcfs_nid2str(peer.nid),
667 lustre_msg_get_status(req->rq_repmsg) : -1,
670 /* update lvbo to return proper attributes.
672 ldlm_resource_getref(res);
673 ldlm_res_lvbo_update(res, NULL, 1);
674 ldlm_resource_putref(res);
678 LDLM_ERROR(lock, "client (nid %s) returned %d "
679 "from %s AST", libcfs_nid2str(peer.nid),
680 (req->rq_repmsg != NULL) ?
681 lustre_msg_get_status(req->rq_repmsg) : 0,
684 ldlm_lock_cancel(lock);
685 /* Server-side AST functions are called from ldlm_reprocess_all,
686 * which needs to be told to please restart its reprocessing. */
693 static int ldlm_cb_interpret(const struct lu_env *env,
694 struct ptlrpc_request *req, void *data, int rc)
696 struct ldlm_cb_async_args *ca = data;
697 struct ldlm_lock *lock = ca->ca_lock;
698 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
701 LASSERT(lock != NULL);
704 case LDLM_GL_CALLBACK:
705 /* Update the LVB from disk if the AST failed
706 * (this is a legal race)
708 * - Glimpse callback of local lock just returns
709 * -ELDLM_NO_LOCK_DATA.
710 * - Glimpse callback of remote lock might return
711 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
713 if (rc == -ELDLM_NO_LOCK_DATA) {
714 LDLM_DEBUG(lock, "lost race - client has a lock but no "
716 ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
717 } else if (rc != 0) {
718 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
720 rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
723 case LDLM_BL_CALLBACK:
725 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
727 case LDLM_CP_CALLBACK:
729 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
732 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
737 /* release extra reference taken in ldlm_ast_fini() */
738 LDLM_LOCK_RELEASE(lock);
741 cfs_atomic_inc(&arg->restart);
746 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
747 struct ldlm_cb_set_arg *arg,
748 struct ldlm_lock *lock,
754 if (unlikely(instant_cancel)) {
755 rc = ptl_send_rpc(req, 1);
756 ptlrpc_req_finished(req);
758 cfs_atomic_inc(&arg->restart);
761 ptlrpc_set_add_req(arg->set, req);
768 * Check if there are requests in the export request list which prevent
769 * the lock canceling and make these requests high priority ones.
771 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
773 struct ptlrpc_request *req;
776 if (lock->l_export == NULL) {
777 LDLM_DEBUG(lock, "client lock: no-op");
781 cfs_spin_lock_bh(&lock->l_export->exp_rpc_lock);
782 cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
784 /* Do not process requests that were not yet added to there
785 * incoming queue or were already removed from there for
787 if (!req->rq_hp && !cfs_list_empty(&req->rq_list) &&
788 req->rq_ops->hpreq_lock_match &&
789 req->rq_ops->hpreq_lock_match(req, lock))
790 ptlrpc_hpreq_reorder(req);
792 cfs_spin_unlock_bh(&lock->l_export->exp_rpc_lock);
797 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
798 * enqueued server lock conflicts with given one.
800 * Sends blocking ast rpc to the client owning that lock; arms timeout timer
801 * to wait for client response.
803 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
804 struct ldlm_lock_desc *desc,
805 void *data, int flag)
807 struct ldlm_cb_async_args *ca;
808 struct ldlm_cb_set_arg *arg = data;
809 struct ldlm_request *body;
810 struct ptlrpc_request *req;
811 int instant_cancel = 0;
815 if (flag == LDLM_CB_CANCELING)
816 /* Don't need to do anything here. */
820 LASSERT(data != NULL);
821 if (lock->l_export->exp_obd->obd_recovering != 0)
822 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
824 ldlm_lock_reorder_req(lock);
826 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
827 &RQF_LDLM_BL_CALLBACK,
828 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
832 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
833 ca = ptlrpc_req_async_args(req);
834 ca->ca_set_arg = arg;
837 req->rq_interpret_reply = ldlm_cb_interpret;
838 req->rq_no_resend = 1;
840 lock_res_and_lock(lock);
841 if (lock->l_granted_mode != lock->l_req_mode) {
842 /* this blocking AST will be communicated as part of the
843 * completion AST instead */
844 unlock_res_and_lock(lock);
846 ptlrpc_req_finished(req);
847 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
851 if (lock->l_destroyed) {
852 /* What's the point? */
853 unlock_res_and_lock(lock);
854 ptlrpc_req_finished(req);
858 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
861 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
862 body->lock_handle[0] = lock->l_remote_handle;
863 body->lock_desc = *desc;
864 body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
866 LDLM_DEBUG(lock, "server preparing blocking AST");
868 ptlrpc_request_set_replen(req);
869 if (instant_cancel) {
870 unlock_res_and_lock(lock);
871 ldlm_lock_cancel(lock);
873 LASSERT(lock->l_granted_mode == lock->l_req_mode);
874 ldlm_add_waiting_lock(lock);
875 unlock_res_and_lock(lock);
878 req->rq_send_state = LUSTRE_IMP_FULL;
879 /* ptlrpc_request_alloc_pack already set timeout */
881 req->rq_timeout = ldlm_get_rq_timeout();
883 if (lock->l_export && lock->l_export->exp_nid_stats &&
884 lock->l_export->exp_nid_stats->nid_ldlm_stats)
885 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
886 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
888 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
893 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
895 struct ldlm_cb_set_arg *arg = data;
896 struct ldlm_request *body;
897 struct ptlrpc_request *req;
898 struct ldlm_cb_async_args *ca;
899 long total_enqueue_wait;
900 int instant_cancel = 0;
904 LASSERT(lock != NULL);
905 LASSERT(data != NULL);
907 total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(),
908 lock->l_last_activity);
910 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
911 &RQF_LDLM_CP_CALLBACK);
915 /* server namespace, doesn't need lock */
916 if (lock->l_resource->lr_lvb_len) {
917 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT,
918 lock->l_resource->lr_lvb_len);
921 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
923 ptlrpc_request_free(req);
927 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
928 ca = ptlrpc_req_async_args(req);
929 ca->ca_set_arg = arg;
932 req->rq_interpret_reply = ldlm_cb_interpret;
933 req->rq_no_resend = 1;
934 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
936 body->lock_handle[0] = lock->l_remote_handle;
937 body->lock_flags = flags;
938 ldlm_lock2desc(lock, &body->lock_desc);
939 if (lock->l_resource->lr_lvb_len) {
940 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
942 lock_res(lock->l_resource);
943 memcpy(lvb, lock->l_resource->lr_lvb_data,
944 lock->l_resource->lr_lvb_len);
945 unlock_res(lock->l_resource);
948 LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)",
951 /* Server-side enqueue wait time estimate, used in
952 __ldlm_add_waiting_lock to set future enqueue timers */
953 if (total_enqueue_wait < ldlm_get_enq_timeout(lock))
954 at_measured(ldlm_lock_to_ns_at(lock),
957 /* bz18618. Don't add lock enqueue time we spend waiting for a
958 previous callback to fail. Locks waiting legitimately will
959 get extended by ldlm_refresh_waiting_lock regardless of the
960 estimate, so it's okay to underestimate here. */
961 LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. "
962 "It is likely that a previous callback timed out.",
964 at_get(ldlm_lock_to_ns_at(lock)));
966 ptlrpc_request_set_replen(req);
968 req->rq_send_state = LUSTRE_IMP_FULL;
969 /* ptlrpc_request_pack already set timeout */
971 req->rq_timeout = ldlm_get_rq_timeout();
973 /* We only send real blocking ASTs after the lock is granted */
974 lock_res_and_lock(lock);
975 if (lock->l_flags & LDLM_FL_AST_SENT) {
976 body->lock_flags |= LDLM_FL_AST_SENT;
977 /* copy ast flags like LDLM_FL_DISCARD_DATA */
978 body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
980 /* We might get here prior to ldlm_handle_enqueue setting
981 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
982 * into waiting list, but this is safe and similar code in
983 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
984 * that would not only cancel the lock, but will also remove
985 * it from waiting list */
986 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
987 unlock_res_and_lock(lock);
988 ldlm_lock_cancel(lock);
990 lock_res_and_lock(lock);
992 /* start the lock-timeout clock */
993 ldlm_add_waiting_lock(lock);
996 unlock_res_and_lock(lock);
998 if (lock->l_export && lock->l_export->exp_nid_stats &&
999 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1000 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1001 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1003 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1008 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1010 struct ldlm_cb_set_arg *arg = data;
1011 struct ldlm_request *body;
1012 struct ptlrpc_request *req;
1013 struct ldlm_cb_async_args *ca;
1017 LASSERT(lock != NULL);
1019 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1020 &RQF_LDLM_GL_CALLBACK,
1021 LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK);
1026 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1027 body->lock_handle[0] = lock->l_remote_handle;
1028 ldlm_lock2desc(lock, &body->lock_desc);
1030 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1031 ca = ptlrpc_req_async_args(req);
1032 ca->ca_set_arg = arg;
1035 /* server namespace, doesn't need lock */
1036 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1037 lock->l_resource->lr_lvb_len);
1038 ptlrpc_request_set_replen(req);
1040 req->rq_send_state = LUSTRE_IMP_FULL;
1041 /* ptlrpc_request_alloc_pack already set timeout */
1043 req->rq_timeout = ldlm_get_rq_timeout();
1045 req->rq_interpret_reply = ldlm_cb_interpret;
1047 if (lock->l_export && lock->l_export->exp_nid_stats &&
1048 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1049 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1050 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1052 rc = ldlm_ast_fini(req, arg, lock, 0);
1057 int ldlm_glimpse_locks(struct ldlm_resource *res, cfs_list_t *gl_work_list)
1062 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1064 if (rc == -ERESTART)
1065 ldlm_reprocess_all(res);
1070 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1071 struct lprocfs_stats *srv_stats)
1073 int lock_type = 0, op = 0;
1075 lock_type = dlm_req->lock_desc.l_resource.lr_type;
1077 switch (lock_type) {
1079 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1082 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1083 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1085 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1088 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1091 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1099 lprocfs_counter_incr(srv_stats, op);
1105 * Main server-side entry point into LDLM. This is called by ptlrpc service
1106 * threads to carry out client lock enqueueing requests.
1108 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1109 struct ptlrpc_request *req,
1110 const struct ldlm_request *dlm_req,
1111 const struct ldlm_callback_suite *cbs)
1113 struct ldlm_reply *dlm_rep;
1115 ldlm_error_t err = ELDLM_OK;
1116 struct ldlm_lock *lock = NULL;
1117 void *cookie = NULL;
1121 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1123 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1124 flags = dlm_req->lock_flags;
1126 LASSERT(req->rq_export);
1128 if (ptlrpc_req2svc(req)->srv_stats != NULL)
1129 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1131 if (req->rq_export && req->rq_export->exp_nid_stats &&
1132 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1133 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1134 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1136 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1137 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1138 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1139 dlm_req->lock_desc.l_resource.lr_type);
1140 GOTO(out, rc = -EFAULT);
1143 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1144 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1145 dlm_req->lock_desc.l_req_mode &
1146 (dlm_req->lock_desc.l_req_mode-1))) {
1147 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1148 dlm_req->lock_desc.l_req_mode);
1149 GOTO(out, rc = -EFAULT);
1152 if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
1153 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1155 DEBUG_REQ(D_ERROR, req,
1156 "PLAIN lock request from IBITS client?");
1157 GOTO(out, rc = -EPROTO);
1159 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1161 DEBUG_REQ(D_ERROR, req,
1162 "IBITS lock request from unaware client?");
1163 GOTO(out, rc = -EPROTO);
1167 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1168 against server's _CONNECT_SUPPORTED flags? (I don't want to use
1169 ibits for mgc/mgs) */
1171 /* INODEBITS_INTEROP: Perform conversion from plain lock to
1172 * inodebits lock if client does not support them. */
1173 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
1174 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1175 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1176 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1177 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1178 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1179 dlm_req->lock_desc.l_req_mode = LCK_CR;
1183 if (unlikely(flags & LDLM_FL_REPLAY)) {
1184 /* Find an existing lock in the per-export lock hash */
1185 /* In the function below, .hs_keycmp resolves to
1186 * ldlm_export_lock_keycmp() */
1187 /* coverity[overrun-buffer-val] */
1188 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1189 (void *)&dlm_req->lock_handle[0]);
1191 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1192 LPX64, lock->l_handle.h_cookie);
1193 GOTO(existing_lock, rc = 0);
1197 /* The lock's callback data might be set in the policy function */
1198 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1199 dlm_req->lock_desc.l_resource.lr_type,
1200 dlm_req->lock_desc.l_req_mode,
1204 GOTO(out, rc = -ENOMEM);
1206 lock->l_last_activity = cfs_time_current_sec();
1207 lock->l_remote_handle = dlm_req->lock_handle[0];
1208 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1210 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1211 /* Don't enqueue a lock onto the export if it is been disonnected
1212 * due to eviction (bug 3822) or server umount (bug 24324).
1213 * Cancel it now instead. */
1214 if (req->rq_export->exp_disconnected) {
1215 LDLM_ERROR(lock, "lock on disconnected export %p",
1217 GOTO(out, rc = -ENOTCONN);
1220 lock->l_export = class_export_lock_get(req->rq_export, lock);
1221 if (lock->l_export->exp_lock_hash)
1222 cfs_hash_add(lock->l_export->exp_lock_hash,
1223 &lock->l_remote_handle,
1228 if (flags & LDLM_FL_HAS_INTENT) {
1229 /* In this case, the reply buffer is allocated deep in
1230 * local_lock_enqueue by the policy function. */
1233 /* based on the assumption that lvb size never changes during
1234 * resource life time otherwise it need resource->lr_lock's
1236 if (lock->l_resource->lr_lvb_len) {
1237 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1239 lock->l_resource->lr_lvb_len);
1242 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1243 GOTO(out, rc = -ENOMEM);
1245 rc = req_capsule_server_pack(&req->rq_pill);
1250 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1251 ldlm_convert_policy_to_local(req->rq_export,
1252 dlm_req->lock_desc.l_resource.lr_type,
1253 &dlm_req->lock_desc.l_policy_data,
1254 &lock->l_policy_data);
1255 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1256 lock->l_req_extent = lock->l_policy_data.l_extent;
1258 err = ldlm_lock_enqueue(ns, &lock, cookie, (int *)&flags);
1262 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1263 dlm_rep->lock_flags = flags;
1265 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1266 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1268 /* We never send a blocking AST until the lock is granted, but
1269 * we can tell it right now */
1270 lock_res_and_lock(lock);
1272 /* Now take into account flags to be inherited from original lock
1273 request both in reply to client and in our own lock flags. */
1274 dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1275 lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1277 /* Don't move a pending lock onto the export if it has already been
1278 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1279 * Cancel it now instead. */
1280 if (unlikely(req->rq_export->exp_disconnected ||
1281 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1282 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1284 } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1285 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
1286 if (lock->l_granted_mode == lock->l_req_mode) {
1288 * Only cancel lock if it was granted, because it would
1289 * be destroyed immediately and would never be granted
1290 * in the future, causing timeouts on client. Not
1291 * granted lock will be cancelled immediately after
1292 * sending completion AST.
1294 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1295 unlock_res_and_lock(lock);
1296 ldlm_lock_cancel(lock);
1297 lock_res_and_lock(lock);
1299 ldlm_add_waiting_lock(lock);
1302 /* Make sure we never ever grant usual metadata locks to liblustre
1304 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1305 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1306 req->rq_export->exp_libclient) {
1307 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1308 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1309 CERROR("Granting sync lock to libclient. "
1310 "req fl %d, rep fl %d, lock fl "LPX64"\n",
1311 dlm_req->lock_flags, dlm_rep->lock_flags,
1313 LDLM_ERROR(lock, "sync lock");
1314 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1315 struct ldlm_intent *it;
1317 it = req_capsule_client_get(&req->rq_pill,
1320 CERROR("This is intent %s ("LPU64")\n",
1321 ldlm_it2str(it->opc), it->opc);
1327 unlock_res_and_lock(lock);
1331 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1332 if (!req->rq_packed_final) {
1333 err = lustre_pack_reply(req, 1, NULL, NULL);
1338 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1339 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1341 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1342 "(err=%d, rc=%d)", err, rc);
1345 if (lock->l_resource->lr_lvb_len > 0) {
1346 /* MDT path won't handle lr_lvb_data, so
1347 * lock/unlock better be contained in the
1351 lvb = req_capsule_server_get(&req->rq_pill,
1353 LASSERTF(lvb != NULL, "req %p, lock %p\n",
1355 lock_res(lock->l_resource);
1356 memcpy(lvb, lock->l_resource->lr_lvb_data,
1357 lock->l_resource->lr_lvb_len);
1358 unlock_res(lock->l_resource);
1361 lock_res_and_lock(lock);
1362 ldlm_resource_unlink_lock(lock);
1363 ldlm_lock_destroy_nolock(lock);
1364 unlock_res_and_lock(lock);
1367 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1368 ldlm_reprocess_all(lock->l_resource);
1370 LDLM_LOCK_RELEASE(lock);
1373 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1379 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1380 ldlm_completion_callback completion_callback,
1381 ldlm_blocking_callback blocking_callback,
1382 ldlm_glimpse_callback glimpse_callback)
1384 struct ldlm_request *dlm_req;
1385 struct ldlm_callback_suite cbs = {
1386 .lcs_completion = completion_callback,
1387 .lcs_blocking = blocking_callback,
1388 .lcs_glimpse = glimpse_callback
1392 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1393 if (dlm_req != NULL) {
1394 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1395 req, dlm_req, &cbs);
1402 int ldlm_handle_convert0(struct ptlrpc_request *req,
1403 const struct ldlm_request *dlm_req)
1405 struct ldlm_reply *dlm_rep;
1406 struct ldlm_lock *lock;
1410 if (req->rq_export && req->rq_export->exp_nid_stats &&
1411 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1412 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1413 LDLM_CONVERT - LDLM_FIRST_OPC);
1415 rc = req_capsule_server_pack(&req->rq_pill);
1419 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1420 dlm_rep->lock_flags = dlm_req->lock_flags;
1422 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1424 req->rq_status = EINVAL;
1428 LDLM_DEBUG(lock, "server-side convert handler START");
1430 lock->l_last_activity = cfs_time_current_sec();
1431 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1432 &dlm_rep->lock_flags);
1434 if (ldlm_del_waiting_lock(lock))
1435 LDLM_DEBUG(lock, "converted waiting lock");
1438 req->rq_status = EDEADLOCK;
1443 if (!req->rq_status)
1444 ldlm_reprocess_all(lock->l_resource);
1445 LDLM_DEBUG(lock, "server-side convert handler END");
1446 LDLM_LOCK_PUT(lock);
1448 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1453 int ldlm_handle_convert(struct ptlrpc_request *req)
1456 struct ldlm_request *dlm_req;
1458 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1459 if (dlm_req != NULL) {
1460 rc = ldlm_handle_convert0(req, dlm_req);
1462 CERROR ("Can't unpack dlm_req\n");
1468 /* Cancel all the locks whos handles are packed into ldlm_request */
1469 int ldlm_request_cancel(struct ptlrpc_request *req,
1470 const struct ldlm_request *dlm_req, int first)
1472 struct ldlm_resource *res, *pres = NULL;
1473 struct ldlm_lock *lock;
1474 int i, count, done = 0;
1477 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1481 /* There is no lock on the server at the replay time,
1482 * skip lock cancelling to make replay tests to pass. */
1483 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1486 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1487 "starting at %d", count, first);
1489 for (i = first; i < count; i++) {
1490 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1492 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1493 "lock (cookie "LPU64")",
1494 dlm_req->lock_handle[i].cookie);
1498 res = lock->l_resource;
1503 ldlm_reprocess_all(pres);
1504 LDLM_RESOURCE_DELREF(pres);
1505 ldlm_resource_putref(pres);
1508 ldlm_resource_getref(res);
1509 LDLM_RESOURCE_ADDREF(res);
1510 ldlm_res_lvbo_update(res, NULL, 1);
1514 ldlm_lock_cancel(lock);
1515 LDLM_LOCK_PUT(lock);
1518 ldlm_reprocess_all(pres);
1519 LDLM_RESOURCE_DELREF(pres);
1520 ldlm_resource_putref(pres);
1522 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1526 int ldlm_handle_cancel(struct ptlrpc_request *req)
1528 struct ldlm_request *dlm_req;
1532 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1533 if (dlm_req == NULL) {
1534 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1538 if (req->rq_export && req->rq_export->exp_nid_stats &&
1539 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1540 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1541 LDLM_CANCEL - LDLM_FIRST_OPC);
1543 rc = req_capsule_server_pack(&req->rq_pill);
1547 if (!ldlm_request_cancel(req, dlm_req, 0))
1548 req->rq_status = ESTALE;
1550 RETURN(ptlrpc_reply(req));
1552 #endif /* HAVE_SERVER_SUPPORT */
1554 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1555 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1560 LDLM_DEBUG(lock, "client blocking AST callback handler");
1562 lock_res_and_lock(lock);
1563 lock->l_flags |= LDLM_FL_CBPENDING;
1565 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1566 lock->l_flags |= LDLM_FL_CANCEL;
1568 do_ast = (!lock->l_readers && !lock->l_writers);
1569 unlock_res_and_lock(lock);
1572 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1573 lock, lock->l_blocking_ast);
1574 if (lock->l_blocking_ast != NULL)
1575 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1578 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1582 LDLM_DEBUG(lock, "client blocking callback handler END");
1583 LDLM_LOCK_RELEASE(lock);
1587 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1588 struct ldlm_namespace *ns,
1589 struct ldlm_request *dlm_req,
1590 struct ldlm_lock *lock)
1593 CFS_LIST_HEAD(ast_list);
1596 LDLM_DEBUG(lock, "client completion callback handler START");
1598 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1599 int to = cfs_time_seconds(1);
1601 cfs_schedule_timeout_and_set_state(
1602 CFS_TASK_INTERRUPTIBLE, to);
1603 if (lock->l_granted_mode == lock->l_req_mode ||
1609 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1611 if (lock->l_lvb_len > 0) {
1612 /* for extent lock, lvb contains ost_lvb{}. */
1613 LASSERT(lock->l_lvb_data != NULL);
1614 LASSERTF(lock->l_lvb_len == lvb_len,
1615 "preallocated %d, actual %d.\n",
1616 lock->l_lvb_len, lvb_len);
1617 } else { /* for layout lock, lvb has variable length */
1620 OBD_ALLOC(lvb_data, lvb_len);
1621 if (lvb_data == NULL)
1622 LDLM_ERROR(lock, "no memory.\n");
1624 lock_res_and_lock(lock);
1625 if (lvb_data == NULL) {
1626 lock->l_flags |= LDLM_FL_FAILED;
1628 LASSERT(lock->l_lvb_data == NULL);
1629 lock->l_lvb_data = lvb_data;
1630 lock->l_lvb_len = lvb_len;
1632 unlock_res_and_lock(lock);
1636 lock_res_and_lock(lock);
1637 if (lock->l_destroyed ||
1638 lock->l_granted_mode == lock->l_req_mode) {
1639 /* bug 11300: the lock has already been granted */
1640 unlock_res_and_lock(lock);
1641 LDLM_DEBUG(lock, "Double grant race happened");
1642 LDLM_LOCK_RELEASE(lock);
1647 /* If we receive the completion AST before the actual enqueue returned,
1648 * then we might need to switch lock modes, resources, or extents. */
1649 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1650 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1651 LDLM_DEBUG(lock, "completion AST, new lock mode");
1654 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1655 ldlm_convert_policy_to_local(req->rq_export,
1656 dlm_req->lock_desc.l_resource.lr_type,
1657 &dlm_req->lock_desc.l_policy_data,
1658 &lock->l_policy_data);
1659 LDLM_DEBUG(lock, "completion AST, new policy data");
1662 ldlm_resource_unlink_lock(lock);
1663 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1664 &lock->l_resource->lr_name,
1665 sizeof(lock->l_resource->lr_name)) != 0) {
1666 unlock_res_and_lock(lock);
1667 if (ldlm_lock_change_resource(ns, lock,
1668 &dlm_req->lock_desc.l_resource.lr_name) != 0) {
1669 LDLM_ERROR(lock, "Failed to allocate resource");
1670 LDLM_LOCK_RELEASE(lock);
1674 LDLM_DEBUG(lock, "completion AST, new resource");
1675 CERROR("change resource!\n");
1676 lock_res_and_lock(lock);
1679 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1680 /* BL_AST locks are not needed in lru.
1681 * let ldlm_cancel_lru() be fast. */
1682 ldlm_lock_remove_from_lru(lock);
1683 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1684 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1687 if (lock->l_lvb_len) {
1688 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
1689 RCL_CLIENT) < lock->l_lvb_len) {
1690 LDLM_ERROR(lock, "completion AST did not contain "
1693 void *lvb = req_capsule_client_get(&req->rq_pill,
1695 memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1699 ldlm_grant_lock(lock, &ast_list);
1700 unlock_res_and_lock(lock);
1702 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1704 /* Let Enqueue to call osc_lock_upcall() and initialize
1706 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1708 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1710 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1712 LDLM_LOCK_RELEASE(lock);
1716 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1717 struct ldlm_namespace *ns,
1718 struct ldlm_request *dlm_req,
1719 struct ldlm_lock *lock)
1724 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1726 if (lock->l_glimpse_ast != NULL)
1727 rc = lock->l_glimpse_ast(lock, req);
1729 if (req->rq_repmsg != NULL) {
1732 req->rq_status = rc;
1736 lock_res_and_lock(lock);
1737 if (lock->l_granted_mode == LCK_PW &&
1738 !lock->l_readers && !lock->l_writers &&
1739 cfs_time_after(cfs_time_current(),
1740 cfs_time_add(lock->l_last_used,
1741 cfs_time_seconds(10)))) {
1742 unlock_res_and_lock(lock);
1743 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1744 ldlm_handle_bl_callback(ns, NULL, lock);
1749 unlock_res_and_lock(lock);
1750 LDLM_LOCK_RELEASE(lock);
1754 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1756 if (req->rq_no_reply)
1759 req->rq_status = rc;
1760 if (!req->rq_packed_final) {
1761 rc = lustre_pack_reply(req, 1, NULL, NULL);
1765 return ptlrpc_reply(req);
1769 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi, int mode)
1771 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1774 cfs_spin_lock(&blp->blp_lock);
1775 if (blwi->blwi_lock && blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) {
1776 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1777 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1779 /* other blocking callbacks are added to the regular list */
1780 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1782 cfs_spin_unlock(&blp->blp_lock);
1784 cfs_waitq_signal(&blp->blp_waitq);
1786 /* can not use blwi->blwi_mode as blwi could be already freed in
1788 if (mode == LDLM_SYNC)
1789 cfs_wait_for_completion(&blwi->blwi_comp);
1794 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1795 struct ldlm_namespace *ns,
1796 struct ldlm_lock_desc *ld,
1797 cfs_list_t *cancels, int count,
1798 struct ldlm_lock *lock,
1801 cfs_init_completion(&blwi->blwi_comp);
1802 CFS_INIT_LIST_HEAD(&blwi->blwi_head);
1804 if (cfs_memory_pressure_get())
1805 blwi->blwi_mem_pressure = 1;
1808 blwi->blwi_mode = mode;
1810 blwi->blwi_ld = *ld;
1812 cfs_list_add(&blwi->blwi_head, cancels);
1813 cfs_list_del_init(cancels);
1814 blwi->blwi_count = count;
1816 blwi->blwi_lock = lock;
1820 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1821 struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
1822 cfs_list_t *cancels, int count, int mode)
1826 if (cancels && count == 0)
1829 if (mode == LDLM_SYNC) {
1830 /* if it is synchronous call do minimum mem alloc, as it could
1831 * be triggered from kernel shrinker
1833 struct ldlm_bl_work_item blwi;
1834 memset(&blwi, 0, sizeof(blwi));
1835 init_blwi(&blwi, ns, ld, cancels, count, lock, LDLM_SYNC);
1836 RETURN(__ldlm_bl_to_thread(&blwi, LDLM_SYNC));
1838 struct ldlm_bl_work_item *blwi;
1839 OBD_ALLOC(blwi, sizeof(*blwi));
1842 init_blwi(blwi, ns, ld, cancels, count, lock, LDLM_ASYNC);
1844 RETURN(__ldlm_bl_to_thread(blwi, LDLM_ASYNC));
1850 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1851 struct ldlm_lock *lock)
1854 RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LDLM_ASYNC));
1860 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1861 cfs_list_t *cancels, int count, int mode)
1864 RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count, mode));
1870 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
1871 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
1873 struct obd_device *obd = req->rq_export->exp_obd;
1880 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
1882 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
1884 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1886 DEBUG_REQ(D_IOCTL, req, "no set_info key");
1889 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
1891 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1893 DEBUG_REQ(D_IOCTL, req, "no set_info val");
1896 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
1899 /* We are responsible for swabbing contents of val */
1901 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
1902 /* Pass it on to mdc (the "export" in this case) */
1903 rc = obd_set_info_async(req->rq_svc_thread->t_env,
1905 sizeof(KEY_HSM_COPYTOOL_SEND),
1906 KEY_HSM_COPYTOOL_SEND,
1909 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
1914 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
1915 const char *msg, int rc,
1916 struct lustre_handle *handle)
1918 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
1919 "%s: [nid %s] [rc %d] [lock "LPX64"]",
1920 msg, libcfs_id2str(req->rq_peer), rc,
1921 handle ? handle->cookie : 0);
1922 if (req->rq_no_reply)
1923 CWARN("No reply was sent, maybe cause bug 21636.\n");
1925 CWARN("Send reply failed, maybe cause bug 21636.\n");
1928 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1929 static int ldlm_callback_handler(struct ptlrpc_request *req)
1931 struct ldlm_namespace *ns;
1932 struct ldlm_request *dlm_req;
1933 struct ldlm_lock *lock;
1937 /* Requests arrive in sender's byte order. The ptlrpc service
1938 * handler has already checked and, if necessary, byte-swapped the
1939 * incoming request message body, but I am responsible for the
1940 * message buffers. */
1942 /* do nothing for sec context finalize */
1943 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
1946 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1948 if (req->rq_export == NULL) {
1949 rc = ldlm_callback_reply(req, -ENOTCONN);
1950 ldlm_callback_errmsg(req, "Operate on unconnected server",
1955 LASSERT(req->rq_export != NULL);
1956 LASSERT(req->rq_export->exp_obd != NULL);
1958 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1959 case LDLM_BL_CALLBACK:
1960 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK))
1963 case LDLM_CP_CALLBACK:
1964 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK))
1967 case LDLM_GL_CALLBACK:
1968 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK))
1972 rc = ldlm_handle_setinfo(req);
1973 ldlm_callback_reply(req, rc);
1975 case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
1976 CERROR("shouldn't be handling OBD_LOG_CANCEL on DLM thread\n");
1977 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1978 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1980 rc = llog_origin_handle_cancel(req);
1981 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
1983 ldlm_callback_reply(req, rc);
1985 case OBD_QC_CALLBACK:
1986 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
1987 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
1989 rc = target_handle_qc_callback(req);
1990 ldlm_callback_reply(req, rc);
1994 /* reply in handler */
1995 req_capsule_set(&req->rq_pill, &RQF_MDS_QUOTA_DQACQ);
1996 rc = target_handle_dqacq_callback(req);
1998 case LLOG_ORIGIN_HANDLE_CREATE:
1999 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2000 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2002 rc = llog_origin_handle_open(req);
2003 ldlm_callback_reply(req, rc);
2005 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2006 req_capsule_set(&req->rq_pill,
2007 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2008 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2010 rc = llog_origin_handle_next_block(req);
2011 ldlm_callback_reply(req, rc);
2013 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2014 req_capsule_set(&req->rq_pill,
2015 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2016 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2018 rc = llog_origin_handle_read_header(req);
2019 ldlm_callback_reply(req, rc);
2021 case LLOG_ORIGIN_HANDLE_CLOSE:
2022 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2024 rc = llog_origin_handle_close(req);
2025 ldlm_callback_reply(req, rc);
2028 CERROR("unknown opcode %u\n",
2029 lustre_msg_get_opc(req->rq_reqmsg));
2030 ldlm_callback_reply(req, -EPROTO);
2034 ns = req->rq_export->exp_obd->obd_namespace;
2035 LASSERT(ns != NULL);
2037 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2039 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2040 if (dlm_req == NULL) {
2041 rc = ldlm_callback_reply(req, -EPROTO);
2042 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2047 /* Force a known safe race, send a cancel to the server for a lock
2048 * which the server has already started a blocking callback on. */
2049 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2050 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2051 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]);
2053 CERROR("ldlm_cli_cancel: %d\n", rc);
2056 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2058 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2059 "disappeared\n", dlm_req->lock_handle[0].cookie);
2060 rc = ldlm_callback_reply(req, -EINVAL);
2061 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2062 &dlm_req->lock_handle[0]);
2066 if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
2067 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2068 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2070 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2071 lock_res_and_lock(lock);
2072 lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
2073 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2074 /* If somebody cancels lock and cache is already dropped,
2075 * or lock is failed before cp_ast received on client,
2076 * we can tell the server we have no lock. Otherwise, we
2077 * should send cancel after dropping the cache. */
2078 if (((lock->l_flags & LDLM_FL_CANCELING) &&
2079 (lock->l_flags & LDLM_FL_BL_DONE)) ||
2080 (lock->l_flags & LDLM_FL_FAILED)) {
2081 LDLM_DEBUG(lock, "callback on lock "
2082 LPX64" - lock disappeared\n",
2083 dlm_req->lock_handle[0].cookie);
2084 unlock_res_and_lock(lock);
2085 LDLM_LOCK_RELEASE(lock);
2086 rc = ldlm_callback_reply(req, -EINVAL);
2087 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2088 &dlm_req->lock_handle[0]);
2091 /* BL_AST locks are not needed in lru.
2092 * let ldlm_cancel_lru() be fast. */
2093 ldlm_lock_remove_from_lru(lock);
2094 lock->l_flags |= LDLM_FL_BL_AST;
2096 unlock_res_and_lock(lock);
2098 /* We want the ost thread to get this reply so that it can respond
2099 * to ost requests (write cache writeback) that might be triggered
2102 * But we'd also like to be able to indicate in the reply that we're
2103 * cancelling right now, because it's unused, or have an intent result
2104 * in the reply, so we might have to push the responsibility for sending
2105 * the reply down into the AST handlers, alas. */
2107 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2108 case LDLM_BL_CALLBACK:
2109 CDEBUG(D_INODE, "blocking ast\n");
2110 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2111 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) {
2112 rc = ldlm_callback_reply(req, 0);
2113 if (req->rq_no_reply || rc)
2114 ldlm_callback_errmsg(req, "Normal process", rc,
2115 &dlm_req->lock_handle[0]);
2117 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2118 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2120 case LDLM_CP_CALLBACK:
2121 CDEBUG(D_INODE, "completion ast\n");
2122 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2123 ldlm_callback_reply(req, 0);
2124 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2126 case LDLM_GL_CALLBACK:
2127 CDEBUG(D_INODE, "glimpse ast\n");
2128 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2129 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2132 LBUG(); /* checked above */
2138 #ifdef HAVE_SERVER_SUPPORT
2139 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2144 /* Requests arrive in sender's byte order. The ptlrpc service
2145 * handler has already checked and, if necessary, byte-swapped the
2146 * incoming request message body, but I am responsible for the
2147 * message buffers. */
2149 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2151 if (req->rq_export == NULL) {
2152 struct ldlm_request *dlm_req;
2154 CERROR("%s from %s arrived at %lu with bad export cookie "
2156 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2157 libcfs_nid2str(req->rq_peer.nid),
2158 req->rq_arrival_time.tv_sec,
2159 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2161 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2162 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2163 dlm_req = req_capsule_client_get(&req->rq_pill,
2165 if (dlm_req != NULL)
2166 ldlm_lock_dump_handle(D_ERROR,
2167 &dlm_req->lock_handle[0]);
2169 ldlm_callback_reply(req, -ENOTCONN);
2173 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2175 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2177 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2178 CDEBUG(D_INODE, "cancel\n");
2179 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
2181 rc = ldlm_handle_cancel(req);
2185 case OBD_LOG_CANCEL:
2186 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2187 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2189 rc = llog_origin_handle_cancel(req);
2190 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2192 ldlm_callback_reply(req, rc);
2195 CERROR("invalid opcode %d\n",
2196 lustre_msg_get_opc(req->rq_reqmsg));
2197 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2198 ldlm_callback_reply(req, -EINVAL);
2204 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2205 struct ldlm_lock *lock)
2207 struct ldlm_request *dlm_req;
2208 struct lustre_handle lockh;
2213 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2214 if (dlm_req == NULL)
2217 ldlm_lock2handle(lock, &lockh);
2218 for (i = 0; i < dlm_req->lock_count; i++) {
2219 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2221 DEBUG_REQ(D_RPCTRACE, req,
2222 "Prio raised by lock "LPX64".", lockh.cookie);
2233 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2235 struct ldlm_request *dlm_req;
2240 /* no prolong in recovery */
2241 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2244 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2245 if (dlm_req == NULL)
2248 for (i = 0; i < dlm_req->lock_count; i++) {
2249 struct ldlm_lock *lock;
2251 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2255 rc = !!(lock->l_flags & LDLM_FL_AST_SENT);
2257 LDLM_DEBUG(lock, "hpreq cancel lock");
2258 LDLM_LOCK_PUT(lock);
2267 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2268 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2269 .hpreq_check = ldlm_cancel_hpreq_check,
2273 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2277 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2279 if (req->rq_export == NULL)
2282 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2283 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2284 req->rq_ops = &ldlm_cancel_hpreq_ops;
2289 int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2290 cfs_hlist_node_t *hnode, void *data)
2293 cfs_list_t *rpc_list = data;
2294 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2296 lock_res_and_lock(lock);
2298 if (lock->l_req_mode != lock->l_granted_mode) {
2299 unlock_res_and_lock(lock);
2303 LASSERT(lock->l_resource);
2304 if (lock->l_resource->lr_type != LDLM_IBITS &&
2305 lock->l_resource->lr_type != LDLM_PLAIN) {
2306 unlock_res_and_lock(lock);
2310 if (lock->l_flags & LDLM_FL_AST_SENT) {
2311 unlock_res_and_lock(lock);
2315 LASSERT(lock->l_blocking_ast);
2316 LASSERT(!lock->l_blocking_lock);
2318 lock->l_flags |= LDLM_FL_AST_SENT;
2319 if (lock->l_export && lock->l_export->exp_lock_hash) {
2320 /* NB: it's safe to call cfs_hash_del() even lock isn't
2321 * in exp_lock_hash. */
2322 /* In the function below, .hs_keycmp resolves to
2323 * ldlm_export_lock_keycmp() */
2324 /* coverity[overrun-buffer-val] */
2325 cfs_hash_del(lock->l_export->exp_lock_hash,
2326 &lock->l_remote_handle, &lock->l_exp_hash);
2329 cfs_list_add_tail(&lock->l_rk_ast, rpc_list);
2330 LDLM_LOCK_GET(lock);
2332 unlock_res_and_lock(lock);
2336 void ldlm_revoke_export_locks(struct obd_export *exp)
2338 cfs_list_t rpc_list;
2341 CFS_INIT_LIST_HEAD(&rpc_list);
2342 cfs_hash_for_each_empty(exp->exp_lock_hash,
2343 ldlm_revoke_lock_cb, &rpc_list);
2344 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2345 LDLM_WORK_REVOKE_AST);
2349 #endif /* HAVE_SERVER_SUPPORT */
2352 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2354 struct ldlm_bl_work_item *blwi = NULL;
2355 static unsigned int num_bl = 0;
2357 cfs_spin_lock(&blp->blp_lock);
2358 /* process a request from the blp_list at least every blp_num_threads */
2359 if (!cfs_list_empty(&blp->blp_list) &&
2360 (cfs_list_empty(&blp->blp_prio_list) || num_bl == 0))
2361 blwi = cfs_list_entry(blp->blp_list.next,
2362 struct ldlm_bl_work_item, blwi_entry);
2364 if (!cfs_list_empty(&blp->blp_prio_list))
2365 blwi = cfs_list_entry(blp->blp_prio_list.next,
2366 struct ldlm_bl_work_item,
2370 if (++num_bl >= cfs_atomic_read(&blp->blp_num_threads))
2372 cfs_list_del(&blwi->blwi_entry);
2374 cfs_spin_unlock(&blp->blp_lock);
2379 /* This only contains temporary data until the thread starts */
2380 struct ldlm_bl_thread_data {
2381 char bltd_name[CFS_CURPROC_COMM_MAX];
2382 struct ldlm_bl_pool *bltd_blp;
2383 cfs_completion_t bltd_comp;
2387 static int ldlm_bl_thread_main(void *arg);
2389 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2391 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2394 cfs_init_completion(&bltd.bltd_comp);
2395 rc = cfs_create_thread(ldlm_bl_thread_main, &bltd, 0);
2397 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
2398 cfs_atomic_read(&blp->blp_num_threads), rc);
2401 cfs_wait_for_completion(&bltd.bltd_comp);
2406 static int ldlm_bl_thread_main(void *arg)
2408 struct ldlm_bl_pool *blp;
2412 struct ldlm_bl_thread_data *bltd = arg;
2414 blp = bltd->bltd_blp;
2417 cfs_atomic_inc_return(&blp->blp_num_threads) - 1;
2418 cfs_atomic_inc(&blp->blp_busy_threads);
2420 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
2421 "ldlm_bl_%02d", bltd->bltd_num);
2422 cfs_daemonize(bltd->bltd_name);
2424 cfs_complete(&bltd->bltd_comp);
2425 /* cannot use bltd after this, it is only on caller's stack */
2429 struct l_wait_info lwi = { 0 };
2430 struct ldlm_bl_work_item *blwi = NULL;
2433 blwi = ldlm_bl_get_work(blp);
2436 cfs_atomic_dec(&blp->blp_busy_threads);
2437 l_wait_event_exclusive(blp->blp_waitq,
2438 (blwi = ldlm_bl_get_work(blp)) != NULL,
2440 busy = cfs_atomic_inc_return(&blp->blp_busy_threads);
2442 busy = cfs_atomic_read(&blp->blp_busy_threads);
2445 if (blwi->blwi_ns == NULL)
2446 /* added by ldlm_cleanup() */
2449 /* Not fatal if racy and have a few too many threads */
2450 if (unlikely(busy < blp->blp_max_threads &&
2451 busy >= cfs_atomic_read(&blp->blp_num_threads) &&
2452 !blwi->blwi_mem_pressure))
2453 /* discard the return value, we tried */
2454 ldlm_bl_thread_start(blp);
2456 if (blwi->blwi_mem_pressure)
2457 cfs_memory_pressure_set();
2459 if (blwi->blwi_count) {
2461 /* The special case when we cancel locks in lru
2462 * asynchronously, we pass the list of locks here.
2463 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2464 * canceled locally yet. */
2465 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2468 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL, 0);
2470 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2473 if (blwi->blwi_mem_pressure)
2474 cfs_memory_pressure_clr();
2476 if (blwi->blwi_mode == LDLM_ASYNC)
2477 OBD_FREE(blwi, sizeof(*blwi));
2479 cfs_complete(&blwi->blwi_comp);
2482 cfs_atomic_dec(&blp->blp_busy_threads);
2483 cfs_atomic_dec(&blp->blp_num_threads);
2484 cfs_complete(&blp->blp_comp);
2490 static int ldlm_setup(void);
2491 static int ldlm_cleanup(void);
2493 int ldlm_get_ref(void)
2497 cfs_mutex_lock(&ldlm_ref_mutex);
2498 if (++ldlm_refcount == 1) {
2503 cfs_mutex_unlock(&ldlm_ref_mutex);
2508 void ldlm_put_ref(void)
2511 cfs_mutex_lock(&ldlm_ref_mutex);
2512 if (ldlm_refcount == 1) {
2513 int rc = ldlm_cleanup();
2515 CERROR("ldlm_cleanup failed: %d\n", rc);
2521 cfs_mutex_unlock(&ldlm_ref_mutex);
2527 * Export handle<->lock hash operations.
2530 ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
2532 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2536 ldlm_export_lock_key(cfs_hlist_node_t *hnode)
2538 struct ldlm_lock *lock;
2540 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2541 return &lock->l_remote_handle;
2545 ldlm_export_lock_keycpy(cfs_hlist_node_t *hnode, void *key)
2547 struct ldlm_lock *lock;
2549 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2550 lock->l_remote_handle = *(struct lustre_handle *)key;
2554 ldlm_export_lock_keycmp(const void *key, cfs_hlist_node_t *hnode)
2556 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2560 ldlm_export_lock_object(cfs_hlist_node_t *hnode)
2562 return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2566 ldlm_export_lock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2568 struct ldlm_lock *lock;
2570 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2571 LDLM_LOCK_GET(lock);
2575 ldlm_export_lock_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2577 struct ldlm_lock *lock;
2579 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2580 LDLM_LOCK_RELEASE(lock);
2583 static cfs_hash_ops_t ldlm_export_lock_ops = {
2584 .hs_hash = ldlm_export_lock_hash,
2585 .hs_key = ldlm_export_lock_key,
2586 .hs_keycmp = ldlm_export_lock_keycmp,
2587 .hs_keycpy = ldlm_export_lock_keycpy,
2588 .hs_object = ldlm_export_lock_object,
2589 .hs_get = ldlm_export_lock_get,
2590 .hs_put = ldlm_export_lock_put,
2591 .hs_put_locked = ldlm_export_lock_put,
2594 int ldlm_init_export(struct obd_export *exp)
2598 exp->exp_lock_hash =
2599 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2600 HASH_EXP_LOCK_CUR_BITS,
2601 HASH_EXP_LOCK_MAX_BITS,
2602 HASH_EXP_LOCK_BKT_BITS, 0,
2603 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2604 &ldlm_export_lock_ops,
2605 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2606 CFS_HASH_NBLK_CHANGE);
2608 if (!exp->exp_lock_hash)
2613 EXPORT_SYMBOL(ldlm_init_export);
2615 void ldlm_destroy_export(struct obd_export *exp)
2618 cfs_hash_putref(exp->exp_lock_hash);
2619 exp->exp_lock_hash = NULL;
2621 ldlm_destroy_flock_export(exp);
2624 EXPORT_SYMBOL(ldlm_destroy_export);
2626 static int ldlm_setup(void)
2628 static struct ptlrpc_service_conf conf;
2629 struct ldlm_bl_pool *blp = NULL;
2636 if (ldlm_state != NULL)
2639 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2640 if (ldlm_state == NULL)
2644 rc = ldlm_proc_setup();
2649 memset(&conf, 0, sizeof(conf));
2650 conf = (typeof(conf)) {
2651 .psc_name = "ldlm_cbd",
2652 .psc_watchdog_factor = 2,
2654 .bc_nbufs = LDLM_NBUFS,
2655 .bc_buf_size = LDLM_BUFSIZE,
2656 .bc_req_max_size = LDLM_MAXREQSIZE,
2657 .bc_rep_max_size = LDLM_MAXREPSIZE,
2658 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2659 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2662 .tc_thr_name = "ldlm_cb",
2663 .tc_thr_factor = LDLM_THR_FACTOR,
2664 .tc_nthrs_init = LDLM_NTHRS_INIT,
2665 .tc_nthrs_base = LDLM_NTHRS_BASE,
2666 .tc_nthrs_max = LDLM_NTHRS_MAX,
2667 .tc_nthrs_user = ldlm_num_threads,
2668 .tc_cpu_affinity = 1,
2669 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2672 .cc_pattern = ldlm_cpts,
2675 .so_req_handler = ldlm_callback_handler,
2678 ldlm_state->ldlm_cb_service = \
2679 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2680 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2681 CERROR("failed to start service\n");
2682 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2683 ldlm_state->ldlm_cb_service = NULL;
2687 #ifdef HAVE_SERVER_SUPPORT
2688 memset(&conf, 0, sizeof(conf));
2689 conf = (typeof(conf)) {
2690 .psc_name = "ldlm_canceld",
2691 .psc_watchdog_factor = 6,
2693 .bc_nbufs = LDLM_NBUFS,
2694 .bc_buf_size = LDLM_BUFSIZE,
2695 .bc_req_max_size = LDLM_MAXREQSIZE,
2696 .bc_rep_max_size = LDLM_MAXREPSIZE,
2697 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
2698 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
2702 .tc_thr_name = "ldlm_cn",
2703 .tc_thr_factor = LDLM_THR_FACTOR,
2704 .tc_nthrs_init = LDLM_NTHRS_INIT,
2705 .tc_nthrs_base = LDLM_NTHRS_BASE,
2706 .tc_nthrs_max = LDLM_NTHRS_MAX,
2707 .tc_nthrs_user = ldlm_num_threads,
2708 .tc_cpu_affinity = 1,
2709 .tc_ctx_tags = LCT_MD_THREAD | \
2714 .cc_pattern = ldlm_cpts,
2717 .so_req_handler = ldlm_cancel_handler,
2718 .so_hpreq_handler = ldlm_hpreq_handler,
2721 ldlm_state->ldlm_cancel_service = \
2722 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2723 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2724 CERROR("failed to start service\n");
2725 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2726 ldlm_state->ldlm_cancel_service = NULL;
2731 OBD_ALLOC(blp, sizeof(*blp));
2733 GOTO(out, rc = -ENOMEM);
2734 ldlm_state->ldlm_bl_pool = blp;
2736 cfs_spin_lock_init(&blp->blp_lock);
2737 CFS_INIT_LIST_HEAD(&blp->blp_list);
2738 CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2739 cfs_waitq_init(&blp->blp_waitq);
2740 cfs_atomic_set(&blp->blp_num_threads, 0);
2741 cfs_atomic_set(&blp->blp_busy_threads, 0);
2744 if (ldlm_num_threads == 0) {
2745 blp->blp_min_threads = LDLM_NTHRS_INIT;
2746 blp->blp_max_threads = LDLM_NTHRS_MAX;
2748 blp->blp_min_threads = blp->blp_max_threads = \
2749 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
2753 for (i = 0; i < blp->blp_min_threads; i++) {
2754 rc = ldlm_bl_thread_start(blp);
2759 # ifdef HAVE_SERVER_SUPPORT
2760 CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2761 expired_lock_thread.elt_state = ELT_STOPPED;
2762 cfs_waitq_init(&expired_lock_thread.elt_waitq);
2764 CFS_INIT_LIST_HEAD(&waiting_locks_list);
2765 cfs_spin_lock_init(&waiting_locks_spinlock);
2766 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2768 rc = cfs_create_thread(expired_lock_main, NULL, CFS_DAEMON_FLAGS);
2770 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2774 cfs_wait_event(expired_lock_thread.elt_waitq,
2775 expired_lock_thread.elt_state == ELT_READY);
2776 # endif /* HAVE_SERVER_SUPPORT */
2778 rc = ldlm_pools_init();
2780 CERROR("Failed to initialize LDLM pools: %d\n", rc);
2791 static int ldlm_cleanup(void)
2795 if (!cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2796 !cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2797 CERROR("ldlm still has namespaces; clean these up first.\n");
2798 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2799 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2806 if (ldlm_state->ldlm_bl_pool != NULL) {
2807 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2809 while (cfs_atomic_read(&blp->blp_num_threads) > 0) {
2810 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2812 cfs_init_completion(&blp->blp_comp);
2814 cfs_spin_lock(&blp->blp_lock);
2815 cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2816 cfs_waitq_signal(&blp->blp_waitq);
2817 cfs_spin_unlock(&blp->blp_lock);
2819 cfs_wait_for_completion(&blp->blp_comp);
2822 OBD_FREE(blp, sizeof(*blp));
2824 #endif /* __KERNEL__ */
2826 if (ldlm_state->ldlm_cb_service != NULL)
2827 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2828 # ifdef HAVE_SERVER_SUPPORT
2829 if (ldlm_state->ldlm_cancel_service != NULL)
2830 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2834 ldlm_proc_cleanup();
2836 # ifdef HAVE_SERVER_SUPPORT
2837 if (expired_lock_thread.elt_state != ELT_STOPPED) {
2838 expired_lock_thread.elt_state = ELT_TERMINATE;
2839 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
2840 cfs_wait_event(expired_lock_thread.elt_waitq,
2841 expired_lock_thread.elt_state == ELT_STOPPED);
2844 #endif /* __KERNEL__ */
2846 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2854 cfs_mutex_init(&ldlm_ref_mutex);
2855 cfs_mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
2856 cfs_mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
2857 ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
2858 sizeof(struct ldlm_resource), 0,
2859 CFS_SLAB_HWCACHE_ALIGN);
2860 if (ldlm_resource_slab == NULL)
2863 ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
2864 sizeof(struct ldlm_lock), 0,
2865 CFS_SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU);
2866 if (ldlm_lock_slab == NULL) {
2867 cfs_mem_cache_destroy(ldlm_resource_slab);
2871 ldlm_interval_slab = cfs_mem_cache_create("interval_node",
2872 sizeof(struct ldlm_interval),
2873 0, CFS_SLAB_HWCACHE_ALIGN);
2874 if (ldlm_interval_slab == NULL) {
2875 cfs_mem_cache_destroy(ldlm_resource_slab);
2876 cfs_mem_cache_destroy(ldlm_lock_slab);
2879 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2880 class_export_dump_hook = ldlm_dump_export_locks;
2885 void ldlm_exit(void)
2889 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
2890 rc = cfs_mem_cache_destroy(ldlm_resource_slab);
2891 LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
2893 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
2894 * synchronize_rcu() to wait a grace period elapsed, so that
2895 * ldlm_lock_free() get a chance to be called. */
2898 rc = cfs_mem_cache_destroy(ldlm_lock_slab);
2899 LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
2900 rc = cfs_mem_cache_destroy(ldlm_interval_slab);
2901 LASSERTF(rc == 0, "couldn't free interval node slab\n");
2905 EXPORT_SYMBOL(ldlm_extent_shift_kms);
2908 #ifdef HAVE_SERVER_SUPPORT
2909 EXPORT_SYMBOL(ldlm_get_processing_policy);
2911 EXPORT_SYMBOL(ldlm_lock2desc);
2912 EXPORT_SYMBOL(ldlm_register_intent);
2913 EXPORT_SYMBOL(ldlm_lockname);
2914 EXPORT_SYMBOL(ldlm_typename);
2915 EXPORT_SYMBOL(ldlm_lock2handle);
2916 EXPORT_SYMBOL(__ldlm_handle2lock);
2917 EXPORT_SYMBOL(ldlm_lock_get);
2918 EXPORT_SYMBOL(ldlm_lock_put);
2919 EXPORT_SYMBOL(ldlm_lock_match);
2920 EXPORT_SYMBOL(ldlm_lock_cancel);
2921 EXPORT_SYMBOL(ldlm_lock_addref);
2922 EXPORT_SYMBOL(ldlm_lock_addref_try);
2923 EXPORT_SYMBOL(ldlm_lock_decref);
2924 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
2925 EXPORT_SYMBOL(ldlm_lock_change_resource);
2926 EXPORT_SYMBOL(ldlm_it2str);
2927 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2928 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
2929 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
2930 EXPORT_SYMBOL(ldlm_lock_allow_match);
2931 EXPORT_SYMBOL(ldlm_lock_downgrade);
2932 EXPORT_SYMBOL(ldlm_lock_convert);
2934 /* ldlm_request.c */
2935 EXPORT_SYMBOL(ldlm_completion_ast_async);
2936 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
2937 EXPORT_SYMBOL(ldlm_completion_ast);
2938 EXPORT_SYMBOL(ldlm_blocking_ast);
2939 EXPORT_SYMBOL(ldlm_glimpse_ast);
2940 EXPORT_SYMBOL(ldlm_expired_completion_wait);
2941 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
2942 EXPORT_SYMBOL(ldlm_prep_elc_req);
2943 EXPORT_SYMBOL(ldlm_cli_convert);
2944 EXPORT_SYMBOL(ldlm_cli_enqueue);
2945 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
2946 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
2947 EXPORT_SYMBOL(ldlm_cli_cancel);
2948 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2949 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2950 EXPORT_SYMBOL(ldlm_cli_cancel_req);
2951 EXPORT_SYMBOL(ldlm_replay_locks);
2952 EXPORT_SYMBOL(ldlm_resource_foreach);
2953 EXPORT_SYMBOL(ldlm_namespace_foreach);
2954 EXPORT_SYMBOL(ldlm_resource_iterate);
2955 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2956 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
2957 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2960 #ifdef HAVE_SERVER_SUPPORT
2961 EXPORT_SYMBOL(ldlm_server_blocking_ast);
2962 EXPORT_SYMBOL(ldlm_server_completion_ast);
2963 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
2964 EXPORT_SYMBOL(ldlm_glimpse_locks);
2965 EXPORT_SYMBOL(ldlm_handle_enqueue);
2966 EXPORT_SYMBOL(ldlm_handle_enqueue0);
2967 EXPORT_SYMBOL(ldlm_handle_cancel);
2968 EXPORT_SYMBOL(ldlm_request_cancel);
2969 EXPORT_SYMBOL(ldlm_handle_convert);
2970 EXPORT_SYMBOL(ldlm_handle_convert0);
2971 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2973 EXPORT_SYMBOL(ldlm_del_waiting_lock);
2974 EXPORT_SYMBOL(ldlm_get_ref);
2975 EXPORT_SYMBOL(ldlm_put_ref);
2976 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
2978 /* ldlm_resource.c */
2979 EXPORT_SYMBOL(ldlm_namespace_new);
2980 EXPORT_SYMBOL(ldlm_namespace_cleanup);
2981 EXPORT_SYMBOL(ldlm_namespace_free);
2982 EXPORT_SYMBOL(ldlm_namespace_dump);
2983 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
2984 EXPORT_SYMBOL(ldlm_resource_get);
2985 EXPORT_SYMBOL(ldlm_resource_putref);
2986 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
2989 EXPORT_SYMBOL(client_import_add_conn);
2990 EXPORT_SYMBOL(client_import_del_conn);
2991 EXPORT_SYMBOL(client_obd_setup);
2992 EXPORT_SYMBOL(client_obd_cleanup);
2993 EXPORT_SYMBOL(client_connect_import);
2994 EXPORT_SYMBOL(client_disconnect_export);
2995 EXPORT_SYMBOL(target_send_reply);
2996 EXPORT_SYMBOL(target_pack_pool_reply);
2998 #ifdef HAVE_SERVER_SUPPORT
2999 EXPORT_SYMBOL(server_disconnect_export);
3000 EXPORT_SYMBOL(target_stop_recovery_thread);
3001 EXPORT_SYMBOL(target_handle_connect);
3002 EXPORT_SYMBOL(target_cleanup_recovery);
3003 EXPORT_SYMBOL(target_destroy_export);
3004 EXPORT_SYMBOL(target_cancel_recovery_timer);
3005 EXPORT_SYMBOL(target_queue_recovery_request);
3006 EXPORT_SYMBOL(target_handle_ping);
3007 EXPORT_SYMBOL(target_handle_disconnect);
3011 EXPORT_SYMBOL(lock_res_and_lock);
3012 EXPORT_SYMBOL(unlock_res_and_lock);