4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_lockd.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LDLM
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 struct mutex 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
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 struct completion 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 struct completion blwi_comp;
127 ldlm_cancel_flags_t blwi_flags;
128 int blwi_mem_pressure;
131 #if defined(HAVE_SERVER_SUPPORT) && defined(__KERNEL__)
134 * Protects both waiting_locks_list and expired_lock_thread.
136 static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */
139 * List for contended locks.
141 * As soon as a lock is contended, it gets placed on this list and
142 * expected time to get a response is filled in the lock. A special
143 * thread walks the list looking for locks that should be released and
144 * schedules client evictions for those that have not been released in
147 * All access to it should be under waiting_locks_spinlock.
149 static cfs_list_t waiting_locks_list;
150 static cfs_timer_t waiting_locks_timer;
152 static struct expired_lock_thread {
153 cfs_waitq_t elt_waitq;
156 cfs_list_t elt_expired_locks;
157 } expired_lock_thread;
159 static inline int have_expired_locks(void)
164 spin_lock_bh(&waiting_locks_spinlock);
165 need_to_run = !cfs_list_empty(&expired_lock_thread.elt_expired_locks);
166 spin_unlock_bh(&waiting_locks_spinlock);
172 * Check expired lock list for expired locks and time them out.
174 static int expired_lock_main(void *arg)
176 cfs_list_t *expired = &expired_lock_thread.elt_expired_locks;
177 struct l_wait_info lwi = { 0 };
181 cfs_daemonize("ldlm_elt");
183 expired_lock_thread.elt_state = ELT_READY;
184 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
187 l_wait_event(expired_lock_thread.elt_waitq,
188 have_expired_locks() ||
189 expired_lock_thread.elt_state == ELT_TERMINATE,
192 spin_lock_bh(&waiting_locks_spinlock);
193 if (expired_lock_thread.elt_dump) {
194 struct libcfs_debug_msg_data msgdata = {
195 .msg_file = __FILE__,
196 .msg_fn = "waiting_locks_callback",
197 .msg_line = expired_lock_thread.elt_dump };
198 spin_unlock_bh(&waiting_locks_spinlock);
200 /* from waiting_locks_callback, but not in timer */
201 libcfs_debug_dumplog();
202 libcfs_run_lbug_upcall(&msgdata);
204 spin_lock_bh(&waiting_locks_spinlock);
205 expired_lock_thread.elt_dump = 0;
210 while (!cfs_list_empty(expired)) {
211 struct obd_export *export;
212 struct ldlm_lock *lock;
214 lock = cfs_list_entry(expired->next, struct ldlm_lock,
216 if ((void *)lock < LP_POISON + CFS_PAGE_SIZE &&
217 (void *)lock >= LP_POISON) {
218 spin_unlock_bh(&waiting_locks_spinlock);
219 CERROR("free lock on elt list %p\n", lock);
222 cfs_list_del_init(&lock->l_pending_chain);
223 if ((void *)lock->l_export < LP_POISON + CFS_PAGE_SIZE &&
224 (void *)lock->l_export >= LP_POISON) {
225 CERROR("lock with free export on elt list %p\n",
227 lock->l_export = NULL;
228 LDLM_ERROR(lock, "free export");
229 /* release extra ref grabbed by
230 * ldlm_add_waiting_lock() or
231 * ldlm_failed_ast() */
232 LDLM_LOCK_RELEASE(lock);
236 if (lock->l_destroyed) {
237 /* release the lock refcount where
238 * waiting_locks_callback() founds */
239 LDLM_LOCK_RELEASE(lock);
242 export = class_export_lock_get(lock->l_export, lock);
243 spin_unlock_bh(&waiting_locks_spinlock);
246 class_fail_export(export);
247 class_export_lock_put(export, lock);
249 /* release extra ref grabbed by ldlm_add_waiting_lock()
250 * or ldlm_failed_ast() */
251 LDLM_LOCK_RELEASE(lock);
253 spin_lock_bh(&waiting_locks_spinlock);
255 spin_unlock_bh(&waiting_locks_spinlock);
257 if (do_dump && obd_dump_on_eviction) {
258 CERROR("dump the log upon eviction\n");
259 libcfs_debug_dumplog();
262 if (expired_lock_thread.elt_state == ELT_TERMINATE)
266 expired_lock_thread.elt_state = ELT_STOPPED;
267 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
271 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
272 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds);
275 * Check if there is a request in the export request list
276 * which prevents the lock canceling.
278 static int ldlm_lock_busy(struct ldlm_lock *lock)
280 struct ptlrpc_request *req;
284 if (lock->l_export == NULL)
287 spin_lock_bh(&lock->l_export->exp_rpc_lock);
288 cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
290 if (req->rq_ops->hpreq_lock_match) {
291 match = req->rq_ops->hpreq_lock_match(req, lock);
296 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
300 /* This is called from within a timer interrupt and cannot schedule */
301 static void waiting_locks_callback(unsigned long unused)
303 struct ldlm_lock *lock;
306 spin_lock_bh(&waiting_locks_spinlock);
307 while (!cfs_list_empty(&waiting_locks_list)) {
308 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
310 if (cfs_time_after(lock->l_callback_timeout,
311 cfs_time_current()) ||
312 (lock->l_req_mode == LCK_GROUP))
315 if (ptlrpc_check_suspend()) {
316 /* there is a case when we talk to one mds, holding
317 * lock from another mds. this way we easily can get
318 * here, if second mds is being recovered. so, we
319 * suspend timeouts. bug 6019 */
321 LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
322 lock->l_export->exp_client_uuid.uuid,
323 lock->l_export->exp_connection->c_remote_uuid.uuid,
324 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
326 cfs_list_del_init(&lock->l_pending_chain);
327 if (lock->l_destroyed) {
328 /* relay the lock refcount decrease to
329 * expired lock thread */
330 cfs_list_add(&lock->l_pending_chain,
331 &expired_lock_thread.elt_expired_locks);
333 __ldlm_add_waiting_lock(lock,
334 ldlm_get_enq_timeout(lock));
339 /* if timeout overlaps the activation time of suspended timeouts
340 * then extend it to give a chance for client to reconnect */
341 if (cfs_time_before(cfs_time_sub(lock->l_callback_timeout,
342 cfs_time_seconds(obd_timeout)/2),
343 ptlrpc_suspend_wakeup_time())) {
344 LDLM_ERROR(lock, "extend timeout due to recovery: %s@%s nid %s ",
345 lock->l_export->exp_client_uuid.uuid,
346 lock->l_export->exp_connection->c_remote_uuid.uuid,
347 libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
349 cfs_list_del_init(&lock->l_pending_chain);
350 if (lock->l_destroyed) {
351 /* relay the lock refcount decrease to
352 * expired lock thread */
353 cfs_list_add(&lock->l_pending_chain,
354 &expired_lock_thread.elt_expired_locks);
356 __ldlm_add_waiting_lock(lock,
357 ldlm_get_enq_timeout(lock));
362 /* Check if we need to prolong timeout */
363 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
364 ldlm_lock_busy(lock)) {
367 if (lock->l_pending_chain.next == &waiting_locks_list)
372 spin_unlock_bh(&waiting_locks_spinlock);
373 LDLM_DEBUG(lock, "prolong the busy lock");
374 ldlm_refresh_waiting_lock(lock,
375 ldlm_get_enq_timeout(lock));
376 spin_lock_bh(&waiting_locks_spinlock);
379 LDLM_LOCK_RELEASE(lock);
383 LDLM_LOCK_RELEASE(lock);
386 ldlm_lock_to_ns(lock)->ns_timeouts++;
387 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
388 "evicting client at %s ",
389 cfs_time_current_sec()- lock->l_last_activity,
391 lock->l_export->exp_connection->c_peer.nid));
393 /* no needs to take an extra ref on the lock since it was in
394 * the waiting_locks_list and ldlm_add_waiting_lock()
395 * already grabbed a ref */
396 cfs_list_del(&lock->l_pending_chain);
397 cfs_list_add(&lock->l_pending_chain,
398 &expired_lock_thread.elt_expired_locks);
402 if (!cfs_list_empty(&expired_lock_thread.elt_expired_locks)) {
403 if (obd_dump_on_timeout && need_dump)
404 expired_lock_thread.elt_dump = __LINE__;
406 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
410 * Make sure the timer will fire again if we have any locks
413 if (!cfs_list_empty(&waiting_locks_list)) {
414 cfs_time_t timeout_rounded;
415 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
417 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
418 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
420 spin_unlock_bh(&waiting_locks_spinlock);
424 * Add lock to the list of contended locks.
426 * Indicate that we're waiting for a client to call us back cancelling a given
427 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
428 * timer to fire appropriately. (We round up to the next second, to avoid
429 * floods of timer firings during periods of high lock contention and traffic).
430 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
431 * if it has been added to the waiting list (1 is returned).
433 * Called with the namespace lock held.
435 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
438 cfs_time_t timeout_rounded;
440 if (!cfs_list_empty(&lock->l_pending_chain))
443 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
444 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
447 timeout = cfs_time_shift(seconds);
448 if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
449 lock->l_callback_timeout = timeout;
451 timeout_rounded = round_timeout(lock->l_callback_timeout);
453 if (cfs_time_before(timeout_rounded,
454 cfs_timer_deadline(&waiting_locks_timer)) ||
455 !cfs_timer_is_armed(&waiting_locks_timer)) {
456 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
458 /* if the new lock has a shorter timeout than something earlier on
459 the list, we'll wait the longer amount of time; no big deal. */
461 cfs_list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
465 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
468 int timeout = ldlm_get_enq_timeout(lock);
470 /* NB: must be called with hold of lock_res_and_lock() */
471 LASSERT(lock->l_res_locked);
474 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
476 spin_lock_bh(&waiting_locks_spinlock);
477 if (lock->l_destroyed) {
478 static cfs_time_t next;
479 spin_unlock_bh(&waiting_locks_spinlock);
480 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
481 if (cfs_time_after(cfs_time_current(), next)) {
482 next = cfs_time_shift(14400);
483 libcfs_debug_dumpstack(NULL);
488 ret = __ldlm_add_waiting_lock(lock, timeout);
490 /* grab ref on the lock if it has been added to the
494 spin_unlock_bh(&waiting_locks_spinlock);
497 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
498 if (cfs_list_empty(&lock->l_exp_list))
499 cfs_list_add(&lock->l_exp_list,
500 &lock->l_export->exp_bl_list);
501 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
504 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
505 ret == 0 ? "not re-" : "", timeout,
506 AT_OFF ? "off" : "on");
511 * Remove a lock from the pending list, likely because it had its cancellation
512 * callback arrive without incident. This adjusts the lock-timeout timer if
513 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
514 * As done by ldlm_del_waiting_lock(), the caller must release the lock
515 * reference when the lock is removed from any list (1 is returned).
517 * Called with namespace lock held.
519 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
521 cfs_list_t *list_next;
523 if (cfs_list_empty(&lock->l_pending_chain))
526 list_next = lock->l_pending_chain.next;
527 if (lock->l_pending_chain.prev == &waiting_locks_list) {
528 /* Removing the head of the list, adjust timer. */
529 if (list_next == &waiting_locks_list) {
530 /* No more, just cancel. */
531 cfs_timer_disarm(&waiting_locks_timer);
533 struct ldlm_lock *next;
534 next = cfs_list_entry(list_next, struct ldlm_lock,
536 cfs_timer_arm(&waiting_locks_timer,
537 round_timeout(next->l_callback_timeout));
540 cfs_list_del_init(&lock->l_pending_chain);
545 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
549 if (lock->l_export == NULL) {
550 /* We don't have a "waiting locks list" on clients. */
551 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
555 spin_lock_bh(&waiting_locks_spinlock);
556 ret = __ldlm_del_waiting_lock(lock);
557 spin_unlock_bh(&waiting_locks_spinlock);
559 /* remove the lock out of export blocking list */
560 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
561 cfs_list_del_init(&lock->l_exp_list);
562 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
565 /* release lock ref if it has indeed been removed
567 LDLM_LOCK_RELEASE(lock);
570 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
573 EXPORT_SYMBOL(ldlm_del_waiting_lock);
576 * Prolong the contended lock waiting time.
578 * Called with namespace lock held.
580 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
582 if (lock->l_export == NULL) {
583 /* We don't have a "waiting locks list" on clients. */
584 LDLM_DEBUG(lock, "client lock: no-op");
588 spin_lock_bh(&waiting_locks_spinlock);
590 if (cfs_list_empty(&lock->l_pending_chain)) {
591 spin_unlock_bh(&waiting_locks_spinlock);
592 LDLM_DEBUG(lock, "wasn't waiting");
596 /* we remove/add the lock to the waiting list, so no needs to
597 * release/take a lock reference */
598 __ldlm_del_waiting_lock(lock);
599 __ldlm_add_waiting_lock(lock, timeout);
600 spin_unlock_bh(&waiting_locks_spinlock);
602 LDLM_DEBUG(lock, "refreshed");
605 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
607 #else /* !HAVE_SERVER_SUPPORT || !__KERNEL__ */
609 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
614 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
619 # ifdef HAVE_SERVER_SUPPORT
620 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
622 LASSERT(lock->l_res_locked);
623 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
628 #endif /* HAVE_SERVER_SUPPORT && __KERNEL__ */
630 #ifdef HAVE_SERVER_SUPPORT
633 * Perform lock cleanup if AST sending failed.
635 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
636 const char *ast_type)
638 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
639 "to a lock %s callback time out: rc %d\n",
640 lock->l_export->exp_obd->obd_name,
641 obd_export_nid2str(lock->l_export), ast_type, rc);
643 if (obd_dump_on_timeout)
644 libcfs_debug_dumplog();
646 spin_lock_bh(&waiting_locks_spinlock);
647 if (__ldlm_del_waiting_lock(lock) == 0)
648 /* the lock was not in any list, grab an extra ref before adding
649 * the lock to the expired list */
651 cfs_list_add(&lock->l_pending_chain,
652 &expired_lock_thread.elt_expired_locks);
653 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
654 spin_unlock_bh(&waiting_locks_spinlock);
656 class_fail_export(lock->l_export);
661 * Perform lock cleanup if AST reply came with error.
663 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
664 struct ptlrpc_request *req, int rc,
665 const char *ast_type)
667 lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
669 if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
670 LASSERT(lock->l_export);
671 if (lock->l_export->exp_libclient) {
672 LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
673 " timeout, just cancelling lock", ast_type,
674 libcfs_nid2str(peer.nid));
675 ldlm_lock_cancel(lock);
677 } else if (lock->l_flags & LDLM_FL_CANCEL) {
678 LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
679 "cancel was received (AST reply lost?)",
680 ast_type, libcfs_nid2str(peer.nid));
681 ldlm_lock_cancel(lock);
684 ldlm_del_waiting_lock(lock);
685 ldlm_failed_ast(lock, rc, ast_type);
689 struct ldlm_resource *res = lock->l_resource;
690 LDLM_DEBUG(lock, "client (nid %s) returned %d"
691 " from %s AST - normal race",
692 libcfs_nid2str(peer.nid),
694 lustre_msg_get_status(req->rq_repmsg) : -1,
697 /* update lvbo to return proper attributes.
699 ldlm_resource_getref(res);
700 ldlm_res_lvbo_update(res, NULL, 1);
701 ldlm_resource_putref(res);
705 LDLM_ERROR(lock, "client (nid %s) returned %d "
706 "from %s AST", libcfs_nid2str(peer.nid),
707 (req->rq_repmsg != NULL) ?
708 lustre_msg_get_status(req->rq_repmsg) : 0,
711 ldlm_lock_cancel(lock);
712 /* Server-side AST functions are called from ldlm_reprocess_all,
713 * which needs to be told to please restart its reprocessing. */
720 static int ldlm_cb_interpret(const struct lu_env *env,
721 struct ptlrpc_request *req, void *data, int rc)
723 struct ldlm_cb_async_args *ca = data;
724 struct ldlm_lock *lock = ca->ca_lock;
725 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
728 LASSERT(lock != NULL);
731 case LDLM_GL_CALLBACK:
732 /* Update the LVB from disk if the AST failed
733 * (this is a legal race)
735 * - Glimpse callback of local lock just returns
736 * -ELDLM_NO_LOCK_DATA.
737 * - Glimpse callback of remote lock might return
738 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
740 if (rc == -ELDLM_NO_LOCK_DATA) {
741 LDLM_DEBUG(lock, "lost race - client has a lock but no "
743 ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
744 } else if (rc != 0) {
745 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
747 rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
750 case LDLM_BL_CALLBACK:
752 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
754 case LDLM_CP_CALLBACK:
756 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
759 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
764 /* release extra reference taken in ldlm_ast_fini() */
765 LDLM_LOCK_RELEASE(lock);
768 cfs_atomic_inc(&arg->restart);
773 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
774 struct ldlm_cb_set_arg *arg,
775 struct ldlm_lock *lock,
781 if (unlikely(instant_cancel)) {
782 rc = ptl_send_rpc(req, 1);
783 ptlrpc_req_finished(req);
785 cfs_atomic_inc(&arg->restart);
788 ptlrpc_set_add_req(arg->set, req);
795 * Check if there are requests in the export request list which prevent
796 * the lock canceling and make these requests high priority ones.
798 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
800 struct ptlrpc_request *req;
803 if (lock->l_export == NULL) {
804 LDLM_DEBUG(lock, "client lock: no-op");
808 spin_lock_bh(&lock->l_export->exp_rpc_lock);
809 cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
811 /* Do not process requests that were not yet added to there
812 * incoming queue or were already removed from there for
813 * processing. We evaluate ptlrpc_request_reorderable() without
814 * holding svcpt->scp_req_lock, and then redo the checks with
815 * the lock held once we need to obtain a reliable result.
817 if (ptlrpc_nrs_req_can_move(req) &&
818 req->rq_ops->hpreq_lock_match &&
819 req->rq_ops->hpreq_lock_match(req, lock))
820 ptlrpc_nrs_req_hp_move(req);
822 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
827 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
828 * enqueued server lock conflicts with given one.
830 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
831 * to wait for client response.
833 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
834 struct ldlm_lock_desc *desc,
835 void *data, int flag)
837 struct ldlm_cb_async_args *ca;
838 struct ldlm_cb_set_arg *arg = data;
839 struct ldlm_request *body;
840 struct ptlrpc_request *req;
841 int instant_cancel = 0;
845 if (flag == LDLM_CB_CANCELING)
846 /* Don't need to do anything here. */
850 LASSERT(data != NULL);
851 if (lock->l_export->exp_obd->obd_recovering != 0)
852 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
854 ldlm_lock_reorder_req(lock);
856 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
857 &RQF_LDLM_BL_CALLBACK,
858 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
862 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
863 ca = ptlrpc_req_async_args(req);
864 ca->ca_set_arg = arg;
867 req->rq_interpret_reply = ldlm_cb_interpret;
868 req->rq_no_resend = 1;
870 lock_res_and_lock(lock);
871 if (lock->l_granted_mode != lock->l_req_mode) {
872 /* this blocking AST will be communicated as part of the
873 * completion AST instead */
874 unlock_res_and_lock(lock);
876 ptlrpc_req_finished(req);
877 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
881 if (lock->l_destroyed) {
882 /* What's the point? */
883 unlock_res_and_lock(lock);
884 ptlrpc_req_finished(req);
888 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
891 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
892 body->lock_handle[0] = lock->l_remote_handle;
893 body->lock_desc = *desc;
894 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_AST_FLAGS);
896 LDLM_DEBUG(lock, "server preparing blocking AST");
898 ptlrpc_request_set_replen(req);
899 if (instant_cancel) {
900 unlock_res_and_lock(lock);
901 ldlm_lock_cancel(lock);
903 LASSERT(lock->l_granted_mode == lock->l_req_mode);
904 ldlm_add_waiting_lock(lock);
905 unlock_res_and_lock(lock);
908 req->rq_send_state = LUSTRE_IMP_FULL;
909 /* ptlrpc_request_alloc_pack already set timeout */
911 req->rq_timeout = ldlm_get_rq_timeout();
913 if (lock->l_export && lock->l_export->exp_nid_stats &&
914 lock->l_export->exp_nid_stats->nid_ldlm_stats)
915 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
916 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
918 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
922 EXPORT_SYMBOL(ldlm_server_blocking_ast);
925 * ->l_completion_ast callback for a remote lock in server namespace.
927 * Sends AST to the client notifying it of lock granting. If initial
928 * lock response was not sent yet, instead of sending another RPC, just
929 * mark the lock as granted and client will understand
931 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
933 struct ldlm_cb_set_arg *arg = data;
934 struct ldlm_request *body;
935 struct ptlrpc_request *req;
936 struct ldlm_cb_async_args *ca;
937 long total_enqueue_wait;
938 int instant_cancel = 0;
943 LASSERT(lock != NULL);
944 LASSERT(data != NULL);
946 total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(),
947 lock->l_last_activity);
949 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
950 &RQF_LDLM_CP_CALLBACK);
954 /* server namespace, doesn't need lock */
955 lvb_len = ldlm_lvbo_size(lock);
956 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
957 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
959 ptlrpc_request_free(req);
963 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
964 ca = ptlrpc_req_async_args(req);
965 ca->ca_set_arg = arg;
968 req->rq_interpret_reply = ldlm_cb_interpret;
969 req->rq_no_resend = 1;
970 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
972 body->lock_handle[0] = lock->l_remote_handle;
973 body->lock_flags = ldlm_flags_to_wire(flags);
974 ldlm_lock2desc(lock, &body->lock_desc);
976 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
978 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
980 /* We still need to send the RPC to wake up the blocked
981 * enqueue thread on the client.
983 * Consider old client, there is no better way to notify
984 * the failure, just zero-sized the LVB, then the client
985 * will fail out as "-EPROTO". */
986 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
990 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
995 LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)",
998 /* Server-side enqueue wait time estimate, used in
999 __ldlm_add_waiting_lock to set future enqueue timers */
1000 if (total_enqueue_wait < ldlm_get_enq_timeout(lock))
1001 at_measured(ldlm_lock_to_ns_at(lock),
1002 total_enqueue_wait);
1004 /* bz18618. Don't add lock enqueue time we spend waiting for a
1005 previous callback to fail. Locks waiting legitimately will
1006 get extended by ldlm_refresh_waiting_lock regardless of the
1007 estimate, so it's okay to underestimate here. */
1008 LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. "
1009 "It is likely that a previous callback timed out.",
1011 at_get(ldlm_lock_to_ns_at(lock)));
1013 ptlrpc_request_set_replen(req);
1015 req->rq_send_state = LUSTRE_IMP_FULL;
1016 /* ptlrpc_request_pack already set timeout */
1018 req->rq_timeout = ldlm_get_rq_timeout();
1020 /* We only send real blocking ASTs after the lock is granted */
1021 lock_res_and_lock(lock);
1022 if (lock->l_flags & LDLM_FL_AST_SENT) {
1023 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1024 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1025 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1028 /* We might get here prior to ldlm_handle_enqueue setting
1029 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1030 * into waiting list, but this is safe and similar code in
1031 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1032 * that would not only cancel the lock, but will also remove
1033 * it from waiting list */
1034 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1035 unlock_res_and_lock(lock);
1036 ldlm_lock_cancel(lock);
1038 lock_res_and_lock(lock);
1040 /* start the lock-timeout clock */
1041 ldlm_add_waiting_lock(lock);
1044 unlock_res_and_lock(lock);
1046 if (lock->l_export && lock->l_export->exp_nid_stats &&
1047 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1048 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1049 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1051 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1053 RETURN(lvb_len < 0 ? lvb_len : rc);
1055 EXPORT_SYMBOL(ldlm_server_completion_ast);
1058 * Server side ->l_glimpse_ast handler for client locks.
1060 * Sends glimpse AST to the client and waits for reply. Then updates
1061 * lvbo with the result.
1063 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1065 struct ldlm_cb_set_arg *arg = data;
1066 struct ldlm_request *body;
1067 struct ptlrpc_request *req;
1068 struct ldlm_cb_async_args *ca;
1070 struct req_format *req_fmt;
1073 LASSERT(lock != NULL);
1075 if (arg->gl_desc != NULL)
1076 /* There is a glimpse descriptor to pack */
1077 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1079 req_fmt = &RQF_LDLM_GL_CALLBACK;
1081 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1082 req_fmt, LUSTRE_DLM_VERSION,
1088 if (arg->gl_desc != NULL) {
1089 /* copy the GL descriptor */
1090 union ldlm_gl_desc *desc;
1091 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1092 *desc = *arg->gl_desc;
1095 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1096 body->lock_handle[0] = lock->l_remote_handle;
1097 ldlm_lock2desc(lock, &body->lock_desc);
1099 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1100 ca = ptlrpc_req_async_args(req);
1101 ca->ca_set_arg = arg;
1104 /* server namespace, doesn't need lock */
1105 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1106 ldlm_lvbo_size(lock));
1107 ptlrpc_request_set_replen(req);
1109 req->rq_send_state = LUSTRE_IMP_FULL;
1110 /* ptlrpc_request_alloc_pack already set timeout */
1112 req->rq_timeout = ldlm_get_rq_timeout();
1114 req->rq_interpret_reply = ldlm_cb_interpret;
1116 if (lock->l_export && lock->l_export->exp_nid_stats &&
1117 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1118 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1119 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1121 rc = ldlm_ast_fini(req, arg, lock, 0);
1125 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1127 int ldlm_glimpse_locks(struct ldlm_resource *res, cfs_list_t *gl_work_list)
1132 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1134 if (rc == -ERESTART)
1135 ldlm_reprocess_all(res);
1139 EXPORT_SYMBOL(ldlm_glimpse_locks);
1141 /* return LDLM lock associated with a lock callback request */
1142 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1144 struct ldlm_cb_async_args *ca;
1145 struct ldlm_lock *lock;
1148 ca = ptlrpc_req_async_args(req);
1151 RETURN(ERR_PTR(-EFAULT));
1155 EXPORT_SYMBOL(ldlm_request_lock);
1157 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1158 struct lprocfs_stats *srv_stats)
1160 int lock_type = 0, op = 0;
1162 lock_type = dlm_req->lock_desc.l_resource.lr_type;
1164 switch (lock_type) {
1166 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1169 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1170 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1172 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1175 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1178 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1186 lprocfs_counter_incr(srv_stats, op);
1192 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1193 * service threads to carry out client lock enqueueing requests.
1195 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1196 struct ptlrpc_request *req,
1197 const struct ldlm_request *dlm_req,
1198 const struct ldlm_callback_suite *cbs)
1200 struct ldlm_reply *dlm_rep;
1202 ldlm_error_t err = ELDLM_OK;
1203 struct ldlm_lock *lock = NULL;
1204 void *cookie = NULL;
1208 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1210 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1211 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1213 LASSERT(req->rq_export);
1215 if (ptlrpc_req2svc(req)->srv_stats != NULL)
1216 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1218 if (req->rq_export && req->rq_export->exp_nid_stats &&
1219 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1220 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1221 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1223 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1224 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1225 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1226 dlm_req->lock_desc.l_resource.lr_type);
1227 GOTO(out, rc = -EFAULT);
1230 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1231 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1232 dlm_req->lock_desc.l_req_mode &
1233 (dlm_req->lock_desc.l_req_mode-1))) {
1234 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1235 dlm_req->lock_desc.l_req_mode);
1236 GOTO(out, rc = -EFAULT);
1239 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) {
1240 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1242 DEBUG_REQ(D_ERROR, req,
1243 "PLAIN lock request from IBITS client?");
1244 GOTO(out, rc = -EPROTO);
1246 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1248 DEBUG_REQ(D_ERROR, req,
1249 "IBITS lock request from unaware client?");
1250 GOTO(out, rc = -EPROTO);
1254 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1255 against server's _CONNECT_SUPPORTED flags? (I don't want to use
1256 ibits for mgc/mgs) */
1258 /* INODEBITS_INTEROP: Perform conversion from plain lock to
1259 * inodebits lock if client does not support them. */
1260 if (!(exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) &&
1261 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1262 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1263 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1264 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1265 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1266 dlm_req->lock_desc.l_req_mode = LCK_CR;
1270 if (unlikely(flags & LDLM_FL_REPLAY)) {
1271 /* Find an existing lock in the per-export lock hash */
1272 /* In the function below, .hs_keycmp resolves to
1273 * ldlm_export_lock_keycmp() */
1274 /* coverity[overrun-buffer-val] */
1275 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1276 (void *)&dlm_req->lock_handle[0]);
1278 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1279 LPX64, lock->l_handle.h_cookie);
1280 GOTO(existing_lock, rc = 0);
1284 /* The lock's callback data might be set in the policy function */
1285 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1286 dlm_req->lock_desc.l_resource.lr_type,
1287 dlm_req->lock_desc.l_req_mode,
1288 cbs, NULL, 0, LVB_T_NONE);
1290 GOTO(out, rc = -ENOMEM);
1292 lock->l_last_activity = cfs_time_current_sec();
1293 lock->l_remote_handle = dlm_req->lock_handle[0];
1294 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1296 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1297 /* Don't enqueue a lock onto the export if it is been disonnected
1298 * due to eviction (bug 3822) or server umount (bug 24324).
1299 * Cancel it now instead. */
1300 if (req->rq_export->exp_disconnected) {
1301 LDLM_ERROR(lock, "lock on disconnected export %p",
1303 GOTO(out, rc = -ENOTCONN);
1306 lock->l_export = class_export_lock_get(req->rq_export, lock);
1307 if (lock->l_export->exp_lock_hash)
1308 cfs_hash_add(lock->l_export->exp_lock_hash,
1309 &lock->l_remote_handle,
1314 if (flags & LDLM_FL_HAS_INTENT) {
1315 /* In this case, the reply buffer is allocated deep in
1316 * local_lock_enqueue by the policy function. */
1319 /* based on the assumption that lvb size never changes during
1320 * resource life time otherwise it need resource->lr_lock's
1322 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1323 RCL_SERVER, ldlm_lvbo_size(lock));
1325 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1326 GOTO(out, rc = -ENOMEM);
1328 rc = req_capsule_server_pack(&req->rq_pill);
1333 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1334 ldlm_convert_policy_to_local(req->rq_export,
1335 dlm_req->lock_desc.l_resource.lr_type,
1336 &dlm_req->lock_desc.l_policy_data,
1337 &lock->l_policy_data);
1338 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1339 lock->l_req_extent = lock->l_policy_data.l_extent;
1341 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1345 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1346 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1348 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1349 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1351 /* We never send a blocking AST until the lock is granted, but
1352 * we can tell it right now */
1353 lock_res_and_lock(lock);
1355 /* Now take into account flags to be inherited from original lock
1356 request both in reply to client and in our own lock flags. */
1357 dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1358 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1359 LDLM_INHERIT_FLAGS);
1361 /* Don't move a pending lock onto the export if it has already been
1362 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1363 * Cancel it now instead. */
1364 if (unlikely(req->rq_export->exp_disconnected ||
1365 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1366 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1368 } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1369 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1370 if (lock->l_granted_mode == lock->l_req_mode) {
1372 * Only cancel lock if it was granted, because it would
1373 * be destroyed immediately and would never be granted
1374 * in the future, causing timeouts on client. Not
1375 * granted lock will be cancelled immediately after
1376 * sending completion AST.
1378 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1379 unlock_res_and_lock(lock);
1380 ldlm_lock_cancel(lock);
1381 lock_res_and_lock(lock);
1383 ldlm_add_waiting_lock(lock);
1386 /* Make sure we never ever grant usual metadata locks to liblustre
1388 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1389 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1390 req->rq_export->exp_libclient) {
1391 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1392 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1393 CERROR("Granting sync lock to libclient. "
1394 "req fl %d, rep fl %d, lock fl "LPX64"\n",
1395 dlm_req->lock_flags, dlm_rep->lock_flags,
1397 LDLM_ERROR(lock, "sync lock");
1398 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1399 struct ldlm_intent *it;
1401 it = req_capsule_client_get(&req->rq_pill,
1404 CERROR("This is intent %s ("LPU64")\n",
1405 ldlm_it2str(it->opc), it->opc);
1411 unlock_res_and_lock(lock);
1415 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1416 if (!req->rq_packed_final) {
1417 err = lustre_pack_reply(req, 1, NULL, NULL);
1422 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1423 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1425 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1426 "(err=%d, rc=%d)", err, rc);
1429 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1431 ldlm_lvbo_size(lock) > 0) {
1435 buf = req_capsule_server_get(&req->rq_pill,
1437 LASSERTF(buf != NULL, "req %p, lock %p\n",
1439 buflen = req_capsule_get_size(&req->rq_pill,
1440 &RMF_DLM_LVB, RCL_SERVER);
1441 buflen = ldlm_lvbo_fill(lock, buf, buflen);
1443 req_capsule_shrink(&req->rq_pill,
1445 buflen, RCL_SERVER);
1450 lock_res_and_lock(lock);
1451 ldlm_resource_unlink_lock(lock);
1452 ldlm_lock_destroy_nolock(lock);
1453 unlock_res_and_lock(lock);
1456 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1457 ldlm_reprocess_all(lock->l_resource);
1459 LDLM_LOCK_RELEASE(lock);
1462 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1467 EXPORT_SYMBOL(ldlm_handle_enqueue0);
1470 * Old-style LDLM main entry point for server code enqueue.
1472 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1473 ldlm_completion_callback completion_callback,
1474 ldlm_blocking_callback blocking_callback,
1475 ldlm_glimpse_callback glimpse_callback)
1477 struct ldlm_request *dlm_req;
1478 struct ldlm_callback_suite cbs = {
1479 .lcs_completion = completion_callback,
1480 .lcs_blocking = blocking_callback,
1481 .lcs_glimpse = glimpse_callback
1485 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1486 if (dlm_req != NULL) {
1487 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1488 req, dlm_req, &cbs);
1494 EXPORT_SYMBOL(ldlm_handle_enqueue);
1497 * Main LDLM entry point for server code to process lock conversion requests.
1499 int ldlm_handle_convert0(struct ptlrpc_request *req,
1500 const struct ldlm_request *dlm_req)
1502 struct ldlm_reply *dlm_rep;
1503 struct ldlm_lock *lock;
1507 if (req->rq_export && req->rq_export->exp_nid_stats &&
1508 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1509 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1510 LDLM_CONVERT - LDLM_FIRST_OPC);
1512 rc = req_capsule_server_pack(&req->rq_pill);
1516 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1517 dlm_rep->lock_flags = dlm_req->lock_flags;
1519 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1521 req->rq_status = EINVAL;
1525 LDLM_DEBUG(lock, "server-side convert handler START");
1527 lock->l_last_activity = cfs_time_current_sec();
1528 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1529 &dlm_rep->lock_flags);
1531 if (ldlm_del_waiting_lock(lock))
1532 LDLM_DEBUG(lock, "converted waiting lock");
1535 req->rq_status = EDEADLOCK;
1540 if (!req->rq_status)
1541 ldlm_reprocess_all(lock->l_resource);
1542 LDLM_DEBUG(lock, "server-side convert handler END");
1543 LDLM_LOCK_PUT(lock);
1545 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1549 EXPORT_SYMBOL(ldlm_handle_convert0);
1552 * Old-style main LDLM entry point for server code to process lock conversion
1555 int ldlm_handle_convert(struct ptlrpc_request *req)
1558 struct ldlm_request *dlm_req;
1560 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1561 if (dlm_req != NULL) {
1562 rc = ldlm_handle_convert0(req, dlm_req);
1564 CERROR ("Can't unpack dlm_req\n");
1569 EXPORT_SYMBOL(ldlm_handle_convert);
1572 * Cancel all the locks whose handles are packed into ldlm_request
1574 * Called by server code expecting such combined cancel activity
1577 int ldlm_request_cancel(struct ptlrpc_request *req,
1578 const struct ldlm_request *dlm_req, int first)
1580 struct ldlm_resource *res, *pres = NULL;
1581 struct ldlm_lock *lock;
1582 int i, count, done = 0;
1585 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1589 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1592 /* There is no lock on the server at the replay time,
1593 * skip lock cancelling to make replay tests to pass. */
1594 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1597 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1598 "starting at %d", count, first);
1600 for (i = first; i < count; i++) {
1601 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1603 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1604 "lock (cookie "LPU64")",
1605 dlm_req->lock_handle[i].cookie);
1609 res = lock->l_resource;
1612 /* This code is an optimization to only attempt lock
1613 * granting on the resource (that could be CPU-expensive)
1614 * after we are done cancelling lock in that resource. */
1617 ldlm_reprocess_all(pres);
1618 LDLM_RESOURCE_DELREF(pres);
1619 ldlm_resource_putref(pres);
1622 ldlm_resource_getref(res);
1623 LDLM_RESOURCE_ADDREF(res);
1624 ldlm_res_lvbo_update(res, NULL, 1);
1628 ldlm_lock_cancel(lock);
1629 LDLM_LOCK_PUT(lock);
1632 ldlm_reprocess_all(pres);
1633 LDLM_RESOURCE_DELREF(pres);
1634 ldlm_resource_putref(pres);
1636 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1639 EXPORT_SYMBOL(ldlm_request_cancel);
1642 * Main LDLM entry point for server code to cancel locks.
1644 * Typically gets called from service handler on LDLM_CANCEL opc.
1646 int ldlm_handle_cancel(struct ptlrpc_request *req)
1648 struct ldlm_request *dlm_req;
1652 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1653 if (dlm_req == NULL) {
1654 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1658 if (req->rq_export && req->rq_export->exp_nid_stats &&
1659 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1660 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1661 LDLM_CANCEL - LDLM_FIRST_OPC);
1663 rc = req_capsule_server_pack(&req->rq_pill);
1667 if (!ldlm_request_cancel(req, dlm_req, 0))
1668 req->rq_status = ESTALE;
1670 RETURN(ptlrpc_reply(req));
1672 EXPORT_SYMBOL(ldlm_handle_cancel);
1673 #endif /* HAVE_SERVER_SUPPORT */
1676 * Callback handler for receiving incoming blocking ASTs.
1678 * This can only happen on client side.
1680 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1681 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1686 LDLM_DEBUG(lock, "client blocking AST callback handler");
1688 lock_res_and_lock(lock);
1689 lock->l_flags |= LDLM_FL_CBPENDING;
1691 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1692 lock->l_flags |= LDLM_FL_CANCEL;
1694 do_ast = (!lock->l_readers && !lock->l_writers);
1695 unlock_res_and_lock(lock);
1698 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1699 lock, lock->l_blocking_ast);
1700 if (lock->l_blocking_ast != NULL)
1701 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1704 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1708 LDLM_DEBUG(lock, "client blocking callback handler END");
1709 LDLM_LOCK_RELEASE(lock);
1714 * Callback handler for receiving incoming completion ASTs.
1716 * This only can happen on client side.
1718 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1719 struct ldlm_namespace *ns,
1720 struct ldlm_request *dlm_req,
1721 struct ldlm_lock *lock)
1724 CFS_LIST_HEAD(ast_list);
1728 LDLM_DEBUG(lock, "client completion callback handler START");
1730 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1731 int to = cfs_time_seconds(1);
1733 cfs_schedule_timeout_and_set_state(
1734 CFS_TASK_INTERRUPTIBLE, to);
1735 if (lock->l_granted_mode == lock->l_req_mode ||
1741 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1743 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1744 GOTO(out, rc = lvb_len);
1745 } else if (lvb_len > 0) {
1746 if (lock->l_lvb_len > 0) {
1747 /* for extent lock, lvb contains ost_lvb{}. */
1748 LASSERT(lock->l_lvb_data != NULL);
1750 if (unlikely(lock->l_lvb_len < lvb_len)) {
1751 LDLM_ERROR(lock, "Replied LVB is larger than "
1752 "expectation, expected = %d, "
1754 lock->l_lvb_len, lvb_len);
1755 GOTO(out, rc = -EINVAL);
1757 } else if (ldlm_has_layout(lock)) { /* for layout lock, lvb has
1758 * variable length */
1761 OBD_ALLOC(lvb_data, lvb_len);
1762 if (lvb_data == NULL) {
1763 LDLM_ERROR(lock, "No memory: %d.\n", lvb_len);
1764 GOTO(out, rc = -ENOMEM);
1767 lock_res_and_lock(lock);
1768 LASSERT(lock->l_lvb_data == NULL);
1769 lock->l_lvb_data = lvb_data;
1770 lock->l_lvb_len = lvb_len;
1771 unlock_res_and_lock(lock);
1775 lock_res_and_lock(lock);
1776 if (lock->l_destroyed ||
1777 lock->l_granted_mode == lock->l_req_mode) {
1778 /* bug 11300: the lock has already been granted */
1779 unlock_res_and_lock(lock);
1780 LDLM_DEBUG(lock, "Double grant race happened");
1784 /* If we receive the completion AST before the actual enqueue returned,
1785 * then we might need to switch lock modes, resources, or extents. */
1786 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1787 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1788 LDLM_DEBUG(lock, "completion AST, new lock mode");
1791 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1792 ldlm_convert_policy_to_local(req->rq_export,
1793 dlm_req->lock_desc.l_resource.lr_type,
1794 &dlm_req->lock_desc.l_policy_data,
1795 &lock->l_policy_data);
1796 LDLM_DEBUG(lock, "completion AST, new policy data");
1799 ldlm_resource_unlink_lock(lock);
1800 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1801 &lock->l_resource->lr_name,
1802 sizeof(lock->l_resource->lr_name)) != 0) {
1803 unlock_res_and_lock(lock);
1804 rc = ldlm_lock_change_resource(ns, lock,
1805 &dlm_req->lock_desc.l_resource.lr_name);
1807 LDLM_ERROR(lock, "Failed to allocate resource");
1810 LDLM_DEBUG(lock, "completion AST, new resource");
1811 CERROR("change resource!\n");
1812 lock_res_and_lock(lock);
1815 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1816 /* BL_AST locks are not needed in LRU.
1817 * Let ldlm_cancel_lru() be fast. */
1818 ldlm_lock_remove_from_lru(lock);
1819 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1820 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1823 if (lock->l_lvb_len > 0) {
1824 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1825 lock->l_lvb_data, lvb_len);
1827 unlock_res_and_lock(lock);
1832 ldlm_grant_lock(lock, &ast_list);
1833 unlock_res_and_lock(lock);
1835 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1837 /* Let Enqueue to call osc_lock_upcall() and initialize
1839 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1841 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1843 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1849 lock_res_and_lock(lock);
1850 lock->l_flags |= LDLM_FL_FAILED;
1851 unlock_res_and_lock(lock);
1852 cfs_waitq_signal(&lock->l_waitq);
1854 LDLM_LOCK_RELEASE(lock);
1858 * Callback handler for receiving incoming glimpse ASTs.
1860 * This only can happen on client side. After handling the glimpse AST
1861 * we also consider dropping the lock here if it is unused locally for a
1864 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1865 struct ldlm_namespace *ns,
1866 struct ldlm_request *dlm_req,
1867 struct ldlm_lock *lock)
1872 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1874 if (lock->l_glimpse_ast != NULL)
1875 rc = lock->l_glimpse_ast(lock, req);
1877 if (req->rq_repmsg != NULL) {
1880 req->rq_status = rc;
1884 lock_res_and_lock(lock);
1885 if (lock->l_granted_mode == LCK_PW &&
1886 !lock->l_readers && !lock->l_writers &&
1887 cfs_time_after(cfs_time_current(),
1888 cfs_time_add(lock->l_last_used,
1889 cfs_time_seconds(10)))) {
1890 unlock_res_and_lock(lock);
1891 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1892 ldlm_handle_bl_callback(ns, NULL, lock);
1897 unlock_res_and_lock(lock);
1898 LDLM_LOCK_RELEASE(lock);
1902 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1904 if (req->rq_no_reply)
1907 req->rq_status = rc;
1908 if (!req->rq_packed_final) {
1909 rc = lustre_pack_reply(req, 1, NULL, NULL);
1913 return ptlrpc_reply(req);
1917 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1918 ldlm_cancel_flags_t cancel_flags)
1920 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1923 spin_lock(&blp->blp_lock);
1924 if (blwi->blwi_lock &&
1925 blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) {
1926 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1927 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1929 /* other blocking callbacks are added to the regular list */
1930 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1932 spin_unlock(&blp->blp_lock);
1934 cfs_waitq_signal(&blp->blp_waitq);
1936 /* can not check blwi->blwi_flags as blwi could be already freed in
1938 if (!(cancel_flags & LCF_ASYNC))
1939 wait_for_completion(&blwi->blwi_comp);
1944 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1945 struct ldlm_namespace *ns,
1946 struct ldlm_lock_desc *ld,
1947 cfs_list_t *cancels, int count,
1948 struct ldlm_lock *lock,
1949 ldlm_cancel_flags_t cancel_flags)
1951 init_completion(&blwi->blwi_comp);
1952 CFS_INIT_LIST_HEAD(&blwi->blwi_head);
1954 if (cfs_memory_pressure_get())
1955 blwi->blwi_mem_pressure = 1;
1958 blwi->blwi_flags = cancel_flags;
1960 blwi->blwi_ld = *ld;
1962 cfs_list_add(&blwi->blwi_head, cancels);
1963 cfs_list_del_init(cancels);
1964 blwi->blwi_count = count;
1966 blwi->blwi_lock = lock;
1971 * Queues a list of locks \a cancels containing \a count locks
1972 * for later processing by a blocking thread. If \a count is zero,
1973 * then the lock referenced as \a lock is queued instead.
1975 * The blocking thread would then call ->l_blocking_ast callback in the lock.
1976 * If list addition fails an error is returned and caller is supposed to
1977 * call ->l_blocking_ast itself.
1979 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1980 struct ldlm_lock_desc *ld,
1981 struct ldlm_lock *lock,
1982 cfs_list_t *cancels, int count,
1983 ldlm_cancel_flags_t cancel_flags)
1987 if (cancels && count == 0)
1990 if (cancel_flags & LCF_ASYNC) {
1991 struct ldlm_bl_work_item *blwi;
1993 OBD_ALLOC(blwi, sizeof(*blwi));
1996 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
1998 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2000 /* if it is synchronous call do minimum mem alloc, as it could
2001 * be triggered from kernel shrinker
2003 struct ldlm_bl_work_item blwi;
2005 memset(&blwi, 0, sizeof(blwi));
2006 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2007 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2013 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2014 struct ldlm_lock *lock)
2017 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2023 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2024 cfs_list_t *cancels, int count,
2025 ldlm_cancel_flags_t cancel_flags)
2028 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2034 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2035 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2037 struct obd_device *obd = req->rq_export->exp_obd;
2044 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2046 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2048 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2050 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2053 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2055 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2057 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2060 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2063 /* We are responsible for swabbing contents of val */
2065 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2066 /* Pass it on to mdc (the "export" in this case) */
2067 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2069 sizeof(KEY_HSM_COPYTOOL_SEND),
2070 KEY_HSM_COPYTOOL_SEND,
2073 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2078 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2079 const char *msg, int rc,
2080 struct lustre_handle *handle)
2082 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2083 "%s: [nid %s] [rc %d] [lock "LPX64"]",
2084 msg, libcfs_id2str(req->rq_peer), rc,
2085 handle ? handle->cookie : 0);
2086 if (req->rq_no_reply)
2087 CWARN("No reply was sent, maybe cause bug 21636.\n");
2089 CWARN("Send reply failed, maybe cause bug 21636.\n");
2092 static int ldlm_handle_qc_callback(struct ptlrpc_request *req)
2094 struct obd_quotactl *oqctl;
2095 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2097 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2098 if (oqctl == NULL) {
2099 CERROR("Can't unpack obd_quotactl\n");
2103 cli->cl_qchk_stat = oqctl->qc_stat;
2107 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2108 static int ldlm_callback_handler(struct ptlrpc_request *req)
2110 struct ldlm_namespace *ns;
2111 struct ldlm_request *dlm_req;
2112 struct ldlm_lock *lock;
2116 /* Requests arrive in sender's byte order. The ptlrpc service
2117 * handler has already checked and, if necessary, byte-swapped the
2118 * incoming request message body, but I am responsible for the
2119 * message buffers. */
2121 /* do nothing for sec context finalize */
2122 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2125 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2127 if (req->rq_export == NULL) {
2128 rc = ldlm_callback_reply(req, -ENOTCONN);
2129 ldlm_callback_errmsg(req, "Operate on unconnected server",
2134 LASSERT(req->rq_export != NULL);
2135 LASSERT(req->rq_export->exp_obd != NULL);
2137 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2138 case LDLM_BL_CALLBACK:
2139 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET))
2142 case LDLM_CP_CALLBACK:
2143 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2146 case LDLM_GL_CALLBACK:
2147 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2151 rc = ldlm_handle_setinfo(req);
2152 ldlm_callback_reply(req, rc);
2154 case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
2155 CERROR("shouldn't be handling OBD_LOG_CANCEL on DLM thread\n");
2156 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2157 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2159 rc = llog_origin_handle_cancel(req);
2160 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2162 ldlm_callback_reply(req, rc);
2164 case LLOG_ORIGIN_HANDLE_CREATE:
2165 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2166 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2168 rc = llog_origin_handle_open(req);
2169 ldlm_callback_reply(req, rc);
2171 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2172 req_capsule_set(&req->rq_pill,
2173 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2174 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2176 rc = llog_origin_handle_next_block(req);
2177 ldlm_callback_reply(req, rc);
2179 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2180 req_capsule_set(&req->rq_pill,
2181 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2182 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2184 rc = llog_origin_handle_read_header(req);
2185 ldlm_callback_reply(req, rc);
2187 case LLOG_ORIGIN_HANDLE_CLOSE:
2188 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2190 rc = llog_origin_handle_close(req);
2191 ldlm_callback_reply(req, rc);
2193 case OBD_QC_CALLBACK:
2194 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
2195 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
2197 rc = ldlm_handle_qc_callback(req);
2198 ldlm_callback_reply(req, rc);
2201 CERROR("unknown opcode %u\n",
2202 lustre_msg_get_opc(req->rq_reqmsg));
2203 ldlm_callback_reply(req, -EPROTO);
2207 ns = req->rq_export->exp_obd->obd_namespace;
2208 LASSERT(ns != NULL);
2210 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2212 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2213 if (dlm_req == NULL) {
2214 rc = ldlm_callback_reply(req, -EPROTO);
2215 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2220 /* Force a known safe race, send a cancel to the server for a lock
2221 * which the server has already started a blocking callback on. */
2222 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2223 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2224 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2226 CERROR("ldlm_cli_cancel: %d\n", rc);
2229 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2231 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2232 "disappeared\n", dlm_req->lock_handle[0].cookie);
2233 rc = ldlm_callback_reply(req, -EINVAL);
2234 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2235 &dlm_req->lock_handle[0]);
2239 if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
2240 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2241 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2243 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2244 lock_res_and_lock(lock);
2245 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2247 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2248 /* If somebody cancels lock and cache is already dropped,
2249 * or lock is failed before cp_ast received on client,
2250 * we can tell the server we have no lock. Otherwise, we
2251 * should send cancel after dropping the cache. */
2252 if (((lock->l_flags & LDLM_FL_CANCELING) &&
2253 (lock->l_flags & LDLM_FL_BL_DONE)) ||
2254 (lock->l_flags & LDLM_FL_FAILED)) {
2255 LDLM_DEBUG(lock, "callback on lock "
2256 LPX64" - lock disappeared\n",
2257 dlm_req->lock_handle[0].cookie);
2258 unlock_res_and_lock(lock);
2259 LDLM_LOCK_RELEASE(lock);
2260 rc = ldlm_callback_reply(req, -EINVAL);
2261 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2262 &dlm_req->lock_handle[0]);
2265 /* BL_AST locks are not needed in LRU.
2266 * Let ldlm_cancel_lru() be fast. */
2267 ldlm_lock_remove_from_lru(lock);
2268 lock->l_flags |= LDLM_FL_BL_AST;
2270 unlock_res_and_lock(lock);
2272 /* We want the ost thread to get this reply so that it can respond
2273 * to ost requests (write cache writeback) that might be triggered
2276 * But we'd also like to be able to indicate in the reply that we're
2277 * cancelling right now, because it's unused, or have an intent result
2278 * in the reply, so we might have to push the responsibility for sending
2279 * the reply down into the AST handlers, alas. */
2281 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2282 case LDLM_BL_CALLBACK:
2283 CDEBUG(D_INODE, "blocking ast\n");
2284 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2285 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) {
2286 rc = ldlm_callback_reply(req, 0);
2287 if (req->rq_no_reply || rc)
2288 ldlm_callback_errmsg(req, "Normal process", rc,
2289 &dlm_req->lock_handle[0]);
2291 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2292 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2294 case LDLM_CP_CALLBACK:
2295 CDEBUG(D_INODE, "completion ast\n");
2296 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2297 ldlm_callback_reply(req, 0);
2298 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2300 case LDLM_GL_CALLBACK:
2301 CDEBUG(D_INODE, "glimpse ast\n");
2302 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2303 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2306 LBUG(); /* checked above */
2312 #ifdef HAVE_SERVER_SUPPORT
2314 * Main handler for canceld thread.
2316 * Separated into its own thread to avoid deadlocks.
2318 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2323 /* Requests arrive in sender's byte order. The ptlrpc service
2324 * handler has already checked and, if necessary, byte-swapped the
2325 * incoming request message body, but I am responsible for the
2326 * message buffers. */
2328 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2330 if (req->rq_export == NULL) {
2331 struct ldlm_request *dlm_req;
2333 CERROR("%s from %s arrived at %lu with bad export cookie "
2335 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2336 libcfs_nid2str(req->rq_peer.nid),
2337 req->rq_arrival_time.tv_sec,
2338 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2340 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2341 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2342 dlm_req = req_capsule_client_get(&req->rq_pill,
2344 if (dlm_req != NULL)
2345 ldlm_lock_dump_handle(D_ERROR,
2346 &dlm_req->lock_handle[0]);
2348 ldlm_callback_reply(req, -ENOTCONN);
2352 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2354 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2356 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2357 CDEBUG(D_INODE, "cancel\n");
2358 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2359 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))
2361 rc = ldlm_handle_cancel(req);
2365 case OBD_LOG_CANCEL:
2366 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2367 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2369 rc = llog_origin_handle_cancel(req);
2370 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2372 ldlm_callback_reply(req, rc);
2375 CERROR("invalid opcode %d\n",
2376 lustre_msg_get_opc(req->rq_reqmsg));
2377 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2378 ldlm_callback_reply(req, -EINVAL);
2384 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2385 struct ldlm_lock *lock)
2387 struct ldlm_request *dlm_req;
2388 struct lustre_handle lockh;
2393 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2394 if (dlm_req == NULL)
2397 ldlm_lock2handle(lock, &lockh);
2398 for (i = 0; i < dlm_req->lock_count; i++) {
2399 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2401 DEBUG_REQ(D_RPCTRACE, req,
2402 "Prio raised by lock "LPX64".", lockh.cookie);
2413 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2415 struct ldlm_request *dlm_req;
2420 /* no prolong in recovery */
2421 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2424 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2425 if (dlm_req == NULL)
2428 for (i = 0; i < dlm_req->lock_count; i++) {
2429 struct ldlm_lock *lock;
2431 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2435 rc = !!(lock->l_flags & LDLM_FL_AST_SENT);
2437 LDLM_DEBUG(lock, "hpreq cancel lock");
2438 LDLM_LOCK_PUT(lock);
2447 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2448 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2449 .hpreq_check = ldlm_cancel_hpreq_check,
2453 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2457 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2459 if (req->rq_export == NULL)
2462 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2463 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2464 req->rq_ops = &ldlm_cancel_hpreq_ops;
2469 int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2470 cfs_hlist_node_t *hnode, void *data)
2473 cfs_list_t *rpc_list = data;
2474 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2476 lock_res_and_lock(lock);
2478 if (lock->l_req_mode != lock->l_granted_mode) {
2479 unlock_res_and_lock(lock);
2483 LASSERT(lock->l_resource);
2484 if (lock->l_resource->lr_type != LDLM_IBITS &&
2485 lock->l_resource->lr_type != LDLM_PLAIN) {
2486 unlock_res_and_lock(lock);
2490 if (lock->l_flags & LDLM_FL_AST_SENT) {
2491 unlock_res_and_lock(lock);
2495 LASSERT(lock->l_blocking_ast);
2496 LASSERT(!lock->l_blocking_lock);
2498 lock->l_flags |= LDLM_FL_AST_SENT;
2499 if (lock->l_export && lock->l_export->exp_lock_hash) {
2500 /* NB: it's safe to call cfs_hash_del() even lock isn't
2501 * in exp_lock_hash. */
2502 /* In the function below, .hs_keycmp resolves to
2503 * ldlm_export_lock_keycmp() */
2504 /* coverity[overrun-buffer-val] */
2505 cfs_hash_del(lock->l_export->exp_lock_hash,
2506 &lock->l_remote_handle, &lock->l_exp_hash);
2509 cfs_list_add_tail(&lock->l_rk_ast, rpc_list);
2510 LDLM_LOCK_GET(lock);
2512 unlock_res_and_lock(lock);
2516 void ldlm_revoke_export_locks(struct obd_export *exp)
2518 cfs_list_t rpc_list;
2521 CFS_INIT_LIST_HEAD(&rpc_list);
2522 cfs_hash_for_each_empty(exp->exp_lock_hash,
2523 ldlm_revoke_lock_cb, &rpc_list);
2524 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2525 LDLM_WORK_REVOKE_AST);
2529 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2530 #endif /* HAVE_SERVER_SUPPORT */
2533 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2535 struct ldlm_bl_work_item *blwi = NULL;
2536 static unsigned int num_bl = 0;
2538 spin_lock(&blp->blp_lock);
2539 /* process a request from the blp_list at least every blp_num_threads */
2540 if (!cfs_list_empty(&blp->blp_list) &&
2541 (cfs_list_empty(&blp->blp_prio_list) || num_bl == 0))
2542 blwi = cfs_list_entry(blp->blp_list.next,
2543 struct ldlm_bl_work_item, blwi_entry);
2545 if (!cfs_list_empty(&blp->blp_prio_list))
2546 blwi = cfs_list_entry(blp->blp_prio_list.next,
2547 struct ldlm_bl_work_item,
2551 if (++num_bl >= cfs_atomic_read(&blp->blp_num_threads))
2553 cfs_list_del(&blwi->blwi_entry);
2555 spin_unlock(&blp->blp_lock);
2560 /* This only contains temporary data until the thread starts */
2561 struct ldlm_bl_thread_data {
2562 char bltd_name[CFS_CURPROC_COMM_MAX];
2563 struct ldlm_bl_pool *bltd_blp;
2564 struct completion bltd_comp;
2568 static int ldlm_bl_thread_main(void *arg);
2570 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2572 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2575 init_completion(&bltd.bltd_comp);
2576 rc = cfs_create_thread(ldlm_bl_thread_main, &bltd, 0);
2578 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
2579 cfs_atomic_read(&blp->blp_num_threads), rc);
2582 wait_for_completion(&bltd.bltd_comp);
2588 * Main blocking requests processing thread.
2590 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2591 * This thread in the end ends up doing actual call to ->l_blocking_ast
2594 static int ldlm_bl_thread_main(void *arg)
2596 struct ldlm_bl_pool *blp;
2600 struct ldlm_bl_thread_data *bltd = arg;
2602 blp = bltd->bltd_blp;
2605 cfs_atomic_inc_return(&blp->blp_num_threads) - 1;
2606 cfs_atomic_inc(&blp->blp_busy_threads);
2608 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
2609 "ldlm_bl_%02d", bltd->bltd_num);
2610 cfs_daemonize(bltd->bltd_name);
2612 complete(&bltd->bltd_comp);
2613 /* cannot use bltd after this, it is only on caller's stack */
2617 struct l_wait_info lwi = { 0 };
2618 struct ldlm_bl_work_item *blwi = NULL;
2621 blwi = ldlm_bl_get_work(blp);
2624 cfs_atomic_dec(&blp->blp_busy_threads);
2625 l_wait_event_exclusive(blp->blp_waitq,
2626 (blwi = ldlm_bl_get_work(blp)) != NULL,
2628 busy = cfs_atomic_inc_return(&blp->blp_busy_threads);
2630 busy = cfs_atomic_read(&blp->blp_busy_threads);
2633 if (blwi->blwi_ns == NULL)
2634 /* added by ldlm_cleanup() */
2637 /* Not fatal if racy and have a few too many threads */
2638 if (unlikely(busy < blp->blp_max_threads &&
2639 busy >= cfs_atomic_read(&blp->blp_num_threads) &&
2640 !blwi->blwi_mem_pressure))
2641 /* discard the return value, we tried */
2642 ldlm_bl_thread_start(blp);
2644 if (blwi->blwi_mem_pressure)
2645 cfs_memory_pressure_set();
2647 if (blwi->blwi_count) {
2649 /* The special case when we cancel locks in LRU
2650 * asynchronously, we pass the list of locks here.
2651 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2652 * canceled locally yet. */
2653 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2656 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2659 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2662 if (blwi->blwi_mem_pressure)
2663 cfs_memory_pressure_clr();
2665 if (blwi->blwi_flags & LCF_ASYNC)
2666 OBD_FREE(blwi, sizeof(*blwi));
2668 complete(&blwi->blwi_comp);
2671 cfs_atomic_dec(&blp->blp_busy_threads);
2672 cfs_atomic_dec(&blp->blp_num_threads);
2673 complete(&blp->blp_comp);
2679 static int ldlm_setup(void);
2680 static int ldlm_cleanup(void);
2682 int ldlm_get_ref(void)
2686 mutex_lock(&ldlm_ref_mutex);
2687 if (++ldlm_refcount == 1) {
2692 mutex_unlock(&ldlm_ref_mutex);
2696 EXPORT_SYMBOL(ldlm_get_ref);
2698 void ldlm_put_ref(void)
2701 mutex_lock(&ldlm_ref_mutex);
2702 if (ldlm_refcount == 1) {
2703 int rc = ldlm_cleanup();
2705 CERROR("ldlm_cleanup failed: %d\n", rc);
2711 mutex_unlock(&ldlm_ref_mutex);
2715 EXPORT_SYMBOL(ldlm_put_ref);
2718 * Export handle<->lock hash operations.
2721 ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
2723 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2727 ldlm_export_lock_key(cfs_hlist_node_t *hnode)
2729 struct ldlm_lock *lock;
2731 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2732 return &lock->l_remote_handle;
2736 ldlm_export_lock_keycpy(cfs_hlist_node_t *hnode, void *key)
2738 struct ldlm_lock *lock;
2740 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2741 lock->l_remote_handle = *(struct lustre_handle *)key;
2745 ldlm_export_lock_keycmp(const void *key, cfs_hlist_node_t *hnode)
2747 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2751 ldlm_export_lock_object(cfs_hlist_node_t *hnode)
2753 return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2757 ldlm_export_lock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2759 struct ldlm_lock *lock;
2761 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2762 LDLM_LOCK_GET(lock);
2766 ldlm_export_lock_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2768 struct ldlm_lock *lock;
2770 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2771 LDLM_LOCK_RELEASE(lock);
2774 static cfs_hash_ops_t ldlm_export_lock_ops = {
2775 .hs_hash = ldlm_export_lock_hash,
2776 .hs_key = ldlm_export_lock_key,
2777 .hs_keycmp = ldlm_export_lock_keycmp,
2778 .hs_keycpy = ldlm_export_lock_keycpy,
2779 .hs_object = ldlm_export_lock_object,
2780 .hs_get = ldlm_export_lock_get,
2781 .hs_put = ldlm_export_lock_put,
2782 .hs_put_locked = ldlm_export_lock_put,
2785 int ldlm_init_export(struct obd_export *exp)
2789 exp->exp_lock_hash =
2790 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2791 HASH_EXP_LOCK_CUR_BITS,
2792 HASH_EXP_LOCK_MAX_BITS,
2793 HASH_EXP_LOCK_BKT_BITS, 0,
2794 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2795 &ldlm_export_lock_ops,
2796 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2797 CFS_HASH_NBLK_CHANGE);
2799 if (!exp->exp_lock_hash)
2804 EXPORT_SYMBOL(ldlm_init_export);
2806 void ldlm_destroy_export(struct obd_export *exp)
2809 cfs_hash_putref(exp->exp_lock_hash);
2810 exp->exp_lock_hash = NULL;
2812 ldlm_destroy_flock_export(exp);
2815 EXPORT_SYMBOL(ldlm_destroy_export);
2817 static int ldlm_setup(void)
2819 static struct ptlrpc_service_conf conf;
2820 struct ldlm_bl_pool *blp = NULL;
2827 if (ldlm_state != NULL)
2830 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2831 if (ldlm_state == NULL)
2835 rc = ldlm_proc_setup();
2840 memset(&conf, 0, sizeof(conf));
2841 conf = (typeof(conf)) {
2842 .psc_name = "ldlm_cbd",
2843 .psc_watchdog_factor = 2,
2845 .bc_nbufs = LDLM_CLIENT_NBUFS,
2846 .bc_buf_size = LDLM_BUFSIZE,
2847 .bc_req_max_size = LDLM_MAXREQSIZE,
2848 .bc_rep_max_size = LDLM_MAXREPSIZE,
2849 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2850 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2853 .tc_thr_name = "ldlm_cb",
2854 .tc_thr_factor = LDLM_THR_FACTOR,
2855 .tc_nthrs_init = LDLM_NTHRS_INIT,
2856 .tc_nthrs_base = LDLM_NTHRS_BASE,
2857 .tc_nthrs_max = LDLM_NTHRS_MAX,
2858 .tc_nthrs_user = ldlm_num_threads,
2859 .tc_cpu_affinity = 1,
2860 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2863 .cc_pattern = ldlm_cpts,
2866 .so_req_handler = ldlm_callback_handler,
2869 ldlm_state->ldlm_cb_service = \
2870 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2871 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2872 CERROR("failed to start service\n");
2873 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2874 ldlm_state->ldlm_cb_service = NULL;
2878 #ifdef HAVE_SERVER_SUPPORT
2879 memset(&conf, 0, sizeof(conf));
2880 conf = (typeof(conf)) {
2881 .psc_name = "ldlm_canceld",
2882 .psc_watchdog_factor = 6,
2884 .bc_nbufs = LDLM_SERVER_NBUFS,
2885 .bc_buf_size = LDLM_BUFSIZE,
2886 .bc_req_max_size = LDLM_MAXREQSIZE,
2887 .bc_rep_max_size = LDLM_MAXREPSIZE,
2888 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
2889 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
2893 .tc_thr_name = "ldlm_cn",
2894 .tc_thr_factor = LDLM_THR_FACTOR,
2895 .tc_nthrs_init = LDLM_NTHRS_INIT,
2896 .tc_nthrs_base = LDLM_NTHRS_BASE,
2897 .tc_nthrs_max = LDLM_NTHRS_MAX,
2898 .tc_nthrs_user = ldlm_num_threads,
2899 .tc_cpu_affinity = 1,
2900 .tc_ctx_tags = LCT_MD_THREAD | \
2905 .cc_pattern = ldlm_cpts,
2908 .so_req_handler = ldlm_cancel_handler,
2909 .so_hpreq_handler = ldlm_hpreq_handler,
2912 ldlm_state->ldlm_cancel_service = \
2913 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2914 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2915 CERROR("failed to start service\n");
2916 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2917 ldlm_state->ldlm_cancel_service = NULL;
2922 OBD_ALLOC(blp, sizeof(*blp));
2924 GOTO(out, rc = -ENOMEM);
2925 ldlm_state->ldlm_bl_pool = blp;
2927 spin_lock_init(&blp->blp_lock);
2928 CFS_INIT_LIST_HEAD(&blp->blp_list);
2929 CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2930 cfs_waitq_init(&blp->blp_waitq);
2931 cfs_atomic_set(&blp->blp_num_threads, 0);
2932 cfs_atomic_set(&blp->blp_busy_threads, 0);
2935 if (ldlm_num_threads == 0) {
2936 blp->blp_min_threads = LDLM_NTHRS_INIT;
2937 blp->blp_max_threads = LDLM_NTHRS_MAX;
2939 blp->blp_min_threads = blp->blp_max_threads = \
2940 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
2944 for (i = 0; i < blp->blp_min_threads; i++) {
2945 rc = ldlm_bl_thread_start(blp);
2950 # ifdef HAVE_SERVER_SUPPORT
2951 CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2952 expired_lock_thread.elt_state = ELT_STOPPED;
2953 cfs_waitq_init(&expired_lock_thread.elt_waitq);
2955 CFS_INIT_LIST_HEAD(&waiting_locks_list);
2956 spin_lock_init(&waiting_locks_spinlock);
2957 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2959 rc = cfs_create_thread(expired_lock_main, NULL, CFS_DAEMON_FLAGS);
2961 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2965 cfs_wait_event(expired_lock_thread.elt_waitq,
2966 expired_lock_thread.elt_state == ELT_READY);
2967 # endif /* HAVE_SERVER_SUPPORT */
2969 rc = ldlm_pools_init();
2971 CERROR("Failed to initialize LDLM pools: %d\n", rc);
2982 static int ldlm_cleanup(void)
2986 if (!cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2987 !cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2988 CERROR("ldlm still has namespaces; clean these up first.\n");
2989 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2990 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2997 if (ldlm_state->ldlm_bl_pool != NULL) {
2998 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3000 while (cfs_atomic_read(&blp->blp_num_threads) > 0) {
3001 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3003 init_completion(&blp->blp_comp);
3005 spin_lock(&blp->blp_lock);
3006 cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3007 cfs_waitq_signal(&blp->blp_waitq);
3008 spin_unlock(&blp->blp_lock);
3010 wait_for_completion(&blp->blp_comp);
3013 OBD_FREE(blp, sizeof(*blp));
3015 #endif /* __KERNEL__ */
3017 if (ldlm_state->ldlm_cb_service != NULL)
3018 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3019 # ifdef HAVE_SERVER_SUPPORT
3020 if (ldlm_state->ldlm_cancel_service != NULL)
3021 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3025 ldlm_proc_cleanup();
3027 # ifdef HAVE_SERVER_SUPPORT
3028 if (expired_lock_thread.elt_state != ELT_STOPPED) {
3029 expired_lock_thread.elt_state = ELT_TERMINATE;
3030 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
3031 cfs_wait_event(expired_lock_thread.elt_waitq,
3032 expired_lock_thread.elt_state == ELT_STOPPED);
3035 #endif /* __KERNEL__ */
3037 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3045 mutex_init(&ldlm_ref_mutex);
3046 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
3047 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
3048 ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
3049 sizeof(struct ldlm_resource), 0,
3050 CFS_SLAB_HWCACHE_ALIGN);
3051 if (ldlm_resource_slab == NULL)
3054 ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
3055 sizeof(struct ldlm_lock), 0,
3056 CFS_SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU);
3057 if (ldlm_lock_slab == NULL) {
3058 cfs_mem_cache_destroy(ldlm_resource_slab);
3062 ldlm_interval_slab = cfs_mem_cache_create("interval_node",
3063 sizeof(struct ldlm_interval),
3064 0, CFS_SLAB_HWCACHE_ALIGN);
3065 if (ldlm_interval_slab == NULL) {
3066 cfs_mem_cache_destroy(ldlm_resource_slab);
3067 cfs_mem_cache_destroy(ldlm_lock_slab);
3070 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3071 class_export_dump_hook = ldlm_dump_export_locks;
3076 void ldlm_exit(void)
3080 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3081 rc = cfs_mem_cache_destroy(ldlm_resource_slab);
3082 LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
3084 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3085 * synchronize_rcu() to wait a grace period elapsed, so that
3086 * ldlm_lock_free() get a chance to be called. */
3089 rc = cfs_mem_cache_destroy(ldlm_lock_slab);
3090 LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
3091 rc = cfs_mem_cache_destroy(ldlm_interval_slab);
3092 LASSERTF(rc == 0, "couldn't free interval node slab\n");