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, 2013, 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 struct kmem_cache *ldlm_resource_slab;
64 extern struct kmem_cache *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 };
182 expired_lock_thread.elt_state = ELT_READY;
183 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
186 l_wait_event(expired_lock_thread.elt_waitq,
187 have_expired_locks() ||
188 expired_lock_thread.elt_state == ELT_TERMINATE,
191 spin_lock_bh(&waiting_locks_spinlock);
192 if (expired_lock_thread.elt_dump) {
193 struct libcfs_debug_msg_data msgdata = {
194 .msg_file = __FILE__,
195 .msg_fn = "waiting_locks_callback",
196 .msg_line = expired_lock_thread.elt_dump };
197 spin_unlock_bh(&waiting_locks_spinlock);
199 /* from waiting_locks_callback, but not in timer */
200 libcfs_debug_dumplog();
201 libcfs_run_lbug_upcall(&msgdata);
203 spin_lock_bh(&waiting_locks_spinlock);
204 expired_lock_thread.elt_dump = 0;
209 while (!cfs_list_empty(expired)) {
210 struct obd_export *export;
211 struct ldlm_lock *lock;
213 lock = cfs_list_entry(expired->next, struct ldlm_lock,
215 if ((void *)lock < LP_POISON + PAGE_CACHE_SIZE &&
216 (void *)lock >= LP_POISON) {
217 spin_unlock_bh(&waiting_locks_spinlock);
218 CERROR("free lock on elt list %p\n", lock);
221 cfs_list_del_init(&lock->l_pending_chain);
222 if ((void *)lock->l_export <
223 LP_POISON + PAGE_CACHE_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_flags & LDLM_FL_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_flags & LDLM_FL_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_flags & LDLM_FL_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_flags & LDLM_FL_RES_LOCKED);
472 lock->l_flags |= LDLM_FL_WAITED;
474 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
476 spin_lock_bh(&waiting_locks_spinlock);
477 if (lock->l_flags & LDLM_FL_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_flags & (LDLM_FL_RES_LOCKED|LDLM_FL_CANCEL_ON_BLOCK))
623 == LDLM_FL_RES_LOCKED);
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_nrs_req_can_move() without
814 * holding svcpt->scp_req_lock, and then redo the check 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_flags & LDLM_FL_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 /* LU-3124 & LU-2187: to not return layout in completion AST because
957 * it may deadlock for LU-2187, or client may not have enough space
958 * for large layout. The layout will be returned to client with an
959 * extra RPC to fetch xattr.lov */
960 if (ldlm_has_layout(lock))
963 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
964 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
966 ptlrpc_request_free(req);
970 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
971 ca = ptlrpc_req_async_args(req);
972 ca->ca_set_arg = arg;
975 req->rq_interpret_reply = ldlm_cb_interpret;
976 req->rq_no_resend = 1;
977 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
979 body->lock_handle[0] = lock->l_remote_handle;
980 body->lock_flags = ldlm_flags_to_wire(flags);
981 ldlm_lock2desc(lock, &body->lock_desc);
983 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
985 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
987 /* We still need to send the RPC to wake up the blocked
988 * enqueue thread on the client.
990 * Consider old client, there is no better way to notify
991 * the failure, just zero-sized the LVB, then the client
992 * will fail out as "-EPROTO". */
993 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
997 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1002 LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)",
1003 total_enqueue_wait);
1005 /* Server-side enqueue wait time estimate, used in
1006 __ldlm_add_waiting_lock to set future enqueue timers */
1007 if (total_enqueue_wait < ldlm_get_enq_timeout(lock))
1008 at_measured(ldlm_lock_to_ns_at(lock),
1009 total_enqueue_wait);
1011 /* bz18618. Don't add lock enqueue time we spend waiting for a
1012 previous callback to fail. Locks waiting legitimately will
1013 get extended by ldlm_refresh_waiting_lock regardless of the
1014 estimate, so it's okay to underestimate here. */
1015 LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. "
1016 "It is likely that a previous callback timed out.",
1018 at_get(ldlm_lock_to_ns_at(lock)));
1020 ptlrpc_request_set_replen(req);
1022 req->rq_send_state = LUSTRE_IMP_FULL;
1023 /* ptlrpc_request_pack already set timeout */
1025 req->rq_timeout = ldlm_get_rq_timeout();
1027 /* We only send real blocking ASTs after the lock is granted */
1028 lock_res_and_lock(lock);
1029 if (lock->l_flags & LDLM_FL_AST_SENT) {
1030 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1031 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1032 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1035 /* We might get here prior to ldlm_handle_enqueue setting
1036 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1037 * into waiting list, but this is safe and similar code in
1038 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1039 * that would not only cancel the lock, but will also remove
1040 * it from waiting list */
1041 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1042 unlock_res_and_lock(lock);
1043 ldlm_lock_cancel(lock);
1045 lock_res_and_lock(lock);
1047 /* start the lock-timeout clock */
1048 ldlm_add_waiting_lock(lock);
1051 unlock_res_and_lock(lock);
1053 if (lock->l_export && lock->l_export->exp_nid_stats &&
1054 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1055 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1056 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1058 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1060 RETURN(lvb_len < 0 ? lvb_len : rc);
1062 EXPORT_SYMBOL(ldlm_server_completion_ast);
1065 * Server side ->l_glimpse_ast handler for client locks.
1067 * Sends glimpse AST to the client and waits for reply. Then updates
1068 * lvbo with the result.
1070 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1072 struct ldlm_cb_set_arg *arg = data;
1073 struct ldlm_request *body;
1074 struct ptlrpc_request *req;
1075 struct ldlm_cb_async_args *ca;
1077 struct req_format *req_fmt;
1080 LASSERT(lock != NULL);
1082 if (arg->gl_desc != NULL)
1083 /* There is a glimpse descriptor to pack */
1084 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1086 req_fmt = &RQF_LDLM_GL_CALLBACK;
1088 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1089 req_fmt, LUSTRE_DLM_VERSION,
1095 if (arg->gl_desc != NULL) {
1096 /* copy the GL descriptor */
1097 union ldlm_gl_desc *desc;
1098 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1099 *desc = *arg->gl_desc;
1102 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1103 body->lock_handle[0] = lock->l_remote_handle;
1104 ldlm_lock2desc(lock, &body->lock_desc);
1106 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1107 ca = ptlrpc_req_async_args(req);
1108 ca->ca_set_arg = arg;
1111 /* server namespace, doesn't need lock */
1112 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1113 ldlm_lvbo_size(lock));
1114 ptlrpc_request_set_replen(req);
1116 req->rq_send_state = LUSTRE_IMP_FULL;
1117 /* ptlrpc_request_alloc_pack already set timeout */
1119 req->rq_timeout = ldlm_get_rq_timeout();
1121 req->rq_interpret_reply = ldlm_cb_interpret;
1123 if (lock->l_export && lock->l_export->exp_nid_stats &&
1124 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1125 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1126 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1128 rc = ldlm_ast_fini(req, arg, lock, 0);
1132 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1134 int ldlm_glimpse_locks(struct ldlm_resource *res, cfs_list_t *gl_work_list)
1139 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1141 if (rc == -ERESTART)
1142 ldlm_reprocess_all(res);
1146 EXPORT_SYMBOL(ldlm_glimpse_locks);
1148 /* return LDLM lock associated with a lock callback request */
1149 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1151 struct ldlm_cb_async_args *ca;
1152 struct ldlm_lock *lock;
1155 ca = ptlrpc_req_async_args(req);
1158 RETURN(ERR_PTR(-EFAULT));
1162 EXPORT_SYMBOL(ldlm_request_lock);
1164 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1165 struct lprocfs_stats *srv_stats)
1167 int lock_type = 0, op = 0;
1169 lock_type = dlm_req->lock_desc.l_resource.lr_type;
1171 switch (lock_type) {
1173 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1176 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1177 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1179 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1182 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1185 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1193 lprocfs_counter_incr(srv_stats, op);
1199 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1200 * service threads to carry out client lock enqueueing requests.
1202 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1203 struct ptlrpc_request *req,
1204 const struct ldlm_request *dlm_req,
1205 const struct ldlm_callback_suite *cbs)
1207 struct ldlm_reply *dlm_rep;
1209 ldlm_error_t err = ELDLM_OK;
1210 struct ldlm_lock *lock = NULL;
1211 void *cookie = NULL;
1215 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1217 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1218 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1220 LASSERT(req->rq_export);
1222 if (ptlrpc_req2svc(req)->srv_stats != NULL)
1223 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1225 if (req->rq_export && req->rq_export->exp_nid_stats &&
1226 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1227 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1228 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1230 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1231 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1232 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1233 dlm_req->lock_desc.l_resource.lr_type);
1234 GOTO(out, rc = -EFAULT);
1237 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1238 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1239 dlm_req->lock_desc.l_req_mode &
1240 (dlm_req->lock_desc.l_req_mode-1))) {
1241 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1242 dlm_req->lock_desc.l_req_mode);
1243 GOTO(out, rc = -EFAULT);
1246 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) {
1247 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1249 DEBUG_REQ(D_ERROR, req,
1250 "PLAIN lock request from IBITS client?");
1251 GOTO(out, rc = -EPROTO);
1253 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1255 DEBUG_REQ(D_ERROR, req,
1256 "IBITS lock request from unaware client?");
1257 GOTO(out, rc = -EPROTO);
1261 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1262 against server's _CONNECT_SUPPORTED flags? (I don't want to use
1263 ibits for mgc/mgs) */
1265 /* INODEBITS_INTEROP: Perform conversion from plain lock to
1266 * inodebits lock if client does not support them. */
1267 if (!(exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) &&
1268 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1269 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1270 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1271 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1272 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1273 dlm_req->lock_desc.l_req_mode = LCK_CR;
1277 if (unlikely(flags & LDLM_FL_REPLAY)) {
1278 /* Find an existing lock in the per-export lock hash */
1279 /* In the function below, .hs_keycmp resolves to
1280 * ldlm_export_lock_keycmp() */
1281 /* coverity[overrun-buffer-val] */
1282 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1283 (void *)&dlm_req->lock_handle[0]);
1285 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1286 LPX64, lock->l_handle.h_cookie);
1287 GOTO(existing_lock, rc = 0);
1291 /* The lock's callback data might be set in the policy function */
1292 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1293 dlm_req->lock_desc.l_resource.lr_type,
1294 dlm_req->lock_desc.l_req_mode,
1295 cbs, NULL, 0, LVB_T_NONE);
1297 GOTO(out, rc = -ENOMEM);
1299 lock->l_last_activity = cfs_time_current_sec();
1300 lock->l_remote_handle = dlm_req->lock_handle[0];
1301 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1303 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1304 /* Don't enqueue a lock onto the export if it is been disonnected
1305 * due to eviction (bug 3822) or server umount (bug 24324).
1306 * Cancel it now instead. */
1307 if (req->rq_export->exp_disconnected) {
1308 LDLM_ERROR(lock, "lock on disconnected export %p",
1310 GOTO(out, rc = -ENOTCONN);
1313 lock->l_export = class_export_lock_get(req->rq_export, lock);
1314 if (lock->l_export->exp_lock_hash)
1315 cfs_hash_add(lock->l_export->exp_lock_hash,
1316 &lock->l_remote_handle,
1321 if (flags & LDLM_FL_HAS_INTENT) {
1322 /* In this case, the reply buffer is allocated deep in
1323 * local_lock_enqueue by the policy function. */
1326 /* based on the assumption that lvb size never changes during
1327 * resource life time otherwise it need resource->lr_lock's
1329 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1330 RCL_SERVER, ldlm_lvbo_size(lock));
1332 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1333 GOTO(out, rc = -ENOMEM);
1335 rc = req_capsule_server_pack(&req->rq_pill);
1340 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1341 ldlm_convert_policy_to_local(req->rq_export,
1342 dlm_req->lock_desc.l_resource.lr_type,
1343 &dlm_req->lock_desc.l_policy_data,
1344 &lock->l_policy_data);
1345 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1346 lock->l_req_extent = lock->l_policy_data.l_extent;
1348 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1355 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1356 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1358 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1359 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1361 /* We never send a blocking AST until the lock is granted, but
1362 * we can tell it right now */
1363 lock_res_and_lock(lock);
1365 /* Now take into account flags to be inherited from original lock
1366 request both in reply to client and in our own lock flags. */
1367 dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1368 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1369 LDLM_INHERIT_FLAGS);
1371 /* Don't move a pending lock onto the export if it has already been
1372 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1373 * Cancel it now instead. */
1374 if (unlikely(req->rq_export->exp_disconnected ||
1375 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1376 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1378 } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1379 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1380 if (lock->l_granted_mode == lock->l_req_mode) {
1382 * Only cancel lock if it was granted, because it would
1383 * be destroyed immediately and would never be granted
1384 * in the future, causing timeouts on client. Not
1385 * granted lock will be cancelled immediately after
1386 * sending completion AST.
1388 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1389 unlock_res_and_lock(lock);
1390 ldlm_lock_cancel(lock);
1391 lock_res_and_lock(lock);
1393 ldlm_add_waiting_lock(lock);
1396 /* Make sure we never ever grant usual metadata locks to liblustre
1398 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1399 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1400 req->rq_export->exp_libclient) {
1401 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1402 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1403 CERROR("Granting sync lock to libclient. "
1404 "req fl %d, rep fl %d, lock fl "LPX64"\n",
1405 dlm_req->lock_flags, dlm_rep->lock_flags,
1407 LDLM_ERROR(lock, "sync lock");
1408 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1409 struct ldlm_intent *it;
1411 it = req_capsule_client_get(&req->rq_pill,
1414 CERROR("This is intent %s ("LPU64")\n",
1415 ldlm_it2str(it->opc), it->opc);
1421 unlock_res_and_lock(lock);
1425 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1426 if (!req->rq_packed_final) {
1427 err = lustre_pack_reply(req, 1, NULL, NULL);
1432 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1433 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1435 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1436 "(err=%d, rc=%d)", err, rc);
1439 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1441 ldlm_lvbo_size(lock) > 0) {
1445 buf = req_capsule_server_get(&req->rq_pill,
1447 LASSERTF(buf != NULL, "req %p, lock %p\n",
1449 buflen = req_capsule_get_size(&req->rq_pill,
1450 &RMF_DLM_LVB, RCL_SERVER);
1451 buflen = ldlm_lvbo_fill(lock, buf, buflen);
1453 req_capsule_shrink(&req->rq_pill,
1455 buflen, RCL_SERVER);
1460 lock_res_and_lock(lock);
1461 ldlm_resource_unlink_lock(lock);
1462 ldlm_lock_destroy_nolock(lock);
1463 unlock_res_and_lock(lock);
1466 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1467 ldlm_reprocess_all(lock->l_resource);
1469 LDLM_LOCK_RELEASE(lock);
1472 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1477 EXPORT_SYMBOL(ldlm_handle_enqueue0);
1480 * Old-style LDLM main entry point for server code enqueue.
1482 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1483 ldlm_completion_callback completion_callback,
1484 ldlm_blocking_callback blocking_callback,
1485 ldlm_glimpse_callback glimpse_callback)
1487 struct ldlm_request *dlm_req;
1488 struct ldlm_callback_suite cbs = {
1489 .lcs_completion = completion_callback,
1490 .lcs_blocking = blocking_callback,
1491 .lcs_glimpse = glimpse_callback
1495 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1496 if (dlm_req != NULL) {
1497 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1498 req, dlm_req, &cbs);
1504 EXPORT_SYMBOL(ldlm_handle_enqueue);
1507 * Main LDLM entry point for server code to process lock conversion requests.
1509 int ldlm_handle_convert0(struct ptlrpc_request *req,
1510 const struct ldlm_request *dlm_req)
1512 struct ldlm_reply *dlm_rep;
1513 struct ldlm_lock *lock;
1517 if (req->rq_export && req->rq_export->exp_nid_stats &&
1518 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1519 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1520 LDLM_CONVERT - LDLM_FIRST_OPC);
1522 rc = req_capsule_server_pack(&req->rq_pill);
1526 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1527 dlm_rep->lock_flags = dlm_req->lock_flags;
1529 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1531 req->rq_status = LUSTRE_EINVAL;
1535 LDLM_DEBUG(lock, "server-side convert handler START");
1537 lock->l_last_activity = cfs_time_current_sec();
1538 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1539 &dlm_rep->lock_flags);
1541 if (ldlm_del_waiting_lock(lock))
1542 LDLM_DEBUG(lock, "converted waiting lock");
1545 req->rq_status = LUSTRE_EDEADLK;
1550 if (!req->rq_status)
1551 ldlm_reprocess_all(lock->l_resource);
1552 LDLM_DEBUG(lock, "server-side convert handler END");
1553 LDLM_LOCK_PUT(lock);
1555 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1559 EXPORT_SYMBOL(ldlm_handle_convert0);
1562 * Old-style main LDLM entry point for server code to process lock conversion
1565 int ldlm_handle_convert(struct ptlrpc_request *req)
1568 struct ldlm_request *dlm_req;
1570 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1571 if (dlm_req != NULL) {
1572 rc = ldlm_handle_convert0(req, dlm_req);
1574 CERROR ("Can't unpack dlm_req\n");
1579 EXPORT_SYMBOL(ldlm_handle_convert);
1582 * Cancel all the locks whose handles are packed into ldlm_request
1584 * Called by server code expecting such combined cancel activity
1587 int ldlm_request_cancel(struct ptlrpc_request *req,
1588 const struct ldlm_request *dlm_req, int first)
1590 struct ldlm_resource *res, *pres = NULL;
1591 struct ldlm_lock *lock;
1592 int i, count, done = 0;
1595 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1599 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1602 /* There is no lock on the server at the replay time,
1603 * skip lock cancelling to make replay tests to pass. */
1604 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1607 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1608 "starting at %d", count, first);
1610 for (i = first; i < count; i++) {
1611 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1613 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1614 "lock (cookie "LPU64")",
1615 dlm_req->lock_handle[i].cookie);
1619 res = lock->l_resource;
1622 /* This code is an optimization to only attempt lock
1623 * granting on the resource (that could be CPU-expensive)
1624 * after we are done cancelling lock in that resource. */
1627 ldlm_reprocess_all(pres);
1628 LDLM_RESOURCE_DELREF(pres);
1629 ldlm_resource_putref(pres);
1632 ldlm_resource_getref(res);
1633 LDLM_RESOURCE_ADDREF(res);
1634 ldlm_res_lvbo_update(res, NULL, 1);
1638 ldlm_lock_cancel(lock);
1639 LDLM_LOCK_PUT(lock);
1642 ldlm_reprocess_all(pres);
1643 LDLM_RESOURCE_DELREF(pres);
1644 ldlm_resource_putref(pres);
1646 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1649 EXPORT_SYMBOL(ldlm_request_cancel);
1652 * Main LDLM entry point for server code to cancel locks.
1654 * Typically gets called from service handler on LDLM_CANCEL opc.
1656 int ldlm_handle_cancel(struct ptlrpc_request *req)
1658 struct ldlm_request *dlm_req;
1662 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1663 if (dlm_req == NULL) {
1664 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1668 if (req->rq_export && req->rq_export->exp_nid_stats &&
1669 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1670 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1671 LDLM_CANCEL - LDLM_FIRST_OPC);
1673 rc = req_capsule_server_pack(&req->rq_pill);
1677 if (!ldlm_request_cancel(req, dlm_req, 0))
1678 req->rq_status = LUSTRE_ESTALE;
1680 RETURN(ptlrpc_reply(req));
1682 EXPORT_SYMBOL(ldlm_handle_cancel);
1683 #endif /* HAVE_SERVER_SUPPORT */
1686 * Callback handler for receiving incoming blocking ASTs.
1688 * This can only happen on client side.
1690 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1691 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1696 LDLM_DEBUG(lock, "client blocking AST callback handler");
1698 lock_res_and_lock(lock);
1699 lock->l_flags |= LDLM_FL_CBPENDING;
1701 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1702 lock->l_flags |= LDLM_FL_CANCEL;
1704 do_ast = (!lock->l_readers && !lock->l_writers);
1705 unlock_res_and_lock(lock);
1708 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1709 lock, lock->l_blocking_ast);
1710 if (lock->l_blocking_ast != NULL)
1711 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1714 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1718 LDLM_DEBUG(lock, "client blocking callback handler END");
1719 LDLM_LOCK_RELEASE(lock);
1724 * Callback handler for receiving incoming completion ASTs.
1726 * This only can happen on client side.
1728 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1729 struct ldlm_namespace *ns,
1730 struct ldlm_request *dlm_req,
1731 struct ldlm_lock *lock)
1734 CFS_LIST_HEAD(ast_list);
1738 LDLM_DEBUG(lock, "client completion callback handler START");
1740 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1741 int to = cfs_time_seconds(1);
1743 cfs_schedule_timeout_and_set_state(
1744 CFS_TASK_INTERRUPTIBLE, to);
1745 if (lock->l_granted_mode == lock->l_req_mode ||
1746 lock->l_flags & LDLM_FL_DESTROYED)
1751 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1753 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1754 GOTO(out, rc = lvb_len);
1755 } else if (lvb_len > 0) {
1756 if (lock->l_lvb_len > 0) {
1757 /* for extent lock, lvb contains ost_lvb{}. */
1758 LASSERT(lock->l_lvb_data != NULL);
1760 if (unlikely(lock->l_lvb_len < lvb_len)) {
1761 LDLM_ERROR(lock, "Replied LVB is larger than "
1762 "expectation, expected = %d, "
1764 lock->l_lvb_len, lvb_len);
1765 GOTO(out, rc = -EINVAL);
1767 } else if (ldlm_has_layout(lock)) { /* for layout lock, lvb has
1768 * variable length */
1771 OBD_ALLOC(lvb_data, lvb_len);
1772 if (lvb_data == NULL) {
1773 LDLM_ERROR(lock, "No memory: %d.\n", lvb_len);
1774 GOTO(out, rc = -ENOMEM);
1777 lock_res_and_lock(lock);
1778 LASSERT(lock->l_lvb_data == NULL);
1779 lock->l_lvb_data = lvb_data;
1780 lock->l_lvb_len = lvb_len;
1781 unlock_res_and_lock(lock);
1785 lock_res_and_lock(lock);
1786 if ((lock->l_flags & LDLM_FL_DESTROYED) ||
1787 lock->l_granted_mode == lock->l_req_mode) {
1788 /* bug 11300: the lock has already been granted */
1789 unlock_res_and_lock(lock);
1790 LDLM_DEBUG(lock, "Double grant race happened");
1794 /* If we receive the completion AST before the actual enqueue returned,
1795 * then we might need to switch lock modes, resources, or extents. */
1796 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1797 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1798 LDLM_DEBUG(lock, "completion AST, new lock mode");
1801 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1802 ldlm_convert_policy_to_local(req->rq_export,
1803 dlm_req->lock_desc.l_resource.lr_type,
1804 &dlm_req->lock_desc.l_policy_data,
1805 &lock->l_policy_data);
1806 LDLM_DEBUG(lock, "completion AST, new policy data");
1809 ldlm_resource_unlink_lock(lock);
1810 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1811 &lock->l_resource->lr_name,
1812 sizeof(lock->l_resource->lr_name)) != 0) {
1813 unlock_res_and_lock(lock);
1814 rc = ldlm_lock_change_resource(ns, lock,
1815 &dlm_req->lock_desc.l_resource.lr_name);
1817 LDLM_ERROR(lock, "Failed to allocate resource");
1820 LDLM_DEBUG(lock, "completion AST, new resource");
1821 CERROR("change resource!\n");
1822 lock_res_and_lock(lock);
1825 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1826 /* BL_AST locks are not needed in LRU.
1827 * Let ldlm_cancel_lru() be fast. */
1828 ldlm_lock_remove_from_lru(lock);
1829 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1830 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1833 if (lock->l_lvb_len > 0) {
1834 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1835 lock->l_lvb_data, lvb_len);
1837 unlock_res_and_lock(lock);
1842 ldlm_grant_lock(lock, &ast_list);
1843 unlock_res_and_lock(lock);
1845 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1847 /* Let Enqueue to call osc_lock_upcall() and initialize
1849 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1851 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1853 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1859 lock_res_and_lock(lock);
1860 lock->l_flags |= LDLM_FL_FAILED;
1861 unlock_res_and_lock(lock);
1862 cfs_waitq_signal(&lock->l_waitq);
1864 LDLM_LOCK_RELEASE(lock);
1868 * Callback handler for receiving incoming glimpse ASTs.
1870 * This only can happen on client side. After handling the glimpse AST
1871 * we also consider dropping the lock here if it is unused locally for a
1874 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1875 struct ldlm_namespace *ns,
1876 struct ldlm_request *dlm_req,
1877 struct ldlm_lock *lock)
1882 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1884 if (lock->l_glimpse_ast != NULL)
1885 rc = lock->l_glimpse_ast(lock, req);
1887 if (req->rq_repmsg != NULL) {
1890 req->rq_status = rc;
1894 lock_res_and_lock(lock);
1895 if (lock->l_granted_mode == LCK_PW &&
1896 !lock->l_readers && !lock->l_writers &&
1897 cfs_time_after(cfs_time_current(),
1898 cfs_time_add(lock->l_last_used,
1899 cfs_time_seconds(10)))) {
1900 unlock_res_and_lock(lock);
1901 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1902 ldlm_handle_bl_callback(ns, NULL, lock);
1907 unlock_res_and_lock(lock);
1908 LDLM_LOCK_RELEASE(lock);
1912 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1914 if (req->rq_no_reply)
1917 req->rq_status = rc;
1918 if (!req->rq_packed_final) {
1919 rc = lustre_pack_reply(req, 1, NULL, NULL);
1923 return ptlrpc_reply(req);
1927 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1928 ldlm_cancel_flags_t cancel_flags)
1930 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1933 spin_lock(&blp->blp_lock);
1934 if (blwi->blwi_lock &&
1935 blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) {
1936 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1937 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1939 /* other blocking callbacks are added to the regular list */
1940 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1942 spin_unlock(&blp->blp_lock);
1944 cfs_waitq_signal(&blp->blp_waitq);
1946 /* can not check blwi->blwi_flags as blwi could be already freed in
1948 if (!(cancel_flags & LCF_ASYNC))
1949 wait_for_completion(&blwi->blwi_comp);
1954 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1955 struct ldlm_namespace *ns,
1956 struct ldlm_lock_desc *ld,
1957 cfs_list_t *cancels, int count,
1958 struct ldlm_lock *lock,
1959 ldlm_cancel_flags_t cancel_flags)
1961 init_completion(&blwi->blwi_comp);
1962 CFS_INIT_LIST_HEAD(&blwi->blwi_head);
1964 if (memory_pressure_get())
1965 blwi->blwi_mem_pressure = 1;
1968 blwi->blwi_flags = cancel_flags;
1970 blwi->blwi_ld = *ld;
1972 cfs_list_add(&blwi->blwi_head, cancels);
1973 cfs_list_del_init(cancels);
1974 blwi->blwi_count = count;
1976 blwi->blwi_lock = lock;
1981 * Queues a list of locks \a cancels containing \a count locks
1982 * for later processing by a blocking thread. If \a count is zero,
1983 * then the lock referenced as \a lock is queued instead.
1985 * The blocking thread would then call ->l_blocking_ast callback in the lock.
1986 * If list addition fails an error is returned and caller is supposed to
1987 * call ->l_blocking_ast itself.
1989 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1990 struct ldlm_lock_desc *ld,
1991 struct ldlm_lock *lock,
1992 cfs_list_t *cancels, int count,
1993 ldlm_cancel_flags_t cancel_flags)
1997 if (cancels && count == 0)
2000 if (cancel_flags & LCF_ASYNC) {
2001 struct ldlm_bl_work_item *blwi;
2003 OBD_ALLOC(blwi, sizeof(*blwi));
2006 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2008 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2010 /* if it is synchronous call do minimum mem alloc, as it could
2011 * be triggered from kernel shrinker
2013 struct ldlm_bl_work_item blwi;
2015 memset(&blwi, 0, sizeof(blwi));
2016 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2017 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2023 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2024 struct ldlm_lock *lock)
2027 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2033 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2034 cfs_list_t *cancels, int count,
2035 ldlm_cancel_flags_t cancel_flags)
2038 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2044 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2045 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2047 struct obd_device *obd = req->rq_export->exp_obd;
2054 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2056 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2058 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2060 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2063 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2065 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2067 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2070 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2073 /* We are responsible for swabbing contents of val */
2075 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2076 /* Pass it on to mdc (the "export" in this case) */
2077 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2079 sizeof(KEY_HSM_COPYTOOL_SEND),
2080 KEY_HSM_COPYTOOL_SEND,
2083 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2088 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2089 const char *msg, int rc,
2090 struct lustre_handle *handle)
2092 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2093 "%s: [nid %s] [rc %d] [lock "LPX64"]",
2094 msg, libcfs_id2str(req->rq_peer), rc,
2095 handle ? handle->cookie : 0);
2096 if (req->rq_no_reply)
2097 CWARN("No reply was sent, maybe cause bug 21636.\n");
2099 CWARN("Send reply failed, maybe cause bug 21636.\n");
2102 static int ldlm_handle_qc_callback(struct ptlrpc_request *req)
2104 struct obd_quotactl *oqctl;
2105 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2107 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2108 if (oqctl == NULL) {
2109 CERROR("Can't unpack obd_quotactl\n");
2113 oqctl->qc_stat = ptlrpc_status_ntoh(oqctl->qc_stat);
2115 cli->cl_qchk_stat = oqctl->qc_stat;
2119 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2120 static int ldlm_callback_handler(struct ptlrpc_request *req)
2122 struct ldlm_namespace *ns;
2123 struct ldlm_request *dlm_req;
2124 struct ldlm_lock *lock;
2128 /* Requests arrive in sender's byte order. The ptlrpc service
2129 * handler has already checked and, if necessary, byte-swapped the
2130 * incoming request message body, but I am responsible for the
2131 * message buffers. */
2133 /* do nothing for sec context finalize */
2134 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2137 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2139 if (req->rq_export == NULL) {
2140 rc = ldlm_callback_reply(req, -ENOTCONN);
2141 ldlm_callback_errmsg(req, "Operate on unconnected server",
2146 LASSERT(req->rq_export != NULL);
2147 LASSERT(req->rq_export->exp_obd != NULL);
2149 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2150 case LDLM_BL_CALLBACK:
2151 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET))
2154 case LDLM_CP_CALLBACK:
2155 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2158 case LDLM_GL_CALLBACK:
2159 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2163 rc = ldlm_handle_setinfo(req);
2164 ldlm_callback_reply(req, rc);
2166 case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
2167 CERROR("shouldn't be handling OBD_LOG_CANCEL on DLM thread\n");
2168 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2169 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2171 rc = llog_origin_handle_cancel(req);
2172 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2174 ldlm_callback_reply(req, rc);
2176 case LLOG_ORIGIN_HANDLE_CREATE:
2177 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2178 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2180 rc = llog_origin_handle_open(req);
2181 ldlm_callback_reply(req, rc);
2183 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2184 req_capsule_set(&req->rq_pill,
2185 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2186 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2188 rc = llog_origin_handle_next_block(req);
2189 ldlm_callback_reply(req, rc);
2191 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2192 req_capsule_set(&req->rq_pill,
2193 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2194 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2196 rc = llog_origin_handle_read_header(req);
2197 ldlm_callback_reply(req, rc);
2199 case LLOG_ORIGIN_HANDLE_CLOSE:
2200 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2202 rc = llog_origin_handle_close(req);
2203 ldlm_callback_reply(req, rc);
2205 case OBD_QC_CALLBACK:
2206 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
2207 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
2209 rc = ldlm_handle_qc_callback(req);
2210 ldlm_callback_reply(req, rc);
2213 CERROR("unknown opcode %u\n",
2214 lustre_msg_get_opc(req->rq_reqmsg));
2215 ldlm_callback_reply(req, -EPROTO);
2219 ns = req->rq_export->exp_obd->obd_namespace;
2220 LASSERT(ns != NULL);
2222 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2224 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2225 if (dlm_req == NULL) {
2226 rc = ldlm_callback_reply(req, -EPROTO);
2227 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2232 /* Force a known safe race, send a cancel to the server for a lock
2233 * which the server has already started a blocking callback on. */
2234 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2235 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2236 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2238 CERROR("ldlm_cli_cancel: %d\n", rc);
2241 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2243 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2244 "disappeared\n", dlm_req->lock_handle[0].cookie);
2245 rc = ldlm_callback_reply(req, -EINVAL);
2246 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2247 &dlm_req->lock_handle[0]);
2251 if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
2252 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2253 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2255 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2256 lock_res_and_lock(lock);
2257 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2259 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2260 /* If somebody cancels lock and cache is already dropped,
2261 * or lock is failed before cp_ast received on client,
2262 * we can tell the server we have no lock. Otherwise, we
2263 * should send cancel after dropping the cache. */
2264 if (((lock->l_flags & LDLM_FL_CANCELING) &&
2265 (lock->l_flags & LDLM_FL_BL_DONE)) ||
2266 (lock->l_flags & LDLM_FL_FAILED)) {
2267 LDLM_DEBUG(lock, "callback on lock "
2268 LPX64" - lock disappeared\n",
2269 dlm_req->lock_handle[0].cookie);
2270 unlock_res_and_lock(lock);
2271 LDLM_LOCK_RELEASE(lock);
2272 rc = ldlm_callback_reply(req, -EINVAL);
2273 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2274 &dlm_req->lock_handle[0]);
2277 /* BL_AST locks are not needed in LRU.
2278 * Let ldlm_cancel_lru() be fast. */
2279 ldlm_lock_remove_from_lru(lock);
2280 lock->l_flags |= LDLM_FL_BL_AST;
2282 unlock_res_and_lock(lock);
2284 /* We want the ost thread to get this reply so that it can respond
2285 * to ost requests (write cache writeback) that might be triggered
2288 * But we'd also like to be able to indicate in the reply that we're
2289 * cancelling right now, because it's unused, or have an intent result
2290 * in the reply, so we might have to push the responsibility for sending
2291 * the reply down into the AST handlers, alas. */
2293 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2294 case LDLM_BL_CALLBACK:
2295 CDEBUG(D_INODE, "blocking ast\n");
2296 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2297 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) {
2298 rc = ldlm_callback_reply(req, 0);
2299 if (req->rq_no_reply || rc)
2300 ldlm_callback_errmsg(req, "Normal process", rc,
2301 &dlm_req->lock_handle[0]);
2303 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2304 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2306 case LDLM_CP_CALLBACK:
2307 CDEBUG(D_INODE, "completion ast\n");
2308 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2309 ldlm_callback_reply(req, 0);
2310 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2312 case LDLM_GL_CALLBACK:
2313 CDEBUG(D_INODE, "glimpse ast\n");
2314 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2315 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2318 LBUG(); /* checked above */
2324 #ifdef HAVE_SERVER_SUPPORT
2326 * Main handler for canceld thread.
2328 * Separated into its own thread to avoid deadlocks.
2330 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2335 /* Requests arrive in sender's byte order. The ptlrpc service
2336 * handler has already checked and, if necessary, byte-swapped the
2337 * incoming request message body, but I am responsible for the
2338 * message buffers. */
2340 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2342 if (req->rq_export == NULL) {
2343 struct ldlm_request *dlm_req;
2345 CERROR("%s from %s arrived at %lu with bad export cookie "
2347 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2348 libcfs_nid2str(req->rq_peer.nid),
2349 req->rq_arrival_time.tv_sec,
2350 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2352 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2353 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2354 dlm_req = req_capsule_client_get(&req->rq_pill,
2356 if (dlm_req != NULL)
2357 ldlm_lock_dump_handle(D_ERROR,
2358 &dlm_req->lock_handle[0]);
2360 ldlm_callback_reply(req, -ENOTCONN);
2364 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2366 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2368 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2369 CDEBUG(D_INODE, "cancel\n");
2370 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2371 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))
2373 rc = ldlm_handle_cancel(req);
2377 case OBD_LOG_CANCEL:
2378 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2379 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2381 rc = llog_origin_handle_cancel(req);
2382 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2384 ldlm_callback_reply(req, rc);
2387 CERROR("invalid opcode %d\n",
2388 lustre_msg_get_opc(req->rq_reqmsg));
2389 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2390 ldlm_callback_reply(req, -EINVAL);
2396 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2397 struct ldlm_lock *lock)
2399 struct ldlm_request *dlm_req;
2400 struct lustre_handle lockh;
2405 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2406 if (dlm_req == NULL)
2409 ldlm_lock2handle(lock, &lockh);
2410 for (i = 0; i < dlm_req->lock_count; i++) {
2411 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2413 DEBUG_REQ(D_RPCTRACE, req,
2414 "Prio raised by lock "LPX64".", lockh.cookie);
2425 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2427 struct ldlm_request *dlm_req;
2432 /* no prolong in recovery */
2433 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2436 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2437 if (dlm_req == NULL)
2440 for (i = 0; i < dlm_req->lock_count; i++) {
2441 struct ldlm_lock *lock;
2443 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2447 rc = !!(lock->l_flags & LDLM_FL_AST_SENT);
2449 LDLM_DEBUG(lock, "hpreq cancel lock");
2450 LDLM_LOCK_PUT(lock);
2459 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2460 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2461 .hpreq_check = ldlm_cancel_hpreq_check,
2465 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2469 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2471 if (req->rq_export == NULL)
2474 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2475 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2476 req->rq_ops = &ldlm_cancel_hpreq_ops;
2481 int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2482 cfs_hlist_node_t *hnode, void *data)
2485 cfs_list_t *rpc_list = data;
2486 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2488 lock_res_and_lock(lock);
2490 if (lock->l_req_mode != lock->l_granted_mode) {
2491 unlock_res_and_lock(lock);
2495 LASSERT(lock->l_resource);
2496 if (lock->l_resource->lr_type != LDLM_IBITS &&
2497 lock->l_resource->lr_type != LDLM_PLAIN) {
2498 unlock_res_and_lock(lock);
2502 if (lock->l_flags & LDLM_FL_AST_SENT) {
2503 unlock_res_and_lock(lock);
2507 LASSERT(lock->l_blocking_ast);
2508 LASSERT(!lock->l_blocking_lock);
2510 lock->l_flags |= LDLM_FL_AST_SENT;
2511 if (lock->l_export && lock->l_export->exp_lock_hash) {
2512 /* NB: it's safe to call cfs_hash_del() even lock isn't
2513 * in exp_lock_hash. */
2514 /* In the function below, .hs_keycmp resolves to
2515 * ldlm_export_lock_keycmp() */
2516 /* coverity[overrun-buffer-val] */
2517 cfs_hash_del(lock->l_export->exp_lock_hash,
2518 &lock->l_remote_handle, &lock->l_exp_hash);
2521 cfs_list_add_tail(&lock->l_rk_ast, rpc_list);
2522 LDLM_LOCK_GET(lock);
2524 unlock_res_and_lock(lock);
2528 void ldlm_revoke_export_locks(struct obd_export *exp)
2530 cfs_list_t rpc_list;
2533 CFS_INIT_LIST_HEAD(&rpc_list);
2534 cfs_hash_for_each_empty(exp->exp_lock_hash,
2535 ldlm_revoke_lock_cb, &rpc_list);
2536 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2537 LDLM_WORK_REVOKE_AST);
2541 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2542 #endif /* HAVE_SERVER_SUPPORT */
2545 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2547 struct ldlm_bl_work_item *blwi = NULL;
2548 static unsigned int num_bl = 0;
2550 spin_lock(&blp->blp_lock);
2551 /* process a request from the blp_list at least every blp_num_threads */
2552 if (!cfs_list_empty(&blp->blp_list) &&
2553 (cfs_list_empty(&blp->blp_prio_list) || num_bl == 0))
2554 blwi = cfs_list_entry(blp->blp_list.next,
2555 struct ldlm_bl_work_item, blwi_entry);
2557 if (!cfs_list_empty(&blp->blp_prio_list))
2558 blwi = cfs_list_entry(blp->blp_prio_list.next,
2559 struct ldlm_bl_work_item,
2563 if (++num_bl >= cfs_atomic_read(&blp->blp_num_threads))
2565 cfs_list_del(&blwi->blwi_entry);
2567 spin_unlock(&blp->blp_lock);
2572 /* This only contains temporary data until the thread starts */
2573 struct ldlm_bl_thread_data {
2574 char bltd_name[CFS_CURPROC_COMM_MAX];
2575 struct ldlm_bl_pool *bltd_blp;
2576 struct completion bltd_comp;
2580 static int ldlm_bl_thread_main(void *arg);
2582 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2584 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2587 init_completion(&bltd.bltd_comp);
2588 bltd.bltd_num = cfs_atomic_read(&blp->blp_num_threads);
2589 snprintf(bltd.bltd_name, sizeof(bltd.bltd_name) - 1,
2590 "ldlm_bl_%02d", bltd.bltd_num);
2591 task = kthread_run(ldlm_bl_thread_main, &bltd, bltd.bltd_name);
2593 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2594 cfs_atomic_read(&blp->blp_num_threads), PTR_ERR(task));
2595 return PTR_ERR(task);
2597 wait_for_completion(&bltd.bltd_comp);
2603 * Main blocking requests processing thread.
2605 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2606 * This thread in the end ends up doing actual call to ->l_blocking_ast
2609 static int ldlm_bl_thread_main(void *arg)
2611 struct ldlm_bl_pool *blp;
2615 struct ldlm_bl_thread_data *bltd = arg;
2617 blp = bltd->bltd_blp;
2619 cfs_atomic_inc(&blp->blp_num_threads);
2620 cfs_atomic_inc(&blp->blp_busy_threads);
2622 complete(&bltd->bltd_comp);
2623 /* cannot use bltd after this, it is only on caller's stack */
2627 struct l_wait_info lwi = { 0 };
2628 struct ldlm_bl_work_item *blwi = NULL;
2631 blwi = ldlm_bl_get_work(blp);
2634 cfs_atomic_dec(&blp->blp_busy_threads);
2635 l_wait_event_exclusive(blp->blp_waitq,
2636 (blwi = ldlm_bl_get_work(blp)) != NULL,
2638 busy = cfs_atomic_inc_return(&blp->blp_busy_threads);
2640 busy = cfs_atomic_read(&blp->blp_busy_threads);
2643 if (blwi->blwi_ns == NULL)
2644 /* added by ldlm_cleanup() */
2647 /* Not fatal if racy and have a few too many threads */
2648 if (unlikely(busy < blp->blp_max_threads &&
2649 busy >= cfs_atomic_read(&blp->blp_num_threads) &&
2650 !blwi->blwi_mem_pressure))
2651 /* discard the return value, we tried */
2652 ldlm_bl_thread_start(blp);
2654 if (blwi->blwi_mem_pressure)
2655 memory_pressure_set();
2657 if (blwi->blwi_count) {
2659 /* The special case when we cancel locks in LRU
2660 * asynchronously, we pass the list of locks here.
2661 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2662 * canceled locally yet. */
2663 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2666 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2669 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2672 if (blwi->blwi_mem_pressure)
2673 memory_pressure_clr();
2675 if (blwi->blwi_flags & LCF_ASYNC)
2676 OBD_FREE(blwi, sizeof(*blwi));
2678 complete(&blwi->blwi_comp);
2681 cfs_atomic_dec(&blp->blp_busy_threads);
2682 cfs_atomic_dec(&blp->blp_num_threads);
2683 complete(&blp->blp_comp);
2689 static int ldlm_setup(void);
2690 static int ldlm_cleanup(void);
2692 int ldlm_get_ref(void)
2696 mutex_lock(&ldlm_ref_mutex);
2697 if (++ldlm_refcount == 1) {
2702 mutex_unlock(&ldlm_ref_mutex);
2706 EXPORT_SYMBOL(ldlm_get_ref);
2708 void ldlm_put_ref(void)
2711 mutex_lock(&ldlm_ref_mutex);
2712 if (ldlm_refcount == 1) {
2713 int rc = ldlm_cleanup();
2715 CERROR("ldlm_cleanup failed: %d\n", rc);
2721 mutex_unlock(&ldlm_ref_mutex);
2725 EXPORT_SYMBOL(ldlm_put_ref);
2728 * Export handle<->lock hash operations.
2731 ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
2733 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2737 ldlm_export_lock_key(cfs_hlist_node_t *hnode)
2739 struct ldlm_lock *lock;
2741 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2742 return &lock->l_remote_handle;
2746 ldlm_export_lock_keycpy(cfs_hlist_node_t *hnode, void *key)
2748 struct ldlm_lock *lock;
2750 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2751 lock->l_remote_handle = *(struct lustre_handle *)key;
2755 ldlm_export_lock_keycmp(const void *key, cfs_hlist_node_t *hnode)
2757 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2761 ldlm_export_lock_object(cfs_hlist_node_t *hnode)
2763 return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2767 ldlm_export_lock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2769 struct ldlm_lock *lock;
2771 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2772 LDLM_LOCK_GET(lock);
2776 ldlm_export_lock_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2778 struct ldlm_lock *lock;
2780 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2781 LDLM_LOCK_RELEASE(lock);
2784 static cfs_hash_ops_t ldlm_export_lock_ops = {
2785 .hs_hash = ldlm_export_lock_hash,
2786 .hs_key = ldlm_export_lock_key,
2787 .hs_keycmp = ldlm_export_lock_keycmp,
2788 .hs_keycpy = ldlm_export_lock_keycpy,
2789 .hs_object = ldlm_export_lock_object,
2790 .hs_get = ldlm_export_lock_get,
2791 .hs_put = ldlm_export_lock_put,
2792 .hs_put_locked = ldlm_export_lock_put,
2795 int ldlm_init_export(struct obd_export *exp)
2799 exp->exp_lock_hash =
2800 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2801 HASH_EXP_LOCK_CUR_BITS,
2802 HASH_EXP_LOCK_MAX_BITS,
2803 HASH_EXP_LOCK_BKT_BITS, 0,
2804 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2805 &ldlm_export_lock_ops,
2806 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2807 CFS_HASH_NBLK_CHANGE);
2809 if (!exp->exp_lock_hash)
2814 EXPORT_SYMBOL(ldlm_init_export);
2816 void ldlm_destroy_export(struct obd_export *exp)
2819 cfs_hash_putref(exp->exp_lock_hash);
2820 exp->exp_lock_hash = NULL;
2822 ldlm_destroy_flock_export(exp);
2825 EXPORT_SYMBOL(ldlm_destroy_export);
2827 static int ldlm_setup(void)
2829 static struct ptlrpc_service_conf conf;
2830 struct ldlm_bl_pool *blp = NULL;
2837 if (ldlm_state != NULL)
2840 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2841 if (ldlm_state == NULL)
2845 rc = ldlm_proc_setup();
2850 memset(&conf, 0, sizeof(conf));
2851 conf = (typeof(conf)) {
2852 .psc_name = "ldlm_cbd",
2853 .psc_watchdog_factor = 2,
2855 .bc_nbufs = LDLM_CLIENT_NBUFS,
2856 .bc_buf_size = LDLM_BUFSIZE,
2857 .bc_req_max_size = LDLM_MAXREQSIZE,
2858 .bc_rep_max_size = LDLM_MAXREPSIZE,
2859 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2860 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2863 .tc_thr_name = "ldlm_cb",
2864 .tc_thr_factor = LDLM_THR_FACTOR,
2865 .tc_nthrs_init = LDLM_NTHRS_INIT,
2866 .tc_nthrs_base = LDLM_NTHRS_BASE,
2867 .tc_nthrs_max = LDLM_NTHRS_MAX,
2868 .tc_nthrs_user = ldlm_num_threads,
2869 .tc_cpu_affinity = 1,
2870 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2873 .cc_pattern = ldlm_cpts,
2876 .so_req_handler = ldlm_callback_handler,
2879 ldlm_state->ldlm_cb_service = \
2880 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2881 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2882 CERROR("failed to start service\n");
2883 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2884 ldlm_state->ldlm_cb_service = NULL;
2888 #ifdef HAVE_SERVER_SUPPORT
2889 memset(&conf, 0, sizeof(conf));
2890 conf = (typeof(conf)) {
2891 .psc_name = "ldlm_canceld",
2892 .psc_watchdog_factor = 6,
2894 .bc_nbufs = LDLM_SERVER_NBUFS,
2895 .bc_buf_size = LDLM_BUFSIZE,
2896 .bc_req_max_size = LDLM_MAXREQSIZE,
2897 .bc_rep_max_size = LDLM_MAXREPSIZE,
2898 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
2899 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
2903 .tc_thr_name = "ldlm_cn",
2904 .tc_thr_factor = LDLM_THR_FACTOR,
2905 .tc_nthrs_init = LDLM_NTHRS_INIT,
2906 .tc_nthrs_base = LDLM_NTHRS_BASE,
2907 .tc_nthrs_max = LDLM_NTHRS_MAX,
2908 .tc_nthrs_user = ldlm_num_threads,
2909 .tc_cpu_affinity = 1,
2910 .tc_ctx_tags = LCT_MD_THREAD | \
2915 .cc_pattern = ldlm_cpts,
2918 .so_req_handler = ldlm_cancel_handler,
2919 .so_hpreq_handler = ldlm_hpreq_handler,
2922 ldlm_state->ldlm_cancel_service = \
2923 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2924 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2925 CERROR("failed to start service\n");
2926 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2927 ldlm_state->ldlm_cancel_service = NULL;
2932 OBD_ALLOC(blp, sizeof(*blp));
2934 GOTO(out, rc = -ENOMEM);
2935 ldlm_state->ldlm_bl_pool = blp;
2937 spin_lock_init(&blp->blp_lock);
2938 CFS_INIT_LIST_HEAD(&blp->blp_list);
2939 CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2940 cfs_waitq_init(&blp->blp_waitq);
2941 cfs_atomic_set(&blp->blp_num_threads, 0);
2942 cfs_atomic_set(&blp->blp_busy_threads, 0);
2945 if (ldlm_num_threads == 0) {
2946 blp->blp_min_threads = LDLM_NTHRS_INIT;
2947 blp->blp_max_threads = LDLM_NTHRS_MAX;
2949 blp->blp_min_threads = blp->blp_max_threads = \
2950 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
2954 for (i = 0; i < blp->blp_min_threads; i++) {
2955 rc = ldlm_bl_thread_start(blp);
2960 # ifdef HAVE_SERVER_SUPPORT
2961 CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2962 expired_lock_thread.elt_state = ELT_STOPPED;
2963 cfs_waitq_init(&expired_lock_thread.elt_waitq);
2965 CFS_INIT_LIST_HEAD(&waiting_locks_list);
2966 spin_lock_init(&waiting_locks_spinlock);
2967 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2969 rc = PTR_ERR(kthread_run(expired_lock_main, NULL, "ldlm_elt"));
2970 if (IS_ERR_VALUE(rc)) {
2971 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2975 cfs_wait_event(expired_lock_thread.elt_waitq,
2976 expired_lock_thread.elt_state == ELT_READY);
2977 # endif /* HAVE_SERVER_SUPPORT */
2979 rc = ldlm_pools_init();
2981 CERROR("Failed to initialize LDLM pools: %d\n", rc);
2992 static int ldlm_cleanup(void)
2996 if (!cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2997 !cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2998 CERROR("ldlm still has namespaces; clean these up first.\n");
2999 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3000 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3007 if (ldlm_state->ldlm_bl_pool != NULL) {
3008 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3010 while (cfs_atomic_read(&blp->blp_num_threads) > 0) {
3011 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3013 init_completion(&blp->blp_comp);
3015 spin_lock(&blp->blp_lock);
3016 cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3017 cfs_waitq_signal(&blp->blp_waitq);
3018 spin_unlock(&blp->blp_lock);
3020 wait_for_completion(&blp->blp_comp);
3023 OBD_FREE(blp, sizeof(*blp));
3025 #endif /* __KERNEL__ */
3027 if (ldlm_state->ldlm_cb_service != NULL)
3028 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3029 # ifdef HAVE_SERVER_SUPPORT
3030 if (ldlm_state->ldlm_cancel_service != NULL)
3031 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3035 ldlm_proc_cleanup();
3037 # ifdef HAVE_SERVER_SUPPORT
3038 if (expired_lock_thread.elt_state != ELT_STOPPED) {
3039 expired_lock_thread.elt_state = ELT_TERMINATE;
3040 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
3041 cfs_wait_event(expired_lock_thread.elt_waitq,
3042 expired_lock_thread.elt_state == ELT_STOPPED);
3045 #endif /* __KERNEL__ */
3047 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3055 mutex_init(&ldlm_ref_mutex);
3056 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
3057 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
3058 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3059 sizeof(struct ldlm_resource), 0,
3060 SLAB_HWCACHE_ALIGN, NULL);
3061 if (ldlm_resource_slab == NULL)
3064 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3065 sizeof(struct ldlm_lock), 0,
3066 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3067 if (ldlm_lock_slab == NULL) {
3068 kmem_cache_destroy(ldlm_resource_slab);
3072 ldlm_interval_slab = kmem_cache_create("interval_node",
3073 sizeof(struct ldlm_interval),
3074 0, SLAB_HWCACHE_ALIGN, NULL);
3075 if (ldlm_interval_slab == NULL) {
3076 kmem_cache_destroy(ldlm_resource_slab);
3077 kmem_cache_destroy(ldlm_lock_slab);
3080 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3081 class_export_dump_hook = ldlm_dump_export_locks;
3086 void ldlm_exit(void)
3089 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3090 kmem_cache_destroy(ldlm_resource_slab);
3092 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3093 * synchronize_rcu() to wait a grace period elapsed, so that
3094 * ldlm_lock_free() get a chance to be called. */
3097 kmem_cache_destroy(ldlm_lock_slab);
3098 kmem_cache_destroy(ldlm_interval_slab);