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, 2015, 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
44 #include <linux/kthread.h>
45 #include <libcfs/libcfs.h>
46 #include <lustre_dlm.h>
47 #include <obd_class.h>
48 #include <libcfs/list.h>
49 #include "ldlm_internal.h"
51 static int ldlm_num_threads;
52 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
53 "number of DLM service threads to start");
55 static char *ldlm_cpts;
56 CFS_MODULE_PARM(ldlm_cpts, "s", charp, 0444,
57 "CPU partitions ldlm threads should run on");
59 static struct mutex ldlm_ref_mutex;
60 static int ldlm_refcount;
62 struct ldlm_cb_async_args {
63 struct ldlm_cb_set_arg *ca_set_arg;
64 struct ldlm_lock *ca_lock;
69 static struct ldlm_state *ldlm_state;
71 inline cfs_time_t round_timeout(cfs_time_t timeout)
73 return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
76 /* timeout for initial callback (AST) reply (bz10399) */
77 static inline unsigned int ldlm_get_rq_timeout(void)
80 unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
82 return timeout < 1 ? 1 : timeout;
87 #define ELT_TERMINATE 2
93 * blp_prio_list is used for callbacks that should be handled
94 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
97 struct list_head blp_prio_list;
100 * blp_list is used for all other callbacks which are likely
101 * to take longer to process.
103 struct list_head blp_list;
105 wait_queue_head_t blp_waitq;
106 struct completion blp_comp;
107 atomic_t blp_num_threads;
108 atomic_t blp_busy_threads;
113 struct ldlm_bl_work_item {
114 struct list_head blwi_entry;
115 struct ldlm_namespace *blwi_ns;
116 struct ldlm_lock_desc blwi_ld;
117 struct ldlm_lock *blwi_lock;
118 struct list_head blwi_head;
120 struct completion blwi_comp;
121 enum ldlm_cancel_flags blwi_flags;
122 int blwi_mem_pressure;
125 #ifdef HAVE_SERVER_SUPPORT
128 * Protects both waiting_locks_list and expired_lock_thread.
130 static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */
133 * List for contended locks.
135 * As soon as a lock is contended, it gets placed on this list and
136 * expected time to get a response is filled in the lock. A special
137 * thread walks the list looking for locks that should be released and
138 * schedules client evictions for those that have not been released in
141 * All access to it should be under waiting_locks_spinlock.
143 static struct list_head waiting_locks_list;
144 static struct timer_list waiting_locks_timer;
146 static struct expired_lock_thread {
147 wait_queue_head_t elt_waitq;
150 struct list_head elt_expired_locks;
151 } expired_lock_thread;
153 static inline int have_expired_locks(void)
158 spin_lock_bh(&waiting_locks_spinlock);
159 need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
160 spin_unlock_bh(&waiting_locks_spinlock);
166 * Check expired lock list for expired locks and time them out.
168 static int expired_lock_main(void *arg)
170 struct list_head *expired = &expired_lock_thread.elt_expired_locks;
171 struct l_wait_info lwi = { 0 };
176 expired_lock_thread.elt_state = ELT_READY;
177 wake_up(&expired_lock_thread.elt_waitq);
180 l_wait_event(expired_lock_thread.elt_waitq,
181 have_expired_locks() ||
182 expired_lock_thread.elt_state == ELT_TERMINATE,
185 spin_lock_bh(&waiting_locks_spinlock);
186 if (expired_lock_thread.elt_dump) {
187 struct libcfs_debug_msg_data msgdata = {
188 .msg_file = __FILE__,
189 .msg_fn = "waiting_locks_callback",
190 .msg_line = expired_lock_thread.elt_dump };
191 spin_unlock_bh(&waiting_locks_spinlock);
193 /* from waiting_locks_callback, but not in timer */
194 libcfs_debug_dumplog();
195 libcfs_run_lbug_upcall(&msgdata);
197 spin_lock_bh(&waiting_locks_spinlock);
198 expired_lock_thread.elt_dump = 0;
203 while (!list_empty(expired)) {
204 struct obd_export *export;
205 struct ldlm_lock *lock;
207 lock = list_entry(expired->next, struct ldlm_lock,
209 if ((void *)lock < LP_POISON + PAGE_CACHE_SIZE &&
210 (void *)lock >= LP_POISON) {
211 spin_unlock_bh(&waiting_locks_spinlock);
212 CERROR("free lock on elt list %p\n", lock);
215 list_del_init(&lock->l_pending_chain);
216 if ((void *)lock->l_export <
217 LP_POISON + PAGE_CACHE_SIZE &&
218 (void *)lock->l_export >= LP_POISON) {
219 CERROR("lock with free export on elt list %p\n",
221 lock->l_export = NULL;
222 LDLM_ERROR(lock, "free export");
223 /* release extra ref grabbed by
224 * ldlm_add_waiting_lock() or
225 * ldlm_failed_ast() */
226 LDLM_LOCK_RELEASE(lock);
230 if (ldlm_is_destroyed(lock)) {
231 /* release the lock refcount where
232 * waiting_locks_callback() founds */
233 LDLM_LOCK_RELEASE(lock);
236 export = class_export_lock_get(lock->l_export, lock);
237 spin_unlock_bh(&waiting_locks_spinlock);
239 spin_lock_bh(&export->exp_bl_list_lock);
240 list_del_init(&lock->l_exp_list);
241 spin_unlock_bh(&export->exp_bl_list_lock);
244 class_fail_export(export);
245 class_export_lock_put(export, lock);
247 /* release extra ref grabbed by ldlm_add_waiting_lock()
248 * or ldlm_failed_ast() */
249 LDLM_LOCK_RELEASE(lock);
251 spin_lock_bh(&waiting_locks_spinlock);
253 spin_unlock_bh(&waiting_locks_spinlock);
255 if (do_dump && obd_dump_on_eviction) {
256 CERROR("dump the log upon eviction\n");
257 libcfs_debug_dumplog();
260 if (expired_lock_thread.elt_state == ELT_TERMINATE)
264 expired_lock_thread.elt_state = ELT_STOPPED;
265 wake_up(&expired_lock_thread.elt_waitq);
269 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
270 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds);
273 * Check if there is a request in the export request list
274 * which prevents the lock canceling.
276 static int ldlm_lock_busy(struct ldlm_lock *lock)
278 struct ptlrpc_request *req;
282 if (lock->l_export == NULL)
285 spin_lock_bh(&lock->l_export->exp_rpc_lock);
286 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
288 if (req->rq_ops->hpreq_lock_match) {
289 match = req->rq_ops->hpreq_lock_match(req, lock);
294 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
298 /* This is called from within a timer interrupt and cannot schedule */
299 static void waiting_locks_callback(unsigned long unused)
301 struct ldlm_lock *lock;
304 spin_lock_bh(&waiting_locks_spinlock);
305 while (!list_empty(&waiting_locks_list)) {
306 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
308 if (cfs_time_after(lock->l_callback_timeout,
309 cfs_time_current()) ||
310 (lock->l_req_mode == LCK_GROUP))
313 /* Check if we need to prolong timeout */
314 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
315 ldlm_lock_busy(lock)) {
318 if (lock->l_pending_chain.next == &waiting_locks_list)
323 spin_unlock_bh(&waiting_locks_spinlock);
324 LDLM_DEBUG(lock, "prolong the busy lock");
325 ldlm_refresh_waiting_lock(lock,
326 ldlm_bl_timeout(lock) >> 1);
327 spin_lock_bh(&waiting_locks_spinlock);
330 LDLM_LOCK_RELEASE(lock);
334 LDLM_LOCK_RELEASE(lock);
337 ldlm_lock_to_ns(lock)->ns_timeouts++;
338 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
339 "evicting client at %s ",
340 cfs_time_current_sec() - lock->l_last_activity,
342 lock->l_export->exp_connection->c_peer.nid));
344 /* no needs to take an extra ref on the lock since it was in
345 * the waiting_locks_list and ldlm_add_waiting_lock()
346 * already grabbed a ref */
347 list_del(&lock->l_pending_chain);
348 list_add(&lock->l_pending_chain,
349 &expired_lock_thread.elt_expired_locks);
353 if (!list_empty(&expired_lock_thread.elt_expired_locks)) {
354 if (obd_dump_on_timeout && need_dump)
355 expired_lock_thread.elt_dump = __LINE__;
357 wake_up(&expired_lock_thread.elt_waitq);
361 * Make sure the timer will fire again if we have any locks
364 if (!list_empty(&waiting_locks_list)) {
365 cfs_time_t timeout_rounded;
366 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
368 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
369 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
371 spin_unlock_bh(&waiting_locks_spinlock);
375 * Add lock to the list of contended locks.
377 * Indicate that we're waiting for a client to call us back cancelling a given
378 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
379 * timer to fire appropriately. (We round up to the next second, to avoid
380 * floods of timer firings during periods of high lock contention and traffic).
381 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
382 * if it has been added to the waiting list (1 is returned).
384 * Called with the namespace lock held.
386 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
389 cfs_time_t timeout_rounded;
391 if (!list_empty(&lock->l_pending_chain))
394 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
395 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
398 timeout = cfs_time_shift(seconds);
399 if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
400 lock->l_callback_timeout = timeout;
402 timeout_rounded = round_timeout(lock->l_callback_timeout);
404 if (cfs_time_before(timeout_rounded,
405 cfs_timer_deadline(&waiting_locks_timer)) ||
406 !cfs_timer_is_armed(&waiting_locks_timer)) {
407 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
409 /* if the new lock has a shorter timeout than something earlier on
410 the list, we'll wait the longer amount of time; no big deal. */
412 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
416 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
418 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
419 if (list_empty(&lock->l_exp_list)) {
420 if (lock->l_granted_mode != lock->l_req_mode)
421 list_add_tail(&lock->l_exp_list,
422 &lock->l_export->exp_bl_list);
424 list_add(&lock->l_exp_list,
425 &lock->l_export->exp_bl_list);
427 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
429 /* A blocked lock is added. Adjust the position in
430 * the stale list if the export is in the list.
431 * If export is stale and not in the list - it is being
432 * processed and will be placed on the right position
433 * on obd_stale_export_put(). */
434 if (!list_empty(&lock->l_export->exp_stale_list))
435 obd_stale_export_adjust(lock->l_export);
438 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
441 int timeout = ldlm_bl_timeout(lock);
443 /* NB: must be called with hold of lock_res_and_lock() */
444 LASSERT(ldlm_is_res_locked(lock));
445 LASSERT(!ldlm_is_cancel_on_block(lock));
447 /* Do not put cross-MDT lock in the waiting list, since we
448 * will not evict it due to timeout for now */
449 if (lock->l_export != NULL &&
450 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
453 spin_lock_bh(&waiting_locks_spinlock);
454 if (ldlm_is_cancel(lock)) {
455 spin_unlock_bh(&waiting_locks_spinlock);
459 if (ldlm_is_destroyed(lock)) {
460 static cfs_time_t next;
462 spin_unlock_bh(&waiting_locks_spinlock);
463 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
464 if (cfs_time_after(cfs_time_current(), next)) {
465 next = cfs_time_shift(14400);
466 libcfs_debug_dumpstack(NULL);
471 ldlm_set_waited(lock);
472 lock->l_last_activity = cfs_time_current_sec();
473 ret = __ldlm_add_waiting_lock(lock, timeout);
475 /* grab ref on the lock if it has been added to the
479 spin_unlock_bh(&waiting_locks_spinlock);
482 ldlm_add_blocked_lock(lock);
484 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
485 ret == 0 ? "not re-" : "", timeout,
486 AT_OFF ? "off" : "on");
491 * Remove a lock from the pending list, likely because it had its cancellation
492 * callback arrive without incident. This adjusts the lock-timeout timer if
493 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
494 * As done by ldlm_del_waiting_lock(), the caller must release the lock
495 * reference when the lock is removed from any list (1 is returned).
497 * Called with namespace lock held.
499 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
501 struct list_head *list_next;
503 if (list_empty(&lock->l_pending_chain))
506 list_next = lock->l_pending_chain.next;
507 if (lock->l_pending_chain.prev == &waiting_locks_list) {
508 /* Removing the head of the list, adjust timer. */
509 if (list_next == &waiting_locks_list) {
510 /* No more, just cancel. */
511 cfs_timer_disarm(&waiting_locks_timer);
513 struct ldlm_lock *next;
514 next = list_entry(list_next, struct ldlm_lock,
516 cfs_timer_arm(&waiting_locks_timer,
517 round_timeout(next->l_callback_timeout));
520 list_del_init(&lock->l_pending_chain);
525 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
529 if (lock->l_export == NULL) {
530 /* We don't have a "waiting locks list" on clients. */
531 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
535 spin_lock_bh(&waiting_locks_spinlock);
536 ret = __ldlm_del_waiting_lock(lock);
537 ldlm_clear_waited(lock);
538 spin_unlock_bh(&waiting_locks_spinlock);
540 /* remove the lock out of export blocking list */
541 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
542 list_del_init(&lock->l_exp_list);
543 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
546 /* release lock ref if it has indeed been removed
548 LDLM_LOCK_RELEASE(lock);
551 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
556 * Prolong the contended lock waiting time.
558 * Called with namespace lock held.
560 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
562 if (lock->l_export == NULL) {
563 /* We don't have a "waiting locks list" on clients. */
564 LDLM_DEBUG(lock, "client lock: no-op");
568 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
569 /* We don't have a "waiting locks list" on OSP. */
570 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
574 spin_lock_bh(&waiting_locks_spinlock);
576 if (list_empty(&lock->l_pending_chain)) {
577 spin_unlock_bh(&waiting_locks_spinlock);
578 LDLM_DEBUG(lock, "wasn't waiting");
582 /* we remove/add the lock to the waiting list, so no needs to
583 * release/take a lock reference */
584 __ldlm_del_waiting_lock(lock);
585 __ldlm_add_waiting_lock(lock, timeout);
586 spin_unlock_bh(&waiting_locks_spinlock);
588 LDLM_DEBUG(lock, "refreshed");
591 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
593 #else /* HAVE_SERVER_SUPPORT */
595 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
600 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
605 #endif /* !HAVE_SERVER_SUPPORT */
607 #ifdef HAVE_SERVER_SUPPORT
610 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
611 * lock cancel, and their replies). Used for lock callback timeout and AST
614 * \param[in] lock lock which is getting the blocking callback
616 * \retval timeout in seconds to wait for the client reply
618 unsigned int ldlm_bl_timeout(struct ldlm_lock *lock)
620 unsigned int timeout;
623 return obd_timeout / 2;
625 /* Since these are non-updating timeouts, we should be conservative.
626 * Take more than usually, 150%
627 * It would be nice to have some kind of "early reply" mechanism for
628 * lock callbacks too... */
629 timeout = at_get(&lock->l_export->exp_bl_lock_at);
630 return max(timeout + (timeout >> 1), ldlm_enqueue_min);
632 EXPORT_SYMBOL(ldlm_bl_timeout);
635 * Perform lock cleanup if AST sending failed.
637 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
638 const char *ast_type)
640 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
641 "to a lock %s callback time out: rc %d\n",
642 lock->l_export->exp_obd->obd_name,
643 obd_export_nid2str(lock->l_export), ast_type, rc);
645 if (obd_dump_on_timeout)
646 libcfs_debug_dumplog();
647 spin_lock_bh(&waiting_locks_spinlock);
648 if (__ldlm_del_waiting_lock(lock) == 0)
649 /* the lock was not in any list, grab an extra ref before adding
650 * the lock to the expired list */
652 list_add(&lock->l_pending_chain,
653 &expired_lock_thread.elt_expired_locks);
654 wake_up(&expired_lock_thread.elt_waitq);
655 spin_unlock_bh(&waiting_locks_spinlock);
659 * Perform lock cleanup if AST reply came with error.
661 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
662 struct ptlrpc_request *req, int rc,
663 const char *ast_type)
665 lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
667 if (!req->rq_replied || (rc && rc != -EINVAL)) {
668 if (lock->l_export && lock->l_export->exp_libclient) {
669 LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
670 " timeout, just cancelling lock", ast_type,
671 libcfs_nid2str(peer.nid));
672 ldlm_lock_cancel(lock);
674 } else if (ldlm_is_cancel(lock)) {
675 LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
676 "cancel was received (AST reply lost?)",
677 ast_type, libcfs_nid2str(peer.nid));
678 ldlm_lock_cancel(lock);
681 LDLM_ERROR(lock, "client (nid %s) %s %s AST "
682 "(req status %d rc %d), evict it",
683 libcfs_nid2str(peer.nid),
684 req->rq_replied ? "returned error from" :
685 "failed to reply to",
687 (req->rq_repmsg != NULL) ?
688 lustre_msg_get_status(req->rq_repmsg) : 0,
690 ldlm_failed_ast(lock, rc, ast_type);
696 struct ldlm_resource *res = lock->l_resource;
698 LDLM_DEBUG(lock, "client (nid %s) returned %d"
699 " from %s AST - normal race",
700 libcfs_nid2str(peer.nid),
702 lustre_msg_get_status(req->rq_repmsg) : -1,
705 /* update lvbo to return proper attributes.
707 ldlm_resource_getref(res);
708 ldlm_res_lvbo_update(res, NULL, 1);
709 ldlm_resource_putref(res);
711 ldlm_lock_cancel(lock);
718 static int ldlm_cb_interpret(const struct lu_env *env,
719 struct ptlrpc_request *req, void *data, int rc)
721 struct ldlm_cb_async_args *ca = data;
722 struct ldlm_lock *lock = ca->ca_lock;
723 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
726 LASSERT(lock != NULL);
729 case LDLM_GL_CALLBACK:
730 /* Update the LVB from disk if the AST failed
731 * (this is a legal race)
733 * - Glimpse callback of local lock just returns
734 * -ELDLM_NO_LOCK_DATA.
735 * - Glimpse callback of remote lock might return
736 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
738 if (rc == -ELDLM_NO_LOCK_DATA) {
739 LDLM_DEBUG(lock, "lost race - client has a lock but no "
741 ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
742 } else if (rc != 0) {
743 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
745 rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
748 case LDLM_BL_CALLBACK:
750 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
752 case LDLM_CP_CALLBACK:
754 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
757 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
762 /* release extra reference taken in ldlm_ast_fini() */
763 LDLM_LOCK_RELEASE(lock);
766 atomic_inc(&arg->restart);
771 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
773 struct ldlm_cb_async_args *ca = data;
774 struct ldlm_lock *lock = ca->ca_lock;
776 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
779 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
780 struct ldlm_cb_set_arg *arg,
781 struct ldlm_lock *lock,
787 if (unlikely(instant_cancel)) {
788 rc = ptl_send_rpc(req, 1);
789 ptlrpc_req_finished(req);
791 atomic_inc(&arg->restart);
794 ptlrpc_set_add_req(arg->set, req);
801 * Check if there are requests in the export request list which prevent
802 * the lock canceling and make these requests high priority ones.
804 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
806 struct ptlrpc_request *req;
809 if (lock->l_export == NULL) {
810 LDLM_DEBUG(lock, "client lock: no-op");
814 spin_lock_bh(&lock->l_export->exp_rpc_lock);
815 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
817 /* Do not process requests that were not yet added to there
818 * incoming queue or were already removed from there for
819 * processing. We evaluate ptlrpc_nrs_req_can_move() without
820 * holding svcpt->scp_req_lock, and then redo the check with
821 * the lock held once we need to obtain a reliable result.
823 if (ptlrpc_nrs_req_can_move(req) &&
824 req->rq_ops->hpreq_lock_match &&
825 req->rq_ops->hpreq_lock_match(req, lock))
826 ptlrpc_nrs_req_hp_move(req);
828 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
833 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
834 * enqueued server lock conflicts with given one.
836 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
837 * to wait for client response.
839 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
840 struct ldlm_lock_desc *desc,
841 void *data, int flag)
843 struct ldlm_cb_async_args *ca;
844 struct ldlm_cb_set_arg *arg = data;
845 struct ldlm_request *body;
846 struct ptlrpc_request *req;
847 int instant_cancel = 0;
851 if (flag == LDLM_CB_CANCELING)
852 /* Don't need to do anything here. */
856 LASSERT(data != NULL);
857 if (lock->l_export->exp_obd->obd_recovering != 0)
858 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
860 ldlm_lock_reorder_req(lock);
862 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
863 &RQF_LDLM_BL_CALLBACK,
864 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
868 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
869 ca = ptlrpc_req_async_args(req);
870 ca->ca_set_arg = arg;
873 req->rq_interpret_reply = ldlm_cb_interpret;
875 lock_res_and_lock(lock);
876 if (ldlm_is_destroyed(lock)) {
877 /* What's the point? */
878 unlock_res_and_lock(lock);
879 ptlrpc_req_finished(req);
883 if (lock->l_granted_mode != lock->l_req_mode) {
884 /* this blocking AST will be communicated as part of the
885 * completion AST instead */
886 ldlm_add_blocked_lock(lock);
887 ldlm_set_waited(lock);
888 unlock_res_and_lock(lock);
890 ptlrpc_req_finished(req);
891 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
895 if (ldlm_is_cancel_on_block(lock))
898 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
899 body->lock_handle[0] = lock->l_remote_handle;
900 body->lock_desc = *desc;
901 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
903 LDLM_DEBUG(lock, "server preparing blocking AST");
905 ptlrpc_request_set_replen(req);
906 ldlm_set_cbpending(lock);
907 if (instant_cancel) {
908 unlock_res_and_lock(lock);
909 ldlm_lock_cancel(lock);
911 req->rq_no_resend = 1;
913 LASSERT(lock->l_granted_mode == lock->l_req_mode);
914 ldlm_add_waiting_lock(lock);
915 unlock_res_and_lock(lock);
917 /* Do not resend after lock callback timeout */
918 req->rq_delay_limit = ldlm_bl_timeout(lock);
919 req->rq_resend_cb = ldlm_update_resend;
922 req->rq_send_state = LUSTRE_IMP_FULL;
923 /* ptlrpc_request_alloc_pack already set timeout */
925 req->rq_timeout = ldlm_get_rq_timeout();
927 lock->l_last_activity = cfs_time_current_sec();
929 if (lock->l_export && lock->l_export->exp_nid_stats &&
930 lock->l_export->exp_nid_stats->nid_ldlm_stats)
931 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
932 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
934 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
940 * ->l_completion_ast callback for a remote lock in server namespace.
942 * Sends AST to the client notifying it of lock granting. If initial
943 * lock response was not sent yet, instead of sending another RPC, just
944 * mark the lock as granted and client will understand
946 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
948 struct ldlm_cb_set_arg *arg = data;
949 struct ldlm_request *body;
950 struct ptlrpc_request *req;
951 struct ldlm_cb_async_args *ca;
952 int instant_cancel = 0;
957 LASSERT(lock != NULL);
958 LASSERT(data != NULL);
960 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
961 LDLM_DEBUG(lock, "dropping CP AST");
965 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
966 &RQF_LDLM_CP_CALLBACK);
970 /* server namespace, doesn't need lock */
971 lvb_len = ldlm_lvbo_size(lock);
972 /* LU-3124 & LU-2187: to not return layout in completion AST because
973 * it may deadlock for LU-2187, or client may not have enough space
974 * for large layout. The layout will be returned to client with an
975 * extra RPC to fetch xattr.lov */
976 if (ldlm_has_layout(lock))
979 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
980 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
982 ptlrpc_request_free(req);
986 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
987 ca = ptlrpc_req_async_args(req);
988 ca->ca_set_arg = arg;
991 req->rq_interpret_reply = ldlm_cb_interpret;
992 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
994 body->lock_handle[0] = lock->l_remote_handle;
995 body->lock_flags = ldlm_flags_to_wire(flags);
996 ldlm_lock2desc(lock, &body->lock_desc);
998 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1000 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
1002 /* We still need to send the RPC to wake up the blocked
1003 * enqueue thread on the client.
1005 * Consider old client, there is no better way to notify
1006 * the failure, just zero-sized the LVB, then the client
1007 * will fail out as "-EPROTO". */
1008 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1012 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1017 lock->l_last_activity = cfs_time_current_sec();
1019 LDLM_DEBUG(lock, "server preparing completion AST");
1021 ptlrpc_request_set_replen(req);
1023 req->rq_send_state = LUSTRE_IMP_FULL;
1024 /* ptlrpc_request_pack already set timeout */
1026 req->rq_timeout = ldlm_get_rq_timeout();
1028 /* We only send real blocking ASTs after the lock is granted */
1029 lock_res_and_lock(lock);
1030 if (ldlm_is_ast_sent(lock)) {
1031 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1032 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1033 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1036 /* We might get here prior to ldlm_handle_enqueue setting
1037 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1038 * into waiting list, but this is safe and similar code in
1039 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1040 * that would not only cancel the lock, but will also remove
1041 * it from waiting list */
1042 if (ldlm_is_cancel_on_block(lock)) {
1043 unlock_res_and_lock(lock);
1044 ldlm_lock_cancel(lock);
1047 req->rq_no_resend = 1;
1049 lock_res_and_lock(lock);
1051 /* start the lock-timeout clock */
1052 ldlm_add_waiting_lock(lock);
1053 /* Do not resend after lock callback timeout */
1054 req->rq_delay_limit = ldlm_bl_timeout(lock);
1055 req->rq_resend_cb = ldlm_update_resend;
1058 unlock_res_and_lock(lock);
1060 if (lock->l_export && lock->l_export->exp_nid_stats &&
1061 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1062 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1063 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1065 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1067 RETURN(lvb_len < 0 ? lvb_len : rc);
1071 * Server side ->l_glimpse_ast handler for client locks.
1073 * Sends glimpse AST to the client and waits for reply. Then updates
1074 * lvbo with the result.
1076 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1078 struct ldlm_cb_set_arg *arg = data;
1079 struct ldlm_request *body;
1080 struct ptlrpc_request *req;
1081 struct ldlm_cb_async_args *ca;
1083 struct req_format *req_fmt;
1086 LASSERT(lock != NULL);
1088 if (arg->gl_desc != NULL)
1089 /* There is a glimpse descriptor to pack */
1090 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1092 req_fmt = &RQF_LDLM_GL_CALLBACK;
1094 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1095 req_fmt, LUSTRE_DLM_VERSION,
1101 if (arg->gl_desc != NULL) {
1102 /* copy the GL descriptor */
1103 union ldlm_gl_desc *desc;
1104 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1105 *desc = *arg->gl_desc;
1108 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1109 body->lock_handle[0] = lock->l_remote_handle;
1110 ldlm_lock2desc(lock, &body->lock_desc);
1112 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1113 ca = ptlrpc_req_async_args(req);
1114 ca->ca_set_arg = arg;
1117 /* server namespace, doesn't need lock */
1118 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1119 ldlm_lvbo_size(lock));
1120 ptlrpc_request_set_replen(req);
1122 req->rq_send_state = LUSTRE_IMP_FULL;
1123 /* ptlrpc_request_alloc_pack already set timeout */
1125 req->rq_timeout = ldlm_get_rq_timeout();
1127 lock->l_last_activity = cfs_time_current_sec();
1129 req->rq_interpret_reply = ldlm_cb_interpret;
1131 if (lock->l_export && lock->l_export->exp_nid_stats &&
1132 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1133 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1134 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1136 rc = ldlm_ast_fini(req, arg, lock, 0);
1141 int ldlm_glimpse_locks(struct ldlm_resource *res,
1142 struct list_head *gl_work_list)
1147 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1149 if (rc == -ERESTART)
1150 ldlm_reprocess_all(res);
1154 EXPORT_SYMBOL(ldlm_glimpse_locks);
1156 /* return LDLM lock associated with a lock callback request */
1157 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1159 struct ldlm_cb_async_args *ca;
1160 struct ldlm_lock *lock;
1163 ca = ptlrpc_req_async_args(req);
1166 RETURN(ERR_PTR(-EFAULT));
1170 EXPORT_SYMBOL(ldlm_request_lock);
1172 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1173 struct lprocfs_stats *srv_stats)
1175 int lock_type = 0, op = 0;
1177 lock_type = dlm_req->lock_desc.l_resource.lr_type;
1179 switch (lock_type) {
1181 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1184 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1185 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1187 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1190 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1193 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1201 lprocfs_counter_incr(srv_stats, op);
1207 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1208 * service threads to carry out client lock enqueueing requests.
1210 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1211 struct ptlrpc_request *req,
1212 const struct ldlm_request *dlm_req,
1213 const struct ldlm_callback_suite *cbs)
1215 struct ldlm_reply *dlm_rep;
1217 enum ldlm_error err = ELDLM_OK;
1218 struct ldlm_lock *lock = NULL;
1219 void *cookie = NULL;
1221 struct ldlm_resource *res = NULL;
1224 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1226 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1227 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1229 LASSERT(req->rq_export);
1231 if (ptlrpc_req2svc(req)->srv_stats != NULL)
1232 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1234 if (req->rq_export && req->rq_export->exp_nid_stats &&
1235 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1236 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1237 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1239 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1240 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1241 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1242 dlm_req->lock_desc.l_resource.lr_type);
1243 GOTO(out, rc = -EFAULT);
1246 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1247 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1248 dlm_req->lock_desc.l_req_mode &
1249 (dlm_req->lock_desc.l_req_mode-1))) {
1250 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1251 dlm_req->lock_desc.l_req_mode);
1252 GOTO(out, rc = -EFAULT);
1255 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) {
1256 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1258 DEBUG_REQ(D_ERROR, req,
1259 "PLAIN lock request from IBITS client?");
1260 GOTO(out, rc = -EPROTO);
1262 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1264 DEBUG_REQ(D_ERROR, req,
1265 "IBITS lock request from unaware client?");
1266 GOTO(out, rc = -EPROTO);
1270 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1271 against server's _CONNECT_SUPPORTED flags? (I don't want to use
1272 ibits for mgc/mgs) */
1274 /* INODEBITS_INTEROP: Perform conversion from plain lock to
1275 * inodebits lock if client does not support them. */
1276 if (!(exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) &&
1277 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1278 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1279 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1280 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1281 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1282 dlm_req->lock_desc.l_req_mode = LCK_CR;
1286 if (unlikely((flags & LDLM_FL_REPLAY) ||
1287 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1288 /* Find an existing lock in the per-export lock hash */
1289 /* In the function below, .hs_keycmp resolves to
1290 * ldlm_export_lock_keycmp() */
1291 /* coverity[overrun-buffer-val] */
1292 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1293 (void *)&dlm_req->lock_handle[0]);
1295 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1296 LPX64, lock->l_handle.h_cookie);
1297 flags |= LDLM_FL_RESENT;
1298 GOTO(existing_lock, rc = 0);
1301 if (ldlm_reclaim_full()) {
1302 DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1303 "reject current enqueue request and let the "
1304 "client retry later.\n");
1305 GOTO(out, rc = -EINPROGRESS);
1309 /* The lock's callback data might be set in the policy function */
1310 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1311 dlm_req->lock_desc.l_resource.lr_type,
1312 dlm_req->lock_desc.l_req_mode,
1313 cbs, NULL, 0, LVB_T_NONE);
1320 lock->l_remote_handle = dlm_req->lock_handle[0];
1321 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1323 /* Initialize resource lvb but not for a lock being replayed since
1324 * Client already got lvb sent in this case.
1325 * This must occur early since some policy methods assume resource
1326 * lvb is available (lr_lvb_data != NULL).
1328 res = lock->l_resource;
1329 if (!(flags & LDLM_FL_REPLAY)) {
1330 /* non-replayed lock, delayed lvb init may need to be done */
1331 rc = ldlm_lvbo_init(res);
1333 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1338 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1339 /* Don't enqueue a lock onto the export if it is been disonnected
1340 * due to eviction (bug 3822) or server umount (bug 24324).
1341 * Cancel it now instead. */
1342 if (req->rq_export->exp_disconnected) {
1343 LDLM_ERROR(lock, "lock on disconnected export %p",
1345 GOTO(out, rc = -ENOTCONN);
1348 lock->l_export = class_export_lock_get(req->rq_export, lock);
1349 if (lock->l_export->exp_lock_hash)
1350 cfs_hash_add(lock->l_export->exp_lock_hash,
1351 &lock->l_remote_handle,
1354 /* Inherit the enqueue flags before the operation, because we do not
1355 * keep the res lock on return and next operations (BL AST) may proceed
1357 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1358 LDLM_FL_INHERIT_MASK);
1361 if (flags & LDLM_FL_HAS_INTENT) {
1362 /* In this case, the reply buffer is allocated deep in
1363 * local_lock_enqueue by the policy function. */
1366 /* based on the assumption that lvb size never changes during
1367 * resource life time otherwise it need resource->lr_lock's
1369 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1370 RCL_SERVER, ldlm_lvbo_size(lock));
1372 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1373 GOTO(out, rc = -ENOMEM);
1375 rc = req_capsule_server_pack(&req->rq_pill);
1380 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1381 ldlm_convert_policy_to_local(req->rq_export,
1382 dlm_req->lock_desc.l_resource.lr_type,
1383 &dlm_req->lock_desc.l_policy_data,
1384 &lock->l_policy_data);
1385 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1386 lock->l_req_extent = lock->l_policy_data.l_extent;
1388 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1395 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1397 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1398 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1400 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1401 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1403 /* We never send a blocking AST until the lock is granted, but
1404 * we can tell it right now */
1405 lock_res_and_lock(lock);
1407 /* Now take into account flags to be inherited from original lock
1408 request both in reply to client and in our own lock flags. */
1409 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1410 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1412 /* Don't move a pending lock onto the export if it has already been
1413 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1414 * Cancel it now instead. */
1415 if (unlikely(req->rq_export->exp_disconnected ||
1416 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1417 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1419 } else if (ldlm_is_ast_sent(lock)) {
1420 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1421 if (lock->l_granted_mode == lock->l_req_mode) {
1423 * Only cancel lock if it was granted, because it would
1424 * be destroyed immediately and would never be granted
1425 * in the future, causing timeouts on client. Not
1426 * granted lock will be cancelled immediately after
1427 * sending completion AST.
1429 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1430 unlock_res_and_lock(lock);
1431 ldlm_lock_cancel(lock);
1432 lock_res_and_lock(lock);
1434 ldlm_add_waiting_lock(lock);
1437 /* Make sure we never ever grant usual metadata locks to liblustre
1439 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1440 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1441 req->rq_export->exp_libclient) {
1442 if (unlikely(!ldlm_is_cancel_on_block(lock) ||
1443 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1444 CERROR("Granting sync lock to libclient. "
1445 "req fl %d, rep fl %d, lock fl "LPX64"\n",
1446 dlm_req->lock_flags, dlm_rep->lock_flags,
1448 LDLM_ERROR(lock, "sync lock");
1449 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1450 struct ldlm_intent *it;
1452 it = req_capsule_client_get(&req->rq_pill,
1455 CERROR("This is intent %s ("LPU64")\n",
1456 ldlm_it2str(it->opc), it->opc);
1462 unlock_res_and_lock(lock);
1466 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1467 if (!req->rq_packed_final) {
1468 err = lustre_pack_reply(req, 1, NULL, NULL);
1473 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1474 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1476 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1477 "(err=%d, rc=%d)", err, rc);
1480 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1482 ldlm_lvbo_size(lock) > 0) {
1486 buf = req_capsule_server_get(&req->rq_pill,
1488 LASSERTF(buf != NULL, "req %p, lock %p\n",
1490 buflen = req_capsule_get_size(&req->rq_pill,
1491 &RMF_DLM_LVB, RCL_SERVER);
1492 /* non-replayed lock, delayed lvb init may
1493 * need to be occur now */
1494 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1495 buflen = ldlm_lvbo_fill(lock, buf,
1501 buflen, RCL_SERVER);
1504 } else if (flags & LDLM_FL_REPLAY) {
1505 /* no LVB resend upon replay */
1519 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1520 if (lock->l_export) {
1521 ldlm_lock_cancel(lock);
1523 lock_res_and_lock(lock);
1524 ldlm_resource_unlink_lock(lock);
1525 ldlm_lock_destroy_nolock(lock);
1526 unlock_res_and_lock(lock);
1531 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1532 ldlm_reprocess_all(lock->l_resource);
1534 LDLM_LOCK_RELEASE(lock);
1537 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1544 * Old-style LDLM main entry point for server code enqueue.
1546 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1547 ldlm_completion_callback completion_callback,
1548 ldlm_blocking_callback blocking_callback,
1549 ldlm_glimpse_callback glimpse_callback)
1551 struct ldlm_request *dlm_req;
1552 struct ldlm_callback_suite cbs = {
1553 .lcs_completion = completion_callback,
1554 .lcs_blocking = blocking_callback,
1555 .lcs_glimpse = glimpse_callback
1559 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1560 if (dlm_req != NULL) {
1561 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1562 req, dlm_req, &cbs);
1570 * Main LDLM entry point for server code to process lock conversion requests.
1572 int ldlm_handle_convert0(struct ptlrpc_request *req,
1573 const struct ldlm_request *dlm_req)
1575 struct ldlm_reply *dlm_rep;
1576 struct ldlm_lock *lock;
1580 if (req->rq_export && req->rq_export->exp_nid_stats &&
1581 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1582 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1583 LDLM_CONVERT - LDLM_FIRST_OPC);
1585 rc = req_capsule_server_pack(&req->rq_pill);
1589 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1590 dlm_rep->lock_flags = dlm_req->lock_flags;
1592 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1594 req->rq_status = LUSTRE_EINVAL;
1598 LDLM_DEBUG(lock, "server-side convert handler START");
1600 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1601 &dlm_rep->lock_flags);
1603 if (ldlm_del_waiting_lock(lock))
1604 LDLM_DEBUG(lock, "converted waiting lock");
1607 req->rq_status = LUSTRE_EDEADLK;
1612 if (!req->rq_status)
1613 ldlm_reprocess_all(lock->l_resource);
1614 LDLM_DEBUG(lock, "server-side convert handler END");
1615 LDLM_LOCK_PUT(lock);
1617 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1623 * Old-style main LDLM entry point for server code to process lock conversion
1626 int ldlm_handle_convert(struct ptlrpc_request *req)
1629 struct ldlm_request *dlm_req;
1631 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1632 if (dlm_req != NULL) {
1633 rc = ldlm_handle_convert0(req, dlm_req);
1635 CERROR ("Can't unpack dlm_req\n");
1642 * Cancel all the locks whose handles are packed into ldlm_request
1644 * Called by server code expecting such combined cancel activity
1647 int ldlm_request_cancel(struct ptlrpc_request *req,
1648 const struct ldlm_request *dlm_req,
1649 int first, enum lustre_at_flags flags)
1651 struct ldlm_resource *res, *pres = NULL;
1652 struct ldlm_lock *lock;
1653 int i, count, done = 0;
1656 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1660 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1663 /* There is no lock on the server at the replay time,
1664 * skip lock cancelling to make replay tests to pass. */
1665 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1668 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1669 "starting at %d", count, first);
1671 for (i = first; i < count; i++) {
1672 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1674 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1675 "lock (cookie "LPU64")",
1676 dlm_req->lock_handle[i].cookie);
1680 res = lock->l_resource;
1683 /* This code is an optimization to only attempt lock
1684 * granting on the resource (that could be CPU-expensive)
1685 * after we are done cancelling lock in that resource. */
1688 ldlm_reprocess_all(pres);
1689 LDLM_RESOURCE_DELREF(pres);
1690 ldlm_resource_putref(pres);
1693 ldlm_resource_getref(res);
1694 LDLM_RESOURCE_ADDREF(res);
1695 ldlm_res_lvbo_update(res, NULL, 1);
1700 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock)) {
1701 long delay = cfs_time_sub(cfs_time_current_sec(),
1702 lock->l_last_activity);
1703 LDLM_DEBUG(lock, "server cancels blocked lock after "
1704 CFS_DURATION_T"s", delay);
1705 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1707 ldlm_lock_cancel(lock);
1708 LDLM_LOCK_PUT(lock);
1711 ldlm_reprocess_all(pres);
1712 LDLM_RESOURCE_DELREF(pres);
1713 ldlm_resource_putref(pres);
1715 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1718 EXPORT_SYMBOL(ldlm_request_cancel);
1721 * Main LDLM entry point for server code to cancel locks.
1723 * Typically gets called from service handler on LDLM_CANCEL opc.
1725 int ldlm_handle_cancel(struct ptlrpc_request *req)
1727 struct ldlm_request *dlm_req;
1731 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1732 if (dlm_req == NULL) {
1733 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1737 if (req->rq_export && req->rq_export->exp_nid_stats &&
1738 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1739 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1740 LDLM_CANCEL - LDLM_FIRST_OPC);
1742 rc = req_capsule_server_pack(&req->rq_pill);
1746 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1747 req->rq_status = LUSTRE_ESTALE;
1749 RETURN(ptlrpc_reply(req));
1751 #endif /* HAVE_SERVER_SUPPORT */
1754 * Callback handler for receiving incoming blocking ASTs.
1756 * This can only happen on client side.
1758 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1759 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1764 LDLM_DEBUG(lock, "client blocking AST callback handler");
1766 lock_res_and_lock(lock);
1767 ldlm_set_cbpending(lock);
1769 if (ldlm_is_cancel_on_block(lock))
1770 ldlm_set_cancel(lock);
1772 do_ast = (!lock->l_readers && !lock->l_writers);
1773 unlock_res_and_lock(lock);
1776 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1777 lock, lock->l_blocking_ast);
1778 if (lock->l_blocking_ast != NULL)
1779 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1782 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1786 LDLM_DEBUG(lock, "client blocking callback handler END");
1787 LDLM_LOCK_RELEASE(lock);
1792 * Callback handler for receiving incoming completion ASTs.
1794 * This only can happen on client side.
1796 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1797 struct ldlm_namespace *ns,
1798 struct ldlm_request *dlm_req,
1799 struct ldlm_lock *lock)
1801 struct list_head ast_list;
1806 LDLM_DEBUG(lock, "client completion callback handler START");
1808 INIT_LIST_HEAD(&ast_list);
1809 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1810 int to = cfs_time_seconds(1);
1812 set_current_state(TASK_INTERRUPTIBLE);
1813 schedule_timeout(to);
1814 if (lock->l_granted_mode == lock->l_req_mode ||
1815 ldlm_is_destroyed(lock))
1820 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1822 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1823 GOTO(out, rc = lvb_len);
1824 } else if (lvb_len > 0) {
1825 if (lock->l_lvb_len > 0) {
1826 /* for extent lock, lvb contains ost_lvb{}. */
1827 LASSERT(lock->l_lvb_data != NULL);
1829 if (unlikely(lock->l_lvb_len < lvb_len)) {
1830 LDLM_ERROR(lock, "Replied LVB is larger than "
1831 "expectation, expected = %d, "
1833 lock->l_lvb_len, lvb_len);
1834 GOTO(out, rc = -EINVAL);
1839 lock_res_and_lock(lock);
1840 if (ldlm_is_destroyed(lock) ||
1841 lock->l_granted_mode == lock->l_req_mode) {
1842 /* bug 11300: the lock has already been granted */
1843 unlock_res_and_lock(lock);
1844 LDLM_DEBUG(lock, "Double grant race happened");
1848 /* If we receive the completion AST before the actual enqueue returned,
1849 * then we might need to switch lock modes, resources, or extents. */
1850 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1851 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1852 LDLM_DEBUG(lock, "completion AST, new lock mode");
1855 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1856 ldlm_convert_policy_to_local(req->rq_export,
1857 dlm_req->lock_desc.l_resource.lr_type,
1858 &dlm_req->lock_desc.l_policy_data,
1859 &lock->l_policy_data);
1860 LDLM_DEBUG(lock, "completion AST, new policy data");
1863 ldlm_resource_unlink_lock(lock);
1864 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1865 &lock->l_resource->lr_name,
1866 sizeof(lock->l_resource->lr_name)) != 0) {
1867 unlock_res_and_lock(lock);
1868 rc = ldlm_lock_change_resource(ns, lock,
1869 &dlm_req->lock_desc.l_resource.lr_name);
1871 LDLM_ERROR(lock, "Failed to allocate resource");
1874 LDLM_DEBUG(lock, "completion AST, new resource");
1875 CERROR("change resource!\n");
1876 lock_res_and_lock(lock);
1879 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1880 /* BL_AST locks are not needed in LRU.
1881 * Let ldlm_cancel_lru() be fast. */
1882 ldlm_lock_remove_from_lru(lock);
1883 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1884 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1887 if (lock->l_lvb_len > 0) {
1888 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1889 lock->l_lvb_data, lvb_len);
1891 unlock_res_and_lock(lock);
1896 ldlm_grant_lock(lock, &ast_list);
1897 unlock_res_and_lock(lock);
1899 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1901 /* Let Enqueue to call osc_lock_upcall() and initialize
1903 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1905 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1907 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1913 lock_res_and_lock(lock);
1914 ldlm_set_failed(lock);
1915 unlock_res_and_lock(lock);
1916 wake_up(&lock->l_waitq);
1918 LDLM_LOCK_RELEASE(lock);
1922 * Callback handler for receiving incoming glimpse ASTs.
1924 * This only can happen on client side. After handling the glimpse AST
1925 * we also consider dropping the lock here if it is unused locally for a
1928 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1929 struct ldlm_namespace *ns,
1930 struct ldlm_request *dlm_req,
1931 struct ldlm_lock *lock)
1936 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1938 if (lock->l_glimpse_ast != NULL)
1939 rc = lock->l_glimpse_ast(lock, req);
1941 if (req->rq_repmsg != NULL) {
1944 req->rq_status = rc;
1948 lock_res_and_lock(lock);
1949 if (lock->l_granted_mode == LCK_PW &&
1950 !lock->l_readers && !lock->l_writers &&
1951 cfs_time_after(cfs_time_current(),
1952 cfs_time_add(lock->l_last_used,
1953 cfs_time_seconds(10)))) {
1954 unlock_res_and_lock(lock);
1955 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1956 ldlm_handle_bl_callback(ns, NULL, lock);
1961 unlock_res_and_lock(lock);
1962 LDLM_LOCK_RELEASE(lock);
1966 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1968 if (req->rq_no_reply)
1971 req->rq_status = rc;
1972 if (!req->rq_packed_final) {
1973 rc = lustre_pack_reply(req, 1, NULL, NULL);
1977 return ptlrpc_reply(req);
1980 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1981 enum ldlm_cancel_flags cancel_flags)
1983 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1986 spin_lock(&blp->blp_lock);
1987 if (blwi->blwi_lock &&
1988 ldlm_is_discard_data(blwi->blwi_lock)) {
1989 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1990 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1992 /* other blocking callbacks are added to the regular list */
1993 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1995 spin_unlock(&blp->blp_lock);
1997 wake_up(&blp->blp_waitq);
1999 /* can not check blwi->blwi_flags as blwi could be already freed in
2001 if (!(cancel_flags & LCF_ASYNC))
2002 wait_for_completion(&blwi->blwi_comp);
2007 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2008 struct ldlm_namespace *ns,
2009 struct ldlm_lock_desc *ld,
2010 struct list_head *cancels, int count,
2011 struct ldlm_lock *lock,
2012 enum ldlm_cancel_flags cancel_flags)
2014 init_completion(&blwi->blwi_comp);
2015 INIT_LIST_HEAD(&blwi->blwi_head);
2017 if (memory_pressure_get())
2018 blwi->blwi_mem_pressure = 1;
2021 blwi->blwi_flags = cancel_flags;
2023 blwi->blwi_ld = *ld;
2025 list_add(&blwi->blwi_head, cancels);
2026 list_del_init(cancels);
2027 blwi->blwi_count = count;
2029 blwi->blwi_lock = lock;
2034 * Queues a list of locks \a cancels containing \a count locks
2035 * for later processing by a blocking thread. If \a count is zero,
2036 * then the lock referenced as \a lock is queued instead.
2038 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2039 * If list addition fails an error is returned and caller is supposed to
2040 * call ->l_blocking_ast itself.
2042 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2043 struct ldlm_lock_desc *ld,
2044 struct ldlm_lock *lock,
2045 struct list_head *cancels, int count,
2046 enum ldlm_cancel_flags cancel_flags)
2050 if (cancels && count == 0)
2053 if (cancel_flags & LCF_ASYNC) {
2054 struct ldlm_bl_work_item *blwi;
2056 OBD_ALLOC(blwi, sizeof(*blwi));
2059 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2061 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2063 /* if it is synchronous call do minimum mem alloc, as it could
2064 * be triggered from kernel shrinker
2066 struct ldlm_bl_work_item blwi;
2068 memset(&blwi, 0, sizeof(blwi));
2069 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2070 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2075 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2076 struct ldlm_lock *lock)
2078 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2081 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2082 struct list_head *cancels, int count,
2083 enum ldlm_cancel_flags cancel_flags)
2085 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2088 int ldlm_bl_thread_wakeup(void)
2090 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2094 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2095 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2097 struct obd_device *obd = req->rq_export->exp_obd;
2104 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2106 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2108 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2110 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2113 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2115 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2117 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2120 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2123 /* We are responsible for swabbing contents of val */
2125 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2126 /* Pass it on to mdc (the "export" in this case) */
2127 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2129 sizeof(KEY_HSM_COPYTOOL_SEND),
2130 KEY_HSM_COPYTOOL_SEND,
2133 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2138 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2139 const char *msg, int rc,
2140 struct lustre_handle *handle)
2142 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2143 "%s: [nid %s] [rc %d] [lock "LPX64"]",
2144 msg, libcfs_id2str(req->rq_peer), rc,
2145 handle ? handle->cookie : 0);
2146 if (req->rq_no_reply)
2147 CWARN("No reply was sent, maybe cause bug 21636.\n");
2149 CWARN("Send reply failed, maybe cause bug 21636.\n");
2152 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2153 static int ldlm_callback_handler(struct ptlrpc_request *req)
2155 struct ldlm_namespace *ns;
2156 struct ldlm_request *dlm_req;
2157 struct ldlm_lock *lock;
2161 /* Requests arrive in sender's byte order. The ptlrpc service
2162 * handler has already checked and, if necessary, byte-swapped the
2163 * incoming request message body, but I am responsible for the
2164 * message buffers. */
2166 /* do nothing for sec context finalize */
2167 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2170 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2172 if (req->rq_export == NULL) {
2173 rc = ldlm_callback_reply(req, -ENOTCONN);
2174 ldlm_callback_errmsg(req, "Operate on unconnected server",
2179 LASSERT(req->rq_export != NULL);
2180 LASSERT(req->rq_export->exp_obd != NULL);
2182 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2183 case LDLM_BL_CALLBACK:
2184 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2186 ldlm_callback_reply(req, -(int)cfs_fail_err);
2190 case LDLM_CP_CALLBACK:
2191 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2194 case LDLM_GL_CALLBACK:
2195 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2199 rc = ldlm_handle_setinfo(req);
2200 ldlm_callback_reply(req, rc);
2202 case LLOG_ORIGIN_HANDLE_CREATE:
2203 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2204 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2206 rc = llog_origin_handle_open(req);
2207 ldlm_callback_reply(req, rc);
2209 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2210 req_capsule_set(&req->rq_pill,
2211 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2212 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2214 rc = llog_origin_handle_next_block(req);
2215 ldlm_callback_reply(req, rc);
2217 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2218 req_capsule_set(&req->rq_pill,
2219 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2220 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2222 rc = llog_origin_handle_read_header(req);
2223 ldlm_callback_reply(req, rc);
2225 case LLOG_ORIGIN_HANDLE_CLOSE:
2226 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2228 rc = llog_origin_handle_close(req);
2229 ldlm_callback_reply(req, rc);
2232 CERROR("unknown opcode %u\n",
2233 lustre_msg_get_opc(req->rq_reqmsg));
2234 ldlm_callback_reply(req, -EPROTO);
2238 ns = req->rq_export->exp_obd->obd_namespace;
2239 LASSERT(ns != NULL);
2241 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2243 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2244 if (dlm_req == NULL) {
2245 rc = ldlm_callback_reply(req, -EPROTO);
2246 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2251 /* Force a known safe race, send a cancel to the server for a lock
2252 * which the server has already started a blocking callback on. */
2253 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2254 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2255 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2257 CERROR("ldlm_cli_cancel: %d\n", rc);
2260 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2262 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2263 "disappeared\n", dlm_req->lock_handle[0].cookie);
2264 rc = ldlm_callback_reply(req, -EINVAL);
2265 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2266 &dlm_req->lock_handle[0]);
2270 if (ldlm_is_fail_loc(lock) &&
2271 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2272 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2274 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2275 lock_res_and_lock(lock);
2276 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2278 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2279 /* If somebody cancels lock and cache is already dropped,
2280 * or lock is failed before cp_ast received on client,
2281 * we can tell the server we have no lock. Otherwise, we
2282 * should send cancel after dropping the cache. */
2283 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2284 ldlm_is_failed(lock)) {
2285 LDLM_DEBUG(lock, "callback on lock "
2286 LPX64" - lock disappeared\n",
2287 dlm_req->lock_handle[0].cookie);
2288 unlock_res_and_lock(lock);
2289 LDLM_LOCK_RELEASE(lock);
2290 rc = ldlm_callback_reply(req, -EINVAL);
2291 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2292 &dlm_req->lock_handle[0]);
2295 /* BL_AST locks are not needed in LRU.
2296 * Let ldlm_cancel_lru() be fast. */
2297 ldlm_lock_remove_from_lru(lock);
2298 ldlm_set_bl_ast(lock);
2300 unlock_res_and_lock(lock);
2302 /* We want the ost thread to get this reply so that it can respond
2303 * to ost requests (write cache writeback) that might be triggered
2306 * But we'd also like to be able to indicate in the reply that we're
2307 * cancelling right now, because it's unused, or have an intent result
2308 * in the reply, so we might have to push the responsibility for sending
2309 * the reply down into the AST handlers, alas. */
2311 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2312 case LDLM_BL_CALLBACK:
2313 CDEBUG(D_INODE, "blocking ast\n");
2314 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2315 if (!ldlm_is_cancel_on_block(lock)) {
2316 rc = ldlm_callback_reply(req, 0);
2317 if (req->rq_no_reply || rc)
2318 ldlm_callback_errmsg(req, "Normal process", rc,
2319 &dlm_req->lock_handle[0]);
2321 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2322 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2324 case LDLM_CP_CALLBACK:
2325 CDEBUG(D_INODE, "completion ast\n");
2326 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2327 ldlm_callback_reply(req, 0);
2328 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2330 case LDLM_GL_CALLBACK:
2331 CDEBUG(D_INODE, "glimpse ast\n");
2332 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2333 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2336 LBUG(); /* checked above */
2342 #ifdef HAVE_SERVER_SUPPORT
2344 * Main handler for canceld thread.
2346 * Separated into its own thread to avoid deadlocks.
2348 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2353 /* Requests arrive in sender's byte order. The ptlrpc service
2354 * handler has already checked and, if necessary, byte-swapped the
2355 * incoming request message body, but I am responsible for the
2356 * message buffers. */
2358 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2360 if (req->rq_export == NULL) {
2361 struct ldlm_request *dlm_req;
2363 CERROR("%s from %s arrived at %lu with bad export cookie "
2365 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2366 libcfs_nid2str(req->rq_peer.nid),
2367 req->rq_arrival_time.tv_sec,
2368 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2370 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2371 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2372 dlm_req = req_capsule_client_get(&req->rq_pill,
2374 if (dlm_req != NULL)
2375 ldlm_lock_dump_handle(D_ERROR,
2376 &dlm_req->lock_handle[0]);
2378 ldlm_callback_reply(req, -ENOTCONN);
2382 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2384 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2386 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2387 CDEBUG(D_INODE, "cancel\n");
2388 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2389 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2390 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2392 rc = ldlm_handle_cancel(req);
2397 CERROR("invalid opcode %d\n",
2398 lustre_msg_get_opc(req->rq_reqmsg));
2399 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2400 ldlm_callback_reply(req, -EINVAL);
2406 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2407 struct ldlm_lock *lock)
2409 struct ldlm_request *dlm_req;
2410 struct lustre_handle lockh;
2415 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2416 if (dlm_req == NULL)
2419 ldlm_lock2handle(lock, &lockh);
2420 for (i = 0; i < dlm_req->lock_count; i++) {
2421 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2423 DEBUG_REQ(D_RPCTRACE, req,
2424 "Prio raised by lock "LPX64".", lockh.cookie);
2435 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2437 struct ldlm_request *dlm_req;
2442 /* no prolong in recovery */
2443 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2446 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2447 if (dlm_req == NULL)
2450 for (i = 0; i < dlm_req->lock_count; i++) {
2451 struct ldlm_lock *lock;
2453 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2457 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2459 LDLM_DEBUG(lock, "hpreq cancel lock");
2460 LDLM_LOCK_PUT(lock);
2469 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2470 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2471 .hpreq_check = ldlm_cancel_hpreq_check,
2475 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2479 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2481 if (req->rq_export == NULL)
2484 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2485 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2486 req->rq_ops = &ldlm_cancel_hpreq_ops;
2491 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2492 struct hlist_node *hnode, void *data)
2495 struct list_head *rpc_list = data;
2496 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2498 lock_res_and_lock(lock);
2500 if (lock->l_req_mode != lock->l_granted_mode) {
2501 unlock_res_and_lock(lock);
2505 LASSERT(lock->l_resource);
2506 if (lock->l_resource->lr_type != LDLM_IBITS &&
2507 lock->l_resource->lr_type != LDLM_PLAIN) {
2508 unlock_res_and_lock(lock);
2512 if (ldlm_is_ast_sent(lock)) {
2513 unlock_res_and_lock(lock);
2517 LASSERT(lock->l_blocking_ast);
2518 LASSERT(!lock->l_blocking_lock);
2520 ldlm_set_ast_sent(lock);
2521 if (lock->l_export && lock->l_export->exp_lock_hash) {
2522 /* NB: it's safe to call cfs_hash_del() even lock isn't
2523 * in exp_lock_hash. */
2524 /* In the function below, .hs_keycmp resolves to
2525 * ldlm_export_lock_keycmp() */
2526 /* coverity[overrun-buffer-val] */
2527 cfs_hash_del(lock->l_export->exp_lock_hash,
2528 &lock->l_remote_handle, &lock->l_exp_hash);
2531 list_add_tail(&lock->l_rk_ast, rpc_list);
2532 LDLM_LOCK_GET(lock);
2534 unlock_res_and_lock(lock);
2538 void ldlm_revoke_export_locks(struct obd_export *exp)
2540 struct list_head rpc_list;
2543 INIT_LIST_HEAD(&rpc_list);
2544 cfs_hash_for_each_empty(exp->exp_lock_hash,
2545 ldlm_revoke_lock_cb, &rpc_list);
2546 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2547 LDLM_WORK_REVOKE_AST);
2551 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2552 #endif /* HAVE_SERVER_SUPPORT */
2554 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2555 struct ldlm_bl_work_item **p_blwi,
2556 struct obd_export **p_exp)
2558 struct ldlm_bl_work_item *blwi = NULL;
2559 static unsigned int num_bl = 0;
2560 static unsigned int num_stale;
2561 int num_th = atomic_read(&blp->blp_num_threads);
2563 *p_exp = obd_stale_export_get();
2565 spin_lock(&blp->blp_lock);
2566 if (*p_exp != NULL) {
2567 if (num_th == 1 || ++num_stale < num_th) {
2568 spin_unlock(&blp->blp_lock);
2575 /* process a request from the blp_list at least every blp_num_threads */
2576 if (!list_empty(&blp->blp_list) &&
2577 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2578 blwi = list_entry(blp->blp_list.next,
2579 struct ldlm_bl_work_item, blwi_entry);
2581 if (!list_empty(&blp->blp_prio_list))
2582 blwi = list_entry(blp->blp_prio_list.next,
2583 struct ldlm_bl_work_item,
2587 if (++num_bl >= num_th)
2589 list_del(&blwi->blwi_entry);
2591 spin_unlock(&blp->blp_lock);
2594 if (*p_exp != NULL && *p_blwi != NULL) {
2595 obd_stale_export_put(*p_exp);
2599 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2602 /* This only contains temporary data until the thread starts */
2603 struct ldlm_bl_thread_data {
2604 struct ldlm_bl_pool *bltd_blp;
2605 struct completion bltd_comp;
2609 static int ldlm_bl_thread_main(void *arg);
2611 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2613 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2614 struct task_struct *task;
2616 init_completion(&bltd.bltd_comp);
2618 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2619 if (bltd.bltd_num >= blp->blp_max_threads) {
2620 atomic_dec(&blp->blp_num_threads);
2624 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2626 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2627 atomic_dec(&blp->blp_num_threads);
2631 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2634 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2635 bltd.bltd_num, PTR_ERR(task));
2636 atomic_dec(&blp->blp_num_threads);
2637 return PTR_ERR(task);
2639 wait_for_completion(&bltd.bltd_comp);
2644 /* Not fatal if racy and have a few too many threads */
2645 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2646 struct ldlm_bl_work_item *blwi)
2648 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2651 if (atomic_read(&blp->blp_busy_threads) <
2652 atomic_read(&blp->blp_num_threads))
2655 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2656 blwi->blwi_mem_pressure))
2662 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2663 struct ldlm_bl_work_item *blwi)
2667 if (blwi->blwi_ns == NULL)
2668 /* added by ldlm_cleanup() */
2669 RETURN(LDLM_ITER_STOP);
2671 if (blwi->blwi_mem_pressure)
2672 memory_pressure_set();
2674 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2676 if (blwi->blwi_count) {
2678 /* The special case when we cancel locks in lru
2679 * asynchronously, we pass the list of locks here.
2680 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2681 * canceled locally yet. */
2682 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2685 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2688 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2691 if (blwi->blwi_mem_pressure)
2692 memory_pressure_clr();
2694 if (blwi->blwi_flags & LCF_ASYNC)
2695 OBD_FREE(blwi, sizeof(*blwi));
2697 complete(&blwi->blwi_comp);
2703 * Cancel stale locks on export. Cancel blocked locks first.
2704 * If the given export has blocked locks, the next in the list may have
2705 * them too, thus cancel not blocked locks only if the current export has
2708 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2709 struct obd_export *exp)
2714 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2716 num = ldlm_export_cancel_blocked_locks(exp);
2718 ldlm_export_cancel_locks(exp);
2720 obd_stale_export_put(exp);
2727 * Main blocking requests processing thread.
2729 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2730 * This thread in the end ends up doing actual call to ->l_blocking_ast
2733 static int ldlm_bl_thread_main(void *arg)
2735 struct ldlm_bl_pool *blp;
2736 struct ldlm_bl_thread_data *bltd = arg;
2739 blp = bltd->bltd_blp;
2741 complete(&bltd->bltd_comp);
2742 /* cannot use bltd after this, it is only on caller's stack */
2745 struct l_wait_info lwi = { 0 };
2746 struct ldlm_bl_work_item *blwi = NULL;
2747 struct obd_export *exp = NULL;
2750 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2753 l_wait_event_exclusive(blp->blp_waitq,
2754 ldlm_bl_get_work(blp, &blwi,
2757 atomic_inc(&blp->blp_busy_threads);
2759 if (ldlm_bl_thread_need_create(blp, blwi))
2760 /* discard the return value, we tried */
2761 ldlm_bl_thread_start(blp, true);
2764 rc = ldlm_bl_thread_exports(blp, exp);
2766 rc = ldlm_bl_thread_blwi(blp, blwi);
2768 atomic_dec(&blp->blp_busy_threads);
2770 if (rc == LDLM_ITER_STOP)
2774 atomic_dec(&blp->blp_num_threads);
2775 complete(&blp->blp_comp);
2780 static int ldlm_setup(void);
2781 static int ldlm_cleanup(void);
2783 int ldlm_get_ref(void)
2787 mutex_lock(&ldlm_ref_mutex);
2788 if (++ldlm_refcount == 1) {
2793 mutex_unlock(&ldlm_ref_mutex);
2798 void ldlm_put_ref(void)
2801 mutex_lock(&ldlm_ref_mutex);
2802 if (ldlm_refcount == 1) {
2803 int rc = ldlm_cleanup();
2805 CERROR("ldlm_cleanup failed: %d\n", rc);
2811 mutex_unlock(&ldlm_ref_mutex);
2817 * Export handle<->lock hash operations.
2820 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2822 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2826 ldlm_export_lock_key(struct hlist_node *hnode)
2828 struct ldlm_lock *lock;
2830 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2831 return &lock->l_remote_handle;
2835 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2837 struct ldlm_lock *lock;
2839 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2840 lock->l_remote_handle = *(struct lustre_handle *)key;
2844 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2846 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2850 ldlm_export_lock_object(struct hlist_node *hnode)
2852 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2856 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2858 struct ldlm_lock *lock;
2860 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2861 LDLM_LOCK_GET(lock);
2865 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2867 struct ldlm_lock *lock;
2869 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2870 LDLM_LOCK_RELEASE(lock);
2873 static struct cfs_hash_ops ldlm_export_lock_ops = {
2874 .hs_hash = ldlm_export_lock_hash,
2875 .hs_key = ldlm_export_lock_key,
2876 .hs_keycmp = ldlm_export_lock_keycmp,
2877 .hs_keycpy = ldlm_export_lock_keycpy,
2878 .hs_object = ldlm_export_lock_object,
2879 .hs_get = ldlm_export_lock_get,
2880 .hs_put = ldlm_export_lock_put,
2881 .hs_put_locked = ldlm_export_lock_put,
2884 int ldlm_init_export(struct obd_export *exp)
2889 exp->exp_lock_hash =
2890 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2891 HASH_EXP_LOCK_CUR_BITS,
2892 HASH_EXP_LOCK_MAX_BITS,
2893 HASH_EXP_LOCK_BKT_BITS, 0,
2894 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2895 &ldlm_export_lock_ops,
2896 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2897 CFS_HASH_NBLK_CHANGE);
2899 if (!exp->exp_lock_hash)
2902 rc = ldlm_init_flock_export(exp);
2908 ldlm_destroy_export(exp);
2911 EXPORT_SYMBOL(ldlm_init_export);
2913 void ldlm_destroy_export(struct obd_export *exp)
2916 cfs_hash_putref(exp->exp_lock_hash);
2917 exp->exp_lock_hash = NULL;
2919 ldlm_destroy_flock_export(exp);
2922 EXPORT_SYMBOL(ldlm_destroy_export);
2924 static int ldlm_setup(void)
2926 static struct ptlrpc_service_conf conf;
2927 struct ldlm_bl_pool *blp = NULL;
2928 #ifdef HAVE_SERVER_SUPPORT
2929 struct task_struct *task;
2930 #endif /* HAVE_SERVER_SUPPORT */
2936 if (ldlm_state != NULL)
2939 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2940 if (ldlm_state == NULL)
2943 #ifdef CONFIG_PROC_FS
2944 rc = ldlm_proc_setup();
2947 #endif /* CONFIG_PROC_FS */
2949 memset(&conf, 0, sizeof(conf));
2950 conf = (typeof(conf)) {
2951 .psc_name = "ldlm_cbd",
2952 .psc_watchdog_factor = 2,
2954 .bc_nbufs = LDLM_CLIENT_NBUFS,
2955 .bc_buf_size = LDLM_BUFSIZE,
2956 .bc_req_max_size = LDLM_MAXREQSIZE,
2957 .bc_rep_max_size = LDLM_MAXREPSIZE,
2958 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2959 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2962 .tc_thr_name = "ldlm_cb",
2963 .tc_thr_factor = LDLM_THR_FACTOR,
2964 .tc_nthrs_init = LDLM_NTHRS_INIT,
2965 .tc_nthrs_base = LDLM_NTHRS_BASE,
2966 .tc_nthrs_max = LDLM_NTHRS_MAX,
2967 .tc_nthrs_user = ldlm_num_threads,
2968 .tc_cpu_affinity = 1,
2969 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2972 .cc_pattern = ldlm_cpts,
2975 .so_req_handler = ldlm_callback_handler,
2978 ldlm_state->ldlm_cb_service = \
2979 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2980 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2981 CERROR("failed to start service\n");
2982 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2983 ldlm_state->ldlm_cb_service = NULL;
2987 #ifdef HAVE_SERVER_SUPPORT
2988 memset(&conf, 0, sizeof(conf));
2989 conf = (typeof(conf)) {
2990 .psc_name = "ldlm_canceld",
2991 .psc_watchdog_factor = 6,
2993 .bc_nbufs = LDLM_SERVER_NBUFS,
2994 .bc_buf_size = LDLM_BUFSIZE,
2995 .bc_req_max_size = LDLM_MAXREQSIZE,
2996 .bc_rep_max_size = LDLM_MAXREPSIZE,
2997 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
2998 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3002 .tc_thr_name = "ldlm_cn",
3003 .tc_thr_factor = LDLM_THR_FACTOR,
3004 .tc_nthrs_init = LDLM_NTHRS_INIT,
3005 .tc_nthrs_base = LDLM_NTHRS_BASE,
3006 .tc_nthrs_max = LDLM_NTHRS_MAX,
3007 .tc_nthrs_user = ldlm_num_threads,
3008 .tc_cpu_affinity = 1,
3009 .tc_ctx_tags = LCT_MD_THREAD | \
3014 .cc_pattern = ldlm_cpts,
3017 .so_req_handler = ldlm_cancel_handler,
3018 .so_hpreq_handler = ldlm_hpreq_handler,
3021 ldlm_state->ldlm_cancel_service = \
3022 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
3023 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3024 CERROR("failed to start service\n");
3025 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3026 ldlm_state->ldlm_cancel_service = NULL;
3029 #endif /* HAVE_SERVER_SUPPORT */
3031 OBD_ALLOC(blp, sizeof(*blp));
3033 GOTO(out, rc = -ENOMEM);
3034 ldlm_state->ldlm_bl_pool = blp;
3036 spin_lock_init(&blp->blp_lock);
3037 INIT_LIST_HEAD(&blp->blp_list);
3038 INIT_LIST_HEAD(&blp->blp_prio_list);
3039 init_waitqueue_head(&blp->blp_waitq);
3040 atomic_set(&blp->blp_num_threads, 0);
3041 atomic_set(&blp->blp_busy_threads, 0);
3043 if (ldlm_num_threads == 0) {
3044 blp->blp_min_threads = LDLM_NTHRS_INIT;
3045 blp->blp_max_threads = LDLM_NTHRS_MAX;
3047 blp->blp_min_threads = blp->blp_max_threads = \
3048 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3052 for (i = 0; i < blp->blp_min_threads; i++) {
3053 rc = ldlm_bl_thread_start(blp, false);
3058 #ifdef HAVE_SERVER_SUPPORT
3059 INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
3060 expired_lock_thread.elt_state = ELT_STOPPED;
3061 init_waitqueue_head(&expired_lock_thread.elt_waitq);
3063 INIT_LIST_HEAD(&waiting_locks_list);
3064 spin_lock_init(&waiting_locks_spinlock);
3065 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, NULL);
3067 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3070 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3074 wait_event(expired_lock_thread.elt_waitq,
3075 expired_lock_thread.elt_state == ELT_READY);
3076 #endif /* HAVE_SERVER_SUPPORT */
3078 rc = ldlm_pools_init();
3080 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3084 rc = ldlm_reclaim_setup();
3086 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3096 static int ldlm_cleanup(void)
3100 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3101 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3102 CERROR("ldlm still has namespaces; clean these up first.\n");
3103 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3104 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3108 ldlm_reclaim_cleanup();
3111 if (ldlm_state->ldlm_bl_pool != NULL) {
3112 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3114 while (atomic_read(&blp->blp_num_threads) > 0) {
3115 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3117 init_completion(&blp->blp_comp);
3119 spin_lock(&blp->blp_lock);
3120 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3121 wake_up(&blp->blp_waitq);
3122 spin_unlock(&blp->blp_lock);
3124 wait_for_completion(&blp->blp_comp);
3127 OBD_FREE(blp, sizeof(*blp));
3130 if (ldlm_state->ldlm_cb_service != NULL)
3131 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3132 #ifdef HAVE_SERVER_SUPPORT
3133 if (ldlm_state->ldlm_cancel_service != NULL)
3134 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3137 ldlm_proc_cleanup();
3139 #ifdef HAVE_SERVER_SUPPORT
3140 if (expired_lock_thread.elt_state != ELT_STOPPED) {
3141 expired_lock_thread.elt_state = ELT_TERMINATE;
3142 wake_up(&expired_lock_thread.elt_waitq);
3143 wait_event(expired_lock_thread.elt_waitq,
3144 expired_lock_thread.elt_state == ELT_STOPPED);
3148 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3156 mutex_init(&ldlm_ref_mutex);
3157 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
3158 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
3160 INIT_LIST_HEAD(&ldlm_srv_namespace_list);
3161 INIT_LIST_HEAD(&ldlm_cli_active_namespace_list);
3162 INIT_LIST_HEAD(&ldlm_cli_inactive_namespace_list);
3164 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3165 sizeof(struct ldlm_resource), 0,
3166 SLAB_HWCACHE_ALIGN, NULL);
3167 if (ldlm_resource_slab == NULL)
3170 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3171 sizeof(struct ldlm_lock), 0,
3172 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3173 if (ldlm_lock_slab == NULL)
3176 ldlm_interval_slab = kmem_cache_create("interval_node",
3177 sizeof(struct ldlm_interval),
3178 0, SLAB_HWCACHE_ALIGN, NULL);
3179 if (ldlm_interval_slab == NULL)
3182 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3183 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3184 0, SLAB_HWCACHE_ALIGN, NULL);
3185 if (ldlm_interval_tree_slab == NULL)
3188 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3189 class_export_dump_hook = ldlm_dump_export_locks;
3194 kmem_cache_destroy(ldlm_interval_slab);
3196 kmem_cache_destroy(ldlm_lock_slab);
3198 kmem_cache_destroy(ldlm_resource_slab);
3203 void ldlm_exit(void)
3206 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3207 kmem_cache_destroy(ldlm_resource_slab);
3208 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3209 * synchronize_rcu() to wait a grace period elapsed, so that
3210 * ldlm_lock_free() get a chance to be called. */
3212 kmem_cache_destroy(ldlm_lock_slab);
3213 kmem_cache_destroy(ldlm_interval_slab);
3214 kmem_cache_destroy(ldlm_interval_tree_slab);