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, 2014, 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 spin_lock_bh(&waiting_locks_spinlock);
448 if (ldlm_is_cancel(lock)) {
449 spin_unlock_bh(&waiting_locks_spinlock);
453 if (ldlm_is_destroyed(lock)) {
454 static cfs_time_t next;
456 spin_unlock_bh(&waiting_locks_spinlock);
457 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
458 if (cfs_time_after(cfs_time_current(), next)) {
459 next = cfs_time_shift(14400);
460 libcfs_debug_dumpstack(NULL);
465 ldlm_set_waited(lock);
466 lock->l_last_activity = cfs_time_current_sec();
467 ret = __ldlm_add_waiting_lock(lock, timeout);
469 /* grab ref on the lock if it has been added to the
473 spin_unlock_bh(&waiting_locks_spinlock);
476 ldlm_add_blocked_lock(lock);
478 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
479 ret == 0 ? "not re-" : "", timeout,
480 AT_OFF ? "off" : "on");
485 * Remove a lock from the pending list, likely because it had its cancellation
486 * callback arrive without incident. This adjusts the lock-timeout timer if
487 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
488 * As done by ldlm_del_waiting_lock(), the caller must release the lock
489 * reference when the lock is removed from any list (1 is returned).
491 * Called with namespace lock held.
493 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
495 struct list_head *list_next;
497 if (list_empty(&lock->l_pending_chain))
500 list_next = lock->l_pending_chain.next;
501 if (lock->l_pending_chain.prev == &waiting_locks_list) {
502 /* Removing the head of the list, adjust timer. */
503 if (list_next == &waiting_locks_list) {
504 /* No more, just cancel. */
505 cfs_timer_disarm(&waiting_locks_timer);
507 struct ldlm_lock *next;
508 next = list_entry(list_next, struct ldlm_lock,
510 cfs_timer_arm(&waiting_locks_timer,
511 round_timeout(next->l_callback_timeout));
514 list_del_init(&lock->l_pending_chain);
519 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
523 if (lock->l_export == NULL) {
524 /* We don't have a "waiting locks list" on clients. */
525 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
529 spin_lock_bh(&waiting_locks_spinlock);
530 ret = __ldlm_del_waiting_lock(lock);
531 ldlm_clear_waited(lock);
532 spin_unlock_bh(&waiting_locks_spinlock);
534 /* remove the lock out of export blocking list */
535 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
536 list_del_init(&lock->l_exp_list);
537 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
540 /* release lock ref if it has indeed been removed
542 LDLM_LOCK_RELEASE(lock);
545 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
550 * Prolong the contended lock waiting time.
552 * Called with namespace lock held.
554 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
556 if (lock->l_export == NULL) {
557 /* We don't have a "waiting locks list" on clients. */
558 LDLM_DEBUG(lock, "client lock: no-op");
562 spin_lock_bh(&waiting_locks_spinlock);
564 if (list_empty(&lock->l_pending_chain)) {
565 spin_unlock_bh(&waiting_locks_spinlock);
566 LDLM_DEBUG(lock, "wasn't waiting");
570 /* we remove/add the lock to the waiting list, so no needs to
571 * release/take a lock reference */
572 __ldlm_del_waiting_lock(lock);
573 __ldlm_add_waiting_lock(lock, timeout);
574 spin_unlock_bh(&waiting_locks_spinlock);
576 LDLM_DEBUG(lock, "refreshed");
579 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
581 #else /* HAVE_SERVER_SUPPORT */
583 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
588 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
593 #endif /* !HAVE_SERVER_SUPPORT */
595 #ifdef HAVE_SERVER_SUPPORT
598 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
599 * lock cancel, and their replies). Used for lock callback timeout and AST
602 * \param[in] lock lock which is getting the blocking callback
604 * \retval timeout in seconds to wait for the client reply
606 unsigned int ldlm_bl_timeout(struct ldlm_lock *lock)
608 unsigned int timeout;
611 return obd_timeout / 2;
613 /* Since these are non-updating timeouts, we should be conservative.
614 * Take more than usually, 150%
615 * It would be nice to have some kind of "early reply" mechanism for
616 * lock callbacks too... */
617 timeout = at_get(&lock->l_export->exp_bl_lock_at);
618 return max(timeout + (timeout >> 1), ldlm_enqueue_min);
620 EXPORT_SYMBOL(ldlm_bl_timeout);
623 * Perform lock cleanup if AST sending failed.
625 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
626 const char *ast_type)
628 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
629 "to a lock %s callback time out: rc %d\n",
630 lock->l_export->exp_obd->obd_name,
631 obd_export_nid2str(lock->l_export), ast_type, rc);
633 if (obd_dump_on_timeout)
634 libcfs_debug_dumplog();
635 spin_lock_bh(&waiting_locks_spinlock);
636 if (__ldlm_del_waiting_lock(lock) == 0)
637 /* the lock was not in any list, grab an extra ref before adding
638 * the lock to the expired list */
640 list_add(&lock->l_pending_chain,
641 &expired_lock_thread.elt_expired_locks);
642 wake_up(&expired_lock_thread.elt_waitq);
643 spin_unlock_bh(&waiting_locks_spinlock);
647 * Perform lock cleanup if AST reply came with error.
649 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
650 struct ptlrpc_request *req, int rc,
651 const char *ast_type)
653 lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
655 if (!req->rq_replied || (rc && rc != -EINVAL)) {
656 if (lock->l_export && lock->l_export->exp_libclient) {
657 LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
658 " timeout, just cancelling lock", ast_type,
659 libcfs_nid2str(peer.nid));
660 ldlm_lock_cancel(lock);
662 } else if (ldlm_is_cancel(lock)) {
663 LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
664 "cancel was received (AST reply lost?)",
665 ast_type, libcfs_nid2str(peer.nid));
666 ldlm_lock_cancel(lock);
669 LDLM_ERROR(lock, "client (nid %s) %s %s AST "
670 "(req status %d rc %d), evict it",
671 libcfs_nid2str(peer.nid),
672 req->rq_replied ? "returned error from" :
673 "failed to reply to",
675 (req->rq_repmsg != NULL) ?
676 lustre_msg_get_status(req->rq_repmsg) : 0,
678 ldlm_failed_ast(lock, rc, ast_type);
684 struct ldlm_resource *res = lock->l_resource;
686 LDLM_DEBUG(lock, "client (nid %s) returned %d"
687 " from %s AST - normal race",
688 libcfs_nid2str(peer.nid),
690 lustre_msg_get_status(req->rq_repmsg) : -1,
693 /* update lvbo to return proper attributes.
695 ldlm_resource_getref(res);
696 ldlm_res_lvbo_update(res, NULL, 1);
697 ldlm_resource_putref(res);
699 ldlm_lock_cancel(lock);
706 static int ldlm_cb_interpret(const struct lu_env *env,
707 struct ptlrpc_request *req, void *data, int rc)
709 struct ldlm_cb_async_args *ca = data;
710 struct ldlm_lock *lock = ca->ca_lock;
711 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
714 LASSERT(lock != NULL);
717 case LDLM_GL_CALLBACK:
718 /* Update the LVB from disk if the AST failed
719 * (this is a legal race)
721 * - Glimpse callback of local lock just returns
722 * -ELDLM_NO_LOCK_DATA.
723 * - Glimpse callback of remote lock might return
724 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
726 if (rc == -ELDLM_NO_LOCK_DATA) {
727 LDLM_DEBUG(lock, "lost race - client has a lock but no "
729 ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
730 } else if (rc != 0) {
731 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
733 rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
736 case LDLM_BL_CALLBACK:
738 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
740 case LDLM_CP_CALLBACK:
742 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
745 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
750 /* release extra reference taken in ldlm_ast_fini() */
751 LDLM_LOCK_RELEASE(lock);
754 atomic_inc(&arg->restart);
759 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
761 struct ldlm_cb_async_args *ca = data;
762 struct ldlm_lock *lock = ca->ca_lock;
764 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
767 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
768 struct ldlm_cb_set_arg *arg,
769 struct ldlm_lock *lock,
775 if (unlikely(instant_cancel)) {
776 rc = ptl_send_rpc(req, 1);
777 ptlrpc_req_finished(req);
779 atomic_inc(&arg->restart);
782 ptlrpc_set_add_req(arg->set, req);
789 * Check if there are requests in the export request list which prevent
790 * the lock canceling and make these requests high priority ones.
792 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
794 struct ptlrpc_request *req;
797 if (lock->l_export == NULL) {
798 LDLM_DEBUG(lock, "client lock: no-op");
802 spin_lock_bh(&lock->l_export->exp_rpc_lock);
803 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
805 /* Do not process requests that were not yet added to there
806 * incoming queue or were already removed from there for
807 * processing. We evaluate ptlrpc_nrs_req_can_move() without
808 * holding svcpt->scp_req_lock, and then redo the check with
809 * the lock held once we need to obtain a reliable result.
811 if (ptlrpc_nrs_req_can_move(req) &&
812 req->rq_ops->hpreq_lock_match &&
813 req->rq_ops->hpreq_lock_match(req, lock))
814 ptlrpc_nrs_req_hp_move(req);
816 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
821 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
822 * enqueued server lock conflicts with given one.
824 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
825 * to wait for client response.
827 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
828 struct ldlm_lock_desc *desc,
829 void *data, int flag)
831 struct ldlm_cb_async_args *ca;
832 struct ldlm_cb_set_arg *arg = data;
833 struct ldlm_request *body;
834 struct ptlrpc_request *req;
835 int instant_cancel = 0;
839 if (flag == LDLM_CB_CANCELING)
840 /* Don't need to do anything here. */
844 LASSERT(data != NULL);
845 if (lock->l_export->exp_obd->obd_recovering != 0)
846 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
848 ldlm_lock_reorder_req(lock);
850 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
851 &RQF_LDLM_BL_CALLBACK,
852 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
856 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
857 ca = ptlrpc_req_async_args(req);
858 ca->ca_set_arg = arg;
861 req->rq_interpret_reply = ldlm_cb_interpret;
863 lock_res_and_lock(lock);
864 if (ldlm_is_destroyed(lock)) {
865 /* What's the point? */
866 unlock_res_and_lock(lock);
867 ptlrpc_req_finished(req);
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 ldlm_add_blocked_lock(lock);
875 ldlm_set_waited(lock);
876 unlock_res_and_lock(lock);
878 ptlrpc_req_finished(req);
879 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
883 if (ldlm_is_cancel_on_block(lock))
886 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
887 body->lock_handle[0] = lock->l_remote_handle;
888 body->lock_desc = *desc;
889 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
891 LDLM_DEBUG(lock, "server preparing blocking AST");
893 ptlrpc_request_set_replen(req);
894 if (instant_cancel) {
895 unlock_res_and_lock(lock);
896 ldlm_lock_cancel(lock);
898 req->rq_no_resend = 1;
900 LASSERT(lock->l_granted_mode == lock->l_req_mode);
901 ldlm_add_waiting_lock(lock);
902 unlock_res_and_lock(lock);
904 /* Do not resend after lock callback timeout */
905 req->rq_delay_limit = ldlm_bl_timeout(lock);
906 req->rq_resend_cb = ldlm_update_resend;
909 req->rq_send_state = LUSTRE_IMP_FULL;
910 /* ptlrpc_request_alloc_pack already set timeout */
912 req->rq_timeout = ldlm_get_rq_timeout();
914 lock->l_last_activity = cfs_time_current_sec();
916 if (lock->l_export && lock->l_export->exp_nid_stats &&
917 lock->l_export->exp_nid_stats->nid_ldlm_stats)
918 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
919 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
921 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
927 * ->l_completion_ast callback for a remote lock in server namespace.
929 * Sends AST to the client notifying it of lock granting. If initial
930 * lock response was not sent yet, instead of sending another RPC, just
931 * mark the lock as granted and client will understand
933 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
935 struct ldlm_cb_set_arg *arg = data;
936 struct ldlm_request *body;
937 struct ptlrpc_request *req;
938 struct ldlm_cb_async_args *ca;
939 int instant_cancel = 0;
944 LASSERT(lock != NULL);
945 LASSERT(data != NULL);
947 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
948 LDLM_DEBUG(lock, "dropping CP AST");
952 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
953 &RQF_LDLM_CP_CALLBACK);
957 /* server namespace, doesn't need lock */
958 lvb_len = ldlm_lvbo_size(lock);
959 /* LU-3124 & LU-2187: to not return layout in completion AST because
960 * it may deadlock for LU-2187, or client may not have enough space
961 * for large layout. The layout will be returned to client with an
962 * extra RPC to fetch xattr.lov */
963 if (ldlm_has_layout(lock))
966 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
967 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
969 ptlrpc_request_free(req);
973 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
974 ca = ptlrpc_req_async_args(req);
975 ca->ca_set_arg = arg;
978 req->rq_interpret_reply = ldlm_cb_interpret;
979 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
981 body->lock_handle[0] = lock->l_remote_handle;
982 body->lock_flags = ldlm_flags_to_wire(flags);
983 ldlm_lock2desc(lock, &body->lock_desc);
985 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
987 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
989 /* We still need to send the RPC to wake up the blocked
990 * enqueue thread on the client.
992 * Consider old client, there is no better way to notify
993 * the failure, just zero-sized the LVB, then the client
994 * will fail out as "-EPROTO". */
995 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
999 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1004 lock->l_last_activity = cfs_time_current_sec();
1006 LDLM_DEBUG(lock, "server preparing completion AST");
1008 ptlrpc_request_set_replen(req);
1010 req->rq_send_state = LUSTRE_IMP_FULL;
1011 /* ptlrpc_request_pack already set timeout */
1013 req->rq_timeout = ldlm_get_rq_timeout();
1015 /* We only send real blocking ASTs after the lock is granted */
1016 lock_res_and_lock(lock);
1017 if (ldlm_is_ast_sent(lock)) {
1018 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1019 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1020 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1023 /* We might get here prior to ldlm_handle_enqueue setting
1024 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1025 * into waiting list, but this is safe and similar code in
1026 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1027 * that would not only cancel the lock, but will also remove
1028 * it from waiting list */
1029 if (ldlm_is_cancel_on_block(lock)) {
1030 unlock_res_and_lock(lock);
1031 ldlm_lock_cancel(lock);
1034 req->rq_no_resend = 1;
1036 lock_res_and_lock(lock);
1038 /* start the lock-timeout clock */
1039 ldlm_add_waiting_lock(lock);
1040 /* Do not resend after lock callback timeout */
1041 req->rq_delay_limit = ldlm_bl_timeout(lock);
1042 req->rq_resend_cb = ldlm_update_resend;
1045 unlock_res_and_lock(lock);
1047 if (lock->l_export && lock->l_export->exp_nid_stats &&
1048 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1049 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1050 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1052 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1054 RETURN(lvb_len < 0 ? lvb_len : rc);
1058 * Server side ->l_glimpse_ast handler for client locks.
1060 * Sends glimpse AST to the client and waits for reply. Then updates
1061 * lvbo with the result.
1063 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1065 struct ldlm_cb_set_arg *arg = data;
1066 struct ldlm_request *body;
1067 struct ptlrpc_request *req;
1068 struct ldlm_cb_async_args *ca;
1070 struct req_format *req_fmt;
1073 LASSERT(lock != NULL);
1075 if (arg->gl_desc != NULL)
1076 /* There is a glimpse descriptor to pack */
1077 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1079 req_fmt = &RQF_LDLM_GL_CALLBACK;
1081 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1082 req_fmt, LUSTRE_DLM_VERSION,
1088 if (arg->gl_desc != NULL) {
1089 /* copy the GL descriptor */
1090 union ldlm_gl_desc *desc;
1091 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1092 *desc = *arg->gl_desc;
1095 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1096 body->lock_handle[0] = lock->l_remote_handle;
1097 ldlm_lock2desc(lock, &body->lock_desc);
1099 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1100 ca = ptlrpc_req_async_args(req);
1101 ca->ca_set_arg = arg;
1104 /* server namespace, doesn't need lock */
1105 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1106 ldlm_lvbo_size(lock));
1107 ptlrpc_request_set_replen(req);
1109 req->rq_send_state = LUSTRE_IMP_FULL;
1110 /* ptlrpc_request_alloc_pack already set timeout */
1112 req->rq_timeout = ldlm_get_rq_timeout();
1114 lock->l_last_activity = cfs_time_current_sec();
1116 req->rq_interpret_reply = ldlm_cb_interpret;
1118 if (lock->l_export && lock->l_export->exp_nid_stats &&
1119 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1120 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1121 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1123 rc = ldlm_ast_fini(req, arg, lock, 0);
1128 int ldlm_glimpse_locks(struct ldlm_resource *res,
1129 struct list_head *gl_work_list)
1134 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1136 if (rc == -ERESTART)
1137 ldlm_reprocess_all(res);
1141 EXPORT_SYMBOL(ldlm_glimpse_locks);
1143 /* return LDLM lock associated with a lock callback request */
1144 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1146 struct ldlm_cb_async_args *ca;
1147 struct ldlm_lock *lock;
1150 ca = ptlrpc_req_async_args(req);
1153 RETURN(ERR_PTR(-EFAULT));
1157 EXPORT_SYMBOL(ldlm_request_lock);
1159 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1160 struct lprocfs_stats *srv_stats)
1162 int lock_type = 0, op = 0;
1164 lock_type = dlm_req->lock_desc.l_resource.lr_type;
1166 switch (lock_type) {
1168 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1171 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1172 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1174 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1177 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1180 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1188 lprocfs_counter_incr(srv_stats, op);
1194 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1195 * service threads to carry out client lock enqueueing requests.
1197 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1198 struct ptlrpc_request *req,
1199 const struct ldlm_request *dlm_req,
1200 const struct ldlm_callback_suite *cbs)
1202 struct ldlm_reply *dlm_rep;
1204 enum ldlm_error err = ELDLM_OK;
1205 struct ldlm_lock *lock = NULL;
1206 void *cookie = NULL;
1208 struct ldlm_resource *res = NULL;
1211 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1213 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1214 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1216 LASSERT(req->rq_export);
1218 if (ptlrpc_req2svc(req)->srv_stats != NULL)
1219 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1221 if (req->rq_export && req->rq_export->exp_nid_stats &&
1222 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1223 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1224 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1226 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1227 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1228 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1229 dlm_req->lock_desc.l_resource.lr_type);
1230 GOTO(out, rc = -EFAULT);
1233 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1234 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1235 dlm_req->lock_desc.l_req_mode &
1236 (dlm_req->lock_desc.l_req_mode-1))) {
1237 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1238 dlm_req->lock_desc.l_req_mode);
1239 GOTO(out, rc = -EFAULT);
1242 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) {
1243 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1245 DEBUG_REQ(D_ERROR, req,
1246 "PLAIN lock request from IBITS client?");
1247 GOTO(out, rc = -EPROTO);
1249 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1251 DEBUG_REQ(D_ERROR, req,
1252 "IBITS lock request from unaware client?");
1253 GOTO(out, rc = -EPROTO);
1257 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1258 against server's _CONNECT_SUPPORTED flags? (I don't want to use
1259 ibits for mgc/mgs) */
1261 /* INODEBITS_INTEROP: Perform conversion from plain lock to
1262 * inodebits lock if client does not support them. */
1263 if (!(exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) &&
1264 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1265 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1266 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1267 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1268 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1269 dlm_req->lock_desc.l_req_mode = LCK_CR;
1273 if (unlikely((flags & LDLM_FL_REPLAY) ||
1274 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1275 /* Find an existing lock in the per-export lock hash */
1276 /* In the function below, .hs_keycmp resolves to
1277 * ldlm_export_lock_keycmp() */
1278 /* coverity[overrun-buffer-val] */
1279 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1280 (void *)&dlm_req->lock_handle[0]);
1282 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1283 LPX64, lock->l_handle.h_cookie);
1284 flags |= LDLM_FL_RESENT;
1285 GOTO(existing_lock, rc = 0);
1288 if (ldlm_reclaim_full()) {
1289 DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1290 "reject current enqueue request and let the "
1291 "client retry later.\n");
1292 GOTO(out, rc = -EINPROGRESS);
1296 /* The lock's callback data might be set in the policy function */
1297 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1298 dlm_req->lock_desc.l_resource.lr_type,
1299 dlm_req->lock_desc.l_req_mode,
1300 cbs, NULL, 0, LVB_T_NONE);
1307 lock->l_remote_handle = dlm_req->lock_handle[0];
1308 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1310 /* Initialize resource lvb but not for a lock being replayed since
1311 * Client already got lvb sent in this case.
1312 * This must occur early since some policy methods assume resource
1313 * lvb is available (lr_lvb_data != NULL).
1315 res = lock->l_resource;
1316 if (!(flags & LDLM_FL_REPLAY)) {
1317 /* non-replayed lock, delayed lvb init may need to be done */
1318 rc = ldlm_lvbo_init(res);
1320 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1325 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1326 /* Don't enqueue a lock onto the export if it is been disonnected
1327 * due to eviction (bug 3822) or server umount (bug 24324).
1328 * Cancel it now instead. */
1329 if (req->rq_export->exp_disconnected) {
1330 LDLM_ERROR(lock, "lock on disconnected export %p",
1332 GOTO(out, rc = -ENOTCONN);
1335 lock->l_export = class_export_lock_get(req->rq_export, lock);
1336 if (lock->l_export->exp_lock_hash)
1337 cfs_hash_add(lock->l_export->exp_lock_hash,
1338 &lock->l_remote_handle,
1341 /* Inherit the enqueue flags before the operation, because we do not
1342 * keep the res lock on return and next operations (BL AST) may proceed
1344 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1345 LDLM_FL_INHERIT_MASK);
1348 if (flags & LDLM_FL_HAS_INTENT) {
1349 /* In this case, the reply buffer is allocated deep in
1350 * local_lock_enqueue by the policy function. */
1353 /* based on the assumption that lvb size never changes during
1354 * resource life time otherwise it need resource->lr_lock's
1356 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1357 RCL_SERVER, ldlm_lvbo_size(lock));
1359 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1360 GOTO(out, rc = -ENOMEM);
1362 rc = req_capsule_server_pack(&req->rq_pill);
1367 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1368 ldlm_convert_policy_to_local(req->rq_export,
1369 dlm_req->lock_desc.l_resource.lr_type,
1370 &dlm_req->lock_desc.l_policy_data,
1371 &lock->l_policy_data);
1372 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1373 lock->l_req_extent = lock->l_policy_data.l_extent;
1375 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1382 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1384 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1385 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1387 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1388 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1390 /* We never send a blocking AST until the lock is granted, but
1391 * we can tell it right now */
1392 lock_res_and_lock(lock);
1394 /* Now take into account flags to be inherited from original lock
1395 request both in reply to client and in our own lock flags. */
1396 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1397 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1399 /* Don't move a pending lock onto the export if it has already been
1400 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1401 * Cancel it now instead. */
1402 if (unlikely(req->rq_export->exp_disconnected ||
1403 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1404 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1406 } else if (ldlm_is_ast_sent(lock)) {
1407 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1408 if (lock->l_granted_mode == lock->l_req_mode) {
1410 * Only cancel lock if it was granted, because it would
1411 * be destroyed immediately and would never be granted
1412 * in the future, causing timeouts on client. Not
1413 * granted lock will be cancelled immediately after
1414 * sending completion AST.
1416 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1417 unlock_res_and_lock(lock);
1418 ldlm_lock_cancel(lock);
1419 lock_res_and_lock(lock);
1421 ldlm_add_waiting_lock(lock);
1424 /* Make sure we never ever grant usual metadata locks to liblustre
1426 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1427 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1428 req->rq_export->exp_libclient) {
1429 if (unlikely(!ldlm_is_cancel_on_block(lock) ||
1430 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1431 CERROR("Granting sync lock to libclient. "
1432 "req fl %d, rep fl %d, lock fl "LPX64"\n",
1433 dlm_req->lock_flags, dlm_rep->lock_flags,
1435 LDLM_ERROR(lock, "sync lock");
1436 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1437 struct ldlm_intent *it;
1439 it = req_capsule_client_get(&req->rq_pill,
1442 CERROR("This is intent %s ("LPU64")\n",
1443 ldlm_it2str(it->opc), it->opc);
1449 unlock_res_and_lock(lock);
1453 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1454 if (!req->rq_packed_final) {
1455 err = lustre_pack_reply(req, 1, NULL, NULL);
1460 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1461 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1463 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1464 "(err=%d, rc=%d)", err, rc);
1467 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1469 ldlm_lvbo_size(lock) > 0) {
1473 buf = req_capsule_server_get(&req->rq_pill,
1475 LASSERTF(buf != NULL, "req %p, lock %p\n",
1477 buflen = req_capsule_get_size(&req->rq_pill,
1478 &RMF_DLM_LVB, RCL_SERVER);
1479 /* non-replayed lock, delayed lvb init may
1480 * need to be occur now */
1481 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1482 buflen = ldlm_lvbo_fill(lock, buf,
1488 buflen, RCL_SERVER);
1491 } else if (flags & LDLM_FL_REPLAY) {
1492 /* no LVB resend upon replay */
1507 if (lock->l_export) {
1508 ldlm_lock_cancel(lock);
1510 lock_res_and_lock(lock);
1511 ldlm_resource_unlink_lock(lock);
1512 ldlm_lock_destroy_nolock(lock);
1513 unlock_res_and_lock(lock);
1518 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1519 ldlm_reprocess_all(lock->l_resource);
1521 LDLM_LOCK_RELEASE(lock);
1524 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1531 * Old-style LDLM main entry point for server code enqueue.
1533 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1534 ldlm_completion_callback completion_callback,
1535 ldlm_blocking_callback blocking_callback,
1536 ldlm_glimpse_callback glimpse_callback)
1538 struct ldlm_request *dlm_req;
1539 struct ldlm_callback_suite cbs = {
1540 .lcs_completion = completion_callback,
1541 .lcs_blocking = blocking_callback,
1542 .lcs_glimpse = glimpse_callback
1546 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1547 if (dlm_req != NULL) {
1548 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1549 req, dlm_req, &cbs);
1557 * Main LDLM entry point for server code to process lock conversion requests.
1559 int ldlm_handle_convert0(struct ptlrpc_request *req,
1560 const struct ldlm_request *dlm_req)
1562 struct ldlm_reply *dlm_rep;
1563 struct ldlm_lock *lock;
1567 if (req->rq_export && req->rq_export->exp_nid_stats &&
1568 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1569 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1570 LDLM_CONVERT - LDLM_FIRST_OPC);
1572 rc = req_capsule_server_pack(&req->rq_pill);
1576 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1577 dlm_rep->lock_flags = dlm_req->lock_flags;
1579 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1581 req->rq_status = LUSTRE_EINVAL;
1585 LDLM_DEBUG(lock, "server-side convert handler START");
1587 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1588 &dlm_rep->lock_flags);
1590 if (ldlm_del_waiting_lock(lock))
1591 LDLM_DEBUG(lock, "converted waiting lock");
1594 req->rq_status = LUSTRE_EDEADLK;
1599 if (!req->rq_status)
1600 ldlm_reprocess_all(lock->l_resource);
1601 LDLM_DEBUG(lock, "server-side convert handler END");
1602 LDLM_LOCK_PUT(lock);
1604 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1610 * Old-style main LDLM entry point for server code to process lock conversion
1613 int ldlm_handle_convert(struct ptlrpc_request *req)
1616 struct ldlm_request *dlm_req;
1618 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1619 if (dlm_req != NULL) {
1620 rc = ldlm_handle_convert0(req, dlm_req);
1622 CERROR ("Can't unpack dlm_req\n");
1629 * Cancel all the locks whose handles are packed into ldlm_request
1631 * Called by server code expecting such combined cancel activity
1634 int ldlm_request_cancel(struct ptlrpc_request *req,
1635 const struct ldlm_request *dlm_req,
1636 int first, enum lustre_at_flags flags)
1638 struct ldlm_resource *res, *pres = NULL;
1639 struct ldlm_lock *lock;
1640 int i, count, done = 0;
1643 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1647 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1650 /* There is no lock on the server at the replay time,
1651 * skip lock cancelling to make replay tests to pass. */
1652 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1655 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1656 "starting at %d", count, first);
1658 for (i = first; i < count; i++) {
1659 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1661 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1662 "lock (cookie "LPU64")",
1663 dlm_req->lock_handle[i].cookie);
1667 res = lock->l_resource;
1670 /* This code is an optimization to only attempt lock
1671 * granting on the resource (that could be CPU-expensive)
1672 * after we are done cancelling lock in that resource. */
1675 ldlm_reprocess_all(pres);
1676 LDLM_RESOURCE_DELREF(pres);
1677 ldlm_resource_putref(pres);
1680 ldlm_resource_getref(res);
1681 LDLM_RESOURCE_ADDREF(res);
1682 ldlm_res_lvbo_update(res, NULL, 1);
1687 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock)) {
1688 long delay = cfs_time_sub(cfs_time_current_sec(),
1689 lock->l_last_activity);
1690 LDLM_DEBUG(lock, "server cancels blocked lock after "
1691 CFS_DURATION_T"s", delay);
1692 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1694 ldlm_lock_cancel(lock);
1695 LDLM_LOCK_PUT(lock);
1698 ldlm_reprocess_all(pres);
1699 LDLM_RESOURCE_DELREF(pres);
1700 ldlm_resource_putref(pres);
1702 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1705 EXPORT_SYMBOL(ldlm_request_cancel);
1708 * Main LDLM entry point for server code to cancel locks.
1710 * Typically gets called from service handler on LDLM_CANCEL opc.
1712 int ldlm_handle_cancel(struct ptlrpc_request *req)
1714 struct ldlm_request *dlm_req;
1718 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1719 if (dlm_req == NULL) {
1720 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1724 if (req->rq_export && req->rq_export->exp_nid_stats &&
1725 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1726 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1727 LDLM_CANCEL - LDLM_FIRST_OPC);
1729 rc = req_capsule_server_pack(&req->rq_pill);
1733 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1734 req->rq_status = LUSTRE_ESTALE;
1736 RETURN(ptlrpc_reply(req));
1738 #endif /* HAVE_SERVER_SUPPORT */
1741 * Callback handler for receiving incoming blocking ASTs.
1743 * This can only happen on client side.
1745 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1746 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1751 LDLM_DEBUG(lock, "client blocking AST callback handler");
1753 lock_res_and_lock(lock);
1754 ldlm_set_cbpending(lock);
1756 if (ldlm_is_cancel_on_block(lock))
1757 ldlm_set_cancel(lock);
1759 do_ast = (!lock->l_readers && !lock->l_writers);
1760 unlock_res_and_lock(lock);
1763 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1764 lock, lock->l_blocking_ast);
1765 if (lock->l_blocking_ast != NULL)
1766 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1769 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1773 LDLM_DEBUG(lock, "client blocking callback handler END");
1774 LDLM_LOCK_RELEASE(lock);
1779 * Callback handler for receiving incoming completion ASTs.
1781 * This only can happen on client side.
1783 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1784 struct ldlm_namespace *ns,
1785 struct ldlm_request *dlm_req,
1786 struct ldlm_lock *lock)
1788 struct list_head ast_list;
1793 LDLM_DEBUG(lock, "client completion callback handler START");
1795 INIT_LIST_HEAD(&ast_list);
1796 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1797 int to = cfs_time_seconds(1);
1799 set_current_state(TASK_INTERRUPTIBLE);
1800 schedule_timeout(to);
1801 if (lock->l_granted_mode == lock->l_req_mode ||
1802 ldlm_is_destroyed(lock))
1807 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1809 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1810 GOTO(out, rc = lvb_len);
1811 } else if (lvb_len > 0) {
1812 if (lock->l_lvb_len > 0) {
1813 /* for extent lock, lvb contains ost_lvb{}. */
1814 LASSERT(lock->l_lvb_data != NULL);
1816 if (unlikely(lock->l_lvb_len < lvb_len)) {
1817 LDLM_ERROR(lock, "Replied LVB is larger than "
1818 "expectation, expected = %d, "
1820 lock->l_lvb_len, lvb_len);
1821 GOTO(out, rc = -EINVAL);
1826 lock_res_and_lock(lock);
1827 if (ldlm_is_destroyed(lock) ||
1828 lock->l_granted_mode == lock->l_req_mode) {
1829 /* bug 11300: the lock has already been granted */
1830 unlock_res_and_lock(lock);
1831 LDLM_DEBUG(lock, "Double grant race happened");
1835 /* If we receive the completion AST before the actual enqueue returned,
1836 * then we might need to switch lock modes, resources, or extents. */
1837 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1838 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1839 LDLM_DEBUG(lock, "completion AST, new lock mode");
1842 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1843 ldlm_convert_policy_to_local(req->rq_export,
1844 dlm_req->lock_desc.l_resource.lr_type,
1845 &dlm_req->lock_desc.l_policy_data,
1846 &lock->l_policy_data);
1847 LDLM_DEBUG(lock, "completion AST, new policy data");
1850 ldlm_resource_unlink_lock(lock);
1851 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1852 &lock->l_resource->lr_name,
1853 sizeof(lock->l_resource->lr_name)) != 0) {
1854 unlock_res_and_lock(lock);
1855 rc = ldlm_lock_change_resource(ns, lock,
1856 &dlm_req->lock_desc.l_resource.lr_name);
1858 LDLM_ERROR(lock, "Failed to allocate resource");
1861 LDLM_DEBUG(lock, "completion AST, new resource");
1862 CERROR("change resource!\n");
1863 lock_res_and_lock(lock);
1866 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1867 /* BL_AST locks are not needed in LRU.
1868 * Let ldlm_cancel_lru() be fast. */
1869 ldlm_lock_remove_from_lru(lock);
1870 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1871 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1874 if (lock->l_lvb_len > 0) {
1875 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1876 lock->l_lvb_data, lvb_len);
1878 unlock_res_and_lock(lock);
1883 ldlm_grant_lock(lock, &ast_list);
1884 unlock_res_and_lock(lock);
1886 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1888 /* Let Enqueue to call osc_lock_upcall() and initialize
1890 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1892 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1894 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1900 lock_res_and_lock(lock);
1901 ldlm_set_failed(lock);
1902 unlock_res_and_lock(lock);
1903 wake_up(&lock->l_waitq);
1905 LDLM_LOCK_RELEASE(lock);
1909 * Callback handler for receiving incoming glimpse ASTs.
1911 * This only can happen on client side. After handling the glimpse AST
1912 * we also consider dropping the lock here if it is unused locally for a
1915 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1916 struct ldlm_namespace *ns,
1917 struct ldlm_request *dlm_req,
1918 struct ldlm_lock *lock)
1923 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1925 if (lock->l_glimpse_ast != NULL)
1926 rc = lock->l_glimpse_ast(lock, req);
1928 if (req->rq_repmsg != NULL) {
1931 req->rq_status = rc;
1935 lock_res_and_lock(lock);
1936 if (lock->l_granted_mode == LCK_PW &&
1937 !lock->l_readers && !lock->l_writers &&
1938 cfs_time_after(cfs_time_current(),
1939 cfs_time_add(lock->l_last_used,
1940 cfs_time_seconds(10)))) {
1941 unlock_res_and_lock(lock);
1942 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1943 ldlm_handle_bl_callback(ns, NULL, lock);
1948 unlock_res_and_lock(lock);
1949 LDLM_LOCK_RELEASE(lock);
1953 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1955 if (req->rq_no_reply)
1958 req->rq_status = rc;
1959 if (!req->rq_packed_final) {
1960 rc = lustre_pack_reply(req, 1, NULL, NULL);
1964 return ptlrpc_reply(req);
1967 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1968 enum ldlm_cancel_flags cancel_flags)
1970 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1973 spin_lock(&blp->blp_lock);
1974 if (blwi->blwi_lock &&
1975 ldlm_is_discard_data(blwi->blwi_lock)) {
1976 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1977 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1979 /* other blocking callbacks are added to the regular list */
1980 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1982 spin_unlock(&blp->blp_lock);
1984 wake_up(&blp->blp_waitq);
1986 /* can not check blwi->blwi_flags as blwi could be already freed in
1988 if (!(cancel_flags & LCF_ASYNC))
1989 wait_for_completion(&blwi->blwi_comp);
1994 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1995 struct ldlm_namespace *ns,
1996 struct ldlm_lock_desc *ld,
1997 struct list_head *cancels, int count,
1998 struct ldlm_lock *lock,
1999 enum ldlm_cancel_flags cancel_flags)
2001 init_completion(&blwi->blwi_comp);
2002 INIT_LIST_HEAD(&blwi->blwi_head);
2004 if (memory_pressure_get())
2005 blwi->blwi_mem_pressure = 1;
2008 blwi->blwi_flags = cancel_flags;
2010 blwi->blwi_ld = *ld;
2012 list_add(&blwi->blwi_head, cancels);
2013 list_del_init(cancels);
2014 blwi->blwi_count = count;
2016 blwi->blwi_lock = lock;
2021 * Queues a list of locks \a cancels containing \a count locks
2022 * for later processing by a blocking thread. If \a count is zero,
2023 * then the lock referenced as \a lock is queued instead.
2025 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2026 * If list addition fails an error is returned and caller is supposed to
2027 * call ->l_blocking_ast itself.
2029 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2030 struct ldlm_lock_desc *ld,
2031 struct ldlm_lock *lock,
2032 struct list_head *cancels, int count,
2033 enum ldlm_cancel_flags cancel_flags)
2037 if (cancels && count == 0)
2040 if (cancel_flags & LCF_ASYNC) {
2041 struct ldlm_bl_work_item *blwi;
2043 OBD_ALLOC(blwi, sizeof(*blwi));
2046 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2048 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2050 /* if it is synchronous call do minimum mem alloc, as it could
2051 * be triggered from kernel shrinker
2053 struct ldlm_bl_work_item blwi;
2055 memset(&blwi, 0, sizeof(blwi));
2056 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2057 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2062 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2063 struct ldlm_lock *lock)
2065 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2068 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2069 struct list_head *cancels, int count,
2070 enum ldlm_cancel_flags cancel_flags)
2072 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2075 int ldlm_bl_thread_wakeup(void)
2077 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2081 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2082 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2084 struct obd_device *obd = req->rq_export->exp_obd;
2091 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2093 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2095 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2097 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2100 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2102 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2104 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2107 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2110 /* We are responsible for swabbing contents of val */
2112 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2113 /* Pass it on to mdc (the "export" in this case) */
2114 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2116 sizeof(KEY_HSM_COPYTOOL_SEND),
2117 KEY_HSM_COPYTOOL_SEND,
2120 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2125 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2126 const char *msg, int rc,
2127 struct lustre_handle *handle)
2129 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2130 "%s: [nid %s] [rc %d] [lock "LPX64"]",
2131 msg, libcfs_id2str(req->rq_peer), rc,
2132 handle ? handle->cookie : 0);
2133 if (req->rq_no_reply)
2134 CWARN("No reply was sent, maybe cause bug 21636.\n");
2136 CWARN("Send reply failed, maybe cause bug 21636.\n");
2139 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2140 static int ldlm_callback_handler(struct ptlrpc_request *req)
2142 struct ldlm_namespace *ns;
2143 struct ldlm_request *dlm_req;
2144 struct ldlm_lock *lock;
2148 /* Requests arrive in sender's byte order. The ptlrpc service
2149 * handler has already checked and, if necessary, byte-swapped the
2150 * incoming request message body, but I am responsible for the
2151 * message buffers. */
2153 /* do nothing for sec context finalize */
2154 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2157 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2159 if (req->rq_export == NULL) {
2160 rc = ldlm_callback_reply(req, -ENOTCONN);
2161 ldlm_callback_errmsg(req, "Operate on unconnected server",
2166 LASSERT(req->rq_export != NULL);
2167 LASSERT(req->rq_export->exp_obd != NULL);
2169 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2170 case LDLM_BL_CALLBACK:
2171 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2173 ldlm_callback_reply(req, -(int)cfs_fail_err);
2177 case LDLM_CP_CALLBACK:
2178 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2181 case LDLM_GL_CALLBACK:
2182 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2186 rc = ldlm_handle_setinfo(req);
2187 ldlm_callback_reply(req, rc);
2189 case LLOG_ORIGIN_HANDLE_CREATE:
2190 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2191 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2193 rc = llog_origin_handle_open(req);
2194 ldlm_callback_reply(req, rc);
2196 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2197 req_capsule_set(&req->rq_pill,
2198 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2199 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2201 rc = llog_origin_handle_next_block(req);
2202 ldlm_callback_reply(req, rc);
2204 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2205 req_capsule_set(&req->rq_pill,
2206 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2207 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2209 rc = llog_origin_handle_read_header(req);
2210 ldlm_callback_reply(req, rc);
2212 case LLOG_ORIGIN_HANDLE_CLOSE:
2213 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2215 rc = llog_origin_handle_close(req);
2216 ldlm_callback_reply(req, rc);
2219 CERROR("unknown opcode %u\n",
2220 lustre_msg_get_opc(req->rq_reqmsg));
2221 ldlm_callback_reply(req, -EPROTO);
2225 ns = req->rq_export->exp_obd->obd_namespace;
2226 LASSERT(ns != NULL);
2228 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2230 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2231 if (dlm_req == NULL) {
2232 rc = ldlm_callback_reply(req, -EPROTO);
2233 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2238 /* Force a known safe race, send a cancel to the server for a lock
2239 * which the server has already started a blocking callback on. */
2240 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2241 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2242 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2244 CERROR("ldlm_cli_cancel: %d\n", rc);
2247 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2249 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2250 "disappeared\n", dlm_req->lock_handle[0].cookie);
2251 rc = ldlm_callback_reply(req, -EINVAL);
2252 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2253 &dlm_req->lock_handle[0]);
2257 if (ldlm_is_fail_loc(lock) &&
2258 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2259 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2261 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2262 lock_res_and_lock(lock);
2263 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2265 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2266 /* If somebody cancels lock and cache is already dropped,
2267 * or lock is failed before cp_ast received on client,
2268 * we can tell the server we have no lock. Otherwise, we
2269 * should send cancel after dropping the cache. */
2270 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2271 ldlm_is_failed(lock)) {
2272 LDLM_DEBUG(lock, "callback on lock "
2273 LPX64" - lock disappeared\n",
2274 dlm_req->lock_handle[0].cookie);
2275 unlock_res_and_lock(lock);
2276 LDLM_LOCK_RELEASE(lock);
2277 rc = ldlm_callback_reply(req, -EINVAL);
2278 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2279 &dlm_req->lock_handle[0]);
2282 /* BL_AST locks are not needed in LRU.
2283 * Let ldlm_cancel_lru() be fast. */
2284 ldlm_lock_remove_from_lru(lock);
2285 ldlm_set_bl_ast(lock);
2287 unlock_res_and_lock(lock);
2289 /* We want the ost thread to get this reply so that it can respond
2290 * to ost requests (write cache writeback) that might be triggered
2293 * But we'd also like to be able to indicate in the reply that we're
2294 * cancelling right now, because it's unused, or have an intent result
2295 * in the reply, so we might have to push the responsibility for sending
2296 * the reply down into the AST handlers, alas. */
2298 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2299 case LDLM_BL_CALLBACK:
2300 CDEBUG(D_INODE, "blocking ast\n");
2301 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2302 if (!ldlm_is_cancel_on_block(lock)) {
2303 rc = ldlm_callback_reply(req, 0);
2304 if (req->rq_no_reply || rc)
2305 ldlm_callback_errmsg(req, "Normal process", rc,
2306 &dlm_req->lock_handle[0]);
2308 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2309 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2311 case LDLM_CP_CALLBACK:
2312 CDEBUG(D_INODE, "completion ast\n");
2313 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2314 ldlm_callback_reply(req, 0);
2315 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2317 case LDLM_GL_CALLBACK:
2318 CDEBUG(D_INODE, "glimpse ast\n");
2319 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2320 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2323 LBUG(); /* checked above */
2329 #ifdef HAVE_SERVER_SUPPORT
2331 * Main handler for canceld thread.
2333 * Separated into its own thread to avoid deadlocks.
2335 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2340 /* Requests arrive in sender's byte order. The ptlrpc service
2341 * handler has already checked and, if necessary, byte-swapped the
2342 * incoming request message body, but I am responsible for the
2343 * message buffers. */
2345 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2347 if (req->rq_export == NULL) {
2348 struct ldlm_request *dlm_req;
2350 CERROR("%s from %s arrived at %lu with bad export cookie "
2352 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2353 libcfs_nid2str(req->rq_peer.nid),
2354 req->rq_arrival_time.tv_sec,
2355 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2357 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2358 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2359 dlm_req = req_capsule_client_get(&req->rq_pill,
2361 if (dlm_req != NULL)
2362 ldlm_lock_dump_handle(D_ERROR,
2363 &dlm_req->lock_handle[0]);
2365 ldlm_callback_reply(req, -ENOTCONN);
2369 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2371 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2373 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2374 CDEBUG(D_INODE, "cancel\n");
2375 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2376 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2377 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2379 rc = ldlm_handle_cancel(req);
2384 CERROR("invalid opcode %d\n",
2385 lustre_msg_get_opc(req->rq_reqmsg));
2386 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2387 ldlm_callback_reply(req, -EINVAL);
2393 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2394 struct ldlm_lock *lock)
2396 struct ldlm_request *dlm_req;
2397 struct lustre_handle lockh;
2402 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2403 if (dlm_req == NULL)
2406 ldlm_lock2handle(lock, &lockh);
2407 for (i = 0; i < dlm_req->lock_count; i++) {
2408 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2410 DEBUG_REQ(D_RPCTRACE, req,
2411 "Prio raised by lock "LPX64".", lockh.cookie);
2422 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2424 struct ldlm_request *dlm_req;
2429 /* no prolong in recovery */
2430 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2433 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2434 if (dlm_req == NULL)
2437 for (i = 0; i < dlm_req->lock_count; i++) {
2438 struct ldlm_lock *lock;
2440 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2444 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2446 LDLM_DEBUG(lock, "hpreq cancel lock");
2447 LDLM_LOCK_PUT(lock);
2456 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2457 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2458 .hpreq_check = ldlm_cancel_hpreq_check,
2462 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2466 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2468 if (req->rq_export == NULL)
2471 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2472 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2473 req->rq_ops = &ldlm_cancel_hpreq_ops;
2478 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2479 struct hlist_node *hnode, void *data)
2482 struct list_head *rpc_list = data;
2483 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2485 lock_res_and_lock(lock);
2487 if (lock->l_req_mode != lock->l_granted_mode) {
2488 unlock_res_and_lock(lock);
2492 LASSERT(lock->l_resource);
2493 if (lock->l_resource->lr_type != LDLM_IBITS &&
2494 lock->l_resource->lr_type != LDLM_PLAIN) {
2495 unlock_res_and_lock(lock);
2499 if (ldlm_is_ast_sent(lock)) {
2500 unlock_res_and_lock(lock);
2504 LASSERT(lock->l_blocking_ast);
2505 LASSERT(!lock->l_blocking_lock);
2507 ldlm_set_ast_sent(lock);
2508 if (lock->l_export && lock->l_export->exp_lock_hash) {
2509 /* NB: it's safe to call cfs_hash_del() even lock isn't
2510 * in exp_lock_hash. */
2511 /* In the function below, .hs_keycmp resolves to
2512 * ldlm_export_lock_keycmp() */
2513 /* coverity[overrun-buffer-val] */
2514 cfs_hash_del(lock->l_export->exp_lock_hash,
2515 &lock->l_remote_handle, &lock->l_exp_hash);
2518 list_add_tail(&lock->l_rk_ast, rpc_list);
2519 LDLM_LOCK_GET(lock);
2521 unlock_res_and_lock(lock);
2525 void ldlm_revoke_export_locks(struct obd_export *exp)
2527 struct list_head rpc_list;
2530 INIT_LIST_HEAD(&rpc_list);
2531 cfs_hash_for_each_empty(exp->exp_lock_hash,
2532 ldlm_revoke_lock_cb, &rpc_list);
2533 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2534 LDLM_WORK_REVOKE_AST);
2538 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2539 #endif /* HAVE_SERVER_SUPPORT */
2541 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2542 struct ldlm_bl_work_item **p_blwi,
2543 struct obd_export **p_exp)
2545 struct ldlm_bl_work_item *blwi = NULL;
2546 static unsigned int num_bl = 0;
2547 static unsigned int num_stale;
2548 int num_th = atomic_read(&blp->blp_num_threads);
2550 *p_exp = obd_stale_export_get();
2552 spin_lock(&blp->blp_lock);
2553 if (*p_exp != NULL) {
2554 if (num_th == 1 || ++num_stale < num_th) {
2555 spin_unlock(&blp->blp_lock);
2562 /* process a request from the blp_list at least every blp_num_threads */
2563 if (!list_empty(&blp->blp_list) &&
2564 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2565 blwi = list_entry(blp->blp_list.next,
2566 struct ldlm_bl_work_item, blwi_entry);
2568 if (!list_empty(&blp->blp_prio_list))
2569 blwi = list_entry(blp->blp_prio_list.next,
2570 struct ldlm_bl_work_item,
2574 if (++num_bl >= num_th)
2576 list_del(&blwi->blwi_entry);
2578 spin_unlock(&blp->blp_lock);
2581 if (*p_exp != NULL && *p_blwi != NULL) {
2582 obd_stale_export_put(*p_exp);
2586 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2589 /* This only contains temporary data until the thread starts */
2590 struct ldlm_bl_thread_data {
2591 char bltd_name[CFS_CURPROC_COMM_MAX];
2592 struct ldlm_bl_pool *bltd_blp;
2593 struct completion bltd_comp;
2597 static int ldlm_bl_thread_main(void *arg);
2599 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2601 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2602 struct task_struct *task;
2604 init_completion(&bltd.bltd_comp);
2605 bltd.bltd_num = atomic_read(&blp->blp_num_threads);
2606 snprintf(bltd.bltd_name, sizeof(bltd.bltd_name) - 1,
2607 "ldlm_bl_%02d", bltd.bltd_num);
2608 task = kthread_run(ldlm_bl_thread_main, &bltd, bltd.bltd_name);
2610 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2611 atomic_read(&blp->blp_num_threads), PTR_ERR(task));
2612 return PTR_ERR(task);
2614 wait_for_completion(&bltd.bltd_comp);
2619 /* Not fatal if racy and have a few too many threads */
2620 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2621 struct ldlm_bl_work_item *blwi)
2623 int busy = atomic_read(&blp->blp_busy_threads);
2625 if (busy >= blp->blp_max_threads)
2628 if (busy < atomic_read(&blp->blp_num_threads))
2631 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2632 blwi->blwi_mem_pressure))
2638 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2639 struct ldlm_bl_work_item *blwi)
2643 if (blwi->blwi_ns == NULL)
2644 /* added by ldlm_cleanup() */
2645 RETURN(LDLM_ITER_STOP);
2647 if (blwi->blwi_mem_pressure)
2648 memory_pressure_set();
2650 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2652 if (blwi->blwi_count) {
2654 /* The special case when we cancel locks in lru
2655 * asynchronously, we pass the list of locks here.
2656 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2657 * canceled locally yet. */
2658 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2661 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2664 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2667 if (blwi->blwi_mem_pressure)
2668 memory_pressure_clr();
2670 if (blwi->blwi_flags & LCF_ASYNC)
2671 OBD_FREE(blwi, sizeof(*blwi));
2673 complete(&blwi->blwi_comp);
2679 * Cancel stale locks on export. Cancel blocked locks first.
2680 * If the given export has blocked locks, the next in the list may have
2681 * them too, thus cancel not blocked locks only if the current export has
2684 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2685 struct obd_export *exp)
2690 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2692 num = ldlm_export_cancel_blocked_locks(exp);
2694 ldlm_export_cancel_locks(exp);
2696 obd_stale_export_put(exp);
2703 * Main blocking requests processing thread.
2705 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2706 * This thread in the end ends up doing actual call to ->l_blocking_ast
2709 static int ldlm_bl_thread_main(void *arg)
2711 struct ldlm_bl_pool *blp;
2712 struct ldlm_bl_thread_data *bltd = arg;
2715 blp = bltd->bltd_blp;
2717 atomic_inc(&blp->blp_num_threads);
2718 atomic_inc(&blp->blp_busy_threads);
2720 complete(&bltd->bltd_comp);
2721 /* cannot use bltd after this, it is only on caller's stack */
2724 struct l_wait_info lwi = { 0 };
2725 struct ldlm_bl_work_item *blwi = NULL;
2726 struct obd_export *exp = NULL;
2729 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2732 atomic_dec(&blp->blp_busy_threads);
2733 l_wait_event_exclusive(blp->blp_waitq,
2734 ldlm_bl_get_work(blp, &blwi,
2737 atomic_inc(&blp->blp_busy_threads);
2740 if (ldlm_bl_thread_need_create(blp, blwi))
2741 /* discard the return value, we tried */
2742 ldlm_bl_thread_start(blp);
2745 rc = ldlm_bl_thread_exports(blp, exp);
2747 rc = ldlm_bl_thread_blwi(blp, blwi);
2749 if (rc == LDLM_ITER_STOP)
2753 atomic_dec(&blp->blp_busy_threads);
2754 atomic_dec(&blp->blp_num_threads);
2755 complete(&blp->blp_comp);
2760 static int ldlm_setup(void);
2761 static int ldlm_cleanup(void);
2763 int ldlm_get_ref(void)
2767 mutex_lock(&ldlm_ref_mutex);
2768 if (++ldlm_refcount == 1) {
2773 mutex_unlock(&ldlm_ref_mutex);
2778 void ldlm_put_ref(void)
2781 mutex_lock(&ldlm_ref_mutex);
2782 if (ldlm_refcount == 1) {
2783 int rc = ldlm_cleanup();
2785 CERROR("ldlm_cleanup failed: %d\n", rc);
2791 mutex_unlock(&ldlm_ref_mutex);
2797 * Export handle<->lock hash operations.
2800 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2802 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2806 ldlm_export_lock_key(struct hlist_node *hnode)
2808 struct ldlm_lock *lock;
2810 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2811 return &lock->l_remote_handle;
2815 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2817 struct ldlm_lock *lock;
2819 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2820 lock->l_remote_handle = *(struct lustre_handle *)key;
2824 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2826 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2830 ldlm_export_lock_object(struct hlist_node *hnode)
2832 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2836 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2838 struct ldlm_lock *lock;
2840 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2841 LDLM_LOCK_GET(lock);
2845 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2847 struct ldlm_lock *lock;
2849 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2850 LDLM_LOCK_RELEASE(lock);
2853 static struct cfs_hash_ops ldlm_export_lock_ops = {
2854 .hs_hash = ldlm_export_lock_hash,
2855 .hs_key = ldlm_export_lock_key,
2856 .hs_keycmp = ldlm_export_lock_keycmp,
2857 .hs_keycpy = ldlm_export_lock_keycpy,
2858 .hs_object = ldlm_export_lock_object,
2859 .hs_get = ldlm_export_lock_get,
2860 .hs_put = ldlm_export_lock_put,
2861 .hs_put_locked = ldlm_export_lock_put,
2864 int ldlm_init_export(struct obd_export *exp)
2869 exp->exp_lock_hash =
2870 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2871 HASH_EXP_LOCK_CUR_BITS,
2872 HASH_EXP_LOCK_MAX_BITS,
2873 HASH_EXP_LOCK_BKT_BITS, 0,
2874 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2875 &ldlm_export_lock_ops,
2876 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2877 CFS_HASH_NBLK_CHANGE);
2879 if (!exp->exp_lock_hash)
2882 rc = ldlm_init_flock_export(exp);
2888 ldlm_destroy_export(exp);
2891 EXPORT_SYMBOL(ldlm_init_export);
2893 void ldlm_destroy_export(struct obd_export *exp)
2896 cfs_hash_putref(exp->exp_lock_hash);
2897 exp->exp_lock_hash = NULL;
2899 ldlm_destroy_flock_export(exp);
2902 EXPORT_SYMBOL(ldlm_destroy_export);
2904 static int ldlm_setup(void)
2906 static struct ptlrpc_service_conf conf;
2907 struct ldlm_bl_pool *blp = NULL;
2908 #ifdef HAVE_SERVER_SUPPORT
2909 struct task_struct *task;
2910 #endif /* HAVE_SERVER_SUPPORT */
2916 if (ldlm_state != NULL)
2919 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2920 if (ldlm_state == NULL)
2923 #ifdef CONFIG_PROC_FS
2924 rc = ldlm_proc_setup();
2927 #endif /* CONFIG_PROC_FS */
2929 memset(&conf, 0, sizeof(conf));
2930 conf = (typeof(conf)) {
2931 .psc_name = "ldlm_cbd",
2932 .psc_watchdog_factor = 2,
2934 .bc_nbufs = LDLM_CLIENT_NBUFS,
2935 .bc_buf_size = LDLM_BUFSIZE,
2936 .bc_req_max_size = LDLM_MAXREQSIZE,
2937 .bc_rep_max_size = LDLM_MAXREPSIZE,
2938 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2939 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2942 .tc_thr_name = "ldlm_cb",
2943 .tc_thr_factor = LDLM_THR_FACTOR,
2944 .tc_nthrs_init = LDLM_NTHRS_INIT,
2945 .tc_nthrs_base = LDLM_NTHRS_BASE,
2946 .tc_nthrs_max = LDLM_NTHRS_MAX,
2947 .tc_nthrs_user = ldlm_num_threads,
2948 .tc_cpu_affinity = 1,
2949 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2952 .cc_pattern = ldlm_cpts,
2955 .so_req_handler = ldlm_callback_handler,
2958 ldlm_state->ldlm_cb_service = \
2959 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2960 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2961 CERROR("failed to start service\n");
2962 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2963 ldlm_state->ldlm_cb_service = NULL;
2967 #ifdef HAVE_SERVER_SUPPORT
2968 memset(&conf, 0, sizeof(conf));
2969 conf = (typeof(conf)) {
2970 .psc_name = "ldlm_canceld",
2971 .psc_watchdog_factor = 6,
2973 .bc_nbufs = LDLM_SERVER_NBUFS,
2974 .bc_buf_size = LDLM_BUFSIZE,
2975 .bc_req_max_size = LDLM_MAXREQSIZE,
2976 .bc_rep_max_size = LDLM_MAXREPSIZE,
2977 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
2978 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
2982 .tc_thr_name = "ldlm_cn",
2983 .tc_thr_factor = LDLM_THR_FACTOR,
2984 .tc_nthrs_init = LDLM_NTHRS_INIT,
2985 .tc_nthrs_base = LDLM_NTHRS_BASE,
2986 .tc_nthrs_max = LDLM_NTHRS_MAX,
2987 .tc_nthrs_user = ldlm_num_threads,
2988 .tc_cpu_affinity = 1,
2989 .tc_ctx_tags = LCT_MD_THREAD | \
2994 .cc_pattern = ldlm_cpts,
2997 .so_req_handler = ldlm_cancel_handler,
2998 .so_hpreq_handler = ldlm_hpreq_handler,
3001 ldlm_state->ldlm_cancel_service = \
3002 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
3003 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3004 CERROR("failed to start service\n");
3005 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3006 ldlm_state->ldlm_cancel_service = NULL;
3009 #endif /* HAVE_SERVER_SUPPORT */
3011 OBD_ALLOC(blp, sizeof(*blp));
3013 GOTO(out, rc = -ENOMEM);
3014 ldlm_state->ldlm_bl_pool = blp;
3016 spin_lock_init(&blp->blp_lock);
3017 INIT_LIST_HEAD(&blp->blp_list);
3018 INIT_LIST_HEAD(&blp->blp_prio_list);
3019 init_waitqueue_head(&blp->blp_waitq);
3020 atomic_set(&blp->blp_num_threads, 0);
3021 atomic_set(&blp->blp_busy_threads, 0);
3023 if (ldlm_num_threads == 0) {
3024 blp->blp_min_threads = LDLM_NTHRS_INIT;
3025 blp->blp_max_threads = LDLM_NTHRS_MAX;
3027 blp->blp_min_threads = blp->blp_max_threads = \
3028 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3032 for (i = 0; i < blp->blp_min_threads; i++) {
3033 rc = ldlm_bl_thread_start(blp);
3038 #ifdef HAVE_SERVER_SUPPORT
3039 INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
3040 expired_lock_thread.elt_state = ELT_STOPPED;
3041 init_waitqueue_head(&expired_lock_thread.elt_waitq);
3043 INIT_LIST_HEAD(&waiting_locks_list);
3044 spin_lock_init(&waiting_locks_spinlock);
3045 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, NULL);
3047 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3050 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3054 wait_event(expired_lock_thread.elt_waitq,
3055 expired_lock_thread.elt_state == ELT_READY);
3056 #endif /* HAVE_SERVER_SUPPORT */
3058 rc = ldlm_pools_init();
3060 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3064 rc = ldlm_reclaim_setup();
3066 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3076 static int ldlm_cleanup(void)
3080 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3081 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3082 CERROR("ldlm still has namespaces; clean these up first.\n");
3083 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3084 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3088 ldlm_reclaim_cleanup();
3091 if (ldlm_state->ldlm_bl_pool != NULL) {
3092 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3094 while (atomic_read(&blp->blp_num_threads) > 0) {
3095 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3097 init_completion(&blp->blp_comp);
3099 spin_lock(&blp->blp_lock);
3100 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3101 wake_up(&blp->blp_waitq);
3102 spin_unlock(&blp->blp_lock);
3104 wait_for_completion(&blp->blp_comp);
3107 OBD_FREE(blp, sizeof(*blp));
3110 if (ldlm_state->ldlm_cb_service != NULL)
3111 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3112 #ifdef HAVE_SERVER_SUPPORT
3113 if (ldlm_state->ldlm_cancel_service != NULL)
3114 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3117 ldlm_proc_cleanup();
3119 #ifdef HAVE_SERVER_SUPPORT
3120 if (expired_lock_thread.elt_state != ELT_STOPPED) {
3121 expired_lock_thread.elt_state = ELT_TERMINATE;
3122 wake_up(&expired_lock_thread.elt_waitq);
3123 wait_event(expired_lock_thread.elt_waitq,
3124 expired_lock_thread.elt_state == ELT_STOPPED);
3128 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3136 mutex_init(&ldlm_ref_mutex);
3137 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
3138 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
3140 INIT_LIST_HEAD(&ldlm_srv_namespace_list);
3141 INIT_LIST_HEAD(&ldlm_cli_active_namespace_list);
3142 INIT_LIST_HEAD(&ldlm_cli_inactive_namespace_list);
3144 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3145 sizeof(struct ldlm_resource), 0,
3146 SLAB_HWCACHE_ALIGN, NULL);
3147 if (ldlm_resource_slab == NULL)
3150 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3151 sizeof(struct ldlm_lock), 0,
3152 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3153 if (ldlm_lock_slab == NULL)
3156 ldlm_interval_slab = kmem_cache_create("interval_node",
3157 sizeof(struct ldlm_interval),
3158 0, SLAB_HWCACHE_ALIGN, NULL);
3159 if (ldlm_interval_slab == NULL)
3162 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3163 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3164 0, SLAB_HWCACHE_ALIGN, NULL);
3165 if (ldlm_interval_tree_slab == NULL)
3168 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3169 class_export_dump_hook = ldlm_dump_export_locks;
3174 kmem_cache_destroy(ldlm_interval_slab);
3176 kmem_cache_destroy(ldlm_lock_slab);
3178 kmem_cache_destroy(ldlm_resource_slab);
3183 void ldlm_exit(void)
3186 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3187 kmem_cache_destroy(ldlm_resource_slab);
3188 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3189 * synchronize_rcu() to wait a grace period elapsed, so that
3190 * ldlm_lock_free() get a chance to be called. */
3192 kmem_cache_destroy(ldlm_lock_slab);
3193 kmem_cache_destroy(ldlm_interval_slab);
3194 kmem_cache_destroy(ldlm_interval_tree_slab);