4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_lockd.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LDLM
45 # include <libcfs/libcfs.h>
47 # include <liblustre.h>
50 #include <lustre_dlm.h>
51 #include <obd_class.h>
52 #include <libcfs/list.h>
53 #include "ldlm_internal.h"
55 static int ldlm_num_threads;
56 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
57 "number of DLM service threads to start");
59 static char *ldlm_cpts;
60 CFS_MODULE_PARM(ldlm_cpts, "s", charp, 0444,
61 "CPU partitions ldlm threads should run on");
63 extern struct kmem_cache *ldlm_resource_slab;
64 extern struct kmem_cache *ldlm_lock_slab;
65 static struct mutex ldlm_ref_mutex;
66 static int ldlm_refcount;
68 struct ldlm_cb_async_args {
69 struct ldlm_cb_set_arg *ca_set_arg;
70 struct ldlm_lock *ca_lock;
75 static struct ldlm_state *ldlm_state;
77 inline cfs_time_t round_timeout(cfs_time_t timeout)
79 return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
82 /* timeout for initial callback (AST) reply (bz10399) */
83 static inline unsigned int ldlm_get_rq_timeout(void)
86 unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
88 return timeout < 1 ? 1 : timeout;
93 #define ELT_TERMINATE 2
99 * blp_prio_list is used for callbacks that should be handled
100 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
103 cfs_list_t blp_prio_list;
106 * blp_list is used for all other callbacks which are likely
107 * to take longer to process.
111 wait_queue_head_t blp_waitq;
112 struct completion blp_comp;
113 cfs_atomic_t blp_num_threads;
114 cfs_atomic_t blp_busy_threads;
119 struct ldlm_bl_work_item {
120 cfs_list_t blwi_entry;
121 struct ldlm_namespace *blwi_ns;
122 struct ldlm_lock_desc blwi_ld;
123 struct ldlm_lock *blwi_lock;
124 cfs_list_t blwi_head;
126 struct completion blwi_comp;
127 ldlm_cancel_flags_t blwi_flags;
128 int blwi_mem_pressure;
131 #if defined(HAVE_SERVER_SUPPORT) && defined(__KERNEL__)
134 * Protects both waiting_locks_list and expired_lock_thread.
136 static spinlock_t waiting_locks_spinlock; /* BH lock (timer) */
139 * List for contended locks.
141 * As soon as a lock is contended, it gets placed on this list and
142 * expected time to get a response is filled in the lock. A special
143 * thread walks the list looking for locks that should be released and
144 * schedules client evictions for those that have not been released in
147 * All access to it should be under waiting_locks_spinlock.
149 static cfs_list_t waiting_locks_list;
150 static struct timer_list waiting_locks_timer;
152 static struct expired_lock_thread {
153 wait_queue_head_t elt_waitq;
156 cfs_list_t elt_expired_locks;
157 } expired_lock_thread;
159 static inline int have_expired_locks(void)
164 spin_lock_bh(&waiting_locks_spinlock);
165 need_to_run = !cfs_list_empty(&expired_lock_thread.elt_expired_locks);
166 spin_unlock_bh(&waiting_locks_spinlock);
172 * Check expired lock list for expired locks and time them out.
174 static int expired_lock_main(void *arg)
176 cfs_list_t *expired = &expired_lock_thread.elt_expired_locks;
177 struct l_wait_info lwi = { 0 };
182 expired_lock_thread.elt_state = ELT_READY;
183 wake_up(&expired_lock_thread.elt_waitq);
186 l_wait_event(expired_lock_thread.elt_waitq,
187 have_expired_locks() ||
188 expired_lock_thread.elt_state == ELT_TERMINATE,
191 spin_lock_bh(&waiting_locks_spinlock);
192 if (expired_lock_thread.elt_dump) {
193 struct libcfs_debug_msg_data msgdata = {
194 .msg_file = __FILE__,
195 .msg_fn = "waiting_locks_callback",
196 .msg_line = expired_lock_thread.elt_dump };
197 spin_unlock_bh(&waiting_locks_spinlock);
199 /* from waiting_locks_callback, but not in timer */
200 libcfs_debug_dumplog();
201 libcfs_run_lbug_upcall(&msgdata);
203 spin_lock_bh(&waiting_locks_spinlock);
204 expired_lock_thread.elt_dump = 0;
209 while (!cfs_list_empty(expired)) {
210 struct obd_export *export;
211 struct ldlm_lock *lock;
213 lock = cfs_list_entry(expired->next, struct ldlm_lock,
215 if ((void *)lock < LP_POISON + PAGE_CACHE_SIZE &&
216 (void *)lock >= LP_POISON) {
217 spin_unlock_bh(&waiting_locks_spinlock);
218 CERROR("free lock on elt list %p\n", lock);
221 cfs_list_del_init(&lock->l_pending_chain);
222 if ((void *)lock->l_export <
223 LP_POISON + PAGE_CACHE_SIZE &&
224 (void *)lock->l_export >= LP_POISON) {
225 CERROR("lock with free export on elt list %p\n",
227 lock->l_export = NULL;
228 LDLM_ERROR(lock, "free export");
229 /* release extra ref grabbed by
230 * ldlm_add_waiting_lock() or
231 * ldlm_failed_ast() */
232 LDLM_LOCK_RELEASE(lock);
236 if (lock->l_flags & LDLM_FL_DESTROYED) {
237 /* release the lock refcount where
238 * waiting_locks_callback() founds */
239 LDLM_LOCK_RELEASE(lock);
242 export = class_export_lock_get(lock->l_export, lock);
243 spin_unlock_bh(&waiting_locks_spinlock);
246 class_fail_export(export);
247 class_export_lock_put(export, lock);
249 /* release extra ref grabbed by ldlm_add_waiting_lock()
250 * or ldlm_failed_ast() */
251 LDLM_LOCK_RELEASE(lock);
253 spin_lock_bh(&waiting_locks_spinlock);
255 spin_unlock_bh(&waiting_locks_spinlock);
257 if (do_dump && obd_dump_on_eviction) {
258 CERROR("dump the log upon eviction\n");
259 libcfs_debug_dumplog();
262 if (expired_lock_thread.elt_state == ELT_TERMINATE)
266 expired_lock_thread.elt_state = ELT_STOPPED;
267 wake_up(&expired_lock_thread.elt_waitq);
271 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
272 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds);
275 * Check if there is a request in the export request list
276 * which prevents the lock canceling.
278 static int ldlm_lock_busy(struct ldlm_lock *lock)
280 struct ptlrpc_request *req;
284 if (lock->l_export == NULL)
287 spin_lock_bh(&lock->l_export->exp_rpc_lock);
288 cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
290 if (req->rq_ops->hpreq_lock_match) {
291 match = req->rq_ops->hpreq_lock_match(req, lock);
296 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
300 /* This is called from within a timer interrupt and cannot schedule */
301 static void waiting_locks_callback(unsigned long unused)
303 struct ldlm_lock *lock;
306 spin_lock_bh(&waiting_locks_spinlock);
307 while (!cfs_list_empty(&waiting_locks_list)) {
308 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
310 if (cfs_time_after(lock->l_callback_timeout,
311 cfs_time_current()) ||
312 (lock->l_req_mode == LCK_GROUP))
315 /* Check if we need to prolong timeout */
316 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
317 ldlm_lock_busy(lock)) {
320 if (lock->l_pending_chain.next == &waiting_locks_list)
325 spin_unlock_bh(&waiting_locks_spinlock);
326 LDLM_DEBUG(lock, "prolong the busy lock");
327 ldlm_refresh_waiting_lock(lock,
328 ldlm_get_enq_timeout(lock));
329 spin_lock_bh(&waiting_locks_spinlock);
332 LDLM_LOCK_RELEASE(lock);
336 LDLM_LOCK_RELEASE(lock);
339 ldlm_lock_to_ns(lock)->ns_timeouts++;
340 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
341 "evicting client at %s ",
342 cfs_time_current_sec() - lock->l_last_activity,
344 lock->l_export->exp_connection->c_peer.nid));
346 /* no needs to take an extra ref on the lock since it was in
347 * the waiting_locks_list and ldlm_add_waiting_lock()
348 * already grabbed a ref */
349 cfs_list_del(&lock->l_pending_chain);
350 cfs_list_add(&lock->l_pending_chain,
351 &expired_lock_thread.elt_expired_locks);
355 if (!cfs_list_empty(&expired_lock_thread.elt_expired_locks)) {
356 if (obd_dump_on_timeout && need_dump)
357 expired_lock_thread.elt_dump = __LINE__;
359 wake_up(&expired_lock_thread.elt_waitq);
363 * Make sure the timer will fire again if we have any locks
366 if (!cfs_list_empty(&waiting_locks_list)) {
367 cfs_time_t timeout_rounded;
368 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
370 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
371 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
373 spin_unlock_bh(&waiting_locks_spinlock);
377 * Add lock to the list of contended locks.
379 * Indicate that we're waiting for a client to call us back cancelling a given
380 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
381 * timer to fire appropriately. (We round up to the next second, to avoid
382 * floods of timer firings during periods of high lock contention and traffic).
383 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
384 * if it has been added to the waiting list (1 is returned).
386 * Called with the namespace lock held.
388 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
391 cfs_time_t timeout_rounded;
393 if (!cfs_list_empty(&lock->l_pending_chain))
396 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
397 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
400 timeout = cfs_time_shift(seconds);
401 if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
402 lock->l_callback_timeout = timeout;
404 timeout_rounded = round_timeout(lock->l_callback_timeout);
406 if (cfs_time_before(timeout_rounded,
407 cfs_timer_deadline(&waiting_locks_timer)) ||
408 !cfs_timer_is_armed(&waiting_locks_timer)) {
409 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
411 /* if the new lock has a shorter timeout than something earlier on
412 the list, we'll wait the longer amount of time; no big deal. */
414 cfs_list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
418 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
421 int timeout = ldlm_get_enq_timeout(lock);
423 /* NB: must be called with hold of lock_res_and_lock() */
424 LASSERT(lock->l_flags & LDLM_FL_RES_LOCKED);
425 lock->l_flags |= LDLM_FL_WAITED;
427 LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
429 spin_lock_bh(&waiting_locks_spinlock);
430 if (lock->l_flags & LDLM_FL_DESTROYED) {
431 static cfs_time_t next;
432 spin_unlock_bh(&waiting_locks_spinlock);
433 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
434 if (cfs_time_after(cfs_time_current(), next)) {
435 next = cfs_time_shift(14400);
436 libcfs_debug_dumpstack(NULL);
441 ret = __ldlm_add_waiting_lock(lock, timeout);
443 /* grab ref on the lock if it has been added to the
447 spin_unlock_bh(&waiting_locks_spinlock);
450 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
451 if (cfs_list_empty(&lock->l_exp_list))
452 cfs_list_add(&lock->l_exp_list,
453 &lock->l_export->exp_bl_list);
454 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
457 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
458 ret == 0 ? "not re-" : "", timeout,
459 AT_OFF ? "off" : "on");
464 * Remove a lock from the pending list, likely because it had its cancellation
465 * callback arrive without incident. This adjusts the lock-timeout timer if
466 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
467 * As done by ldlm_del_waiting_lock(), the caller must release the lock
468 * reference when the lock is removed from any list (1 is returned).
470 * Called with namespace lock held.
472 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
474 cfs_list_t *list_next;
476 if (cfs_list_empty(&lock->l_pending_chain))
479 list_next = lock->l_pending_chain.next;
480 if (lock->l_pending_chain.prev == &waiting_locks_list) {
481 /* Removing the head of the list, adjust timer. */
482 if (list_next == &waiting_locks_list) {
483 /* No more, just cancel. */
484 cfs_timer_disarm(&waiting_locks_timer);
486 struct ldlm_lock *next;
487 next = cfs_list_entry(list_next, struct ldlm_lock,
489 cfs_timer_arm(&waiting_locks_timer,
490 round_timeout(next->l_callback_timeout));
493 cfs_list_del_init(&lock->l_pending_chain);
498 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
502 if (lock->l_export == NULL) {
503 /* We don't have a "waiting locks list" on clients. */
504 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
508 spin_lock_bh(&waiting_locks_spinlock);
509 ret = __ldlm_del_waiting_lock(lock);
510 spin_unlock_bh(&waiting_locks_spinlock);
512 /* remove the lock out of export blocking list */
513 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
514 cfs_list_del_init(&lock->l_exp_list);
515 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
518 /* release lock ref if it has indeed been removed
520 LDLM_LOCK_RELEASE(lock);
523 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
526 EXPORT_SYMBOL(ldlm_del_waiting_lock);
529 * Prolong the contended lock waiting time.
531 * Called with namespace lock held.
533 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
535 if (lock->l_export == NULL) {
536 /* We don't have a "waiting locks list" on clients. */
537 LDLM_DEBUG(lock, "client lock: no-op");
541 spin_lock_bh(&waiting_locks_spinlock);
543 if (cfs_list_empty(&lock->l_pending_chain)) {
544 spin_unlock_bh(&waiting_locks_spinlock);
545 LDLM_DEBUG(lock, "wasn't waiting");
549 /* we remove/add the lock to the waiting list, so no needs to
550 * release/take a lock reference */
551 __ldlm_del_waiting_lock(lock);
552 __ldlm_add_waiting_lock(lock, timeout);
553 spin_unlock_bh(&waiting_locks_spinlock);
555 LDLM_DEBUG(lock, "refreshed");
558 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
560 #else /* !HAVE_SERVER_SUPPORT || !__KERNEL__ */
562 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
567 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
572 # ifdef HAVE_SERVER_SUPPORT
573 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
575 LASSERT((lock->l_flags & (LDLM_FL_RES_LOCKED|LDLM_FL_CANCEL_ON_BLOCK))
576 == LDLM_FL_RES_LOCKED);
581 #endif /* HAVE_SERVER_SUPPORT && __KERNEL__ */
583 #ifdef HAVE_SERVER_SUPPORT
586 * Perform lock cleanup if AST sending failed.
588 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
589 const char *ast_type)
591 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
592 "to a lock %s callback time out: rc %d\n",
593 lock->l_export->exp_obd->obd_name,
594 obd_export_nid2str(lock->l_export), ast_type, rc);
596 if (obd_dump_on_timeout)
597 libcfs_debug_dumplog();
599 spin_lock_bh(&waiting_locks_spinlock);
600 if (__ldlm_del_waiting_lock(lock) == 0)
601 /* the lock was not in any list, grab an extra ref before adding
602 * the lock to the expired list */
604 cfs_list_add(&lock->l_pending_chain,
605 &expired_lock_thread.elt_expired_locks);
606 wake_up(&expired_lock_thread.elt_waitq);
607 spin_unlock_bh(&waiting_locks_spinlock);
609 class_fail_export(lock->l_export);
614 * Perform lock cleanup if AST reply came with error.
616 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
617 struct ptlrpc_request *req, int rc,
618 const char *ast_type)
620 lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
622 if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
623 LASSERT(lock->l_export);
624 if (lock->l_export->exp_libclient) {
625 LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
626 " timeout, just cancelling lock", ast_type,
627 libcfs_nid2str(peer.nid));
628 ldlm_lock_cancel(lock);
630 } else if (lock->l_flags & LDLM_FL_CANCEL) {
631 LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
632 "cancel was received (AST reply lost?)",
633 ast_type, libcfs_nid2str(peer.nid));
634 ldlm_lock_cancel(lock);
637 ldlm_del_waiting_lock(lock);
638 ldlm_failed_ast(lock, rc, ast_type);
642 struct ldlm_resource *res = lock->l_resource;
643 LDLM_DEBUG(lock, "client (nid %s) returned %d"
644 " from %s AST - normal race",
645 libcfs_nid2str(peer.nid),
647 lustre_msg_get_status(req->rq_repmsg) : -1,
650 /* update lvbo to return proper attributes.
652 ldlm_resource_getref(res);
653 ldlm_res_lvbo_update(res, NULL, 1);
654 ldlm_resource_putref(res);
658 LDLM_ERROR(lock, "client (nid %s) returned %d "
659 "from %s AST", libcfs_nid2str(peer.nid),
660 (req->rq_repmsg != NULL) ?
661 lustre_msg_get_status(req->rq_repmsg) : 0,
664 ldlm_lock_cancel(lock);
665 /* Server-side AST functions are called from ldlm_reprocess_all,
666 * which needs to be told to please restart its reprocessing. */
673 static int ldlm_cb_interpret(const struct lu_env *env,
674 struct ptlrpc_request *req, void *data, int rc)
676 struct ldlm_cb_async_args *ca = data;
677 struct ldlm_lock *lock = ca->ca_lock;
678 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
681 LASSERT(lock != NULL);
684 case LDLM_GL_CALLBACK:
685 /* Update the LVB from disk if the AST failed
686 * (this is a legal race)
688 * - Glimpse callback of local lock just returns
689 * -ELDLM_NO_LOCK_DATA.
690 * - Glimpse callback of remote lock might return
691 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
693 if (rc == -ELDLM_NO_LOCK_DATA) {
694 LDLM_DEBUG(lock, "lost race - client has a lock but no "
696 ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
697 } else if (rc != 0) {
698 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
700 rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
703 case LDLM_BL_CALLBACK:
705 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
707 case LDLM_CP_CALLBACK:
709 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
712 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
717 /* release extra reference taken in ldlm_ast_fini() */
718 LDLM_LOCK_RELEASE(lock);
721 cfs_atomic_inc(&arg->restart);
726 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
727 struct ldlm_cb_set_arg *arg,
728 struct ldlm_lock *lock,
734 if (unlikely(instant_cancel)) {
735 rc = ptl_send_rpc(req, 1);
736 ptlrpc_req_finished(req);
738 cfs_atomic_inc(&arg->restart);
741 ptlrpc_set_add_req(arg->set, req);
748 * Check if there are requests in the export request list which prevent
749 * the lock canceling and make these requests high priority ones.
751 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
753 struct ptlrpc_request *req;
756 if (lock->l_export == NULL) {
757 LDLM_DEBUG(lock, "client lock: no-op");
761 spin_lock_bh(&lock->l_export->exp_rpc_lock);
762 cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
764 /* Do not process requests that were not yet added to there
765 * incoming queue or were already removed from there for
766 * processing. We evaluate ptlrpc_nrs_req_can_move() without
767 * holding svcpt->scp_req_lock, and then redo the check with
768 * the lock held once we need to obtain a reliable result.
770 if (ptlrpc_nrs_req_can_move(req) &&
771 req->rq_ops->hpreq_lock_match &&
772 req->rq_ops->hpreq_lock_match(req, lock))
773 ptlrpc_nrs_req_hp_move(req);
775 spin_unlock_bh(&lock->l_export->exp_rpc_lock);
780 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
781 * enqueued server lock conflicts with given one.
783 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
784 * to wait for client response.
786 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
787 struct ldlm_lock_desc *desc,
788 void *data, int flag)
790 struct ldlm_cb_async_args *ca;
791 struct ldlm_cb_set_arg *arg = data;
792 struct ldlm_request *body;
793 struct ptlrpc_request *req;
794 int instant_cancel = 0;
798 if (flag == LDLM_CB_CANCELING)
799 /* Don't need to do anything here. */
803 LASSERT(data != NULL);
804 if (lock->l_export->exp_obd->obd_recovering != 0)
805 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
807 ldlm_lock_reorder_req(lock);
809 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
810 &RQF_LDLM_BL_CALLBACK,
811 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
815 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
816 ca = ptlrpc_req_async_args(req);
817 ca->ca_set_arg = arg;
820 req->rq_interpret_reply = ldlm_cb_interpret;
821 req->rq_no_resend = 1;
823 lock_res_and_lock(lock);
824 if (lock->l_granted_mode != lock->l_req_mode) {
825 /* this blocking AST will be communicated as part of the
826 * completion AST instead */
827 unlock_res_and_lock(lock);
829 ptlrpc_req_finished(req);
830 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
834 if (lock->l_flags & LDLM_FL_DESTROYED) {
835 /* What's the point? */
836 unlock_res_and_lock(lock);
837 ptlrpc_req_finished(req);
841 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
844 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
845 body->lock_handle[0] = lock->l_remote_handle;
846 body->lock_desc = *desc;
847 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_AST_FLAGS);
849 LDLM_DEBUG(lock, "server preparing blocking AST");
851 ptlrpc_request_set_replen(req);
852 if (instant_cancel) {
853 unlock_res_and_lock(lock);
854 ldlm_lock_cancel(lock);
856 LASSERT(lock->l_granted_mode == lock->l_req_mode);
857 ldlm_add_waiting_lock(lock);
858 unlock_res_and_lock(lock);
861 req->rq_send_state = LUSTRE_IMP_FULL;
862 /* ptlrpc_request_alloc_pack already set timeout */
864 req->rq_timeout = ldlm_get_rq_timeout();
866 lock->l_last_activity = cfs_time_current_sec();
868 if (lock->l_export && lock->l_export->exp_nid_stats &&
869 lock->l_export->exp_nid_stats->nid_ldlm_stats)
870 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
871 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
873 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
877 EXPORT_SYMBOL(ldlm_server_blocking_ast);
880 * ->l_completion_ast callback for a remote lock in server namespace.
882 * Sends AST to the client notifying it of lock granting. If initial
883 * lock response was not sent yet, instead of sending another RPC, just
884 * mark the lock as granted and client will understand
886 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
888 struct ldlm_cb_set_arg *arg = data;
889 struct ldlm_request *body;
890 struct ptlrpc_request *req;
891 struct ldlm_cb_async_args *ca;
892 long total_enqueue_wait;
893 int instant_cancel = 0;
898 LASSERT(lock != NULL);
899 LASSERT(data != NULL);
901 total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(),
902 lock->l_last_activity);
904 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
905 &RQF_LDLM_CP_CALLBACK);
909 /* server namespace, doesn't need lock */
910 lvb_len = ldlm_lvbo_size(lock);
911 /* LU-3124 & LU-2187: to not return layout in completion AST because
912 * it may deadlock for LU-2187, or client may not have enough space
913 * for large layout. The layout will be returned to client with an
914 * extra RPC to fetch xattr.lov */
915 if (ldlm_has_layout(lock))
918 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
919 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
921 ptlrpc_request_free(req);
925 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
926 ca = ptlrpc_req_async_args(req);
927 ca->ca_set_arg = arg;
930 req->rq_interpret_reply = ldlm_cb_interpret;
931 req->rq_no_resend = 1;
932 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
934 body->lock_handle[0] = lock->l_remote_handle;
935 body->lock_flags = ldlm_flags_to_wire(flags);
936 ldlm_lock2desc(lock, &body->lock_desc);
938 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
940 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
942 /* We still need to send the RPC to wake up the blocked
943 * enqueue thread on the client.
945 * Consider old client, there is no better way to notify
946 * the failure, just zero-sized the LVB, then the client
947 * will fail out as "-EPROTO". */
948 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
952 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
957 LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)",
960 lock->l_last_activity = cfs_time_current_sec();
962 /* Server-side enqueue wait time estimate, used in
963 __ldlm_add_waiting_lock to set future enqueue timers */
964 if (total_enqueue_wait < ldlm_get_enq_timeout(lock))
965 at_measured(ldlm_lock_to_ns_at(lock),
968 /* bz18618. Don't add lock enqueue time we spend waiting for a
969 previous callback to fail. Locks waiting legitimately will
970 get extended by ldlm_refresh_waiting_lock regardless of the
971 estimate, so it's okay to underestimate here. */
972 LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. "
973 "It is likely that a previous callback timed out.",
975 at_get(ldlm_lock_to_ns_at(lock)));
977 ptlrpc_request_set_replen(req);
979 req->rq_send_state = LUSTRE_IMP_FULL;
980 /* ptlrpc_request_pack already set timeout */
982 req->rq_timeout = ldlm_get_rq_timeout();
984 /* We only send real blocking ASTs after the lock is granted */
985 lock_res_and_lock(lock);
986 if (lock->l_flags & LDLM_FL_AST_SENT) {
987 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
988 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
989 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
992 /* We might get here prior to ldlm_handle_enqueue setting
993 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
994 * into waiting list, but this is safe and similar code in
995 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
996 * that would not only cancel the lock, but will also remove
997 * it from waiting list */
998 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
999 unlock_res_and_lock(lock);
1000 ldlm_lock_cancel(lock);
1002 lock_res_and_lock(lock);
1004 /* start the lock-timeout clock */
1005 ldlm_add_waiting_lock(lock);
1008 unlock_res_and_lock(lock);
1010 if (lock->l_export && lock->l_export->exp_nid_stats &&
1011 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1012 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1013 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1015 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1017 RETURN(lvb_len < 0 ? lvb_len : rc);
1019 EXPORT_SYMBOL(ldlm_server_completion_ast);
1022 * Server side ->l_glimpse_ast handler for client locks.
1024 * Sends glimpse AST to the client and waits for reply. Then updates
1025 * lvbo with the result.
1027 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1029 struct ldlm_cb_set_arg *arg = data;
1030 struct ldlm_request *body;
1031 struct ptlrpc_request *req;
1032 struct ldlm_cb_async_args *ca;
1034 struct req_format *req_fmt;
1037 LASSERT(lock != NULL);
1039 if (arg->gl_desc != NULL)
1040 /* There is a glimpse descriptor to pack */
1041 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1043 req_fmt = &RQF_LDLM_GL_CALLBACK;
1045 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1046 req_fmt, LUSTRE_DLM_VERSION,
1052 if (arg->gl_desc != NULL) {
1053 /* copy the GL descriptor */
1054 union ldlm_gl_desc *desc;
1055 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1056 *desc = *arg->gl_desc;
1059 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1060 body->lock_handle[0] = lock->l_remote_handle;
1061 ldlm_lock2desc(lock, &body->lock_desc);
1063 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1064 ca = ptlrpc_req_async_args(req);
1065 ca->ca_set_arg = arg;
1068 /* server namespace, doesn't need lock */
1069 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1070 ldlm_lvbo_size(lock));
1071 ptlrpc_request_set_replen(req);
1073 req->rq_send_state = LUSTRE_IMP_FULL;
1074 /* ptlrpc_request_alloc_pack already set timeout */
1076 req->rq_timeout = ldlm_get_rq_timeout();
1078 lock->l_last_activity = cfs_time_current_sec();
1080 req->rq_interpret_reply = ldlm_cb_interpret;
1082 if (lock->l_export && lock->l_export->exp_nid_stats &&
1083 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1084 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1085 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1087 rc = ldlm_ast_fini(req, arg, lock, 0);
1091 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1093 int ldlm_glimpse_locks(struct ldlm_resource *res, cfs_list_t *gl_work_list)
1098 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1100 if (rc == -ERESTART)
1101 ldlm_reprocess_all(res);
1105 EXPORT_SYMBOL(ldlm_glimpse_locks);
1107 /* return LDLM lock associated with a lock callback request */
1108 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1110 struct ldlm_cb_async_args *ca;
1111 struct ldlm_lock *lock;
1114 ca = ptlrpc_req_async_args(req);
1117 RETURN(ERR_PTR(-EFAULT));
1121 EXPORT_SYMBOL(ldlm_request_lock);
1123 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1124 struct lprocfs_stats *srv_stats)
1126 int lock_type = 0, op = 0;
1128 lock_type = dlm_req->lock_desc.l_resource.lr_type;
1130 switch (lock_type) {
1132 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1135 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1136 op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1138 op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1141 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1144 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1152 lprocfs_counter_incr(srv_stats, op);
1158 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1159 * service threads to carry out client lock enqueueing requests.
1161 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1162 struct ptlrpc_request *req,
1163 const struct ldlm_request *dlm_req,
1164 const struct ldlm_callback_suite *cbs)
1166 struct ldlm_reply *dlm_rep;
1168 ldlm_error_t err = ELDLM_OK;
1169 struct ldlm_lock *lock = NULL;
1170 void *cookie = NULL;
1174 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1176 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1177 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1179 LASSERT(req->rq_export);
1181 if (ptlrpc_req2svc(req)->srv_stats != NULL)
1182 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1184 if (req->rq_export && req->rq_export->exp_nid_stats &&
1185 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1186 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1187 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1189 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1190 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1191 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1192 dlm_req->lock_desc.l_resource.lr_type);
1193 GOTO(out, rc = -EFAULT);
1196 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1197 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1198 dlm_req->lock_desc.l_req_mode &
1199 (dlm_req->lock_desc.l_req_mode-1))) {
1200 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1201 dlm_req->lock_desc.l_req_mode);
1202 GOTO(out, rc = -EFAULT);
1205 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) {
1206 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1208 DEBUG_REQ(D_ERROR, req,
1209 "PLAIN lock request from IBITS client?");
1210 GOTO(out, rc = -EPROTO);
1212 } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1214 DEBUG_REQ(D_ERROR, req,
1215 "IBITS lock request from unaware client?");
1216 GOTO(out, rc = -EPROTO);
1220 /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1221 against server's _CONNECT_SUPPORTED flags? (I don't want to use
1222 ibits for mgc/mgs) */
1224 /* INODEBITS_INTEROP: Perform conversion from plain lock to
1225 * inodebits lock if client does not support them. */
1226 if (!(exp_connect_flags(req->rq_export) & OBD_CONNECT_IBITS) &&
1227 (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1228 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1229 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1230 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1231 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1232 dlm_req->lock_desc.l_req_mode = LCK_CR;
1236 if (unlikely(flags & LDLM_FL_REPLAY)) {
1237 /* Find an existing lock in the per-export lock hash */
1238 /* In the function below, .hs_keycmp resolves to
1239 * ldlm_export_lock_keycmp() */
1240 /* coverity[overrun-buffer-val] */
1241 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1242 (void *)&dlm_req->lock_handle[0]);
1244 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1245 LPX64, lock->l_handle.h_cookie);
1246 GOTO(existing_lock, rc = 0);
1250 /* The lock's callback data might be set in the policy function */
1251 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1252 dlm_req->lock_desc.l_resource.lr_type,
1253 dlm_req->lock_desc.l_req_mode,
1254 cbs, NULL, 0, LVB_T_NONE);
1256 GOTO(out, rc = -ENOMEM);
1258 lock->l_last_activity = cfs_time_current_sec();
1259 lock->l_remote_handle = dlm_req->lock_handle[0];
1260 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1262 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1263 /* Don't enqueue a lock onto the export if it is been disonnected
1264 * due to eviction (bug 3822) or server umount (bug 24324).
1265 * Cancel it now instead. */
1266 if (req->rq_export->exp_disconnected) {
1267 LDLM_ERROR(lock, "lock on disconnected export %p",
1269 GOTO(out, rc = -ENOTCONN);
1272 lock->l_export = class_export_lock_get(req->rq_export, lock);
1273 if (lock->l_export->exp_lock_hash)
1274 cfs_hash_add(lock->l_export->exp_lock_hash,
1275 &lock->l_remote_handle,
1280 if (flags & LDLM_FL_HAS_INTENT) {
1281 /* In this case, the reply buffer is allocated deep in
1282 * local_lock_enqueue by the policy function. */
1285 /* based on the assumption that lvb size never changes during
1286 * resource life time otherwise it need resource->lr_lock's
1288 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1289 RCL_SERVER, ldlm_lvbo_size(lock));
1291 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1292 GOTO(out, rc = -ENOMEM);
1294 rc = req_capsule_server_pack(&req->rq_pill);
1299 if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1300 ldlm_convert_policy_to_local(req->rq_export,
1301 dlm_req->lock_desc.l_resource.lr_type,
1302 &dlm_req->lock_desc.l_policy_data,
1303 &lock->l_policy_data);
1304 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1305 lock->l_req_extent = lock->l_policy_data.l_extent;
1307 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1314 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1315 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1317 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1318 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1320 /* We never send a blocking AST until the lock is granted, but
1321 * we can tell it right now */
1322 lock_res_and_lock(lock);
1324 /* Now take into account flags to be inherited from original lock
1325 request both in reply to client and in our own lock flags. */
1326 dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1327 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1328 LDLM_INHERIT_FLAGS);
1330 /* Don't move a pending lock onto the export if it has already been
1331 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1332 * Cancel it now instead. */
1333 if (unlikely(req->rq_export->exp_disconnected ||
1334 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1335 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1337 } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1338 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1339 if (lock->l_granted_mode == lock->l_req_mode) {
1341 * Only cancel lock if it was granted, because it would
1342 * be destroyed immediately and would never be granted
1343 * in the future, causing timeouts on client. Not
1344 * granted lock will be cancelled immediately after
1345 * sending completion AST.
1347 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1348 unlock_res_and_lock(lock);
1349 ldlm_lock_cancel(lock);
1350 lock_res_and_lock(lock);
1352 ldlm_add_waiting_lock(lock);
1355 /* Make sure we never ever grant usual metadata locks to liblustre
1357 if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1358 dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1359 req->rq_export->exp_libclient) {
1360 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1361 !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1362 CERROR("Granting sync lock to libclient. "
1363 "req fl %d, rep fl %d, lock fl "LPX64"\n",
1364 dlm_req->lock_flags, dlm_rep->lock_flags,
1366 LDLM_ERROR(lock, "sync lock");
1367 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1368 struct ldlm_intent *it;
1370 it = req_capsule_client_get(&req->rq_pill,
1373 CERROR("This is intent %s ("LPU64")\n",
1374 ldlm_it2str(it->opc), it->opc);
1380 unlock_res_and_lock(lock);
1384 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1385 if (!req->rq_packed_final) {
1386 err = lustre_pack_reply(req, 1, NULL, NULL);
1391 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1392 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1394 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1395 "(err=%d, rc=%d)", err, rc);
1398 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1400 ldlm_lvbo_size(lock) > 0) {
1404 buf = req_capsule_server_get(&req->rq_pill,
1406 LASSERTF(buf != NULL, "req %p, lock %p\n",
1408 buflen = req_capsule_get_size(&req->rq_pill,
1409 &RMF_DLM_LVB, RCL_SERVER);
1411 buflen = ldlm_lvbo_fill(lock, buf,
1417 buflen, RCL_SERVER);
1425 lock_res_and_lock(lock);
1426 ldlm_resource_unlink_lock(lock);
1427 ldlm_lock_destroy_nolock(lock);
1428 unlock_res_and_lock(lock);
1431 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1432 ldlm_reprocess_all(lock->l_resource);
1434 LDLM_LOCK_RELEASE(lock);
1437 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1442 EXPORT_SYMBOL(ldlm_handle_enqueue0);
1445 * Old-style LDLM main entry point for server code enqueue.
1447 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1448 ldlm_completion_callback completion_callback,
1449 ldlm_blocking_callback blocking_callback,
1450 ldlm_glimpse_callback glimpse_callback)
1452 struct ldlm_request *dlm_req;
1453 struct ldlm_callback_suite cbs = {
1454 .lcs_completion = completion_callback,
1455 .lcs_blocking = blocking_callback,
1456 .lcs_glimpse = glimpse_callback
1460 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1461 if (dlm_req != NULL) {
1462 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1463 req, dlm_req, &cbs);
1469 EXPORT_SYMBOL(ldlm_handle_enqueue);
1472 * Main LDLM entry point for server code to process lock conversion requests.
1474 int ldlm_handle_convert0(struct ptlrpc_request *req,
1475 const struct ldlm_request *dlm_req)
1477 struct ldlm_reply *dlm_rep;
1478 struct ldlm_lock *lock;
1482 if (req->rq_export && req->rq_export->exp_nid_stats &&
1483 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1484 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1485 LDLM_CONVERT - LDLM_FIRST_OPC);
1487 rc = req_capsule_server_pack(&req->rq_pill);
1491 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1492 dlm_rep->lock_flags = dlm_req->lock_flags;
1494 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1496 req->rq_status = LUSTRE_EINVAL;
1500 LDLM_DEBUG(lock, "server-side convert handler START");
1502 lock->l_last_activity = cfs_time_current_sec();
1503 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1504 &dlm_rep->lock_flags);
1506 if (ldlm_del_waiting_lock(lock))
1507 LDLM_DEBUG(lock, "converted waiting lock");
1510 req->rq_status = LUSTRE_EDEADLK;
1515 if (!req->rq_status)
1516 ldlm_reprocess_all(lock->l_resource);
1517 LDLM_DEBUG(lock, "server-side convert handler END");
1518 LDLM_LOCK_PUT(lock);
1520 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1524 EXPORT_SYMBOL(ldlm_handle_convert0);
1527 * Old-style main LDLM entry point for server code to process lock conversion
1530 int ldlm_handle_convert(struct ptlrpc_request *req)
1533 struct ldlm_request *dlm_req;
1535 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1536 if (dlm_req != NULL) {
1537 rc = ldlm_handle_convert0(req, dlm_req);
1539 CERROR ("Can't unpack dlm_req\n");
1544 EXPORT_SYMBOL(ldlm_handle_convert);
1547 * Cancel all the locks whose handles are packed into ldlm_request
1549 * Called by server code expecting such combined cancel activity
1552 int ldlm_request_cancel(struct ptlrpc_request *req,
1553 const struct ldlm_request *dlm_req, int first)
1555 struct ldlm_resource *res, *pres = NULL;
1556 struct ldlm_lock *lock;
1557 int i, count, done = 0;
1560 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1564 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1567 /* There is no lock on the server at the replay time,
1568 * skip lock cancelling to make replay tests to pass. */
1569 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1572 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1573 "starting at %d", count, first);
1575 for (i = first; i < count; i++) {
1576 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1578 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1579 "lock (cookie "LPU64")",
1580 dlm_req->lock_handle[i].cookie);
1584 res = lock->l_resource;
1587 /* This code is an optimization to only attempt lock
1588 * granting on the resource (that could be CPU-expensive)
1589 * after we are done cancelling lock in that resource. */
1592 ldlm_reprocess_all(pres);
1593 LDLM_RESOURCE_DELREF(pres);
1594 ldlm_resource_putref(pres);
1597 ldlm_resource_getref(res);
1598 LDLM_RESOURCE_ADDREF(res);
1599 ldlm_res_lvbo_update(res, NULL, 1);
1603 ldlm_lock_cancel(lock);
1604 LDLM_LOCK_PUT(lock);
1607 ldlm_reprocess_all(pres);
1608 LDLM_RESOURCE_DELREF(pres);
1609 ldlm_resource_putref(pres);
1611 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1614 EXPORT_SYMBOL(ldlm_request_cancel);
1617 * Main LDLM entry point for server code to cancel locks.
1619 * Typically gets called from service handler on LDLM_CANCEL opc.
1621 int ldlm_handle_cancel(struct ptlrpc_request *req)
1623 struct ldlm_request *dlm_req;
1627 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1628 if (dlm_req == NULL) {
1629 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1633 if (req->rq_export && req->rq_export->exp_nid_stats &&
1634 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1635 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1636 LDLM_CANCEL - LDLM_FIRST_OPC);
1638 rc = req_capsule_server_pack(&req->rq_pill);
1642 if (!ldlm_request_cancel(req, dlm_req, 0))
1643 req->rq_status = LUSTRE_ESTALE;
1645 RETURN(ptlrpc_reply(req));
1647 EXPORT_SYMBOL(ldlm_handle_cancel);
1648 #endif /* HAVE_SERVER_SUPPORT */
1651 * Callback handler for receiving incoming blocking ASTs.
1653 * This can only happen on client side.
1655 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1656 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1661 LDLM_DEBUG(lock, "client blocking AST callback handler");
1663 lock_res_and_lock(lock);
1664 lock->l_flags |= LDLM_FL_CBPENDING;
1666 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1667 lock->l_flags |= LDLM_FL_CANCEL;
1669 do_ast = (!lock->l_readers && !lock->l_writers);
1670 unlock_res_and_lock(lock);
1673 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1674 lock, lock->l_blocking_ast);
1675 if (lock->l_blocking_ast != NULL)
1676 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1679 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1683 LDLM_DEBUG(lock, "client blocking callback handler END");
1684 LDLM_LOCK_RELEASE(lock);
1689 * Callback handler for receiving incoming completion ASTs.
1691 * This only can happen on client side.
1693 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1694 struct ldlm_namespace *ns,
1695 struct ldlm_request *dlm_req,
1696 struct ldlm_lock *lock)
1699 CFS_LIST_HEAD(ast_list);
1703 LDLM_DEBUG(lock, "client completion callback handler START");
1705 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1706 int to = cfs_time_seconds(1);
1708 schedule_timeout_and_set_state(
1709 TASK_INTERRUPTIBLE, to);
1710 if (lock->l_granted_mode == lock->l_req_mode ||
1711 lock->l_flags & LDLM_FL_DESTROYED)
1716 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1718 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1719 GOTO(out, rc = lvb_len);
1720 } else if (lvb_len > 0) {
1721 if (lock->l_lvb_len > 0) {
1722 /* for extent lock, lvb contains ost_lvb{}. */
1723 LASSERT(lock->l_lvb_data != NULL);
1725 if (unlikely(lock->l_lvb_len < lvb_len)) {
1726 LDLM_ERROR(lock, "Replied LVB is larger than "
1727 "expectation, expected = %d, "
1729 lock->l_lvb_len, lvb_len);
1730 GOTO(out, rc = -EINVAL);
1732 } else if (ldlm_has_layout(lock)) { /* for layout lock, lvb has
1733 * variable length */
1736 OBD_ALLOC_LARGE(lvb_data, lvb_len);
1737 if (lvb_data == NULL) {
1738 LDLM_ERROR(lock, "No memory: %d.\n", lvb_len);
1739 GOTO(out, rc = -ENOMEM);
1742 lock_res_and_lock(lock);
1743 LASSERT(lock->l_lvb_data == NULL);
1744 lock->l_lvb_type = LVB_T_LAYOUT;
1745 lock->l_lvb_data = lvb_data;
1746 lock->l_lvb_len = lvb_len;
1747 unlock_res_and_lock(lock);
1751 lock_res_and_lock(lock);
1752 if ((lock->l_flags & LDLM_FL_DESTROYED) ||
1753 lock->l_granted_mode == lock->l_req_mode) {
1754 /* bug 11300: the lock has already been granted */
1755 unlock_res_and_lock(lock);
1756 LDLM_DEBUG(lock, "Double grant race happened");
1760 /* If we receive the completion AST before the actual enqueue returned,
1761 * then we might need to switch lock modes, resources, or extents. */
1762 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1763 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1764 LDLM_DEBUG(lock, "completion AST, new lock mode");
1767 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1768 ldlm_convert_policy_to_local(req->rq_export,
1769 dlm_req->lock_desc.l_resource.lr_type,
1770 &dlm_req->lock_desc.l_policy_data,
1771 &lock->l_policy_data);
1772 LDLM_DEBUG(lock, "completion AST, new policy data");
1775 ldlm_resource_unlink_lock(lock);
1776 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1777 &lock->l_resource->lr_name,
1778 sizeof(lock->l_resource->lr_name)) != 0) {
1779 unlock_res_and_lock(lock);
1780 rc = ldlm_lock_change_resource(ns, lock,
1781 &dlm_req->lock_desc.l_resource.lr_name);
1783 LDLM_ERROR(lock, "Failed to allocate resource");
1786 LDLM_DEBUG(lock, "completion AST, new resource");
1787 CERROR("change resource!\n");
1788 lock_res_and_lock(lock);
1791 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1792 /* BL_AST locks are not needed in LRU.
1793 * Let ldlm_cancel_lru() be fast. */
1794 ldlm_lock_remove_from_lru(lock);
1795 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1796 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1799 if (lock->l_lvb_len > 0) {
1800 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1801 lock->l_lvb_data, lvb_len);
1803 unlock_res_and_lock(lock);
1808 ldlm_grant_lock(lock, &ast_list);
1809 unlock_res_and_lock(lock);
1811 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1813 /* Let Enqueue to call osc_lock_upcall() and initialize
1815 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1817 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1819 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1825 lock_res_and_lock(lock);
1826 lock->l_flags |= LDLM_FL_FAILED;
1827 unlock_res_and_lock(lock);
1828 wake_up(&lock->l_waitq);
1830 LDLM_LOCK_RELEASE(lock);
1834 * Callback handler for receiving incoming glimpse ASTs.
1836 * This only can happen on client side. After handling the glimpse AST
1837 * we also consider dropping the lock here if it is unused locally for a
1840 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1841 struct ldlm_namespace *ns,
1842 struct ldlm_request *dlm_req,
1843 struct ldlm_lock *lock)
1848 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1850 if (lock->l_glimpse_ast != NULL)
1851 rc = lock->l_glimpse_ast(lock, req);
1853 if (req->rq_repmsg != NULL) {
1856 req->rq_status = rc;
1860 lock_res_and_lock(lock);
1861 if (lock->l_granted_mode == LCK_PW &&
1862 !lock->l_readers && !lock->l_writers &&
1863 cfs_time_after(cfs_time_current(),
1864 cfs_time_add(lock->l_last_used,
1865 cfs_time_seconds(10)))) {
1866 unlock_res_and_lock(lock);
1867 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1868 ldlm_handle_bl_callback(ns, NULL, lock);
1873 unlock_res_and_lock(lock);
1874 LDLM_LOCK_RELEASE(lock);
1878 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1880 if (req->rq_no_reply)
1883 req->rq_status = rc;
1884 if (!req->rq_packed_final) {
1885 rc = lustre_pack_reply(req, 1, NULL, NULL);
1889 return ptlrpc_reply(req);
1893 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1894 ldlm_cancel_flags_t cancel_flags)
1896 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1899 spin_lock(&blp->blp_lock);
1900 if (blwi->blwi_lock &&
1901 blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) {
1902 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1903 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1905 /* other blocking callbacks are added to the regular list */
1906 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1908 spin_unlock(&blp->blp_lock);
1910 wake_up(&blp->blp_waitq);
1912 /* can not check blwi->blwi_flags as blwi could be already freed in
1914 if (!(cancel_flags & LCF_ASYNC))
1915 wait_for_completion(&blwi->blwi_comp);
1920 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1921 struct ldlm_namespace *ns,
1922 struct ldlm_lock_desc *ld,
1923 cfs_list_t *cancels, int count,
1924 struct ldlm_lock *lock,
1925 ldlm_cancel_flags_t cancel_flags)
1927 init_completion(&blwi->blwi_comp);
1928 CFS_INIT_LIST_HEAD(&blwi->blwi_head);
1930 if (memory_pressure_get())
1931 blwi->blwi_mem_pressure = 1;
1934 blwi->blwi_flags = cancel_flags;
1936 blwi->blwi_ld = *ld;
1938 cfs_list_add(&blwi->blwi_head, cancels);
1939 cfs_list_del_init(cancels);
1940 blwi->blwi_count = count;
1942 blwi->blwi_lock = lock;
1947 * Queues a list of locks \a cancels containing \a count locks
1948 * for later processing by a blocking thread. If \a count is zero,
1949 * then the lock referenced as \a lock is queued instead.
1951 * The blocking thread would then call ->l_blocking_ast callback in the lock.
1952 * If list addition fails an error is returned and caller is supposed to
1953 * call ->l_blocking_ast itself.
1955 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1956 struct ldlm_lock_desc *ld,
1957 struct ldlm_lock *lock,
1958 cfs_list_t *cancels, int count,
1959 ldlm_cancel_flags_t cancel_flags)
1963 if (cancels && count == 0)
1966 if (cancel_flags & LCF_ASYNC) {
1967 struct ldlm_bl_work_item *blwi;
1969 OBD_ALLOC(blwi, sizeof(*blwi));
1972 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
1974 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
1976 /* if it is synchronous call do minimum mem alloc, as it could
1977 * be triggered from kernel shrinker
1979 struct ldlm_bl_work_item blwi;
1981 memset(&blwi, 0, sizeof(blwi));
1982 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
1983 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
1989 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1990 struct ldlm_lock *lock)
1993 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
1999 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2000 cfs_list_t *cancels, int count,
2001 ldlm_cancel_flags_t cancel_flags)
2004 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2010 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2011 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2013 struct obd_device *obd = req->rq_export->exp_obd;
2020 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2022 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2024 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2026 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2029 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2031 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2033 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2036 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2039 /* We are responsible for swabbing contents of val */
2041 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2042 /* Pass it on to mdc (the "export" in this case) */
2043 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2045 sizeof(KEY_HSM_COPYTOOL_SEND),
2046 KEY_HSM_COPYTOOL_SEND,
2049 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2054 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2055 const char *msg, int rc,
2056 struct lustre_handle *handle)
2058 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2059 "%s: [nid %s] [rc %d] [lock "LPX64"]",
2060 msg, libcfs_id2str(req->rq_peer), rc,
2061 handle ? handle->cookie : 0);
2062 if (req->rq_no_reply)
2063 CWARN("No reply was sent, maybe cause bug 21636.\n");
2065 CWARN("Send reply failed, maybe cause bug 21636.\n");
2068 static int ldlm_handle_qc_callback(struct ptlrpc_request *req)
2070 struct obd_quotactl *oqctl;
2071 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2073 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2074 if (oqctl == NULL) {
2075 CERROR("Can't unpack obd_quotactl\n");
2079 oqctl->qc_stat = ptlrpc_status_ntoh(oqctl->qc_stat);
2081 cli->cl_qchk_stat = oqctl->qc_stat;
2085 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2086 static int ldlm_callback_handler(struct ptlrpc_request *req)
2088 struct ldlm_namespace *ns;
2089 struct ldlm_request *dlm_req;
2090 struct ldlm_lock *lock;
2094 /* Requests arrive in sender's byte order. The ptlrpc service
2095 * handler has already checked and, if necessary, byte-swapped the
2096 * incoming request message body, but I am responsible for the
2097 * message buffers. */
2099 /* do nothing for sec context finalize */
2100 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2103 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2105 if (req->rq_export == NULL) {
2106 rc = ldlm_callback_reply(req, -ENOTCONN);
2107 ldlm_callback_errmsg(req, "Operate on unconnected server",
2112 LASSERT(req->rq_export != NULL);
2113 LASSERT(req->rq_export->exp_obd != NULL);
2115 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2116 case LDLM_BL_CALLBACK:
2117 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET))
2120 case LDLM_CP_CALLBACK:
2121 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2124 case LDLM_GL_CALLBACK:
2125 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2129 rc = ldlm_handle_setinfo(req);
2130 ldlm_callback_reply(req, rc);
2132 case LLOG_ORIGIN_HANDLE_CREATE:
2133 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2134 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2136 rc = llog_origin_handle_open(req);
2137 ldlm_callback_reply(req, rc);
2139 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2140 req_capsule_set(&req->rq_pill,
2141 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2142 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2144 rc = llog_origin_handle_next_block(req);
2145 ldlm_callback_reply(req, rc);
2147 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2148 req_capsule_set(&req->rq_pill,
2149 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2150 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2152 rc = llog_origin_handle_read_header(req);
2153 ldlm_callback_reply(req, rc);
2155 case LLOG_ORIGIN_HANDLE_CLOSE:
2156 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2158 rc = llog_origin_handle_close(req);
2159 ldlm_callback_reply(req, rc);
2161 case OBD_QC_CALLBACK:
2162 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
2163 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
2165 rc = ldlm_handle_qc_callback(req);
2166 ldlm_callback_reply(req, rc);
2169 CERROR("unknown opcode %u\n",
2170 lustre_msg_get_opc(req->rq_reqmsg));
2171 ldlm_callback_reply(req, -EPROTO);
2175 ns = req->rq_export->exp_obd->obd_namespace;
2176 LASSERT(ns != NULL);
2178 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2180 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2181 if (dlm_req == NULL) {
2182 rc = ldlm_callback_reply(req, -EPROTO);
2183 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2188 /* Force a known safe race, send a cancel to the server for a lock
2189 * which the server has already started a blocking callback on. */
2190 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2191 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2192 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2194 CERROR("ldlm_cli_cancel: %d\n", rc);
2197 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2199 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2200 "disappeared\n", dlm_req->lock_handle[0].cookie);
2201 rc = ldlm_callback_reply(req, -EINVAL);
2202 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2203 &dlm_req->lock_handle[0]);
2207 if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
2208 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2209 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2211 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2212 lock_res_and_lock(lock);
2213 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2215 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2216 /* If somebody cancels lock and cache is already dropped,
2217 * or lock is failed before cp_ast received on client,
2218 * we can tell the server we have no lock. Otherwise, we
2219 * should send cancel after dropping the cache. */
2220 if (((lock->l_flags & LDLM_FL_CANCELING) &&
2221 (lock->l_flags & LDLM_FL_BL_DONE)) ||
2222 (lock->l_flags & LDLM_FL_FAILED)) {
2223 LDLM_DEBUG(lock, "callback on lock "
2224 LPX64" - lock disappeared\n",
2225 dlm_req->lock_handle[0].cookie);
2226 unlock_res_and_lock(lock);
2227 LDLM_LOCK_RELEASE(lock);
2228 rc = ldlm_callback_reply(req, -EINVAL);
2229 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2230 &dlm_req->lock_handle[0]);
2233 /* BL_AST locks are not needed in LRU.
2234 * Let ldlm_cancel_lru() be fast. */
2235 ldlm_lock_remove_from_lru(lock);
2236 lock->l_flags |= LDLM_FL_BL_AST;
2238 unlock_res_and_lock(lock);
2240 /* We want the ost thread to get this reply so that it can respond
2241 * to ost requests (write cache writeback) that might be triggered
2244 * But we'd also like to be able to indicate in the reply that we're
2245 * cancelling right now, because it's unused, or have an intent result
2246 * in the reply, so we might have to push the responsibility for sending
2247 * the reply down into the AST handlers, alas. */
2249 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2250 case LDLM_BL_CALLBACK:
2251 CDEBUG(D_INODE, "blocking ast\n");
2252 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2253 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) {
2254 rc = ldlm_callback_reply(req, 0);
2255 if (req->rq_no_reply || rc)
2256 ldlm_callback_errmsg(req, "Normal process", rc,
2257 &dlm_req->lock_handle[0]);
2259 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2260 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2262 case LDLM_CP_CALLBACK:
2263 CDEBUG(D_INODE, "completion ast\n");
2264 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2265 ldlm_callback_reply(req, 0);
2266 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2268 case LDLM_GL_CALLBACK:
2269 CDEBUG(D_INODE, "glimpse ast\n");
2270 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2271 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2274 LBUG(); /* checked above */
2280 #ifdef HAVE_SERVER_SUPPORT
2282 * Main handler for canceld thread.
2284 * Separated into its own thread to avoid deadlocks.
2286 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2291 /* Requests arrive in sender's byte order. The ptlrpc service
2292 * handler has already checked and, if necessary, byte-swapped the
2293 * incoming request message body, but I am responsible for the
2294 * message buffers. */
2296 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2298 if (req->rq_export == NULL) {
2299 struct ldlm_request *dlm_req;
2301 CERROR("%s from %s arrived at %lu with bad export cookie "
2303 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2304 libcfs_nid2str(req->rq_peer.nid),
2305 req->rq_arrival_time.tv_sec,
2306 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2308 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2309 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2310 dlm_req = req_capsule_client_get(&req->rq_pill,
2312 if (dlm_req != NULL)
2313 ldlm_lock_dump_handle(D_ERROR,
2314 &dlm_req->lock_handle[0]);
2316 ldlm_callback_reply(req, -ENOTCONN);
2320 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2322 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2324 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2325 CDEBUG(D_INODE, "cancel\n");
2326 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2327 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))
2329 rc = ldlm_handle_cancel(req);
2334 CERROR("invalid opcode %d\n",
2335 lustre_msg_get_opc(req->rq_reqmsg));
2336 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2337 ldlm_callback_reply(req, -EINVAL);
2343 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2344 struct ldlm_lock *lock)
2346 struct ldlm_request *dlm_req;
2347 struct lustre_handle lockh;
2352 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2353 if (dlm_req == NULL)
2356 ldlm_lock2handle(lock, &lockh);
2357 for (i = 0; i < dlm_req->lock_count; i++) {
2358 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2360 DEBUG_REQ(D_RPCTRACE, req,
2361 "Prio raised by lock "LPX64".", lockh.cookie);
2372 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2374 struct ldlm_request *dlm_req;
2379 /* no prolong in recovery */
2380 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2383 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2384 if (dlm_req == NULL)
2387 for (i = 0; i < dlm_req->lock_count; i++) {
2388 struct ldlm_lock *lock;
2390 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2394 rc = !!(lock->l_flags & LDLM_FL_AST_SENT);
2396 LDLM_DEBUG(lock, "hpreq cancel lock");
2397 LDLM_LOCK_PUT(lock);
2406 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2407 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2408 .hpreq_check = ldlm_cancel_hpreq_check,
2412 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2416 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2418 if (req->rq_export == NULL)
2421 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2422 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2423 req->rq_ops = &ldlm_cancel_hpreq_ops;
2428 int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2429 cfs_hlist_node_t *hnode, void *data)
2432 cfs_list_t *rpc_list = data;
2433 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2435 lock_res_and_lock(lock);
2437 if (lock->l_req_mode != lock->l_granted_mode) {
2438 unlock_res_and_lock(lock);
2442 LASSERT(lock->l_resource);
2443 if (lock->l_resource->lr_type != LDLM_IBITS &&
2444 lock->l_resource->lr_type != LDLM_PLAIN) {
2445 unlock_res_and_lock(lock);
2449 if (lock->l_flags & LDLM_FL_AST_SENT) {
2450 unlock_res_and_lock(lock);
2454 LASSERT(lock->l_blocking_ast);
2455 LASSERT(!lock->l_blocking_lock);
2457 lock->l_flags |= LDLM_FL_AST_SENT;
2458 if (lock->l_export && lock->l_export->exp_lock_hash) {
2459 /* NB: it's safe to call cfs_hash_del() even lock isn't
2460 * in exp_lock_hash. */
2461 /* In the function below, .hs_keycmp resolves to
2462 * ldlm_export_lock_keycmp() */
2463 /* coverity[overrun-buffer-val] */
2464 cfs_hash_del(lock->l_export->exp_lock_hash,
2465 &lock->l_remote_handle, &lock->l_exp_hash);
2468 cfs_list_add_tail(&lock->l_rk_ast, rpc_list);
2469 LDLM_LOCK_GET(lock);
2471 unlock_res_and_lock(lock);
2475 void ldlm_revoke_export_locks(struct obd_export *exp)
2477 cfs_list_t rpc_list;
2480 CFS_INIT_LIST_HEAD(&rpc_list);
2481 cfs_hash_for_each_empty(exp->exp_lock_hash,
2482 ldlm_revoke_lock_cb, &rpc_list);
2483 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2484 LDLM_WORK_REVOKE_AST);
2488 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2489 #endif /* HAVE_SERVER_SUPPORT */
2492 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2494 struct ldlm_bl_work_item *blwi = NULL;
2495 static unsigned int num_bl = 0;
2497 spin_lock(&blp->blp_lock);
2498 /* process a request from the blp_list at least every blp_num_threads */
2499 if (!cfs_list_empty(&blp->blp_list) &&
2500 (cfs_list_empty(&blp->blp_prio_list) || num_bl == 0))
2501 blwi = cfs_list_entry(blp->blp_list.next,
2502 struct ldlm_bl_work_item, blwi_entry);
2504 if (!cfs_list_empty(&blp->blp_prio_list))
2505 blwi = cfs_list_entry(blp->blp_prio_list.next,
2506 struct ldlm_bl_work_item,
2510 if (++num_bl >= cfs_atomic_read(&blp->blp_num_threads))
2512 cfs_list_del(&blwi->blwi_entry);
2514 spin_unlock(&blp->blp_lock);
2519 /* This only contains temporary data until the thread starts */
2520 struct ldlm_bl_thread_data {
2521 char bltd_name[CFS_CURPROC_COMM_MAX];
2522 struct ldlm_bl_pool *bltd_blp;
2523 struct completion bltd_comp;
2527 static int ldlm_bl_thread_main(void *arg);
2529 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2531 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2532 struct task_struct *task;
2534 init_completion(&bltd.bltd_comp);
2535 bltd.bltd_num = cfs_atomic_read(&blp->blp_num_threads);
2536 snprintf(bltd.bltd_name, sizeof(bltd.bltd_name) - 1,
2537 "ldlm_bl_%02d", bltd.bltd_num);
2538 task = kthread_run(ldlm_bl_thread_main, &bltd, bltd.bltd_name);
2540 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2541 cfs_atomic_read(&blp->blp_num_threads), PTR_ERR(task));
2542 return PTR_ERR(task);
2544 wait_for_completion(&bltd.bltd_comp);
2550 * Main blocking requests processing thread.
2552 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2553 * This thread in the end ends up doing actual call to ->l_blocking_ast
2556 static int ldlm_bl_thread_main(void *arg)
2558 struct ldlm_bl_pool *blp;
2562 struct ldlm_bl_thread_data *bltd = arg;
2564 blp = bltd->bltd_blp;
2566 cfs_atomic_inc(&blp->blp_num_threads);
2567 cfs_atomic_inc(&blp->blp_busy_threads);
2569 complete(&bltd->bltd_comp);
2570 /* cannot use bltd after this, it is only on caller's stack */
2574 struct l_wait_info lwi = { 0 };
2575 struct ldlm_bl_work_item *blwi = NULL;
2578 blwi = ldlm_bl_get_work(blp);
2581 cfs_atomic_dec(&blp->blp_busy_threads);
2582 l_wait_event_exclusive(blp->blp_waitq,
2583 (blwi = ldlm_bl_get_work(blp)) != NULL,
2585 busy = cfs_atomic_inc_return(&blp->blp_busy_threads);
2587 busy = cfs_atomic_read(&blp->blp_busy_threads);
2590 if (blwi->blwi_ns == NULL)
2591 /* added by ldlm_cleanup() */
2594 /* Not fatal if racy and have a few too many threads */
2595 if (unlikely(busy < blp->blp_max_threads &&
2596 busy >= cfs_atomic_read(&blp->blp_num_threads) &&
2597 !blwi->blwi_mem_pressure))
2598 /* discard the return value, we tried */
2599 ldlm_bl_thread_start(blp);
2601 if (blwi->blwi_mem_pressure)
2602 memory_pressure_set();
2604 if (blwi->blwi_count) {
2606 /* The special case when we cancel locks in LRU
2607 * asynchronously, we pass the list of locks here.
2608 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2609 * canceled locally yet. */
2610 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2613 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2616 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2619 if (blwi->blwi_mem_pressure)
2620 memory_pressure_clr();
2622 if (blwi->blwi_flags & LCF_ASYNC)
2623 OBD_FREE(blwi, sizeof(*blwi));
2625 complete(&blwi->blwi_comp);
2628 cfs_atomic_dec(&blp->blp_busy_threads);
2629 cfs_atomic_dec(&blp->blp_num_threads);
2630 complete(&blp->blp_comp);
2636 static int ldlm_setup(void);
2637 static int ldlm_cleanup(void);
2639 int ldlm_get_ref(void)
2643 mutex_lock(&ldlm_ref_mutex);
2644 if (++ldlm_refcount == 1) {
2649 mutex_unlock(&ldlm_ref_mutex);
2653 EXPORT_SYMBOL(ldlm_get_ref);
2655 void ldlm_put_ref(void)
2658 mutex_lock(&ldlm_ref_mutex);
2659 if (ldlm_refcount == 1) {
2660 int rc = ldlm_cleanup();
2662 CERROR("ldlm_cleanup failed: %d\n", rc);
2668 mutex_unlock(&ldlm_ref_mutex);
2672 EXPORT_SYMBOL(ldlm_put_ref);
2675 * Export handle<->lock hash operations.
2678 ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
2680 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2684 ldlm_export_lock_key(cfs_hlist_node_t *hnode)
2686 struct ldlm_lock *lock;
2688 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2689 return &lock->l_remote_handle;
2693 ldlm_export_lock_keycpy(cfs_hlist_node_t *hnode, void *key)
2695 struct ldlm_lock *lock;
2697 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2698 lock->l_remote_handle = *(struct lustre_handle *)key;
2702 ldlm_export_lock_keycmp(const void *key, cfs_hlist_node_t *hnode)
2704 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2708 ldlm_export_lock_object(cfs_hlist_node_t *hnode)
2710 return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2714 ldlm_export_lock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2716 struct ldlm_lock *lock;
2718 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2719 LDLM_LOCK_GET(lock);
2723 ldlm_export_lock_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2725 struct ldlm_lock *lock;
2727 lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2728 LDLM_LOCK_RELEASE(lock);
2731 static cfs_hash_ops_t ldlm_export_lock_ops = {
2732 .hs_hash = ldlm_export_lock_hash,
2733 .hs_key = ldlm_export_lock_key,
2734 .hs_keycmp = ldlm_export_lock_keycmp,
2735 .hs_keycpy = ldlm_export_lock_keycpy,
2736 .hs_object = ldlm_export_lock_object,
2737 .hs_get = ldlm_export_lock_get,
2738 .hs_put = ldlm_export_lock_put,
2739 .hs_put_locked = ldlm_export_lock_put,
2742 int ldlm_init_export(struct obd_export *exp)
2747 exp->exp_lock_hash =
2748 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2749 HASH_EXP_LOCK_CUR_BITS,
2750 HASH_EXP_LOCK_MAX_BITS,
2751 HASH_EXP_LOCK_BKT_BITS, 0,
2752 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2753 &ldlm_export_lock_ops,
2754 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2755 CFS_HASH_NBLK_CHANGE);
2757 if (!exp->exp_lock_hash)
2760 rc = ldlm_init_flock_export(exp);
2766 ldlm_destroy_export(exp);
2769 EXPORT_SYMBOL(ldlm_init_export);
2771 void ldlm_destroy_export(struct obd_export *exp)
2774 cfs_hash_putref(exp->exp_lock_hash);
2775 exp->exp_lock_hash = NULL;
2777 ldlm_destroy_flock_export(exp);
2780 EXPORT_SYMBOL(ldlm_destroy_export);
2782 static int ldlm_setup(void)
2784 static struct ptlrpc_service_conf conf;
2785 struct ldlm_bl_pool *blp = NULL;
2787 # ifdef HAVE_SERVER_SUPPORT
2788 struct task_struct *task;
2796 if (ldlm_state != NULL)
2799 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2800 if (ldlm_state == NULL)
2804 rc = ldlm_proc_setup();
2809 memset(&conf, 0, sizeof(conf));
2810 conf = (typeof(conf)) {
2811 .psc_name = "ldlm_cbd",
2812 .psc_watchdog_factor = 2,
2814 .bc_nbufs = LDLM_CLIENT_NBUFS,
2815 .bc_buf_size = LDLM_BUFSIZE,
2816 .bc_req_max_size = LDLM_MAXREQSIZE,
2817 .bc_rep_max_size = LDLM_MAXREPSIZE,
2818 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2819 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2822 .tc_thr_name = "ldlm_cb",
2823 .tc_thr_factor = LDLM_THR_FACTOR,
2824 .tc_nthrs_init = LDLM_NTHRS_INIT,
2825 .tc_nthrs_base = LDLM_NTHRS_BASE,
2826 .tc_nthrs_max = LDLM_NTHRS_MAX,
2827 .tc_nthrs_user = ldlm_num_threads,
2828 .tc_cpu_affinity = 1,
2829 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2832 .cc_pattern = ldlm_cpts,
2835 .so_req_handler = ldlm_callback_handler,
2838 ldlm_state->ldlm_cb_service = \
2839 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2840 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2841 CERROR("failed to start service\n");
2842 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2843 ldlm_state->ldlm_cb_service = NULL;
2847 #ifdef HAVE_SERVER_SUPPORT
2848 memset(&conf, 0, sizeof(conf));
2849 conf = (typeof(conf)) {
2850 .psc_name = "ldlm_canceld",
2851 .psc_watchdog_factor = 6,
2853 .bc_nbufs = LDLM_SERVER_NBUFS,
2854 .bc_buf_size = LDLM_BUFSIZE,
2855 .bc_req_max_size = LDLM_MAXREQSIZE,
2856 .bc_rep_max_size = LDLM_MAXREPSIZE,
2857 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
2858 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
2862 .tc_thr_name = "ldlm_cn",
2863 .tc_thr_factor = LDLM_THR_FACTOR,
2864 .tc_nthrs_init = LDLM_NTHRS_INIT,
2865 .tc_nthrs_base = LDLM_NTHRS_BASE,
2866 .tc_nthrs_max = LDLM_NTHRS_MAX,
2867 .tc_nthrs_user = ldlm_num_threads,
2868 .tc_cpu_affinity = 1,
2869 .tc_ctx_tags = LCT_MD_THREAD | \
2874 .cc_pattern = ldlm_cpts,
2877 .so_req_handler = ldlm_cancel_handler,
2878 .so_hpreq_handler = ldlm_hpreq_handler,
2881 ldlm_state->ldlm_cancel_service = \
2882 ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2883 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2884 CERROR("failed to start service\n");
2885 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2886 ldlm_state->ldlm_cancel_service = NULL;
2891 OBD_ALLOC(blp, sizeof(*blp));
2893 GOTO(out, rc = -ENOMEM);
2894 ldlm_state->ldlm_bl_pool = blp;
2896 spin_lock_init(&blp->blp_lock);
2897 CFS_INIT_LIST_HEAD(&blp->blp_list);
2898 CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2899 init_waitqueue_head(&blp->blp_waitq);
2900 cfs_atomic_set(&blp->blp_num_threads, 0);
2901 cfs_atomic_set(&blp->blp_busy_threads, 0);
2904 if (ldlm_num_threads == 0) {
2905 blp->blp_min_threads = LDLM_NTHRS_INIT;
2906 blp->blp_max_threads = LDLM_NTHRS_MAX;
2908 blp->blp_min_threads = blp->blp_max_threads = \
2909 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
2913 for (i = 0; i < blp->blp_min_threads; i++) {
2914 rc = ldlm_bl_thread_start(blp);
2919 # ifdef HAVE_SERVER_SUPPORT
2920 CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2921 expired_lock_thread.elt_state = ELT_STOPPED;
2922 init_waitqueue_head(&expired_lock_thread.elt_waitq);
2924 CFS_INIT_LIST_HEAD(&waiting_locks_list);
2925 spin_lock_init(&waiting_locks_spinlock);
2926 cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2928 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
2931 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2935 wait_event(expired_lock_thread.elt_waitq,
2936 expired_lock_thread.elt_state == ELT_READY);
2937 # endif /* HAVE_SERVER_SUPPORT */
2939 rc = ldlm_pools_init();
2941 CERROR("Failed to initialize LDLM pools: %d\n", rc);
2952 static int ldlm_cleanup(void)
2956 if (!cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2957 !cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2958 CERROR("ldlm still has namespaces; clean these up first.\n");
2959 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2960 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2967 if (ldlm_state->ldlm_bl_pool != NULL) {
2968 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2970 while (cfs_atomic_read(&blp->blp_num_threads) > 0) {
2971 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2973 init_completion(&blp->blp_comp);
2975 spin_lock(&blp->blp_lock);
2976 cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2977 wake_up(&blp->blp_waitq);
2978 spin_unlock(&blp->blp_lock);
2980 wait_for_completion(&blp->blp_comp);
2983 OBD_FREE(blp, sizeof(*blp));
2985 #endif /* __KERNEL__ */
2987 if (ldlm_state->ldlm_cb_service != NULL)
2988 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2989 # ifdef HAVE_SERVER_SUPPORT
2990 if (ldlm_state->ldlm_cancel_service != NULL)
2991 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2995 ldlm_proc_cleanup();
2997 # ifdef HAVE_SERVER_SUPPORT
2998 if (expired_lock_thread.elt_state != ELT_STOPPED) {
2999 expired_lock_thread.elt_state = ELT_TERMINATE;
3000 wake_up(&expired_lock_thread.elt_waitq);
3001 wait_event(expired_lock_thread.elt_waitq,
3002 expired_lock_thread.elt_state == ELT_STOPPED);
3005 #endif /* __KERNEL__ */
3007 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3015 mutex_init(&ldlm_ref_mutex);
3016 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
3017 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
3018 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3019 sizeof(struct ldlm_resource), 0,
3020 SLAB_HWCACHE_ALIGN, NULL);
3021 if (ldlm_resource_slab == NULL)
3024 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3025 sizeof(struct ldlm_lock), 0,
3026 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3027 if (ldlm_lock_slab == NULL) {
3028 kmem_cache_destroy(ldlm_resource_slab);
3032 ldlm_interval_slab = kmem_cache_create("interval_node",
3033 sizeof(struct ldlm_interval),
3034 0, SLAB_HWCACHE_ALIGN, NULL);
3035 if (ldlm_interval_slab == NULL) {
3036 kmem_cache_destroy(ldlm_resource_slab);
3037 kmem_cache_destroy(ldlm_lock_slab);
3040 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3041 class_export_dump_hook = ldlm_dump_export_locks;
3046 void ldlm_exit(void)
3049 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3050 kmem_cache_destroy(ldlm_resource_slab);
3052 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3053 * synchronize_rcu() to wait a grace period elapsed, so that
3054 * ldlm_lock_free() get a chance to be called. */
3057 kmem_cache_destroy(ldlm_lock_slab);
3058 kmem_cache_destroy(ldlm_interval_slab);