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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lockd.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <linux/kthread.h>
41 #include <linux/list.h>
42 #include <libcfs/libcfs.h>
43 #include <lustre_errno.h>
44 #include <lustre_dlm.h>
45 #include <obd_class.h>
46 #include "ldlm_internal.h"
48 static int ldlm_num_threads;
49 module_param(ldlm_num_threads, int, 0444);
50 MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start");
52 static unsigned int ldlm_cpu_bind = 1;
53 module_param(ldlm_cpu_bind, uint, 0444);
54 MODULE_PARM_DESC(ldlm_cpu_bind,
55 "bind DLM service threads to particular CPU partitions");
57 static char *ldlm_cpts;
58 module_param(ldlm_cpts, charp, 0444);
59 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
61 static DEFINE_MUTEX(ldlm_ref_mutex);
62 static int ldlm_refcount;
64 struct kobject *ldlm_kobj;
65 struct kset *ldlm_ns_kset;
66 struct kset *ldlm_svc_kset;
70 static struct ldlm_state *ldlm_state;
72 /* timeout for initial callback (AST) reply (bz10399)
73 * Due to having to send a 32 bit time value over the
74 * wire return it as time_t instead of time64_t
76 static inline time_t ldlm_get_rq_timeout(void)
79 time_t timeout = min(ldlm_timeout, obd_timeout / 3);
81 return timeout < 1 ? 1 : timeout;
88 * blp_prio_list is used for callbacks that should be handled
89 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
92 struct list_head blp_prio_list;
95 * blp_list is used for all other callbacks which are likely
96 * to take longer to process.
98 struct list_head blp_list;
100 wait_queue_head_t blp_waitq;
101 struct completion blp_comp;
102 atomic_t blp_num_threads;
103 atomic_t blp_busy_threads;
108 struct ldlm_bl_work_item {
109 struct list_head blwi_entry;
110 struct ldlm_namespace *blwi_ns;
111 struct ldlm_lock_desc blwi_ld;
112 struct ldlm_lock *blwi_lock;
113 struct list_head blwi_head;
115 struct completion blwi_comp;
116 enum ldlm_cancel_flags blwi_flags;
117 int blwi_mem_pressure;
120 #ifdef HAVE_SERVER_SUPPORT
123 * Protects both waiting_locks_list and expired_lock_thread.
125 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
128 * List for contended locks.
130 * As soon as a lock is contended, it gets placed on this list and
131 * expected time to get a response is filled in the lock. A special
132 * thread walks the list looking for locks that should be released and
133 * schedules client evictions for those that have not been released in
136 * All access to it should be under waiting_locks_spinlock.
138 static LIST_HEAD(waiting_locks_list);
139 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
140 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
148 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
149 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
150 static int expired_lock_dump;
151 static LIST_HEAD(expired_lock_list);
153 static int ldlm_lock_busy(struct ldlm_lock *lock);
154 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout);
155 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds);
157 static inline int have_expired_locks(void)
162 spin_lock_bh(&waiting_locks_spinlock);
163 need_to_run = !list_empty(&expired_lock_list);
164 spin_unlock_bh(&waiting_locks_spinlock);
170 * Check expired lock list for expired locks and time them out.
172 static int expired_lock_main(void *arg)
174 struct list_head *expired = &expired_lock_list;
175 struct l_wait_info lwi = { 0 };
180 expired_lock_thread_state = ELT_READY;
181 wake_up(&expired_lock_wait_queue);
184 l_wait_event(expired_lock_wait_queue,
185 have_expired_locks() ||
186 expired_lock_thread_state == ELT_TERMINATE,
189 spin_lock_bh(&waiting_locks_spinlock);
190 if (expired_lock_dump) {
191 spin_unlock_bh(&waiting_locks_spinlock);
193 /* from waiting_locks_callback, but not in timer */
194 libcfs_debug_dumplog();
196 spin_lock_bh(&waiting_locks_spinlock);
197 expired_lock_dump = 0;
202 while (!list_empty(expired)) {
203 struct obd_export *export;
204 struct ldlm_lock *lock;
206 lock = list_entry(expired->next, struct ldlm_lock,
208 if ((void *)lock < LP_POISON + PAGE_SIZE &&
209 (void *)lock >= LP_POISON) {
210 spin_unlock_bh(&waiting_locks_spinlock);
211 CERROR("free lock on elt list %p\n", lock);
214 list_del_init(&lock->l_pending_chain);
215 if ((void *)lock->l_export <
216 LP_POISON + PAGE_SIZE &&
217 (void *)lock->l_export >= LP_POISON) {
218 CERROR("lock with free export on elt list %p\n",
220 lock->l_export = NULL;
221 LDLM_ERROR(lock, "free export");
222 /* release extra ref grabbed by
223 * ldlm_add_waiting_lock() or
224 * ldlm_failed_ast() */
225 LDLM_LOCK_RELEASE(lock);
229 if (ldlm_is_destroyed(lock)) {
230 /* release the lock refcount where
231 * waiting_locks_callback() founds */
232 LDLM_LOCK_RELEASE(lock);
235 export = class_export_lock_get(lock->l_export, lock);
236 spin_unlock_bh(&waiting_locks_spinlock);
238 /* Check if we need to prolong timeout */
239 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
240 lock->l_callback_timeout != 0 && /* not AST error */
241 ldlm_lock_busy(lock)) {
242 LDLM_DEBUG(lock, "prolong the busy lock");
243 lock_res_and_lock(lock);
244 ldlm_add_waiting_lock(lock,
245 ldlm_bl_timeout(lock) >> 1);
246 unlock_res_and_lock(lock);
248 spin_lock_bh(&export->exp_bl_list_lock);
249 list_del_init(&lock->l_exp_list);
250 spin_unlock_bh(&export->exp_bl_list_lock);
253 "lock callback timer expired after %llds: evicting client at %s ",
254 ktime_get_real_seconds() -
256 obd_export_nid2str(export));
257 ldlm_lock_to_ns(lock)->ns_timeouts++;
259 class_fail_export(export);
261 class_export_lock_put(export, lock);
262 /* release extra ref grabbed by ldlm_add_waiting_lock()
263 * or ldlm_failed_ast() */
264 LDLM_LOCK_RELEASE(lock);
266 spin_lock_bh(&waiting_locks_spinlock);
268 spin_unlock_bh(&waiting_locks_spinlock);
270 if (do_dump && obd_dump_on_eviction) {
271 CERROR("dump the log upon eviction\n");
272 libcfs_debug_dumplog();
275 if (expired_lock_thread_state == ELT_TERMINATE)
279 expired_lock_thread_state = ELT_STOPPED;
280 wake_up(&expired_lock_wait_queue);
285 * Check if there is a request in the export request list
286 * which prevents the lock canceling.
288 static int ldlm_lock_busy(struct ldlm_lock *lock)
290 struct ptlrpc_request *req;
294 if (lock->l_export == NULL)
297 spin_lock(&lock->l_export->exp_rpc_lock);
298 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
300 if (req->rq_ops->hpreq_lock_match) {
301 match = req->rq_ops->hpreq_lock_match(req, lock);
306 spin_unlock(&lock->l_export->exp_rpc_lock);
310 /* This is called from within a timer interrupt and cannot schedule */
311 static void waiting_locks_callback(TIMER_DATA_TYPE unused)
313 struct ldlm_lock *lock;
316 spin_lock_bh(&waiting_locks_spinlock);
317 while (!list_empty(&waiting_locks_list)) {
318 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
320 if (lock->l_callback_timeout > ktime_get_seconds() ||
321 lock->l_req_mode == LCK_GROUP)
324 /* no needs to take an extra ref on the lock since it was in
325 * the waiting_locks_list and ldlm_add_waiting_lock()
326 * already grabbed a ref */
327 list_del(&lock->l_pending_chain);
328 list_add(&lock->l_pending_chain, &expired_lock_list);
332 if (!list_empty(&expired_lock_list)) {
333 if (obd_dump_on_timeout && need_dump)
334 expired_lock_dump = __LINE__;
336 wake_up(&expired_lock_wait_queue);
340 * Make sure the timer will fire again if we have any locks
343 if (!list_empty(&waiting_locks_list)) {
344 unsigned long timeout_jiffies;
346 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
348 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
349 mod_timer(&waiting_locks_timer, timeout_jiffies);
351 spin_unlock_bh(&waiting_locks_spinlock);
355 * Add lock to the list of contended locks.
357 * Indicate that we're waiting for a client to call us back cancelling a given
358 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
359 * timer to fire appropriately. (We round up to the next second, to avoid
360 * floods of timer firings during periods of high lock contention and traffic).
361 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
362 * if it has been added to the waiting list (1 is returned).
364 * Called with the namespace lock held.
366 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds)
368 unsigned long timeout_jiffies;
371 if (!list_empty(&lock->l_pending_chain))
374 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
375 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
378 timeout = ktime_get_seconds() + seconds;
379 if (likely(timeout > lock->l_callback_timeout))
380 lock->l_callback_timeout = timeout;
382 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
384 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
385 !timer_pending(&waiting_locks_timer))
386 mod_timer(&waiting_locks_timer, timeout_jiffies);
388 /* if the new lock has a shorter timeout than something earlier on
389 * the list, we'll wait the longer amount of time; no big deal.
392 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
396 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
398 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
399 if (list_empty(&lock->l_exp_list)) {
400 if (!ldlm_is_granted(lock))
401 list_add_tail(&lock->l_exp_list,
402 &lock->l_export->exp_bl_list);
404 list_add(&lock->l_exp_list,
405 &lock->l_export->exp_bl_list);
407 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
409 /* A blocked lock is added. Adjust the position in
410 * the stale list if the export is in the list.
411 * If export is stale and not in the list - it is being
412 * processed and will be placed on the right position
413 * on obd_stale_export_put(). */
414 if (!list_empty(&lock->l_export->exp_stale_list))
415 obd_stale_export_adjust(lock->l_export);
418 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
422 /* NB: must be called with hold of lock_res_and_lock() */
423 LASSERT(ldlm_is_res_locked(lock));
424 LASSERT(!ldlm_is_cancel_on_block(lock));
426 /* Do not put cross-MDT lock in the waiting list, since we
427 * will not evict it due to timeout for now */
428 if (lock->l_export != NULL &&
429 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
432 spin_lock_bh(&waiting_locks_spinlock);
433 if (ldlm_is_cancel(lock)) {
434 spin_unlock_bh(&waiting_locks_spinlock);
438 if (ldlm_is_destroyed(lock)) {
439 static time64_t next;
441 spin_unlock_bh(&waiting_locks_spinlock);
442 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
443 if (ktime_get_seconds() > next) {
444 next = ktime_get_seconds() + 14400;
445 libcfs_debug_dumpstack(NULL);
450 ldlm_set_waited(lock);
451 lock->l_blast_sent = ktime_get_real_seconds();
452 ret = __ldlm_add_waiting_lock(lock, timeout);
454 /* grab ref on the lock if it has been added to the
458 spin_unlock_bh(&waiting_locks_spinlock);
461 ldlm_add_blocked_lock(lock);
463 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)",
464 ret == 0 ? "not re-" : "", timeout,
465 AT_OFF ? "off" : "on");
470 * Remove a lock from the pending list, likely because it had its cancellation
471 * callback arrive without incident. This adjusts the lock-timeout timer if
472 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
473 * As done by ldlm_del_waiting_lock(), the caller must release the lock
474 * reference when the lock is removed from any list (1 is returned).
476 * Called with namespace lock held.
478 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
480 struct list_head *list_next;
482 if (list_empty(&lock->l_pending_chain))
485 list_next = lock->l_pending_chain.next;
486 if (lock->l_pending_chain.prev == &waiting_locks_list) {
487 /* Removing the head of the list, adjust timer. */
488 if (list_next == &waiting_locks_list) {
489 /* No more, just cancel. */
490 del_timer(&waiting_locks_timer);
492 struct ldlm_lock *next;
494 next = list_entry(list_next, struct ldlm_lock,
496 mod_timer(&waiting_locks_timer,
497 cfs_time_seconds(next->l_callback_timeout));
500 list_del_init(&lock->l_pending_chain);
505 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
509 if (lock->l_export == NULL) {
510 /* We don't have a "waiting locks list" on clients. */
511 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
515 spin_lock_bh(&waiting_locks_spinlock);
516 ret = __ldlm_del_waiting_lock(lock);
517 ldlm_clear_waited(lock);
518 spin_unlock_bh(&waiting_locks_spinlock);
520 /* remove the lock out of export blocking list */
521 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
522 list_del_init(&lock->l_exp_list);
523 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
526 /* release lock ref if it has indeed been removed
528 LDLM_LOCK_RELEASE(lock);
531 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
536 * Prolong the contended lock waiting time.
538 * Called with namespace lock held.
540 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
542 if (lock->l_export == NULL) {
543 /* We don't have a "waiting locks list" on clients. */
544 LDLM_DEBUG(lock, "client lock: no-op");
548 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
549 /* We don't have a "waiting locks list" on OSP. */
550 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
554 spin_lock_bh(&waiting_locks_spinlock);
556 if (list_empty(&lock->l_pending_chain)) {
557 spin_unlock_bh(&waiting_locks_spinlock);
558 LDLM_DEBUG(lock, "wasn't waiting");
562 /* we remove/add the lock to the waiting list, so no needs to
563 * release/take a lock reference */
564 __ldlm_del_waiting_lock(lock);
565 __ldlm_add_waiting_lock(lock, timeout);
566 spin_unlock_bh(&waiting_locks_spinlock);
568 LDLM_DEBUG(lock, "refreshed");
571 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
573 #else /* HAVE_SERVER_SUPPORT */
575 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
580 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
585 #endif /* !HAVE_SERVER_SUPPORT */
587 #ifdef HAVE_SERVER_SUPPORT
590 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
591 * lock cancel, and their replies). Used for lock callback timeout and AST
594 * \param[in] lock lock which is getting the blocking callback
596 * \retval timeout in seconds to wait for the client reply
598 time64_t ldlm_bl_timeout(struct ldlm_lock *lock)
603 return obd_timeout / 2;
605 /* Since these are non-updating timeouts, we should be conservative.
606 * Take more than usually, 150%
607 * It would be nice to have some kind of "early reply" mechanism for
608 * lock callbacks too... */
609 timeout = at_get(&lock->l_export->exp_bl_lock_at);
610 return max(timeout + (timeout >> 1), (time64_t)ldlm_enqueue_min);
612 EXPORT_SYMBOL(ldlm_bl_timeout);
615 * Perform lock cleanup if AST sending failed.
617 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
618 const char *ast_type)
620 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
621 "to a lock %s callback time out: rc %d\n",
622 lock->l_export->exp_obd->obd_name,
623 obd_export_nid2str(lock->l_export), ast_type, rc);
625 if (obd_dump_on_timeout)
626 libcfs_debug_dumplog();
627 spin_lock_bh(&waiting_locks_spinlock);
628 if (__ldlm_del_waiting_lock(lock) == 0)
629 /* the lock was not in any list, grab an extra ref before adding
630 * the lock to the expired list */
632 lock->l_callback_timeout = 0; /* differentiate it from expired locks */
633 list_add(&lock->l_pending_chain, &expired_lock_list);
634 wake_up(&expired_lock_wait_queue);
635 spin_unlock_bh(&waiting_locks_spinlock);
639 * Perform lock cleanup if AST reply came with error.
641 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
642 struct ptlrpc_request *req, int rc,
643 const char *ast_type)
645 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
647 if (!req->rq_replied || (rc && rc != -EINVAL)) {
648 if (ldlm_is_cancel(lock)) {
650 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
651 ast_type, req, req->rq_xid,
652 libcfs_nid2str(peer.nid));
653 ldlm_lock_cancel(lock);
655 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
657 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
658 /* Upon umount process the AST fails because cannot be
659 * sent. This shouldn't lead to the client eviction.
660 * -ENODEV error is returned by ptl_send_rpc() for
661 * new request in such import.
662 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
663 * if imp_invalid is set or obd_no_recov.
664 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
665 * in ptlrpc_import_delay_req() as well with -EIO code.
666 * In all such cases errors are ignored.
668 LDLM_DEBUG(lock, "%s AST can't be sent due to a server"
669 " %s failure or umount process: rc = %d\n",
671 req->rq_import->imp_obd->obd_name, rc);
674 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
675 libcfs_nid2str(peer.nid),
676 req->rq_replied ? "returned error from" :
677 "failed to reply to",
678 ast_type, req, req->rq_xid,
679 (req->rq_repmsg != NULL) ?
680 lustre_msg_get_status(req->rq_repmsg) : 0,
682 ldlm_failed_ast(lock, rc, ast_type);
688 struct ldlm_resource *res = lock->l_resource;
691 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
692 libcfs_nid2str(peer.nid),
694 lustre_msg_get_status(req->rq_repmsg) : -1,
695 ast_type, req, req->rq_xid);
697 /* update lvbo to return proper attributes.
699 ldlm_resource_getref(res);
700 ldlm_lvbo_update(res, lock, NULL, 1);
701 ldlm_resource_putref(res);
703 ldlm_lock_cancel(lock);
710 static int ldlm_cb_interpret(const struct lu_env *env,
711 struct ptlrpc_request *req, void *args, int rc)
713 struct ldlm_cb_async_args *ca = args;
714 struct ldlm_lock *lock = ca->ca_lock;
715 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
718 LASSERT(lock != NULL);
721 case LDLM_GL_CALLBACK:
722 /* Update the LVB from disk if the AST failed
723 * (this is a legal race)
725 * - Glimpse callback of local lock just returns
726 * -ELDLM_NO_LOCK_DATA.
727 * - Glimpse callback of remote lock might return
728 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
730 if (unlikely(arg->gl_interpret_reply)) {
731 rc = arg->gl_interpret_reply(NULL, req, args, rc);
732 } else if (rc == -ELDLM_NO_LOCK_DATA) {
734 "lost race - client has a lock but no inode");
735 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
736 } else if (rc != 0) {
737 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
739 rc = ldlm_lvbo_update(lock->l_resource,
743 case LDLM_BL_CALLBACK:
745 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
747 case LDLM_CP_CALLBACK:
749 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
752 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
757 /* release extra reference taken in ldlm_ast_fini() */
758 LDLM_LOCK_RELEASE(lock);
761 atomic_inc(&arg->restart);
766 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
768 struct ldlm_cb_async_args *ca = data;
769 struct ldlm_lock *lock = ca->ca_lock;
771 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
774 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
775 struct ldlm_cb_set_arg *arg,
776 struct ldlm_lock *lock,
782 if (unlikely(instant_cancel)) {
783 rc = ptl_send_rpc(req, 1);
784 ptlrpc_req_finished(req);
786 atomic_inc(&arg->restart);
789 ptlrpc_set_add_req(arg->set, req);
796 * Check if there are requests in the export request list which prevent
797 * the lock canceling and make these requests high priority ones.
799 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
801 struct ptlrpc_request *req;
804 if (lock->l_export == NULL) {
805 LDLM_DEBUG(lock, "client lock: no-op");
809 spin_lock(&lock->l_export->exp_rpc_lock);
810 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
812 /* Do not process requests that were not yet added to there
813 * incoming queue or were already removed from there for
814 * processing. We evaluate ptlrpc_nrs_req_can_move() without
815 * holding svcpt->scp_req_lock, and then redo the check with
816 * the lock held once we need to obtain a reliable result.
818 if (ptlrpc_nrs_req_can_move(req) &&
819 req->rq_ops->hpreq_lock_match &&
820 req->rq_ops->hpreq_lock_match(req, lock))
821 ptlrpc_nrs_req_hp_move(req);
823 spin_unlock(&lock->l_export->exp_rpc_lock);
828 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
829 * enqueued server lock conflicts with given one.
831 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
832 * to wait for client response.
834 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
835 struct ldlm_lock_desc *desc,
836 void *data, int flag)
838 struct ldlm_cb_async_args *ca;
839 struct ldlm_cb_set_arg *arg = data;
840 struct ldlm_request *body;
841 struct ptlrpc_request *req;
842 int instant_cancel = 0;
846 if (flag == LDLM_CB_CANCELING)
847 /* Don't need to do anything here. */
850 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
851 LDLM_DEBUG(lock, "dropping BL AST");
856 LASSERT(data != NULL);
857 if (lock->l_export->exp_obd->obd_recovering != 0)
858 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
860 ldlm_lock_reorder_req(lock);
862 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
863 &RQF_LDLM_BL_CALLBACK,
864 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
868 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
869 ca = ptlrpc_req_async_args(req);
870 ca->ca_set_arg = arg;
873 req->rq_interpret_reply = ldlm_cb_interpret;
875 lock_res_and_lock(lock);
876 if (ldlm_is_destroyed(lock)) {
877 /* What's the point? */
878 unlock_res_and_lock(lock);
879 ptlrpc_req_finished(req);
883 if (!ldlm_is_granted(lock)) {
884 /* this blocking AST will be communicated as part of the
885 * completion AST instead */
886 ldlm_add_blocked_lock(lock);
887 ldlm_set_waited(lock);
888 unlock_res_and_lock(lock);
890 ptlrpc_req_finished(req);
891 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
895 if (ldlm_is_cancel_on_block(lock))
898 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
899 body->lock_handle[0] = lock->l_remote_handle;
900 body->lock_desc = *desc;
901 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
903 LDLM_DEBUG(lock, "server preparing blocking AST");
905 ptlrpc_request_set_replen(req);
906 ldlm_set_cbpending(lock);
907 if (instant_cancel) {
908 unlock_res_and_lock(lock);
909 ldlm_lock_cancel(lock);
911 req->rq_no_resend = 1;
913 LASSERT(ldlm_is_granted(lock));
914 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
915 unlock_res_and_lock(lock);
917 /* Do not resend after lock callback timeout */
918 req->rq_delay_limit = ldlm_bl_timeout(lock);
919 req->rq_resend_cb = ldlm_update_resend;
922 req->rq_send_state = LUSTRE_IMP_FULL;
923 /* ptlrpc_request_alloc_pack already set timeout */
925 req->rq_timeout = ldlm_get_rq_timeout();
927 if (lock->l_export && lock->l_export->exp_nid_stats &&
928 lock->l_export->exp_nid_stats->nid_ldlm_stats)
929 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
930 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
932 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
938 * ->l_completion_ast callback for a remote lock in server namespace.
940 * Sends AST to the client notifying it of lock granting. If initial
941 * lock response was not sent yet, instead of sending another RPC, just
942 * mark the lock as granted and client will understand
944 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
946 struct ldlm_cb_set_arg *arg = data;
947 struct ldlm_request *body;
948 struct ptlrpc_request *req;
949 struct ldlm_cb_async_args *ca;
950 int instant_cancel = 0;
955 LASSERT(lock != NULL);
956 LASSERT(data != NULL);
958 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
959 LDLM_DEBUG(lock, "dropping CP AST");
963 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
964 &RQF_LDLM_CP_CALLBACK);
968 /* server namespace, doesn't need lock */
969 lvb_len = ldlm_lvbo_size(lock);
970 /* LU-3124 & LU-2187: to not return layout in completion AST because
971 * it may deadlock for LU-2187, or client may not have enough space
972 * for large layout. The layout will be returned to client with an
973 * extra RPC to fetch xattr.lov */
974 if (ldlm_has_layout(lock))
977 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
978 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
980 ptlrpc_request_free(req);
984 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
985 ca = ptlrpc_req_async_args(req);
986 ca->ca_set_arg = arg;
989 req->rq_interpret_reply = ldlm_cb_interpret;
990 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
992 body->lock_handle[0] = lock->l_remote_handle;
993 body->lock_flags = ldlm_flags_to_wire(flags);
994 ldlm_lock2desc(lock, &body->lock_desc);
996 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
997 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
999 /* We still need to send the RPC to wake up the blocked
1000 * enqueue thread on the client.
1002 * Consider old client, there is no better way to notify
1003 * the failure, just zero-sized the LVB, then the client
1004 * will fail out as "-EPROTO". */
1005 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1009 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1014 LDLM_DEBUG(lock, "server preparing completion AST");
1016 ptlrpc_request_set_replen(req);
1018 req->rq_send_state = LUSTRE_IMP_FULL;
1019 /* ptlrpc_request_pack already set timeout */
1021 req->rq_timeout = ldlm_get_rq_timeout();
1023 /* We only send real blocking ASTs after the lock is granted */
1024 lock_res_and_lock(lock);
1025 if (ldlm_is_ast_sent(lock)) {
1026 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1027 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1028 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1031 /* We might get here prior to ldlm_handle_enqueue setting
1032 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1033 * into waiting list, but this is safe and similar code in
1034 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1035 * that would not only cancel the lock, but will also remove
1036 * it from waiting list */
1037 if (ldlm_is_cancel_on_block(lock)) {
1038 unlock_res_and_lock(lock);
1039 ldlm_lock_cancel(lock);
1042 req->rq_no_resend = 1;
1044 lock_res_and_lock(lock);
1046 /* start the lock-timeout clock */
1047 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1048 /* Do not resend after lock callback timeout */
1049 req->rq_delay_limit = ldlm_bl_timeout(lock);
1050 req->rq_resend_cb = ldlm_update_resend;
1053 unlock_res_and_lock(lock);
1055 if (lock->l_export && lock->l_export->exp_nid_stats &&
1056 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1057 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1058 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1060 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1062 RETURN(lvb_len < 0 ? lvb_len : rc);
1066 * Server side ->l_glimpse_ast handler for client locks.
1068 * Sends glimpse AST to the client and waits for reply. Then updates
1069 * lvbo with the result.
1071 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1073 struct ldlm_cb_set_arg *arg = data;
1074 struct ldlm_request *body;
1075 struct ptlrpc_request *req;
1076 struct ldlm_cb_async_args *ca;
1078 struct req_format *req_fmt;
1081 LASSERT(lock != NULL);
1083 if (arg->gl_desc != NULL)
1084 /* There is a glimpse descriptor to pack */
1085 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1087 req_fmt = &RQF_LDLM_GL_CALLBACK;
1089 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1090 req_fmt, LUSTRE_DLM_VERSION,
1096 if (arg->gl_desc != NULL) {
1097 /* copy the GL descriptor */
1098 union ldlm_gl_desc *desc;
1099 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1100 *desc = *arg->gl_desc;
1103 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1104 body->lock_handle[0] = lock->l_remote_handle;
1105 ldlm_lock2desc(lock, &body->lock_desc);
1107 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1108 ca = ptlrpc_req_async_args(req);
1109 ca->ca_set_arg = arg;
1112 /* server namespace, doesn't need lock */
1113 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1114 ldlm_lvbo_size(lock));
1115 ptlrpc_request_set_replen(req);
1117 req->rq_send_state = LUSTRE_IMP_FULL;
1118 /* ptlrpc_request_alloc_pack already set timeout */
1120 req->rq_timeout = ldlm_get_rq_timeout();
1122 req->rq_interpret_reply = ldlm_cb_interpret;
1124 if (lock->l_export && lock->l_export->exp_nid_stats &&
1125 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1126 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1127 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1129 rc = ldlm_ast_fini(req, arg, lock, 0);
1133 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1135 int ldlm_glimpse_locks(struct ldlm_resource *res,
1136 struct list_head *gl_work_list)
1141 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1143 if (rc == -ERESTART)
1144 ldlm_reprocess_all(res, NULL);
1148 EXPORT_SYMBOL(ldlm_glimpse_locks);
1150 /* return LDLM lock associated with a lock callback request */
1151 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1153 struct ldlm_cb_async_args *ca;
1154 struct ldlm_lock *lock;
1157 ca = ptlrpc_req_async_args(req);
1160 RETURN(ERR_PTR(-EFAULT));
1164 EXPORT_SYMBOL(ldlm_request_lock);
1167 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1168 * service threads to carry out client lock enqueueing requests.
1170 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1171 struct ptlrpc_request *req,
1172 const struct ldlm_request *dlm_req,
1173 const struct ldlm_callback_suite *cbs)
1175 struct ldlm_reply *dlm_rep;
1177 enum ldlm_error err = ELDLM_OK;
1178 struct ldlm_lock *lock = NULL;
1179 void *cookie = NULL;
1181 struct ldlm_resource *res = NULL;
1182 const struct lu_env *env = req->rq_svc_thread->t_env;
1185 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1187 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1188 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1190 LASSERT(req->rq_export);
1192 /* for intent enqueue the stat will be updated inside intent policy */
1193 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1194 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1195 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1197 if (req->rq_export && req->rq_export->exp_nid_stats &&
1198 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1199 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1200 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1202 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1203 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1204 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1205 dlm_req->lock_desc.l_resource.lr_type);
1206 GOTO(out, rc = -EFAULT);
1209 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1210 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1211 dlm_req->lock_desc.l_req_mode &
1212 (dlm_req->lock_desc.l_req_mode-1))) {
1213 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1214 dlm_req->lock_desc.l_req_mode);
1215 GOTO(out, rc = -EFAULT);
1218 if (unlikely((flags & LDLM_FL_REPLAY) ||
1219 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1220 /* Find an existing lock in the per-export lock hash */
1221 /* In the function below, .hs_keycmp resolves to
1222 * ldlm_export_lock_keycmp() */
1223 /* coverity[overrun-buffer-val] */
1224 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1225 (void *)&dlm_req->lock_handle[0]);
1227 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie %#llx",
1228 lock->l_handle.h_cookie);
1229 flags |= LDLM_FL_RESENT;
1230 GOTO(existing_lock, rc = 0);
1233 if (ldlm_reclaim_full()) {
1234 DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1235 "reject current enqueue request and let the "
1236 "client retry later.\n");
1237 GOTO(out, rc = -EINPROGRESS);
1241 /* The lock's callback data might be set in the policy function */
1242 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1243 dlm_req->lock_desc.l_resource.lr_type,
1244 dlm_req->lock_desc.l_req_mode,
1245 cbs, NULL, 0, LVB_T_NONE);
1252 lock->l_remote_handle = dlm_req->lock_handle[0];
1253 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1255 /* Initialize resource lvb but not for a lock being replayed since
1256 * Client already got lvb sent in this case.
1257 * This must occur early since some policy methods assume resource
1258 * lvb is available (lr_lvb_data != NULL).
1260 res = lock->l_resource;
1261 if (!(flags & LDLM_FL_REPLAY)) {
1262 /* non-replayed lock, delayed lvb init may need to be done */
1263 rc = ldlm_lvbo_init(res);
1265 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1270 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1271 /* Don't enqueue a lock onto the export if it is been disonnected
1272 * due to eviction (bug 3822) or server umount (bug 24324).
1273 * Cancel it now instead. */
1274 if (req->rq_export->exp_disconnected) {
1275 LDLM_ERROR(lock, "lock on disconnected export %p",
1277 GOTO(out, rc = -ENOTCONN);
1280 lock->l_export = class_export_lock_get(req->rq_export, lock);
1281 if (lock->l_export->exp_lock_hash)
1282 cfs_hash_add(lock->l_export->exp_lock_hash,
1283 &lock->l_remote_handle,
1286 /* Inherit the enqueue flags before the operation, because we do not
1287 * keep the res lock on return and next operations (BL AST) may proceed
1289 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1290 LDLM_FL_INHERIT_MASK);
1292 ldlm_convert_policy_to_local(req->rq_export,
1293 dlm_req->lock_desc.l_resource.lr_type,
1294 &dlm_req->lock_desc.l_policy_data,
1295 &lock->l_policy_data);
1296 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1297 lock->l_req_extent = lock->l_policy_data.l_extent;
1298 else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS)
1299 lock->l_policy_data.l_inodebits.try_bits =
1300 dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1303 if (flags & LDLM_FL_HAS_INTENT) {
1304 /* In this case, the reply buffer is allocated deep in
1305 * local_lock_enqueue by the policy function. */
1308 /* based on the assumption that lvb size never changes during
1309 * resource life time otherwise it need resource->lr_lock's
1311 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1312 RCL_SERVER, ldlm_lvbo_size(lock));
1314 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1315 GOTO(out, rc = -ENOMEM);
1317 rc = req_capsule_server_pack(&req->rq_pill);
1322 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1329 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1331 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1332 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1334 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1335 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1337 /* We never send a blocking AST until the lock is granted, but
1338 * we can tell it right now */
1339 lock_res_and_lock(lock);
1341 /* Now take into account flags to be inherited from original lock
1342 request both in reply to client and in our own lock flags. */
1343 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1344 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1346 /* Don't move a pending lock onto the export if it has already been
1347 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1348 * Cancel it now instead. */
1349 if (unlikely(req->rq_export->exp_disconnected ||
1350 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1351 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1353 } else if (ldlm_is_ast_sent(lock)) {
1354 /* fill lock desc for possible lock convert */
1355 if (lock->l_blocking_lock &&
1356 lock->l_resource->lr_type == LDLM_IBITS) {
1357 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1358 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1361 "save blocking bits %llx in granted lock",
1362 bl_lock->l_policy_data.l_inodebits.bits);
1363 /* If lock is blocked then save blocking ibits
1364 * in returned lock policy for the possible lock
1365 * convert on a client.
1367 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1368 bl_lock->l_policy_data.l_inodebits.bits;
1370 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1371 if (ldlm_is_granted(lock)) {
1373 * Only cancel lock if it was granted, because it would
1374 * be destroyed immediately and would never be granted
1375 * in the future, causing timeouts on client. Not
1376 * granted lock will be cancelled immediately after
1377 * sending completion AST.
1379 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1380 unlock_res_and_lock(lock);
1381 ldlm_lock_cancel(lock);
1382 lock_res_and_lock(lock);
1384 ldlm_add_waiting_lock(lock,
1385 ldlm_bl_timeout(lock));
1389 unlock_res_and_lock(lock);
1393 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1394 if (!req->rq_packed_final) {
1395 err = lustre_pack_reply(req, 1, NULL, NULL);
1400 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1401 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1403 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1404 "(err=%d, rc=%d)", err, rc);
1407 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1409 ldlm_lvbo_size(lock) > 0) {
1414 buf = req_capsule_server_get(&req->rq_pill,
1416 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1417 buflen = req_capsule_get_size(&req->rq_pill,
1418 &RMF_DLM_LVB, RCL_SERVER);
1419 /* non-replayed lock, delayed lvb init may
1420 * need to be occur now
1422 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1425 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1427 req_capsule_shrink(&req->rq_pill,
1430 } else if (rc2 == -ERANGE) {
1431 rc2 = req_capsule_server_grow(
1433 &RMF_DLM_LVB, buflen);
1437 /* if we can't grow the buffer,
1438 * it's ok to return empty lvb
1449 } else if (flags & LDLM_FL_REPLAY) {
1450 /* no LVB resend upon replay */
1452 req_capsule_shrink(&req->rq_pill,
1462 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1463 if (lock->l_export) {
1464 ldlm_lock_cancel(lock);
1466 lock_res_and_lock(lock);
1467 ldlm_resource_unlink_lock(lock);
1468 ldlm_lock_destroy_nolock(lock);
1469 unlock_res_and_lock(lock);
1474 if (!err && !ldlm_is_cbpending(lock) &&
1475 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1476 ldlm_reprocess_all(lock->l_resource, lock);
1478 LDLM_LOCK_RELEASE(lock);
1481 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1487 /* Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1488 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1490 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1492 if (lock->l_blocking_lock) {
1493 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1494 lock->l_blocking_lock = NULL;
1498 /* A lock can be converted to new ibits or mode and should be considered
1499 * as new lock. Clear all states related to a previous blocking AST
1500 * processing so new conflicts will cause new blocking ASTs.
1502 * This is used during lock convert below and lock downgrade to COS mode in
1503 * ldlm_lock_mode_downgrade().
1505 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1507 ldlm_clear_ast_sent(lock);
1508 lock->l_bl_ast_run = 0;
1509 ldlm_clear_blocking_lock(lock);
1513 * Main LDLM entry point for server code to process lock conversion requests.
1515 int ldlm_handle_convert0(struct ptlrpc_request *req,
1516 const struct ldlm_request *dlm_req)
1518 struct obd_export *exp = req->rq_export;
1519 struct ldlm_reply *dlm_rep;
1520 struct ldlm_lock *lock;
1527 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1528 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1529 LDLM_CONVERT - LDLM_FIRST_OPC);
1531 rc = req_capsule_server_pack(&req->rq_pill);
1535 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1536 dlm_rep->lock_flags = dlm_req->lock_flags;
1538 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1540 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1541 req->rq_status = ELDLM_NO_LOCK_DATA;
1545 LDLM_DEBUG(lock, "server-side convert handler START");
1547 lock_res_and_lock(lock);
1548 bits = lock->l_policy_data.l_inodebits.bits;
1549 new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1551 if (ldlm_is_cancel(lock)) {
1552 LDLM_DEBUG(lock, "convert on canceled lock!");
1553 unlock_res_and_lock(lock);
1554 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1557 if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1558 LDLM_ERROR(lock, "lock mode differs!");
1559 unlock_res_and_lock(lock);
1560 GOTO(out_put, rc = -EPROTO);
1563 if (bits == new_bits) {
1565 * This can be valid situation if CONVERT RPCs are
1566 * re-ordered. Just finish silently
1568 LDLM_DEBUG(lock, "lock is converted already!");
1569 unlock_res_and_lock(lock);
1571 if (ldlm_is_waited(lock))
1572 ldlm_del_waiting_lock(lock);
1574 ldlm_clear_cbpending(lock);
1575 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1576 ldlm_inodebits_drop(lock, bits & ~new_bits);
1578 ldlm_clear_blocking_data(lock);
1579 unlock_res_and_lock(lock);
1581 ldlm_reprocess_all(lock->l_resource, NULL);
1584 dlm_rep->lock_handle = lock->l_remote_handle;
1585 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1586 &dlm_rep->lock_desc.l_policy_data);
1590 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1591 LDLM_LOCK_PUT(lock);
1592 req->rq_status = rc;
1597 * Cancel all the locks whose handles are packed into ldlm_request
1599 * Called by server code expecting such combined cancel activity
1602 int ldlm_request_cancel(struct ptlrpc_request *req,
1603 const struct ldlm_request *dlm_req,
1604 int first, enum lustre_at_flags flags)
1606 struct ldlm_resource *res, *pres = NULL;
1607 struct ldlm_lock *lock;
1608 int i, count, done = 0;
1613 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1614 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1615 (size - offsetof(struct ldlm_request, lock_handle)) /
1616 sizeof(struct lustre_handle) < dlm_req->lock_count)
1619 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1623 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1626 /* There is no lock on the server at the replay time,
1627 * skip lock cancelling to make replay tests to pass. */
1628 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1631 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1632 "starting at %d", count, first);
1634 for (i = first; i < count; i++) {
1635 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1637 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1638 "lock (cookie %llu)",
1639 dlm_req->lock_handle[i].cookie);
1643 res = lock->l_resource;
1646 /* This code is an optimization to only attempt lock
1647 * granting on the resource (that could be CPU-expensive)
1648 * after we are done cancelling lock in that resource.
1652 ldlm_reprocess_all(pres, NULL);
1653 LDLM_RESOURCE_DELREF(pres);
1654 ldlm_resource_putref(pres);
1657 ldlm_resource_getref(res);
1658 LDLM_RESOURCE_ADDREF(res);
1660 if (!ldlm_is_discard_data(lock))
1661 ldlm_lvbo_update(res, lock,
1667 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1668 lock->l_blast_sent != 0) {
1669 time64_t delay = ktime_get_real_seconds() -
1671 LDLM_DEBUG(lock, "server cancels blocked lock after %llds",
1673 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1675 ldlm_lock_cancel(lock);
1676 LDLM_LOCK_PUT(lock);
1679 ldlm_reprocess_all(pres, NULL);
1680 LDLM_RESOURCE_DELREF(pres);
1681 ldlm_resource_putref(pres);
1683 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1686 EXPORT_SYMBOL(ldlm_request_cancel);
1689 * Main LDLM entry point for server code to cancel locks.
1691 * Typically gets called from service handler on LDLM_CANCEL opc.
1693 int ldlm_handle_cancel(struct ptlrpc_request *req)
1695 struct ldlm_request *dlm_req;
1699 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1700 if (dlm_req == NULL) {
1701 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1705 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1706 offsetof(struct ldlm_request, lock_handle[1]))
1709 if (req->rq_export && req->rq_export->exp_nid_stats &&
1710 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1711 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1712 LDLM_CANCEL - LDLM_FIRST_OPC);
1714 rc = req_capsule_server_pack(&req->rq_pill);
1718 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1719 req->rq_status = LUSTRE_ESTALE;
1721 RETURN(ptlrpc_reply(req));
1723 #endif /* HAVE_SERVER_SUPPORT */
1726 * Server may pass additional information about blocking lock.
1727 * For IBITS locks it is conflicting bits which can be used for
1728 * lock convert instead of cancel.
1730 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1732 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1734 check_res_locked(lock->l_resource);
1735 if (ns_is_client(ns) && ld &&
1736 (lock->l_resource->lr_type == LDLM_IBITS)) {
1738 * Lock description contains policy of blocking lock,
1739 * and its cancel_bits is used to pass conflicting bits.
1740 * NOTE: ld can be NULL or can be not NULL but zeroed if
1741 * passed from ldlm_bl_thread_blwi(), check below used bits
1742 * in ld to make sure it is valid description.
1744 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1745 ldlm_res_eq(&ld->l_resource.lr_name,
1746 &lock->l_resource->lr_name) &&
1747 !(ldlm_is_cbpending(lock) &&
1748 lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1749 /* always combine conflicting ibits */
1750 lock->l_policy_data.l_inodebits.cancel_bits |=
1751 ld->l_policy_data.l_inodebits.cancel_bits;
1753 /* If cancel_bits are not obtained or
1754 * if the lock is already CBPENDING and
1755 * has no cancel_bits set
1756 * - the full lock is to be cancelled
1758 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1764 * Callback handler for receiving incoming blocking ASTs.
1766 * This can only happen on client side.
1768 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1769 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1775 LDLM_DEBUG(lock, "client blocking AST callback handler");
1777 lock_res_and_lock(lock);
1779 /* get extra information from desc if any */
1780 ldlm_bl_desc2lock(ld, lock);
1781 ldlm_set_cbpending(lock);
1783 if (ldlm_is_cancel_on_block(lock))
1784 ldlm_set_cancel(lock);
1786 do_ast = (!lock->l_readers && !lock->l_writers);
1787 unlock_res_and_lock(lock);
1790 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1791 lock, lock->l_blocking_ast);
1792 if (lock->l_blocking_ast != NULL)
1793 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1796 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1800 LDLM_DEBUG(lock, "client blocking callback handler END");
1801 LDLM_LOCK_RELEASE(lock);
1806 * Callback handler for receiving incoming completion ASTs.
1808 * This only can happen on client side.
1810 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1811 struct ldlm_namespace *ns,
1812 struct ldlm_request *dlm_req,
1813 struct ldlm_lock *lock)
1815 struct list_head ast_list;
1820 LDLM_DEBUG(lock, "client completion callback handler START");
1822 INIT_LIST_HEAD(&ast_list);
1823 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1824 long to = cfs_time_seconds(1);
1827 set_current_state(TASK_INTERRUPTIBLE);
1828 schedule_timeout(to);
1829 if (ldlm_is_granted(lock) ||
1830 ldlm_is_destroyed(lock))
1835 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1837 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1838 GOTO(out, rc = lvb_len);
1839 } else if (lvb_len > 0) {
1840 if (lock->l_lvb_len > 0) {
1841 /* for extent lock, lvb contains ost_lvb{}. */
1842 LASSERT(lock->l_lvb_data != NULL);
1844 if (unlikely(lock->l_lvb_len < lvb_len)) {
1845 LDLM_ERROR(lock, "Replied LVB is larger than "
1846 "expectation, expected = %d, "
1848 lock->l_lvb_len, lvb_len);
1849 GOTO(out, rc = -EINVAL);
1854 lock_res_and_lock(lock);
1856 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1857 &lock->l_resource->lr_name)) {
1858 ldlm_resource_unlink_lock(lock);
1859 unlock_res_and_lock(lock);
1860 rc = ldlm_lock_change_resource(ns, lock,
1861 &dlm_req->lock_desc.l_resource.lr_name);
1863 LDLM_ERROR(lock, "Failed to allocate resource");
1866 LDLM_DEBUG(lock, "completion AST, new resource");
1867 lock_res_and_lock(lock);
1870 if (ldlm_is_destroyed(lock) ||
1871 ldlm_is_granted(lock)) {
1872 /* bug 11300: the lock has already been granted */
1873 unlock_res_and_lock(lock);
1874 LDLM_DEBUG(lock, "Double grant race happened");
1878 /* If we receive the completion AST before the actual enqueue returned,
1879 * then we might need to switch lock modes, resources, or extents. */
1880 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1881 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1882 LDLM_DEBUG(lock, "completion AST, new lock mode");
1885 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1886 ldlm_convert_policy_to_local(req->rq_export,
1887 dlm_req->lock_desc.l_resource.lr_type,
1888 &dlm_req->lock_desc.l_policy_data,
1889 &lock->l_policy_data);
1890 LDLM_DEBUG(lock, "completion AST, new policy data");
1893 ldlm_resource_unlink_lock(lock);
1895 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1897 * BL_AST locks are not needed in LRU.
1898 * Let ldlm_cancel_lru() be fast.
1900 ldlm_lock_remove_from_lru(lock);
1901 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
1902 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1903 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1906 if (lock->l_lvb_len > 0) {
1907 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1908 lock->l_lvb_data, lvb_len);
1910 unlock_res_and_lock(lock);
1915 ldlm_grant_lock(lock, &ast_list);
1916 unlock_res_and_lock(lock);
1918 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1920 /* Let Enqueue to call osc_lock_upcall() and initialize
1922 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1924 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1926 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1932 lock_res_and_lock(lock);
1933 ldlm_set_failed(lock);
1934 unlock_res_and_lock(lock);
1935 wake_up(&lock->l_waitq);
1937 LDLM_LOCK_RELEASE(lock);
1941 * Callback handler for receiving incoming glimpse ASTs.
1943 * This only can happen on client side. After handling the glimpse AST
1944 * we also consider dropping the lock here if it is unused locally for a
1947 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1948 struct ldlm_namespace *ns,
1949 struct ldlm_request *dlm_req,
1950 struct ldlm_lock *lock)
1952 struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
1957 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1959 if (lock->l_glimpse_ast != NULL)
1960 rc = lock->l_glimpse_ast(lock, req);
1962 if (req->rq_repmsg != NULL) {
1965 req->rq_status = rc;
1969 lock_res_and_lock(lock);
1970 if (lock->l_granted_mode == LCK_PW &&
1971 !lock->l_readers && !lock->l_writers &&
1972 ktime_after(ktime_get(),
1973 ktime_add(lock->l_last_used,
1974 ktime_set(ns->ns_dirty_age_limit, 0)))) {
1975 unlock_res_and_lock(lock);
1977 /* For MDS glimpse it is always DOM lock, set corresponding
1978 * cancel_bits to perform lock convert if needed
1980 if (lock->l_resource->lr_type == LDLM_IBITS)
1981 ld->l_policy_data.l_inodebits.cancel_bits =
1983 if (ldlm_bl_to_thread_lock(ns, ld, lock))
1984 ldlm_handle_bl_callback(ns, ld, lock);
1989 unlock_res_and_lock(lock);
1990 LDLM_LOCK_RELEASE(lock);
1994 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1996 if (req->rq_no_reply)
1999 req->rq_status = rc;
2000 if (!req->rq_packed_final) {
2001 rc = lustre_pack_reply(req, 1, NULL, NULL);
2005 return ptlrpc_reply(req);
2008 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2009 enum ldlm_cancel_flags cancel_flags)
2011 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2014 spin_lock(&blp->blp_lock);
2015 if (blwi->blwi_lock &&
2016 ldlm_is_discard_data(blwi->blwi_lock)) {
2017 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2018 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2020 /* other blocking callbacks are added to the regular list */
2021 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2023 spin_unlock(&blp->blp_lock);
2025 wake_up(&blp->blp_waitq);
2027 /* can not check blwi->blwi_flags as blwi could be already freed in
2029 if (!(cancel_flags & LCF_ASYNC))
2030 wait_for_completion(&blwi->blwi_comp);
2035 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2036 struct ldlm_namespace *ns,
2037 struct ldlm_lock_desc *ld,
2038 struct list_head *cancels, int count,
2039 struct ldlm_lock *lock,
2040 enum ldlm_cancel_flags cancel_flags)
2042 init_completion(&blwi->blwi_comp);
2043 INIT_LIST_HEAD(&blwi->blwi_head);
2045 if (memory_pressure_get())
2046 blwi->blwi_mem_pressure = 1;
2049 blwi->blwi_flags = cancel_flags;
2051 blwi->blwi_ld = *ld;
2053 list_add(&blwi->blwi_head, cancels);
2054 list_del_init(cancels);
2055 blwi->blwi_count = count;
2057 blwi->blwi_lock = lock;
2062 * Queues a list of locks \a cancels containing \a count locks
2063 * for later processing by a blocking thread. If \a count is zero,
2064 * then the lock referenced as \a lock is queued instead.
2066 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2067 * If list addition fails an error is returned and caller is supposed to
2068 * call ->l_blocking_ast itself.
2070 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2071 struct ldlm_lock_desc *ld,
2072 struct ldlm_lock *lock,
2073 struct list_head *cancels, int count,
2074 enum ldlm_cancel_flags cancel_flags)
2078 if (cancels && count == 0)
2081 if (cancel_flags & LCF_ASYNC) {
2082 struct ldlm_bl_work_item *blwi;
2084 OBD_ALLOC(blwi, sizeof(*blwi));
2087 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2089 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2091 /* if it is synchronous call do minimum mem alloc, as it could
2092 * be triggered from kernel shrinker
2094 struct ldlm_bl_work_item blwi;
2096 memset(&blwi, 0, sizeof(blwi));
2097 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2098 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2103 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2104 struct ldlm_lock *lock)
2106 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2109 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2110 struct list_head *cancels, int count,
2111 enum ldlm_cancel_flags cancel_flags)
2113 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2116 int ldlm_bl_thread_wakeup(void)
2118 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2122 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2123 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2125 struct obd_device *obd = req->rq_export->exp_obd;
2132 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2134 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2136 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2138 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2141 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2143 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2145 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2148 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2151 /* We are responsible for swabbing contents of val */
2153 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2154 /* Pass it on to mdc (the "export" in this case) */
2155 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2157 sizeof(KEY_HSM_COPYTOOL_SEND),
2158 KEY_HSM_COPYTOOL_SEND,
2161 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2166 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2167 const char *msg, int rc,
2168 const struct lustre_handle *handle)
2170 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2171 "%s: [nid %s] [rc %d] [lock %#llx]",
2172 msg, libcfs_id2str(req->rq_peer), rc,
2173 handle ? handle->cookie : 0);
2174 if (req->rq_no_reply)
2175 CWARN("No reply was sent, maybe cause bug 21636.\n");
2177 CWARN("Send reply failed, maybe cause bug 21636.\n");
2180 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2181 static int ldlm_callback_handler(struct ptlrpc_request *req)
2183 struct ldlm_namespace *ns;
2184 struct ldlm_request *dlm_req;
2185 struct ldlm_lock *lock;
2189 /* Requests arrive in sender's byte order. The ptlrpc service
2190 * handler has already checked and, if necessary, byte-swapped the
2191 * incoming request message body, but I am responsible for the
2192 * message buffers. */
2194 /* do nothing for sec context finalize */
2195 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2198 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2200 if (req->rq_export == NULL) {
2201 rc = ldlm_callback_reply(req, -ENOTCONN);
2202 ldlm_callback_errmsg(req, "Operate on unconnected server",
2207 LASSERT(req->rq_export != NULL);
2208 LASSERT(req->rq_export->exp_obd != NULL);
2210 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2211 case LDLM_BL_CALLBACK:
2212 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2214 ldlm_callback_reply(req, -(int)cfs_fail_err);
2218 case LDLM_CP_CALLBACK:
2219 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2222 case LDLM_GL_CALLBACK:
2223 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2227 rc = ldlm_handle_setinfo(req);
2228 ldlm_callback_reply(req, rc);
2231 CERROR("unknown opcode %u\n",
2232 lustre_msg_get_opc(req->rq_reqmsg));
2233 ldlm_callback_reply(req, -EPROTO);
2237 ns = req->rq_export->exp_obd->obd_namespace;
2238 LASSERT(ns != NULL);
2240 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2242 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2243 if (dlm_req == NULL) {
2244 rc = ldlm_callback_reply(req, -EPROTO);
2245 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2250 /* Force a known safe race, send a cancel to the server for a lock
2251 * which the server has already started a blocking callback on. */
2252 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2253 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2254 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2256 CERROR("ldlm_cli_cancel: %d\n", rc);
2259 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2261 CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock "
2262 "disappeared\n", dlm_req->lock_handle[0].cookie);
2263 rc = ldlm_callback_reply(req, -EINVAL);
2264 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2265 &dlm_req->lock_handle[0]);
2269 if (ldlm_is_fail_loc(lock) &&
2270 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2271 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2273 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2274 lock_res_and_lock(lock);
2275 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2277 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2278 /* If somebody cancels lock and cache is already dropped,
2279 * or lock is failed before cp_ast received on client,
2280 * we can tell the server we have no lock. Otherwise, we
2281 * should send cancel after dropping the cache. */
2282 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2283 ldlm_is_failed(lock)) {
2284 LDLM_DEBUG(lock, "callback on lock %llx - lock disappeared",
2285 dlm_req->lock_handle[0].cookie);
2286 unlock_res_and_lock(lock);
2287 LDLM_LOCK_RELEASE(lock);
2288 rc = ldlm_callback_reply(req, -EINVAL);
2289 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2290 &dlm_req->lock_handle[0]);
2293 /* BL_AST locks are not needed in LRU.
2294 * Let ldlm_cancel_lru() be fast. */
2295 ldlm_lock_remove_from_lru(lock);
2296 ldlm_set_bl_ast(lock);
2298 unlock_res_and_lock(lock);
2300 /* We want the ost thread to get this reply so that it can respond
2301 * to ost requests (write cache writeback) that might be triggered
2304 * But we'd also like to be able to indicate in the reply that we're
2305 * cancelling right now, because it's unused, or have an intent result
2306 * in the reply, so we might have to push the responsibility for sending
2307 * the reply down into the AST handlers, alas. */
2309 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2310 case LDLM_BL_CALLBACK:
2311 CDEBUG(D_INODE, "blocking ast\n");
2312 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2313 if (!ldlm_is_cancel_on_block(lock)) {
2314 rc = ldlm_callback_reply(req, 0);
2315 if (req->rq_no_reply || rc)
2316 ldlm_callback_errmsg(req, "Normal process", rc,
2317 &dlm_req->lock_handle[0]);
2319 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2320 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2322 case LDLM_CP_CALLBACK:
2323 CDEBUG(D_INODE, "completion ast\n");
2324 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2325 ldlm_callback_reply(req, 0);
2326 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2328 case LDLM_GL_CALLBACK:
2329 CDEBUG(D_INODE, "glimpse ast\n");
2330 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2331 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2334 LBUG(); /* checked above */
2340 #ifdef HAVE_SERVER_SUPPORT
2342 * Main handler for canceld thread.
2344 * Separated into its own thread to avoid deadlocks.
2346 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2352 /* Requests arrive in sender's byte order. The ptlrpc service
2353 * handler has already checked and, if necessary, byte-swapped the
2354 * incoming request message body, but I am responsible for the
2355 * message buffers. */
2357 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2359 if (req->rq_export == NULL) {
2360 struct ldlm_request *dlm_req;
2362 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2363 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2364 libcfs_nid2str(req->rq_peer.nid),
2365 (unsigned long long)req->rq_arrival_time.tv_sec,
2366 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2368 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2369 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2370 dlm_req = req_capsule_client_get(&req->rq_pill,
2372 if (dlm_req != NULL)
2373 ldlm_lock_dump_handle(D_ERROR,
2374 &dlm_req->lock_handle[0]);
2376 ldlm_callback_reply(req, -ENOTCONN);
2380 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2381 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2383 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2384 CDEBUG(D_INODE, "cancel\n");
2385 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2386 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2387 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2389 rc = ldlm_handle_cancel(req);
2393 struct ldlm_request *dlm_req;
2395 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2396 CDEBUG(D_INODE, "convert\n");
2398 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2399 if (dlm_req == NULL) {
2400 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2401 rc = ldlm_callback_reply(req, -EPROTO);
2403 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2404 rc = ptlrpc_reply(req);
2409 CERROR("invalid opcode %d\n",
2410 lustre_msg_get_opc(req->rq_reqmsg));
2411 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2412 rc = ldlm_callback_reply(req, -EINVAL);
2418 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2419 struct ldlm_lock *lock)
2421 struct ldlm_request *dlm_req;
2422 struct lustre_handle lockh;
2428 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2429 if (dlm_req == NULL)
2432 ldlm_lock2handle(lock, &lockh);
2433 for (i = 0; i < dlm_req->lock_count; i++) {
2434 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2436 DEBUG_REQ(D_RPCTRACE, req,
2437 "Prio raised by lock %#llx.", lockh.cookie);
2446 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2448 struct ldlm_request *dlm_req;
2455 /* no prolong in recovery */
2456 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2459 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2460 if (dlm_req == NULL)
2463 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2464 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2465 (size - offsetof(struct ldlm_request, lock_handle)) /
2466 sizeof(struct lustre_handle) < dlm_req->lock_count)
2469 for (i = 0; i < dlm_req->lock_count; i++) {
2470 struct ldlm_lock *lock;
2472 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2476 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2478 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2479 LDLM_LOCK_PUT(lock);
2488 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2489 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2490 .hpreq_check = ldlm_cancel_hpreq_check,
2494 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2498 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2500 if (req->rq_export == NULL)
2503 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2504 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2505 req->rq_ops = &ldlm_cancel_hpreq_ops;
2506 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2507 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2508 req->rq_ops = &ldlm_cancel_hpreq_ops;
2513 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2514 struct hlist_node *hnode, void *data)
2517 struct list_head *rpc_list = data;
2518 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2520 lock_res_and_lock(lock);
2522 if (!ldlm_is_granted(lock)) {
2523 unlock_res_and_lock(lock);
2527 LASSERT(lock->l_resource);
2528 if (lock->l_resource->lr_type != LDLM_IBITS &&
2529 lock->l_resource->lr_type != LDLM_PLAIN) {
2530 unlock_res_and_lock(lock);
2534 if (ldlm_is_ast_sent(lock)) {
2535 unlock_res_and_lock(lock);
2539 LASSERT(lock->l_blocking_ast);
2540 LASSERT(!lock->l_blocking_lock);
2542 ldlm_set_ast_sent(lock);
2543 if (lock->l_export && lock->l_export->exp_lock_hash) {
2544 /* NB: it's safe to call cfs_hash_del() even lock isn't
2545 * in exp_lock_hash. */
2546 /* In the function below, .hs_keycmp resolves to
2547 * ldlm_export_lock_keycmp() */
2548 /* coverity[overrun-buffer-val] */
2549 cfs_hash_del(lock->l_export->exp_lock_hash,
2550 &lock->l_remote_handle, &lock->l_exp_hash);
2553 list_add_tail(&lock->l_rk_ast, rpc_list);
2554 LDLM_LOCK_GET(lock);
2556 unlock_res_and_lock(lock);
2560 void ldlm_revoke_export_locks(struct obd_export *exp)
2562 struct list_head rpc_list;
2565 INIT_LIST_HEAD(&rpc_list);
2566 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2567 ldlm_revoke_lock_cb, &rpc_list, 0);
2568 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2569 LDLM_WORK_REVOKE_AST);
2573 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2574 #endif /* HAVE_SERVER_SUPPORT */
2576 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2577 struct ldlm_bl_work_item **p_blwi,
2578 struct obd_export **p_exp)
2580 struct ldlm_bl_work_item *blwi = NULL;
2581 static unsigned int num_bl = 0;
2582 static unsigned int num_stale;
2583 int num_th = atomic_read(&blp->blp_num_threads);
2585 *p_exp = obd_stale_export_get();
2587 spin_lock(&blp->blp_lock);
2588 if (*p_exp != NULL) {
2589 if (num_th == 1 || ++num_stale < num_th) {
2590 spin_unlock(&blp->blp_lock);
2597 /* process a request from the blp_list at least every blp_num_threads */
2598 if (!list_empty(&blp->blp_list) &&
2599 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2600 blwi = list_entry(blp->blp_list.next,
2601 struct ldlm_bl_work_item, blwi_entry);
2603 if (!list_empty(&blp->blp_prio_list))
2604 blwi = list_entry(blp->blp_prio_list.next,
2605 struct ldlm_bl_work_item,
2609 if (++num_bl >= num_th)
2611 list_del(&blwi->blwi_entry);
2613 spin_unlock(&blp->blp_lock);
2616 if (*p_exp != NULL && *p_blwi != NULL) {
2617 obd_stale_export_put(*p_exp);
2621 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2624 /* This only contains temporary data until the thread starts */
2625 struct ldlm_bl_thread_data {
2626 struct ldlm_bl_pool *bltd_blp;
2627 struct completion bltd_comp;
2631 static int ldlm_bl_thread_main(void *arg);
2633 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2635 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2636 struct task_struct *task;
2638 init_completion(&bltd.bltd_comp);
2640 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2641 if (bltd.bltd_num >= blp->blp_max_threads) {
2642 atomic_dec(&blp->blp_num_threads);
2646 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2648 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2649 atomic_dec(&blp->blp_num_threads);
2653 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2656 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2657 bltd.bltd_num, PTR_ERR(task));
2658 atomic_dec(&blp->blp_num_threads);
2659 return PTR_ERR(task);
2661 wait_for_completion(&bltd.bltd_comp);
2666 /* Not fatal if racy and have a few too many threads */
2667 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2668 struct ldlm_bl_work_item *blwi)
2670 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2673 if (atomic_read(&blp->blp_busy_threads) <
2674 atomic_read(&blp->blp_num_threads))
2677 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2678 blwi->blwi_mem_pressure))
2684 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2685 struct ldlm_bl_work_item *blwi)
2689 if (blwi->blwi_ns == NULL)
2690 /* added by ldlm_cleanup() */
2691 RETURN(LDLM_ITER_STOP);
2693 if (blwi->blwi_mem_pressure)
2694 memory_pressure_set();
2696 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2698 if (blwi->blwi_count) {
2700 /* The special case when we cancel locks in lru
2701 * asynchronously, we pass the list of locks here.
2702 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2703 * canceled locally yet. */
2704 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2707 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2710 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2713 if (blwi->blwi_mem_pressure)
2714 memory_pressure_clr();
2716 if (blwi->blwi_flags & LCF_ASYNC)
2717 OBD_FREE(blwi, sizeof(*blwi));
2719 complete(&blwi->blwi_comp);
2725 * Cancel stale locks on export. Cancel blocked locks first.
2726 * If the given export has blocked locks, the next in the list may have
2727 * them too, thus cancel not blocked locks only if the current export has
2730 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2731 struct obd_export *exp)
2736 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2738 num = ldlm_export_cancel_blocked_locks(exp);
2740 ldlm_export_cancel_locks(exp);
2742 obd_stale_export_put(exp);
2749 * Main blocking requests processing thread.
2751 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2752 * This thread in the end ends up doing actual call to ->l_blocking_ast
2755 static int ldlm_bl_thread_main(void *arg)
2758 struct ldlm_bl_pool *blp;
2759 struct ldlm_bl_thread_data *bltd = arg;
2767 rc = lu_env_init(env, LCT_DT_THREAD);
2770 rc = lu_env_add(env);
2772 GOTO(out_env_fini, rc);
2774 blp = bltd->bltd_blp;
2776 complete(&bltd->bltd_comp);
2777 /* cannot use bltd after this, it is only on caller's stack */
2780 struct l_wait_info lwi = { 0 };
2781 struct ldlm_bl_work_item *blwi = NULL;
2782 struct obd_export *exp = NULL;
2785 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2788 l_wait_event_exclusive(blp->blp_waitq,
2789 ldlm_bl_get_work(blp, &blwi,
2792 atomic_inc(&blp->blp_busy_threads);
2794 if (ldlm_bl_thread_need_create(blp, blwi))
2795 /* discard the return value, we tried */
2796 ldlm_bl_thread_start(blp, true);
2799 rc = ldlm_bl_thread_exports(blp, exp);
2801 rc = ldlm_bl_thread_blwi(blp, blwi);
2803 atomic_dec(&blp->blp_busy_threads);
2805 if (rc == LDLM_ITER_STOP)
2808 /* If there are many namespaces, we will not sleep waiting for
2809 * work, and must do a cond_resched to avoid holding the CPU
2814 atomic_dec(&blp->blp_num_threads);
2815 complete(&blp->blp_comp);
2826 static int ldlm_setup(void);
2827 static int ldlm_cleanup(void);
2829 int ldlm_get_ref(void)
2833 mutex_lock(&ldlm_ref_mutex);
2834 if (++ldlm_refcount == 1) {
2839 mutex_unlock(&ldlm_ref_mutex);
2844 void ldlm_put_ref(void)
2847 mutex_lock(&ldlm_ref_mutex);
2848 if (ldlm_refcount == 1) {
2849 int rc = ldlm_cleanup();
2851 CERROR("ldlm_cleanup failed: %d\n", rc);
2857 mutex_unlock(&ldlm_ref_mutex);
2863 * Export handle<->lock hash operations.
2866 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2868 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2872 ldlm_export_lock_key(struct hlist_node *hnode)
2874 struct ldlm_lock *lock;
2876 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2877 return &lock->l_remote_handle;
2881 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2883 struct ldlm_lock *lock;
2885 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2886 lock->l_remote_handle = *(struct lustre_handle *)key;
2890 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2892 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2896 ldlm_export_lock_object(struct hlist_node *hnode)
2898 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2902 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2904 struct ldlm_lock *lock;
2906 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2907 LDLM_LOCK_GET(lock);
2911 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2913 struct ldlm_lock *lock;
2915 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2916 LDLM_LOCK_RELEASE(lock);
2919 static struct cfs_hash_ops ldlm_export_lock_ops = {
2920 .hs_hash = ldlm_export_lock_hash,
2921 .hs_key = ldlm_export_lock_key,
2922 .hs_keycmp = ldlm_export_lock_keycmp,
2923 .hs_keycpy = ldlm_export_lock_keycpy,
2924 .hs_object = ldlm_export_lock_object,
2925 .hs_get = ldlm_export_lock_get,
2926 .hs_put = ldlm_export_lock_put,
2927 .hs_put_locked = ldlm_export_lock_put,
2930 int ldlm_init_export(struct obd_export *exp)
2935 exp->exp_lock_hash =
2936 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2937 HASH_EXP_LOCK_CUR_BITS,
2938 HASH_EXP_LOCK_MAX_BITS,
2939 HASH_EXP_LOCK_BKT_BITS, 0,
2940 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2941 &ldlm_export_lock_ops,
2942 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2943 CFS_HASH_NBLK_CHANGE);
2945 if (!exp->exp_lock_hash)
2948 rc = ldlm_init_flock_export(exp);
2954 ldlm_destroy_export(exp);
2957 EXPORT_SYMBOL(ldlm_init_export);
2959 void ldlm_destroy_export(struct obd_export *exp)
2962 cfs_hash_putref(exp->exp_lock_hash);
2963 exp->exp_lock_hash = NULL;
2965 ldlm_destroy_flock_export(exp);
2968 EXPORT_SYMBOL(ldlm_destroy_export);
2970 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
2971 struct attribute *attr,
2974 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
2977 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
2978 struct attribute *attr,
2985 rc = kstrtoul(buffer, 10, &val);
2989 ldlm_cancel_unused_locks_before_replay = val;
2993 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
2995 static struct attribute *ldlm_attrs[] = {
2996 &lustre_attr_cancel_unused_locks_before_replay.attr,
3000 static struct attribute_group ldlm_attr_group = {
3001 .attrs = ldlm_attrs,
3004 static int ldlm_setup(void)
3006 static struct ptlrpc_service_conf conf;
3007 struct ldlm_bl_pool *blp = NULL;
3008 #ifdef HAVE_SERVER_SUPPORT
3009 struct task_struct *task;
3010 #endif /* HAVE_SERVER_SUPPORT */
3016 if (ldlm_state != NULL)
3019 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3020 if (ldlm_state == NULL)
3023 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3027 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3031 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3035 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3039 rc = ldlm_debugfs_setup();
3043 memset(&conf, 0, sizeof(conf));
3044 conf = (typeof(conf)) {
3045 .psc_name = "ldlm_cbd",
3046 .psc_watchdog_factor = 2,
3048 .bc_nbufs = LDLM_CLIENT_NBUFS,
3049 .bc_buf_size = LDLM_BUFSIZE,
3050 .bc_req_max_size = LDLM_MAXREQSIZE,
3051 .bc_rep_max_size = LDLM_MAXREPSIZE,
3052 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3053 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3056 .tc_thr_name = "ldlm_cb",
3057 .tc_thr_factor = LDLM_THR_FACTOR,
3058 .tc_nthrs_init = LDLM_NTHRS_INIT,
3059 .tc_nthrs_base = LDLM_NTHRS_BASE,
3060 .tc_nthrs_max = LDLM_NTHRS_MAX,
3061 .tc_nthrs_user = ldlm_num_threads,
3062 .tc_cpu_bind = ldlm_cpu_bind,
3063 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3066 .cc_pattern = ldlm_cpts,
3067 .cc_affinity = true,
3070 .so_req_handler = ldlm_callback_handler,
3073 ldlm_state->ldlm_cb_service = \
3074 ptlrpc_register_service(&conf, ldlm_svc_kset,
3075 ldlm_svc_debugfs_dir);
3076 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3077 CERROR("failed to start service\n");
3078 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3079 ldlm_state->ldlm_cb_service = NULL;
3083 #ifdef HAVE_SERVER_SUPPORT
3084 memset(&conf, 0, sizeof(conf));
3085 conf = (typeof(conf)) {
3086 .psc_name = "ldlm_canceld",
3087 .psc_watchdog_factor = 6,
3089 .bc_nbufs = LDLM_SERVER_NBUFS,
3090 .bc_buf_size = LDLM_BUFSIZE,
3091 .bc_req_max_size = LDLM_MAXREQSIZE,
3092 .bc_rep_max_size = LDLM_MAXREPSIZE,
3093 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3094 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3098 .tc_thr_name = "ldlm_cn",
3099 .tc_thr_factor = LDLM_THR_FACTOR,
3100 .tc_nthrs_init = LDLM_NTHRS_INIT,
3101 .tc_nthrs_base = LDLM_NTHRS_BASE,
3102 .tc_nthrs_max = LDLM_NTHRS_MAX,
3103 .tc_nthrs_user = ldlm_num_threads,
3104 .tc_cpu_bind = ldlm_cpu_bind,
3105 .tc_ctx_tags = LCT_MD_THREAD | \
3110 .cc_pattern = ldlm_cpts,
3111 .cc_affinity = true,
3114 .so_req_handler = ldlm_cancel_handler,
3115 .so_hpreq_handler = ldlm_hpreq_handler,
3118 ldlm_state->ldlm_cancel_service = \
3119 ptlrpc_register_service(&conf, ldlm_svc_kset,
3120 ldlm_svc_debugfs_dir);
3121 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3122 CERROR("failed to start service\n");
3123 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3124 ldlm_state->ldlm_cancel_service = NULL;
3127 #endif /* HAVE_SERVER_SUPPORT */
3129 OBD_ALLOC(blp, sizeof(*blp));
3131 GOTO(out, rc = -ENOMEM);
3132 ldlm_state->ldlm_bl_pool = blp;
3134 spin_lock_init(&blp->blp_lock);
3135 INIT_LIST_HEAD(&blp->blp_list);
3136 INIT_LIST_HEAD(&blp->blp_prio_list);
3137 init_waitqueue_head(&blp->blp_waitq);
3138 atomic_set(&blp->blp_num_threads, 0);
3139 atomic_set(&blp->blp_busy_threads, 0);
3141 if (ldlm_num_threads == 0) {
3142 blp->blp_min_threads = LDLM_NTHRS_INIT;
3143 blp->blp_max_threads = LDLM_NTHRS_MAX;
3145 blp->blp_min_threads = blp->blp_max_threads = \
3146 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3150 for (i = 0; i < blp->blp_min_threads; i++) {
3151 rc = ldlm_bl_thread_start(blp, false);
3156 #ifdef HAVE_SERVER_SUPPORT
3157 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3160 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3164 wait_event(expired_lock_wait_queue,
3165 expired_lock_thread_state == ELT_READY);
3166 #endif /* HAVE_SERVER_SUPPORT */
3168 rc = ldlm_pools_init();
3170 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3174 rc = ldlm_reclaim_setup();
3176 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3186 static int ldlm_cleanup(void)
3190 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3191 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3192 CERROR("ldlm still has namespaces; clean these up first.\n");
3193 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3194 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3198 ldlm_reclaim_cleanup();
3201 if (ldlm_state->ldlm_bl_pool != NULL) {
3202 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3204 while (atomic_read(&blp->blp_num_threads) > 0) {
3205 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3207 init_completion(&blp->blp_comp);
3209 spin_lock(&blp->blp_lock);
3210 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3211 wake_up(&blp->blp_waitq);
3212 spin_unlock(&blp->blp_lock);
3214 wait_for_completion(&blp->blp_comp);
3217 OBD_FREE(blp, sizeof(*blp));
3220 if (ldlm_state->ldlm_cb_service != NULL)
3221 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3222 #ifdef HAVE_SERVER_SUPPORT
3223 if (ldlm_state->ldlm_cancel_service != NULL)
3224 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3228 kset_unregister(ldlm_ns_kset);
3230 kset_unregister(ldlm_svc_kset);
3232 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3233 kobject_put(ldlm_kobj);
3236 ldlm_debugfs_cleanup();
3238 #ifdef HAVE_SERVER_SUPPORT
3239 if (expired_lock_thread_state != ELT_STOPPED) {
3240 expired_lock_thread_state = ELT_TERMINATE;
3241 wake_up(&expired_lock_wait_queue);
3242 wait_event(expired_lock_wait_queue,
3243 expired_lock_thread_state == ELT_STOPPED);
3247 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3255 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3256 sizeof(struct ldlm_resource), 0,
3257 SLAB_HWCACHE_ALIGN, NULL);
3258 if (ldlm_resource_slab == NULL)
3261 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3262 sizeof(struct ldlm_lock), 0,
3263 SLAB_HWCACHE_ALIGN, NULL);
3264 if (ldlm_lock_slab == NULL)
3267 ldlm_interval_slab = kmem_cache_create("interval_node",
3268 sizeof(struct ldlm_interval),
3269 0, SLAB_HWCACHE_ALIGN, NULL);
3270 if (ldlm_interval_slab == NULL)
3273 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3274 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3275 0, SLAB_HWCACHE_ALIGN, NULL);
3276 if (ldlm_interval_tree_slab == NULL)
3279 #ifdef HAVE_SERVER_SUPPORT
3280 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3281 sizeof(struct ldlm_ibits_node),
3282 0, SLAB_HWCACHE_ALIGN, NULL);
3283 if (ldlm_inodebits_slab == NULL)
3284 goto out_interval_tree;
3286 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3287 sizeof(struct ldlm_glimpse_work),
3289 if (ldlm_glimpse_work_kmem == NULL)
3293 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3294 class_export_dump_hook = ldlm_dump_export_locks;
3297 #ifdef HAVE_SERVER_SUPPORT
3299 kmem_cache_destroy(ldlm_inodebits_slab);
3301 kmem_cache_destroy(ldlm_interval_tree_slab);
3304 kmem_cache_destroy(ldlm_interval_slab);
3306 kmem_cache_destroy(ldlm_lock_slab);
3308 kmem_cache_destroy(ldlm_resource_slab);
3313 void ldlm_exit(void)
3316 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3317 kmem_cache_destroy(ldlm_resource_slab);
3319 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3320 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3321 * so that ldlm_lock_free() get a chance to be called.
3324 kmem_cache_destroy(ldlm_lock_slab);
3325 kmem_cache_destroy(ldlm_interval_slab);
3326 kmem_cache_destroy(ldlm_interval_tree_slab);
3327 #ifdef HAVE_SERVER_SUPPORT
3328 kmem_cache_destroy(ldlm_inodebits_slab);
3329 kmem_cache_destroy(ldlm_glimpse_work_kmem);