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 <libcfs/linux/linux-mem.h>
44 #include <lustre_errno.h>
45 #include <lustre_dlm.h>
46 #include <obd_class.h>
47 #include "ldlm_internal.h"
49 static int ldlm_num_threads;
50 module_param(ldlm_num_threads, int, 0444);
51 MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start");
53 static unsigned int ldlm_cpu_bind = 1;
54 module_param(ldlm_cpu_bind, uint, 0444);
55 MODULE_PARM_DESC(ldlm_cpu_bind,
56 "bind DLM service threads to particular CPU partitions");
58 static char *ldlm_cpts;
59 module_param(ldlm_cpts, charp, 0444);
60 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
62 static DEFINE_MUTEX(ldlm_ref_mutex);
63 static int ldlm_refcount;
65 struct kobject *ldlm_kobj;
66 struct kset *ldlm_ns_kset;
67 struct kset *ldlm_svc_kset;
71 static struct ldlm_state *ldlm_state;
74 * timeout for initial callback (AST) reply (bz10399)
75 * Due to having to send a 32 bit time value over the
76 * wire return it as timeout_t instead of time64_t
78 static inline timeout_t ldlm_get_rq_timeout(void)
81 timeout_t timeout = min(ldlm_timeout, obd_timeout / 3);
83 return timeout < 1 ? 1 : timeout;
90 * blp_prio_list is used for callbacks that should be handled
91 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
94 struct list_head blp_prio_list;
97 * blp_list is used for all other callbacks which are likely
98 * to take longer to process.
100 struct list_head blp_list;
102 wait_queue_head_t blp_waitq;
103 struct completion blp_comp;
104 atomic_t blp_num_threads;
105 atomic_t blp_busy_threads;
110 struct ldlm_bl_work_item {
111 struct list_head blwi_entry;
112 struct ldlm_namespace *blwi_ns;
113 struct ldlm_lock_desc blwi_ld;
114 struct ldlm_lock *blwi_lock;
115 struct list_head blwi_head;
117 struct completion blwi_comp;
118 enum ldlm_cancel_flags blwi_flags;
119 int blwi_mem_pressure;
122 #ifdef HAVE_SERVER_SUPPORT
125 * Protects both waiting_locks_list and expired_lock_thread.
127 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
130 * List for contended locks.
132 * As soon as a lock is contended, it gets placed on this list and
133 * expected time to get a response is filled in the lock. A special
134 * thread walks the list looking for locks that should be released and
135 * schedules client evictions for those that have not been released in
138 * All access to it should be under waiting_locks_spinlock.
140 static LIST_HEAD(waiting_locks_list);
141 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
142 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
150 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
151 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
152 static int expired_lock_dump;
153 static LIST_HEAD(expired_lock_list);
155 static int ldlm_lock_busy(struct ldlm_lock *lock);
156 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout);
157 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout);
159 static inline int have_expired_locks(void)
164 spin_lock_bh(&waiting_locks_spinlock);
165 need_to_run = !list_empty(&expired_lock_list);
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 struct list_head *expired = &expired_lock_list;
181 expired_lock_thread_state = ELT_READY;
182 wake_up(&expired_lock_wait_queue);
185 wait_event_idle(expired_lock_wait_queue,
186 have_expired_locks() ||
187 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");
223 * release extra ref grabbed by
224 * ldlm_add_waiting_lock() or
227 LDLM_LOCK_RELEASE(lock);
231 if (ldlm_is_destroyed(lock)) {
233 * release the lock refcount where
234 * waiting_locks_callback() founds
236 LDLM_LOCK_RELEASE(lock);
239 export = class_export_lock_get(lock->l_export, lock);
240 spin_unlock_bh(&waiting_locks_spinlock);
242 /* Check if we need to prolong timeout */
243 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
244 lock->l_callback_timestamp != 0 && /* not AST error */
245 ldlm_lock_busy(lock)) {
246 LDLM_DEBUG(lock, "prolong the busy lock");
247 lock_res_and_lock(lock);
248 ldlm_add_waiting_lock(lock,
249 ldlm_bl_timeout(lock) >> 1);
250 unlock_res_and_lock(lock);
252 spin_lock_bh(&export->exp_bl_list_lock);
253 list_del_init(&lock->l_exp_list);
254 spin_unlock_bh(&export->exp_bl_list_lock);
257 "lock callback timer expired after %llds: evicting client at %s ",
258 ktime_get_real_seconds() -
260 obd_export_nid2str(export));
261 ldlm_lock_to_ns(lock)->ns_timeouts++;
263 class_fail_export(export);
265 class_export_lock_put(export, lock);
267 * release extra ref grabbed by ldlm_add_waiting_lock()
268 * or ldlm_failed_ast()
270 LDLM_LOCK_RELEASE(lock);
272 spin_lock_bh(&waiting_locks_spinlock);
274 spin_unlock_bh(&waiting_locks_spinlock);
276 if (do_dump && obd_dump_on_eviction) {
277 CERROR("dump the log upon eviction\n");
278 libcfs_debug_dumplog();
281 if (expired_lock_thread_state == ELT_TERMINATE)
285 expired_lock_thread_state = ELT_STOPPED;
286 wake_up(&expired_lock_wait_queue);
291 * Check if there is a request in the export request list
292 * which prevents the lock canceling.
294 static int ldlm_lock_busy(struct ldlm_lock *lock)
296 struct ptlrpc_request *req;
301 if (lock->l_export == NULL)
304 spin_lock(&lock->l_export->exp_rpc_lock);
305 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
307 if (req->rq_ops->hpreq_lock_match) {
308 match = req->rq_ops->hpreq_lock_match(req, lock);
313 spin_unlock(&lock->l_export->exp_rpc_lock);
317 /* This is called from within a timer interrupt and cannot schedule */
318 static void waiting_locks_callback(TIMER_DATA_TYPE unused)
320 struct ldlm_lock *lock;
323 spin_lock_bh(&waiting_locks_spinlock);
324 while (!list_empty(&waiting_locks_list)) {
325 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
327 if (lock->l_callback_timestamp > ktime_get_seconds() ||
328 lock->l_req_mode == LCK_GROUP)
332 * no needs to take an extra ref on the lock since it was in
333 * the waiting_locks_list and ldlm_add_waiting_lock()
334 * already grabbed a ref
336 list_move(&lock->l_pending_chain, &expired_lock_list);
340 if (!list_empty(&expired_lock_list)) {
341 if (obd_dump_on_timeout && need_dump)
342 expired_lock_dump = __LINE__;
344 wake_up(&expired_lock_wait_queue);
348 * Make sure the timer will fire again if we have any locks
351 if (!list_empty(&waiting_locks_list)) {
352 time64_t now = ktime_get_seconds();
355 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
357 if (lock->l_callback_timestamp - now > 0)
358 delta = lock->l_callback_timestamp - now;
359 mod_timer(&waiting_locks_timer,
360 jiffies + cfs_time_seconds(delta));
362 spin_unlock_bh(&waiting_locks_spinlock);
366 * Add lock to the list of contended locks.
368 * Indicate that we're waiting for a client to call us back cancelling a given
369 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
370 * timer to fire appropriately. (We round up to the next second, to avoid
371 * floods of timer firings during periods of high lock contention and traffic).
372 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
373 * if it has been added to the waiting list (1 is returned).
375 * Called with the namespace lock held.
377 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t delay)
379 unsigned long timeout_jiffies = jiffies;
380 time64_t now = ktime_get_seconds();
384 if (!list_empty(&lock->l_pending_chain))
387 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
388 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
391 deadline = now + delay;
392 if (likely(deadline > lock->l_callback_timestamp))
393 lock->l_callback_timestamp = deadline;
395 timeout = clamp_t(timeout_t, lock->l_callback_timestamp - now,
397 timeout_jiffies += cfs_time_seconds(timeout);
399 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
400 !timer_pending(&waiting_locks_timer))
401 mod_timer(&waiting_locks_timer, timeout_jiffies);
404 * if the new lock has a shorter timeout than something earlier on
405 * the list, we'll wait the longer amount of time; no big deal.
408 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
412 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
414 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
415 if (list_empty(&lock->l_exp_list)) {
416 if (!ldlm_is_granted(lock))
417 list_add_tail(&lock->l_exp_list,
418 &lock->l_export->exp_bl_list);
420 list_add(&lock->l_exp_list,
421 &lock->l_export->exp_bl_list);
423 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
426 * A blocked lock is added. Adjust the position in
427 * the stale list if the export is in the list.
428 * If export is stale and not in the list - it is being
429 * processed and will be placed on the right position
430 * on obd_stale_export_put().
432 if (!list_empty(&lock->l_export->exp_stale_list))
433 obd_stale_export_adjust(lock->l_export);
436 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
440 /* NB: must be called with hold of lock_res_and_lock() */
441 LASSERT(ldlm_is_res_locked(lock));
442 LASSERT(!ldlm_is_cancel_on_block(lock));
445 * Do not put cross-MDT lock in the waiting list, since we
446 * will not evict it due to timeout for now
448 if (lock->l_export != NULL &&
449 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
452 spin_lock_bh(&waiting_locks_spinlock);
453 if (ldlm_is_cancel(lock)) {
454 spin_unlock_bh(&waiting_locks_spinlock);
458 if (ldlm_is_destroyed(lock)) {
459 static time64_t next;
461 spin_unlock_bh(&waiting_locks_spinlock);
462 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
463 if (ktime_get_seconds() > next) {
464 next = ktime_get_seconds() + 14400;
465 libcfs_debug_dumpstack(NULL);
470 ldlm_set_waited(lock);
471 lock->l_blast_sent = ktime_get_real_seconds();
472 ret = __ldlm_add_waiting_lock(lock, timeout);
475 * grab ref on the lock if it has been added to the
480 spin_unlock_bh(&waiting_locks_spinlock);
483 ldlm_add_blocked_lock(lock);
485 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
486 ret == 0 ? "not re-" : "", timeout,
487 AT_OFF ? "off" : "on");
492 * Remove a lock from the pending list, likely because it had its cancellation
493 * callback arrive without incident. This adjusts the lock-timeout timer if
494 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
495 * As done by ldlm_del_waiting_lock(), the caller must release the lock
496 * reference when the lock is removed from any list (1 is returned).
498 * Called with namespace lock held.
500 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
502 struct list_head *list_next;
504 if (list_empty(&lock->l_pending_chain))
507 list_next = lock->l_pending_chain.next;
508 if (lock->l_pending_chain.prev == &waiting_locks_list) {
509 /* Removing the head of the list, adjust timer. */
510 if (list_next == &waiting_locks_list) {
511 /* No more, just cancel. */
512 del_timer(&waiting_locks_timer);
514 time64_t now = ktime_get_seconds();
515 struct ldlm_lock *next;
518 next = list_entry(list_next, struct ldlm_lock,
520 if (next->l_callback_timestamp - now > 0)
521 delta = lock->l_callback_timestamp - now;
523 mod_timer(&waiting_locks_timer,
524 jiffies + cfs_time_seconds(delta));
527 list_del_init(&lock->l_pending_chain);
532 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
536 if (lock->l_export == NULL) {
537 /* We don't have a "waiting locks list" on clients. */
538 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
542 spin_lock_bh(&waiting_locks_spinlock);
543 ret = __ldlm_del_waiting_lock(lock);
544 ldlm_clear_waited(lock);
545 spin_unlock_bh(&waiting_locks_spinlock);
547 /* remove the lock out of export blocking list */
548 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
549 list_del_init(&lock->l_exp_list);
550 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
554 * release lock ref if it has indeed been removed
557 LDLM_LOCK_RELEASE(lock);
560 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
565 * Prolong the contended lock waiting time.
567 * Called with namespace lock held.
569 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
571 if (lock->l_export == NULL) {
572 /* We don't have a "waiting locks list" on clients. */
573 LDLM_DEBUG(lock, "client lock: no-op");
577 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
578 /* We don't have a "waiting locks list" on OSP. */
579 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
583 spin_lock_bh(&waiting_locks_spinlock);
585 if (list_empty(&lock->l_pending_chain)) {
586 spin_unlock_bh(&waiting_locks_spinlock);
587 LDLM_DEBUG(lock, "wasn't waiting");
592 * we remove/add the lock to the waiting list, so no needs to
593 * release/take a lock reference
595 __ldlm_del_waiting_lock(lock);
596 __ldlm_add_waiting_lock(lock, timeout);
597 spin_unlock_bh(&waiting_locks_spinlock);
599 LDLM_DEBUG(lock, "refreshed");
602 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
604 #else /* HAVE_SERVER_SUPPORT */
606 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
611 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
616 #endif /* !HAVE_SERVER_SUPPORT */
618 #ifdef HAVE_SERVER_SUPPORT
621 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
622 * lock cancel, and their replies). Used for lock callback timeout and AST
625 * \param[in] lock lock which is getting the blocking callback
627 * \retval timeout in seconds to wait for the client reply
629 timeout_t ldlm_bl_timeout(struct ldlm_lock *lock)
634 return obd_timeout / 2;
637 * Since these are non-updating timeouts, we should be conservative.
638 * Take more than usually, 150%
639 * It would be nice to have some kind of "early reply" mechanism for
640 * lock callbacks too...
642 timeout = at_get(&lock->l_export->exp_bl_lock_at);
643 return max_t(timeout_t, timeout + (timeout >> 1),
644 (timeout_t)ldlm_enqueue_min);
646 EXPORT_SYMBOL(ldlm_bl_timeout);
649 * Perform lock cleanup if AST sending failed.
651 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
652 const char *ast_type)
654 LCONSOLE_ERROR_MSG(0x138,
655 "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
656 lock->l_export->exp_obd->obd_name,
657 obd_export_nid2str(lock->l_export), ast_type, rc);
659 if (obd_dump_on_timeout)
660 libcfs_debug_dumplog();
661 spin_lock_bh(&waiting_locks_spinlock);
662 if (__ldlm_del_waiting_lock(lock) == 0)
664 * the lock was not in any list, grab an extra ref before adding
665 * the lock to the expired list
668 /* differentiate it from expired locks */
669 lock->l_callback_timestamp = 0;
670 list_add(&lock->l_pending_chain, &expired_lock_list);
671 wake_up(&expired_lock_wait_queue);
672 spin_unlock_bh(&waiting_locks_spinlock);
676 * Perform lock cleanup if AST reply came with error.
678 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
679 struct ptlrpc_request *req, int rc,
680 const char *ast_type)
682 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
684 if (!req->rq_replied || (rc && rc != -EINVAL)) {
685 if (ldlm_is_cancel(lock)) {
687 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
688 ast_type, req, req->rq_xid,
689 libcfs_nid2str(peer.nid));
690 ldlm_lock_cancel(lock);
692 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
694 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
696 * Upon umount process the AST fails because cannot be
697 * sent. This shouldn't lead to the client eviction.
698 * -ENODEV error is returned by ptl_send_rpc() for
699 * new request in such import.
700 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
701 * if imp_invalid is set or obd_no_recov.
702 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
703 * in ptlrpc_import_delay_req() as well with -EIO code.
704 * In all such cases errors are ignored.
707 "%s AST can't be sent due to a server %s failure or umount process: rc = %d\n",
709 req->rq_import->imp_obd->obd_name, rc);
712 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
713 libcfs_nid2str(peer.nid),
714 req->rq_replied ? "returned error from" :
715 "failed to reply to",
716 ast_type, req, req->rq_xid,
717 (req->rq_repmsg != NULL) ?
718 lustre_msg_get_status(req->rq_repmsg) : 0,
720 ldlm_failed_ast(lock, rc, ast_type);
726 struct ldlm_resource *res = lock->l_resource;
729 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
730 libcfs_nid2str(peer.nid),
732 lustre_msg_get_status(req->rq_repmsg) : -1,
733 ast_type, req, req->rq_xid);
736 * update lvbo to return proper attributes.
739 ldlm_resource_getref(res);
740 ldlm_lvbo_update(res, lock, NULL, 1);
741 ldlm_resource_putref(res);
743 ldlm_lock_cancel(lock);
750 static int ldlm_cb_interpret(const struct lu_env *env,
751 struct ptlrpc_request *req, void *args, int rc)
753 struct ldlm_cb_async_args *ca = args;
754 struct ldlm_lock *lock = ca->ca_lock;
755 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
759 LASSERT(lock != NULL);
762 case LDLM_GL_CALLBACK:
764 * Update the LVB from disk if the AST failed
765 * (this is a legal race)
767 * - Glimpse callback of local lock just returns
768 * -ELDLM_NO_LOCK_DATA.
769 * - Glimpse callback of remote lock might return
770 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
772 if (unlikely(arg->gl_interpret_reply)) {
773 rc = arg->gl_interpret_reply(NULL, req, args, rc);
774 } else if (rc == -ELDLM_NO_LOCK_DATA) {
776 "lost race - client has a lock but no inode");
777 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
778 } else if (rc != 0) {
779 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
781 rc = ldlm_lvbo_update(lock->l_resource,
785 case LDLM_BL_CALLBACK:
787 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
789 case LDLM_CP_CALLBACK:
791 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
794 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
799 /* release extra reference taken in ldlm_ast_fini() */
800 LDLM_LOCK_RELEASE(lock);
803 atomic_inc(&arg->restart);
808 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
810 struct ldlm_cb_async_args *ca = data;
811 struct ldlm_lock *lock = ca->ca_lock;
813 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
816 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
817 struct ldlm_cb_set_arg *arg,
818 struct ldlm_lock *lock,
825 if (unlikely(instant_cancel)) {
826 rc = ptl_send_rpc(req, 1);
827 ptlrpc_req_finished(req);
829 atomic_inc(&arg->restart);
832 ptlrpc_set_add_req(arg->set, req);
839 * Check if there are requests in the export request list which prevent
840 * the lock canceling and make these requests high priority ones.
842 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
844 struct ptlrpc_request *req;
848 if (lock->l_export == NULL) {
849 LDLM_DEBUG(lock, "client lock: no-op");
853 spin_lock(&lock->l_export->exp_rpc_lock);
854 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
857 * Do not process requests that were not yet added to there
858 * incoming queue or were already removed from there for
859 * processing. We evaluate ptlrpc_nrs_req_can_move() without
860 * holding svcpt->scp_req_lock, and then redo the check with
861 * the lock held once we need to obtain a reliable result.
863 if (ptlrpc_nrs_req_can_move(req) &&
864 req->rq_ops->hpreq_lock_match &&
865 req->rq_ops->hpreq_lock_match(req, lock))
866 ptlrpc_nrs_req_hp_move(req);
868 spin_unlock(&lock->l_export->exp_rpc_lock);
873 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
874 * enqueued server lock conflicts with given one.
876 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
877 * to wait for client response.
879 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
880 struct ldlm_lock_desc *desc,
881 void *data, int flag)
883 struct ldlm_cb_async_args *ca;
884 struct ldlm_cb_set_arg *arg = data;
885 struct ldlm_request *body;
886 struct ptlrpc_request *req;
887 int instant_cancel = 0;
892 if (flag == LDLM_CB_CANCELING)
893 /* Don't need to do anything here. */
896 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
897 LDLM_DEBUG(lock, "dropping BL AST");
902 LASSERT(data != NULL);
903 if (lock->l_export->exp_obd->obd_recovering != 0)
904 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
906 ldlm_lock_reorder_req(lock);
908 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
909 &RQF_LDLM_BL_CALLBACK,
910 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
914 ca = ptlrpc_req_async_args(ca, req);
915 ca->ca_set_arg = arg;
918 req->rq_interpret_reply = ldlm_cb_interpret;
920 lock_res_and_lock(lock);
921 if (ldlm_is_destroyed(lock)) {
922 /* What's the point? */
923 unlock_res_and_lock(lock);
924 ptlrpc_req_finished(req);
928 if (!ldlm_is_granted(lock)) {
930 * this blocking AST will be communicated as part of the
931 * completion AST instead
933 ldlm_add_blocked_lock(lock);
934 ldlm_set_waited(lock);
935 unlock_res_and_lock(lock);
937 ptlrpc_req_finished(req);
938 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
942 if (ldlm_is_cancel_on_block(lock))
945 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
946 body->lock_handle[0] = lock->l_remote_handle;
947 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
948 body->lock_desc = *desc;
949 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
951 LDLM_DEBUG(lock, "server preparing blocking AST");
953 ptlrpc_request_set_replen(req);
954 ldlm_set_cbpending(lock);
955 if (instant_cancel) {
956 unlock_res_and_lock(lock);
957 ldlm_lock_cancel(lock);
959 req->rq_no_resend = 1;
961 LASSERT(ldlm_is_granted(lock));
962 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
963 unlock_res_and_lock(lock);
965 /* Do not resend after lock callback timeout */
966 req->rq_delay_limit = ldlm_bl_timeout(lock);
967 req->rq_resend_cb = ldlm_update_resend;
970 req->rq_send_state = LUSTRE_IMP_FULL;
971 /* ptlrpc_request_alloc_pack already set timeout */
973 req->rq_timeout = ldlm_get_rq_timeout();
975 if (lock->l_export && lock->l_export->exp_nid_stats &&
976 lock->l_export->exp_nid_stats->nid_ldlm_stats)
977 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
978 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
980 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
986 * ->l_completion_ast callback for a remote lock in server namespace.
988 * Sends AST to the client notifying it of lock granting. If initial
989 * lock response was not sent yet, instead of sending another RPC, just
990 * mark the lock as granted and client will understand
992 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
994 struct ldlm_cb_set_arg *arg = data;
995 struct ldlm_request *body;
996 struct ptlrpc_request *req;
997 struct ldlm_cb_async_args *ca;
998 int instant_cancel = 0;
1004 LASSERT(lock != NULL);
1005 LASSERT(data != NULL);
1007 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1008 LDLM_DEBUG(lock, "dropping CP AST");
1012 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1013 &RQF_LDLM_CP_CALLBACK);
1017 /* server namespace, doesn't need lock */
1018 lvb_len = ldlm_lvbo_size(lock);
1020 * LU-3124 & LU-2187: to not return layout in completion AST because
1021 * it may deadlock for LU-2187, or client may not have enough space
1022 * for large layout. The layout will be returned to client with an
1023 * extra RPC to fetch xattr.lov
1025 if (ldlm_has_layout(lock))
1028 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1029 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1031 ptlrpc_request_free(req);
1035 ca = ptlrpc_req_async_args(ca, req);
1036 ca->ca_set_arg = arg;
1039 req->rq_interpret_reply = ldlm_cb_interpret;
1040 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1042 body->lock_handle[0] = lock->l_remote_handle;
1043 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
1044 body->lock_flags = ldlm_flags_to_wire(flags);
1045 ldlm_lock2desc(lock, &body->lock_desc);
1047 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1048 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1051 * We still need to send the RPC to wake up the blocked
1052 * enqueue thread on the client.
1054 * Consider old client, there is no better way to notify
1055 * the failure, just zero-sized the LVB, then the client
1056 * will fail out as "-EPROTO".
1058 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1062 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1067 LDLM_DEBUG(lock, "server preparing completion AST");
1069 ptlrpc_request_set_replen(req);
1071 req->rq_send_state = LUSTRE_IMP_FULL;
1072 /* ptlrpc_request_pack already set timeout */
1074 req->rq_timeout = ldlm_get_rq_timeout();
1076 /* We only send real blocking ASTs after the lock is granted */
1077 lock_res_and_lock(lock);
1078 if (ldlm_is_ast_sent(lock)) {
1079 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1080 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1081 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1085 * We might get here prior to ldlm_handle_enqueue setting
1086 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1087 * into waiting list, but this is safe and similar code in
1088 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1089 * that would not only cancel the lock, but will also remove
1090 * it from waiting list
1092 if (ldlm_is_cancel_on_block(lock)) {
1093 unlock_res_and_lock(lock);
1094 ldlm_lock_cancel(lock);
1097 req->rq_no_resend = 1;
1099 lock_res_and_lock(lock);
1101 /* start the lock-timeout clock */
1102 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1103 /* Do not resend after lock callback timeout */
1104 req->rq_delay_limit = ldlm_bl_timeout(lock);
1105 req->rq_resend_cb = ldlm_update_resend;
1108 unlock_res_and_lock(lock);
1110 if (lock->l_export && lock->l_export->exp_nid_stats &&
1111 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1112 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1113 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1115 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1117 RETURN(lvb_len < 0 ? lvb_len : rc);
1121 * Server side ->l_glimpse_ast handler for client locks.
1123 * Sends glimpse AST to the client and waits for reply. Then updates
1124 * lvbo with the result.
1126 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1128 struct ldlm_cb_set_arg *arg = data;
1129 struct ldlm_request *body;
1130 struct ptlrpc_request *req;
1131 struct ldlm_cb_async_args *ca;
1133 struct req_format *req_fmt;
1137 LASSERT(lock != NULL);
1139 if (arg->gl_desc != NULL)
1140 /* There is a glimpse descriptor to pack */
1141 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1143 req_fmt = &RQF_LDLM_GL_CALLBACK;
1145 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1146 req_fmt, LUSTRE_DLM_VERSION,
1152 if (arg->gl_desc != NULL) {
1153 /* copy the GL descriptor */
1154 union ldlm_gl_desc *desc;
1156 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1157 *desc = *arg->gl_desc;
1160 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1161 body->lock_handle[0] = lock->l_remote_handle;
1162 ldlm_lock2desc(lock, &body->lock_desc);
1164 ca = ptlrpc_req_async_args(ca, req);
1165 ca->ca_set_arg = arg;
1168 /* server namespace, doesn't need lock */
1169 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1170 ldlm_lvbo_size(lock));
1171 ptlrpc_request_set_replen(req);
1173 req->rq_send_state = LUSTRE_IMP_FULL;
1174 /* ptlrpc_request_alloc_pack already set timeout */
1176 req->rq_timeout = ldlm_get_rq_timeout();
1178 req->rq_interpret_reply = ldlm_cb_interpret;
1180 if (lock->l_export && lock->l_export->exp_nid_stats) {
1181 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1183 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1184 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1187 rc = ldlm_ast_fini(req, arg, lock, 0);
1191 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1193 int ldlm_glimpse_locks(struct ldlm_resource *res,
1194 struct list_head *gl_work_list)
1200 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1202 if (rc == -ERESTART)
1203 ldlm_reprocess_all(res, NULL);
1207 EXPORT_SYMBOL(ldlm_glimpse_locks);
1209 /* return LDLM lock associated with a lock callback request */
1210 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1212 struct ldlm_cb_async_args *ca;
1213 struct ldlm_lock *lock;
1217 ca = ptlrpc_req_async_args(ca, req);
1220 RETURN(ERR_PTR(-EFAULT));
1224 EXPORT_SYMBOL(ldlm_request_lock);
1227 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1228 * service threads to carry out client lock enqueueing requests.
1230 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1231 struct ptlrpc_request *req,
1232 const struct ldlm_request *dlm_req,
1233 const struct ldlm_callback_suite *cbs)
1235 struct ldlm_reply *dlm_rep;
1237 enum ldlm_error err = ELDLM_OK;
1238 struct ldlm_lock *lock = NULL;
1239 void *cookie = NULL;
1241 struct ldlm_resource *res = NULL;
1242 const struct lu_env *env = req->rq_svc_thread->t_env;
1246 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1248 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1249 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1251 LASSERT(req->rq_export);
1253 /* for intent enqueue the stat will be updated inside intent policy */
1254 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1255 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1256 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1258 if (req->rq_export && req->rq_export->exp_nid_stats &&
1259 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1260 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1261 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1263 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1264 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1265 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1266 dlm_req->lock_desc.l_resource.lr_type);
1267 GOTO(out, rc = -EFAULT);
1270 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1271 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1272 dlm_req->lock_desc.l_req_mode &
1273 (dlm_req->lock_desc.l_req_mode-1))) {
1274 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1275 dlm_req->lock_desc.l_req_mode);
1276 GOTO(out, rc = -EFAULT);
1279 if (unlikely((flags & LDLM_FL_REPLAY) ||
1280 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1281 /* Find an existing lock in the per-export lock hash */
1283 * In the function below, .hs_keycmp resolves to
1284 * ldlm_export_lock_keycmp()
1286 /* coverity[overrun-buffer-val] */
1287 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1288 (void *)&dlm_req->lock_handle[0]);
1290 DEBUG_REQ(D_DLMTRACE, req,
1291 "found existing lock cookie %#llx",
1292 lock->l_handle.h_cookie);
1293 flags |= LDLM_FL_RESENT;
1294 GOTO(existing_lock, rc = 0);
1297 if (ldlm_reclaim_full()) {
1298 DEBUG_REQ(D_DLMTRACE, req,
1299 "Too many granted locks, reject current enqueue request and let the client retry later");
1300 GOTO(out, rc = -EINPROGRESS);
1304 /* The lock's callback data might be set in the policy function */
1305 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1306 dlm_req->lock_desc.l_resource.lr_type,
1307 dlm_req->lock_desc.l_req_mode,
1308 cbs, NULL, 0, LVB_T_NONE);
1315 lock->l_remote_handle = dlm_req->lock_handle[0];
1316 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1319 * Initialize resource lvb but not for a lock being replayed since
1320 * Client already got lvb sent in this case.
1321 * This must occur early since some policy methods assume resource
1322 * lvb is available (lr_lvb_data != NULL).
1324 res = lock->l_resource;
1325 if (!(flags & LDLM_FL_REPLAY)) {
1326 /* non-replayed lock, delayed lvb init may need to be done */
1327 rc = ldlm_lvbo_init(res);
1329 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1334 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1336 * Don't enqueue a lock onto the export if it is been disonnected
1337 * due to eviction (b=3822) or server umount (b=24324).
1338 * Cancel it now instead.
1340 if (req->rq_export->exp_disconnected) {
1341 LDLM_ERROR(lock, "lock on disconnected export %p",
1343 GOTO(out, rc = -ENOTCONN);
1346 lock->l_export = class_export_lock_get(req->rq_export, lock);
1347 if (lock->l_export->exp_lock_hash)
1348 cfs_hash_add(lock->l_export->exp_lock_hash,
1349 &lock->l_remote_handle,
1353 * Inherit the enqueue flags before the operation, because we do not
1354 * keep the res lock on return and next operations (BL AST) may proceed
1357 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1358 LDLM_FL_INHERIT_MASK);
1360 ldlm_convert_policy_to_local(req->rq_export,
1361 dlm_req->lock_desc.l_resource.lr_type,
1362 &dlm_req->lock_desc.l_policy_data,
1363 &lock->l_policy_data);
1364 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1365 lock->l_req_extent = lock->l_policy_data.l_extent;
1366 else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS)
1367 lock->l_policy_data.l_inodebits.try_bits =
1368 dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1372 if (!(flags & LDLM_FL_HAS_INTENT)) {
1373 /* based on the assumption that lvb size never changes during
1374 * resource life time otherwise it need resource->lr_lock's
1376 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1377 RCL_SERVER, ldlm_lvbo_size(lock));
1379 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1380 GOTO(out, rc = -ENOMEM);
1382 rc = req_capsule_server_pack(&req->rq_pill);
1387 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1394 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1396 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1397 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1399 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1400 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1403 * We never send a blocking AST until the lock is granted, but
1404 * we can tell it right now
1406 lock_res_and_lock(lock);
1409 * Now take into account flags to be inherited from original lock
1410 * request both in reply to client and in our own lock flags.
1412 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1413 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1416 * Don't move a pending lock onto the export if it has already been
1417 * disconnected due to eviction (b=5683) or server umount (b=24324).
1418 * Cancel it now instead.
1420 if (unlikely(req->rq_export->exp_disconnected ||
1421 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1422 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1424 } else if (ldlm_is_ast_sent(lock)) {
1425 /* fill lock desc for possible lock convert */
1426 if (lock->l_blocking_lock &&
1427 lock->l_resource->lr_type == LDLM_IBITS) {
1428 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1429 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1432 "save blocking bits %llx in granted lock",
1433 bl_lock->l_policy_data.l_inodebits.bits);
1435 * If lock is blocked then save blocking ibits
1436 * in returned lock policy for the possible lock
1437 * convert on a client.
1439 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1440 bl_lock->l_policy_data.l_inodebits.bits;
1442 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1443 if (ldlm_is_granted(lock)) {
1445 * Only cancel lock if it was granted, because it would
1446 * be destroyed immediately and would never be granted
1447 * in the future, causing timeouts on client. Not
1448 * granted lock will be cancelled immediately after
1449 * sending completion AST.
1451 if (ldlm_is_cancel_on_block(lock)) {
1452 unlock_res_and_lock(lock);
1453 ldlm_lock_cancel(lock);
1454 lock_res_and_lock(lock);
1456 ldlm_add_waiting_lock(lock,
1457 ldlm_bl_timeout(lock));
1461 unlock_res_and_lock(lock);
1465 req->rq_status = rc ?: err; /* return either error - b=11190 */
1466 if (!req->rq_packed_final) {
1467 err = lustre_pack_reply(req, 1, NULL, NULL);
1473 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1474 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1478 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1482 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1484 ldlm_lvbo_size(lock) > 0) {
1489 buf = req_capsule_server_get(&req->rq_pill,
1491 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1492 buflen = req_capsule_get_size(&req->rq_pill,
1493 &RMF_DLM_LVB, RCL_SERVER);
1495 * non-replayed lock, delayed lvb init may
1496 * need to be occur now
1498 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1501 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1503 req_capsule_shrink(&req->rq_pill,
1506 } else if (rc2 == -ERANGE) {
1507 rc2 = req_capsule_server_grow(
1509 &RMF_DLM_LVB, buflen);
1514 * if we can't grow the buffer,
1515 * it's ok to return empty lvb
1526 } else if (flags & LDLM_FL_REPLAY) {
1527 /* no LVB resend upon replay */
1529 req_capsule_shrink(&req->rq_pill,
1539 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1540 if (lock->l_export) {
1541 ldlm_lock_cancel(lock);
1543 lock_res_and_lock(lock);
1544 ldlm_resource_unlink_lock(lock);
1545 ldlm_lock_destroy_nolock(lock);
1546 unlock_res_and_lock(lock);
1551 if (!err && !ldlm_is_cbpending(lock) &&
1552 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1553 ldlm_reprocess_all(lock->l_resource, lock);
1555 LDLM_LOCK_RELEASE(lock);
1558 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1565 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1566 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1568 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1570 if (lock->l_blocking_lock) {
1571 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1572 lock->l_blocking_lock = NULL;
1576 /* A lock can be converted to new ibits or mode and should be considered
1577 * as new lock. Clear all states related to a previous blocking AST
1578 * processing so new conflicts will cause new blocking ASTs.
1580 * This is used during lock convert below and lock downgrade to COS mode in
1581 * ldlm_lock_mode_downgrade().
1583 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1585 ldlm_clear_ast_sent(lock);
1586 lock->l_bl_ast_run = 0;
1587 ldlm_clear_blocking_lock(lock);
1591 * Main LDLM entry point for server code to process lock conversion requests.
1593 int ldlm_handle_convert0(struct ptlrpc_request *req,
1594 const struct ldlm_request *dlm_req)
1596 struct obd_export *exp = req->rq_export;
1597 struct ldlm_reply *dlm_rep;
1598 struct ldlm_lock *lock;
1605 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1606 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1607 LDLM_CONVERT - LDLM_FIRST_OPC);
1609 rc = req_capsule_server_pack(&req->rq_pill);
1613 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1614 dlm_rep->lock_flags = dlm_req->lock_flags;
1616 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1618 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1619 req->rq_status = ELDLM_NO_LOCK_DATA;
1623 LDLM_DEBUG(lock, "server-side convert handler START");
1625 lock_res_and_lock(lock);
1626 bits = lock->l_policy_data.l_inodebits.bits;
1627 new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1629 if (ldlm_is_cancel(lock)) {
1630 LDLM_DEBUG(lock, "convert on canceled lock!");
1631 unlock_res_and_lock(lock);
1632 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1635 if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1636 LDLM_ERROR(lock, "lock mode differs!");
1637 unlock_res_and_lock(lock);
1638 GOTO(out_put, rc = -EPROTO);
1641 if (bits == new_bits) {
1643 * This can be valid situation if CONVERT RPCs are
1644 * re-ordered. Just finish silently
1646 LDLM_DEBUG(lock, "lock is converted already!");
1647 unlock_res_and_lock(lock);
1649 if (ldlm_is_waited(lock))
1650 ldlm_del_waiting_lock(lock);
1652 ldlm_clear_cbpending(lock);
1653 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1654 ldlm_inodebits_drop(lock, bits & ~new_bits);
1656 ldlm_clear_blocking_data(lock);
1657 unlock_res_and_lock(lock);
1659 ldlm_reprocess_all(lock->l_resource, NULL);
1662 dlm_rep->lock_handle = lock->l_remote_handle;
1663 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1664 &dlm_rep->lock_desc.l_policy_data);
1668 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1669 LDLM_LOCK_PUT(lock);
1670 req->rq_status = rc;
1675 * Cancel all the locks whose handles are packed into ldlm_request
1677 * Called by server code expecting such combined cancel activity
1680 int ldlm_request_cancel(struct ptlrpc_request *req,
1681 const struct ldlm_request *dlm_req,
1682 int first, enum lustre_at_flags flags)
1684 struct ldlm_resource *res, *pres = NULL;
1685 struct ldlm_lock *lock;
1686 int i, count, done = 0;
1691 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1692 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1693 (size - offsetof(struct ldlm_request, lock_handle)) /
1694 sizeof(struct lustre_handle) < dlm_req->lock_count)
1697 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1701 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1705 * There is no lock on the server at the replay time,
1706 * skip lock cancelling to make replay tests to pass.
1708 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1711 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1714 for (i = first; i < count; i++) {
1715 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1717 /* below message checked in replay-single.sh test_36 */
1718 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1719 dlm_req->lock_handle[i].cookie);
1723 res = lock->l_resource;
1727 * This code is an optimization to only attempt lock
1728 * granting on the resource (that could be CPU-expensive)
1729 * after we are done cancelling lock in that resource.
1733 ldlm_reprocess_all(pres, NULL);
1734 LDLM_RESOURCE_DELREF(pres);
1735 ldlm_resource_putref(pres);
1738 ldlm_resource_getref(res);
1739 LDLM_RESOURCE_ADDREF(res);
1741 if (!ldlm_is_discard_data(lock))
1742 ldlm_lvbo_update(res, lock,
1748 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1749 lock->l_blast_sent != 0) {
1750 timeout_t delay = 0;
1752 if (ktime_get_real_seconds() > lock->l_blast_sent)
1753 delay = ktime_get_real_seconds() -
1756 "server cancels blocked lock after %ds",
1758 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1760 ldlm_lock_cancel(lock);
1761 LDLM_LOCK_PUT(lock);
1764 ldlm_reprocess_all(pres, NULL);
1765 LDLM_RESOURCE_DELREF(pres);
1766 ldlm_resource_putref(pres);
1768 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1771 EXPORT_SYMBOL(ldlm_request_cancel);
1774 * Main LDLM entry point for server code to cancel locks.
1776 * Typically gets called from service handler on LDLM_CANCEL opc.
1778 int ldlm_handle_cancel(struct ptlrpc_request *req)
1780 struct ldlm_request *dlm_req;
1785 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1786 if (dlm_req == NULL) {
1787 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1791 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1792 offsetof(struct ldlm_request, lock_handle[1]))
1795 if (req->rq_export && req->rq_export->exp_nid_stats &&
1796 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1797 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1798 LDLM_CANCEL - LDLM_FIRST_OPC);
1800 rc = req_capsule_server_pack(&req->rq_pill);
1804 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1805 req->rq_status = LUSTRE_ESTALE;
1807 RETURN(ptlrpc_reply(req));
1809 #endif /* HAVE_SERVER_SUPPORT */
1812 * Server may pass additional information about blocking lock.
1813 * For IBITS locks it is conflicting bits which can be used for
1814 * lock convert instead of cancel.
1816 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1818 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1820 check_res_locked(lock->l_resource);
1821 if (ns_is_client(ns) && ld &&
1822 (lock->l_resource->lr_type == LDLM_IBITS)) {
1824 * Lock description contains policy of blocking lock,
1825 * and its cancel_bits is used to pass conflicting bits.
1826 * NOTE: ld can be NULL or can be not NULL but zeroed if
1827 * passed from ldlm_bl_thread_blwi(), check below used bits
1828 * in ld to make sure it is valid description.
1830 * If server may replace lock resource keeping the same cookie,
1831 * never use cancel bits from different resource, full cancel
1834 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1835 ldlm_res_eq(&ld->l_resource.lr_name,
1836 &lock->l_resource->lr_name) &&
1837 !(ldlm_is_cbpending(lock) &&
1838 lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1839 /* always combine conflicting ibits */
1840 lock->l_policy_data.l_inodebits.cancel_bits |=
1841 ld->l_policy_data.l_inodebits.cancel_bits;
1843 /* If cancel_bits are not obtained or
1844 * if the lock is already CBPENDING and
1845 * has no cancel_bits set
1846 * - the full lock is to be cancelled
1848 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1854 * Callback handler for receiving incoming blocking ASTs.
1856 * This can only happen on client side.
1858 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1859 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1865 LDLM_DEBUG(lock, "client blocking AST callback handler");
1867 lock_res_and_lock(lock);
1869 /* get extra information from desc if any */
1870 ldlm_bl_desc2lock(ld, lock);
1871 ldlm_set_cbpending(lock);
1873 do_ast = (!lock->l_readers && !lock->l_writers);
1874 unlock_res_and_lock(lock);
1878 "Lock %p already unused, calling callback (%p)\n",
1879 lock, lock->l_blocking_ast);
1880 if (lock->l_blocking_ast != NULL)
1881 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1885 "Lock %p is referenced, will be cancelled later\n",
1889 LDLM_DEBUG(lock, "client blocking callback handler END");
1890 LDLM_LOCK_RELEASE(lock);
1894 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1896 if (req->rq_no_reply)
1899 req->rq_status = rc;
1900 if (!req->rq_packed_final) {
1901 rc = lustre_pack_reply(req, 1, NULL, NULL);
1905 return ptlrpc_reply(req);
1909 * Callback handler for receiving incoming completion ASTs.
1911 * This only can happen on client side.
1913 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1914 struct ldlm_namespace *ns,
1915 struct ldlm_request *dlm_req,
1916 struct ldlm_lock *lock)
1918 LIST_HEAD(ast_list);
1924 LDLM_DEBUG(lock, "client completion callback handler START");
1926 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1927 long to = cfs_time_seconds(1);
1929 ldlm_callback_reply(req, 0);
1932 to = schedule_timeout_interruptible(to);
1933 if (ldlm_is_granted(lock) ||
1934 ldlm_is_destroyed(lock))
1939 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1941 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1942 GOTO(out, rc = lvb_len);
1943 } else if (lvb_len > 0) {
1944 if (lock->l_lvb_len > 0) {
1945 /* for extent lock, lvb contains ost_lvb{}. */
1946 LASSERT(lock->l_lvb_data != NULL);
1948 if (unlikely(lock->l_lvb_len < lvb_len)) {
1950 "Replied LVB is larger than expectation, expected = %d, replied = %d",
1951 lock->l_lvb_len, lvb_len);
1952 GOTO(out, rc = -EINVAL);
1957 lock_res_and_lock(lock);
1959 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1960 &lock->l_resource->lr_name)) {
1961 ldlm_resource_unlink_lock(lock);
1962 unlock_res_and_lock(lock);
1963 rc = ldlm_lock_change_resource(ns, lock,
1964 &dlm_req->lock_desc.l_resource.lr_name);
1966 LDLM_ERROR(lock, "Failed to allocate resource");
1969 LDLM_DEBUG(lock, "completion AST, new resource");
1970 lock_res_and_lock(lock);
1973 if (ldlm_is_failed(lock)) {
1974 unlock_res_and_lock(lock);
1975 LDLM_LOCK_RELEASE(lock);
1979 if (ldlm_is_destroyed(lock) ||
1980 ldlm_is_granted(lock)) {
1981 /* b=11300: the lock has already been granted */
1982 unlock_res_and_lock(lock);
1983 LDLM_DEBUG(lock, "Double grant race happened");
1988 * If we receive the completion AST before the actual enqueue returned,
1989 * then we might need to switch lock modes, resources, or extents.
1991 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1992 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1993 LDLM_DEBUG(lock, "completion AST, new lock mode");
1996 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1997 ldlm_convert_policy_to_local(req->rq_export,
1998 dlm_req->lock_desc.l_resource.lr_type,
1999 &dlm_req->lock_desc.l_policy_data,
2000 &lock->l_policy_data);
2001 LDLM_DEBUG(lock, "completion AST, new policy data");
2004 ldlm_resource_unlink_lock(lock);
2006 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2008 * BL_AST locks are not needed in LRU.
2009 * Let ldlm_cancel_lru() be fast.
2011 ldlm_lock_remove_from_lru(lock);
2012 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2013 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2014 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2017 if (lock->l_lvb_len > 0) {
2018 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2019 lock->l_lvb_data, lvb_len);
2021 unlock_res_and_lock(lock);
2026 ldlm_grant_lock(lock, &ast_list);
2027 unlock_res_and_lock(lock);
2029 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2032 * Let Enqueue to call osc_lock_upcall() and initialize
2035 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2037 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2039 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2045 lock_res_and_lock(lock);
2046 ldlm_set_failed(lock);
2047 unlock_res_and_lock(lock);
2048 wake_up(&lock->l_waitq);
2050 LDLM_LOCK_RELEASE(lock);
2056 * Callback handler for receiving incoming glimpse ASTs.
2058 * This only can happen on client side. After handling the glimpse AST
2059 * we also consider dropping the lock here if it is unused locally for a
2062 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2063 struct ldlm_namespace *ns,
2064 struct ldlm_request *dlm_req,
2065 struct ldlm_lock *lock)
2067 struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2072 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2074 if (lock->l_glimpse_ast != NULL)
2075 rc = lock->l_glimpse_ast(lock, req);
2077 if (req->rq_repmsg != NULL) {
2080 req->rq_status = rc;
2084 lock_res_and_lock(lock);
2085 if (lock->l_granted_mode == LCK_PW &&
2086 !lock->l_readers && !lock->l_writers &&
2087 ktime_after(ktime_get(),
2088 ktime_add(lock->l_last_used, ns->ns_dirty_age_limit))) {
2089 unlock_res_and_lock(lock);
2091 /* For MDS glimpse it is always DOM lock, set corresponding
2092 * cancel_bits to perform lock convert if needed
2094 if (lock->l_resource->lr_type == LDLM_IBITS)
2095 ld->l_policy_data.l_inodebits.cancel_bits =
2097 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2098 ldlm_handle_bl_callback(ns, ld, lock);
2103 unlock_res_and_lock(lock);
2104 LDLM_LOCK_RELEASE(lock);
2108 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2109 enum ldlm_cancel_flags cancel_flags)
2111 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2115 spin_lock(&blp->blp_lock);
2116 if (blwi->blwi_lock &&
2117 ldlm_is_discard_data(blwi->blwi_lock)) {
2118 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2119 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2121 /* other blocking callbacks are added to the regular list */
2122 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2124 spin_unlock(&blp->blp_lock);
2126 wake_up(&blp->blp_waitq);
2129 * can not check blwi->blwi_flags as blwi could be already freed in
2132 if (!(cancel_flags & LCF_ASYNC))
2133 wait_for_completion(&blwi->blwi_comp);
2138 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2139 struct ldlm_namespace *ns,
2140 struct ldlm_lock_desc *ld,
2141 struct list_head *cancels, int count,
2142 struct ldlm_lock *lock,
2143 enum ldlm_cancel_flags cancel_flags)
2145 init_completion(&blwi->blwi_comp);
2146 INIT_LIST_HEAD(&blwi->blwi_head);
2148 if (current->flags & PF_MEMALLOC)
2149 blwi->blwi_mem_pressure = 1;
2152 blwi->blwi_flags = cancel_flags;
2154 blwi->blwi_ld = *ld;
2156 list_splice_init(cancels, &blwi->blwi_head);
2157 blwi->blwi_count = count;
2159 blwi->blwi_lock = lock;
2164 * Queues a list of locks \a cancels containing \a count locks
2165 * for later processing by a blocking thread. If \a count is zero,
2166 * then the lock referenced as \a lock is queued instead.
2168 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2169 * If list addition fails an error is returned and caller is supposed to
2170 * call ->l_blocking_ast itself.
2172 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2173 struct ldlm_lock_desc *ld,
2174 struct ldlm_lock *lock,
2175 struct list_head *cancels, int count,
2176 enum ldlm_cancel_flags cancel_flags)
2180 if (cancels && count == 0)
2183 if (cancel_flags & LCF_ASYNC) {
2184 struct ldlm_bl_work_item *blwi;
2186 OBD_ALLOC(blwi, sizeof(*blwi));
2189 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2191 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2194 * if it is synchronous call do minimum mem alloc, as it could
2195 * be triggered from kernel shrinker
2197 struct ldlm_bl_work_item blwi;
2199 memset(&blwi, 0, sizeof(blwi));
2200 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2201 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2206 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2207 struct ldlm_lock *lock)
2209 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2212 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2213 struct list_head *cancels, int count,
2214 enum ldlm_cancel_flags cancel_flags)
2216 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2219 int ldlm_bl_to_thread_ns(struct ldlm_namespace *ns)
2221 return ldlm_bl_to_thread(ns, NULL, NULL, NULL, 0, LCF_ASYNC);
2224 int ldlm_bl_thread_wakeup(void)
2226 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2230 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2231 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2233 struct obd_device *obd = req->rq_export->exp_obd;
2241 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2243 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2245 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2247 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2250 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2252 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2254 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2257 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2260 /* We are responsible for swabbing contents of val */
2262 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2263 /* Pass it on to mdc (the "export" in this case) */
2264 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2266 sizeof(KEY_HSM_COPYTOOL_SEND),
2267 KEY_HSM_COPYTOOL_SEND,
2270 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2275 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2276 const char *msg, int rc,
2277 const struct lustre_handle *handle)
2279 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2280 "%s, NID=%s lock=%#llx: rc = %d",
2281 msg, libcfs_id2str(req->rq_peer),
2282 handle ? handle->cookie : 0, rc);
2283 if (req->rq_no_reply)
2284 CWARN("No reply was sent, maybe cause b=21636.\n");
2286 CWARN("Send reply failed, maybe cause b=21636.\n");
2289 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2290 static int ldlm_callback_handler(struct ptlrpc_request *req)
2292 struct ldlm_namespace *ns;
2293 struct ldlm_request *dlm_req;
2294 struct ldlm_lock *lock;
2300 * Requests arrive in sender's byte order. The ptlrpc service
2301 * handler has already checked and, if necessary, byte-swapped the
2302 * incoming request message body, but I am responsible for the
2306 /* do nothing for sec context finalize */
2307 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2310 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2312 if (req->rq_export == NULL) {
2313 rc = ldlm_callback_reply(req, -ENOTCONN);
2314 ldlm_callback_errmsg(req, "Operate on unconnected server",
2319 LASSERT(req->rq_export != NULL);
2320 LASSERT(req->rq_export->exp_obd != NULL);
2322 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2323 case LDLM_BL_CALLBACK:
2324 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2326 ldlm_callback_reply(req, -(int)cfs_fail_err);
2330 case LDLM_CP_CALLBACK:
2331 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2334 case LDLM_GL_CALLBACK:
2335 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2339 rc = ldlm_handle_setinfo(req);
2340 ldlm_callback_reply(req, rc);
2343 CERROR("unknown opcode %u\n",
2344 lustre_msg_get_opc(req->rq_reqmsg));
2345 ldlm_callback_reply(req, -EPROTO);
2349 ns = req->rq_export->exp_obd->obd_namespace;
2350 LASSERT(ns != NULL);
2352 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2354 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2355 if (dlm_req == NULL) {
2356 rc = ldlm_callback_reply(req, -EPROTO);
2357 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2363 * Force a known safe race, send a cancel to the server for a lock
2364 * which the server has already started a blocking callback on.
2366 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2367 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2368 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2370 CERROR("ldlm_cli_cancel: %d\n", rc);
2373 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2376 "callback on lock %#llx - lock disappeared\n",
2377 dlm_req->lock_handle[0].cookie);
2378 rc = ldlm_callback_reply(req, -EINVAL);
2379 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2380 &dlm_req->lock_handle[0]);
2384 if (ldlm_is_fail_loc(lock) &&
2385 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2386 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2388 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2389 lock_res_and_lock(lock);
2390 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2392 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2394 * If somebody cancels lock and cache is already dropped,
2395 * or lock is failed before cp_ast received on client,
2396 * we can tell the server we have no lock. Otherwise, we
2397 * should send cancel after dropping the cache.
2399 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2400 ldlm_is_failed(lock)) {
2402 "callback on lock %llx - lock disappeared",
2403 dlm_req->lock_handle[0].cookie);
2404 unlock_res_and_lock(lock);
2405 LDLM_LOCK_RELEASE(lock);
2406 rc = ldlm_callback_reply(req, -EINVAL);
2407 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2408 &dlm_req->lock_handle[0]);
2412 * BL_AST locks are not needed in LRU.
2413 * Let ldlm_cancel_lru() be fast.
2415 ldlm_lock_remove_from_lru(lock);
2416 ldlm_set_bl_ast(lock);
2418 if (lock->l_remote_handle.cookie == 0)
2419 lock->l_remote_handle = dlm_req->lock_handle[1];
2420 unlock_res_and_lock(lock);
2423 * We want the ost thread to get this reply so that it can respond
2424 * to ost requests (write cache writeback) that might be triggered
2427 * But we'd also like to be able to indicate in the reply that we're
2428 * cancelling right now, because it's unused, or have an intent result
2429 * in the reply, so we might have to push the responsibility for sending
2430 * the reply down into the AST handlers, alas.
2433 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2434 case LDLM_BL_CALLBACK:
2435 CDEBUG(D_INODE, "blocking ast\n");
2436 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2437 if (!ldlm_is_cancel_on_block(lock)) {
2438 rc = ldlm_callback_reply(req, 0);
2439 if (req->rq_no_reply || rc)
2440 ldlm_callback_errmsg(req, "Normal process", rc,
2441 &dlm_req->lock_handle[0]);
2443 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2444 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2446 case LDLM_CP_CALLBACK:
2447 CDEBUG(D_INODE, "completion ast\n");
2448 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2449 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2450 if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2451 ldlm_callback_reply(req, rc);
2453 case LDLM_GL_CALLBACK:
2454 CDEBUG(D_INODE, "glimpse ast\n");
2455 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2456 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2459 LBUG(); /* checked above */
2465 #ifdef HAVE_SERVER_SUPPORT
2467 * Main handler for canceld thread.
2469 * Separated into its own thread to avoid deadlocks.
2471 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2478 * Requests arrive in sender's byte order. The ptlrpc service
2479 * handler has already checked and, if necessary, byte-swapped the
2480 * incoming request message body, but I am responsible for the
2484 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2486 if (req->rq_export == NULL) {
2487 struct ldlm_request *dlm_req;
2489 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2490 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2491 libcfs_nid2str(req->rq_peer.nid),
2492 (unsigned long long)req->rq_arrival_time.tv_sec,
2493 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2495 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2496 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2497 dlm_req = req_capsule_client_get(&req->rq_pill,
2499 if (dlm_req != NULL)
2500 ldlm_lock_dump_handle(D_ERROR,
2501 &dlm_req->lock_handle[0]);
2503 ldlm_callback_reply(req, -ENOTCONN);
2507 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2508 /* XXX FIXME move this back to mds/handler.c, b=249 */
2510 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2511 CDEBUG(D_INODE, "cancel\n");
2512 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2513 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2514 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2516 rc = ldlm_handle_cancel(req);
2520 struct ldlm_request *dlm_req;
2522 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2523 CDEBUG(D_INODE, "convert\n");
2525 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2526 if (dlm_req == NULL) {
2527 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2528 rc = ldlm_callback_reply(req, -EPROTO);
2530 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2531 rc = ptlrpc_reply(req);
2536 CERROR("invalid opcode %d\n",
2537 lustre_msg_get_opc(req->rq_reqmsg));
2538 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2539 rc = ldlm_callback_reply(req, -EINVAL);
2545 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2546 struct ldlm_lock *lock)
2548 struct ldlm_request *dlm_req;
2549 struct lustre_handle lockh;
2555 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2556 if (dlm_req == NULL)
2559 ldlm_lock2handle(lock, &lockh);
2560 for (i = 0; i < dlm_req->lock_count; i++) {
2561 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2563 DEBUG_REQ(D_RPCTRACE, req,
2564 "Prio raised by lock %#llx", lockh.cookie);
2573 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2575 struct ldlm_request *dlm_req;
2582 /* no prolong in recovery */
2583 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2586 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2587 if (dlm_req == NULL)
2590 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2591 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2592 (size - offsetof(struct ldlm_request, lock_handle)) /
2593 sizeof(struct lustre_handle) < dlm_req->lock_count)
2596 for (i = 0; i < dlm_req->lock_count; i++) {
2597 struct ldlm_lock *lock;
2599 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2603 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2605 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2606 LDLM_LOCK_PUT(lock);
2615 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2616 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2617 .hpreq_check = ldlm_cancel_hpreq_check,
2621 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2625 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2627 if (req->rq_export == NULL)
2630 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2631 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2632 req->rq_ops = &ldlm_cancel_hpreq_ops;
2633 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2634 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2635 req->rq_ops = &ldlm_cancel_hpreq_ops;
2640 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2641 struct hlist_node *hnode, void *data)
2644 struct list_head *rpc_list = data;
2645 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2647 lock_res_and_lock(lock);
2649 if (!ldlm_is_granted(lock)) {
2650 unlock_res_and_lock(lock);
2654 LASSERT(lock->l_resource);
2655 if (lock->l_resource->lr_type != LDLM_IBITS &&
2656 lock->l_resource->lr_type != LDLM_PLAIN) {
2657 unlock_res_and_lock(lock);
2661 if (ldlm_is_ast_sent(lock)) {
2662 unlock_res_and_lock(lock);
2666 LASSERT(lock->l_blocking_ast);
2667 LASSERT(!lock->l_blocking_lock);
2669 ldlm_set_ast_sent(lock);
2670 if (lock->l_export && lock->l_export->exp_lock_hash) {
2672 * NB: it's safe to call cfs_hash_del() even lock isn't
2676 * In the function below, .hs_keycmp resolves to
2677 * ldlm_export_lock_keycmp()
2679 /* coverity[overrun-buffer-val] */
2680 cfs_hash_del(lock->l_export->exp_lock_hash,
2681 &lock->l_remote_handle, &lock->l_exp_hash);
2684 list_add_tail(&lock->l_rk_ast, rpc_list);
2685 LDLM_LOCK_GET(lock);
2687 unlock_res_and_lock(lock);
2691 void ldlm_revoke_export_locks(struct obd_export *exp)
2693 LIST_HEAD(rpc_list);
2697 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2698 ldlm_revoke_lock_cb, &rpc_list, 0);
2699 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2700 LDLM_WORK_REVOKE_AST);
2704 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2705 #endif /* HAVE_SERVER_SUPPORT */
2707 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2708 struct ldlm_bl_work_item **p_blwi,
2709 struct obd_export **p_exp)
2711 struct ldlm_bl_work_item *blwi = NULL;
2712 static unsigned int num_bl;
2713 static unsigned int num_stale;
2714 int num_th = atomic_read(&blp->blp_num_threads);
2716 *p_exp = obd_stale_export_get();
2718 spin_lock(&blp->blp_lock);
2719 if (*p_exp != NULL) {
2720 if (num_th == 1 || ++num_stale < num_th) {
2721 spin_unlock(&blp->blp_lock);
2727 /* process a request from the blp_list at least every blp_num_threads */
2728 if (!list_empty(&blp->blp_list) &&
2729 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2730 blwi = list_entry(blp->blp_list.next,
2731 struct ldlm_bl_work_item, blwi_entry);
2733 if (!list_empty(&blp->blp_prio_list))
2734 blwi = list_entry(blp->blp_prio_list.next,
2735 struct ldlm_bl_work_item,
2739 if (++num_bl >= num_th)
2741 list_del(&blwi->blwi_entry);
2743 spin_unlock(&blp->blp_lock);
2746 if (*p_exp != NULL && *p_blwi != NULL) {
2747 obd_stale_export_put(*p_exp);
2751 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2754 /* This only contains temporary data until the thread starts */
2755 struct ldlm_bl_thread_data {
2756 struct ldlm_bl_pool *bltd_blp;
2757 struct completion bltd_comp;
2761 static int ldlm_bl_thread_main(void *arg);
2763 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2765 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2766 struct task_struct *task;
2768 init_completion(&bltd.bltd_comp);
2770 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2771 if (bltd.bltd_num >= blp->blp_max_threads) {
2772 atomic_dec(&blp->blp_num_threads);
2776 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2778 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2779 atomic_dec(&blp->blp_num_threads);
2783 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2786 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2787 bltd.bltd_num, PTR_ERR(task));
2788 atomic_dec(&blp->blp_num_threads);
2789 return PTR_ERR(task);
2791 wait_for_completion(&bltd.bltd_comp);
2796 /* Not fatal if racy and have a few too many threads */
2797 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2798 struct ldlm_bl_work_item *blwi)
2800 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2803 if (atomic_read(&blp->blp_busy_threads) <
2804 atomic_read(&blp->blp_num_threads))
2807 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2808 blwi->blwi_mem_pressure))
2814 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2815 struct ldlm_bl_work_item *blwi)
2817 /* '1' for consistency with code that checks !mpflag to restore */
2818 unsigned int mpflags = 1;
2822 if (blwi->blwi_ns == NULL)
2823 /* added by ldlm_cleanup() */
2824 RETURN(LDLM_ITER_STOP);
2826 if (blwi->blwi_mem_pressure)
2827 mpflags = memalloc_noreclaim_save();
2829 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2831 if (blwi->blwi_count) {
2834 * The special case when we cancel locks in lru
2835 * asynchronously, we pass the list of locks here.
2836 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2837 * canceled locally yet.
2839 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2842 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2844 } else if (blwi->blwi_lock) {
2845 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2848 ldlm_pool_recalc(&blwi->blwi_ns->ns_pool, true);
2849 spin_lock(&blwi->blwi_ns->ns_lock);
2850 blwi->blwi_ns->ns_rpc_recalc = 0;
2851 spin_unlock(&blwi->blwi_ns->ns_lock);
2852 ldlm_namespace_put(blwi->blwi_ns);
2855 if (blwi->blwi_mem_pressure)
2856 memalloc_noreclaim_restore(mpflags);
2858 if (blwi->blwi_flags & LCF_ASYNC)
2859 OBD_FREE(blwi, sizeof(*blwi));
2861 complete(&blwi->blwi_comp);
2867 * Cancel stale locks on export. Cancel blocked locks first.
2868 * If the given export has blocked locks, the next in the list may have
2869 * them too, thus cancel not blocked locks only if the current export has
2872 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2873 struct obd_export *exp)
2879 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2881 num = ldlm_export_cancel_blocked_locks(exp);
2883 ldlm_export_cancel_locks(exp);
2885 obd_stale_export_put(exp);
2892 * Main blocking requests processing thread.
2894 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2895 * This thread in the end ends up doing actual call to ->l_blocking_ast
2898 static int ldlm_bl_thread_main(void *arg)
2901 struct ldlm_bl_pool *blp;
2902 struct ldlm_bl_thread_data *bltd = arg;
2910 rc = lu_env_init(env, LCT_DT_THREAD);
2913 rc = lu_env_add(env);
2915 GOTO(out_env_fini, rc);
2917 blp = bltd->bltd_blp;
2919 complete(&bltd->bltd_comp);
2920 /* cannot use bltd after this, it is only on caller's stack */
2923 struct ldlm_bl_work_item *blwi = NULL;
2924 struct obd_export *exp = NULL;
2927 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2930 wait_event_idle_exclusive(blp->blp_waitq,
2931 ldlm_bl_get_work(blp, &blwi,
2933 atomic_inc(&blp->blp_busy_threads);
2935 if (ldlm_bl_thread_need_create(blp, blwi))
2936 /* discard the return value, we tried */
2937 ldlm_bl_thread_start(blp, true);
2940 rc = ldlm_bl_thread_exports(blp, exp);
2942 rc = ldlm_bl_thread_blwi(blp, blwi);
2944 atomic_dec(&blp->blp_busy_threads);
2946 if (rc == LDLM_ITER_STOP)
2950 * If there are many namespaces, we will not sleep waiting for
2951 * work, and must do a cond_resched to avoid holding the CPU
2957 atomic_dec(&blp->blp_num_threads);
2958 complete(&blp->blp_comp);
2969 static int ldlm_setup(void);
2970 static int ldlm_cleanup(void);
2972 int ldlm_get_ref(void)
2977 mutex_lock(&ldlm_ref_mutex);
2978 if (++ldlm_refcount == 1) {
2983 mutex_unlock(&ldlm_ref_mutex);
2988 void ldlm_put_ref(void)
2991 mutex_lock(&ldlm_ref_mutex);
2992 if (ldlm_refcount == 1) {
2993 int rc = ldlm_cleanup();
2996 CERROR("ldlm_cleanup failed: %d\n", rc);
3002 mutex_unlock(&ldlm_ref_mutex);
3008 * Export handle<->lock hash operations.
3011 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
3013 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
3017 ldlm_export_lock_key(struct hlist_node *hnode)
3019 struct ldlm_lock *lock;
3021 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3022 return &lock->l_remote_handle;
3026 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3028 struct ldlm_lock *lock;
3030 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3031 lock->l_remote_handle = *(struct lustre_handle *)key;
3035 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3037 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3041 ldlm_export_lock_object(struct hlist_node *hnode)
3043 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3047 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3049 struct ldlm_lock *lock;
3051 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3052 LDLM_LOCK_GET(lock);
3056 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3058 struct ldlm_lock *lock;
3060 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3061 LDLM_LOCK_RELEASE(lock);
3064 static struct cfs_hash_ops ldlm_export_lock_ops = {
3065 .hs_hash = ldlm_export_lock_hash,
3066 .hs_key = ldlm_export_lock_key,
3067 .hs_keycmp = ldlm_export_lock_keycmp,
3068 .hs_keycpy = ldlm_export_lock_keycpy,
3069 .hs_object = ldlm_export_lock_object,
3070 .hs_get = ldlm_export_lock_get,
3071 .hs_put = ldlm_export_lock_put,
3072 .hs_put_locked = ldlm_export_lock_put,
3075 int ldlm_init_export(struct obd_export *exp)
3081 exp->exp_lock_hash =
3082 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3083 HASH_EXP_LOCK_CUR_BITS,
3084 HASH_EXP_LOCK_MAX_BITS,
3085 HASH_EXP_LOCK_BKT_BITS, 0,
3086 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3087 &ldlm_export_lock_ops,
3088 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3089 CFS_HASH_NBLK_CHANGE);
3091 if (!exp->exp_lock_hash)
3094 rc = ldlm_init_flock_export(exp);
3100 ldlm_destroy_export(exp);
3103 EXPORT_SYMBOL(ldlm_init_export);
3105 void ldlm_destroy_export(struct obd_export *exp)
3108 cfs_hash_putref(exp->exp_lock_hash);
3109 exp->exp_lock_hash = NULL;
3111 ldlm_destroy_flock_export(exp);
3114 EXPORT_SYMBOL(ldlm_destroy_export);
3116 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3117 struct attribute *attr,
3120 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3123 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3124 struct attribute *attr,
3131 rc = kstrtoul(buffer, 10, &val);
3135 ldlm_cancel_unused_locks_before_replay = val;
3139 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3141 static struct attribute *ldlm_attrs[] = {
3142 &lustre_attr_cancel_unused_locks_before_replay.attr,
3146 static struct attribute_group ldlm_attr_group = {
3147 .attrs = ldlm_attrs,
3150 static int ldlm_setup(void)
3152 static struct ptlrpc_service_conf conf;
3153 struct ldlm_bl_pool *blp = NULL;
3154 #ifdef HAVE_SERVER_SUPPORT
3155 struct task_struct *task;
3156 #endif /* HAVE_SERVER_SUPPORT */
3162 if (ldlm_state != NULL)
3165 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3166 if (ldlm_state == NULL)
3169 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3173 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3177 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3181 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3185 rc = ldlm_debugfs_setup();
3189 memset(&conf, 0, sizeof(conf));
3190 conf = (typeof(conf)) {
3191 .psc_name = "ldlm_cbd",
3192 .psc_watchdog_factor = 2,
3194 .bc_nbufs = LDLM_CLIENT_NBUFS,
3195 .bc_buf_size = LDLM_BUFSIZE,
3196 .bc_req_max_size = LDLM_MAXREQSIZE,
3197 .bc_rep_max_size = LDLM_MAXREPSIZE,
3198 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3199 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3202 .tc_thr_name = "ldlm_cb",
3203 .tc_thr_factor = LDLM_THR_FACTOR,
3204 .tc_nthrs_init = LDLM_NTHRS_INIT,
3205 .tc_nthrs_base = LDLM_NTHRS_BASE,
3206 .tc_nthrs_max = LDLM_NTHRS_MAX,
3207 .tc_nthrs_user = ldlm_num_threads,
3208 .tc_cpu_bind = ldlm_cpu_bind,
3209 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3212 .cc_pattern = ldlm_cpts,
3213 .cc_affinity = true,
3216 .so_req_handler = ldlm_callback_handler,
3219 ldlm_state->ldlm_cb_service = \
3220 ptlrpc_register_service(&conf, ldlm_svc_kset,
3221 ldlm_svc_debugfs_dir);
3222 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3223 CERROR("failed to start service\n");
3224 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3225 ldlm_state->ldlm_cb_service = NULL;
3229 #ifdef HAVE_SERVER_SUPPORT
3230 memset(&conf, 0, sizeof(conf));
3231 conf = (typeof(conf)) {
3232 .psc_name = "ldlm_canceld",
3233 .psc_watchdog_factor = 6,
3235 .bc_nbufs = LDLM_SERVER_NBUFS,
3236 .bc_buf_size = LDLM_BUFSIZE,
3237 .bc_req_max_size = LDLM_MAXREQSIZE,
3238 .bc_rep_max_size = LDLM_MAXREPSIZE,
3239 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3240 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3244 .tc_thr_name = "ldlm_cn",
3245 .tc_thr_factor = LDLM_THR_FACTOR,
3246 .tc_nthrs_init = LDLM_NTHRS_INIT,
3247 .tc_nthrs_base = LDLM_NTHRS_BASE,
3248 .tc_nthrs_max = LDLM_NTHRS_MAX,
3249 .tc_nthrs_user = ldlm_num_threads,
3250 .tc_cpu_bind = ldlm_cpu_bind,
3251 .tc_ctx_tags = LCT_MD_THREAD | \
3256 .cc_pattern = ldlm_cpts,
3257 .cc_affinity = true,
3260 .so_req_handler = ldlm_cancel_handler,
3261 .so_hpreq_handler = ldlm_hpreq_handler,
3264 ldlm_state->ldlm_cancel_service = \
3265 ptlrpc_register_service(&conf, ldlm_svc_kset,
3266 ldlm_svc_debugfs_dir);
3267 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3268 CERROR("failed to start service\n");
3269 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3270 ldlm_state->ldlm_cancel_service = NULL;
3273 #endif /* HAVE_SERVER_SUPPORT */
3275 OBD_ALLOC(blp, sizeof(*blp));
3277 GOTO(out, rc = -ENOMEM);
3278 ldlm_state->ldlm_bl_pool = blp;
3280 spin_lock_init(&blp->blp_lock);
3281 INIT_LIST_HEAD(&blp->blp_list);
3282 INIT_LIST_HEAD(&blp->blp_prio_list);
3283 init_waitqueue_head(&blp->blp_waitq);
3284 atomic_set(&blp->blp_num_threads, 0);
3285 atomic_set(&blp->blp_busy_threads, 0);
3287 if (ldlm_num_threads == 0) {
3288 blp->blp_min_threads = LDLM_NTHRS_INIT;
3289 blp->blp_max_threads = LDLM_NTHRS_MAX;
3291 blp->blp_min_threads = blp->blp_max_threads = \
3292 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3296 for (i = 0; i < blp->blp_min_threads; i++) {
3297 rc = ldlm_bl_thread_start(blp, false);
3302 #ifdef HAVE_SERVER_SUPPORT
3303 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3306 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3310 wait_event(expired_lock_wait_queue,
3311 expired_lock_thread_state == ELT_READY);
3312 #endif /* HAVE_SERVER_SUPPORT */
3314 rc = ldlm_pools_init();
3316 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3320 rc = ldlm_reclaim_setup();
3322 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3332 static int ldlm_cleanup(void)
3336 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3337 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3338 CERROR("ldlm still has namespaces; clean these up first.\n");
3339 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3340 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3344 ldlm_reclaim_cleanup();
3347 if (ldlm_state->ldlm_bl_pool != NULL) {
3348 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3350 while (atomic_read(&blp->blp_num_threads) > 0) {
3351 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3353 init_completion(&blp->blp_comp);
3355 spin_lock(&blp->blp_lock);
3356 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3357 wake_up(&blp->blp_waitq);
3358 spin_unlock(&blp->blp_lock);
3360 wait_for_completion(&blp->blp_comp);
3363 OBD_FREE(blp, sizeof(*blp));
3366 if (ldlm_state->ldlm_cb_service != NULL)
3367 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3368 #ifdef HAVE_SERVER_SUPPORT
3369 if (ldlm_state->ldlm_cancel_service != NULL)
3370 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3374 kset_unregister(ldlm_ns_kset);
3376 kset_unregister(ldlm_svc_kset);
3378 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3379 kobject_put(ldlm_kobj);
3382 ldlm_debugfs_cleanup();
3384 #ifdef HAVE_SERVER_SUPPORT
3385 if (expired_lock_thread_state != ELT_STOPPED) {
3386 expired_lock_thread_state = ELT_TERMINATE;
3387 wake_up(&expired_lock_wait_queue);
3388 wait_event(expired_lock_wait_queue,
3389 expired_lock_thread_state == ELT_STOPPED);
3393 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3401 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3402 sizeof(struct ldlm_resource), 0,
3403 SLAB_HWCACHE_ALIGN, NULL);
3404 if (ldlm_resource_slab == NULL)
3407 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3408 sizeof(struct ldlm_lock), 0,
3409 SLAB_HWCACHE_ALIGN, NULL);
3410 if (ldlm_lock_slab == NULL)
3413 ldlm_interval_slab = kmem_cache_create("interval_node",
3414 sizeof(struct ldlm_interval),
3415 0, SLAB_HWCACHE_ALIGN, NULL);
3416 if (ldlm_interval_slab == NULL)
3419 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3420 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3421 0, SLAB_HWCACHE_ALIGN, NULL);
3422 if (ldlm_interval_tree_slab == NULL)
3425 #ifdef HAVE_SERVER_SUPPORT
3426 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3427 sizeof(struct ldlm_ibits_node),
3428 0, SLAB_HWCACHE_ALIGN, NULL);
3429 if (ldlm_inodebits_slab == NULL)
3430 goto out_interval_tree;
3432 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3433 sizeof(struct ldlm_glimpse_work),
3435 if (ldlm_glimpse_work_kmem == NULL)
3439 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3440 class_export_dump_hook = ldlm_dump_export_locks;
3443 #ifdef HAVE_SERVER_SUPPORT
3445 kmem_cache_destroy(ldlm_inodebits_slab);
3447 kmem_cache_destroy(ldlm_interval_tree_slab);
3450 kmem_cache_destroy(ldlm_interval_slab);
3452 kmem_cache_destroy(ldlm_lock_slab);
3454 kmem_cache_destroy(ldlm_resource_slab);
3459 void ldlm_exit(void)
3462 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3463 kmem_cache_destroy(ldlm_resource_slab);
3465 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3466 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3467 * so that ldlm_lock_free() get a chance to be called.
3470 kmem_cache_destroy(ldlm_lock_slab);
3471 kmem_cache_destroy(ldlm_interval_slab);
3472 kmem_cache_destroy(ldlm_interval_tree_slab);
3473 #ifdef HAVE_SERVER_SUPPORT
3474 kmem_cache_destroy(ldlm_inodebits_slab);
3475 kmem_cache_destroy(ldlm_glimpse_work_kmem);