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_desc = *desc;
948 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
950 LDLM_DEBUG(lock, "server preparing blocking AST");
952 ptlrpc_request_set_replen(req);
953 ldlm_set_cbpending(lock);
954 if (instant_cancel) {
955 unlock_res_and_lock(lock);
956 ldlm_lock_cancel(lock);
958 req->rq_no_resend = 1;
960 LASSERT(ldlm_is_granted(lock));
961 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
962 unlock_res_and_lock(lock);
964 /* Do not resend after lock callback timeout */
965 req->rq_delay_limit = ldlm_bl_timeout(lock);
966 req->rq_resend_cb = ldlm_update_resend;
969 req->rq_send_state = LUSTRE_IMP_FULL;
970 /* ptlrpc_request_alloc_pack already set timeout */
972 req->rq_timeout = ldlm_get_rq_timeout();
974 if (lock->l_export && lock->l_export->exp_nid_stats &&
975 lock->l_export->exp_nid_stats->nid_ldlm_stats)
976 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
977 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
979 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
985 * ->l_completion_ast callback for a remote lock in server namespace.
987 * Sends AST to the client notifying it of lock granting. If initial
988 * lock response was not sent yet, instead of sending another RPC, just
989 * mark the lock as granted and client will understand
991 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
993 struct ldlm_cb_set_arg *arg = data;
994 struct ldlm_request *body;
995 struct ptlrpc_request *req;
996 struct ldlm_cb_async_args *ca;
997 int instant_cancel = 0;
1003 LASSERT(lock != NULL);
1004 LASSERT(data != NULL);
1006 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1007 LDLM_DEBUG(lock, "dropping CP AST");
1011 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1012 &RQF_LDLM_CP_CALLBACK);
1016 /* server namespace, doesn't need lock */
1017 lvb_len = ldlm_lvbo_size(lock);
1019 * LU-3124 & LU-2187: to not return layout in completion AST because
1020 * it may deadlock for LU-2187, or client may not have enough space
1021 * for large layout. The layout will be returned to client with an
1022 * extra RPC to fetch xattr.lov
1024 if (ldlm_has_layout(lock))
1027 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1028 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1030 ptlrpc_request_free(req);
1034 ca = ptlrpc_req_async_args(ca, req);
1035 ca->ca_set_arg = arg;
1038 req->rq_interpret_reply = ldlm_cb_interpret;
1039 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1041 body->lock_handle[0] = lock->l_remote_handle;
1042 body->lock_flags = ldlm_flags_to_wire(flags);
1043 ldlm_lock2desc(lock, &body->lock_desc);
1045 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1046 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1049 * We still need to send the RPC to wake up the blocked
1050 * enqueue thread on the client.
1052 * Consider old client, there is no better way to notify
1053 * the failure, just zero-sized the LVB, then the client
1054 * will fail out as "-EPROTO".
1056 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1060 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1065 LDLM_DEBUG(lock, "server preparing completion AST");
1067 ptlrpc_request_set_replen(req);
1069 req->rq_send_state = LUSTRE_IMP_FULL;
1070 /* ptlrpc_request_pack already set timeout */
1072 req->rq_timeout = ldlm_get_rq_timeout();
1074 /* We only send real blocking ASTs after the lock is granted */
1075 lock_res_and_lock(lock);
1076 if (ldlm_is_ast_sent(lock)) {
1077 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1078 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1079 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1083 * We might get here prior to ldlm_handle_enqueue setting
1084 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1085 * into waiting list, but this is safe and similar code in
1086 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1087 * that would not only cancel the lock, but will also remove
1088 * it from waiting list
1090 if (ldlm_is_cancel_on_block(lock)) {
1091 unlock_res_and_lock(lock);
1092 ldlm_lock_cancel(lock);
1095 req->rq_no_resend = 1;
1097 lock_res_and_lock(lock);
1099 /* start the lock-timeout clock */
1100 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1101 /* Do not resend after lock callback timeout */
1102 req->rq_delay_limit = ldlm_bl_timeout(lock);
1103 req->rq_resend_cb = ldlm_update_resend;
1106 unlock_res_and_lock(lock);
1108 if (lock->l_export && lock->l_export->exp_nid_stats &&
1109 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1110 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1111 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1113 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1115 RETURN(lvb_len < 0 ? lvb_len : rc);
1119 * Server side ->l_glimpse_ast handler for client locks.
1121 * Sends glimpse AST to the client and waits for reply. Then updates
1122 * lvbo with the result.
1124 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1126 struct ldlm_cb_set_arg *arg = data;
1127 struct ldlm_request *body;
1128 struct ptlrpc_request *req;
1129 struct ldlm_cb_async_args *ca;
1131 struct req_format *req_fmt;
1135 LASSERT(lock != NULL);
1137 if (arg->gl_desc != NULL)
1138 /* There is a glimpse descriptor to pack */
1139 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1141 req_fmt = &RQF_LDLM_GL_CALLBACK;
1143 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1144 req_fmt, LUSTRE_DLM_VERSION,
1150 if (arg->gl_desc != NULL) {
1151 /* copy the GL descriptor */
1152 union ldlm_gl_desc *desc;
1154 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1155 *desc = *arg->gl_desc;
1158 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1159 body->lock_handle[0] = lock->l_remote_handle;
1160 ldlm_lock2desc(lock, &body->lock_desc);
1162 ca = ptlrpc_req_async_args(ca, req);
1163 ca->ca_set_arg = arg;
1166 /* server namespace, doesn't need lock */
1167 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1168 ldlm_lvbo_size(lock));
1169 ptlrpc_request_set_replen(req);
1171 req->rq_send_state = LUSTRE_IMP_FULL;
1172 /* ptlrpc_request_alloc_pack already set timeout */
1174 req->rq_timeout = ldlm_get_rq_timeout();
1176 req->rq_interpret_reply = ldlm_cb_interpret;
1178 if (lock->l_export && lock->l_export->exp_nid_stats) {
1179 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1181 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1182 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1185 rc = ldlm_ast_fini(req, arg, lock, 0);
1189 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1191 int ldlm_glimpse_locks(struct ldlm_resource *res,
1192 struct list_head *gl_work_list)
1198 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1200 if (rc == -ERESTART)
1201 ldlm_reprocess_all(res, NULL);
1205 EXPORT_SYMBOL(ldlm_glimpse_locks);
1207 /* return LDLM lock associated with a lock callback request */
1208 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1210 struct ldlm_cb_async_args *ca;
1211 struct ldlm_lock *lock;
1215 ca = ptlrpc_req_async_args(ca, req);
1218 RETURN(ERR_PTR(-EFAULT));
1222 EXPORT_SYMBOL(ldlm_request_lock);
1225 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1226 * service threads to carry out client lock enqueueing requests.
1228 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1229 struct ptlrpc_request *req,
1230 const struct ldlm_request *dlm_req,
1231 const struct ldlm_callback_suite *cbs)
1233 struct ldlm_reply *dlm_rep;
1235 enum ldlm_error err = ELDLM_OK;
1236 struct ldlm_lock *lock = NULL;
1237 void *cookie = NULL;
1239 struct ldlm_resource *res = NULL;
1240 const struct lu_env *env = req->rq_svc_thread->t_env;
1244 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1246 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1247 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1249 LASSERT(req->rq_export);
1251 /* for intent enqueue the stat will be updated inside intent policy */
1252 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1253 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1254 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1256 if (req->rq_export && req->rq_export->exp_nid_stats &&
1257 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1258 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1259 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1261 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1262 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1263 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1264 dlm_req->lock_desc.l_resource.lr_type);
1265 GOTO(out, rc = -EFAULT);
1268 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1269 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1270 dlm_req->lock_desc.l_req_mode &
1271 (dlm_req->lock_desc.l_req_mode-1))) {
1272 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1273 dlm_req->lock_desc.l_req_mode);
1274 GOTO(out, rc = -EFAULT);
1277 if (unlikely((flags & LDLM_FL_REPLAY) ||
1278 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1279 /* Find an existing lock in the per-export lock hash */
1281 * In the function below, .hs_keycmp resolves to
1282 * ldlm_export_lock_keycmp()
1284 /* coverity[overrun-buffer-val] */
1285 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1286 (void *)&dlm_req->lock_handle[0]);
1288 DEBUG_REQ(D_DLMTRACE, req,
1289 "found existing lock cookie %#llx",
1290 lock->l_handle.h_cookie);
1291 flags |= LDLM_FL_RESENT;
1292 GOTO(existing_lock, rc = 0);
1295 if (ldlm_reclaim_full()) {
1296 DEBUG_REQ(D_DLMTRACE, req,
1297 "Too many granted locks, reject current enqueue request and let the client retry later");
1298 GOTO(out, rc = -EINPROGRESS);
1302 /* The lock's callback data might be set in the policy function */
1303 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1304 dlm_req->lock_desc.l_resource.lr_type,
1305 dlm_req->lock_desc.l_req_mode,
1306 cbs, NULL, 0, LVB_T_NONE);
1313 lock->l_remote_handle = dlm_req->lock_handle[0];
1314 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1317 * Initialize resource lvb but not for a lock being replayed since
1318 * Client already got lvb sent in this case.
1319 * This must occur early since some policy methods assume resource
1320 * lvb is available (lr_lvb_data != NULL).
1322 res = lock->l_resource;
1323 if (!(flags & LDLM_FL_REPLAY)) {
1324 /* non-replayed lock, delayed lvb init may need to be done */
1325 rc = ldlm_lvbo_init(res);
1327 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1332 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1334 * Don't enqueue a lock onto the export if it is been disonnected
1335 * due to eviction (b=3822) or server umount (b=24324).
1336 * Cancel it now instead.
1338 if (req->rq_export->exp_disconnected) {
1339 LDLM_ERROR(lock, "lock on disconnected export %p",
1341 GOTO(out, rc = -ENOTCONN);
1344 lock->l_export = class_export_lock_get(req->rq_export, lock);
1345 if (lock->l_export->exp_lock_hash)
1346 cfs_hash_add(lock->l_export->exp_lock_hash,
1347 &lock->l_remote_handle,
1351 * Inherit the enqueue flags before the operation, because we do not
1352 * keep the res lock on return and next operations (BL AST) may proceed
1355 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1356 LDLM_FL_INHERIT_MASK);
1358 ldlm_convert_policy_to_local(req->rq_export,
1359 dlm_req->lock_desc.l_resource.lr_type,
1360 &dlm_req->lock_desc.l_policy_data,
1361 &lock->l_policy_data);
1362 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1363 lock->l_req_extent = lock->l_policy_data.l_extent;
1364 else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS)
1365 lock->l_policy_data.l_inodebits.try_bits =
1366 dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1370 if (!(flags & LDLM_FL_HAS_INTENT)) {
1371 /* based on the assumption that lvb size never changes during
1372 * resource life time otherwise it need resource->lr_lock's
1374 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1375 RCL_SERVER, ldlm_lvbo_size(lock));
1377 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1378 GOTO(out, rc = -ENOMEM);
1380 rc = req_capsule_server_pack(&req->rq_pill);
1385 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1392 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1394 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1395 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1397 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1398 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1401 * We never send a blocking AST until the lock is granted, but
1402 * we can tell it right now
1404 lock_res_and_lock(lock);
1407 * Now take into account flags to be inherited from original lock
1408 * request both in reply to client and in our own lock flags.
1410 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1411 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1414 * Don't move a pending lock onto the export if it has already been
1415 * disconnected due to eviction (b=5683) or server umount (b=24324).
1416 * Cancel it now instead.
1418 if (unlikely(req->rq_export->exp_disconnected ||
1419 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1420 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1422 } else if (ldlm_is_ast_sent(lock)) {
1423 /* fill lock desc for possible lock convert */
1424 if (lock->l_blocking_lock &&
1425 lock->l_resource->lr_type == LDLM_IBITS) {
1426 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1427 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1430 "save blocking bits %llx in granted lock",
1431 bl_lock->l_policy_data.l_inodebits.bits);
1433 * If lock is blocked then save blocking ibits
1434 * in returned lock policy for the possible lock
1435 * convert on a client.
1437 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1438 bl_lock->l_policy_data.l_inodebits.bits;
1440 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1441 if (ldlm_is_granted(lock)) {
1443 * Only cancel lock if it was granted, because it would
1444 * be destroyed immediately and would never be granted
1445 * in the future, causing timeouts on client. Not
1446 * granted lock will be cancelled immediately after
1447 * sending completion AST.
1449 if (ldlm_is_cancel_on_block(lock)) {
1450 unlock_res_and_lock(lock);
1451 ldlm_lock_cancel(lock);
1452 lock_res_and_lock(lock);
1454 ldlm_add_waiting_lock(lock,
1455 ldlm_bl_timeout(lock));
1459 unlock_res_and_lock(lock);
1463 req->rq_status = rc ?: err; /* return either error - b=11190 */
1464 if (!req->rq_packed_final) {
1465 err = lustre_pack_reply(req, 1, NULL, NULL);
1471 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1472 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1476 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1480 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1482 ldlm_lvbo_size(lock) > 0) {
1487 buf = req_capsule_server_get(&req->rq_pill,
1489 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1490 buflen = req_capsule_get_size(&req->rq_pill,
1491 &RMF_DLM_LVB, RCL_SERVER);
1493 * non-replayed lock, delayed lvb init may
1494 * need to be occur now
1496 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1499 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1501 req_capsule_shrink(&req->rq_pill,
1504 } else if (rc2 == -ERANGE) {
1505 rc2 = req_capsule_server_grow(
1507 &RMF_DLM_LVB, buflen);
1512 * if we can't grow the buffer,
1513 * it's ok to return empty lvb
1524 } else if (flags & LDLM_FL_REPLAY) {
1525 /* no LVB resend upon replay */
1527 req_capsule_shrink(&req->rq_pill,
1537 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1538 if (lock->l_export) {
1539 ldlm_lock_cancel(lock);
1541 lock_res_and_lock(lock);
1542 ldlm_resource_unlink_lock(lock);
1543 ldlm_lock_destroy_nolock(lock);
1544 unlock_res_and_lock(lock);
1549 if (!err && !ldlm_is_cbpending(lock) &&
1550 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1551 ldlm_reprocess_all(lock->l_resource, lock);
1553 LDLM_LOCK_RELEASE(lock);
1556 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1563 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1564 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1566 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1568 if (lock->l_blocking_lock) {
1569 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1570 lock->l_blocking_lock = NULL;
1574 /* A lock can be converted to new ibits or mode and should be considered
1575 * as new lock. Clear all states related to a previous blocking AST
1576 * processing so new conflicts will cause new blocking ASTs.
1578 * This is used during lock convert below and lock downgrade to COS mode in
1579 * ldlm_lock_mode_downgrade().
1581 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1583 ldlm_clear_ast_sent(lock);
1584 lock->l_bl_ast_run = 0;
1585 ldlm_clear_blocking_lock(lock);
1589 * Main LDLM entry point for server code to process lock conversion requests.
1591 int ldlm_handle_convert0(struct ptlrpc_request *req,
1592 const struct ldlm_request *dlm_req)
1594 struct obd_export *exp = req->rq_export;
1595 struct ldlm_reply *dlm_rep;
1596 struct ldlm_lock *lock;
1603 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1604 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1605 LDLM_CONVERT - LDLM_FIRST_OPC);
1607 rc = req_capsule_server_pack(&req->rq_pill);
1611 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1612 dlm_rep->lock_flags = dlm_req->lock_flags;
1614 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1616 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1617 req->rq_status = ELDLM_NO_LOCK_DATA;
1621 LDLM_DEBUG(lock, "server-side convert handler START");
1623 lock_res_and_lock(lock);
1624 bits = lock->l_policy_data.l_inodebits.bits;
1625 new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1627 if (ldlm_is_cancel(lock)) {
1628 LDLM_DEBUG(lock, "convert on canceled lock!");
1629 unlock_res_and_lock(lock);
1630 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1633 if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1634 LDLM_ERROR(lock, "lock mode differs!");
1635 unlock_res_and_lock(lock);
1636 GOTO(out_put, rc = -EPROTO);
1639 if (bits == new_bits) {
1641 * This can be valid situation if CONVERT RPCs are
1642 * re-ordered. Just finish silently
1644 LDLM_DEBUG(lock, "lock is converted already!");
1645 unlock_res_and_lock(lock);
1647 if (ldlm_is_waited(lock))
1648 ldlm_del_waiting_lock(lock);
1650 ldlm_clear_cbpending(lock);
1651 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1652 ldlm_inodebits_drop(lock, bits & ~new_bits);
1654 ldlm_clear_blocking_data(lock);
1655 unlock_res_and_lock(lock);
1657 ldlm_reprocess_all(lock->l_resource, NULL);
1660 dlm_rep->lock_handle = lock->l_remote_handle;
1661 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1662 &dlm_rep->lock_desc.l_policy_data);
1666 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1667 LDLM_LOCK_PUT(lock);
1668 req->rq_status = rc;
1673 * Cancel all the locks whose handles are packed into ldlm_request
1675 * Called by server code expecting such combined cancel activity
1678 int ldlm_request_cancel(struct ptlrpc_request *req,
1679 const struct ldlm_request *dlm_req,
1680 int first, enum lustre_at_flags flags)
1682 struct ldlm_resource *res, *pres = NULL;
1683 struct ldlm_lock *lock;
1684 int i, count, done = 0;
1689 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1690 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1691 (size - offsetof(struct ldlm_request, lock_handle)) /
1692 sizeof(struct lustre_handle) < dlm_req->lock_count)
1695 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1699 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1703 * There is no lock on the server at the replay time,
1704 * skip lock cancelling to make replay tests to pass.
1706 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1709 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1712 for (i = first; i < count; i++) {
1713 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1715 /* below message checked in replay-single.sh test_36 */
1716 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1717 dlm_req->lock_handle[i].cookie);
1721 res = lock->l_resource;
1725 * This code is an optimization to only attempt lock
1726 * granting on the resource (that could be CPU-expensive)
1727 * after we are done cancelling lock in that resource.
1731 ldlm_reprocess_all(pres, NULL);
1732 LDLM_RESOURCE_DELREF(pres);
1733 ldlm_resource_putref(pres);
1736 ldlm_resource_getref(res);
1737 LDLM_RESOURCE_ADDREF(res);
1739 if (!ldlm_is_discard_data(lock))
1740 ldlm_lvbo_update(res, lock,
1746 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1747 lock->l_blast_sent != 0) {
1748 timeout_t delay = 0;
1750 if (ktime_get_real_seconds() > lock->l_blast_sent)
1751 delay = ktime_get_real_seconds() -
1754 "server cancels blocked lock after %ds",
1756 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1758 ldlm_lock_cancel(lock);
1759 LDLM_LOCK_PUT(lock);
1762 ldlm_reprocess_all(pres, NULL);
1763 LDLM_RESOURCE_DELREF(pres);
1764 ldlm_resource_putref(pres);
1766 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1769 EXPORT_SYMBOL(ldlm_request_cancel);
1772 * Main LDLM entry point for server code to cancel locks.
1774 * Typically gets called from service handler on LDLM_CANCEL opc.
1776 int ldlm_handle_cancel(struct ptlrpc_request *req)
1778 struct ldlm_request *dlm_req;
1783 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1784 if (dlm_req == NULL) {
1785 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1789 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1790 offsetof(struct ldlm_request, lock_handle[1]))
1793 if (req->rq_export && req->rq_export->exp_nid_stats &&
1794 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1795 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1796 LDLM_CANCEL - LDLM_FIRST_OPC);
1798 rc = req_capsule_server_pack(&req->rq_pill);
1802 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1803 req->rq_status = LUSTRE_ESTALE;
1805 RETURN(ptlrpc_reply(req));
1807 #endif /* HAVE_SERVER_SUPPORT */
1810 * Server may pass additional information about blocking lock.
1811 * For IBITS locks it is conflicting bits which can be used for
1812 * lock convert instead of cancel.
1814 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1816 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1818 check_res_locked(lock->l_resource);
1819 if (ns_is_client(ns) && ld &&
1820 (lock->l_resource->lr_type == LDLM_IBITS)) {
1822 * Lock description contains policy of blocking lock,
1823 * and its cancel_bits is used to pass conflicting bits.
1824 * NOTE: ld can be NULL or can be not NULL but zeroed if
1825 * passed from ldlm_bl_thread_blwi(), check below used bits
1826 * in ld to make sure it is valid description.
1828 * If server may replace lock resource keeping the same cookie,
1829 * never use cancel bits from different resource, full cancel
1832 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1833 ldlm_res_eq(&ld->l_resource.lr_name,
1834 &lock->l_resource->lr_name) &&
1835 !(ldlm_is_cbpending(lock) &&
1836 lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1837 /* always combine conflicting ibits */
1838 lock->l_policy_data.l_inodebits.cancel_bits |=
1839 ld->l_policy_data.l_inodebits.cancel_bits;
1841 /* If cancel_bits are not obtained or
1842 * if the lock is already CBPENDING and
1843 * has no cancel_bits set
1844 * - the full lock is to be cancelled
1846 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1852 * Callback handler for receiving incoming blocking ASTs.
1854 * This can only happen on client side.
1856 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1857 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1863 LDLM_DEBUG(lock, "client blocking AST callback handler");
1865 lock_res_and_lock(lock);
1867 /* get extra information from desc if any */
1868 ldlm_bl_desc2lock(ld, lock);
1869 ldlm_set_cbpending(lock);
1871 do_ast = (!lock->l_readers && !lock->l_writers);
1872 unlock_res_and_lock(lock);
1876 "Lock %p already unused, calling callback (%p)\n",
1877 lock, lock->l_blocking_ast);
1878 if (lock->l_blocking_ast != NULL)
1879 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1883 "Lock %p is referenced, will be cancelled later\n",
1887 LDLM_DEBUG(lock, "client blocking callback handler END");
1888 LDLM_LOCK_RELEASE(lock);
1892 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1894 if (req->rq_no_reply)
1897 req->rq_status = rc;
1898 if (!req->rq_packed_final) {
1899 rc = lustre_pack_reply(req, 1, NULL, NULL);
1903 return ptlrpc_reply(req);
1907 * Callback handler for receiving incoming completion ASTs.
1909 * This only can happen on client side.
1911 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1912 struct ldlm_namespace *ns,
1913 struct ldlm_request *dlm_req,
1914 struct ldlm_lock *lock)
1916 LIST_HEAD(ast_list);
1922 LDLM_DEBUG(lock, "client completion callback handler START");
1924 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1925 long to = cfs_time_seconds(1);
1927 ldlm_callback_reply(req, 0);
1930 schedule_timeout_interruptible(to);
1931 if (ldlm_is_granted(lock) ||
1932 ldlm_is_destroyed(lock))
1937 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1939 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1940 GOTO(out, rc = lvb_len);
1941 } else if (lvb_len > 0) {
1942 if (lock->l_lvb_len > 0) {
1943 /* for extent lock, lvb contains ost_lvb{}. */
1944 LASSERT(lock->l_lvb_data != NULL);
1946 if (unlikely(lock->l_lvb_len < lvb_len)) {
1948 "Replied LVB is larger than expectation, expected = %d, replied = %d",
1949 lock->l_lvb_len, lvb_len);
1950 GOTO(out, rc = -EINVAL);
1955 lock_res_and_lock(lock);
1957 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1958 &lock->l_resource->lr_name)) {
1959 ldlm_resource_unlink_lock(lock);
1960 unlock_res_and_lock(lock);
1961 rc = ldlm_lock_change_resource(ns, lock,
1962 &dlm_req->lock_desc.l_resource.lr_name);
1964 LDLM_ERROR(lock, "Failed to allocate resource");
1967 LDLM_DEBUG(lock, "completion AST, new resource");
1968 lock_res_and_lock(lock);
1971 if (ldlm_is_failed(lock)) {
1972 unlock_res_and_lock(lock);
1973 LDLM_LOCK_RELEASE(lock);
1977 if (ldlm_is_destroyed(lock) ||
1978 ldlm_is_granted(lock)) {
1979 /* b=11300: the lock has already been granted */
1980 unlock_res_and_lock(lock);
1981 LDLM_DEBUG(lock, "Double grant race happened");
1986 * If we receive the completion AST before the actual enqueue returned,
1987 * then we might need to switch lock modes, resources, or extents.
1989 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1990 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1991 LDLM_DEBUG(lock, "completion AST, new lock mode");
1994 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1995 ldlm_convert_policy_to_local(req->rq_export,
1996 dlm_req->lock_desc.l_resource.lr_type,
1997 &dlm_req->lock_desc.l_policy_data,
1998 &lock->l_policy_data);
1999 LDLM_DEBUG(lock, "completion AST, new policy data");
2002 ldlm_resource_unlink_lock(lock);
2004 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2006 * BL_AST locks are not needed in LRU.
2007 * Let ldlm_cancel_lru() be fast.
2009 ldlm_lock_remove_from_lru(lock);
2010 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2011 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2012 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2015 if (lock->l_lvb_len > 0) {
2016 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2017 lock->l_lvb_data, lvb_len);
2019 unlock_res_and_lock(lock);
2024 ldlm_grant_lock(lock, &ast_list);
2025 unlock_res_and_lock(lock);
2027 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2030 * Let Enqueue to call osc_lock_upcall() and initialize
2033 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2035 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2037 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2043 lock_res_and_lock(lock);
2044 ldlm_set_failed(lock);
2045 unlock_res_and_lock(lock);
2046 wake_up(&lock->l_waitq);
2048 LDLM_LOCK_RELEASE(lock);
2054 * Callback handler for receiving incoming glimpse ASTs.
2056 * This only can happen on client side. After handling the glimpse AST
2057 * we also consider dropping the lock here if it is unused locally for a
2060 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2061 struct ldlm_namespace *ns,
2062 struct ldlm_request *dlm_req,
2063 struct ldlm_lock *lock)
2065 struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2070 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2072 if (lock->l_glimpse_ast != NULL)
2073 rc = lock->l_glimpse_ast(lock, req);
2075 if (req->rq_repmsg != NULL) {
2078 req->rq_status = rc;
2082 lock_res_and_lock(lock);
2083 if (lock->l_granted_mode == LCK_PW &&
2084 !lock->l_readers && !lock->l_writers &&
2085 ktime_after(ktime_get(),
2086 ktime_add(lock->l_last_used, ns->ns_dirty_age_limit))) {
2087 unlock_res_and_lock(lock);
2089 /* For MDS glimpse it is always DOM lock, set corresponding
2090 * cancel_bits to perform lock convert if needed
2092 if (lock->l_resource->lr_type == LDLM_IBITS)
2093 ld->l_policy_data.l_inodebits.cancel_bits =
2095 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2096 ldlm_handle_bl_callback(ns, ld, lock);
2101 unlock_res_and_lock(lock);
2102 LDLM_LOCK_RELEASE(lock);
2106 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2107 enum ldlm_cancel_flags cancel_flags)
2109 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2113 spin_lock(&blp->blp_lock);
2114 if (blwi->blwi_lock &&
2115 ldlm_is_discard_data(blwi->blwi_lock)) {
2116 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
2117 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2119 /* other blocking callbacks are added to the regular list */
2120 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2122 spin_unlock(&blp->blp_lock);
2124 wake_up(&blp->blp_waitq);
2127 * can not check blwi->blwi_flags as blwi could be already freed in
2130 if (!(cancel_flags & LCF_ASYNC))
2131 wait_for_completion(&blwi->blwi_comp);
2136 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2137 struct ldlm_namespace *ns,
2138 struct ldlm_lock_desc *ld,
2139 struct list_head *cancels, int count,
2140 struct ldlm_lock *lock,
2141 enum ldlm_cancel_flags cancel_flags)
2143 init_completion(&blwi->blwi_comp);
2144 INIT_LIST_HEAD(&blwi->blwi_head);
2146 if (current->flags & PF_MEMALLOC)
2147 blwi->blwi_mem_pressure = 1;
2150 blwi->blwi_flags = cancel_flags;
2152 blwi->blwi_ld = *ld;
2154 list_splice_init(cancels, &blwi->blwi_head);
2155 blwi->blwi_count = count;
2157 blwi->blwi_lock = lock;
2162 * Queues a list of locks \a cancels containing \a count locks
2163 * for later processing by a blocking thread. If \a count is zero,
2164 * then the lock referenced as \a lock is queued instead.
2166 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2167 * If list addition fails an error is returned and caller is supposed to
2168 * call ->l_blocking_ast itself.
2170 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2171 struct ldlm_lock_desc *ld,
2172 struct ldlm_lock *lock,
2173 struct list_head *cancels, int count,
2174 enum ldlm_cancel_flags cancel_flags)
2178 if (cancels && count == 0)
2181 if (cancel_flags & LCF_ASYNC) {
2182 struct ldlm_bl_work_item *blwi;
2184 OBD_ALLOC(blwi, sizeof(*blwi));
2187 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2189 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2192 * if it is synchronous call do minimum mem alloc, as it could
2193 * be triggered from kernel shrinker
2195 struct ldlm_bl_work_item blwi;
2197 memset(&blwi, 0, sizeof(blwi));
2198 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2199 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2204 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2205 struct ldlm_lock *lock)
2207 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2210 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2211 struct list_head *cancels, int count,
2212 enum ldlm_cancel_flags cancel_flags)
2214 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2217 int ldlm_bl_to_thread_ns(struct ldlm_namespace *ns)
2219 return ldlm_bl_to_thread(ns, NULL, NULL, NULL, 0, LCF_ASYNC);
2222 int ldlm_bl_thread_wakeup(void)
2224 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2228 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2229 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2231 struct obd_device *obd = req->rq_export->exp_obd;
2239 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2241 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2243 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2245 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2248 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2250 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2252 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2255 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2258 /* We are responsible for swabbing contents of val */
2260 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2261 /* Pass it on to mdc (the "export" in this case) */
2262 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2264 sizeof(KEY_HSM_COPYTOOL_SEND),
2265 KEY_HSM_COPYTOOL_SEND,
2268 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2273 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2274 const char *msg, int rc,
2275 const struct lustre_handle *handle)
2277 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2278 "%s, NID=%s lock=%#llx: rc = %d",
2279 msg, libcfs_id2str(req->rq_peer),
2280 handle ? handle->cookie : 0, rc);
2281 if (req->rq_no_reply)
2282 CWARN("No reply was sent, maybe cause b=21636.\n");
2284 CWARN("Send reply failed, maybe cause b=21636.\n");
2287 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2288 static int ldlm_callback_handler(struct ptlrpc_request *req)
2290 struct ldlm_namespace *ns;
2291 struct ldlm_request *dlm_req;
2292 struct ldlm_lock *lock;
2298 * Requests arrive in sender's byte order. The ptlrpc service
2299 * handler has already checked and, if necessary, byte-swapped the
2300 * incoming request message body, but I am responsible for the
2304 /* do nothing for sec context finalize */
2305 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2308 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2310 if (req->rq_export == NULL) {
2311 rc = ldlm_callback_reply(req, -ENOTCONN);
2312 ldlm_callback_errmsg(req, "Operate on unconnected server",
2317 LASSERT(req->rq_export != NULL);
2318 LASSERT(req->rq_export->exp_obd != NULL);
2320 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2321 case LDLM_BL_CALLBACK:
2322 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2324 ldlm_callback_reply(req, -(int)cfs_fail_err);
2328 case LDLM_CP_CALLBACK:
2329 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2332 case LDLM_GL_CALLBACK:
2333 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2337 rc = ldlm_handle_setinfo(req);
2338 ldlm_callback_reply(req, rc);
2341 CERROR("unknown opcode %u\n",
2342 lustre_msg_get_opc(req->rq_reqmsg));
2343 ldlm_callback_reply(req, -EPROTO);
2347 ns = req->rq_export->exp_obd->obd_namespace;
2348 LASSERT(ns != NULL);
2350 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2352 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2353 if (dlm_req == NULL) {
2354 rc = ldlm_callback_reply(req, -EPROTO);
2355 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2361 * Force a known safe race, send a cancel to the server for a lock
2362 * which the server has already started a blocking callback on.
2364 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2365 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2366 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2368 CERROR("ldlm_cli_cancel: %d\n", rc);
2371 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2374 "callback on lock %#llx - lock disappeared\n",
2375 dlm_req->lock_handle[0].cookie);
2376 rc = ldlm_callback_reply(req, -EINVAL);
2377 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2378 &dlm_req->lock_handle[0]);
2382 if (ldlm_is_fail_loc(lock) &&
2383 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2384 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2386 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2387 lock_res_and_lock(lock);
2388 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2390 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2392 * If somebody cancels lock and cache is already dropped,
2393 * or lock is failed before cp_ast received on client,
2394 * we can tell the server we have no lock. Otherwise, we
2395 * should send cancel after dropping the cache.
2397 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2398 ldlm_is_failed(lock)) {
2400 "callback on lock %llx - lock disappeared",
2401 dlm_req->lock_handle[0].cookie);
2402 unlock_res_and_lock(lock);
2403 LDLM_LOCK_RELEASE(lock);
2404 rc = ldlm_callback_reply(req, -EINVAL);
2405 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2406 &dlm_req->lock_handle[0]);
2410 * BL_AST locks are not needed in LRU.
2411 * Let ldlm_cancel_lru() be fast.
2413 ldlm_lock_remove_from_lru(lock);
2414 ldlm_set_bl_ast(lock);
2416 unlock_res_and_lock(lock);
2419 * We want the ost thread to get this reply so that it can respond
2420 * to ost requests (write cache writeback) that might be triggered
2423 * But we'd also like to be able to indicate in the reply that we're
2424 * cancelling right now, because it's unused, or have an intent result
2425 * in the reply, so we might have to push the responsibility for sending
2426 * the reply down into the AST handlers, alas.
2429 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2430 case LDLM_BL_CALLBACK:
2431 CDEBUG(D_INODE, "blocking ast\n");
2432 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2433 if (!ldlm_is_cancel_on_block(lock)) {
2434 rc = ldlm_callback_reply(req, 0);
2435 if (req->rq_no_reply || rc)
2436 ldlm_callback_errmsg(req, "Normal process", rc,
2437 &dlm_req->lock_handle[0]);
2439 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2440 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2442 case LDLM_CP_CALLBACK:
2443 CDEBUG(D_INODE, "completion ast\n");
2444 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2445 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2446 if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2447 ldlm_callback_reply(req, rc);
2449 case LDLM_GL_CALLBACK:
2450 CDEBUG(D_INODE, "glimpse ast\n");
2451 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2452 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2455 LBUG(); /* checked above */
2461 #ifdef HAVE_SERVER_SUPPORT
2463 * Main handler for canceld thread.
2465 * Separated into its own thread to avoid deadlocks.
2467 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2474 * Requests arrive in sender's byte order. The ptlrpc service
2475 * handler has already checked and, if necessary, byte-swapped the
2476 * incoming request message body, but I am responsible for the
2480 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2482 if (req->rq_export == NULL) {
2483 struct ldlm_request *dlm_req;
2485 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2486 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2487 libcfs_nid2str(req->rq_peer.nid),
2488 (unsigned long long)req->rq_arrival_time.tv_sec,
2489 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2491 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2492 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2493 dlm_req = req_capsule_client_get(&req->rq_pill,
2495 if (dlm_req != NULL)
2496 ldlm_lock_dump_handle(D_ERROR,
2497 &dlm_req->lock_handle[0]);
2499 ldlm_callback_reply(req, -ENOTCONN);
2503 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2504 /* XXX FIXME move this back to mds/handler.c, b=249 */
2506 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2507 CDEBUG(D_INODE, "cancel\n");
2508 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2509 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2510 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2512 rc = ldlm_handle_cancel(req);
2516 struct ldlm_request *dlm_req;
2518 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2519 CDEBUG(D_INODE, "convert\n");
2521 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2522 if (dlm_req == NULL) {
2523 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2524 rc = ldlm_callback_reply(req, -EPROTO);
2526 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2527 rc = ptlrpc_reply(req);
2532 CERROR("invalid opcode %d\n",
2533 lustre_msg_get_opc(req->rq_reqmsg));
2534 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2535 rc = ldlm_callback_reply(req, -EINVAL);
2541 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2542 struct ldlm_lock *lock)
2544 struct ldlm_request *dlm_req;
2545 struct lustre_handle lockh;
2551 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2552 if (dlm_req == NULL)
2555 ldlm_lock2handle(lock, &lockh);
2556 for (i = 0; i < dlm_req->lock_count; i++) {
2557 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2559 DEBUG_REQ(D_RPCTRACE, req,
2560 "Prio raised by lock %#llx", lockh.cookie);
2569 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2571 struct ldlm_request *dlm_req;
2578 /* no prolong in recovery */
2579 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2582 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2583 if (dlm_req == NULL)
2586 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2587 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2588 (size - offsetof(struct ldlm_request, lock_handle)) /
2589 sizeof(struct lustre_handle) < dlm_req->lock_count)
2592 for (i = 0; i < dlm_req->lock_count; i++) {
2593 struct ldlm_lock *lock;
2595 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2599 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2601 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2602 LDLM_LOCK_PUT(lock);
2611 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2612 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2613 .hpreq_check = ldlm_cancel_hpreq_check,
2617 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2621 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2623 if (req->rq_export == NULL)
2626 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2627 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2628 req->rq_ops = &ldlm_cancel_hpreq_ops;
2629 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2630 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2631 req->rq_ops = &ldlm_cancel_hpreq_ops;
2636 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2637 struct hlist_node *hnode, void *data)
2640 struct list_head *rpc_list = data;
2641 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2643 lock_res_and_lock(lock);
2645 if (!ldlm_is_granted(lock)) {
2646 unlock_res_and_lock(lock);
2650 LASSERT(lock->l_resource);
2651 if (lock->l_resource->lr_type != LDLM_IBITS &&
2652 lock->l_resource->lr_type != LDLM_PLAIN) {
2653 unlock_res_and_lock(lock);
2657 if (ldlm_is_ast_sent(lock)) {
2658 unlock_res_and_lock(lock);
2662 LASSERT(lock->l_blocking_ast);
2663 LASSERT(!lock->l_blocking_lock);
2665 ldlm_set_ast_sent(lock);
2666 if (lock->l_export && lock->l_export->exp_lock_hash) {
2668 * NB: it's safe to call cfs_hash_del() even lock isn't
2672 * In the function below, .hs_keycmp resolves to
2673 * ldlm_export_lock_keycmp()
2675 /* coverity[overrun-buffer-val] */
2676 cfs_hash_del(lock->l_export->exp_lock_hash,
2677 &lock->l_remote_handle, &lock->l_exp_hash);
2680 list_add_tail(&lock->l_rk_ast, rpc_list);
2681 LDLM_LOCK_GET(lock);
2683 unlock_res_and_lock(lock);
2687 void ldlm_revoke_export_locks(struct obd_export *exp)
2689 LIST_HEAD(rpc_list);
2693 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2694 ldlm_revoke_lock_cb, &rpc_list, 0);
2695 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2696 LDLM_WORK_REVOKE_AST);
2700 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2701 #endif /* HAVE_SERVER_SUPPORT */
2703 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2704 struct ldlm_bl_work_item **p_blwi,
2705 struct obd_export **p_exp)
2707 struct ldlm_bl_work_item *blwi = NULL;
2708 static unsigned int num_bl;
2709 static unsigned int num_stale;
2710 int num_th = atomic_read(&blp->blp_num_threads);
2712 *p_exp = obd_stale_export_get();
2714 spin_lock(&blp->blp_lock);
2715 if (*p_exp != NULL) {
2716 if (num_th == 1 || ++num_stale < num_th) {
2717 spin_unlock(&blp->blp_lock);
2723 /* process a request from the blp_list at least every blp_num_threads */
2724 if (!list_empty(&blp->blp_list) &&
2725 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2726 blwi = list_entry(blp->blp_list.next,
2727 struct ldlm_bl_work_item, blwi_entry);
2729 if (!list_empty(&blp->blp_prio_list))
2730 blwi = list_entry(blp->blp_prio_list.next,
2731 struct ldlm_bl_work_item,
2735 if (++num_bl >= num_th)
2737 list_del(&blwi->blwi_entry);
2739 spin_unlock(&blp->blp_lock);
2742 if (*p_exp != NULL && *p_blwi != NULL) {
2743 obd_stale_export_put(*p_exp);
2747 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2750 /* This only contains temporary data until the thread starts */
2751 struct ldlm_bl_thread_data {
2752 struct ldlm_bl_pool *bltd_blp;
2753 struct completion bltd_comp;
2757 static int ldlm_bl_thread_main(void *arg);
2759 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2761 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2762 struct task_struct *task;
2764 init_completion(&bltd.bltd_comp);
2766 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2767 if (bltd.bltd_num >= blp->blp_max_threads) {
2768 atomic_dec(&blp->blp_num_threads);
2772 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2774 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2775 atomic_dec(&blp->blp_num_threads);
2779 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2782 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2783 bltd.bltd_num, PTR_ERR(task));
2784 atomic_dec(&blp->blp_num_threads);
2785 return PTR_ERR(task);
2787 wait_for_completion(&bltd.bltd_comp);
2792 /* Not fatal if racy and have a few too many threads */
2793 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2794 struct ldlm_bl_work_item *blwi)
2796 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2799 if (atomic_read(&blp->blp_busy_threads) <
2800 atomic_read(&blp->blp_num_threads))
2803 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2804 blwi->blwi_mem_pressure))
2810 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2811 struct ldlm_bl_work_item *blwi)
2813 /* '1' for consistency with code that checks !mpflag to restore */
2814 unsigned int mpflags = 1;
2818 if (blwi->blwi_ns == NULL)
2819 /* added by ldlm_cleanup() */
2820 RETURN(LDLM_ITER_STOP);
2822 if (blwi->blwi_mem_pressure)
2823 mpflags = memalloc_noreclaim_save();
2825 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2827 if (blwi->blwi_count) {
2830 * The special case when we cancel locks in lru
2831 * asynchronously, we pass the list of locks here.
2832 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2833 * canceled locally yet.
2835 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2838 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2840 } else if (blwi->blwi_lock) {
2841 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2844 ldlm_pool_recalc(&blwi->blwi_ns->ns_pool, true);
2845 spin_lock(&blwi->blwi_ns->ns_lock);
2846 blwi->blwi_ns->ns_rpc_recalc = 0;
2847 spin_unlock(&blwi->blwi_ns->ns_lock);
2848 ldlm_namespace_put(blwi->blwi_ns);
2851 if (blwi->blwi_mem_pressure)
2852 memalloc_noreclaim_restore(mpflags);
2854 if (blwi->blwi_flags & LCF_ASYNC)
2855 OBD_FREE(blwi, sizeof(*blwi));
2857 complete(&blwi->blwi_comp);
2863 * Cancel stale locks on export. Cancel blocked locks first.
2864 * If the given export has blocked locks, the next in the list may have
2865 * them too, thus cancel not blocked locks only if the current export has
2868 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2869 struct obd_export *exp)
2875 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2877 num = ldlm_export_cancel_blocked_locks(exp);
2879 ldlm_export_cancel_locks(exp);
2881 obd_stale_export_put(exp);
2888 * Main blocking requests processing thread.
2890 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2891 * This thread in the end ends up doing actual call to ->l_blocking_ast
2894 static int ldlm_bl_thread_main(void *arg)
2897 struct ldlm_bl_pool *blp;
2898 struct ldlm_bl_thread_data *bltd = arg;
2906 rc = lu_env_init(env, LCT_DT_THREAD);
2909 rc = lu_env_add(env);
2911 GOTO(out_env_fini, rc);
2913 blp = bltd->bltd_blp;
2915 complete(&bltd->bltd_comp);
2916 /* cannot use bltd after this, it is only on caller's stack */
2919 struct ldlm_bl_work_item *blwi = NULL;
2920 struct obd_export *exp = NULL;
2923 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2926 wait_event_idle_exclusive(blp->blp_waitq,
2927 ldlm_bl_get_work(blp, &blwi,
2929 atomic_inc(&blp->blp_busy_threads);
2931 if (ldlm_bl_thread_need_create(blp, blwi))
2932 /* discard the return value, we tried */
2933 ldlm_bl_thread_start(blp, true);
2936 rc = ldlm_bl_thread_exports(blp, exp);
2938 rc = ldlm_bl_thread_blwi(blp, blwi);
2940 atomic_dec(&blp->blp_busy_threads);
2942 if (rc == LDLM_ITER_STOP)
2946 * If there are many namespaces, we will not sleep waiting for
2947 * work, and must do a cond_resched to avoid holding the CPU
2953 atomic_dec(&blp->blp_num_threads);
2954 complete(&blp->blp_comp);
2965 static int ldlm_setup(void);
2966 static int ldlm_cleanup(void);
2968 int ldlm_get_ref(void)
2973 mutex_lock(&ldlm_ref_mutex);
2974 if (++ldlm_refcount == 1) {
2979 mutex_unlock(&ldlm_ref_mutex);
2984 void ldlm_put_ref(void)
2987 mutex_lock(&ldlm_ref_mutex);
2988 if (ldlm_refcount == 1) {
2989 int rc = ldlm_cleanup();
2992 CERROR("ldlm_cleanup failed: %d\n", rc);
2998 mutex_unlock(&ldlm_ref_mutex);
3004 * Export handle<->lock hash operations.
3007 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
3009 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
3013 ldlm_export_lock_key(struct hlist_node *hnode)
3015 struct ldlm_lock *lock;
3017 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3018 return &lock->l_remote_handle;
3022 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3024 struct ldlm_lock *lock;
3026 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3027 lock->l_remote_handle = *(struct lustre_handle *)key;
3031 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3033 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3037 ldlm_export_lock_object(struct hlist_node *hnode)
3039 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3043 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3045 struct ldlm_lock *lock;
3047 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3048 LDLM_LOCK_GET(lock);
3052 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3054 struct ldlm_lock *lock;
3056 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3057 LDLM_LOCK_RELEASE(lock);
3060 static struct cfs_hash_ops ldlm_export_lock_ops = {
3061 .hs_hash = ldlm_export_lock_hash,
3062 .hs_key = ldlm_export_lock_key,
3063 .hs_keycmp = ldlm_export_lock_keycmp,
3064 .hs_keycpy = ldlm_export_lock_keycpy,
3065 .hs_object = ldlm_export_lock_object,
3066 .hs_get = ldlm_export_lock_get,
3067 .hs_put = ldlm_export_lock_put,
3068 .hs_put_locked = ldlm_export_lock_put,
3071 int ldlm_init_export(struct obd_export *exp)
3077 exp->exp_lock_hash =
3078 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3079 HASH_EXP_LOCK_CUR_BITS,
3080 HASH_EXP_LOCK_MAX_BITS,
3081 HASH_EXP_LOCK_BKT_BITS, 0,
3082 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3083 &ldlm_export_lock_ops,
3084 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3085 CFS_HASH_NBLK_CHANGE);
3087 if (!exp->exp_lock_hash)
3090 rc = ldlm_init_flock_export(exp);
3096 ldlm_destroy_export(exp);
3099 EXPORT_SYMBOL(ldlm_init_export);
3101 void ldlm_destroy_export(struct obd_export *exp)
3104 cfs_hash_putref(exp->exp_lock_hash);
3105 exp->exp_lock_hash = NULL;
3107 ldlm_destroy_flock_export(exp);
3110 EXPORT_SYMBOL(ldlm_destroy_export);
3112 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3113 struct attribute *attr,
3116 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3119 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3120 struct attribute *attr,
3127 rc = kstrtoul(buffer, 10, &val);
3131 ldlm_cancel_unused_locks_before_replay = val;
3135 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3137 static struct attribute *ldlm_attrs[] = {
3138 &lustre_attr_cancel_unused_locks_before_replay.attr,
3142 static struct attribute_group ldlm_attr_group = {
3143 .attrs = ldlm_attrs,
3146 static int ldlm_setup(void)
3148 static struct ptlrpc_service_conf conf;
3149 struct ldlm_bl_pool *blp = NULL;
3150 #ifdef HAVE_SERVER_SUPPORT
3151 struct task_struct *task;
3152 #endif /* HAVE_SERVER_SUPPORT */
3158 if (ldlm_state != NULL)
3161 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3162 if (ldlm_state == NULL)
3165 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3169 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3173 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3177 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3181 rc = ldlm_debugfs_setup();
3185 memset(&conf, 0, sizeof(conf));
3186 conf = (typeof(conf)) {
3187 .psc_name = "ldlm_cbd",
3188 .psc_watchdog_factor = 2,
3190 .bc_nbufs = LDLM_CLIENT_NBUFS,
3191 .bc_buf_size = LDLM_BUFSIZE,
3192 .bc_req_max_size = LDLM_MAXREQSIZE,
3193 .bc_rep_max_size = LDLM_MAXREPSIZE,
3194 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3195 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3198 .tc_thr_name = "ldlm_cb",
3199 .tc_thr_factor = LDLM_THR_FACTOR,
3200 .tc_nthrs_init = LDLM_NTHRS_INIT,
3201 .tc_nthrs_base = LDLM_NTHRS_BASE,
3202 .tc_nthrs_max = LDLM_NTHRS_MAX,
3203 .tc_nthrs_user = ldlm_num_threads,
3204 .tc_cpu_bind = ldlm_cpu_bind,
3205 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3208 .cc_pattern = ldlm_cpts,
3209 .cc_affinity = true,
3212 .so_req_handler = ldlm_callback_handler,
3215 ldlm_state->ldlm_cb_service = \
3216 ptlrpc_register_service(&conf, ldlm_svc_kset,
3217 ldlm_svc_debugfs_dir);
3218 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3219 CERROR("failed to start service\n");
3220 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3221 ldlm_state->ldlm_cb_service = NULL;
3225 #ifdef HAVE_SERVER_SUPPORT
3226 memset(&conf, 0, sizeof(conf));
3227 conf = (typeof(conf)) {
3228 .psc_name = "ldlm_canceld",
3229 .psc_watchdog_factor = 6,
3231 .bc_nbufs = LDLM_SERVER_NBUFS,
3232 .bc_buf_size = LDLM_BUFSIZE,
3233 .bc_req_max_size = LDLM_MAXREQSIZE,
3234 .bc_rep_max_size = LDLM_MAXREPSIZE,
3235 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3236 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3240 .tc_thr_name = "ldlm_cn",
3241 .tc_thr_factor = LDLM_THR_FACTOR,
3242 .tc_nthrs_init = LDLM_NTHRS_INIT,
3243 .tc_nthrs_base = LDLM_NTHRS_BASE,
3244 .tc_nthrs_max = LDLM_NTHRS_MAX,
3245 .tc_nthrs_user = ldlm_num_threads,
3246 .tc_cpu_bind = ldlm_cpu_bind,
3247 .tc_ctx_tags = LCT_MD_THREAD | \
3252 .cc_pattern = ldlm_cpts,
3253 .cc_affinity = true,
3256 .so_req_handler = ldlm_cancel_handler,
3257 .so_hpreq_handler = ldlm_hpreq_handler,
3260 ldlm_state->ldlm_cancel_service = \
3261 ptlrpc_register_service(&conf, ldlm_svc_kset,
3262 ldlm_svc_debugfs_dir);
3263 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3264 CERROR("failed to start service\n");
3265 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3266 ldlm_state->ldlm_cancel_service = NULL;
3269 #endif /* HAVE_SERVER_SUPPORT */
3271 OBD_ALLOC(blp, sizeof(*blp));
3273 GOTO(out, rc = -ENOMEM);
3274 ldlm_state->ldlm_bl_pool = blp;
3276 spin_lock_init(&blp->blp_lock);
3277 INIT_LIST_HEAD(&blp->blp_list);
3278 INIT_LIST_HEAD(&blp->blp_prio_list);
3279 init_waitqueue_head(&blp->blp_waitq);
3280 atomic_set(&blp->blp_num_threads, 0);
3281 atomic_set(&blp->blp_busy_threads, 0);
3283 if (ldlm_num_threads == 0) {
3284 blp->blp_min_threads = LDLM_NTHRS_INIT;
3285 blp->blp_max_threads = LDLM_NTHRS_MAX;
3287 blp->blp_min_threads = blp->blp_max_threads = \
3288 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3292 for (i = 0; i < blp->blp_min_threads; i++) {
3293 rc = ldlm_bl_thread_start(blp, false);
3298 #ifdef HAVE_SERVER_SUPPORT
3299 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3302 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3306 wait_event(expired_lock_wait_queue,
3307 expired_lock_thread_state == ELT_READY);
3308 #endif /* HAVE_SERVER_SUPPORT */
3310 rc = ldlm_pools_init();
3312 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3316 rc = ldlm_reclaim_setup();
3318 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3328 static int ldlm_cleanup(void)
3332 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3333 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3334 CERROR("ldlm still has namespaces; clean these up first.\n");
3335 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3336 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3340 ldlm_reclaim_cleanup();
3343 if (ldlm_state->ldlm_bl_pool != NULL) {
3344 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3346 while (atomic_read(&blp->blp_num_threads) > 0) {
3347 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3349 init_completion(&blp->blp_comp);
3351 spin_lock(&blp->blp_lock);
3352 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3353 wake_up(&blp->blp_waitq);
3354 spin_unlock(&blp->blp_lock);
3356 wait_for_completion(&blp->blp_comp);
3359 OBD_FREE(blp, sizeof(*blp));
3362 if (ldlm_state->ldlm_cb_service != NULL)
3363 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3364 #ifdef HAVE_SERVER_SUPPORT
3365 if (ldlm_state->ldlm_cancel_service != NULL)
3366 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3370 kset_unregister(ldlm_ns_kset);
3372 kset_unregister(ldlm_svc_kset);
3374 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3375 kobject_put(ldlm_kobj);
3378 ldlm_debugfs_cleanup();
3380 #ifdef HAVE_SERVER_SUPPORT
3381 if (expired_lock_thread_state != ELT_STOPPED) {
3382 expired_lock_thread_state = ELT_TERMINATE;
3383 wake_up(&expired_lock_wait_queue);
3384 wait_event(expired_lock_wait_queue,
3385 expired_lock_thread_state == ELT_STOPPED);
3389 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3397 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3398 sizeof(struct ldlm_resource), 0,
3399 SLAB_HWCACHE_ALIGN, NULL);
3400 if (ldlm_resource_slab == NULL)
3403 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3404 sizeof(struct ldlm_lock), 0,
3405 SLAB_HWCACHE_ALIGN, NULL);
3406 if (ldlm_lock_slab == NULL)
3409 ldlm_interval_slab = kmem_cache_create("interval_node",
3410 sizeof(struct ldlm_interval),
3411 0, SLAB_HWCACHE_ALIGN, NULL);
3412 if (ldlm_interval_slab == NULL)
3415 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3416 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3417 0, SLAB_HWCACHE_ALIGN, NULL);
3418 if (ldlm_interval_tree_slab == NULL)
3421 #ifdef HAVE_SERVER_SUPPORT
3422 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3423 sizeof(struct ldlm_ibits_node),
3424 0, SLAB_HWCACHE_ALIGN, NULL);
3425 if (ldlm_inodebits_slab == NULL)
3426 goto out_interval_tree;
3428 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3429 sizeof(struct ldlm_glimpse_work),
3431 if (ldlm_glimpse_work_kmem == NULL)
3435 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3436 class_export_dump_hook = ldlm_dump_export_locks;
3439 #ifdef HAVE_SERVER_SUPPORT
3441 kmem_cache_destroy(ldlm_inodebits_slab);
3443 kmem_cache_destroy(ldlm_interval_tree_slab);
3446 kmem_cache_destroy(ldlm_interval_slab);
3448 kmem_cache_destroy(ldlm_lock_slab);
3450 kmem_cache_destroy(ldlm_resource_slab);
3455 void ldlm_exit(void)
3458 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3459 kmem_cache_destroy(ldlm_resource_slab);
3461 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3462 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3463 * so that ldlm_lock_free() get a chance to be called.
3466 kmem_cache_destroy(ldlm_lock_slab);
3467 kmem_cache_destroy(ldlm_interval_slab);
3468 kmem_cache_destroy(ldlm_interval_tree_slab);
3469 #ifdef HAVE_SERVER_SUPPORT
3470 kmem_cache_destroy(ldlm_inodebits_slab);
3471 kmem_cache_destroy(ldlm_glimpse_work_kmem);