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/
31 * lustre/ldlm/ldlm_lockd.c
33 * Author: Peter Braam <braam@clusterfs.com>
34 * Author: Phil Schwan <phil@clusterfs.com>
37 #define DEBUG_SUBSYSTEM S_LDLM
39 #include <linux/kthread.h>
40 #include <linux/list.h>
41 #include <libcfs/libcfs.h>
42 #include <libcfs/linux/linux-mem.h>
43 #include <lustre_errno.h>
44 #include <lustre_dlm.h>
45 #include <obd_class.h>
46 #include "ldlm_internal.h"
48 static int ldlm_num_threads;
49 module_param(ldlm_num_threads, int, 0444);
50 MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start");
52 static unsigned int ldlm_cpu_bind = 1;
53 module_param(ldlm_cpu_bind, uint, 0444);
54 MODULE_PARM_DESC(ldlm_cpu_bind,
55 "bind DLM service threads to particular CPU partitions");
57 static char *ldlm_cpts;
58 module_param(ldlm_cpts, charp, 0444);
59 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
61 static DEFINE_MUTEX(ldlm_ref_mutex);
62 static int ldlm_refcount;
64 struct kobject *ldlm_kobj;
65 struct kset *ldlm_ns_kset;
66 struct kset *ldlm_svc_kset;
70 static struct ldlm_state *ldlm_state;
73 * timeout for initial callback (AST) reply (bz10399)
74 * Due to having to send a 32 bit time value over the
75 * wire return it as timeout_t instead of time64_t
77 static inline timeout_t ldlm_get_rq_timeout(void)
80 timeout_t timeout = min(ldlm_timeout, obd_timeout / 3);
82 return timeout < 1 ? 1 : timeout;
89 * blp_prio_list is used for callbacks that should be handled
90 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
93 struct list_head blp_prio_list;
96 * blp_list is used for all other callbacks which are likely
97 * to take longer to process.
99 struct list_head blp_list;
101 wait_queue_head_t blp_waitq;
102 struct completion blp_comp;
103 atomic_t blp_num_threads;
104 atomic_t blp_busy_threads;
111 struct ldlm_bl_work_item {
112 struct list_head blwi_entry;
113 struct ldlm_namespace *blwi_ns;
114 struct ldlm_lock_desc blwi_ld;
115 struct ldlm_lock *blwi_lock;
116 struct list_head blwi_head;
118 struct completion blwi_comp;
119 enum ldlm_cancel_flags blwi_flags;
120 int blwi_mem_pressure;
123 #ifdef HAVE_SERVER_SUPPORT
126 * Protects both waiting_locks_list and expired_lock_thread.
128 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
131 * List for contended locks.
133 * As soon as a lock is contended, it gets placed on this list and
134 * expected time to get a response is filled in the lock. A special
135 * thread walks the list looking for locks that should be released and
136 * schedules client evictions for those that have not been released in
139 * All access to it should be under waiting_locks_spinlock.
141 static LIST_HEAD(waiting_locks_list);
142 static void waiting_locks_callback(TIMER_DATA_TYPE unused);
143 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
151 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
152 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
153 static int expired_lock_dump;
154 static LIST_HEAD(expired_lock_list);
156 static int ldlm_lock_busy(struct ldlm_lock *lock);
157 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout);
158 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout);
160 static inline int have_expired_locks(void)
165 spin_lock_bh(&waiting_locks_spinlock);
166 need_to_run = !list_empty(&expired_lock_list);
167 spin_unlock_bh(&waiting_locks_spinlock);
173 * Check expired lock list for expired locks and time them out.
175 static int expired_lock_main(void *arg)
177 struct list_head *expired = &expired_lock_list;
182 expired_lock_thread_state = ELT_READY;
183 wake_up(&expired_lock_wait_queue);
186 wait_event_idle(expired_lock_wait_queue,
187 have_expired_locks() ||
188 expired_lock_thread_state == ELT_TERMINATE);
190 spin_lock_bh(&waiting_locks_spinlock);
191 if (expired_lock_dump) {
192 spin_unlock_bh(&waiting_locks_spinlock);
194 /* from waiting_locks_callback, but not in timer */
195 libcfs_debug_dumplog();
197 spin_lock_bh(&waiting_locks_spinlock);
198 expired_lock_dump = 0;
203 while (!list_empty(expired)) {
204 struct obd_export *export;
205 struct ldlm_lock *lock;
207 lock = list_entry(expired->next, struct ldlm_lock,
209 if ((void *)lock < LP_POISON + PAGE_SIZE &&
210 (void *)lock >= LP_POISON) {
211 spin_unlock_bh(&waiting_locks_spinlock);
212 CERROR("free lock on elt list %p\n", lock);
215 list_del_init(&lock->l_pending_chain);
216 if ((void *)lock->l_export <
217 LP_POISON + PAGE_SIZE &&
218 (void *)lock->l_export >= LP_POISON) {
219 CERROR("lock with free export on elt list %p\n",
221 lock->l_export = NULL;
222 LDLM_ERROR(lock, "free export");
224 * release extra ref grabbed by
225 * ldlm_add_waiting_lock() or
228 LDLM_LOCK_RELEASE(lock);
232 if (ldlm_is_destroyed(lock)) {
234 * release the lock refcount where
235 * waiting_locks_callback() founds
237 LDLM_LOCK_RELEASE(lock);
240 export = class_export_lock_get(lock->l_export, lock);
241 spin_unlock_bh(&waiting_locks_spinlock);
243 /* Check if we need to prolong timeout */
244 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
245 lock->l_callback_timestamp != 0 && /* not AST error */
246 ldlm_lock_busy(lock)) {
247 LDLM_DEBUG(lock, "prolong the busy lock");
248 lock_res_and_lock(lock);
249 ldlm_add_waiting_lock(lock,
250 ldlm_bl_timeout(lock) >> 1);
251 unlock_res_and_lock(lock);
253 spin_lock_bh(&export->exp_bl_list_lock);
254 list_del_init(&lock->l_exp_list);
255 spin_unlock_bh(&export->exp_bl_list_lock);
258 "lock callback timer expired after %llds: evicting client at %s ",
259 ktime_get_seconds() -
261 obd_export_nid2str(export));
262 ldlm_lock_to_ns(lock)->ns_timeouts++;
263 if (do_dump_on_eviction(export->exp_obd))
265 class_fail_export(export);
267 class_export_lock_put(export, lock);
269 * release extra ref grabbed by ldlm_add_waiting_lock()
270 * or ldlm_failed_ast()
272 LDLM_LOCK_RELEASE(lock);
274 spin_lock_bh(&waiting_locks_spinlock);
276 spin_unlock_bh(&waiting_locks_spinlock);
279 CERROR("dump the log upon eviction\n");
280 libcfs_debug_dumplog();
283 if (expired_lock_thread_state == ELT_TERMINATE)
287 expired_lock_thread_state = ELT_STOPPED;
288 wake_up(&expired_lock_wait_queue);
293 * Check if there is a request in the export request list
294 * which prevents the lock canceling.
296 static int ldlm_lock_busy(struct ldlm_lock *lock)
298 struct ptlrpc_request *req;
303 if (lock->l_export == NULL)
306 spin_lock(&lock->l_export->exp_rpc_lock);
307 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
309 if (req->rq_ops->hpreq_lock_match) {
310 match = req->rq_ops->hpreq_lock_match(req, lock);
315 spin_unlock(&lock->l_export->exp_rpc_lock);
319 /* This is called from within a timer interrupt and cannot schedule */
320 static void waiting_locks_callback(TIMER_DATA_TYPE unused)
322 struct ldlm_lock *lock;
325 spin_lock_bh(&waiting_locks_spinlock);
326 while (!list_empty(&waiting_locks_list)) {
327 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
329 if (lock->l_callback_timestamp > ktime_get_seconds() ||
330 lock->l_req_mode == LCK_GROUP)
334 * no needs to take an extra ref on the lock since it was in
335 * the waiting_locks_list and ldlm_add_waiting_lock()
336 * already grabbed a ref
338 list_move(&lock->l_pending_chain, &expired_lock_list);
342 if (!list_empty(&expired_lock_list)) {
343 if (obd_dump_on_timeout && need_dump)
344 expired_lock_dump = __LINE__;
346 wake_up(&expired_lock_wait_queue);
350 * Make sure the timer will fire again if we have any locks
353 if (!list_empty(&waiting_locks_list)) {
354 time64_t now = ktime_get_seconds();
357 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
359 if (lock->l_callback_timestamp - now > 0)
360 delta = lock->l_callback_timestamp - now;
361 mod_timer(&waiting_locks_timer,
362 jiffies + cfs_time_seconds(delta));
364 spin_unlock_bh(&waiting_locks_spinlock);
368 * Add lock to the list of contended locks.
370 * Indicate that we're waiting for a client to call us back cancelling a given
371 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
372 * timer to fire appropriately. (We round up to the next second, to avoid
373 * floods of timer firings during periods of high lock contention and traffic).
374 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
375 * if it has been added to the waiting list (1 is returned).
377 * Called with the namespace lock held.
379 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t delay)
381 unsigned long timeout_jiffies = jiffies;
385 lock->l_blast_sent = ktime_get_seconds();
386 if (!list_empty(&lock->l_pending_chain))
389 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
390 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
393 deadline = lock->l_blast_sent + delay;
394 if (likely(deadline > lock->l_callback_timestamp))
395 lock->l_callback_timestamp = deadline;
397 timeout = clamp_t(timeout_t,
398 lock->l_callback_timestamp - lock->l_blast_sent,
400 timeout_jiffies += cfs_time_seconds(timeout);
402 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
403 !timer_pending(&waiting_locks_timer))
404 mod_timer(&waiting_locks_timer, timeout_jiffies);
407 * if the new lock has a shorter timeout than something earlier on
408 * the list, we'll wait the longer amount of time; no big deal.
411 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
415 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
417 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
418 if (list_empty(&lock->l_exp_list)) {
419 if (!ldlm_is_granted(lock))
420 list_add_tail(&lock->l_exp_list,
421 &lock->l_export->exp_bl_list);
423 list_add(&lock->l_exp_list,
424 &lock->l_export->exp_bl_list);
426 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
429 * A blocked lock is added. Adjust the position in
430 * the stale list if the export is in the list.
431 * If export is stale and not in the list - it is being
432 * processed and will be placed on the right position
433 * on obd_stale_export_put().
435 if (!list_empty(&lock->l_export->exp_stale_list))
436 obd_stale_export_adjust(lock->l_export);
439 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
443 /* NB: must be called with hold of lock_res_and_lock() */
444 LASSERT(ldlm_is_res_locked(lock));
445 LASSERT(!ldlm_is_cancel_on_block(lock));
448 * Do not put cross-MDT lock in the waiting list, since we
449 * will not evict it due to timeout for now
451 if (lock->l_export != NULL &&
452 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
455 spin_lock_bh(&waiting_locks_spinlock);
456 if (ldlm_is_cancel(lock)) {
457 spin_unlock_bh(&waiting_locks_spinlock);
461 if (ldlm_is_destroyed(lock)) {
462 static time64_t next;
464 spin_unlock_bh(&waiting_locks_spinlock);
465 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
466 if (ktime_get_seconds() > next) {
467 next = ktime_get_seconds() + 14400;
468 libcfs_debug_dumpstack(NULL);
473 ldlm_set_waited(lock);
474 ret = __ldlm_add_waiting_lock(lock, timeout);
477 * grab ref on the lock if it has been added to the
482 spin_unlock_bh(&waiting_locks_spinlock);
485 ldlm_add_blocked_lock(lock);
487 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
488 ret == 0 ? "not re-" : "", timeout,
489 AT_OFF ? "off" : "on");
494 * Remove a lock from the pending list, likely because it had its cancellation
495 * callback arrive without incident. This adjusts the lock-timeout timer if
496 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
497 * As done by ldlm_del_waiting_lock(), the caller must release the lock
498 * reference when the lock is removed from any list (1 is returned).
500 * Called with namespace lock held.
502 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
504 struct list_head *list_next;
506 if (list_empty(&lock->l_pending_chain))
509 list_next = lock->l_pending_chain.next;
510 if (lock->l_pending_chain.prev == &waiting_locks_list) {
511 /* Removing the head of the list, adjust timer. */
512 if (list_next == &waiting_locks_list) {
513 /* No more, just cancel. */
514 timer_delete(&waiting_locks_timer);
516 time64_t now = ktime_get_seconds();
517 struct ldlm_lock *next;
520 next = list_entry(list_next, struct ldlm_lock,
522 if (next->l_callback_timestamp - now > 0)
523 delta = lock->l_callback_timestamp - now;
525 mod_timer(&waiting_locks_timer,
526 jiffies + cfs_time_seconds(delta));
529 list_del_init(&lock->l_pending_chain);
534 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
538 if (lock->l_export == NULL) {
539 /* We don't have a "waiting locks list" on clients. */
540 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
544 spin_lock_bh(&waiting_locks_spinlock);
545 ret = __ldlm_del_waiting_lock(lock);
546 ldlm_clear_waited(lock);
547 spin_unlock_bh(&waiting_locks_spinlock);
549 /* remove the lock out of export blocking list */
550 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
551 list_del_init(&lock->l_exp_list);
552 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
556 * release lock ref if it has indeed been removed
559 LDLM_LOCK_RELEASE(lock);
562 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
567 * Prolong the contended lock waiting time.
569 * Called with namespace lock held.
571 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
573 if (lock->l_export == NULL) {
574 /* We don't have a "waiting locks list" on clients. */
575 LDLM_DEBUG(lock, "client lock: no-op");
579 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
580 /* We don't have a "waiting locks list" on OSP. */
581 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
585 spin_lock_bh(&waiting_locks_spinlock);
587 if (list_empty(&lock->l_pending_chain)) {
588 spin_unlock_bh(&waiting_locks_spinlock);
589 LDLM_DEBUG(lock, "wasn't waiting");
594 * we remove/add the lock to the waiting list, so no needs to
595 * release/take a lock reference
597 __ldlm_del_waiting_lock(lock);
598 __ldlm_add_waiting_lock(lock, timeout);
599 spin_unlock_bh(&waiting_locks_spinlock);
601 LDLM_DEBUG(lock, "refreshed to %ds", timeout);
604 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
606 #else /* HAVE_SERVER_SUPPORT */
608 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
613 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
618 #endif /* !HAVE_SERVER_SUPPORT */
620 #ifdef HAVE_SERVER_SUPPORT
623 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
624 * lock cancel, and their replies). Used for lock callback timeout and AST
627 * \param[in] lock lock which is getting the blocking callback
629 * \retval timeout in seconds to wait for the client reply
631 timeout_t ldlm_bl_timeout(struct ldlm_lock *lock)
636 return obd_timeout / 2;
639 * Since these are non-updating timeouts, we should be conservative.
640 * Take more than usually, 150%
641 * It would be nice to have some kind of "early reply" mechanism for
642 * lock callbacks too...
644 timeout = at_get(&lock->l_export->exp_bl_lock_at);
645 return max_t(timeout_t, timeout + (timeout >> 1),
646 (timeout_t)ldlm_enqueue_min);
648 EXPORT_SYMBOL(ldlm_bl_timeout);
651 * Perform lock cleanup if AST sending failed.
653 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
654 const char *ast_type)
656 LCONSOLE_ERROR_MSG(0x138,
657 "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
658 lock->l_export->exp_obd->obd_name,
659 obd_export_nid2str(lock->l_export), ast_type, rc);
661 if (obd_dump_on_timeout)
662 libcfs_debug_dumplog();
663 spin_lock_bh(&waiting_locks_spinlock);
664 if (__ldlm_del_waiting_lock(lock) == 0)
666 * the lock was not in any list, grab an extra ref before adding
667 * the lock to the expired list
670 /* differentiate it from expired locks */
671 lock->l_callback_timestamp = 0;
672 list_add(&lock->l_pending_chain, &expired_lock_list);
673 wake_up(&expired_lock_wait_queue);
674 spin_unlock_bh(&waiting_locks_spinlock);
678 * Perform lock cleanup if AST reply came with error.
680 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
681 struct ptlrpc_request *req, int rc,
682 const char *ast_type)
684 struct lnet_processid *peer = &req->rq_import->imp_connection->c_peer;
686 if (!req->rq_replied || (rc && rc != -EINVAL)) {
687 if (ldlm_is_cancel(lock)) {
689 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
690 ast_type, req, req->rq_xid,
691 libcfs_nidstr(&peer->nid));
692 ldlm_lock_cancel(lock);
694 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
696 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
698 * Upon umount process the AST fails because cannot be
699 * sent. This shouldn't lead to the client eviction.
700 * -ENODEV error is returned by ptl_send_rpc() for
701 * new request in such import.
702 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
703 * if imp_invalid is set or obd_no_recov.
704 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
705 * in ptlrpc_import_delay_req() as well with -EIO code.
706 * In all such cases errors are ignored.
709 "%s AST can't be sent due to a server %s failure or umount process: rc = %d\n",
711 req->rq_import->imp_obd->obd_name, rc);
714 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
715 libcfs_nidstr(&peer->nid),
716 req->rq_replied ? "returned error from" :
717 "failed to reply to",
718 ast_type, req, req->rq_xid,
719 (req->rq_repmsg != NULL) ?
720 lustre_msg_get_status(req->rq_repmsg) : 0,
722 ldlm_failed_ast(lock, rc, ast_type);
728 struct ldlm_resource *res = lock->l_resource;
731 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
732 libcfs_nidstr(&peer->nid),
734 lustre_msg_get_status(req->rq_repmsg) : -1,
735 ast_type, req, req->rq_xid);
738 * update lvbo to return proper attributes.
741 ldlm_resource_getref(res);
742 ldlm_lvbo_update(res, lock, NULL, 1);
743 ldlm_resource_putref(res);
745 ldlm_lock_cancel(lock);
752 static int ldlm_cb_interpret(const struct lu_env *env,
753 struct ptlrpc_request *req, void *args, int rc)
755 struct ldlm_cb_async_args *ca = args;
756 struct ldlm_lock *lock = ca->ca_lock;
757 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
761 LASSERT(lock != NULL);
764 case LDLM_GL_CALLBACK:
766 * Update the LVB from disk if the AST failed
767 * (this is a legal race)
769 * - Glimpse callback of local lock just returns
770 * -ELDLM_NO_LOCK_DATA.
771 * - Glimpse callback of remote lock might return
772 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
774 if (unlikely(arg->gl_interpret_reply)) {
775 rc = arg->gl_interpret_reply(NULL, req, args, rc);
776 } else if (rc == -ELDLM_NO_LOCK_DATA) {
778 "lost race - client has a lock but no inode");
779 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
780 } else if (rc != 0) {
781 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
783 rc = ldlm_lvbo_update(lock->l_resource,
787 case LDLM_BL_CALLBACK:
789 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
791 case LDLM_CP_CALLBACK:
793 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
796 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
801 /* release extra reference taken in ldlm_ast_fini() */
802 LDLM_LOCK_RELEASE(lock);
805 atomic_inc(&arg->restart);
810 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
812 struct ldlm_cb_async_args *ca = data;
813 struct ldlm_lock *lock = ca->ca_lock;
815 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
818 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
819 struct ldlm_cb_set_arg *arg,
820 struct ldlm_lock *lock,
827 if (unlikely(instant_cancel)) {
828 rc = ptl_send_rpc(req, 1);
829 ptlrpc_req_finished(req);
831 atomic_inc(&arg->restart);
834 ptlrpc_set_add_req(arg->set, req);
841 * Check if there are requests in the export request list which prevent
842 * the lock canceling and make these requests high priority ones.
844 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
846 struct ptlrpc_request *req;
850 if (lock->l_export == NULL) {
851 LDLM_DEBUG(lock, "client lock: no-op");
855 spin_lock(&lock->l_export->exp_rpc_lock);
856 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
859 * Do not process requests that were not yet added to there
860 * incoming queue or were already removed from there for
861 * processing. We evaluate ptlrpc_nrs_req_can_move() without
862 * holding svcpt->scp_req_lock, and then redo the check with
863 * the lock held once we need to obtain a reliable result.
865 if (ptlrpc_nrs_req_can_move(req) &&
866 req->rq_ops->hpreq_lock_match &&
867 req->rq_ops->hpreq_lock_match(req, lock))
868 ptlrpc_nrs_req_hp_move(req);
870 spin_unlock(&lock->l_export->exp_rpc_lock);
875 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
876 * enqueued server lock conflicts with given one.
878 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
879 * to wait for client response.
881 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
882 struct ldlm_lock_desc *desc,
883 void *data, int flag)
885 struct ldlm_cb_async_args *ca;
886 struct ldlm_cb_set_arg *arg = data;
887 struct ldlm_request *body;
888 struct ptlrpc_request *req;
889 int instant_cancel = 0;
894 if (flag == LDLM_CB_CANCELING)
895 /* Don't need to do anything here. */
898 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
899 LDLM_DEBUG(lock, "dropping BL AST");
904 LASSERT(data != NULL);
905 if (lock->l_export->exp_obd->obd_recovering != 0)
906 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
908 ldlm_lock_reorder_req(lock);
910 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
911 &RQF_LDLM_BL_CALLBACK,
912 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
916 ca = ptlrpc_req_async_args(ca, req);
917 ca->ca_set_arg = arg;
920 req->rq_interpret_reply = ldlm_cb_interpret;
922 lock_res_and_lock(lock);
923 if (ldlm_is_destroyed(lock)) {
924 /* What's the point? */
925 unlock_res_and_lock(lock);
926 ptlrpc_req_finished(req);
930 if (!ldlm_is_granted(lock)) {
932 * this blocking AST will be communicated as part of the
933 * completion AST instead
935 ldlm_add_blocked_lock(lock);
936 ldlm_set_waited(lock);
937 unlock_res_and_lock(lock);
939 ptlrpc_req_finished(req);
940 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
944 if (ldlm_is_cancel_on_block(lock))
947 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
948 body->lock_handle[0] = lock->l_remote_handle;
949 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
950 body->lock_desc = *desc;
951 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
953 LDLM_DEBUG(lock, "server preparing blocking AST");
955 ptlrpc_request_set_replen(req);
956 ldlm_set_cbpending(lock);
957 if (instant_cancel) {
958 unlock_res_and_lock(lock);
959 ldlm_lock_cancel(lock);
961 req->rq_no_resend = 1;
963 LASSERT(ldlm_is_granted(lock));
964 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
965 unlock_res_and_lock(lock);
967 /* Do not resend after lock callback timeout */
968 req->rq_delay_limit = ldlm_bl_timeout(lock);
969 req->rq_resend_cb = ldlm_update_resend;
972 req->rq_send_state = LUSTRE_IMP_FULL;
973 /* ptlrpc_request_alloc_pack already set timeout */
975 req->rq_timeout = ldlm_get_rq_timeout();
977 if (lock->l_export && lock->l_export->exp_nid_stats &&
978 lock->l_export->exp_nid_stats->nid_ldlm_stats)
979 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
980 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
982 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
988 * ->l_completion_ast callback for a remote lock in server namespace.
990 * Sends AST to the client notifying it of lock granting. If initial
991 * lock response was not sent yet, instead of sending another RPC, just
992 * mark the lock as granted and client will understand
994 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
996 struct ldlm_cb_set_arg *arg = data;
997 struct ldlm_request *body;
998 struct ptlrpc_request *req;
999 struct ldlm_cb_async_args *ca;
1000 int instant_cancel = 0;
1006 LASSERT(lock != NULL);
1007 LASSERT(data != NULL);
1009 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1010 LDLM_DEBUG(lock, "dropping CP AST");
1014 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1015 &RQF_LDLM_CP_CALLBACK);
1019 /* server namespace, doesn't need lock */
1020 lvb_len = ldlm_lvbo_size(lock);
1022 * LU-3124 & LU-2187: to not return layout in completion AST because
1023 * it may deadlock for LU-2187, or client may not have enough space
1024 * for large layout. The layout will be returned to client with an
1025 * extra RPC to fetch xattr.lov
1027 if (ldlm_has_layout(lock))
1030 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1031 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1033 ptlrpc_request_free(req);
1037 ca = ptlrpc_req_async_args(ca, req);
1038 ca->ca_set_arg = arg;
1041 req->rq_interpret_reply = ldlm_cb_interpret;
1042 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1044 body->lock_handle[0] = lock->l_remote_handle;
1045 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
1046 body->lock_flags = ldlm_flags_to_wire(flags);
1047 ldlm_lock2desc(lock, &body->lock_desc);
1049 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1050 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1053 * We still need to send the RPC to wake up the blocked
1054 * enqueue thread on the client.
1056 * Consider old client, there is no better way to notify
1057 * the failure, just zero-sized the LVB, then the client
1058 * will fail out as "-EPROTO".
1060 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1064 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1069 LDLM_DEBUG(lock, "server preparing completion AST");
1071 ptlrpc_request_set_replen(req);
1073 req->rq_send_state = LUSTRE_IMP_FULL;
1074 /* ptlrpc_request_pack already set timeout */
1076 req->rq_timeout = ldlm_get_rq_timeout();
1078 /* We only send real blocking ASTs after the lock is granted */
1079 lock_res_and_lock(lock);
1080 if (ldlm_is_ast_sent(lock)) {
1081 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1082 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1083 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1087 * We might get here prior to ldlm_handle_enqueue setting
1088 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1089 * into waiting list, but this is safe and similar code in
1090 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1091 * that would not only cancel the lock, but will also remove
1092 * it from waiting list
1094 if (ldlm_is_cancel_on_block(lock)) {
1095 unlock_res_and_lock(lock);
1096 ldlm_lock_cancel(lock);
1099 req->rq_no_resend = 1;
1101 lock_res_and_lock(lock);
1103 /* start the lock-timeout clock */
1104 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1105 /* Do not resend after lock callback timeout */
1106 req->rq_delay_limit = ldlm_bl_timeout(lock);
1107 req->rq_resend_cb = ldlm_update_resend;
1110 unlock_res_and_lock(lock);
1112 if (lock->l_export && lock->l_export->exp_nid_stats &&
1113 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1114 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1115 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1117 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1119 RETURN(lvb_len < 0 ? lvb_len : rc);
1123 * Server side ->l_glimpse_ast handler for client locks.
1125 * Sends glimpse AST to the client and waits for reply. Then updates
1126 * lvbo with the result.
1128 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1130 struct ldlm_cb_set_arg *arg = data;
1131 struct ldlm_request *body;
1132 struct ptlrpc_request *req;
1133 struct ldlm_cb_async_args *ca;
1135 struct req_format *req_fmt;
1139 LASSERT(lock != NULL);
1141 if (arg->gl_desc != NULL)
1142 /* There is a glimpse descriptor to pack */
1143 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1145 req_fmt = &RQF_LDLM_GL_CALLBACK;
1147 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1148 req_fmt, LUSTRE_DLM_VERSION,
1154 if (arg->gl_desc != NULL) {
1155 /* copy the GL descriptor */
1156 union ldlm_gl_desc *desc;
1158 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1159 *desc = *arg->gl_desc;
1162 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1163 body->lock_handle[0] = lock->l_remote_handle;
1164 ldlm_lock2desc(lock, &body->lock_desc);
1166 ca = ptlrpc_req_async_args(ca, req);
1167 ca->ca_set_arg = arg;
1170 /* server namespace, doesn't need lock */
1171 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1172 ldlm_lvbo_size(lock));
1173 ptlrpc_request_set_replen(req);
1175 req->rq_send_state = LUSTRE_IMP_FULL;
1176 /* ptlrpc_request_alloc_pack already set timeout */
1178 req->rq_timeout = ldlm_get_rq_timeout();
1180 req->rq_interpret_reply = ldlm_cb_interpret;
1182 if (lock->l_export && lock->l_export->exp_nid_stats) {
1183 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1185 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1186 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1189 rc = ldlm_ast_fini(req, arg, lock, 0);
1193 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1195 int ldlm_glimpse_locks(struct ldlm_resource *res,
1196 struct list_head *gl_work_list)
1202 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1204 if (rc == -ERESTART)
1205 ldlm_reprocess_all(res, 0);
1209 EXPORT_SYMBOL(ldlm_glimpse_locks);
1211 /* return LDLM lock associated with a lock callback request */
1212 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1214 struct ldlm_cb_async_args *ca;
1215 struct ldlm_lock *lock;
1219 ca = ptlrpc_req_async_args(ca, req);
1222 RETURN(ERR_PTR(-EFAULT));
1226 EXPORT_SYMBOL(ldlm_request_lock);
1229 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1230 * service threads to carry out client lock enqueueing requests.
1232 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1233 struct ptlrpc_request *req,
1234 const struct ldlm_request *dlm_req,
1235 const struct ldlm_callback_suite *cbs)
1237 struct ldlm_reply *dlm_rep;
1239 enum ldlm_error err = ELDLM_OK;
1240 struct ldlm_lock *lock = NULL;
1241 void *cookie = NULL;
1243 struct ldlm_resource *res = NULL;
1244 const struct lu_env *env = req->rq_svc_thread->t_env;
1248 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1250 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1251 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1253 LASSERT(req->rq_export);
1255 /* for intent enqueue the stat will be updated inside intent policy */
1256 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1257 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1258 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1260 if (req->rq_export && req->rq_export->exp_nid_stats &&
1261 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1262 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1263 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1265 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1266 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1267 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1268 dlm_req->lock_desc.l_resource.lr_type);
1269 GOTO(out, rc = -EFAULT);
1272 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1273 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1274 dlm_req->lock_desc.l_req_mode &
1275 (dlm_req->lock_desc.l_req_mode-1))) {
1276 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1277 dlm_req->lock_desc.l_req_mode);
1278 GOTO(out, rc = -EFAULT);
1281 if (unlikely((flags & LDLM_FL_REPLAY) ||
1282 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1283 /* Find an existing lock in the per-export lock hash */
1285 * In the function below, .hs_keycmp resolves to
1286 * ldlm_export_lock_keycmp()
1288 /* coverity[overrun-buffer-val] */
1289 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1290 (void *)&dlm_req->lock_handle[0]);
1292 DEBUG_REQ(D_DLMTRACE, req,
1293 "found existing lock cookie %#llx",
1294 lock->l_handle.h_cookie);
1295 flags |= LDLM_FL_RESENT;
1296 GOTO(existing_lock, rc = 0);
1299 if (ldlm_reclaim_full()) {
1300 DEBUG_REQ(D_DLMTRACE, req,
1301 "Too many granted locks, reject current enqueue request and let the client retry later");
1302 GOTO(out, rc = -EINPROGRESS);
1306 /* The lock's callback data might be set in the policy function */
1307 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1308 dlm_req->lock_desc.l_resource.lr_type,
1309 dlm_req->lock_desc.l_req_mode,
1310 cbs, NULL, 0, LVB_T_NONE);
1317 lock->l_remote_handle = dlm_req->lock_handle[0];
1318 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1321 * Initialize resource lvb but not for a lock being replayed since
1322 * Client already got lvb sent in this case.
1323 * This must occur early since some policy methods assume resource
1324 * lvb is available (lr_lvb_data != NULL).
1326 res = lock->l_resource;
1327 if (!(flags & LDLM_FL_REPLAY)) {
1328 /* non-replayed lock, delayed lvb init may need to be done */
1329 rc = ldlm_lvbo_init(res);
1331 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1336 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1338 * Don't enqueue a lock onto the export if it is been disonnected
1339 * due to eviction (b=3822) or server umount (b=24324).
1340 * Cancel it now instead.
1342 if (req->rq_export->exp_disconnected) {
1343 LDLM_ERROR(lock, "lock on disconnected export %p",
1345 GOTO(out, rc = -ENOTCONN);
1348 lock->l_export = class_export_lock_get(req->rq_export, lock);
1349 if (lock->l_export->exp_lock_hash)
1350 cfs_hash_add(lock->l_export->exp_lock_hash,
1351 &lock->l_remote_handle,
1355 * Inherit the enqueue flags before the operation, because we do not
1356 * keep the res lock on return and next operations (BL AST) may proceed
1359 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1360 LDLM_FL_INHERIT_MASK);
1362 ldlm_convert_policy_to_local(req->rq_export,
1363 dlm_req->lock_desc.l_resource.lr_type,
1364 &dlm_req->lock_desc.l_policy_data,
1365 &lock->l_policy_data);
1366 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
1367 lock->l_req_extent = lock->l_policy_data.l_extent;
1368 } else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) {
1369 lock->l_policy_data.l_inodebits.try_bits =
1370 dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1371 lock->l_policy_data.l_inodebits.li_gid =
1372 dlm_req->lock_desc.l_policy_data.l_inodebits.li_gid;
1377 if (!(flags & LDLM_FL_HAS_INTENT)) {
1378 /* based on the assumption that lvb size never changes during
1379 * resource life time otherwise it need resource->lr_lock's
1381 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1382 RCL_SERVER, ldlm_lvbo_size(lock));
1384 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1385 GOTO(out, rc = -ENOMEM);
1387 rc = req_capsule_server_pack(&req->rq_pill);
1392 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1399 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1401 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1402 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1404 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1405 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1408 * We never send a blocking AST until the lock is granted, but
1409 * we can tell it right now
1411 lock_res_and_lock(lock);
1414 * Now take into account flags to be inherited from original lock
1415 * request both in reply to client and in our own lock flags.
1417 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1418 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1421 * Don't move a pending lock onto the export if it has already been
1422 * disconnected due to eviction (b=5683) or server umount (b=24324).
1423 * Cancel it now instead.
1425 if (unlikely(req->rq_export->exp_disconnected ||
1426 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1427 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1429 } else if (ldlm_is_ast_sent(lock)) {
1430 /* fill lock desc for possible lock convert */
1431 if (lock->l_blocking_lock &&
1432 lock->l_resource->lr_type == LDLM_IBITS) {
1433 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1434 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1437 "save blocking bits %llx in granted lock",
1438 bl_lock->l_policy_data.l_inodebits.bits);
1440 * If lock is blocked then save blocking ibits
1441 * in returned lock policy for the possible lock
1442 * convert on a client.
1444 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1445 bl_lock->l_policy_data.l_inodebits.bits;
1447 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1448 if (ldlm_is_granted(lock)) {
1450 * Only cancel lock if it was granted, because it would
1451 * be destroyed immediately and would never be granted
1452 * in the future, causing timeouts on client. Not
1453 * granted lock will be cancelled immediately after
1454 * sending completion AST.
1456 if (ldlm_is_cancel_on_block(lock)) {
1457 unlock_res_and_lock(lock);
1458 ldlm_lock_cancel(lock);
1459 lock_res_and_lock(lock);
1461 ldlm_add_waiting_lock(lock,
1462 ldlm_bl_timeout(lock));
1466 unlock_res_and_lock(lock);
1470 req->rq_status = rc ?: err; /* return either error - b=11190 */
1471 if (!req->rq_packed_final) {
1472 int rc1 = lustre_pack_reply(req, 1, NULL, NULL);
1478 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1479 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1483 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1487 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1489 ldlm_lvbo_size(lock) > 0) {
1494 buf = req_capsule_server_get(&req->rq_pill,
1496 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1497 buflen = req_capsule_get_size(&req->rq_pill,
1498 &RMF_DLM_LVB, RCL_SERVER);
1500 * non-replayed lock, delayed lvb init may
1501 * need to be occur now
1503 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1506 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1508 req_capsule_shrink(&req->rq_pill,
1511 } else if (rc2 == -ERANGE) {
1512 rc2 = req_capsule_server_grow(
1514 &RMF_DLM_LVB, buflen);
1519 * if we can't grow the buffer,
1520 * it's ok to return empty lvb
1531 } else if (flags & LDLM_FL_REPLAY) {
1532 /* no LVB resend upon replay */
1534 req_capsule_shrink(&req->rq_pill,
1544 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1545 if (lock->l_export) {
1546 ldlm_lock_cancel(lock);
1548 lock_res_and_lock(lock);
1549 ldlm_resource_unlink_lock(lock);
1550 ldlm_lock_destroy_nolock(lock);
1551 unlock_res_and_lock(lock);
1553 ldlm_reprocess_all(lock->l_resource, lock->l_policy_data.l_inodebits.bits);
1556 if (!err && !ldlm_is_cbpending(lock) &&
1557 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1558 ldlm_reprocess_all(lock->l_resource,
1559 lock->l_policy_data.l_inodebits.bits);
1561 LDLM_LOCK_RELEASE(lock);
1564 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1571 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1572 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1574 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1576 if (lock->l_blocking_lock) {
1577 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1578 lock->l_blocking_lock = NULL;
1582 /* A lock can be converted to new ibits or mode and should be considered
1583 * as new lock. Clear all states related to a previous blocking AST
1584 * processing so new conflicts will cause new blocking ASTs.
1586 * This is used during lock convert below and lock downgrade to COS mode in
1587 * ldlm_lock_mode_downgrade().
1589 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1591 ldlm_clear_ast_sent(lock);
1592 lock->l_bl_ast_run = 0;
1593 ldlm_clear_blocking_lock(lock);
1597 * Main LDLM entry point for server code to process lock conversion requests.
1599 int ldlm_handle_convert0(struct ptlrpc_request *req,
1600 const struct ldlm_request *dlm_req)
1602 struct obd_export *exp = req->rq_export;
1603 struct ldlm_reply *dlm_rep;
1604 struct ldlm_lock *lock;
1611 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1612 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1613 LDLM_CONVERT - LDLM_FIRST_OPC);
1615 rc = req_capsule_server_pack(&req->rq_pill);
1619 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1620 dlm_rep->lock_flags = dlm_req->lock_flags;
1622 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1624 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1625 req->rq_status = ELDLM_NO_LOCK_DATA;
1629 LDLM_DEBUG(lock, "server-side convert handler START");
1631 lock_res_and_lock(lock);
1632 bits = lock->l_policy_data.l_inodebits.bits;
1633 new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1635 if (ldlm_is_cancel(lock)) {
1636 LDLM_DEBUG(lock, "convert on canceled lock!");
1637 unlock_res_and_lock(lock);
1638 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1641 if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1642 LDLM_ERROR(lock, "lock mode differs!");
1643 unlock_res_and_lock(lock);
1644 GOTO(out_put, rc = -EPROTO);
1647 if (bits == new_bits) {
1649 * This can be valid situation if CONVERT RPCs are
1650 * re-ordered. Just finish silently
1652 LDLM_DEBUG(lock, "lock is converted already!");
1653 unlock_res_and_lock(lock);
1655 if (ldlm_is_waited(lock))
1656 ldlm_del_waiting_lock(lock);
1658 ldlm_clear_cbpending(lock);
1659 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1660 ldlm_inodebits_drop(lock, bits & ~new_bits);
1662 ldlm_clear_blocking_data(lock);
1663 unlock_res_and_lock(lock);
1665 /* All old bits should be reprocessed to send new BL AST if
1666 * it wasn't sent earlier due to LDLM_FL_AST_SENT bit set.
1668 ldlm_reprocess_all(lock->l_resource, bits);
1671 dlm_rep->lock_handle = lock->l_remote_handle;
1672 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1673 &dlm_rep->lock_desc.l_policy_data);
1677 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1678 LDLM_LOCK_PUT(lock);
1679 req->rq_status = rc;
1684 * Cancel all the locks whose handles are packed into ldlm_request
1686 * Called by server code expecting such combined cancel activity
1689 int ldlm_request_cancel(struct ptlrpc_request *req,
1690 const struct ldlm_request *dlm_req,
1691 int first, enum lustre_at_flags flags)
1693 struct ldlm_resource *res, *pres = NULL;
1694 struct ldlm_lock *lock;
1695 int i, count, done = 0;
1700 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1701 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1702 (size - offsetof(struct ldlm_request, lock_handle)) /
1703 sizeof(struct lustre_handle) < dlm_req->lock_count)
1706 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1710 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1714 * There is no lock on the server at the replay time,
1715 * skip lock cancelling to make replay tests to pass.
1717 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1720 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1723 for (i = first; i < count; i++) {
1724 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1726 /* below message checked in replay-single.sh test_36 */
1727 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1728 dlm_req->lock_handle[i].cookie);
1732 res = lock->l_resource;
1736 * This code is an optimization to only attempt lock
1737 * granting on the resource (that could be CPU-expensive)
1738 * after we are done cancelling lock in that resource.
1742 ldlm_reprocess_all(pres, 0);
1743 LDLM_RESOURCE_DELREF(pres);
1744 ldlm_resource_putref(pres);
1747 ldlm_resource_getref(res);
1748 LDLM_RESOURCE_ADDREF(res);
1750 if (!ldlm_is_discard_data(lock))
1751 ldlm_lvbo_update(res, lock,
1757 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1758 lock->l_blast_sent != 0) {
1759 timeout_t delay = 0;
1761 if (ktime_get_seconds() > lock->l_blast_sent)
1762 delay = ktime_get_seconds() -
1765 "server cancels blocked lock after %ds",
1767 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1769 ldlm_lock_cancel(lock);
1770 LDLM_LOCK_PUT(lock);
1773 ldlm_reprocess_all(pres, 0);
1774 LDLM_RESOURCE_DELREF(pres);
1775 ldlm_resource_putref(pres);
1777 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1780 EXPORT_SYMBOL(ldlm_request_cancel);
1783 * Main LDLM entry point for server code to cancel locks.
1785 * Typically gets called from service handler on LDLM_CANCEL opc.
1787 int ldlm_handle_cancel(struct ptlrpc_request *req)
1789 struct ldlm_request *dlm_req;
1794 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1795 if (dlm_req == NULL) {
1796 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1800 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1801 offsetof(struct ldlm_request, lock_handle[1]))
1804 if (req->rq_export && req->rq_export->exp_nid_stats &&
1805 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1806 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1807 LDLM_CANCEL - LDLM_FIRST_OPC);
1809 rc = req_capsule_server_pack(&req->rq_pill);
1813 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1814 req->rq_status = LUSTRE_ESTALE;
1816 RETURN(ptlrpc_reply(req));
1818 #endif /* HAVE_SERVER_SUPPORT */
1821 * Server may pass additional information about blocking lock.
1822 * For IBITS locks it is conflicting bits which can be used for
1823 * lock convert instead of cancel.
1825 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1827 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1829 check_res_locked(lock->l_resource);
1830 if (ns_is_client(ns) && ld &&
1831 (lock->l_resource->lr_type == LDLM_IBITS)) {
1833 * Lock description contains policy of blocking lock,
1834 * and its cancel_bits is used to pass conflicting bits.
1835 * NOTE: ld can be NULL or can be not NULL but zeroed if
1836 * passed from ldlm_bl_thread_blwi(), check below used bits
1837 * in ld to make sure it is valid description.
1839 * If server may replace lock resource keeping the same cookie,
1840 * never use cancel bits from different resource, full cancel
1843 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1844 ldlm_res_eq(&ld->l_resource.lr_name,
1845 &lock->l_resource->lr_name) &&
1846 !(ldlm_is_cbpending(lock) &&
1847 lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1848 /* always combine conflicting ibits */
1849 lock->l_policy_data.l_inodebits.cancel_bits |=
1850 ld->l_policy_data.l_inodebits.cancel_bits;
1852 /* If cancel_bits are not obtained or
1853 * if the lock is already CBPENDING and
1854 * has no cancel_bits set
1855 * - the full lock is to be cancelled
1857 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1863 * Callback handler for receiving incoming blocking ASTs.
1865 * This can only happen on client side.
1867 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1868 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1874 LDLM_DEBUG(lock, "client blocking AST callback handler");
1876 lock_res_and_lock(lock);
1878 /* get extra information from desc if any */
1879 ldlm_bl_desc2lock(ld, lock);
1880 ldlm_set_cbpending(lock);
1882 do_ast = (!lock->l_readers && !lock->l_writers);
1883 unlock_res_and_lock(lock);
1887 "Lock %p already unused, calling callback (%p)\n",
1888 lock, lock->l_blocking_ast);
1889 if (lock->l_blocking_ast != NULL)
1890 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1894 "Lock %p is referenced, will be cancelled later\n",
1898 LDLM_DEBUG(lock, "client blocking callback handler END");
1899 LDLM_LOCK_RELEASE(lock);
1903 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1905 if (req->rq_no_reply)
1908 req->rq_status = rc;
1909 if (!req->rq_packed_final) {
1910 rc = lustre_pack_reply(req, 1, NULL, NULL);
1914 return ptlrpc_reply(req);
1918 * Callback handler for receiving incoming completion ASTs.
1920 * This only can happen on client side.
1922 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1923 struct ldlm_namespace *ns,
1924 struct ldlm_request *dlm_req,
1925 struct ldlm_lock *lock)
1927 LIST_HEAD(ast_list);
1933 LDLM_DEBUG(lock, "client completion callback handler START");
1935 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1936 long to = cfs_time_seconds(1);
1938 ldlm_callback_reply(req, 0);
1941 to = schedule_timeout_interruptible(to);
1942 if (ldlm_is_granted(lock) ||
1943 ldlm_is_destroyed(lock))
1948 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1950 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1951 GOTO(out, rc = lvb_len);
1952 } else if (lvb_len > 0) {
1953 if (lock->l_lvb_len > 0) {
1954 /* for extent lock, lvb contains ost_lvb{}. */
1955 LASSERT(lock->l_lvb_data != NULL);
1957 if (unlikely(lock->l_lvb_len < lvb_len)) {
1959 "Replied LVB is larger than expectation, expected = %d, replied = %d",
1960 lock->l_lvb_len, lvb_len);
1961 GOTO(out, rc = -EINVAL);
1966 lock_res_and_lock(lock);
1968 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
1969 &lock->l_resource->lr_name)) {
1970 ldlm_resource_unlink_lock(lock);
1971 unlock_res_and_lock(lock);
1972 rc = ldlm_lock_change_resource(ns, lock,
1973 &dlm_req->lock_desc.l_resource.lr_name);
1975 LDLM_ERROR(lock, "Failed to allocate resource");
1978 LDLM_DEBUG(lock, "completion AST, new resource");
1979 lock_res_and_lock(lock);
1982 if (ldlm_is_failed(lock)) {
1983 unlock_res_and_lock(lock);
1984 LDLM_LOCK_RELEASE(lock);
1988 if (ldlm_is_destroyed(lock) ||
1989 ldlm_is_granted(lock)) {
1990 /* b=11300: the lock has already been granted */
1991 unlock_res_and_lock(lock);
1992 LDLM_DEBUG(lock, "Double grant race happened");
1997 * If we receive the completion AST before the actual enqueue returned,
1998 * then we might need to switch lock modes, resources, or extents.
2000 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
2001 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
2002 LDLM_DEBUG(lock, "completion AST, new lock mode");
2005 if (lock->l_resource->lr_type != LDLM_PLAIN) {
2006 ldlm_convert_policy_to_local(req->rq_export,
2007 dlm_req->lock_desc.l_resource.lr_type,
2008 &dlm_req->lock_desc.l_policy_data,
2009 &lock->l_policy_data);
2010 LDLM_DEBUG(lock, "completion AST, new policy data");
2013 ldlm_resource_unlink_lock(lock);
2015 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2017 * BL_AST locks are not needed in LRU.
2018 * Let ldlm_cancel_lru() be fast.
2020 ldlm_lock_remove_from_lru(lock);
2021 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2022 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2023 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2026 if (lock->l_lvb_len > 0) {
2027 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2028 lock->l_lvb_data, lvb_len);
2030 unlock_res_and_lock(lock);
2035 ldlm_grant_lock(lock, &ast_list);
2036 unlock_res_and_lock(lock);
2038 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2041 * Let Enqueue to call osc_lock_upcall() and initialize
2044 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2046 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2048 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2054 lock_res_and_lock(lock);
2055 ldlm_set_failed(lock);
2056 unlock_res_and_lock(lock);
2057 wake_up(&lock->l_waitq);
2059 LDLM_LOCK_RELEASE(lock);
2065 * Callback handler for receiving incoming glimpse ASTs.
2067 * This only can happen on client side. After handling the glimpse AST
2068 * we also consider dropping the lock here if it is unused locally for a
2071 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2072 struct ldlm_namespace *ns,
2073 struct ldlm_request *dlm_req,
2074 struct ldlm_lock *lock)
2076 struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2081 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2083 if (lock->l_glimpse_ast != NULL)
2084 rc = lock->l_glimpse_ast(lock, req);
2086 if (req->rq_repmsg != NULL) {
2089 req->rq_status = rc;
2093 lock_res_and_lock(lock);
2094 if (lock->l_granted_mode == LCK_PW &&
2095 !lock->l_readers && !lock->l_writers &&
2096 ktime_after(ktime_get(),
2097 ktime_add(lock->l_last_used, ns->ns_dirty_age_limit))) {
2098 unlock_res_and_lock(lock);
2100 /* For MDS glimpse it is always DOM lock, set corresponding
2101 * cancel_bits to perform lock convert if needed
2103 if (lock->l_resource->lr_type == LDLM_IBITS)
2104 ld->l_policy_data.l_inodebits.cancel_bits =
2106 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2107 ldlm_handle_bl_callback(ns, ld, lock);
2112 unlock_res_and_lock(lock);
2113 LDLM_LOCK_RELEASE(lock);
2117 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2118 enum ldlm_cancel_flags cancel_flags)
2120 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2121 char *prio = "regular";
2126 spin_lock(&blp->blp_lock);
2127 /* cannot access blwi after added to list and lock is dropped */
2128 count = blwi->blwi_lock ? 1 : blwi->blwi_count;
2130 /* if the server is waiting on a lock to be cancelled (bl_ast), this is
2131 * an urgent request and should go in the priority queue so it doesn't
2132 * get stuck behind non-priority work (eg, lru size management)
2134 * We also prioritize discard_data, which is for eviction handling
2136 if (blwi->blwi_lock &&
2137 (ldlm_is_discard_data(blwi->blwi_lock) ||
2138 ldlm_is_bl_ast(blwi->blwi_lock))) {
2139 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2142 /* other blocking callbacks are added to the regular list */
2143 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2145 blp->blp_total_locks += count;
2146 blp->blp_total_blwis++;
2147 spin_unlock(&blp->blp_lock);
2149 wake_up(&blp->blp_waitq);
2151 /* unlocked read of blp values is intentional - OK for debug */
2153 "added %d/%d locks to %s blp list, %d blwis in pool\n",
2154 count, blp->blp_total_locks, prio, blp->blp_total_blwis);
2157 * can not check blwi->blwi_flags as blwi could be already freed in
2160 if (!(cancel_flags & LCF_ASYNC))
2161 wait_for_completion(&blwi->blwi_comp);
2166 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2167 struct ldlm_namespace *ns,
2168 struct ldlm_lock_desc *ld,
2169 struct list_head *cancels, int count,
2170 struct ldlm_lock *lock,
2171 enum ldlm_cancel_flags cancel_flags)
2173 init_completion(&blwi->blwi_comp);
2174 INIT_LIST_HEAD(&blwi->blwi_head);
2176 if (current->flags & PF_MEMALLOC)
2177 blwi->blwi_mem_pressure = 1;
2180 blwi->blwi_flags = cancel_flags;
2182 blwi->blwi_ld = *ld;
2184 list_splice_init(cancels, &blwi->blwi_head);
2185 blwi->blwi_count = count;
2187 blwi->blwi_lock = lock;
2192 * Queues a list of locks \a cancels containing \a count locks
2193 * for later processing by a blocking thread. If \a count is zero,
2194 * then the lock referenced as \a lock is queued instead.
2196 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2197 * If list addition fails an error is returned and caller is supposed to
2198 * call ->l_blocking_ast itself.
2200 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2201 struct ldlm_lock_desc *ld,
2202 struct ldlm_lock *lock,
2203 struct list_head *cancels, int count,
2204 enum ldlm_cancel_flags cancel_flags)
2208 if (cancels && count == 0)
2211 if (cancel_flags & LCF_ASYNC) {
2212 struct ldlm_bl_work_item *blwi;
2214 OBD_ALLOC(blwi, sizeof(*blwi));
2217 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2219 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2222 * if it is synchronous call do minimum mem alloc, as it could
2223 * be triggered from kernel shrinker
2225 struct ldlm_bl_work_item blwi;
2227 memset(&blwi, 0, sizeof(blwi));
2228 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2229 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2234 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2235 struct ldlm_lock *lock)
2237 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2240 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2241 struct list_head *cancels, int count,
2242 enum ldlm_cancel_flags cancel_flags)
2244 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2247 int ldlm_bl_to_thread_ns(struct ldlm_namespace *ns)
2249 return ldlm_bl_to_thread(ns, NULL, NULL, NULL, 0, LCF_ASYNC);
2252 int ldlm_bl_thread_wakeup(void)
2254 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2258 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2259 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2261 struct obd_device *obd = req->rq_export->exp_obd;
2269 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2271 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2273 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2275 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2278 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2280 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2282 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2285 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2288 /* We are responsible for swabbing contents of val */
2290 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2291 /* Pass it on to mdc (the "export" in this case) */
2292 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2294 sizeof(KEY_HSM_COPYTOOL_SEND),
2295 KEY_HSM_COPYTOOL_SEND,
2298 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2303 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2304 const char *msg, int rc,
2305 const struct lustre_handle *handle)
2307 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2308 "%s, NID=%s lock=%#llx: rc = %d",
2309 msg, libcfs_id2str(req->rq_peer),
2310 handle ? handle->cookie : 0, rc);
2311 if (req->rq_no_reply)
2312 CWARN("No reply was sent, maybe cause b=21636.\n");
2314 CWARN("Send reply failed, maybe cause b=21636.\n");
2317 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2318 static int ldlm_callback_handler(struct ptlrpc_request *req)
2320 struct ldlm_namespace *ns;
2321 struct ldlm_request *dlm_req;
2322 struct ldlm_lock *lock;
2328 * Requests arrive in sender's byte order. The ptlrpc service
2329 * handler has already checked and, if necessary, byte-swapped the
2330 * incoming request message body, but I am responsible for the
2334 /* do nothing for sec context finalize */
2335 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2338 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2340 if (req->rq_export == NULL) {
2341 rc = ldlm_callback_reply(req, -ENOTCONN);
2342 ldlm_callback_errmsg(req, "Operate on unconnected server",
2347 LASSERT(req->rq_export != NULL);
2348 LASSERT(req->rq_export->exp_obd != NULL);
2350 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2351 case LDLM_BL_CALLBACK:
2352 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2354 ldlm_callback_reply(req, -(int)cfs_fail_err);
2358 case LDLM_CP_CALLBACK:
2359 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2362 case LDLM_GL_CALLBACK:
2363 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2367 rc = ldlm_handle_setinfo(req);
2368 ldlm_callback_reply(req, rc);
2371 CERROR("unknown opcode %u\n",
2372 lustre_msg_get_opc(req->rq_reqmsg));
2373 ldlm_callback_reply(req, -EPROTO);
2377 ns = req->rq_export->exp_obd->obd_namespace;
2378 LASSERT(ns != NULL);
2380 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2382 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2383 if (dlm_req == NULL) {
2384 rc = ldlm_callback_reply(req, -EPROTO);
2385 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2391 * Force a known safe race, send a cancel to the server for a lock
2392 * which the server has already started a blocking callback on.
2394 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2395 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2396 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2398 CERROR("ldlm_cli_cancel: %d\n", rc);
2401 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2404 "callback on lock %#llx - lock disappeared\n",
2405 dlm_req->lock_handle[0].cookie);
2406 rc = ldlm_callback_reply(req, -EINVAL);
2407 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2408 &dlm_req->lock_handle[0]);
2412 if (ldlm_is_fail_loc(lock) &&
2413 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2414 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2416 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2417 lock_res_and_lock(lock);
2418 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2420 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2422 * If somebody cancels lock and cache is already dropped,
2423 * or lock is failed before cp_ast received on client,
2424 * we can tell the server we have no lock. Otherwise, we
2425 * should send cancel after dropping the cache.
2427 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2428 ldlm_is_failed(lock)) {
2430 "callback on lock %llx - lock disappeared",
2431 dlm_req->lock_handle[0].cookie);
2432 unlock_res_and_lock(lock);
2433 LDLM_LOCK_RELEASE(lock);
2434 rc = ldlm_callback_reply(req, -EINVAL);
2435 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2436 &dlm_req->lock_handle[0]);
2440 * BL_AST locks are not needed in LRU.
2441 * Let ldlm_cancel_lru() be fast.
2443 ldlm_lock_remove_from_lru(lock);
2444 ldlm_set_bl_ast(lock);
2446 if (lock->l_remote_handle.cookie == 0)
2447 lock->l_remote_handle = dlm_req->lock_handle[1];
2448 unlock_res_and_lock(lock);
2451 * We want the ost thread to get this reply so that it can respond
2452 * to ost requests (write cache writeback) that might be triggered
2455 * But we'd also like to be able to indicate in the reply that we're
2456 * cancelling right now, because it's unused, or have an intent result
2457 * in the reply, so we might have to push the responsibility for sending
2458 * the reply down into the AST handlers, alas.
2461 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2462 case LDLM_BL_CALLBACK:
2463 CDEBUG(D_INODE, "blocking ast\n");
2464 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2465 if (!ldlm_is_cancel_on_block(lock)) {
2466 rc = ldlm_callback_reply(req, 0);
2467 if (req->rq_no_reply || rc)
2468 ldlm_callback_errmsg(req, "Normal process", rc,
2469 &dlm_req->lock_handle[0]);
2471 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2472 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2474 case LDLM_CP_CALLBACK:
2475 CDEBUG(D_INODE, "completion ast\n");
2476 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2477 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2478 if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2479 ldlm_callback_reply(req, rc);
2481 case LDLM_GL_CALLBACK:
2482 CDEBUG(D_INODE, "glimpse ast\n");
2483 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2484 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2487 LBUG(); /* checked above */
2493 #ifdef HAVE_SERVER_SUPPORT
2495 * Main handler for canceld thread.
2497 * Separated into its own thread to avoid deadlocks.
2499 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2506 * Requests arrive in sender's byte order. The ptlrpc service
2507 * handler has already checked and, if necessary, byte-swapped the
2508 * incoming request message body, but I am responsible for the
2512 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2514 if (req->rq_export == NULL) {
2515 struct ldlm_request *dlm_req;
2517 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2518 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2519 libcfs_nid2str(req->rq_peer.nid),
2520 (unsigned long long)req->rq_arrival_time.tv_sec,
2521 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2523 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2524 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2525 dlm_req = req_capsule_client_get(&req->rq_pill,
2527 if (dlm_req != NULL)
2528 ldlm_lock_dump_handle(D_ERROR,
2529 &dlm_req->lock_handle[0]);
2531 ldlm_callback_reply(req, -ENOTCONN);
2535 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2536 /* XXX FIXME move this back to mds/handler.c, b=249 */
2538 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2539 CDEBUG(D_INODE, "cancel\n");
2540 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2541 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2542 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2544 rc = ldlm_handle_cancel(req);
2548 struct ldlm_request *dlm_req;
2550 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2551 CDEBUG(D_INODE, "convert\n");
2553 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2554 if (dlm_req == NULL) {
2555 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2556 rc = ldlm_callback_reply(req, -EPROTO);
2558 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2559 rc = ptlrpc_reply(req);
2564 CERROR("invalid opcode %d\n",
2565 lustre_msg_get_opc(req->rq_reqmsg));
2566 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2567 rc = ldlm_callback_reply(req, -EINVAL);
2573 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2574 struct ldlm_lock *lock)
2576 struct ldlm_request *dlm_req;
2577 struct lustre_handle lockh;
2583 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2584 if (dlm_req == NULL)
2587 ldlm_lock2handle(lock, &lockh);
2588 for (i = 0; i < dlm_req->lock_count; i++) {
2589 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2591 DEBUG_REQ(D_RPCTRACE, req,
2592 "Prio raised by lock %#llx", lockh.cookie);
2601 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2603 struct ldlm_request *dlm_req;
2610 /* no prolong in recovery */
2611 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2614 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2615 if (dlm_req == NULL)
2618 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2619 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2620 (size - offsetof(struct ldlm_request, lock_handle)) /
2621 sizeof(struct lustre_handle) < dlm_req->lock_count)
2624 for (i = 0; i < dlm_req->lock_count; i++) {
2625 struct ldlm_lock *lock;
2627 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2631 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2633 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2634 LDLM_LOCK_PUT(lock);
2643 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2644 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2645 .hpreq_check = ldlm_cancel_hpreq_check,
2649 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2653 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2655 if (req->rq_export == NULL)
2658 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2659 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2660 req->rq_ops = &ldlm_cancel_hpreq_ops;
2661 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2662 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2663 req->rq_ops = &ldlm_cancel_hpreq_ops;
2668 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2669 struct hlist_node *hnode, void *data)
2672 struct list_head *rpc_list = data;
2673 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2675 lock_res_and_lock(lock);
2677 if (!ldlm_is_granted(lock)) {
2678 unlock_res_and_lock(lock);
2682 LASSERT(lock->l_resource);
2683 if (lock->l_resource->lr_type != LDLM_IBITS &&
2684 lock->l_resource->lr_type != LDLM_PLAIN) {
2685 unlock_res_and_lock(lock);
2689 if (ldlm_is_ast_sent(lock)) {
2690 unlock_res_and_lock(lock);
2694 LASSERT(lock->l_blocking_ast);
2695 LASSERT(!lock->l_blocking_lock);
2697 ldlm_set_ast_sent(lock);
2698 if (lock->l_export && lock->l_export->exp_lock_hash) {
2700 * NB: it's safe to call cfs_hash_del() even lock isn't
2704 * In the function below, .hs_keycmp resolves to
2705 * ldlm_export_lock_keycmp()
2707 /* coverity[overrun-buffer-val] */
2708 cfs_hash_del(lock->l_export->exp_lock_hash,
2709 &lock->l_remote_handle, &lock->l_exp_hash);
2712 list_add_tail(&lock->l_rk_ast, rpc_list);
2713 LDLM_LOCK_GET(lock);
2715 unlock_res_and_lock(lock);
2719 void ldlm_revoke_export_locks(struct obd_export *exp)
2722 LIST_HEAD(rpc_list);
2725 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2726 ldlm_revoke_lock_cb, &rpc_list, 0);
2727 rc = ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2728 LDLM_WORK_REVOKE_AST);
2730 if (rc == -ERESTART)
2731 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2735 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2736 #endif /* HAVE_SERVER_SUPPORT */
2738 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2739 struct ldlm_bl_work_item **p_blwi,
2740 struct obd_export **p_exp)
2742 struct ldlm_bl_work_item *blwi = NULL;
2743 static unsigned int num_bl;
2744 static unsigned int num_stale;
2745 int num_th = atomic_read(&blp->blp_num_threads);
2747 *p_exp = obd_stale_export_get();
2749 spin_lock(&blp->blp_lock);
2750 if (*p_exp != NULL) {
2751 if (num_th == 1 || ++num_stale < num_th) {
2752 spin_unlock(&blp->blp_lock);
2758 /* process a request from the blp_list at least every blp_num_threads */
2759 if (!list_empty(&blp->blp_list) &&
2760 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2761 blwi = list_entry(blp->blp_list.next,
2762 struct ldlm_bl_work_item, blwi_entry);
2764 if (!list_empty(&blp->blp_prio_list))
2765 blwi = list_entry(blp->blp_prio_list.next,
2766 struct ldlm_bl_work_item,
2770 if (++num_bl >= num_th)
2772 list_del(&blwi->blwi_entry);
2773 blp->blp_total_locks -= blwi->blwi_lock ? 1 : blwi->blwi_count;
2774 blp->blp_total_blwis--;
2776 spin_unlock(&blp->blp_lock);
2779 /* intentional unlocked read of blp values - OK for debug */
2782 "Got %d locks of %d total in blp. (%d blwis in pool)\n",
2783 blwi->blwi_lock ? 1 : blwi->blwi_count,
2784 blp->blp_total_locks, blp->blp_total_blwis);
2787 "No blwi found in queue (no bl locks in queue)\n");
2790 if (*p_exp != NULL && *p_blwi != NULL) {
2791 obd_stale_export_put(*p_exp);
2795 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2798 /* This only contains temporary data until the thread starts */
2799 struct ldlm_bl_thread_data {
2800 struct ldlm_bl_pool *bltd_blp;
2801 struct completion bltd_comp;
2805 static int ldlm_bl_thread_main(void *arg);
2807 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2809 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2810 struct task_struct *task;
2812 init_completion(&bltd.bltd_comp);
2814 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2815 if (bltd.bltd_num >= blp->blp_max_threads) {
2816 atomic_dec(&blp->blp_num_threads);
2820 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2822 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2823 atomic_dec(&blp->blp_num_threads);
2827 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2830 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2831 bltd.bltd_num, PTR_ERR(task));
2832 atomic_dec(&blp->blp_num_threads);
2833 return PTR_ERR(task);
2835 wait_for_completion(&bltd.bltd_comp);
2840 /* Not fatal if racy and have a few too many threads */
2841 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2842 struct ldlm_bl_work_item *blwi)
2844 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2847 if (atomic_read(&blp->blp_busy_threads) <
2848 atomic_read(&blp->blp_num_threads))
2851 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2852 blwi->blwi_mem_pressure))
2858 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2859 struct ldlm_bl_work_item *blwi)
2861 /* '1' for consistency with code that checks !mpflag to restore */
2862 unsigned int mpflags = 1;
2866 if (blwi->blwi_ns == NULL)
2867 /* added by ldlm_cleanup() */
2868 RETURN(LDLM_ITER_STOP);
2870 if (blwi->blwi_mem_pressure)
2871 mpflags = memalloc_noreclaim_save();
2873 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2875 if (blwi->blwi_count) {
2878 * The special case when we cancel locks in lru
2879 * asynchronously, we pass the list of locks here.
2880 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2881 * canceled locally yet.
2883 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2886 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2888 } else if (blwi->blwi_lock) {
2889 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2892 ldlm_pool_recalc(&blwi->blwi_ns->ns_pool, true);
2893 spin_lock(&blwi->blwi_ns->ns_lock);
2894 blwi->blwi_ns->ns_rpc_recalc = 0;
2895 spin_unlock(&blwi->blwi_ns->ns_lock);
2896 ldlm_namespace_put(blwi->blwi_ns);
2899 if (blwi->blwi_mem_pressure)
2900 memalloc_noreclaim_restore(mpflags);
2902 if (blwi->blwi_flags & LCF_ASYNC)
2903 OBD_FREE(blwi, sizeof(*blwi));
2905 complete(&blwi->blwi_comp);
2911 * Cancel stale locks on export. Cancel blocked locks first.
2912 * If the given export has blocked locks, the next in the list may have
2913 * them too, thus cancel not blocked locks only if the current export has
2916 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2917 struct obd_export *exp)
2923 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2925 num = ldlm_export_cancel_blocked_locks(exp);
2927 ldlm_export_cancel_locks(exp);
2929 obd_stale_export_put(exp);
2936 * Main blocking requests processing thread.
2938 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2939 * This thread in the end ends up doing actual call to ->l_blocking_ast
2942 static int ldlm_bl_thread_main(void *arg)
2945 struct ldlm_bl_pool *blp;
2946 struct ldlm_bl_thread_data *bltd = arg;
2954 rc = lu_env_init(env, LCT_DT_THREAD);
2957 rc = lu_env_add(env);
2959 GOTO(out_env_fini, rc);
2961 blp = bltd->bltd_blp;
2963 complete(&bltd->bltd_comp);
2964 /* cannot use bltd after this, it is only on caller's stack */
2967 struct ldlm_bl_work_item *blwi = NULL;
2968 struct obd_export *exp = NULL;
2971 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2974 wait_event_idle_exclusive(blp->blp_waitq,
2975 ldlm_bl_get_work(blp, &blwi,
2977 atomic_inc(&blp->blp_busy_threads);
2979 if (ldlm_bl_thread_need_create(blp, blwi))
2980 /* discard the return value, we tried */
2981 ldlm_bl_thread_start(blp, true);
2984 rc = ldlm_bl_thread_exports(blp, exp);
2986 rc = ldlm_bl_thread_blwi(blp, blwi);
2988 atomic_dec(&blp->blp_busy_threads);
2990 if (rc == LDLM_ITER_STOP)
2994 * If there are many namespaces, we will not sleep waiting for
2995 * work, and must do a cond_resched to avoid holding the CPU
3001 atomic_dec(&blp->blp_num_threads);
3002 complete(&blp->blp_comp);
3013 static int ldlm_setup(void);
3014 static int ldlm_cleanup(void);
3016 int ldlm_get_ref(void)
3021 mutex_lock(&ldlm_ref_mutex);
3022 if (++ldlm_refcount == 1) {
3027 mutex_unlock(&ldlm_ref_mutex);
3032 void ldlm_put_ref(void)
3035 mutex_lock(&ldlm_ref_mutex);
3036 if (ldlm_refcount == 1) {
3037 int rc = ldlm_cleanup();
3040 CERROR("ldlm_cleanup failed: %d\n", rc);
3046 mutex_unlock(&ldlm_ref_mutex);
3052 * Export handle<->lock hash operations.
3055 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
3057 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
3061 ldlm_export_lock_key(struct hlist_node *hnode)
3063 struct ldlm_lock *lock;
3065 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3066 return &lock->l_remote_handle;
3070 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3072 struct ldlm_lock *lock;
3074 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3075 lock->l_remote_handle = *(struct lustre_handle *)key;
3079 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3081 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3085 ldlm_export_lock_object(struct hlist_node *hnode)
3087 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3091 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3093 struct ldlm_lock *lock;
3095 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3096 LDLM_LOCK_GET(lock);
3100 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3102 struct ldlm_lock *lock;
3104 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3105 LDLM_LOCK_RELEASE(lock);
3108 static struct cfs_hash_ops ldlm_export_lock_ops = {
3109 .hs_hash = ldlm_export_lock_hash,
3110 .hs_key = ldlm_export_lock_key,
3111 .hs_keycmp = ldlm_export_lock_keycmp,
3112 .hs_keycpy = ldlm_export_lock_keycpy,
3113 .hs_object = ldlm_export_lock_object,
3114 .hs_get = ldlm_export_lock_get,
3115 .hs_put = ldlm_export_lock_put,
3116 .hs_put_locked = ldlm_export_lock_put,
3119 int ldlm_init_export(struct obd_export *exp)
3125 exp->exp_lock_hash =
3126 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3127 HASH_EXP_LOCK_CUR_BITS,
3128 HASH_EXP_LOCK_MAX_BITS,
3129 HASH_EXP_LOCK_BKT_BITS, 0,
3130 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3131 &ldlm_export_lock_ops,
3132 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3133 CFS_HASH_NBLK_CHANGE);
3135 if (!exp->exp_lock_hash)
3138 rc = ldlm_init_flock_export(exp);
3144 ldlm_destroy_export(exp);
3147 EXPORT_SYMBOL(ldlm_init_export);
3149 void ldlm_destroy_export(struct obd_export *exp)
3152 cfs_hash_putref(exp->exp_lock_hash);
3153 exp->exp_lock_hash = NULL;
3155 ldlm_destroy_flock_export(exp);
3158 EXPORT_SYMBOL(ldlm_destroy_export);
3160 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3161 struct attribute *attr,
3164 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3167 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3168 struct attribute *attr,
3175 rc = kstrtoul(buffer, 10, &val);
3179 ldlm_cancel_unused_locks_before_replay = val;
3183 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3185 static struct attribute *ldlm_attrs[] = {
3186 &lustre_attr_cancel_unused_locks_before_replay.attr,
3190 static struct attribute_group ldlm_attr_group = {
3191 .attrs = ldlm_attrs,
3194 static int ldlm_setup(void)
3196 static struct ptlrpc_service_conf conf;
3197 struct ldlm_bl_pool *blp = NULL;
3198 #ifdef HAVE_SERVER_SUPPORT
3199 struct task_struct *task;
3200 #endif /* HAVE_SERVER_SUPPORT */
3206 if (ldlm_state != NULL)
3209 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3210 if (ldlm_state == NULL)
3213 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3217 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3221 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3225 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3229 rc = ldlm_debugfs_setup();
3233 memset(&conf, 0, sizeof(conf));
3234 conf = (typeof(conf)) {
3235 .psc_name = "ldlm_cbd",
3236 .psc_watchdog_factor = 2,
3238 .bc_nbufs = LDLM_CLIENT_NBUFS,
3239 .bc_buf_size = LDLM_BUFSIZE,
3240 .bc_req_max_size = LDLM_MAXREQSIZE,
3241 .bc_rep_max_size = LDLM_MAXREPSIZE,
3242 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3243 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3246 .tc_thr_name = "ldlm_cb",
3247 .tc_thr_factor = LDLM_THR_FACTOR,
3248 .tc_nthrs_init = LDLM_NTHRS_INIT,
3249 .tc_nthrs_base = LDLM_NTHRS_BASE,
3250 .tc_nthrs_max = LDLM_NTHRS_MAX,
3251 .tc_nthrs_user = ldlm_num_threads,
3252 .tc_cpu_bind = ldlm_cpu_bind,
3253 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3256 .cc_pattern = ldlm_cpts,
3257 .cc_affinity = true,
3260 .so_req_handler = ldlm_callback_handler,
3263 ldlm_state->ldlm_cb_service = \
3264 ptlrpc_register_service(&conf, ldlm_svc_kset,
3265 ldlm_svc_debugfs_dir);
3266 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3267 CERROR("failed to start service\n");
3268 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3269 ldlm_state->ldlm_cb_service = NULL;
3273 #ifdef HAVE_SERVER_SUPPORT
3274 memset(&conf, 0, sizeof(conf));
3275 conf = (typeof(conf)) {
3276 .psc_name = "ldlm_canceld",
3277 .psc_watchdog_factor = 6,
3279 .bc_nbufs = LDLM_SERVER_NBUFS,
3280 .bc_buf_size = LDLM_BUFSIZE,
3281 .bc_req_max_size = LDLM_MAXREQSIZE,
3282 .bc_rep_max_size = LDLM_MAXREPSIZE,
3283 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3284 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3288 .tc_thr_name = "ldlm_cn",
3289 .tc_thr_factor = LDLM_THR_FACTOR,
3290 .tc_nthrs_init = LDLM_NTHRS_INIT,
3291 .tc_nthrs_base = LDLM_NTHRS_BASE,
3292 .tc_nthrs_max = LDLM_NTHRS_MAX,
3293 .tc_nthrs_user = ldlm_num_threads,
3294 .tc_cpu_bind = ldlm_cpu_bind,
3295 .tc_ctx_tags = LCT_MD_THREAD | \
3300 .cc_pattern = ldlm_cpts,
3301 .cc_affinity = true,
3304 .so_req_handler = ldlm_cancel_handler,
3305 .so_hpreq_handler = ldlm_hpreq_handler,
3308 ldlm_state->ldlm_cancel_service = \
3309 ptlrpc_register_service(&conf, ldlm_svc_kset,
3310 ldlm_svc_debugfs_dir);
3311 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3312 CERROR("failed to start service\n");
3313 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3314 ldlm_state->ldlm_cancel_service = NULL;
3317 #endif /* HAVE_SERVER_SUPPORT */
3319 OBD_ALLOC(blp, sizeof(*blp));
3321 GOTO(out, rc = -ENOMEM);
3322 ldlm_state->ldlm_bl_pool = blp;
3324 spin_lock_init(&blp->blp_lock);
3325 INIT_LIST_HEAD(&blp->blp_list);
3326 INIT_LIST_HEAD(&blp->blp_prio_list);
3327 init_waitqueue_head(&blp->blp_waitq);
3328 atomic_set(&blp->blp_num_threads, 0);
3329 atomic_set(&blp->blp_busy_threads, 0);
3330 blp->blp_total_locks = 0;
3331 blp->blp_total_blwis = 0;
3333 if (ldlm_num_threads == 0) {
3334 blp->blp_min_threads = LDLM_NTHRS_INIT;
3335 blp->blp_max_threads = LDLM_NTHRS_MAX;
3337 blp->blp_min_threads = blp->blp_max_threads = \
3338 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3342 for (i = 0; i < blp->blp_min_threads; i++) {
3343 rc = ldlm_bl_thread_start(blp, false);
3348 #ifdef HAVE_SERVER_SUPPORT
3349 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3352 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3356 wait_event(expired_lock_wait_queue,
3357 expired_lock_thread_state == ELT_READY);
3358 #endif /* HAVE_SERVER_SUPPORT */
3360 rc = ldlm_pools_init();
3362 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3366 rc = ldlm_reclaim_setup();
3368 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3378 static int ldlm_cleanup(void)
3382 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3383 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3384 CERROR("ldlm still has namespaces; clean these up first.\n");
3385 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3386 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3390 ldlm_reclaim_cleanup();
3393 if (ldlm_state->ldlm_bl_pool != NULL) {
3394 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3396 while (atomic_read(&blp->blp_num_threads) > 0) {
3397 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3399 init_completion(&blp->blp_comp);
3401 spin_lock(&blp->blp_lock);
3402 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3403 wake_up(&blp->blp_waitq);
3404 spin_unlock(&blp->blp_lock);
3406 wait_for_completion(&blp->blp_comp);
3409 OBD_FREE(blp, sizeof(*blp));
3412 if (ldlm_state->ldlm_cb_service != NULL)
3413 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3414 #ifdef HAVE_SERVER_SUPPORT
3415 if (ldlm_state->ldlm_cancel_service != NULL)
3416 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3420 kset_unregister(ldlm_ns_kset);
3422 kset_unregister(ldlm_svc_kset);
3424 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3425 kobject_put(ldlm_kobj);
3428 ldlm_debugfs_cleanup();
3430 #ifdef HAVE_SERVER_SUPPORT
3431 if (expired_lock_thread_state != ELT_STOPPED) {
3432 expired_lock_thread_state = ELT_TERMINATE;
3433 wake_up(&expired_lock_wait_queue);
3434 wait_event(expired_lock_wait_queue,
3435 expired_lock_thread_state == ELT_STOPPED);
3439 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3447 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3448 sizeof(struct ldlm_resource), 0,
3449 SLAB_HWCACHE_ALIGN, NULL);
3450 if (ldlm_resource_slab == NULL)
3453 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3454 sizeof(struct ldlm_lock), 0,
3455 SLAB_HWCACHE_ALIGN, NULL);
3456 if (ldlm_lock_slab == NULL)
3459 ldlm_interval_slab = kmem_cache_create("interval_node",
3460 sizeof(struct ldlm_interval),
3461 0, SLAB_HWCACHE_ALIGN, NULL);
3462 if (ldlm_interval_slab == NULL)
3465 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3466 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3467 0, SLAB_HWCACHE_ALIGN, NULL);
3468 if (ldlm_interval_tree_slab == NULL)
3471 #ifdef HAVE_SERVER_SUPPORT
3472 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3473 sizeof(struct ldlm_ibits_node),
3474 0, SLAB_HWCACHE_ALIGN, NULL);
3475 if (ldlm_inodebits_slab == NULL)
3476 goto out_interval_tree;
3478 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3479 sizeof(struct ldlm_glimpse_work),
3481 if (ldlm_glimpse_work_kmem == NULL)
3485 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3486 class_export_dump_hook = ldlm_dump_export_locks;
3489 #ifdef HAVE_SERVER_SUPPORT
3491 kmem_cache_destroy(ldlm_inodebits_slab);
3493 kmem_cache_destroy(ldlm_interval_tree_slab);
3496 kmem_cache_destroy(ldlm_interval_slab);
3498 kmem_cache_destroy(ldlm_lock_slab);
3500 kmem_cache_destroy(ldlm_resource_slab);
3505 void ldlm_exit(void)
3508 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3510 kmem_cache_destroy(ldlm_resource_slab);
3512 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3513 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3514 * so that ldlm_lock_free() get a chance to be called.
3517 kmem_cache_destroy(ldlm_lock_slab);
3518 kmem_cache_destroy(ldlm_interval_slab);
3519 kmem_cache_destroy(ldlm_interval_tree_slab);
3520 #ifdef HAVE_SERVER_SUPPORT
3521 kmem_cache_destroy(ldlm_inodebits_slab);
3522 kmem_cache_destroy(ldlm_glimpse_work_kmem);