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 del_timer(&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 * Calculate the per-export Blocking timeout by the given RPC (covering the
652 * reply to this RPC and the next RPC). The next RPC could be still not CANCEL,
653 * but having the lock refresh mechanism it is enough.
655 * Used for lock refresh timeout when we are in the middle of the process -
656 * BL AST is sent, CANCEL is ahead - it is still 1 reply for the current RPC
657 * and at least 1 RPC (which will trigger another refresh if it will be not
658 * CANCEL) - but more accurate than ldlm_bl_timeout as the timeout is taken
659 * from the RPC (i.e. the view of the client on the current AT) is taken into
662 * \param[in] req req which export needs the timeout calculation
664 * \retval timeout in seconds to wait for the next client's RPC
666 timeout_t ldlm_bl_timeout_by_rpc(struct ptlrpc_request *req)
668 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
669 timeout_t timeout, req_timeout, at_timeout, netl;
672 return obd_timeout / 2;
674 /* A blocked lock means somebody in the cluster is waiting, and we
675 * should not consider the worst ever case, consisting of a chain of
676 * failures on each step, however this timeout should survive a
677 * recovery of at least 1 failure, let this one to be the worst one:
678 * in case a server NID is dead first re-connect is done through the
679 * same router and also times out.
681 * Either this on the next RPC times out, take the max.
682 * Considering the current RPC, take just the left time.
684 netl = at_get(&req->rq_export->exp_imp_reverse->imp_at.iat_net_latency);
685 req_timeout = req->rq_deadline - ktime_get_real_seconds() + netl;
686 at_timeout = at_est2timeout(at_get(&svcpt->scp_at_estimate)) + netl;
687 req_timeout = max(req_timeout, at_timeout);
689 /* Take 1 re-connect failure and 1 re-connect success into account. */
690 timeout = at_timeout + INITIAL_CONNECT_TIMEOUT + netl + req_timeout;
692 /* Client's timeout is calculated as at_est2timeout(), let's be a bit
693 * more conservative than client */
694 return max(timeout + (timeout >> 4), (timeout_t)ldlm_enqueue_min);
696 EXPORT_SYMBOL(ldlm_bl_timeout_by_rpc);
699 * Perform lock cleanup if AST sending failed.
701 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
702 const char *ast_type)
704 LCONSOLE_ERROR_MSG(0x138,
705 "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
706 lock->l_export->exp_obd->obd_name,
707 obd_export_nid2str(lock->l_export), ast_type, rc);
709 if (obd_dump_on_timeout)
710 libcfs_debug_dumplog();
711 spin_lock_bh(&waiting_locks_spinlock);
712 if (__ldlm_del_waiting_lock(lock) == 0)
714 * the lock was not in any list, grab an extra ref before adding
715 * the lock to the expired list
718 /* differentiate it from expired locks */
719 lock->l_callback_timestamp = 0;
720 list_add(&lock->l_pending_chain, &expired_lock_list);
721 wake_up(&expired_lock_wait_queue);
722 spin_unlock_bh(&waiting_locks_spinlock);
726 * Perform lock cleanup if AST reply came with error.
728 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
729 struct ptlrpc_request *req, int rc,
730 const char *ast_type)
732 struct lnet_processid *peer = &req->rq_import->imp_connection->c_peer;
734 if (!req->rq_replied || (rc && rc != -EINVAL)) {
735 if (ldlm_is_cancel(lock)) {
737 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
738 ast_type, req, req->rq_xid,
739 libcfs_nidstr(&peer->nid));
740 ldlm_lock_cancel(lock);
742 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
744 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
746 * Upon umount process the AST fails because cannot be
747 * sent. This shouldn't lead to the client eviction.
748 * -ENODEV error is returned by ptl_send_rpc() for
749 * new request in such import.
750 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
751 * if imp_invalid is set or obd_no_recov.
752 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
753 * in ptlrpc_import_delay_req() as well with -EIO code.
754 * In all such cases errors are ignored.
757 "%s AST can't be sent due to a server %s failure or umount process: rc = %d",
759 req->rq_import->imp_obd->obd_name, rc);
762 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
763 libcfs_nidstr(&peer->nid),
764 req->rq_replied ? "returned error from" :
765 "failed to reply to",
766 ast_type, req, req->rq_xid,
767 (req->rq_repmsg != NULL) ?
768 lustre_msg_get_status(req->rq_repmsg) : 0,
770 ldlm_failed_ast(lock, rc, ast_type);
776 struct ldlm_resource *res = lock->l_resource;
779 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
780 libcfs_nidstr(&peer->nid),
782 lustre_msg_get_status(req->rq_repmsg) : -1,
783 ast_type, req, req->rq_xid);
786 * update lvbo to return proper attributes.
789 ldlm_resource_getref(res);
790 ldlm_lvbo_update(res, lock, NULL, 1);
791 ldlm_resource_putref(res);
793 ldlm_lock_cancel(lock);
800 static int ldlm_cb_interpret(const struct lu_env *env,
801 struct ptlrpc_request *req, void *args, int rc)
803 struct ldlm_cb_async_args *ca = args;
804 struct ldlm_lock *lock = ca->ca_lock;
805 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
809 LASSERT(lock != NULL);
812 case LDLM_GL_CALLBACK:
814 * Update the LVB from disk if the AST failed
815 * (this is a legal race)
817 * - Glimpse callback of local lock just returns
818 * -ELDLM_NO_LOCK_DATA.
819 * - Glimpse callback of remote lock might return
820 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
822 if (unlikely(arg->gl_interpret_reply)) {
823 rc = arg->gl_interpret_reply(NULL, req, args, rc);
824 } else if (rc == -ELDLM_NO_LOCK_DATA) {
826 "lost race - client has a lock but no inode");
827 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
828 } else if (rc != 0) {
829 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
831 rc = ldlm_lvbo_update(lock->l_resource,
835 case LDLM_BL_CALLBACK:
837 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
839 case LDLM_CP_CALLBACK:
841 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
844 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
849 /* release extra reference taken in ldlm_ast_fini() */
850 LDLM_LOCK_RELEASE(lock);
853 atomic_inc(&arg->restart);
858 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
860 struct ldlm_cb_async_args *ca = data;
861 struct ldlm_lock *lock = ca->ca_lock;
863 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
866 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
867 struct ldlm_cb_set_arg *arg,
868 struct ldlm_lock *lock,
875 if (unlikely(instant_cancel)) {
876 rc = ptl_send_rpc(req, 1);
877 ptlrpc_req_finished(req);
879 atomic_inc(&arg->restart);
882 ptlrpc_set_add_req(arg->set, req);
889 * Check if there are requests in the export request list which prevent
890 * the lock canceling and make these requests high priority ones.
892 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
894 struct ptlrpc_request *req;
898 if (lock->l_export == NULL) {
899 LDLM_DEBUG(lock, "client lock: no-op");
903 spin_lock(&lock->l_export->exp_rpc_lock);
904 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
907 * Do not process requests that were not yet added to there
908 * incoming queue or were already removed from there for
909 * processing. We evaluate ptlrpc_nrs_req_can_move() without
910 * holding svcpt->scp_req_lock, and then redo the check with
911 * the lock held once we need to obtain a reliable result.
913 if (ptlrpc_nrs_req_can_move(req) &&
914 req->rq_ops->hpreq_lock_match &&
915 req->rq_ops->hpreq_lock_match(req, lock))
916 ptlrpc_nrs_req_hp_move(req);
918 spin_unlock(&lock->l_export->exp_rpc_lock);
923 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
924 * enqueued server lock conflicts with given one.
926 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
927 * to wait for client response.
929 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
930 struct ldlm_lock_desc *desc,
931 void *data, int flag)
933 struct ldlm_cb_async_args *ca;
934 struct ldlm_cb_set_arg *arg = data;
935 struct ldlm_request *body;
936 struct ptlrpc_request *req;
937 int instant_cancel = 0;
942 if (flag == LDLM_CB_CANCELING)
943 /* Don't need to do anything here. */
946 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
947 LDLM_DEBUG(lock, "dropping BL AST");
952 LASSERT(data != NULL);
953 if (lock->l_export->exp_obd->obd_recovering != 0)
954 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
956 ldlm_lock_reorder_req(lock);
958 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
959 &RQF_LDLM_BL_CALLBACK,
960 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
964 ca = ptlrpc_req_async_args(ca, req);
965 ca->ca_set_arg = arg;
968 req->rq_interpret_reply = ldlm_cb_interpret;
970 lock_res_and_lock(lock);
971 if (ldlm_is_destroyed(lock)) {
972 /* What's the point? */
973 unlock_res_and_lock(lock);
974 ptlrpc_req_finished(req);
978 if (!ldlm_is_granted(lock)) {
980 * this blocking AST will be communicated as part of the
981 * completion AST instead
983 ldlm_add_blocked_lock(lock);
984 ldlm_set_waited(lock);
985 unlock_res_and_lock(lock);
987 ptlrpc_req_finished(req);
988 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
992 if (ldlm_is_cancel_on_block(lock))
995 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
996 body->lock_handle[0] = lock->l_remote_handle;
997 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
998 body->lock_desc = *desc;
999 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
1001 LDLM_DEBUG(lock, "server preparing blocking AST");
1003 ptlrpc_request_set_replen(req);
1004 ldlm_set_cbpending(lock);
1005 if (instant_cancel) {
1006 unlock_res_and_lock(lock);
1007 ldlm_lock_cancel(lock);
1009 req->rq_no_resend = 1;
1011 LASSERT(ldlm_is_granted(lock));
1012 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1013 unlock_res_and_lock(lock);
1015 /* Do not resend after lock callback timeout */
1016 req->rq_delay_limit = ldlm_bl_timeout(lock);
1017 req->rq_resend_cb = ldlm_update_resend;
1020 req->rq_send_state = LUSTRE_IMP_FULL;
1021 /* ptlrpc_request_alloc_pack already set timeout */
1023 req->rq_timeout = ldlm_get_rq_timeout();
1025 if (lock->l_export && lock->l_export->exp_nid_stats &&
1026 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1027 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1028 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
1030 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1036 * ->l_completion_ast callback for a remote lock in server namespace.
1038 * Sends AST to the client notifying it of lock granting. If initial
1039 * lock response was not sent yet, instead of sending another RPC, just
1040 * mark the lock as granted and client will understand
1042 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
1044 struct ldlm_cb_set_arg *arg = data;
1045 struct ldlm_request *body;
1046 struct ptlrpc_request *req;
1047 struct ldlm_cb_async_args *ca;
1048 int instant_cancel = 0;
1054 LASSERT(lock != NULL);
1055 LASSERT(data != NULL);
1057 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1058 LDLM_DEBUG(lock, "dropping CP AST");
1062 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1063 &RQF_LDLM_CP_CALLBACK);
1067 /* server namespace, doesn't need lock */
1068 lvb_len = ldlm_lvbo_size(lock);
1070 * LU-3124 & LU-2187: to not return layout in completion AST because
1071 * it may deadlock for LU-2187, or client may not have enough space
1072 * for large layout. The layout will be returned to client with an
1073 * extra RPC to fetch xattr.lov
1075 if (ldlm_has_layout(lock))
1078 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1079 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1081 ptlrpc_request_free(req);
1085 ca = ptlrpc_req_async_args(ca, req);
1086 ca->ca_set_arg = arg;
1089 req->rq_interpret_reply = ldlm_cb_interpret;
1090 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1092 body->lock_handle[0] = lock->l_remote_handle;
1093 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
1094 body->lock_flags = ldlm_flags_to_wire(flags);
1095 ldlm_lock2desc(lock, &body->lock_desc);
1097 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1098 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1101 * We still need to send the RPC to wake up the blocked
1102 * enqueue thread on the client.
1104 * Consider old client, there is no better way to notify
1105 * the failure, just zero-sized the LVB, then the client
1106 * will fail out as "-EPROTO".
1108 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1112 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1117 LDLM_DEBUG(lock, "server preparing completion AST");
1119 ptlrpc_request_set_replen(req);
1121 req->rq_send_state = LUSTRE_IMP_FULL;
1122 /* ptlrpc_request_pack already set timeout */
1124 req->rq_timeout = ldlm_get_rq_timeout();
1126 /* We only send real blocking ASTs after the lock is granted */
1127 lock_res_and_lock(lock);
1128 if (ldlm_is_ast_sent(lock)) {
1129 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1130 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1131 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1135 * We might get here prior to ldlm_handle_enqueue setting
1136 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1137 * into waiting list, but this is safe and similar code in
1138 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1139 * that would not only cancel the lock, but will also remove
1140 * it from waiting list
1142 if (ldlm_is_cancel_on_block(lock)) {
1143 unlock_res_and_lock(lock);
1144 ldlm_lock_cancel(lock);
1147 req->rq_no_resend = 1;
1149 lock_res_and_lock(lock);
1151 /* start the lock-timeout clock */
1152 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1153 /* Do not resend after lock callback timeout */
1154 req->rq_delay_limit = ldlm_bl_timeout(lock);
1155 req->rq_resend_cb = ldlm_update_resend;
1158 unlock_res_and_lock(lock);
1160 if (lock->l_export && lock->l_export->exp_nid_stats &&
1161 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1162 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1163 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1165 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1167 RETURN(lvb_len < 0 ? lvb_len : rc);
1171 * Server side ->l_glimpse_ast handler for client locks.
1173 * Sends glimpse AST to the client and waits for reply. Then updates
1174 * lvbo with the result.
1176 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1178 struct ldlm_cb_set_arg *arg = data;
1179 struct ldlm_request *body;
1180 struct ptlrpc_request *req;
1181 struct ldlm_cb_async_args *ca;
1183 struct req_format *req_fmt;
1187 LASSERT(lock != NULL);
1189 if (arg->gl_desc != NULL)
1190 /* There is a glimpse descriptor to pack */
1191 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1193 req_fmt = &RQF_LDLM_GL_CALLBACK;
1195 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1196 req_fmt, LUSTRE_DLM_VERSION,
1202 if (arg->gl_desc != NULL) {
1203 /* copy the GL descriptor */
1204 union ldlm_gl_desc *desc;
1206 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1207 *desc = *arg->gl_desc;
1210 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1211 body->lock_handle[0] = lock->l_remote_handle;
1212 ldlm_lock2desc(lock, &body->lock_desc);
1214 ca = ptlrpc_req_async_args(ca, req);
1215 ca->ca_set_arg = arg;
1218 /* server namespace, doesn't need lock */
1219 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1220 ldlm_lvbo_size(lock));
1221 ptlrpc_request_set_replen(req);
1223 req->rq_send_state = LUSTRE_IMP_FULL;
1224 /* ptlrpc_request_alloc_pack already set timeout */
1226 req->rq_timeout = ldlm_get_rq_timeout();
1228 req->rq_interpret_reply = ldlm_cb_interpret;
1230 if (lock->l_export && lock->l_export->exp_nid_stats) {
1231 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1233 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1234 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1237 rc = ldlm_ast_fini(req, arg, lock, 0);
1241 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1243 int ldlm_glimpse_locks(struct ldlm_resource *res,
1244 struct list_head *gl_work_list)
1250 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1252 if (rc == -ERESTART)
1253 ldlm_reprocess_all(res, 0);
1257 EXPORT_SYMBOL(ldlm_glimpse_locks);
1259 /* return LDLM lock associated with a lock callback request */
1260 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1262 struct ldlm_cb_async_args *ca;
1263 struct ldlm_lock *lock;
1267 ca = ptlrpc_req_async_args(ca, req);
1270 RETURN(ERR_PTR(-EFAULT));
1274 EXPORT_SYMBOL(ldlm_request_lock);
1277 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1278 * service threads to carry out client lock enqueueing requests.
1280 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1281 struct ptlrpc_request *req,
1282 const struct ldlm_request *dlm_req,
1283 const struct ldlm_callback_suite *cbs)
1285 struct ldlm_reply *dlm_rep;
1287 enum ldlm_error err = ELDLM_OK;
1288 struct ldlm_lock *lock = NULL;
1289 void *cookie = NULL;
1291 struct ldlm_resource *res = NULL;
1292 const struct lu_env *env = req->rq_svc_thread->t_env;
1296 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1298 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1299 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1301 LASSERT(req->rq_export);
1303 /* for intent enqueue the stat will be updated inside intent policy */
1304 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1305 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1306 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1308 if (req->rq_export && req->rq_export->exp_nid_stats &&
1309 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1310 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1311 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1313 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1314 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1315 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1316 dlm_req->lock_desc.l_resource.lr_type);
1317 GOTO(out, rc = -EFAULT);
1320 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1321 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1322 dlm_req->lock_desc.l_req_mode &
1323 (dlm_req->lock_desc.l_req_mode-1))) {
1324 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1325 dlm_req->lock_desc.l_req_mode);
1326 GOTO(out, rc = -EFAULT);
1329 if (unlikely((flags & LDLM_FL_REPLAY) ||
1330 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1331 /* Find an existing lock in the per-export lock hash */
1333 * In the function below, .hs_keycmp resolves to
1334 * ldlm_export_lock_keycmp()
1336 /* coverity[overrun-buffer-val] */
1337 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1338 (void *)&dlm_req->lock_handle[0]);
1340 DEBUG_REQ(D_DLMTRACE, req,
1341 "found existing lock cookie %#llx",
1342 lock->l_handle.h_cookie);
1343 flags |= LDLM_FL_RESENT;
1344 GOTO(existing_lock, rc = 0);
1347 if (ldlm_reclaim_full()) {
1348 DEBUG_REQ(D_DLMTRACE, req,
1349 "Too many granted locks, reject current enqueue request and let the client retry later");
1350 GOTO(out, rc = -EINPROGRESS);
1354 /* The lock's callback data might be set in the policy function */
1355 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1356 dlm_req->lock_desc.l_resource.lr_type,
1357 dlm_req->lock_desc.l_req_mode,
1358 cbs, NULL, 0, LVB_T_NONE);
1365 lock->l_remote_handle = dlm_req->lock_handle[0];
1366 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1369 * Initialize resource lvb but not for a lock being replayed since
1370 * Client already got lvb sent in this case.
1371 * This must occur early since some policy methods assume resource
1372 * lvb is available (lr_lvb_data != NULL).
1374 res = lock->l_resource;
1375 if (!(flags & LDLM_FL_REPLAY)) {
1376 /* non-replayed lock, delayed lvb init may need to be done */
1377 rc = ldlm_lvbo_init(res);
1379 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1384 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1386 * Don't enqueue a lock onto the export if it is been disonnected
1387 * due to eviction (b=3822) or server umount (b=24324).
1388 * Cancel it now instead.
1390 if (req->rq_export->exp_disconnected) {
1391 LDLM_ERROR(lock, "lock on disconnected export %p",
1393 GOTO(out, rc = -ENOTCONN);
1396 lock->l_export = class_export_lock_get(req->rq_export, lock);
1397 if (lock->l_export->exp_lock_hash)
1398 cfs_hash_add(lock->l_export->exp_lock_hash,
1399 &lock->l_remote_handle,
1403 * Inherit the enqueue flags before the operation, because we do not
1404 * keep the res lock on return and next operations (BL AST) may proceed
1407 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1408 LDLM_FL_INHERIT_MASK);
1410 ldlm_convert_policy_to_local(req->rq_export,
1411 dlm_req->lock_desc.l_resource.lr_type,
1412 &dlm_req->lock_desc.l_policy_data,
1413 &lock->l_policy_data);
1414 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
1415 lock->l_req_extent = lock->l_policy_data.l_extent;
1416 } else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) {
1417 lock->l_policy_data.l_inodebits.try_bits =
1418 dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1419 lock->l_policy_data.l_inodebits.li_gid =
1420 dlm_req->lock_desc.l_policy_data.l_inodebits.li_gid;
1425 if (!(flags & LDLM_FL_HAS_INTENT)) {
1426 /* based on the assumption that lvb size never changes during
1427 * resource life time otherwise it need resource->lr_lock's
1429 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1430 RCL_SERVER, ldlm_lvbo_size(lock));
1432 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1433 GOTO(out, rc = -ENOMEM);
1435 rc = req_capsule_server_pack(&req->rq_pill);
1440 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1447 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1449 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1450 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1452 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1453 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1456 * We never send a blocking AST until the lock is granted, but
1457 * we can tell it right now
1459 lock_res_and_lock(lock);
1462 * Now take into account flags to be inherited from original lock
1463 * request both in reply to client and in our own lock flags.
1465 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1466 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1469 * Don't move a pending lock onto the export if it has already been
1470 * disconnected due to eviction (b=5683) or server umount (b=24324).
1471 * Cancel it now instead.
1473 if (unlikely(req->rq_export->exp_disconnected ||
1474 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1475 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1477 } else if (ldlm_is_ast_sent(lock)) {
1478 /* fill lock desc for possible lock convert */
1479 if (lock->l_blocking_lock &&
1480 lock->l_resource->lr_type == LDLM_IBITS) {
1481 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1482 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1485 "save blocking bits %llx in granted lock",
1486 bl_lock->l_policy_data.l_inodebits.bits);
1488 * If lock is blocked then save blocking ibits
1489 * in returned lock policy for the possible lock
1490 * convert on a client.
1492 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1493 bl_lock->l_policy_data.l_inodebits.bits;
1495 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1496 if (ldlm_is_granted(lock)) {
1498 * Only cancel lock if it was granted, because it would
1499 * be destroyed immediately and would never be granted
1500 * in the future, causing timeouts on client. Not
1501 * granted lock will be cancelled immediately after
1502 * sending completion AST.
1504 if (ldlm_is_cancel_on_block(lock)) {
1505 unlock_res_and_lock(lock);
1506 ldlm_lock_cancel(lock);
1507 lock_res_and_lock(lock);
1509 ldlm_add_waiting_lock(lock,
1510 ldlm_bl_timeout(lock));
1514 unlock_res_and_lock(lock);
1518 req->rq_status = rc ?: err; /* return either error - b=11190 */
1519 if (!req->rq_packed_final) {
1520 int rc1 = lustre_pack_reply(req, 1, NULL, NULL);
1526 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1527 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1531 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1535 req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1537 ldlm_lvbo_size(lock) > 0) {
1542 buf = req_capsule_server_get(&req->rq_pill,
1544 LASSERTF(buf != NULL, "req %p, lock %p\n", req, lock);
1545 buflen = req_capsule_get_size(&req->rq_pill,
1546 &RMF_DLM_LVB, RCL_SERVER);
1548 * non-replayed lock, delayed lvb init may
1549 * need to be occur now
1551 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1554 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1556 req_capsule_shrink(&req->rq_pill,
1559 } else if (rc2 == -ERANGE) {
1560 rc2 = req_capsule_server_grow(
1562 &RMF_DLM_LVB, buflen);
1567 * if we can't grow the buffer,
1568 * it's ok to return empty lvb
1579 } else if (flags & LDLM_FL_REPLAY) {
1580 /* no LVB resend upon replay */
1582 req_capsule_shrink(&req->rq_pill,
1592 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1593 if (lock->l_export) {
1594 ldlm_lock_cancel(lock);
1596 lock_res_and_lock(lock);
1597 ldlm_resource_unlink_lock(lock);
1598 ldlm_lock_destroy_nolock(lock);
1599 unlock_res_and_lock(lock);
1601 ldlm_reprocess_all(lock->l_resource, lock->l_policy_data.l_inodebits.bits);
1604 if (!err && !ldlm_is_cbpending(lock) &&
1605 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1606 ldlm_reprocess_all(lock->l_resource,
1607 lock->l_policy_data.l_inodebits.bits);
1609 LDLM_LOCK_RELEASE(lock);
1612 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1619 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1620 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1622 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1624 if (lock->l_blocking_lock) {
1625 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1626 lock->l_blocking_lock = NULL;
1630 /* A lock can be converted to new ibits or mode and should be considered
1631 * as new lock. Clear all states related to a previous blocking AST
1632 * processing so new conflicts will cause new blocking ASTs.
1634 * This is used during lock convert below and lock downgrade to COS mode in
1635 * ldlm_lock_mode_downgrade().
1637 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1639 ldlm_clear_ast_sent(lock);
1640 lock->l_bl_ast_run = 0;
1641 ldlm_clear_blocking_lock(lock);
1645 * Main LDLM entry point for server code to process lock conversion requests.
1647 int ldlm_handle_convert0(struct ptlrpc_request *req,
1648 const struct ldlm_request *dlm_req)
1650 struct obd_export *exp = req->rq_export;
1651 struct ldlm_reply *dlm_rep;
1652 struct ldlm_lock *lock;
1659 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1660 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1661 LDLM_CONVERT - LDLM_FIRST_OPC);
1663 rc = req_capsule_server_pack(&req->rq_pill);
1667 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1668 dlm_rep->lock_flags = dlm_req->lock_flags;
1670 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1672 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1673 req->rq_status = ELDLM_NO_LOCK_DATA;
1677 LDLM_DEBUG(lock, "server-side convert handler START");
1679 lock_res_and_lock(lock);
1680 bits = lock->l_policy_data.l_inodebits.bits;
1681 new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1683 if (ldlm_is_cancel(lock)) {
1684 LDLM_DEBUG(lock, "convert on canceled lock!");
1685 unlock_res_and_lock(lock);
1686 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1689 if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1690 LDLM_ERROR(lock, "lock mode differs!");
1691 unlock_res_and_lock(lock);
1692 GOTO(out_put, rc = -EPROTO);
1695 if (bits == new_bits) {
1697 * This can be valid situation if CONVERT RPCs are
1698 * re-ordered. Just finish silently
1700 LDLM_DEBUG(lock, "lock is converted already!");
1701 unlock_res_and_lock(lock);
1703 if (ldlm_is_waited(lock))
1704 ldlm_del_waiting_lock(lock);
1706 ldlm_clear_cbpending(lock);
1707 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1708 ldlm_inodebits_drop(lock, bits & ~new_bits);
1710 ldlm_clear_blocking_data(lock);
1711 unlock_res_and_lock(lock);
1713 /* All old bits should be reprocessed to send new BL AST if
1714 * it wasn't sent earlier due to LDLM_FL_AST_SENT bit set.
1716 ldlm_reprocess_all(lock->l_resource, bits);
1719 dlm_rep->lock_handle = lock->l_remote_handle;
1720 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1721 &dlm_rep->lock_desc.l_policy_data);
1725 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1726 LDLM_LOCK_PUT(lock);
1727 req->rq_status = rc;
1732 * Cancel all the locks whose handles are packed into ldlm_request
1734 * Called by server code expecting such combined cancel activity
1737 int ldlm_request_cancel(struct ptlrpc_request *req,
1738 const struct ldlm_request *dlm_req,
1739 int first, enum lustre_at_flags flags)
1741 struct ldlm_resource *res, *pres = NULL;
1742 struct ldlm_lock *lock;
1743 int i, count, done = 0;
1748 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1749 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1750 (size - offsetof(struct ldlm_request, lock_handle)) /
1751 sizeof(struct lustre_handle) < dlm_req->lock_count)
1754 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1758 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1762 * There is no lock on the server at the replay time,
1763 * skip lock cancelling to make replay tests to pass.
1765 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1768 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1771 for (i = first; i < count; i++) {
1772 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1774 /* below message checked in replay-single.sh test_36 */
1775 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1776 dlm_req->lock_handle[i].cookie);
1780 res = lock->l_resource;
1784 * This code is an optimization to only attempt lock
1785 * granting on the resource (that could be CPU-expensive)
1786 * after we are done cancelling lock in that resource.
1790 ldlm_reprocess_all(pres, 0);
1791 LDLM_RESOURCE_DELREF(pres);
1792 ldlm_resource_putref(pres);
1795 ldlm_resource_getref(res);
1796 LDLM_RESOURCE_ADDREF(res);
1798 if (!ldlm_is_discard_data(lock))
1799 ldlm_lvbo_update(res, lock,
1805 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1806 lock->l_blast_sent != 0) {
1807 timeout_t delay = 0;
1809 if (ktime_get_seconds() > lock->l_blast_sent)
1810 delay = ktime_get_seconds() -
1813 "server cancels blocked lock after %ds",
1815 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1817 ldlm_lock_cancel(lock);
1818 LDLM_LOCK_PUT(lock);
1821 ldlm_reprocess_all(pres, 0);
1822 LDLM_RESOURCE_DELREF(pres);
1823 ldlm_resource_putref(pres);
1825 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1828 EXPORT_SYMBOL(ldlm_request_cancel);
1831 * Main LDLM entry point for server code to cancel locks.
1833 * Typically gets called from service handler on LDLM_CANCEL opc.
1835 int ldlm_handle_cancel(struct ptlrpc_request *req)
1837 struct ldlm_request *dlm_req;
1842 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1843 if (dlm_req == NULL) {
1844 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1848 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1849 offsetof(struct ldlm_request, lock_handle[1]))
1852 if (req->rq_export && req->rq_export->exp_nid_stats &&
1853 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1854 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1855 LDLM_CANCEL - LDLM_FIRST_OPC);
1857 rc = req_capsule_server_pack(&req->rq_pill);
1861 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1862 req->rq_status = LUSTRE_ESTALE;
1864 RETURN(ptlrpc_reply(req));
1866 #endif /* HAVE_SERVER_SUPPORT */
1869 * Server may pass additional information about blocking lock.
1870 * For IBITS locks it is conflicting bits which can be used for
1871 * lock convert instead of cancel.
1873 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1875 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1877 check_res_locked(lock->l_resource);
1878 if (ns_is_client(ns) && ld &&
1879 (lock->l_resource->lr_type == LDLM_IBITS)) {
1881 * Lock description contains policy of blocking lock,
1882 * and its cancel_bits is used to pass conflicting bits.
1883 * NOTE: ld can be NULL or can be not NULL but zeroed if
1884 * passed from ldlm_bl_thread_blwi(), check below used bits
1885 * in ld to make sure it is valid description.
1887 * If server may replace lock resource keeping the same cookie,
1888 * never use cancel bits from different resource, full cancel
1891 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1892 ldlm_res_eq(&ld->l_resource.lr_name,
1893 &lock->l_resource->lr_name) &&
1894 !(ldlm_is_cbpending(lock) &&
1895 lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1896 /* always combine conflicting ibits */
1897 lock->l_policy_data.l_inodebits.cancel_bits |=
1898 ld->l_policy_data.l_inodebits.cancel_bits;
1900 /* If cancel_bits are not obtained or
1901 * if the lock is already CBPENDING and
1902 * has no cancel_bits set
1903 * - the full lock is to be cancelled
1905 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1911 * Callback handler for receiving incoming blocking ASTs.
1913 * This can only happen on client side.
1915 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1916 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1922 LDLM_DEBUG(lock, "client blocking AST callback handler");
1924 lock_res_and_lock(lock);
1926 /* get extra information from desc if any */
1927 ldlm_bl_desc2lock(ld, lock);
1928 ldlm_set_cbpending(lock);
1930 do_ast = (!lock->l_readers && !lock->l_writers);
1931 unlock_res_and_lock(lock);
1935 "Lock %p already unused, calling callback (%p)\n",
1936 lock, lock->l_blocking_ast);
1937 if (lock->l_blocking_ast != NULL)
1938 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1942 "Lock %p is referenced, will be cancelled later\n",
1946 LDLM_DEBUG(lock, "client blocking callback handler END");
1947 LDLM_LOCK_RELEASE(lock);
1951 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1953 if (req->rq_no_reply)
1956 req->rq_status = rc;
1957 if (!req->rq_packed_final) {
1958 rc = lustre_pack_reply(req, 1, NULL, NULL);
1962 return ptlrpc_reply(req);
1966 * Callback handler for receiving incoming completion ASTs.
1968 * This only can happen on client side.
1970 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1971 struct ldlm_namespace *ns,
1972 struct ldlm_request *dlm_req,
1973 struct ldlm_lock *lock)
1975 LIST_HEAD(ast_list);
1981 LDLM_DEBUG(lock, "client completion callback handler START");
1983 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1984 long to = cfs_time_seconds(1);
1986 ldlm_callback_reply(req, 0);
1989 to = schedule_timeout_interruptible(to);
1990 if (ldlm_is_granted(lock) ||
1991 ldlm_is_destroyed(lock))
1996 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1998 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1999 GOTO(out, rc = lvb_len);
2000 } else if (lvb_len > 0) {
2001 if (lock->l_lvb_len > 0) {
2002 /* for extent lock, lvb contains ost_lvb{}. */
2003 LASSERT(lock->l_lvb_data != NULL);
2005 if (unlikely(lock->l_lvb_len < lvb_len)) {
2007 "Replied LVB is larger than expectation, expected = %d, replied = %d",
2008 lock->l_lvb_len, lvb_len);
2009 GOTO(out, rc = -EINVAL);
2014 lock_res_and_lock(lock);
2016 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
2017 &lock->l_resource->lr_name)) {
2018 ldlm_resource_unlink_lock(lock);
2019 unlock_res_and_lock(lock);
2020 rc = ldlm_lock_change_resource(ns, lock,
2021 &dlm_req->lock_desc.l_resource.lr_name);
2023 LDLM_ERROR(lock, "Failed to allocate resource");
2026 LDLM_DEBUG(lock, "completion AST, new resource");
2027 lock_res_and_lock(lock);
2030 if (ldlm_is_failed(lock)) {
2031 unlock_res_and_lock(lock);
2032 LDLM_LOCK_RELEASE(lock);
2036 if (ldlm_is_destroyed(lock) ||
2037 ldlm_is_granted(lock)) {
2038 /* b=11300: the lock has already been granted */
2039 unlock_res_and_lock(lock);
2040 LDLM_DEBUG(lock, "Double grant race happened");
2045 * If we receive the completion AST before the actual enqueue returned,
2046 * then we might need to switch lock modes, resources, or extents.
2048 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
2049 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
2050 LDLM_DEBUG(lock, "completion AST, new lock mode");
2053 if (lock->l_resource->lr_type != LDLM_PLAIN) {
2054 ldlm_convert_policy_to_local(req->rq_export,
2055 dlm_req->lock_desc.l_resource.lr_type,
2056 &dlm_req->lock_desc.l_policy_data,
2057 &lock->l_policy_data);
2058 LDLM_DEBUG(lock, "completion AST, new policy data");
2061 ldlm_resource_unlink_lock(lock);
2063 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2065 * BL_AST locks are not needed in LRU.
2066 * Let ldlm_cancel_lru() be fast.
2068 ldlm_lock_remove_from_lru(lock);
2069 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2070 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2071 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2074 if (lock->l_lvb_len > 0) {
2075 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2076 lock->l_lvb_data, lvb_len);
2078 unlock_res_and_lock(lock);
2083 ldlm_grant_lock(lock, &ast_list);
2084 unlock_res_and_lock(lock);
2086 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2089 * Let Enqueue to call osc_lock_upcall() and initialize
2092 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2094 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2096 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2102 lock_res_and_lock(lock);
2103 ldlm_set_failed(lock);
2104 unlock_res_and_lock(lock);
2105 wake_up(&lock->l_waitq);
2107 LDLM_LOCK_RELEASE(lock);
2113 * Callback handler for receiving incoming glimpse ASTs.
2115 * This only can happen on client side. After handling the glimpse AST
2116 * we also consider dropping the lock here if it is unused locally for a
2119 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2120 struct ldlm_namespace *ns,
2121 struct ldlm_request *dlm_req,
2122 struct ldlm_lock *lock)
2124 struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2129 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2131 if (lock->l_glimpse_ast != NULL)
2132 rc = lock->l_glimpse_ast(lock, req);
2134 if (req->rq_repmsg != NULL) {
2137 req->rq_status = rc;
2141 lock_res_and_lock(lock);
2142 if (lock->l_granted_mode == LCK_PW &&
2143 !lock->l_readers && !lock->l_writers &&
2144 ktime_after(ktime_get(),
2145 ktime_add(lock->l_last_used, ns->ns_dirty_age_limit))) {
2146 unlock_res_and_lock(lock);
2148 /* For MDS glimpse it is always DOM lock, set corresponding
2149 * cancel_bits to perform lock convert if needed
2151 if (lock->l_resource->lr_type == LDLM_IBITS)
2152 ld->l_policy_data.l_inodebits.cancel_bits =
2154 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2155 ldlm_handle_bl_callback(ns, ld, lock);
2160 unlock_res_and_lock(lock);
2161 LDLM_LOCK_RELEASE(lock);
2165 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2166 enum ldlm_cancel_flags cancel_flags)
2168 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2169 char *prio = "regular";
2174 spin_lock(&blp->blp_lock);
2175 /* cannot access blwi after added to list and lock is dropped */
2176 count = blwi->blwi_lock ? 1 : blwi->blwi_count;
2178 /* if the server is waiting on a lock to be cancelled (bl_ast), this is
2179 * an urgent request and should go in the priority queue so it doesn't
2180 * get stuck behind non-priority work (eg, lru size management)
2182 * We also prioritize discard_data, which is for eviction handling
2184 if (blwi->blwi_lock &&
2185 (ldlm_is_discard_data(blwi->blwi_lock) ||
2186 ldlm_is_bl_ast(blwi->blwi_lock))) {
2187 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2190 /* other blocking callbacks are added to the regular list */
2191 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2193 blp->blp_total_locks += count;
2194 blp->blp_total_blwis++;
2195 spin_unlock(&blp->blp_lock);
2197 wake_up(&blp->blp_waitq);
2199 /* unlocked read of blp values is intentional - OK for debug */
2201 "added %d/%d locks to %s blp list, %d blwis in pool\n",
2202 count, blp->blp_total_locks, prio, blp->blp_total_blwis);
2205 * can not check blwi->blwi_flags as blwi could be already freed in
2208 if (!(cancel_flags & LCF_ASYNC))
2209 wait_for_completion(&blwi->blwi_comp);
2214 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2215 struct ldlm_namespace *ns,
2216 struct ldlm_lock_desc *ld,
2217 struct list_head *cancels, int count,
2218 struct ldlm_lock *lock,
2219 enum ldlm_cancel_flags cancel_flags)
2221 init_completion(&blwi->blwi_comp);
2222 INIT_LIST_HEAD(&blwi->blwi_head);
2224 if (current->flags & PF_MEMALLOC)
2225 blwi->blwi_mem_pressure = 1;
2228 blwi->blwi_flags = cancel_flags;
2230 blwi->blwi_ld = *ld;
2232 list_splice_init(cancels, &blwi->blwi_head);
2233 blwi->blwi_count = count;
2235 blwi->blwi_lock = lock;
2240 * Queues a list of locks \a cancels containing \a count locks
2241 * for later processing by a blocking thread. If \a count is zero,
2242 * then the lock referenced as \a lock is queued instead.
2244 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2245 * If list addition fails an error is returned and caller is supposed to
2246 * call ->l_blocking_ast itself.
2248 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2249 struct ldlm_lock_desc *ld,
2250 struct ldlm_lock *lock,
2251 struct list_head *cancels, int count,
2252 enum ldlm_cancel_flags cancel_flags)
2256 if (cancels && count == 0)
2259 if (cancel_flags & LCF_ASYNC) {
2260 struct ldlm_bl_work_item *blwi;
2262 OBD_ALLOC(blwi, sizeof(*blwi));
2265 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2267 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2270 * if it is synchronous call do minimum mem alloc, as it could
2271 * be triggered from kernel shrinker
2273 struct ldlm_bl_work_item blwi;
2275 memset(&blwi, 0, sizeof(blwi));
2276 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2277 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2282 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2283 struct ldlm_lock *lock)
2285 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2288 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2289 struct list_head *cancels, int count,
2290 enum ldlm_cancel_flags cancel_flags)
2292 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2295 int ldlm_bl_to_thread_ns(struct ldlm_namespace *ns)
2297 return ldlm_bl_to_thread(ns, NULL, NULL, NULL, 0, LCF_ASYNC);
2300 int ldlm_bl_thread_wakeup(void)
2302 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2306 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2307 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2309 struct obd_device *obd = req->rq_export->exp_obd;
2317 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2319 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2321 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2323 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2326 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2328 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2330 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2333 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2336 /* We are responsible for swabbing contents of val */
2338 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2339 /* Pass it on to mdc (the "export" in this case) */
2340 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2342 sizeof(KEY_HSM_COPYTOOL_SEND),
2343 KEY_HSM_COPYTOOL_SEND,
2346 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2351 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2352 const char *msg, int rc,
2353 const struct lustre_handle *handle)
2355 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2356 "%s, NID=%s lock=%#llx: rc = %d",
2357 msg, libcfs_idstr(&req->rq_peer),
2358 handle ? handle->cookie : 0, rc);
2359 if (req->rq_no_reply)
2360 CWARN("No reply was sent, maybe cause b=21636.\n");
2362 CWARN("Send reply failed, maybe cause b=21636.\n");
2365 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2366 static int ldlm_callback_handler(struct ptlrpc_request *req)
2368 struct ldlm_namespace *ns;
2369 struct ldlm_request *dlm_req;
2370 struct ldlm_lock *lock;
2376 * Requests arrive in sender's byte order. The ptlrpc service
2377 * handler has already checked and, if necessary, byte-swapped the
2378 * incoming request message body, but I am responsible for the
2382 /* do nothing for sec context finalize */
2383 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2386 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2388 if (req->rq_export == NULL) {
2389 rc = ldlm_callback_reply(req, -ENOTCONN);
2390 ldlm_callback_errmsg(req, "Operate on unconnected server",
2395 LASSERT(req->rq_export != NULL);
2396 LASSERT(req->rq_export->exp_obd != NULL);
2398 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2399 case LDLM_BL_CALLBACK:
2400 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2402 ldlm_callback_reply(req, -(int)cfs_fail_err);
2406 case LDLM_CP_CALLBACK:
2407 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2410 case LDLM_GL_CALLBACK:
2411 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2415 rc = ldlm_handle_setinfo(req);
2416 ldlm_callback_reply(req, rc);
2419 CERROR("unknown opcode %u\n",
2420 lustre_msg_get_opc(req->rq_reqmsg));
2421 ldlm_callback_reply(req, -EPROTO);
2425 ns = req->rq_export->exp_obd->obd_namespace;
2426 LASSERT(ns != NULL);
2428 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2430 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2431 if (dlm_req == NULL) {
2432 rc = ldlm_callback_reply(req, -EPROTO);
2433 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2439 * Force a known safe race, send a cancel to the server for a lock
2440 * which the server has already started a blocking callback on.
2442 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2443 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2444 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2446 CERROR("ldlm_cli_cancel: %d\n", rc);
2449 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2452 "callback on lock %#llx - lock disappeared\n",
2453 dlm_req->lock_handle[0].cookie);
2454 rc = ldlm_callback_reply(req, -EINVAL);
2455 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2456 &dlm_req->lock_handle[0]);
2460 if (ldlm_is_fail_loc(lock) &&
2461 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2462 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2464 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2465 lock_res_and_lock(lock);
2466 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2468 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2470 * If somebody cancels lock and cache is already dropped,
2471 * or lock is failed before cp_ast received on client,
2472 * we can tell the server we have no lock. Otherwise, we
2473 * should send cancel after dropping the cache.
2475 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2476 ldlm_is_failed(lock)) {
2478 "callback on lock %llx - lock disappeared",
2479 dlm_req->lock_handle[0].cookie);
2480 unlock_res_and_lock(lock);
2481 LDLM_LOCK_RELEASE(lock);
2482 rc = ldlm_callback_reply(req, -EINVAL);
2483 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2484 &dlm_req->lock_handle[0]);
2488 * BL_AST locks are not needed in LRU.
2489 * Let ldlm_cancel_lru() be fast.
2491 ldlm_lock_remove_from_lru(lock);
2492 ldlm_set_bl_ast(lock);
2494 if (lock->l_remote_handle.cookie == 0)
2495 lock->l_remote_handle = dlm_req->lock_handle[1];
2496 unlock_res_and_lock(lock);
2499 * We want the ost thread to get this reply so that it can respond
2500 * to ost requests (write cache writeback) that might be triggered
2503 * But we'd also like to be able to indicate in the reply that we're
2504 * cancelling right now, because it's unused, or have an intent result
2505 * in the reply, so we might have to push the responsibility for sending
2506 * the reply down into the AST handlers, alas.
2509 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2510 case LDLM_BL_CALLBACK:
2511 CDEBUG(D_INODE, "blocking ast\n");
2512 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2513 if (!ldlm_is_cancel_on_block(lock)) {
2514 rc = ldlm_callback_reply(req, 0);
2515 if (req->rq_no_reply || rc)
2516 ldlm_callback_errmsg(req, "Normal process", rc,
2517 &dlm_req->lock_handle[0]);
2519 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2520 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2522 case LDLM_CP_CALLBACK:
2523 CDEBUG(D_INODE, "completion ast\n");
2524 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2525 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2526 if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2527 ldlm_callback_reply(req, rc);
2529 case LDLM_GL_CALLBACK:
2530 CDEBUG(D_INODE, "glimpse ast\n");
2531 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2532 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2535 LBUG(); /* checked above */
2541 #ifdef HAVE_SERVER_SUPPORT
2543 * Main handler for canceld thread.
2545 * Separated into its own thread to avoid deadlocks.
2547 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2554 * Requests arrive in sender's byte order. The ptlrpc service
2555 * handler has already checked and, if necessary, byte-swapped the
2556 * incoming request message body, but I am responsible for the
2560 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2562 if (req->rq_export == NULL) {
2563 struct ldlm_request *dlm_req;
2565 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2566 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2567 libcfs_nidstr(&req->rq_peer.nid),
2568 (unsigned long long)req->rq_arrival_time.tv_sec,
2569 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2571 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2572 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2573 dlm_req = req_capsule_client_get(&req->rq_pill,
2575 if (dlm_req != NULL)
2576 ldlm_lock_dump_handle(D_ERROR,
2577 &dlm_req->lock_handle[0]);
2579 ldlm_callback_reply(req, -ENOTCONN);
2583 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2584 /* XXX FIXME move this back to mds/handler.c, b=249 */
2586 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2587 CDEBUG(D_INODE, "cancel\n");
2588 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2589 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2590 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2592 rc = ldlm_handle_cancel(req);
2596 struct ldlm_request *dlm_req;
2598 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2599 CDEBUG(D_INODE, "convert\n");
2601 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2602 if (dlm_req == NULL) {
2603 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2604 rc = ldlm_callback_reply(req, -EPROTO);
2606 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2607 rc = ptlrpc_reply(req);
2612 CERROR("invalid opcode %d\n",
2613 lustre_msg_get_opc(req->rq_reqmsg));
2614 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2615 rc = ldlm_callback_reply(req, -EINVAL);
2621 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2622 struct ldlm_lock *lock)
2624 struct ldlm_request *dlm_req;
2625 struct lustre_handle lockh;
2631 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2632 if (dlm_req == NULL)
2635 ldlm_lock2handle(lock, &lockh);
2636 for (i = 0; i < dlm_req->lock_count; i++) {
2637 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2639 DEBUG_REQ(D_RPCTRACE, req,
2640 "Prio raised by lock %#llx", lockh.cookie);
2649 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2651 struct ldlm_request *dlm_req;
2658 /* no prolong in recovery */
2659 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2662 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2663 if (dlm_req == NULL)
2666 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2667 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2668 (size - offsetof(struct ldlm_request, lock_handle)) /
2669 sizeof(struct lustre_handle) < dlm_req->lock_count)
2672 for (i = 0; i < dlm_req->lock_count; i++) {
2673 struct ldlm_lock *lock;
2675 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2679 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2681 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2682 LDLM_LOCK_PUT(lock);
2691 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2692 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2693 .hpreq_check = ldlm_cancel_hpreq_check,
2697 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2701 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2703 if (req->rq_export == NULL)
2706 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2707 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2708 req->rq_ops = &ldlm_cancel_hpreq_ops;
2709 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2710 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2711 req->rq_ops = &ldlm_cancel_hpreq_ops;
2716 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2717 struct hlist_node *hnode, void *data)
2720 struct list_head *rpc_list = data;
2721 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2723 lock_res_and_lock(lock);
2725 if (!ldlm_is_granted(lock)) {
2726 unlock_res_and_lock(lock);
2730 LASSERT(lock->l_resource);
2731 if (lock->l_resource->lr_type != LDLM_IBITS &&
2732 lock->l_resource->lr_type != LDLM_PLAIN) {
2733 unlock_res_and_lock(lock);
2737 if (ldlm_is_ast_sent(lock)) {
2738 unlock_res_and_lock(lock);
2742 LASSERT(lock->l_blocking_ast);
2743 LASSERT(!lock->l_blocking_lock);
2745 ldlm_set_ast_sent(lock);
2746 if (lock->l_export && lock->l_export->exp_lock_hash) {
2748 * NB: it's safe to call cfs_hash_del() even lock isn't
2752 * In the function below, .hs_keycmp resolves to
2753 * ldlm_export_lock_keycmp()
2755 /* coverity[overrun-buffer-val] */
2756 cfs_hash_del(lock->l_export->exp_lock_hash,
2757 &lock->l_remote_handle, &lock->l_exp_hash);
2760 list_add_tail(&lock->l_rk_ast, rpc_list);
2761 LDLM_LOCK_GET(lock);
2763 unlock_res_and_lock(lock);
2767 void ldlm_revoke_export_locks(struct obd_export *exp)
2770 LIST_HEAD(rpc_list);
2773 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2774 ldlm_revoke_lock_cb, &rpc_list, 0);
2775 rc = ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2776 LDLM_WORK_REVOKE_AST);
2778 if (rc == -ERESTART)
2779 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2783 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2784 #endif /* HAVE_SERVER_SUPPORT */
2786 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2787 struct ldlm_bl_work_item **p_blwi,
2788 struct obd_export **p_exp)
2790 struct ldlm_bl_work_item *blwi = NULL;
2791 static unsigned int num_bl;
2792 static unsigned int num_stale;
2793 int num_th = atomic_read(&blp->blp_num_threads);
2795 *p_exp = obd_stale_export_get();
2797 spin_lock(&blp->blp_lock);
2798 if (*p_exp != NULL) {
2799 if (num_th == 1 || ++num_stale < num_th) {
2800 spin_unlock(&blp->blp_lock);
2806 /* process a request from the blp_list at least every blp_num_threads */
2807 if (!list_empty(&blp->blp_list) &&
2808 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2809 blwi = list_entry(blp->blp_list.next,
2810 struct ldlm_bl_work_item, blwi_entry);
2812 if (!list_empty(&blp->blp_prio_list))
2813 blwi = list_entry(blp->blp_prio_list.next,
2814 struct ldlm_bl_work_item,
2818 if (++num_bl >= num_th)
2820 list_del(&blwi->blwi_entry);
2821 blp->blp_total_locks -= blwi->blwi_lock ? 1 : blwi->blwi_count;
2822 blp->blp_total_blwis--;
2824 spin_unlock(&blp->blp_lock);
2827 /* intentional unlocked read of blp values - OK for debug */
2830 "Got %d locks of %d total in blp. (%d blwis in pool)\n",
2831 blwi->blwi_lock ? 1 : blwi->blwi_count,
2832 blp->blp_total_locks, blp->blp_total_blwis);
2835 "No blwi found in queue (no bl locks in queue)\n");
2838 if (*p_exp != NULL && *p_blwi != NULL) {
2839 obd_stale_export_put(*p_exp);
2843 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2846 /* This only contains temporary data until the thread starts */
2847 struct ldlm_bl_thread_data {
2848 struct ldlm_bl_pool *bltd_blp;
2849 struct completion bltd_comp;
2853 static int ldlm_bl_thread_main(void *arg);
2855 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2857 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2858 struct task_struct *task;
2860 init_completion(&bltd.bltd_comp);
2862 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2863 if (bltd.bltd_num >= blp->blp_max_threads) {
2864 atomic_dec(&blp->blp_num_threads);
2868 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2870 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2871 atomic_dec(&blp->blp_num_threads);
2875 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2878 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2879 bltd.bltd_num, PTR_ERR(task));
2880 atomic_dec(&blp->blp_num_threads);
2881 return PTR_ERR(task);
2883 wait_for_completion(&bltd.bltd_comp);
2888 /* Not fatal if racy and have a few too many threads */
2889 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2890 struct ldlm_bl_work_item *blwi)
2892 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2895 if (atomic_read(&blp->blp_busy_threads) <
2896 atomic_read(&blp->blp_num_threads))
2899 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2900 blwi->blwi_mem_pressure))
2906 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2907 struct ldlm_bl_work_item *blwi)
2909 /* '1' for consistency with code that checks !mpflag to restore */
2910 unsigned int mpflags = 1;
2914 if (blwi->blwi_ns == NULL)
2915 /* added by ldlm_cleanup() */
2916 RETURN(LDLM_ITER_STOP);
2918 if (blwi->blwi_mem_pressure)
2919 mpflags = memalloc_noreclaim_save();
2921 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2923 if (blwi->blwi_count) {
2926 * The special case when we cancel locks in lru
2927 * asynchronously, we pass the list of locks here.
2928 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2929 * canceled locally yet.
2931 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2934 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2936 } else if (blwi->blwi_lock) {
2937 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2940 ldlm_pool_recalc(&blwi->blwi_ns->ns_pool, true);
2941 spin_lock(&blwi->blwi_ns->ns_lock);
2942 blwi->blwi_ns->ns_rpc_recalc = 0;
2943 spin_unlock(&blwi->blwi_ns->ns_lock);
2944 ldlm_namespace_put(blwi->blwi_ns);
2947 if (blwi->blwi_mem_pressure)
2948 memalloc_noreclaim_restore(mpflags);
2950 if (blwi->blwi_flags & LCF_ASYNC)
2951 OBD_FREE(blwi, sizeof(*blwi));
2953 complete(&blwi->blwi_comp);
2959 * Cancel stale locks on export. Cancel blocked locks first.
2960 * If the given export has blocked locks, the next in the list may have
2961 * them too, thus cancel not blocked locks only if the current export has
2964 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2965 struct obd_export *exp)
2971 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2973 num = ldlm_export_cancel_blocked_locks(exp);
2975 ldlm_export_cancel_locks(exp);
2977 obd_stale_export_put(exp);
2984 * Main blocking requests processing thread.
2986 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2987 * This thread in the end ends up doing actual call to ->l_blocking_ast
2990 static int ldlm_bl_thread_main(void *arg)
2993 struct ldlm_bl_pool *blp;
2994 struct ldlm_bl_thread_data *bltd = arg;
3002 rc = lu_env_init(env, LCT_DT_THREAD);
3005 rc = lu_env_add(env);
3007 GOTO(out_env_fini, rc);
3009 blp = bltd->bltd_blp;
3011 complete(&bltd->bltd_comp);
3012 /* cannot use bltd after this, it is only on caller's stack */
3015 struct ldlm_bl_work_item *blwi = NULL;
3016 struct obd_export *exp = NULL;
3019 rc = ldlm_bl_get_work(blp, &blwi, &exp);
3022 wait_event_idle_exclusive(blp->blp_waitq,
3023 ldlm_bl_get_work(blp, &blwi,
3025 atomic_inc(&blp->blp_busy_threads);
3027 if (ldlm_bl_thread_need_create(blp, blwi))
3028 /* discard the return value, we tried */
3029 ldlm_bl_thread_start(blp, true);
3032 rc = ldlm_bl_thread_exports(blp, exp);
3034 rc = ldlm_bl_thread_blwi(blp, blwi);
3036 atomic_dec(&blp->blp_busy_threads);
3038 if (rc == LDLM_ITER_STOP)
3042 * If there are many namespaces, we will not sleep waiting for
3043 * work, and must do a cond_resched to avoid holding the CPU
3049 atomic_dec(&blp->blp_num_threads);
3050 complete(&blp->blp_comp);
3061 static int ldlm_setup(void);
3062 static int ldlm_cleanup(void);
3064 int ldlm_get_ref(void)
3069 mutex_lock(&ldlm_ref_mutex);
3070 if (++ldlm_refcount == 1) {
3075 mutex_unlock(&ldlm_ref_mutex);
3080 void ldlm_put_ref(void)
3083 mutex_lock(&ldlm_ref_mutex);
3084 if (ldlm_refcount == 1) {
3085 int rc = ldlm_cleanup();
3088 CERROR("ldlm_cleanup failed: %d\n", rc);
3094 mutex_unlock(&ldlm_ref_mutex);
3100 * Export handle<->lock hash operations.
3103 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned int mask)
3105 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
3109 ldlm_export_lock_key(struct hlist_node *hnode)
3111 struct ldlm_lock *lock;
3113 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3114 return &lock->l_remote_handle;
3118 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3120 struct ldlm_lock *lock;
3122 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3123 lock->l_remote_handle = *(struct lustre_handle *)key;
3127 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3129 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3133 ldlm_export_lock_object(struct hlist_node *hnode)
3135 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3139 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3141 struct ldlm_lock *lock;
3143 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3144 LDLM_LOCK_GET(lock);
3148 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3150 struct ldlm_lock *lock;
3152 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3153 LDLM_LOCK_RELEASE(lock);
3156 static struct cfs_hash_ops ldlm_export_lock_ops = {
3157 .hs_hash = ldlm_export_lock_hash,
3158 .hs_key = ldlm_export_lock_key,
3159 .hs_keycmp = ldlm_export_lock_keycmp,
3160 .hs_keycpy = ldlm_export_lock_keycpy,
3161 .hs_object = ldlm_export_lock_object,
3162 .hs_get = ldlm_export_lock_get,
3163 .hs_put = ldlm_export_lock_put,
3164 .hs_put_locked = ldlm_export_lock_put,
3167 int ldlm_init_export(struct obd_export *exp)
3173 exp->exp_lock_hash =
3174 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3175 HASH_EXP_LOCK_CUR_BITS,
3176 HASH_EXP_LOCK_MAX_BITS,
3177 HASH_EXP_LOCK_BKT_BITS, 0,
3178 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3179 &ldlm_export_lock_ops,
3180 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3181 CFS_HASH_NBLK_CHANGE);
3183 if (!exp->exp_lock_hash)
3186 rc = ldlm_init_flock_export(exp);
3192 ldlm_destroy_export(exp);
3195 EXPORT_SYMBOL(ldlm_init_export);
3197 void ldlm_destroy_export(struct obd_export *exp)
3200 cfs_hash_putref(exp->exp_lock_hash);
3201 exp->exp_lock_hash = NULL;
3203 ldlm_destroy_flock_export(exp);
3206 EXPORT_SYMBOL(ldlm_destroy_export);
3208 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3209 struct attribute *attr,
3212 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3215 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3216 struct attribute *attr,
3223 rc = kstrtoul(buffer, 10, &val);
3227 ldlm_cancel_unused_locks_before_replay = val;
3231 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3233 static struct attribute *ldlm_attrs[] = {
3234 &lustre_attr_cancel_unused_locks_before_replay.attr,
3238 static struct attribute_group ldlm_attr_group = {
3239 .attrs = ldlm_attrs,
3242 static int ldlm_setup(void)
3244 static struct ptlrpc_service_conf conf;
3245 struct ldlm_bl_pool *blp = NULL;
3246 #ifdef HAVE_SERVER_SUPPORT
3247 struct task_struct *task;
3248 #endif /* HAVE_SERVER_SUPPORT */
3254 if (ldlm_state != NULL)
3257 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3258 if (ldlm_state == NULL)
3261 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3265 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3269 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3273 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3277 rc = ldlm_debugfs_setup();
3281 memset(&conf, 0, sizeof(conf));
3282 conf = (typeof(conf)) {
3283 .psc_name = "ldlm_cbd",
3284 .psc_watchdog_factor = 2,
3286 .bc_nbufs = LDLM_CLIENT_NBUFS,
3287 .bc_buf_size = LDLM_BUFSIZE,
3288 .bc_req_max_size = LDLM_MAXREQSIZE,
3289 .bc_rep_max_size = LDLM_MAXREPSIZE,
3290 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3291 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3294 .tc_thr_name = "ldlm_cb",
3295 .tc_thr_factor = LDLM_THR_FACTOR,
3296 .tc_nthrs_init = LDLM_NTHRS_INIT,
3297 .tc_nthrs_base = LDLM_NTHRS_BASE,
3298 .tc_nthrs_max = LDLM_NTHRS_MAX,
3299 .tc_nthrs_user = ldlm_num_threads,
3300 .tc_cpu_bind = ldlm_cpu_bind,
3301 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3304 .cc_pattern = ldlm_cpts,
3305 .cc_affinity = true,
3308 .so_req_handler = ldlm_callback_handler,
3311 ldlm_state->ldlm_cb_service = \
3312 ptlrpc_register_service(&conf, ldlm_svc_kset,
3313 ldlm_svc_debugfs_dir);
3314 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3315 CERROR("failed to start service\n");
3316 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3317 ldlm_state->ldlm_cb_service = NULL;
3321 #ifdef HAVE_SERVER_SUPPORT
3322 memset(&conf, 0, sizeof(conf));
3323 conf = (typeof(conf)) {
3324 .psc_name = "ldlm_canceld",
3325 .psc_watchdog_factor = 6,
3327 .bc_nbufs = LDLM_SERVER_NBUFS,
3328 .bc_buf_size = LDLM_BUFSIZE,
3329 .bc_req_max_size = LDLM_MAXREQSIZE,
3330 .bc_rep_max_size = LDLM_MAXREPSIZE,
3331 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3332 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3336 .tc_thr_name = "ldlm_cn",
3337 .tc_thr_factor = LDLM_THR_FACTOR,
3338 .tc_nthrs_init = LDLM_NTHRS_INIT,
3339 .tc_nthrs_base = LDLM_NTHRS_BASE,
3340 .tc_nthrs_max = LDLM_NTHRS_MAX,
3341 .tc_nthrs_user = ldlm_num_threads,
3342 .tc_cpu_bind = ldlm_cpu_bind,
3343 .tc_ctx_tags = LCT_MD_THREAD | \
3348 .cc_pattern = ldlm_cpts,
3349 .cc_affinity = true,
3352 .so_req_handler = ldlm_cancel_handler,
3353 .so_hpreq_handler = ldlm_hpreq_handler,
3356 ldlm_state->ldlm_cancel_service = \
3357 ptlrpc_register_service(&conf, ldlm_svc_kset,
3358 ldlm_svc_debugfs_dir);
3359 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3360 CERROR("failed to start service\n");
3361 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3362 ldlm_state->ldlm_cancel_service = NULL;
3365 #endif /* HAVE_SERVER_SUPPORT */
3367 OBD_ALLOC(blp, sizeof(*blp));
3369 GOTO(out, rc = -ENOMEM);
3370 ldlm_state->ldlm_bl_pool = blp;
3372 spin_lock_init(&blp->blp_lock);
3373 INIT_LIST_HEAD(&blp->blp_list);
3374 INIT_LIST_HEAD(&blp->blp_prio_list);
3375 init_waitqueue_head(&blp->blp_waitq);
3376 atomic_set(&blp->blp_num_threads, 0);
3377 atomic_set(&blp->blp_busy_threads, 0);
3378 blp->blp_total_locks = 0;
3379 blp->blp_total_blwis = 0;
3381 if (ldlm_num_threads == 0) {
3382 blp->blp_min_threads = LDLM_NTHRS_INIT;
3383 blp->blp_max_threads = LDLM_NTHRS_MAX;
3385 blp->blp_min_threads = blp->blp_max_threads = \
3386 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3390 for (i = 0; i < blp->blp_min_threads; i++) {
3391 rc = ldlm_bl_thread_start(blp, false);
3396 #ifdef HAVE_SERVER_SUPPORT
3397 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3400 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3404 wait_event(expired_lock_wait_queue,
3405 expired_lock_thread_state == ELT_READY);
3406 #endif /* HAVE_SERVER_SUPPORT */
3408 rc = ldlm_pools_init();
3410 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3414 rc = ldlm_reclaim_setup();
3416 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3426 static int ldlm_cleanup(void)
3430 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3431 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3432 CERROR("ldlm still has namespaces; clean these up first.\n");
3433 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3434 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3438 ldlm_reclaim_cleanup();
3441 if (ldlm_state->ldlm_bl_pool != NULL) {
3442 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3444 while (atomic_read(&blp->blp_num_threads) > 0) {
3445 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3447 init_completion(&blp->blp_comp);
3449 spin_lock(&blp->blp_lock);
3450 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3451 wake_up(&blp->blp_waitq);
3452 spin_unlock(&blp->blp_lock);
3454 wait_for_completion(&blp->blp_comp);
3457 OBD_FREE(blp, sizeof(*blp));
3460 if (ldlm_state->ldlm_cb_service != NULL)
3461 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3462 #ifdef HAVE_SERVER_SUPPORT
3463 if (ldlm_state->ldlm_cancel_service != NULL)
3464 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3468 kset_unregister(ldlm_ns_kset);
3470 kset_unregister(ldlm_svc_kset);
3472 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3473 kobject_put(ldlm_kobj);
3476 ldlm_debugfs_cleanup();
3478 #ifdef HAVE_SERVER_SUPPORT
3479 if (expired_lock_thread_state != ELT_STOPPED) {
3480 expired_lock_thread_state = ELT_TERMINATE;
3481 wake_up(&expired_lock_wait_queue);
3482 wait_event(expired_lock_wait_queue,
3483 expired_lock_thread_state == ELT_STOPPED);
3487 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3495 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3496 sizeof(struct ldlm_resource), 0,
3497 SLAB_HWCACHE_ALIGN, NULL);
3498 if (ldlm_resource_slab == NULL)
3501 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3502 sizeof(struct ldlm_lock), 0,
3503 SLAB_HWCACHE_ALIGN, NULL);
3504 if (ldlm_lock_slab == NULL)
3507 ldlm_interval_slab = kmem_cache_create("interval_node",
3508 sizeof(struct ldlm_interval),
3509 0, SLAB_HWCACHE_ALIGN, NULL);
3510 if (ldlm_interval_slab == NULL)
3513 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3514 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3515 0, SLAB_HWCACHE_ALIGN, NULL);
3516 if (ldlm_interval_tree_slab == NULL)
3519 #ifdef HAVE_SERVER_SUPPORT
3520 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3521 sizeof(struct ldlm_ibits_node),
3522 0, SLAB_HWCACHE_ALIGN, NULL);
3523 if (ldlm_inodebits_slab == NULL)
3524 goto out_interval_tree;
3526 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3527 sizeof(struct ldlm_glimpse_work),
3529 if (ldlm_glimpse_work_kmem == NULL)
3533 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3534 class_export_dump_hook = ldlm_dump_export_locks;
3537 #ifdef HAVE_SERVER_SUPPORT
3539 kmem_cache_destroy(ldlm_inodebits_slab);
3541 kmem_cache_destroy(ldlm_interval_tree_slab);
3544 kmem_cache_destroy(ldlm_interval_slab);
3546 kmem_cache_destroy(ldlm_lock_slab);
3548 kmem_cache_destroy(ldlm_resource_slab);
3553 void ldlm_exit(void)
3556 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3558 kmem_cache_destroy(ldlm_resource_slab);
3560 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3561 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3562 * so that ldlm_lock_free() get a chance to be called.
3565 kmem_cache_destroy(ldlm_lock_slab);
3566 kmem_cache_destroy(ldlm_interval_slab);
3567 kmem_cache_destroy(ldlm_interval_tree_slab);
3568 #ifdef HAVE_SERVER_SUPPORT
3569 kmem_cache_destroy(ldlm_inodebits_slab);
3570 kmem_cache_destroy(ldlm_glimpse_work_kmem);