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_first_entry(expired, 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 (!CFS_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_first_entry(&waiting_locks_list, 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_first_entry(&waiting_locks_list, 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 (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
390 CFS_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)
441 struct obd_device *obd = NULL;
444 /* NB: must be called with hold of lock_res_and_lock() */
445 LASSERT(ldlm_is_res_locked(lock));
446 LASSERT(!ldlm_is_cancel_on_block(lock));
449 * Do not put cross-MDT lock in the waiting list, since we
450 * will not evict it due to timeout for now
452 if (lock->l_export != NULL) {
453 obd = lock->l_export->exp_obd;
455 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS)
459 spin_lock_bh(&waiting_locks_spinlock);
460 if (ldlm_is_cancel(lock)) {
461 spin_unlock_bh(&waiting_locks_spinlock);
465 if (ldlm_is_destroyed(lock)) {
466 static time64_t next;
468 spin_unlock_bh(&waiting_locks_spinlock);
469 LDLM_ERROR(lock, "not waiting on destroyed lock (b=5653)");
470 if (ktime_get_seconds() > next) {
471 next = ktime_get_seconds() + 14400;
477 ldlm_set_waited(lock);
478 ret = __ldlm_add_waiting_lock(lock, timeout);
481 * grab ref on the lock if it has been added to the
486 spin_unlock_bh(&waiting_locks_spinlock);
489 ldlm_add_blocked_lock(lock);
491 at_off = obd_at_off(obd);
492 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
493 ret == 0 ? "not re-" : "", timeout,
494 at_off ? "off" : "on");
499 * Remove a lock from the pending list, likely because it had its cancellation
500 * callback arrive without incident. This adjusts the lock-timeout timer if
501 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
502 * As done by ldlm_del_waiting_lock(), the caller must release the lock
503 * reference when the lock is removed from any list (1 is returned).
505 * Called with namespace lock held.
507 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
509 struct list_head *list_next;
511 if (list_empty(&lock->l_pending_chain))
514 list_next = lock->l_pending_chain.next;
515 if (lock->l_pending_chain.prev == &waiting_locks_list) {
516 /* Removing the head of the list, adjust timer. */
517 if (list_next == &waiting_locks_list) {
518 /* No more, just cancel. */
519 timer_delete(&waiting_locks_timer);
521 time64_t now = ktime_get_seconds();
522 struct ldlm_lock *next;
525 next = list_entry(list_next, struct ldlm_lock,
527 if (next->l_callback_timestamp - now > 0)
528 delta = lock->l_callback_timestamp - now;
530 mod_timer(&waiting_locks_timer,
531 jiffies + cfs_time_seconds(delta));
534 list_del_init(&lock->l_pending_chain);
539 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
543 if (lock->l_export == NULL) {
544 /* We don't have a "waiting locks list" on clients. */
545 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
549 spin_lock_bh(&waiting_locks_spinlock);
550 ret = __ldlm_del_waiting_lock(lock);
551 ldlm_clear_waited(lock);
552 spin_unlock_bh(&waiting_locks_spinlock);
554 /* remove the lock out of export blocking list */
555 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
556 list_del_init(&lock->l_exp_list);
557 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
561 * release lock ref if it has indeed been removed
564 LDLM_LOCK_RELEASE(lock);
567 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
572 * Prolong the contended lock waiting time.
574 * Called with namespace lock held.
576 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
578 if (lock->l_export == NULL) {
579 /* We don't have a "waiting locks list" on clients. */
580 LDLM_DEBUG(lock, "client lock: no-op");
584 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
585 /* We don't have a "waiting locks list" on OSP. */
586 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
590 spin_lock_bh(&waiting_locks_spinlock);
592 if (list_empty(&lock->l_pending_chain)) {
593 spin_unlock_bh(&waiting_locks_spinlock);
594 LDLM_DEBUG(lock, "wasn't waiting");
599 * we remove/add the lock to the waiting list, so no needs to
600 * release/take a lock reference
602 __ldlm_del_waiting_lock(lock);
603 __ldlm_add_waiting_lock(lock, timeout);
604 spin_unlock_bh(&waiting_locks_spinlock);
606 LDLM_DEBUG(lock, "refreshed to %ds", timeout);
609 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
611 #else /* HAVE_SERVER_SUPPORT */
613 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
618 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, timeout_t timeout)
623 #endif /* !HAVE_SERVER_SUPPORT */
625 #ifdef HAVE_SERVER_SUPPORT
628 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
629 * lock cancel, and their replies). Used for lock callback timeout and AST
632 * \param[in] lock lock which is getting the blocking callback
634 * \retval timeout in seconds to wait for the client reply
636 timeout_t ldlm_bl_timeout(struct ldlm_lock *lock)
639 struct obd_device *obd = lock->l_export->exp_obd;
642 return obd_timeout / 2;
645 * Since these are non-updating timeouts, we should be conservative.
646 * Take more than usually, 150%
647 * It would be nice to have some kind of "early reply" mechanism for
648 * lock callbacks too...
650 timeout = obd_at_get(obd, &lock->l_export->exp_bl_lock_at);
651 return max_t(timeout_t, timeout + (timeout >> 1),
652 (timeout_t)obd_get_ldlm_enqueue_min(obd));
654 EXPORT_SYMBOL(ldlm_bl_timeout);
657 * Calculate the per-export Blocking timeout by the given RPC (covering the
658 * reply to this RPC and the next RPC). The next RPC could be still not CANCEL,
659 * but having the lock refresh mechanism it is enough.
661 * Used for lock refresh timeout when we are in the middle of the process -
662 * BL AST is sent, CANCEL is ahead - it is still 1 reply for the current RPC
663 * and at least 1 RPC (which will trigger another refresh if it will be not
664 * CANCEL) - but more accurate than ldlm_bl_timeout as the timeout is taken
665 * from the RPC (i.e. the view of the client on the current AT) is taken into
668 * \param[in] req req which export needs the timeout calculation
670 * \retval timeout in seconds to wait for the next client's RPC
672 timeout_t ldlm_bl_timeout_by_rpc(struct ptlrpc_request *req)
674 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
675 timeout_t timeout, req_timeout, at_timeout, netl;
676 struct obd_device *obd = req->rq_export->exp_obd;
679 return obd_timeout / 2;
681 /* A blocked lock means somebody in the cluster is waiting, and we
682 * should not consider the worst ever case, consisting of a chain of
683 * failures on each step, however this timeout should survive a
684 * recovery of at least 1 failure, let this one to be the worst one:
685 * in case a server NID is dead first re-connect is done through the
686 * same router and also times out.
688 * Either this on the next RPC times out, take the max.
689 * Considering the current RPC, take just the left time.
691 netl = obd_at_get(obd,
692 &req->rq_export->exp_imp_reverse->imp_at.iat_net_latency);
693 req_timeout = req->rq_deadline - ktime_get_real_seconds() + netl;
694 at_timeout = at_est2timeout(obd_at_get(obd, &svcpt->scp_at_estimate))
696 req_timeout = max(req_timeout, at_timeout);
698 /* Take 1 re-connect failure and 1 re-connect success into account. */
699 timeout = at_timeout + INITIAL_CONNECT_TIMEOUT + netl + req_timeout;
701 /* Client's timeout is calculated as at_est2timeout(), let's be a bit
702 * more conservative than client */
703 return max(timeout + (timeout >> 4),
704 (timeout_t)obd_get_ldlm_enqueue_min(obd));
706 EXPORT_SYMBOL(ldlm_bl_timeout_by_rpc);
709 * Perform lock cleanup if AST sending failed.
711 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
712 const char *ast_type)
714 LCONSOLE_ERROR_MSG(0x138,
715 "%s: A client on nid %s was evicted due to a lock %s callback time out: rc %d\n",
716 lock->l_export->exp_obd->obd_name,
717 obd_export_nid2str(lock->l_export), ast_type, rc);
719 if (obd_dump_on_timeout)
720 libcfs_debug_dumplog();
721 spin_lock_bh(&waiting_locks_spinlock);
722 if (__ldlm_del_waiting_lock(lock) == 0)
724 * the lock was not in any list, grab an extra ref before adding
725 * the lock to the expired list
728 /* differentiate it from expired locks */
729 lock->l_callback_timestamp = 0;
730 list_add(&lock->l_pending_chain, &expired_lock_list);
731 wake_up(&expired_lock_wait_queue);
732 spin_unlock_bh(&waiting_locks_spinlock);
736 * Perform lock cleanup if AST reply came with error.
738 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
739 struct ptlrpc_request *req, int rc,
740 const char *ast_type)
742 struct lnet_processid *peer = &req->rq_import->imp_connection->c_peer;
744 if (!req->rq_replied || (rc && rc != -EINVAL)) {
745 if (ldlm_is_cancel(lock)) {
747 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
748 ast_type, req, req->rq_xid,
749 libcfs_nidstr(&peer->nid));
750 ldlm_lock_cancel(lock);
752 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
754 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
756 * Upon umount process the AST fails because cannot be
757 * sent. This shouldn't lead to the client eviction.
758 * -ENODEV error is returned by ptl_send_rpc() for
759 * new request in such import.
760 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
761 * if imp_invalid is set or obd_no_recov.
762 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
763 * in ptlrpc_import_delay_req() as well with -EIO code.
764 * In all such cases errors are ignored.
767 "%s AST can't be sent due to a server %s failure or umount process: rc = %d",
769 req->rq_import->imp_obd->obd_name, rc);
772 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
773 libcfs_nidstr(&peer->nid),
774 req->rq_replied ? "returned error from" :
775 "failed to reply to",
776 ast_type, req, req->rq_xid,
777 (req->rq_repmsg != NULL) ?
778 lustre_msg_get_status(req->rq_repmsg) : 0,
780 ldlm_failed_ast(lock, rc, ast_type);
786 struct ldlm_resource *res = lock->l_resource;
789 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
790 libcfs_nidstr(&peer->nid),
792 lustre_msg_get_status(req->rq_repmsg) : -1,
793 ast_type, req, req->rq_xid);
796 * update lvbo to return proper attributes.
799 ldlm_resource_getref(res);
800 ldlm_lvbo_update(res, lock, NULL, 1);
801 ldlm_resource_putref(res);
803 ldlm_lock_cancel(lock);
810 static int ldlm_cb_interpret(const struct lu_env *env,
811 struct ptlrpc_request *req, void *args, int rc)
813 struct ldlm_cb_async_args *ca = args;
814 struct ldlm_lock *lock = ca->ca_lock;
815 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
819 LASSERT(lock != NULL);
822 case LDLM_GL_CALLBACK:
824 * Update the LVB from disk if the AST failed
825 * (this is a legal race)
827 * - Glimpse callback of local lock just returns
828 * -ELDLM_NO_LOCK_DATA.
829 * - Glimpse callback of remote lock might return
830 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
832 if (unlikely(arg->gl_interpret_reply)) {
833 rc = arg->gl_interpret_reply(NULL, req, args, rc);
834 } else if (rc == -ELDLM_NO_LOCK_DATA) {
836 "lost race - client has a lock but no inode");
837 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
838 } else if (rc != 0) {
839 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
841 rc = ldlm_lvbo_update(lock->l_resource,
845 case LDLM_BL_CALLBACK:
847 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
849 case LDLM_CP_CALLBACK:
851 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
854 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
859 /* release extra reference taken in ldlm_ast_fini() */
860 LDLM_LOCK_RELEASE(lock);
863 atomic_inc(&arg->restart);
868 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
870 struct ldlm_cb_async_args *ca = data;
871 struct ldlm_lock *lock = ca->ca_lock;
873 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
876 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
877 struct ldlm_cb_set_arg *arg,
878 struct ldlm_lock *lock,
885 if (unlikely(instant_cancel)) {
886 rc = ptl_send_rpc(req, 1);
887 ptlrpc_req_finished(req);
889 atomic_inc(&arg->restart);
892 ptlrpc_set_add_req(arg->set, req);
899 * Check if there are requests in the export request list which prevent
900 * the lock canceling and make these requests high priority ones.
902 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
904 struct ptlrpc_request *req;
908 if (lock->l_export == NULL) {
909 LDLM_DEBUG(lock, "client lock: no-op");
913 spin_lock(&lock->l_export->exp_rpc_lock);
914 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
917 * Do not process requests that were not yet added to there
918 * incoming queue or were already removed from there for
919 * processing. We evaluate ptlrpc_nrs_req_can_move() without
920 * holding svcpt->scp_req_lock, and then redo the check with
921 * the lock held once we need to obtain a reliable result.
923 if (ptlrpc_nrs_req_can_move(req) &&
924 req->rq_ops->hpreq_lock_match &&
925 req->rq_ops->hpreq_lock_match(req, lock))
926 ptlrpc_nrs_req_hp_move(req);
928 spin_unlock(&lock->l_export->exp_rpc_lock);
933 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
934 * enqueued server lock conflicts with given one.
936 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
937 * to wait for client response.
939 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
940 struct ldlm_lock_desc *desc,
941 void *data, int flag)
943 struct ldlm_cb_async_args *ca;
944 struct ldlm_cb_set_arg *arg = data;
945 struct ldlm_request *body;
946 struct ptlrpc_request *req;
947 int instant_cancel = 0;
949 struct obd_device *obd;
953 if (flag == LDLM_CB_CANCELING)
954 /* Don't need to do anything here. */
957 if (CFS_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
958 LDLM_DEBUG(lock, "dropping BL AST");
963 LASSERT(data != NULL);
965 obd = lock->l_export->exp_obd;
966 if (obd->obd_recovering != 0)
967 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
969 ldlm_lock_reorder_req(lock);
971 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
972 &RQF_LDLM_BL_CALLBACK,
973 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
977 ca = ptlrpc_req_async_args(ca, req);
978 ca->ca_set_arg = arg;
981 req->rq_interpret_reply = ldlm_cb_interpret;
983 lock_res_and_lock(lock);
984 if (ldlm_is_destroyed(lock)) {
985 /* What's the point? */
986 unlock_res_and_lock(lock);
987 ptlrpc_req_finished(req);
991 if (!ldlm_is_granted(lock)) {
993 * this blocking AST will be communicated as part of the
994 * completion AST instead
996 ldlm_add_blocked_lock(lock);
997 ldlm_set_waited(lock);
998 unlock_res_and_lock(lock);
1000 ptlrpc_req_finished(req);
1001 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
1005 if (ldlm_is_cancel_on_block(lock))
1008 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1009 body->lock_handle[0] = lock->l_remote_handle;
1010 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
1011 body->lock_desc = *desc;
1012 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
1014 LDLM_DEBUG(lock, "server preparing blocking AST");
1016 ptlrpc_request_set_replen(req);
1017 ldlm_set_cbpending(lock);
1018 if (instant_cancel) {
1019 unlock_res_and_lock(lock);
1020 ldlm_lock_cancel(lock);
1022 req->rq_no_resend = 1;
1024 LASSERT(ldlm_is_granted(lock));
1025 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1026 unlock_res_and_lock(lock);
1028 /* Do not resend after lock callback timeout */
1029 req->rq_delay_limit = ldlm_bl_timeout(lock);
1030 req->rq_resend_cb = ldlm_update_resend;
1033 req->rq_send_state = LUSTRE_IMP_FULL;
1034 /* ptlrpc_request_alloc_pack already set timeout */
1035 if (obd_at_off(obd))
1036 req->rq_timeout = ldlm_get_rq_timeout();
1038 if (lock->l_export && lock->l_export->exp_nid_stats &&
1039 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1040 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1041 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
1043 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1049 * ->l_completion_ast callback for a remote lock in server namespace.
1051 * Sends AST to the client notifying it of lock granting. If initial
1052 * lock response was not sent yet, instead of sending another RPC, just
1053 * mark the lock as granted and client will understand
1055 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
1057 struct ldlm_cb_set_arg *arg = data;
1058 struct ldlm_request *body;
1059 struct ptlrpc_request *req;
1060 struct ldlm_cb_async_args *ca;
1061 int instant_cancel = 0;
1064 struct obd_device *obd;
1068 LASSERT(lock != NULL);
1069 LASSERT(data != NULL);
1071 if (CFS_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
1072 LDLM_DEBUG(lock, "dropping CP AST");
1076 obd = lock->l_export->exp_obd;
1077 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
1078 &RQF_LDLM_CP_CALLBACK);
1082 /* server namespace, doesn't need lock */
1083 lvb_len = ldlm_lvbo_size(lock);
1085 * LU-3124 & LU-2187: to not return layout in completion AST because
1086 * it may deadlock for LU-2187, or client may not have enough space
1087 * for large layout. The layout will be returned to client with an
1088 * extra RPC to fetch xattr.lov
1090 if (ldlm_has_layout(lock))
1093 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
1094 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
1096 ptlrpc_request_free(req);
1100 ca = ptlrpc_req_async_args(ca, req);
1101 ca->ca_set_arg = arg;
1104 req->rq_interpret_reply = ldlm_cb_interpret;
1105 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1107 body->lock_handle[0] = lock->l_remote_handle;
1108 body->lock_handle[1].cookie = lock->l_handle.h_cookie;
1109 body->lock_flags = ldlm_flags_to_wire(flags);
1110 ldlm_lock2desc(lock, &body->lock_desc);
1112 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
1113 lvb_len = ldlm_lvbo_fill(lock, lvb, &lvb_len);
1116 * We still need to send the RPC to wake up the blocked
1117 * enqueue thread on the client.
1119 * Consider old client, there is no better way to notify
1120 * the failure, just zero-sized the LVB, then the client
1121 * will fail out as "-EPROTO".
1123 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1127 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1132 LDLM_DEBUG(lock, "server preparing completion AST");
1134 ptlrpc_request_set_replen(req);
1136 req->rq_send_state = LUSTRE_IMP_FULL;
1137 /* ptlrpc_request_pack already set timeout */
1138 if (obd_at_off(obd))
1139 req->rq_timeout = ldlm_get_rq_timeout();
1141 /* We only send real blocking ASTs after the lock is granted */
1142 lock_res_and_lock(lock);
1143 if (ldlm_is_ast_sent(lock)) {
1144 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1145 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1146 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1150 * We might get here prior to ldlm_handle_enqueue setting
1151 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1152 * into waiting list, but this is safe and similar code in
1153 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1154 * that would not only cancel the lock, but will also remove
1155 * it from waiting list
1157 if (ldlm_is_cancel_on_block(lock)) {
1158 unlock_res_and_lock(lock);
1159 ldlm_lock_cancel(lock);
1162 req->rq_no_resend = 1;
1164 lock_res_and_lock(lock);
1166 /* start the lock-timeout clock */
1167 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1168 /* Do not resend after lock callback timeout */
1169 req->rq_delay_limit = ldlm_bl_timeout(lock);
1170 req->rq_resend_cb = ldlm_update_resend;
1173 unlock_res_and_lock(lock);
1175 if (lock->l_export && lock->l_export->exp_nid_stats &&
1176 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1177 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1178 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1180 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1182 RETURN(lvb_len < 0 ? lvb_len : rc);
1184 EXPORT_SYMBOL(ldlm_server_completion_ast);
1187 * Server side ->l_glimpse_ast handler for client locks.
1189 * Sends glimpse AST to the client and waits for reply. Then updates
1190 * lvbo with the result.
1192 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1194 struct ldlm_cb_set_arg *arg = data;
1195 struct ldlm_request *body;
1196 struct ptlrpc_request *req;
1197 struct ldlm_cb_async_args *ca;
1199 struct req_format *req_fmt;
1200 struct obd_device *obd = lock->l_export->exp_obd;
1204 LASSERT(lock != NULL);
1206 if (arg->gl_desc != NULL)
1207 /* There is a glimpse descriptor to pack */
1208 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1210 req_fmt = &RQF_LDLM_GL_CALLBACK;
1212 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1213 req_fmt, LUSTRE_DLM_VERSION,
1219 if (arg->gl_desc != NULL) {
1220 /* copy the GL descriptor */
1221 union ldlm_gl_desc *desc;
1223 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1224 *desc = *arg->gl_desc;
1227 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1228 body->lock_handle[0] = lock->l_remote_handle;
1229 ldlm_lock2desc(lock, &body->lock_desc);
1231 ca = ptlrpc_req_async_args(ca, req);
1232 ca->ca_set_arg = arg;
1235 /* server namespace, doesn't need lock */
1236 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1237 ldlm_lvbo_size(lock));
1238 ptlrpc_request_set_replen(req);
1240 req->rq_send_state = LUSTRE_IMP_FULL;
1241 /* ptlrpc_request_alloc_pack already set timeout */
1242 if (obd_at_off(obd))
1243 req->rq_timeout = ldlm_get_rq_timeout();
1245 req->rq_interpret_reply = ldlm_cb_interpret;
1247 if (lock->l_export && lock->l_export->exp_nid_stats) {
1248 struct nid_stat *nid_stats = lock->l_export->exp_nid_stats;
1250 lprocfs_counter_incr(nid_stats->nid_ldlm_stats,
1251 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1254 rc = ldlm_ast_fini(req, arg, lock, 0);
1258 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1260 int ldlm_glimpse_locks(struct ldlm_resource *res,
1261 struct list_head *gl_work_list)
1267 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1269 if (rc == -ERESTART)
1270 ldlm_reprocess_all(res, 0);
1274 EXPORT_SYMBOL(ldlm_glimpse_locks);
1276 /* return LDLM lock associated with a lock callback request */
1277 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1279 struct ldlm_cb_async_args *ca;
1280 struct ldlm_lock *lock;
1284 ca = ptlrpc_req_async_args(ca, req);
1287 RETURN(ERR_PTR(-EFAULT));
1291 EXPORT_SYMBOL(ldlm_request_lock);
1294 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1295 * service threads to carry out client lock enqueueing requests.
1297 int ldlm_handle_enqueue(struct ldlm_namespace *ns,
1298 struct req_capsule *pill,
1299 const struct ldlm_request *dlm_req,
1300 const struct ldlm_callback_suite *cbs)
1302 struct ldlm_reply *dlm_rep;
1304 enum ldlm_error err = ELDLM_OK;
1305 struct ldlm_lock *lock = NULL;
1306 void *cookie = NULL;
1308 struct ldlm_resource *res = NULL;
1309 struct ptlrpc_request *req = pill->rc_req;
1310 const struct lu_env *env = req->rq_svc_thread->t_env;
1314 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1316 LASSERT(req && req->rq_export);
1318 if (req_capsule_ptlreq(pill))
1319 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF,
1322 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1324 /* for intent enqueue the stat will be updated inside intent policy */
1325 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1326 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1327 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1329 if (req->rq_export->exp_nid_stats &&
1330 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1331 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1332 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1334 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1335 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1336 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1337 dlm_req->lock_desc.l_resource.lr_type);
1338 GOTO(out, rc = -EFAULT);
1341 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1342 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1343 dlm_req->lock_desc.l_req_mode &
1344 (dlm_req->lock_desc.l_req_mode-1))) {
1345 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1346 dlm_req->lock_desc.l_req_mode);
1347 GOTO(out, rc = -EFAULT);
1350 if (unlikely((flags & LDLM_FL_REPLAY) ||
1351 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1352 /* Find an existing lock in the per-export lock hash */
1354 * In the function below, .hs_keycmp resolves to
1355 * ldlm_export_lock_keycmp()
1357 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1358 (void *)&dlm_req->lock_handle[0]);
1360 DEBUG_REQ(D_DLMTRACE, req,
1361 "found existing lock cookie %#llx",
1362 lock->l_handle.h_cookie);
1363 flags |= LDLM_FL_RESENT;
1364 GOTO(existing_lock, rc = 0);
1367 if (ldlm_reclaim_full()) {
1368 DEBUG_REQ(D_DLMTRACE, req,
1369 "Too many granted locks, reject current enqueue request and let the client retry later");
1370 GOTO(out, rc = -EINPROGRESS);
1374 /* The lock's callback data might be set in the policy function */
1375 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1376 dlm_req->lock_desc.l_resource.lr_type,
1377 dlm_req->lock_desc.l_req_mode,
1378 cbs, NULL, 0, LVB_T_NONE);
1385 lock->l_remote_handle = dlm_req->lock_handle[0];
1386 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1389 * Initialize resource lvb but not for a lock being replayed since
1390 * Client already got lvb sent in this case.
1391 * This must occur early since some policy methods assume resource
1392 * lvb is available (lr_lvb_data != NULL).
1394 res = lock->l_resource;
1395 if (!(flags & LDLM_FL_REPLAY)) {
1396 /* non-replayed lock, delayed lvb init may need to be done */
1397 rc = ldlm_lvbo_init(res);
1399 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1404 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1406 * Don't enqueue a lock onto the export if it is been disonnected
1407 * due to eviction (b=3822) or server umount (b=24324).
1408 * Cancel it now instead.
1410 if (req->rq_export->exp_disconnected) {
1411 LDLM_ERROR(lock, "lock on disconnected export %p",
1413 GOTO(out, rc = -ENOTCONN);
1416 lock->l_export = class_export_lock_get(req->rq_export, lock);
1417 if (lock->l_export->exp_lock_hash)
1418 cfs_hash_add(lock->l_export->exp_lock_hash,
1419 &lock->l_remote_handle,
1423 * Inherit the enqueue flags before the operation, because we do not
1424 * keep the res lock on return and next operations (BL AST) may proceed
1427 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1428 LDLM_FL_INHERIT_MASK);
1430 ldlm_convert_policy_to_local(req->rq_export,
1431 dlm_req->lock_desc.l_resource.lr_type,
1432 &dlm_req->lock_desc.l_policy_data,
1433 &lock->l_policy_data);
1434 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
1435 lock->l_req_extent = lock->l_policy_data.l_extent;
1436 } else if (dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) {
1437 lock->l_policy_data.l_inodebits.try_bits =
1438 dlm_req->lock_desc.l_policy_data.l_inodebits.try_bits;
1439 lock->l_policy_data.l_inodebits.li_gid =
1440 dlm_req->lock_desc.l_policy_data.l_inodebits.li_gid;
1445 if (!(flags & LDLM_FL_HAS_INTENT)) {
1446 /* based on the assumption that lvb size never changes during
1447 * resource life time otherwise it need resource->lr_lock's
1449 req_capsule_set_size(pill, &RMF_DLM_LVB,
1450 RCL_SERVER, ldlm_lvbo_size(lock));
1452 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1453 GOTO(out, rc = -ENOMEM);
1455 rc = req_capsule_server_pack(pill);
1460 err = ldlm_lock_enqueue(env, ns, &lock, cookie, &flags);
1467 dlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1469 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1470 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1472 if (lock->l_resource->lr_type == LDLM_EXTENT)
1473 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1476 * We never send a blocking AST until the lock is granted, but
1477 * we can tell it right now
1479 lock_res_and_lock(lock);
1482 * Now take into account flags to be inherited from original lock
1483 * request both in reply to client and in our own lock flags.
1485 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1486 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1489 * Don't move a pending lock onto the export if it has already been
1490 * disconnected due to eviction (b=5683) or server umount (b=24324).
1491 * Cancel it now instead.
1493 if (unlikely(req->rq_export->exp_disconnected ||
1494 CFS_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1495 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1497 } else if (ldlm_is_ast_sent(lock)) {
1498 /* fill lock desc for possible lock convert */
1499 if (lock->l_blocking_lock &&
1500 lock->l_resource->lr_type == LDLM_IBITS) {
1501 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1502 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1505 "save blocking bits %llx in granted lock",
1506 bl_lock->l_policy_data.l_inodebits.bits);
1508 * If lock is blocked then save blocking ibits
1509 * in returned lock policy for the possible lock
1510 * convert on a client.
1512 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1513 bl_lock->l_policy_data.l_inodebits.bits;
1515 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1516 if (ldlm_is_granted(lock)) {
1518 * Only cancel lock if it was granted, because it would
1519 * be destroyed immediately and would never be granted
1520 * in the future, causing timeouts on client. Not
1521 * granted lock will be cancelled immediately after
1522 * sending completion AST.
1524 if (ldlm_is_cancel_on_block(lock)) {
1525 unlock_res_and_lock(lock);
1526 ldlm_lock_cancel(lock);
1527 lock_res_and_lock(lock);
1529 ldlm_add_waiting_lock(lock,
1530 ldlm_bl_timeout(lock));
1534 unlock_res_and_lock(lock);
1538 if (req_capsule_ptlreq(pill)) {
1539 req->rq_status = rc ?: err; /* return either error - b=11190 */
1540 if (!req->rq_packed_final) {
1541 int rc1 = lustre_pack_reply(req, 1, NULL, NULL);
1548 * The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1549 * ldlm_reprocess_all. If this moves, revisit that code. -phil
1553 "server-side enqueue handler, sending reply (err=%d, rc=%d)",
1557 req_capsule_has_field(pill, &RMF_DLM_LVB,
1559 ldlm_lvbo_size(lock) > 0) {
1564 buf = req_capsule_server_get(pill, &RMF_DLM_LVB);
1565 LASSERTF(buf != NULL, "req %px, lock %px\n", req, lock);
1566 buflen = req_capsule_get_size(pill, &RMF_DLM_LVB,
1569 * non-replayed lock, delayed lvb init may
1570 * need to be occur now
1572 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1575 rc2 = ldlm_lvbo_fill(lock, buf, &buflen);
1577 req_capsule_shrink(pill, &RMF_DLM_LVB,
1579 } else if (rc2 == -ERANGE) {
1580 rc2 = req_capsule_server_grow(
1587 * if we can't grow the buffer,
1588 * it's ok to return empty lvb
1592 pill, &RMF_DLM_LVB, 0,
1598 } else if (flags & LDLM_FL_REPLAY) {
1599 /* no LVB resend upon replay */
1601 req_capsule_shrink(pill, &RMF_DLM_LVB,
1610 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1611 if (lock->l_export) {
1612 ldlm_lock_cancel(lock);
1614 lock_res_and_lock(lock);
1615 ldlm_resource_unlink_lock(lock);
1616 ldlm_lock_destroy_nolock(lock);
1617 unlock_res_and_lock(lock);
1619 ldlm_reprocess_all(lock->l_resource, lock->l_policy_data.l_inodebits.bits);
1622 if (!err && !ldlm_is_cbpending(lock) &&
1623 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1624 ldlm_reprocess_all(lock->l_resource,
1625 lock->l_policy_data.l_inodebits.bits);
1627 LDLM_LOCK_RELEASE(lock);
1630 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1635 EXPORT_SYMBOL(ldlm_handle_enqueue);
1638 * Clear the blocking lock, the race is possible between ldlm_handle_convert0()
1639 * and ldlm_work_bl_ast_lock(), so this is done under lock with check for NULL.
1641 void ldlm_clear_blocking_lock(struct ldlm_lock *lock)
1643 if (lock->l_blocking_lock) {
1644 LDLM_LOCK_RELEASE(lock->l_blocking_lock);
1645 lock->l_blocking_lock = NULL;
1649 /* A lock can be converted to new ibits or mode and should be considered
1650 * as new lock. Clear all states related to a previous blocking AST
1651 * processing so new conflicts will cause new blocking ASTs.
1653 * This is used during lock convert below and lock downgrade to COS mode in
1654 * ldlm_lock_mode_downgrade().
1656 void ldlm_clear_blocking_data(struct ldlm_lock *lock)
1658 ldlm_clear_ast_sent(lock);
1659 lock->l_bl_ast_run = 0;
1660 ldlm_clear_blocking_lock(lock);
1664 * Main LDLM entry point for server code to process lock conversion requests.
1666 int ldlm_handle_convert0(struct ptlrpc_request *req,
1667 const struct ldlm_request *dlm_req)
1669 struct obd_export *exp = req->rq_export;
1670 struct ldlm_reply *dlm_rep;
1671 struct ldlm_lock *lock;
1678 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1679 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1680 LDLM_CONVERT - LDLM_FIRST_OPC);
1682 rc = req_capsule_server_pack(&req->rq_pill);
1686 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1687 dlm_rep->lock_flags = dlm_req->lock_flags;
1689 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1691 LDLM_DEBUG_NOLOCK("server lock is canceled already");
1692 req->rq_status = ELDLM_NO_LOCK_DATA;
1696 LDLM_DEBUG(lock, "server-side convert handler START");
1698 lock_res_and_lock(lock);
1699 bits = lock->l_policy_data.l_inodebits.bits;
1700 new_bits = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1702 if (ldlm_is_cancel(lock)) {
1703 LDLM_DEBUG(lock, "convert on canceled lock!");
1704 unlock_res_and_lock(lock);
1705 GOTO(out_put, rc = ELDLM_NO_LOCK_DATA);
1708 if (dlm_req->lock_desc.l_req_mode != lock->l_granted_mode) {
1709 LDLM_ERROR(lock, "lock mode differs!");
1710 unlock_res_and_lock(lock);
1711 GOTO(out_put, rc = -EPROTO);
1714 if (bits == new_bits) {
1716 * This can be valid situation if CONVERT RPCs are
1717 * re-ordered. Just finish silently
1719 LDLM_DEBUG(lock, "lock is converted already!");
1720 unlock_res_and_lock(lock);
1722 if (ldlm_is_waited(lock))
1723 ldlm_del_waiting_lock(lock);
1725 ldlm_clear_cbpending(lock);
1726 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1727 ldlm_inodebits_drop(lock, bits & ~new_bits);
1729 ldlm_clear_blocking_data(lock);
1730 unlock_res_and_lock(lock);
1732 /* All old bits should be reprocessed to send new BL AST if
1733 * it wasn't sent earlier due to LDLM_FL_AST_SENT bit set.
1735 ldlm_reprocess_all(lock->l_resource, bits);
1738 dlm_rep->lock_handle = lock->l_remote_handle;
1739 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1740 &dlm_rep->lock_desc.l_policy_data);
1744 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d", rc);
1745 LDLM_LOCK_PUT(lock);
1746 req->rq_status = rc;
1751 * Cancel all the locks whose handles are packed into ldlm_request
1753 * Called by server code expecting such combined cancel activity
1756 int ldlm_request_cancel(struct ptlrpc_request *req,
1757 const struct ldlm_request *dlm_req,
1758 int first, enum lustre_at_flags flags)
1760 struct ldlm_resource *res, *pres = NULL;
1761 struct ldlm_lock *lock;
1762 int i, count, done = 0;
1767 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
1768 if (size <= offsetof(struct ldlm_request, lock_handle) ||
1769 (size - offsetof(struct ldlm_request, lock_handle)) /
1770 sizeof(struct lustre_handle) < dlm_req->lock_count)
1773 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1777 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1781 * There is no lock on the server at the replay time,
1782 * skip lock cancelling to make replay tests to pass.
1784 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1787 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, starting at %d",
1790 for (i = first; i < count; i++) {
1791 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1793 /* below message checked in replay-single.sh test_36 */
1794 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (cookie %llu)",
1795 dlm_req->lock_handle[i].cookie);
1799 res = lock->l_resource;
1803 * This code is an optimization to only attempt lock
1804 * granting on the resource (that could be CPU-expensive)
1805 * after we are done cancelling lock in that resource.
1809 ldlm_reprocess_all(pres, 0);
1810 LDLM_RESOURCE_DELREF(pres);
1811 ldlm_resource_putref(pres);
1814 ldlm_resource_getref(res);
1815 LDLM_RESOURCE_ADDREF(res);
1817 if (!ldlm_is_discard_data(lock))
1818 ldlm_lvbo_update(res, lock,
1824 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1825 lock->l_blast_sent != 0) {
1826 timeout_t delay = 0;
1828 if (ktime_get_seconds() > lock->l_blast_sent)
1829 delay = ktime_get_seconds() -
1832 "server cancels blocked lock after %ds",
1834 obd_at_measure(lock->l_export->exp_obd,
1835 &lock->l_export->exp_bl_lock_at,
1838 ldlm_lock_cancel(lock);
1839 LDLM_LOCK_PUT(lock);
1842 ldlm_reprocess_all(pres, 0);
1843 LDLM_RESOURCE_DELREF(pres);
1844 ldlm_resource_putref(pres);
1846 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1849 EXPORT_SYMBOL(ldlm_request_cancel);
1852 * Main LDLM entry point for server code to cancel locks.
1854 * Typically gets called from service handler on LDLM_CANCEL opc.
1856 int ldlm_handle_cancel(struct ptlrpc_request *req)
1858 struct ldlm_request *dlm_req;
1863 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1864 if (dlm_req == NULL) {
1865 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1869 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) <
1870 offsetof(struct ldlm_request, lock_handle[1]))
1873 if (req->rq_export && req->rq_export->exp_nid_stats &&
1874 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1875 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1876 LDLM_CANCEL - LDLM_FIRST_OPC);
1878 rc = req_capsule_server_pack(&req->rq_pill);
1882 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1883 req->rq_status = LUSTRE_ESTALE;
1885 RETURN(ptlrpc_reply(req));
1887 #endif /* HAVE_SERVER_SUPPORT */
1890 * Server may pass additional information about blocking lock.
1891 * For IBITS locks it is conflicting bits which can be used for
1892 * lock convert instead of cancel.
1894 void ldlm_bl_desc2lock(const struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1896 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
1898 check_res_locked(lock->l_resource);
1899 if (ns_is_client(ns) && ld &&
1900 (lock->l_resource->lr_type == LDLM_IBITS)) {
1902 * Lock description contains policy of blocking lock,
1903 * and its cancel_bits is used to pass conflicting bits.
1904 * NOTE: ld can be NULL or can be not NULL but zeroed if
1905 * passed from ldlm_bl_thread_blwi(), check below used bits
1906 * in ld to make sure it is valid description.
1908 * If server may replace lock resource keeping the same cookie,
1909 * never use cancel bits from different resource, full cancel
1912 if (ld->l_policy_data.l_inodebits.cancel_bits &&
1913 ldlm_res_eq(&ld->l_resource.lr_name,
1914 &lock->l_resource->lr_name) &&
1915 !(ldlm_is_cbpending(lock) &&
1916 lock->l_policy_data.l_inodebits.cancel_bits == 0)) {
1917 /* always combine conflicting ibits */
1918 lock->l_policy_data.l_inodebits.cancel_bits |=
1919 ld->l_policy_data.l_inodebits.cancel_bits;
1921 /* If cancel_bits are not obtained or
1922 * if the lock is already CBPENDING and
1923 * has no cancel_bits set
1924 * - the full lock is to be cancelled
1926 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1932 * Callback handler for receiving incoming blocking ASTs.
1934 * This can only happen on client side.
1936 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1937 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1943 LDLM_DEBUG(lock, "client blocking AST callback handler");
1945 lock_res_and_lock(lock);
1947 /* get extra information from desc if any */
1948 ldlm_bl_desc2lock(ld, lock);
1949 ldlm_set_cbpending(lock);
1951 do_ast = (!lock->l_readers && !lock->l_writers);
1952 unlock_res_and_lock(lock);
1956 "Lock %p already unused, calling callback (%p)\n",
1957 lock, lock->l_blocking_ast);
1958 if (lock->l_blocking_ast != NULL)
1959 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1963 "Lock %p is referenced, will be cancelled later\n",
1967 LDLM_DEBUG(lock, "client blocking callback handler END");
1968 LDLM_LOCK_RELEASE(lock);
1972 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1974 if (req->rq_no_reply)
1977 req->rq_status = rc;
1978 if (!req->rq_packed_final) {
1979 rc = lustre_pack_reply(req, 1, NULL, NULL);
1983 return ptlrpc_reply(req);
1987 * Callback handler for receiving incoming completion ASTs.
1989 * This only can happen on client side.
1991 static int ldlm_handle_cp_callback(struct ptlrpc_request *req,
1992 struct ldlm_namespace *ns,
1993 struct ldlm_request *dlm_req,
1994 struct ldlm_lock *lock)
1996 LIST_HEAD(ast_list);
2002 LDLM_DEBUG(lock, "client completion callback handler START");
2004 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
2005 long to = cfs_time_seconds(1);
2007 ldlm_callback_reply(req, 0);
2010 to = schedule_timeout_interruptible(to);
2011 if (ldlm_is_granted(lock) ||
2012 ldlm_is_destroyed(lock))
2017 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
2019 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
2020 GOTO(out, rc = lvb_len);
2021 } else if (lvb_len > 0) {
2022 if (lock->l_lvb_len > 0) {
2023 /* for extent lock, lvb contains ost_lvb{}. */
2024 LASSERT(lock->l_lvb_data != NULL);
2026 if (unlikely(lock->l_lvb_len < lvb_len)) {
2028 "Replied LVB is larger than expectation, expected = %d, replied = %d",
2029 lock->l_lvb_len, lvb_len);
2030 GOTO(out, rc = -EINVAL);
2035 lock_res_and_lock(lock);
2037 if (!ldlm_res_eq(&dlm_req->lock_desc.l_resource.lr_name,
2038 &lock->l_resource->lr_name)) {
2039 ldlm_resource_unlink_lock(lock);
2040 unlock_res_and_lock(lock);
2041 rc = ldlm_lock_change_resource(ns, lock,
2042 &dlm_req->lock_desc.l_resource.lr_name);
2044 LDLM_ERROR(lock, "Failed to allocate resource");
2047 LDLM_DEBUG(lock, "completion AST, new resource");
2048 lock_res_and_lock(lock);
2051 if (ldlm_is_failed(lock)) {
2052 unlock_res_and_lock(lock);
2053 LDLM_LOCK_RELEASE(lock);
2057 if (ldlm_is_destroyed(lock) ||
2058 ldlm_is_granted(lock)) {
2059 /* b=11300: the lock has already been granted */
2060 unlock_res_and_lock(lock);
2061 LDLM_DEBUG(lock, "Double grant race happened");
2066 * If we receive the completion AST before the actual enqueue returned,
2067 * then we might need to switch lock modes, resources, or extents.
2069 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
2070 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
2071 LDLM_DEBUG(lock, "completion AST, new lock mode");
2074 if (lock->l_resource->lr_type != LDLM_PLAIN) {
2075 ldlm_convert_policy_to_local(req->rq_export,
2076 dlm_req->lock_desc.l_resource.lr_type,
2077 &dlm_req->lock_desc.l_policy_data,
2078 &lock->l_policy_data);
2079 LDLM_DEBUG(lock, "completion AST, new policy data");
2082 ldlm_resource_unlink_lock(lock);
2084 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
2086 * BL_AST locks are not needed in LRU.
2087 * Let ldlm_cancel_lru() be fast.
2089 ldlm_lock_remove_from_lru(lock);
2090 ldlm_bl_desc2lock(&dlm_req->lock_desc, lock);
2091 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
2092 LDLM_DEBUG(lock, "completion AST includes blocking AST");
2095 if (lock->l_lvb_len > 0) {
2096 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
2097 lock->l_lvb_data, lvb_len);
2099 unlock_res_and_lock(lock);
2104 ldlm_grant_lock(lock, &ast_list);
2105 unlock_res_and_lock(lock);
2107 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
2110 * Let Enqueue to call osc_lock_upcall() and initialize
2113 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
2115 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
2117 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
2123 lock_res_and_lock(lock);
2124 ldlm_set_failed(lock);
2125 unlock_res_and_lock(lock);
2126 wake_up(&lock->l_waitq);
2128 LDLM_LOCK_RELEASE(lock);
2134 * Callback handler for receiving incoming glimpse ASTs.
2136 * This only can happen on client side. After handling the glimpse AST
2137 * we also consider dropping the lock here if it is unused locally for a
2140 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
2141 struct ldlm_namespace *ns,
2142 struct ldlm_request *dlm_req,
2143 struct ldlm_lock *lock)
2145 struct ldlm_lock_desc *ld = &dlm_req->lock_desc;
2150 LDLM_DEBUG(lock, "client glimpse AST callback handler");
2152 if (lock->l_glimpse_ast != NULL)
2153 rc = lock->l_glimpse_ast(lock, req);
2155 if (req->rq_repmsg != NULL) {
2158 req->rq_status = rc;
2162 lock_res_and_lock(lock);
2163 if (lock->l_granted_mode == LCK_PW &&
2164 !lock->l_readers && !lock->l_writers &&
2165 ktime_after(ktime_get(),
2166 ktime_add(lock->l_last_used, ns->ns_dirty_age_limit))) {
2167 unlock_res_and_lock(lock);
2169 /* For MDS glimpse it is always DOM lock, set corresponding
2170 * cancel_bits to perform lock convert if needed
2172 if (lock->l_resource->lr_type == LDLM_IBITS)
2173 ld->l_policy_data.l_inodebits.cancel_bits =
2175 if (ldlm_bl_to_thread_lock(ns, ld, lock))
2176 ldlm_handle_bl_callback(ns, ld, lock);
2181 unlock_res_and_lock(lock);
2182 LDLM_LOCK_RELEASE(lock);
2186 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
2187 enum ldlm_cancel_flags cancel_flags)
2189 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2190 char *prio = "regular";
2195 spin_lock(&blp->blp_lock);
2196 /* cannot access blwi after added to list and lock is dropped */
2197 count = blwi->blwi_lock ? 1 : blwi->blwi_count;
2199 /* if the server is waiting on a lock to be cancelled (bl_ast), this is
2200 * an urgent request and should go in the priority queue so it doesn't
2201 * get stuck behind non-priority work (eg, lru size management)
2203 * We also prioritize discard_data, which is for eviction handling
2205 if (blwi->blwi_lock &&
2206 (ldlm_is_discard_data(blwi->blwi_lock) ||
2207 ldlm_is_bl_ast(blwi->blwi_lock))) {
2208 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
2211 /* other blocking callbacks are added to the regular list */
2212 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
2214 blp->blp_total_locks += count;
2215 blp->blp_total_blwis++;
2216 spin_unlock(&blp->blp_lock);
2218 wake_up(&blp->blp_waitq);
2220 /* unlocked read of blp values is intentional - OK for debug */
2222 "added %d/%d locks to %s blp list, %d blwis in pool\n",
2223 count, blp->blp_total_locks, prio, blp->blp_total_blwis);
2226 * can not check blwi->blwi_flags as blwi could be already freed in
2229 if (!(cancel_flags & LCF_ASYNC))
2230 wait_for_completion(&blwi->blwi_comp);
2235 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
2236 struct ldlm_namespace *ns,
2237 struct ldlm_lock_desc *ld,
2238 struct list_head *cancels, int count,
2239 struct ldlm_lock *lock,
2240 enum ldlm_cancel_flags cancel_flags)
2242 init_completion(&blwi->blwi_comp);
2243 INIT_LIST_HEAD(&blwi->blwi_head);
2245 if (current->flags & PF_MEMALLOC)
2246 blwi->blwi_mem_pressure = 1;
2249 blwi->blwi_flags = cancel_flags;
2251 blwi->blwi_ld = *ld;
2253 list_splice_init(cancels, &blwi->blwi_head);
2254 blwi->blwi_count = count;
2256 blwi->blwi_lock = lock;
2261 * Queues a list of locks \a cancels containing \a count locks
2262 * for later processing by a blocking thread. If \a count is zero,
2263 * then the lock referenced as \a lock is queued instead.
2265 * The blocking thread would then call ->l_blocking_ast callback in the lock.
2266 * If list addition fails an error is returned and caller is supposed to
2267 * call ->l_blocking_ast itself.
2269 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
2270 struct ldlm_lock_desc *ld,
2271 struct ldlm_lock *lock,
2272 struct list_head *cancels, int count,
2273 enum ldlm_cancel_flags cancel_flags)
2277 if (cancels && count == 0)
2280 if (cancel_flags & LCF_ASYNC) {
2281 struct ldlm_bl_work_item *blwi;
2283 OBD_ALLOC(blwi, sizeof(*blwi));
2286 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
2288 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
2291 * if it is synchronous call do minimum mem alloc, as it could
2292 * be triggered from kernel shrinker
2294 struct ldlm_bl_work_item blwi;
2296 memset(&blwi, 0, sizeof(blwi));
2297 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
2298 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
2303 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2304 struct ldlm_lock *lock)
2306 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2309 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2310 struct list_head *cancels, int count,
2311 enum ldlm_cancel_flags cancel_flags)
2313 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2316 int ldlm_bl_to_thread_ns(struct ldlm_namespace *ns)
2318 return ldlm_bl_to_thread(ns, NULL, NULL, NULL, 0, LCF_ASYNC);
2321 int ldlm_bl_thread_wakeup(void)
2323 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2327 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2328 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2330 struct obd_device *obd = req->rq_export->exp_obd;
2338 DEBUG_REQ(D_HSM, req, "%s: handle setinfo", obd->obd_name);
2340 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2342 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2344 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2347 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2349 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2351 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2354 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2357 /* We are responsible for swabbing contents of val */
2359 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2360 /* Pass it on to mdc (the "export" in this case) */
2361 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2363 sizeof(KEY_HSM_COPYTOOL_SEND),
2364 KEY_HSM_COPYTOOL_SEND,
2367 DEBUG_REQ(D_WARNING, req, "ignoring unknown key '%s'", key);
2372 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2373 const char *msg, int rc,
2374 const struct lustre_handle *handle)
2376 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2377 "%s, NID=%s lock=%#llx: rc = %d",
2378 msg, libcfs_idstr(&req->rq_peer),
2379 handle ? handle->cookie : 0, rc);
2380 if (req->rq_no_reply)
2381 CWARN("No reply was sent, maybe cause b=21636.\n");
2383 CWARN("Send reply failed, maybe cause b=21636.\n");
2386 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2387 static int ldlm_callback_handler(struct ptlrpc_request *req)
2389 struct ldlm_namespace *ns;
2390 struct ldlm_request *dlm_req;
2391 struct ldlm_lock *lock;
2397 * Requests arrive in sender's byte order. The ptlrpc service
2398 * handler has already checked and, if necessary, byte-swapped the
2399 * incoming request message body, but I am responsible for the
2403 /* do nothing for sec context finalize */
2404 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2407 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2409 if (req->rq_export == NULL) {
2410 rc = ldlm_callback_reply(req, -ENOTCONN);
2411 ldlm_callback_errmsg(req, "Operate on unconnected server",
2416 LASSERT(req->rq_export != NULL);
2417 LASSERT(req->rq_export->exp_obd != NULL);
2419 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2420 case LDLM_BL_CALLBACK:
2421 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2423 ldlm_callback_reply(req, -(int)cfs_fail_err);
2427 case LDLM_CP_CALLBACK:
2428 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2431 case LDLM_GL_CALLBACK:
2432 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2436 rc = ldlm_handle_setinfo(req);
2437 ldlm_callback_reply(req, rc);
2440 CERROR("unknown opcode %u\n",
2441 lustre_msg_get_opc(req->rq_reqmsg));
2442 ldlm_callback_reply(req, -EPROTO);
2446 ns = req->rq_export->exp_obd->obd_namespace;
2447 LASSERT(ns != NULL);
2449 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2451 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2452 if (dlm_req == NULL) {
2453 rc = ldlm_callback_reply(req, -EPROTO);
2454 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2460 * Force a known safe race, send a cancel to the server for a lock
2461 * which the server has already started a blocking callback on.
2463 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2464 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2465 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2467 CERROR("ldlm_cli_cancel: %d\n", rc);
2470 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2473 "callback on lock %#llx - lock disappeared\n",
2474 dlm_req->lock_handle[0].cookie);
2475 rc = ldlm_callback_reply(req, -EINVAL);
2476 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2477 &dlm_req->lock_handle[0]);
2481 if (ldlm_is_fail_loc(lock) &&
2482 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2483 CFS_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2485 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2486 lock_res_and_lock(lock);
2487 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2489 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2491 * If somebody cancels lock and cache is already dropped,
2492 * or lock is failed before cp_ast received on client,
2493 * we can tell the server we have no lock. Otherwise, we
2494 * should send cancel after dropping the cache.
2496 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2497 ldlm_is_failed(lock)) {
2499 "callback on lock %llx - lock disappeared",
2500 dlm_req->lock_handle[0].cookie);
2501 unlock_res_and_lock(lock);
2502 LDLM_LOCK_RELEASE(lock);
2503 rc = ldlm_callback_reply(req, -EINVAL);
2504 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2505 &dlm_req->lock_handle[0]);
2509 * BL_AST locks are not needed in LRU.
2510 * Let ldlm_cancel_lru() be fast.
2512 ldlm_lock_remove_from_lru(lock);
2513 ldlm_set_bl_ast(lock);
2515 if (lock->l_remote_handle.cookie == 0)
2516 lock->l_remote_handle = dlm_req->lock_handle[1];
2517 unlock_res_and_lock(lock);
2520 * We want the ost thread to get this reply so that it can respond
2521 * to ost requests (write cache writeback) that might be triggered
2524 * But we'd also like to be able to indicate in the reply that we're
2525 * cancelling right now, because it's unused, or have an intent result
2526 * in the reply, so we might have to push the responsibility for sending
2527 * the reply down into the AST handlers, alas.
2530 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2531 case LDLM_BL_CALLBACK:
2532 LDLM_DEBUG(lock, "blocking ast");
2533 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2534 if (!ldlm_is_cancel_on_block(lock)) {
2535 rc = ldlm_callback_reply(req, 0);
2536 if (req->rq_no_reply || rc)
2537 ldlm_callback_errmsg(req, "Normal process", rc,
2538 &dlm_req->lock_handle[0]);
2540 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2541 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2543 case LDLM_CP_CALLBACK:
2544 LDLM_DEBUG(lock, "completion ast");
2545 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2546 rc = ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2547 if (!CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE))
2548 ldlm_callback_reply(req, rc);
2550 case LDLM_GL_CALLBACK:
2551 LDLM_DEBUG(lock, "glimpse ast");
2552 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2553 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2556 LBUG(); /* checked above */
2562 #ifdef HAVE_SERVER_SUPPORT
2564 * Main handler for canceld thread.
2566 * Separated into its own thread to avoid deadlocks.
2568 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2575 * Requests arrive in sender's byte order. The ptlrpc service
2576 * handler has already checked and, if necessary, byte-swapped the
2577 * incoming request message body, but I am responsible for the
2581 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2583 if (req->rq_export == NULL) {
2584 struct ldlm_request *dlm_req;
2586 CERROR("%s from %s arrived at %llu with bad export cookie %llu\n",
2587 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2588 libcfs_nidstr(&req->rq_peer.nid),
2589 (unsigned long long)req->rq_arrival_time.tv_sec,
2590 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2592 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2593 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2594 dlm_req = req_capsule_client_get(&req->rq_pill,
2596 if (dlm_req != NULL)
2597 ldlm_lock_dump_handle(D_ERROR,
2598 &dlm_req->lock_handle[0]);
2600 ldlm_callback_reply(req, -ENOTCONN);
2604 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2605 /* XXX FIXME move this back to mds/handler.c, b=249 */
2607 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2608 CDEBUG(D_INODE, "cancel\n");
2609 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2610 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2611 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2613 rc = ldlm_handle_cancel(req);
2617 struct ldlm_request *dlm_req;
2619 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2620 CDEBUG(D_INODE, "convert\n");
2622 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2623 if (dlm_req == NULL) {
2624 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2625 rc = ldlm_callback_reply(req, -EPROTO);
2627 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2628 rc = ptlrpc_reply(req);
2633 CERROR("invalid opcode %d\n",
2634 lustre_msg_get_opc(req->rq_reqmsg));
2635 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2636 rc = ldlm_callback_reply(req, -EINVAL);
2642 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2643 struct ldlm_lock *lock)
2645 struct ldlm_request *dlm_req;
2646 struct lustre_handle lockh;
2652 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2653 if (dlm_req == NULL)
2656 ldlm_lock2handle(lock, &lockh);
2657 for (i = 0; i < dlm_req->lock_count; i++) {
2658 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2660 DEBUG_REQ(D_RPCTRACE, req,
2661 "Prio raised by lock %#llx", lockh.cookie);
2670 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2672 struct ldlm_request *dlm_req;
2679 /* no prolong in recovery */
2680 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2683 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2684 if (dlm_req == NULL)
2687 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT);
2688 if (size <= offsetof(struct ldlm_request, lock_handle) ||
2689 (size - offsetof(struct ldlm_request, lock_handle)) /
2690 sizeof(struct lustre_handle) < dlm_req->lock_count)
2693 for (i = 0; i < dlm_req->lock_count; i++) {
2694 struct ldlm_lock *lock;
2696 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2700 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2702 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2703 LDLM_LOCK_PUT(lock);
2712 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2713 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2714 .hpreq_check = ldlm_cancel_hpreq_check,
2718 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2722 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2724 if (req->rq_export == NULL)
2727 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2728 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2729 req->rq_ops = &ldlm_cancel_hpreq_ops;
2730 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2731 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2732 req->rq_ops = &ldlm_cancel_hpreq_ops;
2737 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2738 struct hlist_node *hnode, void *data)
2741 struct list_head *rpc_list = data;
2742 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2744 lock_res_and_lock(lock);
2746 if (!ldlm_is_granted(lock)) {
2747 unlock_res_and_lock(lock);
2751 LASSERT(lock->l_resource);
2752 if (lock->l_resource->lr_type != LDLM_IBITS &&
2753 lock->l_resource->lr_type != LDLM_PLAIN) {
2754 unlock_res_and_lock(lock);
2758 if (ldlm_is_ast_sent(lock)) {
2759 unlock_res_and_lock(lock);
2763 LASSERT(lock->l_blocking_ast);
2764 LASSERT(!lock->l_blocking_lock);
2766 ldlm_set_ast_sent(lock);
2767 if (lock->l_export && lock->l_export->exp_lock_hash) {
2769 * NB: it's safe to call cfs_hash_del() even lock isn't
2773 * In the function below, .hs_keycmp resolves to
2774 * ldlm_export_lock_keycmp()
2776 cfs_hash_del(lock->l_export->exp_lock_hash,
2777 &lock->l_remote_handle, &lock->l_exp_hash);
2780 list_add_tail(&lock->l_rk_ast, rpc_list);
2781 LDLM_LOCK_GET(lock);
2783 unlock_res_and_lock(lock);
2787 void ldlm_revoke_export_locks(struct obd_export *exp)
2790 LIST_HEAD(rpc_list);
2793 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2794 ldlm_revoke_lock_cb, &rpc_list, 0);
2795 rc = ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2796 LDLM_WORK_REVOKE_AST);
2798 if (rc == -ERESTART)
2799 ldlm_reprocess_recovery_done(exp->exp_obd->obd_namespace);
2803 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2804 #endif /* HAVE_SERVER_SUPPORT */
2806 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2807 struct ldlm_bl_work_item **p_blwi,
2808 struct obd_export **p_exp)
2810 struct ldlm_bl_work_item *blwi = NULL;
2811 static unsigned int num_bl;
2812 static unsigned int num_stale;
2813 int num_th = atomic_read(&blp->blp_num_threads);
2815 *p_exp = obd_stale_export_get();
2817 spin_lock(&blp->blp_lock);
2818 if (*p_exp != NULL) {
2819 if (num_th == 1 || ++num_stale < num_th) {
2820 spin_unlock(&blp->blp_lock);
2826 /* process a request from the blp_list at least every blp_num_threads */
2827 if (!list_empty(&blp->blp_list) &&
2828 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2829 blwi = list_first_entry(&blp->blp_list,
2830 struct ldlm_bl_work_item, blwi_entry);
2832 if (!list_empty(&blp->blp_prio_list))
2833 blwi = list_first_entry(&blp->blp_prio_list,
2834 struct ldlm_bl_work_item,
2838 if (++num_bl >= num_th)
2840 list_del(&blwi->blwi_entry);
2841 blp->blp_total_locks -= blwi->blwi_lock ? 1 : blwi->blwi_count;
2842 blp->blp_total_blwis--;
2844 spin_unlock(&blp->blp_lock);
2847 /* intentional unlocked read of blp values - OK for debug */
2850 "Got %d locks of %d total in blp. (%d blwis in pool)\n",
2851 blwi->blwi_lock ? 1 : blwi->blwi_count,
2852 blp->blp_total_locks, blp->blp_total_blwis);
2855 "No blwi found in queue (no bl locks in queue)\n");
2858 if (*p_exp != NULL && *p_blwi != NULL) {
2859 obd_stale_export_put(*p_exp);
2863 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2866 /* This only contains temporary data until the thread starts */
2867 struct ldlm_bl_thread_data {
2868 struct ldlm_bl_pool *bltd_blp;
2869 struct completion bltd_comp;
2873 static int ldlm_bl_thread_main(void *arg);
2875 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2877 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2878 struct task_struct *task;
2880 init_completion(&bltd.bltd_comp);
2882 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2883 if (bltd.bltd_num >= blp->blp_max_threads) {
2884 atomic_dec(&blp->blp_num_threads);
2888 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2890 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2891 atomic_dec(&blp->blp_num_threads);
2895 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2898 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2899 bltd.bltd_num, PTR_ERR(task));
2900 atomic_dec(&blp->blp_num_threads);
2901 return PTR_ERR(task);
2903 wait_for_completion(&bltd.bltd_comp);
2908 /* Not fatal if racy and have a few too many threads */
2909 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2910 struct ldlm_bl_work_item *blwi)
2912 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2915 if (atomic_read(&blp->blp_busy_threads) <
2916 atomic_read(&blp->blp_num_threads))
2919 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2920 blwi->blwi_mem_pressure))
2926 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2927 struct ldlm_bl_work_item *blwi)
2929 /* '1' for consistency with code that checks !mpflag to restore */
2930 unsigned int mpflags = 1;
2934 if (blwi->blwi_ns == NULL)
2935 /* added by ldlm_cleanup() */
2936 RETURN(LDLM_ITER_STOP);
2938 if (blwi->blwi_mem_pressure)
2939 mpflags = memalloc_noreclaim_save();
2941 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2943 if (blwi->blwi_count) {
2946 * The special case when we cancel locks in lru
2947 * asynchronously, we pass the list of locks here.
2948 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2949 * canceled locally yet.
2951 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2954 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2956 } else if (blwi->blwi_lock) {
2957 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2960 ldlm_pool_recalc(&blwi->blwi_ns->ns_pool, true);
2961 spin_lock(&blwi->blwi_ns->ns_lock);
2962 blwi->blwi_ns->ns_rpc_recalc = 0;
2963 spin_unlock(&blwi->blwi_ns->ns_lock);
2964 ldlm_namespace_put(blwi->blwi_ns);
2967 if (blwi->blwi_mem_pressure)
2968 memalloc_noreclaim_restore(mpflags);
2970 if (blwi->blwi_flags & LCF_ASYNC)
2971 OBD_FREE(blwi, sizeof(*blwi));
2973 complete(&blwi->blwi_comp);
2979 * Cancel stale locks on export. Cancel blocked locks first.
2980 * If the given export has blocked locks, the next in the list may have
2981 * them too, thus cancel not blocked locks only if the current export has
2984 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2985 struct obd_export *exp)
2991 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2993 num = ldlm_export_cancel_blocked_locks(exp);
2995 ldlm_export_cancel_locks(exp);
2997 obd_stale_export_put(exp);
3004 * Main blocking requests processing thread.
3006 * Callers put locks into its queue by calling ldlm_bl_to_thread.
3007 * This thread in the end ends up doing actual call to ->l_blocking_ast
3010 static int ldlm_bl_thread_main(void *arg)
3013 struct ldlm_bl_pool *blp;
3014 struct ldlm_bl_thread_data *bltd = arg;
3022 rc = lu_env_init(env, LCT_DT_THREAD);
3025 rc = lu_env_add(env);
3027 GOTO(out_env_fini, rc);
3029 blp = bltd->bltd_blp;
3031 complete(&bltd->bltd_comp);
3032 /* cannot use bltd after this, it is only on caller's stack */
3035 struct ldlm_bl_work_item *blwi = NULL;
3036 struct obd_export *exp = NULL;
3039 rc = ldlm_bl_get_work(blp, &blwi, &exp);
3042 wait_event_idle_exclusive(blp->blp_waitq,
3043 ldlm_bl_get_work(blp, &blwi,
3045 atomic_inc(&blp->blp_busy_threads);
3047 if (ldlm_bl_thread_need_create(blp, blwi))
3048 /* discard the return value, we tried */
3049 ldlm_bl_thread_start(blp, true);
3052 rc = ldlm_bl_thread_exports(blp, exp);
3054 rc = ldlm_bl_thread_blwi(blp, blwi);
3056 atomic_dec(&blp->blp_busy_threads);
3058 if (rc == LDLM_ITER_STOP)
3062 * If there are many namespaces, we will not sleep waiting for
3063 * work, and must do a cond_resched to avoid holding the CPU
3069 atomic_dec(&blp->blp_num_threads);
3070 complete(&blp->blp_comp);
3081 static int ldlm_setup(void);
3082 static int ldlm_cleanup(void);
3084 int ldlm_get_ref(void)
3089 mutex_lock(&ldlm_ref_mutex);
3090 if (++ldlm_refcount == 1) {
3095 mutex_unlock(&ldlm_ref_mutex);
3100 void ldlm_put_ref(void)
3103 mutex_lock(&ldlm_ref_mutex);
3104 if (ldlm_refcount == 1) {
3105 int rc = ldlm_cleanup();
3108 CERROR("ldlm_cleanup failed: %d\n", rc);
3114 mutex_unlock(&ldlm_ref_mutex);
3120 * Export handle<->lock hash operations.
3123 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key,
3124 const unsigned int bits)
3126 return cfs_hash_64(((struct lustre_handle *)key)->cookie, bits);
3130 ldlm_export_lock_key(struct hlist_node *hnode)
3132 struct ldlm_lock *lock;
3134 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3135 return &lock->l_remote_handle;
3139 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
3141 struct ldlm_lock *lock;
3143 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3144 lock->l_remote_handle = *(struct lustre_handle *)key;
3148 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
3150 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
3154 ldlm_export_lock_object(struct hlist_node *hnode)
3156 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3160 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
3162 struct ldlm_lock *lock;
3164 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3165 LDLM_LOCK_GET(lock);
3169 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
3171 struct ldlm_lock *lock;
3173 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
3174 LDLM_LOCK_RELEASE(lock);
3177 static struct cfs_hash_ops ldlm_export_lock_ops = {
3178 .hs_hash = ldlm_export_lock_hash,
3179 .hs_key = ldlm_export_lock_key,
3180 .hs_keycmp = ldlm_export_lock_keycmp,
3181 .hs_keycpy = ldlm_export_lock_keycpy,
3182 .hs_object = ldlm_export_lock_object,
3183 .hs_get = ldlm_export_lock_get,
3184 .hs_put = ldlm_export_lock_put,
3185 .hs_put_locked = ldlm_export_lock_put,
3188 int ldlm_init_export(struct obd_export *exp)
3194 exp->exp_lock_hash =
3195 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
3196 HASH_EXP_LOCK_CUR_BITS,
3197 HASH_EXP_LOCK_MAX_BITS,
3198 HASH_EXP_LOCK_BKT_BITS, 0,
3199 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
3200 &ldlm_export_lock_ops,
3201 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
3202 CFS_HASH_NBLK_CHANGE);
3204 if (!exp->exp_lock_hash)
3207 rc = ldlm_init_flock_export(exp);
3213 ldlm_destroy_export(exp);
3216 EXPORT_SYMBOL(ldlm_init_export);
3218 void ldlm_destroy_export(struct obd_export *exp)
3221 cfs_hash_putref(exp->exp_lock_hash);
3222 exp->exp_lock_hash = NULL;
3224 ldlm_destroy_flock_export(exp);
3227 EXPORT_SYMBOL(ldlm_destroy_export);
3229 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
3230 struct attribute *attr,
3233 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
3236 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
3237 struct attribute *attr,
3244 rc = kstrtoul(buffer, 10, &val);
3248 ldlm_cancel_unused_locks_before_replay = val;
3252 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
3254 static struct attribute *ldlm_attrs[] = {
3255 &lustre_attr_cancel_unused_locks_before_replay.attr,
3259 static struct attribute_group ldlm_attr_group = {
3260 .attrs = ldlm_attrs,
3263 static int ldlm_setup(void)
3265 static struct ptlrpc_service_conf conf;
3266 struct ldlm_bl_pool *blp = NULL;
3267 #ifdef HAVE_SERVER_SUPPORT
3268 struct task_struct *task;
3269 #endif /* HAVE_SERVER_SUPPORT */
3275 if (ldlm_state != NULL)
3278 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
3279 if (ldlm_state == NULL)
3282 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
3286 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
3290 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
3294 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
3298 rc = ldlm_debugfs_setup();
3302 memset(&conf, 0, sizeof(conf));
3303 conf = (typeof(conf)) {
3304 .psc_name = "ldlm_cbd",
3305 .psc_watchdog_factor = 2,
3307 .bc_nbufs = LDLM_CLIENT_NBUFS,
3308 .bc_buf_size = LDLM_BUFSIZE,
3309 .bc_req_max_size = LDLM_MAXREQSIZE,
3310 .bc_rep_max_size = LDLM_MAXREPSIZE,
3311 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
3312 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
3315 .tc_thr_name = "ldlm_cb",
3316 .tc_thr_factor = LDLM_THR_FACTOR,
3317 .tc_nthrs_init = LDLM_NTHRS_INIT,
3318 .tc_nthrs_base = LDLM_NTHRS_BASE,
3319 .tc_nthrs_max = LDLM_NTHRS_MAX,
3320 .tc_nthrs_user = ldlm_num_threads,
3321 .tc_cpu_bind = ldlm_cpu_bind,
3322 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
3325 .cc_pattern = ldlm_cpts,
3326 .cc_affinity = true,
3329 .so_req_handler = ldlm_callback_handler,
3332 ldlm_state->ldlm_cb_service = \
3333 ptlrpc_register_service(&conf, ldlm_svc_kset,
3334 ldlm_svc_debugfs_dir);
3335 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
3336 CERROR("failed to start service\n");
3337 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
3338 ldlm_state->ldlm_cb_service = NULL;
3342 #ifdef HAVE_SERVER_SUPPORT
3343 memset(&conf, 0, sizeof(conf));
3344 conf = (typeof(conf)) {
3345 .psc_name = "ldlm_canceld",
3346 .psc_watchdog_factor = 6,
3348 .bc_nbufs = LDLM_SERVER_NBUFS,
3349 .bc_buf_size = LDLM_BUFSIZE,
3350 .bc_req_max_size = LDLM_MAXREQSIZE,
3351 .bc_rep_max_size = LDLM_MAXREPSIZE,
3352 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
3353 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
3357 .tc_thr_name = "ldlm_cn",
3358 .tc_thr_factor = LDLM_THR_FACTOR,
3359 .tc_nthrs_init = LDLM_NTHRS_INIT,
3360 .tc_nthrs_base = LDLM_NTHRS_BASE,
3361 .tc_nthrs_max = LDLM_NTHRS_MAX,
3362 .tc_nthrs_user = ldlm_num_threads,
3363 .tc_cpu_bind = ldlm_cpu_bind,
3364 .tc_ctx_tags = LCT_MD_THREAD | \
3369 .cc_pattern = ldlm_cpts,
3370 .cc_affinity = true,
3373 .so_req_handler = ldlm_cancel_handler,
3374 .so_hpreq_handler = ldlm_hpreq_handler,
3377 ldlm_state->ldlm_cancel_service = \
3378 ptlrpc_register_service(&conf, ldlm_svc_kset,
3379 ldlm_svc_debugfs_dir);
3380 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3381 CERROR("failed to start service\n");
3382 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3383 ldlm_state->ldlm_cancel_service = NULL;
3386 #endif /* HAVE_SERVER_SUPPORT */
3388 OBD_ALLOC(blp, sizeof(*blp));
3390 GOTO(out, rc = -ENOMEM);
3391 ldlm_state->ldlm_bl_pool = blp;
3393 spin_lock_init(&blp->blp_lock);
3394 INIT_LIST_HEAD(&blp->blp_list);
3395 INIT_LIST_HEAD(&blp->blp_prio_list);
3396 init_waitqueue_head(&blp->blp_waitq);
3397 atomic_set(&blp->blp_num_threads, 0);
3398 atomic_set(&blp->blp_busy_threads, 0);
3399 blp->blp_total_locks = 0;
3400 blp->blp_total_blwis = 0;
3402 if (ldlm_num_threads == 0) {
3403 blp->blp_min_threads = LDLM_NTHRS_INIT;
3404 blp->blp_max_threads = LDLM_NTHRS_MAX;
3406 blp->blp_min_threads = blp->blp_max_threads = \
3407 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3411 for (i = 0; i < blp->blp_min_threads; i++) {
3412 rc = ldlm_bl_thread_start(blp, false);
3417 #ifdef HAVE_SERVER_SUPPORT
3418 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3421 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3425 wait_event(expired_lock_wait_queue,
3426 expired_lock_thread_state == ELT_READY);
3427 #endif /* HAVE_SERVER_SUPPORT */
3429 rc = ldlm_pools_init();
3431 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3435 rc = ldlm_reclaim_setup();
3437 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3447 static int ldlm_cleanup(void)
3451 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3452 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3453 CERROR("ldlm still has namespaces; clean these up first.\n");
3454 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3455 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3459 ldlm_reclaim_cleanup();
3462 if (ldlm_state->ldlm_bl_pool != NULL) {
3463 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3465 while (atomic_read(&blp->blp_num_threads) > 0) {
3466 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3468 init_completion(&blp->blp_comp);
3470 spin_lock(&blp->blp_lock);
3471 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3472 wake_up(&blp->blp_waitq);
3473 spin_unlock(&blp->blp_lock);
3475 wait_for_completion(&blp->blp_comp);
3478 OBD_FREE(blp, sizeof(*blp));
3481 if (ldlm_state->ldlm_cb_service != NULL)
3482 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3483 #ifdef HAVE_SERVER_SUPPORT
3484 if (ldlm_state->ldlm_cancel_service != NULL)
3485 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3489 kset_unregister(ldlm_ns_kset);
3491 kset_unregister(ldlm_svc_kset);
3493 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3494 kobject_put(ldlm_kobj);
3497 ldlm_debugfs_cleanup();
3499 #ifdef HAVE_SERVER_SUPPORT
3500 if (expired_lock_thread_state != ELT_STOPPED) {
3501 expired_lock_thread_state = ELT_TERMINATE;
3502 wake_up(&expired_lock_wait_queue);
3503 wait_event(expired_lock_wait_queue,
3504 expired_lock_thread_state == ELT_STOPPED);
3508 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3516 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3517 sizeof(struct ldlm_resource), 0,
3518 SLAB_HWCACHE_ALIGN, NULL);
3519 if (ldlm_resource_slab == NULL)
3522 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3523 sizeof(struct ldlm_lock), 0,
3524 SLAB_HWCACHE_ALIGN, NULL);
3525 if (ldlm_lock_slab == NULL)
3528 ldlm_interval_slab = kmem_cache_create("interval_node",
3529 sizeof(struct ldlm_interval),
3530 0, SLAB_HWCACHE_ALIGN, NULL);
3531 if (ldlm_interval_slab == NULL)
3534 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3535 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3536 0, SLAB_HWCACHE_ALIGN, NULL);
3537 if (ldlm_interval_tree_slab == NULL)
3540 #ifdef HAVE_SERVER_SUPPORT
3541 ldlm_inodebits_slab = kmem_cache_create("ldlm_ibits_node",
3542 sizeof(struct ldlm_ibits_node),
3543 0, SLAB_HWCACHE_ALIGN, NULL);
3544 if (ldlm_inodebits_slab == NULL)
3545 goto out_interval_tree;
3547 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3548 sizeof(struct ldlm_glimpse_work),
3550 if (ldlm_glimpse_work_kmem == NULL)
3554 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3555 class_export_dump_hook = ldlm_dump_export_locks;
3558 #ifdef HAVE_SERVER_SUPPORT
3560 kmem_cache_destroy(ldlm_inodebits_slab);
3562 kmem_cache_destroy(ldlm_interval_tree_slab);
3565 kmem_cache_destroy(ldlm_interval_slab);
3567 kmem_cache_destroy(ldlm_lock_slab);
3569 kmem_cache_destroy(ldlm_resource_slab);
3574 void ldlm_exit(void)
3577 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3579 kmem_cache_destroy(ldlm_resource_slab);
3581 * ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3582 * rcu_barrier() to wait all outstanding RCU callbacks to complete,
3583 * so that ldlm_lock_free() get a chance to be called.
3586 kmem_cache_destroy(ldlm_lock_slab);
3587 kmem_cache_destroy(ldlm_interval_slab);
3588 kmem_cache_destroy(ldlm_interval_tree_slab);
3589 #ifdef HAVE_SERVER_SUPPORT
3590 kmem_cache_destroy(ldlm_inodebits_slab);
3591 kmem_cache_destroy(ldlm_glimpse_work_kmem);