4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lockd.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <linux/kthread.h>
41 #include <linux/list.h>
42 #include <libcfs/libcfs.h>
43 #include <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 char *ldlm_cpts;
53 module_param(ldlm_cpts, charp, 0444);
54 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
56 static DEFINE_MUTEX(ldlm_ref_mutex);
57 static int ldlm_refcount;
59 struct kobject *ldlm_kobj;
60 struct kset *ldlm_ns_kset;
61 struct kset *ldlm_svc_kset;
65 static struct ldlm_state *ldlm_state;
67 /* timeout for initial callback (AST) reply (bz10399)
68 * Due to having to send a 32 bit time value over the
69 * wire return it as time_t instead of time64_t
71 static inline time_t ldlm_get_rq_timeout(void)
74 time_t timeout = min(ldlm_timeout, obd_timeout / 3);
76 return timeout < 1 ? 1 : timeout;
83 * blp_prio_list is used for callbacks that should be handled
84 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
87 struct list_head blp_prio_list;
90 * blp_list is used for all other callbacks which are likely
91 * to take longer to process.
93 struct list_head blp_list;
95 wait_queue_head_t blp_waitq;
96 struct completion blp_comp;
97 atomic_t blp_num_threads;
98 atomic_t blp_busy_threads;
103 struct ldlm_bl_work_item {
104 struct list_head blwi_entry;
105 struct ldlm_namespace *blwi_ns;
106 struct ldlm_lock_desc blwi_ld;
107 struct ldlm_lock *blwi_lock;
108 struct list_head blwi_head;
110 struct completion blwi_comp;
111 enum ldlm_cancel_flags blwi_flags;
112 int blwi_mem_pressure;
115 #ifdef HAVE_SERVER_SUPPORT
118 * Protects both waiting_locks_list and expired_lock_thread.
120 static DEFINE_SPINLOCK(waiting_locks_spinlock); /* BH lock (timer) */
123 * List for contended locks.
125 * As soon as a lock is contended, it gets placed on this list and
126 * expected time to get a response is filled in the lock. A special
127 * thread walks the list looking for locks that should be released and
128 * schedules client evictions for those that have not been released in
131 * All access to it should be under waiting_locks_spinlock.
133 static LIST_HEAD(waiting_locks_list);
134 static void waiting_locks_callback(cfs_timer_cb_arg_t unused);
135 static CFS_DEFINE_TIMER(waiting_locks_timer, waiting_locks_callback, 0, 0);
143 static DECLARE_WAIT_QUEUE_HEAD(expired_lock_wait_queue);
144 static enum elt_state expired_lock_thread_state = ELT_STOPPED;
145 static int expired_lock_dump;
146 static LIST_HEAD(expired_lock_list);
148 static int ldlm_lock_busy(struct ldlm_lock *lock);
149 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout);
150 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds);
152 static inline int have_expired_locks(void)
157 spin_lock_bh(&waiting_locks_spinlock);
158 need_to_run = !list_empty(&expired_lock_list);
159 spin_unlock_bh(&waiting_locks_spinlock);
165 * Check expired lock list for expired locks and time them out.
167 static int expired_lock_main(void *arg)
169 struct list_head *expired = &expired_lock_list;
170 struct l_wait_info lwi = { 0 };
175 expired_lock_thread_state = ELT_READY;
176 wake_up(&expired_lock_wait_queue);
179 l_wait_event(expired_lock_wait_queue,
180 have_expired_locks() ||
181 expired_lock_thread_state == ELT_TERMINATE,
184 spin_lock_bh(&waiting_locks_spinlock);
185 if (expired_lock_dump) {
186 spin_unlock_bh(&waiting_locks_spinlock);
188 /* from waiting_locks_callback, but not in timer */
189 libcfs_debug_dumplog();
191 spin_lock_bh(&waiting_locks_spinlock);
192 expired_lock_dump = 0;
197 while (!list_empty(expired)) {
198 struct obd_export *export;
199 struct ldlm_lock *lock;
201 lock = list_entry(expired->next, struct ldlm_lock,
203 if ((void *)lock < LP_POISON + PAGE_SIZE &&
204 (void *)lock >= LP_POISON) {
205 spin_unlock_bh(&waiting_locks_spinlock);
206 CERROR("free lock on elt list %p\n", lock);
209 list_del_init(&lock->l_pending_chain);
210 if ((void *)lock->l_export <
211 LP_POISON + PAGE_SIZE &&
212 (void *)lock->l_export >= LP_POISON) {
213 CERROR("lock with free export on elt list %p\n",
215 lock->l_export = NULL;
216 LDLM_ERROR(lock, "free export");
217 /* release extra ref grabbed by
218 * ldlm_add_waiting_lock() or
219 * ldlm_failed_ast() */
220 LDLM_LOCK_RELEASE(lock);
224 if (ldlm_is_destroyed(lock)) {
225 /* release the lock refcount where
226 * waiting_locks_callback() founds */
227 LDLM_LOCK_RELEASE(lock);
230 export = class_export_lock_get(lock->l_export, lock);
231 spin_unlock_bh(&waiting_locks_spinlock);
233 /* Check if we need to prolong timeout */
234 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
235 lock->l_callback_timeout != 0 && /* not AST error */
236 ldlm_lock_busy(lock)) {
237 LDLM_DEBUG(lock, "prolong the busy lock");
238 lock_res_and_lock(lock);
239 ldlm_add_waiting_lock(lock,
240 ldlm_bl_timeout(lock) >> 1);
241 unlock_res_and_lock(lock);
243 spin_lock_bh(&export->exp_bl_list_lock);
244 list_del_init(&lock->l_exp_list);
245 spin_unlock_bh(&export->exp_bl_list_lock);
248 "lock callback timer expired after %llds: evicting client at %s ",
249 ktime_get_real_seconds() -
251 obd_export_nid2str(export));
252 ldlm_lock_to_ns(lock)->ns_timeouts++;
254 class_fail_export(export);
256 class_export_lock_put(export, lock);
257 /* release extra ref grabbed by ldlm_add_waiting_lock()
258 * or ldlm_failed_ast() */
259 LDLM_LOCK_RELEASE(lock);
261 spin_lock_bh(&waiting_locks_spinlock);
263 spin_unlock_bh(&waiting_locks_spinlock);
265 if (do_dump && obd_dump_on_eviction) {
266 CERROR("dump the log upon eviction\n");
267 libcfs_debug_dumplog();
270 if (expired_lock_thread_state == ELT_TERMINATE)
274 expired_lock_thread_state = ELT_STOPPED;
275 wake_up(&expired_lock_wait_queue);
280 * Check if there is a request in the export request list
281 * which prevents the lock canceling.
283 static int ldlm_lock_busy(struct ldlm_lock *lock)
285 struct ptlrpc_request *req;
289 if (lock->l_export == NULL)
292 spin_lock(&lock->l_export->exp_rpc_lock);
293 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
295 if (req->rq_ops->hpreq_lock_match) {
296 match = req->rq_ops->hpreq_lock_match(req, lock);
301 spin_unlock(&lock->l_export->exp_rpc_lock);
305 /* This is called from within a timer interrupt and cannot schedule */
306 static void waiting_locks_callback(cfs_timer_cb_arg_t unused)
308 struct ldlm_lock *lock;
311 spin_lock_bh(&waiting_locks_spinlock);
312 while (!list_empty(&waiting_locks_list)) {
313 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
315 if (lock->l_callback_timeout > ktime_get_seconds() ||
316 lock->l_req_mode == LCK_GROUP)
319 /* no needs to take an extra ref on the lock since it was in
320 * the waiting_locks_list and ldlm_add_waiting_lock()
321 * already grabbed a ref */
322 list_del(&lock->l_pending_chain);
323 list_add(&lock->l_pending_chain, &expired_lock_list);
327 if (!list_empty(&expired_lock_list)) {
328 if (obd_dump_on_timeout && need_dump)
329 expired_lock_dump = __LINE__;
331 wake_up(&expired_lock_wait_queue);
335 * Make sure the timer will fire again if we have any locks
338 if (!list_empty(&waiting_locks_list)) {
339 unsigned long timeout_jiffies;
341 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
343 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
344 mod_timer(&waiting_locks_timer, timeout_jiffies);
346 spin_unlock_bh(&waiting_locks_spinlock);
350 * Add lock to the list of contended locks.
352 * Indicate that we're waiting for a client to call us back cancelling a given
353 * lock. We add it to the pending-callback chain, and schedule the lock-timeout
354 * timer to fire appropriately. (We round up to the next second, to avoid
355 * floods of timer firings during periods of high lock contention and traffic).
356 * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
357 * if it has been added to the waiting list (1 is returned).
359 * Called with the namespace lock held.
361 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t seconds)
363 unsigned long timeout_jiffies;
366 if (!list_empty(&lock->l_pending_chain))
369 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
370 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
373 timeout = ktime_get_seconds() + seconds;
374 if (likely(timeout > lock->l_callback_timeout))
375 lock->l_callback_timeout = timeout;
377 timeout_jiffies = cfs_time_seconds(lock->l_callback_timeout);
379 if (time_before(timeout_jiffies, waiting_locks_timer.expires) ||
380 !timer_pending(&waiting_locks_timer))
381 mod_timer(&waiting_locks_timer, timeout_jiffies);
383 /* if the new lock has a shorter timeout than something earlier on
384 * the list, we'll wait the longer amount of time; no big deal.
387 list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
391 static void ldlm_add_blocked_lock(struct ldlm_lock *lock)
393 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
394 if (list_empty(&lock->l_exp_list)) {
395 if (lock->l_granted_mode != lock->l_req_mode)
396 list_add_tail(&lock->l_exp_list,
397 &lock->l_export->exp_bl_list);
399 list_add(&lock->l_exp_list,
400 &lock->l_export->exp_bl_list);
402 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
404 /* A blocked lock is added. Adjust the position in
405 * the stale list if the export is in the list.
406 * If export is stale and not in the list - it is being
407 * processed and will be placed on the right position
408 * on obd_stale_export_put(). */
409 if (!list_empty(&lock->l_export->exp_stale_list))
410 obd_stale_export_adjust(lock->l_export);
413 static int ldlm_add_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
417 /* NB: must be called with hold of lock_res_and_lock() */
418 LASSERT(ldlm_is_res_locked(lock));
419 LASSERT(!ldlm_is_cancel_on_block(lock));
421 /* Do not put cross-MDT lock in the waiting list, since we
422 * will not evict it due to timeout for now */
423 if (lock->l_export != NULL &&
424 (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS))
427 spin_lock_bh(&waiting_locks_spinlock);
428 if (ldlm_is_cancel(lock)) {
429 spin_unlock_bh(&waiting_locks_spinlock);
433 if (ldlm_is_destroyed(lock)) {
434 static time64_t next;
436 spin_unlock_bh(&waiting_locks_spinlock);
437 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
438 if (ktime_get_seconds() > next) {
439 next = ktime_get_seconds() + 14400;
440 libcfs_debug_dumpstack(NULL);
445 ldlm_set_waited(lock);
446 lock->l_blast_sent = ktime_get_real_seconds();
447 ret = __ldlm_add_waiting_lock(lock, timeout);
449 /* grab ref on the lock if it has been added to the
453 spin_unlock_bh(&waiting_locks_spinlock);
456 ldlm_add_blocked_lock(lock);
458 LDLM_DEBUG(lock, "%sadding to wait list(timeout: %lld, AT: %s)",
459 ret == 0 ? "not re-" : "", timeout,
460 AT_OFF ? "off" : "on");
465 * Remove a lock from the pending list, likely because it had its cancellation
466 * callback arrive without incident. This adjusts the lock-timeout timer if
467 * needed. Returns 0 if the lock wasn't pending after all, 1 if it was.
468 * As done by ldlm_del_waiting_lock(), the caller must release the lock
469 * reference when the lock is removed from any list (1 is returned).
471 * Called with namespace lock held.
473 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
475 struct list_head *list_next;
477 if (list_empty(&lock->l_pending_chain))
480 list_next = lock->l_pending_chain.next;
481 if (lock->l_pending_chain.prev == &waiting_locks_list) {
482 /* Removing the head of the list, adjust timer. */
483 if (list_next == &waiting_locks_list) {
484 /* No more, just cancel. */
485 del_timer(&waiting_locks_timer);
487 struct ldlm_lock *next;
489 next = list_entry(list_next, struct ldlm_lock,
491 mod_timer(&waiting_locks_timer,
492 cfs_time_seconds(next->l_callback_timeout));
495 list_del_init(&lock->l_pending_chain);
500 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
504 if (lock->l_export == NULL) {
505 /* We don't have a "waiting locks list" on clients. */
506 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
510 spin_lock_bh(&waiting_locks_spinlock);
511 ret = __ldlm_del_waiting_lock(lock);
512 ldlm_clear_waited(lock);
513 spin_unlock_bh(&waiting_locks_spinlock);
515 /* remove the lock out of export blocking list */
516 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
517 list_del_init(&lock->l_exp_list);
518 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
521 /* release lock ref if it has indeed been removed
523 LDLM_LOCK_RELEASE(lock);
526 LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
531 * Prolong the contended lock waiting time.
533 * Called with namespace lock held.
535 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
537 if (lock->l_export == NULL) {
538 /* We don't have a "waiting locks list" on clients. */
539 LDLM_DEBUG(lock, "client lock: no-op");
543 if (exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) {
544 /* We don't have a "waiting locks list" on OSP. */
545 LDLM_DEBUG(lock, "MDS-MDS lock: no-op");
549 spin_lock_bh(&waiting_locks_spinlock);
551 if (list_empty(&lock->l_pending_chain)) {
552 spin_unlock_bh(&waiting_locks_spinlock);
553 LDLM_DEBUG(lock, "wasn't waiting");
557 /* we remove/add the lock to the waiting list, so no needs to
558 * release/take a lock reference */
559 __ldlm_del_waiting_lock(lock);
560 __ldlm_add_waiting_lock(lock, timeout);
561 spin_unlock_bh(&waiting_locks_spinlock);
563 LDLM_DEBUG(lock, "refreshed");
566 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
568 #else /* HAVE_SERVER_SUPPORT */
570 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
575 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, time64_t timeout)
580 #endif /* !HAVE_SERVER_SUPPORT */
582 #ifdef HAVE_SERVER_SUPPORT
585 * Calculate the per-export Blocking timeout (covering BL AST, data flush,
586 * lock cancel, and their replies). Used for lock callback timeout and AST
589 * \param[in] lock lock which is getting the blocking callback
591 * \retval timeout in seconds to wait for the client reply
593 time64_t ldlm_bl_timeout(struct ldlm_lock *lock)
598 return obd_timeout / 2;
600 /* Since these are non-updating timeouts, we should be conservative.
601 * Take more than usually, 150%
602 * It would be nice to have some kind of "early reply" mechanism for
603 * lock callbacks too... */
604 timeout = at_get(&lock->l_export->exp_bl_lock_at);
605 return max(timeout + (timeout >> 1), (time64_t)ldlm_enqueue_min);
607 EXPORT_SYMBOL(ldlm_bl_timeout);
610 * Perform lock cleanup if AST sending failed.
612 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
613 const char *ast_type)
615 LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
616 "to a lock %s callback time out: rc %d\n",
617 lock->l_export->exp_obd->obd_name,
618 obd_export_nid2str(lock->l_export), ast_type, rc);
620 if (obd_dump_on_timeout)
621 libcfs_debug_dumplog();
622 spin_lock_bh(&waiting_locks_spinlock);
623 if (__ldlm_del_waiting_lock(lock) == 0)
624 /* the lock was not in any list, grab an extra ref before adding
625 * the lock to the expired list */
627 lock->l_callback_timeout = 0; /* differentiate it from expired locks */
628 list_add(&lock->l_pending_chain, &expired_lock_list);
629 wake_up(&expired_lock_wait_queue);
630 spin_unlock_bh(&waiting_locks_spinlock);
634 * Perform lock cleanup if AST reply came with error.
636 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
637 struct ptlrpc_request *req, int rc,
638 const char *ast_type)
640 struct lnet_process_id peer = req->rq_import->imp_connection->c_peer;
642 if (!req->rq_replied || (rc && rc != -EINVAL)) {
643 if (ldlm_is_cancel(lock)) {
645 "%s AST (req@%p x%llu) timeout from nid %s, but cancel was received (AST reply lost?)",
646 ast_type, req, req->rq_xid,
647 libcfs_nid2str(peer.nid));
648 ldlm_lock_cancel(lock);
650 } else if (rc == -ENODEV || rc == -ESHUTDOWN ||
652 req->rq_import->imp_state == LUSTRE_IMP_CLOSED)) {
653 /* Upon umount process the AST fails because cannot be
654 * sent. This shouldn't lead to the client eviction.
655 * -ENODEV error is returned by ptl_send_rpc() for
656 * new request in such import.
657 * -SHUTDOWN is returned by ptlrpc_import_delay_req()
658 * if imp_invalid is set or obd_no_recov.
659 * Meanwhile there is also check for LUSTRE_IMP_CLOSED
660 * in ptlrpc_import_delay_req() as well with -EIO code.
661 * In all such cases errors are ignored.
663 LDLM_DEBUG(lock, "%s AST can't be sent due to a server"
664 " %s failure or umount process: rc = %d\n",
666 req->rq_import->imp_obd->obd_name, rc);
669 "client (nid %s) %s %s AST (req@%p x%llu status %d rc %d), evict it",
670 libcfs_nid2str(peer.nid),
671 req->rq_replied ? "returned error from" :
672 "failed to reply to",
673 ast_type, req, req->rq_xid,
674 (req->rq_repmsg != NULL) ?
675 lustre_msg_get_status(req->rq_repmsg) : 0,
677 ldlm_failed_ast(lock, rc, ast_type);
683 struct ldlm_resource *res = lock->l_resource;
686 "client (nid %s) returned %d from %s AST (req@%p x%llu) - normal race",
687 libcfs_nid2str(peer.nid),
689 lustre_msg_get_status(req->rq_repmsg) : -1,
690 ast_type, req, req->rq_xid);
692 /* update lvbo to return proper attributes.
694 ldlm_resource_getref(res);
695 ldlm_lvbo_update(res, lock, NULL, 1);
696 ldlm_resource_putref(res);
698 ldlm_lock_cancel(lock);
705 static int ldlm_cb_interpret(const struct lu_env *env,
706 struct ptlrpc_request *req, void *data, int rc)
708 struct ldlm_cb_async_args *ca = data;
709 struct ldlm_lock *lock = ca->ca_lock;
710 struct ldlm_cb_set_arg *arg = ca->ca_set_arg;
713 LASSERT(lock != NULL);
716 case LDLM_GL_CALLBACK:
717 /* Update the LVB from disk if the AST failed
718 * (this is a legal race)
720 * - Glimpse callback of local lock just returns
721 * -ELDLM_NO_LOCK_DATA.
722 * - Glimpse callback of remote lock might return
723 * -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
725 if (unlikely(arg->gl_interpret_reply)) {
726 rc = arg->gl_interpret_reply(env, req, data, rc);
727 } else if (rc == -ELDLM_NO_LOCK_DATA) {
728 LDLM_DEBUG(lock, "lost race - client has a lock but no "
730 ldlm_lvbo_update(lock->l_resource, lock, NULL, 1);
731 } else if (rc != 0) {
732 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
734 rc = ldlm_lvbo_update(lock->l_resource, lock, req, 1);
737 case LDLM_BL_CALLBACK:
739 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
741 case LDLM_CP_CALLBACK:
743 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
746 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
751 /* release extra reference taken in ldlm_ast_fini() */
752 LDLM_LOCK_RELEASE(lock);
755 atomic_inc(&arg->restart);
760 static void ldlm_update_resend(struct ptlrpc_request *req, void *data)
762 struct ldlm_cb_async_args *ca = data;
763 struct ldlm_lock *lock = ca->ca_lock;
765 ldlm_refresh_waiting_lock(lock, ldlm_bl_timeout(lock));
768 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
769 struct ldlm_cb_set_arg *arg,
770 struct ldlm_lock *lock,
776 if (unlikely(instant_cancel)) {
777 rc = ptl_send_rpc(req, 1);
778 ptlrpc_req_finished(req);
780 atomic_inc(&arg->restart);
783 ptlrpc_set_add_req(arg->set, req);
790 * Check if there are requests in the export request list which prevent
791 * the lock canceling and make these requests high priority ones.
793 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
795 struct ptlrpc_request *req;
798 if (lock->l_export == NULL) {
799 LDLM_DEBUG(lock, "client lock: no-op");
803 spin_lock(&lock->l_export->exp_rpc_lock);
804 list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
806 /* Do not process requests that were not yet added to there
807 * incoming queue or were already removed from there for
808 * processing. We evaluate ptlrpc_nrs_req_can_move() without
809 * holding svcpt->scp_req_lock, and then redo the check with
810 * the lock held once we need to obtain a reliable result.
812 if (ptlrpc_nrs_req_can_move(req) &&
813 req->rq_ops->hpreq_lock_match &&
814 req->rq_ops->hpreq_lock_match(req, lock))
815 ptlrpc_nrs_req_hp_move(req);
817 spin_unlock(&lock->l_export->exp_rpc_lock);
822 * ->l_blocking_ast() method for server-side locks. This is invoked when newly
823 * enqueued server lock conflicts with given one.
825 * Sends blocking AST RPC to the client owning that lock; arms timeout timer
826 * to wait for client response.
828 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
829 struct ldlm_lock_desc *desc,
830 void *data, int flag)
832 struct ldlm_cb_async_args *ca;
833 struct ldlm_cb_set_arg *arg = data;
834 struct ldlm_request *body;
835 struct ptlrpc_request *req;
836 int instant_cancel = 0;
840 if (flag == LDLM_CB_CANCELING)
841 /* Don't need to do anything here. */
844 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_BL_AST)) {
845 LDLM_DEBUG(lock, "dropping BL AST");
850 LASSERT(data != NULL);
851 if (lock->l_export->exp_obd->obd_recovering != 0)
852 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
854 ldlm_lock_reorder_req(lock);
856 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
857 &RQF_LDLM_BL_CALLBACK,
858 LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
862 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
863 ca = ptlrpc_req_async_args(req);
864 ca->ca_set_arg = arg;
867 req->rq_interpret_reply = ldlm_cb_interpret;
869 lock_res_and_lock(lock);
870 if (ldlm_is_destroyed(lock)) {
871 /* What's the point? */
872 unlock_res_and_lock(lock);
873 ptlrpc_req_finished(req);
877 if (lock->l_granted_mode != lock->l_req_mode) {
878 /* this blocking AST will be communicated as part of the
879 * completion AST instead */
880 ldlm_add_blocked_lock(lock);
881 ldlm_set_waited(lock);
882 unlock_res_and_lock(lock);
884 ptlrpc_req_finished(req);
885 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
889 if (ldlm_is_cancel_on_block(lock))
892 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
893 body->lock_handle[0] = lock->l_remote_handle;
894 body->lock_desc = *desc;
895 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_FL_AST_MASK);
897 LDLM_DEBUG(lock, "server preparing blocking AST");
899 ptlrpc_request_set_replen(req);
900 ldlm_set_cbpending(lock);
901 if (instant_cancel) {
902 unlock_res_and_lock(lock);
903 ldlm_lock_cancel(lock);
905 req->rq_no_resend = 1;
907 LASSERT(lock->l_granted_mode == lock->l_req_mode);
908 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
909 unlock_res_and_lock(lock);
911 /* Do not resend after lock callback timeout */
912 req->rq_delay_limit = ldlm_bl_timeout(lock);
913 req->rq_resend_cb = ldlm_update_resend;
916 req->rq_send_state = LUSTRE_IMP_FULL;
917 /* ptlrpc_request_alloc_pack already set timeout */
919 req->rq_timeout = ldlm_get_rq_timeout();
921 if (lock->l_export && lock->l_export->exp_nid_stats &&
922 lock->l_export->exp_nid_stats->nid_ldlm_stats)
923 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
924 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
926 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
932 * ->l_completion_ast callback for a remote lock in server namespace.
934 * Sends AST to the client notifying it of lock granting. If initial
935 * lock response was not sent yet, instead of sending another RPC, just
936 * mark the lock as granted and client will understand
938 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
940 struct ldlm_cb_set_arg *arg = data;
941 struct ldlm_request *body;
942 struct ptlrpc_request *req;
943 struct ldlm_cb_async_args *ca;
944 int instant_cancel = 0;
949 LASSERT(lock != NULL);
950 LASSERT(data != NULL);
952 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
953 LDLM_DEBUG(lock, "dropping CP AST");
957 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
958 &RQF_LDLM_CP_CALLBACK);
962 /* server namespace, doesn't need lock */
963 lvb_len = ldlm_lvbo_size(lock);
964 /* LU-3124 & LU-2187: to not return layout in completion AST because
965 * it may deadlock for LU-2187, or client may not have enough space
966 * for large layout. The layout will be returned to client with an
967 * extra RPC to fetch xattr.lov */
968 if (ldlm_has_layout(lock))
971 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
972 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
974 ptlrpc_request_free(req);
978 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
979 ca = ptlrpc_req_async_args(req);
980 ca->ca_set_arg = arg;
983 req->rq_interpret_reply = ldlm_cb_interpret;
984 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
986 body->lock_handle[0] = lock->l_remote_handle;
987 body->lock_flags = ldlm_flags_to_wire(flags);
988 ldlm_lock2desc(lock, &body->lock_desc);
990 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
992 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
994 /* We still need to send the RPC to wake up the blocked
995 * enqueue thread on the client.
997 * Consider old client, there is no better way to notify
998 * the failure, just zero-sized the LVB, then the client
999 * will fail out as "-EPROTO". */
1000 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1004 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1009 LDLM_DEBUG(lock, "server preparing completion AST");
1011 ptlrpc_request_set_replen(req);
1013 req->rq_send_state = LUSTRE_IMP_FULL;
1014 /* ptlrpc_request_pack already set timeout */
1016 req->rq_timeout = ldlm_get_rq_timeout();
1018 /* We only send real blocking ASTs after the lock is granted */
1019 lock_res_and_lock(lock);
1020 if (ldlm_is_ast_sent(lock)) {
1021 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1022 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1023 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1026 /* We might get here prior to ldlm_handle_enqueue setting
1027 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1028 * into waiting list, but this is safe and similar code in
1029 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1030 * that would not only cancel the lock, but will also remove
1031 * it from waiting list */
1032 if (ldlm_is_cancel_on_block(lock)) {
1033 unlock_res_and_lock(lock);
1034 ldlm_lock_cancel(lock);
1037 req->rq_no_resend = 1;
1039 lock_res_and_lock(lock);
1041 /* start the lock-timeout clock */
1042 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1043 /* Do not resend after lock callback timeout */
1044 req->rq_delay_limit = ldlm_bl_timeout(lock);
1045 req->rq_resend_cb = ldlm_update_resend;
1048 unlock_res_and_lock(lock);
1050 if (lock->l_export && lock->l_export->exp_nid_stats &&
1051 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1052 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1053 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1055 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1057 RETURN(lvb_len < 0 ? lvb_len : rc);
1061 * Server side ->l_glimpse_ast handler for client locks.
1063 * Sends glimpse AST to the client and waits for reply. Then updates
1064 * lvbo with the result.
1066 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1068 struct ldlm_cb_set_arg *arg = data;
1069 struct ldlm_request *body;
1070 struct ptlrpc_request *req;
1071 struct ldlm_cb_async_args *ca;
1073 struct req_format *req_fmt;
1076 LASSERT(lock != NULL);
1078 if (arg->gl_desc != NULL)
1079 /* There is a glimpse descriptor to pack */
1080 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1082 req_fmt = &RQF_LDLM_GL_CALLBACK;
1084 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1085 req_fmt, LUSTRE_DLM_VERSION,
1091 if (arg->gl_desc != NULL) {
1092 /* copy the GL descriptor */
1093 union ldlm_gl_desc *desc;
1094 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1095 *desc = *arg->gl_desc;
1098 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1099 body->lock_handle[0] = lock->l_remote_handle;
1100 ldlm_lock2desc(lock, &body->lock_desc);
1102 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1103 ca = ptlrpc_req_async_args(req);
1104 ca->ca_set_arg = arg;
1107 /* server namespace, doesn't need lock */
1108 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1109 ldlm_lvbo_size(lock));
1110 ptlrpc_request_set_replen(req);
1112 req->rq_send_state = LUSTRE_IMP_FULL;
1113 /* ptlrpc_request_alloc_pack already set timeout */
1115 req->rq_timeout = ldlm_get_rq_timeout();
1117 req->rq_interpret_reply = ldlm_cb_interpret;
1119 if (lock->l_export && lock->l_export->exp_nid_stats &&
1120 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1121 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1122 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1124 rc = ldlm_ast_fini(req, arg, lock, 0);
1128 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1130 int ldlm_glimpse_locks(struct ldlm_resource *res,
1131 struct list_head *gl_work_list)
1136 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1138 if (rc == -ERESTART)
1139 ldlm_reprocess_all(res);
1143 EXPORT_SYMBOL(ldlm_glimpse_locks);
1145 /* return LDLM lock associated with a lock callback request */
1146 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1148 struct ldlm_cb_async_args *ca;
1149 struct ldlm_lock *lock;
1152 ca = ptlrpc_req_async_args(req);
1155 RETURN(ERR_PTR(-EFAULT));
1159 EXPORT_SYMBOL(ldlm_request_lock);
1162 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1163 * service threads to carry out client lock enqueueing requests.
1165 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1166 struct ptlrpc_request *req,
1167 const struct ldlm_request *dlm_req,
1168 const struct ldlm_callback_suite *cbs)
1170 struct ldlm_reply *dlm_rep;
1172 enum ldlm_error err = ELDLM_OK;
1173 struct ldlm_lock *lock = NULL;
1174 void *cookie = NULL;
1176 struct ldlm_resource *res = NULL;
1179 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1181 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1182 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1184 LASSERT(req->rq_export);
1186 /* for intent enqueue the stat will be updated inside intent policy */
1187 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1188 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1189 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1191 if (req->rq_export && req->rq_export->exp_nid_stats &&
1192 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1193 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1194 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1196 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1197 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1198 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1199 dlm_req->lock_desc.l_resource.lr_type);
1200 GOTO(out, rc = -EFAULT);
1203 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1204 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1205 dlm_req->lock_desc.l_req_mode &
1206 (dlm_req->lock_desc.l_req_mode-1))) {
1207 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1208 dlm_req->lock_desc.l_req_mode);
1209 GOTO(out, rc = -EFAULT);
1212 if (unlikely((flags & LDLM_FL_REPLAY) ||
1213 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1214 /* Find an existing lock in the per-export lock hash */
1215 /* In the function below, .hs_keycmp resolves to
1216 * ldlm_export_lock_keycmp() */
1217 /* coverity[overrun-buffer-val] */
1218 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1219 (void *)&dlm_req->lock_handle[0]);
1221 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie %#llx",
1222 lock->l_handle.h_cookie);
1223 flags |= LDLM_FL_RESENT;
1224 GOTO(existing_lock, rc = 0);
1227 if (ldlm_reclaim_full()) {
1228 DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1229 "reject current enqueue request and let the "
1230 "client retry later.\n");
1231 GOTO(out, rc = -EINPROGRESS);
1235 /* The lock's callback data might be set in the policy function */
1236 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1237 dlm_req->lock_desc.l_resource.lr_type,
1238 dlm_req->lock_desc.l_req_mode,
1239 cbs, NULL, 0, LVB_T_NONE);
1246 lock->l_remote_handle = dlm_req->lock_handle[0];
1247 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1249 /* Initialize resource lvb but not for a lock being replayed since
1250 * Client already got lvb sent in this case.
1251 * This must occur early since some policy methods assume resource
1252 * lvb is available (lr_lvb_data != NULL).
1254 res = lock->l_resource;
1255 if (!(flags & LDLM_FL_REPLAY)) {
1256 /* non-replayed lock, delayed lvb init may need to be done */
1257 rc = ldlm_lvbo_init(res);
1259 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1264 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1265 /* Don't enqueue a lock onto the export if it is been disonnected
1266 * due to eviction (bug 3822) or server umount (bug 24324).
1267 * Cancel it now instead. */
1268 if (req->rq_export->exp_disconnected) {
1269 LDLM_ERROR(lock, "lock on disconnected export %p",
1271 GOTO(out, rc = -ENOTCONN);
1274 lock->l_export = class_export_lock_get(req->rq_export, lock);
1275 if (lock->l_export->exp_lock_hash)
1276 cfs_hash_add(lock->l_export->exp_lock_hash,
1277 &lock->l_remote_handle,
1280 /* Inherit the enqueue flags before the operation, because we do not
1281 * keep the res lock on return and next operations (BL AST) may proceed
1283 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1284 LDLM_FL_INHERIT_MASK);
1286 ldlm_convert_policy_to_local(req->rq_export,
1287 dlm_req->lock_desc.l_resource.lr_type,
1288 &dlm_req->lock_desc.l_policy_data,
1289 &lock->l_policy_data);
1290 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1291 lock->l_req_extent = lock->l_policy_data.l_extent;
1294 if (flags & LDLM_FL_HAS_INTENT) {
1295 /* In this case, the reply buffer is allocated deep in
1296 * local_lock_enqueue by the policy function. */
1299 /* based on the assumption that lvb size never changes during
1300 * resource life time otherwise it need resource->lr_lock's
1302 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1303 RCL_SERVER, ldlm_lvbo_size(lock));
1305 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1306 GOTO(out, rc = -ENOMEM);
1308 rc = req_capsule_server_pack(&req->rq_pill);
1313 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1320 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1322 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1323 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1325 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1326 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1328 /* We never send a blocking AST until the lock is granted, but
1329 * we can tell it right now */
1330 lock_res_and_lock(lock);
1332 /* Now take into account flags to be inherited from original lock
1333 request both in reply to client and in our own lock flags. */
1334 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1335 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1337 /* Don't move a pending lock onto the export if it has already been
1338 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1339 * Cancel it now instead. */
1340 if (unlikely(req->rq_export->exp_disconnected ||
1341 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1342 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1344 } else if (ldlm_is_ast_sent(lock)) {
1345 /* fill lock desc for possible lock convert */
1346 if (lock->l_blocking_lock &&
1347 lock->l_resource->lr_type == LDLM_IBITS) {
1348 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1349 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1352 "save blocking bits %llx in granted lock",
1353 bl_lock->l_policy_data.l_inodebits.bits);
1354 /* If lock is blocked then save blocking ibits
1355 * in returned lock policy for the possible lock
1356 * convert on a client.
1358 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1359 bl_lock->l_policy_data.l_inodebits.bits;
1361 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1362 if (lock->l_granted_mode == lock->l_req_mode) {
1364 * Only cancel lock if it was granted, because it would
1365 * be destroyed immediately and would never be granted
1366 * in the future, causing timeouts on client. Not
1367 * granted lock will be cancelled immediately after
1368 * sending completion AST.
1370 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1371 unlock_res_and_lock(lock);
1372 ldlm_lock_cancel(lock);
1373 lock_res_and_lock(lock);
1375 ldlm_add_waiting_lock(lock,
1376 ldlm_bl_timeout(lock));
1380 unlock_res_and_lock(lock);
1384 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1385 if (!req->rq_packed_final) {
1386 err = lustre_pack_reply(req, 1, NULL, NULL);
1391 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1392 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1394 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1395 "(err=%d, rc=%d)", err, rc);
1398 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1400 ldlm_lvbo_size(lock) > 0) {
1404 buf = req_capsule_server_get(&req->rq_pill,
1406 LASSERTF(buf != NULL, "req %p, lock %p\n",
1408 buflen = req_capsule_get_size(&req->rq_pill,
1409 &RMF_DLM_LVB, RCL_SERVER);
1410 /* non-replayed lock, delayed lvb init may
1411 * need to be occur now */
1412 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1413 buflen = ldlm_lvbo_fill(lock, buf,
1419 buflen, RCL_SERVER);
1422 } else if (flags & LDLM_FL_REPLAY) {
1423 /* no LVB resend upon replay */
1437 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1438 if (lock->l_export) {
1439 ldlm_lock_cancel(lock);
1441 lock_res_and_lock(lock);
1442 ldlm_resource_unlink_lock(lock);
1443 ldlm_lock_destroy_nolock(lock);
1444 unlock_res_and_lock(lock);
1449 if (!err && !ldlm_is_cbpending(lock) &&
1450 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1451 ldlm_reprocess_all(lock->l_resource);
1453 LDLM_LOCK_RELEASE(lock);
1456 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1463 * Main LDLM entry point for server code to process lock conversion requests.
1465 int ldlm_handle_convert0(struct ptlrpc_request *req,
1466 const struct ldlm_request *dlm_req)
1468 struct obd_export *exp = req->rq_export;
1469 struct ldlm_reply *dlm_rep;
1470 struct ldlm_lock *lock;
1475 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1476 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1477 LDLM_CONVERT - LDLM_FIRST_OPC);
1479 rc = req_capsule_server_pack(&req->rq_pill);
1483 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1484 dlm_rep->lock_flags = dlm_req->lock_flags;
1486 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1491 bits = lock->l_policy_data.l_inodebits.bits;
1492 new = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1493 LDLM_DEBUG(lock, "server-side convert handler START");
1495 if (ldlm_is_cancel(lock)) {
1496 LDLM_ERROR(lock, "convert on canceled lock!");
1497 rc = ELDLM_NO_LOCK_DATA;
1498 } else if (dlm_req->lock_desc.l_req_mode !=
1499 lock->l_granted_mode) {
1500 LDLM_ERROR(lock, "lock mode differs!");
1501 rc = ELDLM_NO_LOCK_DATA;
1502 } else if (bits == new) {
1503 /* This can be valid situation if CONVERT RPCs are
1504 * re-ordered. Just finish silently */
1505 LDLM_DEBUG(lock, "lock is converted already!");
1508 lock_res_and_lock(lock);
1509 if (ldlm_is_waited(lock))
1510 ldlm_del_waiting_lock(lock);
1512 ldlm_clear_cbpending(lock);
1513 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1514 ldlm_inodebits_drop(lock, bits & ~new);
1515 lock->l_bl_ast_run = 0;
1516 ldlm_clear_ast_sent(lock);
1517 unlock_res_and_lock(lock);
1519 ldlm_reprocess_all(lock->l_resource);
1523 if (rc == ELDLM_OK) {
1524 dlm_rep->lock_handle = lock->l_remote_handle;
1525 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1526 &dlm_rep->lock_desc.l_policy_data);
1529 LDLM_DEBUG(lock, "server-side convert handler END, rc = %d",
1531 LDLM_LOCK_PUT(lock);
1533 rc = ELDLM_NO_LOCK_DATA;
1534 LDLM_DEBUG_NOLOCK("server-side convert handler END, rc = %d",
1538 req->rq_status = rc;
1544 * Cancel all the locks whose handles are packed into ldlm_request
1546 * Called by server code expecting such combined cancel activity
1549 int ldlm_request_cancel(struct ptlrpc_request *req,
1550 const struct ldlm_request *dlm_req,
1551 int first, enum lustre_at_flags flags)
1553 struct ldlm_resource *res, *pres = NULL;
1554 struct ldlm_lock *lock;
1555 int i, count, done = 0;
1558 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1562 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1565 /* There is no lock on the server at the replay time,
1566 * skip lock cancelling to make replay tests to pass. */
1567 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1570 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1571 "starting at %d", count, first);
1573 for (i = first; i < count; i++) {
1574 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1576 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1577 "lock (cookie %llu)",
1578 dlm_req->lock_handle[i].cookie);
1582 res = lock->l_resource;
1585 /* This code is an optimization to only attempt lock
1586 * granting on the resource (that could be CPU-expensive)
1587 * after we are done cancelling lock in that resource. */
1590 ldlm_reprocess_all(pres);
1591 LDLM_RESOURCE_DELREF(pres);
1592 ldlm_resource_putref(pres);
1595 ldlm_resource_getref(res);
1596 LDLM_RESOURCE_ADDREF(res);
1598 if (!ldlm_is_discard_data(lock))
1599 ldlm_lvbo_update(res, lock, NULL, 1);
1604 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock) &&
1605 lock->l_blast_sent != 0) {
1606 time64_t delay = ktime_get_real_seconds() -
1608 LDLM_DEBUG(lock, "server cancels blocked lock after %llds",
1610 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1612 ldlm_lock_cancel(lock);
1613 LDLM_LOCK_PUT(lock);
1616 ldlm_reprocess_all(pres);
1617 LDLM_RESOURCE_DELREF(pres);
1618 ldlm_resource_putref(pres);
1620 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1623 EXPORT_SYMBOL(ldlm_request_cancel);
1626 * Main LDLM entry point for server code to cancel locks.
1628 * Typically gets called from service handler on LDLM_CANCEL opc.
1630 int ldlm_handle_cancel(struct ptlrpc_request *req)
1632 struct ldlm_request *dlm_req;
1636 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1637 if (dlm_req == NULL) {
1638 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1642 if (req->rq_export && req->rq_export->exp_nid_stats &&
1643 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1644 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1645 LDLM_CANCEL - LDLM_FIRST_OPC);
1647 rc = req_capsule_server_pack(&req->rq_pill);
1651 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1652 req->rq_status = LUSTRE_ESTALE;
1654 RETURN(ptlrpc_reply(req));
1656 #endif /* HAVE_SERVER_SUPPORT */
1659 * Callback handler for receiving incoming blocking ASTs.
1661 * This can only happen on client side.
1663 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1664 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1669 LDLM_DEBUG(lock, "client blocking AST callback handler");
1671 lock_res_and_lock(lock);
1673 /* set bits to cancel for this lock for possible lock convert */
1674 if (ns_is_client(ns) && (lock->l_resource->lr_type == LDLM_IBITS)) {
1675 /* Lock description contains policy of blocking lock,
1676 * and its cancel_bits is used to pass conflicting bits.
1677 * NOTE: ld can be NULL or can be not NULL but zeroed if
1678 * passed from ldlm_bl_thread_blwi(), check below used bits
1679 * in ld to make sure it is valid description.
1681 if (ld && ld->l_policy_data.l_inodebits.bits)
1682 lock->l_policy_data.l_inodebits.cancel_bits =
1683 ld->l_policy_data.l_inodebits.cancel_bits;
1684 /* if there is no valid ld and lock is cbpending already
1685 * then cancel_bits should be kept, otherwise it is zeroed.
1687 else if (!ldlm_is_cbpending(lock))
1688 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1690 ldlm_set_cbpending(lock);
1692 if (ldlm_is_cancel_on_block(lock))
1693 ldlm_set_cancel(lock);
1695 do_ast = (!lock->l_readers && !lock->l_writers);
1696 unlock_res_and_lock(lock);
1699 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1700 lock, lock->l_blocking_ast);
1701 if (lock->l_blocking_ast != NULL)
1702 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1705 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1709 LDLM_DEBUG(lock, "client blocking callback handler END");
1710 LDLM_LOCK_RELEASE(lock);
1715 * Callback handler for receiving incoming completion ASTs.
1717 * This only can happen on client side.
1719 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1720 struct ldlm_namespace *ns,
1721 struct ldlm_request *dlm_req,
1722 struct ldlm_lock *lock)
1724 struct list_head ast_list;
1729 LDLM_DEBUG(lock, "client completion callback handler START");
1731 INIT_LIST_HEAD(&ast_list);
1732 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1733 long to = cfs_time_seconds(1);
1736 set_current_state(TASK_INTERRUPTIBLE);
1737 schedule_timeout(to);
1738 if (lock->l_granted_mode == lock->l_req_mode ||
1739 ldlm_is_destroyed(lock))
1744 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1746 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1747 GOTO(out, rc = lvb_len);
1748 } else if (lvb_len > 0) {
1749 if (lock->l_lvb_len > 0) {
1750 /* for extent lock, lvb contains ost_lvb{}. */
1751 LASSERT(lock->l_lvb_data != NULL);
1753 if (unlikely(lock->l_lvb_len < lvb_len)) {
1754 LDLM_ERROR(lock, "Replied LVB is larger than "
1755 "expectation, expected = %d, "
1757 lock->l_lvb_len, lvb_len);
1758 GOTO(out, rc = -EINVAL);
1763 lock_res_and_lock(lock);
1764 if (ldlm_is_destroyed(lock) ||
1765 lock->l_granted_mode == lock->l_req_mode) {
1766 /* bug 11300: the lock has already been granted */
1767 unlock_res_and_lock(lock);
1768 LDLM_DEBUG(lock, "Double grant race happened");
1772 /* If we receive the completion AST before the actual enqueue returned,
1773 * then we might need to switch lock modes, resources, or extents. */
1774 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1775 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1776 LDLM_DEBUG(lock, "completion AST, new lock mode");
1779 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1780 ldlm_convert_policy_to_local(req->rq_export,
1781 dlm_req->lock_desc.l_resource.lr_type,
1782 &dlm_req->lock_desc.l_policy_data,
1783 &lock->l_policy_data);
1784 LDLM_DEBUG(lock, "completion AST, new policy data");
1787 ldlm_resource_unlink_lock(lock);
1788 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1789 &lock->l_resource->lr_name,
1790 sizeof(lock->l_resource->lr_name)) != 0) {
1791 unlock_res_and_lock(lock);
1792 rc = ldlm_lock_change_resource(ns, lock,
1793 &dlm_req->lock_desc.l_resource.lr_name);
1795 LDLM_ERROR(lock, "Failed to allocate resource");
1798 LDLM_DEBUG(lock, "completion AST, new resource");
1799 CERROR("change resource!\n");
1800 lock_res_and_lock(lock);
1803 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1804 /* BL_AST locks are not needed in LRU.
1805 * Let ldlm_cancel_lru() be fast. */
1806 ldlm_lock_remove_from_lru(lock);
1807 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1808 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1811 if (lock->l_lvb_len > 0) {
1812 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1813 lock->l_lvb_data, lvb_len);
1815 unlock_res_and_lock(lock);
1820 ldlm_grant_lock(lock, &ast_list);
1821 unlock_res_and_lock(lock);
1823 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1825 /* Let Enqueue to call osc_lock_upcall() and initialize
1827 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1829 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1831 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1837 lock_res_and_lock(lock);
1838 ldlm_set_failed(lock);
1839 unlock_res_and_lock(lock);
1840 wake_up(&lock->l_waitq);
1842 LDLM_LOCK_RELEASE(lock);
1846 * Callback handler for receiving incoming glimpse ASTs.
1848 * This only can happen on client side. After handling the glimpse AST
1849 * we also consider dropping the lock here if it is unused locally for a
1852 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1853 struct ldlm_namespace *ns,
1854 struct ldlm_request *dlm_req,
1855 struct ldlm_lock *lock)
1860 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1862 if (lock->l_glimpse_ast != NULL)
1863 rc = lock->l_glimpse_ast(lock, req);
1865 if (req->rq_repmsg != NULL) {
1868 req->rq_status = rc;
1872 lock_res_and_lock(lock);
1873 if (lock->l_granted_mode == LCK_PW &&
1874 !lock->l_readers && !lock->l_writers &&
1875 ktime_after(ktime_get(),
1876 ktime_add(lock->l_last_used,
1877 ktime_set(ns->ns_dirty_age_limit, 0)))) {
1878 unlock_res_and_lock(lock);
1879 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1880 ldlm_handle_bl_callback(ns, NULL, lock);
1885 unlock_res_and_lock(lock);
1886 LDLM_LOCK_RELEASE(lock);
1890 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1892 if (req->rq_no_reply)
1895 req->rq_status = rc;
1896 if (!req->rq_packed_final) {
1897 rc = lustre_pack_reply(req, 1, NULL, NULL);
1901 return ptlrpc_reply(req);
1904 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1905 enum ldlm_cancel_flags cancel_flags)
1907 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1910 spin_lock(&blp->blp_lock);
1911 if (blwi->blwi_lock &&
1912 ldlm_is_discard_data(blwi->blwi_lock)) {
1913 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1914 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1916 /* other blocking callbacks are added to the regular list */
1917 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1919 spin_unlock(&blp->blp_lock);
1921 wake_up(&blp->blp_waitq);
1923 /* can not check blwi->blwi_flags as blwi could be already freed in
1925 if (!(cancel_flags & LCF_ASYNC))
1926 wait_for_completion(&blwi->blwi_comp);
1931 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1932 struct ldlm_namespace *ns,
1933 struct ldlm_lock_desc *ld,
1934 struct list_head *cancels, int count,
1935 struct ldlm_lock *lock,
1936 enum ldlm_cancel_flags cancel_flags)
1938 init_completion(&blwi->blwi_comp);
1939 INIT_LIST_HEAD(&blwi->blwi_head);
1941 if (memory_pressure_get())
1942 blwi->blwi_mem_pressure = 1;
1945 blwi->blwi_flags = cancel_flags;
1947 blwi->blwi_ld = *ld;
1949 list_add(&blwi->blwi_head, cancels);
1950 list_del_init(cancels);
1951 blwi->blwi_count = count;
1953 blwi->blwi_lock = lock;
1958 * Queues a list of locks \a cancels containing \a count locks
1959 * for later processing by a blocking thread. If \a count is zero,
1960 * then the lock referenced as \a lock is queued instead.
1962 * The blocking thread would then call ->l_blocking_ast callback in the lock.
1963 * If list addition fails an error is returned and caller is supposed to
1964 * call ->l_blocking_ast itself.
1966 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1967 struct ldlm_lock_desc *ld,
1968 struct ldlm_lock *lock,
1969 struct list_head *cancels, int count,
1970 enum ldlm_cancel_flags cancel_flags)
1974 if (cancels && count == 0)
1977 if (cancel_flags & LCF_ASYNC) {
1978 struct ldlm_bl_work_item *blwi;
1980 OBD_ALLOC(blwi, sizeof(*blwi));
1983 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
1985 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
1987 /* if it is synchronous call do minimum mem alloc, as it could
1988 * be triggered from kernel shrinker
1990 struct ldlm_bl_work_item blwi;
1992 memset(&blwi, 0, sizeof(blwi));
1993 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
1994 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
1999 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2000 struct ldlm_lock *lock)
2002 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2005 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2006 struct list_head *cancels, int count,
2007 enum ldlm_cancel_flags cancel_flags)
2009 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2012 int ldlm_bl_thread_wakeup(void)
2014 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2018 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2019 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2021 struct obd_device *obd = req->rq_export->exp_obd;
2028 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2030 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2032 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2034 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2037 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2039 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2041 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2044 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2047 /* We are responsible for swabbing contents of val */
2049 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2050 /* Pass it on to mdc (the "export" in this case) */
2051 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2053 sizeof(KEY_HSM_COPYTOOL_SEND),
2054 KEY_HSM_COPYTOOL_SEND,
2057 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2062 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2063 const char *msg, int rc,
2064 const struct lustre_handle *handle)
2066 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2067 "%s: [nid %s] [rc %d] [lock %#llx]",
2068 msg, libcfs_id2str(req->rq_peer), rc,
2069 handle ? handle->cookie : 0);
2070 if (req->rq_no_reply)
2071 CWARN("No reply was sent, maybe cause bug 21636.\n");
2073 CWARN("Send reply failed, maybe cause bug 21636.\n");
2076 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2077 static int ldlm_callback_handler(struct ptlrpc_request *req)
2079 struct ldlm_namespace *ns;
2080 struct ldlm_request *dlm_req;
2081 struct ldlm_lock *lock;
2085 /* Requests arrive in sender's byte order. The ptlrpc service
2086 * handler has already checked and, if necessary, byte-swapped the
2087 * incoming request message body, but I am responsible for the
2088 * message buffers. */
2090 /* do nothing for sec context finalize */
2091 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2094 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2096 if (req->rq_export == NULL) {
2097 rc = ldlm_callback_reply(req, -ENOTCONN);
2098 ldlm_callback_errmsg(req, "Operate on unconnected server",
2103 LASSERT(req->rq_export != NULL);
2104 LASSERT(req->rq_export->exp_obd != NULL);
2106 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2107 case LDLM_BL_CALLBACK:
2108 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2110 ldlm_callback_reply(req, -(int)cfs_fail_err);
2114 case LDLM_CP_CALLBACK:
2115 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2118 case LDLM_GL_CALLBACK:
2119 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2123 rc = ldlm_handle_setinfo(req);
2124 ldlm_callback_reply(req, rc);
2127 CERROR("unknown opcode %u\n",
2128 lustre_msg_get_opc(req->rq_reqmsg));
2129 ldlm_callback_reply(req, -EPROTO);
2133 ns = req->rq_export->exp_obd->obd_namespace;
2134 LASSERT(ns != NULL);
2136 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2138 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2139 if (dlm_req == NULL) {
2140 rc = ldlm_callback_reply(req, -EPROTO);
2141 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2146 /* Force a known safe race, send a cancel to the server for a lock
2147 * which the server has already started a blocking callback on. */
2148 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2149 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2150 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2152 CERROR("ldlm_cli_cancel: %d\n", rc);
2155 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2157 CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock "
2158 "disappeared\n", dlm_req->lock_handle[0].cookie);
2159 rc = ldlm_callback_reply(req, -EINVAL);
2160 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2161 &dlm_req->lock_handle[0]);
2165 if (ldlm_is_fail_loc(lock) &&
2166 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2167 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2169 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2170 lock_res_and_lock(lock);
2171 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2173 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2174 /* If somebody cancels lock and cache is already dropped,
2175 * or lock is failed before cp_ast received on client,
2176 * we can tell the server we have no lock. Otherwise, we
2177 * should send cancel after dropping the cache. */
2178 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2179 ldlm_is_failed(lock)) {
2180 LDLM_DEBUG(lock, "callback on lock %llx - lock disappeared",
2181 dlm_req->lock_handle[0].cookie);
2182 unlock_res_and_lock(lock);
2183 LDLM_LOCK_RELEASE(lock);
2184 rc = ldlm_callback_reply(req, -EINVAL);
2185 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2186 &dlm_req->lock_handle[0]);
2189 /* BL_AST locks are not needed in LRU.
2190 * Let ldlm_cancel_lru() be fast. */
2191 ldlm_lock_remove_from_lru(lock);
2192 ldlm_set_bl_ast(lock);
2194 unlock_res_and_lock(lock);
2196 /* We want the ost thread to get this reply so that it can respond
2197 * to ost requests (write cache writeback) that might be triggered
2200 * But we'd also like to be able to indicate in the reply that we're
2201 * cancelling right now, because it's unused, or have an intent result
2202 * in the reply, so we might have to push the responsibility for sending
2203 * the reply down into the AST handlers, alas. */
2205 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2206 case LDLM_BL_CALLBACK:
2207 CDEBUG(D_INODE, "blocking ast\n");
2208 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2209 if (!ldlm_is_cancel_on_block(lock)) {
2210 rc = ldlm_callback_reply(req, 0);
2211 if (req->rq_no_reply || rc)
2212 ldlm_callback_errmsg(req, "Normal process", rc,
2213 &dlm_req->lock_handle[0]);
2215 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2216 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2218 case LDLM_CP_CALLBACK:
2219 CDEBUG(D_INODE, "completion ast\n");
2220 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2221 ldlm_callback_reply(req, 0);
2222 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2224 case LDLM_GL_CALLBACK:
2225 CDEBUG(D_INODE, "glimpse ast\n");
2226 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2227 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2230 LBUG(); /* checked above */
2236 #ifdef HAVE_SERVER_SUPPORT
2238 * Main handler for canceld thread.
2240 * Separated into its own thread to avoid deadlocks.
2242 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2248 /* Requests arrive in sender's byte order. The ptlrpc service
2249 * handler has already checked and, if necessary, byte-swapped the
2250 * incoming request message body, but I am responsible for the
2251 * message buffers. */
2253 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2255 if (req->rq_export == NULL) {
2256 struct ldlm_request *dlm_req;
2258 CERROR("%s from %s arrived at %lu with bad export cookie "
2260 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2261 libcfs_nid2str(req->rq_peer.nid),
2262 req->rq_arrival_time.tv_sec,
2263 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2265 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2266 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2267 dlm_req = req_capsule_client_get(&req->rq_pill,
2269 if (dlm_req != NULL)
2270 ldlm_lock_dump_handle(D_ERROR,
2271 &dlm_req->lock_handle[0]);
2273 ldlm_callback_reply(req, -ENOTCONN);
2277 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2278 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2280 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2281 CDEBUG(D_INODE, "cancel\n");
2282 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2283 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2284 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2286 rc = ldlm_handle_cancel(req);
2290 struct ldlm_request *dlm_req;
2292 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2293 CDEBUG(D_INODE, "convert\n");
2295 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2296 if (dlm_req == NULL) {
2297 CDEBUG(D_INFO, "bad request buffer for cancel\n");
2298 rc = ldlm_callback_reply(req, -EPROTO);
2300 req->rq_status = ldlm_handle_convert0(req, dlm_req);
2301 rc = ptlrpc_reply(req);
2306 CERROR("invalid opcode %d\n",
2307 lustre_msg_get_opc(req->rq_reqmsg));
2308 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2309 rc = ldlm_callback_reply(req, -EINVAL);
2315 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2316 struct ldlm_lock *lock)
2318 struct ldlm_request *dlm_req;
2319 struct lustre_handle lockh;
2325 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2326 if (dlm_req == NULL)
2329 ldlm_lock2handle(lock, &lockh);
2330 for (i = 0; i < dlm_req->lock_count; i++) {
2331 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2333 DEBUG_REQ(D_RPCTRACE, req,
2334 "Prio raised by lock %#llx.", lockh.cookie);
2343 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2345 struct ldlm_request *dlm_req;
2351 /* no prolong in recovery */
2352 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2355 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2356 if (dlm_req == NULL)
2359 for (i = 0; i < dlm_req->lock_count; i++) {
2360 struct ldlm_lock *lock;
2362 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2366 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2368 LDLM_DEBUG(lock, "hpreq cancel/convert lock");
2369 LDLM_LOCK_PUT(lock);
2378 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2379 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2380 .hpreq_check = ldlm_cancel_hpreq_check,
2384 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2388 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2390 if (req->rq_export == NULL)
2393 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2394 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2395 req->rq_ops = &ldlm_cancel_hpreq_ops;
2396 } else if (LDLM_CONVERT == lustre_msg_get_opc(req->rq_reqmsg)) {
2397 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2398 req->rq_ops = &ldlm_cancel_hpreq_ops;
2403 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2404 struct hlist_node *hnode, void *data)
2407 struct list_head *rpc_list = data;
2408 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2410 lock_res_and_lock(lock);
2412 if (lock->l_req_mode != lock->l_granted_mode) {
2413 unlock_res_and_lock(lock);
2417 LASSERT(lock->l_resource);
2418 if (lock->l_resource->lr_type != LDLM_IBITS &&
2419 lock->l_resource->lr_type != LDLM_PLAIN) {
2420 unlock_res_and_lock(lock);
2424 if (ldlm_is_ast_sent(lock)) {
2425 unlock_res_and_lock(lock);
2429 LASSERT(lock->l_blocking_ast);
2430 LASSERT(!lock->l_blocking_lock);
2432 ldlm_set_ast_sent(lock);
2433 if (lock->l_export && lock->l_export->exp_lock_hash) {
2434 /* NB: it's safe to call cfs_hash_del() even lock isn't
2435 * in exp_lock_hash. */
2436 /* In the function below, .hs_keycmp resolves to
2437 * ldlm_export_lock_keycmp() */
2438 /* coverity[overrun-buffer-val] */
2439 cfs_hash_del(lock->l_export->exp_lock_hash,
2440 &lock->l_remote_handle, &lock->l_exp_hash);
2443 list_add_tail(&lock->l_rk_ast, rpc_list);
2444 LDLM_LOCK_GET(lock);
2446 unlock_res_and_lock(lock);
2450 void ldlm_revoke_export_locks(struct obd_export *exp)
2452 struct list_head rpc_list;
2455 INIT_LIST_HEAD(&rpc_list);
2456 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2457 ldlm_revoke_lock_cb, &rpc_list, 0);
2458 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2459 LDLM_WORK_REVOKE_AST);
2463 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2464 #endif /* HAVE_SERVER_SUPPORT */
2466 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2467 struct ldlm_bl_work_item **p_blwi,
2468 struct obd_export **p_exp)
2470 struct ldlm_bl_work_item *blwi = NULL;
2471 static unsigned int num_bl = 0;
2472 static unsigned int num_stale;
2473 int num_th = atomic_read(&blp->blp_num_threads);
2475 *p_exp = obd_stale_export_get();
2477 spin_lock(&blp->blp_lock);
2478 if (*p_exp != NULL) {
2479 if (num_th == 1 || ++num_stale < num_th) {
2480 spin_unlock(&blp->blp_lock);
2487 /* process a request from the blp_list at least every blp_num_threads */
2488 if (!list_empty(&blp->blp_list) &&
2489 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2490 blwi = list_entry(blp->blp_list.next,
2491 struct ldlm_bl_work_item, blwi_entry);
2493 if (!list_empty(&blp->blp_prio_list))
2494 blwi = list_entry(blp->blp_prio_list.next,
2495 struct ldlm_bl_work_item,
2499 if (++num_bl >= num_th)
2501 list_del(&blwi->blwi_entry);
2503 spin_unlock(&blp->blp_lock);
2506 if (*p_exp != NULL && *p_blwi != NULL) {
2507 obd_stale_export_put(*p_exp);
2511 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2514 /* This only contains temporary data until the thread starts */
2515 struct ldlm_bl_thread_data {
2516 struct ldlm_bl_pool *bltd_blp;
2517 struct completion bltd_comp;
2521 static int ldlm_bl_thread_main(void *arg);
2523 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2525 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2526 struct task_struct *task;
2528 init_completion(&bltd.bltd_comp);
2530 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2531 if (bltd.bltd_num >= blp->blp_max_threads) {
2532 atomic_dec(&blp->blp_num_threads);
2536 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2538 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2539 atomic_dec(&blp->blp_num_threads);
2543 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2546 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2547 bltd.bltd_num, PTR_ERR(task));
2548 atomic_dec(&blp->blp_num_threads);
2549 return PTR_ERR(task);
2551 wait_for_completion(&bltd.bltd_comp);
2556 /* Not fatal if racy and have a few too many threads */
2557 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2558 struct ldlm_bl_work_item *blwi)
2560 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2563 if (atomic_read(&blp->blp_busy_threads) <
2564 atomic_read(&blp->blp_num_threads))
2567 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2568 blwi->blwi_mem_pressure))
2574 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2575 struct ldlm_bl_work_item *blwi)
2579 if (blwi->blwi_ns == NULL)
2580 /* added by ldlm_cleanup() */
2581 RETURN(LDLM_ITER_STOP);
2583 if (blwi->blwi_mem_pressure)
2584 memory_pressure_set();
2586 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2588 if (blwi->blwi_count) {
2590 /* The special case when we cancel locks in lru
2591 * asynchronously, we pass the list of locks here.
2592 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2593 * canceled locally yet. */
2594 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2597 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2600 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2603 if (blwi->blwi_mem_pressure)
2604 memory_pressure_clr();
2606 if (blwi->blwi_flags & LCF_ASYNC)
2607 OBD_FREE(blwi, sizeof(*blwi));
2609 complete(&blwi->blwi_comp);
2615 * Cancel stale locks on export. Cancel blocked locks first.
2616 * If the given export has blocked locks, the next in the list may have
2617 * them too, thus cancel not blocked locks only if the current export has
2620 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2621 struct obd_export *exp)
2626 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2628 num = ldlm_export_cancel_blocked_locks(exp);
2630 ldlm_export_cancel_locks(exp);
2632 obd_stale_export_put(exp);
2639 * Main blocking requests processing thread.
2641 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2642 * This thread in the end ends up doing actual call to ->l_blocking_ast
2645 static int ldlm_bl_thread_main(void *arg)
2647 struct ldlm_bl_pool *blp;
2648 struct ldlm_bl_thread_data *bltd = arg;
2651 blp = bltd->bltd_blp;
2653 complete(&bltd->bltd_comp);
2654 /* cannot use bltd after this, it is only on caller's stack */
2657 struct l_wait_info lwi = { 0 };
2658 struct ldlm_bl_work_item *blwi = NULL;
2659 struct obd_export *exp = NULL;
2662 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2665 l_wait_event_exclusive(blp->blp_waitq,
2666 ldlm_bl_get_work(blp, &blwi,
2669 atomic_inc(&blp->blp_busy_threads);
2671 if (ldlm_bl_thread_need_create(blp, blwi))
2672 /* discard the return value, we tried */
2673 ldlm_bl_thread_start(blp, true);
2676 rc = ldlm_bl_thread_exports(blp, exp);
2678 rc = ldlm_bl_thread_blwi(blp, blwi);
2680 atomic_dec(&blp->blp_busy_threads);
2682 if (rc == LDLM_ITER_STOP)
2685 /* If there are many namespaces, we will not sleep waiting for
2686 * work, and must do a cond_resched to avoid holding the CPU
2691 atomic_dec(&blp->blp_num_threads);
2692 complete(&blp->blp_comp);
2697 static int ldlm_setup(void);
2698 static int ldlm_cleanup(void);
2700 int ldlm_get_ref(void)
2704 mutex_lock(&ldlm_ref_mutex);
2705 if (++ldlm_refcount == 1) {
2710 mutex_unlock(&ldlm_ref_mutex);
2715 void ldlm_put_ref(void)
2718 mutex_lock(&ldlm_ref_mutex);
2719 if (ldlm_refcount == 1) {
2720 int rc = ldlm_cleanup();
2722 CERROR("ldlm_cleanup failed: %d\n", rc);
2728 mutex_unlock(&ldlm_ref_mutex);
2734 * Export handle<->lock hash operations.
2737 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2739 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2743 ldlm_export_lock_key(struct hlist_node *hnode)
2745 struct ldlm_lock *lock;
2747 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2748 return &lock->l_remote_handle;
2752 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2754 struct ldlm_lock *lock;
2756 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2757 lock->l_remote_handle = *(struct lustre_handle *)key;
2761 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2763 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2767 ldlm_export_lock_object(struct hlist_node *hnode)
2769 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2773 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
2775 struct ldlm_lock *lock;
2777 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2778 LDLM_LOCK_GET(lock);
2782 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2784 struct ldlm_lock *lock;
2786 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2787 LDLM_LOCK_RELEASE(lock);
2790 static struct cfs_hash_ops ldlm_export_lock_ops = {
2791 .hs_hash = ldlm_export_lock_hash,
2792 .hs_key = ldlm_export_lock_key,
2793 .hs_keycmp = ldlm_export_lock_keycmp,
2794 .hs_keycpy = ldlm_export_lock_keycpy,
2795 .hs_object = ldlm_export_lock_object,
2796 .hs_get = ldlm_export_lock_get,
2797 .hs_put = ldlm_export_lock_put,
2798 .hs_put_locked = ldlm_export_lock_put,
2801 int ldlm_init_export(struct obd_export *exp)
2806 exp->exp_lock_hash =
2807 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2808 HASH_EXP_LOCK_CUR_BITS,
2809 HASH_EXP_LOCK_MAX_BITS,
2810 HASH_EXP_LOCK_BKT_BITS, 0,
2811 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2812 &ldlm_export_lock_ops,
2813 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2814 CFS_HASH_NBLK_CHANGE);
2816 if (!exp->exp_lock_hash)
2819 rc = ldlm_init_flock_export(exp);
2825 ldlm_destroy_export(exp);
2828 EXPORT_SYMBOL(ldlm_init_export);
2830 void ldlm_destroy_export(struct obd_export *exp)
2833 cfs_hash_putref(exp->exp_lock_hash);
2834 exp->exp_lock_hash = NULL;
2836 ldlm_destroy_flock_export(exp);
2839 EXPORT_SYMBOL(ldlm_destroy_export);
2841 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
2842 struct attribute *attr,
2845 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
2848 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
2849 struct attribute *attr,
2856 rc = kstrtoul(buffer, 10, &val);
2860 ldlm_cancel_unused_locks_before_replay = val;
2864 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
2866 static struct attribute *ldlm_attrs[] = {
2867 &lustre_attr_cancel_unused_locks_before_replay.attr,
2871 static struct attribute_group ldlm_attr_group = {
2872 .attrs = ldlm_attrs,
2875 static int ldlm_setup(void)
2877 static struct ptlrpc_service_conf conf;
2878 struct ldlm_bl_pool *blp = NULL;
2879 #ifdef HAVE_SERVER_SUPPORT
2880 struct task_struct *task;
2881 #endif /* HAVE_SERVER_SUPPORT */
2887 if (ldlm_state != NULL)
2890 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2891 if (ldlm_state == NULL)
2894 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
2898 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
2902 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
2906 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
2910 rc = ldlm_debugfs_setup();
2914 memset(&conf, 0, sizeof(conf));
2915 conf = (typeof(conf)) {
2916 .psc_name = "ldlm_cbd",
2917 .psc_watchdog_factor = 2,
2919 .bc_nbufs = LDLM_CLIENT_NBUFS,
2920 .bc_buf_size = LDLM_BUFSIZE,
2921 .bc_req_max_size = LDLM_MAXREQSIZE,
2922 .bc_rep_max_size = LDLM_MAXREPSIZE,
2923 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2924 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2927 .tc_thr_name = "ldlm_cb",
2928 .tc_thr_factor = LDLM_THR_FACTOR,
2929 .tc_nthrs_init = LDLM_NTHRS_INIT,
2930 .tc_nthrs_base = LDLM_NTHRS_BASE,
2931 .tc_nthrs_max = LDLM_NTHRS_MAX,
2932 .tc_nthrs_user = ldlm_num_threads,
2933 .tc_cpu_affinity = 1,
2934 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2937 .cc_pattern = ldlm_cpts,
2940 .so_req_handler = ldlm_callback_handler,
2943 ldlm_state->ldlm_cb_service = \
2944 ptlrpc_register_service(&conf, ldlm_svc_kset,
2945 ldlm_svc_debugfs_dir);
2946 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2947 CERROR("failed to start service\n");
2948 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2949 ldlm_state->ldlm_cb_service = NULL;
2953 #ifdef HAVE_SERVER_SUPPORT
2954 memset(&conf, 0, sizeof(conf));
2955 conf = (typeof(conf)) {
2956 .psc_name = "ldlm_canceld",
2957 .psc_watchdog_factor = 6,
2959 .bc_nbufs = LDLM_SERVER_NBUFS,
2960 .bc_buf_size = LDLM_BUFSIZE,
2961 .bc_req_max_size = LDLM_MAXREQSIZE,
2962 .bc_rep_max_size = LDLM_MAXREPSIZE,
2963 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
2964 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
2968 .tc_thr_name = "ldlm_cn",
2969 .tc_thr_factor = LDLM_THR_FACTOR,
2970 .tc_nthrs_init = LDLM_NTHRS_INIT,
2971 .tc_nthrs_base = LDLM_NTHRS_BASE,
2972 .tc_nthrs_max = LDLM_NTHRS_MAX,
2973 .tc_nthrs_user = ldlm_num_threads,
2974 .tc_cpu_affinity = 1,
2975 .tc_ctx_tags = LCT_MD_THREAD | \
2980 .cc_pattern = ldlm_cpts,
2983 .so_req_handler = ldlm_cancel_handler,
2984 .so_hpreq_handler = ldlm_hpreq_handler,
2987 ldlm_state->ldlm_cancel_service = \
2988 ptlrpc_register_service(&conf, ldlm_svc_kset,
2989 ldlm_svc_debugfs_dir);
2990 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2991 CERROR("failed to start service\n");
2992 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2993 ldlm_state->ldlm_cancel_service = NULL;
2996 #endif /* HAVE_SERVER_SUPPORT */
2998 OBD_ALLOC(blp, sizeof(*blp));
3000 GOTO(out, rc = -ENOMEM);
3001 ldlm_state->ldlm_bl_pool = blp;
3003 spin_lock_init(&blp->blp_lock);
3004 INIT_LIST_HEAD(&blp->blp_list);
3005 INIT_LIST_HEAD(&blp->blp_prio_list);
3006 init_waitqueue_head(&blp->blp_waitq);
3007 atomic_set(&blp->blp_num_threads, 0);
3008 atomic_set(&blp->blp_busy_threads, 0);
3010 if (ldlm_num_threads == 0) {
3011 blp->blp_min_threads = LDLM_NTHRS_INIT;
3012 blp->blp_max_threads = LDLM_NTHRS_MAX;
3014 blp->blp_min_threads = blp->blp_max_threads = \
3015 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3019 for (i = 0; i < blp->blp_min_threads; i++) {
3020 rc = ldlm_bl_thread_start(blp, false);
3025 #ifdef HAVE_SERVER_SUPPORT
3026 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3029 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3033 wait_event(expired_lock_wait_queue,
3034 expired_lock_thread_state == ELT_READY);
3035 #endif /* HAVE_SERVER_SUPPORT */
3037 rc = ldlm_pools_init();
3039 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3043 rc = ldlm_reclaim_setup();
3045 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3055 static int ldlm_cleanup(void)
3059 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3060 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3061 CERROR("ldlm still has namespaces; clean these up first.\n");
3062 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3063 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3067 ldlm_reclaim_cleanup();
3070 if (ldlm_state->ldlm_bl_pool != NULL) {
3071 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3073 while (atomic_read(&blp->blp_num_threads) > 0) {
3074 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3076 init_completion(&blp->blp_comp);
3078 spin_lock(&blp->blp_lock);
3079 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3080 wake_up(&blp->blp_waitq);
3081 spin_unlock(&blp->blp_lock);
3083 wait_for_completion(&blp->blp_comp);
3086 OBD_FREE(blp, sizeof(*blp));
3089 if (ldlm_state->ldlm_cb_service != NULL)
3090 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3091 #ifdef HAVE_SERVER_SUPPORT
3092 if (ldlm_state->ldlm_cancel_service != NULL)
3093 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3097 kset_unregister(ldlm_ns_kset);
3099 kset_unregister(ldlm_svc_kset);
3101 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3102 kobject_put(ldlm_kobj);
3105 ldlm_debugfs_cleanup();
3107 #ifdef HAVE_SERVER_SUPPORT
3108 if (expired_lock_thread_state != ELT_STOPPED) {
3109 expired_lock_thread_state = ELT_TERMINATE;
3110 wake_up(&expired_lock_wait_queue);
3111 wait_event(expired_lock_wait_queue,
3112 expired_lock_thread_state == ELT_STOPPED);
3116 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3124 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3125 sizeof(struct ldlm_resource), 0,
3126 SLAB_HWCACHE_ALIGN, NULL);
3127 if (ldlm_resource_slab == NULL)
3130 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3131 sizeof(struct ldlm_lock), 0,
3132 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3133 if (ldlm_lock_slab == NULL)
3136 ldlm_interval_slab = kmem_cache_create("interval_node",
3137 sizeof(struct ldlm_interval),
3138 0, SLAB_HWCACHE_ALIGN, NULL);
3139 if (ldlm_interval_slab == NULL)
3142 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3143 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3144 0, SLAB_HWCACHE_ALIGN, NULL);
3145 if (ldlm_interval_tree_slab == NULL)
3148 #ifdef HAVE_SERVER_SUPPORT
3149 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3150 sizeof(struct ldlm_glimpse_work),
3152 if (ldlm_glimpse_work_kmem == NULL)
3153 goto out_interval_tree;
3156 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3157 class_export_dump_hook = ldlm_dump_export_locks;
3160 #ifdef HAVE_SERVER_SUPPORT
3162 kmem_cache_destroy(ldlm_interval_tree_slab);
3165 kmem_cache_destroy(ldlm_interval_slab);
3167 kmem_cache_destroy(ldlm_lock_slab);
3169 kmem_cache_destroy(ldlm_resource_slab);
3174 void ldlm_exit(void)
3177 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3178 kmem_cache_destroy(ldlm_resource_slab);
3179 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3180 * synchronize_rcu() to wait a grace period elapsed, so that
3181 * ldlm_lock_free() get a chance to be called. */
3183 kmem_cache_destroy(ldlm_lock_slab);
3184 kmem_cache_destroy(ldlm_interval_slab);
3185 kmem_cache_destroy(ldlm_interval_tree_slab);
3186 #ifdef HAVE_SERVER_SUPPORT
3187 kmem_cache_destroy(ldlm_glimpse_work_kmem);