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() -
250 lock->l_last_activity,
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_last_activity = 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 lock->l_last_activity = ktime_get_real_seconds();
923 if (lock->l_export && lock->l_export->exp_nid_stats &&
924 lock->l_export->exp_nid_stats->nid_ldlm_stats)
925 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
926 LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
928 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
934 * ->l_completion_ast callback for a remote lock in server namespace.
936 * Sends AST to the client notifying it of lock granting. If initial
937 * lock response was not sent yet, instead of sending another RPC, just
938 * mark the lock as granted and client will understand
940 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
942 struct ldlm_cb_set_arg *arg = data;
943 struct ldlm_request *body;
944 struct ptlrpc_request *req;
945 struct ldlm_cb_async_args *ca;
946 int instant_cancel = 0;
951 LASSERT(lock != NULL);
952 LASSERT(data != NULL);
954 if (OBD_FAIL_PRECHECK(OBD_FAIL_LDLM_SRV_CP_AST)) {
955 LDLM_DEBUG(lock, "dropping CP AST");
959 req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
960 &RQF_LDLM_CP_CALLBACK);
964 /* server namespace, doesn't need lock */
965 lvb_len = ldlm_lvbo_size(lock);
966 /* LU-3124 & LU-2187: to not return layout in completion AST because
967 * it may deadlock for LU-2187, or client may not have enough space
968 * for large layout. The layout will be returned to client with an
969 * extra RPC to fetch xattr.lov */
970 if (ldlm_has_layout(lock))
973 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT, lvb_len);
974 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
976 ptlrpc_request_free(req);
980 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
981 ca = ptlrpc_req_async_args(req);
982 ca->ca_set_arg = arg;
985 req->rq_interpret_reply = ldlm_cb_interpret;
986 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
988 body->lock_handle[0] = lock->l_remote_handle;
989 body->lock_flags = ldlm_flags_to_wire(flags);
990 ldlm_lock2desc(lock, &body->lock_desc);
992 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
994 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
996 /* We still need to send the RPC to wake up the blocked
997 * enqueue thread on the client.
999 * Consider old client, there is no better way to notify
1000 * the failure, just zero-sized the LVB, then the client
1001 * will fail out as "-EPROTO". */
1002 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, 0,
1006 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len,
1011 lock->l_last_activity = ktime_get_real_seconds();
1013 LDLM_DEBUG(lock, "server preparing completion AST");
1015 ptlrpc_request_set_replen(req);
1017 req->rq_send_state = LUSTRE_IMP_FULL;
1018 /* ptlrpc_request_pack already set timeout */
1020 req->rq_timeout = ldlm_get_rq_timeout();
1022 /* We only send real blocking ASTs after the lock is granted */
1023 lock_res_and_lock(lock);
1024 if (ldlm_is_ast_sent(lock)) {
1025 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1026 /* Copy AST flags like LDLM_FL_DISCARD_DATA. */
1027 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
1030 /* We might get here prior to ldlm_handle_enqueue setting
1031 * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
1032 * into waiting list, but this is safe and similar code in
1033 * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
1034 * that would not only cancel the lock, but will also remove
1035 * it from waiting list */
1036 if (ldlm_is_cancel_on_block(lock)) {
1037 unlock_res_and_lock(lock);
1038 ldlm_lock_cancel(lock);
1041 req->rq_no_resend = 1;
1043 lock_res_and_lock(lock);
1045 /* start the lock-timeout clock */
1046 ldlm_add_waiting_lock(lock, ldlm_bl_timeout(lock));
1047 /* Do not resend after lock callback timeout */
1048 req->rq_delay_limit = ldlm_bl_timeout(lock);
1049 req->rq_resend_cb = ldlm_update_resend;
1052 unlock_res_and_lock(lock);
1054 if (lock->l_export && lock->l_export->exp_nid_stats &&
1055 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1056 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1057 LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1059 rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1061 RETURN(lvb_len < 0 ? lvb_len : rc);
1065 * Server side ->l_glimpse_ast handler for client locks.
1067 * Sends glimpse AST to the client and waits for reply. Then updates
1068 * lvbo with the result.
1070 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1072 struct ldlm_cb_set_arg *arg = data;
1073 struct ldlm_request *body;
1074 struct ptlrpc_request *req;
1075 struct ldlm_cb_async_args *ca;
1077 struct req_format *req_fmt;
1080 LASSERT(lock != NULL);
1082 if (arg->gl_desc != NULL)
1083 /* There is a glimpse descriptor to pack */
1084 req_fmt = &RQF_LDLM_GL_CALLBACK_DESC;
1086 req_fmt = &RQF_LDLM_GL_CALLBACK;
1088 req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1089 req_fmt, LUSTRE_DLM_VERSION,
1095 if (arg->gl_desc != NULL) {
1096 /* copy the GL descriptor */
1097 union ldlm_gl_desc *desc;
1098 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1099 *desc = *arg->gl_desc;
1102 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1103 body->lock_handle[0] = lock->l_remote_handle;
1104 ldlm_lock2desc(lock, &body->lock_desc);
1106 CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1107 ca = ptlrpc_req_async_args(req);
1108 ca->ca_set_arg = arg;
1111 /* server namespace, doesn't need lock */
1112 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1113 ldlm_lvbo_size(lock));
1114 ptlrpc_request_set_replen(req);
1116 req->rq_send_state = LUSTRE_IMP_FULL;
1117 /* ptlrpc_request_alloc_pack already set timeout */
1119 req->rq_timeout = ldlm_get_rq_timeout();
1121 lock->l_last_activity = ktime_get_real_seconds();
1123 req->rq_interpret_reply = ldlm_cb_interpret;
1125 if (lock->l_export && lock->l_export->exp_nid_stats &&
1126 lock->l_export->exp_nid_stats->nid_ldlm_stats)
1127 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1128 LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1130 rc = ldlm_ast_fini(req, arg, lock, 0);
1134 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1136 int ldlm_glimpse_locks(struct ldlm_resource *res,
1137 struct list_head *gl_work_list)
1142 rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1144 if (rc == -ERESTART)
1145 ldlm_reprocess_all(res);
1149 EXPORT_SYMBOL(ldlm_glimpse_locks);
1151 /* return LDLM lock associated with a lock callback request */
1152 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1154 struct ldlm_cb_async_args *ca;
1155 struct ldlm_lock *lock;
1158 ca = ptlrpc_req_async_args(req);
1161 RETURN(ERR_PTR(-EFAULT));
1165 EXPORT_SYMBOL(ldlm_request_lock);
1168 * Main server-side entry point into LDLM for enqueue. This is called by ptlrpc
1169 * service threads to carry out client lock enqueueing requests.
1171 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1172 struct ptlrpc_request *req,
1173 const struct ldlm_request *dlm_req,
1174 const struct ldlm_callback_suite *cbs)
1176 struct ldlm_reply *dlm_rep;
1178 enum ldlm_error err = ELDLM_OK;
1179 struct ldlm_lock *lock = NULL;
1180 void *cookie = NULL;
1182 struct ldlm_resource *res = NULL;
1185 LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1187 ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF, LATF_SKIP);
1188 flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1190 LASSERT(req->rq_export);
1192 /* for intent enqueue the stat will be updated inside intent policy */
1193 if (ptlrpc_req2svc(req)->srv_stats != NULL &&
1194 !(dlm_req->lock_flags & LDLM_FL_HAS_INTENT))
1195 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1197 if (req->rq_export && req->rq_export->exp_nid_stats &&
1198 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1199 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1200 LDLM_ENQUEUE - LDLM_FIRST_OPC);
1202 if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1203 dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1204 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1205 dlm_req->lock_desc.l_resource.lr_type);
1206 GOTO(out, rc = -EFAULT);
1209 if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1210 dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1211 dlm_req->lock_desc.l_req_mode &
1212 (dlm_req->lock_desc.l_req_mode-1))) {
1213 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1214 dlm_req->lock_desc.l_req_mode);
1215 GOTO(out, rc = -EFAULT);
1218 if (unlikely((flags & LDLM_FL_REPLAY) ||
1219 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))) {
1220 /* Find an existing lock in the per-export lock hash */
1221 /* In the function below, .hs_keycmp resolves to
1222 * ldlm_export_lock_keycmp() */
1223 /* coverity[overrun-buffer-val] */
1224 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1225 (void *)&dlm_req->lock_handle[0]);
1227 DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie %#llx",
1228 lock->l_handle.h_cookie);
1229 flags |= LDLM_FL_RESENT;
1230 GOTO(existing_lock, rc = 0);
1233 if (ldlm_reclaim_full()) {
1234 DEBUG_REQ(D_DLMTRACE, req, "Too many granted locks, "
1235 "reject current enqueue request and let the "
1236 "client retry later.\n");
1237 GOTO(out, rc = -EINPROGRESS);
1241 /* The lock's callback data might be set in the policy function */
1242 lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1243 dlm_req->lock_desc.l_resource.lr_type,
1244 dlm_req->lock_desc.l_req_mode,
1245 cbs, NULL, 0, LVB_T_NONE);
1252 lock->l_remote_handle = dlm_req->lock_handle[0];
1253 LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1255 /* Initialize resource lvb but not for a lock being replayed since
1256 * Client already got lvb sent in this case.
1257 * This must occur early since some policy methods assume resource
1258 * lvb is available (lr_lvb_data != NULL).
1260 res = lock->l_resource;
1261 if (!(flags & LDLM_FL_REPLAY)) {
1262 /* non-replayed lock, delayed lvb init may need to be done */
1263 rc = ldlm_lvbo_init(res);
1265 LDLM_DEBUG(lock, "delayed lvb init failed (rc %d)", rc);
1270 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1271 /* Don't enqueue a lock onto the export if it is been disonnected
1272 * due to eviction (bug 3822) or server umount (bug 24324).
1273 * Cancel it now instead. */
1274 if (req->rq_export->exp_disconnected) {
1275 LDLM_ERROR(lock, "lock on disconnected export %p",
1277 GOTO(out, rc = -ENOTCONN);
1280 lock->l_export = class_export_lock_get(req->rq_export, lock);
1281 if (lock->l_export->exp_lock_hash)
1282 cfs_hash_add(lock->l_export->exp_lock_hash,
1283 &lock->l_remote_handle,
1286 /* Inherit the enqueue flags before the operation, because we do not
1287 * keep the res lock on return and next operations (BL AST) may proceed
1289 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1290 LDLM_FL_INHERIT_MASK);
1292 ldlm_convert_policy_to_local(req->rq_export,
1293 dlm_req->lock_desc.l_resource.lr_type,
1294 &dlm_req->lock_desc.l_policy_data,
1295 &lock->l_policy_data);
1296 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1297 lock->l_req_extent = lock->l_policy_data.l_extent;
1300 if (flags & LDLM_FL_HAS_INTENT) {
1301 /* In this case, the reply buffer is allocated deep in
1302 * local_lock_enqueue by the policy function. */
1305 /* based on the assumption that lvb size never changes during
1306 * resource life time otherwise it need resource->lr_lock's
1308 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1309 RCL_SERVER, ldlm_lvbo_size(lock));
1311 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1312 GOTO(out, rc = -ENOMEM);
1314 rc = req_capsule_server_pack(&req->rq_pill);
1319 err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1326 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1328 ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1329 ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1331 if (lock && lock->l_resource->lr_type == LDLM_EXTENT)
1332 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 6);
1334 /* We never send a blocking AST until the lock is granted, but
1335 * we can tell it right now */
1336 lock_res_and_lock(lock);
1338 /* Now take into account flags to be inherited from original lock
1339 request both in reply to client and in our own lock flags. */
1340 dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1341 lock->l_flags |= flags & LDLM_FL_INHERIT_MASK;
1343 /* Don't move a pending lock onto the export if it has already been
1344 * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1345 * Cancel it now instead. */
1346 if (unlikely(req->rq_export->exp_disconnected ||
1347 OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1348 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1350 } else if (ldlm_is_ast_sent(lock)) {
1351 /* fill lock desc for possible lock convert */
1352 if (lock->l_blocking_lock &&
1353 lock->l_resource->lr_type == LDLM_IBITS) {
1354 struct ldlm_lock *bl_lock = lock->l_blocking_lock;
1355 struct ldlm_lock_desc *rep_desc = &dlm_rep->lock_desc;
1358 "save blocking bits %llx in granted lock",
1359 bl_lock->l_policy_data.l_inodebits.bits);
1360 /* If lock is blocked then save blocking ibits
1361 * in returned lock policy for the possible lock
1362 * convert on a client.
1364 rep_desc->l_policy_data.l_inodebits.cancel_bits =
1365 bl_lock->l_policy_data.l_inodebits.bits;
1367 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1368 if (lock->l_granted_mode == lock->l_req_mode) {
1370 * Only cancel lock if it was granted, because it would
1371 * be destroyed immediately and would never be granted
1372 * in the future, causing timeouts on client. Not
1373 * granted lock will be cancelled immediately after
1374 * sending completion AST.
1376 if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1377 unlock_res_and_lock(lock);
1378 ldlm_lock_cancel(lock);
1379 lock_res_and_lock(lock);
1381 ldlm_add_waiting_lock(lock,
1382 ldlm_bl_timeout(lock));
1386 unlock_res_and_lock(lock);
1390 req->rq_status = rc ?: err; /* return either error - bug 11190 */
1391 if (!req->rq_packed_final) {
1392 err = lustre_pack_reply(req, 1, NULL, NULL);
1397 /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1398 * ldlm_reprocess_all. If this moves, revisit that code. -phil */
1400 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1401 "(err=%d, rc=%d)", err, rc);
1404 if (req_capsule_has_field(&req->rq_pill, &RMF_DLM_LVB,
1406 ldlm_lvbo_size(lock) > 0) {
1410 buf = req_capsule_server_get(&req->rq_pill,
1412 LASSERTF(buf != NULL, "req %p, lock %p\n",
1414 buflen = req_capsule_get_size(&req->rq_pill,
1415 &RMF_DLM_LVB, RCL_SERVER);
1416 /* non-replayed lock, delayed lvb init may
1417 * need to be occur now */
1418 if ((buflen > 0) && !(flags & LDLM_FL_REPLAY)) {
1419 buflen = ldlm_lvbo_fill(lock, buf,
1425 buflen, RCL_SERVER);
1428 } else if (flags & LDLM_FL_REPLAY) {
1429 /* no LVB resend upon replay */
1443 if (rc != 0 && !(flags & LDLM_FL_RESENT)) {
1444 if (lock->l_export) {
1445 ldlm_lock_cancel(lock);
1447 lock_res_and_lock(lock);
1448 ldlm_resource_unlink_lock(lock);
1449 ldlm_lock_destroy_nolock(lock);
1450 unlock_res_and_lock(lock);
1455 if (!err && !ldlm_is_cbpending(lock) &&
1456 dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1457 ldlm_reprocess_all(lock->l_resource);
1459 LDLM_LOCK_RELEASE(lock);
1462 LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1469 * Main LDLM entry point for server code to process lock conversion requests.
1471 int ldlm_handle_convert0(struct ptlrpc_request *req,
1472 const struct ldlm_request *dlm_req)
1474 struct obd_export *exp = req->rq_export;
1475 struct ldlm_reply *dlm_rep;
1476 struct ldlm_lock *lock;
1481 if (exp && exp->exp_nid_stats && exp->exp_nid_stats->nid_ldlm_stats)
1482 lprocfs_counter_incr(exp->exp_nid_stats->nid_ldlm_stats,
1483 LDLM_CONVERT - LDLM_FIRST_OPC);
1485 rc = req_capsule_server_pack(&req->rq_pill);
1489 dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1490 dlm_rep->lock_flags = dlm_req->lock_flags;
1492 lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1497 bits = lock->l_policy_data.l_inodebits.bits;
1498 new = dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1499 LDLM_DEBUG(lock, "server-side convert handler START");
1501 if (ldlm_is_cancel(lock)) {
1502 LDLM_ERROR(lock, "convert on canceled lock!");
1503 req->rq_status = LUSTRE_EINVAL;
1504 } else if (dlm_req->lock_desc.l_req_mode !=
1505 lock->l_granted_mode) {
1506 LDLM_ERROR(lock, "lock mode differs!");
1507 req->rq_status = LUSTRE_EINVAL;
1508 } else if (bits == new) {
1509 /* This can be valid situation if CONVERT RPCs are
1510 * re-ordered. Just finish silently*/
1511 LDLM_DEBUG(lock, "lock is converted already!");
1514 lock_res_and_lock(lock);
1515 if (ldlm_is_waited(lock))
1516 ldlm_del_waiting_lock(lock);
1518 ldlm_clear_cbpending(lock);
1519 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1520 ldlm_inodebits_drop(lock, bits & ~new);
1521 lock->l_bl_ast_run = 0;
1522 ldlm_clear_ast_sent(lock);
1523 unlock_res_and_lock(lock);
1526 dlm_rep->lock_handle = lock->l_remote_handle;
1527 ldlm_ibits_policy_local_to_wire(&lock->l_policy_data,
1528 &dlm_rep->lock_desc.l_policy_data);
1530 ldlm_reprocess_all(lock->l_resource);
1532 LDLM_DEBUG(lock, "server-side convert handler END");
1533 LDLM_LOCK_PUT(lock);
1535 req->rq_status = LUSTRE_EINVAL;
1536 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1543 * Cancel all the locks whose handles are packed into ldlm_request
1545 * Called by server code expecting such combined cancel activity
1548 int ldlm_request_cancel(struct ptlrpc_request *req,
1549 const struct ldlm_request *dlm_req,
1550 int first, enum lustre_at_flags flags)
1552 struct ldlm_resource *res, *pres = NULL;
1553 struct ldlm_lock *lock;
1554 int i, count, done = 0;
1557 count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1561 if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1564 /* There is no lock on the server at the replay time,
1565 * skip lock cancelling to make replay tests to pass. */
1566 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1569 LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1570 "starting at %d", count, first);
1572 for (i = first; i < count; i++) {
1573 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1575 LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1576 "lock (cookie %llu)",
1577 dlm_req->lock_handle[i].cookie);
1581 res = lock->l_resource;
1584 /* This code is an optimization to only attempt lock
1585 * granting on the resource (that could be CPU-expensive)
1586 * after we are done cancelling lock in that resource. */
1589 ldlm_reprocess_all(pres);
1590 LDLM_RESOURCE_DELREF(pres);
1591 ldlm_resource_putref(pres);
1594 ldlm_resource_getref(res);
1595 LDLM_RESOURCE_ADDREF(res);
1597 if (!ldlm_is_discard_data(lock))
1598 ldlm_lvbo_update(res, lock, NULL, 1);
1603 if ((flags & LATF_STATS) && ldlm_is_ast_sent(lock)) {
1604 time64_t delay = ktime_get_real_seconds() -
1605 lock->l_last_activity;
1606 LDLM_DEBUG(lock, "server cancels blocked lock after %llds",
1608 at_measured(&lock->l_export->exp_bl_lock_at, delay);
1610 ldlm_lock_cancel(lock);
1611 LDLM_LOCK_PUT(lock);
1614 ldlm_reprocess_all(pres);
1615 LDLM_RESOURCE_DELREF(pres);
1616 ldlm_resource_putref(pres);
1618 LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1621 EXPORT_SYMBOL(ldlm_request_cancel);
1624 * Main LDLM entry point for server code to cancel locks.
1626 * Typically gets called from service handler on LDLM_CANCEL opc.
1628 int ldlm_handle_cancel(struct ptlrpc_request *req)
1630 struct ldlm_request *dlm_req;
1634 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1635 if (dlm_req == NULL) {
1636 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1640 if (req->rq_export && req->rq_export->exp_nid_stats &&
1641 req->rq_export->exp_nid_stats->nid_ldlm_stats)
1642 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1643 LDLM_CANCEL - LDLM_FIRST_OPC);
1645 rc = req_capsule_server_pack(&req->rq_pill);
1649 if (!ldlm_request_cancel(req, dlm_req, 0, LATF_STATS))
1650 req->rq_status = LUSTRE_ESTALE;
1652 RETURN(ptlrpc_reply(req));
1654 #endif /* HAVE_SERVER_SUPPORT */
1657 * Callback handler for receiving incoming blocking ASTs.
1659 * This can only happen on client side.
1661 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1662 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1667 LDLM_DEBUG(lock, "client blocking AST callback handler");
1669 lock_res_and_lock(lock);
1671 /* set bits to cancel for this lock for possible lock convert */
1672 if (ns_is_client(ns) && (lock->l_resource->lr_type == LDLM_IBITS)) {
1673 /* Lock description contains policy of blocking lock,
1674 * and its cancel_bits is used to pass conflicting bits.
1675 * NOTE: ld can be NULL or can be not NULL but zeroed if
1676 * passed from ldlm_bl_thread_blwi(), check below used bits
1677 * in ld to make sure it is valid description.
1679 if (ld && ld->l_policy_data.l_inodebits.bits)
1680 lock->l_policy_data.l_inodebits.cancel_bits =
1681 ld->l_policy_data.l_inodebits.cancel_bits;
1682 /* if there is no valid ld and lock is cbpending already
1683 * then cancel_bits should be kept, otherwise it is zeroed.
1685 else if (!ldlm_is_cbpending(lock))
1686 lock->l_policy_data.l_inodebits.cancel_bits = 0;
1688 ldlm_set_cbpending(lock);
1690 if (ldlm_is_cancel_on_block(lock))
1691 ldlm_set_cancel(lock);
1693 do_ast = (!lock->l_readers && !lock->l_writers);
1694 unlock_res_and_lock(lock);
1697 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1698 lock, lock->l_blocking_ast);
1699 if (lock->l_blocking_ast != NULL)
1700 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1703 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1707 LDLM_DEBUG(lock, "client blocking callback handler END");
1708 LDLM_LOCK_RELEASE(lock);
1713 * Callback handler for receiving incoming completion ASTs.
1715 * This only can happen on client side.
1717 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1718 struct ldlm_namespace *ns,
1719 struct ldlm_request *dlm_req,
1720 struct ldlm_lock *lock)
1722 struct list_head ast_list;
1727 LDLM_DEBUG(lock, "client completion callback handler START");
1729 INIT_LIST_HEAD(&ast_list);
1730 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1731 long to = cfs_time_seconds(1);
1734 set_current_state(TASK_INTERRUPTIBLE);
1735 schedule_timeout(to);
1736 if (lock->l_granted_mode == lock->l_req_mode ||
1737 ldlm_is_destroyed(lock))
1742 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1744 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
1745 GOTO(out, rc = lvb_len);
1746 } else if (lvb_len > 0) {
1747 if (lock->l_lvb_len > 0) {
1748 /* for extent lock, lvb contains ost_lvb{}. */
1749 LASSERT(lock->l_lvb_data != NULL);
1751 if (unlikely(lock->l_lvb_len < lvb_len)) {
1752 LDLM_ERROR(lock, "Replied LVB is larger than "
1753 "expectation, expected = %d, "
1755 lock->l_lvb_len, lvb_len);
1756 GOTO(out, rc = -EINVAL);
1761 lock_res_and_lock(lock);
1762 if (ldlm_is_destroyed(lock) ||
1763 lock->l_granted_mode == lock->l_req_mode) {
1764 /* bug 11300: the lock has already been granted */
1765 unlock_res_and_lock(lock);
1766 LDLM_DEBUG(lock, "Double grant race happened");
1770 /* If we receive the completion AST before the actual enqueue returned,
1771 * then we might need to switch lock modes, resources, or extents. */
1772 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1773 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1774 LDLM_DEBUG(lock, "completion AST, new lock mode");
1777 if (lock->l_resource->lr_type != LDLM_PLAIN) {
1778 ldlm_convert_policy_to_local(req->rq_export,
1779 dlm_req->lock_desc.l_resource.lr_type,
1780 &dlm_req->lock_desc.l_policy_data,
1781 &lock->l_policy_data);
1782 LDLM_DEBUG(lock, "completion AST, new policy data");
1785 ldlm_resource_unlink_lock(lock);
1786 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1787 &lock->l_resource->lr_name,
1788 sizeof(lock->l_resource->lr_name)) != 0) {
1789 unlock_res_and_lock(lock);
1790 rc = ldlm_lock_change_resource(ns, lock,
1791 &dlm_req->lock_desc.l_resource.lr_name);
1793 LDLM_ERROR(lock, "Failed to allocate resource");
1796 LDLM_DEBUG(lock, "completion AST, new resource");
1797 CERROR("change resource!\n");
1798 lock_res_and_lock(lock);
1801 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1802 /* BL_AST locks are not needed in LRU.
1803 * Let ldlm_cancel_lru() be fast. */
1804 ldlm_lock_remove_from_lru(lock);
1805 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1806 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1809 if (lock->l_lvb_len > 0) {
1810 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
1811 lock->l_lvb_data, lvb_len);
1813 unlock_res_and_lock(lock);
1818 ldlm_grant_lock(lock, &ast_list);
1819 unlock_res_and_lock(lock);
1821 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1823 /* Let Enqueue to call osc_lock_upcall() and initialize
1825 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1827 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1829 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1835 lock_res_and_lock(lock);
1836 ldlm_set_failed(lock);
1837 unlock_res_and_lock(lock);
1838 wake_up(&lock->l_waitq);
1840 LDLM_LOCK_RELEASE(lock);
1844 * Callback handler for receiving incoming glimpse ASTs.
1846 * This only can happen on client side. After handling the glimpse AST
1847 * we also consider dropping the lock here if it is unused locally for a
1850 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1851 struct ldlm_namespace *ns,
1852 struct ldlm_request *dlm_req,
1853 struct ldlm_lock *lock)
1858 LDLM_DEBUG(lock, "client glimpse AST callback handler");
1860 if (lock->l_glimpse_ast != NULL)
1861 rc = lock->l_glimpse_ast(lock, req);
1863 if (req->rq_repmsg != NULL) {
1866 req->rq_status = rc;
1870 lock_res_and_lock(lock);
1871 if (lock->l_granted_mode == LCK_PW &&
1872 !lock->l_readers && !lock->l_writers &&
1873 ktime_after(ktime_get(),
1874 ktime_add(lock->l_last_used,
1875 ktime_set(ns->ns_dirty_age_limit, 0)))) {
1876 unlock_res_and_lock(lock);
1877 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1878 ldlm_handle_bl_callback(ns, NULL, lock);
1883 unlock_res_and_lock(lock);
1884 LDLM_LOCK_RELEASE(lock);
1888 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1890 if (req->rq_no_reply)
1893 req->rq_status = rc;
1894 if (!req->rq_packed_final) {
1895 rc = lustre_pack_reply(req, 1, NULL, NULL);
1899 return ptlrpc_reply(req);
1902 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
1903 enum ldlm_cancel_flags cancel_flags)
1905 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1908 spin_lock(&blp->blp_lock);
1909 if (blwi->blwi_lock &&
1910 ldlm_is_discard_data(blwi->blwi_lock)) {
1911 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1912 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1914 /* other blocking callbacks are added to the regular list */
1915 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1917 spin_unlock(&blp->blp_lock);
1919 wake_up(&blp->blp_waitq);
1921 /* can not check blwi->blwi_flags as blwi could be already freed in
1923 if (!(cancel_flags & LCF_ASYNC))
1924 wait_for_completion(&blwi->blwi_comp);
1929 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1930 struct ldlm_namespace *ns,
1931 struct ldlm_lock_desc *ld,
1932 struct list_head *cancels, int count,
1933 struct ldlm_lock *lock,
1934 enum ldlm_cancel_flags cancel_flags)
1936 init_completion(&blwi->blwi_comp);
1937 INIT_LIST_HEAD(&blwi->blwi_head);
1939 if (memory_pressure_get())
1940 blwi->blwi_mem_pressure = 1;
1943 blwi->blwi_flags = cancel_flags;
1945 blwi->blwi_ld = *ld;
1947 list_add(&blwi->blwi_head, cancels);
1948 list_del_init(cancels);
1949 blwi->blwi_count = count;
1951 blwi->blwi_lock = lock;
1956 * Queues a list of locks \a cancels containing \a count locks
1957 * for later processing by a blocking thread. If \a count is zero,
1958 * then the lock referenced as \a lock is queued instead.
1960 * The blocking thread would then call ->l_blocking_ast callback in the lock.
1961 * If list addition fails an error is returned and caller is supposed to
1962 * call ->l_blocking_ast itself.
1964 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1965 struct ldlm_lock_desc *ld,
1966 struct ldlm_lock *lock,
1967 struct list_head *cancels, int count,
1968 enum ldlm_cancel_flags cancel_flags)
1972 if (cancels && count == 0)
1975 if (cancel_flags & LCF_ASYNC) {
1976 struct ldlm_bl_work_item *blwi;
1978 OBD_ALLOC(blwi, sizeof(*blwi));
1981 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
1983 RETURN(__ldlm_bl_to_thread(blwi, cancel_flags));
1985 /* if it is synchronous call do minimum mem alloc, as it could
1986 * be triggered from kernel shrinker
1988 struct ldlm_bl_work_item blwi;
1990 memset(&blwi, 0, sizeof(blwi));
1991 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
1992 RETURN(__ldlm_bl_to_thread(&blwi, cancel_flags));
1997 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1998 struct ldlm_lock *lock)
2000 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
2003 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
2004 struct list_head *cancels, int count,
2005 enum ldlm_cancel_flags cancel_flags)
2007 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
2010 int ldlm_bl_thread_wakeup(void)
2012 wake_up(&ldlm_state->ldlm_bl_pool->blp_waitq);
2016 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
2017 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
2019 struct obd_device *obd = req->rq_export->exp_obd;
2026 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
2028 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2030 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2032 DEBUG_REQ(D_IOCTL, req, "no set_info key");
2035 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
2037 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2039 DEBUG_REQ(D_IOCTL, req, "no set_info val");
2042 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
2045 /* We are responsible for swabbing contents of val */
2047 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
2048 /* Pass it on to mdc (the "export" in this case) */
2049 rc = obd_set_info_async(req->rq_svc_thread->t_env,
2051 sizeof(KEY_HSM_COPYTOOL_SEND),
2052 KEY_HSM_COPYTOOL_SEND,
2055 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
2060 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
2061 const char *msg, int rc,
2062 const struct lustre_handle *handle)
2064 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
2065 "%s: [nid %s] [rc %d] [lock %#llx]",
2066 msg, libcfs_id2str(req->rq_peer), rc,
2067 handle ? handle->cookie : 0);
2068 if (req->rq_no_reply)
2069 CWARN("No reply was sent, maybe cause bug 21636.\n");
2071 CWARN("Send reply failed, maybe cause bug 21636.\n");
2074 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2075 static int ldlm_callback_handler(struct ptlrpc_request *req)
2077 struct ldlm_namespace *ns;
2078 struct ldlm_request *dlm_req;
2079 struct ldlm_lock *lock;
2083 /* Requests arrive in sender's byte order. The ptlrpc service
2084 * handler has already checked and, if necessary, byte-swapped the
2085 * incoming request message body, but I am responsible for the
2086 * message buffers. */
2088 /* do nothing for sec context finalize */
2089 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
2092 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2094 if (req->rq_export == NULL) {
2095 rc = ldlm_callback_reply(req, -ENOTCONN);
2096 ldlm_callback_errmsg(req, "Operate on unconnected server",
2101 LASSERT(req->rq_export != NULL);
2102 LASSERT(req->rq_export->exp_obd != NULL);
2104 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2105 case LDLM_BL_CALLBACK:
2106 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
2108 ldlm_callback_reply(req, -(int)cfs_fail_err);
2112 case LDLM_CP_CALLBACK:
2113 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2116 case LDLM_GL_CALLBACK:
2117 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2121 rc = ldlm_handle_setinfo(req);
2122 ldlm_callback_reply(req, rc);
2124 case LLOG_ORIGIN_HANDLE_CREATE:
2125 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2126 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2128 rc = llog_origin_handle_open(req);
2129 ldlm_callback_reply(req, rc);
2131 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2132 req_capsule_set(&req->rq_pill,
2133 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2134 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2136 rc = llog_origin_handle_next_block(req);
2137 ldlm_callback_reply(req, rc);
2139 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2140 req_capsule_set(&req->rq_pill,
2141 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2142 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2144 rc = llog_origin_handle_read_header(req);
2145 ldlm_callback_reply(req, rc);
2147 case LLOG_ORIGIN_HANDLE_CLOSE:
2148 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2150 rc = llog_origin_handle_close(req);
2151 ldlm_callback_reply(req, rc);
2154 CERROR("unknown opcode %u\n",
2155 lustre_msg_get_opc(req->rq_reqmsg));
2156 ldlm_callback_reply(req, -EPROTO);
2160 ns = req->rq_export->exp_obd->obd_namespace;
2161 LASSERT(ns != NULL);
2163 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2165 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2166 if (dlm_req == NULL) {
2167 rc = ldlm_callback_reply(req, -EPROTO);
2168 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2173 /* Force a known safe race, send a cancel to the server for a lock
2174 * which the server has already started a blocking callback on. */
2175 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2176 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2177 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
2179 CERROR("ldlm_cli_cancel: %d\n", rc);
2182 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2184 CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock "
2185 "disappeared\n", dlm_req->lock_handle[0].cookie);
2186 rc = ldlm_callback_reply(req, -EINVAL);
2187 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2188 &dlm_req->lock_handle[0]);
2192 if (ldlm_is_fail_loc(lock) &&
2193 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2194 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2196 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2197 lock_res_and_lock(lock);
2198 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2200 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2201 /* If somebody cancels lock and cache is already dropped,
2202 * or lock is failed before cp_ast received on client,
2203 * we can tell the server we have no lock. Otherwise, we
2204 * should send cancel after dropping the cache. */
2205 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
2206 ldlm_is_failed(lock)) {
2207 LDLM_DEBUG(lock, "callback on lock %llx - lock disappeared",
2208 dlm_req->lock_handle[0].cookie);
2209 unlock_res_and_lock(lock);
2210 LDLM_LOCK_RELEASE(lock);
2211 rc = ldlm_callback_reply(req, -EINVAL);
2212 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2213 &dlm_req->lock_handle[0]);
2216 /* BL_AST locks are not needed in LRU.
2217 * Let ldlm_cancel_lru() be fast. */
2218 ldlm_lock_remove_from_lru(lock);
2219 ldlm_set_bl_ast(lock);
2221 unlock_res_and_lock(lock);
2223 /* We want the ost thread to get this reply so that it can respond
2224 * to ost requests (write cache writeback) that might be triggered
2227 * But we'd also like to be able to indicate in the reply that we're
2228 * cancelling right now, because it's unused, or have an intent result
2229 * in the reply, so we might have to push the responsibility for sending
2230 * the reply down into the AST handlers, alas. */
2232 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2233 case LDLM_BL_CALLBACK:
2234 CDEBUG(D_INODE, "blocking ast\n");
2235 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2236 if (!ldlm_is_cancel_on_block(lock)) {
2237 rc = ldlm_callback_reply(req, 0);
2238 if (req->rq_no_reply || rc)
2239 ldlm_callback_errmsg(req, "Normal process", rc,
2240 &dlm_req->lock_handle[0]);
2242 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2243 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2245 case LDLM_CP_CALLBACK:
2246 CDEBUG(D_INODE, "completion ast\n");
2247 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2248 ldlm_callback_reply(req, 0);
2249 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2251 case LDLM_GL_CALLBACK:
2252 CDEBUG(D_INODE, "glimpse ast\n");
2253 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2254 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2257 LBUG(); /* checked above */
2263 #ifdef HAVE_SERVER_SUPPORT
2265 * Main handler for canceld thread.
2267 * Separated into its own thread to avoid deadlocks.
2269 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2274 /* Requests arrive in sender's byte order. The ptlrpc service
2275 * handler has already checked and, if necessary, byte-swapped the
2276 * incoming request message body, but I am responsible for the
2277 * message buffers. */
2279 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2281 if (req->rq_export == NULL) {
2282 struct ldlm_request *dlm_req;
2284 CERROR("%s from %s arrived at %lu with bad export cookie "
2286 ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2287 libcfs_nid2str(req->rq_peer.nid),
2288 req->rq_arrival_time.tv_sec,
2289 lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2291 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2292 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2293 dlm_req = req_capsule_client_get(&req->rq_pill,
2295 if (dlm_req != NULL)
2296 ldlm_lock_dump_handle(D_ERROR,
2297 &dlm_req->lock_handle[0]);
2299 ldlm_callback_reply(req, -ENOTCONN);
2303 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2305 /* XXX FIXME move this back to mds/handler.c, bug 249 */
2307 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2308 CDEBUG(D_INODE, "cancel\n");
2309 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2310 CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND) ||
2311 CFS_FAIL_CHECK(OBD_FAIL_LDLM_BL_EVICT))
2313 rc = ldlm_handle_cancel(req);
2318 CERROR("invalid opcode %d\n",
2319 lustre_msg_get_opc(req->rq_reqmsg));
2320 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2321 ldlm_callback_reply(req, -EINVAL);
2327 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2328 struct ldlm_lock *lock)
2330 struct ldlm_request *dlm_req;
2331 struct lustre_handle lockh;
2336 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2337 if (dlm_req == NULL)
2340 ldlm_lock2handle(lock, &lockh);
2341 for (i = 0; i < dlm_req->lock_count; i++) {
2342 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2344 DEBUG_REQ(D_RPCTRACE, req,
2345 "Prio raised by lock %#llx.", lockh.cookie);
2356 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2358 struct ldlm_request *dlm_req;
2363 /* no prolong in recovery */
2364 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2367 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2368 if (dlm_req == NULL)
2371 for (i = 0; i < dlm_req->lock_count; i++) {
2372 struct ldlm_lock *lock;
2374 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2378 rc = ldlm_is_ast_sent(lock) ? 1 : 0;
2380 LDLM_DEBUG(lock, "hpreq cancel lock");
2381 LDLM_LOCK_PUT(lock);
2390 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2391 .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2392 .hpreq_check = ldlm_cancel_hpreq_check,
2396 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2400 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2402 if (req->rq_export == NULL)
2405 if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2406 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2407 req->rq_ops = &ldlm_cancel_hpreq_ops;
2412 static int ldlm_revoke_lock_cb(struct cfs_hash *hs, struct cfs_hash_bd *bd,
2413 struct hlist_node *hnode, void *data)
2416 struct list_head *rpc_list = data;
2417 struct ldlm_lock *lock = cfs_hash_object(hs, hnode);
2419 lock_res_and_lock(lock);
2421 if (lock->l_req_mode != lock->l_granted_mode) {
2422 unlock_res_and_lock(lock);
2426 LASSERT(lock->l_resource);
2427 if (lock->l_resource->lr_type != LDLM_IBITS &&
2428 lock->l_resource->lr_type != LDLM_PLAIN) {
2429 unlock_res_and_lock(lock);
2433 if (ldlm_is_ast_sent(lock)) {
2434 unlock_res_and_lock(lock);
2438 LASSERT(lock->l_blocking_ast);
2439 LASSERT(!lock->l_blocking_lock);
2441 ldlm_set_ast_sent(lock);
2442 if (lock->l_export && lock->l_export->exp_lock_hash) {
2443 /* NB: it's safe to call cfs_hash_del() even lock isn't
2444 * in exp_lock_hash. */
2445 /* In the function below, .hs_keycmp resolves to
2446 * ldlm_export_lock_keycmp() */
2447 /* coverity[overrun-buffer-val] */
2448 cfs_hash_del(lock->l_export->exp_lock_hash,
2449 &lock->l_remote_handle, &lock->l_exp_hash);
2452 list_add_tail(&lock->l_rk_ast, rpc_list);
2453 LDLM_LOCK_GET(lock);
2455 unlock_res_and_lock(lock);
2459 void ldlm_revoke_export_locks(struct obd_export *exp)
2461 struct list_head rpc_list;
2464 INIT_LIST_HEAD(&rpc_list);
2465 cfs_hash_for_each_nolock(exp->exp_lock_hash,
2466 ldlm_revoke_lock_cb, &rpc_list, 0);
2467 ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2468 LDLM_WORK_REVOKE_AST);
2472 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2473 #endif /* HAVE_SERVER_SUPPORT */
2475 static int ldlm_bl_get_work(struct ldlm_bl_pool *blp,
2476 struct ldlm_bl_work_item **p_blwi,
2477 struct obd_export **p_exp)
2479 struct ldlm_bl_work_item *blwi = NULL;
2480 static unsigned int num_bl = 0;
2481 static unsigned int num_stale;
2482 int num_th = atomic_read(&blp->blp_num_threads);
2484 *p_exp = obd_stale_export_get();
2486 spin_lock(&blp->blp_lock);
2487 if (*p_exp != NULL) {
2488 if (num_th == 1 || ++num_stale < num_th) {
2489 spin_unlock(&blp->blp_lock);
2496 /* process a request from the blp_list at least every blp_num_threads */
2497 if (!list_empty(&blp->blp_list) &&
2498 (list_empty(&blp->blp_prio_list) || num_bl == 0))
2499 blwi = list_entry(blp->blp_list.next,
2500 struct ldlm_bl_work_item, blwi_entry);
2502 if (!list_empty(&blp->blp_prio_list))
2503 blwi = list_entry(blp->blp_prio_list.next,
2504 struct ldlm_bl_work_item,
2508 if (++num_bl >= num_th)
2510 list_del(&blwi->blwi_entry);
2512 spin_unlock(&blp->blp_lock);
2515 if (*p_exp != NULL && *p_blwi != NULL) {
2516 obd_stale_export_put(*p_exp);
2520 return (*p_blwi != NULL || *p_exp != NULL) ? 1 : 0;
2523 /* This only contains temporary data until the thread starts */
2524 struct ldlm_bl_thread_data {
2525 struct ldlm_bl_pool *bltd_blp;
2526 struct completion bltd_comp;
2530 static int ldlm_bl_thread_main(void *arg);
2532 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp, bool check_busy)
2534 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2535 struct task_struct *task;
2537 init_completion(&bltd.bltd_comp);
2539 bltd.bltd_num = atomic_inc_return(&blp->blp_num_threads);
2540 if (bltd.bltd_num >= blp->blp_max_threads) {
2541 atomic_dec(&blp->blp_num_threads);
2545 LASSERTF(bltd.bltd_num > 0, "thread num:%d\n", bltd.bltd_num);
2547 atomic_read(&blp->blp_busy_threads) < (bltd.bltd_num - 1)) {
2548 atomic_dec(&blp->blp_num_threads);
2552 task = kthread_run(ldlm_bl_thread_main, &bltd, "ldlm_bl_%02d",
2555 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
2556 bltd.bltd_num, PTR_ERR(task));
2557 atomic_dec(&blp->blp_num_threads);
2558 return PTR_ERR(task);
2560 wait_for_completion(&bltd.bltd_comp);
2565 /* Not fatal if racy and have a few too many threads */
2566 static int ldlm_bl_thread_need_create(struct ldlm_bl_pool *blp,
2567 struct ldlm_bl_work_item *blwi)
2569 if (atomic_read(&blp->blp_num_threads) >= blp->blp_max_threads)
2572 if (atomic_read(&blp->blp_busy_threads) <
2573 atomic_read(&blp->blp_num_threads))
2576 if (blwi != NULL && (blwi->blwi_ns == NULL ||
2577 blwi->blwi_mem_pressure))
2583 static int ldlm_bl_thread_blwi(struct ldlm_bl_pool *blp,
2584 struct ldlm_bl_work_item *blwi)
2588 if (blwi->blwi_ns == NULL)
2589 /* added by ldlm_cleanup() */
2590 RETURN(LDLM_ITER_STOP);
2592 if (blwi->blwi_mem_pressure)
2593 memory_pressure_set();
2595 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL2, 4);
2597 if (blwi->blwi_count) {
2599 /* The special case when we cancel locks in lru
2600 * asynchronously, we pass the list of locks here.
2601 * Thus locks are marked LDLM_FL_CANCELING, but NOT
2602 * canceled locally yet. */
2603 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2606 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
2609 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2612 if (blwi->blwi_mem_pressure)
2613 memory_pressure_clr();
2615 if (blwi->blwi_flags & LCF_ASYNC)
2616 OBD_FREE(blwi, sizeof(*blwi));
2618 complete(&blwi->blwi_comp);
2624 * Cancel stale locks on export. Cancel blocked locks first.
2625 * If the given export has blocked locks, the next in the list may have
2626 * them too, thus cancel not blocked locks only if the current export has
2629 static int ldlm_bl_thread_exports(struct ldlm_bl_pool *blp,
2630 struct obd_export *exp)
2635 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_BL_EVICT, 4);
2637 num = ldlm_export_cancel_blocked_locks(exp);
2639 ldlm_export_cancel_locks(exp);
2641 obd_stale_export_put(exp);
2648 * Main blocking requests processing thread.
2650 * Callers put locks into its queue by calling ldlm_bl_to_thread.
2651 * This thread in the end ends up doing actual call to ->l_blocking_ast
2654 static int ldlm_bl_thread_main(void *arg)
2656 struct ldlm_bl_pool *blp;
2657 struct ldlm_bl_thread_data *bltd = arg;
2660 blp = bltd->bltd_blp;
2662 complete(&bltd->bltd_comp);
2663 /* cannot use bltd after this, it is only on caller's stack */
2666 struct l_wait_info lwi = { 0 };
2667 struct ldlm_bl_work_item *blwi = NULL;
2668 struct obd_export *exp = NULL;
2671 rc = ldlm_bl_get_work(blp, &blwi, &exp);
2674 l_wait_event_exclusive(blp->blp_waitq,
2675 ldlm_bl_get_work(blp, &blwi,
2678 atomic_inc(&blp->blp_busy_threads);
2680 if (ldlm_bl_thread_need_create(blp, blwi))
2681 /* discard the return value, we tried */
2682 ldlm_bl_thread_start(blp, true);
2685 rc = ldlm_bl_thread_exports(blp, exp);
2687 rc = ldlm_bl_thread_blwi(blp, blwi);
2689 atomic_dec(&blp->blp_busy_threads);
2691 if (rc == LDLM_ITER_STOP)
2694 /* If there are many namespaces, we will not sleep waiting for
2695 * work, and must do a cond_resched to avoid holding the CPU
2700 atomic_dec(&blp->blp_num_threads);
2701 complete(&blp->blp_comp);
2706 static int ldlm_setup(void);
2707 static int ldlm_cleanup(void);
2709 int ldlm_get_ref(void)
2713 mutex_lock(&ldlm_ref_mutex);
2714 if (++ldlm_refcount == 1) {
2719 mutex_unlock(&ldlm_ref_mutex);
2724 void ldlm_put_ref(void)
2727 mutex_lock(&ldlm_ref_mutex);
2728 if (ldlm_refcount == 1) {
2729 int rc = ldlm_cleanup();
2731 CERROR("ldlm_cleanup failed: %d\n", rc);
2737 mutex_unlock(&ldlm_ref_mutex);
2743 * Export handle<->lock hash operations.
2746 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
2748 return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2752 ldlm_export_lock_key(struct hlist_node *hnode)
2754 struct ldlm_lock *lock;
2756 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2757 return &lock->l_remote_handle;
2761 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
2763 struct ldlm_lock *lock;
2765 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2766 lock->l_remote_handle = *(struct lustre_handle *)key;
2770 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
2772 return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2776 ldlm_export_lock_object(struct hlist_node *hnode)
2778 return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2782 ldlm_export_lock_get(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_GET(lock);
2791 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
2793 struct ldlm_lock *lock;
2795 lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2796 LDLM_LOCK_RELEASE(lock);
2799 static struct cfs_hash_ops ldlm_export_lock_ops = {
2800 .hs_hash = ldlm_export_lock_hash,
2801 .hs_key = ldlm_export_lock_key,
2802 .hs_keycmp = ldlm_export_lock_keycmp,
2803 .hs_keycpy = ldlm_export_lock_keycpy,
2804 .hs_object = ldlm_export_lock_object,
2805 .hs_get = ldlm_export_lock_get,
2806 .hs_put = ldlm_export_lock_put,
2807 .hs_put_locked = ldlm_export_lock_put,
2810 int ldlm_init_export(struct obd_export *exp)
2815 exp->exp_lock_hash =
2816 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2817 HASH_EXP_LOCK_CUR_BITS,
2818 HASH_EXP_LOCK_MAX_BITS,
2819 HASH_EXP_LOCK_BKT_BITS, 0,
2820 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2821 &ldlm_export_lock_ops,
2822 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2823 CFS_HASH_NBLK_CHANGE);
2825 if (!exp->exp_lock_hash)
2828 rc = ldlm_init_flock_export(exp);
2834 ldlm_destroy_export(exp);
2837 EXPORT_SYMBOL(ldlm_init_export);
2839 void ldlm_destroy_export(struct obd_export *exp)
2842 cfs_hash_putref(exp->exp_lock_hash);
2843 exp->exp_lock_hash = NULL;
2845 ldlm_destroy_flock_export(exp);
2848 EXPORT_SYMBOL(ldlm_destroy_export);
2850 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
2851 struct attribute *attr,
2854 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
2857 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
2858 struct attribute *attr,
2865 rc = kstrtoul(buffer, 10, &val);
2869 ldlm_cancel_unused_locks_before_replay = val;
2873 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
2875 static struct attribute *ldlm_attrs[] = {
2876 &lustre_attr_cancel_unused_locks_before_replay.attr,
2880 static struct attribute_group ldlm_attr_group = {
2881 .attrs = ldlm_attrs,
2884 static int ldlm_setup(void)
2886 static struct ptlrpc_service_conf conf;
2887 struct ldlm_bl_pool *blp = NULL;
2888 #ifdef HAVE_SERVER_SUPPORT
2889 struct task_struct *task;
2890 #endif /* HAVE_SERVER_SUPPORT */
2896 if (ldlm_state != NULL)
2899 OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2900 if (ldlm_state == NULL)
2903 ldlm_kobj = kobject_create_and_add("ldlm", &lustre_kset->kobj);
2907 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
2911 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
2915 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
2919 rc = ldlm_debugfs_setup();
2923 memset(&conf, 0, sizeof(conf));
2924 conf = (typeof(conf)) {
2925 .psc_name = "ldlm_cbd",
2926 .psc_watchdog_factor = 2,
2928 .bc_nbufs = LDLM_CLIENT_NBUFS,
2929 .bc_buf_size = LDLM_BUFSIZE,
2930 .bc_req_max_size = LDLM_MAXREQSIZE,
2931 .bc_rep_max_size = LDLM_MAXREPSIZE,
2932 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
2933 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
2936 .tc_thr_name = "ldlm_cb",
2937 .tc_thr_factor = LDLM_THR_FACTOR,
2938 .tc_nthrs_init = LDLM_NTHRS_INIT,
2939 .tc_nthrs_base = LDLM_NTHRS_BASE,
2940 .tc_nthrs_max = LDLM_NTHRS_MAX,
2941 .tc_nthrs_user = ldlm_num_threads,
2942 .tc_cpu_affinity = 1,
2943 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
2946 .cc_pattern = ldlm_cpts,
2949 .so_req_handler = ldlm_callback_handler,
2952 ldlm_state->ldlm_cb_service = \
2953 ptlrpc_register_service(&conf, ldlm_svc_kset,
2954 ldlm_svc_debugfs_dir);
2955 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2956 CERROR("failed to start service\n");
2957 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2958 ldlm_state->ldlm_cb_service = NULL;
2962 #ifdef HAVE_SERVER_SUPPORT
2963 memset(&conf, 0, sizeof(conf));
2964 conf = (typeof(conf)) {
2965 .psc_name = "ldlm_canceld",
2966 .psc_watchdog_factor = 6,
2968 .bc_nbufs = LDLM_SERVER_NBUFS,
2969 .bc_buf_size = LDLM_BUFSIZE,
2970 .bc_req_max_size = LDLM_MAXREQSIZE,
2971 .bc_rep_max_size = LDLM_MAXREPSIZE,
2972 .bc_req_portal = LDLM_CANCEL_REQUEST_PORTAL,
2973 .bc_rep_portal = LDLM_CANCEL_REPLY_PORTAL,
2977 .tc_thr_name = "ldlm_cn",
2978 .tc_thr_factor = LDLM_THR_FACTOR,
2979 .tc_nthrs_init = LDLM_NTHRS_INIT,
2980 .tc_nthrs_base = LDLM_NTHRS_BASE,
2981 .tc_nthrs_max = LDLM_NTHRS_MAX,
2982 .tc_nthrs_user = ldlm_num_threads,
2983 .tc_cpu_affinity = 1,
2984 .tc_ctx_tags = LCT_MD_THREAD | \
2989 .cc_pattern = ldlm_cpts,
2992 .so_req_handler = ldlm_cancel_handler,
2993 .so_hpreq_handler = ldlm_hpreq_handler,
2996 ldlm_state->ldlm_cancel_service = \
2997 ptlrpc_register_service(&conf, ldlm_svc_kset,
2998 ldlm_svc_debugfs_dir);
2999 if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
3000 CERROR("failed to start service\n");
3001 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
3002 ldlm_state->ldlm_cancel_service = NULL;
3005 #endif /* HAVE_SERVER_SUPPORT */
3007 OBD_ALLOC(blp, sizeof(*blp));
3009 GOTO(out, rc = -ENOMEM);
3010 ldlm_state->ldlm_bl_pool = blp;
3012 spin_lock_init(&blp->blp_lock);
3013 INIT_LIST_HEAD(&blp->blp_list);
3014 INIT_LIST_HEAD(&blp->blp_prio_list);
3015 init_waitqueue_head(&blp->blp_waitq);
3016 atomic_set(&blp->blp_num_threads, 0);
3017 atomic_set(&blp->blp_busy_threads, 0);
3019 if (ldlm_num_threads == 0) {
3020 blp->blp_min_threads = LDLM_NTHRS_INIT;
3021 blp->blp_max_threads = LDLM_NTHRS_MAX;
3023 blp->blp_min_threads = blp->blp_max_threads = \
3024 min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
3028 for (i = 0; i < blp->blp_min_threads; i++) {
3029 rc = ldlm_bl_thread_start(blp, false);
3034 #ifdef HAVE_SERVER_SUPPORT
3035 task = kthread_run(expired_lock_main, NULL, "ldlm_elt");
3038 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
3042 wait_event(expired_lock_wait_queue,
3043 expired_lock_thread_state == ELT_READY);
3044 #endif /* HAVE_SERVER_SUPPORT */
3046 rc = ldlm_pools_init();
3048 CERROR("Failed to initialize LDLM pools: %d\n", rc);
3052 rc = ldlm_reclaim_setup();
3054 CERROR("Failed to setup reclaim thread: rc = %d\n", rc);
3064 static int ldlm_cleanup(void)
3068 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
3069 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
3070 CERROR("ldlm still has namespaces; clean these up first.\n");
3071 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
3072 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
3076 ldlm_reclaim_cleanup();
3079 if (ldlm_state->ldlm_bl_pool != NULL) {
3080 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
3082 while (atomic_read(&blp->blp_num_threads) > 0) {
3083 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
3085 init_completion(&blp->blp_comp);
3087 spin_lock(&blp->blp_lock);
3088 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
3089 wake_up(&blp->blp_waitq);
3090 spin_unlock(&blp->blp_lock);
3092 wait_for_completion(&blp->blp_comp);
3095 OBD_FREE(blp, sizeof(*blp));
3098 if (ldlm_state->ldlm_cb_service != NULL)
3099 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
3100 #ifdef HAVE_SERVER_SUPPORT
3101 if (ldlm_state->ldlm_cancel_service != NULL)
3102 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
3106 kset_unregister(ldlm_ns_kset);
3108 kset_unregister(ldlm_svc_kset);
3110 sysfs_remove_group(ldlm_kobj, &ldlm_attr_group);
3111 kobject_put(ldlm_kobj);
3114 ldlm_debugfs_cleanup();
3116 #ifdef HAVE_SERVER_SUPPORT
3117 if (expired_lock_thread_state != ELT_STOPPED) {
3118 expired_lock_thread_state = ELT_TERMINATE;
3119 wake_up(&expired_lock_wait_queue);
3120 wait_event(expired_lock_wait_queue,
3121 expired_lock_thread_state == ELT_STOPPED);
3125 OBD_FREE(ldlm_state, sizeof(*ldlm_state));
3133 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
3134 sizeof(struct ldlm_resource), 0,
3135 SLAB_HWCACHE_ALIGN, NULL);
3136 if (ldlm_resource_slab == NULL)
3139 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
3140 sizeof(struct ldlm_lock), 0,
3141 SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
3142 if (ldlm_lock_slab == NULL)
3145 ldlm_interval_slab = kmem_cache_create("interval_node",
3146 sizeof(struct ldlm_interval),
3147 0, SLAB_HWCACHE_ALIGN, NULL);
3148 if (ldlm_interval_slab == NULL)
3151 ldlm_interval_tree_slab = kmem_cache_create("interval_tree",
3152 sizeof(struct ldlm_interval_tree) * LCK_MODE_NUM,
3153 0, SLAB_HWCACHE_ALIGN, NULL);
3154 if (ldlm_interval_tree_slab == NULL)
3157 #ifdef HAVE_SERVER_SUPPORT
3158 ldlm_glimpse_work_kmem = kmem_cache_create("ldlm_glimpse_work_kmem",
3159 sizeof(struct ldlm_glimpse_work),
3161 if (ldlm_glimpse_work_kmem == NULL)
3162 goto out_interval_tree;
3165 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3166 class_export_dump_hook = ldlm_dump_export_locks;
3169 #ifdef HAVE_SERVER_SUPPORT
3171 kmem_cache_destroy(ldlm_interval_tree_slab);
3174 kmem_cache_destroy(ldlm_interval_slab);
3176 kmem_cache_destroy(ldlm_lock_slab);
3178 kmem_cache_destroy(ldlm_resource_slab);
3183 void ldlm_exit(void)
3186 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
3187 kmem_cache_destroy(ldlm_resource_slab);
3188 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
3189 * synchronize_rcu() to wait a grace period elapsed, so that
3190 * ldlm_lock_free() get a chance to be called. */
3192 kmem_cache_destroy(ldlm_lock_slab);
3193 kmem_cache_destroy(ldlm_interval_slab);
3194 kmem_cache_destroy(ldlm_interval_tree_slab);
3195 #ifdef HAVE_SERVER_SUPPORT
3196 kmem_cache_destroy(ldlm_glimpse_work_kmem);