1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <liblustre.h>
43 #include <lustre_dlm.h>
44 #include <obd_class.h>
47 #include "ldlm_internal.h"
49 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
50 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
51 "lock enqueue timeout minimum");
53 /* in client side, whether the cached locks will be canceled before replay */
54 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
56 static void interrupted_completion_wait(void *data)
60 struct lock_wait_data {
61 struct ldlm_lock *lwd_lock;
65 struct ldlm_async_args {
66 struct lustre_handle lock_handle;
69 int ldlm_expired_completion_wait(void *data)
71 struct lock_wait_data *lwd = data;
72 struct ldlm_lock *lock = lwd->lwd_lock;
73 struct obd_import *imp;
74 struct obd_device *obd;
77 if (lock->l_conn_export == NULL) {
78 static cfs_time_t next_dump = 0, last_dump = 0;
80 if (ptlrpc_check_suspend())
83 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
84 CFS_DURATION_T"s ago); not entering recovery in "
85 "server code, just going back to sleep",
86 lock->l_last_activity,
87 cfs_time_sub(cfs_time_current_sec(),
88 lock->l_last_activity));
89 if (cfs_time_after(cfs_time_current(), next_dump)) {
90 last_dump = next_dump;
91 next_dump = cfs_time_shift(300);
92 ldlm_namespace_dump(D_DLMTRACE,
93 ldlm_lock_to_ns(lock));
95 libcfs_debug_dumplog();
100 obd = lock->l_conn_export->exp_obd;
101 imp = obd->u.cli.cl_import;
102 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
103 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
104 CFS_DURATION_T"s ago), entering recovery for %s@%s",
105 lock->l_last_activity,
106 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
107 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
112 /* We use the same basis for both server side and client side functions
113 from a single node. */
114 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
116 int timeout = at_get(ldlm_lock_to_ns_at(lock));
118 return obd_timeout / 2;
119 /* Since these are non-updating timeouts, we should be conservative.
120 It would be nice to have some kind of "early reply" mechanism for
121 lock callbacks too... */
122 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
123 return max(timeout, ldlm_enqueue_min);
125 EXPORT_SYMBOL(ldlm_get_enq_timeout);
128 * Helper function for ldlm_completion_ast(), updating timings when lock is
131 static int ldlm_completion_tail(struct ldlm_lock *lock)
136 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
137 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
140 delay = cfs_time_sub(cfs_time_current_sec(),
141 lock->l_last_activity);
142 LDLM_DEBUG(lock, "client-side enqueue: granted after "
143 CFS_DURATION_T"s", delay);
145 /* Update our time estimate */
146 at_measured(ldlm_lock_to_ns_at(lock),
154 * Implementation of ->l_completion_ast() for a client, that doesn't wait
155 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
156 * other threads that cannot block for long.
158 int ldlm_completion_ast_async(struct ldlm_lock *lock, int flags, void *data)
162 if (flags == LDLM_FL_WAIT_NOREPROC) {
163 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
167 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
168 LDLM_FL_BLOCK_CONV))) {
169 cfs_waitq_signal(&lock->l_waitq);
170 RETURN(ldlm_completion_tail(lock));
173 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
175 ldlm_lock_dump(D_OTHER, lock, 0);
176 ldlm_reprocess_all(lock->l_resource);
181 * Client side LDLM "completion" AST. This is called in several cases:
183 * - when a reply to an ENQUEUE rpc is received from the server
184 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
185 * this point (determined by flags);
187 * - when LDLM_CP_CALLBACK rpc comes to client to notify it that lock has
190 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
193 * - to force all locks when resource is destroyed (cleanup_resource());
195 * - during lock conversion (not used currently).
197 * If lock is not granted in the first case, this function waits until second
198 * or penultimate cases happen in some other thread.
201 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
203 /* XXX ALLOCATE - 160 bytes */
204 struct lock_wait_data lwd;
205 struct obd_device *obd;
206 struct obd_import *imp = NULL;
207 struct l_wait_info lwi;
212 if (flags == LDLM_FL_WAIT_NOREPROC) {
213 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
217 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
218 LDLM_FL_BLOCK_CONV))) {
219 cfs_waitq_signal(&lock->l_waitq);
223 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
225 ldlm_lock_dump(D_OTHER, lock, 0);
229 obd = class_exp2obd(lock->l_conn_export);
231 /* if this is a local lock, then there is no import */
233 imp = obd->u.cli.cl_import;
236 /* Wait a long time for enqueue - server may have to callback a
237 lock from another client. Server will evict the other client if it
238 doesn't respond reasonably, and then give us the lock. */
239 timeout = ldlm_get_enq_timeout(lock) * 2;
243 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
244 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
245 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
247 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
248 ldlm_expired_completion_wait,
249 interrupted_completion_wait, &lwd);
253 cfs_spin_lock(&imp->imp_lock);
254 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
255 cfs_spin_unlock(&imp->imp_lock);
258 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
259 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
260 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
261 lock->l_flags |= LDLM_FL_FAIL_LOC;
264 /* Go to sleep until the lock is granted or cancelled. */
265 rc = l_wait_event(lock->l_waitq,
266 is_granted_or_cancelled(lock), &lwi);
270 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
275 RETURN(ldlm_completion_tail(lock));
279 * A helper to build a blocking ast function
281 * Perform a common operation for blocking asts:
282 * defferred lock cancellation.
284 * \param lock the lock blocking or canceling ast was called on
286 * \see mdt_blocking_ast
287 * \see ldlm_blocking_ast
289 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
294 lock->l_flags |= LDLM_FL_CBPENDING;
295 do_ast = (!lock->l_readers && !lock->l_writers);
296 unlock_res_and_lock(lock);
299 struct lustre_handle lockh;
302 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
303 ldlm_lock2handle(lock, &lockh);
304 rc = ldlm_cli_cancel(&lockh);
306 CERROR("ldlm_cli_cancel: %d\n", rc);
308 LDLM_DEBUG(lock, "Lock still has references, will be "
315 * Server blocking AST
317 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
320 * \param lock the lock which blocks a request or cancelling lock
323 * \param flag indicates whether this cancelling or blocking callback
325 * \see ldlm_blocking_ast_nocheck
327 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
328 void *data, int flag)
332 if (flag == LDLM_CB_CANCELING) {
333 /* Don't need to do anything here. */
337 lock_res_and_lock(lock);
338 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
339 * that ldlm_blocking_ast is called just before intent_policy method
340 * takes the lr_lock, then by the time we get the lock, we might not
341 * be the correct blocking function anymore. So check, and return
343 if (lock->l_blocking_ast != ldlm_blocking_ast) {
344 unlock_res_and_lock(lock);
347 RETURN(ldlm_blocking_ast_nocheck(lock));
351 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
352 * comment in filter_intent_policy() on why you may need this.
354 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
357 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
358 * that is rather subtle: with OST-side locking, it may so happen that
359 * _all_ extent locks are held by the OST. If client wants to obtain
360 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
361 * on the server), dummy glimpse callback fires and does
362 * nothing. Client still receives correct file size due to the
363 * following fragment in filter_intent_policy():
365 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
366 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
367 * res->lr_namespace->ns_lvbo->lvbo_update) {
368 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
371 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
372 * returns correct file size to the client.
374 return -ELDLM_NO_LOCK_DATA;
377 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
378 const struct ldlm_res_id *res_id,
379 ldlm_type_t type, ldlm_policy_data_t *policy,
380 ldlm_mode_t mode, int *flags,
381 ldlm_blocking_callback blocking,
382 ldlm_completion_callback completion,
383 ldlm_glimpse_callback glimpse,
384 void *data, __u32 lvb_len,
385 const __u64 *client_cookie,
386 struct lustre_handle *lockh)
388 struct ldlm_lock *lock;
390 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
391 .lcs_blocking = blocking,
392 .lcs_glimpse = glimpse,
396 LASSERT(!(*flags & LDLM_FL_REPLAY));
397 if (unlikely(ns_is_client(ns))) {
398 CERROR("Trying to enqueue local lock in a shadow namespace\n");
402 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len);
404 GOTO(out_nolock, err = -ENOMEM);
406 ldlm_lock2handle(lock, lockh);
408 /* NB: we don't have any lock now (lock_res_and_lock)
409 * because it's a new lock */
410 ldlm_lock_addref_internal_nolock(lock, mode);
411 lock->l_flags |= LDLM_FL_LOCAL;
412 if (*flags & LDLM_FL_ATOMIC_CB)
413 lock->l_flags |= LDLM_FL_ATOMIC_CB;
416 lock->l_policy_data = *policy;
417 if (client_cookie != NULL)
418 lock->l_client_cookie = *client_cookie;
419 if (type == LDLM_EXTENT)
420 lock->l_req_extent = policy->l_extent;
422 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
423 if (unlikely(err != ELDLM_OK))
427 *policy = lock->l_policy_data;
429 if (lock->l_completion_ast)
430 lock->l_completion_ast(lock, *flags, NULL);
432 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
435 LDLM_LOCK_RELEASE(lock);
440 static void failed_lock_cleanup(struct ldlm_namespace *ns,
441 struct ldlm_lock *lock, int mode)
445 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
446 lock_res_and_lock(lock);
447 /* Check that lock is not granted or failed, we might race. */
448 if ((lock->l_req_mode != lock->l_granted_mode) &&
449 !(lock->l_flags & LDLM_FL_FAILED)) {
450 /* Make sure that this lock will not be found by raced
451 * bl_ast and -EINVAL reply is sent to server anyways.
453 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
454 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
457 unlock_res_and_lock(lock);
461 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
462 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
464 LDLM_DEBUG(lock, "lock was granted or failed in race");
466 ldlm_lock_decref_internal(lock, mode);
468 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
469 * from llite/file.c/ll_file_flock(). */
470 /* This code makes for the fact that we do not have blocking handler on
471 * a client for flock locks. As such this is the place where we must
472 * completely kill failed locks. (interrupted and those that
473 * were waiting to be granted when server evicted us. */
474 if (lock->l_resource->lr_type == LDLM_FLOCK) {
475 lock_res_and_lock(lock);
476 ldlm_resource_unlink_lock(lock);
477 ldlm_lock_destroy_nolock(lock);
478 unlock_res_and_lock(lock);
482 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
483 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
484 int *flags, void *lvb, __u32 lvb_len,
485 struct lustre_handle *lockh,int rc)
487 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
488 int is_replay = *flags & LDLM_FL_REPLAY;
489 struct ldlm_lock *lock;
490 struct ldlm_reply *reply;
491 int cleanup_phase = 1;
494 lock = ldlm_handle2lock(lockh);
495 /* ldlm_cli_enqueue is holding a reference on this lock. */
497 LASSERT(type == LDLM_FLOCK);
501 if (rc != ELDLM_OK) {
503 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
504 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
505 if (rc == ELDLM_LOCK_ABORTED) {
506 /* Before we return, swab the reply */
507 reply = req_capsule_server_get(&req->rq_pill,
512 struct ost_lvb *tmplvb;
514 req_capsule_set_size(&req->rq_pill,
515 &RMF_DLM_LVB, RCL_SERVER,
517 tmplvb = req_capsule_server_get(&req->rq_pill,
520 GOTO(cleanup, rc = -EPROTO);
522 memcpy(lvb, tmplvb, lvb_len);
528 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
530 GOTO(cleanup, rc = -EPROTO);
532 /* lock enqueued on the server */
535 lock_res_and_lock(lock);
536 /* Key change rehash lock in per-export hash with new key */
537 if (exp->exp_lock_hash) {
538 cfs_hash_rehash_key(exp->exp_lock_hash,
539 &lock->l_remote_handle,
543 lock->l_remote_handle = reply->lock_handle;
546 *flags = reply->lock_flags;
547 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
548 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
549 * to wait with no timeout as well */
550 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
551 unlock_res_and_lock(lock);
553 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
554 lock, reply->lock_handle.cookie, *flags);
556 /* If enqueue returned a blocked lock but the completion handler has
557 * already run, then it fixed up the resource and we don't need to do it
559 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
560 int newmode = reply->lock_desc.l_req_mode;
562 if (newmode && newmode != lock->l_req_mode) {
563 LDLM_DEBUG(lock, "server returned different mode %s",
564 ldlm_lockname[newmode]);
565 lock->l_req_mode = newmode;
568 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
569 lock->l_resource->lr_name.name,
570 sizeof(struct ldlm_res_id))) {
571 CDEBUG(D_INFO, "remote intent success, locking "
572 "(%ld,%ld,%ld) instead of "
574 (long)reply->lock_desc.l_resource.lr_name.name[0],
575 (long)reply->lock_desc.l_resource.lr_name.name[1],
576 (long)reply->lock_desc.l_resource.lr_name.name[2],
577 (long)lock->l_resource->lr_name.name[0],
578 (long)lock->l_resource->lr_name.name[1],
579 (long)lock->l_resource->lr_name.name[2]);
581 rc = ldlm_lock_change_resource(ns, lock,
582 &reply->lock_desc.l_resource.lr_name);
583 if (rc || lock->l_resource == NULL)
584 GOTO(cleanup, rc = -ENOMEM);
585 LDLM_DEBUG(lock, "client-side enqueue, new resource");
588 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
590 lock->l_policy_data =
591 reply->lock_desc.l_policy_data;
592 if (type != LDLM_PLAIN)
593 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
596 if ((*flags) & LDLM_FL_AST_SENT ||
597 /* Cancel extent locks as soon as possible on a liblustre client,
598 * because it cannot handle asynchronous ASTs robustly (see
600 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
601 lock_res_and_lock(lock);
602 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
603 unlock_res_and_lock(lock);
604 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
607 /* If the lock has already been granted by a completion AST, don't
608 * clobber the LVB with an older one. */
609 if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
612 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
614 tmplvb = req_capsule_server_get(&req->rq_pill,
617 GOTO(cleanup, rc = -EPROTO);
618 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
622 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
623 if (lock->l_completion_ast != NULL) {
624 int err = lock->l_completion_ast(lock, *flags, NULL);
632 if (lvb_len && lvb != NULL) {
633 /* Copy the LVB here, and not earlier, because the completion
634 * AST (if any) can override what we got in the reply */
635 memcpy(lvb, lock->l_lvb_data, lvb_len);
638 LDLM_DEBUG(lock, "client-side enqueue END");
641 if (cleanup_phase == 1 && rc)
642 failed_lock_cleanup(ns, lock, mode);
643 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
645 LDLM_LOCK_RELEASE(lock);
649 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
650 * a single page on the send/receive side. XXX: 512 should be changed
651 * to more adequate value. */
652 static inline int ldlm_req_handles_avail(int req_size, int off)
656 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
657 if (likely(avail >= 0))
658 avail /= (int)sizeof(struct lustre_handle);
661 avail += LDLM_LOCKREQ_HANDLES - off;
666 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
667 enum req_location loc,
670 int size = req_capsule_msg_size(pill, loc);
671 return ldlm_req_handles_avail(size, off);
674 static inline int ldlm_format_handles_avail(struct obd_import *imp,
675 const struct req_format *fmt,
676 enum req_location loc, int off)
678 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
679 return ldlm_req_handles_avail(size, off);
682 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
683 * @count locks in @cancels. */
684 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
685 int version, int opc, int canceloff,
686 cfs_list_t *cancels, int count)
688 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
689 struct req_capsule *pill = &req->rq_pill;
690 struct ldlm_request *dlm = NULL;
691 int flags, avail, to_free, bufcount, pack = 0;
698 if (exp_connect_cancelset(exp)) {
699 /* Estimate the amount of available space in the request. */
700 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
701 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
703 flags = ns_connect_lru_resize(ns) ?
704 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
705 to_free = !ns_connect_lru_resize(ns) &&
706 opc == LDLM_ENQUEUE ? 1 : 0;
708 /* Cancel lru locks here _only_ if the server supports
709 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
710 * rpc, what will make us slower. */
712 count += ldlm_cancel_lru_local(ns, cancels, to_free,
713 avail - count, 0, flags);
718 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
719 ldlm_request_bufsize(pack, opc));
722 rc = ptlrpc_request_pack(req, version, opc);
724 ldlm_lock_list_put(cancels, l_bl_ast, count);
728 if (exp_connect_cancelset(exp)) {
730 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
732 /* Skip first lock handler in ldlm_request_pack(),
733 * this method will incrment @lock_count according
734 * to the lock handle amount actually written to
736 dlm->lock_count = canceloff;
738 /* Pack into the request @pack lock handles. */
739 ldlm_cli_cancel_list(cancels, pack, req, 0);
740 /* Prepare and send separate cancel rpc for others. */
741 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
743 ldlm_lock_list_put(cancels, l_bl_ast, count);
748 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
749 cfs_list_t *cancels, int count)
751 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
752 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
755 /* If a request has some specific initialisation it is passed in @reqp,
756 * otherwise it is created in ldlm_cli_enqueue.
758 * Supports sync and async requests, pass @async flag accordingly. If a
759 * request was created in ldlm_cli_enqueue and it is the async request,
760 * pass it to the caller in @reqp. */
761 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
762 struct ldlm_enqueue_info *einfo,
763 const struct ldlm_res_id *res_id,
764 ldlm_policy_data_t const *policy, int *flags,
765 void *lvb, __u32 lvb_len, struct lustre_handle *lockh,
768 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
769 struct ldlm_lock *lock;
770 struct ldlm_request *body;
771 int is_replay = *flags & LDLM_FL_REPLAY;
772 int req_passed_in = 1;
774 struct ptlrpc_request *req;
777 LASSERT(exp != NULL);
779 /* If we're replaying this lock, just check some invariants.
780 * If we're creating a new lock, get everything all setup nice. */
782 lock = ldlm_handle2lock_long(lockh, 0);
783 LASSERT(lock != NULL);
784 LDLM_DEBUG(lock, "client-side enqueue START");
785 LASSERT(exp == lock->l_conn_export);
787 const struct ldlm_callback_suite cbs = {
788 .lcs_completion = einfo->ei_cb_cp,
789 .lcs_blocking = einfo->ei_cb_bl,
790 .lcs_glimpse = einfo->ei_cb_gl,
791 .lcs_weigh = einfo->ei_cb_wg
793 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
794 einfo->ei_mode, &cbs, einfo->ei_cbdata,
798 /* for the local lock, add the reference */
799 ldlm_lock_addref_internal(lock, einfo->ei_mode);
800 ldlm_lock2handle(lock, lockh);
801 if (policy != NULL) {
802 /* INODEBITS_INTEROP: If the server does not support
803 * inodebits, we will request a plain lock in the
804 * descriptor (ldlm_lock2desc() below) but use an
805 * inodebits lock internally with both bits set.
807 if (einfo->ei_type == LDLM_IBITS &&
808 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
809 lock->l_policy_data.l_inodebits.bits =
810 MDS_INODELOCK_LOOKUP |
811 MDS_INODELOCK_UPDATE;
813 lock->l_policy_data = *policy;
816 if (einfo->ei_type == LDLM_EXTENT)
817 lock->l_req_extent = policy->l_extent;
818 LDLM_DEBUG(lock, "client-side enqueue START");
821 /* lock not sent to server yet */
823 if (reqp == NULL || *reqp == NULL) {
824 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
829 failed_lock_cleanup(ns, lock, einfo->ei_mode);
830 LDLM_LOCK_RELEASE(lock);
840 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
842 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
843 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
846 lock->l_conn_export = exp;
847 lock->l_export = NULL;
848 lock->l_blocking_ast = einfo->ei_cb_bl;
850 /* Dump lock data into the request buffer */
851 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
852 ldlm_lock2desc(lock, &body->lock_desc);
853 body->lock_flags = *flags;
854 body->lock_handle[0] = *lockh;
856 /* Continue as normal. */
857 if (!req_passed_in) {
859 req_capsule_extend(&req->rq_pill,
860 &RQF_LDLM_ENQUEUE_LVB);
861 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
862 RCL_SERVER, lvb_len);
864 ptlrpc_request_set_replen(req);
868 * Liblustre client doesn't get extent locks, except for O_APPEND case
869 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
870 * [i_size, OBD_OBJECT_EOF] lock is taken.
872 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
873 policy->l_extent.end == OBD_OBJECT_EOF));
876 LASSERT(reqp != NULL);
880 LDLM_DEBUG(lock, "sending request");
882 rc = ptlrpc_queue_wait(req);
884 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
885 einfo->ei_mode, flags, lvb, lvb_len,
888 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
889 * one reference that we took */
891 LDLM_LOCK_RELEASE(lock);
895 if (!req_passed_in && req != NULL) {
896 ptlrpc_req_finished(req);
904 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
907 struct ldlm_resource *res;
910 if (ns_is_client(ldlm_lock_to_ns(lock))) {
911 CERROR("Trying to cancel local lock\n");
914 LDLM_DEBUG(lock, "client-side local convert");
916 res = ldlm_lock_convert(lock, new_mode, flags);
918 ldlm_reprocess_all(res);
923 LDLM_DEBUG(lock, "client-side local convert handler END");
928 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
929 * conversion of locks which are on the waiting or converting queue */
930 /* Caller of this code is supposed to take care of lock readers/writers
932 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
934 struct ldlm_request *body;
935 struct ldlm_reply *reply;
936 struct ldlm_lock *lock;
937 struct ldlm_resource *res;
938 struct ptlrpc_request *req;
942 lock = ldlm_handle2lock(lockh);
949 if (lock->l_conn_export == NULL)
950 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
952 LDLM_DEBUG(lock, "client-side convert");
954 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
955 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
962 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
963 body->lock_handle[0] = lock->l_remote_handle;
965 body->lock_desc.l_req_mode = new_mode;
966 body->lock_flags = *flags;
969 ptlrpc_request_set_replen(req);
970 rc = ptlrpc_queue_wait(req);
974 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
976 GOTO(out, rc = -EPROTO);
979 GOTO(out, rc = req->rq_status);
981 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
983 ldlm_reprocess_all(res);
984 /* Go to sleep until the lock is granted. */
985 /* FIXME: or cancelled. */
986 if (lock->l_completion_ast) {
987 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
998 ptlrpc_req_finished(req);
1002 /* Cancel locks locally.
1004 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
1005 * LDLM_FL_CANCELING otherwise;
1006 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
1007 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
1009 int rc = LDLM_FL_LOCAL_ONLY;
1012 if (lock->l_conn_export) {
1015 LDLM_DEBUG(lock, "client-side cancel");
1016 /* Set this flag to prevent others from getting new references*/
1017 lock_res_and_lock(lock);
1018 lock->l_flags |= LDLM_FL_CBPENDING;
1019 local_only = (lock->l_flags &
1020 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1021 ldlm_cancel_callback(lock);
1022 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1023 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1024 unlock_res_and_lock(lock);
1027 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1029 rc = LDLM_FL_LOCAL_ONLY;
1031 ldlm_lock_cancel(lock);
1033 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1034 LDLM_ERROR(lock, "Trying to cancel local lock");
1037 LDLM_DEBUG(lock, "server-side local cancel");
1038 ldlm_lock_cancel(lock);
1039 ldlm_reprocess_all(lock->l_resource);
1045 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
1046 of the request @req. */
1047 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1048 cfs_list_t *head, int count)
1050 struct ldlm_request *dlm;
1051 struct ldlm_lock *lock;
1052 int max, packed = 0;
1055 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1056 LASSERT(dlm != NULL);
1058 /* Check the room in the request buffer. */
1059 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1060 sizeof(struct ldlm_request);
1061 max /= sizeof(struct lustre_handle);
1062 max += LDLM_LOCKREQ_HANDLES;
1063 LASSERT(max >= dlm->lock_count + count);
1065 /* XXX: it would be better to pack lock handles grouped by resource.
1066 * so that the server cancel would call filter_lvbo_update() less
1068 cfs_list_for_each_entry(lock, head, l_bl_ast) {
1071 LASSERT(lock->l_conn_export);
1072 /* Pack the lock handle to the given request buffer. */
1073 LDLM_DEBUG(lock, "packing");
1074 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1077 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1081 /* Prepare and send a batched cancel rpc, it will include count lock handles
1082 * of locks given in @head. */
1083 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1084 int count, ldlm_cancel_flags_t flags)
1086 struct ptlrpc_request *req = NULL;
1087 struct obd_import *imp;
1092 LASSERT(exp != NULL);
1095 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
1097 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1100 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1101 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1108 imp = class_exp2cliimp(exp);
1109 if (imp == NULL || imp->imp_invalid) {
1111 "skipping cancel on invalid import %p\n", imp);
1115 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1117 GOTO(out, rc = -ENOMEM);
1119 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1120 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1121 ldlm_request_bufsize(count, LDLM_CANCEL));
1123 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1125 ptlrpc_request_free(req);
1128 req->rq_no_resend = 1;
1129 req->rq_no_delay = 1;
1131 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1132 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1133 ptlrpc_at_set_req_timeout(req);
1135 ldlm_cancel_pack(req, cancels, count);
1137 ptlrpc_request_set_replen(req);
1138 if (flags & LCF_ASYNC) {
1139 ptlrpcd_add_req(req, PSCOPE_OTHER);
1143 rc = ptlrpc_queue_wait(req);
1146 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1147 "out of sync -- not fatal\n",
1148 libcfs_nid2str(req->rq_import->
1149 imp_connection->c_peer.nid));
1151 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1152 req->rq_import_generation == imp->imp_generation) {
1153 ptlrpc_req_finished(req);
1155 } else if (rc != ELDLM_OK) {
1156 CERROR("Got rc %d from cancel RPC: canceling "
1164 ptlrpc_req_finished(req);
1167 return sent ? sent : rc;
1170 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1172 LASSERT(imp != NULL);
1173 return &imp->imp_obd->obd_namespace->ns_pool;
1177 * Update client's obd pool related fields with new SLV and Limit from \a req.
1179 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1181 struct obd_device *obd;
1182 __u64 old_slv, new_slv;
1185 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1186 !imp_connect_lru_resize(req->rq_import)))
1189 * Do nothing for corner cases.
1195 * In some cases RPC may contain slv and limit zeroed out. This is
1196 * the case when server does not support lru resize feature. This is
1197 * also possible in some recovery cases when server side reqs have no
1198 * ref to obd export and thus access to server side namespace is no
1201 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1202 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1203 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1204 "(SLV: "LPU64", Limit: %u)",
1205 lustre_msg_get_slv(req->rq_repmsg),
1206 lustre_msg_get_limit(req->rq_repmsg));
1210 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1211 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1212 obd = req->rq_import->imp_obd;
1215 * Set new SLV and Limit to obd fields to make accessible for pool
1216 * thread. We do not access obd_namespace and pool directly here
1217 * as there is no reliable way to make sure that they are still
1218 * alive in cleanup time. Evil races are possible which may cause
1219 * oops in that time.
1221 cfs_write_lock(&obd->obd_pool_lock);
1222 old_slv = obd->obd_pool_slv;
1223 obd->obd_pool_slv = new_slv;
1224 obd->obd_pool_limit = new_limit;
1225 cfs_write_unlock(&obd->obd_pool_lock);
1229 EXPORT_SYMBOL(ldlm_cli_update_pool);
1231 int ldlm_cli_cancel(struct lustre_handle *lockh)
1233 struct obd_export *exp;
1234 int avail, flags, count = 1, rc = 0;
1235 struct ldlm_namespace *ns;
1236 struct ldlm_lock *lock;
1237 CFS_LIST_HEAD(cancels);
1240 /* concurrent cancels on the same handle can happen */
1241 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1243 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1247 rc = ldlm_cli_cancel_local(lock);
1248 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1249 LDLM_LOCK_RELEASE(lock);
1250 RETURN(rc < 0 ? rc : 0);
1252 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1253 * rpc which goes to canceld portal, so we can cancel other lru locks
1254 * here and send them all as one LDLM_CANCEL rpc. */
1255 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1256 cfs_list_add(&lock->l_bl_ast, &cancels);
1258 exp = lock->l_conn_export;
1259 if (exp_connect_cancelset(exp)) {
1260 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1265 ns = ldlm_lock_to_ns(lock);
1266 flags = ns_connect_lru_resize(ns) ?
1267 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1268 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1271 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1275 /* XXX until we will have compound requests and can cut cancels from generic rpc
1276 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1277 int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
1278 ldlm_cancel_flags_t flags)
1280 CFS_LIST_HEAD(head);
1281 struct ldlm_lock *lock, *next;
1282 int left = 0, bl_ast = 0, rc;
1285 cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1289 if (flags & LCF_LOCAL) {
1290 rc = LDLM_FL_LOCAL_ONLY;
1291 ldlm_lock_cancel(lock);
1293 rc = ldlm_cli_cancel_local(lock);
1295 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1296 LDLM_DEBUG(lock, "Cancel lock separately");
1297 cfs_list_del_init(&lock->l_bl_ast);
1298 cfs_list_add(&lock->l_bl_ast, &head);
1302 if (rc == LDLM_FL_LOCAL_ONLY) {
1303 /* CANCEL RPC should not be sent to server. */
1304 cfs_list_del_init(&lock->l_bl_ast);
1305 LDLM_LOCK_RELEASE(lock);
1312 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1319 * Cancel as many locks as possible w/o sending any rpcs (e.g. to write back
1320 * dirty data, to close a file, ...) or waiting for any rpcs in-flight (e.g.
1321 * readahead requests, ...)
1323 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1324 struct ldlm_lock *lock,
1325 int unused, int added,
1328 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1329 ldlm_cancel_for_recovery cb = ns->ns_cancel_for_recovery;
1330 lock_res_and_lock(lock);
1332 /* don't check added & count since we want to process all locks
1333 * from unused list */
1334 switch (lock->l_resource->lr_type) {
1340 result = LDLM_POLICY_SKIP_LOCK;
1341 lock->l_flags |= LDLM_FL_SKIPPED;
1345 unlock_res_and_lock(lock);
1350 * Callback function for lru-resize policy. Makes decision whether to keep
1351 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1352 * \a added and number of locks to be preferably canceled \a count.
1354 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1356 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1358 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1359 struct ldlm_lock *lock,
1360 int unused, int added,
1363 cfs_time_t cur = cfs_time_current();
1364 struct ldlm_pool *pl = &ns->ns_pool;
1369 * Stop lru processing when we reached passed @count or checked all
1372 if (count && added >= count)
1373 return LDLM_POLICY_KEEP_LOCK;
1375 slv = ldlm_pool_get_slv(pl);
1376 lvf = ldlm_pool_get_lvf(pl);
1377 la = cfs_duration_sec(cfs_time_sub(cur,
1378 lock->l_last_used));
1381 * Stop when slv is not yet come from server or lv is smaller than
1384 lv = lvf * la * unused;
1387 * Inform pool about current CLV to see it via proc.
1389 ldlm_pool_set_clv(pl, lv);
1390 return (slv == 0 || lv < slv) ?
1391 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1395 * Callback function for proc used policy. Makes decision whether to keep
1396 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1397 * \a added and number of locks to be preferably canceled \a count.
1399 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1401 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1403 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1404 struct ldlm_lock *lock,
1405 int unused, int added,
1409 * Stop lru processing when we reached passed @count or checked all
1412 return (added >= count) ?
1413 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1417 * Callback function for aged policy. Makes decision whether to keep
1418 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1419 * \a added and number of locks to be preferably canceled \a count.
1421 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1423 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1425 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1426 struct ldlm_lock *lock,
1427 int unused, int added,
1431 * Stop lru processing if young lock is found and we reached passed
1434 return ((added >= count) &&
1435 cfs_time_before(cfs_time_current(),
1436 cfs_time_add(lock->l_last_used,
1438 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1442 * Callback function for default policy. Makes decision whether to keep
1443 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1444 * \a added and number of locks to be preferably canceled \a count.
1446 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1448 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1450 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1451 struct ldlm_lock *lock,
1452 int unused, int added,
1456 * Stop lru processing when we reached passed @count or checked all
1459 return (added >= count) ?
1460 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1463 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1464 struct ldlm_lock *, int,
1467 static ldlm_cancel_lru_policy_t
1468 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1470 if (flags & LDLM_CANCEL_NO_WAIT)
1471 return ldlm_cancel_no_wait_policy;
1473 if (ns_connect_lru_resize(ns)) {
1474 if (flags & LDLM_CANCEL_SHRINK)
1475 /* We kill passed number of old locks. */
1476 return ldlm_cancel_passed_policy;
1477 else if (flags & LDLM_CANCEL_LRUR)
1478 return ldlm_cancel_lrur_policy;
1479 else if (flags & LDLM_CANCEL_PASSED)
1480 return ldlm_cancel_passed_policy;
1482 if (flags & LDLM_CANCEL_AGED)
1483 return ldlm_cancel_aged_policy;
1486 return ldlm_cancel_default_policy;
1489 /* - Free space in lru for @count new locks,
1490 * redundant unused locks are canceled locally;
1491 * - also cancel locally unused aged locks;
1492 * - do not cancel more than @max locks;
1493 * - GET the found locks and add them into the @cancels list.
1495 * A client lock can be added to the l_bl_ast list only when it is
1496 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1497 * There are the following use cases: ldlm_cancel_resource_local(),
1498 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1499 * flag properly. As any attempt to cancel a lock rely on this flag,
1500 * l_bl_ast list is accessed later without any special locking.
1502 * Calling policies for enabled lru resize:
1503 * ----------------------------------------
1504 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1505 * cancel not more than @count locks;
1507 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1508 * the beginning of lru list);
1510 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1511 * memory pressre policy function;
1513 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1515 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1516 * (typically before replaying locks) w/o
1517 * sending any rpcs or waiting for any
1518 * outstanding rpc to complete.
1520 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
1521 int count, int max, int flags)
1523 ldlm_cancel_lru_policy_t pf;
1524 struct ldlm_lock *lock, *next;
1525 int added = 0, unused, remained;
1528 cfs_spin_lock(&ns->ns_lock);
1529 unused = ns->ns_nr_unused;
1532 if (!ns_connect_lru_resize(ns))
1533 count += unused - ns->ns_max_unused;
1535 pf = ldlm_cancel_lru_policy(ns, flags);
1536 LASSERT(pf != NULL);
1538 while (!cfs_list_empty(&ns->ns_unused_list)) {
1539 ldlm_policy_res_t result;
1541 /* all unused locks */
1542 if (remained-- <= 0)
1545 /* For any flags, stop scanning if @max is reached. */
1546 if (max && added >= max)
1549 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1551 /* No locks which got blocking requests. */
1552 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1554 if (flags & LDLM_CANCEL_NO_WAIT &&
1555 lock->l_flags & LDLM_FL_SKIPPED)
1556 /* already processed */
1559 /* Somebody is already doing CANCEL. No need in this
1560 * lock in lru, do not traverse it again. */
1561 if (!(lock->l_flags & LDLM_FL_CANCELING))
1564 ldlm_lock_remove_from_lru_nolock(lock);
1566 if (&lock->l_lru == &ns->ns_unused_list)
1569 LDLM_LOCK_GET(lock);
1570 cfs_spin_unlock(&ns->ns_lock);
1571 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1573 /* Pass the lock through the policy filter and see if it
1574 * should stay in lru.
1576 * Even for shrinker policy we stop scanning if
1577 * we find a lock that should stay in the cache.
1578 * We should take into account lock age anyway
1579 * as new lock even if it is small of weight is
1580 * valuable resource.
1582 * That is, for shrinker policy we drop only
1583 * old locks, but additionally chose them by
1584 * their weight. Big extent locks will stay in
1586 result = pf(ns, lock, unused, added, count);
1587 if (result == LDLM_POLICY_KEEP_LOCK) {
1588 lu_ref_del(&lock->l_reference,
1589 __FUNCTION__, cfs_current());
1590 LDLM_LOCK_RELEASE(lock);
1591 cfs_spin_lock(&ns->ns_lock);
1594 if (result == LDLM_POLICY_SKIP_LOCK) {
1595 lu_ref_del(&lock->l_reference,
1596 __FUNCTION__, cfs_current());
1597 LDLM_LOCK_RELEASE(lock);
1598 cfs_spin_lock(&ns->ns_lock);
1602 lock_res_and_lock(lock);
1603 /* Check flags again under the lock. */
1604 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1605 (ldlm_lock_remove_from_lru(lock) == 0)) {
1606 /* other thread is removing lock from lru or
1607 * somebody is already doing CANCEL or
1608 * there is a blocking request which will send
1609 * cancel by itseft or the lock is matched
1610 * is already not unused. */
1611 unlock_res_and_lock(lock);
1612 lu_ref_del(&lock->l_reference,
1613 __FUNCTION__, cfs_current());
1614 LDLM_LOCK_RELEASE(lock);
1615 cfs_spin_lock(&ns->ns_lock);
1618 LASSERT(!lock->l_readers && !lock->l_writers);
1620 /* If we have chosen to cancel this lock voluntarily, we
1621 * better send cancel notification to server, so that it
1622 * frees appropriate state. This might lead to a race
1623 * where while we are doing cancel here, server is also
1624 * silently cancelling this lock. */
1625 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1627 /* Setting the CBPENDING flag is a little misleading,
1628 * but prevents an important race; namely, once
1629 * CBPENDING is set, the lock can accumulate no more
1630 * readers/writers. Since readers and writers are
1631 * already zero here, ldlm_lock_decref() won't see
1632 * this flag and call l_blocking_ast */
1633 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1635 /* We can't re-add to l_lru as it confuses the
1636 * refcounting in ldlm_lock_remove_from_lru() if an AST
1637 * arrives after we drop lr_lock below. We use l_bl_ast
1638 * and can't use l_pending_chain as it is used both on
1639 * server and client nevertheless bug 5666 says it is
1640 * used only on server */
1641 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1642 cfs_list_add(&lock->l_bl_ast, cancels);
1643 unlock_res_and_lock(lock);
1644 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1645 cfs_spin_lock(&ns->ns_lock);
1649 cfs_spin_unlock(&ns->ns_lock);
1653 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1654 int count, int max, ldlm_cancel_flags_t cancel_flags,
1658 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1661 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1664 /* when called with LDLM_ASYNC the blocking callback will be handled
1665 * in a thread and this function will return after the thread has been
1666 * asked to call the callback. when called with LDLM_SYNC the blocking
1667 * callback will be performed in this function. */
1668 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t mode,
1671 CFS_LIST_HEAD(cancels);
1676 mode = LDLM_SYNC; /* force to be sync in user space */
1678 /* Just prepare the list of locks, do not actually cancel them yet.
1679 * Locks are cancelled later in a separate thread. */
1680 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1681 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, mode);
1688 /* Find and cancel locally unused locks found on resource, matched to the
1689 * given policy, mode. GET the found locks and add them into the @cancels
1691 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1692 cfs_list_t *cancels,
1693 ldlm_policy_data_t *policy,
1694 ldlm_mode_t mode, int lock_flags,
1695 ldlm_cancel_flags_t cancel_flags, void *opaque)
1697 struct ldlm_lock *lock;
1702 cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1703 if (opaque != NULL && lock->l_ast_data != opaque) {
1704 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1705 lock->l_ast_data, opaque);
1710 if (lock->l_readers || lock->l_writers)
1713 /* If somebody is already doing CANCEL, or blocking ast came,
1714 * skip this lock. */
1715 if (lock->l_flags & LDLM_FL_BL_AST ||
1716 lock->l_flags & LDLM_FL_CANCELING)
1719 if (lockmode_compat(lock->l_granted_mode, mode))
1722 /* If policy is given and this is IBITS lock, add to list only
1723 * those locks that match by policy. */
1724 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1725 !(lock->l_policy_data.l_inodebits.bits &
1726 policy->l_inodebits.bits))
1729 /* See CBPENDING comment in ldlm_cancel_lru */
1730 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1733 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1734 cfs_list_add(&lock->l_bl_ast, cancels);
1735 LDLM_LOCK_GET(lock);
1740 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1743 /* If @req is NULL, send CANCEL request to server with handles of locks
1744 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1745 * separately per lock.
1746 * If @req is not NULL, put handles of locks in @cancels into the request
1747 * buffer at the offset @off.
1748 * Destroy @cancels at the end. */
1749 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1750 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1752 struct ldlm_lock *lock;
1756 if (cfs_list_empty(cancels) || count == 0)
1759 /* XXX: requests (both batched and not) could be sent in parallel.
1760 * Usually it is enough to have just 1 RPC, but it is possible that
1761 * there are to many locks to be cancelled in LRU or on a resource.
1762 * It would also speed up the case when the server does not support
1765 LASSERT(!cfs_list_empty(cancels));
1766 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1768 LASSERT(lock->l_conn_export);
1770 if (exp_connect_cancelset(lock->l_conn_export)) {
1773 ldlm_cancel_pack(req, cancels, count);
1775 res = ldlm_cli_cancel_req(lock->l_conn_export,
1779 res = ldlm_cli_cancel_req(lock->l_conn_export,
1784 CERROR("ldlm_cli_cancel_list: %d\n", res);
1789 ldlm_lock_list_put(cancels, l_bl_ast, res);
1791 LASSERT(count == 0);
1795 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1796 const struct ldlm_res_id *res_id,
1797 ldlm_policy_data_t *policy,
1799 ldlm_cancel_flags_t flags,
1802 struct ldlm_resource *res;
1803 CFS_LIST_HEAD(cancels);
1808 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1810 /* This is not a problem. */
1811 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1815 LDLM_RESOURCE_ADDREF(res);
1816 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1817 0, flags | LCF_BL_AST, opaque);
1818 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1820 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1822 LDLM_RESOURCE_DELREF(res);
1823 ldlm_resource_putref(res);
1827 struct ldlm_cli_cancel_arg {
1832 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1833 cfs_hlist_node_t *hnode, void *arg)
1835 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1836 struct ldlm_cli_cancel_arg *lc = arg;
1839 rc = ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1841 lc->lc_flags, lc->lc_opaque);
1843 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1844 res->lr_name.name[0], rc);
1846 /* must return 0 for hash iteration */
1850 /* Cancel all locks on a namespace (or a specific resource, if given)
1851 * that have 0 readers/writers.
1853 * If flags & LCF_LOCAL, throw the locks away without trying
1854 * to notify the server. */
1855 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1856 const struct ldlm_res_id *res_id,
1857 ldlm_cancel_flags_t flags, void *opaque)
1859 struct ldlm_cli_cancel_arg arg = {
1861 .lc_opaque = opaque,
1869 if (res_id != NULL) {
1870 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1874 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1875 ldlm_cli_hash_cancel_unused, &arg);
1880 /* Lock iterators. */
1882 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1885 cfs_list_t *tmp, *next;
1886 struct ldlm_lock *lock;
1887 int rc = LDLM_ITER_CONTINUE;
1892 RETURN(LDLM_ITER_CONTINUE);
1895 cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
1896 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1898 if (iter(lock, closure) == LDLM_ITER_STOP)
1899 GOTO(out, rc = LDLM_ITER_STOP);
1902 cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
1903 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1905 if (iter(lock, closure) == LDLM_ITER_STOP)
1906 GOTO(out, rc = LDLM_ITER_STOP);
1909 cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
1910 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1912 if (iter(lock, closure) == LDLM_ITER_STOP)
1913 GOTO(out, rc = LDLM_ITER_STOP);
1920 struct iter_helper_data {
1921 ldlm_iterator_t iter;
1925 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1927 struct iter_helper_data *helper = closure;
1928 return helper->iter(lock, helper->closure);
1931 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1932 cfs_hlist_node_t *hnode, void *arg)
1935 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1937 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
1941 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
1942 ldlm_iterator_t iter, void *closure)
1945 struct iter_helper_data helper = { iter: iter, closure: closure };
1947 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1948 ldlm_res_iter_helper, &helper);
1952 /* non-blocking function to manipulate a lock whose cb_data is being put away.
1953 * return 0: find no resource
1954 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
1957 int ldlm_resource_iterate(struct ldlm_namespace *ns,
1958 const struct ldlm_res_id *res_id,
1959 ldlm_iterator_t iter, void *data)
1961 struct ldlm_resource *res;
1966 CERROR("must pass in namespace\n");
1970 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1974 LDLM_RESOURCE_ADDREF(res);
1975 rc = ldlm_resource_foreach(res, iter, data);
1976 LDLM_RESOURCE_DELREF(res);
1977 ldlm_resource_putref(res);
1983 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1985 cfs_list_t *list = closure;
1987 /* we use l_pending_chain here, because it's unused on clients. */
1988 LASSERTF(cfs_list_empty(&lock->l_pending_chain),
1989 "lock %p next %p prev %p\n",
1990 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1991 /* bug 9573: don't replay locks left after eviction, or
1992 * bug 17614: locks being actively cancelled. Get a reference
1993 * on a lock so that it does not disapear under us (e.g. due to cancel)
1995 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
1996 cfs_list_add(&lock->l_pending_chain, list);
1997 LDLM_LOCK_GET(lock);
2000 return LDLM_ITER_CONTINUE;
2003 static int replay_lock_interpret(const struct lu_env *env,
2004 struct ptlrpc_request *req,
2005 struct ldlm_async_args *aa, int rc)
2007 struct ldlm_lock *lock;
2008 struct ldlm_reply *reply;
2009 struct obd_export *exp;
2012 cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
2017 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2019 GOTO(out, rc = -EPROTO);
2021 lock = ldlm_handle2lock(&aa->lock_handle);
2023 CERROR("received replay ack for unknown local cookie "LPX64
2024 " remote cookie "LPX64 " from server %s id %s\n",
2025 aa->lock_handle.cookie, reply->lock_handle.cookie,
2026 req->rq_export->exp_client_uuid.uuid,
2027 libcfs_id2str(req->rq_peer));
2028 GOTO(out, rc = -ESTALE);
2031 /* Key change rehash lock in per-export hash with new key */
2032 exp = req->rq_export;
2033 if (exp && exp->exp_lock_hash) {
2034 cfs_hash_rehash_key(exp->exp_lock_hash,
2035 &lock->l_remote_handle,
2036 &reply->lock_handle,
2039 lock->l_remote_handle = reply->lock_handle;
2042 LDLM_DEBUG(lock, "replayed lock:");
2043 ptlrpc_import_recovery_state_machine(req->rq_import);
2044 LDLM_LOCK_PUT(lock);
2047 ptlrpc_connect_import(req->rq_import, NULL);
2052 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2054 struct ptlrpc_request *req;
2055 struct ldlm_async_args *aa;
2056 struct ldlm_request *body;
2061 /* Bug 11974: Do not replay a lock which is actively being canceled */
2062 if (lock->l_flags & LDLM_FL_CANCELING) {
2063 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2067 /* If this is reply-less callback lock, we cannot replay it, since
2068 * server might have long dropped it, but notification of that event was
2069 * lost by network. (and server granted conflicting lock already) */
2070 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2071 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2072 ldlm_lock_cancel(lock);
2076 * If granted mode matches the requested mode, this lock is granted.
2078 * If they differ, but we have a granted mode, then we were granted
2079 * one mode and now want another: ergo, converting.
2081 * If we haven't been granted anything and are on a resource list,
2082 * then we're blocked/waiting.
2084 * If we haven't been granted anything and we're NOT on a resource list,
2085 * then we haven't got a reply yet and don't have a known disposition.
2086 * This happens whenever a lock enqueue is the request that triggers
2089 if (lock->l_granted_mode == lock->l_req_mode)
2090 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2091 else if (lock->l_granted_mode)
2092 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2093 else if (!cfs_list_empty(&lock->l_res_link))
2094 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2096 flags = LDLM_FL_REPLAY;
2098 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2099 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2103 /* We're part of recovery, so don't wait for it. */
2104 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2106 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2107 ldlm_lock2desc(lock, &body->lock_desc);
2108 body->lock_flags = flags;
2110 ldlm_lock2handle(lock, &body->lock_handle[0]);
2111 if (lock->l_lvb_len != 0) {
2112 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2113 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2116 ptlrpc_request_set_replen(req);
2117 /* notify the server we've replayed all requests.
2118 * also, we mark the request to be put on a dedicated
2119 * queue to be processed after all request replayes.
2121 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2123 LDLM_DEBUG(lock, "replaying lock:");
2125 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2126 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2127 aa = ptlrpc_req_async_args(req);
2128 aa->lock_handle = body->lock_handle[0];
2129 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2130 ptlrpcd_add_req(req, PSCOPE_OTHER);
2136 * Cancel as many unused locks as possible before replay. since we are
2137 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2140 * Called only in recovery before replaying locks. there is no need to
2141 * replay locks that are unused. since the clients may hold thousands of
2142 * cached unused locks, dropping the unused locks can greatly reduce the
2143 * load on the servers at recovery time.
2145 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2148 CFS_LIST_HEAD(cancels);
2150 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2151 "replay for namespace %s (%d)\n",
2152 ldlm_ns_name(ns), ns->ns_nr_unused);
2154 /* We don't need to care whether or not LRU resize is enabled
2155 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2156 * count parameter */
2157 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2158 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2160 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2161 canceled, ldlm_ns_name(ns));
2164 int ldlm_replay_locks(struct obd_import *imp)
2166 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2167 CFS_LIST_HEAD(list);
2168 struct ldlm_lock *lock, *next;
2173 LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
2175 /* don't replay locks if import failed recovery */
2176 if (imp->imp_vbr_failed)
2179 /* ensure this doesn't fall to 0 before all have been queued */
2180 cfs_atomic_inc(&imp->imp_replay_inflight);
2182 if (ldlm_cancel_unused_locks_before_replay)
2183 ldlm_cancel_unused_locks_for_replay(ns);
2185 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2187 cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2188 cfs_list_del_init(&lock->l_pending_chain);
2190 LDLM_LOCK_RELEASE(lock);
2191 continue; /* or try to do the rest? */
2193 rc = replay_one_lock(imp, lock);
2194 LDLM_LOCK_RELEASE(lock);
2197 cfs_atomic_dec(&imp->imp_replay_inflight);