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 2008 Sun Microsystems, Inc. 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 static void interrupted_completion_wait(void *data)
57 struct lock_wait_data {
58 struct ldlm_lock *lwd_lock;
62 struct ldlm_async_args {
63 struct lustre_handle lock_handle;
66 int ldlm_expired_completion_wait(void *data)
68 struct lock_wait_data *lwd = data;
69 struct ldlm_lock *lock = lwd->lwd_lock;
70 struct obd_import *imp;
71 struct obd_device *obd;
74 if (lock->l_conn_export == NULL) {
75 static cfs_time_t next_dump = 0, last_dump = 0;
77 if (ptlrpc_check_suspend())
80 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
81 CFS_DURATION_T"s ago); not entering recovery in "
82 "server code, just going back to sleep",
83 lock->l_enqueued_time.tv_sec,
84 cfs_time_sub(cfs_time_current_sec(),
85 lock->l_enqueued_time.tv_sec));
86 if (cfs_time_after(cfs_time_current(), next_dump)) {
87 last_dump = next_dump;
88 next_dump = cfs_time_shift(300);
89 ldlm_namespace_dump(D_DLMTRACE,
90 lock->l_resource->lr_namespace);
92 libcfs_debug_dumplog();
97 obd = lock->l_conn_export->exp_obd;
98 imp = obd->u.cli.cl_import;
99 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
100 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
101 CFS_DURATION_T"s ago), entering recovery for %s@%s",
102 lock->l_enqueued_time.tv_sec,
103 cfs_time_sub(cfs_time_current_sec(),
104 lock->l_enqueued_time.tv_sec),
105 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
110 /* We use the same basis for both server side and client side functions
111 from a single node. */
112 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
114 int timeout = at_get(&lock->l_resource->lr_namespace->ns_at_estimate);
116 return obd_timeout / 2;
117 /* Since these are non-updating timeouts, we should be conservative.
118 It would be nice to have some kind of "early reply" mechanism for
119 lock callbacks too... */
120 timeout = timeout + (timeout >> 1); /* 150% */
121 return max(timeout, ldlm_enqueue_min);
124 static int is_granted_or_cancelled(struct ldlm_lock *lock)
128 lock_res_and_lock(lock);
129 if (((lock->l_req_mode == lock->l_granted_mode) &&
130 !(lock->l_flags & LDLM_FL_CP_REQD)) ||
131 (lock->l_flags & LDLM_FL_FAILED))
133 unlock_res_and_lock(lock);
138 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
140 /* XXX ALLOCATE - 160 bytes */
141 struct lock_wait_data lwd;
142 struct obd_device *obd;
143 struct obd_import *imp = NULL;
144 struct l_wait_info lwi;
149 if (flags == LDLM_FL_WAIT_NOREPROC) {
150 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
154 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
155 LDLM_FL_BLOCK_CONV))) {
156 cfs_waitq_signal(&lock->l_waitq);
160 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
162 ldlm_lock_dump(D_OTHER, lock, 0);
163 ldlm_reprocess_all(lock->l_resource);
167 obd = class_exp2obd(lock->l_conn_export);
169 /* if this is a local lock, then there is no import */
171 imp = obd->u.cli.cl_import;
174 /* Wait a long time for enqueue - server may have to callback a
175 lock from another client. Server will evict the other client if it
176 doesn't respond reasonably, and then give us the lock. */
177 timeout = ldlm_get_enq_timeout(lock) * 2;
181 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
182 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
183 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
185 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
186 ldlm_expired_completion_wait,
187 interrupted_completion_wait, &lwd);
191 spin_lock(&imp->imp_lock);
192 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
193 spin_unlock(&imp->imp_lock);
196 /* Go to sleep until the lock is granted or cancelled. */
197 rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
199 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
200 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
205 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
210 LDLM_DEBUG(lock, "client-side enqueue waking up: granted after "
212 cfs_time_sub(cfs_time_current_sec(),
213 lock->l_enqueued_time.tv_sec));
215 /* Update our time estimate */
216 at_add(&lock->l_resource->lr_namespace->ns_at_estimate,
217 cfs_time_current_sec() - lock->l_enqueued_time.tv_sec);
223 * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
225 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
226 void *data, int flag)
231 if (flag == LDLM_CB_CANCELING) {
232 /* Don't need to do anything here. */
236 lock_res_and_lock(lock);
237 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
238 * that ldlm_blocking_ast is called just before intent_policy method
239 * takes the ns_lock, then by the time we get the lock, we might not
240 * be the correct blocking function anymore. So check, and return
242 if (lock->l_blocking_ast != ldlm_blocking_ast) {
243 unlock_res_and_lock(lock);
247 lock->l_flags |= LDLM_FL_CBPENDING;
248 do_ast = (!lock->l_readers && !lock->l_writers);
249 unlock_res_and_lock(lock);
252 struct lustre_handle lockh;
255 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
256 ldlm_lock2handle(lock, &lockh);
257 rc = ldlm_cli_cancel(&lockh);
259 CERROR("ldlm_cli_cancel: %d\n", rc);
261 LDLM_DEBUG(lock, "Lock still has references, will be "
268 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
269 * comment in filter_intent_policy() on why you may need this.
271 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
274 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
275 * that is rather subtle: with OST-side locking, it may so happen that
276 * _all_ extent locks are held by the OST. If client wants to obtain
277 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
278 * on the server), dummy glimpse callback fires and does
279 * nothing. Client still receives correct file size due to the
280 * following fragment in filter_intent_policy():
282 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
283 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
284 * res->lr_namespace->ns_lvbo->lvbo_update) {
285 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
288 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
289 * returns correct file size to the client.
291 return -ELDLM_NO_LOCK_DATA;
294 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
295 const struct ldlm_res_id *res_id,
296 ldlm_type_t type, ldlm_policy_data_t *policy,
297 ldlm_mode_t mode, int *flags,
298 ldlm_blocking_callback blocking,
299 ldlm_completion_callback completion,
300 ldlm_glimpse_callback glimpse,
301 void *data, __u32 lvb_len, void *lvb_swabber,
302 struct lustre_handle *lockh)
304 struct ldlm_lock *lock;
308 LASSERT(!(*flags & LDLM_FL_REPLAY));
309 if (unlikely(ns_is_client(ns))) {
310 CERROR("Trying to enqueue local lock in a shadow namespace\n");
314 lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
315 completion, glimpse, data, lvb_len);
317 GOTO(out_nolock, err = -ENOMEM);
318 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
320 ldlm_lock_addref_internal(lock, mode);
321 ldlm_lock2handle(lock, lockh);
322 lock_res_and_lock(lock);
323 lock->l_flags |= LDLM_FL_LOCAL;
324 if (*flags & LDLM_FL_ATOMIC_CB)
325 lock->l_flags |= LDLM_FL_ATOMIC_CB;
326 lock->l_lvb_swabber = lvb_swabber;
327 unlock_res_and_lock(lock);
329 lock->l_policy_data = *policy;
330 if (type == LDLM_EXTENT)
331 lock->l_req_extent = policy->l_extent;
333 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
334 if (unlikely(err != ELDLM_OK))
338 *policy = lock->l_policy_data;
340 LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
343 if (lock->l_completion_ast)
344 lock->l_completion_ast(lock, *flags, NULL);
346 LDLM_DEBUG(lock, "client-side local enqueue END");
354 static void failed_lock_cleanup(struct ldlm_namespace *ns,
355 struct ldlm_lock *lock,
356 struct lustre_handle *lockh, int mode)
358 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
359 lock_res_and_lock(lock);
360 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
361 unlock_res_and_lock(lock);
362 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
364 ldlm_lock_decref_and_cancel(lockh, mode);
366 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
367 * from llite/file.c/ll_file_flock(). */
368 if (lock->l_resource->lr_type == LDLM_FLOCK) {
369 ldlm_lock_destroy(lock);
373 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
374 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
375 int *flags, void *lvb, __u32 lvb_len,
376 void *lvb_swabber, struct lustre_handle *lockh,int rc)
378 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
379 int is_replay = *flags & LDLM_FL_REPLAY;
380 struct ldlm_lock *lock;
381 struct ldlm_reply *reply;
382 int cleanup_phase = 1;
385 lock = ldlm_handle2lock(lockh);
386 /* ldlm_cli_enqueue is holding a reference on this lock. */
388 LASSERT(type == LDLM_FLOCK);
392 if (rc != ELDLM_OK) {
394 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
395 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
396 if (rc == ELDLM_LOCK_ABORTED) {
397 /* Before we return, swab the reply */
398 reply = req_capsule_server_get(&req->rq_pill,
403 struct ost_lvb *tmplvb;
405 req_capsule_set_size(&req->rq_pill,
406 &RMF_DLM_LVB, RCL_SERVER,
408 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
412 GOTO(cleanup, rc = -EPROTO);
414 memcpy(lvb, tmplvb, lvb_len);
420 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
422 GOTO(cleanup, rc = -EPROTO);
424 /* lock enqueued on the server */
427 lock_res_and_lock(lock);
428 lock->l_remote_handle = reply->lock_handle;
429 *flags = reply->lock_flags;
430 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
431 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
432 * to wait with no timeout as well */
433 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
434 unlock_res_and_lock(lock);
436 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
437 lock, reply->lock_handle.cookie, *flags);
439 /* If enqueue returned a blocked lock but the completion handler has
440 * already run, then it fixed up the resource and we don't need to do it
442 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
443 int newmode = reply->lock_desc.l_req_mode;
445 if (newmode && newmode != lock->l_req_mode) {
446 LDLM_DEBUG(lock, "server returned different mode %s",
447 ldlm_lockname[newmode]);
448 lock->l_req_mode = newmode;
451 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
452 lock->l_resource->lr_name.name,
453 sizeof(struct ldlm_res_id))) {
454 CDEBUG(D_INFO, "remote intent success, locking "
455 "(%ld,%ld,%ld) instead of "
457 (long)reply->lock_desc.l_resource.lr_name.name[0],
458 (long)reply->lock_desc.l_resource.lr_name.name[1],
459 (long)reply->lock_desc.l_resource.lr_name.name[2],
460 (long)lock->l_resource->lr_name.name[0],
461 (long)lock->l_resource->lr_name.name[1],
462 (long)lock->l_resource->lr_name.name[2]);
464 rc = ldlm_lock_change_resource(ns, lock,
465 &reply->lock_desc.l_resource.lr_name);
466 if (rc || lock->l_resource == NULL)
467 GOTO(cleanup, rc = -ENOMEM);
468 LDLM_DEBUG(lock, "client-side enqueue, new resource");
471 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
473 lock->l_policy_data =
474 reply->lock_desc.l_policy_data;
475 if (type != LDLM_PLAIN)
476 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
479 if ((*flags) & LDLM_FL_AST_SENT ||
480 /* Cancel extent locks as soon as possible on a liblustre client,
481 * because it cannot handle asynchronous ASTs robustly (see
483 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
484 lock_res_and_lock(lock);
485 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
486 unlock_res_and_lock(lock);
487 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
490 /* If the lock has already been granted by a completion AST, don't
491 * clobber the LVB with an older one. */
492 if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
495 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
497 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
501 GOTO(cleanup, rc = -EPROTO);
502 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
506 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
507 if (lock->l_completion_ast != NULL) {
508 int err = lock->l_completion_ast(lock, *flags, NULL);
511 if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
516 if (lvb_len && lvb != NULL) {
517 /* Copy the LVB here, and not earlier, because the completion
518 * AST (if any) can override what we got in the reply */
519 memcpy(lvb, lock->l_lvb_data, lvb_len);
522 LDLM_DEBUG(lock, "client-side enqueue END");
525 if (cleanup_phase == 1 && rc)
526 failed_lock_cleanup(ns, lock, lockh, mode);
527 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
533 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
534 * a single page on the send/receive side. XXX: 512 should be changed
535 * to more adequate value. */
536 static inline int ldlm_req_handles_avail(int req_size, int off)
540 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
541 avail /= sizeof(struct lustre_handle);
542 avail += LDLM_LOCKREQ_HANDLES - off;
547 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
548 enum req_location loc,
551 int size = req_capsule_msg_size(pill, loc);
552 return ldlm_req_handles_avail(size, off);
555 static inline int ldlm_format_handles_avail(struct obd_import *imp,
556 const struct req_format *fmt,
557 enum req_location loc, int off)
559 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
560 return ldlm_req_handles_avail(size, off);
563 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
564 * @count locks in @cancels. */
565 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
566 int version, int opc, int canceloff,
567 struct list_head *cancels, int count)
569 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
570 struct req_capsule *pill = &req->rq_pill;
571 struct ldlm_request *dlm = NULL;
572 int flags, avail, to_free, bufcount, pack = 0;
579 if (exp_connect_cancelset(exp)) {
580 /* Estimate the amount of available space in the request. */
581 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
582 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
584 flags = ns_connect_lru_resize(ns) ?
585 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
586 to_free = !ns_connect_lru_resize(ns) &&
587 opc == LDLM_ENQUEUE ? 1 : 0;
589 /* Cancel lru locks here _only_ if the server supports
590 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
591 * rpc, what will make us slower. */
593 count += ldlm_cancel_lru_local(ns, cancels, to_free,
594 avail - count, 0, flags);
599 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
600 ldlm_request_bufsize(pack, opc));
603 rc = ptlrpc_request_pack(req, version, opc);
605 ldlm_lock_list_put(cancels, l_bl_ast, count);
609 if (exp_connect_cancelset(exp)) {
611 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
613 /* Skip first lock handler in ldlm_request_pack(),
614 * this method will incrment @lock_count according
615 * to the lock handle amount actually written to
617 dlm->lock_count = canceloff;
619 /* Pack into the request @pack lock handles. */
620 ldlm_cli_cancel_list(cancels, pack, req, 0);
621 /* Prepare and send separate cancel rpc for others. */
622 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
624 ldlm_lock_list_put(cancels, l_bl_ast, count);
629 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
630 struct list_head *cancels, int count)
632 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
633 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
636 /* If a request has some specific initialisation it is passed in @reqp,
637 * otherwise it is created in ldlm_cli_enqueue.
639 * Supports sync and async requests, pass @async flag accordingly. If a
640 * request was created in ldlm_cli_enqueue and it is the async request,
641 * pass it to the caller in @reqp. */
642 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
643 struct ldlm_enqueue_info *einfo,
644 const struct ldlm_res_id *res_id,
645 ldlm_policy_data_t *policy, int *flags,
646 void *lvb, __u32 lvb_len, void *lvb_swabber,
647 struct lustre_handle *lockh, int async)
649 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
650 struct ldlm_lock *lock;
651 struct ldlm_request *body;
652 int is_replay = *flags & LDLM_FL_REPLAY;
653 int req_passed_in = 1;
655 struct ptlrpc_request *req;
658 LASSERT(exp != NULL);
660 /* If we're replaying this lock, just check some invariants.
661 * If we're creating a new lock, get everything all setup nice. */
663 lock = ldlm_handle2lock(lockh);
664 LASSERT(lock != NULL);
665 LDLM_DEBUG(lock, "client-side enqueue START");
666 LASSERT(exp == lock->l_conn_export);
668 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
669 einfo->ei_mode, einfo->ei_cb_bl,
670 einfo->ei_cb_cp, einfo->ei_cb_gl,
671 einfo->ei_cbdata, lvb_len);
674 /* for the local lock, add the reference */
675 ldlm_lock_addref_internal(lock, einfo->ei_mode);
676 ldlm_lock2handle(lock, lockh);
677 lock->l_lvb_swabber = lvb_swabber;
678 if (policy != NULL) {
679 /* INODEBITS_INTEROP: If the server does not support
680 * inodebits, we will request a plain lock in the
681 * descriptor (ldlm_lock2desc() below) but use an
682 * inodebits lock internally with both bits set.
684 if (einfo->ei_type == LDLM_IBITS &&
685 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
686 lock->l_policy_data.l_inodebits.bits =
687 MDS_INODELOCK_LOOKUP |
688 MDS_INODELOCK_UPDATE;
690 lock->l_policy_data = *policy;
693 if (einfo->ei_type == LDLM_EXTENT)
694 lock->l_req_extent = policy->l_extent;
695 LDLM_DEBUG(lock, "client-side enqueue START");
698 /* lock not sent to server yet */
700 if (reqp == NULL || *reqp == NULL) {
701 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
706 failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
717 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
719 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
720 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
723 lock->l_conn_export = exp;
724 lock->l_export = NULL;
725 lock->l_blocking_ast = einfo->ei_cb_bl;
727 /* Dump lock data into the request buffer */
728 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
729 ldlm_lock2desc(lock, &body->lock_desc);
730 body->lock_flags = *flags;
731 body->lock_handle[0] = *lockh;
733 /* Continue as normal. */
734 if (!req_passed_in) {
736 req_capsule_extend(&req->rq_pill,
737 &RQF_LDLM_ENQUEUE_LVB);
738 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
739 RCL_SERVER, lvb_len);
741 ptlrpc_request_set_replen(req);
745 * Liblustre client doesn't get extent locks, except for O_APPEND case
746 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
747 * [i_size, OBD_OBJECT_EOF] lock is taken.
749 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
750 policy->l_extent.end == OBD_OBJECT_EOF));
753 LASSERT(reqp != NULL);
757 LDLM_DEBUG(lock, "sending request");
758 rc = ptlrpc_queue_wait(req);
759 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
760 einfo->ei_mode, flags, lvb, lvb_len,
761 lvb_swabber, lockh, rc);
763 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
764 * one reference that we took */
770 if (!req_passed_in && req != NULL) {
771 ptlrpc_req_finished(req);
779 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
782 struct ldlm_resource *res;
785 if (ns_is_client(lock->l_resource->lr_namespace)) {
786 CERROR("Trying to cancel local lock\n");
789 LDLM_DEBUG(lock, "client-side local convert");
791 res = ldlm_lock_convert(lock, new_mode, flags);
793 ldlm_reprocess_all(res);
798 LDLM_DEBUG(lock, "client-side local convert handler END");
803 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
804 * conversion of locks which are on the waiting or converting queue */
805 /* Caller of this code is supposed to take care of lock readers/writers
807 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
809 struct ldlm_request *body;
810 struct ldlm_reply *reply;
811 struct ldlm_lock *lock;
812 struct ldlm_resource *res;
813 struct ptlrpc_request *req;
817 lock = ldlm_handle2lock(lockh);
824 if (lock->l_conn_export == NULL)
825 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
827 LDLM_DEBUG(lock, "client-side convert");
829 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
830 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
837 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
838 body->lock_handle[0] = lock->l_remote_handle;
840 body->lock_desc.l_req_mode = new_mode;
841 body->lock_flags = *flags;
844 ptlrpc_request_set_replen(req);
845 rc = ptlrpc_queue_wait(req);
849 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
851 GOTO(out, rc = -EPROTO);
854 GOTO(out, rc = req->rq_status);
856 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
858 ldlm_reprocess_all(res);
859 /* Go to sleep until the lock is granted. */
860 /* FIXME: or cancelled. */
861 if (lock->l_completion_ast) {
862 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
873 ptlrpc_req_finished(req);
877 /* Cancel locks locally.
879 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
880 * LDLM_FL_CANCELING otherwise;
881 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
882 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
884 int rc = LDLM_FL_LOCAL_ONLY;
887 if (lock->l_conn_export) {
890 LDLM_DEBUG(lock, "client-side cancel");
891 /* Set this flag to prevent others from getting new references*/
892 lock_res_and_lock(lock);
893 lock->l_flags |= LDLM_FL_CBPENDING;
894 local_only = (lock->l_flags &
895 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
896 ldlm_cancel_callback(lock);
897 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
898 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
899 unlock_res_and_lock(lock);
902 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
904 rc = LDLM_FL_LOCAL_ONLY;
906 ldlm_lock_cancel(lock);
908 if (ns_is_client(lock->l_resource->lr_namespace)) {
909 LDLM_ERROR(lock, "Trying to cancel local lock");
912 LDLM_DEBUG(lock, "server-side local cancel");
913 ldlm_lock_cancel(lock);
914 ldlm_reprocess_all(lock->l_resource);
915 LDLM_DEBUG(lock, "server-side local cancel handler END");
921 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
922 of the request @req. */
923 static void ldlm_cancel_pack(struct ptlrpc_request *req,
924 struct list_head *head, int count)
926 struct ldlm_request *dlm;
927 struct ldlm_lock *lock;
931 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
932 LASSERT(dlm != NULL);
934 /* Check the room in the request buffer. */
935 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
936 sizeof(struct ldlm_request);
937 max /= sizeof(struct lustre_handle);
938 max += LDLM_LOCKREQ_HANDLES;
939 LASSERT(max >= dlm->lock_count + count);
941 /* XXX: it would be better to pack lock handles grouped by resource.
942 * so that the server cancel would call filter_lvbo_update() less
944 list_for_each_entry(lock, head, l_bl_ast) {
947 LASSERT(lock->l_conn_export);
948 /* Pack the lock handle to the given request buffer. */
949 LDLM_DEBUG(lock, "packing");
950 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
953 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
957 /* Prepare and send a batched cancel rpc, it will include count lock handles
958 * of locks given in @head. */
959 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
960 int count, int flags)
962 struct ptlrpc_request *req = NULL;
963 struct obd_import *imp;
968 LASSERT(exp != NULL);
971 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
973 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
976 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
977 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
984 imp = class_exp2cliimp(exp);
985 if (imp == NULL || imp->imp_invalid) {
987 "skipping cancel on invalid import %p\n", imp);
991 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
993 GOTO(out, rc = -ENOMEM);
995 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
996 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
997 ldlm_request_bufsize(count, LDLM_CANCEL));
999 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1001 ptlrpc_request_free(req);
1004 req->rq_no_resend = 1;
1005 req->rq_no_delay = 1;
1007 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1008 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1009 ptlrpc_at_set_req_timeout(req);
1011 ldlm_cancel_pack(req, cancels, count);
1013 ptlrpc_request_set_replen(req);
1014 if (flags & LDLM_FL_ASYNC) {
1015 ptlrpcd_add_req(req);
1019 rc = ptlrpc_queue_wait(req);
1022 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1023 "out of sync -- not fatal\n",
1024 libcfs_nid2str(req->rq_import->
1025 imp_connection->c_peer.nid));
1027 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1028 req->rq_import_generation == imp->imp_generation) {
1029 ptlrpc_req_finished(req);
1031 } else if (rc != ELDLM_OK) {
1032 CERROR("Got rc %d from cancel RPC: canceling "
1040 ptlrpc_req_finished(req);
1043 return sent ? sent : rc;
1046 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1048 LASSERT(imp != NULL);
1049 return &imp->imp_obd->obd_namespace->ns_pool;
1053 * Update client's obd pool related fields with new SLV and Limit from \a req.
1055 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1057 struct obd_device *obd;
1058 __u64 old_slv, new_slv;
1062 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1063 !imp_connect_lru_resize(req->rq_import)))
1066 * Do nothing for corner cases.
1072 * In some cases RPC may contain slv and limit zeroed out. This is
1073 * the case when server does not support lru resize feature. This is
1074 * also possible in some recovery cases when server side reqs have no
1075 * ref to obd export and thus access to server side namespace is no
1078 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1079 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1080 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1081 "(SLV: "LPU64", Limit: %u)",
1082 lustre_msg_get_slv(req->rq_repmsg),
1083 lustre_msg_get_limit(req->rq_repmsg));
1087 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1088 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1089 obd = req->rq_import->imp_obd;
1092 * Set new SLV and Limit to obd fields to make accessible for pool
1093 * thread. We do not access obd_namespace and pool directly here
1094 * as there is no reliable way to make sure that they are still
1095 * alive in cleanup time. Evil races are possible which may cause
1096 * oops in that time.
1098 write_lock(&obd->obd_pool_lock);
1099 old_slv = obd->obd_pool_slv;
1100 obd->obd_pool_slv = new_slv;
1101 obd->obd_pool_limit = new_limit;
1102 write_unlock(&obd->obd_pool_lock);
1105 * Check if we need to wakeup pools thread for fast SLV change.
1106 * This is only done when threads period is noticably long like
1109 #if defined(__KERNEL__) && (LDLM_POOLS_THREAD_PERIOD >= 10)
1111 __u64 fast_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE;
1112 do_div(fast_change, 100);
1115 * Wake up pools thread only if SLV has changed more than
1116 * 50% since last update. In this case we want to react asap.
1117 * Otherwise it is no sense to wake up pools as they are
1118 * re-calculated every LDLM_POOLS_THREAD_PERIOD anyways.
1120 if (old_slv > new_slv && old_slv - new_slv > fast_change)
1121 ldlm_pools_wakeup();
1126 EXPORT_SYMBOL(ldlm_cli_update_pool);
1128 int ldlm_cli_cancel(struct lustre_handle *lockh)
1130 struct obd_export *exp;
1131 int avail, flags, count = 1, rc = 0;
1132 struct ldlm_namespace *ns;
1133 struct ldlm_lock *lock;
1134 CFS_LIST_HEAD(cancels);
1137 /* concurrent cancels on the same handle can happen */
1138 lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
1140 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1144 rc = ldlm_cli_cancel_local(lock);
1145 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1146 LDLM_LOCK_PUT(lock);
1147 RETURN(rc < 0 ? rc : 0);
1149 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1150 * rpc which goes to canceld portal, so we can cancel other lru locks
1151 * here and send them all as one LDLM_CANCEL rpc. */
1152 LASSERT(list_empty(&lock->l_bl_ast));
1153 list_add(&lock->l_bl_ast, &cancels);
1155 exp = lock->l_conn_export;
1156 if (exp_connect_cancelset(exp)) {
1157 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1162 ns = lock->l_resource->lr_namespace;
1163 flags = ns_connect_lru_resize(ns) ?
1164 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1165 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1166 LDLM_FL_BL_AST, flags);
1168 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1172 /* XXX until we will have compound requests and can cut cancels from generic rpc
1173 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1174 static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
1176 CFS_LIST_HEAD(head);
1177 struct ldlm_lock *lock, *next;
1178 int left = 0, bl_ast = 0, rc;
1181 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1185 if (flags & LDLM_FL_LOCAL_ONLY) {
1186 rc = LDLM_FL_LOCAL_ONLY;
1187 ldlm_lock_cancel(lock);
1189 rc = ldlm_cli_cancel_local(lock);
1191 if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1192 LDLM_DEBUG(lock, "Cancel lock separately");
1193 list_del_init(&lock->l_bl_ast);
1194 list_add(&lock->l_bl_ast, &head);
1198 if (rc == LDLM_FL_LOCAL_ONLY) {
1199 /* CANCEL RPC should not be sent to server. */
1200 list_del_init(&lock->l_bl_ast);
1201 LDLM_LOCK_PUT(lock);
1208 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1215 * Callback function for shrink policy. Makes decision whether to keep
1216 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1217 * \a added and number of locks to be preferably canceled \a count.
1219 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1221 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1223 static ldlm_policy_res_t ldlm_cancel_shrink_policy(struct ldlm_namespace *ns,
1224 struct ldlm_lock *lock,
1225 int unused, int added,
1232 * Stop lru processing when we reached passed @count or checked all
1235 if (count && added >= count)
1236 return LDLM_POLICY_KEEP_LOCK;
1238 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1239 struct ldlm_extent *l_extent;
1242 * For all extent locks cost is 1 + number of pages in
1245 l_extent = &lock->l_policy_data.l_extent;
1246 page_nr = (l_extent->end - l_extent->start);
1247 do_div(page_nr, CFS_PAGE_SIZE);
1251 * XXX: In fact this is evil hack, we can't access inode
1252 * here. For doing it right we need somehow to have number
1253 * of covered by lock. This should be fixed later when 10718
1256 if (lock->l_ast_data != NULL) {
1257 struct inode *inode = lock->l_ast_data;
1258 if (page_nr > inode->i_mapping->nrpages)
1259 page_nr = inode->i_mapping->nrpages;
1262 lock_cost = 1 + page_nr;
1265 * For all locks which are not extent ones cost is 1
1271 * Keep all expensive locks in lru for the memory pressure time
1272 * cancel policy. They anyways may be canceled by lru resize
1273 * pplicy if they have not small enough CLV.
1275 return lock_cost > ns->ns_shrink_thumb ?
1276 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1280 * Callback function for lru-resize policy. Makes decision whether to keep
1281 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1282 * \a added and number of locks to be preferably canceled \a count.
1284 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1286 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1288 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1289 struct ldlm_lock *lock,
1290 int unused, int added,
1293 cfs_time_t cur = cfs_time_current();
1294 struct ldlm_pool *pl = &ns->ns_pool;
1299 * Stop lru processing when we reached passed @count or checked all
1302 if (count && added >= count)
1303 return LDLM_POLICY_KEEP_LOCK;
1305 slv = ldlm_pool_get_slv(pl);
1306 lvf = ldlm_pool_get_lvf(pl);
1307 la = cfs_duration_sec(cfs_time_sub(cur,
1308 lock->l_last_used));
1311 * Stop when slv is not yet come from server or lv is smaller than
1314 lv = lvf * la * unused;
1317 * Inform pool about current CLV to see it via proc.
1319 ldlm_pool_set_clv(pl, lv);
1320 return (slv == 1 || lv < slv) ?
1321 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1325 * Callback function for proc used policy. Makes decision whether to keep
1326 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1327 * \a added and number of locks to be preferably canceled \a count.
1329 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1331 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1333 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1334 struct ldlm_lock *lock,
1335 int unused, int added,
1339 * Stop lru processing when we reached passed @count or checked all
1342 return (added >= count) ?
1343 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1347 * Callback function for aged policy. Makes decision whether to keep
1348 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1349 * \a added and number of locks to be preferably canceled \a count.
1351 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1353 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1355 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1356 struct ldlm_lock *lock,
1357 int unused, int added,
1361 * Stop lru processing if young lock is found and we reached passed
1364 return ((added >= count) &&
1365 cfs_time_before(cfs_time_current(),
1366 cfs_time_add(lock->l_last_used,
1368 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1372 * Callback function for default policy. Makes decision whether to keep
1373 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1374 * \a added and number of locks to be preferably canceled \a count.
1376 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1378 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1380 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1381 struct ldlm_lock *lock,
1382 int unused, int added,
1386 * Stop lru processing when we reached passed @count or checked all
1389 return (added >= count) ?
1390 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1393 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1394 struct ldlm_lock *, int,
1397 static ldlm_cancel_lru_policy_t
1398 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1400 if (ns_connect_lru_resize(ns)) {
1401 if (flags & LDLM_CANCEL_SHRINK)
1402 return ldlm_cancel_shrink_policy;
1403 else if (flags & LDLM_CANCEL_LRUR)
1404 return ldlm_cancel_lrur_policy;
1405 else if (flags & LDLM_CANCEL_PASSED)
1406 return ldlm_cancel_passed_policy;
1408 if (flags & LDLM_CANCEL_AGED)
1409 return ldlm_cancel_aged_policy;
1412 return ldlm_cancel_default_policy;
1415 /* - Free space in lru for @count new locks,
1416 * redundant unused locks are canceled locally;
1417 * - also cancel locally unused aged locks;
1418 * - do not cancel more than @max locks;
1419 * - GET the found locks and add them into the @cancels list.
1421 * A client lock can be added to the l_bl_ast list only when it is
1422 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1423 * There are the following use cases: ldlm_cancel_resource_local(),
1424 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1425 * flag properly. As any attempt to cancel a lock rely on this flag,
1426 * l_bl_ast list is accessed later without any special locking.
1428 * Calling policies for enabled lru resize:
1429 * ----------------------------------------
1430 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1431 * cancel not more than @count locks;
1433 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1434 * the beginning of lru list);
1436 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1437 * memory pressre policy function;
1439 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1441 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1442 int count, int max, int cancel_flags, int flags)
1444 ldlm_cancel_lru_policy_t pf;
1445 struct ldlm_lock *lock, *next;
1446 int added = 0, unused;
1449 spin_lock(&ns->ns_unused_lock);
1450 unused = ns->ns_nr_unused;
1452 if (!ns_connect_lru_resize(ns))
1453 count += unused - ns->ns_max_unused;
1455 pf = ldlm_cancel_lru_policy(ns, flags);
1456 LASSERT(pf != NULL);
1458 while (!list_empty(&ns->ns_unused_list)) {
1459 /* For any flags, stop scanning if @max is reached. */
1460 if (max && added >= max)
1463 list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru){
1464 /* No locks which got blocking requests. */
1465 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1467 /* Somebody is already doing CANCEL. No need in this
1468 * lock in lru, do not traverse it again. */
1469 if (!(lock->l_flags & LDLM_FL_CANCELING))
1472 ldlm_lock_remove_from_lru_nolock(lock);
1474 if (&lock->l_lru == &ns->ns_unused_list)
1477 /* Pass the lock through the policy filter and see if it
1478 * should stay in lru.
1480 * Even for shrinker policy we stop scanning if
1481 * we find a lock that should stay in the cache.
1482 * We should take into account lock age anyway
1483 * as new lock even if it is small of weight is
1484 * valuable resource.
1486 * That is, for shrinker policy we drop only
1487 * old locks, but additionally chose them by
1488 * their weight. Big extent locks will stay in
1490 if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
1493 LDLM_LOCK_GET(lock); /* dropped by bl thread */
1494 spin_unlock(&ns->ns_unused_lock);
1496 lock_res_and_lock(lock);
1497 /* Check flags again under the lock. */
1498 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1499 (ldlm_lock_remove_from_lru(lock) == 0)) {
1500 /* other thread is removing lock from lru or
1501 * somebody is already doing CANCEL or
1502 * there is a blocking request which will send
1503 * cancel by itseft or the lock is matched
1504 * is already not unused. */
1505 unlock_res_and_lock(lock);
1506 LDLM_LOCK_PUT(lock);
1507 spin_lock(&ns->ns_unused_lock);
1510 LASSERT(!lock->l_readers && !lock->l_writers);
1512 /* If we have chosen to cancel this lock voluntarily, we
1513 * better send cancel notification to server, so that it
1514 * frees appropriate state. This might lead to a race
1515 * where while we are doing cancel here, server is also
1516 * silently cancelling this lock. */
1517 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1519 /* Setting the CBPENDING flag is a little misleading,
1520 * but prevents an important race; namely, once
1521 * CBPENDING is set, the lock can accumulate no more
1522 * readers/writers. Since readers and writers are
1523 * already zero here, ldlm_lock_decref() won't see
1524 * this flag and call l_blocking_ast */
1525 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1527 /* We can't re-add to l_lru as it confuses the
1528 * refcounting in ldlm_lock_remove_from_lru() if an AST
1529 * arrives after we drop ns_lock below. We use l_bl_ast
1530 * and can't use l_pending_chain as it is used both on
1531 * server and client nevertheless bug 5666 says it is
1532 * used only on server */
1533 LASSERT(list_empty(&lock->l_bl_ast));
1534 list_add(&lock->l_bl_ast, cancels);
1535 unlock_res_and_lock(lock);
1536 spin_lock(&ns->ns_unused_lock);
1540 spin_unlock(&ns->ns_unused_lock);
1541 RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
1544 /* Returns number of locks which could be canceled next time when
1545 * ldlm_cancel_lru() is called. Used from locks pool shrinker. */
1546 int ldlm_cancel_lru_estimate(struct ldlm_namespace *ns,
1547 int count, int max, int flags)
1549 ldlm_cancel_lru_policy_t pf;
1550 struct ldlm_lock *lock;
1551 int added = 0, unused;
1554 pf = ldlm_cancel_lru_policy(ns, flags);
1555 LASSERT(pf != NULL);
1556 spin_lock(&ns->ns_unused_lock);
1557 unused = ns->ns_nr_unused;
1559 list_for_each_entry(lock, &ns->ns_unused_list, l_lru) {
1560 /* For any flags, stop scanning if @max is reached. */
1561 if (max && added >= max)
1564 /* Somebody is already doing CANCEL or there is a
1565 * blocking request will send cancel. Let's not count
1567 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1568 (lock->l_flags & LDLM_FL_BL_AST))
1571 /* Pass the lock through the policy filter and see if it
1572 * should stay in lru. */
1573 if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
1579 spin_unlock(&ns->ns_unused_lock);
1583 /* when called with LDLM_ASYNC the blocking callback will be handled
1584 * in a thread and this function will return after the thread has been
1585 * asked to call the callback. when called with LDLM_SYNC the blocking
1586 * callback will be performed in this function. */
1587 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1590 CFS_LIST_HEAD(cancels);
1595 sync = LDLM_SYNC; /* force to be sync in user space */
1597 count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
1598 if (sync == LDLM_ASYNC) {
1599 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1604 /* If an error occured in ASYNC mode, or
1605 * this is SYNC mode, cancel the list. */
1606 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1610 /* Find and cancel locally unused locks found on resource, matched to the
1611 * given policy, mode. GET the found locks and add them into the @cancels
1613 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1614 struct list_head *cancels,
1615 ldlm_policy_data_t *policy,
1616 ldlm_mode_t mode, int lock_flags,
1617 int cancel_flags, void *opaque)
1619 struct ldlm_lock *lock;
1624 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1625 if (opaque != NULL && lock->l_ast_data != opaque) {
1626 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1627 lock->l_ast_data, opaque);
1632 if (lock->l_readers || lock->l_writers) {
1633 if (cancel_flags & LDLM_FL_WARN) {
1634 LDLM_ERROR(lock, "lock in use");
1640 /* If somebody is already doing CANCEL, or blocking ast came,
1641 * skip this lock. */
1642 if (lock->l_flags & LDLM_FL_BL_AST ||
1643 lock->l_flags & LDLM_FL_CANCELING)
1646 if (lockmode_compat(lock->l_granted_mode, mode))
1649 /* If policy is given and this is IBITS lock, add to list only
1650 * those locks that match by policy. */
1651 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1652 !(lock->l_policy_data.l_inodebits.bits &
1653 policy->l_inodebits.bits))
1656 /* See CBPENDING comment in ldlm_cancel_lru */
1657 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1660 LASSERT(list_empty(&lock->l_bl_ast));
1661 list_add(&lock->l_bl_ast, cancels);
1662 LDLM_LOCK_GET(lock);
1667 RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
1670 /* If @req is NULL, send CANCEL request to server with handles of locks
1671 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1672 * separately per lock.
1673 * If @req is not NULL, put handles of locks in @cancels into the request
1674 * buffer at the offset @off.
1675 * Destroy @cancels at the end. */
1676 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1677 struct ptlrpc_request *req, int flags)
1679 struct ldlm_lock *lock;
1683 if (list_empty(cancels) || count == 0)
1686 /* XXX: requests (both batched and not) could be sent in parallel.
1687 * Usually it is enough to have just 1 RPC, but it is possible that
1688 * there are to many locks to be cancelled in LRU or on a resource.
1689 * It would also speed up the case when the server does not support
1692 LASSERT(!list_empty(cancels));
1693 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1694 LASSERT(lock->l_conn_export);
1696 if (exp_connect_cancelset(lock->l_conn_export)) {
1699 ldlm_cancel_pack(req, cancels, count);
1701 res = ldlm_cli_cancel_req(lock->l_conn_export,
1705 res = ldlm_cli_cancel_req(lock->l_conn_export,
1710 CERROR("ldlm_cli_cancel_list: %d\n", res);
1715 ldlm_lock_list_put(cancels, l_bl_ast, res);
1717 LASSERT(count == 0);
1721 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1722 const struct ldlm_res_id *res_id,
1723 ldlm_policy_data_t *policy,
1724 ldlm_mode_t mode, int flags, void *opaque)
1726 struct ldlm_resource *res;
1727 CFS_LIST_HEAD(cancels);
1732 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1734 /* This is not a problem. */
1735 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1739 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1741 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1743 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1745 ldlm_resource_putref(res);
1749 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1751 int no_resource = 0;
1753 spin_lock(&ns->ns_hash_lock);
1754 if (ns->ns_resources == 0)
1756 spin_unlock(&ns->ns_hash_lock);
1758 RETURN(no_resource);
1761 /* Cancel all locks on a namespace (or a specific resource, if given)
1762 * that have 0 readers/writers.
1764 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1765 * to notify the server. */
1766 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1767 const struct ldlm_res_id *res_id,
1768 int flags, void *opaque)
1777 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1781 spin_lock(&ns->ns_hash_lock);
1782 for (i = 0; i < RES_HASH_SIZE; i++) {
1783 struct list_head *tmp;
1784 tmp = ns->ns_hash[i].next;
1785 while (tmp != &(ns->ns_hash[i])) {
1786 struct ldlm_resource *res;
1789 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1790 ldlm_resource_getref(res);
1791 spin_unlock(&ns->ns_hash_lock);
1793 rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1798 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1799 res->lr_name.name[0], rc);
1801 spin_lock(&ns->ns_hash_lock);
1803 ldlm_resource_putref_locked(res);
1806 spin_unlock(&ns->ns_hash_lock);
1811 /* join/split resource locks to/from lru list */
1812 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
1813 const struct ldlm_res_id *res_id, int join)
1815 struct ldlm_resource *res;
1816 struct ldlm_lock *lock, *n;
1820 LASSERT(ns_is_client(ns));
1822 res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
1825 LASSERT(res->lr_type == LDLM_EXTENT);
1831 list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1832 if (list_empty(&lock->l_lru) &&
1833 !lock->l_readers && !lock->l_writers &&
1834 !(lock->l_flags & LDLM_FL_LOCAL) &&
1835 !(lock->l_flags & LDLM_FL_CBPENDING) &&
1836 !(lock->l_flags & LDLM_FL_BL_AST)) {
1837 ldlm_lock_add_to_lru(lock);
1838 lock->l_flags &= ~LDLM_FL_NO_LRU;
1839 LDLM_DEBUG(lock, "join lock to lru");
1845 spin_lock(&ns->ns_unused_lock);
1846 list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1847 if (lock->l_resource == res) {
1848 ldlm_lock_remove_from_lru_nolock(lock);
1849 lock->l_flags |= LDLM_FL_NO_LRU;
1850 LDLM_DEBUG(lock, "split lock from lru");
1854 spin_unlock(&ns->ns_unused_lock);
1857 ldlm_resource_putref(res);
1861 /* Lock iterators. */
1863 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1866 struct list_head *tmp, *next;
1867 struct ldlm_lock *lock;
1868 int rc = LDLM_ITER_CONTINUE;
1873 RETURN(LDLM_ITER_CONTINUE);
1876 list_for_each_safe(tmp, next, &res->lr_granted) {
1877 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1879 if (iter(lock, closure) == LDLM_ITER_STOP)
1880 GOTO(out, rc = LDLM_ITER_STOP);
1883 list_for_each_safe(tmp, next, &res->lr_converting) {
1884 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1886 if (iter(lock, closure) == LDLM_ITER_STOP)
1887 GOTO(out, rc = LDLM_ITER_STOP);
1890 list_for_each_safe(tmp, next, &res->lr_waiting) {
1891 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1893 if (iter(lock, closure) == LDLM_ITER_STOP)
1894 GOTO(out, rc = LDLM_ITER_STOP);
1901 struct iter_helper_data {
1902 ldlm_iterator_t iter;
1906 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1908 struct iter_helper_data *helper = closure;
1909 return helper->iter(lock, helper->closure);
1912 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1914 return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1917 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1920 struct iter_helper_data helper = { iter: iter, closure: closure };
1921 return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1924 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1925 ldlm_res_iterator_t iter, void *closure)
1927 int i, rc = LDLM_ITER_CONTINUE;
1928 struct ldlm_resource *res;
1929 struct list_head *tmp;
1932 spin_lock(&ns->ns_hash_lock);
1933 for (i = 0; i < RES_HASH_SIZE; i++) {
1934 tmp = ns->ns_hash[i].next;
1935 while (tmp != &(ns->ns_hash[i])) {
1936 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1937 ldlm_resource_getref(res);
1938 spin_unlock(&ns->ns_hash_lock);
1940 rc = iter(res, closure);
1942 spin_lock(&ns->ns_hash_lock);
1944 ldlm_resource_putref_locked(res);
1945 if (rc == LDLM_ITER_STOP)
1950 spin_unlock(&ns->ns_hash_lock);
1954 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1955 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1956 const struct ldlm_res_id *res_id,
1957 ldlm_iterator_t iter, void *data)
1959 struct ldlm_resource *res;
1963 CERROR("must pass in namespace\n");
1967 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1973 ldlm_resource_foreach(res, iter, data);
1974 ldlm_resource_putref(res);
1980 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1982 struct list_head *list = closure;
1984 /* we use l_pending_chain here, because it's unused on clients. */
1985 LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1986 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1987 /* bug 9573: don't replay locks left after eviction */
1988 if (!(lock->l_flags & LDLM_FL_FAILED))
1989 list_add(&lock->l_pending_chain, list);
1990 return LDLM_ITER_CONTINUE;
1993 static int replay_lock_interpret(struct ptlrpc_request *req,
1994 struct ldlm_async_args *aa, int rc)
1996 struct ldlm_lock *lock;
1997 struct ldlm_reply *reply;
2000 atomic_dec(&req->rq_import->imp_replay_inflight);
2005 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2007 GOTO(out, rc = -EPROTO);
2009 lock = ldlm_handle2lock(&aa->lock_handle);
2011 CERROR("received replay ack for unknown local cookie "LPX64
2012 " remote cookie "LPX64 " from server %s id %s\n",
2013 aa->lock_handle.cookie, reply->lock_handle.cookie,
2014 req->rq_export->exp_client_uuid.uuid,
2015 libcfs_id2str(req->rq_peer));
2016 GOTO(out, rc = -ESTALE);
2019 lock->l_remote_handle = reply->lock_handle;
2020 LDLM_DEBUG(lock, "replayed lock:");
2021 ptlrpc_import_recovery_state_machine(req->rq_import);
2022 LDLM_LOCK_PUT(lock);
2025 ptlrpc_connect_import(req->rq_import, NULL);
2031 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2033 struct ptlrpc_request *req;
2034 struct ldlm_async_args *aa;
2035 struct ldlm_request *body;
2040 /* Bug 11974: Do not replay a lock which is actively being canceled */
2041 if (lock->l_flags & LDLM_FL_CANCELING) {
2042 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2046 /* If this is reply-less callback lock, we cannot replay it, since
2047 * server might have long dropped it, but notification of that event was
2048 * lost by network. (and server granted conflicting lock already) */
2049 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2050 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2051 ldlm_lock_cancel(lock);
2055 * If granted mode matches the requested mode, this lock is granted.
2057 * If they differ, but we have a granted mode, then we were granted
2058 * one mode and now want another: ergo, converting.
2060 * If we haven't been granted anything and are on a resource list,
2061 * then we're blocked/waiting.
2063 * If we haven't been granted anything and we're NOT on a resource list,
2064 * then we haven't got a reply yet and don't have a known disposition.
2065 * This happens whenever a lock enqueue is the request that triggers
2068 if (lock->l_granted_mode == lock->l_req_mode)
2069 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2070 else if (lock->l_granted_mode)
2071 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2072 else if (!list_empty(&lock->l_res_link))
2073 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2075 flags = LDLM_FL_REPLAY;
2077 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2078 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2082 /* We're part of recovery, so don't wait for it. */
2083 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2085 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2086 ldlm_lock2desc(lock, &body->lock_desc);
2087 body->lock_flags = flags;
2089 ldlm_lock2handle(lock, &body->lock_handle[0]);
2090 if (lock->l_lvb_len != 0) {
2091 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2092 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2095 ptlrpc_request_set_replen(req);
2096 /* notify the server we've replayed all requests.
2097 * also, we mark the request to be put on a dedicated
2098 * queue to be processed after all request replayes.
2100 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2102 LDLM_DEBUG(lock, "replaying lock:");
2104 atomic_inc(&req->rq_import->imp_replay_inflight);
2105 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2106 aa = ptlrpc_req_async_args(req);
2107 aa->lock_handle = body->lock_handle[0];
2108 req->rq_interpret_reply = replay_lock_interpret;
2109 ptlrpcd_add_req(req);
2114 int ldlm_replay_locks(struct obd_import *imp)
2116 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2117 CFS_LIST_HEAD(list);
2118 struct ldlm_lock *lock, *next;
2123 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2125 /* ensure this doesn't fall to 0 before all have been queued */
2126 atomic_inc(&imp->imp_replay_inflight);
2128 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2130 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2131 list_del_init(&lock->l_pending_chain);
2133 continue; /* or try to do the rest? */
2134 rc = replay_one_lock(imp, lock);
2137 atomic_dec(&imp->imp_replay_inflight);