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 [sun.com URL with a
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_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_current_sec() - lock->l_enqueued_time.tv_sec,
104 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
109 /* We use the same basis for both server side and client side functions
110 from a single node. */
111 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
113 int timeout = at_get(&lock->l_resource->lr_namespace->ns_at_estimate);
115 return obd_timeout / 2;
116 /* Since these are non-updating timeouts, we should be conservative.
117 It would be nice to have some kind of "early reply" mechanism for
118 lock callbacks too... */
119 timeout = timeout + (timeout >> 1); /* 150% */
120 return max(timeout, ldlm_enqueue_min);
123 static int is_granted_or_cancelled(struct ldlm_lock *lock)
127 lock_res_and_lock(lock);
128 if (((lock->l_req_mode == lock->l_granted_mode) &&
129 !(lock->l_flags & LDLM_FL_CP_REQD)) ||
130 (lock->l_flags & LDLM_FL_FAILED))
132 unlock_res_and_lock(lock);
137 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
139 /* XXX ALLOCATE - 160 bytes */
140 struct lock_wait_data lwd;
141 struct obd_device *obd;
142 struct obd_import *imp = NULL;
143 struct l_wait_info lwi;
148 if (flags == LDLM_FL_WAIT_NOREPROC) {
149 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
153 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
154 LDLM_FL_BLOCK_CONV))) {
155 cfs_waitq_signal(&lock->l_waitq);
159 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
161 ldlm_lock_dump(D_OTHER, lock, 0);
162 ldlm_reprocess_all(lock->l_resource);
166 obd = class_exp2obd(lock->l_conn_export);
168 /* if this is a local lock, then there is no import */
170 imp = obd->u.cli.cl_import;
173 /* Wait a long time for enqueue - server may have to callback a
174 lock from another client. Server will evict the other client if it
175 doesn't respond reasonably, and then give us the lock. */
176 timeout = ldlm_get_enq_timeout(lock) * 2;
180 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
181 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
182 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
184 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
185 ldlm_expired_completion_wait,
186 interrupted_completion_wait, &lwd);
190 spin_lock(&imp->imp_lock);
191 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
192 spin_unlock(&imp->imp_lock);
195 /* Go to sleep until the lock is granted or cancelled. */
196 rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
198 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
199 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
204 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
209 LDLM_DEBUG(lock, "client-side enqueue waking up: granted after %lds",
210 cfs_time_current_sec() - lock->l_enqueued_time.tv_sec);
212 /* Update our time estimate */
213 at_add(&lock->l_resource->lr_namespace->ns_at_estimate,
214 cfs_time_current_sec() - lock->l_enqueued_time.tv_sec);
220 * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
222 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
223 void *data, int flag)
228 if (flag == LDLM_CB_CANCELING) {
229 /* Don't need to do anything here. */
233 lock_res_and_lock(lock);
234 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
235 * that ldlm_blocking_ast is called just before intent_policy method
236 * takes the ns_lock, then by the time we get the lock, we might not
237 * be the correct blocking function anymore. So check, and return
239 if (lock->l_blocking_ast != ldlm_blocking_ast) {
240 unlock_res_and_lock(lock);
244 lock->l_flags |= LDLM_FL_CBPENDING;
245 do_ast = (!lock->l_readers && !lock->l_writers);
246 unlock_res_and_lock(lock);
249 struct lustre_handle lockh;
252 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
253 ldlm_lock2handle(lock, &lockh);
254 rc = ldlm_cli_cancel(&lockh);
256 CERROR("ldlm_cli_cancel: %d\n", rc);
258 LDLM_DEBUG(lock, "Lock still has references, will be "
265 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
266 * comment in filter_intent_policy() on why you may need this.
268 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
271 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
272 * that is rather subtle: with OST-side locking, it may so happen that
273 * _all_ extent locks are held by the OST. If client wants to obtain
274 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
275 * on the server), dummy glimpse callback fires and does
276 * nothing. Client still receives correct file size due to the
277 * following fragment in filter_intent_policy():
279 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
280 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
281 * res->lr_namespace->ns_lvbo->lvbo_update) {
282 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
285 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
286 * returns correct file size to the client.
288 return -ELDLM_NO_LOCK_DATA;
291 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
292 const struct ldlm_res_id *res_id,
293 ldlm_type_t type, ldlm_policy_data_t *policy,
294 ldlm_mode_t mode, int *flags,
295 ldlm_blocking_callback blocking,
296 ldlm_completion_callback completion,
297 ldlm_glimpse_callback glimpse,
298 void *data, __u32 lvb_len, void *lvb_swabber,
299 struct lustre_handle *lockh)
301 struct ldlm_lock *lock;
305 LASSERT(!(*flags & LDLM_FL_REPLAY));
306 if (unlikely(ns_is_client(ns))) {
307 CERROR("Trying to enqueue local lock in a shadow namespace\n");
311 lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
312 completion, glimpse, data, lvb_len);
314 GOTO(out_nolock, err = -ENOMEM);
315 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
317 ldlm_lock_addref_internal(lock, mode);
318 ldlm_lock2handle(lock, lockh);
319 lock_res_and_lock(lock);
320 lock->l_flags |= LDLM_FL_LOCAL;
321 if (*flags & LDLM_FL_ATOMIC_CB)
322 lock->l_flags |= LDLM_FL_ATOMIC_CB;
323 lock->l_lvb_swabber = lvb_swabber;
324 unlock_res_and_lock(lock);
326 lock->l_policy_data = *policy;
327 if (type == LDLM_EXTENT)
328 lock->l_req_extent = policy->l_extent;
330 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
331 if (unlikely(err != ELDLM_OK))
335 *policy = lock->l_policy_data;
337 LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
340 if (lock->l_completion_ast)
341 lock->l_completion_ast(lock, *flags, NULL);
343 LDLM_DEBUG(lock, "client-side local enqueue END");
351 static void failed_lock_cleanup(struct ldlm_namespace *ns,
352 struct ldlm_lock *lock,
353 struct lustre_handle *lockh, int mode)
355 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
356 lock_res_and_lock(lock);
357 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
358 unlock_res_and_lock(lock);
359 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
361 ldlm_lock_decref_and_cancel(lockh, mode);
363 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
364 * from llite/file.c/ll_file_flock(). */
365 if (lock->l_resource->lr_type == LDLM_FLOCK) {
366 ldlm_lock_destroy(lock);
370 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
371 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
372 int *flags, void *lvb, __u32 lvb_len,
373 void *lvb_swabber, struct lustre_handle *lockh,int rc)
375 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
376 int is_replay = *flags & LDLM_FL_REPLAY;
377 struct ldlm_lock *lock;
378 struct ldlm_reply *reply;
379 int cleanup_phase = 1;
382 lock = ldlm_handle2lock(lockh);
383 /* ldlm_cli_enqueue is holding a reference on this lock. */
385 LASSERT(type == LDLM_FLOCK);
389 if (rc != ELDLM_OK) {
391 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
392 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
393 if (rc == ELDLM_LOCK_ABORTED) {
394 /* Before we return, swab the reply */
395 reply = req_capsule_server_get(&req->rq_pill,
400 struct ost_lvb *tmplvb;
402 req_capsule_set_size(&req->rq_pill,
403 &RMF_DLM_LVB, RCL_SERVER,
405 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
409 GOTO(cleanup, rc = -EPROTO);
411 memcpy(lvb, tmplvb, lvb_len);
417 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
419 GOTO(cleanup, rc = -EPROTO);
421 /* lock enqueued on the server */
424 lock_res_and_lock(lock);
425 lock->l_remote_handle = reply->lock_handle;
426 *flags = reply->lock_flags;
427 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
428 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
429 * to wait with no timeout as well */
430 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
431 unlock_res_and_lock(lock);
433 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
434 lock, reply->lock_handle.cookie, *flags);
436 /* If enqueue returned a blocked lock but the completion handler has
437 * already run, then it fixed up the resource and we don't need to do it
439 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
440 int newmode = reply->lock_desc.l_req_mode;
442 if (newmode && newmode != lock->l_req_mode) {
443 LDLM_DEBUG(lock, "server returned different mode %s",
444 ldlm_lockname[newmode]);
445 lock->l_req_mode = newmode;
448 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
449 lock->l_resource->lr_name.name,
450 sizeof(struct ldlm_res_id))) {
451 CDEBUG(D_INFO, "remote intent success, locking "
452 "(%ld,%ld,%ld) instead of "
454 (long)reply->lock_desc.l_resource.lr_name.name[0],
455 (long)reply->lock_desc.l_resource.lr_name.name[1],
456 (long)reply->lock_desc.l_resource.lr_name.name[2],
457 (long)lock->l_resource->lr_name.name[0],
458 (long)lock->l_resource->lr_name.name[1],
459 (long)lock->l_resource->lr_name.name[2]);
461 rc = ldlm_lock_change_resource(ns, lock,
462 &reply->lock_desc.l_resource.lr_name);
463 if (rc || lock->l_resource == NULL)
464 GOTO(cleanup, rc = -ENOMEM);
465 LDLM_DEBUG(lock, "client-side enqueue, new resource");
468 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
470 lock->l_policy_data =
471 reply->lock_desc.l_policy_data;
472 if (type != LDLM_PLAIN)
473 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
476 if ((*flags) & LDLM_FL_AST_SENT ||
477 /* Cancel extent locks as soon as possible on a liblustre client,
478 * because it cannot handle asynchronous ASTs robustly (see
480 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
481 lock_res_and_lock(lock);
482 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
483 unlock_res_and_lock(lock);
484 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
487 /* If the lock has already been granted by a completion AST, don't
488 * clobber the LVB with an older one. */
489 if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
492 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
494 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
498 GOTO(cleanup, rc = -EPROTO);
499 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
503 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
504 if (lock->l_completion_ast != NULL) {
505 int err = lock->l_completion_ast(lock, *flags, NULL);
508 if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
513 if (lvb_len && lvb != NULL) {
514 /* Copy the LVB here, and not earlier, because the completion
515 * AST (if any) can override what we got in the reply */
516 memcpy(lvb, lock->l_lvb_data, lvb_len);
519 LDLM_DEBUG(lock, "client-side enqueue END");
522 if (cleanup_phase == 1 && rc)
523 failed_lock_cleanup(ns, lock, lockh, mode);
524 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
530 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
531 * a single page on the send/receive side. XXX: 512 should be changed
532 * to more adequate value. */
533 static inline int ldlm_req_handles_avail(int req_size, int off)
537 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
538 avail /= sizeof(struct lustre_handle);
539 avail += LDLM_LOCKREQ_HANDLES - off;
544 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
545 enum req_location loc,
548 int size = req_capsule_msg_size(pill, loc);
549 return ldlm_req_handles_avail(size, off);
552 static inline int ldlm_format_handles_avail(struct obd_import *imp,
553 const struct req_format *fmt,
554 enum req_location loc, int off)
556 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
557 return ldlm_req_handles_avail(size, off);
560 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
561 * @count locks in @cancels. */
562 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
563 int version, int opc, int canceloff,
564 struct list_head *cancels, int count)
566 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
567 struct req_capsule *pill = &req->rq_pill;
568 struct ldlm_request *dlm = NULL;
569 int flags, avail, to_free, bufcount, pack = 0;
576 if (exp_connect_cancelset(exp)) {
577 /* Estimate the amount of available space in the request. */
578 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
579 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
581 flags = ns_connect_lru_resize(ns) ?
582 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
583 to_free = !ns_connect_lru_resize(ns) &&
584 opc == LDLM_ENQUEUE ? 1 : 0;
586 /* Cancel lru locks here _only_ if the server supports
587 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
588 * rpc, what will make us slower. */
590 count += ldlm_cancel_lru_local(ns, cancels, to_free,
591 avail - count, 0, flags);
596 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
597 ldlm_request_bufsize(pack, opc));
600 rc = ptlrpc_request_pack(req, version, opc);
602 ldlm_lock_list_put(cancels, l_bl_ast, count);
606 if (exp_connect_cancelset(exp)) {
608 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
610 /* Skip first lock handler in ldlm_request_pack(),
611 * this method will incrment @lock_count according
612 * to the lock handle amount actually written to
614 dlm->lock_count = canceloff;
616 /* Pack into the request @pack lock handles. */
617 ldlm_cli_cancel_list(cancels, pack, req, 0);
618 /* Prepare and send separate cancel rpc for others. */
619 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
621 ldlm_lock_list_put(cancels, l_bl_ast, count);
626 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
627 struct list_head *cancels, int count)
629 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
630 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
633 /* If a request has some specific initialisation it is passed in @reqp,
634 * otherwise it is created in ldlm_cli_enqueue.
636 * Supports sync and async requests, pass @async flag accordingly. If a
637 * request was created in ldlm_cli_enqueue and it is the async request,
638 * pass it to the caller in @reqp. */
639 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
640 struct ldlm_enqueue_info *einfo,
641 const struct ldlm_res_id *res_id,
642 ldlm_policy_data_t *policy, int *flags,
643 void *lvb, __u32 lvb_len, void *lvb_swabber,
644 struct lustre_handle *lockh, int async)
646 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
647 struct ldlm_lock *lock;
648 struct ldlm_request *body;
649 int is_replay = *flags & LDLM_FL_REPLAY;
650 int req_passed_in = 1;
652 struct ptlrpc_request *req;
655 LASSERT(exp != NULL);
657 /* If we're replaying this lock, just check some invariants.
658 * If we're creating a new lock, get everything all setup nice. */
660 lock = ldlm_handle2lock(lockh);
661 LASSERT(lock != NULL);
662 LDLM_DEBUG(lock, "client-side enqueue START");
663 LASSERT(exp == lock->l_conn_export);
665 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
666 einfo->ei_mode, einfo->ei_cb_bl,
667 einfo->ei_cb_cp, einfo->ei_cb_gl,
668 einfo->ei_cbdata, lvb_len);
671 /* for the local lock, add the reference */
672 ldlm_lock_addref_internal(lock, einfo->ei_mode);
673 ldlm_lock2handle(lock, lockh);
674 lock->l_lvb_swabber = lvb_swabber;
675 if (policy != NULL) {
676 /* INODEBITS_INTEROP: If the server does not support
677 * inodebits, we will request a plain lock in the
678 * descriptor (ldlm_lock2desc() below) but use an
679 * inodebits lock internally with both bits set.
681 if (einfo->ei_type == LDLM_IBITS &&
682 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
683 lock->l_policy_data.l_inodebits.bits =
684 MDS_INODELOCK_LOOKUP |
685 MDS_INODELOCK_UPDATE;
687 lock->l_policy_data = *policy;
690 if (einfo->ei_type == LDLM_EXTENT)
691 lock->l_req_extent = policy->l_extent;
692 LDLM_DEBUG(lock, "client-side enqueue START");
695 /* lock not sent to server yet */
697 if (reqp == NULL || *reqp == NULL) {
698 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
703 failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
714 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
716 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
717 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
720 lock->l_conn_export = exp;
721 lock->l_export = NULL;
722 lock->l_blocking_ast = einfo->ei_cb_bl;
724 /* Dump lock data into the request buffer */
725 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
726 ldlm_lock2desc(lock, &body->lock_desc);
727 body->lock_flags = *flags;
728 body->lock_handle[0] = *lockh;
730 /* Continue as normal. */
731 if (!req_passed_in) {
733 req_capsule_extend(&req->rq_pill,
734 &RQF_LDLM_ENQUEUE_LVB);
735 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
736 RCL_SERVER, lvb_len);
738 ptlrpc_request_set_replen(req);
742 * Liblustre client doesn't get extent locks, except for O_APPEND case
743 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
744 * [i_size, OBD_OBJECT_EOF] lock is taken.
746 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
747 policy->l_extent.end == OBD_OBJECT_EOF));
750 LASSERT(reqp != NULL);
754 LDLM_DEBUG(lock, "sending request");
755 rc = ptlrpc_queue_wait(req);
756 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
757 einfo->ei_mode, flags, lvb, lvb_len,
758 lvb_swabber, lockh, rc);
760 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
761 * one reference that we took */
767 if (!req_passed_in && req != NULL) {
768 ptlrpc_req_finished(req);
776 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
779 struct ldlm_resource *res;
782 if (ns_is_client(lock->l_resource->lr_namespace)) {
783 CERROR("Trying to cancel local lock\n");
786 LDLM_DEBUG(lock, "client-side local convert");
788 res = ldlm_lock_convert(lock, new_mode, flags);
790 ldlm_reprocess_all(res);
795 LDLM_DEBUG(lock, "client-side local convert handler END");
800 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
801 * conversion of locks which are on the waiting or converting queue */
802 /* Caller of this code is supposed to take care of lock readers/writers
804 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
806 struct ldlm_request *body;
807 struct ldlm_reply *reply;
808 struct ldlm_lock *lock;
809 struct ldlm_resource *res;
810 struct ptlrpc_request *req;
814 lock = ldlm_handle2lock(lockh);
821 if (lock->l_conn_export == NULL)
822 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
824 LDLM_DEBUG(lock, "client-side convert");
826 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
827 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
834 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
835 body->lock_handle[0] = lock->l_remote_handle;
837 body->lock_desc.l_req_mode = new_mode;
838 body->lock_flags = *flags;
841 ptlrpc_request_set_replen(req);
842 rc = ptlrpc_queue_wait(req);
846 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
848 GOTO(out, rc = -EPROTO);
851 GOTO(out, rc = req->rq_status);
853 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
855 ldlm_reprocess_all(res);
856 /* Go to sleep until the lock is granted. */
857 /* FIXME: or cancelled. */
858 if (lock->l_completion_ast) {
859 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
870 ptlrpc_req_finished(req);
874 /* Cancel locks locally.
876 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
877 * LDLM_FL_CANCELING otherwise;
878 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
879 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
881 int rc = LDLM_FL_LOCAL_ONLY;
884 if (lock->l_conn_export) {
887 LDLM_DEBUG(lock, "client-side cancel");
888 /* Set this flag to prevent others from getting new references*/
889 lock_res_and_lock(lock);
890 lock->l_flags |= LDLM_FL_CBPENDING;
891 local_only = (lock->l_flags &
892 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
893 ldlm_cancel_callback(lock);
894 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
895 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
896 unlock_res_and_lock(lock);
899 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
901 rc = LDLM_FL_LOCAL_ONLY;
903 ldlm_lock_cancel(lock);
905 if (ns_is_client(lock->l_resource->lr_namespace)) {
906 LDLM_ERROR(lock, "Trying to cancel local lock");
909 LDLM_DEBUG(lock, "server-side local cancel");
910 ldlm_lock_cancel(lock);
911 ldlm_reprocess_all(lock->l_resource);
912 LDLM_DEBUG(lock, "server-side local cancel handler END");
918 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
919 of the request @req. */
920 static void ldlm_cancel_pack(struct ptlrpc_request *req,
921 struct list_head *head, int count)
923 struct ldlm_request *dlm;
924 struct ldlm_lock *lock;
928 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
929 LASSERT(dlm != NULL);
931 /* Check the room in the request buffer. */
932 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
933 sizeof(struct ldlm_request);
934 max /= sizeof(struct lustre_handle);
935 max += LDLM_LOCKREQ_HANDLES;
936 LASSERT(max >= dlm->lock_count + count);
938 /* XXX: it would be better to pack lock handles grouped by resource.
939 * so that the server cancel would call filter_lvbo_update() less
941 list_for_each_entry(lock, head, l_bl_ast) {
944 LASSERT(lock->l_conn_export);
945 /* Pack the lock handle to the given request buffer. */
946 LDLM_DEBUG(lock, "packing");
947 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
950 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
954 /* Prepare and send a batched cancel rpc, it will include count lock handles
955 * of locks given in @head. */
956 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
957 int count, int flags)
959 struct ptlrpc_request *req = NULL;
960 struct obd_import *imp;
965 LASSERT(exp != NULL);
968 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
970 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
973 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
974 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
981 imp = class_exp2cliimp(exp);
982 if (imp == NULL || imp->imp_invalid) {
984 "skipping cancel on invalid import %p\n", imp);
988 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
990 GOTO(out, rc = -ENOMEM);
992 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
993 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
994 ldlm_request_bufsize(count, LDLM_CANCEL));
996 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
998 ptlrpc_request_free(req);
1001 req->rq_no_resend = 1;
1002 req->rq_no_delay = 1;
1004 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1005 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1006 ptlrpc_at_set_req_timeout(req);
1008 ldlm_cancel_pack(req, cancels, count);
1010 ptlrpc_request_set_replen(req);
1011 if (flags & LDLM_FL_ASYNC) {
1012 ptlrpcd_add_req(req);
1016 rc = ptlrpc_queue_wait(req);
1019 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1020 "out of sync -- not fatal\n",
1021 libcfs_nid2str(req->rq_import->
1022 imp_connection->c_peer.nid));
1024 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1025 req->rq_import_generation == imp->imp_generation) {
1026 ptlrpc_req_finished(req);
1028 } else if (rc != ELDLM_OK) {
1029 CERROR("Got rc %d from cancel RPC: canceling "
1037 ptlrpc_req_finished(req);
1040 return sent ? sent : rc;
1043 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1045 LASSERT(imp != NULL);
1046 return &imp->imp_obd->obd_namespace->ns_pool;
1050 * Update client's obd pool related fields with new SLV and Limit from \a req.
1052 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1054 struct obd_device *obd;
1055 __u64 old_slv, new_slv;
1059 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1060 !imp_connect_lru_resize(req->rq_import)))
1063 * Do nothing for corner cases.
1069 * In some cases RPC may contain slv and limit zeroed out. This is
1070 * the case when server does not support lru resize feature. This is
1071 * also possible in some recovery cases when server side reqs have no
1072 * ref to obd export and thus access to server side namespace is no
1075 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1076 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1077 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1078 "(SLV: "LPU64", Limit: %u)",
1079 lustre_msg_get_slv(req->rq_repmsg),
1080 lustre_msg_get_limit(req->rq_repmsg));
1084 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1085 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1086 obd = req->rq_import->imp_obd;
1089 * Set new SLV and Limit to obd fields to make accessible for pool
1090 * thread. We do not access obd_namespace and pool directly here
1091 * as there is no reliable way to make sure that they are still
1092 * alive in cleanup time. Evil races are possible which may cause
1093 * oops in that time.
1095 write_lock(&obd->obd_pool_lock);
1096 old_slv = obd->obd_pool_slv;
1097 obd->obd_pool_slv = new_slv;
1098 obd->obd_pool_limit = new_limit;
1099 write_unlock(&obd->obd_pool_lock);
1102 * Check if we need to wakeup pools thread for fast SLV change.
1103 * This is only done when threads period is noticably long like
1106 #if defined(__KERNEL__) && (LDLM_POOLS_THREAD_PERIOD >= 10)
1108 __u64 fast_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE;
1109 do_div(fast_change, 100);
1112 * Wake up pools thread only if SLV has changed more than
1113 * 50% since last update. In this case we want to react asap.
1114 * Otherwise it is no sense to wake up pools as they are
1115 * re-calculated every LDLM_POOLS_THREAD_PERIOD anyways.
1117 if (old_slv > new_slv && old_slv - new_slv > fast_change)
1118 ldlm_pools_wakeup();
1123 EXPORT_SYMBOL(ldlm_cli_update_pool);
1125 int ldlm_cli_cancel(struct lustre_handle *lockh)
1127 struct obd_export *exp;
1128 int avail, flags, count = 1, rc = 0;
1129 struct ldlm_namespace *ns;
1130 struct ldlm_lock *lock;
1131 CFS_LIST_HEAD(cancels);
1134 /* concurrent cancels on the same handle can happen */
1135 lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
1137 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1141 rc = ldlm_cli_cancel_local(lock);
1142 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1143 LDLM_LOCK_PUT(lock);
1144 RETURN(rc < 0 ? rc : 0);
1146 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1147 * rpc which goes to canceld portal, so we can cancel other lru locks
1148 * here and send them all as one LDLM_CANCEL rpc. */
1149 LASSERT(list_empty(&lock->l_bl_ast));
1150 list_add(&lock->l_bl_ast, &cancels);
1152 exp = lock->l_conn_export;
1153 if (exp_connect_cancelset(exp)) {
1154 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1159 ns = lock->l_resource->lr_namespace;
1160 flags = ns_connect_lru_resize(ns) ?
1161 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1162 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1163 LDLM_FL_BL_AST, flags);
1165 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1169 /* XXX until we will have compound requests and can cut cancels from generic rpc
1170 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1171 static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
1173 CFS_LIST_HEAD(head);
1174 struct ldlm_lock *lock, *next;
1175 int left = 0, bl_ast = 0, rc;
1178 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1182 if (flags & LDLM_FL_LOCAL_ONLY) {
1183 rc = LDLM_FL_LOCAL_ONLY;
1184 ldlm_lock_cancel(lock);
1186 rc = ldlm_cli_cancel_local(lock);
1188 if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1189 LDLM_DEBUG(lock, "Cancel lock separately");
1190 list_del_init(&lock->l_bl_ast);
1191 list_add(&lock->l_bl_ast, &head);
1195 if (rc == LDLM_FL_LOCAL_ONLY) {
1196 /* CANCEL RPC should not be sent to server. */
1197 list_del_init(&lock->l_bl_ast);
1198 LDLM_LOCK_PUT(lock);
1205 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1212 * Callback function for shrink policy. Makes decision whether to keep
1213 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1214 * \a added and number of locks to be preferably canceled \a count.
1216 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1218 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1220 static ldlm_policy_res_t ldlm_cancel_shrink_policy(struct ldlm_namespace *ns,
1221 struct ldlm_lock *lock,
1222 int unused, int added,
1229 * Stop lru processing when we reached passed @count or checked all
1232 if (count && added >= count)
1233 return LDLM_POLICY_KEEP_LOCK;
1235 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1236 struct ldlm_extent *l_extent;
1239 * For all extent locks cost is 1 + number of pages in
1242 l_extent = &lock->l_policy_data.l_extent;
1243 page_nr = (l_extent->end - l_extent->start);
1244 do_div(page_nr, CFS_PAGE_SIZE);
1248 * XXX: In fact this is evil hack, we can't access inode
1249 * here. For doing it right we need somehow to have number
1250 * of covered by lock. This should be fixed later when 10718
1253 if (lock->l_ast_data != NULL) {
1254 struct inode *inode = lock->l_ast_data;
1255 if (page_nr > inode->i_mapping->nrpages)
1256 page_nr = inode->i_mapping->nrpages;
1259 lock_cost = 1 + page_nr;
1262 * For all locks which are not extent ones cost is 1
1268 * Keep all expensive locks in lru for the memory pressure time
1269 * cancel policy. They anyways may be canceled by lru resize
1270 * pplicy if they have not small enough CLV.
1272 return lock_cost > ns->ns_shrink_thumb ?
1273 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1277 * Callback function for lru-resize policy. Makes decision whether to keep
1278 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1279 * \a added and number of locks to be preferably canceled \a count.
1281 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1283 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1285 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1286 struct ldlm_lock *lock,
1287 int unused, int added,
1290 cfs_time_t cur = cfs_time_current();
1291 struct ldlm_pool *pl = &ns->ns_pool;
1296 * Stop lru processing when we reached passed @count or checked all
1299 if (count && added >= count)
1300 return LDLM_POLICY_KEEP_LOCK;
1302 slv = ldlm_pool_get_slv(pl);
1303 lvf = ldlm_pool_get_lvf(pl);
1304 la = cfs_duration_sec(cfs_time_sub(cur,
1305 lock->l_last_used));
1308 * Stop when slv is not yet come from server or lv is smaller than
1311 lv = lvf * la * unused;
1314 * Inform pool about current CLV to see it via proc.
1316 ldlm_pool_set_clv(pl, lv);
1317 return (slv == 1 || lv < slv) ?
1318 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1322 * Callback function for proc used policy. Makes decision whether to keep
1323 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1324 * \a added and number of locks to be preferably canceled \a count.
1326 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1328 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1330 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1331 struct ldlm_lock *lock,
1332 int unused, int added,
1336 * Stop lru processing when we reached passed @count or checked all
1339 return (added >= count) ?
1340 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1344 * Callback function for aged policy. Makes decision whether to keep
1345 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1346 * \a added and number of locks to be preferably canceled \a count.
1348 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1350 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1352 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1353 struct ldlm_lock *lock,
1354 int unused, int added,
1358 * Stop lru processing if young lock is found and we reached passed
1361 return ((added >= count) &&
1362 cfs_time_before(cfs_time_current(),
1363 cfs_time_add(lock->l_last_used,
1365 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1369 * Callback function for default policy. Makes decision whether to keep
1370 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1371 * \a added and number of locks to be preferably canceled \a count.
1373 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1375 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1377 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1378 struct ldlm_lock *lock,
1379 int unused, int added,
1383 * Stop lru processing when we reached passed @count or checked all
1386 return (added >= count) ?
1387 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1390 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1391 struct ldlm_lock *, int,
1394 static ldlm_cancel_lru_policy_t
1395 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1397 if (ns_connect_lru_resize(ns)) {
1398 if (flags & LDLM_CANCEL_SHRINK)
1399 return ldlm_cancel_shrink_policy;
1400 else if (flags & LDLM_CANCEL_LRUR)
1401 return ldlm_cancel_lrur_policy;
1402 else if (flags & LDLM_CANCEL_PASSED)
1403 return ldlm_cancel_passed_policy;
1405 if (flags & LDLM_CANCEL_AGED)
1406 return ldlm_cancel_aged_policy;
1409 return ldlm_cancel_default_policy;
1412 /* - Free space in lru for @count new locks,
1413 * redundant unused locks are canceled locally;
1414 * - also cancel locally unused aged locks;
1415 * - do not cancel more than @max locks;
1416 * - GET the found locks and add them into the @cancels list.
1418 * A client lock can be added to the l_bl_ast list only when it is
1419 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1420 * There are the following use cases: ldlm_cancel_resource_local(),
1421 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1422 * flag properly. As any attempt to cancel a lock rely on this flag,
1423 * l_bl_ast list is accessed later without any special locking.
1425 * Calling policies for enabled lru resize:
1426 * ----------------------------------------
1427 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1428 * cancel not more than @count locks;
1430 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1431 * the beginning of lru list);
1433 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1434 * memory pressre policy function;
1436 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1438 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1439 int count, int max, int cancel_flags, int flags)
1441 ldlm_cancel_lru_policy_t pf;
1442 struct ldlm_lock *lock, *next;
1443 int added = 0, unused;
1446 spin_lock(&ns->ns_unused_lock);
1447 unused = ns->ns_nr_unused;
1449 if (!ns_connect_lru_resize(ns))
1450 count += unused - ns->ns_max_unused;
1452 pf = ldlm_cancel_lru_policy(ns, flags);
1453 LASSERT(pf != NULL);
1455 while (!list_empty(&ns->ns_unused_list)) {
1456 /* For any flags, stop scanning if @max is reached. */
1457 if (max && added >= max)
1460 list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru){
1461 /* No locks which got blocking requests. */
1462 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1464 /* Somebody is already doing CANCEL. No need in this
1465 * lock in lru, do not traverse it again. */
1466 if (!(lock->l_flags & LDLM_FL_CANCELING))
1469 ldlm_lock_remove_from_lru_nolock(lock);
1471 if (&lock->l_lru == &ns->ns_unused_list)
1474 /* Pass the lock through the policy filter and see if it
1475 * should stay in lru.
1477 * Even for shrinker policy we stop scanning if
1478 * we find a lock that should stay in the cache.
1479 * We should take into account lock age anyway
1480 * as new lock even if it is small of weight is
1481 * valuable resource.
1483 * That is, for shrinker policy we drop only
1484 * old locks, but additionally chose them by
1485 * their weight. Big extent locks will stay in
1487 if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
1490 LDLM_LOCK_GET(lock); /* dropped by bl thread */
1491 spin_unlock(&ns->ns_unused_lock);
1493 lock_res_and_lock(lock);
1494 /* Check flags again under the lock. */
1495 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1496 (ldlm_lock_remove_from_lru(lock) == 0)) {
1497 /* other thread is removing lock from lru or
1498 * somebody is already doing CANCEL or
1499 * there is a blocking request which will send
1500 * cancel by itseft or the lock is matched
1501 * is already not unused. */
1502 unlock_res_and_lock(lock);
1503 LDLM_LOCK_PUT(lock);
1504 spin_lock(&ns->ns_unused_lock);
1507 LASSERT(!lock->l_readers && !lock->l_writers);
1509 /* If we have chosen to cancel this lock voluntarily, we
1510 * better send cancel notification to server, so that it
1511 * frees appropriate state. This might lead to a race
1512 * where while we are doing cancel here, server is also
1513 * silently cancelling this lock. */
1514 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1516 /* Setting the CBPENDING flag is a little misleading,
1517 * but prevents an important race; namely, once
1518 * CBPENDING is set, the lock can accumulate no more
1519 * readers/writers. Since readers and writers are
1520 * already zero here, ldlm_lock_decref() won't see
1521 * this flag and call l_blocking_ast */
1522 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1524 /* We can't re-add to l_lru as it confuses the
1525 * refcounting in ldlm_lock_remove_from_lru() if an AST
1526 * arrives after we drop ns_lock below. We use l_bl_ast
1527 * and can't use l_pending_chain as it is used both on
1528 * server and client nevertheless bug 5666 says it is
1529 * used only on server */
1530 LASSERT(list_empty(&lock->l_bl_ast));
1531 list_add(&lock->l_bl_ast, cancels);
1532 unlock_res_and_lock(lock);
1533 spin_lock(&ns->ns_unused_lock);
1537 spin_unlock(&ns->ns_unused_lock);
1538 RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
1541 /* Returns number of locks which could be canceled next time when
1542 * ldlm_cancel_lru() is called. Used from locks pool shrinker. */
1543 int ldlm_cancel_lru_estimate(struct ldlm_namespace *ns,
1544 int count, int max, int flags)
1546 ldlm_cancel_lru_policy_t pf;
1547 struct ldlm_lock *lock;
1548 int added = 0, unused;
1551 pf = ldlm_cancel_lru_policy(ns, flags);
1552 LASSERT(pf != NULL);
1553 spin_lock(&ns->ns_unused_lock);
1554 unused = ns->ns_nr_unused;
1556 list_for_each_entry(lock, &ns->ns_unused_list, l_lru) {
1557 /* For any flags, stop scanning if @max is reached. */
1558 if (max && added >= max)
1561 /* Somebody is already doing CANCEL or there is a
1562 * blocking request will send cancel. Let's not count
1564 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1565 (lock->l_flags & LDLM_FL_BL_AST))
1568 /* Pass the lock through the policy filter and see if it
1569 * should stay in lru. */
1570 if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
1576 spin_unlock(&ns->ns_unused_lock);
1580 /* when called with LDLM_ASYNC the blocking callback will be handled
1581 * in a thread and this function will return after the thread has been
1582 * asked to call the callback. when called with LDLM_SYNC the blocking
1583 * callback will be performed in this function. */
1584 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1587 CFS_LIST_HEAD(cancels);
1592 sync = LDLM_SYNC; /* force to be sync in user space */
1594 count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
1595 if (sync == LDLM_ASYNC) {
1596 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1601 /* If an error occured in ASYNC mode, or
1602 * this is SYNC mode, cancel the list. */
1603 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1607 /* Find and cancel locally unused locks found on resource, matched to the
1608 * given policy, mode. GET the found locks and add them into the @cancels
1610 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1611 struct list_head *cancels,
1612 ldlm_policy_data_t *policy,
1613 ldlm_mode_t mode, int lock_flags,
1614 int cancel_flags, void *opaque)
1616 struct ldlm_lock *lock;
1621 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1622 if (opaque != NULL && lock->l_ast_data != opaque) {
1623 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1624 lock->l_ast_data, opaque);
1629 if (lock->l_readers || lock->l_writers) {
1630 if (cancel_flags & LDLM_FL_WARN) {
1631 LDLM_ERROR(lock, "lock in use");
1637 /* If somebody is already doing CANCEL, or blocking ast came,
1638 * skip this lock. */
1639 if (lock->l_flags & LDLM_FL_BL_AST ||
1640 lock->l_flags & LDLM_FL_CANCELING)
1643 if (lockmode_compat(lock->l_granted_mode, mode))
1646 /* If policy is given and this is IBITS lock, add to list only
1647 * those locks that match by policy. */
1648 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1649 !(lock->l_policy_data.l_inodebits.bits &
1650 policy->l_inodebits.bits))
1653 /* See CBPENDING comment in ldlm_cancel_lru */
1654 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1657 LASSERT(list_empty(&lock->l_bl_ast));
1658 list_add(&lock->l_bl_ast, cancels);
1659 LDLM_LOCK_GET(lock);
1664 RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
1667 /* If @req is NULL, send CANCEL request to server with handles of locks
1668 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1669 * separately per lock.
1670 * If @req is not NULL, put handles of locks in @cancels into the request
1671 * buffer at the offset @off.
1672 * Destroy @cancels at the end. */
1673 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1674 struct ptlrpc_request *req, int flags)
1676 struct ldlm_lock *lock;
1680 if (list_empty(cancels) || count == 0)
1683 /* XXX: requests (both batched and not) could be sent in parallel.
1684 * Usually it is enough to have just 1 RPC, but it is possible that
1685 * there are to many locks to be cancelled in LRU or on a resource.
1686 * It would also speed up the case when the server does not support
1689 LASSERT(!list_empty(cancels));
1690 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1691 LASSERT(lock->l_conn_export);
1693 if (exp_connect_cancelset(lock->l_conn_export)) {
1696 ldlm_cancel_pack(req, cancels, count);
1698 res = ldlm_cli_cancel_req(lock->l_conn_export,
1702 res = ldlm_cli_cancel_req(lock->l_conn_export,
1707 CERROR("ldlm_cli_cancel_list: %d\n", res);
1712 ldlm_lock_list_put(cancels, l_bl_ast, res);
1714 LASSERT(count == 0);
1718 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1719 const struct ldlm_res_id *res_id,
1720 ldlm_policy_data_t *policy,
1721 ldlm_mode_t mode, int flags, void *opaque)
1723 struct ldlm_resource *res;
1724 CFS_LIST_HEAD(cancels);
1729 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1731 /* This is not a problem. */
1732 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1736 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1738 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1740 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1742 ldlm_resource_putref(res);
1746 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1748 int no_resource = 0;
1750 spin_lock(&ns->ns_hash_lock);
1751 if (ns->ns_resources == 0)
1753 spin_unlock(&ns->ns_hash_lock);
1755 RETURN(no_resource);
1758 /* Cancel all locks on a namespace (or a specific resource, if given)
1759 * that have 0 readers/writers.
1761 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1762 * to notify the server. */
1763 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1764 const struct ldlm_res_id *res_id,
1765 int flags, void *opaque)
1774 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1778 spin_lock(&ns->ns_hash_lock);
1779 for (i = 0; i < RES_HASH_SIZE; i++) {
1780 struct list_head *tmp;
1781 tmp = ns->ns_hash[i].next;
1782 while (tmp != &(ns->ns_hash[i])) {
1783 struct ldlm_resource *res;
1786 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1787 ldlm_resource_getref(res);
1788 spin_unlock(&ns->ns_hash_lock);
1790 rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1795 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1796 res->lr_name.name[0], rc);
1798 spin_lock(&ns->ns_hash_lock);
1800 ldlm_resource_putref_locked(res);
1803 spin_unlock(&ns->ns_hash_lock);
1808 /* join/split resource locks to/from lru list */
1809 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
1810 const struct ldlm_res_id *res_id, int join)
1812 struct ldlm_resource *res;
1813 struct ldlm_lock *lock, *n;
1817 LASSERT(ns_is_client(ns));
1819 res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
1822 LASSERT(res->lr_type == LDLM_EXTENT);
1828 list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1829 if (list_empty(&lock->l_lru) &&
1830 !lock->l_readers && !lock->l_writers &&
1831 !(lock->l_flags & LDLM_FL_LOCAL) &&
1832 !(lock->l_flags & LDLM_FL_CBPENDING) &&
1833 !(lock->l_flags & LDLM_FL_BL_AST)) {
1834 ldlm_lock_add_to_lru(lock);
1835 lock->l_flags &= ~LDLM_FL_NO_LRU;
1836 LDLM_DEBUG(lock, "join lock to lru");
1842 spin_lock(&ns->ns_unused_lock);
1843 list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1844 if (lock->l_resource == res) {
1845 ldlm_lock_remove_from_lru_nolock(lock);
1846 lock->l_flags |= LDLM_FL_NO_LRU;
1847 LDLM_DEBUG(lock, "split lock from lru");
1851 spin_unlock(&ns->ns_unused_lock);
1854 ldlm_resource_putref(res);
1858 /* Lock iterators. */
1860 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1863 struct list_head *tmp, *next;
1864 struct ldlm_lock *lock;
1865 int rc = LDLM_ITER_CONTINUE;
1870 RETURN(LDLM_ITER_CONTINUE);
1873 list_for_each_safe(tmp, next, &res->lr_granted) {
1874 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1876 if (iter(lock, closure) == LDLM_ITER_STOP)
1877 GOTO(out, rc = LDLM_ITER_STOP);
1880 list_for_each_safe(tmp, next, &res->lr_converting) {
1881 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1883 if (iter(lock, closure) == LDLM_ITER_STOP)
1884 GOTO(out, rc = LDLM_ITER_STOP);
1887 list_for_each_safe(tmp, next, &res->lr_waiting) {
1888 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1890 if (iter(lock, closure) == LDLM_ITER_STOP)
1891 GOTO(out, rc = LDLM_ITER_STOP);
1898 struct iter_helper_data {
1899 ldlm_iterator_t iter;
1903 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1905 struct iter_helper_data *helper = closure;
1906 return helper->iter(lock, helper->closure);
1909 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1911 return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1914 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1917 struct iter_helper_data helper = { iter: iter, closure: closure };
1918 return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1921 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1922 ldlm_res_iterator_t iter, void *closure)
1924 int i, rc = LDLM_ITER_CONTINUE;
1925 struct ldlm_resource *res;
1926 struct list_head *tmp;
1929 spin_lock(&ns->ns_hash_lock);
1930 for (i = 0; i < RES_HASH_SIZE; i++) {
1931 tmp = ns->ns_hash[i].next;
1932 while (tmp != &(ns->ns_hash[i])) {
1933 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1934 ldlm_resource_getref(res);
1935 spin_unlock(&ns->ns_hash_lock);
1937 rc = iter(res, closure);
1939 spin_lock(&ns->ns_hash_lock);
1941 ldlm_resource_putref_locked(res);
1942 if (rc == LDLM_ITER_STOP)
1947 spin_unlock(&ns->ns_hash_lock);
1951 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1952 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1953 const struct ldlm_res_id *res_id,
1954 ldlm_iterator_t iter, void *data)
1956 struct ldlm_resource *res;
1960 CERROR("must pass in namespace\n");
1964 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1970 ldlm_resource_foreach(res, iter, data);
1971 ldlm_resource_putref(res);
1977 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1979 struct list_head *list = closure;
1981 /* we use l_pending_chain here, because it's unused on clients. */
1982 LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1983 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1984 /* bug 9573: don't replay locks left after eviction */
1985 if (!(lock->l_flags & LDLM_FL_FAILED))
1986 list_add(&lock->l_pending_chain, list);
1987 return LDLM_ITER_CONTINUE;
1990 static int replay_lock_interpret(struct ptlrpc_request *req,
1991 struct ldlm_async_args *aa, int rc)
1993 struct ldlm_lock *lock;
1994 struct ldlm_reply *reply;
1997 atomic_dec(&req->rq_import->imp_replay_inflight);
2002 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2004 GOTO(out, rc = -EPROTO);
2006 lock = ldlm_handle2lock(&aa->lock_handle);
2008 CERROR("received replay ack for unknown local cookie "LPX64
2009 " remote cookie "LPX64 " from server %s id %s\n",
2010 aa->lock_handle.cookie, reply->lock_handle.cookie,
2011 req->rq_export->exp_client_uuid.uuid,
2012 libcfs_id2str(req->rq_peer));
2013 GOTO(out, rc = -ESTALE);
2016 lock->l_remote_handle = reply->lock_handle;
2017 LDLM_DEBUG(lock, "replayed lock:");
2018 ptlrpc_import_recovery_state_machine(req->rq_import);
2019 LDLM_LOCK_PUT(lock);
2022 ptlrpc_connect_import(req->rq_import, NULL);
2028 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2030 struct ptlrpc_request *req;
2031 struct ldlm_async_args *aa;
2032 struct ldlm_request *body;
2037 /* Bug 11974: Do not replay a lock which is actively being canceled */
2038 if (lock->l_flags & LDLM_FL_CANCELING) {
2039 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2043 /* If this is reply-less callback lock, we cannot replay it, since
2044 * server might have long dropped it, but notification of that event was
2045 * lost by network. (and server granted conflicting lock already) */
2046 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2047 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2048 ldlm_lock_cancel(lock);
2052 * If granted mode matches the requested mode, this lock is granted.
2054 * If they differ, but we have a granted mode, then we were granted
2055 * one mode and now want another: ergo, converting.
2057 * If we haven't been granted anything and are on a resource list,
2058 * then we're blocked/waiting.
2060 * If we haven't been granted anything and we're NOT on a resource list,
2061 * then we haven't got a reply yet and don't have a known disposition.
2062 * This happens whenever a lock enqueue is the request that triggers
2065 if (lock->l_granted_mode == lock->l_req_mode)
2066 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2067 else if (lock->l_granted_mode)
2068 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2069 else if (!list_empty(&lock->l_res_link))
2070 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2072 flags = LDLM_FL_REPLAY;
2074 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2075 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2079 /* We're part of recovery, so don't wait for it. */
2080 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2082 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2083 ldlm_lock2desc(lock, &body->lock_desc);
2084 body->lock_flags = flags;
2086 ldlm_lock2handle(lock, &body->lock_handle[0]);
2087 if (lock->l_lvb_len != 0) {
2088 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2089 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2092 ptlrpc_request_set_replen(req);
2093 /* notify the server we've replayed all requests.
2094 * also, we mark the request to be put on a dedicated
2095 * queue to be processed after all request replayes.
2097 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2099 LDLM_DEBUG(lock, "replaying lock:");
2101 atomic_inc(&req->rq_import->imp_replay_inflight);
2102 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2103 aa = ptlrpc_req_async_args(req);
2104 aa->lock_handle = body->lock_handle[0];
2105 req->rq_interpret_reply = replay_lock_interpret;
2106 ptlrpcd_add_req(req);
2111 int ldlm_replay_locks(struct obd_import *imp)
2113 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2114 CFS_LIST_HEAD(list);
2115 struct ldlm_lock *lock, *next;
2120 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2122 /* ensure this doesn't fall to 0 before all have been queued */
2123 atomic_inc(&imp->imp_replay_inflight);
2125 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2127 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2128 list_del_init(&lock->l_pending_chain);
2130 continue; /* or try to do the rest? */
2131 rc = replay_one_lock(imp, lock);
2134 atomic_dec(&imp->imp_replay_inflight);