1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
6 * This file is part of the Lustre file system, http://www.lustre.org
7 * Lustre is a trademark of Cluster File Systems, Inc.
9 * You may have signed or agreed to another license before downloading
10 * this software. If so, you are bound by the terms and conditions
11 * of that agreement, and the following does not apply to you. See the
12 * LICENSE file included with this distribution for more information.
14 * If you did not agree to a different license, then this copy of Lustre
15 * is open source software; you can redistribute it and/or modify it
16 * under the terms of version 2 of the GNU General Public License as
17 * published by the Free Software Foundation.
19 * In either case, Lustre is distributed in the hope that it will be
20 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
21 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
22 * license text for more details.
25 #define DEBUG_SUBSYSTEM S_LDLM
28 #include <liblustre.h>
31 #include <lustre_dlm.h>
32 #include <obd_class.h>
35 #include "ldlm_internal.h"
37 static void interrupted_completion_wait(void *data)
41 struct lock_wait_data {
42 struct ldlm_lock *lwd_lock;
46 struct ldlm_async_args {
47 struct lustre_handle lock_handle;
50 int ldlm_expired_completion_wait(void *data)
52 struct lock_wait_data *lwd = data;
53 struct ldlm_lock *lock = lwd->lwd_lock;
54 struct obd_import *imp;
55 struct obd_device *obd;
58 if (lock->l_conn_export == NULL) {
59 static cfs_time_t next_dump = 0, last_dump = 0;
61 if (ptlrpc_check_suspend())
64 LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago); "
65 "not entering recovery in server code, just going "
66 "back to sleep", lock->l_enqueued_time.tv_sec,
67 cfs_time_current_sec() - lock->l_enqueued_time.tv_sec);
68 if (cfs_time_after(cfs_time_current(), next_dump)) {
69 last_dump = next_dump;
70 next_dump = cfs_time_shift(300);
71 ldlm_namespace_dump(D_DLMTRACE,
72 lock->l_resource->lr_namespace);
74 libcfs_debug_dumplog();
79 obd = lock->l_conn_export->exp_obd;
80 imp = obd->u.cli.cl_import;
81 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
82 LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago), entering "
83 "recovery for %s@%s", lock->l_enqueued_time.tv_sec,
84 cfs_time_current_sec() - lock->l_enqueued_time.tv_sec,
85 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
90 static int is_granted_or_cancelled(struct ldlm_lock *lock)
94 lock_res_and_lock(lock);
95 if (((lock->l_req_mode == lock->l_granted_mode) &&
96 !(lock->l_flags & LDLM_FL_CP_REQD)) ||
97 (lock->l_flags & LDLM_FL_FAILED))
99 unlock_res_and_lock(lock);
104 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
106 /* XXX ALLOCATE - 160 bytes */
107 struct lock_wait_data lwd;
108 struct obd_device *obd;
109 struct obd_import *imp = NULL;
110 struct l_wait_info lwi;
114 if (flags == LDLM_FL_WAIT_NOREPROC) {
115 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
119 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
120 LDLM_FL_BLOCK_CONV))) {
121 cfs_waitq_signal(&lock->l_waitq);
125 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
127 ldlm_lock_dump(D_OTHER, lock, 0);
128 ldlm_reprocess_all(lock->l_resource);
132 obd = class_exp2obd(lock->l_conn_export);
134 /* if this is a local lock, then there is no import */
136 imp = obd->u.cli.cl_import;
140 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
141 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
142 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
144 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
145 ldlm_expired_completion_wait,
146 interrupted_completion_wait, &lwd);
150 spin_lock(&imp->imp_lock);
151 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
152 spin_unlock(&imp->imp_lock);
155 /* Go to sleep until the lock is granted or cancelled. */
156 rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
158 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
159 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
164 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
169 LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
174 * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
176 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
177 void *data, int flag)
182 if (flag == LDLM_CB_CANCELING) {
183 /* Don't need to do anything here. */
187 lock_res_and_lock(lock);
188 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
189 * that ldlm_blocking_ast is called just before intent_policy method
190 * takes the ns_lock, then by the time we get the lock, we might not
191 * be the correct blocking function anymore. So check, and return
193 if (lock->l_blocking_ast != ldlm_blocking_ast) {
194 unlock_res_and_lock(lock);
198 lock->l_flags |= LDLM_FL_CBPENDING;
199 do_ast = (!lock->l_readers && !lock->l_writers);
200 unlock_res_and_lock(lock);
203 struct lustre_handle lockh;
206 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
207 ldlm_lock2handle(lock, &lockh);
208 rc = ldlm_cli_cancel(&lockh);
210 CERROR("ldlm_cli_cancel: %d\n", rc);
212 LDLM_DEBUG(lock, "Lock still has references, will be "
219 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
220 * comment in filter_intent_policy() on why you may need this.
222 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
225 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
226 * that is rather subtle: with OST-side locking, it may so happen that
227 * _all_ extent locks are held by the OST. If client wants to obtain
228 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
229 * on the server), dummy glimpse callback fires and does
230 * nothing. Client still receives correct file size due to the
231 * following fragment in filter_intent_policy():
233 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
234 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
235 * res->lr_namespace->ns_lvbo->lvbo_update) {
236 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
239 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
240 * returns correct file size to the client.
242 return -ELDLM_NO_LOCK_DATA;
245 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
246 const struct ldlm_res_id *res_id,
247 ldlm_type_t type, ldlm_policy_data_t *policy,
248 ldlm_mode_t mode, int *flags,
249 ldlm_blocking_callback blocking,
250 ldlm_completion_callback completion,
251 ldlm_glimpse_callback glimpse,
252 void *data, __u32 lvb_len, void *lvb_swabber,
253 struct lustre_handle *lockh)
255 struct ldlm_lock *lock;
259 LASSERT(!(*flags & LDLM_FL_REPLAY));
260 if (unlikely(ns_is_client(ns))) {
261 CERROR("Trying to enqueue local lock in a shadow namespace\n");
265 lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
266 completion, glimpse, data, lvb_len);
268 GOTO(out_nolock, err = -ENOMEM);
269 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
271 ldlm_lock_addref_internal(lock, mode);
272 ldlm_lock2handle(lock, lockh);
273 lock_res_and_lock(lock);
274 lock->l_flags |= LDLM_FL_LOCAL;
275 if (*flags & LDLM_FL_ATOMIC_CB)
276 lock->l_flags |= LDLM_FL_ATOMIC_CB;
277 lock->l_lvb_swabber = lvb_swabber;
278 unlock_res_and_lock(lock);
280 lock->l_policy_data = *policy;
281 if (type == LDLM_EXTENT)
282 lock->l_req_extent = policy->l_extent;
284 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
285 if (unlikely(err != ELDLM_OK))
289 *policy = lock->l_policy_data;
291 LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
294 if (lock->l_completion_ast)
295 lock->l_completion_ast(lock, *flags, NULL);
297 LDLM_DEBUG(lock, "client-side local enqueue END");
305 static void failed_lock_cleanup(struct ldlm_namespace *ns,
306 struct ldlm_lock *lock,
307 struct lustre_handle *lockh, int mode)
309 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
310 lock_res_and_lock(lock);
311 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
312 unlock_res_and_lock(lock);
313 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
315 ldlm_lock_decref_and_cancel(lockh, mode);
317 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
318 * from llite/file.c/ll_file_flock(). */
319 if (lock->l_resource->lr_type == LDLM_FLOCK) {
320 ldlm_lock_destroy(lock);
324 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
325 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
326 int *flags, void *lvb, __u32 lvb_len,
327 void *lvb_swabber, struct lustre_handle *lockh,int rc)
329 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
330 int is_replay = *flags & LDLM_FL_REPLAY;
331 struct ldlm_lock *lock;
332 struct ldlm_reply *reply;
333 int cleanup_phase = 1;
336 lock = ldlm_handle2lock(lockh);
337 /* ldlm_cli_enqueue is holding a reference on this lock. */
339 LASSERT(type == LDLM_FLOCK);
343 if (rc != ELDLM_OK) {
345 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
346 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
347 if (rc == ELDLM_LOCK_ABORTED) {
348 /* Before we return, swab the reply */
349 reply = req_capsule_server_get(&req->rq_pill,
354 struct ost_lvb *tmplvb;
356 req_capsule_set_size(&req->rq_pill,
357 &RMF_DLM_LVB, RCL_SERVER,
359 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
363 GOTO(cleanup, rc = -EPROTO);
365 memcpy(lvb, tmplvb, lvb_len);
371 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
373 GOTO(cleanup, rc = -EPROTO);
375 /* lock enqueued on the server */
378 lock_res_and_lock(lock);
379 lock->l_remote_handle = reply->lock_handle;
380 *flags = reply->lock_flags;
381 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
382 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
383 * to wait with no timeout as well */
384 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
385 unlock_res_and_lock(lock);
387 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
388 lock, reply->lock_handle.cookie, *flags);
390 /* If enqueue returned a blocked lock but the completion handler has
391 * already run, then it fixed up the resource and we don't need to do it
393 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
394 int newmode = reply->lock_desc.l_req_mode;
396 if (newmode && newmode != lock->l_req_mode) {
397 LDLM_DEBUG(lock, "server returned different mode %s",
398 ldlm_lockname[newmode]);
399 lock->l_req_mode = newmode;
402 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
403 lock->l_resource->lr_name.name,
404 sizeof(struct ldlm_res_id))) {
405 CDEBUG(D_INFO, "remote intent success, locking "
406 "(%ld,%ld,%ld) instead of "
408 (long)reply->lock_desc.l_resource.lr_name.name[0],
409 (long)reply->lock_desc.l_resource.lr_name.name[1],
410 (long)reply->lock_desc.l_resource.lr_name.name[2],
411 (long)lock->l_resource->lr_name.name[0],
412 (long)lock->l_resource->lr_name.name[1],
413 (long)lock->l_resource->lr_name.name[2]);
415 rc = ldlm_lock_change_resource(ns, lock,
416 &reply->lock_desc.l_resource.lr_name);
417 if (rc || lock->l_resource == NULL)
418 GOTO(cleanup, rc = -ENOMEM);
419 LDLM_DEBUG(lock, "client-side enqueue, new resource");
422 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
424 lock->l_policy_data =
425 reply->lock_desc.l_policy_data;
426 if (type != LDLM_PLAIN)
427 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
430 if ((*flags) & LDLM_FL_AST_SENT ||
431 /* Cancel extent locks as soon as possible on a liblustre client,
432 * because it cannot handle asynchronous ASTs robustly (see
434 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
435 lock_res_and_lock(lock);
436 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
437 unlock_res_and_lock(lock);
438 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
441 /* If the lock has already been granted by a completion AST, don't
442 * clobber the LVB with an older one. */
443 if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
446 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
448 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
452 GOTO(cleanup, rc = -EPROTO);
453 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
457 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
458 if (lock->l_completion_ast != NULL) {
459 int err = lock->l_completion_ast(lock, *flags, NULL);
462 if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
467 if (lvb_len && lvb != NULL) {
468 /* Copy the LVB here, and not earlier, because the completion
469 * AST (if any) can override what we got in the reply */
470 memcpy(lvb, lock->l_lvb_data, lvb_len);
473 LDLM_DEBUG(lock, "client-side enqueue END");
476 if (cleanup_phase == 1 && rc)
477 failed_lock_cleanup(ns, lock, lockh, mode);
478 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
484 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
485 * a single page on the send/receive side. XXX: 512 should be changed
486 * to more adequate value. */
487 static inline int ldlm_req_handles_avail(int req_size, int off)
491 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512) - req_size;
492 avail /= sizeof(struct lustre_handle);
493 avail += LDLM_LOCKREQ_HANDLES - off;
498 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
499 enum req_location loc,
502 int size = req_capsule_msg_size(pill, loc);
503 return ldlm_req_handles_avail(size, off);
506 static inline int ldlm_format_handles_avail(struct obd_import *imp,
507 const struct req_format *fmt,
508 enum req_location loc, int off)
510 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
511 return ldlm_req_handles_avail(size, off);
514 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
515 * @count locks in @cancels. */
516 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
517 int version, int opc, int canceloff,
518 struct list_head *cancels, int count)
520 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
521 struct req_capsule *pill = &req->rq_pill;
522 struct ldlm_request *dlm = NULL;
523 int flags, avail, to_free, bufcount, pack = 0;
530 if (exp_connect_cancelset(exp)) {
531 /* Estimate the amount of available space in the request. */
532 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
533 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
535 flags = ns_connect_lru_resize(ns) ?
536 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
537 to_free = !ns_connect_lru_resize(ns) &&
538 opc == LDLM_ENQUEUE ? 1 : 0;
540 /* Cancel lru locks here _only_ if the server supports
541 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
542 * rpc, what will make us slower. */
544 count += ldlm_cancel_lru_local(ns, cancels, to_free,
545 avail - count, 0, flags);
550 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
551 ldlm_request_bufsize(pack, opc));
554 rc = ptlrpc_request_pack(req, version, opc);
556 ldlm_lock_list_put(cancels, l_bl_ast, count);
560 if (exp_connect_cancelset(exp)) {
562 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
564 /* Skip first lock handler in ldlm_request_pack(),
565 * this method will incrment @lock_count according
566 * to the lock handle amount actually written to
568 dlm->lock_count = canceloff;
570 /* Pack into the request @pack lock handles. */
571 ldlm_cli_cancel_list(cancels, pack, req, 0);
572 /* Prepare and send separate cancel rpc for others. */
573 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
575 ldlm_lock_list_put(cancels, l_bl_ast, count);
580 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
581 struct list_head *cancels, int count)
583 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
584 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
587 /* If a request has some specific initialisation it is passed in @reqp,
588 * otherwise it is created in ldlm_cli_enqueue.
590 * Supports sync and async requests, pass @async flag accordingly. If a
591 * request was created in ldlm_cli_enqueue and it is the async request,
592 * pass it to the caller in @reqp. */
593 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
594 struct ldlm_enqueue_info *einfo,
595 const struct ldlm_res_id *res_id,
596 ldlm_policy_data_t *policy, int *flags,
597 void *lvb, __u32 lvb_len, void *lvb_swabber,
598 struct lustre_handle *lockh, int async)
600 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
601 struct ldlm_lock *lock;
602 struct ldlm_request *body;
603 int is_replay = *flags & LDLM_FL_REPLAY;
604 int req_passed_in = 1;
606 struct ptlrpc_request *req;
609 LASSERT(exp != NULL);
611 /* If we're replaying this lock, just check some invariants.
612 * If we're creating a new lock, get everything all setup nice. */
614 lock = ldlm_handle2lock(lockh);
615 LASSERT(lock != NULL);
616 LDLM_DEBUG(lock, "client-side enqueue START");
617 LASSERT(exp == lock->l_conn_export);
619 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
620 einfo->ei_mode, einfo->ei_cb_bl,
621 einfo->ei_cb_cp, einfo->ei_cb_gl,
622 einfo->ei_cbdata, lvb_len);
625 /* for the local lock, add the reference */
626 ldlm_lock_addref_internal(lock, einfo->ei_mode);
627 ldlm_lock2handle(lock, lockh);
628 lock->l_lvb_swabber = lvb_swabber;
629 if (policy != NULL) {
630 /* INODEBITS_INTEROP: If the server does not support
631 * inodebits, we will request a plain lock in the
632 * descriptor (ldlm_lock2desc() below) but use an
633 * inodebits lock internally with both bits set.
635 if (einfo->ei_type == LDLM_IBITS &&
636 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
637 lock->l_policy_data.l_inodebits.bits =
638 MDS_INODELOCK_LOOKUP |
639 MDS_INODELOCK_UPDATE;
641 lock->l_policy_data = *policy;
644 if (einfo->ei_type == LDLM_EXTENT)
645 lock->l_req_extent = policy->l_extent;
646 LDLM_DEBUG(lock, "client-side enqueue START");
649 /* lock not sent to server yet */
651 if (reqp == NULL || *reqp == NULL) {
652 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
657 failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
668 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
670 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
671 DLM_LOCKREQ_OFF, len, sizeof(*body));
674 lock->l_conn_export = exp;
675 lock->l_export = NULL;
676 lock->l_blocking_ast = einfo->ei_cb_bl;
678 /* Dump lock data into the request buffer */
679 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
680 ldlm_lock2desc(lock, &body->lock_desc);
681 body->lock_flags = *flags;
682 body->lock_handle[0] = *lockh;
684 /* Continue as normal. */
685 if (!req_passed_in) {
687 req_capsule_extend(&req->rq_pill,
688 &RQF_LDLM_ENQUEUE_LVB);
689 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
690 RCL_SERVER, lvb_len);
692 ptlrpc_request_set_replen(req);
696 * Liblustre client doesn't get extent locks, except for O_APPEND case
697 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
698 * [i_size, OBD_OBJECT_EOF] lock is taken.
700 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
701 policy->l_extent.end == OBD_OBJECT_EOF));
704 LASSERT(reqp != NULL);
708 LDLM_DEBUG(lock, "sending request");
709 rc = ptlrpc_queue_wait(req);
710 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
711 einfo->ei_mode, flags, lvb, lvb_len,
712 lvb_swabber, lockh, rc);
714 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
715 * one reference that we took */
721 if (!req_passed_in && req != NULL) {
722 ptlrpc_req_finished(req);
730 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
733 struct ldlm_resource *res;
736 if (ns_is_client(lock->l_resource->lr_namespace)) {
737 CERROR("Trying to cancel local lock\n");
740 LDLM_DEBUG(lock, "client-side local convert");
742 res = ldlm_lock_convert(lock, new_mode, flags);
744 ldlm_reprocess_all(res);
749 LDLM_DEBUG(lock, "client-side local convert handler END");
754 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
755 * conversion of locks which are on the waiting or converting queue */
756 /* Caller of this code is supposed to take care of lock readers/writers
758 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
760 struct ldlm_request *body;
761 struct ldlm_reply *reply;
762 struct ldlm_lock *lock;
763 struct ldlm_resource *res;
764 struct ptlrpc_request *req;
768 lock = ldlm_handle2lock(lockh);
775 if (lock->l_conn_export == NULL)
776 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
778 LDLM_DEBUG(lock, "client-side convert");
780 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
781 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
788 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
789 body->lock_handle[0] = lock->l_remote_handle;
791 body->lock_desc.l_req_mode = new_mode;
792 body->lock_flags = *flags;
795 ptlrpc_request_set_replen(req);
796 rc = ptlrpc_queue_wait(req);
800 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
802 GOTO(out, rc = -EPROTO);
805 GOTO(out, rc = req->rq_status);
807 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
809 ldlm_reprocess_all(res);
810 /* Go to sleep until the lock is granted. */
811 /* FIXME: or cancelled. */
812 if (lock->l_completion_ast) {
813 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
824 ptlrpc_req_finished(req);
828 /* Cancel locks locally.
830 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
831 * LDLM_FL_CANCELING otherwise;
832 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
833 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
835 int rc = LDLM_FL_LOCAL_ONLY;
838 if (lock->l_conn_export) {
841 LDLM_DEBUG(lock, "client-side cancel");
842 /* Set this flag to prevent others from getting new references*/
843 lock_res_and_lock(lock);
844 lock->l_flags |= LDLM_FL_CBPENDING;
845 local_only = (lock->l_flags &
846 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
847 ldlm_cancel_callback(lock);
848 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
849 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
850 unlock_res_and_lock(lock);
853 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
855 rc = LDLM_FL_LOCAL_ONLY;
857 ldlm_lock_cancel(lock);
859 if (ns_is_client(lock->l_resource->lr_namespace)) {
860 LDLM_ERROR(lock, "Trying to cancel local lock");
863 LDLM_DEBUG(lock, "server-side local cancel");
864 ldlm_lock_cancel(lock);
865 ldlm_reprocess_all(lock->l_resource);
866 LDLM_DEBUG(lock, "server-side local cancel handler END");
872 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
873 of the request @req. */
874 static void ldlm_cancel_pack(struct ptlrpc_request *req,
875 struct list_head *head, int count)
877 struct ldlm_request *dlm;
878 struct ldlm_lock *lock;
882 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
883 LASSERT(dlm != NULL);
885 /* Check the room in the request buffer. */
886 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
887 sizeof(struct ldlm_request);
888 max /= sizeof(struct lustre_handle);
889 max += LDLM_LOCKREQ_HANDLES;
890 LASSERT(max >= dlm->lock_count + count);
892 /* XXX: it would be better to pack lock handles grouped by resource.
893 * so that the server cancel would call filter_lvbo_update() less
895 list_for_each_entry(lock, head, l_bl_ast) {
898 LASSERT(lock->l_conn_export);
899 /* Pack the lock handle to the given request buffer. */
900 LDLM_DEBUG(lock, "packing");
901 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
904 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
908 /* Prepare and send a batched cancel rpc, it will include count lock handles
909 * of locks given in @head. */
910 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
911 int count, int flags)
913 struct ptlrpc_request *req = NULL;
914 struct obd_import *imp;
919 LASSERT(exp != NULL);
922 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
925 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
926 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
933 imp = class_exp2cliimp(exp);
934 if (imp == NULL || imp->imp_invalid) {
936 "skipping cancel on invalid import %p\n", imp);
940 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
942 GOTO(out, rc = -ENOMEM);
944 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
945 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
946 ldlm_request_bufsize(count, LDLM_CANCEL));
948 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
950 ptlrpc_request_free(req);
953 req->rq_no_resend = 1;
954 req->rq_no_delay = 1;
956 /* XXX FIXME bug 249 */
957 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
958 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
960 ldlm_cancel_pack(req, cancels, count);
962 ptlrpc_request_set_replen(req);
963 if (flags & LDLM_FL_ASYNC) {
964 ptlrpcd_add_req(req);
968 rc = ptlrpc_queue_wait(req);
971 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
972 "out of sync -- not fatal\n",
973 libcfs_nid2str(req->rq_import->
974 imp_connection->c_peer.nid));
976 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
977 req->rq_import_generation == imp->imp_generation) {
978 ptlrpc_req_finished(req);
980 } else if (rc != ELDLM_OK) {
981 CERROR("Got rc %d from cancel RPC: canceling "
989 ptlrpc_req_finished(req);
992 return sent ? sent : rc;
995 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
997 LASSERT(imp != NULL);
998 return &imp->imp_obd->obd_namespace->ns_pool;
1001 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1003 __u64 old_slv, new_slv;
1004 struct ldlm_pool *pl;
1008 if (!imp_connect_lru_resize(req->rq_import))
1011 /* In some cases RPC may contain slv and limit zeroed out. This is
1012 * the case when server does not support lru resize feature. This is
1013 * also possible in some recovery cases when server side reqs have no
1014 * ref to obd export and thus access to server side namespace is no
1016 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1017 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1018 DEBUG_REQ(D_HA, req, "zero SLV or Limit found "
1019 "(SLV: "LPU64", Limit: %u)",
1020 lustre_msg_get_slv(req->rq_repmsg),
1021 lustre_msg_get_limit(req->rq_repmsg));
1025 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1026 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1027 pl = ldlm_imp2pl(req->rq_import);
1029 spin_lock(&pl->pl_lock);
1030 old_slv = ldlm_pool_get_slv(pl);
1031 ldlm_pool_set_slv(pl, new_slv);
1032 ldlm_pool_set_limit(pl, new_limit);
1034 /* Check if we need to wakeup pools thread for fast SLV change.
1035 * This is only done when threads period is noticably long like
1037 #if defined(__KERNEL__) && (LDLM_POOLS_THREAD_PERIOD >= 10)
1039 __u64 fast_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE;
1040 do_div(fast_change, 100);
1042 /* Wake up pools thread only if SLV has changed more than
1043 * 50% since last update. In this case we want to react asap.
1044 * Otherwise it is no sense to wake up pools as they are
1045 * re-calculated every LDLM_POOLS_THREAD_PERIOD anyways. */
1046 if (old_slv > new_slv && old_slv - new_slv > fast_change)
1047 ldlm_pools_wakeup();
1050 spin_unlock(&pl->pl_lock);
1053 EXPORT_SYMBOL(ldlm_cli_update_pool);
1055 int ldlm_cli_cancel(struct lustre_handle *lockh)
1057 struct obd_export *exp;
1058 int avail, flags, count = 1, rc = 0;
1059 struct ldlm_namespace *ns;
1060 struct ldlm_lock *lock;
1061 CFS_LIST_HEAD(cancels);
1064 /* concurrent cancels on the same handle can happen */
1065 lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
1067 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1071 rc = ldlm_cli_cancel_local(lock);
1072 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1073 LDLM_LOCK_PUT(lock);
1074 RETURN(rc < 0 ? rc : 0);
1076 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1077 * rpc which goes to canceld portal, so we can cancel other lru locks
1078 * here and send them all as one LDLM_CANCEL rpc. */
1079 LASSERT(list_empty(&lock->l_bl_ast));
1080 list_add(&lock->l_bl_ast, &cancels);
1082 exp = lock->l_conn_export;
1083 if (exp_connect_cancelset(exp)) {
1084 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1089 ns = lock->l_resource->lr_namespace;
1090 flags = ns_connect_lru_resize(ns) ?
1091 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1092 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1093 LDLM_FL_BL_AST, flags);
1095 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1099 /* XXX until we will have compound requests and can cut cancels from generic rpc
1100 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1101 static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
1103 CFS_LIST_HEAD(head);
1104 struct ldlm_lock *lock, *next;
1105 int left = 0, bl_ast = 0, rc;
1108 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1112 if (flags & LDLM_FL_LOCAL_ONLY) {
1113 rc = LDLM_FL_LOCAL_ONLY;
1114 ldlm_lock_cancel(lock);
1116 rc = ldlm_cli_cancel_local(lock);
1118 if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1119 LDLM_DEBUG(lock, "Cancel lock separately");
1120 list_del_init(&lock->l_bl_ast);
1121 list_add(&lock->l_bl_ast, &head);
1125 if (rc == LDLM_FL_LOCAL_ONLY) {
1126 /* CANCEL RPC should not be sent to server. */
1127 list_del_init(&lock->l_bl_ast);
1128 LDLM_LOCK_PUT(lock);
1135 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1141 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1143 static ldlm_policy_res_t ldlm_cancel_shrink_policy(struct ldlm_namespace *ns,
1144 struct ldlm_lock *lock,
1145 int unused, int added,
1151 /* Stop lru processing when we reached passed @count or checked all
1153 if (count && added >= count)
1154 return LDLM_POLICY_KEEP_LOCK;
1156 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1157 struct ldlm_extent *l_extent;
1159 /* For all extent locks cost is 1 + number of pages in
1161 l_extent = &lock->l_policy_data.l_extent;
1162 page_nr = (l_extent->end - l_extent->start);
1163 do_div(page_nr, CFS_PAGE_SIZE);
1166 /* XXX: In fact this is evil hack, we can't access inode
1167 * here. For doing it right we need somehow to have number
1168 * of covered by lock. This should be fixed later when 10718
1170 if (lock->l_ast_data != NULL) {
1171 struct inode *inode = lock->l_ast_data;
1172 if (page_nr > inode->i_mapping->nrpages)
1173 page_nr = inode->i_mapping->nrpages;
1176 lock_cost = 1 + page_nr;
1178 /* For all locks which are not extent ones cost is 1 */
1182 /* Keep all expensive locks in lru for the memory pressure time
1183 * cancel policy. They anyways may be canceled by lru resize
1184 * pplicy if they have not small enough CLV. */
1185 return lock_cost > ns->ns_shrink_thumb ?
1186 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1189 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1191 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1192 struct ldlm_lock *lock,
1193 int unused, int added,
1196 cfs_time_t cur = cfs_time_current();
1197 struct ldlm_pool *pl = &ns->ns_pool;
1201 /* Stop lru processing when we reached passed @count or checked all
1203 if (count && added >= count)
1204 return LDLM_POLICY_KEEP_LOCK;
1206 spin_lock(&pl->pl_lock);
1207 slv = ldlm_pool_get_slv(pl);
1208 lvf = atomic_read(&pl->pl_lock_volume_factor);
1209 spin_unlock(&pl->pl_lock);
1211 la = cfs_duration_sec(cfs_time_sub(cur,
1212 lock->l_last_used));
1214 /* Stop when slv is not yet come from server or
1215 * lv is smaller than it is. */
1216 lv = lvf * la * unused;
1217 return (slv == 1 || lv < slv) ?
1218 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1221 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1223 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1224 struct ldlm_lock *lock,
1225 int unused, int added,
1228 /* Stop lru processing when we reached passed @count or checked all
1230 return (added >= count) ?
1231 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1234 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1236 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1237 struct ldlm_lock *lock,
1238 int unused, int added,
1241 /* Stop lru processing if young lock is found and we reached passed
1243 return ((added >= count) &&
1244 cfs_time_before(cfs_time_current(),
1245 cfs_time_add(lock->l_last_used,
1247 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1250 /* Return 1 to stop lru processing and keep current lock cached. Return zero
1252 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1253 struct ldlm_lock *lock,
1254 int unused, int added,
1257 /* Stop lru processing when we reached passed @count or checked all
1259 return (added >= count) ?
1260 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1263 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1264 struct ldlm_lock *, int,
1267 static ldlm_cancel_lru_policy_t
1268 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1270 if (ns_connect_lru_resize(ns)) {
1271 if (flags & LDLM_CANCEL_SHRINK)
1272 return ldlm_cancel_shrink_policy;
1273 else if (flags & LDLM_CANCEL_LRUR)
1274 return ldlm_cancel_lrur_policy;
1275 else if (flags & LDLM_CANCEL_PASSED)
1276 return ldlm_cancel_passed_policy;
1278 if (flags & LDLM_CANCEL_AGED)
1279 return ldlm_cancel_aged_policy;
1282 return ldlm_cancel_default_policy;
1285 /* - Free space in lru for @count new locks,
1286 * redundant unused locks are canceled locally;
1287 * - also cancel locally unused aged locks;
1288 * - do not cancel more than @max locks;
1289 * - GET the found locks and add them into the @cancels list.
1291 * A client lock can be added to the l_bl_ast list only when it is
1292 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1293 * There are the following use cases: ldlm_cancel_resource_local(),
1294 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1295 * flag properly. As any attempt to cancel a lock rely on this flag,
1296 * l_bl_ast list is accessed later without any special locking.
1298 * Calling policies for enabled lru resize:
1299 * ----------------------------------------
1300 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1301 * cancel not more than @count locks;
1303 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1304 * the beginning of lru list);
1306 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1307 * memory pressre policy function;
1309 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1311 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1312 int count, int max, int cancel_flags, int flags)
1314 ldlm_cancel_lru_policy_t pf;
1315 struct ldlm_lock *lock, *next;
1316 int added = 0, unused;
1319 spin_lock(&ns->ns_unused_lock);
1320 unused = ns->ns_nr_unused;
1322 if (!ns_connect_lru_resize(ns))
1323 count += unused - ns->ns_max_unused;
1325 pf = ldlm_cancel_lru_policy(ns, flags);
1326 LASSERT(pf != NULL);
1328 while (!list_empty(&ns->ns_unused_list)) {
1329 /* For any flags, stop scanning if @max is reached. */
1330 if (max && added >= max)
1333 list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru){
1334 /* No locks which got blocking requests. */
1335 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1337 /* Somebody is already doing CANCEL. No need in this
1338 * lock in lru, do not traverse it again. */
1339 if (!(lock->l_flags & LDLM_FL_CANCELING))
1342 ldlm_lock_remove_from_lru_nolock(lock);
1344 if (&lock->l_lru == &ns->ns_unused_list)
1347 /* Pass the lock through the policy filter and see if it
1348 * should stay in lru.
1350 * Even for shrinker policy we stop scanning if
1351 * we find a lock that should stay in the cache.
1352 * We should take into account lock age anyway
1353 * as new lock even if it is small of weight is
1354 * valuable resource.
1356 * That is, for shrinker policy we drop only
1357 * old locks, but additionally chose them by
1358 * their weight. Big extent locks will stay in
1360 if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
1363 LDLM_LOCK_GET(lock); /* dropped by bl thread */
1364 spin_unlock(&ns->ns_unused_lock);
1366 lock_res_and_lock(lock);
1367 /* Check flags again under the lock. */
1368 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1369 (ldlm_lock_remove_from_lru(lock) == 0)) {
1370 /* other thread is removing lock from lru or
1371 * somebody is already doing CANCEL or
1372 * there is a blocking request which will send
1373 * cancel by itseft or the lock is matched
1374 * is already not unused. */
1375 unlock_res_and_lock(lock);
1376 LDLM_LOCK_PUT(lock);
1377 spin_lock(&ns->ns_unused_lock);
1380 LASSERT(!lock->l_readers && !lock->l_writers);
1382 /* If we have chosen to cancel this lock voluntarily, we
1383 * better send cancel notification to server, so that it
1384 * frees appropriate state. This might lead to a race
1385 * where while we are doing cancel here, server is also
1386 * silently cancelling this lock. */
1387 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1389 /* Setting the CBPENDING flag is a little misleading,
1390 * but prevents an important race; namely, once
1391 * CBPENDING is set, the lock can accumulate no more
1392 * readers/writers. Since readers and writers are
1393 * already zero here, ldlm_lock_decref() won't see
1394 * this flag and call l_blocking_ast */
1395 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1397 /* We can't re-add to l_lru as it confuses the
1398 * refcounting in ldlm_lock_remove_from_lru() if an AST
1399 * arrives after we drop ns_lock below. We use l_bl_ast
1400 * and can't use l_pending_chain as it is used both on
1401 * server and client nevertheless bug 5666 says it is
1402 * used only on server */
1403 LASSERT(list_empty(&lock->l_bl_ast));
1404 list_add(&lock->l_bl_ast, cancels);
1405 unlock_res_and_lock(lock);
1406 spin_lock(&ns->ns_unused_lock);
1410 spin_unlock(&ns->ns_unused_lock);
1411 RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
1414 /* Returns number of locks which could be canceled next time when
1415 * ldlm_cancel_lru() is called. Used from locks pool shrinker. */
1416 int ldlm_cancel_lru_estimate(struct ldlm_namespace *ns,
1417 int count, int max, int flags)
1419 ldlm_cancel_lru_policy_t pf;
1420 struct ldlm_lock *lock;
1421 int added = 0, unused;
1424 pf = ldlm_cancel_lru_policy(ns, flags);
1425 LASSERT(pf != NULL);
1426 spin_lock(&ns->ns_unused_lock);
1427 unused = ns->ns_nr_unused;
1429 list_for_each_entry(lock, &ns->ns_unused_list, l_lru) {
1430 /* For any flags, stop scanning if @max is reached. */
1431 if (max && added >= max)
1434 /* Somebody is already doing CANCEL or there is a
1435 * blocking request will send cancel. Let's not count
1437 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1438 (lock->l_flags & LDLM_FL_BL_AST))
1441 /* Pass the lock through the policy filter and see if it
1442 * should stay in lru. */
1443 if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
1449 spin_unlock(&ns->ns_unused_lock);
1453 /* when called with LDLM_ASYNC the blocking callback will be handled
1454 * in a thread and this function will return after the thread has been
1455 * asked to call the callback. when called with LDLM_SYNC the blocking
1456 * callback will be performed in this function. */
1457 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1460 CFS_LIST_HEAD(cancels);
1465 sync = LDLM_SYNC; /* force to be sync in user space */
1467 count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
1468 if (sync == LDLM_ASYNC) {
1469 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1474 /* If an error occured in ASYNC mode, or
1475 * this is SYNC mode, cancel the list. */
1476 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1480 /* Find and cancel locally unused locks found on resource, matched to the
1481 * given policy, mode. GET the found locks and add them into the @cancels
1483 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1484 struct list_head *cancels,
1485 ldlm_policy_data_t *policy,
1486 ldlm_mode_t mode, int lock_flags,
1487 int cancel_flags, void *opaque)
1489 struct ldlm_lock *lock;
1494 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1495 if (opaque != NULL && lock->l_ast_data != opaque) {
1496 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1497 lock->l_ast_data, opaque);
1502 if (lock->l_readers || lock->l_writers) {
1503 if (cancel_flags & LDLM_FL_WARN) {
1504 LDLM_ERROR(lock, "lock in use");
1510 /* If somebody is already doing CANCEL, or blocking ast came,
1511 * skip this lock. */
1512 if (lock->l_flags & LDLM_FL_BL_AST ||
1513 lock->l_flags & LDLM_FL_CANCELING)
1516 if (lockmode_compat(lock->l_granted_mode, mode))
1519 /* If policy is given and this is IBITS lock, add to list only
1520 * those locks that match by policy. */
1521 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1522 !(lock->l_policy_data.l_inodebits.bits &
1523 policy->l_inodebits.bits))
1526 /* See CBPENDING comment in ldlm_cancel_lru */
1527 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1530 LASSERT(list_empty(&lock->l_bl_ast));
1531 list_add(&lock->l_bl_ast, cancels);
1532 LDLM_LOCK_GET(lock);
1537 RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
1540 /* If @req is NULL, send CANCEL request to server with handles of locks
1541 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1542 * separately per lock.
1543 * If @req is not NULL, put handles of locks in @cancels into the request
1544 * buffer at the offset @off.
1545 * Destroy @cancels at the end. */
1546 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1547 struct ptlrpc_request *req, int flags)
1549 struct ldlm_lock *lock;
1553 if (list_empty(cancels) || count == 0)
1556 /* XXX: requests (both batched and not) could be sent in parallel.
1557 * Usually it is enough to have just 1 RPC, but it is possible that
1558 * there are to many locks to be cancelled in LRU or on a resource.
1559 * It would also speed up the case when the server does not support
1562 LASSERT(!list_empty(cancels));
1563 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1564 LASSERT(lock->l_conn_export);
1566 if (exp_connect_cancelset(lock->l_conn_export)) {
1569 ldlm_cancel_pack(req, cancels, count);
1571 res = ldlm_cli_cancel_req(lock->l_conn_export,
1575 res = ldlm_cli_cancel_req(lock->l_conn_export,
1580 CERROR("ldlm_cli_cancel_list: %d\n", res);
1585 ldlm_lock_list_put(cancels, l_bl_ast, res);
1587 LASSERT(count == 0);
1591 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1592 const struct ldlm_res_id *res_id,
1593 ldlm_policy_data_t *policy,
1594 ldlm_mode_t mode, int flags, void *opaque)
1596 struct ldlm_resource *res;
1597 CFS_LIST_HEAD(cancels);
1602 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1604 /* This is not a problem. */
1605 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1609 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1611 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1613 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1615 ldlm_resource_putref(res);
1619 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1621 int no_resource = 0;
1623 spin_lock(&ns->ns_hash_lock);
1624 if (ns->ns_resources == 0)
1626 spin_unlock(&ns->ns_hash_lock);
1628 RETURN(no_resource);
1631 /* Cancel all locks on a namespace (or a specific resource, if given)
1632 * that have 0 readers/writers.
1634 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1635 * to notify the server. */
1636 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1637 const struct ldlm_res_id *res_id,
1638 int flags, void *opaque)
1647 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1651 spin_lock(&ns->ns_hash_lock);
1652 for (i = 0; i < RES_HASH_SIZE; i++) {
1653 struct list_head *tmp;
1654 tmp = ns->ns_hash[i].next;
1655 while (tmp != &(ns->ns_hash[i])) {
1656 struct ldlm_resource *res;
1659 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1660 ldlm_resource_getref(res);
1661 spin_unlock(&ns->ns_hash_lock);
1663 rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1668 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1669 res->lr_name.name[0], rc);
1671 spin_lock(&ns->ns_hash_lock);
1673 ldlm_resource_putref_locked(res);
1676 spin_unlock(&ns->ns_hash_lock);
1681 /* join/split resource locks to/from lru list */
1682 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
1683 const struct ldlm_res_id *res_id, int join)
1685 struct ldlm_resource *res;
1686 struct ldlm_lock *lock, *n;
1690 LASSERT(ns_is_client(ns));
1692 res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
1695 LASSERT(res->lr_type == LDLM_EXTENT);
1701 list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1702 if (list_empty(&lock->l_lru) &&
1703 !lock->l_readers && !lock->l_writers &&
1704 !(lock->l_flags & LDLM_FL_LOCAL) &&
1705 !(lock->l_flags & LDLM_FL_CBPENDING) &&
1706 !(lock->l_flags & LDLM_FL_BL_AST)) {
1707 ldlm_lock_add_to_lru(lock);
1708 lock->l_flags &= ~LDLM_FL_NO_LRU;
1709 LDLM_DEBUG(lock, "join lock to lru");
1715 spin_lock(&ns->ns_unused_lock);
1716 list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1717 if (lock->l_resource == res) {
1718 ldlm_lock_remove_from_lru_nolock(lock);
1719 lock->l_flags |= LDLM_FL_NO_LRU;
1720 LDLM_DEBUG(lock, "split lock from lru");
1724 spin_unlock(&ns->ns_unused_lock);
1727 ldlm_resource_putref(res);
1731 /* Lock iterators. */
1733 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1736 struct list_head *tmp, *next;
1737 struct ldlm_lock *lock;
1738 int rc = LDLM_ITER_CONTINUE;
1743 RETURN(LDLM_ITER_CONTINUE);
1746 list_for_each_safe(tmp, next, &res->lr_granted) {
1747 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1749 if (iter(lock, closure) == LDLM_ITER_STOP)
1750 GOTO(out, rc = LDLM_ITER_STOP);
1753 list_for_each_safe(tmp, next, &res->lr_converting) {
1754 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1756 if (iter(lock, closure) == LDLM_ITER_STOP)
1757 GOTO(out, rc = LDLM_ITER_STOP);
1760 list_for_each_safe(tmp, next, &res->lr_waiting) {
1761 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1763 if (iter(lock, closure) == LDLM_ITER_STOP)
1764 GOTO(out, rc = LDLM_ITER_STOP);
1771 struct iter_helper_data {
1772 ldlm_iterator_t iter;
1776 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1778 struct iter_helper_data *helper = closure;
1779 return helper->iter(lock, helper->closure);
1782 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1784 return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1787 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1790 struct iter_helper_data helper = { iter: iter, closure: closure };
1791 return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1794 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1795 ldlm_res_iterator_t iter, void *closure)
1797 int i, rc = LDLM_ITER_CONTINUE;
1798 struct ldlm_resource *res;
1799 struct list_head *tmp;
1802 spin_lock(&ns->ns_hash_lock);
1803 for (i = 0; i < RES_HASH_SIZE; i++) {
1804 tmp = ns->ns_hash[i].next;
1805 while (tmp != &(ns->ns_hash[i])) {
1806 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1807 ldlm_resource_getref(res);
1808 spin_unlock(&ns->ns_hash_lock);
1810 rc = iter(res, closure);
1812 spin_lock(&ns->ns_hash_lock);
1814 ldlm_resource_putref_locked(res);
1815 if (rc == LDLM_ITER_STOP)
1820 spin_unlock(&ns->ns_hash_lock);
1824 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1825 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1826 const struct ldlm_res_id *res_id,
1827 ldlm_iterator_t iter, void *data)
1829 struct ldlm_resource *res;
1833 CERROR("must pass in namespace\n");
1837 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1843 ldlm_resource_foreach(res, iter, data);
1844 ldlm_resource_putref(res);
1850 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1852 struct list_head *list = closure;
1854 /* we use l_pending_chain here, because it's unused on clients. */
1855 LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1856 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1857 /* bug 9573: don't replay locks left after eviction */
1858 if (!(lock->l_flags & LDLM_FL_FAILED))
1859 list_add(&lock->l_pending_chain, list);
1860 return LDLM_ITER_CONTINUE;
1863 static int replay_lock_interpret(struct ptlrpc_request *req,
1864 struct ldlm_async_args *aa, int rc)
1866 struct ldlm_lock *lock;
1867 struct ldlm_reply *reply;
1870 atomic_dec(&req->rq_import->imp_replay_inflight);
1875 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1877 GOTO(out, rc = -EPROTO);
1879 lock = ldlm_handle2lock(&aa->lock_handle);
1881 CERROR("received replay ack for unknown local cookie "LPX64
1882 " remote cookie "LPX64 " from server %s id %s\n",
1883 aa->lock_handle.cookie, reply->lock_handle.cookie,
1884 req->rq_export->exp_client_uuid.uuid,
1885 libcfs_id2str(req->rq_peer));
1886 GOTO(out, rc = -ESTALE);
1889 lock->l_remote_handle = reply->lock_handle;
1890 LDLM_DEBUG(lock, "replayed lock:");
1891 ptlrpc_import_recovery_state_machine(req->rq_import);
1892 LDLM_LOCK_PUT(lock);
1895 ptlrpc_connect_import(req->rq_import, NULL);
1901 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1903 struct ptlrpc_request *req;
1904 struct ldlm_async_args *aa;
1905 struct ldlm_request *body;
1910 /* Bug 11974: Do not replay a lock which is actively being canceled */
1911 if (lock->l_flags & LDLM_FL_CANCELING) {
1912 LDLM_DEBUG(lock, "Not replaying canceled lock:");
1916 /* If this is reply-less callback lock, we cannot replay it, since
1917 * server might have long dropped it, but notification of that event was
1918 * lost by network. (and server granted conflicting lock already) */
1919 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1920 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
1921 ldlm_lock_cancel(lock);
1925 * If granted mode matches the requested mode, this lock is granted.
1927 * If they differ, but we have a granted mode, then we were granted
1928 * one mode and now want another: ergo, converting.
1930 * If we haven't been granted anything and are on a resource list,
1931 * then we're blocked/waiting.
1933 * If we haven't been granted anything and we're NOT on a resource list,
1934 * then we haven't got a reply yet and don't have a known disposition.
1935 * This happens whenever a lock enqueue is the request that triggers
1938 if (lock->l_granted_mode == lock->l_req_mode)
1939 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1940 else if (lock->l_granted_mode)
1941 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1942 else if (!list_empty(&lock->l_res_link))
1943 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1945 flags = LDLM_FL_REPLAY;
1947 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
1948 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
1952 /* We're part of recovery, so don't wait for it. */
1953 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1955 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1956 ldlm_lock2desc(lock, &body->lock_desc);
1957 body->lock_flags = flags;
1959 ldlm_lock2handle(lock, &body->lock_handle[0]);
1960 if (lock->l_lvb_len != 0) {
1961 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
1962 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1965 ptlrpc_request_set_replen(req);
1966 /* notify the server we've replayed all requests.
1967 * also, we mark the request to be put on a dedicated
1968 * queue to be processed after all request replayes.
1970 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
1972 LDLM_DEBUG(lock, "replaying lock:");
1974 atomic_inc(&req->rq_import->imp_replay_inflight);
1975 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1976 aa = (struct ldlm_async_args *)&req->rq_async_args;
1977 aa->lock_handle = body->lock_handle[0];
1978 req->rq_interpret_reply = replay_lock_interpret;
1979 ptlrpcd_add_req(req);
1984 int ldlm_replay_locks(struct obd_import *imp)
1986 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1987 struct list_head list;
1988 struct ldlm_lock *lock, *next;
1992 CFS_INIT_LIST_HEAD(&list);
1994 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1996 /* ensure this doesn't fall to 0 before all have been queued */
1997 atomic_inc(&imp->imp_replay_inflight);
1999 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2001 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2002 list_del_init(&lock->l_pending_chain);
2004 continue; /* or try to do the rest? */
2005 rc = replay_one_lock(imp, lock);
2008 atomic_dec(&imp->imp_replay_inflight);