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 int ldlm_expired_completion_wait(void *data)
48 struct lock_wait_data *lwd = data;
49 struct ldlm_lock *lock = lwd->lwd_lock;
50 struct obd_import *imp;
51 struct obd_device *obd;
54 if (lock->l_conn_export == NULL) {
55 static cfs_time_t next_dump = 0, last_dump = 0;
57 LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago); "
58 "not entering recovery in server code, just going "
59 "back to sleep", lock->l_enqueued_time.tv_sec,
60 CURRENT_SECONDS - lock->l_enqueued_time.tv_sec);
61 if (cfs_time_after(cfs_time_current(), next_dump)) {
62 last_dump = next_dump;
63 next_dump = cfs_time_shift(300);
64 ldlm_namespace_dump(D_DLMTRACE,
65 lock->l_resource->lr_namespace);
67 libcfs_debug_dumplog();
72 obd = lock->l_conn_export->exp_obd;
73 imp = obd->u.cli.cl_import;
74 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
75 LDLM_ERROR(lock, "lock timed out (enqueued at %lu, %lus ago), entering "
76 "recovery for %s@%s", lock->l_enqueued_time.tv_sec,
77 CURRENT_SECONDS - lock->l_enqueued_time.tv_sec,
78 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
83 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
85 /* XXX ALLOCATE - 160 bytes */
86 struct lock_wait_data lwd;
87 struct obd_device *obd;
88 struct obd_import *imp = NULL;
89 struct l_wait_info lwi;
93 if (flags == LDLM_FL_WAIT_NOREPROC) {
94 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
98 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
99 LDLM_FL_BLOCK_CONV))) {
100 cfs_waitq_signal(&lock->l_waitq);
104 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
106 ldlm_lock_dump(D_OTHER, lock, 0);
107 ldlm_reprocess_all(lock->l_resource);
111 obd = class_exp2obd(lock->l_conn_export);
113 /* if this is a local lock, then there is no import */
115 imp = obd->u.cli.cl_import;
119 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
120 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
121 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
123 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(obd_timeout),
124 ldlm_expired_completion_wait,
125 interrupted_completion_wait, &lwd);
129 spin_lock(&imp->imp_lock);
130 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
131 spin_unlock(&imp->imp_lock);
134 /* Go to sleep until the lock is granted or cancelled. */
135 rc = l_wait_event(lock->l_waitq,
136 ((lock->l_req_mode == lock->l_granted_mode) ||
137 (lock->l_flags & LDLM_FL_FAILED)), &lwi);
139 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
140 LDLM_DEBUG(lock, "client-side enqueue waking up: destroyed");
145 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
150 LDLM_DEBUG(lock, "client-side enqueue waking up: granted");
155 * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
157 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
158 void *data, int flag)
163 if (flag == LDLM_CB_CANCELING) {
164 /* Don't need to do anything here. */
168 lock_res_and_lock(lock);
169 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
170 * that ldlm_blocking_ast is called just before intent_policy method
171 * takes the ns_lock, then by the time we get the lock, we might not
172 * be the correct blocking function anymore. So check, and return
174 if (lock->l_blocking_ast != ldlm_blocking_ast) {
175 unlock_res_and_lock(lock);
179 lock->l_flags |= LDLM_FL_CBPENDING;
180 do_ast = (!lock->l_readers && !lock->l_writers);
181 unlock_res_and_lock(lock);
184 struct lustre_handle lockh;
187 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
188 ldlm_lock2handle(lock, &lockh);
189 rc = ldlm_cli_cancel(&lockh);
191 CERROR("ldlm_cli_cancel: %d\n", rc);
193 LDLM_DEBUG(lock, "Lock still has references, will be "
200 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
201 * comment in filter_intent_policy() on why you may need this.
203 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
206 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
207 * that is rather subtle: with OST-side locking, it may so happen that
208 * _all_ extent locks are held by the OST. If client wants to obtain
209 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
210 * on the server), dummy glimpse callback fires and does
211 * nothing. Client still receives correct file size due to the
212 * following fragment in filter_intent_policy():
214 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
215 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
216 * res->lr_namespace->ns_lvbo->lvbo_update) {
217 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
220 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
221 * returns correct file size to the client.
223 return -ELDLM_NO_LOCK_DATA;
226 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns, struct ldlm_res_id res_id,
227 ldlm_type_t type, ldlm_policy_data_t *policy,
228 ldlm_mode_t mode, int *flags,
229 ldlm_blocking_callback blocking,
230 ldlm_completion_callback completion,
231 ldlm_glimpse_callback glimpse,
232 void *data, __u32 lvb_len, void *lvb_swabber,
233 struct lustre_handle *lockh)
235 struct ldlm_lock *lock;
239 LASSERT(!(*flags & LDLM_FL_REPLAY));
241 CERROR("Trying to enqueue local lock in a shadow namespace\n");
245 lock = ldlm_lock_create(ns, res_id, type, mode, blocking,
246 completion, glimpse, data, lvb_len);
248 GOTO(out_nolock, err = -ENOMEM);
249 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
251 ldlm_lock_addref_internal(lock, mode);
252 ldlm_lock2handle(lock, lockh);
253 lock_res_and_lock(lock);
254 lock->l_flags |= LDLM_FL_LOCAL;
255 if (*flags & LDLM_FL_ATOMIC_CB)
256 lock->l_flags |= LDLM_FL_ATOMIC_CB;
257 lock->l_lvb_swabber = lvb_swabber;
258 unlock_res_and_lock(lock);
260 lock->l_policy_data = *policy;
261 if (type == LDLM_EXTENT)
262 lock->l_req_extent = policy->l_extent;
264 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
269 *policy = lock->l_policy_data;
270 if ((*flags) & LDLM_FL_LOCK_CHANGED)
271 res_id = lock->l_resource->lr_name;
273 LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
276 if (lock->l_completion_ast)
277 lock->l_completion_ast(lock, *flags, NULL);
279 LDLM_DEBUG(lock, "client-side local enqueue END");
287 static void failed_lock_cleanup(struct ldlm_namespace *ns,
288 struct ldlm_lock *lock,
289 struct lustre_handle *lockh, int mode)
291 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
292 lock_res_and_lock(lock);
293 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
294 unlock_res_and_lock(lock);
295 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
297 ldlm_lock_decref_and_cancel(lockh, mode);
299 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
300 * from llite/file.c/ll_file_flock(). */
301 if (lock->l_resource->lr_type == LDLM_FLOCK) {
302 ldlm_lock_destroy(lock);
306 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
307 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
308 int *flags, void *lvb, __u32 lvb_len,
309 void *lvb_swabber, struct lustre_handle *lockh,int rc)
311 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
312 int is_replay = *flags & LDLM_FL_REPLAY;
313 struct ldlm_lock *lock;
314 struct ldlm_reply *reply;
315 int cleanup_phase = 1;
318 lock = ldlm_handle2lock(lockh);
319 /* ldlm_cli_enqueue is holding a reference on this lock. */
321 LASSERT(type == LDLM_FLOCK);
325 if (rc != ELDLM_OK) {
327 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
328 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
329 if (rc == ELDLM_LOCK_ABORTED) {
330 /* Before we return, swab the reply */
331 reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF,
333 lustre_swab_ldlm_reply);
335 CERROR("Can't unpack ldlm_reply\n");
340 tmplvb = lustre_swab_repbuf(req,
345 GOTO(cleanup, rc = -EPROTO);
347 memcpy(lvb, tmplvb, lvb_len);
353 reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
354 lustre_swab_ldlm_reply);
356 CERROR("Can't unpack ldlm_reply\n");
357 GOTO(cleanup, rc = -EPROTO);
360 /* lock enqueued on the server */
363 lock_res_and_lock(lock);
364 lock->l_remote_handle = reply->lock_handle;
365 *flags = reply->lock_flags;
366 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
367 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
368 * to wait with no timeout as well */
369 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
370 unlock_res_and_lock(lock);
372 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
373 lock, reply->lock_handle.cookie, *flags);
375 /* If enqueue returned a blocked lock but the completion handler has
376 * already run, then it fixed up the resource and we don't need to do it
378 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
379 int newmode = reply->lock_desc.l_req_mode;
381 if (newmode && newmode != lock->l_req_mode) {
382 LDLM_DEBUG(lock, "server returned different mode %s",
383 ldlm_lockname[newmode]);
384 lock->l_req_mode = newmode;
387 if (reply->lock_desc.l_resource.lr_name.name[0] !=
388 lock->l_resource->lr_name.name[0]) {
389 CDEBUG(D_INFO, "remote intent success, locking %ld "
391 (long)reply->lock_desc.l_resource.lr_name.name[0],
392 (long)lock->l_resource->lr_name.name[0]);
394 ldlm_lock_change_resource(ns, lock,
395 reply->lock_desc.l_resource.lr_name);
396 if (lock->l_resource == NULL) {
398 GOTO(cleanup, rc = -ENOMEM);
400 LDLM_DEBUG(lock, "client-side enqueue, new resource");
403 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
405 lock->l_policy_data =
406 reply->lock_desc.l_policy_data;
407 if (type != LDLM_PLAIN)
408 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
411 if ((*flags) & LDLM_FL_AST_SENT ||
412 /* Cancel extent locks as soon as possible on a liblustre client,
413 * because it cannot handle asynchronous ASTs robustly (see
415 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
416 lock_res_and_lock(lock);
417 lock->l_flags |= LDLM_FL_CBPENDING;
418 unlock_res_and_lock(lock);
419 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
422 /* If the lock has already been granted by a completion AST, don't
423 * clobber the LVB with an older one. */
424 if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
426 tmplvb = lustre_swab_repbuf(req, DLM_REPLY_REC_OFF, lvb_len,
429 GOTO(cleanup, rc = -EPROTO);
430 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
434 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
435 if (lock->l_completion_ast != NULL) {
436 int err = lock->l_completion_ast(lock, *flags, NULL);
439 if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
444 if (lvb_len && lvb != NULL) {
445 /* Copy the LVB here, and not earlier, because the completion
446 * AST (if any) can override what we got in the reply */
447 memcpy(lvb, lock->l_lvb_data, lvb_len);
450 LDLM_DEBUG(lock, "client-side enqueue END");
453 if (cleanup_phase == 1 && rc)
454 failed_lock_cleanup(ns, lock, lockh, mode);
455 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
461 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
462 * a single page on the send/receive side. XXX: 512 should be changed
463 * to more adequate value. */
464 #define ldlm_req_handles_avail(exp, size, bufcount, off) \
466 int _avail = min_t(int, LDLM_MAXREQSIZE, PAGE_SIZE - 512); \
467 int _s = size[DLM_LOCKREQ_OFF]; \
468 size[DLM_LOCKREQ_OFF] = sizeof(struct ldlm_request); \
469 _avail -= lustre_msg_size(class_exp2cliimp(exp)->imp_msg_magic, \
471 _avail /= sizeof(struct lustre_handle); \
472 _avail += LDLM_LOCKREQ_HANDLES - off; \
473 size[DLM_LOCKREQ_OFF] = _s; \
477 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
478 * @count locks in @cancels. */
479 struct ptlrpc_request *ldlm_prep_enqueue_req(struct obd_export *exp,
480 int bufcount, int *size,
481 struct list_head *cancels,
484 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
485 struct ldlm_request *dlm = NULL;
486 struct ptlrpc_request *req;
492 if (exp_connect_cancelset(exp)) {
493 /* Estimate the amount of free space in the request. */
494 int avail = ldlm_req_handles_avail(exp, size, bufcount,
495 LDLM_ENQUEUE_CANCEL_OFF);
496 LASSERT(avail >= count);
498 /* Cancel lru locks here _only_ if the server supports
499 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
500 * rpc right on enqueue, what will make it slower, vs.
501 * asynchronous rpc in blocking thread. */
502 count += ldlm_cancel_lru_local(ns, cancels, 1, avail - count,
504 size[DLM_LOCKREQ_OFF] =
505 ldlm_request_bufsize(count, LDLM_ENQUEUE);
507 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
508 LDLM_ENQUEUE, bufcount, size, NULL);
509 if (exp_connect_cancelset(exp) && req) {
510 dlm = lustre_msg_buf(req->rq_reqmsg,
511 DLM_LOCKREQ_OFF, sizeof(*dlm));
512 /* Skip first lock handler in ldlm_request_pack(), this method
513 * will incrment @lock_count according to the lock handle amount
514 * actually written to the buffer. */
515 dlm->lock_count = LDLM_ENQUEUE_CANCEL_OFF;
516 ldlm_cli_cancel_list(cancels, count, req, DLM_LOCKREQ_OFF);
518 ldlm_lock_list_put(cancels, l_bl_ast, count);
523 /* If a request has some specific initialisation it is passed in @reqp,
524 * otherwise it is created in ldlm_cli_enqueue.
526 * Supports sync and async requests, pass @async flag accordingly. If a
527 * request was created in ldlm_cli_enqueue and it is the async request,
528 * pass it to the caller in @reqp. */
529 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
530 struct ldlm_enqueue_info *einfo, struct ldlm_res_id res_id,
531 ldlm_policy_data_t *policy, int *flags,
532 void *lvb, __u32 lvb_len, void *lvb_swabber,
533 struct lustre_handle *lockh, int async)
535 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
536 struct ldlm_lock *lock;
537 struct ldlm_request *body;
538 struct ldlm_reply *reply;
539 int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
540 [DLM_LOCKREQ_OFF] = sizeof(*body),
541 [DLM_REPLY_REC_OFF] = lvb_len };
542 int is_replay = *flags & LDLM_FL_REPLAY;
543 int req_passed_in = 1, rc, err;
544 struct ptlrpc_request *req;
547 LASSERT(exp != NULL);
549 /* If we're replaying this lock, just check some invariants.
550 * If we're creating a new lock, get everything all setup nice. */
552 lock = ldlm_handle2lock(lockh);
553 LASSERT(lock != NULL);
554 LDLM_DEBUG(lock, "client-side enqueue START");
555 LASSERT(exp == lock->l_conn_export);
557 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
558 einfo->ei_mode, einfo->ei_cb_bl,
559 einfo->ei_cb_cp, einfo->ei_cb_gl,
560 einfo->ei_cbdata, lvb_len);
563 /* for the local lock, add the reference */
564 ldlm_lock_addref_internal(lock, einfo->ei_mode);
565 ldlm_lock2handle(lock, lockh);
566 lock->l_lvb_swabber = lvb_swabber;
567 if (policy != NULL) {
568 /* INODEBITS_INTEROP: If the server does not support
569 * inodebits, we will request a plain lock in the
570 * descriptor (ldlm_lock2desc() below) but use an
571 * inodebits lock internally with both bits set.
573 if (einfo->ei_type == LDLM_IBITS &&
574 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
575 lock->l_policy_data.l_inodebits.bits =
576 MDS_INODELOCK_LOOKUP |
577 MDS_INODELOCK_UPDATE;
579 lock->l_policy_data = *policy;
582 if (einfo->ei_type == LDLM_EXTENT)
583 lock->l_req_extent = policy->l_extent;
584 LDLM_DEBUG(lock, "client-side enqueue START");
587 /* lock not sent to server yet */
589 if (reqp == NULL || *reqp == NULL) {
590 req = ldlm_prep_enqueue_req(exp, 2, size, NULL, 0);
592 failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
601 LASSERTF(lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF) >=
602 sizeof(*body), "buflen[%d] = %d, not "LPSZ"\n",
604 lustre_msg_buflen(req->rq_reqmsg, DLM_LOCKREQ_OFF),
608 lock->l_conn_export = exp;
609 lock->l_export = NULL;
610 lock->l_blocking_ast = einfo->ei_cb_bl;
612 /* Dump lock data into the request buffer */
613 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
614 ldlm_lock2desc(lock, &body->lock_desc);
615 body->lock_flags = *flags;
616 body->lock_handle[0] = *lockh;
618 /* Continue as normal. */
619 if (!req_passed_in) {
620 size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
621 ptlrpc_req_set_repsize(req, 2 + (lvb_len > 0), size);
625 * Liblustre client doesn't get extent locks, except for O_APPEND case
626 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
627 * [i_size, OBD_OBJECT_EOF] lock is taken.
629 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
630 policy->l_extent.end == OBD_OBJECT_EOF));
633 LASSERT(reqp != NULL);
637 LDLM_DEBUG(lock, "sending request");
638 rc = ptlrpc_queue_wait(req);
639 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
640 einfo->ei_mode, flags, lvb, lvb_len,
641 lvb_swabber, lockh, rc);
643 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
644 * one reference that we took */
650 if (!req_passed_in && req != NULL) {
651 ptlrpc_req_finished(req);
659 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
662 struct ldlm_resource *res;
665 if (lock->l_resource->lr_namespace->ns_client) {
666 CERROR("Trying to cancel local lock\n");
669 LDLM_DEBUG(lock, "client-side local convert");
671 res = ldlm_lock_convert(lock, new_mode, flags);
673 ldlm_reprocess_all(res);
678 LDLM_DEBUG(lock, "client-side local convert handler END");
683 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
684 * conversion of locks which are on the waiting or converting queue */
685 /* Caller of this code is supposed to take care of lock readers/writers
687 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, int *flags)
689 struct ldlm_request *body;
690 struct ldlm_reply *reply;
691 struct ldlm_lock *lock;
692 struct ldlm_resource *res;
693 struct ptlrpc_request *req;
694 int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
695 [DLM_LOCKREQ_OFF] = sizeof(*body) };
699 lock = ldlm_handle2lock(lockh);
706 if (lock->l_conn_export == NULL)
707 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
709 LDLM_DEBUG(lock, "client-side convert");
711 req = ptlrpc_prep_req(class_exp2cliimp(lock->l_conn_export),
712 LUSTRE_DLM_VERSION, LDLM_CONVERT, 2, size, NULL);
714 GOTO(out, rc = -ENOMEM);
716 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
717 body->lock_handle[0] = lock->l_remote_handle;
719 body->lock_desc.l_req_mode = new_mode;
720 body->lock_flags = *flags;
722 size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
723 ptlrpc_req_set_repsize(req, 2, size);
725 rc = ptlrpc_queue_wait(req);
729 reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
730 lustre_swab_ldlm_reply);
732 CERROR ("Can't unpack ldlm_reply\n");
733 GOTO (out, rc = -EPROTO);
737 GOTO(out, rc = req->rq_status);
739 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
741 ldlm_reprocess_all(res);
742 /* Go to sleep until the lock is granted. */
743 /* FIXME: or cancelled. */
744 if (lock->l_completion_ast) {
745 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
756 ptlrpc_req_finished(req);
760 /* Cancel locks locally.
762 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
763 * LDLM_FL_CANCELING otherwise;
764 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
765 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
767 int rc = LDLM_FL_LOCAL_ONLY;
770 if (lock->l_conn_export) {
773 LDLM_DEBUG(lock, "client-side cancel");
774 /* Set this flag to prevent others from getting new references*/
775 lock_res_and_lock(lock);
776 lock->l_flags |= LDLM_FL_CBPENDING;
777 local_only = (lock->l_flags &
778 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
779 ldlm_cancel_callback(lock);
780 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
781 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
782 unlock_res_and_lock(lock);
785 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
787 rc = LDLM_FL_LOCAL_ONLY;
789 ldlm_lock_cancel(lock);
791 if (lock->l_resource->lr_namespace->ns_client) {
792 LDLM_ERROR(lock, "Trying to cancel local lock");
795 LDLM_DEBUG(lock, "server-side local cancel");
796 ldlm_lock_cancel(lock);
797 ldlm_reprocess_all(lock->l_resource);
798 LDLM_DEBUG(lock, "server-side local cancel handler END");
804 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
805 of the request @req. */
806 static void ldlm_cancel_pack(struct ptlrpc_request *req, int off,
807 struct list_head *head, int count)
809 struct ldlm_request *dlm;
810 struct ldlm_lock *lock;
814 dlm = lustre_msg_buf(req->rq_reqmsg, off, sizeof(*dlm));
815 LASSERT(dlm != NULL);
817 /* Check the room in the request buffer. */
818 max = lustre_msg_buflen(req->rq_reqmsg, off) -
819 sizeof(struct ldlm_request);
820 max /= sizeof(struct lustre_handle);
821 max += LDLM_LOCKREQ_HANDLES;
822 LASSERT(max >= dlm->lock_count + count);
824 /* XXX: it would be better to pack lock handles grouped by resource.
825 * so that the server cancel would call filter_lvbo_update() less
827 list_for_each_entry(lock, head, l_bl_ast) {
830 LASSERT(lock->l_conn_export);
831 /* Pack the lock handle to the given request buffer. */
832 LDLM_DEBUG(lock, "packing");
833 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
836 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
840 /* Prepare and send a batched cancel rpc, it will include count lock handles
841 * of locks given in @head. */
842 int ldlm_cli_cancel_req(struct obd_export *exp,
843 struct list_head *cancels, int count)
845 struct ptlrpc_request *req = NULL;
846 struct ldlm_request *body;
847 int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
848 [DLM_LOCKREQ_OFF] = sizeof(*body) };
849 struct obd_import *imp;
854 LASSERT(exp != NULL);
857 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
860 free = ldlm_req_handles_avail(exp, size, 2, 0);
864 size[DLM_LOCKREQ_OFF] = ldlm_request_bufsize(count, LDLM_CANCEL);
866 imp = class_exp2cliimp(exp);
867 if (imp == NULL || imp->imp_invalid) {
869 "skipping cancel on invalid import %p\n", imp);
873 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_CANCEL, 2,
876 GOTO(out, rc = -ENOMEM);
878 req->rq_no_resend = 1;
880 /* XXX FIXME bug 249 */
881 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
882 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
884 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF,
886 ldlm_cancel_pack(req, DLM_LOCKREQ_OFF, cancels, count);
888 ptlrpc_req_set_repsize(req, 1, NULL);
889 rc = ptlrpc_queue_wait(req);
892 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
893 "out of sync -- not fatal\n",
894 libcfs_nid2str(req->rq_import->
895 imp_connection->c_peer.nid));
896 } else if (rc == -ETIMEDOUT) {
897 ptlrpc_req_finished(req);
899 } else if (rc != ELDLM_OK) {
900 CERROR("Got rc %d from cancel RPC: canceling "
908 ptlrpc_req_finished(req);
911 return sent ? sent : rc;
914 int ldlm_cli_cancel(struct lustre_handle *lockh)
916 struct ldlm_lock *lock;
921 /* concurrent cancels on the same handle can happen */
922 lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
924 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
928 rc = ldlm_cli_cancel_local(lock);
929 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY)
932 list_add(&lock->l_bl_ast, &head);
933 rc = ldlm_cli_cancel_req(lock->l_conn_export, &head, 1);
937 return rc < 0 ? rc : 0;
940 /* - Free space in lru for @count new locks,
941 * redundant unused locks are canceled locally;
942 * - also cancel locally unused aged locks;
943 * - do not cancel more than @max locks;
944 * - GET the found locks and add them into the @cancels list.
946 * A client lock can be added to the l_bl_ast list only when it is
947 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
948 * There are the following use cases: ldlm_cancel_resource_local(),
949 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
950 * flag properly. As any attempt to cancel a lock rely on this flag,
951 * l_bl_ast list is accessed later without any special locking. */
952 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
953 int count, int max, int flags)
955 cfs_time_t cur = cfs_time_current();
956 struct ldlm_lock *lock, *next;
957 int rc, added = 0, left;
960 spin_lock(&ns->ns_unused_lock);
961 count += ns->ns_nr_unused - ns->ns_max_unused;
962 while (!list_empty(&ns->ns_unused_list)) {
963 if (max && added >= max)
966 list_for_each_entry(lock, &ns->ns_unused_list, l_lru) {
967 /* somebody is already doing CANCEL or there is a
968 * blocking request will send cancel. */
969 if (!(lock->l_flags & LDLM_FL_CANCELING) &&
970 !(lock->l_flags & LDLM_FL_BL_AST))
973 if (&lock->l_lru == &ns->ns_unused_list)
976 if ((added >= count) &&
977 (!(flags & LDLM_CANCEL_AGED) ||
978 cfs_time_before_64(cur, (__u64)ns->ns_max_age +
982 LDLM_LOCK_GET(lock); /* dropped by bl thread */
983 spin_unlock(&ns->ns_unused_lock);
985 lock_res_and_lock(lock);
986 /* Check flags again under the lock. */
987 if ((lock->l_flags & LDLM_FL_CANCELING) ||
988 (lock->l_flags & LDLM_FL_BL_AST) ||
989 (ldlm_lock_remove_from_lru(lock) == 0)) {
990 /* other thread is removing lock from lru or
991 * somebody is already doing CANCEL or
992 * there is a blocking request which will send
993 * cancel by itseft. */
994 unlock_res_and_lock(lock);
996 spin_lock(&ns->ns_unused_lock);
999 LASSERT(!lock->l_readers && !lock->l_writers);
1001 /* If we have chosen to canecl this lock voluntarily, we better
1002 send cancel notification to server, so that it frees
1003 appropriate state. This might lead to a race where while
1004 we are doing cancel here, server is also silently
1005 cancelling this lock. */
1006 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1008 /* Setting the CBPENDING flag is a little misleading, but
1009 * prevents an important race; namely, once CBPENDING is set,
1010 * the lock can accumulate no more readers/writers. Since
1011 * readers and writers are already zero here, ldlm_lock_decref
1012 * won't see this flag and call l_blocking_ast */
1013 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1014 /* We can't re-add to l_lru as it confuses the refcounting in
1015 * ldlm_lock_remove_from_lru() if an AST arrives after we drop
1016 * ns_lock below. We use l_bl_ast and can't use l_pending_chain
1017 * as it is used both on server and client nevertheles bug 5666
1018 * says it is used only on server. --umka */
1020 LASSERT(list_empty(&lock->l_bl_ast));
1021 list_add(&lock->l_bl_ast, cancels);
1022 unlock_res_and_lock(lock);
1023 spin_lock(&ns->ns_unused_lock);
1026 spin_unlock(&ns->ns_unused_lock);
1028 /* Handle only @added inserted locks. */
1030 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1034 rc = ldlm_cli_cancel_local(lock);
1035 if (rc == LDLM_FL_BL_AST) {
1036 CFS_LIST_HEAD(head);
1038 LDLM_DEBUG(lock, "Cancel lock separately");
1039 list_del_init(&lock->l_bl_ast);
1040 list_add(&lock->l_bl_ast, &head);
1041 ldlm_cli_cancel_req(lock->l_conn_export, &head, 1);
1042 rc = LDLM_FL_LOCAL_ONLY;
1044 if (rc == LDLM_FL_LOCAL_ONLY) {
1045 /* CANCEL RPC should not be sent to server. */
1046 list_del_init(&lock->l_bl_ast);
1047 LDLM_LOCK_PUT(lock);
1055 /* when called with LDLM_ASYNC the blocking callback will be handled
1056 * in a thread and this function will return after the thread has been
1057 * asked to call the callback. when called with LDLM_SYNC the blocking
1058 * callback will be performed in this function. */
1059 int ldlm_cancel_lru(struct ldlm_namespace *ns, ldlm_sync_t sync)
1061 CFS_LIST_HEAD(cancels);
1066 sync = LDLM_SYNC; /* force to be sync in user space */
1068 count = ldlm_cancel_lru_local(ns, &cancels, 0, 0, 0);
1069 if (sync == LDLM_ASYNC) {
1070 struct ldlm_lock *lock, *next;
1071 list_for_each_entry_safe(lock, next, &cancels, l_bl_ast) {
1072 /* Remove from the list to allow blocking thread to
1073 * re-use l_bl_ast. */
1074 list_del_init(&lock->l_bl_ast);
1075 rc = ldlm_bl_to_thread(ns, NULL, lock,
1078 list_add_tail(&lock->l_bl_ast, &next->l_bl_ast);
1082 /* If some locks are left in the list in ASYNC mode, or
1083 * this is SYNC mode, cancel the list. */
1084 ldlm_cli_cancel_list(&cancels, count, NULL, DLM_LOCKREQ_OFF);
1088 /* Find and cancel locally unused locks found on resource, matched to the
1089 * given policy, mode. GET the found locks and add them into the @cancels
1091 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1092 struct list_head *cancels,
1093 ldlm_policy_data_t *policy,
1094 ldlm_mode_t mode, int lock_flags,
1095 int flags, void *opaque)
1097 struct ldlm_lock *lock, *next;
1098 int count = 0, left;
1102 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1103 if (opaque != NULL && lock->l_ast_data != opaque) {
1104 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1105 lock->l_ast_data, opaque);
1110 if (lock->l_readers || lock->l_writers) {
1111 if (flags & LDLM_FL_WARN) {
1112 LDLM_ERROR(lock, "lock in use");
1118 /* If somebody is already doing CANCEL, or blocking ast came,
1119 * skip this lock. */
1120 if (lock->l_flags & LDLM_FL_BL_AST ||
1121 lock->l_flags & LDLM_FL_CANCELING)
1124 if (lockmode_compat(lock->l_granted_mode, mode))
1127 /* If policy is given and this is IBITS lock, add to list only
1128 * those locks that match by policy. */
1129 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1130 !(lock->l_policy_data.l_inodebits.bits &
1131 policy->l_inodebits.bits))
1134 /* See CBPENDING comment in ldlm_cancel_lru */
1135 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1138 LASSERT(list_empty(&lock->l_bl_ast));
1139 list_add(&lock->l_bl_ast, cancels);
1140 LDLM_LOCK_GET(lock);
1145 /* Handle only @count inserted locks. */
1147 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1148 int rc = LDLM_FL_LOCAL_ONLY;
1152 if (flags & LDLM_FL_LOCAL_ONLY)
1153 ldlm_lock_cancel(lock);
1155 rc = ldlm_cli_cancel_local(lock);
1157 if (rc == LDLM_FL_BL_AST) {
1158 CFS_LIST_HEAD(head);
1160 LDLM_DEBUG(lock, "Cancel lock separately");
1161 list_del_init(&lock->l_bl_ast);
1162 list_add(&lock->l_bl_ast, &head);
1163 ldlm_cli_cancel_req(lock->l_conn_export, &head, 1);
1164 rc = LDLM_FL_LOCAL_ONLY;
1166 if (rc == LDLM_FL_LOCAL_ONLY) {
1167 /* CANCEL RPC should not be sent to server. */
1168 list_del_init(&lock->l_bl_ast);
1169 LDLM_LOCK_PUT(lock);
1176 /* If @req is NULL, send CANCEL request to server with handles of locks
1177 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1178 * separately per lock.
1179 * If @req is not NULL, put handles of locks in @cancels into the request
1180 * buffer at the offset @off.
1181 * Destroy @cancels at the end. */
1182 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1183 struct ptlrpc_request *req, int off)
1185 struct ldlm_lock *lock;
1189 if (list_empty(cancels) || count == 0)
1193 LASSERT(!list_empty(cancels));
1194 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1195 LASSERT(lock->l_conn_export);
1197 if (exp_connect_cancelset(lock->l_conn_export)) {
1200 ldlm_cancel_pack(req, off, cancels, count);
1202 res = ldlm_cli_cancel_req(lock->l_conn_export,
1205 res = ldlm_cli_cancel_req(lock->l_conn_export,
1210 CERROR("ldlm_cli_cancel_list: %d\n", res);
1215 ldlm_lock_list_put(cancels, l_bl_ast, res);
1217 LASSERT(list_empty(cancels));
1221 static int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1222 struct ldlm_res_id res_id,
1223 int flags, void *opaque)
1225 struct ldlm_resource *res;
1226 CFS_LIST_HEAD(cancels);
1231 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1233 /* This is not a problem. */
1234 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id.name[0]);
1238 count = ldlm_cancel_resource_local(res, &cancels, NULL, LCK_MINMODE,
1240 rc = ldlm_cli_cancel_list(&cancels, count, NULL, DLM_LOCKREQ_OFF);
1242 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1244 ldlm_resource_putref(res);
1248 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1250 int no_resource = 0;
1252 spin_lock(&ns->ns_hash_lock);
1253 if (ns->ns_resources == 0)
1255 spin_unlock(&ns->ns_hash_lock);
1257 RETURN(no_resource);
1260 /* Cancel all locks on a namespace (or a specific resource, if given)
1261 * that have 0 readers/writers.
1263 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1264 * to notify the server. */
1265 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1266 struct ldlm_res_id *res_id, int flags, void *opaque)
1275 RETURN(ldlm_cli_cancel_unused_resource(ns, *res_id, flags,
1278 spin_lock(&ns->ns_hash_lock);
1279 for (i = 0; i < RES_HASH_SIZE; i++) {
1280 struct list_head *tmp;
1281 tmp = ns->ns_hash[i].next;
1282 while (tmp != &(ns->ns_hash[i])) {
1283 struct ldlm_resource *res;
1286 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1287 ldlm_resource_getref(res);
1288 spin_unlock(&ns->ns_hash_lock);
1290 rc = ldlm_cli_cancel_unused_resource(ns, res->lr_name,
1294 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1295 res->lr_name.name[0], rc);
1297 spin_lock(&ns->ns_hash_lock);
1299 ldlm_resource_putref_locked(res);
1302 spin_unlock(&ns->ns_hash_lock);
1307 /* join/split resource locks to/from lru list */
1308 int ldlm_cli_join_lru(struct ldlm_namespace *ns,
1309 struct ldlm_res_id *res_id, int join)
1311 struct ldlm_resource *res;
1312 struct ldlm_lock *lock, *n;
1316 LASSERT(ns->ns_client == LDLM_NAMESPACE_CLIENT);
1318 res = ldlm_resource_get(ns, NULL, *res_id, LDLM_EXTENT, 0);
1321 LASSERT(res->lr_type == LDLM_EXTENT);
1327 list_for_each_entry_safe (lock, n, &res->lr_granted, l_res_link) {
1328 if (list_empty(&lock->l_lru) &&
1329 !lock->l_readers && !lock->l_writers &&
1330 !(lock->l_flags & LDLM_FL_LOCAL) &&
1331 !(lock->l_flags & LDLM_FL_CBPENDING)) {
1332 lock->l_last_used = cfs_time_current();
1333 spin_lock(&ns->ns_unused_lock);
1334 LASSERT(ns->ns_nr_unused >= 0);
1335 list_add_tail(&lock->l_lru, &ns->ns_unused_list);
1337 spin_unlock(&ns->ns_unused_lock);
1338 lock->l_flags &= ~LDLM_FL_NO_LRU;
1339 LDLM_DEBUG(lock, "join lock to lru");
1345 spin_lock(&ns->ns_unused_lock);
1346 list_for_each_entry_safe (lock, n, &ns->ns_unused_list, l_lru) {
1347 if (lock->l_resource == res) {
1348 ldlm_lock_remove_from_lru_nolock(lock);
1349 lock->l_flags |= LDLM_FL_NO_LRU;
1350 LDLM_DEBUG(lock, "split lock from lru");
1354 spin_unlock(&ns->ns_unused_lock);
1357 ldlm_resource_putref(res);
1361 /* Lock iterators. */
1363 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1366 struct list_head *tmp, *next;
1367 struct ldlm_lock *lock;
1368 int rc = LDLM_ITER_CONTINUE;
1373 RETURN(LDLM_ITER_CONTINUE);
1376 list_for_each_safe(tmp, next, &res->lr_granted) {
1377 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1379 if (iter(lock, closure) == LDLM_ITER_STOP)
1380 GOTO(out, rc = LDLM_ITER_STOP);
1383 list_for_each_safe(tmp, next, &res->lr_converting) {
1384 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1386 if (iter(lock, closure) == LDLM_ITER_STOP)
1387 GOTO(out, rc = LDLM_ITER_STOP);
1390 list_for_each_safe(tmp, next, &res->lr_waiting) {
1391 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1393 if (iter(lock, closure) == LDLM_ITER_STOP)
1394 GOTO(out, rc = LDLM_ITER_STOP);
1401 struct iter_helper_data {
1402 ldlm_iterator_t iter;
1406 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1408 struct iter_helper_data *helper = closure;
1409 return helper->iter(lock, helper->closure);
1412 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1414 return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1417 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1420 struct iter_helper_data helper = { iter: iter, closure: closure };
1421 return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1424 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1425 ldlm_res_iterator_t iter, void *closure)
1427 int i, rc = LDLM_ITER_CONTINUE;
1428 struct ldlm_resource *res;
1429 struct list_head *tmp;
1432 spin_lock(&ns->ns_hash_lock);
1433 for (i = 0; i < RES_HASH_SIZE; i++) {
1434 tmp = ns->ns_hash[i].next;
1435 while (tmp != &(ns->ns_hash[i])) {
1436 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1437 ldlm_resource_getref(res);
1438 spin_unlock(&ns->ns_hash_lock);
1440 rc = iter(res, closure);
1442 spin_lock(&ns->ns_hash_lock);
1444 ldlm_resource_putref_locked(res);
1445 if (rc == LDLM_ITER_STOP)
1450 spin_unlock(&ns->ns_hash_lock);
1454 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1455 void ldlm_resource_iterate(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1456 ldlm_iterator_t iter, void *data)
1458 struct ldlm_resource *res;
1462 CERROR("must pass in namespace\n");
1466 res = ldlm_resource_get(ns, NULL, *res_id, 0, 0);
1472 ldlm_resource_foreach(res, iter, data);
1473 ldlm_resource_putref(res);
1479 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1481 struct list_head *list = closure;
1483 /* we use l_pending_chain here, because it's unused on clients. */
1484 LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1485 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1486 /* bug 9573: don't replay locks left after eviction */
1487 if (!(lock->l_flags & LDLM_FL_FAILED))
1488 list_add(&lock->l_pending_chain, list);
1489 return LDLM_ITER_CONTINUE;
1492 static int replay_lock_interpret(struct ptlrpc_request *req,
1493 void * data, int rc)
1495 struct ldlm_lock *lock;
1496 struct ldlm_reply *reply;
1499 atomic_dec(&req->rq_import->imp_replay_inflight);
1503 lock = req->rq_async_args.pointer_arg[0];
1504 LASSERT(lock != NULL);
1506 reply = lustre_swab_repbuf(req, DLM_LOCKREPLY_OFF, sizeof(*reply),
1507 lustre_swab_ldlm_reply);
1508 if (reply == NULL) {
1509 CERROR("Can't unpack ldlm_reply\n");
1510 GOTO (out, rc = -EPROTO);
1513 lock->l_remote_handle = reply->lock_handle;
1514 LDLM_DEBUG(lock, "replayed lock:");
1515 ptlrpc_import_recovery_state_machine(req->rq_import);
1518 ptlrpc_connect_import(req->rq_import, NULL);
1524 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
1526 struct ptlrpc_request *req;
1527 struct ldlm_request *body;
1528 struct ldlm_reply *reply;
1530 int size[3] = { sizeof(struct ptlrpc_body) };
1534 /* If this is reply-less callback lock, we cannot replay it, since
1535 * server might have long dropped it, but notification of that event was
1536 * lost by network. (and server granted conflicting lock already) */
1537 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1538 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
1539 ldlm_lock_cancel(lock);
1543 * If granted mode matches the requested mode, this lock is granted.
1545 * If they differ, but we have a granted mode, then we were granted
1546 * one mode and now want another: ergo, converting.
1548 * If we haven't been granted anything and are on a resource list,
1549 * then we're blocked/waiting.
1551 * If we haven't been granted anything and we're NOT on a resource list,
1552 * then we haven't got a reply yet and don't have a known disposition.
1553 * This happens whenever a lock enqueue is the request that triggers
1556 if (lock->l_granted_mode == lock->l_req_mode)
1557 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
1558 else if (lock->l_granted_mode)
1559 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
1560 else if (!list_empty(&lock->l_res_link))
1561 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
1563 flags = LDLM_FL_REPLAY;
1565 size[DLM_LOCKREQ_OFF] = sizeof(*body);
1566 req = ptlrpc_prep_req(imp, LUSTRE_DLM_VERSION, LDLM_ENQUEUE, 2, size,
1571 /* We're part of recovery, so don't wait for it. */
1572 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
1574 body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
1575 ldlm_lock2desc(lock, &body->lock_desc);
1576 body->lock_flags = flags;
1578 ldlm_lock2handle(lock, &body->lock_handle[0]);
1579 size[DLM_LOCKREPLY_OFF] = sizeof(*reply);
1580 if (lock->l_lvb_len != 0) {
1582 size[DLM_REPLY_REC_OFF] = lock->l_lvb_len;
1584 ptlrpc_req_set_repsize(req, buffers, size);
1586 LDLM_DEBUG(lock, "replaying lock:");
1588 atomic_inc(&req->rq_import->imp_replay_inflight);
1589 req->rq_async_args.pointer_arg[0] = lock;
1590 req->rq_interpret_reply = replay_lock_interpret;
1591 ptlrpcd_add_req(req);
1596 int ldlm_replay_locks(struct obd_import *imp)
1598 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
1599 struct list_head list;
1600 struct ldlm_lock *lock, *next;
1604 CFS_INIT_LIST_HEAD(&list);
1606 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1608 /* ensure this doesn't fall to 0 before all have been queued */
1609 atomic_inc(&imp->imp_replay_inflight);
1611 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
1613 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
1614 list_del_init(&lock->l_pending_chain);
1616 continue; /* or try to do the rest? */
1617 rc = replay_one_lock(imp, lock);
1620 atomic_dec(&imp->imp_replay_inflight);