1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <liblustre.h>
43 #include <lustre_dlm.h>
44 #include <obd_class.h>
47 #include "ldlm_internal.h"
49 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
50 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
51 "lock enqueue timeout minimum");
53 static void interrupted_completion_wait(void *data)
57 struct lock_wait_data {
58 struct ldlm_lock *lwd_lock;
62 struct ldlm_async_args {
63 struct lustre_handle lock_handle;
66 int ldlm_expired_completion_wait(void *data)
68 struct lock_wait_data *lwd = data;
69 struct ldlm_lock *lock = lwd->lwd_lock;
70 struct obd_import *imp;
71 struct obd_device *obd;
74 if (lock->l_conn_export == NULL) {
75 static cfs_time_t next_dump = 0, last_dump = 0;
77 if (ptlrpc_check_suspend())
80 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
81 CFS_DURATION_T"s ago); not entering recovery in "
82 "server code, just going back to sleep",
83 lock->l_last_activity,
84 cfs_time_sub(cfs_time_current_sec(),
85 lock->l_last_activity));
86 if (cfs_time_after(cfs_time_current(), next_dump)) {
87 last_dump = next_dump;
88 next_dump = cfs_time_shift(300);
89 ldlm_namespace_dump(D_DLMTRACE,
90 lock->l_resource->lr_namespace);
92 libcfs_debug_dumplog();
97 obd = lock->l_conn_export->exp_obd;
98 imp = obd->u.cli.cl_import;
99 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
100 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
101 CFS_DURATION_T"s ago), entering recovery for %s@%s",
102 lock->l_last_activity,
103 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
104 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
109 /* We use the same basis for both server side and client side functions
110 from a single node. */
111 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
113 int timeout = at_get(&lock->l_resource->lr_namespace->ns_at_estimate);
115 return obd_timeout / 2;
116 /* Since these are non-updating timeouts, we should be conservative.
117 It would be nice to have some kind of "early reply" mechanism for
118 lock callbacks too... */
119 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
120 return max(timeout, ldlm_enqueue_min);
122 EXPORT_SYMBOL(ldlm_get_enq_timeout);
125 * Helper function for ldlm_completion_ast(), updating timings when lock is
128 static int ldlm_completion_tail(struct ldlm_lock *lock)
133 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
134 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
137 delay = cfs_time_sub(cfs_time_current_sec(),
138 lock->l_last_activity);
139 LDLM_DEBUG(lock, "client-side enqueue: granted after "
140 CFS_DURATION_T"s", delay);
142 /* Update our time estimate */
143 at_measured(&lock->l_resource->lr_namespace->ns_at_estimate,
151 * Implementation of ->l_completion_ast() for a client, that doesn't wait
152 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
153 * other threads that cannot block for long.
155 int ldlm_completion_ast_async(struct ldlm_lock *lock, int flags, void *data)
159 if (flags == LDLM_FL_WAIT_NOREPROC) {
160 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
164 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
165 LDLM_FL_BLOCK_CONV))) {
166 cfs_waitq_signal(&lock->l_waitq);
167 RETURN(ldlm_completion_tail(lock));
170 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
172 ldlm_lock_dump(D_OTHER, lock, 0);
173 ldlm_reprocess_all(lock->l_resource);
178 * Client side LDLM "completion" AST. This is called in several cases:
180 * - when a reply to an ENQUEUE rpc is received from the server
181 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
182 * this point (determined by flags);
184 * - when LDLM_CP_CALLBACK rpc comes to client to notify it that lock has
187 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
190 * - to force all locks when resource is destroyed (cleanup_resource());
192 * - during lock conversion (not used currently).
194 * If lock is not granted in the first case, this function waits until second
195 * or penultimate cases happen in some other thread.
198 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
200 /* XXX ALLOCATE - 160 bytes */
201 struct lock_wait_data lwd;
202 struct obd_device *obd;
203 struct obd_import *imp = NULL;
204 struct l_wait_info lwi;
209 if (flags == LDLM_FL_WAIT_NOREPROC) {
210 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
214 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
215 LDLM_FL_BLOCK_CONV))) {
216 cfs_waitq_signal(&lock->l_waitq);
220 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
222 ldlm_lock_dump(D_OTHER, lock, 0);
226 obd = class_exp2obd(lock->l_conn_export);
228 /* if this is a local lock, then there is no import */
230 imp = obd->u.cli.cl_import;
233 /* Wait a long time for enqueue - server may have to callback a
234 lock from another client. Server will evict the other client if it
235 doesn't respond reasonably, and then give us the lock. */
236 timeout = ldlm_get_enq_timeout(lock) * 2;
240 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
241 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
242 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
244 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
245 ldlm_expired_completion_wait,
246 interrupted_completion_wait, &lwd);
250 cfs_spin_lock(&imp->imp_lock);
251 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
252 cfs_spin_unlock(&imp->imp_lock);
255 if (ns_is_client(lock->l_resource->lr_namespace) &&
256 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
257 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
258 lock->l_flags |= LDLM_FL_FAIL_LOC;
261 /* Go to sleep until the lock is granted or cancelled. */
262 rc = l_wait_event(lock->l_waitq,
263 is_granted_or_cancelled(lock), &lwi);
267 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
272 RETURN(ldlm_completion_tail(lock));
276 * A helper to build a blocking ast function
278 * Perform a common operation for blocking asts:
279 * defferred lock cancellation.
281 * \param lock the lock blocking or canceling ast was called on
283 * \see mdt_blocking_ast
284 * \see ldlm_blocking_ast
286 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
291 lock->l_flags |= LDLM_FL_CBPENDING;
292 do_ast = (!lock->l_readers && !lock->l_writers);
293 unlock_res_and_lock(lock);
296 struct lustre_handle lockh;
299 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
300 ldlm_lock2handle(lock, &lockh);
301 rc = ldlm_cli_cancel(&lockh);
303 CERROR("ldlm_cli_cancel: %d\n", rc);
305 LDLM_DEBUG(lock, "Lock still has references, will be "
312 * Server blocking AST
314 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
317 * \param lock the lock which blocks a request or cancelling lock
320 * \param flag indicates whether this cancelling or blocking callback
322 * \see ldlm_blocking_ast_nocheck
324 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
325 void *data, int flag)
329 if (flag == LDLM_CB_CANCELING) {
330 /* Don't need to do anything here. */
334 lock_res_and_lock(lock);
335 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
336 * that ldlm_blocking_ast is called just before intent_policy method
337 * takes the ns_lock, then by the time we get the lock, we might not
338 * be the correct blocking function anymore. So check, and return
340 if (lock->l_blocking_ast != ldlm_blocking_ast) {
341 unlock_res_and_lock(lock);
344 RETURN(ldlm_blocking_ast_nocheck(lock));
348 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
349 * comment in filter_intent_policy() on why you may need this.
351 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
354 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
355 * that is rather subtle: with OST-side locking, it may so happen that
356 * _all_ extent locks are held by the OST. If client wants to obtain
357 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
358 * on the server), dummy glimpse callback fires and does
359 * nothing. Client still receives correct file size due to the
360 * following fragment in filter_intent_policy():
362 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
363 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
364 * res->lr_namespace->ns_lvbo->lvbo_update) {
365 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
368 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
369 * returns correct file size to the client.
371 return -ELDLM_NO_LOCK_DATA;
374 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
375 const struct ldlm_res_id *res_id,
376 ldlm_type_t type, ldlm_policy_data_t *policy,
377 ldlm_mode_t mode, int *flags,
378 ldlm_blocking_callback blocking,
379 ldlm_completion_callback completion,
380 ldlm_glimpse_callback glimpse,
381 void *data, __u32 lvb_len,
382 const __u64 *client_cookie,
383 struct lustre_handle *lockh)
385 struct ldlm_lock *lock;
387 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
388 .lcs_blocking = blocking,
389 .lcs_glimpse = glimpse,
393 LASSERT(!(*flags & LDLM_FL_REPLAY));
394 if (unlikely(ns_is_client(ns))) {
395 CERROR("Trying to enqueue local lock in a shadow namespace\n");
399 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len);
401 GOTO(out_nolock, err = -ENOMEM);
403 ldlm_lock_addref_internal(lock, mode);
404 ldlm_lock2handle(lock, lockh);
405 lock_res_and_lock(lock);
406 lock->l_flags |= LDLM_FL_LOCAL;
407 if (*flags & LDLM_FL_ATOMIC_CB)
408 lock->l_flags |= LDLM_FL_ATOMIC_CB;
409 unlock_res_and_lock(lock);
411 lock->l_policy_data = *policy;
412 if (client_cookie != NULL)
413 lock->l_client_cookie = *client_cookie;
414 if (type == LDLM_EXTENT)
415 lock->l_req_extent = policy->l_extent;
417 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
418 if (unlikely(err != ELDLM_OK))
422 *policy = lock->l_policy_data;
424 if (lock->l_completion_ast)
425 lock->l_completion_ast(lock, *flags, NULL);
427 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
430 LDLM_LOCK_RELEASE(lock);
435 static void failed_lock_cleanup(struct ldlm_namespace *ns,
436 struct ldlm_lock *lock,
437 struct lustre_handle *lockh, int mode)
441 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
442 lock_res_and_lock(lock);
443 /* Check that lock is not granted or failed, we might race. */
444 if ((lock->l_req_mode != lock->l_granted_mode) &&
445 !(lock->l_flags & LDLM_FL_FAILED)) {
446 /* Make sure that this lock will not be found by raced
447 * bl_ast and -EINVAL reply is sent to server anyways.
449 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
453 unlock_res_and_lock(lock);
457 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
458 "LDLM_FL_ATOMIC_CB");
459 ldlm_lock_decref_and_cancel(lockh, mode);
461 LDLM_DEBUG(lock, "lock was granted or failed in race");
462 ldlm_lock_decref(lockh, mode);
465 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
466 * from llite/file.c/ll_file_flock(). */
467 if (lock->l_resource->lr_type == LDLM_FLOCK) {
468 ldlm_lock_destroy(lock);
472 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
473 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
474 int *flags, void *lvb, __u32 lvb_len,
475 struct lustre_handle *lockh,int rc)
477 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
478 int is_replay = *flags & LDLM_FL_REPLAY;
479 struct lustre_handle old_hash_key;
480 struct ldlm_lock *lock;
481 struct ldlm_reply *reply;
482 int cleanup_phase = 1;
485 lock = ldlm_handle2lock(lockh);
486 /* ldlm_cli_enqueue is holding a reference on this lock. */
488 LASSERT(type == LDLM_FLOCK);
492 if (rc != ELDLM_OK) {
494 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
495 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
496 if (rc == ELDLM_LOCK_ABORTED) {
497 /* Before we return, swab the reply */
498 reply = req_capsule_server_get(&req->rq_pill,
503 struct ost_lvb *tmplvb;
505 req_capsule_set_size(&req->rq_pill,
506 &RMF_DLM_LVB, RCL_SERVER,
508 tmplvb = req_capsule_server_get(&req->rq_pill,
511 GOTO(cleanup, rc = -EPROTO);
513 memcpy(lvb, tmplvb, lvb_len);
519 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
521 GOTO(cleanup, rc = -EPROTO);
523 /* lock enqueued on the server */
526 lock_res_and_lock(lock);
527 old_hash_key = lock->l_remote_handle;
528 lock->l_remote_handle = reply->lock_handle;
530 /* Key change rehash lock in per-export hash with new key */
531 if (exp->exp_lock_hash)
532 cfs_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
533 &lock->l_remote_handle,
536 *flags = reply->lock_flags;
537 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
538 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
539 * to wait with no timeout as well */
540 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
541 unlock_res_and_lock(lock);
543 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
544 lock, reply->lock_handle.cookie, *flags);
546 /* If enqueue returned a blocked lock but the completion handler has
547 * already run, then it fixed up the resource and we don't need to do it
549 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
550 int newmode = reply->lock_desc.l_req_mode;
552 if (newmode && newmode != lock->l_req_mode) {
553 LDLM_DEBUG(lock, "server returned different mode %s",
554 ldlm_lockname[newmode]);
555 lock->l_req_mode = newmode;
558 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
559 lock->l_resource->lr_name.name,
560 sizeof(struct ldlm_res_id))) {
561 CDEBUG(D_INFO, "remote intent success, locking "
562 "(%ld,%ld,%ld) instead of "
564 (long)reply->lock_desc.l_resource.lr_name.name[0],
565 (long)reply->lock_desc.l_resource.lr_name.name[1],
566 (long)reply->lock_desc.l_resource.lr_name.name[2],
567 (long)lock->l_resource->lr_name.name[0],
568 (long)lock->l_resource->lr_name.name[1],
569 (long)lock->l_resource->lr_name.name[2]);
571 rc = ldlm_lock_change_resource(ns, lock,
572 &reply->lock_desc.l_resource.lr_name);
573 if (rc || lock->l_resource == NULL)
574 GOTO(cleanup, rc = -ENOMEM);
575 LDLM_DEBUG(lock, "client-side enqueue, new resource");
578 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
580 lock->l_policy_data =
581 reply->lock_desc.l_policy_data;
582 if (type != LDLM_PLAIN)
583 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
586 if ((*flags) & LDLM_FL_AST_SENT ||
587 /* Cancel extent locks as soon as possible on a liblustre client,
588 * because it cannot handle asynchronous ASTs robustly (see
590 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
591 lock_res_and_lock(lock);
592 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
593 unlock_res_and_lock(lock);
594 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
597 /* If the lock has already been granted by a completion AST, don't
598 * clobber the LVB with an older one. */
599 if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
602 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
604 tmplvb = req_capsule_server_get(&req->rq_pill,
607 GOTO(cleanup, rc = -EPROTO);
608 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
612 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
613 if (lock->l_completion_ast != NULL) {
614 int err = lock->l_completion_ast(lock, *flags, NULL);
617 if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
622 if (lvb_len && lvb != NULL) {
623 /* Copy the LVB here, and not earlier, because the completion
624 * AST (if any) can override what we got in the reply */
625 memcpy(lvb, lock->l_lvb_data, lvb_len);
628 LDLM_DEBUG(lock, "client-side enqueue END");
631 if (cleanup_phase == 1 && rc)
632 failed_lock_cleanup(ns, lock, lockh, mode);
633 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
635 LDLM_LOCK_RELEASE(lock);
639 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
640 * a single page on the send/receive side. XXX: 512 should be changed
641 * to more adequate value. */
642 static inline int ldlm_req_handles_avail(int req_size, int off)
646 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
647 if (likely(avail >= 0))
648 avail /= (int)sizeof(struct lustre_handle);
651 avail += LDLM_LOCKREQ_HANDLES - off;
656 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
657 enum req_location loc,
660 int size = req_capsule_msg_size(pill, loc);
661 return ldlm_req_handles_avail(size, off);
664 static inline int ldlm_format_handles_avail(struct obd_import *imp,
665 const struct req_format *fmt,
666 enum req_location loc, int off)
668 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
669 return ldlm_req_handles_avail(size, off);
672 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
673 * @count locks in @cancels. */
674 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
675 int version, int opc, int canceloff,
676 cfs_list_t *cancels, int count)
678 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
679 struct req_capsule *pill = &req->rq_pill;
680 struct ldlm_request *dlm = NULL;
681 int flags, avail, to_free, bufcount, pack = 0;
688 if (exp_connect_cancelset(exp)) {
689 /* Estimate the amount of available space in the request. */
690 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
691 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
693 flags = ns_connect_lru_resize(ns) ?
694 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
695 to_free = !ns_connect_lru_resize(ns) &&
696 opc == LDLM_ENQUEUE ? 1 : 0;
698 /* Cancel lru locks here _only_ if the server supports
699 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
700 * rpc, what will make us slower. */
702 count += ldlm_cancel_lru_local(ns, cancels, to_free,
703 avail - count, 0, flags);
708 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
709 ldlm_request_bufsize(pack, opc));
712 rc = ptlrpc_request_pack(req, version, opc);
714 ldlm_lock_list_put(cancels, l_bl_ast, count);
718 if (exp_connect_cancelset(exp)) {
720 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
722 /* Skip first lock handler in ldlm_request_pack(),
723 * this method will incrment @lock_count according
724 * to the lock handle amount actually written to
726 dlm->lock_count = canceloff;
728 /* Pack into the request @pack lock handles. */
729 ldlm_cli_cancel_list(cancels, pack, req, 0);
730 /* Prepare and send separate cancel rpc for others. */
731 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
733 ldlm_lock_list_put(cancels, l_bl_ast, count);
738 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
739 cfs_list_t *cancels, int count)
741 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
742 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
745 /* If a request has some specific initialisation it is passed in @reqp,
746 * otherwise it is created in ldlm_cli_enqueue.
748 * Supports sync and async requests, pass @async flag accordingly. If a
749 * request was created in ldlm_cli_enqueue and it is the async request,
750 * pass it to the caller in @reqp. */
751 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
752 struct ldlm_enqueue_info *einfo,
753 const struct ldlm_res_id *res_id,
754 ldlm_policy_data_t const *policy, int *flags,
755 void *lvb, __u32 lvb_len, struct lustre_handle *lockh,
758 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
759 struct ldlm_lock *lock;
760 struct ldlm_request *body;
761 int is_replay = *flags & LDLM_FL_REPLAY;
762 int req_passed_in = 1;
764 struct ptlrpc_request *req;
767 LASSERT(exp != NULL);
769 /* If we're replaying this lock, just check some invariants.
770 * If we're creating a new lock, get everything all setup nice. */
772 lock = ldlm_handle2lock_long(lockh, 0);
773 LASSERT(lock != NULL);
774 LDLM_DEBUG(lock, "client-side enqueue START");
775 LASSERT(exp == lock->l_conn_export);
777 const struct ldlm_callback_suite cbs = {
778 .lcs_completion = einfo->ei_cb_cp,
779 .lcs_blocking = einfo->ei_cb_bl,
780 .lcs_glimpse = einfo->ei_cb_gl,
781 .lcs_weigh = einfo->ei_cb_wg
783 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
784 einfo->ei_mode, &cbs, einfo->ei_cbdata,
788 /* for the local lock, add the reference */
789 ldlm_lock_addref_internal(lock, einfo->ei_mode);
790 ldlm_lock2handle(lock, lockh);
791 if (policy != NULL) {
792 /* INODEBITS_INTEROP: If the server does not support
793 * inodebits, we will request a plain lock in the
794 * descriptor (ldlm_lock2desc() below) but use an
795 * inodebits lock internally with both bits set.
797 if (einfo->ei_type == LDLM_IBITS &&
798 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
799 lock->l_policy_data.l_inodebits.bits =
800 MDS_INODELOCK_LOOKUP |
801 MDS_INODELOCK_UPDATE;
803 lock->l_policy_data = *policy;
806 if (einfo->ei_type == LDLM_EXTENT)
807 lock->l_req_extent = policy->l_extent;
808 LDLM_DEBUG(lock, "client-side enqueue START");
811 /* lock not sent to server yet */
813 if (reqp == NULL || *reqp == NULL) {
814 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
819 failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
820 LDLM_LOCK_RELEASE(lock);
830 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
832 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
833 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
836 lock->l_conn_export = exp;
837 lock->l_export = NULL;
838 lock->l_blocking_ast = einfo->ei_cb_bl;
840 /* Dump lock data into the request buffer */
841 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
842 ldlm_lock2desc(lock, &body->lock_desc);
843 body->lock_flags = *flags;
844 body->lock_handle[0] = *lockh;
846 /* Continue as normal. */
847 if (!req_passed_in) {
849 req_capsule_extend(&req->rq_pill,
850 &RQF_LDLM_ENQUEUE_LVB);
851 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
852 RCL_SERVER, lvb_len);
854 ptlrpc_request_set_replen(req);
858 * Liblustre client doesn't get extent locks, except for O_APPEND case
859 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
860 * [i_size, OBD_OBJECT_EOF] lock is taken.
862 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
863 policy->l_extent.end == OBD_OBJECT_EOF));
866 LASSERT(reqp != NULL);
870 LDLM_DEBUG(lock, "sending request");
872 rc = ptlrpc_queue_wait(req);
874 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
875 einfo->ei_mode, flags, lvb, lvb_len,
878 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
879 * one reference that we took */
881 LDLM_LOCK_RELEASE(lock);
885 if (!req_passed_in && req != NULL) {
886 ptlrpc_req_finished(req);
894 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
897 struct ldlm_resource *res;
900 if (ns_is_client(lock->l_resource->lr_namespace)) {
901 CERROR("Trying to cancel local lock\n");
904 LDLM_DEBUG(lock, "client-side local convert");
906 res = ldlm_lock_convert(lock, new_mode, flags);
908 ldlm_reprocess_all(res);
913 LDLM_DEBUG(lock, "client-side local convert handler END");
918 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
919 * conversion of locks which are on the waiting or converting queue */
920 /* Caller of this code is supposed to take care of lock readers/writers
922 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
924 struct ldlm_request *body;
925 struct ldlm_reply *reply;
926 struct ldlm_lock *lock;
927 struct ldlm_resource *res;
928 struct ptlrpc_request *req;
932 lock = ldlm_handle2lock(lockh);
939 if (lock->l_conn_export == NULL)
940 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
942 LDLM_DEBUG(lock, "client-side convert");
944 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
945 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
952 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
953 body->lock_handle[0] = lock->l_remote_handle;
955 body->lock_desc.l_req_mode = new_mode;
956 body->lock_flags = *flags;
959 ptlrpc_request_set_replen(req);
960 rc = ptlrpc_queue_wait(req);
964 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
966 GOTO(out, rc = -EPROTO);
969 GOTO(out, rc = req->rq_status);
971 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
973 ldlm_reprocess_all(res);
974 /* Go to sleep until the lock is granted. */
975 /* FIXME: or cancelled. */
976 if (lock->l_completion_ast) {
977 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
988 ptlrpc_req_finished(req);
992 /* Cancel locks locally.
994 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
995 * LDLM_FL_CANCELING otherwise;
996 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
997 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
999 int rc = LDLM_FL_LOCAL_ONLY;
1002 if (lock->l_conn_export) {
1005 LDLM_DEBUG(lock, "client-side cancel");
1006 /* Set this flag to prevent others from getting new references*/
1007 lock_res_and_lock(lock);
1008 lock->l_flags |= LDLM_FL_CBPENDING;
1009 local_only = (lock->l_flags &
1010 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1011 ldlm_cancel_callback(lock);
1012 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1013 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1014 unlock_res_and_lock(lock);
1017 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1019 rc = LDLM_FL_LOCAL_ONLY;
1021 ldlm_lock_cancel(lock);
1023 if (ns_is_client(lock->l_resource->lr_namespace)) {
1024 LDLM_ERROR(lock, "Trying to cancel local lock");
1027 LDLM_DEBUG(lock, "server-side local cancel");
1028 ldlm_lock_cancel(lock);
1029 ldlm_reprocess_all(lock->l_resource);
1035 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
1036 of the request @req. */
1037 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1038 cfs_list_t *head, int count)
1040 struct ldlm_request *dlm;
1041 struct ldlm_lock *lock;
1042 int max, packed = 0;
1045 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1046 LASSERT(dlm != NULL);
1048 /* Check the room in the request buffer. */
1049 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1050 sizeof(struct ldlm_request);
1051 max /= sizeof(struct lustre_handle);
1052 max += LDLM_LOCKREQ_HANDLES;
1053 LASSERT(max >= dlm->lock_count + count);
1055 /* XXX: it would be better to pack lock handles grouped by resource.
1056 * so that the server cancel would call filter_lvbo_update() less
1058 cfs_list_for_each_entry(lock, head, l_bl_ast) {
1061 LASSERT(lock->l_conn_export);
1062 /* Pack the lock handle to the given request buffer. */
1063 LDLM_DEBUG(lock, "packing");
1064 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1067 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1071 /* Prepare and send a batched cancel rpc, it will include count lock handles
1072 * of locks given in @head. */
1073 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1074 int count, int flags)
1076 struct ptlrpc_request *req = NULL;
1077 struct obd_import *imp;
1082 LASSERT(exp != NULL);
1085 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
1087 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1090 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1091 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1098 imp = class_exp2cliimp(exp);
1099 if (imp == NULL || imp->imp_invalid) {
1101 "skipping cancel on invalid import %p\n", imp);
1105 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1107 GOTO(out, rc = -ENOMEM);
1109 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1110 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1111 ldlm_request_bufsize(count, LDLM_CANCEL));
1113 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1115 ptlrpc_request_free(req);
1118 req->rq_no_resend = 1;
1119 req->rq_no_delay = 1;
1121 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1122 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1123 ptlrpc_at_set_req_timeout(req);
1125 ldlm_cancel_pack(req, cancels, count);
1127 ptlrpc_request_set_replen(req);
1128 if (flags & LDLM_FL_ASYNC) {
1129 ptlrpcd_add_req(req, PSCOPE_OTHER);
1133 rc = ptlrpc_queue_wait(req);
1136 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1137 "out of sync -- not fatal\n",
1138 libcfs_nid2str(req->rq_import->
1139 imp_connection->c_peer.nid));
1141 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1142 req->rq_import_generation == imp->imp_generation) {
1143 ptlrpc_req_finished(req);
1145 } else if (rc != ELDLM_OK) {
1146 CERROR("Got rc %d from cancel RPC: canceling "
1154 ptlrpc_req_finished(req);
1157 return sent ? sent : rc;
1160 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1162 LASSERT(imp != NULL);
1163 return &imp->imp_obd->obd_namespace->ns_pool;
1167 * Update client's obd pool related fields with new SLV and Limit from \a req.
1169 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1171 struct obd_device *obd;
1172 __u64 old_slv, new_slv;
1175 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1176 !imp_connect_lru_resize(req->rq_import)))
1179 * Do nothing for corner cases.
1185 * In some cases RPC may contain slv and limit zeroed out. This is
1186 * the case when server does not support lru resize feature. This is
1187 * also possible in some recovery cases when server side reqs have no
1188 * ref to obd export and thus access to server side namespace is no
1191 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1192 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1193 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1194 "(SLV: "LPU64", Limit: %u)",
1195 lustre_msg_get_slv(req->rq_repmsg),
1196 lustre_msg_get_limit(req->rq_repmsg));
1200 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1201 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1202 obd = req->rq_import->imp_obd;
1205 * Set new SLV and Limit to obd fields to make accessible for pool
1206 * thread. We do not access obd_namespace and pool directly here
1207 * as there is no reliable way to make sure that they are still
1208 * alive in cleanup time. Evil races are possible which may cause
1209 * oops in that time.
1211 cfs_write_lock(&obd->obd_pool_lock);
1212 old_slv = obd->obd_pool_slv;
1213 obd->obd_pool_slv = new_slv;
1214 obd->obd_pool_limit = new_limit;
1215 cfs_write_unlock(&obd->obd_pool_lock);
1219 EXPORT_SYMBOL(ldlm_cli_update_pool);
1221 int ldlm_cli_cancel(struct lustre_handle *lockh)
1223 struct obd_export *exp;
1224 int avail, flags, count = 1, rc = 0;
1225 struct ldlm_namespace *ns;
1226 struct ldlm_lock *lock;
1227 CFS_LIST_HEAD(cancels);
1230 /* concurrent cancels on the same handle can happen */
1231 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1233 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1237 rc = ldlm_cli_cancel_local(lock);
1238 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1239 LDLM_LOCK_RELEASE(lock);
1240 RETURN(rc < 0 ? rc : 0);
1242 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1243 * rpc which goes to canceld portal, so we can cancel other lru locks
1244 * here and send them all as one LDLM_CANCEL rpc. */
1245 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1246 cfs_list_add(&lock->l_bl_ast, &cancels);
1248 exp = lock->l_conn_export;
1249 if (exp_connect_cancelset(exp)) {
1250 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1255 ns = lock->l_resource->lr_namespace;
1256 flags = ns_connect_lru_resize(ns) ?
1257 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1258 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1259 LDLM_FL_BL_AST, flags);
1261 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1265 /* XXX until we will have compound requests and can cut cancels from generic rpc
1266 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1267 static int ldlm_cancel_list(cfs_list_t *cancels, int count, int flags)
1269 CFS_LIST_HEAD(head);
1270 struct ldlm_lock *lock, *next;
1271 int left = 0, bl_ast = 0, rc;
1274 cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1278 if (flags & LDLM_FL_LOCAL_ONLY) {
1279 rc = LDLM_FL_LOCAL_ONLY;
1280 ldlm_lock_cancel(lock);
1282 rc = ldlm_cli_cancel_local(lock);
1284 if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1285 LDLM_DEBUG(lock, "Cancel lock separately");
1286 cfs_list_del_init(&lock->l_bl_ast);
1287 cfs_list_add(&lock->l_bl_ast, &head);
1291 if (rc == LDLM_FL_LOCAL_ONLY) {
1292 /* CANCEL RPC should not be sent to server. */
1293 cfs_list_del_init(&lock->l_bl_ast);
1294 LDLM_LOCK_RELEASE(lock);
1301 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1308 * Callback function for lru-resize policy. Makes decision whether to keep
1309 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1310 * \a added and number of locks to be preferably canceled \a count.
1312 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1314 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1316 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1317 struct ldlm_lock *lock,
1318 int unused, int added,
1321 cfs_time_t cur = cfs_time_current();
1322 struct ldlm_pool *pl = &ns->ns_pool;
1327 * Stop lru processing when we reached passed @count or checked all
1330 if (count && added >= count)
1331 return LDLM_POLICY_KEEP_LOCK;
1333 slv = ldlm_pool_get_slv(pl);
1334 lvf = ldlm_pool_get_lvf(pl);
1335 la = cfs_duration_sec(cfs_time_sub(cur,
1336 lock->l_last_used));
1339 * Stop when slv is not yet come from server or lv is smaller than
1342 lv = lvf * la * unused;
1345 * Inform pool about current CLV to see it via proc.
1347 ldlm_pool_set_clv(pl, lv);
1348 return (slv == 0 || lv < slv) ?
1349 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1353 * Callback function for proc used policy. Makes decision whether to keep
1354 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1355 * \a added and number of locks to be preferably canceled \a count.
1357 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1359 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1361 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1362 struct ldlm_lock *lock,
1363 int unused, int added,
1367 * Stop lru processing when we reached passed @count or checked all
1370 return (added >= count) ?
1371 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1375 * Callback function for aged policy. Makes decision whether to keep
1376 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1377 * \a added and number of locks to be preferably canceled \a count.
1379 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1381 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1383 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1384 struct ldlm_lock *lock,
1385 int unused, int added,
1389 * Stop lru processing if young lock is found and we reached passed
1392 return ((added >= count) &&
1393 cfs_time_before(cfs_time_current(),
1394 cfs_time_add(lock->l_last_used,
1396 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1400 * Callback function for default policy. Makes decision whether to keep
1401 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1402 * \a added and number of locks to be preferably canceled \a count.
1404 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1406 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1408 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1409 struct ldlm_lock *lock,
1410 int unused, int added,
1414 * Stop lru processing when we reached passed @count or checked all
1417 return (added >= count) ?
1418 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1421 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1422 struct ldlm_lock *, int,
1425 static ldlm_cancel_lru_policy_t
1426 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1428 if (ns_connect_lru_resize(ns)) {
1429 if (flags & LDLM_CANCEL_SHRINK)
1430 /* We kill passed number of old locks. */
1431 return ldlm_cancel_passed_policy;
1432 else if (flags & LDLM_CANCEL_LRUR)
1433 return ldlm_cancel_lrur_policy;
1434 else if (flags & LDLM_CANCEL_PASSED)
1435 return ldlm_cancel_passed_policy;
1437 if (flags & LDLM_CANCEL_AGED)
1438 return ldlm_cancel_aged_policy;
1441 return ldlm_cancel_default_policy;
1444 /* - Free space in lru for @count new locks,
1445 * redundant unused locks are canceled locally;
1446 * - also cancel locally unused aged locks;
1447 * - do not cancel more than @max locks;
1448 * - GET the found locks and add them into the @cancels list.
1450 * A client lock can be added to the l_bl_ast list only when it is
1451 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1452 * There are the following use cases: ldlm_cancel_resource_local(),
1453 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1454 * flag properly. As any attempt to cancel a lock rely on this flag,
1455 * l_bl_ast list is accessed later without any special locking.
1457 * Calling policies for enabled lru resize:
1458 * ----------------------------------------
1459 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1460 * cancel not more than @count locks;
1462 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1463 * the beginning of lru list);
1465 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1466 * memory pressre policy function;
1468 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1470 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1471 int count, int max, int cancel_flags, int flags)
1473 ldlm_cancel_lru_policy_t pf;
1474 struct ldlm_lock *lock, *next;
1475 int added = 0, unused;
1478 cfs_spin_lock(&ns->ns_unused_lock);
1479 unused = ns->ns_nr_unused;
1481 if (!ns_connect_lru_resize(ns))
1482 count += unused - ns->ns_max_unused;
1484 pf = ldlm_cancel_lru_policy(ns, flags);
1485 LASSERT(pf != NULL);
1487 while (!cfs_list_empty(&ns->ns_unused_list)) {
1488 /* For any flags, stop scanning if @max is reached. */
1489 if (max && added >= max)
1492 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1494 /* No locks which got blocking requests. */
1495 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1497 /* Somebody is already doing CANCEL. No need in this
1498 * lock in lru, do not traverse it again. */
1499 if (!(lock->l_flags & LDLM_FL_CANCELING))
1502 ldlm_lock_remove_from_lru_nolock(lock);
1504 if (&lock->l_lru == &ns->ns_unused_list)
1507 LDLM_LOCK_GET(lock);
1508 cfs_spin_unlock(&ns->ns_unused_lock);
1509 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1511 /* Pass the lock through the policy filter and see if it
1512 * should stay in lru.
1514 * Even for shrinker policy we stop scanning if
1515 * we find a lock that should stay in the cache.
1516 * We should take into account lock age anyway
1517 * as new lock even if it is small of weight is
1518 * valuable resource.
1520 * That is, for shrinker policy we drop only
1521 * old locks, but additionally chose them by
1522 * their weight. Big extent locks will stay in
1524 if (pf(ns, lock, unused, added, count) ==
1525 LDLM_POLICY_KEEP_LOCK) {
1526 lu_ref_del(&lock->l_reference,
1527 __FUNCTION__, cfs_current());
1528 LDLM_LOCK_RELEASE(lock);
1529 cfs_spin_lock(&ns->ns_unused_lock);
1533 lock_res_and_lock(lock);
1534 /* Check flags again under the lock. */
1535 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1536 (ldlm_lock_remove_from_lru(lock) == 0)) {
1537 /* other thread is removing lock from lru or
1538 * somebody is already doing CANCEL or
1539 * there is a blocking request which will send
1540 * cancel by itseft or the lock is matched
1541 * is already not unused. */
1542 unlock_res_and_lock(lock);
1543 lu_ref_del(&lock->l_reference,
1544 __FUNCTION__, cfs_current());
1545 LDLM_LOCK_RELEASE(lock);
1546 cfs_spin_lock(&ns->ns_unused_lock);
1549 LASSERT(!lock->l_readers && !lock->l_writers);
1551 /* If we have chosen to cancel this lock voluntarily, we
1552 * better send cancel notification to server, so that it
1553 * frees appropriate state. This might lead to a race
1554 * where while we are doing cancel here, server is also
1555 * silently cancelling this lock. */
1556 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1558 /* Setting the CBPENDING flag is a little misleading,
1559 * but prevents an important race; namely, once
1560 * CBPENDING is set, the lock can accumulate no more
1561 * readers/writers. Since readers and writers are
1562 * already zero here, ldlm_lock_decref() won't see
1563 * this flag and call l_blocking_ast */
1564 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1566 /* We can't re-add to l_lru as it confuses the
1567 * refcounting in ldlm_lock_remove_from_lru() if an AST
1568 * arrives after we drop ns_lock below. We use l_bl_ast
1569 * and can't use l_pending_chain as it is used both on
1570 * server and client nevertheless bug 5666 says it is
1571 * used only on server */
1572 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1573 cfs_list_add(&lock->l_bl_ast, cancels);
1574 unlock_res_and_lock(lock);
1575 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1576 cfs_spin_lock(&ns->ns_unused_lock);
1580 cfs_spin_unlock(&ns->ns_unused_lock);
1581 RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
1584 /* when called with LDLM_ASYNC the blocking callback will be handled
1585 * in a thread and this function will return after the thread has been
1586 * asked to call the callback. when called with LDLM_SYNC the blocking
1587 * callback will be performed in this function. */
1588 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1591 CFS_LIST_HEAD(cancels);
1596 sync = LDLM_SYNC; /* force to be sync in user space */
1598 count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
1599 if (sync == LDLM_ASYNC) {
1600 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1605 /* If an error occured in ASYNC mode, or this is SYNC mode,
1606 * cancel the list. */
1607 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1611 /* Find and cancel locally unused locks found on resource, matched to the
1612 * given policy, mode. GET the found locks and add them into the @cancels
1614 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1615 cfs_list_t *cancels,
1616 ldlm_policy_data_t *policy,
1617 ldlm_mode_t mode, int lock_flags,
1618 int cancel_flags, void *opaque)
1620 struct ldlm_lock *lock;
1625 cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1626 if (opaque != NULL && lock->l_ast_data != opaque) {
1627 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1628 lock->l_ast_data, opaque);
1633 if (lock->l_readers || lock->l_writers) {
1634 if (cancel_flags & LDLM_FL_WARN) {
1635 LDLM_ERROR(lock, "lock in use");
1641 /* If somebody is already doing CANCEL, or blocking ast came,
1642 * skip this lock. */
1643 if (lock->l_flags & LDLM_FL_BL_AST ||
1644 lock->l_flags & LDLM_FL_CANCELING)
1647 if (lockmode_compat(lock->l_granted_mode, mode))
1650 /* If policy is given and this is IBITS lock, add to list only
1651 * those locks that match by policy. */
1652 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1653 !(lock->l_policy_data.l_inodebits.bits &
1654 policy->l_inodebits.bits))
1657 /* See CBPENDING comment in ldlm_cancel_lru */
1658 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1661 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1662 cfs_list_add(&lock->l_bl_ast, cancels);
1663 LDLM_LOCK_GET(lock);
1668 RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
1671 /* If @req is NULL, send CANCEL request to server with handles of locks
1672 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1673 * separately per lock.
1674 * If @req is not NULL, put handles of locks in @cancels into the request
1675 * buffer at the offset @off.
1676 * Destroy @cancels at the end. */
1677 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1678 struct ptlrpc_request *req, int flags)
1680 struct ldlm_lock *lock;
1684 if (cfs_list_empty(cancels) || count == 0)
1687 /* XXX: requests (both batched and not) could be sent in parallel.
1688 * Usually it is enough to have just 1 RPC, but it is possible that
1689 * there are to many locks to be cancelled in LRU or on a resource.
1690 * It would also speed up the case when the server does not support
1693 LASSERT(!cfs_list_empty(cancels));
1694 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1696 LASSERT(lock->l_conn_export);
1698 if (exp_connect_cancelset(lock->l_conn_export)) {
1701 ldlm_cancel_pack(req, cancels, count);
1703 res = ldlm_cli_cancel_req(lock->l_conn_export,
1707 res = ldlm_cli_cancel_req(lock->l_conn_export,
1712 CERROR("ldlm_cli_cancel_list: %d\n", res);
1717 ldlm_lock_list_put(cancels, l_bl_ast, res);
1719 LASSERT(count == 0);
1723 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1724 const struct ldlm_res_id *res_id,
1725 ldlm_policy_data_t *policy,
1726 ldlm_mode_t mode, int flags, void *opaque)
1728 struct ldlm_resource *res;
1729 CFS_LIST_HEAD(cancels);
1734 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1736 /* This is not a problem. */
1737 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1741 LDLM_RESOURCE_ADDREF(res);
1742 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1744 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1746 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1748 LDLM_RESOURCE_DELREF(res);
1749 ldlm_resource_putref(res);
1753 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1755 int no_resource = 0;
1757 cfs_spin_lock(&ns->ns_hash_lock);
1758 if (ns->ns_resources == 0)
1760 cfs_spin_unlock(&ns->ns_hash_lock);
1762 RETURN(no_resource);
1765 /* Cancel all locks on a namespace (or a specific resource, if given)
1766 * that have 0 readers/writers.
1768 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1769 * to notify the server. */
1770 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1771 const struct ldlm_res_id *res_id,
1772 int flags, void *opaque)
1781 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1785 cfs_spin_lock(&ns->ns_hash_lock);
1786 for (i = 0; i < RES_HASH_SIZE; i++) {
1788 tmp = ns->ns_hash[i].next;
1789 while (tmp != &(ns->ns_hash[i])) {
1790 struct ldlm_resource *res;
1793 res = cfs_list_entry(tmp, struct ldlm_resource,
1795 ldlm_resource_getref(res);
1796 cfs_spin_unlock(&ns->ns_hash_lock);
1798 LDLM_RESOURCE_ADDREF(res);
1799 rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1804 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1805 res->lr_name.name[0], rc);
1807 LDLM_RESOURCE_DELREF(res);
1808 cfs_spin_lock(&ns->ns_hash_lock);
1810 ldlm_resource_putref_locked(res);
1813 cfs_spin_unlock(&ns->ns_hash_lock);
1818 /* Lock iterators. */
1820 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1823 cfs_list_t *tmp, *next;
1824 struct ldlm_lock *lock;
1825 int rc = LDLM_ITER_CONTINUE;
1830 RETURN(LDLM_ITER_CONTINUE);
1833 cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
1834 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1836 if (iter(lock, closure) == LDLM_ITER_STOP)
1837 GOTO(out, rc = LDLM_ITER_STOP);
1840 cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
1841 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1843 if (iter(lock, closure) == LDLM_ITER_STOP)
1844 GOTO(out, rc = LDLM_ITER_STOP);
1847 cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
1848 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1850 if (iter(lock, closure) == LDLM_ITER_STOP)
1851 GOTO(out, rc = LDLM_ITER_STOP);
1858 struct iter_helper_data {
1859 ldlm_iterator_t iter;
1863 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1865 struct iter_helper_data *helper = closure;
1866 return helper->iter(lock, helper->closure);
1869 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1871 return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1874 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1877 struct iter_helper_data helper = { iter: iter, closure: closure };
1878 return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1881 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1882 ldlm_res_iterator_t iter, void *closure)
1884 int i, rc = LDLM_ITER_CONTINUE;
1885 struct ldlm_resource *res;
1889 cfs_spin_lock(&ns->ns_hash_lock);
1890 for (i = 0; i < RES_HASH_SIZE; i++) {
1891 tmp = ns->ns_hash[i].next;
1892 while (tmp != &(ns->ns_hash[i])) {
1893 res = cfs_list_entry(tmp, struct ldlm_resource,
1895 ldlm_resource_getref(res);
1896 cfs_spin_unlock(&ns->ns_hash_lock);
1897 LDLM_RESOURCE_ADDREF(res);
1899 rc = iter(res, closure);
1901 LDLM_RESOURCE_DELREF(res);
1902 cfs_spin_lock(&ns->ns_hash_lock);
1904 ldlm_resource_putref_locked(res);
1905 if (rc == LDLM_ITER_STOP)
1910 cfs_spin_unlock(&ns->ns_hash_lock);
1914 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1915 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1916 const struct ldlm_res_id *res_id,
1917 ldlm_iterator_t iter, void *data)
1919 struct ldlm_resource *res;
1923 CERROR("must pass in namespace\n");
1927 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1933 LDLM_RESOURCE_ADDREF(res);
1934 ldlm_resource_foreach(res, iter, data);
1935 LDLM_RESOURCE_DELREF(res);
1936 ldlm_resource_putref(res);
1942 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1944 cfs_list_t *list = closure;
1946 /* we use l_pending_chain here, because it's unused on clients. */
1947 LASSERTF(cfs_list_empty(&lock->l_pending_chain),
1948 "lock %p next %p prev %p\n",
1949 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1950 /* bug 9573: don't replay locks left after eviction, or
1951 * bug 17614: locks being actively cancelled. Get a reference
1952 * on a lock so that it does not disapear under us (e.g. due to cancel)
1954 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
1955 cfs_list_add(&lock->l_pending_chain, list);
1956 LDLM_LOCK_GET(lock);
1959 return LDLM_ITER_CONTINUE;
1962 static int replay_lock_interpret(const struct lu_env *env,
1963 struct ptlrpc_request *req,
1964 struct ldlm_async_args *aa, int rc)
1966 struct lustre_handle old_hash_key;
1967 struct ldlm_lock *lock;
1968 struct ldlm_reply *reply;
1969 struct obd_export *exp;
1972 cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
1977 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1979 GOTO(out, rc = -EPROTO);
1981 lock = ldlm_handle2lock(&aa->lock_handle);
1983 CERROR("received replay ack for unknown local cookie "LPX64
1984 " remote cookie "LPX64 " from server %s id %s\n",
1985 aa->lock_handle.cookie, reply->lock_handle.cookie,
1986 req->rq_export->exp_client_uuid.uuid,
1987 libcfs_id2str(req->rq_peer));
1988 GOTO(out, rc = -ESTALE);
1991 old_hash_key = lock->l_remote_handle;
1992 lock->l_remote_handle = reply->lock_handle;
1994 /* Key change rehash lock in per-export hash with new key */
1995 exp = req->rq_export;
1996 if (exp && exp->exp_lock_hash)
1997 cfs_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
1998 &lock->l_remote_handle,
2001 LDLM_DEBUG(lock, "replayed lock:");
2002 ptlrpc_import_recovery_state_machine(req->rq_import);
2003 LDLM_LOCK_PUT(lock);
2006 ptlrpc_connect_import(req->rq_import, NULL);
2011 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2013 struct ptlrpc_request *req;
2014 struct ldlm_async_args *aa;
2015 struct ldlm_request *body;
2020 /* Bug 11974: Do not replay a lock which is actively being canceled */
2021 if (lock->l_flags & LDLM_FL_CANCELING) {
2022 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2026 /* If this is reply-less callback lock, we cannot replay it, since
2027 * server might have long dropped it, but notification of that event was
2028 * lost by network. (and server granted conflicting lock already) */
2029 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2030 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2031 ldlm_lock_cancel(lock);
2035 * If granted mode matches the requested mode, this lock is granted.
2037 * If they differ, but we have a granted mode, then we were granted
2038 * one mode and now want another: ergo, converting.
2040 * If we haven't been granted anything and are on a resource list,
2041 * then we're blocked/waiting.
2043 * If we haven't been granted anything and we're NOT on a resource list,
2044 * then we haven't got a reply yet and don't have a known disposition.
2045 * This happens whenever a lock enqueue is the request that triggers
2048 if (lock->l_granted_mode == lock->l_req_mode)
2049 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2050 else if (lock->l_granted_mode)
2051 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2052 else if (!cfs_list_empty(&lock->l_res_link))
2053 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2055 flags = LDLM_FL_REPLAY;
2057 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2058 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2062 /* We're part of recovery, so don't wait for it. */
2063 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2065 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2066 ldlm_lock2desc(lock, &body->lock_desc);
2067 body->lock_flags = flags;
2069 ldlm_lock2handle(lock, &body->lock_handle[0]);
2070 if (lock->l_lvb_len != 0) {
2071 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2072 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2075 ptlrpc_request_set_replen(req);
2076 /* notify the server we've replayed all requests.
2077 * also, we mark the request to be put on a dedicated
2078 * queue to be processed after all request replayes.
2080 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2082 LDLM_DEBUG(lock, "replaying lock:");
2084 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2085 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2086 aa = ptlrpc_req_async_args(req);
2087 aa->lock_handle = body->lock_handle[0];
2088 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2089 ptlrpcd_add_req(req, PSCOPE_OTHER);
2094 int ldlm_replay_locks(struct obd_import *imp)
2096 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2097 CFS_LIST_HEAD(list);
2098 struct ldlm_lock *lock, *next;
2103 LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
2105 /* ensure this doesn't fall to 0 before all have been queued */
2106 cfs_atomic_inc(&imp->imp_replay_inflight);
2108 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2110 cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2111 cfs_list_del_init(&lock->l_pending_chain);
2113 LDLM_LOCK_PUT(lock);
2114 continue; /* or try to do the rest? */
2116 rc = replay_one_lock(imp, lock);
2117 LDLM_LOCK_PUT(lock);
2120 cfs_atomic_dec(&imp->imp_replay_inflight);