1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_LDLM
40 #include <liblustre.h>
43 #include <lustre_dlm.h>
44 #include <obd_class.h>
47 #include "ldlm_internal.h"
49 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
50 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
51 "lock enqueue timeout minimum");
53 static void interrupted_completion_wait(void *data)
57 struct lock_wait_data {
58 struct ldlm_lock *lwd_lock;
62 struct ldlm_async_args {
63 struct lustre_handle lock_handle;
66 int ldlm_expired_completion_wait(void *data)
68 struct lock_wait_data *lwd = data;
69 struct ldlm_lock *lock = lwd->lwd_lock;
70 struct obd_import *imp;
71 struct obd_device *obd;
74 if (lock->l_conn_export == NULL) {
75 static cfs_time_t next_dump = 0, last_dump = 0;
77 if (ptlrpc_check_suspend())
80 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
81 CFS_DURATION_T"s ago); not entering recovery in "
82 "server code, just going back to sleep",
83 lock->l_enqueued_time.tv_sec,
84 cfs_time_sub(cfs_time_current_sec(),
85 lock->l_enqueued_time.tv_sec));
86 if (cfs_time_after(cfs_time_current(), next_dump)) {
87 last_dump = next_dump;
88 next_dump = cfs_time_shift(300);
89 ldlm_namespace_dump(D_DLMTRACE,
90 lock->l_resource->lr_namespace);
92 libcfs_debug_dumplog();
97 obd = lock->l_conn_export->exp_obd;
98 imp = obd->u.cli.cl_import;
99 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
100 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
101 CFS_DURATION_T"s ago), entering recovery for %s@%s",
102 lock->l_enqueued_time.tv_sec,
103 cfs_time_sub(cfs_time_current_sec(),
104 lock->l_enqueued_time.tv_sec),
105 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
110 /* We use the same basis for both server side and client side functions
111 from a single node. */
112 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
114 int timeout = at_get(&lock->l_resource->lr_namespace->ns_at_estimate);
116 return obd_timeout / 2;
117 /* Since these are non-updating timeouts, we should be conservative.
118 It would be nice to have some kind of "early reply" mechanism for
119 lock callbacks too... */
120 timeout = timeout + (timeout >> 1); /* 150% */
121 return max(timeout, ldlm_enqueue_min);
123 EXPORT_SYMBOL(ldlm_get_enq_timeout);
126 * Helper function for ldlm_completion_ast(), updating timings when lock is
129 static int ldlm_completion_tail(struct ldlm_lock *lock)
134 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
135 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
138 delay = cfs_time_sub(cfs_time_current_sec(),
139 lock->l_enqueued_time.tv_sec);
140 LDLM_DEBUG(lock, "client-side enqueue: granted after "
141 CFS_DURATION_T"s", delay);
143 /* Update our time estimate */
144 at_add(&lock->l_resource->lr_namespace->ns_at_estimate, delay);
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 spin_lock(&imp->imp_lock);
251 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
252 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, void *lvb_swabber,
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);
402 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
404 ldlm_lock_addref_internal(lock, mode);
405 ldlm_lock2handle(lock, lockh);
406 lock_res_and_lock(lock);
407 lock->l_flags |= LDLM_FL_LOCAL;
408 if (*flags & LDLM_FL_ATOMIC_CB)
409 lock->l_flags |= LDLM_FL_ATOMIC_CB;
410 lock->l_lvb_swabber = lvb_swabber;
411 unlock_res_and_lock(lock);
413 lock->l_policy_data = *policy;
414 if (client_cookie != NULL)
415 lock->l_client_cookie = *client_cookie;
416 if (type == LDLM_EXTENT)
417 lock->l_req_extent = policy->l_extent;
419 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
420 if (unlikely(err != ELDLM_OK))
424 *policy = lock->l_policy_data;
426 LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
429 if (lock->l_completion_ast)
430 lock->l_completion_ast(lock, *flags, NULL);
432 LDLM_DEBUG(lock, "client-side local enqueue END");
435 LDLM_LOCK_RELEASE(lock);
440 static void failed_lock_cleanup(struct ldlm_namespace *ns,
441 struct ldlm_lock *lock,
442 struct lustre_handle *lockh, int mode)
446 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
447 lock_res_and_lock(lock);
448 /* Check that lock is not granted or failed, we might race. */
449 if ((lock->l_req_mode != lock->l_granted_mode) &&
450 !(lock->l_flags & LDLM_FL_FAILED)) {
451 /* Make sure that this lock will not be found by raced
452 * bl_ast and -EINVAL reply is sent to server anyways.
454 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
458 unlock_res_and_lock(lock);
462 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
463 "LDLM_FL_ATOMIC_CB");
464 ldlm_lock_decref_and_cancel(lockh, mode);
466 LDLM_DEBUG(lock, "lock was granted or failed in race");
467 ldlm_lock_decref(lockh, mode);
470 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
471 * from llite/file.c/ll_file_flock(). */
472 if (lock->l_resource->lr_type == LDLM_FLOCK) {
473 ldlm_lock_destroy(lock);
477 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
478 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
479 int *flags, void *lvb, __u32 lvb_len,
480 void *lvb_swabber, struct lustre_handle *lockh,int rc)
482 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
483 int is_replay = *flags & LDLM_FL_REPLAY;
484 struct lustre_handle old_hash_key;
485 struct ldlm_lock *lock;
486 struct ldlm_reply *reply;
487 int cleanup_phase = 1;
490 lock = ldlm_handle2lock(lockh);
491 /* ldlm_cli_enqueue is holding a reference on this lock. */
493 LASSERT(type == LDLM_FLOCK);
497 if (rc != ELDLM_OK) {
499 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
500 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
501 if (rc == ELDLM_LOCK_ABORTED) {
502 /* Before we return, swab the reply */
503 reply = req_capsule_server_get(&req->rq_pill,
508 struct ost_lvb *tmplvb;
510 req_capsule_set_size(&req->rq_pill,
511 &RMF_DLM_LVB, RCL_SERVER,
513 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
517 GOTO(cleanup, rc = -EPROTO);
519 memcpy(lvb, tmplvb, lvb_len);
525 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
527 GOTO(cleanup, rc = -EPROTO);
529 /* lock enqueued on the server */
532 lock_res_and_lock(lock);
533 old_hash_key = lock->l_remote_handle;
534 lock->l_remote_handle = reply->lock_handle;
536 /* Key change rehash lock in per-export hash with new key */
537 if (exp->exp_lock_hash)
538 lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
539 &lock->l_remote_handle,
542 *flags = reply->lock_flags;
543 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
544 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
545 * to wait with no timeout as well */
546 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
547 unlock_res_and_lock(lock);
549 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
550 lock, reply->lock_handle.cookie, *flags);
552 /* If enqueue returned a blocked lock but the completion handler has
553 * already run, then it fixed up the resource and we don't need to do it
555 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
556 int newmode = reply->lock_desc.l_req_mode;
558 if (newmode && newmode != lock->l_req_mode) {
559 LDLM_DEBUG(lock, "server returned different mode %s",
560 ldlm_lockname[newmode]);
561 lock->l_req_mode = newmode;
564 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
565 lock->l_resource->lr_name.name,
566 sizeof(struct ldlm_res_id))) {
567 CDEBUG(D_INFO, "remote intent success, locking "
568 "(%ld,%ld,%ld) instead of "
570 (long)reply->lock_desc.l_resource.lr_name.name[0],
571 (long)reply->lock_desc.l_resource.lr_name.name[1],
572 (long)reply->lock_desc.l_resource.lr_name.name[2],
573 (long)lock->l_resource->lr_name.name[0],
574 (long)lock->l_resource->lr_name.name[1],
575 (long)lock->l_resource->lr_name.name[2]);
577 rc = ldlm_lock_change_resource(ns, lock,
578 &reply->lock_desc.l_resource.lr_name);
579 if (rc || lock->l_resource == NULL)
580 GOTO(cleanup, rc = -ENOMEM);
581 LDLM_DEBUG(lock, "client-side enqueue, new resource");
584 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
586 lock->l_policy_data =
587 reply->lock_desc.l_policy_data;
588 if (type != LDLM_PLAIN)
589 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
592 if ((*flags) & LDLM_FL_AST_SENT ||
593 /* Cancel extent locks as soon as possible on a liblustre client,
594 * because it cannot handle asynchronous ASTs robustly (see
596 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
597 lock_res_and_lock(lock);
598 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
599 unlock_res_and_lock(lock);
600 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
603 /* If the lock has already been granted by a completion AST, don't
604 * clobber the LVB with an older one. */
605 if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
608 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
610 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
614 GOTO(cleanup, rc = -EPROTO);
615 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
619 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
620 if (lock->l_completion_ast != NULL) {
621 int err = lock->l_completion_ast(lock, *flags, NULL);
624 if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
629 if (lvb_len && lvb != NULL) {
630 /* Copy the LVB here, and not earlier, because the completion
631 * AST (if any) can override what we got in the reply */
632 memcpy(lvb, lock->l_lvb_data, lvb_len);
635 LDLM_DEBUG(lock, "client-side enqueue END");
638 if (cleanup_phase == 1 && rc)
639 failed_lock_cleanup(ns, lock, lockh, mode);
640 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
642 LDLM_LOCK_RELEASE(lock);
646 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
647 * a single page on the send/receive side. XXX: 512 should be changed
648 * to more adequate value. */
649 static inline int ldlm_req_handles_avail(int req_size, int off)
653 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
654 if (likely(avail >= 0))
655 avail /= (int)sizeof(struct lustre_handle);
658 avail += LDLM_LOCKREQ_HANDLES - off;
663 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
664 enum req_location loc,
667 int size = req_capsule_msg_size(pill, loc);
668 return ldlm_req_handles_avail(size, off);
671 static inline int ldlm_format_handles_avail(struct obd_import *imp,
672 const struct req_format *fmt,
673 enum req_location loc, int off)
675 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
676 return ldlm_req_handles_avail(size, off);
679 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
680 * @count locks in @cancels. */
681 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
682 int version, int opc, int canceloff,
683 struct list_head *cancels, int count)
685 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
686 struct req_capsule *pill = &req->rq_pill;
687 struct ldlm_request *dlm = NULL;
688 int flags, avail, to_free, bufcount, pack = 0;
695 if (exp_connect_cancelset(exp)) {
696 /* Estimate the amount of available space in the request. */
697 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
698 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
700 flags = ns_connect_lru_resize(ns) ?
701 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
702 to_free = !ns_connect_lru_resize(ns) &&
703 opc == LDLM_ENQUEUE ? 1 : 0;
705 /* Cancel lru locks here _only_ if the server supports
706 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
707 * rpc, what will make us slower. */
709 count += ldlm_cancel_lru_local(ns, cancels, to_free,
710 avail - count, 0, flags);
715 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
716 ldlm_request_bufsize(pack, opc));
719 rc = ptlrpc_request_pack(req, version, opc);
721 ldlm_lock_list_put(cancels, l_bl_ast, count);
725 if (exp_connect_cancelset(exp)) {
727 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
729 /* Skip first lock handler in ldlm_request_pack(),
730 * this method will incrment @lock_count according
731 * to the lock handle amount actually written to
733 dlm->lock_count = canceloff;
735 /* Pack into the request @pack lock handles. */
736 ldlm_cli_cancel_list(cancels, pack, req, 0);
737 /* Prepare and send separate cancel rpc for others. */
738 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
740 ldlm_lock_list_put(cancels, l_bl_ast, count);
745 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
746 struct list_head *cancels, int count)
748 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
749 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
752 /* If a request has some specific initialisation it is passed in @reqp,
753 * otherwise it is created in ldlm_cli_enqueue.
755 * Supports sync and async requests, pass @async flag accordingly. If a
756 * request was created in ldlm_cli_enqueue and it is the async request,
757 * pass it to the caller in @reqp. */
758 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
759 struct ldlm_enqueue_info *einfo,
760 const struct ldlm_res_id *res_id,
761 ldlm_policy_data_t *policy, int *flags,
762 void *lvb, __u32 lvb_len, void *lvb_swabber,
763 struct lustre_handle *lockh, int async)
765 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
766 struct ldlm_lock *lock;
767 struct ldlm_request *body;
768 int is_replay = *flags & LDLM_FL_REPLAY;
769 int req_passed_in = 1;
771 struct ptlrpc_request *req;
774 LASSERT(exp != NULL);
776 /* If we're replaying this lock, just check some invariants.
777 * If we're creating a new lock, get everything all setup nice. */
779 lock = ldlm_handle2lock_long(lockh, 0);
780 LASSERT(lock != NULL);
781 LDLM_DEBUG(lock, "client-side enqueue START");
782 LASSERT(exp == lock->l_conn_export);
784 const struct ldlm_callback_suite cbs = {
785 .lcs_completion = einfo->ei_cb_cp,
786 .lcs_blocking = einfo->ei_cb_bl,
787 .lcs_glimpse = einfo->ei_cb_gl,
788 .lcs_weigh = einfo->ei_cb_wg
790 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
791 einfo->ei_mode, &cbs, einfo->ei_cbdata,
795 /* for the local lock, add the reference */
796 ldlm_lock_addref_internal(lock, einfo->ei_mode);
797 ldlm_lock2handle(lock, lockh);
798 lock->l_lvb_swabber = lvb_swabber;
799 if (policy != NULL) {
800 /* INODEBITS_INTEROP: If the server does not support
801 * inodebits, we will request a plain lock in the
802 * descriptor (ldlm_lock2desc() below) but use an
803 * inodebits lock internally with both bits set.
805 if (einfo->ei_type == LDLM_IBITS &&
806 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
807 lock->l_policy_data.l_inodebits.bits =
808 MDS_INODELOCK_LOOKUP |
809 MDS_INODELOCK_UPDATE;
811 lock->l_policy_data = *policy;
814 if (einfo->ei_type == LDLM_EXTENT)
815 lock->l_req_extent = policy->l_extent;
816 LDLM_DEBUG(lock, "client-side enqueue START");
819 /* lock not sent to server yet */
821 if (reqp == NULL || *reqp == NULL) {
822 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
827 failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
828 LDLM_LOCK_RELEASE(lock);
838 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
840 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
841 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
844 lock->l_conn_export = exp;
845 lock->l_export = NULL;
846 lock->l_blocking_ast = einfo->ei_cb_bl;
848 /* Dump lock data into the request buffer */
849 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
850 ldlm_lock2desc(lock, &body->lock_desc);
851 body->lock_flags = *flags;
852 body->lock_handle[0] = *lockh;
854 /* Continue as normal. */
855 if (!req_passed_in) {
857 req_capsule_extend(&req->rq_pill,
858 &RQF_LDLM_ENQUEUE_LVB);
859 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
860 RCL_SERVER, lvb_len);
862 ptlrpc_request_set_replen(req);
866 * Liblustre client doesn't get extent locks, except for O_APPEND case
867 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
868 * [i_size, OBD_OBJECT_EOF] lock is taken.
870 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
871 policy->l_extent.end == OBD_OBJECT_EOF));
874 LASSERT(reqp != NULL);
878 LDLM_DEBUG(lock, "sending request");
880 rc = ptlrpc_queue_wait(req);
882 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
883 einfo->ei_mode, flags, lvb, lvb_len,
884 lvb_swabber, lockh, rc);
886 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
887 * one reference that we took */
889 LDLM_LOCK_RELEASE(lock);
893 if (!req_passed_in && req != NULL) {
894 ptlrpc_req_finished(req);
902 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
905 struct ldlm_resource *res;
908 if (ns_is_client(lock->l_resource->lr_namespace)) {
909 CERROR("Trying to cancel local lock\n");
912 LDLM_DEBUG(lock, "client-side local convert");
914 res = ldlm_lock_convert(lock, new_mode, flags);
916 ldlm_reprocess_all(res);
921 LDLM_DEBUG(lock, "client-side local convert handler END");
926 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
927 * conversion of locks which are on the waiting or converting queue */
928 /* Caller of this code is supposed to take care of lock readers/writers
930 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
932 struct ldlm_request *body;
933 struct ldlm_reply *reply;
934 struct ldlm_lock *lock;
935 struct ldlm_resource *res;
936 struct ptlrpc_request *req;
940 lock = ldlm_handle2lock(lockh);
947 if (lock->l_conn_export == NULL)
948 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
950 LDLM_DEBUG(lock, "client-side convert");
952 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
953 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
960 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
961 body->lock_handle[0] = lock->l_remote_handle;
963 body->lock_desc.l_req_mode = new_mode;
964 body->lock_flags = *flags;
967 ptlrpc_request_set_replen(req);
968 rc = ptlrpc_queue_wait(req);
972 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
974 GOTO(out, rc = -EPROTO);
977 GOTO(out, rc = req->rq_status);
979 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
981 ldlm_reprocess_all(res);
982 /* Go to sleep until the lock is granted. */
983 /* FIXME: or cancelled. */
984 if (lock->l_completion_ast) {
985 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
996 ptlrpc_req_finished(req);
1000 /* Cancel locks locally.
1002 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
1003 * LDLM_FL_CANCELING otherwise;
1004 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
1005 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
1007 int rc = LDLM_FL_LOCAL_ONLY;
1010 if (lock->l_conn_export) {
1013 LDLM_DEBUG(lock, "client-side cancel");
1014 /* Set this flag to prevent others from getting new references*/
1015 lock_res_and_lock(lock);
1016 lock->l_flags |= LDLM_FL_CBPENDING;
1017 local_only = (lock->l_flags &
1018 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1019 ldlm_cancel_callback(lock);
1020 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1021 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1022 unlock_res_and_lock(lock);
1025 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1027 rc = LDLM_FL_LOCAL_ONLY;
1029 ldlm_lock_cancel(lock);
1031 if (ns_is_client(lock->l_resource->lr_namespace)) {
1032 LDLM_ERROR(lock, "Trying to cancel local lock");
1035 LDLM_DEBUG(lock, "server-side local cancel");
1036 ldlm_lock_cancel(lock);
1037 ldlm_reprocess_all(lock->l_resource);
1038 LDLM_DEBUG(lock, "server-side local cancel handler END");
1044 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
1045 of the request @req. */
1046 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1047 struct list_head *head, int count)
1049 struct ldlm_request *dlm;
1050 struct ldlm_lock *lock;
1051 int max, packed = 0;
1054 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1055 LASSERT(dlm != NULL);
1057 /* Check the room in the request buffer. */
1058 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1059 sizeof(struct ldlm_request);
1060 max /= sizeof(struct lustre_handle);
1061 max += LDLM_LOCKREQ_HANDLES;
1062 LASSERT(max >= dlm->lock_count + count);
1064 /* XXX: it would be better to pack lock handles grouped by resource.
1065 * so that the server cancel would call filter_lvbo_update() less
1067 list_for_each_entry(lock, head, l_bl_ast) {
1070 LASSERT(lock->l_conn_export);
1071 /* Pack the lock handle to the given request buffer. */
1072 LDLM_DEBUG(lock, "packing");
1073 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1076 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1080 /* Prepare and send a batched cancel rpc, it will include count lock handles
1081 * of locks given in @head. */
1082 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1083 int count, int flags)
1085 struct ptlrpc_request *req = NULL;
1086 struct obd_import *imp;
1091 LASSERT(exp != NULL);
1094 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
1096 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1099 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1100 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1107 imp = class_exp2cliimp(exp);
1108 if (imp == NULL || imp->imp_invalid) {
1110 "skipping cancel on invalid import %p\n", imp);
1114 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1116 GOTO(out, rc = -ENOMEM);
1118 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1119 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1120 ldlm_request_bufsize(count, LDLM_CANCEL));
1122 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1124 ptlrpc_request_free(req);
1127 req->rq_no_resend = 1;
1128 req->rq_no_delay = 1;
1130 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1131 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1132 ptlrpc_at_set_req_timeout(req);
1134 ldlm_cancel_pack(req, cancels, count);
1136 ptlrpc_request_set_replen(req);
1137 if (flags & LDLM_FL_ASYNC) {
1138 ptlrpcd_add_req(req, PSCOPE_OTHER);
1142 rc = ptlrpc_queue_wait(req);
1145 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1146 "out of sync -- not fatal\n",
1147 libcfs_nid2str(req->rq_import->
1148 imp_connection->c_peer.nid));
1150 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1151 req->rq_import_generation == imp->imp_generation) {
1152 ptlrpc_req_finished(req);
1154 } else if (rc != ELDLM_OK) {
1155 CERROR("Got rc %d from cancel RPC: canceling "
1163 ptlrpc_req_finished(req);
1166 return sent ? sent : rc;
1169 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1171 LASSERT(imp != NULL);
1172 return &imp->imp_obd->obd_namespace->ns_pool;
1176 * Update client's obd pool related fields with new SLV and Limit from \a req.
1178 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1180 struct obd_device *obd;
1181 __u64 old_slv, new_slv;
1184 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1185 !imp_connect_lru_resize(req->rq_import)))
1188 * Do nothing for corner cases.
1194 * In some cases RPC may contain slv and limit zeroed out. This is
1195 * the case when server does not support lru resize feature. This is
1196 * also possible in some recovery cases when server side reqs have no
1197 * ref to obd export and thus access to server side namespace is no
1200 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1201 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1202 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1203 "(SLV: "LPU64", Limit: %u)",
1204 lustre_msg_get_slv(req->rq_repmsg),
1205 lustre_msg_get_limit(req->rq_repmsg));
1209 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1210 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1211 obd = req->rq_import->imp_obd;
1214 * Set new SLV and Limit to obd fields to make accessible for pool
1215 * thread. We do not access obd_namespace and pool directly here
1216 * as there is no reliable way to make sure that they are still
1217 * alive in cleanup time. Evil races are possible which may cause
1218 * oops in that time.
1220 write_lock(&obd->obd_pool_lock);
1221 old_slv = obd->obd_pool_slv;
1222 obd->obd_pool_slv = new_slv;
1223 obd->obd_pool_limit = new_limit;
1224 write_unlock(&obd->obd_pool_lock);
1228 EXPORT_SYMBOL(ldlm_cli_update_pool);
1230 int ldlm_cli_cancel(struct lustre_handle *lockh)
1232 struct obd_export *exp;
1233 int avail, flags, count = 1, rc = 0;
1234 struct ldlm_namespace *ns;
1235 struct ldlm_lock *lock;
1236 CFS_LIST_HEAD(cancels);
1239 /* concurrent cancels on the same handle can happen */
1240 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1242 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1246 rc = ldlm_cli_cancel_local(lock);
1247 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1248 LDLM_LOCK_RELEASE(lock);
1249 RETURN(rc < 0 ? rc : 0);
1251 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1252 * rpc which goes to canceld portal, so we can cancel other lru locks
1253 * here and send them all as one LDLM_CANCEL rpc. */
1254 LASSERT(list_empty(&lock->l_bl_ast));
1255 list_add(&lock->l_bl_ast, &cancels);
1257 exp = lock->l_conn_export;
1258 if (exp_connect_cancelset(exp)) {
1259 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1264 ns = lock->l_resource->lr_namespace;
1265 flags = ns_connect_lru_resize(ns) ?
1266 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1267 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1268 LDLM_FL_BL_AST, flags);
1270 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1274 /* XXX until we will have compound requests and can cut cancels from generic rpc
1275 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1276 static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
1278 CFS_LIST_HEAD(head);
1279 struct ldlm_lock *lock, *next;
1280 int left = 0, bl_ast = 0, rc;
1283 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1287 if (flags & LDLM_FL_LOCAL_ONLY) {
1288 rc = LDLM_FL_LOCAL_ONLY;
1289 ldlm_lock_cancel(lock);
1291 rc = ldlm_cli_cancel_local(lock);
1293 if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1294 LDLM_DEBUG(lock, "Cancel lock separately");
1295 list_del_init(&lock->l_bl_ast);
1296 list_add(&lock->l_bl_ast, &head);
1300 if (rc == LDLM_FL_LOCAL_ONLY) {
1301 /* CANCEL RPC should not be sent to server. */
1302 list_del_init(&lock->l_bl_ast);
1303 LDLM_LOCK_RELEASE(lock);
1310 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1317 * Callback function for lru-resize policy. Makes decision whether to keep
1318 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1319 * \a added and number of locks to be preferably canceled \a count.
1321 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1323 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1325 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1326 struct ldlm_lock *lock,
1327 int unused, int added,
1330 cfs_time_t cur = cfs_time_current();
1331 struct ldlm_pool *pl = &ns->ns_pool;
1336 * Stop lru processing when we reached passed @count or checked all
1339 if (count && added >= count)
1340 return LDLM_POLICY_KEEP_LOCK;
1342 slv = ldlm_pool_get_slv(pl);
1343 lvf = ldlm_pool_get_lvf(pl);
1344 la = cfs_duration_sec(cfs_time_sub(cur,
1345 lock->l_last_used));
1348 * Stop when slv is not yet come from server or lv is smaller than
1351 lv = lvf * la * unused;
1354 * Inform pool about current CLV to see it via proc.
1356 ldlm_pool_set_clv(pl, lv);
1357 return (slv == 1 || lv < slv) ?
1358 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1362 * Callback function for proc used policy. Makes decision whether to keep
1363 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1364 * \a added and number of locks to be preferably canceled \a count.
1366 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1368 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1370 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1371 struct ldlm_lock *lock,
1372 int unused, int added,
1376 * Stop lru processing when we reached passed @count or checked all
1379 return (added >= count) ?
1380 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1384 * Callback function for aged policy. Makes decision whether to keep
1385 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1386 * \a added and number of locks to be preferably canceled \a count.
1388 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1390 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1392 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1393 struct ldlm_lock *lock,
1394 int unused, int added,
1398 * Stop lru processing if young lock is found and we reached passed
1401 return ((added >= count) &&
1402 cfs_time_before(cfs_time_current(),
1403 cfs_time_add(lock->l_last_used,
1405 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1409 * Callback function for default policy. Makes decision whether to keep
1410 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1411 * \a added and number of locks to be preferably canceled \a count.
1413 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1415 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1417 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1418 struct ldlm_lock *lock,
1419 int unused, int added,
1423 * Stop lru processing when we reached passed @count or checked all
1426 return (added >= count) ?
1427 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1430 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1431 struct ldlm_lock *, int,
1434 static ldlm_cancel_lru_policy_t
1435 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1437 if (ns_connect_lru_resize(ns)) {
1438 if (flags & LDLM_CANCEL_SHRINK)
1439 /* We kill passed number of old locks. */
1440 return ldlm_cancel_passed_policy;
1441 else if (flags & LDLM_CANCEL_LRUR)
1442 return ldlm_cancel_lrur_policy;
1443 else if (flags & LDLM_CANCEL_PASSED)
1444 return ldlm_cancel_passed_policy;
1446 if (flags & LDLM_CANCEL_AGED)
1447 return ldlm_cancel_aged_policy;
1450 return ldlm_cancel_default_policy;
1453 /* - Free space in lru for @count new locks,
1454 * redundant unused locks are canceled locally;
1455 * - also cancel locally unused aged locks;
1456 * - do not cancel more than @max locks;
1457 * - GET the found locks and add them into the @cancels list.
1459 * A client lock can be added to the l_bl_ast list only when it is
1460 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1461 * There are the following use cases: ldlm_cancel_resource_local(),
1462 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1463 * flag properly. As any attempt to cancel a lock rely on this flag,
1464 * l_bl_ast list is accessed later without any special locking.
1466 * Calling policies for enabled lru resize:
1467 * ----------------------------------------
1468 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1469 * cancel not more than @count locks;
1471 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1472 * the beginning of lru list);
1474 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1475 * memory pressre policy function;
1477 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1479 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1480 int count, int max, int cancel_flags, int flags)
1482 ldlm_cancel_lru_policy_t pf;
1483 struct ldlm_lock *lock, *next;
1484 int added = 0, unused;
1487 spin_lock(&ns->ns_unused_lock);
1488 unused = ns->ns_nr_unused;
1490 if (!ns_connect_lru_resize(ns))
1491 count += unused - ns->ns_max_unused;
1493 pf = ldlm_cancel_lru_policy(ns, flags);
1494 LASSERT(pf != NULL);
1496 while (!list_empty(&ns->ns_unused_list)) {
1497 /* For any flags, stop scanning if @max is reached. */
1498 if (max && added >= max)
1501 list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru){
1502 /* No locks which got blocking requests. */
1503 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1505 /* Somebody is already doing CANCEL. No need in this
1506 * lock in lru, do not traverse it again. */
1507 if (!(lock->l_flags & LDLM_FL_CANCELING))
1510 ldlm_lock_remove_from_lru_nolock(lock);
1512 if (&lock->l_lru == &ns->ns_unused_list)
1515 LDLM_LOCK_GET(lock);
1516 spin_unlock(&ns->ns_unused_lock);
1517 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1519 /* Pass the lock through the policy filter and see if it
1520 * should stay in lru.
1522 * Even for shrinker policy we stop scanning if
1523 * we find a lock that should stay in the cache.
1524 * We should take into account lock age anyway
1525 * as new lock even if it is small of weight is
1526 * valuable resource.
1528 * That is, for shrinker policy we drop only
1529 * old locks, but additionally chose them by
1530 * their weight. Big extent locks will stay in
1532 if (pf(ns, lock, unused, added, count) ==
1533 LDLM_POLICY_KEEP_LOCK) {
1534 lu_ref_del(&lock->l_reference,
1535 __FUNCTION__, cfs_current());
1536 LDLM_LOCK_RELEASE(lock);
1537 spin_lock(&ns->ns_unused_lock);
1541 lock_res_and_lock(lock);
1542 /* Check flags again under the lock. */
1543 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1544 (ldlm_lock_remove_from_lru(lock) == 0)) {
1545 /* other thread is removing lock from lru or
1546 * somebody is already doing CANCEL or
1547 * there is a blocking request which will send
1548 * cancel by itseft or the lock is matched
1549 * is already not unused. */
1550 unlock_res_and_lock(lock);
1551 lu_ref_del(&lock->l_reference,
1552 __FUNCTION__, cfs_current());
1553 LDLM_LOCK_RELEASE(lock);
1554 spin_lock(&ns->ns_unused_lock);
1557 LASSERT(!lock->l_readers && !lock->l_writers);
1559 /* If we have chosen to cancel this lock voluntarily, we
1560 * better send cancel notification to server, so that it
1561 * frees appropriate state. This might lead to a race
1562 * where while we are doing cancel here, server is also
1563 * silently cancelling this lock. */
1564 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1566 /* Setting the CBPENDING flag is a little misleading,
1567 * but prevents an important race; namely, once
1568 * CBPENDING is set, the lock can accumulate no more
1569 * readers/writers. Since readers and writers are
1570 * already zero here, ldlm_lock_decref() won't see
1571 * this flag and call l_blocking_ast */
1572 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1574 /* We can't re-add to l_lru as it confuses the
1575 * refcounting in ldlm_lock_remove_from_lru() if an AST
1576 * arrives after we drop ns_lock below. We use l_bl_ast
1577 * and can't use l_pending_chain as it is used both on
1578 * server and client nevertheless bug 5666 says it is
1579 * used only on server */
1580 LASSERT(list_empty(&lock->l_bl_ast));
1581 list_add(&lock->l_bl_ast, cancels);
1582 unlock_res_and_lock(lock);
1583 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1584 spin_lock(&ns->ns_unused_lock);
1588 spin_unlock(&ns->ns_unused_lock);
1589 RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
1592 /* when called with LDLM_ASYNC the blocking callback will be handled
1593 * in a thread and this function will return after the thread has been
1594 * asked to call the callback. when called with LDLM_SYNC the blocking
1595 * callback will be performed in this function. */
1596 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1599 CFS_LIST_HEAD(cancels);
1604 sync = LDLM_SYNC; /* force to be sync in user space */
1606 count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
1607 if (sync == LDLM_ASYNC) {
1608 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1613 /* If an error occured in ASYNC mode, or this is SYNC mode,
1614 * cancel the list. */
1615 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1619 /* Find and cancel locally unused locks found on resource, matched to the
1620 * given policy, mode. GET the found locks and add them into the @cancels
1622 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1623 struct list_head *cancels,
1624 ldlm_policy_data_t *policy,
1625 ldlm_mode_t mode, int lock_flags,
1626 int cancel_flags, void *opaque)
1628 struct ldlm_lock *lock;
1633 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1634 if (opaque != NULL && lock->l_ast_data != opaque) {
1635 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1636 lock->l_ast_data, opaque);
1641 if (lock->l_readers || lock->l_writers) {
1642 if (cancel_flags & LDLM_FL_WARN) {
1643 LDLM_ERROR(lock, "lock in use");
1649 /* If somebody is already doing CANCEL, or blocking ast came,
1650 * skip this lock. */
1651 if (lock->l_flags & LDLM_FL_BL_AST ||
1652 lock->l_flags & LDLM_FL_CANCELING)
1655 if (lockmode_compat(lock->l_granted_mode, mode))
1658 /* If policy is given and this is IBITS lock, add to list only
1659 * those locks that match by policy. */
1660 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1661 !(lock->l_policy_data.l_inodebits.bits &
1662 policy->l_inodebits.bits))
1665 /* See CBPENDING comment in ldlm_cancel_lru */
1666 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1669 LASSERT(list_empty(&lock->l_bl_ast));
1670 list_add(&lock->l_bl_ast, cancels);
1671 LDLM_LOCK_GET(lock);
1676 RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
1679 /* If @req is NULL, send CANCEL request to server with handles of locks
1680 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1681 * separately per lock.
1682 * If @req is not NULL, put handles of locks in @cancels into the request
1683 * buffer at the offset @off.
1684 * Destroy @cancels at the end. */
1685 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1686 struct ptlrpc_request *req, int flags)
1688 struct ldlm_lock *lock;
1692 if (list_empty(cancels) || count == 0)
1695 /* XXX: requests (both batched and not) could be sent in parallel.
1696 * Usually it is enough to have just 1 RPC, but it is possible that
1697 * there are to many locks to be cancelled in LRU or on a resource.
1698 * It would also speed up the case when the server does not support
1701 LASSERT(!list_empty(cancels));
1702 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1703 LASSERT(lock->l_conn_export);
1705 if (exp_connect_cancelset(lock->l_conn_export)) {
1708 ldlm_cancel_pack(req, cancels, count);
1710 res = ldlm_cli_cancel_req(lock->l_conn_export,
1714 res = ldlm_cli_cancel_req(lock->l_conn_export,
1719 CERROR("ldlm_cli_cancel_list: %d\n", res);
1724 ldlm_lock_list_put(cancels, l_bl_ast, res);
1726 LASSERT(count == 0);
1730 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1731 const struct ldlm_res_id *res_id,
1732 ldlm_policy_data_t *policy,
1733 ldlm_mode_t mode, int flags, void *opaque)
1735 struct ldlm_resource *res;
1736 CFS_LIST_HEAD(cancels);
1741 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1743 /* This is not a problem. */
1744 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1748 LDLM_RESOURCE_ADDREF(res);
1749 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1751 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1753 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1755 LDLM_RESOURCE_DELREF(res);
1756 ldlm_resource_putref(res);
1760 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1762 int no_resource = 0;
1764 spin_lock(&ns->ns_hash_lock);
1765 if (ns->ns_resources == 0)
1767 spin_unlock(&ns->ns_hash_lock);
1769 RETURN(no_resource);
1772 /* Cancel all locks on a namespace (or a specific resource, if given)
1773 * that have 0 readers/writers.
1775 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1776 * to notify the server. */
1777 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1778 const struct ldlm_res_id *res_id,
1779 int flags, void *opaque)
1788 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1792 spin_lock(&ns->ns_hash_lock);
1793 for (i = 0; i < RES_HASH_SIZE; i++) {
1794 struct list_head *tmp;
1795 tmp = ns->ns_hash[i].next;
1796 while (tmp != &(ns->ns_hash[i])) {
1797 struct ldlm_resource *res;
1800 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1801 ldlm_resource_getref(res);
1802 spin_unlock(&ns->ns_hash_lock);
1804 LDLM_RESOURCE_ADDREF(res);
1805 rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1810 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1811 res->lr_name.name[0], rc);
1813 LDLM_RESOURCE_DELREF(res);
1814 spin_lock(&ns->ns_hash_lock);
1816 ldlm_resource_putref_locked(res);
1819 spin_unlock(&ns->ns_hash_lock);
1824 /* Lock iterators. */
1826 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1829 struct list_head *tmp, *next;
1830 struct ldlm_lock *lock;
1831 int rc = LDLM_ITER_CONTINUE;
1836 RETURN(LDLM_ITER_CONTINUE);
1839 list_for_each_safe(tmp, next, &res->lr_granted) {
1840 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1842 if (iter(lock, closure) == LDLM_ITER_STOP)
1843 GOTO(out, rc = LDLM_ITER_STOP);
1846 list_for_each_safe(tmp, next, &res->lr_converting) {
1847 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1849 if (iter(lock, closure) == LDLM_ITER_STOP)
1850 GOTO(out, rc = LDLM_ITER_STOP);
1853 list_for_each_safe(tmp, next, &res->lr_waiting) {
1854 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1856 if (iter(lock, closure) == LDLM_ITER_STOP)
1857 GOTO(out, rc = LDLM_ITER_STOP);
1864 struct iter_helper_data {
1865 ldlm_iterator_t iter;
1869 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1871 struct iter_helper_data *helper = closure;
1872 return helper->iter(lock, helper->closure);
1875 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1877 return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1880 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1883 struct iter_helper_data helper = { iter: iter, closure: closure };
1884 return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1887 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1888 ldlm_res_iterator_t iter, void *closure)
1890 int i, rc = LDLM_ITER_CONTINUE;
1891 struct ldlm_resource *res;
1892 struct list_head *tmp;
1895 spin_lock(&ns->ns_hash_lock);
1896 for (i = 0; i < RES_HASH_SIZE; i++) {
1897 tmp = ns->ns_hash[i].next;
1898 while (tmp != &(ns->ns_hash[i])) {
1899 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1900 ldlm_resource_getref(res);
1901 spin_unlock(&ns->ns_hash_lock);
1902 LDLM_RESOURCE_ADDREF(res);
1904 rc = iter(res, closure);
1906 LDLM_RESOURCE_DELREF(res);
1907 spin_lock(&ns->ns_hash_lock);
1909 ldlm_resource_putref_locked(res);
1910 if (rc == LDLM_ITER_STOP)
1915 spin_unlock(&ns->ns_hash_lock);
1919 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1920 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1921 const struct ldlm_res_id *res_id,
1922 ldlm_iterator_t iter, void *data)
1924 struct ldlm_resource *res;
1928 CERROR("must pass in namespace\n");
1932 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1938 LDLM_RESOURCE_ADDREF(res);
1939 ldlm_resource_foreach(res, iter, data);
1940 LDLM_RESOURCE_DELREF(res);
1941 ldlm_resource_putref(res);
1947 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1949 struct list_head *list = closure;
1951 /* we use l_pending_chain here, because it's unused on clients. */
1952 LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1953 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1954 /* bug 9573: don't replay locks left after eviction */
1955 if (!(lock->l_flags & LDLM_FL_FAILED))
1956 list_add(&lock->l_pending_chain, list);
1957 return LDLM_ITER_CONTINUE;
1960 static int replay_lock_interpret(const struct lu_env *env,
1961 struct ptlrpc_request *req,
1962 struct ldlm_async_args *aa, int rc)
1964 struct lustre_handle old_hash_key;
1965 struct ldlm_lock *lock;
1966 struct ldlm_reply *reply;
1967 struct obd_export *exp;
1970 atomic_dec(&req->rq_import->imp_replay_inflight);
1975 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1977 GOTO(out, rc = -EPROTO);
1979 lock = ldlm_handle2lock(&aa->lock_handle);
1981 CERROR("received replay ack for unknown local cookie "LPX64
1982 " remote cookie "LPX64 " from server %s id %s\n",
1983 aa->lock_handle.cookie, reply->lock_handle.cookie,
1984 req->rq_export->exp_client_uuid.uuid,
1985 libcfs_id2str(req->rq_peer));
1986 GOTO(out, rc = -ESTALE);
1989 old_hash_key = lock->l_remote_handle;
1990 lock->l_remote_handle = reply->lock_handle;
1992 /* Key change rehash lock in per-export hash with new key */
1993 exp = req->rq_export;
1994 if (exp && exp->exp_lock_hash)
1995 lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
1996 &lock->l_remote_handle,
1999 LDLM_DEBUG(lock, "replayed lock:");
2000 ptlrpc_import_recovery_state_machine(req->rq_import);
2001 LDLM_LOCK_PUT(lock);
2004 ptlrpc_connect_import(req->rq_import, NULL);
2009 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2011 struct ptlrpc_request *req;
2012 struct ldlm_async_args *aa;
2013 struct ldlm_request *body;
2018 /* Bug 11974: Do not replay a lock which is actively being canceled */
2019 if (lock->l_flags & LDLM_FL_CANCELING) {
2020 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2024 /* If this is reply-less callback lock, we cannot replay it, since
2025 * server might have long dropped it, but notification of that event was
2026 * lost by network. (and server granted conflicting lock already) */
2027 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2028 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2029 ldlm_lock_cancel(lock);
2033 * If granted mode matches the requested mode, this lock is granted.
2035 * If they differ, but we have a granted mode, then we were granted
2036 * one mode and now want another: ergo, converting.
2038 * If we haven't been granted anything and are on a resource list,
2039 * then we're blocked/waiting.
2041 * If we haven't been granted anything and we're NOT on a resource list,
2042 * then we haven't got a reply yet and don't have a known disposition.
2043 * This happens whenever a lock enqueue is the request that triggers
2046 if (lock->l_granted_mode == lock->l_req_mode)
2047 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2048 else if (lock->l_granted_mode)
2049 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2050 else if (!list_empty(&lock->l_res_link))
2051 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2053 flags = LDLM_FL_REPLAY;
2055 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2056 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2060 /* We're part of recovery, so don't wait for it. */
2061 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2063 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2064 ldlm_lock2desc(lock, &body->lock_desc);
2065 body->lock_flags = flags;
2067 ldlm_lock2handle(lock, &body->lock_handle[0]);
2068 if (lock->l_lvb_len != 0) {
2069 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2070 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2073 ptlrpc_request_set_replen(req);
2074 /* notify the server we've replayed all requests.
2075 * also, we mark the request to be put on a dedicated
2076 * queue to be processed after all request replayes.
2078 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2080 LDLM_DEBUG(lock, "replaying lock:");
2082 atomic_inc(&req->rq_import->imp_replay_inflight);
2083 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2084 aa = ptlrpc_req_async_args(req);
2085 aa->lock_handle = body->lock_handle[0];
2086 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2087 ptlrpcd_add_req(req, PSCOPE_OTHER);
2092 int ldlm_replay_locks(struct obd_import *imp)
2094 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2095 CFS_LIST_HEAD(list);
2096 struct ldlm_lock *lock, *next;
2101 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2103 /* ensure this doesn't fall to 0 before all have been queued */
2104 atomic_inc(&imp->imp_replay_inflight);
2106 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2108 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2109 list_del_init(&lock->l_pending_chain);
2111 continue; /* or try to do the rest? */
2112 rc = replay_one_lock(imp, lock);
2115 atomic_dec(&imp->imp_replay_inflight);