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, int mode)
440 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
441 lock_res_and_lock(lock);
442 /* Check that lock is not granted or failed, we might race. */
443 if ((lock->l_req_mode != lock->l_granted_mode) &&
444 !(lock->l_flags & LDLM_FL_FAILED)) {
445 /* Make sure that this lock will not be found by raced
446 * bl_ast and -EINVAL reply is sent to server anyways.
448 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
449 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
452 unlock_res_and_lock(lock);
456 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
457 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
459 LDLM_DEBUG(lock, "lock was granted or failed in race");
461 ldlm_lock_decref_internal(lock, mode);
463 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
464 * from llite/file.c/ll_file_flock(). */
465 /* This code makes for the fact that we do not have blocking handler on
466 * a client for flock locks. As such this is the place where we must
467 * completely kill failed locks. (interrupted and those that
468 * were waiting to be granted when server evicted us. */
469 if (lock->l_resource->lr_type == LDLM_FLOCK) {
470 lock_res_and_lock(lock);
471 ldlm_resource_unlink_lock(lock);
472 ldlm_lock_destroy_nolock(lock);
473 unlock_res_and_lock(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 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_get(&req->rq_pill,
516 GOTO(cleanup, rc = -EPROTO);
518 memcpy(lvb, tmplvb, lvb_len);
524 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
526 GOTO(cleanup, rc = -EPROTO);
528 /* lock enqueued on the server */
531 lock_res_and_lock(lock);
532 old_hash_key = lock->l_remote_handle;
533 lock->l_remote_handle = reply->lock_handle;
535 /* Key change rehash lock in per-export hash with new key */
536 if (exp->exp_lock_hash)
537 cfs_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
538 &lock->l_remote_handle,
541 *flags = reply->lock_flags;
542 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
543 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
544 * to wait with no timeout as well */
545 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
546 unlock_res_and_lock(lock);
548 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
549 lock, reply->lock_handle.cookie, *flags);
551 /* If enqueue returned a blocked lock but the completion handler has
552 * already run, then it fixed up the resource and we don't need to do it
554 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
555 int newmode = reply->lock_desc.l_req_mode;
557 if (newmode && newmode != lock->l_req_mode) {
558 LDLM_DEBUG(lock, "server returned different mode %s",
559 ldlm_lockname[newmode]);
560 lock->l_req_mode = newmode;
563 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
564 lock->l_resource->lr_name.name,
565 sizeof(struct ldlm_res_id))) {
566 CDEBUG(D_INFO, "remote intent success, locking "
567 "(%ld,%ld,%ld) instead of "
569 (long)reply->lock_desc.l_resource.lr_name.name[0],
570 (long)reply->lock_desc.l_resource.lr_name.name[1],
571 (long)reply->lock_desc.l_resource.lr_name.name[2],
572 (long)lock->l_resource->lr_name.name[0],
573 (long)lock->l_resource->lr_name.name[1],
574 (long)lock->l_resource->lr_name.name[2]);
576 rc = ldlm_lock_change_resource(ns, lock,
577 &reply->lock_desc.l_resource.lr_name);
578 if (rc || lock->l_resource == NULL)
579 GOTO(cleanup, rc = -ENOMEM);
580 LDLM_DEBUG(lock, "client-side enqueue, new resource");
583 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
585 lock->l_policy_data =
586 reply->lock_desc.l_policy_data;
587 if (type != LDLM_PLAIN)
588 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
591 if ((*flags) & LDLM_FL_AST_SENT ||
592 /* Cancel extent locks as soon as possible on a liblustre client,
593 * because it cannot handle asynchronous ASTs robustly (see
595 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
596 lock_res_and_lock(lock);
597 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
598 unlock_res_and_lock(lock);
599 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
602 /* If the lock has already been granted by a completion AST, don't
603 * clobber the LVB with an older one. */
604 if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
607 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
609 tmplvb = req_capsule_server_get(&req->rq_pill,
612 GOTO(cleanup, rc = -EPROTO);
613 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
617 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
618 if (lock->l_completion_ast != NULL) {
619 int err = lock->l_completion_ast(lock, *flags, NULL);
627 if (lvb_len && lvb != NULL) {
628 /* Copy the LVB here, and not earlier, because the completion
629 * AST (if any) can override what we got in the reply */
630 memcpy(lvb, lock->l_lvb_data, lvb_len);
633 LDLM_DEBUG(lock, "client-side enqueue END");
636 if (cleanup_phase == 1 && rc)
637 failed_lock_cleanup(ns, lock, mode);
638 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
640 LDLM_LOCK_RELEASE(lock);
644 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
645 * a single page on the send/receive side. XXX: 512 should be changed
646 * to more adequate value. */
647 static inline int ldlm_req_handles_avail(int req_size, int off)
651 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
652 if (likely(avail >= 0))
653 avail /= (int)sizeof(struct lustre_handle);
656 avail += LDLM_LOCKREQ_HANDLES - off;
661 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
662 enum req_location loc,
665 int size = req_capsule_msg_size(pill, loc);
666 return ldlm_req_handles_avail(size, off);
669 static inline int ldlm_format_handles_avail(struct obd_import *imp,
670 const struct req_format *fmt,
671 enum req_location loc, int off)
673 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
674 return ldlm_req_handles_avail(size, off);
677 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
678 * @count locks in @cancels. */
679 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
680 int version, int opc, int canceloff,
681 cfs_list_t *cancels, int count)
683 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
684 struct req_capsule *pill = &req->rq_pill;
685 struct ldlm_request *dlm = NULL;
686 int flags, avail, to_free, bufcount, pack = 0;
693 if (exp_connect_cancelset(exp)) {
694 /* Estimate the amount of available space in the request. */
695 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
696 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
698 flags = ns_connect_lru_resize(ns) ?
699 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
700 to_free = !ns_connect_lru_resize(ns) &&
701 opc == LDLM_ENQUEUE ? 1 : 0;
703 /* Cancel lru locks here _only_ if the server supports
704 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
705 * rpc, what will make us slower. */
707 count += ldlm_cancel_lru_local(ns, cancels, to_free,
708 avail - count, 0, flags);
713 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
714 ldlm_request_bufsize(pack, opc));
717 rc = ptlrpc_request_pack(req, version, opc);
719 ldlm_lock_list_put(cancels, l_bl_ast, count);
723 if (exp_connect_cancelset(exp)) {
725 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
727 /* Skip first lock handler in ldlm_request_pack(),
728 * this method will incrment @lock_count according
729 * to the lock handle amount actually written to
731 dlm->lock_count = canceloff;
733 /* Pack into the request @pack lock handles. */
734 ldlm_cli_cancel_list(cancels, pack, req, 0);
735 /* Prepare and send separate cancel rpc for others. */
736 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
738 ldlm_lock_list_put(cancels, l_bl_ast, count);
743 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
744 cfs_list_t *cancels, int count)
746 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
747 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
750 /* If a request has some specific initialisation it is passed in @reqp,
751 * otherwise it is created in ldlm_cli_enqueue.
753 * Supports sync and async requests, pass @async flag accordingly. If a
754 * request was created in ldlm_cli_enqueue and it is the async request,
755 * pass it to the caller in @reqp. */
756 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
757 struct ldlm_enqueue_info *einfo,
758 const struct ldlm_res_id *res_id,
759 ldlm_policy_data_t const *policy, int *flags,
760 void *lvb, __u32 lvb_len, struct lustre_handle *lockh,
763 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
764 struct ldlm_lock *lock;
765 struct ldlm_request *body;
766 int is_replay = *flags & LDLM_FL_REPLAY;
767 int req_passed_in = 1;
769 struct ptlrpc_request *req;
772 LASSERT(exp != NULL);
774 /* If we're replaying this lock, just check some invariants.
775 * If we're creating a new lock, get everything all setup nice. */
777 lock = ldlm_handle2lock_long(lockh, 0);
778 LASSERT(lock != NULL);
779 LDLM_DEBUG(lock, "client-side enqueue START");
780 LASSERT(exp == lock->l_conn_export);
782 const struct ldlm_callback_suite cbs = {
783 .lcs_completion = einfo->ei_cb_cp,
784 .lcs_blocking = einfo->ei_cb_bl,
785 .lcs_glimpse = einfo->ei_cb_gl,
786 .lcs_weigh = einfo->ei_cb_wg
788 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
789 einfo->ei_mode, &cbs, einfo->ei_cbdata,
793 /* for the local lock, add the reference */
794 ldlm_lock_addref_internal(lock, einfo->ei_mode);
795 ldlm_lock2handle(lock, lockh);
796 if (policy != NULL) {
797 /* INODEBITS_INTEROP: If the server does not support
798 * inodebits, we will request a plain lock in the
799 * descriptor (ldlm_lock2desc() below) but use an
800 * inodebits lock internally with both bits set.
802 if (einfo->ei_type == LDLM_IBITS &&
803 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
804 lock->l_policy_data.l_inodebits.bits =
805 MDS_INODELOCK_LOOKUP |
806 MDS_INODELOCK_UPDATE;
808 lock->l_policy_data = *policy;
811 if (einfo->ei_type == LDLM_EXTENT)
812 lock->l_req_extent = policy->l_extent;
813 LDLM_DEBUG(lock, "client-side enqueue START");
816 /* lock not sent to server yet */
818 if (reqp == NULL || *reqp == NULL) {
819 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
824 failed_lock_cleanup(ns, lock, einfo->ei_mode);
825 LDLM_LOCK_RELEASE(lock);
835 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
837 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
838 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
841 lock->l_conn_export = exp;
842 lock->l_export = NULL;
843 lock->l_blocking_ast = einfo->ei_cb_bl;
845 /* Dump lock data into the request buffer */
846 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
847 ldlm_lock2desc(lock, &body->lock_desc);
848 body->lock_flags = *flags;
849 body->lock_handle[0] = *lockh;
851 /* Continue as normal. */
852 if (!req_passed_in) {
854 req_capsule_extend(&req->rq_pill,
855 &RQF_LDLM_ENQUEUE_LVB);
856 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
857 RCL_SERVER, lvb_len);
859 ptlrpc_request_set_replen(req);
863 * Liblustre client doesn't get extent locks, except for O_APPEND case
864 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
865 * [i_size, OBD_OBJECT_EOF] lock is taken.
867 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
868 policy->l_extent.end == OBD_OBJECT_EOF));
871 LASSERT(reqp != NULL);
875 LDLM_DEBUG(lock, "sending request");
877 rc = ptlrpc_queue_wait(req);
879 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
880 einfo->ei_mode, flags, lvb, lvb_len,
883 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
884 * one reference that we took */
886 LDLM_LOCK_RELEASE(lock);
890 if (!req_passed_in && req != NULL) {
891 ptlrpc_req_finished(req);
899 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
902 struct ldlm_resource *res;
905 if (ns_is_client(lock->l_resource->lr_namespace)) {
906 CERROR("Trying to cancel local lock\n");
909 LDLM_DEBUG(lock, "client-side local convert");
911 res = ldlm_lock_convert(lock, new_mode, flags);
913 ldlm_reprocess_all(res);
918 LDLM_DEBUG(lock, "client-side local convert handler END");
923 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
924 * conversion of locks which are on the waiting or converting queue */
925 /* Caller of this code is supposed to take care of lock readers/writers
927 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
929 struct ldlm_request *body;
930 struct ldlm_reply *reply;
931 struct ldlm_lock *lock;
932 struct ldlm_resource *res;
933 struct ptlrpc_request *req;
937 lock = ldlm_handle2lock(lockh);
944 if (lock->l_conn_export == NULL)
945 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
947 LDLM_DEBUG(lock, "client-side convert");
949 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
950 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
957 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
958 body->lock_handle[0] = lock->l_remote_handle;
960 body->lock_desc.l_req_mode = new_mode;
961 body->lock_flags = *flags;
964 ptlrpc_request_set_replen(req);
965 rc = ptlrpc_queue_wait(req);
969 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
971 GOTO(out, rc = -EPROTO);
974 GOTO(out, rc = req->rq_status);
976 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
978 ldlm_reprocess_all(res);
979 /* Go to sleep until the lock is granted. */
980 /* FIXME: or cancelled. */
981 if (lock->l_completion_ast) {
982 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
993 ptlrpc_req_finished(req);
997 /* Cancel locks locally.
999 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
1000 * LDLM_FL_CANCELING otherwise;
1001 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
1002 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
1004 int rc = LDLM_FL_LOCAL_ONLY;
1007 if (lock->l_conn_export) {
1010 LDLM_DEBUG(lock, "client-side cancel");
1011 /* Set this flag to prevent others from getting new references*/
1012 lock_res_and_lock(lock);
1013 lock->l_flags |= LDLM_FL_CBPENDING;
1014 local_only = (lock->l_flags &
1015 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1016 ldlm_cancel_callback(lock);
1017 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
1018 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1019 unlock_res_and_lock(lock);
1022 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1024 rc = LDLM_FL_LOCAL_ONLY;
1026 ldlm_lock_cancel(lock);
1028 if (ns_is_client(lock->l_resource->lr_namespace)) {
1029 LDLM_ERROR(lock, "Trying to cancel local lock");
1032 LDLM_DEBUG(lock, "server-side local cancel");
1033 ldlm_lock_cancel(lock);
1034 ldlm_reprocess_all(lock->l_resource);
1040 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
1041 of the request @req. */
1042 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1043 cfs_list_t *head, int count)
1045 struct ldlm_request *dlm;
1046 struct ldlm_lock *lock;
1047 int max, packed = 0;
1050 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1051 LASSERT(dlm != NULL);
1053 /* Check the room in the request buffer. */
1054 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1055 sizeof(struct ldlm_request);
1056 max /= sizeof(struct lustre_handle);
1057 max += LDLM_LOCKREQ_HANDLES;
1058 LASSERT(max >= dlm->lock_count + count);
1060 /* XXX: it would be better to pack lock handles grouped by resource.
1061 * so that the server cancel would call filter_lvbo_update() less
1063 cfs_list_for_each_entry(lock, head, l_bl_ast) {
1066 LASSERT(lock->l_conn_export);
1067 /* Pack the lock handle to the given request buffer. */
1068 LDLM_DEBUG(lock, "packing");
1069 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1072 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1076 /* Prepare and send a batched cancel rpc, it will include count lock handles
1077 * of locks given in @head. */
1078 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1079 int count, ldlm_cancel_flags_t flags)
1081 struct ptlrpc_request *req = NULL;
1082 struct obd_import *imp;
1087 LASSERT(exp != NULL);
1090 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
1092 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1095 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1096 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1103 imp = class_exp2cliimp(exp);
1104 if (imp == NULL || imp->imp_invalid) {
1106 "skipping cancel on invalid import %p\n", imp);
1110 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1112 GOTO(out, rc = -ENOMEM);
1114 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1115 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1116 ldlm_request_bufsize(count, LDLM_CANCEL));
1118 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1120 ptlrpc_request_free(req);
1123 req->rq_no_resend = 1;
1124 req->rq_no_delay = 1;
1126 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1127 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1128 ptlrpc_at_set_req_timeout(req);
1130 ldlm_cancel_pack(req, cancels, count);
1132 ptlrpc_request_set_replen(req);
1133 if (flags & LCF_ASYNC) {
1134 ptlrpcd_add_req(req, PSCOPE_OTHER);
1138 rc = ptlrpc_queue_wait(req);
1141 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1142 "out of sync -- not fatal\n",
1143 libcfs_nid2str(req->rq_import->
1144 imp_connection->c_peer.nid));
1146 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1147 req->rq_import_generation == imp->imp_generation) {
1148 ptlrpc_req_finished(req);
1150 } else if (rc != ELDLM_OK) {
1151 CERROR("Got rc %d from cancel RPC: canceling "
1159 ptlrpc_req_finished(req);
1162 return sent ? sent : rc;
1165 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1167 LASSERT(imp != NULL);
1168 return &imp->imp_obd->obd_namespace->ns_pool;
1172 * Update client's obd pool related fields with new SLV and Limit from \a req.
1174 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1176 struct obd_device *obd;
1177 __u64 old_slv, new_slv;
1180 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1181 !imp_connect_lru_resize(req->rq_import)))
1184 * Do nothing for corner cases.
1190 * In some cases RPC may contain slv and limit zeroed out. This is
1191 * the case when server does not support lru resize feature. This is
1192 * also possible in some recovery cases when server side reqs have no
1193 * ref to obd export and thus access to server side namespace is no
1196 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1197 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1198 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1199 "(SLV: "LPU64", Limit: %u)",
1200 lustre_msg_get_slv(req->rq_repmsg),
1201 lustre_msg_get_limit(req->rq_repmsg));
1205 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1206 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1207 obd = req->rq_import->imp_obd;
1210 * Set new SLV and Limit to obd fields to make accessible for pool
1211 * thread. We do not access obd_namespace and pool directly here
1212 * as there is no reliable way to make sure that they are still
1213 * alive in cleanup time. Evil races are possible which may cause
1214 * oops in that time.
1216 cfs_write_lock(&obd->obd_pool_lock);
1217 old_slv = obd->obd_pool_slv;
1218 obd->obd_pool_slv = new_slv;
1219 obd->obd_pool_limit = new_limit;
1220 cfs_write_unlock(&obd->obd_pool_lock);
1224 EXPORT_SYMBOL(ldlm_cli_update_pool);
1226 int ldlm_cli_cancel(struct lustre_handle *lockh)
1228 struct obd_export *exp;
1229 int avail, flags, count = 1, rc = 0;
1230 struct ldlm_namespace *ns;
1231 struct ldlm_lock *lock;
1232 CFS_LIST_HEAD(cancels);
1235 /* concurrent cancels on the same handle can happen */
1236 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1238 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1242 rc = ldlm_cli_cancel_local(lock);
1243 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1244 LDLM_LOCK_RELEASE(lock);
1245 RETURN(rc < 0 ? rc : 0);
1247 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1248 * rpc which goes to canceld portal, so we can cancel other lru locks
1249 * here and send them all as one LDLM_CANCEL rpc. */
1250 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1251 cfs_list_add(&lock->l_bl_ast, &cancels);
1253 exp = lock->l_conn_export;
1254 if (exp_connect_cancelset(exp)) {
1255 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1260 ns = lock->l_resource->lr_namespace;
1261 flags = ns_connect_lru_resize(ns) ?
1262 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1263 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1266 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1270 /* XXX until we will have compound requests and can cut cancels from generic rpc
1271 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1272 static int ldlm_cancel_list(cfs_list_t *cancels, int count,
1273 ldlm_cancel_flags_t flags)
1275 CFS_LIST_HEAD(head);
1276 struct ldlm_lock *lock, *next;
1277 int left = 0, bl_ast = 0, rc;
1280 cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1284 if (flags & LCF_LOCAL) {
1285 rc = LDLM_FL_LOCAL_ONLY;
1286 ldlm_lock_cancel(lock);
1288 rc = ldlm_cli_cancel_local(lock);
1290 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1291 LDLM_DEBUG(lock, "Cancel lock separately");
1292 cfs_list_del_init(&lock->l_bl_ast);
1293 cfs_list_add(&lock->l_bl_ast, &head);
1297 if (rc == LDLM_FL_LOCAL_ONLY) {
1298 /* CANCEL RPC should not be sent to server. */
1299 cfs_list_del_init(&lock->l_bl_ast);
1300 LDLM_LOCK_RELEASE(lock);
1307 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1314 * Callback function for lru-resize policy. Makes decision whether to keep
1315 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1316 * \a added and number of locks to be preferably canceled \a count.
1318 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1320 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1322 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1323 struct ldlm_lock *lock,
1324 int unused, int added,
1327 cfs_time_t cur = cfs_time_current();
1328 struct ldlm_pool *pl = &ns->ns_pool;
1333 * Stop lru processing when we reached passed @count or checked all
1336 if (count && added >= count)
1337 return LDLM_POLICY_KEEP_LOCK;
1339 slv = ldlm_pool_get_slv(pl);
1340 lvf = ldlm_pool_get_lvf(pl);
1341 la = cfs_duration_sec(cfs_time_sub(cur,
1342 lock->l_last_used));
1345 * Stop when slv is not yet come from server or lv is smaller than
1348 lv = lvf * la * unused;
1351 * Inform pool about current CLV to see it via proc.
1353 ldlm_pool_set_clv(pl, lv);
1354 return (slv == 0 || lv < slv) ?
1355 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1359 * Callback function for proc used policy. Makes decision whether to keep
1360 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1361 * \a added and number of locks to be preferably canceled \a count.
1363 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1365 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1367 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1368 struct ldlm_lock *lock,
1369 int unused, int added,
1373 * Stop lru processing when we reached passed @count or checked all
1376 return (added >= count) ?
1377 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1381 * Callback function for aged policy. Makes decision whether to keep
1382 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1383 * \a added and number of locks to be preferably canceled \a count.
1385 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1387 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1389 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1390 struct ldlm_lock *lock,
1391 int unused, int added,
1395 * Stop lru processing if young lock is found and we reached passed
1398 return ((added >= count) &&
1399 cfs_time_before(cfs_time_current(),
1400 cfs_time_add(lock->l_last_used,
1402 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1406 * Callback function for default policy. Makes decision whether to keep
1407 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1408 * \a added and number of locks to be preferably canceled \a count.
1410 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1412 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1414 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1415 struct ldlm_lock *lock,
1416 int unused, int added,
1420 * Stop lru processing when we reached passed @count or checked all
1423 return (added >= count) ?
1424 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1427 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1428 struct ldlm_lock *, int,
1431 static ldlm_cancel_lru_policy_t
1432 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1434 if (ns_connect_lru_resize(ns)) {
1435 if (flags & LDLM_CANCEL_SHRINK)
1436 /* We kill passed number of old locks. */
1437 return ldlm_cancel_passed_policy;
1438 else if (flags & LDLM_CANCEL_LRUR)
1439 return ldlm_cancel_lrur_policy;
1440 else if (flags & LDLM_CANCEL_PASSED)
1441 return ldlm_cancel_passed_policy;
1443 if (flags & LDLM_CANCEL_AGED)
1444 return ldlm_cancel_aged_policy;
1447 return ldlm_cancel_default_policy;
1450 /* - Free space in lru for @count new locks,
1451 * redundant unused locks are canceled locally;
1452 * - also cancel locally unused aged locks;
1453 * - do not cancel more than @max locks;
1454 * - GET the found locks and add them into the @cancels list.
1456 * A client lock can be added to the l_bl_ast list only when it is
1457 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1458 * There are the following use cases: ldlm_cancel_resource_local(),
1459 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1460 * flag properly. As any attempt to cancel a lock rely on this flag,
1461 * l_bl_ast list is accessed later without any special locking.
1463 * Calling policies for enabled lru resize:
1464 * ----------------------------------------
1465 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1466 * cancel not more than @count locks;
1468 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1469 * the beginning of lru list);
1471 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1472 * memory pressre policy function;
1474 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1476 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1477 int count, int max, ldlm_cancel_flags_t cancel_flags,
1480 ldlm_cancel_lru_policy_t pf;
1481 struct ldlm_lock *lock, *next;
1482 int added = 0, unused;
1485 cfs_spin_lock(&ns->ns_unused_lock);
1486 unused = ns->ns_nr_unused;
1488 if (!ns_connect_lru_resize(ns))
1489 count += unused - ns->ns_max_unused;
1491 pf = ldlm_cancel_lru_policy(ns, flags);
1492 LASSERT(pf != NULL);
1494 while (!cfs_list_empty(&ns->ns_unused_list)) {
1495 /* For any flags, stop scanning if @max is reached. */
1496 if (max && added >= max)
1499 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1501 /* No locks which got blocking requests. */
1502 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1504 /* Somebody is already doing CANCEL. No need in this
1505 * lock in lru, do not traverse it again. */
1506 if (!(lock->l_flags & LDLM_FL_CANCELING))
1509 ldlm_lock_remove_from_lru_nolock(lock);
1511 if (&lock->l_lru == &ns->ns_unused_list)
1514 LDLM_LOCK_GET(lock);
1515 cfs_spin_unlock(&ns->ns_unused_lock);
1516 lu_ref_add(&lock->l_reference, __FUNCTION__, cfs_current());
1518 /* Pass the lock through the policy filter and see if it
1519 * should stay in lru.
1521 * Even for shrinker policy we stop scanning if
1522 * we find a lock that should stay in the cache.
1523 * We should take into account lock age anyway
1524 * as new lock even if it is small of weight is
1525 * valuable resource.
1527 * That is, for shrinker policy we drop only
1528 * old locks, but additionally chose them by
1529 * their weight. Big extent locks will stay in
1531 if (pf(ns, lock, unused, added, count) ==
1532 LDLM_POLICY_KEEP_LOCK) {
1533 lu_ref_del(&lock->l_reference,
1534 __FUNCTION__, cfs_current());
1535 LDLM_LOCK_RELEASE(lock);
1536 cfs_spin_lock(&ns->ns_unused_lock);
1540 lock_res_and_lock(lock);
1541 /* Check flags again under the lock. */
1542 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1543 (ldlm_lock_remove_from_lru(lock) == 0)) {
1544 /* other thread is removing lock from lru or
1545 * somebody is already doing CANCEL or
1546 * there is a blocking request which will send
1547 * cancel by itseft or the lock is matched
1548 * is already not unused. */
1549 unlock_res_and_lock(lock);
1550 lu_ref_del(&lock->l_reference,
1551 __FUNCTION__, cfs_current());
1552 LDLM_LOCK_RELEASE(lock);
1553 cfs_spin_lock(&ns->ns_unused_lock);
1556 LASSERT(!lock->l_readers && !lock->l_writers);
1558 /* If we have chosen to cancel this lock voluntarily, we
1559 * better send cancel notification to server, so that it
1560 * frees appropriate state. This might lead to a race
1561 * where while we are doing cancel here, server is also
1562 * silently cancelling this lock. */
1563 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1565 /* Setting the CBPENDING flag is a little misleading,
1566 * but prevents an important race; namely, once
1567 * CBPENDING is set, the lock can accumulate no more
1568 * readers/writers. Since readers and writers are
1569 * already zero here, ldlm_lock_decref() won't see
1570 * this flag and call l_blocking_ast */
1571 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1573 /* We can't re-add to l_lru as it confuses the
1574 * refcounting in ldlm_lock_remove_from_lru() if an AST
1575 * arrives after we drop ns_lock below. We use l_bl_ast
1576 * and can't use l_pending_chain as it is used both on
1577 * server and client nevertheless bug 5666 says it is
1578 * used only on server */
1579 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1580 cfs_list_add(&lock->l_bl_ast, cancels);
1581 unlock_res_and_lock(lock);
1582 lu_ref_del(&lock->l_reference, __FUNCTION__, cfs_current());
1583 cfs_spin_lock(&ns->ns_unused_lock);
1587 cfs_spin_unlock(&ns->ns_unused_lock);
1588 RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
1591 /* when called with LDLM_ASYNC the blocking callback will be handled
1592 * in a thread and this function will return after the thread has been
1593 * asked to call the callback. when called with LDLM_SYNC the blocking
1594 * callback will be performed in this function. */
1595 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1598 CFS_LIST_HEAD(cancels);
1603 sync = LDLM_SYNC; /* force to be sync in user space */
1605 count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
1606 if (sync == LDLM_ASYNC) {
1607 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1612 /* If an error occured in ASYNC mode, or this is SYNC mode,
1613 * cancel the list. */
1614 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1618 /* Find and cancel locally unused locks found on resource, matched to the
1619 * given policy, mode. GET the found locks and add them into the @cancels
1621 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1622 cfs_list_t *cancels,
1623 ldlm_policy_data_t *policy,
1624 ldlm_mode_t mode, int lock_flags,
1625 ldlm_cancel_flags_t cancel_flags, void *opaque)
1627 struct ldlm_lock *lock;
1632 cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1633 if (opaque != NULL && lock->l_ast_data != opaque) {
1634 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1635 lock->l_ast_data, opaque);
1640 if (lock->l_readers || lock->l_writers)
1643 /* If somebody is already doing CANCEL, or blocking ast came,
1644 * skip this lock. */
1645 if (lock->l_flags & LDLM_FL_BL_AST ||
1646 lock->l_flags & LDLM_FL_CANCELING)
1649 if (lockmode_compat(lock->l_granted_mode, mode))
1652 /* If policy is given and this is IBITS lock, add to list only
1653 * those locks that match by policy. */
1654 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1655 !(lock->l_policy_data.l_inodebits.bits &
1656 policy->l_inodebits.bits))
1659 /* See CBPENDING comment in ldlm_cancel_lru */
1660 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1663 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1664 cfs_list_add(&lock->l_bl_ast, cancels);
1665 LDLM_LOCK_GET(lock);
1670 RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
1673 /* If @req is NULL, send CANCEL request to server with handles of locks
1674 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1675 * separately per lock.
1676 * If @req is not NULL, put handles of locks in @cancels into the request
1677 * buffer at the offset @off.
1678 * Destroy @cancels at the end. */
1679 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1680 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1682 struct ldlm_lock *lock;
1686 if (cfs_list_empty(cancels) || count == 0)
1689 /* XXX: requests (both batched and not) could be sent in parallel.
1690 * Usually it is enough to have just 1 RPC, but it is possible that
1691 * there are to many locks to be cancelled in LRU or on a resource.
1692 * It would also speed up the case when the server does not support
1695 LASSERT(!cfs_list_empty(cancels));
1696 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1698 LASSERT(lock->l_conn_export);
1700 if (exp_connect_cancelset(lock->l_conn_export)) {
1703 ldlm_cancel_pack(req, cancels, count);
1705 res = ldlm_cli_cancel_req(lock->l_conn_export,
1709 res = ldlm_cli_cancel_req(lock->l_conn_export,
1714 CERROR("ldlm_cli_cancel_list: %d\n", res);
1719 ldlm_lock_list_put(cancels, l_bl_ast, res);
1721 LASSERT(count == 0);
1725 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1726 const struct ldlm_res_id *res_id,
1727 ldlm_policy_data_t *policy,
1729 ldlm_cancel_flags_t flags,
1732 struct ldlm_resource *res;
1733 CFS_LIST_HEAD(cancels);
1738 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1740 /* This is not a problem. */
1741 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1745 LDLM_RESOURCE_ADDREF(res);
1746 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1748 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1750 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1752 LDLM_RESOURCE_DELREF(res);
1753 ldlm_resource_putref(res);
1757 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1759 int no_resource = 0;
1761 cfs_spin_lock(&ns->ns_hash_lock);
1762 if (ns->ns_resources == 0)
1764 cfs_spin_unlock(&ns->ns_hash_lock);
1766 RETURN(no_resource);
1769 /* Cancel all locks on a namespace (or a specific resource, if given)
1770 * that have 0 readers/writers.
1772 * If flags & LCF_LOCAL, throw the locks away without trying
1773 * to notify the server. */
1774 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1775 const struct ldlm_res_id *res_id,
1776 ldlm_cancel_flags_t flags, void *opaque)
1785 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1789 cfs_spin_lock(&ns->ns_hash_lock);
1790 for (i = 0; i < RES_HASH_SIZE; i++) {
1792 tmp = ns->ns_hash[i].next;
1793 while (tmp != &(ns->ns_hash[i])) {
1794 struct ldlm_resource *res;
1797 res = cfs_list_entry(tmp, struct ldlm_resource,
1799 ldlm_resource_getref(res);
1800 cfs_spin_unlock(&ns->ns_hash_lock);
1802 LDLM_RESOURCE_ADDREF(res);
1803 rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1808 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1809 res->lr_name.name[0], rc);
1811 LDLM_RESOURCE_DELREF(res);
1812 cfs_spin_lock(&ns->ns_hash_lock);
1814 ldlm_resource_putref_locked(res);
1817 cfs_spin_unlock(&ns->ns_hash_lock);
1822 /* Lock iterators. */
1824 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1827 cfs_list_t *tmp, *next;
1828 struct ldlm_lock *lock;
1829 int rc = LDLM_ITER_CONTINUE;
1834 RETURN(LDLM_ITER_CONTINUE);
1837 cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
1838 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1840 if (iter(lock, closure) == LDLM_ITER_STOP)
1841 GOTO(out, rc = LDLM_ITER_STOP);
1844 cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
1845 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1847 if (iter(lock, closure) == LDLM_ITER_STOP)
1848 GOTO(out, rc = LDLM_ITER_STOP);
1851 cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
1852 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
1854 if (iter(lock, closure) == LDLM_ITER_STOP)
1855 GOTO(out, rc = LDLM_ITER_STOP);
1862 struct iter_helper_data {
1863 ldlm_iterator_t iter;
1867 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1869 struct iter_helper_data *helper = closure;
1870 return helper->iter(lock, helper->closure);
1873 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1875 return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1878 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1881 struct iter_helper_data helper = { iter: iter, closure: closure };
1882 return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1885 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1886 ldlm_res_iterator_t iter, void *closure)
1888 int i, rc = LDLM_ITER_CONTINUE;
1889 struct ldlm_resource *res;
1893 cfs_spin_lock(&ns->ns_hash_lock);
1894 for (i = 0; i < RES_HASH_SIZE; i++) {
1895 tmp = ns->ns_hash[i].next;
1896 while (tmp != &(ns->ns_hash[i])) {
1897 res = cfs_list_entry(tmp, struct ldlm_resource,
1899 ldlm_resource_getref(res);
1900 cfs_spin_unlock(&ns->ns_hash_lock);
1901 LDLM_RESOURCE_ADDREF(res);
1903 rc = iter(res, closure);
1905 LDLM_RESOURCE_DELREF(res);
1906 cfs_spin_lock(&ns->ns_hash_lock);
1908 ldlm_resource_putref_locked(res);
1909 if (rc == LDLM_ITER_STOP)
1914 cfs_spin_unlock(&ns->ns_hash_lock);
1918 /* non-blocking function to manipulate a lock whose cb_data is being put away.
1919 * return 0: find no resource
1920 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
1923 int ldlm_resource_iterate(struct ldlm_namespace *ns,
1924 const struct ldlm_res_id *res_id,
1925 ldlm_iterator_t iter, void *data)
1927 struct ldlm_resource *res;
1932 CERROR("must pass in namespace\n");
1936 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1940 LDLM_RESOURCE_ADDREF(res);
1941 rc = ldlm_resource_foreach(res, iter, data);
1942 LDLM_RESOURCE_DELREF(res);
1943 ldlm_resource_putref(res);
1949 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1951 cfs_list_t *list = closure;
1953 /* we use l_pending_chain here, because it's unused on clients. */
1954 LASSERTF(cfs_list_empty(&lock->l_pending_chain),
1955 "lock %p next %p prev %p\n",
1956 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1957 /* bug 9573: don't replay locks left after eviction, or
1958 * bug 17614: locks being actively cancelled. Get a reference
1959 * on a lock so that it does not disapear under us (e.g. due to cancel)
1961 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
1962 cfs_list_add(&lock->l_pending_chain, list);
1963 LDLM_LOCK_GET(lock);
1966 return LDLM_ITER_CONTINUE;
1969 static int replay_lock_interpret(const struct lu_env *env,
1970 struct ptlrpc_request *req,
1971 struct ldlm_async_args *aa, int rc)
1973 struct lustre_handle old_hash_key;
1974 struct ldlm_lock *lock;
1975 struct ldlm_reply *reply;
1976 struct obd_export *exp;
1979 cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
1984 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1986 GOTO(out, rc = -EPROTO);
1988 lock = ldlm_handle2lock(&aa->lock_handle);
1990 CERROR("received replay ack for unknown local cookie "LPX64
1991 " remote cookie "LPX64 " from server %s id %s\n",
1992 aa->lock_handle.cookie, reply->lock_handle.cookie,
1993 req->rq_export->exp_client_uuid.uuid,
1994 libcfs_id2str(req->rq_peer));
1995 GOTO(out, rc = -ESTALE);
1998 old_hash_key = lock->l_remote_handle;
1999 lock->l_remote_handle = reply->lock_handle;
2001 /* Key change rehash lock in per-export hash with new key */
2002 exp = req->rq_export;
2003 if (exp && exp->exp_lock_hash)
2004 cfs_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
2005 &lock->l_remote_handle,
2008 LDLM_DEBUG(lock, "replayed lock:");
2009 ptlrpc_import_recovery_state_machine(req->rq_import);
2010 LDLM_LOCK_PUT(lock);
2013 ptlrpc_connect_import(req->rq_import, NULL);
2018 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2020 struct ptlrpc_request *req;
2021 struct ldlm_async_args *aa;
2022 struct ldlm_request *body;
2027 /* Bug 11974: Do not replay a lock which is actively being canceled */
2028 if (lock->l_flags & LDLM_FL_CANCELING) {
2029 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2033 /* If this is reply-less callback lock, we cannot replay it, since
2034 * server might have long dropped it, but notification of that event was
2035 * lost by network. (and server granted conflicting lock already) */
2036 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2037 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2038 ldlm_lock_cancel(lock);
2042 * If granted mode matches the requested mode, this lock is granted.
2044 * If they differ, but we have a granted mode, then we were granted
2045 * one mode and now want another: ergo, converting.
2047 * If we haven't been granted anything and are on a resource list,
2048 * then we're blocked/waiting.
2050 * If we haven't been granted anything and we're NOT on a resource list,
2051 * then we haven't got a reply yet and don't have a known disposition.
2052 * This happens whenever a lock enqueue is the request that triggers
2055 if (lock->l_granted_mode == lock->l_req_mode)
2056 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2057 else if (lock->l_granted_mode)
2058 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2059 else if (!cfs_list_empty(&lock->l_res_link))
2060 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2062 flags = LDLM_FL_REPLAY;
2064 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2065 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2069 /* We're part of recovery, so don't wait for it. */
2070 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2072 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2073 ldlm_lock2desc(lock, &body->lock_desc);
2074 body->lock_flags = flags;
2076 ldlm_lock2handle(lock, &body->lock_handle[0]);
2077 if (lock->l_lvb_len != 0) {
2078 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2079 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2082 ptlrpc_request_set_replen(req);
2083 /* notify the server we've replayed all requests.
2084 * also, we mark the request to be put on a dedicated
2085 * queue to be processed after all request replayes.
2087 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2089 LDLM_DEBUG(lock, "replaying lock:");
2091 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2092 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2093 aa = ptlrpc_req_async_args(req);
2094 aa->lock_handle = body->lock_handle[0];
2095 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2096 ptlrpcd_add_req(req, PSCOPE_OTHER);
2101 int ldlm_replay_locks(struct obd_import *imp)
2103 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2104 CFS_LIST_HEAD(list);
2105 struct ldlm_lock *lock, *next;
2110 LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
2112 /* don't replay locks if import failed recovery */
2113 if (imp->imp_vbr_failed)
2116 /* ensure this doesn't fall to 0 before all have been queued */
2117 cfs_atomic_inc(&imp->imp_replay_inflight);
2119 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2121 cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2122 cfs_list_del_init(&lock->l_pending_chain);
2124 LDLM_LOCK_PUT(lock);
2125 continue; /* or try to do the rest? */
2127 rc = replay_one_lock(imp, lock);
2128 LDLM_LOCK_PUT(lock);
2131 cfs_atomic_dec(&imp->imp_replay_inflight);