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);
124 static int is_granted_or_cancelled(struct ldlm_lock *lock)
128 lock_res_and_lock(lock);
129 if (((lock->l_req_mode == lock->l_granted_mode) &&
130 !(lock->l_flags & LDLM_FL_CP_REQD)) ||
131 (lock->l_flags & LDLM_FL_FAILED))
133 unlock_res_and_lock(lock);
139 * Helper function for ldlm_completion_ast(), updating timings when lock is
142 static int ldlm_completion_tail(struct ldlm_lock *lock)
147 if (lock->l_destroyed || lock->l_flags & LDLM_FL_FAILED) {
148 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
151 delay = cfs_time_sub(cfs_time_current_sec(),
152 lock->l_enqueued_time.tv_sec);
153 LDLM_DEBUG(lock, "client-side enqueue: granted after "
154 CFS_DURATION_T"s", delay);
156 /* Update our time estimate */
157 at_add(&lock->l_resource->lr_namespace->ns_at_estimate, delay);
164 * Implementation of ->l_completion_ast() for a client that doesn't wait
165 * until lock is granted. Suitable for locks enqueued through ptlrpcd or
166 * other threads that cannot block for long.
168 int ldlm_completion_ast_async(struct ldlm_lock *lock, int flags, void *data)
172 if (flags == LDLM_FL_WAIT_NOREPROC) {
173 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
177 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
178 LDLM_FL_BLOCK_CONV))) {
179 cfs_waitq_signal(&lock->l_waitq);
180 RETURN(ldlm_completion_tail(lock));
183 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
185 ldlm_lock_dump(D_OTHER, lock, 0);
190 * Client side LDLM "completion" AST. This is called in several cases:
192 * - when a reply to an ENQUEUE rpc is received from the server
193 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
194 * this point (determined by flags);
196 * - when LDLM_CP_CALLBACK rpc comes to client to notify it that lock has
199 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
202 * - to force all locks when resource is destroyed (cleanup_resource());
204 * - during lock conversion (not used currently).
206 * If lock is not granted in the first case, this function waits until second
207 * or penultimate cases happen in some other thread.
210 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data)
212 /* XXX ALLOCATE - 160 bytes */
213 struct lock_wait_data lwd;
214 struct obd_device *obd;
215 struct obd_import *imp = NULL;
216 struct l_wait_info lwi;
221 if (flags == LDLM_FL_WAIT_NOREPROC) {
222 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
226 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
227 LDLM_FL_BLOCK_CONV))) {
228 cfs_waitq_signal(&lock->l_waitq);
232 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
234 ldlm_lock_dump(D_OTHER, lock, 0);
238 obd = class_exp2obd(lock->l_conn_export);
240 /* if this is a local lock, then there is no import */
242 imp = obd->u.cli.cl_import;
245 /* Wait a long time for enqueue - server may have to callback a
246 lock from another client. Server will evict the other client if it
247 doesn't respond reasonably, and then give us the lock. */
248 timeout = ldlm_get_enq_timeout(lock) * 2;
252 if (lock->l_flags & LDLM_FL_NO_TIMEOUT) {
253 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
254 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
256 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
257 ldlm_expired_completion_wait,
258 interrupted_completion_wait, &lwd);
262 spin_lock(&imp->imp_lock);
263 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
264 spin_unlock(&imp->imp_lock);
267 /* Go to sleep until the lock is granted or cancelled. */
268 rc = l_wait_event(lock->l_waitq, is_granted_or_cancelled(lock), &lwi);
271 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
276 RETURN(ldlm_completion_tail(lock));
280 * ->l_blocking_ast() callback for LDLM locks acquired by server-side OBDs.
282 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
283 void *data, int flag)
288 if (flag == LDLM_CB_CANCELING) {
289 /* Don't need to do anything here. */
293 lock_res_and_lock(lock);
294 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
295 * that ldlm_blocking_ast is called just before intent_policy method
296 * takes the ns_lock, then by the time we get the lock, we might not
297 * be the correct blocking function anymore. So check, and return
299 if (lock->l_blocking_ast != ldlm_blocking_ast) {
300 unlock_res_and_lock(lock);
304 lock->l_flags |= LDLM_FL_CBPENDING;
305 do_ast = (!lock->l_readers && !lock->l_writers);
306 unlock_res_and_lock(lock);
309 struct lustre_handle lockh;
312 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
313 ldlm_lock2handle(lock, &lockh);
314 rc = ldlm_cli_cancel(&lockh);
316 CERROR("ldlm_cli_cancel: %d\n", rc);
318 LDLM_DEBUG(lock, "Lock still has references, will be "
325 * ->l_glimpse_ast() for DLM extent locks acquired on the server-side. See
326 * comment in filter_intent_policy() on why you may need this.
328 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
331 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for
332 * that is rather subtle: with OST-side locking, it may so happen that
333 * _all_ extent locks are held by the OST. If client wants to obtain
334 * current file size it calls ll{,u}_glimpse_size(), and (as locks are
335 * on the server), dummy glimpse callback fires and does
336 * nothing. Client still receives correct file size due to the
337 * following fragment in filter_intent_policy():
339 * rc = l->l_glimpse_ast(l, NULL); // this will update the LVB
340 * if (rc != 0 && res->lr_namespace->ns_lvbo &&
341 * res->lr_namespace->ns_lvbo->lvbo_update) {
342 * res->lr_namespace->ns_lvbo->lvbo_update(res, NULL, 0, 1);
345 * that is, after glimpse_ast() fails, filter_lvbo_update() runs, and
346 * returns correct file size to the client.
348 return -ELDLM_NO_LOCK_DATA;
351 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
352 const struct ldlm_res_id *res_id,
353 ldlm_type_t type, ldlm_policy_data_t *policy,
354 ldlm_mode_t mode, int *flags,
355 ldlm_blocking_callback blocking,
356 ldlm_completion_callback completion,
357 ldlm_glimpse_callback glimpse,
358 void *data, __u32 lvb_len, void *lvb_swabber,
359 struct lustre_handle *lockh)
361 struct ldlm_lock *lock;
363 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
364 .lcs_blocking = blocking,
365 .lcs_glimpse = glimpse,
369 LASSERT(!(*flags & LDLM_FL_REPLAY));
370 if (unlikely(ns_is_client(ns))) {
371 CERROR("Trying to enqueue local lock in a shadow namespace\n");
375 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len);
377 GOTO(out_nolock, err = -ENOMEM);
378 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
380 ldlm_lock_addref_internal(lock, mode);
381 ldlm_lock2handle(lock, lockh);
382 lock_res_and_lock(lock);
383 lock->l_flags |= LDLM_FL_LOCAL;
384 if (*flags & LDLM_FL_ATOMIC_CB)
385 lock->l_flags |= LDLM_FL_ATOMIC_CB;
386 lock->l_lvb_swabber = lvb_swabber;
387 unlock_res_and_lock(lock);
389 lock->l_policy_data = *policy;
390 if (type == LDLM_EXTENT)
391 lock->l_req_extent = policy->l_extent;
393 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
394 if (unlikely(err != ELDLM_OK))
398 *policy = lock->l_policy_data;
400 LDLM_DEBUG_NOLOCK("client-side local enqueue handler END (lock %p)",
403 if (lock->l_completion_ast)
404 lock->l_completion_ast(lock, *flags, NULL);
406 LDLM_DEBUG(lock, "client-side local enqueue END");
414 static void failed_lock_cleanup(struct ldlm_namespace *ns,
415 struct ldlm_lock *lock,
416 struct lustre_handle *lockh, int mode)
418 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
419 lock_res_and_lock(lock);
420 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
421 unlock_res_and_lock(lock);
422 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
424 ldlm_lock_decref_and_cancel(lockh, mode);
426 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
427 * from llite/file.c/ll_file_flock(). */
428 if (lock->l_resource->lr_type == LDLM_FLOCK) {
429 ldlm_lock_destroy(lock);
433 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
434 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
435 int *flags, void *lvb, __u32 lvb_len,
436 void *lvb_swabber, struct lustre_handle *lockh,int rc)
438 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
439 int is_replay = *flags & LDLM_FL_REPLAY;
440 struct lustre_handle old_hash_key;
441 struct ldlm_lock *lock;
442 struct ldlm_reply *reply;
443 int cleanup_phase = 1;
446 lock = ldlm_handle2lock(lockh);
447 /* ldlm_cli_enqueue is holding a reference on this lock. */
449 LASSERT(type == LDLM_FLOCK);
453 if (rc != ELDLM_OK) {
455 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
456 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
457 if (rc == ELDLM_LOCK_ABORTED) {
458 /* Before we return, swab the reply */
459 reply = req_capsule_server_get(&req->rq_pill,
464 struct ost_lvb *tmplvb;
466 req_capsule_set_size(&req->rq_pill,
467 &RMF_DLM_LVB, RCL_SERVER,
469 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
473 GOTO(cleanup, rc = -EPROTO);
475 memcpy(lvb, tmplvb, lvb_len);
481 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
483 GOTO(cleanup, rc = -EPROTO);
485 /* lock enqueued on the server */
488 lock_res_and_lock(lock);
489 old_hash_key = lock->l_remote_handle;
490 lock->l_remote_handle = reply->lock_handle;
492 /* Key change rehash lock in per-export hash with new key */
493 if (exp->exp_lock_hash)
494 lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
495 &lock->l_remote_handle,
498 *flags = reply->lock_flags;
499 lock->l_flags |= reply->lock_flags & LDLM_INHERIT_FLAGS;
500 /* move NO_TIMEOUT flag to the lock to force ldlm_lock_match()
501 * to wait with no timeout as well */
502 lock->l_flags |= reply->lock_flags & LDLM_FL_NO_TIMEOUT;
503 unlock_res_and_lock(lock);
505 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: 0x%x\n",
506 lock, reply->lock_handle.cookie, *flags);
508 /* If enqueue returned a blocked lock but the completion handler has
509 * already run, then it fixed up the resource and we don't need to do it
511 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
512 int newmode = reply->lock_desc.l_req_mode;
514 if (newmode && newmode != lock->l_req_mode) {
515 LDLM_DEBUG(lock, "server returned different mode %s",
516 ldlm_lockname[newmode]);
517 lock->l_req_mode = newmode;
520 if (memcmp(reply->lock_desc.l_resource.lr_name.name,
521 lock->l_resource->lr_name.name,
522 sizeof(struct ldlm_res_id))) {
523 CDEBUG(D_INFO, "remote intent success, locking "
524 "(%ld,%ld,%ld) instead of "
526 (long)reply->lock_desc.l_resource.lr_name.name[0],
527 (long)reply->lock_desc.l_resource.lr_name.name[1],
528 (long)reply->lock_desc.l_resource.lr_name.name[2],
529 (long)lock->l_resource->lr_name.name[0],
530 (long)lock->l_resource->lr_name.name[1],
531 (long)lock->l_resource->lr_name.name[2]);
533 rc = ldlm_lock_change_resource(ns, lock,
534 &reply->lock_desc.l_resource.lr_name);
535 if (rc || lock->l_resource == NULL)
536 GOTO(cleanup, rc = -ENOMEM);
537 LDLM_DEBUG(lock, "client-side enqueue, new resource");
540 if (!(type == LDLM_IBITS && !(exp->exp_connect_flags &
542 lock->l_policy_data =
543 reply->lock_desc.l_policy_data;
544 if (type != LDLM_PLAIN)
545 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
548 if ((*flags) & LDLM_FL_AST_SENT ||
549 /* Cancel extent locks as soon as possible on a liblustre client,
550 * because it cannot handle asynchronous ASTs robustly (see
552 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
553 lock_res_and_lock(lock);
554 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
555 unlock_res_and_lock(lock);
556 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
559 /* If the lock has already been granted by a completion AST, don't
560 * clobber the LVB with an older one. */
561 if (lvb_len && (lock->l_req_mode != lock->l_granted_mode)) {
564 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
566 tmplvb = req_capsule_server_swab_get(&req->rq_pill,
570 GOTO(cleanup, rc = -EPROTO);
571 memcpy(lock->l_lvb_data, tmplvb, lvb_len);
575 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
576 if (lock->l_completion_ast != NULL) {
577 int err = lock->l_completion_ast(lock, *flags, NULL);
580 if (rc && type != LDLM_FLOCK) /* bug 9425, bug 10250 */
585 if (lvb_len && lvb != NULL) {
586 /* Copy the LVB here, and not earlier, because the completion
587 * AST (if any) can override what we got in the reply */
588 memcpy(lvb, lock->l_lvb_data, lvb_len);
591 LDLM_DEBUG(lock, "client-side enqueue END");
594 if (cleanup_phase == 1 && rc)
595 failed_lock_cleanup(ns, lock, lockh, mode);
596 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
602 /* PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
603 * a single page on the send/receive side. XXX: 512 should be changed
604 * to more adequate value. */
605 static inline int ldlm_req_handles_avail(int req_size, int off)
609 avail = min_t(int, LDLM_MAXREQSIZE, CFS_PAGE_SIZE - 512) - req_size;
610 avail /= sizeof(struct lustre_handle);
611 avail += LDLM_LOCKREQ_HANDLES - off;
616 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
617 enum req_location loc,
620 int size = req_capsule_msg_size(pill, loc);
621 return ldlm_req_handles_avail(size, off);
624 static inline int ldlm_format_handles_avail(struct obd_import *imp,
625 const struct req_format *fmt,
626 enum req_location loc, int off)
628 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
629 return ldlm_req_handles_avail(size, off);
632 /* Cancel lru locks and pack them into the enqueue request. Pack there the given
633 * @count locks in @cancels. */
634 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
635 int version, int opc, int canceloff,
636 struct list_head *cancels, int count)
638 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
639 struct req_capsule *pill = &req->rq_pill;
640 struct ldlm_request *dlm = NULL;
641 int flags, avail, to_free, bufcount, pack = 0;
648 if (exp_connect_cancelset(exp)) {
649 /* Estimate the amount of available space in the request. */
650 bufcount = req_capsule_filled_sizes(pill, RCL_CLIENT);
651 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
653 flags = ns_connect_lru_resize(ns) ?
654 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
655 to_free = !ns_connect_lru_resize(ns) &&
656 opc == LDLM_ENQUEUE ? 1 : 0;
658 /* Cancel lru locks here _only_ if the server supports
659 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
660 * rpc, what will make us slower. */
662 count += ldlm_cancel_lru_local(ns, cancels, to_free,
663 avail - count, 0, flags);
668 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
669 ldlm_request_bufsize(pack, opc));
672 rc = ptlrpc_request_pack(req, version, opc);
674 ldlm_lock_list_put(cancels, l_bl_ast, count);
678 if (exp_connect_cancelset(exp)) {
680 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
682 /* Skip first lock handler in ldlm_request_pack(),
683 * this method will incrment @lock_count according
684 * to the lock handle amount actually written to
686 dlm->lock_count = canceloff;
688 /* Pack into the request @pack lock handles. */
689 ldlm_cli_cancel_list(cancels, pack, req, 0);
690 /* Prepare and send separate cancel rpc for others. */
691 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
693 ldlm_lock_list_put(cancels, l_bl_ast, count);
698 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
699 struct list_head *cancels, int count)
701 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
702 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
705 /* If a request has some specific initialisation it is passed in @reqp,
706 * otherwise it is created in ldlm_cli_enqueue.
708 * Supports sync and async requests, pass @async flag accordingly. If a
709 * request was created in ldlm_cli_enqueue and it is the async request,
710 * pass it to the caller in @reqp. */
711 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
712 struct ldlm_enqueue_info *einfo,
713 const struct ldlm_res_id *res_id,
714 ldlm_policy_data_t *policy, int *flags,
715 void *lvb, __u32 lvb_len, void *lvb_swabber,
716 struct lustre_handle *lockh, int async)
718 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
719 struct ldlm_lock *lock;
720 struct ldlm_request *body;
721 int is_replay = *flags & LDLM_FL_REPLAY;
722 int req_passed_in = 1;
724 struct ptlrpc_request *req;
727 LASSERT(exp != NULL);
729 /* If we're replaying this lock, just check some invariants.
730 * If we're creating a new lock, get everything all setup nice. */
732 lock = ldlm_handle2lock_long(lockh, 0);
733 LASSERT(lock != NULL);
734 LDLM_DEBUG(lock, "client-side enqueue START");
735 LASSERT(exp == lock->l_conn_export);
737 const struct ldlm_callback_suite cbs = {
738 .lcs_completion = einfo->ei_cb_cp,
739 .lcs_blocking = einfo->ei_cb_bl,
740 .lcs_glimpse = einfo->ei_cb_gl,
741 .lcs_weigh = einfo->ei_cb_wg
743 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
744 einfo->ei_mode, &cbs, einfo->ei_cbdata,
748 /* for the local lock, add the reference */
749 ldlm_lock_addref_internal(lock, einfo->ei_mode);
750 ldlm_lock2handle(lock, lockh);
751 lock->l_lvb_swabber = lvb_swabber;
752 if (policy != NULL) {
753 /* INODEBITS_INTEROP: If the server does not support
754 * inodebits, we will request a plain lock in the
755 * descriptor (ldlm_lock2desc() below) but use an
756 * inodebits lock internally with both bits set.
758 if (einfo->ei_type == LDLM_IBITS &&
759 !(exp->exp_connect_flags & OBD_CONNECT_IBITS))
760 lock->l_policy_data.l_inodebits.bits =
761 MDS_INODELOCK_LOOKUP |
762 MDS_INODELOCK_UPDATE;
764 lock->l_policy_data = *policy;
767 if (einfo->ei_type == LDLM_EXTENT)
768 lock->l_req_extent = policy->l_extent;
769 LDLM_DEBUG(lock, "client-side enqueue START");
772 /* lock not sent to server yet */
774 if (reqp == NULL || *reqp == NULL) {
775 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
780 failed_lock_cleanup(ns, lock, lockh, einfo->ei_mode);
791 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
793 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
794 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
797 lock->l_conn_export = exp;
798 lock->l_export = NULL;
799 lock->l_blocking_ast = einfo->ei_cb_bl;
801 /* Dump lock data into the request buffer */
802 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
803 ldlm_lock2desc(lock, &body->lock_desc);
804 body->lock_flags = *flags;
805 body->lock_handle[0] = *lockh;
807 /* Continue as normal. */
808 if (!req_passed_in) {
810 req_capsule_extend(&req->rq_pill,
811 &RQF_LDLM_ENQUEUE_LVB);
812 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
813 RCL_SERVER, lvb_len);
815 ptlrpc_request_set_replen(req);
819 * Liblustre client doesn't get extent locks, except for O_APPEND case
820 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
821 * [i_size, OBD_OBJECT_EOF] lock is taken.
823 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
824 policy->l_extent.end == OBD_OBJECT_EOF));
827 LASSERT(reqp != NULL);
831 LDLM_DEBUG(lock, "sending request");
832 rc = ptlrpc_queue_wait(req);
833 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
834 einfo->ei_mode, flags, lvb, lvb_len,
835 lvb_swabber, lockh, rc);
837 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
838 * one reference that we took */
844 if (!req_passed_in && req != NULL) {
845 ptlrpc_req_finished(req);
853 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
856 struct ldlm_resource *res;
859 if (ns_is_client(lock->l_resource->lr_namespace)) {
860 CERROR("Trying to cancel local lock\n");
863 LDLM_DEBUG(lock, "client-side local convert");
865 res = ldlm_lock_convert(lock, new_mode, flags);
867 ldlm_reprocess_all(res);
872 LDLM_DEBUG(lock, "client-side local convert handler END");
877 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
878 * conversion of locks which are on the waiting or converting queue */
879 /* Caller of this code is supposed to take care of lock readers/writers
881 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
883 struct ldlm_request *body;
884 struct ldlm_reply *reply;
885 struct ldlm_lock *lock;
886 struct ldlm_resource *res;
887 struct ptlrpc_request *req;
891 lock = ldlm_handle2lock(lockh);
898 if (lock->l_conn_export == NULL)
899 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
901 LDLM_DEBUG(lock, "client-side convert");
903 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
904 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
911 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
912 body->lock_handle[0] = lock->l_remote_handle;
914 body->lock_desc.l_req_mode = new_mode;
915 body->lock_flags = *flags;
918 ptlrpc_request_set_replen(req);
919 rc = ptlrpc_queue_wait(req);
923 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
925 GOTO(out, rc = -EPROTO);
928 GOTO(out, rc = req->rq_status);
930 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
932 ldlm_reprocess_all(res);
933 /* Go to sleep until the lock is granted. */
934 /* FIXME: or cancelled. */
935 if (lock->l_completion_ast) {
936 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
947 ptlrpc_req_finished(req);
951 /* Cancel locks locally.
953 * LDLM_FL_LOCAL_ONLY if tere is no need in a CANCEL rpc to the server;
954 * LDLM_FL_CANCELING otherwise;
955 * LDLM_FL_BL_AST if there is a need in a separate CANCEL rpc. */
956 static int ldlm_cli_cancel_local(struct ldlm_lock *lock)
958 int rc = LDLM_FL_LOCAL_ONLY;
961 if (lock->l_conn_export) {
964 LDLM_DEBUG(lock, "client-side cancel");
965 /* Set this flag to prevent others from getting new references*/
966 lock_res_and_lock(lock);
967 lock->l_flags |= LDLM_FL_CBPENDING;
968 local_only = (lock->l_flags &
969 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
970 ldlm_cancel_callback(lock);
971 rc = (lock->l_flags & LDLM_FL_BL_AST) ?
972 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
973 unlock_res_and_lock(lock);
976 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
978 rc = LDLM_FL_LOCAL_ONLY;
980 ldlm_lock_cancel(lock);
982 if (ns_is_client(lock->l_resource->lr_namespace)) {
983 LDLM_ERROR(lock, "Trying to cancel local lock");
986 LDLM_DEBUG(lock, "server-side local cancel");
987 ldlm_lock_cancel(lock);
988 ldlm_reprocess_all(lock->l_resource);
989 LDLM_DEBUG(lock, "server-side local cancel handler END");
995 /* Pack @count locks in @head into ldlm_request buffer at the offset @off,
996 of the request @req. */
997 static void ldlm_cancel_pack(struct ptlrpc_request *req,
998 struct list_head *head, int count)
1000 struct ldlm_request *dlm;
1001 struct ldlm_lock *lock;
1002 int max, packed = 0;
1005 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1006 LASSERT(dlm != NULL);
1008 /* Check the room in the request buffer. */
1009 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1010 sizeof(struct ldlm_request);
1011 max /= sizeof(struct lustre_handle);
1012 max += LDLM_LOCKREQ_HANDLES;
1013 LASSERT(max >= dlm->lock_count + count);
1015 /* XXX: it would be better to pack lock handles grouped by resource.
1016 * so that the server cancel would call filter_lvbo_update() less
1018 list_for_each_entry(lock, head, l_bl_ast) {
1021 LASSERT(lock->l_conn_export);
1022 /* Pack the lock handle to the given request buffer. */
1023 LDLM_DEBUG(lock, "packing");
1024 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1027 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1031 /* Prepare and send a batched cancel rpc, it will include count lock handles
1032 * of locks given in @head. */
1033 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1034 int count, int flags)
1036 struct ptlrpc_request *req = NULL;
1037 struct obd_import *imp;
1042 LASSERT(exp != NULL);
1045 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, obd_fail_val);
1047 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1050 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1051 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1058 imp = class_exp2cliimp(exp);
1059 if (imp == NULL || imp->imp_invalid) {
1061 "skipping cancel on invalid import %p\n", imp);
1065 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1067 GOTO(out, rc = -ENOMEM);
1069 bufcount = req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1070 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1071 ldlm_request_bufsize(count, LDLM_CANCEL));
1073 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1075 ptlrpc_request_free(req);
1078 req->rq_no_resend = 1;
1079 req->rq_no_delay = 1;
1081 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1082 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1083 ptlrpc_at_set_req_timeout(req);
1085 ldlm_cancel_pack(req, cancels, count);
1087 ptlrpc_request_set_replen(req);
1088 if (flags & LDLM_FL_ASYNC) {
1089 ptlrpcd_add_req(req);
1093 rc = ptlrpc_queue_wait(req);
1096 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1097 "out of sync -- not fatal\n",
1098 libcfs_nid2str(req->rq_import->
1099 imp_connection->c_peer.nid));
1101 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1102 req->rq_import_generation == imp->imp_generation) {
1103 ptlrpc_req_finished(req);
1105 } else if (rc != ELDLM_OK) {
1106 CERROR("Got rc %d from cancel RPC: canceling "
1114 ptlrpc_req_finished(req);
1117 return sent ? sent : rc;
1120 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1122 LASSERT(imp != NULL);
1123 return &imp->imp_obd->obd_namespace->ns_pool;
1127 * Update client's obd pool related fields with new SLV and Limit from \a req.
1129 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1131 struct obd_device *obd;
1132 __u64 old_slv, new_slv;
1136 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1137 !imp_connect_lru_resize(req->rq_import)))
1140 * Do nothing for corner cases.
1146 * In some cases RPC may contain slv and limit zeroed out. This is
1147 * the case when server does not support lru resize feature. This is
1148 * also possible in some recovery cases when server side reqs have no
1149 * ref to obd export and thus access to server side namespace is no
1152 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1153 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1154 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1155 "(SLV: "LPU64", Limit: %u)",
1156 lustre_msg_get_slv(req->rq_repmsg),
1157 lustre_msg_get_limit(req->rq_repmsg));
1161 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1162 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1163 obd = req->rq_import->imp_obd;
1166 * Set new SLV and Limit to obd fields to make accessible for pool
1167 * thread. We do not access obd_namespace and pool directly here
1168 * as there is no reliable way to make sure that they are still
1169 * alive in cleanup time. Evil races are possible which may cause
1170 * oops in that time.
1172 write_lock(&obd->obd_pool_lock);
1173 old_slv = obd->obd_pool_slv;
1174 obd->obd_pool_slv = new_slv;
1175 obd->obd_pool_limit = new_limit;
1176 write_unlock(&obd->obd_pool_lock);
1180 EXPORT_SYMBOL(ldlm_cli_update_pool);
1182 int ldlm_cli_cancel(struct lustre_handle *lockh)
1184 struct obd_export *exp;
1185 int avail, flags, count = 1, rc = 0;
1186 struct ldlm_namespace *ns;
1187 struct ldlm_lock *lock;
1188 CFS_LIST_HEAD(cancels);
1191 /* concurrent cancels on the same handle can happen */
1192 lock = __ldlm_handle2lock(lockh, LDLM_FL_CANCELING);
1194 LDLM_DEBUG_NOLOCK("lock is already being destroyed\n");
1198 rc = ldlm_cli_cancel_local(lock);
1199 if (rc < 0 || rc == LDLM_FL_LOCAL_ONLY) {
1200 LDLM_LOCK_PUT(lock);
1201 RETURN(rc < 0 ? rc : 0);
1203 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1204 * rpc which goes to canceld portal, so we can cancel other lru locks
1205 * here and send them all as one LDLM_CANCEL rpc. */
1206 LASSERT(list_empty(&lock->l_bl_ast));
1207 list_add(&lock->l_bl_ast, &cancels);
1209 exp = lock->l_conn_export;
1210 if (exp_connect_cancelset(exp)) {
1211 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1216 ns = lock->l_resource->lr_namespace;
1217 flags = ns_connect_lru_resize(ns) ?
1218 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1219 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1220 LDLM_FL_BL_AST, flags);
1222 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1226 /* XXX until we will have compound requests and can cut cancels from generic rpc
1227 * we need send cancels with LDLM_FL_BL_AST flag as separate rpc */
1228 static int ldlm_cancel_list(struct list_head *cancels, int count, int flags)
1230 CFS_LIST_HEAD(head);
1231 struct ldlm_lock *lock, *next;
1232 int left = 0, bl_ast = 0, rc;
1235 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1239 if (flags & LDLM_FL_LOCAL_ONLY) {
1240 rc = LDLM_FL_LOCAL_ONLY;
1241 ldlm_lock_cancel(lock);
1243 rc = ldlm_cli_cancel_local(lock);
1245 if (!(flags & LDLM_FL_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1246 LDLM_DEBUG(lock, "Cancel lock separately");
1247 list_del_init(&lock->l_bl_ast);
1248 list_add(&lock->l_bl_ast, &head);
1252 if (rc == LDLM_FL_LOCAL_ONLY) {
1253 /* CANCEL RPC should not be sent to server. */
1254 list_del_init(&lock->l_bl_ast);
1255 LDLM_LOCK_PUT(lock);
1262 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1269 * Callback function for shrink policy. Makes decision whether to keep
1270 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1271 * \a added and number of locks to be preferably canceled \a count.
1273 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1275 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1277 static ldlm_policy_res_t ldlm_cancel_shrink_policy(struct ldlm_namespace *ns,
1278 struct ldlm_lock *lock,
1279 int unused, int added,
1286 * Stop lru processing when we reached passed @count or checked all
1289 if (count && added >= count)
1290 return LDLM_POLICY_KEEP_LOCK;
1292 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1293 struct ldlm_extent *l_extent;
1296 * For all extent locks cost is 1 + number of pages in
1299 l_extent = &lock->l_policy_data.l_extent;
1300 page_nr = (l_extent->end - l_extent->start);
1301 do_div(page_nr, CFS_PAGE_SIZE);
1305 * XXX: In fact this is evil hack, we can't access inode
1306 * here. For doing it right we need somehow to have number
1307 * of covered by lock. This should be fixed later when 10718
1310 if (lock->l_ast_data != NULL) {
1311 struct inode *inode = lock->l_ast_data;
1312 if (page_nr > inode->i_mapping->nrpages)
1313 page_nr = inode->i_mapping->nrpages;
1316 lock_cost = 1 + page_nr;
1319 * For all locks which are not extent ones cost is 1
1325 * Keep all expensive locks in lru for the memory pressure time
1326 * cancel policy. They anyways may be canceled by lru resize
1327 * pplicy if they have not small enough CLV.
1329 return lock_cost > ns->ns_shrink_thumb ?
1330 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1334 * Callback function for lru-resize policy. Makes decision whether to keep
1335 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1336 * \a added and number of locks to be preferably canceled \a count.
1338 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1340 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1342 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1343 struct ldlm_lock *lock,
1344 int unused, int added,
1347 cfs_time_t cur = cfs_time_current();
1348 struct ldlm_pool *pl = &ns->ns_pool;
1353 * Stop lru processing when we reached passed @count or checked all
1356 if (count && added >= count)
1357 return LDLM_POLICY_KEEP_LOCK;
1359 slv = ldlm_pool_get_slv(pl);
1360 lvf = ldlm_pool_get_lvf(pl);
1361 la = cfs_duration_sec(cfs_time_sub(cur,
1362 lock->l_last_used));
1365 * Stop when slv is not yet come from server or lv is smaller than
1368 lv = lvf * la * unused;
1371 * Inform pool about current CLV to see it via proc.
1373 ldlm_pool_set_clv(pl, lv);
1374 return (slv == 1 || lv < slv) ?
1375 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1379 * Callback function for proc used policy. Makes decision whether to keep
1380 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1381 * \a added and number of locks to be preferably canceled \a count.
1383 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1385 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1387 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1388 struct ldlm_lock *lock,
1389 int unused, int added,
1393 * Stop lru processing when we reached passed @count or checked all
1396 return (added >= count) ?
1397 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1401 * Callback function for aged policy. Makes decision whether to keep
1402 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1403 * \a added and number of locks to be preferably canceled \a count.
1405 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1407 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1409 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1410 struct ldlm_lock *lock,
1411 int unused, int added,
1415 * Stop lru processing if young lock is found and we reached passed
1418 return ((added >= count) &&
1419 cfs_time_before(cfs_time_current(),
1420 cfs_time_add(lock->l_last_used,
1422 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1426 * Callback function for default policy. Makes decision whether to keep
1427 * \a lock in LRU for current \a LRU size \a unused, added in current scan
1428 * \a added and number of locks to be preferably canceled \a count.
1430 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1432 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1434 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1435 struct ldlm_lock *lock,
1436 int unused, int added,
1440 * Stop lru processing when we reached passed @count or checked all
1443 return (added >= count) ?
1444 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1447 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1448 struct ldlm_lock *, int,
1451 static ldlm_cancel_lru_policy_t
1452 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1454 if (ns_connect_lru_resize(ns)) {
1455 if (flags & LDLM_CANCEL_SHRINK)
1456 return ldlm_cancel_shrink_policy;
1457 else if (flags & LDLM_CANCEL_LRUR)
1458 return ldlm_cancel_lrur_policy;
1459 else if (flags & LDLM_CANCEL_PASSED)
1460 return ldlm_cancel_passed_policy;
1462 if (flags & LDLM_CANCEL_AGED)
1463 return ldlm_cancel_aged_policy;
1466 return ldlm_cancel_default_policy;
1469 /* - Free space in lru for @count new locks,
1470 * redundant unused locks are canceled locally;
1471 * - also cancel locally unused aged locks;
1472 * - do not cancel more than @max locks;
1473 * - GET the found locks and add them into the @cancels list.
1475 * A client lock can be added to the l_bl_ast list only when it is
1476 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing CANCEL.
1477 * There are the following use cases: ldlm_cancel_resource_local(),
1478 * ldlm_cancel_lru_local() and ldlm_cli_cancel(), which check&set this
1479 * flag properly. As any attempt to cancel a lock rely on this flag,
1480 * l_bl_ast list is accessed later without any special locking.
1482 * Calling policies for enabled lru resize:
1483 * ----------------------------------------
1484 * flags & LDLM_CANCEL_LRUR - use lru resize policy (SLV from server) to
1485 * cancel not more than @count locks;
1487 * flags & LDLM_CANCEL_PASSED - cancel @count number of old locks (located at
1488 * the beginning of lru list);
1490 * flags & LDLM_CANCEL_SHRINK - cancel not more than @count locks according to
1491 * memory pressre policy function;
1493 * flags & LDLM_CANCEL_AGED - cancel alocks according to "aged policy".
1495 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1496 int count, int max, int cancel_flags, int flags)
1498 ldlm_cancel_lru_policy_t pf;
1499 struct ldlm_lock *lock, *next;
1500 int added = 0, unused;
1503 spin_lock(&ns->ns_unused_lock);
1504 unused = ns->ns_nr_unused;
1506 if (!ns_connect_lru_resize(ns))
1507 count += unused - ns->ns_max_unused;
1509 pf = ldlm_cancel_lru_policy(ns, flags);
1510 LASSERT(pf != NULL);
1512 while (!list_empty(&ns->ns_unused_list)) {
1513 /* For any flags, stop scanning if @max is reached. */
1514 if (max && added >= max)
1517 list_for_each_entry_safe(lock, next, &ns->ns_unused_list, l_lru){
1518 /* No locks which got blocking requests. */
1519 LASSERT(!(lock->l_flags & LDLM_FL_BL_AST));
1521 /* Somebody is already doing CANCEL. No need in this
1522 * lock in lru, do not traverse it again. */
1523 if (!(lock->l_flags & LDLM_FL_CANCELING))
1526 ldlm_lock_remove_from_lru_nolock(lock);
1528 if (&lock->l_lru == &ns->ns_unused_list)
1531 /* Pass the lock through the policy filter and see if it
1532 * should stay in lru.
1534 * Even for shrinker policy we stop scanning if
1535 * we find a lock that should stay in the cache.
1536 * We should take into account lock age anyway
1537 * as new lock even if it is small of weight is
1538 * valuable resource.
1540 * That is, for shrinker policy we drop only
1541 * old locks, but additionally chose them by
1542 * their weight. Big extent locks will stay in
1544 if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
1547 LDLM_LOCK_GET(lock); /* dropped by bl thread */
1548 spin_unlock(&ns->ns_unused_lock);
1550 lock_res_and_lock(lock);
1551 /* Check flags again under the lock. */
1552 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1553 (ldlm_lock_remove_from_lru(lock) == 0)) {
1554 /* other thread is removing lock from lru or
1555 * somebody is already doing CANCEL or
1556 * there is a blocking request which will send
1557 * cancel by itseft or the lock is matched
1558 * is already not unused. */
1559 unlock_res_and_lock(lock);
1560 LDLM_LOCK_PUT(lock);
1561 spin_lock(&ns->ns_unused_lock);
1564 LASSERT(!lock->l_readers && !lock->l_writers);
1566 /* If we have chosen to cancel this lock voluntarily, we
1567 * better send cancel notification to server, so that it
1568 * frees appropriate state. This might lead to a race
1569 * where while we are doing cancel here, server is also
1570 * silently cancelling this lock. */
1571 lock->l_flags &= ~LDLM_FL_CANCEL_ON_BLOCK;
1573 /* Setting the CBPENDING flag is a little misleading,
1574 * but prevents an important race; namely, once
1575 * CBPENDING is set, the lock can accumulate no more
1576 * readers/writers. Since readers and writers are
1577 * already zero here, ldlm_lock_decref() won't see
1578 * this flag and call l_blocking_ast */
1579 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1581 /* We can't re-add to l_lru as it confuses the
1582 * refcounting in ldlm_lock_remove_from_lru() if an AST
1583 * arrives after we drop ns_lock below. We use l_bl_ast
1584 * and can't use l_pending_chain as it is used both on
1585 * server and client nevertheless bug 5666 says it is
1586 * used only on server */
1587 LASSERT(list_empty(&lock->l_bl_ast));
1588 list_add(&lock->l_bl_ast, cancels);
1589 unlock_res_and_lock(lock);
1590 spin_lock(&ns->ns_unused_lock);
1594 spin_unlock(&ns->ns_unused_lock);
1595 RETURN(ldlm_cancel_list(cancels, added, cancel_flags));
1598 /* Returns number of locks which could be canceled next time when
1599 * ldlm_cancel_lru() is called. Used from locks pool shrinker. */
1600 int ldlm_cancel_lru_estimate(struct ldlm_namespace *ns,
1601 int count, int max, int flags)
1603 ldlm_cancel_lru_policy_t pf;
1604 struct ldlm_lock *lock;
1605 int added = 0, unused;
1608 pf = ldlm_cancel_lru_policy(ns, flags);
1609 LASSERT(pf != NULL);
1610 spin_lock(&ns->ns_unused_lock);
1611 unused = ns->ns_nr_unused;
1613 list_for_each_entry(lock, &ns->ns_unused_list, l_lru) {
1614 /* For any flags, stop scanning if @max is reached. */
1615 if (max && added >= max)
1618 /* Somebody is already doing CANCEL or there is a
1619 * blocking request will send cancel. Let's not count
1621 if ((lock->l_flags & LDLM_FL_CANCELING) ||
1622 (lock->l_flags & LDLM_FL_BL_AST))
1625 /* Pass the lock through the policy filter and see if it
1626 * should stay in lru. */
1627 if (pf(ns, lock, unused, added, count) == LDLM_POLICY_KEEP_LOCK)
1633 spin_unlock(&ns->ns_unused_lock);
1637 /* when called with LDLM_ASYNC the blocking callback will be handled
1638 * in a thread and this function will return after the thread has been
1639 * asked to call the callback. when called with LDLM_SYNC the blocking
1640 * callback will be performed in this function. */
1641 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr, ldlm_sync_t sync,
1644 CFS_LIST_HEAD(cancels);
1649 sync = LDLM_SYNC; /* force to be sync in user space */
1651 count = ldlm_cancel_lru_local(ns, &cancels, nr, 0, 0, flags);
1652 if (sync == LDLM_ASYNC) {
1653 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count);
1658 /* If an error occured in ASYNC mode, or
1659 * this is SYNC mode, cancel the list. */
1660 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
1664 /* Find and cancel locally unused locks found on resource, matched to the
1665 * given policy, mode. GET the found locks and add them into the @cancels
1667 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1668 struct list_head *cancels,
1669 ldlm_policy_data_t *policy,
1670 ldlm_mode_t mode, int lock_flags,
1671 int cancel_flags, void *opaque)
1673 struct ldlm_lock *lock;
1678 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1679 if (opaque != NULL && lock->l_ast_data != opaque) {
1680 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1681 lock->l_ast_data, opaque);
1686 if (lock->l_readers || lock->l_writers) {
1687 if (cancel_flags & LDLM_FL_WARN) {
1688 LDLM_ERROR(lock, "lock in use");
1694 /* If somebody is already doing CANCEL, or blocking ast came,
1695 * skip this lock. */
1696 if (lock->l_flags & LDLM_FL_BL_AST ||
1697 lock->l_flags & LDLM_FL_CANCELING)
1700 if (lockmode_compat(lock->l_granted_mode, mode))
1703 /* If policy is given and this is IBITS lock, add to list only
1704 * those locks that match by policy. */
1705 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1706 !(lock->l_policy_data.l_inodebits.bits &
1707 policy->l_inodebits.bits))
1710 /* See CBPENDING comment in ldlm_cancel_lru */
1711 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1714 LASSERT(list_empty(&lock->l_bl_ast));
1715 list_add(&lock->l_bl_ast, cancels);
1716 LDLM_LOCK_GET(lock);
1721 RETURN(ldlm_cancel_list(cancels, count, cancel_flags));
1724 /* If @req is NULL, send CANCEL request to server with handles of locks
1725 * in the @cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1726 * separately per lock.
1727 * If @req is not NULL, put handles of locks in @cancels into the request
1728 * buffer at the offset @off.
1729 * Destroy @cancels at the end. */
1730 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1731 struct ptlrpc_request *req, int flags)
1733 struct ldlm_lock *lock;
1737 if (list_empty(cancels) || count == 0)
1740 /* XXX: requests (both batched and not) could be sent in parallel.
1741 * Usually it is enough to have just 1 RPC, but it is possible that
1742 * there are to many locks to be cancelled in LRU or on a resource.
1743 * It would also speed up the case when the server does not support
1746 LASSERT(!list_empty(cancels));
1747 lock = list_entry(cancels->next, struct ldlm_lock, l_bl_ast);
1748 LASSERT(lock->l_conn_export);
1750 if (exp_connect_cancelset(lock->l_conn_export)) {
1753 ldlm_cancel_pack(req, cancels, count);
1755 res = ldlm_cli_cancel_req(lock->l_conn_export,
1759 res = ldlm_cli_cancel_req(lock->l_conn_export,
1764 CERROR("ldlm_cli_cancel_list: %d\n", res);
1769 ldlm_lock_list_put(cancels, l_bl_ast, res);
1771 LASSERT(count == 0);
1775 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1776 const struct ldlm_res_id *res_id,
1777 ldlm_policy_data_t *policy,
1778 ldlm_mode_t mode, int flags, void *opaque)
1780 struct ldlm_resource *res;
1781 CFS_LIST_HEAD(cancels);
1786 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1788 /* This is not a problem. */
1789 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1793 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1795 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1797 CERROR("ldlm_cli_cancel_unused_resource: %d\n", rc);
1799 ldlm_resource_putref(res);
1803 static inline int have_no_nsresource(struct ldlm_namespace *ns)
1805 int no_resource = 0;
1807 spin_lock(&ns->ns_hash_lock);
1808 if (ns->ns_resources == 0)
1810 spin_unlock(&ns->ns_hash_lock);
1812 RETURN(no_resource);
1815 /* Cancel all locks on a namespace (or a specific resource, if given)
1816 * that have 0 readers/writers.
1818 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1819 * to notify the server. */
1820 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1821 const struct ldlm_res_id *res_id,
1822 int flags, void *opaque)
1831 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1835 spin_lock(&ns->ns_hash_lock);
1836 for (i = 0; i < RES_HASH_SIZE; i++) {
1837 struct list_head *tmp;
1838 tmp = ns->ns_hash[i].next;
1839 while (tmp != &(ns->ns_hash[i])) {
1840 struct ldlm_resource *res;
1843 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1844 ldlm_resource_getref(res);
1845 spin_unlock(&ns->ns_hash_lock);
1847 rc = ldlm_cli_cancel_unused_resource(ns, &res->lr_name,
1852 CERROR("ldlm_cli_cancel_unused ("LPU64"): %d\n",
1853 res->lr_name.name[0], rc);
1855 spin_lock(&ns->ns_hash_lock);
1857 ldlm_resource_putref_locked(res);
1860 spin_unlock(&ns->ns_hash_lock);
1865 /* Lock iterators. */
1867 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1870 struct list_head *tmp, *next;
1871 struct ldlm_lock *lock;
1872 int rc = LDLM_ITER_CONTINUE;
1877 RETURN(LDLM_ITER_CONTINUE);
1880 list_for_each_safe(tmp, next, &res->lr_granted) {
1881 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1883 if (iter(lock, closure) == LDLM_ITER_STOP)
1884 GOTO(out, rc = LDLM_ITER_STOP);
1887 list_for_each_safe(tmp, next, &res->lr_converting) {
1888 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1890 if (iter(lock, closure) == LDLM_ITER_STOP)
1891 GOTO(out, rc = LDLM_ITER_STOP);
1894 list_for_each_safe(tmp, next, &res->lr_waiting) {
1895 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1897 if (iter(lock, closure) == LDLM_ITER_STOP)
1898 GOTO(out, rc = LDLM_ITER_STOP);
1905 struct iter_helper_data {
1906 ldlm_iterator_t iter;
1910 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
1912 struct iter_helper_data *helper = closure;
1913 return helper->iter(lock, helper->closure);
1916 static int ldlm_res_iter_helper(struct ldlm_resource *res, void *closure)
1918 return ldlm_resource_foreach(res, ldlm_iter_helper, closure);
1921 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1924 struct iter_helper_data helper = { iter: iter, closure: closure };
1925 return ldlm_namespace_foreach_res(ns, ldlm_res_iter_helper, &helper);
1928 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
1929 ldlm_res_iterator_t iter, void *closure)
1931 int i, rc = LDLM_ITER_CONTINUE;
1932 struct ldlm_resource *res;
1933 struct list_head *tmp;
1936 spin_lock(&ns->ns_hash_lock);
1937 for (i = 0; i < RES_HASH_SIZE; i++) {
1938 tmp = ns->ns_hash[i].next;
1939 while (tmp != &(ns->ns_hash[i])) {
1940 res = list_entry(tmp, struct ldlm_resource, lr_hash);
1941 ldlm_resource_getref(res);
1942 spin_unlock(&ns->ns_hash_lock);
1944 rc = iter(res, closure);
1946 spin_lock(&ns->ns_hash_lock);
1948 ldlm_resource_putref_locked(res);
1949 if (rc == LDLM_ITER_STOP)
1954 spin_unlock(&ns->ns_hash_lock);
1958 /* non-blocking function to manipulate a lock whose cb_data is being put away.*/
1959 void ldlm_resource_iterate(struct ldlm_namespace *ns,
1960 const struct ldlm_res_id *res_id,
1961 ldlm_iterator_t iter, void *data)
1963 struct ldlm_resource *res;
1967 CERROR("must pass in namespace\n");
1971 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1977 ldlm_resource_foreach(res, iter, data);
1978 ldlm_resource_putref(res);
1984 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
1986 struct list_head *list = closure;
1988 /* we use l_pending_chain here, because it's unused on clients. */
1989 LASSERTF(list_empty(&lock->l_pending_chain),"lock %p next %p prev %p\n",
1990 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
1991 /* bug 9573: don't replay locks left after eviction */
1992 if (!(lock->l_flags & LDLM_FL_FAILED))
1993 list_add(&lock->l_pending_chain, list);
1994 return LDLM_ITER_CONTINUE;
1997 static int replay_lock_interpret(struct ptlrpc_request *req,
1998 struct ldlm_async_args *aa, int rc)
2000 struct lustre_handle old_hash_key;
2001 struct ldlm_lock *lock;
2002 struct ldlm_reply *reply;
2003 struct obd_export *exp;
2006 atomic_dec(&req->rq_import->imp_replay_inflight);
2011 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2013 GOTO(out, rc = -EPROTO);
2015 lock = ldlm_handle2lock(&aa->lock_handle);
2017 CERROR("received replay ack for unknown local cookie "LPX64
2018 " remote cookie "LPX64 " from server %s id %s\n",
2019 aa->lock_handle.cookie, reply->lock_handle.cookie,
2020 req->rq_export->exp_client_uuid.uuid,
2021 libcfs_id2str(req->rq_peer));
2022 GOTO(out, rc = -ESTALE);
2025 old_hash_key = lock->l_remote_handle;
2026 lock->l_remote_handle = reply->lock_handle;
2028 /* Key change rehash lock in per-export hash with new key */
2029 exp = req->rq_export;
2030 if (exp && exp->exp_lock_hash)
2031 lustre_hash_rehash_key(exp->exp_lock_hash, &old_hash_key,
2032 &lock->l_remote_handle,
2035 LDLM_DEBUG(lock, "replayed lock:");
2036 ptlrpc_import_recovery_state_machine(req->rq_import);
2037 LDLM_LOCK_PUT(lock);
2040 ptlrpc_connect_import(req->rq_import, NULL);
2045 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2047 struct ptlrpc_request *req;
2048 struct ldlm_async_args *aa;
2049 struct ldlm_request *body;
2054 /* Bug 11974: Do not replay a lock which is actively being canceled */
2055 if (lock->l_flags & LDLM_FL_CANCELING) {
2056 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2060 /* If this is reply-less callback lock, we cannot replay it, since
2061 * server might have long dropped it, but notification of that event was
2062 * lost by network. (and server granted conflicting lock already) */
2063 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
2064 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2065 ldlm_lock_cancel(lock);
2069 * If granted mode matches the requested mode, this lock is granted.
2071 * If they differ, but we have a granted mode, then we were granted
2072 * one mode and now want another: ergo, converting.
2074 * If we haven't been granted anything and are on a resource list,
2075 * then we're blocked/waiting.
2077 * If we haven't been granted anything and we're NOT on a resource list,
2078 * then we haven't got a reply yet and don't have a known disposition.
2079 * This happens whenever a lock enqueue is the request that triggers
2082 if (lock->l_granted_mode == lock->l_req_mode)
2083 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2084 else if (lock->l_granted_mode)
2085 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2086 else if (!list_empty(&lock->l_res_link))
2087 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2089 flags = LDLM_FL_REPLAY;
2091 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2092 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2096 /* We're part of recovery, so don't wait for it. */
2097 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2099 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2100 ldlm_lock2desc(lock, &body->lock_desc);
2101 body->lock_flags = flags;
2103 ldlm_lock2handle(lock, &body->lock_handle[0]);
2104 if (lock->l_lvb_len != 0) {
2105 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2106 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2109 ptlrpc_request_set_replen(req);
2110 /* notify the server we've replayed all requests.
2111 * also, we mark the request to be put on a dedicated
2112 * queue to be processed after all request replayes.
2114 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2116 LDLM_DEBUG(lock, "replaying lock:");
2118 atomic_inc(&req->rq_import->imp_replay_inflight);
2119 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2120 aa = ptlrpc_req_async_args(req);
2121 aa->lock_handle = body->lock_handle[0];
2122 req->rq_interpret_reply = replay_lock_interpret;
2123 ptlrpcd_add_req(req);
2128 int ldlm_replay_locks(struct obd_import *imp)
2130 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2131 CFS_LIST_HEAD(list);
2132 struct ldlm_lock *lock, *next;
2137 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2139 /* ensure this doesn't fall to 0 before all have been queued */
2140 atomic_inc(&imp->imp_replay_inflight);
2142 (void)ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2144 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2145 list_del_init(&lock->l_pending_chain);
2147 continue; /* or try to do the rest? */
2148 rc = replay_one_lock(imp, lock);
2151 atomic_dec(&imp->imp_replay_inflight);