4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 * This file contains Asynchronous System Trap (AST) handlers and related
38 * LDLM request-processing routines.
40 * An AST is a callback issued on a lock when its state is changed. There are
41 * several different types of ASTs (callbacks) registered for each lock:
43 * - completion AST: when a lock is enqueued by some process, but cannot be
44 * granted immediately due to other conflicting locks on the same resource,
45 * the completion AST is sent to notify the caller when the lock is
48 * - blocking AST: when a lock is granted to some process, if another process
49 * enqueues a conflicting (blocking) lock on a resource, a blocking AST is
50 * sent to notify the holder(s) of the lock(s) of the conflicting lock
51 * request. The lock holder(s) must release their lock(s) on that resource in
52 * a timely manner or be evicted by the server.
54 * - glimpse AST: this is used when a process wants information about a lock
55 * (i.e. the lock value block (LVB)) but does not necessarily require holding
56 * the lock. If the resource is locked, the lock holder(s) are sent glimpse
57 * ASTs and the LVB is returned to the caller, and lock holder(s) may CANCEL
58 * their lock(s) if they are idle. If the resource is not locked, the server
62 #define DEBUG_SUBSYSTEM S_LDLM
65 #include <liblustre.h>
68 #include <lustre_dlm.h>
69 #include <obd_class.h>
72 #include "ldlm_internal.h"
74 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
75 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
76 "lock enqueue timeout minimum");
78 /* in client side, whether the cached locks will be canceled before replay */
79 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
81 static void interrupted_completion_wait(void *data)
85 struct lock_wait_data {
86 struct ldlm_lock *lwd_lock;
90 struct ldlm_async_args {
91 struct lustre_handle lock_handle;
94 int ldlm_expired_completion_wait(void *data)
96 struct lock_wait_data *lwd = data;
97 struct ldlm_lock *lock = lwd->lwd_lock;
98 struct obd_import *imp;
99 struct obd_device *obd;
102 if (lock->l_conn_export == NULL) {
103 static cfs_time_t next_dump = 0, last_dump = 0;
105 LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", "
106 CFS_DURATION_T"s ago)\n",
107 lock->l_last_activity,
108 cfs_time_sub(cfs_time_current_sec(),
109 lock->l_last_activity));
110 LDLM_DEBUG(lock, "lock timed out (enqueued at "CFS_TIME_T", "
111 CFS_DURATION_T"s ago); not entering recovery in "
112 "server code, just going back to sleep",
113 lock->l_last_activity,
114 cfs_time_sub(cfs_time_current_sec(),
115 lock->l_last_activity));
116 if (cfs_time_after(cfs_time_current(), next_dump)) {
117 last_dump = next_dump;
118 next_dump = cfs_time_shift(300);
119 ldlm_namespace_dump(D_DLMTRACE,
120 ldlm_lock_to_ns(lock));
122 libcfs_debug_dumplog();
127 obd = lock->l_conn_export->exp_obd;
128 imp = obd->u.cli.cl_import;
129 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
130 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
131 CFS_DURATION_T"s ago), entering recovery for %s@%s",
132 lock->l_last_activity,
133 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
134 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
138 EXPORT_SYMBOL(ldlm_expired_completion_wait);
140 /* We use the same basis for both server side and client side functions
141 from a single node. */
142 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
144 int timeout = at_get(ldlm_lock_to_ns_at(lock));
146 return obd_timeout / 2;
147 /* Since these are non-updating timeouts, we should be conservative.
148 It would be nice to have some kind of "early reply" mechanism for
149 lock callbacks too... */
150 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
151 return max(timeout, ldlm_enqueue_min);
153 EXPORT_SYMBOL(ldlm_get_enq_timeout);
156 * Helper function for ldlm_completion_ast(), updating timings when lock is
159 static int ldlm_completion_tail(struct ldlm_lock *lock)
164 if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
165 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
168 delay = cfs_time_sub(cfs_time_current_sec(),
169 lock->l_last_activity);
170 LDLM_DEBUG(lock, "client-side enqueue: granted after "
171 CFS_DURATION_T"s", delay);
173 /* Update our time estimate */
174 at_measured(ldlm_lock_to_ns_at(lock),
182 * Implementation of ->l_completion_ast() for a client, that doesn't wait
183 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
184 * other threads that cannot block for long.
186 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
190 if (flags == LDLM_FL_WAIT_NOREPROC) {
191 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
195 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
196 LDLM_FL_BLOCK_CONV))) {
197 wake_up(&lock->l_waitq);
198 RETURN(ldlm_completion_tail(lock));
201 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
203 ldlm_reprocess_all(lock->l_resource);
206 EXPORT_SYMBOL(ldlm_completion_ast_async);
209 * Generic LDLM "completion" AST. This is called in several cases:
211 * - when a reply to an ENQUEUE RPC is received from the server
212 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
213 * this point (determined by flags);
215 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
218 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
221 * - to force all locks when resource is destroyed (cleanup_resource());
223 * - during lock conversion (not used currently).
225 * If lock is not granted in the first case, this function waits until second
226 * or penultimate cases happen in some other thread.
229 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
231 /* XXX ALLOCATE - 160 bytes */
232 struct lock_wait_data lwd;
233 struct obd_device *obd;
234 struct obd_import *imp = NULL;
235 struct l_wait_info lwi;
240 if (flags == LDLM_FL_WAIT_NOREPROC) {
241 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
245 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
246 LDLM_FL_BLOCK_CONV))) {
247 wake_up(&lock->l_waitq);
251 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
256 obd = class_exp2obd(lock->l_conn_export);
258 /* if this is a local lock, then there is no import */
260 imp = obd->u.cli.cl_import;
263 /* Wait a long time for enqueue - server may have to callback a
264 lock from another client. Server will evict the other client if it
265 doesn't respond reasonably, and then give us the lock. */
266 timeout = ldlm_get_enq_timeout(lock) * 2;
270 if (ldlm_is_no_timeout(lock)) {
271 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
272 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
274 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
275 ldlm_expired_completion_wait,
276 interrupted_completion_wait, &lwd);
280 spin_lock(&imp->imp_lock);
281 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
282 spin_unlock(&imp->imp_lock);
285 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
286 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
287 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
288 ldlm_set_fail_loc(lock);
291 /* Go to sleep until the lock is granted or cancelled. */
292 rc = l_wait_event(lock->l_waitq,
293 is_granted_or_cancelled(lock), &lwi);
297 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
302 RETURN(ldlm_completion_tail(lock));
304 EXPORT_SYMBOL(ldlm_completion_ast);
307 * A helper to build a blocking AST function
309 * Perform a common operation for blocking ASTs:
310 * defferred lock cancellation.
312 * \param lock the lock blocking or canceling AST was called on
314 * \see mdt_blocking_ast
315 * \see ldlm_blocking_ast
317 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
322 ldlm_set_cbpending(lock);
323 do_ast = (!lock->l_readers && !lock->l_writers);
324 unlock_res_and_lock(lock);
327 struct lustre_handle lockh;
330 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
331 ldlm_lock2handle(lock, &lockh);
332 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
334 CERROR("ldlm_cli_cancel: %d\n", rc);
336 LDLM_DEBUG(lock, "Lock still has references, will be "
341 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
344 * Server blocking AST
346 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
349 * \param lock the lock which blocks a request or cancelling lock
352 * \param flag indicates whether this cancelling or blocking callback
354 * \see ldlm_blocking_ast_nocheck
356 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
357 void *data, int flag)
361 if (flag == LDLM_CB_CANCELING) {
362 /* Don't need to do anything here. */
366 lock_res_and_lock(lock);
367 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
368 * that ldlm_blocking_ast is called just before intent_policy method
369 * takes the lr_lock, then by the time we get the lock, we might not
370 * be the correct blocking function anymore. So check, and return
372 if (lock->l_blocking_ast != ldlm_blocking_ast) {
373 unlock_res_and_lock(lock);
376 RETURN(ldlm_blocking_ast_nocheck(lock));
378 EXPORT_SYMBOL(ldlm_blocking_ast);
381 * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server.
383 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is
384 * rather subtle: with OST-side locking, it may so happen that _all_ extent
385 * locks are held by the OST. If client wants to obtain the current file size
386 * it calls ll_glimpse_size(), and (as all locks are held only on the server),
387 * this dummy glimpse callback fires and does nothing. The client still
388 * receives the correct file size due to the following fragment of code in
389 * ldlm_cb_interpret():
391 * if (rc == -ELDLM_NO_LOCK_DATA) {
392 * LDLM_DEBUG(lock, "lost race - client has a lock but no"
394 * ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
397 * That is, after the glimpse returns this error, ofd_lvbo_update() is called
398 * and returns the updated file attributes from the inode to the client.
400 * See also comment in ofd_intent_policy() on why servers must set a non-NULL
401 * l_glimpse_ast when grabbing DLM locks. Otherwise, the server will assume
402 * that the object is in the process of being destroyed.
404 * \param[in] lock DLM lock being glimpsed, unused
405 * \param[in] reqp pointer to ptlrpc_request, unused
407 * \retval -ELDLM_NO_LOCK_DATA to get attributes from disk object
409 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
411 return -ELDLM_NO_LOCK_DATA;
413 EXPORT_SYMBOL(ldlm_glimpse_ast);
416 * Enqueue a local lock (typically on a server).
418 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
419 const struct ldlm_res_id *res_id,
420 ldlm_type_t type, ldlm_policy_data_t *policy,
421 ldlm_mode_t mode, __u64 *flags,
422 ldlm_blocking_callback blocking,
423 ldlm_completion_callback completion,
424 ldlm_glimpse_callback glimpse,
425 void *data, __u32 lvb_len, enum lvb_type lvb_type,
426 const __u64 *client_cookie,
427 struct lustre_handle *lockh)
429 struct ldlm_lock *lock;
431 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
432 .lcs_blocking = blocking,
433 .lcs_glimpse = glimpse,
437 LASSERT(!(*flags & LDLM_FL_REPLAY));
438 if (unlikely(ns_is_client(ns))) {
439 CERROR("Trying to enqueue local lock in a shadow namespace\n");
443 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
446 GOTO(out_nolock, err = PTR_ERR(lock));
448 ldlm_lock2handle(lock, lockh);
450 /* NB: we don't have any lock now (lock_res_and_lock)
451 * because it's a new lock */
452 ldlm_lock_addref_internal_nolock(lock, mode);
453 ldlm_set_local(lock);
454 if (*flags & LDLM_FL_ATOMIC_CB)
455 ldlm_set_atomic_cb(lock);
458 lock->l_policy_data = *policy;
459 if (client_cookie != NULL)
460 lock->l_client_cookie = *client_cookie;
461 if (type == LDLM_EXTENT) {
462 /* extent lock without policy is a bug */
466 lock->l_req_extent = policy->l_extent;
469 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
470 if (unlikely(err != ELDLM_OK))
474 *policy = lock->l_policy_data;
476 if (lock->l_completion_ast)
477 lock->l_completion_ast(lock, *flags, NULL);
479 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
482 LDLM_LOCK_RELEASE(lock);
486 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
488 static void failed_lock_cleanup(struct ldlm_namespace *ns,
489 struct ldlm_lock *lock, int mode)
493 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
494 lock_res_and_lock(lock);
495 /* Check that lock is not granted or failed, we might race. */
496 if ((lock->l_req_mode != lock->l_granted_mode) &&
497 !ldlm_is_failed(lock)) {
498 /* Make sure that this lock will not be found by raced
499 * bl_ast and -EINVAL reply is sent to server anyways.
501 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
502 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
505 unlock_res_and_lock(lock);
509 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
510 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
512 LDLM_DEBUG(lock, "lock was granted or failed in race");
514 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
515 * from llite/file.c/ll_file_flock(). */
516 /* This code makes for the fact that we do not have blocking handler on
517 * a client for flock locks. As such this is the place where we must
518 * completely kill failed locks. (interrupted and those that
519 * were waiting to be granted when server evicted us. */
520 if (lock->l_resource->lr_type == LDLM_FLOCK) {
521 lock_res_and_lock(lock);
522 if (!ldlm_is_destroyed(lock)) {
523 ldlm_resource_unlink_lock(lock);
524 ldlm_lock_decref_internal_nolock(lock, mode);
525 ldlm_lock_destroy_nolock(lock);
527 unlock_res_and_lock(lock);
529 ldlm_lock_decref_internal(lock, mode);
534 * Finishing portion of client lock enqueue code.
536 * Called after receiving reply from server.
538 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
539 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
540 __u64 *flags, void *lvb, __u32 lvb_len,
541 struct lustre_handle *lockh,int rc)
543 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
544 int is_replay = *flags & LDLM_FL_REPLAY;
545 struct ldlm_lock *lock;
546 struct ldlm_reply *reply;
547 int cleanup_phase = 1;
551 lock = ldlm_handle2lock(lockh);
552 /* ldlm_cli_enqueue is holding a reference on this lock. */
554 LASSERT(type == LDLM_FLOCK);
558 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
559 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
561 if (rc != ELDLM_OK) {
563 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
564 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
566 if (rc != ELDLM_LOCK_ABORTED)
570 /* Before we return, swab the reply */
571 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
573 GOTO(cleanup, rc = -EPROTO);
576 LASSERT(lvb != NULL);
578 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
581 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
582 GOTO(cleanup, rc = size);
583 } else if (unlikely(size > lvb_len)) {
584 LDLM_ERROR(lock, "Replied LVB is larger than "
585 "expectation, expected = %d, replied = %d",
587 GOTO(cleanup, rc = -EINVAL);
591 if (rc == ELDLM_LOCK_ABORTED) {
593 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
595 GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED));
598 /* lock enqueued on the server */
601 lock_res_and_lock(lock);
602 /* Key change rehash lock in per-export hash with new key */
603 if (exp->exp_lock_hash) {
604 /* In the function below, .hs_keycmp resolves to
605 * ldlm_export_lock_keycmp() */
606 /* coverity[overrun-buffer-val] */
607 cfs_hash_rehash_key(exp->exp_lock_hash,
608 &lock->l_remote_handle,
612 lock->l_remote_handle = reply->lock_handle;
615 *flags = ldlm_flags_from_wire(reply->lock_flags);
616 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
617 LDLM_FL_INHERIT_MASK);
618 unlock_res_and_lock(lock);
620 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: "LPX64"\n",
621 lock, reply->lock_handle.cookie, *flags);
623 /* If enqueue returned a blocked lock but the completion handler has
624 * already run, then it fixed up the resource and we don't need to do it
626 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
627 int newmode = reply->lock_desc.l_req_mode;
629 if (newmode && newmode != lock->l_req_mode) {
630 LDLM_DEBUG(lock, "server returned different mode %s",
631 ldlm_lockname[newmode]);
632 lock->l_req_mode = newmode;
635 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
636 &lock->l_resource->lr_name)) {
637 CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
638 " instead of "DLDLMRES"\n",
639 PLDLMRES(&reply->lock_desc.l_resource),
640 PLDLMRES(lock->l_resource));
642 rc = ldlm_lock_change_resource(ns, lock,
643 &reply->lock_desc.l_resource.lr_name);
644 if (rc || lock->l_resource == NULL)
645 GOTO(cleanup, rc = -ENOMEM);
646 LDLM_DEBUG(lock, "client-side enqueue, new resource");
649 if (!(type == LDLM_IBITS &&
650 !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
651 /* We assume lock type cannot change on server*/
652 ldlm_convert_policy_to_local(exp,
653 lock->l_resource->lr_type,
654 &reply->lock_desc.l_policy_data,
655 &lock->l_policy_data);
656 if (type != LDLM_PLAIN)
657 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
660 if ((*flags) & LDLM_FL_AST_SENT ||
661 /* Cancel extent locks as soon as possible on a liblustre client,
662 * because it cannot handle asynchronous ASTs robustly (see
664 (LIBLUSTRE_CLIENT && type == LDLM_EXTENT)) {
665 lock_res_and_lock(lock);
666 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
667 unlock_res_and_lock(lock);
668 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
671 /* If the lock has already been granted by a completion AST, don't
672 * clobber the LVB with an older one. */
674 /* We must lock or a racing completion might update lvb without
675 * letting us know and we'll clobber the correct value.
676 * Cannot unlock after the check either, a that still leaves
677 * a tiny window for completion to get in */
678 lock_res_and_lock(lock);
679 if (lock->l_req_mode != lock->l_granted_mode)
680 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
681 lock->l_lvb_data, size);
682 unlock_res_and_lock(lock);
690 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
691 if (lock->l_completion_ast != NULL) {
692 int err = lock->l_completion_ast(lock, *flags, NULL);
700 if (lvb_len && lvb != NULL) {
701 /* Copy the LVB here, and not earlier, because the completion
702 * AST (if any) can override what we got in the reply */
703 memcpy(lvb, lock->l_lvb_data, lvb_len);
706 LDLM_DEBUG(lock, "client-side enqueue END");
709 if (cleanup_phase == 1 && rc)
710 failed_lock_cleanup(ns, lock, mode);
711 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
713 LDLM_LOCK_RELEASE(lock);
716 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
719 * Estimate number of lock handles that would fit into request of given
720 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
721 * a single page on the send/receive side. XXX: 512 should be changed to
722 * more adequate value.
724 static inline int ldlm_req_handles_avail(int req_size, int off)
728 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
729 if (likely(avail >= 0))
730 avail /= (int)sizeof(struct lustre_handle);
733 avail += LDLM_LOCKREQ_HANDLES - off;
738 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
739 enum req_location loc,
742 int size = req_capsule_msg_size(pill, loc);
743 return ldlm_req_handles_avail(size, off);
746 static inline int ldlm_format_handles_avail(struct obd_import *imp,
747 const struct req_format *fmt,
748 enum req_location loc, int off)
750 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
751 return ldlm_req_handles_avail(size, off);
755 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
756 * \a count locks in \a cancels.
758 * This is to be called by functions preparing their own requests that
759 * might contain lists of locks to cancel in addition to actual operation
760 * that needs to be performed.
762 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
763 int version, int opc, int canceloff,
764 struct list_head *cancels, int count)
766 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
767 struct req_capsule *pill = &req->rq_pill;
768 struct ldlm_request *dlm = NULL;
769 struct list_head head = LIST_HEAD_INIT(head);
770 int flags, avail, to_free, pack = 0;
776 if (ns_connect_cancelset(ns)) {
777 /* Estimate the amount of available space in the request. */
778 req_capsule_filled_sizes(pill, RCL_CLIENT);
779 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
781 flags = ns_connect_lru_resize(ns) ?
782 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
783 to_free = !ns_connect_lru_resize(ns) &&
784 opc == LDLM_ENQUEUE ? 1 : 0;
786 /* Cancel LRU locks here _only_ if the server supports
787 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
788 * RPC, which will make us slower. */
790 count += ldlm_cancel_lru_local(ns, cancels, to_free,
791 avail - count, 0, flags);
796 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
797 ldlm_request_bufsize(pack, opc));
800 rc = ptlrpc_request_pack(req, version, opc);
802 ldlm_lock_list_put(cancels, l_bl_ast, count);
806 if (ns_connect_cancelset(ns)) {
808 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
810 /* Skip first lock handler in ldlm_request_pack(),
811 * this method will incrment @lock_count according
812 * to the lock handle amount actually written to
814 dlm->lock_count = canceloff;
816 /* Pack into the request @pack lock handles. */
817 ldlm_cli_cancel_list(cancels, pack, req, 0);
818 /* Prepare and send separate cancel RPC for others. */
819 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
821 ldlm_lock_list_put(cancels, l_bl_ast, count);
825 EXPORT_SYMBOL(ldlm_prep_elc_req);
827 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
828 cfs_list_t *cancels, int count)
830 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
831 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
833 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
835 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
837 struct ptlrpc_request *req;
841 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
843 RETURN(ERR_PTR(-ENOMEM));
845 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
847 ptlrpc_request_free(req);
851 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
852 ptlrpc_request_set_replen(req);
855 EXPORT_SYMBOL(ldlm_enqueue_pack);
858 * Client-side lock enqueue.
860 * If a request has some specific initialisation it is passed in \a reqp,
861 * otherwise it is created in ldlm_cli_enqueue.
863 * Supports sync and async requests, pass \a async flag accordingly. If a
864 * request was created in ldlm_cli_enqueue and it is the async request,
865 * pass it to the caller in \a reqp.
867 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
868 struct ldlm_enqueue_info *einfo,
869 const struct ldlm_res_id *res_id,
870 ldlm_policy_data_t const *policy, __u64 *flags,
871 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
872 struct lustre_handle *lockh, int async)
874 struct ldlm_namespace *ns;
875 struct ldlm_lock *lock;
876 struct ldlm_request *body;
877 int is_replay = *flags & LDLM_FL_REPLAY;
878 int req_passed_in = 1;
880 struct ptlrpc_request *req;
883 LASSERT(exp != NULL);
885 ns = exp->exp_obd->obd_namespace;
887 /* If we're replaying this lock, just check some invariants.
888 * If we're creating a new lock, get everything all setup nice. */
890 lock = ldlm_handle2lock_long(lockh, 0);
891 LASSERT(lock != NULL);
892 LDLM_DEBUG(lock, "client-side enqueue START");
893 LASSERT(exp == lock->l_conn_export);
895 const struct ldlm_callback_suite cbs = {
896 .lcs_completion = einfo->ei_cb_cp,
897 .lcs_blocking = einfo->ei_cb_bl,
898 .lcs_glimpse = einfo->ei_cb_gl
900 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
901 einfo->ei_mode, &cbs, einfo->ei_cbdata,
904 RETURN(PTR_ERR(lock));
905 /* for the local lock, add the reference */
906 ldlm_lock_addref_internal(lock, einfo->ei_mode);
907 ldlm_lock2handle(lock, lockh);
909 lock->l_policy_data = *policy;
911 if (einfo->ei_type == LDLM_EXTENT) {
912 /* extent lock without policy is a bug */
916 lock->l_req_extent = policy->l_extent;
918 LDLM_DEBUG(lock, "client-side enqueue START, flags "LPX64"\n",
922 lock->l_conn_export = exp;
923 lock->l_export = NULL;
924 lock->l_blocking_ast = einfo->ei_cb_bl;
925 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
927 /* lock not sent to server yet */
929 if (reqp == NULL || *reqp == NULL) {
930 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
935 failed_lock_cleanup(ns, lock, einfo->ei_mode);
936 LDLM_LOCK_RELEASE(lock);
946 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
948 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
949 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
952 /* Dump lock data into the request buffer */
953 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
954 ldlm_lock2desc(lock, &body->lock_desc);
955 body->lock_flags = ldlm_flags_to_wire(*flags);
956 body->lock_handle[0] = *lockh;
958 /* Continue as normal. */
959 if (!req_passed_in) {
961 req_capsule_extend(&req->rq_pill,
962 &RQF_LDLM_ENQUEUE_LVB);
963 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
965 ptlrpc_request_set_replen(req);
969 * Liblustre client doesn't get extent locks, except for O_APPEND case
970 * where [0, OBD_OBJECT_EOF] lock is taken, or truncate, where
971 * [i_size, OBD_OBJECT_EOF] lock is taken.
973 LASSERT(ergo(LIBLUSTRE_CLIENT, einfo->ei_type != LDLM_EXTENT ||
974 policy->l_extent.end == OBD_OBJECT_EOF));
977 LASSERT(reqp != NULL);
981 LDLM_DEBUG(lock, "sending request");
983 rc = ptlrpc_queue_wait(req);
985 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
986 einfo->ei_mode, flags, lvb, lvb_len,
989 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
990 * one reference that we took */
992 LDLM_LOCK_RELEASE(lock);
996 if (!req_passed_in && req != NULL) {
997 ptlrpc_req_finished(req);
1004 EXPORT_SYMBOL(ldlm_cli_enqueue);
1006 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
1009 struct ldlm_resource *res;
1012 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1013 CERROR("Trying to cancel local lock\n");
1016 LDLM_DEBUG(lock, "client-side local convert");
1018 res = ldlm_lock_convert(lock, new_mode, flags);
1020 ldlm_reprocess_all(res);
1023 rc = LUSTRE_EDEADLK;
1025 LDLM_DEBUG(lock, "client-side local convert handler END");
1026 LDLM_LOCK_PUT(lock);
1030 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1031 * conversion of locks which are on the waiting or converting queue */
1032 /* Caller of this code is supposed to take care of lock readers/writers
1034 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1036 struct ldlm_request *body;
1037 struct ldlm_reply *reply;
1038 struct ldlm_lock *lock;
1039 struct ldlm_resource *res;
1040 struct ptlrpc_request *req;
1044 lock = ldlm_handle2lock(lockh);
1051 if (lock->l_conn_export == NULL)
1052 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1054 LDLM_DEBUG(lock, "client-side convert");
1056 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1057 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1060 LDLM_LOCK_PUT(lock);
1064 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1065 body->lock_handle[0] = lock->l_remote_handle;
1067 body->lock_desc.l_req_mode = new_mode;
1068 body->lock_flags = ldlm_flags_to_wire(*flags);
1071 ptlrpc_request_set_replen(req);
1072 rc = ptlrpc_queue_wait(req);
1076 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1078 GOTO(out, rc = -EPROTO);
1081 GOTO(out, rc = req->rq_status);
1083 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1085 ldlm_reprocess_all(res);
1086 /* Go to sleep until the lock is granted. */
1087 /* FIXME: or cancelled. */
1088 if (lock->l_completion_ast) {
1089 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1095 rc = LUSTRE_EDEADLK;
1099 LDLM_LOCK_PUT(lock);
1100 ptlrpc_req_finished(req);
1103 EXPORT_SYMBOL(ldlm_cli_convert);
1106 * Cancel locks locally.
1108 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1109 * \retval LDLM_FL_CANCELING otherwise;
1110 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1112 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1114 __u64 rc = LDLM_FL_LOCAL_ONLY;
1117 if (lock->l_conn_export) {
1120 LDLM_DEBUG(lock, "client-side cancel");
1121 /* Set this flag to prevent others from getting new references*/
1122 lock_res_and_lock(lock);
1123 ldlm_set_cbpending(lock);
1124 local_only = !!(lock->l_flags &
1125 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1126 ldlm_cancel_callback(lock);
1127 rc = (ldlm_is_bl_ast(lock)) ?
1128 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1129 unlock_res_and_lock(lock);
1132 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1134 rc = LDLM_FL_LOCAL_ONLY;
1136 ldlm_lock_cancel(lock);
1138 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1139 LDLM_ERROR(lock, "Trying to cancel local lock");
1142 LDLM_DEBUG(lock, "server-side local cancel");
1143 ldlm_lock_cancel(lock);
1144 ldlm_reprocess_all(lock->l_resource);
1151 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1153 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1154 cfs_list_t *head, int count)
1156 struct ldlm_request *dlm;
1157 struct ldlm_lock *lock;
1158 int max, packed = 0;
1161 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1162 LASSERT(dlm != NULL);
1164 /* Check the room in the request buffer. */
1165 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1166 sizeof(struct ldlm_request);
1167 max /= sizeof(struct lustre_handle);
1168 max += LDLM_LOCKREQ_HANDLES;
1169 LASSERT(max >= dlm->lock_count + count);
1171 /* XXX: it would be better to pack lock handles grouped by resource.
1172 * so that the server cancel would call filter_lvbo_update() less
1174 cfs_list_for_each_entry(lock, head, l_bl_ast) {
1177 LASSERT(lock->l_conn_export);
1178 /* Pack the lock handle to the given request buffer. */
1179 LDLM_DEBUG(lock, "packing");
1180 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1183 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1188 * Prepare and send a batched cancel RPC. It will include \a count lock
1189 * handles of locks given in \a cancels list. */
1190 int ldlm_cli_cancel_req(struct obd_export *exp, cfs_list_t *cancels,
1191 int count, ldlm_cancel_flags_t flags)
1193 struct ptlrpc_request *req = NULL;
1194 struct obd_import *imp;
1199 LASSERT(exp != NULL);
1202 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1204 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1207 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1208 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1213 imp = class_exp2cliimp(exp);
1214 if (imp == NULL || imp->imp_invalid) {
1216 "skipping cancel on invalid import %p\n", imp);
1220 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1222 GOTO(out, rc = -ENOMEM);
1224 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1225 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1226 ldlm_request_bufsize(count, LDLM_CANCEL));
1228 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1230 ptlrpc_request_free(req);
1234 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1235 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1236 ptlrpc_at_set_req_timeout(req);
1238 ldlm_cancel_pack(req, cancels, count);
1240 ptlrpc_request_set_replen(req);
1241 if (flags & LCF_ASYNC) {
1242 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1246 rc = ptlrpc_queue_wait(req);
1248 if (rc == LUSTRE_ESTALE) {
1249 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1250 "out of sync -- not fatal\n",
1251 libcfs_nid2str(req->rq_import->
1252 imp_connection->c_peer.nid));
1254 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1255 req->rq_import_generation == imp->imp_generation) {
1256 ptlrpc_req_finished(req);
1258 } else if (rc != ELDLM_OK) {
1259 /* -ESHUTDOWN is common on umount */
1260 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1261 "Got rc %d from cancel RPC: "
1262 "canceling anyway\n", rc);
1269 ptlrpc_req_finished(req);
1272 return sent ? sent : rc;
1274 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1276 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1278 LASSERT(imp != NULL);
1279 return &imp->imp_obd->obd_namespace->ns_pool;
1283 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1285 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1287 struct obd_device *obd;
1291 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1292 !imp_connect_lru_resize(req->rq_import)))
1295 * Do nothing for corner cases.
1300 /* In some cases RPC may contain SLV and limit zeroed out. This
1301 * is the case when server does not support LRU resize feature.
1302 * This is also possible in some recovery cases when server-side
1303 * reqs have no reference to the OBD export and thus access to
1304 * server-side namespace is not possible. */
1305 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1306 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1307 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1308 "(SLV: "LPU64", Limit: %u)",
1309 lustre_msg_get_slv(req->rq_repmsg),
1310 lustre_msg_get_limit(req->rq_repmsg));
1314 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1315 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1316 obd = req->rq_import->imp_obd;
1318 /* Set new SLV and limit in OBD fields to make them accessible
1319 * to the pool thread. We do not access obd_namespace and pool
1320 * directly here as there is no reliable way to make sure that
1321 * they are still alive at cleanup time. Evil races are possible
1322 * which may cause Oops at that time. */
1323 write_lock(&obd->obd_pool_lock);
1324 obd->obd_pool_slv = new_slv;
1325 obd->obd_pool_limit = new_limit;
1326 write_unlock(&obd->obd_pool_lock);
1330 EXPORT_SYMBOL(ldlm_cli_update_pool);
1333 * Client side lock cancel.
1335 * Lock must not have any readers or writers by this time.
1337 int ldlm_cli_cancel(struct lustre_handle *lockh,
1338 ldlm_cancel_flags_t cancel_flags)
1340 struct obd_export *exp;
1341 int avail, flags, count = 1;
1343 struct ldlm_namespace *ns;
1344 struct ldlm_lock *lock;
1345 struct list_head cancels = LIST_HEAD_INIT(cancels);
1348 /* concurrent cancels on the same handle can happen */
1349 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1351 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1355 rc = ldlm_cli_cancel_local(lock);
1356 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1357 LDLM_LOCK_RELEASE(lock);
1360 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1361 * RPC which goes to canceld portal, so we can cancel other LRU locks
1362 * here and send them all as one LDLM_CANCEL RPC. */
1363 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1364 cfs_list_add(&lock->l_bl_ast, &cancels);
1366 exp = lock->l_conn_export;
1367 if (exp_connect_cancelset(exp)) {
1368 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1373 ns = ldlm_lock_to_ns(lock);
1374 flags = ns_connect_lru_resize(ns) ?
1375 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1376 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1379 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1382 EXPORT_SYMBOL(ldlm_cli_cancel);
1385 * Locally cancel up to \a count locks in list \a cancels.
1386 * Return the number of cancelled locks.
1388 int ldlm_cli_cancel_list_local(cfs_list_t *cancels, int count,
1389 ldlm_cancel_flags_t flags)
1391 struct list_head head = LIST_HEAD_INIT(head);
1392 struct ldlm_lock *lock, *next;
1393 int left = 0, bl_ast = 0;
1397 cfs_list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1401 if (flags & LCF_LOCAL) {
1402 rc = LDLM_FL_LOCAL_ONLY;
1403 ldlm_lock_cancel(lock);
1405 rc = ldlm_cli_cancel_local(lock);
1407 /* Until we have compound requests and can send LDLM_CANCEL
1408 * requests batched with generic RPCs, we need to send cancels
1409 * with the LDLM_FL_BL_AST flag in a separate RPC from
1410 * the one being generated now. */
1411 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1412 LDLM_DEBUG(lock, "Cancel lock separately");
1413 cfs_list_del_init(&lock->l_bl_ast);
1414 cfs_list_add(&lock->l_bl_ast, &head);
1418 if (rc == LDLM_FL_LOCAL_ONLY) {
1419 /* CANCEL RPC should not be sent to server. */
1420 cfs_list_del_init(&lock->l_bl_ast);
1421 LDLM_LOCK_RELEASE(lock);
1427 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1432 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1435 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1436 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1437 * readahead requests, ...)
1439 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1440 struct ldlm_lock *lock,
1441 int unused, int added,
1444 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1446 /* don't check added & count since we want to process all locks
1448 * It's fine to not take lock to access lock->l_resource since
1449 * the lock has already been granted so it won't change. */
1450 switch (lock->l_resource->lr_type) {
1453 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1456 result = LDLM_POLICY_SKIP_LOCK;
1457 lock_res_and_lock(lock);
1458 ldlm_set_skipped(lock);
1459 unlock_res_and_lock(lock);
1467 * Callback function for LRU-resize policy. Decides whether to keep
1468 * \a lock in LRU for current \a LRU size \a unused, added in current
1469 * scan \a added and number of locks to be preferably canceled \a count.
1471 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1473 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1475 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1476 struct ldlm_lock *lock,
1477 int unused, int added,
1480 cfs_time_t cur = cfs_time_current();
1481 struct ldlm_pool *pl = &ns->ns_pool;
1485 /* Stop LRU processing when we reach past @count or have checked all
1487 if (count && added >= count)
1488 return LDLM_POLICY_KEEP_LOCK;
1490 slv = ldlm_pool_get_slv(pl);
1491 lvf = ldlm_pool_get_lvf(pl);
1492 la = cfs_duration_sec(cfs_time_sub(cur,
1493 lock->l_last_used));
1494 lv = lvf * la * unused;
1496 /* Inform pool about current CLV to see it via proc. */
1497 ldlm_pool_set_clv(pl, lv);
1499 /* Stop when SLV is not yet come from server or lv is smaller than
1501 if (slv == 0 || lv < slv)
1502 return LDLM_POLICY_KEEP_LOCK;
1504 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) == 0)
1505 return LDLM_POLICY_KEEP_LOCK;
1507 return LDLM_POLICY_CANCEL_LOCK;
1511 * Callback function for proc used policy. Makes decision whether to keep
1512 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1513 * added and number of locks to be preferably canceled \a count.
1515 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1517 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1519 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1520 struct ldlm_lock *lock,
1521 int unused, int added,
1524 /* Stop LRU processing when we reach past @count or have checked all
1526 return (added >= count) ?
1527 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1531 * Callback function for aged policy. Makes decision whether to keep \a lock in
1532 * LRU for current LRU size \a unused, added in current scan \a added and
1533 * number of locks to be preferably canceled \a count.
1535 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1537 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1539 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1540 struct ldlm_lock *lock,
1541 int unused, int added,
1545 return LDLM_POLICY_KEEP_LOCK;
1547 if (cfs_time_before(cfs_time_current(),
1548 cfs_time_add(lock->l_last_used, ns->ns_max_age)))
1549 return LDLM_POLICY_KEEP_LOCK;
1551 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) == 0)
1552 return LDLM_POLICY_KEEP_LOCK;
1554 return LDLM_POLICY_CANCEL_LOCK;
1558 * Callback function for default policy. Makes decision whether to keep \a lock
1559 * in LRU for current LRU size \a unused, added in current scan \a added and
1560 * number of locks to be preferably canceled \a count.
1562 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1564 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1566 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1567 struct ldlm_lock *lock,
1568 int unused, int added,
1571 /* Stop LRU processing when we reach past count or have checked all
1573 return (added >= count) ?
1574 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1577 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1578 struct ldlm_lock *, int,
1581 static ldlm_cancel_lru_policy_t
1582 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1584 if (flags & LDLM_CANCEL_NO_WAIT)
1585 return ldlm_cancel_no_wait_policy;
1587 if (ns_connect_lru_resize(ns)) {
1588 if (flags & LDLM_CANCEL_SHRINK)
1589 /* We kill passed number of old locks. */
1590 return ldlm_cancel_passed_policy;
1591 else if (flags & LDLM_CANCEL_LRUR)
1592 return ldlm_cancel_lrur_policy;
1593 else if (flags & LDLM_CANCEL_PASSED)
1594 return ldlm_cancel_passed_policy;
1596 if (flags & LDLM_CANCEL_AGED)
1597 return ldlm_cancel_aged_policy;
1600 return ldlm_cancel_default_policy;
1604 * - Free space in LRU for \a count new locks,
1605 * redundant unused locks are canceled locally;
1606 * - also cancel locally unused aged locks;
1607 * - do not cancel more than \a max locks;
1608 * - GET the found locks and add them into the \a cancels list.
1610 * A client lock can be added to the l_bl_ast list only when it is
1611 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1612 * CANCEL. There are the following use cases:
1613 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1614 * ldlm_cli_cancel(), which check and set this flag properly. As any
1615 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1616 * later without any special locking.
1618 * Calling policies for enabled LRU resize:
1619 * ----------------------------------------
1620 * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1621 * cancel not more than \a count locks;
1623 * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1624 * the beginning of LRU list);
1626 * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1627 * memory pressre policy function;
1629 * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1631 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1632 * (typically before replaying locks) w/o
1633 * sending any RPCs or waiting for any
1634 * outstanding RPC to complete.
1636 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns, cfs_list_t *cancels,
1637 int count, int max, int flags)
1639 ldlm_cancel_lru_policy_t pf;
1640 struct ldlm_lock *lock, *next;
1641 int added = 0, unused, remained;
1644 spin_lock(&ns->ns_lock);
1645 unused = ns->ns_nr_unused;
1648 if (!ns_connect_lru_resize(ns))
1649 count += unused - ns->ns_max_unused;
1651 pf = ldlm_cancel_lru_policy(ns, flags);
1652 LASSERT(pf != NULL);
1654 while (!cfs_list_empty(&ns->ns_unused_list)) {
1655 ldlm_policy_res_t result;
1657 /* all unused locks */
1658 if (remained-- <= 0)
1661 /* For any flags, stop scanning if @max is reached. */
1662 if (max && added >= max)
1665 cfs_list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1667 /* No locks which got blocking requests. */
1668 LASSERT(!ldlm_is_bl_ast(lock));
1670 if (flags & LDLM_CANCEL_NO_WAIT &&
1671 ldlm_is_skipped(lock))
1672 /* already processed */
1675 /* Somebody is already doing CANCEL. No need for this
1676 * lock in LRU, do not traverse it again. */
1677 if (!ldlm_is_canceling(lock))
1680 ldlm_lock_remove_from_lru_nolock(lock);
1682 if (&lock->l_lru == &ns->ns_unused_list)
1685 LDLM_LOCK_GET(lock);
1686 spin_unlock(&ns->ns_lock);
1687 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1689 /* Pass the lock through the policy filter and see if it
1690 * should stay in LRU.
1692 * Even for shrinker policy we stop scanning if
1693 * we find a lock that should stay in the cache.
1694 * We should take into account lock age anyway
1695 * as a new lock is a valuable resource even if
1696 * it has a low weight.
1698 * That is, for shrinker policy we drop only
1699 * old locks, but additionally choose them by
1700 * their weight. Big extent locks will stay in
1702 result = pf(ns, lock, unused, added, count);
1703 if (result == LDLM_POLICY_KEEP_LOCK) {
1704 lu_ref_del(&lock->l_reference,
1705 __FUNCTION__, current);
1706 LDLM_LOCK_RELEASE(lock);
1707 spin_lock(&ns->ns_lock);
1710 if (result == LDLM_POLICY_SKIP_LOCK) {
1711 lu_ref_del(&lock->l_reference,
1713 LDLM_LOCK_RELEASE(lock);
1714 spin_lock(&ns->ns_lock);
1718 lock_res_and_lock(lock);
1719 /* Check flags again under the lock. */
1720 if (ldlm_is_canceling(lock) ||
1721 (ldlm_lock_remove_from_lru(lock) == 0)) {
1722 /* Another thread is removing lock from LRU, or
1723 * somebody is already doing CANCEL, or there
1724 * is a blocking request which will send cancel
1725 * by itself, or the lock is no longer unused. */
1726 unlock_res_and_lock(lock);
1727 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1728 LDLM_LOCK_RELEASE(lock);
1729 spin_lock(&ns->ns_lock);
1732 LASSERT(!lock->l_readers && !lock->l_writers);
1734 /* If we have chosen to cancel this lock voluntarily, we
1735 * better send cancel notification to server, so that it
1736 * frees appropriate state. This might lead to a race
1737 * where while we are doing cancel here, server is also
1738 * silently cancelling this lock. */
1739 ldlm_clear_cancel_on_block(lock);
1741 /* Setting the CBPENDING flag is a little misleading,
1742 * but prevents an important race; namely, once
1743 * CBPENDING is set, the lock can accumulate no more
1744 * readers/writers. Since readers and writers are
1745 * already zero here, ldlm_lock_decref() won't see
1746 * this flag and call l_blocking_ast */
1747 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1749 /* We can't re-add to l_lru as it confuses the
1750 * refcounting in ldlm_lock_remove_from_lru() if an AST
1751 * arrives after we drop lr_lock below. We use l_bl_ast
1752 * and can't use l_pending_chain as it is used both on
1753 * server and client nevertheless bug 5666 says it is
1754 * used only on server */
1755 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1756 cfs_list_add(&lock->l_bl_ast, cancels);
1757 unlock_res_and_lock(lock);
1758 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1759 spin_lock(&ns->ns_lock);
1763 spin_unlock(&ns->ns_lock);
1767 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, cfs_list_t *cancels,
1768 int count, int max, ldlm_cancel_flags_t cancel_flags,
1772 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1775 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1779 * Cancel at least \a nr locks from given namespace LRU.
1781 * When called with LCF_ASYNC the blocking callback will be handled
1782 * in a thread and this function will return after the thread has been
1783 * asked to call the callback. When called with LCF_ASYNC the blocking
1784 * callback will be performed in this function.
1786 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1787 ldlm_cancel_flags_t cancel_flags,
1790 struct list_head cancels = LIST_HEAD_INIT(cancels);
1795 cancel_flags &= ~LCF_ASYNC; /* force to be sync in user space */
1797 /* Just prepare the list of locks, do not actually cancel them yet.
1798 * Locks are cancelled later in a separate thread. */
1799 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1800 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1808 * Find and cancel locally unused locks found on resource, matched to the
1809 * given policy, mode. GET the found locks and add them into the \a cancels
1812 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1813 cfs_list_t *cancels,
1814 ldlm_policy_data_t *policy,
1815 ldlm_mode_t mode, __u64 lock_flags,
1816 ldlm_cancel_flags_t cancel_flags, void *opaque)
1818 struct ldlm_lock *lock;
1823 cfs_list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1824 if (opaque != NULL && lock->l_ast_data != opaque) {
1825 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1826 lock->l_ast_data, opaque);
1831 if (lock->l_readers || lock->l_writers)
1834 /* If somebody is already doing CANCEL, or blocking AST came,
1835 * skip this lock. */
1836 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
1839 if (lockmode_compat(lock->l_granted_mode, mode))
1842 /* If policy is given and this is IBITS lock, add to list only
1843 * those locks that match by policy. */
1844 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1845 !(lock->l_policy_data.l_inodebits.bits &
1846 policy->l_inodebits.bits))
1849 /* See CBPENDING comment in ldlm_cancel_lru */
1850 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1853 LASSERT(cfs_list_empty(&lock->l_bl_ast));
1854 cfs_list_add(&lock->l_bl_ast, cancels);
1855 LDLM_LOCK_GET(lock);
1860 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1862 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1865 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1867 * If \a req is NULL, send CANCEL request to server with handles of locks
1868 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1869 * separately per lock.
1870 * If \a req is not NULL, put handles of locks in \a cancels into the request
1871 * buffer at the offset \a off.
1872 * Destroy \a cancels at the end.
1874 int ldlm_cli_cancel_list(cfs_list_t *cancels, int count,
1875 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1877 struct ldlm_lock *lock;
1881 if (cfs_list_empty(cancels) || count == 0)
1884 /* XXX: requests (both batched and not) could be sent in parallel.
1885 * Usually it is enough to have just 1 RPC, but it is possible that
1886 * there are too many locks to be cancelled in LRU or on a resource.
1887 * It would also speed up the case when the server does not support
1890 LASSERT(!cfs_list_empty(cancels));
1891 lock = cfs_list_entry(cancels->next, struct ldlm_lock,
1893 LASSERT(lock->l_conn_export);
1895 if (exp_connect_cancelset(lock->l_conn_export)) {
1898 ldlm_cancel_pack(req, cancels, count);
1900 res = ldlm_cli_cancel_req(lock->l_conn_export,
1904 res = ldlm_cli_cancel_req(lock->l_conn_export,
1909 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1910 "ldlm_cli_cancel_list: %d\n", res);
1915 ldlm_lock_list_put(cancels, l_bl_ast, res);
1917 LASSERT(count == 0);
1920 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1923 * Cancel all locks on a resource that have 0 readers/writers.
1925 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1926 * to notify the server. */
1927 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1928 const struct ldlm_res_id *res_id,
1929 ldlm_policy_data_t *policy,
1931 ldlm_cancel_flags_t flags,
1934 struct ldlm_resource *res;
1935 struct list_head cancels = LIST_HEAD_INIT(cancels);
1940 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1942 /* This is not a problem. */
1943 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1947 LDLM_RESOURCE_ADDREF(res);
1948 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1949 0, flags | LCF_BL_AST, opaque);
1950 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1952 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
1955 LDLM_RESOURCE_DELREF(res);
1956 ldlm_resource_putref(res);
1959 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1961 struct ldlm_cli_cancel_arg {
1966 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1967 cfs_hlist_node_t *hnode, void *arg)
1969 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1970 struct ldlm_cli_cancel_arg *lc = arg;
1972 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1973 NULL, LCK_MINMODE, lc->lc_flags,
1975 /* must return 0 for hash iteration */
1980 * Cancel all locks on a namespace (or a specific resource, if given)
1981 * that have 0 readers/writers.
1983 * If flags & LCF_LOCAL, throw the locks away without trying
1984 * to notify the server. */
1985 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1986 const struct ldlm_res_id *res_id,
1987 ldlm_cancel_flags_t flags, void *opaque)
1989 struct ldlm_cli_cancel_arg arg = {
1991 .lc_opaque = opaque,
1999 if (res_id != NULL) {
2000 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
2004 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2005 ldlm_cli_hash_cancel_unused, &arg);
2009 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2011 /* Lock iterators. */
2013 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2016 cfs_list_t *tmp, *next;
2017 struct ldlm_lock *lock;
2018 int rc = LDLM_ITER_CONTINUE;
2023 RETURN(LDLM_ITER_CONTINUE);
2026 cfs_list_for_each_safe(tmp, next, &res->lr_granted) {
2027 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2029 if (iter(lock, closure) == LDLM_ITER_STOP)
2030 GOTO(out, rc = LDLM_ITER_STOP);
2033 cfs_list_for_each_safe(tmp, next, &res->lr_converting) {
2034 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2036 if (iter(lock, closure) == LDLM_ITER_STOP)
2037 GOTO(out, rc = LDLM_ITER_STOP);
2040 cfs_list_for_each_safe(tmp, next, &res->lr_waiting) {
2041 lock = cfs_list_entry(tmp, struct ldlm_lock, l_res_link);
2043 if (iter(lock, closure) == LDLM_ITER_STOP)
2044 GOTO(out, rc = LDLM_ITER_STOP);
2050 EXPORT_SYMBOL(ldlm_resource_foreach);
2052 struct iter_helper_data {
2053 ldlm_iterator_t iter;
2057 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2059 struct iter_helper_data *helper = closure;
2060 return helper->iter(lock, helper->closure);
2063 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2064 cfs_hlist_node_t *hnode, void *arg)
2067 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2069 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2073 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2074 ldlm_iterator_t iter, void *closure)
2077 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2079 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2080 ldlm_res_iter_helper, &helper);
2083 EXPORT_SYMBOL(ldlm_namespace_foreach);
2085 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2086 * return 0: find no resource
2087 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2090 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2091 const struct ldlm_res_id *res_id,
2092 ldlm_iterator_t iter, void *data)
2094 struct ldlm_resource *res;
2098 LASSERTF(ns != NULL, "must pass in namespace\n");
2100 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2104 LDLM_RESOURCE_ADDREF(res);
2105 rc = ldlm_resource_foreach(res, iter, data);
2106 LDLM_RESOURCE_DELREF(res);
2107 ldlm_resource_putref(res);
2110 EXPORT_SYMBOL(ldlm_resource_iterate);
2114 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2116 cfs_list_t *list = closure;
2118 /* we use l_pending_chain here, because it's unused on clients. */
2119 LASSERTF(cfs_list_empty(&lock->l_pending_chain),
2120 "lock %p next %p prev %p\n",
2121 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2122 /* bug 9573: don't replay locks left after eviction, or
2123 * bug 17614: locks being actively cancelled. Get a reference
2124 * on a lock so that it does not disapear under us (e.g. due to cancel)
2126 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2127 cfs_list_add(&lock->l_pending_chain, list);
2128 LDLM_LOCK_GET(lock);
2131 return LDLM_ITER_CONTINUE;
2134 static int replay_lock_interpret(const struct lu_env *env,
2135 struct ptlrpc_request *req,
2136 struct ldlm_async_args *aa, int rc)
2138 struct ldlm_lock *lock;
2139 struct ldlm_reply *reply;
2140 struct obd_export *exp;
2143 atomic_dec(&req->rq_import->imp_replay_inflight);
2147 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2149 GOTO(out, rc = -EPROTO);
2151 lock = ldlm_handle2lock(&aa->lock_handle);
2153 CERROR("received replay ack for unknown local cookie "LPX64
2154 " remote cookie "LPX64 " from server %s id %s\n",
2155 aa->lock_handle.cookie, reply->lock_handle.cookie,
2156 req->rq_export->exp_client_uuid.uuid,
2157 libcfs_id2str(req->rq_peer));
2158 GOTO(out, rc = -ESTALE);
2161 /* Key change rehash lock in per-export hash with new key */
2162 exp = req->rq_export;
2163 if (exp && exp->exp_lock_hash) {
2164 /* In the function below, .hs_keycmp resolves to
2165 * ldlm_export_lock_keycmp() */
2166 /* coverity[overrun-buffer-val] */
2167 cfs_hash_rehash_key(exp->exp_lock_hash,
2168 &lock->l_remote_handle,
2169 &reply->lock_handle,
2172 lock->l_remote_handle = reply->lock_handle;
2175 LDLM_DEBUG(lock, "replayed lock:");
2176 ptlrpc_import_recovery_state_machine(req->rq_import);
2177 LDLM_LOCK_PUT(lock);
2180 ptlrpc_connect_import(req->rq_import);
2185 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2187 struct ptlrpc_request *req;
2188 struct ldlm_async_args *aa;
2189 struct ldlm_request *body;
2194 /* Bug 11974: Do not replay a lock which is actively being canceled */
2195 if (ldlm_is_canceling(lock)) {
2196 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2200 /* If this is reply-less callback lock, we cannot replay it, since
2201 * server might have long dropped it, but notification of that event was
2202 * lost by network. (and server granted conflicting lock already) */
2203 if (ldlm_is_cancel_on_block(lock)) {
2204 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2205 ldlm_lock_cancel(lock);
2210 * If granted mode matches the requested mode, this lock is granted.
2212 * If they differ, but we have a granted mode, then we were granted
2213 * one mode and now want another: ergo, converting.
2215 * If we haven't been granted anything and are on a resource list,
2216 * then we're blocked/waiting.
2218 * If we haven't been granted anything and we're NOT on a resource list,
2219 * then we haven't got a reply yet and don't have a known disposition.
2220 * This happens whenever a lock enqueue is the request that triggers
2223 if (lock->l_granted_mode == lock->l_req_mode)
2224 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2225 else if (lock->l_granted_mode)
2226 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2227 else if (!cfs_list_empty(&lock->l_res_link))
2228 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2230 flags = LDLM_FL_REPLAY;
2232 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2233 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2237 /* We're part of recovery, so don't wait for it. */
2238 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2240 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2241 ldlm_lock2desc(lock, &body->lock_desc);
2242 body->lock_flags = ldlm_flags_to_wire(flags);
2244 ldlm_lock2handle(lock, &body->lock_handle[0]);
2245 if (lock->l_lvb_len > 0)
2246 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2247 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2249 ptlrpc_request_set_replen(req);
2250 /* notify the server we've replayed all requests.
2251 * also, we mark the request to be put on a dedicated
2252 * queue to be processed after all request replayes.
2254 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2256 LDLM_DEBUG(lock, "replaying lock:");
2258 atomic_inc(&req->rq_import->imp_replay_inflight);
2259 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2260 aa = ptlrpc_req_async_args(req);
2261 aa->lock_handle = body->lock_handle[0];
2262 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2263 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2269 * Cancel as many unused locks as possible before replay. since we are
2270 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2273 * Called only in recovery before replaying locks. there is no need to
2274 * replay locks that are unused. since the clients may hold thousands of
2275 * cached unused locks, dropping the unused locks can greatly reduce the
2276 * load on the servers at recovery time.
2278 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2281 struct list_head cancels = LIST_HEAD_INIT(cancels);
2283 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2284 "replay for namespace %s (%d)\n",
2285 ldlm_ns_name(ns), ns->ns_nr_unused);
2287 /* We don't need to care whether or not LRU resize is enabled
2288 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2289 * count parameter */
2290 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2291 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2293 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2294 canceled, ldlm_ns_name(ns));
2297 int ldlm_replay_locks(struct obd_import *imp)
2299 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2300 struct list_head list = LIST_HEAD_INIT(list);
2301 struct ldlm_lock *lock, *next;
2306 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2308 /* don't replay locks if import failed recovery */
2309 if (imp->imp_vbr_failed)
2312 /* ensure this doesn't fall to 0 before all have been queued */
2313 atomic_inc(&imp->imp_replay_inflight);
2315 if (ldlm_cancel_unused_locks_before_replay)
2316 ldlm_cancel_unused_locks_for_replay(ns);
2318 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2320 cfs_list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2321 cfs_list_del_init(&lock->l_pending_chain);
2323 LDLM_LOCK_RELEASE(lock);
2324 continue; /* or try to do the rest? */
2326 rc = replay_one_lock(imp, lock);
2327 LDLM_LOCK_RELEASE(lock);
2330 atomic_dec(&imp->imp_replay_inflight);
2334 EXPORT_SYMBOL(ldlm_replay_locks);