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
64 #include <lustre_dlm.h>
65 #include <obd_class.h>
68 #include "ldlm_internal.h"
70 int ldlm_enqueue_min = OBD_TIMEOUT_DEFAULT;
71 CFS_MODULE_PARM(ldlm_enqueue_min, "i", int, 0644,
72 "lock enqueue timeout minimum");
74 /* in client side, whether the cached locks will be canceled before replay */
75 unsigned int ldlm_cancel_unused_locks_before_replay = 1;
77 static void interrupted_completion_wait(void *data)
81 struct lock_wait_data {
82 struct ldlm_lock *lwd_lock;
86 struct ldlm_async_args {
87 struct lustre_handle lock_handle;
90 int ldlm_expired_completion_wait(void *data)
92 struct lock_wait_data *lwd = data;
93 struct ldlm_lock *lock = lwd->lwd_lock;
94 struct obd_import *imp;
95 struct obd_device *obd;
98 if (lock->l_conn_export == NULL) {
99 static cfs_time_t next_dump = 0, last_dump = 0;
101 LCONSOLE_WARN("lock timed out (enqueued at "CFS_TIME_T", "
102 CFS_DURATION_T"s ago)\n",
103 lock->l_last_activity,
104 cfs_time_sub(cfs_time_current_sec(),
105 lock->l_last_activity));
106 LDLM_DEBUG(lock, "lock timed out (enqueued at "CFS_TIME_T", "
107 CFS_DURATION_T"s ago); not entering recovery in "
108 "server code, just going back to sleep",
109 lock->l_last_activity,
110 cfs_time_sub(cfs_time_current_sec(),
111 lock->l_last_activity));
112 if (cfs_time_after(cfs_time_current(), next_dump)) {
113 last_dump = next_dump;
114 next_dump = cfs_time_shift(300);
115 ldlm_namespace_dump(D_DLMTRACE,
116 ldlm_lock_to_ns(lock));
118 libcfs_debug_dumplog();
123 obd = lock->l_conn_export->exp_obd;
124 imp = obd->u.cli.cl_import;
125 ptlrpc_fail_import(imp, lwd->lwd_conn_cnt);
126 LDLM_ERROR(lock, "lock timed out (enqueued at "CFS_TIME_T", "
127 CFS_DURATION_T"s ago), entering recovery for %s@%s",
128 lock->l_last_activity,
129 cfs_time_sub(cfs_time_current_sec(), lock->l_last_activity),
130 obd2cli_tgt(obd), imp->imp_connection->c_remote_uuid.uuid);
134 EXPORT_SYMBOL(ldlm_expired_completion_wait);
136 /* We use the same basis for both server side and client side functions
137 from a single node. */
138 int ldlm_get_enq_timeout(struct ldlm_lock *lock)
140 int timeout = at_get(ldlm_lock_to_ns_at(lock));
142 return obd_timeout / 2;
143 /* Since these are non-updating timeouts, we should be conservative.
144 It would be nice to have some kind of "early reply" mechanism for
145 lock callbacks too... */
146 timeout = min_t(int, at_max, timeout + (timeout >> 1)); /* 150% */
147 return max(timeout, ldlm_enqueue_min);
149 EXPORT_SYMBOL(ldlm_get_enq_timeout);
152 * Helper function for ldlm_completion_ast(), updating timings when lock is
155 static int ldlm_completion_tail(struct ldlm_lock *lock)
160 if (ldlm_is_destroyed(lock) || ldlm_is_failed(lock)) {
161 LDLM_DEBUG(lock, "client-side enqueue: destroyed");
164 delay = cfs_time_sub(cfs_time_current_sec(),
165 lock->l_last_activity);
166 LDLM_DEBUG(lock, "client-side enqueue: granted after "
167 CFS_DURATION_T"s", delay);
169 /* Update our time estimate */
170 at_measured(ldlm_lock_to_ns_at(lock),
178 * Implementation of ->l_completion_ast() for a client, that doesn't wait
179 * until lock is granted. Suitable for locks enqueued through ptlrpcd, of
180 * other threads that cannot block for long.
182 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data)
186 if (flags == LDLM_FL_WAIT_NOREPROC) {
187 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
191 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
192 LDLM_FL_BLOCK_CONV))) {
193 wake_up(&lock->l_waitq);
194 RETURN(ldlm_completion_tail(lock));
197 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
199 ldlm_reprocess_all(lock->l_resource);
202 EXPORT_SYMBOL(ldlm_completion_ast_async);
205 * Generic LDLM "completion" AST. This is called in several cases:
207 * - when a reply to an ENQUEUE RPC is received from the server
208 * (ldlm_cli_enqueue_fini()). Lock might be granted or not granted at
209 * this point (determined by flags);
211 * - when LDLM_CP_CALLBACK RPC comes to client to notify it that lock has
214 * - when ldlm_lock_match(LDLM_FL_LVB_READY) is about to wait until lock
217 * - to force all locks when resource is destroyed (cleanup_resource());
219 * - during lock conversion (not used currently).
221 * If lock is not granted in the first case, this function waits until second
222 * or penultimate cases happen in some other thread.
225 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
227 /* XXX ALLOCATE - 160 bytes */
228 struct lock_wait_data lwd;
229 struct obd_device *obd;
230 struct obd_import *imp = NULL;
231 struct l_wait_info lwi;
236 if (flags == LDLM_FL_WAIT_NOREPROC) {
237 LDLM_DEBUG(lock, "client-side enqueue waiting on pending lock");
241 if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
242 LDLM_FL_BLOCK_CONV))) {
243 wake_up(&lock->l_waitq);
247 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
252 obd = class_exp2obd(lock->l_conn_export);
254 /* if this is a local lock, then there is no import */
256 imp = obd->u.cli.cl_import;
259 /* Wait a long time for enqueue - server may have to callback a
260 lock from another client. Server will evict the other client if it
261 doesn't respond reasonably, and then give us the lock. */
262 timeout = ldlm_get_enq_timeout(lock) * 2;
266 if (ldlm_is_no_timeout(lock)) {
267 LDLM_DEBUG(lock, "waiting indefinitely because of NO_TIMEOUT");
268 lwi = LWI_INTR(interrupted_completion_wait, &lwd);
270 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout),
271 ldlm_expired_completion_wait,
272 interrupted_completion_wait, &lwd);
276 spin_lock(&imp->imp_lock);
277 lwd.lwd_conn_cnt = imp->imp_conn_cnt;
278 spin_unlock(&imp->imp_lock);
281 if (ns_is_client(ldlm_lock_to_ns(lock)) &&
282 OBD_FAIL_CHECK_RESET(OBD_FAIL_LDLM_INTR_CP_AST,
283 OBD_FAIL_LDLM_CP_BL_RACE | OBD_FAIL_ONCE)) {
284 ldlm_set_fail_loc(lock);
287 /* Go to sleep until the lock is granted or cancelled. */
288 rc = l_wait_event(lock->l_waitq,
289 is_granted_or_cancelled(lock), &lwi);
293 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
298 RETURN(ldlm_completion_tail(lock));
300 EXPORT_SYMBOL(ldlm_completion_ast);
303 * A helper to build a blocking AST function
305 * Perform a common operation for blocking ASTs:
306 * defferred lock cancellation.
308 * \param lock the lock blocking or canceling AST was called on
310 * \see mdt_blocking_ast
311 * \see ldlm_blocking_ast
313 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock)
318 ldlm_set_cbpending(lock);
319 do_ast = (!lock->l_readers && !lock->l_writers);
320 unlock_res_and_lock(lock);
323 struct lustre_handle lockh;
326 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
327 ldlm_lock2handle(lock, &lockh);
328 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
330 CERROR("ldlm_cli_cancel: %d\n", rc);
332 LDLM_DEBUG(lock, "Lock still has references, will be "
337 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
340 * Server blocking AST
342 * ->l_blocking_ast() callback for LDLM locks acquired by server-side
345 * \param lock the lock which blocks a request or cancelling lock
348 * \param flag indicates whether this cancelling or blocking callback
350 * \see ldlm_blocking_ast_nocheck
352 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
353 void *data, int flag)
357 if (flag == LDLM_CB_CANCELING) {
358 /* Don't need to do anything here. */
362 lock_res_and_lock(lock);
363 /* Get this: if ldlm_blocking_ast is racing with intent_policy, such
364 * that ldlm_blocking_ast is called just before intent_policy method
365 * takes the lr_lock, then by the time we get the lock, we might not
366 * be the correct blocking function anymore. So check, and return
368 if (lock->l_blocking_ast != ldlm_blocking_ast) {
369 unlock_res_and_lock(lock);
372 RETURN(ldlm_blocking_ast_nocheck(lock));
374 EXPORT_SYMBOL(ldlm_blocking_ast);
377 * Implements ldlm_lock::l_glimpse_ast for extent locks acquired on the server.
379 * Returning -ELDLM_NO_LOCK_DATA actually works, but the reason for that is
380 * rather subtle: with OST-side locking, it may so happen that _all_ extent
381 * locks are held by the OST. If client wants to obtain the current file size
382 * it calls ll_glimpse_size(), and (as all locks are held only on the server),
383 * this dummy glimpse callback fires and does nothing. The client still
384 * receives the correct file size due to the following fragment of code in
385 * ldlm_cb_interpret():
387 * if (rc == -ELDLM_NO_LOCK_DATA) {
388 * LDLM_DEBUG(lock, "lost race - client has a lock but no"
390 * ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
393 * That is, after the glimpse returns this error, ofd_lvbo_update() is called
394 * and returns the updated file attributes from the inode to the client.
396 * See also comment in ofd_intent_policy() on why servers must set a non-NULL
397 * l_glimpse_ast when grabbing DLM locks. Otherwise, the server will assume
398 * that the object is in the process of being destroyed.
400 * \param[in] lock DLM lock being glimpsed, unused
401 * \param[in] reqp pointer to ptlrpc_request, unused
403 * \retval -ELDLM_NO_LOCK_DATA to get attributes from disk object
405 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp)
407 return -ELDLM_NO_LOCK_DATA;
409 EXPORT_SYMBOL(ldlm_glimpse_ast);
412 * Enqueue a local lock (typically on a server).
414 int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
415 const struct ldlm_res_id *res_id,
416 ldlm_type_t type, ldlm_policy_data_t *policy,
417 ldlm_mode_t mode, __u64 *flags,
418 ldlm_blocking_callback blocking,
419 ldlm_completion_callback completion,
420 ldlm_glimpse_callback glimpse,
421 void *data, __u32 lvb_len, enum lvb_type lvb_type,
422 const __u64 *client_cookie,
423 struct lustre_handle *lockh)
425 struct ldlm_lock *lock;
427 const struct ldlm_callback_suite cbs = { .lcs_completion = completion,
428 .lcs_blocking = blocking,
429 .lcs_glimpse = glimpse,
433 LASSERT(!(*flags & LDLM_FL_REPLAY));
434 if (unlikely(ns_is_client(ns))) {
435 CERROR("Trying to enqueue local lock in a shadow namespace\n");
439 lock = ldlm_lock_create(ns, res_id, type, mode, &cbs, data, lvb_len,
442 GOTO(out_nolock, err = PTR_ERR(lock));
444 err = ldlm_lvbo_init(lock->l_resource);
446 LDLM_ERROR(lock, "delayed lvb init failed (rc %d)", err);
450 ldlm_lock2handle(lock, lockh);
452 /* NB: we don't have any lock now (lock_res_and_lock)
453 * because it's a new lock */
454 ldlm_lock_addref_internal_nolock(lock, mode);
455 ldlm_set_local(lock);
456 if (*flags & LDLM_FL_ATOMIC_CB)
457 ldlm_set_atomic_cb(lock);
460 lock->l_policy_data = *policy;
461 if (client_cookie != NULL)
462 lock->l_client_cookie = *client_cookie;
463 if (type == LDLM_EXTENT) {
464 /* extent lock without policy is a bug */
468 lock->l_req_extent = policy->l_extent;
471 err = ldlm_lock_enqueue(ns, &lock, policy, flags);
472 if (unlikely(err != ELDLM_OK))
476 *policy = lock->l_policy_data;
478 if (lock->l_completion_ast)
479 lock->l_completion_ast(lock, *flags, NULL);
481 LDLM_DEBUG(lock, "client-side local enqueue handler, new lock created");
484 LDLM_LOCK_RELEASE(lock);
488 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
490 static void failed_lock_cleanup(struct ldlm_namespace *ns,
491 struct ldlm_lock *lock, int mode)
495 /* Set a flag to prevent us from sending a CANCEL (bug 407) */
496 lock_res_and_lock(lock);
497 /* Check that lock is not granted or failed, we might race. */
498 if ((lock->l_req_mode != lock->l_granted_mode) &&
499 !ldlm_is_failed(lock)) {
500 /* Make sure that this lock will not be found by raced
501 * bl_ast and -EINVAL reply is sent to server anyways.
503 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_FAILED |
504 LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING;
507 unlock_res_and_lock(lock);
511 "setting FL_LOCAL_ONLY | LDLM_FL_FAILED | "
512 "LDLM_FL_ATOMIC_CB | LDLM_FL_CBPENDING");
514 LDLM_DEBUG(lock, "lock was granted or failed in race");
516 /* XXX - HACK because we shouldn't call ldlm_lock_destroy()
517 * from llite/file.c/ll_file_flock(). */
518 /* This code makes for the fact that we do not have blocking handler on
519 * a client for flock locks. As such this is the place where we must
520 * completely kill failed locks. (interrupted and those that
521 * were waiting to be granted when server evicted us. */
522 if (lock->l_resource->lr_type == LDLM_FLOCK) {
523 lock_res_and_lock(lock);
524 if (!ldlm_is_destroyed(lock)) {
525 ldlm_resource_unlink_lock(lock);
526 ldlm_lock_decref_internal_nolock(lock, mode);
527 ldlm_lock_destroy_nolock(lock);
529 unlock_res_and_lock(lock);
531 ldlm_lock_decref_internal(lock, mode);
536 * Finishing portion of client lock enqueue code.
538 * Called after receiving reply from server.
540 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
541 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
542 __u64 *flags, void *lvb, __u32 lvb_len,
543 struct lustre_handle *lockh,int rc)
545 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
546 int is_replay = *flags & LDLM_FL_REPLAY;
547 struct ldlm_lock *lock;
548 struct ldlm_reply *reply;
549 int cleanup_phase = 1;
553 lock = ldlm_handle2lock(lockh);
554 /* ldlm_cli_enqueue is holding a reference on this lock. */
556 LASSERT(type == LDLM_FLOCK);
560 LASSERTF(ergo(lvb_len != 0, lvb_len == lock->l_lvb_len),
561 "lvb_len = %d, l_lvb_len = %d\n", lvb_len, lock->l_lvb_len);
563 if (rc != ELDLM_OK) {
565 LDLM_DEBUG(lock, "client-side enqueue END (%s)",
566 rc == ELDLM_LOCK_ABORTED ? "ABORTED" : "FAILED");
568 if (rc != ELDLM_LOCK_ABORTED)
572 /* Before we return, swab the reply */
573 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
575 GOTO(cleanup, rc = -EPROTO);
578 LASSERT(lvb != NULL);
580 size = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
583 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", size);
584 GOTO(cleanup, rc = size);
585 } else if (unlikely(size > lvb_len)) {
586 LDLM_ERROR(lock, "Replied LVB is larger than "
587 "expectation, expected = %d, replied = %d",
589 GOTO(cleanup, rc = -EINVAL);
593 if (rc == ELDLM_LOCK_ABORTED) {
595 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
597 GOTO(cleanup, rc = (rc != 0 ? rc : ELDLM_LOCK_ABORTED));
600 /* lock enqueued on the server */
603 lock_res_and_lock(lock);
604 /* Key change rehash lock in per-export hash with new key */
605 if (exp->exp_lock_hash) {
606 /* In the function below, .hs_keycmp resolves to
607 * ldlm_export_lock_keycmp() */
608 /* coverity[overrun-buffer-val] */
609 cfs_hash_rehash_key(exp->exp_lock_hash,
610 &lock->l_remote_handle,
614 lock->l_remote_handle = reply->lock_handle;
617 *flags = ldlm_flags_from_wire(reply->lock_flags);
618 lock->l_flags |= ldlm_flags_from_wire(reply->lock_flags &
619 LDLM_FL_INHERIT_MASK);
620 unlock_res_and_lock(lock);
622 CDEBUG(D_INFO, "local: %p, remote cookie: "LPX64", flags: "LPX64"\n",
623 lock, reply->lock_handle.cookie, *flags);
625 /* If enqueue returned a blocked lock but the completion handler has
626 * already run, then it fixed up the resource and we don't need to do it
628 if ((*flags) & LDLM_FL_LOCK_CHANGED) {
629 int newmode = reply->lock_desc.l_req_mode;
631 if (newmode && newmode != lock->l_req_mode) {
632 LDLM_DEBUG(lock, "server returned different mode %s",
633 ldlm_lockname[newmode]);
634 lock->l_req_mode = newmode;
637 if (!ldlm_res_eq(&reply->lock_desc.l_resource.lr_name,
638 &lock->l_resource->lr_name)) {
639 CDEBUG(D_INFO, "remote intent success, locking "DLDLMRES
640 " instead of "DLDLMRES"\n",
641 PLDLMRES(&reply->lock_desc.l_resource),
642 PLDLMRES(lock->l_resource));
644 rc = ldlm_lock_change_resource(ns, lock,
645 &reply->lock_desc.l_resource.lr_name);
646 if (rc || lock->l_resource == NULL)
647 GOTO(cleanup, rc = -ENOMEM);
648 LDLM_DEBUG(lock, "client-side enqueue, new resource");
651 if (!(type == LDLM_IBITS &&
652 !(exp_connect_flags(exp) & OBD_CONNECT_IBITS)))
653 /* We assume lock type cannot change on server*/
654 ldlm_convert_policy_to_local(exp,
655 lock->l_resource->lr_type,
656 &reply->lock_desc.l_policy_data,
657 &lock->l_policy_data);
658 if (type != LDLM_PLAIN)
659 LDLM_DEBUG(lock,"client-side enqueue, new policy data");
662 if ((*flags) & LDLM_FL_AST_SENT) {
663 lock_res_and_lock(lock);
664 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
665 unlock_res_and_lock(lock);
666 LDLM_DEBUG(lock, "enqueue reply includes blocking AST");
669 /* If the lock has already been granted by a completion AST, don't
670 * clobber the LVB with an older one. */
672 /* We must lock or a racing completion might update lvb without
673 * letting us know and we'll clobber the correct value.
674 * Cannot unlock after the check either, a that still leaves
675 * a tiny window for completion to get in */
676 lock_res_and_lock(lock);
677 if (lock->l_req_mode != lock->l_granted_mode)
678 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_SERVER,
679 lock->l_lvb_data, size);
680 unlock_res_and_lock(lock);
688 rc = ldlm_lock_enqueue(ns, &lock, NULL, flags);
689 if (lock->l_completion_ast != NULL) {
690 int err = lock->l_completion_ast(lock, *flags, NULL);
698 if (lvb_len && lvb != NULL) {
699 /* Copy the LVB here, and not earlier, because the completion
700 * AST (if any) can override what we got in the reply */
701 memcpy(lvb, lock->l_lvb_data, lvb_len);
704 LDLM_DEBUG(lock, "client-side enqueue END");
707 if (cleanup_phase == 1 && rc)
708 failed_lock_cleanup(ns, lock, mode);
709 /* Put lock 2 times, the second reference is held by ldlm_cli_enqueue */
711 LDLM_LOCK_RELEASE(lock);
714 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
717 * Estimate number of lock handles that would fit into request of given
718 * size. PAGE_SIZE-512 is to allow TCP/IP and LNET headers to fit into
719 * a single page on the send/receive side. XXX: 512 should be changed to
720 * more adequate value.
722 static inline int ldlm_req_handles_avail(int req_size, int off)
726 avail = min_t(int, LDLM_MAXREQSIZE, PAGE_CACHE_SIZE - 512) - req_size;
727 if (likely(avail >= 0))
728 avail /= (int)sizeof(struct lustre_handle);
731 avail += LDLM_LOCKREQ_HANDLES - off;
736 static inline int ldlm_capsule_handles_avail(struct req_capsule *pill,
737 enum req_location loc,
740 int size = req_capsule_msg_size(pill, loc);
741 return ldlm_req_handles_avail(size, off);
744 static inline int ldlm_format_handles_avail(struct obd_import *imp,
745 const struct req_format *fmt,
746 enum req_location loc, int off)
748 int size = req_capsule_fmt_size(imp->imp_msg_magic, fmt, loc);
749 return ldlm_req_handles_avail(size, off);
753 * Cancel LRU locks and pack them into the enqueue request. Pack there the given
754 * \a count locks in \a cancels.
756 * This is to be called by functions preparing their own requests that
757 * might contain lists of locks to cancel in addition to actual operation
758 * that needs to be performed.
760 int ldlm_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
761 int version, int opc, int canceloff,
762 struct list_head *cancels, int count)
764 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
765 struct req_capsule *pill = &req->rq_pill;
766 struct ldlm_request *dlm = NULL;
767 struct list_head head = LIST_HEAD_INIT(head);
768 int flags, avail, to_free, pack = 0;
774 if (ns_connect_cancelset(ns)) {
775 /* Estimate the amount of available space in the request. */
776 req_capsule_filled_sizes(pill, RCL_CLIENT);
777 avail = ldlm_capsule_handles_avail(pill, RCL_CLIENT, canceloff);
779 flags = ns_connect_lru_resize(ns) ?
780 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
781 to_free = !ns_connect_lru_resize(ns) &&
782 opc == LDLM_ENQUEUE ? 1 : 0;
784 /* Cancel LRU locks here _only_ if the server supports
785 * EARLY_CANCEL. Otherwise we have to send extra CANCEL
786 * RPC, which will make us slower. */
788 count += ldlm_cancel_lru_local(ns, cancels, to_free,
789 avail - count, 0, flags);
794 req_capsule_set_size(pill, &RMF_DLM_REQ, RCL_CLIENT,
795 ldlm_request_bufsize(pack, opc));
798 rc = ptlrpc_request_pack(req, version, opc);
800 ldlm_lock_list_put(cancels, l_bl_ast, count);
804 if (ns_connect_cancelset(ns)) {
806 dlm = req_capsule_client_get(pill, &RMF_DLM_REQ);
808 /* Skip first lock handler in ldlm_request_pack(),
809 * this method will incrment @lock_count according
810 * to the lock handle amount actually written to
812 dlm->lock_count = canceloff;
814 /* Pack into the request @pack lock handles. */
815 ldlm_cli_cancel_list(cancels, pack, req, 0);
816 /* Prepare and send separate cancel RPC for others. */
817 ldlm_cli_cancel_list(cancels, count - pack, NULL, 0);
819 ldlm_lock_list_put(cancels, l_bl_ast, count);
823 EXPORT_SYMBOL(ldlm_prep_elc_req);
825 int ldlm_prep_enqueue_req(struct obd_export *exp, struct ptlrpc_request *req,
826 struct list_head *cancels, int count)
828 return ldlm_prep_elc_req(exp, req, LUSTRE_DLM_VERSION, LDLM_ENQUEUE,
829 LDLM_ENQUEUE_CANCEL_OFF, cancels, count);
831 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
833 struct ptlrpc_request *ldlm_enqueue_pack(struct obd_export *exp, int lvb_len)
835 struct ptlrpc_request *req;
839 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
841 RETURN(ERR_PTR(-ENOMEM));
843 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
845 ptlrpc_request_free(req);
849 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
850 ptlrpc_request_set_replen(req);
853 EXPORT_SYMBOL(ldlm_enqueue_pack);
856 * Client-side lock enqueue.
858 * If a request has some specific initialisation it is passed in \a reqp,
859 * otherwise it is created in ldlm_cli_enqueue.
861 * Supports sync and async requests, pass \a async flag accordingly. If a
862 * request was created in ldlm_cli_enqueue and it is the async request,
863 * pass it to the caller in \a reqp.
865 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
866 struct ldlm_enqueue_info *einfo,
867 const struct ldlm_res_id *res_id,
868 ldlm_policy_data_t const *policy, __u64 *flags,
869 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
870 struct lustre_handle *lockh, int async)
872 struct ldlm_namespace *ns;
873 struct ldlm_lock *lock;
874 struct ldlm_request *body;
875 int is_replay = *flags & LDLM_FL_REPLAY;
876 int req_passed_in = 1;
878 struct ptlrpc_request *req;
881 LASSERT(exp != NULL);
883 ns = exp->exp_obd->obd_namespace;
885 /* If we're replaying this lock, just check some invariants.
886 * If we're creating a new lock, get everything all setup nice. */
888 lock = ldlm_handle2lock_long(lockh, 0);
889 LASSERT(lock != NULL);
890 LDLM_DEBUG(lock, "client-side enqueue START");
891 LASSERT(exp == lock->l_conn_export);
893 const struct ldlm_callback_suite cbs = {
894 .lcs_completion = einfo->ei_cb_cp,
895 .lcs_blocking = einfo->ei_cb_bl,
896 .lcs_glimpse = einfo->ei_cb_gl
898 lock = ldlm_lock_create(ns, res_id, einfo->ei_type,
899 einfo->ei_mode, &cbs, einfo->ei_cbdata,
902 RETURN(PTR_ERR(lock));
903 /* for the local lock, add the reference */
904 ldlm_lock_addref_internal(lock, einfo->ei_mode);
905 ldlm_lock2handle(lock, lockh);
907 lock->l_policy_data = *policy;
909 if (einfo->ei_type == LDLM_EXTENT) {
910 /* extent lock without policy is a bug */
914 lock->l_req_extent = policy->l_extent;
916 LDLM_DEBUG(lock, "client-side enqueue START, flags "LPX64"\n",
920 lock->l_conn_export = exp;
921 lock->l_export = NULL;
922 lock->l_blocking_ast = einfo->ei_cb_bl;
923 lock->l_flags |= (*flags & (LDLM_FL_NO_LRU | LDLM_FL_EXCL));
925 /* lock not sent to server yet */
927 if (reqp == NULL || *reqp == NULL) {
928 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
933 failed_lock_cleanup(ns, lock, einfo->ei_mode);
934 LDLM_LOCK_RELEASE(lock);
944 len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ,
946 LASSERTF(len >= sizeof(*body), "buflen[%d] = %d, not %d\n",
947 DLM_LOCKREQ_OFF, len, (int)sizeof(*body));
950 /* Dump lock data into the request buffer */
951 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
952 ldlm_lock2desc(lock, &body->lock_desc);
953 body->lock_flags = ldlm_flags_to_wire(*flags);
954 body->lock_handle[0] = *lockh;
956 /* Continue as normal. */
957 if (!req_passed_in) {
959 req_capsule_extend(&req->rq_pill,
960 &RQF_LDLM_ENQUEUE_LVB);
961 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
963 ptlrpc_request_set_replen(req);
967 LASSERT(reqp != NULL);
971 LDLM_DEBUG(lock, "sending request");
973 rc = ptlrpc_queue_wait(req);
975 err = ldlm_cli_enqueue_fini(exp, req, einfo->ei_type, policy ? 1 : 0,
976 einfo->ei_mode, flags, lvb, lvb_len,
979 /* If ldlm_cli_enqueue_fini did not find the lock, we need to free
980 * one reference that we took */
982 LDLM_LOCK_RELEASE(lock);
986 if (!req_passed_in && req != NULL) {
987 ptlrpc_req_finished(req);
994 EXPORT_SYMBOL(ldlm_cli_enqueue);
996 static int ldlm_cli_convert_local(struct ldlm_lock *lock, int new_mode,
999 struct ldlm_resource *res;
1002 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1003 CERROR("Trying to cancel local lock\n");
1006 LDLM_DEBUG(lock, "client-side local convert");
1008 res = ldlm_lock_convert(lock, new_mode, flags);
1010 ldlm_reprocess_all(res);
1013 rc = LUSTRE_EDEADLK;
1015 LDLM_DEBUG(lock, "client-side local convert handler END");
1016 LDLM_LOCK_PUT(lock);
1020 /* FIXME: one of ldlm_cli_convert or the server side should reject attempted
1021 * conversion of locks which are on the waiting or converting queue */
1022 /* Caller of this code is supposed to take care of lock readers/writers
1024 int ldlm_cli_convert(struct lustre_handle *lockh, int new_mode, __u32 *flags)
1026 struct ldlm_request *body;
1027 struct ldlm_reply *reply;
1028 struct ldlm_lock *lock;
1029 struct ldlm_resource *res;
1030 struct ptlrpc_request *req;
1034 lock = ldlm_handle2lock(lockh);
1041 if (lock->l_conn_export == NULL)
1042 RETURN(ldlm_cli_convert_local(lock, new_mode, flags));
1044 LDLM_DEBUG(lock, "client-side convert");
1046 req = ptlrpc_request_alloc_pack(class_exp2cliimp(lock->l_conn_export),
1047 &RQF_LDLM_CONVERT, LUSTRE_DLM_VERSION,
1050 LDLM_LOCK_PUT(lock);
1054 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1055 body->lock_handle[0] = lock->l_remote_handle;
1057 body->lock_desc.l_req_mode = new_mode;
1058 body->lock_flags = ldlm_flags_to_wire(*flags);
1061 ptlrpc_request_set_replen(req);
1062 rc = ptlrpc_queue_wait(req);
1066 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1068 GOTO(out, rc = -EPROTO);
1071 GOTO(out, rc = req->rq_status);
1073 res = ldlm_lock_convert(lock, new_mode, &reply->lock_flags);
1075 ldlm_reprocess_all(res);
1076 /* Go to sleep until the lock is granted. */
1077 /* FIXME: or cancelled. */
1078 if (lock->l_completion_ast) {
1079 rc = lock->l_completion_ast(lock, LDLM_FL_WAIT_NOREPROC,
1085 rc = LUSTRE_EDEADLK;
1089 LDLM_LOCK_PUT(lock);
1090 ptlrpc_req_finished(req);
1093 EXPORT_SYMBOL(ldlm_cli_convert);
1096 * Cancel locks locally.
1098 * \retval LDLM_FL_LOCAL_ONLY if there is no need for a CANCEL RPC to the server
1099 * \retval LDLM_FL_CANCELING otherwise;
1100 * \retval LDLM_FL_BL_AST if there is a need for a separate CANCEL RPC.
1102 static __u64 ldlm_cli_cancel_local(struct ldlm_lock *lock)
1104 __u64 rc = LDLM_FL_LOCAL_ONLY;
1107 if (lock->l_conn_export) {
1110 LDLM_DEBUG(lock, "client-side cancel");
1111 /* Set this flag to prevent others from getting new references*/
1112 lock_res_and_lock(lock);
1113 ldlm_set_cbpending(lock);
1114 local_only = !!(lock->l_flags &
1115 (LDLM_FL_LOCAL_ONLY|LDLM_FL_CANCEL_ON_BLOCK));
1116 ldlm_cancel_callback(lock);
1117 rc = (ldlm_is_bl_ast(lock)) ?
1118 LDLM_FL_BL_AST : LDLM_FL_CANCELING;
1119 unlock_res_and_lock(lock);
1122 CDEBUG(D_DLMTRACE, "not sending request (at caller's "
1124 rc = LDLM_FL_LOCAL_ONLY;
1126 ldlm_lock_cancel(lock);
1128 if (ns_is_client(ldlm_lock_to_ns(lock))) {
1129 LDLM_ERROR(lock, "Trying to cancel local lock");
1132 LDLM_DEBUG(lock, "server-side local cancel");
1133 ldlm_lock_cancel(lock);
1134 ldlm_reprocess_all(lock->l_resource);
1141 * Pack \a count locks in \a head into ldlm_request buffer of request \a req.
1143 static void ldlm_cancel_pack(struct ptlrpc_request *req,
1144 struct list_head *head, int count)
1146 struct ldlm_request *dlm;
1147 struct ldlm_lock *lock;
1148 int max, packed = 0;
1151 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1152 LASSERT(dlm != NULL);
1154 /* Check the room in the request buffer. */
1155 max = req_capsule_get_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT) -
1156 sizeof(struct ldlm_request);
1157 max /= sizeof(struct lustre_handle);
1158 max += LDLM_LOCKREQ_HANDLES;
1159 LASSERT(max >= dlm->lock_count + count);
1161 /* XXX: it would be better to pack lock handles grouped by resource.
1162 * so that the server cancel would call filter_lvbo_update() less
1164 list_for_each_entry(lock, head, l_bl_ast) {
1167 LASSERT(lock->l_conn_export);
1168 /* Pack the lock handle to the given request buffer. */
1169 LDLM_DEBUG(lock, "packing");
1170 dlm->lock_handle[dlm->lock_count++] = lock->l_remote_handle;
1173 CDEBUG(D_DLMTRACE, "%d locks packed\n", packed);
1178 * Prepare and send a batched cancel RPC. It will include \a count lock
1179 * handles of locks given in \a cancels list. */
1180 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *cancels,
1181 int count, ldlm_cancel_flags_t flags)
1183 struct ptlrpc_request *req = NULL;
1184 struct obd_import *imp;
1189 LASSERT(exp != NULL);
1192 CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_PAUSE_CANCEL, cfs_fail_val);
1194 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_RACE))
1197 free = ldlm_format_handles_avail(class_exp2cliimp(exp),
1198 &RQF_LDLM_CANCEL, RCL_CLIENT, 0);
1203 imp = class_exp2cliimp(exp);
1204 if (imp == NULL || imp->imp_invalid) {
1206 "skipping cancel on invalid import %p\n", imp);
1210 req = ptlrpc_request_alloc(imp, &RQF_LDLM_CANCEL);
1212 GOTO(out, rc = -ENOMEM);
1214 req_capsule_filled_sizes(&req->rq_pill, RCL_CLIENT);
1215 req_capsule_set_size(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT,
1216 ldlm_request_bufsize(count, LDLM_CANCEL));
1218 rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CANCEL);
1220 ptlrpc_request_free(req);
1224 req->rq_request_portal = LDLM_CANCEL_REQUEST_PORTAL;
1225 req->rq_reply_portal = LDLM_CANCEL_REPLY_PORTAL;
1226 ptlrpc_at_set_req_timeout(req);
1228 ldlm_cancel_pack(req, cancels, count);
1230 ptlrpc_request_set_replen(req);
1231 if (flags & LCF_ASYNC) {
1232 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
1236 rc = ptlrpc_queue_wait(req);
1238 if (rc == LUSTRE_ESTALE) {
1239 CDEBUG(D_DLMTRACE, "client/server (nid %s) "
1240 "out of sync -- not fatal\n",
1241 libcfs_nid2str(req->rq_import->
1242 imp_connection->c_peer.nid));
1244 } else if (rc == -ETIMEDOUT && /* check there was no reconnect*/
1245 req->rq_import_generation == imp->imp_generation) {
1246 ptlrpc_req_finished(req);
1248 } else if (rc != ELDLM_OK) {
1249 /* -ESHUTDOWN is common on umount */
1250 CDEBUG_LIMIT(rc == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1251 "Got rc %d from cancel RPC: "
1252 "canceling anyway\n", rc);
1259 ptlrpc_req_finished(req);
1262 return sent ? sent : rc;
1264 EXPORT_SYMBOL(ldlm_cli_cancel_req);
1266 static inline struct ldlm_pool *ldlm_imp2pl(struct obd_import *imp)
1268 LASSERT(imp != NULL);
1269 return &imp->imp_obd->obd_namespace->ns_pool;
1273 * Update client's OBD pool related fields with new SLV and Limit from \a req.
1275 int ldlm_cli_update_pool(struct ptlrpc_request *req)
1277 struct obd_device *obd;
1281 if (unlikely(!req->rq_import || !req->rq_import->imp_obd ||
1282 !imp_connect_lru_resize(req->rq_import)))
1285 * Do nothing for corner cases.
1290 /* In some cases RPC may contain SLV and limit zeroed out. This
1291 * is the case when server does not support LRU resize feature.
1292 * This is also possible in some recovery cases when server-side
1293 * reqs have no reference to the OBD export and thus access to
1294 * server-side namespace is not possible. */
1295 if (lustre_msg_get_slv(req->rq_repmsg) == 0 ||
1296 lustre_msg_get_limit(req->rq_repmsg) == 0) {
1297 DEBUG_REQ(D_HA, req, "Zero SLV or Limit found "
1298 "(SLV: "LPU64", Limit: %u)",
1299 lustre_msg_get_slv(req->rq_repmsg),
1300 lustre_msg_get_limit(req->rq_repmsg));
1304 new_limit = lustre_msg_get_limit(req->rq_repmsg);
1305 new_slv = lustre_msg_get_slv(req->rq_repmsg);
1306 obd = req->rq_import->imp_obd;
1308 /* Set new SLV and limit in OBD fields to make them accessible
1309 * to the pool thread. We do not access obd_namespace and pool
1310 * directly here as there is no reliable way to make sure that
1311 * they are still alive at cleanup time. Evil races are possible
1312 * which may cause Oops at that time. */
1313 write_lock(&obd->obd_pool_lock);
1314 obd->obd_pool_slv = new_slv;
1315 obd->obd_pool_limit = new_limit;
1316 write_unlock(&obd->obd_pool_lock);
1320 EXPORT_SYMBOL(ldlm_cli_update_pool);
1323 * Client side lock cancel.
1325 * Lock must not have any readers or writers by this time.
1327 int ldlm_cli_cancel(struct lustre_handle *lockh,
1328 ldlm_cancel_flags_t cancel_flags)
1330 struct obd_export *exp;
1331 int avail, flags, count = 1;
1333 struct ldlm_namespace *ns;
1334 struct ldlm_lock *lock;
1335 struct list_head cancels = LIST_HEAD_INIT(cancels);
1338 /* concurrent cancels on the same handle can happen */
1339 lock = ldlm_handle2lock_long(lockh, LDLM_FL_CANCELING);
1341 LDLM_DEBUG_NOLOCK("lock is already being destroyed");
1345 rc = ldlm_cli_cancel_local(lock);
1346 if (rc == LDLM_FL_LOCAL_ONLY || cancel_flags & LCF_LOCAL) {
1347 LDLM_LOCK_RELEASE(lock);
1350 /* Even if the lock is marked as LDLM_FL_BL_AST, this is a LDLM_CANCEL
1351 * RPC which goes to canceld portal, so we can cancel other LRU locks
1352 * here and send them all as one LDLM_CANCEL RPC. */
1353 LASSERT(list_empty(&lock->l_bl_ast));
1354 list_add(&lock->l_bl_ast, &cancels);
1356 exp = lock->l_conn_export;
1357 if (exp_connect_cancelset(exp)) {
1358 avail = ldlm_format_handles_avail(class_exp2cliimp(exp),
1363 ns = ldlm_lock_to_ns(lock);
1364 flags = ns_connect_lru_resize(ns) ?
1365 LDLM_CANCEL_LRUR : LDLM_CANCEL_AGED;
1366 count += ldlm_cancel_lru_local(ns, &cancels, 0, avail - 1,
1369 ldlm_cli_cancel_list(&cancels, count, NULL, cancel_flags);
1372 EXPORT_SYMBOL(ldlm_cli_cancel);
1375 * Locally cancel up to \a count locks in list \a cancels.
1376 * Return the number of cancelled locks.
1378 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1379 ldlm_cancel_flags_t flags)
1381 struct list_head head = LIST_HEAD_INIT(head);
1382 struct ldlm_lock *lock, *next;
1383 int left = 0, bl_ast = 0;
1387 list_for_each_entry_safe(lock, next, cancels, l_bl_ast) {
1391 if (flags & LCF_LOCAL) {
1392 rc = LDLM_FL_LOCAL_ONLY;
1393 ldlm_lock_cancel(lock);
1395 rc = ldlm_cli_cancel_local(lock);
1397 /* Until we have compound requests and can send LDLM_CANCEL
1398 * requests batched with generic RPCs, we need to send cancels
1399 * with the LDLM_FL_BL_AST flag in a separate RPC from
1400 * the one being generated now. */
1401 if (!(flags & LCF_BL_AST) && (rc == LDLM_FL_BL_AST)) {
1402 LDLM_DEBUG(lock, "Cancel lock separately");
1403 list_del_init(&lock->l_bl_ast);
1404 list_add(&lock->l_bl_ast, &head);
1408 if (rc == LDLM_FL_LOCAL_ONLY) {
1409 /* CANCEL RPC should not be sent to server. */
1410 list_del_init(&lock->l_bl_ast);
1411 LDLM_LOCK_RELEASE(lock);
1417 ldlm_cli_cancel_list(&head, bl_ast, NULL, 0);
1422 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
1425 * Cancel as many locks as possible w/o sending any RPCs (e.g. to write back
1426 * dirty data, to close a file, ...) or waiting for any RPCs in-flight (e.g.
1427 * readahead requests, ...)
1429 static ldlm_policy_res_t ldlm_cancel_no_wait_policy(struct ldlm_namespace *ns,
1430 struct ldlm_lock *lock,
1431 int unused, int added,
1434 ldlm_policy_res_t result = LDLM_POLICY_CANCEL_LOCK;
1436 /* don't check added & count since we want to process all locks
1438 * It's fine to not take lock to access lock->l_resource since
1439 * the lock has already been granted so it won't change. */
1440 switch (lock->l_resource->lr_type) {
1443 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) != 0)
1446 result = LDLM_POLICY_SKIP_LOCK;
1447 lock_res_and_lock(lock);
1448 ldlm_set_skipped(lock);
1449 unlock_res_and_lock(lock);
1457 * Callback function for LRU-resize policy. Decides whether to keep
1458 * \a lock in LRU for current \a LRU size \a unused, added in current
1459 * scan \a added and number of locks to be preferably canceled \a count.
1461 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1463 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1465 static ldlm_policy_res_t ldlm_cancel_lrur_policy(struct ldlm_namespace *ns,
1466 struct ldlm_lock *lock,
1467 int unused, int added,
1470 cfs_time_t cur = cfs_time_current();
1471 struct ldlm_pool *pl = &ns->ns_pool;
1475 /* Stop LRU processing when we reach past @count or have checked all
1477 if (count && added >= count)
1478 return LDLM_POLICY_KEEP_LOCK;
1480 slv = ldlm_pool_get_slv(pl);
1481 lvf = ldlm_pool_get_lvf(pl);
1482 la = cfs_duration_sec(cfs_time_sub(cur,
1483 lock->l_last_used));
1484 lv = lvf * la * unused;
1486 /* Inform pool about current CLV to see it via proc. */
1487 ldlm_pool_set_clv(pl, lv);
1489 /* Stop when SLV is not yet come from server or lv is smaller than
1491 if (slv == 0 || lv < slv)
1492 return LDLM_POLICY_KEEP_LOCK;
1494 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) == 0)
1495 return LDLM_POLICY_KEEP_LOCK;
1497 return LDLM_POLICY_CANCEL_LOCK;
1501 * Callback function for proc used policy. Makes decision whether to keep
1502 * \a lock in LRU for current \a LRU size \a unused, added in current scan \a
1503 * added and number of locks to be preferably canceled \a count.
1505 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1507 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1509 static ldlm_policy_res_t ldlm_cancel_passed_policy(struct ldlm_namespace *ns,
1510 struct ldlm_lock *lock,
1511 int unused, int added,
1514 /* Stop LRU processing when we reach past @count or have checked all
1516 return (added >= count) ?
1517 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1521 * Callback function for aged policy. Makes decision whether to keep \a lock in
1522 * LRU for current LRU size \a unused, added in current scan \a added and
1523 * number of locks to be preferably canceled \a count.
1525 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1527 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1529 static ldlm_policy_res_t ldlm_cancel_aged_policy(struct ldlm_namespace *ns,
1530 struct ldlm_lock *lock,
1531 int unused, int added,
1535 return LDLM_POLICY_KEEP_LOCK;
1537 if (cfs_time_before(cfs_time_current(),
1538 cfs_time_add(lock->l_last_used, ns->ns_max_age)))
1539 return LDLM_POLICY_KEEP_LOCK;
1541 if (ns->ns_cancel != NULL && ns->ns_cancel(lock) == 0)
1542 return LDLM_POLICY_KEEP_LOCK;
1544 return LDLM_POLICY_CANCEL_LOCK;
1548 * Callback function for default policy. Makes decision whether to keep \a lock
1549 * in LRU for current LRU size \a unused, added in current scan \a added and
1550 * number of locks to be preferably canceled \a count.
1552 * \retval LDLM_POLICY_KEEP_LOCK keep lock in LRU in stop scanning
1554 * \retval LDLM_POLICY_CANCEL_LOCK cancel lock from LRU
1556 static ldlm_policy_res_t ldlm_cancel_default_policy(struct ldlm_namespace *ns,
1557 struct ldlm_lock *lock,
1558 int unused, int added,
1561 /* Stop LRU processing when we reach past count or have checked all
1563 return (added >= count) ?
1564 LDLM_POLICY_KEEP_LOCK : LDLM_POLICY_CANCEL_LOCK;
1567 typedef ldlm_policy_res_t (*ldlm_cancel_lru_policy_t)(struct ldlm_namespace *,
1568 struct ldlm_lock *, int,
1571 static ldlm_cancel_lru_policy_t
1572 ldlm_cancel_lru_policy(struct ldlm_namespace *ns, int flags)
1574 if (flags & LDLM_CANCEL_NO_WAIT)
1575 return ldlm_cancel_no_wait_policy;
1577 if (ns_connect_lru_resize(ns)) {
1578 if (flags & LDLM_CANCEL_SHRINK)
1579 /* We kill passed number of old locks. */
1580 return ldlm_cancel_passed_policy;
1581 else if (flags & LDLM_CANCEL_LRUR)
1582 return ldlm_cancel_lrur_policy;
1583 else if (flags & LDLM_CANCEL_PASSED)
1584 return ldlm_cancel_passed_policy;
1586 if (flags & LDLM_CANCEL_AGED)
1587 return ldlm_cancel_aged_policy;
1590 return ldlm_cancel_default_policy;
1594 * - Free space in LRU for \a count new locks,
1595 * redundant unused locks are canceled locally;
1596 * - also cancel locally unused aged locks;
1597 * - do not cancel more than \a max locks;
1598 * - GET the found locks and add them into the \a cancels list.
1600 * A client lock can be added to the l_bl_ast list only when it is
1601 * marked LDLM_FL_CANCELING. Otherwise, somebody is already doing
1602 * CANCEL. There are the following use cases:
1603 * ldlm_cancel_resource_local(), ldlm_cancel_lru_local() and
1604 * ldlm_cli_cancel(), which check and set this flag properly. As any
1605 * attempt to cancel a lock rely on this flag, l_bl_ast list is accessed
1606 * later without any special locking.
1608 * Calling policies for enabled LRU resize:
1609 * ----------------------------------------
1610 * flags & LDLM_CANCEL_LRUR - use LRU resize policy (SLV from server) to
1611 * cancel not more than \a count locks;
1613 * flags & LDLM_CANCEL_PASSED - cancel \a count number of old locks (located at
1614 * the beginning of LRU list);
1616 * flags & LDLM_CANCEL_SHRINK - cancel not more than \a count locks according to
1617 * memory pressre policy function;
1619 * flags & LDLM_CANCEL_AGED - cancel \a count locks according to "aged policy".
1621 * flags & LDLM_CANCEL_NO_WAIT - cancel as many unused locks as possible
1622 * (typically before replaying locks) w/o
1623 * sending any RPCs or waiting for any
1624 * outstanding RPC to complete.
1626 static int ldlm_prepare_lru_list(struct ldlm_namespace *ns,
1627 struct list_head *cancels, int count, int max,
1630 ldlm_cancel_lru_policy_t pf;
1631 struct ldlm_lock *lock, *next;
1632 int added = 0, unused, remained;
1635 spin_lock(&ns->ns_lock);
1636 unused = ns->ns_nr_unused;
1639 if (!ns_connect_lru_resize(ns))
1640 count += unused - ns->ns_max_unused;
1642 pf = ldlm_cancel_lru_policy(ns, flags);
1643 LASSERT(pf != NULL);
1645 while (!list_empty(&ns->ns_unused_list)) {
1646 ldlm_policy_res_t result;
1648 /* all unused locks */
1649 if (remained-- <= 0)
1652 /* For any flags, stop scanning if @max is reached. */
1653 if (max && added >= max)
1656 list_for_each_entry_safe(lock, next, &ns->ns_unused_list,
1658 /* No locks which got blocking requests. */
1659 LASSERT(!ldlm_is_bl_ast(lock));
1661 if (flags & LDLM_CANCEL_NO_WAIT &&
1662 ldlm_is_skipped(lock))
1663 /* already processed */
1666 /* Somebody is already doing CANCEL. No need for this
1667 * lock in LRU, do not traverse it again. */
1668 if (!ldlm_is_canceling(lock))
1671 ldlm_lock_remove_from_lru_nolock(lock);
1673 if (&lock->l_lru == &ns->ns_unused_list)
1676 LDLM_LOCK_GET(lock);
1677 spin_unlock(&ns->ns_lock);
1678 lu_ref_add(&lock->l_reference, __FUNCTION__, current);
1680 /* Pass the lock through the policy filter and see if it
1681 * should stay in LRU.
1683 * Even for shrinker policy we stop scanning if
1684 * we find a lock that should stay in the cache.
1685 * We should take into account lock age anyway
1686 * as a new lock is a valuable resource even if
1687 * it has a low weight.
1689 * That is, for shrinker policy we drop only
1690 * old locks, but additionally choose them by
1691 * their weight. Big extent locks will stay in
1693 result = pf(ns, lock, unused, added, count);
1694 if (result == LDLM_POLICY_KEEP_LOCK) {
1695 lu_ref_del(&lock->l_reference,
1696 __FUNCTION__, current);
1697 LDLM_LOCK_RELEASE(lock);
1698 spin_lock(&ns->ns_lock);
1701 if (result == LDLM_POLICY_SKIP_LOCK) {
1702 lu_ref_del(&lock->l_reference,
1704 LDLM_LOCK_RELEASE(lock);
1705 spin_lock(&ns->ns_lock);
1709 lock_res_and_lock(lock);
1710 /* Check flags again under the lock. */
1711 if (ldlm_is_canceling(lock) ||
1712 (ldlm_lock_remove_from_lru(lock) == 0)) {
1713 /* Another thread is removing lock from LRU, or
1714 * somebody is already doing CANCEL, or there
1715 * is a blocking request which will send cancel
1716 * by itself, or the lock is no longer unused. */
1717 unlock_res_and_lock(lock);
1718 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1719 LDLM_LOCK_RELEASE(lock);
1720 spin_lock(&ns->ns_lock);
1723 LASSERT(!lock->l_readers && !lock->l_writers);
1725 /* If we have chosen to cancel this lock voluntarily, we
1726 * better send cancel notification to server, so that it
1727 * frees appropriate state. This might lead to a race
1728 * where while we are doing cancel here, server is also
1729 * silently cancelling this lock. */
1730 ldlm_clear_cancel_on_block(lock);
1732 /* Setting the CBPENDING flag is a little misleading,
1733 * but prevents an important race; namely, once
1734 * CBPENDING is set, the lock can accumulate no more
1735 * readers/writers. Since readers and writers are
1736 * already zero here, ldlm_lock_decref() won't see
1737 * this flag and call l_blocking_ast */
1738 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING;
1740 /* We can't re-add to l_lru as it confuses the
1741 * refcounting in ldlm_lock_remove_from_lru() if an AST
1742 * arrives after we drop lr_lock below. We use l_bl_ast
1743 * and can't use l_pending_chain as it is used both on
1744 * server and client nevertheless bug 5666 says it is
1745 * used only on server */
1746 LASSERT(list_empty(&lock->l_bl_ast));
1747 list_add(&lock->l_bl_ast, cancels);
1748 unlock_res_and_lock(lock);
1749 lu_ref_del(&lock->l_reference, __FUNCTION__, current);
1750 spin_lock(&ns->ns_lock);
1754 spin_unlock(&ns->ns_lock);
1758 int ldlm_cancel_lru_local(struct ldlm_namespace *ns, struct list_head *cancels,
1759 int count, int max, ldlm_cancel_flags_t cancel_flags,
1763 added = ldlm_prepare_lru_list(ns, cancels, count, max, flags);
1766 return ldlm_cli_cancel_list_local(cancels, added, cancel_flags);
1770 * Cancel at least \a nr locks from given namespace LRU.
1772 * When called with LCF_ASYNC the blocking callback will be handled
1773 * in a thread and this function will return after the thread has been
1774 * asked to call the callback. When called with LCF_ASYNC the blocking
1775 * callback will be performed in this function.
1777 int ldlm_cancel_lru(struct ldlm_namespace *ns, int nr,
1778 ldlm_cancel_flags_t cancel_flags,
1781 struct list_head cancels = LIST_HEAD_INIT(cancels);
1785 /* Just prepare the list of locks, do not actually cancel them yet.
1786 * Locks are cancelled later in a separate thread. */
1787 count = ldlm_prepare_lru_list(ns, &cancels, nr, 0, flags);
1788 rc = ldlm_bl_to_thread_list(ns, NULL, &cancels, count, cancel_flags);
1796 * Find and cancel locally unused locks found on resource, matched to the
1797 * given policy, mode. GET the found locks and add them into the \a cancels
1800 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1801 struct list_head *cancels,
1802 ldlm_policy_data_t *policy,
1803 ldlm_mode_t mode, __u64 lock_flags,
1804 ldlm_cancel_flags_t cancel_flags, void *opaque)
1806 struct ldlm_lock *lock;
1811 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
1812 if (opaque != NULL && lock->l_ast_data != opaque) {
1813 LDLM_ERROR(lock, "data %p doesn't match opaque %p",
1814 lock->l_ast_data, opaque);
1819 if (lock->l_readers || lock->l_writers)
1822 /* If somebody is already doing CANCEL, or blocking AST came,
1823 * skip this lock. */
1824 if (ldlm_is_bl_ast(lock) || ldlm_is_canceling(lock))
1827 if (lockmode_compat(lock->l_granted_mode, mode))
1830 /* If policy is given and this is IBITS lock, add to list only
1831 * those locks that match by policy. */
1832 if (policy && (lock->l_resource->lr_type == LDLM_IBITS) &&
1833 !(lock->l_policy_data.l_inodebits.bits &
1834 policy->l_inodebits.bits))
1837 /* See CBPENDING comment in ldlm_cancel_lru */
1838 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_CANCELING |
1841 LASSERT(list_empty(&lock->l_bl_ast));
1842 list_add(&lock->l_bl_ast, cancels);
1843 LDLM_LOCK_GET(lock);
1848 RETURN(ldlm_cli_cancel_list_local(cancels, count, cancel_flags));
1850 EXPORT_SYMBOL(ldlm_cancel_resource_local);
1853 * Cancel client-side locks from a list and send/prepare cancel RPCs to the
1855 * If \a req is NULL, send CANCEL request to server with handles of locks
1856 * in the \a cancels. If EARLY_CANCEL is not supported, send CANCEL requests
1857 * separately per lock.
1858 * If \a req is not NULL, put handles of locks in \a cancels into the request
1859 * buffer at the offset \a off.
1860 * Destroy \a cancels at the end.
1862 int ldlm_cli_cancel_list(struct list_head *cancels, int count,
1863 struct ptlrpc_request *req, ldlm_cancel_flags_t flags)
1865 struct ldlm_lock *lock;
1869 if (list_empty(cancels) || count == 0)
1872 /* XXX: requests (both batched and not) could be sent in parallel.
1873 * Usually it is enough to have just 1 RPC, but it is possible that
1874 * there are too many locks to be cancelled in LRU or on a resource.
1875 * It would also speed up the case when the server does not support
1878 LASSERT(!list_empty(cancels));
1879 lock = list_entry(cancels->next, struct ldlm_lock,
1881 LASSERT(lock->l_conn_export);
1883 if (exp_connect_cancelset(lock->l_conn_export)) {
1886 ldlm_cancel_pack(req, cancels, count);
1888 res = ldlm_cli_cancel_req(lock->l_conn_export,
1892 res = ldlm_cli_cancel_req(lock->l_conn_export,
1897 CDEBUG_LIMIT(res == -ESHUTDOWN ? D_DLMTRACE : D_ERROR,
1898 "ldlm_cli_cancel_list: %d\n", res);
1903 ldlm_lock_list_put(cancels, l_bl_ast, res);
1905 LASSERT(count == 0);
1908 EXPORT_SYMBOL(ldlm_cli_cancel_list);
1911 * Cancel all locks on a resource that have 0 readers/writers.
1913 * If flags & LDLM_FL_LOCAL_ONLY, throw the locks away without trying
1914 * to notify the server. */
1915 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1916 const struct ldlm_res_id *res_id,
1917 ldlm_policy_data_t *policy,
1919 ldlm_cancel_flags_t flags,
1922 struct ldlm_resource *res;
1923 struct list_head cancels = LIST_HEAD_INIT(cancels);
1928 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
1930 /* This is not a problem. */
1931 CDEBUG(D_INFO, "No resource "LPU64"\n", res_id->name[0]);
1935 LDLM_RESOURCE_ADDREF(res);
1936 count = ldlm_cancel_resource_local(res, &cancels, policy, mode,
1937 0, flags | LCF_BL_AST, opaque);
1938 rc = ldlm_cli_cancel_list(&cancels, count, NULL, flags);
1940 CERROR("canceling unused lock "DLDLMRES": rc = %d\n",
1943 LDLM_RESOURCE_DELREF(res);
1944 ldlm_resource_putref(res);
1947 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
1949 struct ldlm_cli_cancel_arg {
1954 static int ldlm_cli_hash_cancel_unused(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1955 struct hlist_node *hnode, void *arg)
1957 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1958 struct ldlm_cli_cancel_arg *lc = arg;
1960 ldlm_cli_cancel_unused_resource(ldlm_res_to_ns(res), &res->lr_name,
1961 NULL, LCK_MINMODE, lc->lc_flags,
1963 /* must return 0 for hash iteration */
1968 * Cancel all locks on a namespace (or a specific resource, if given)
1969 * that have 0 readers/writers.
1971 * If flags & LCF_LOCAL, throw the locks away without trying
1972 * to notify the server. */
1973 int ldlm_cli_cancel_unused(struct ldlm_namespace *ns,
1974 const struct ldlm_res_id *res_id,
1975 ldlm_cancel_flags_t flags, void *opaque)
1977 struct ldlm_cli_cancel_arg arg = {
1979 .lc_opaque = opaque,
1987 if (res_id != NULL) {
1988 RETURN(ldlm_cli_cancel_unused_resource(ns, res_id, NULL,
1992 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1993 ldlm_cli_hash_cancel_unused, &arg);
1997 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1999 /* Lock iterators. */
2001 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
2004 struct list_head *tmp, *next;
2005 struct ldlm_lock *lock;
2006 int rc = LDLM_ITER_CONTINUE;
2011 RETURN(LDLM_ITER_CONTINUE);
2014 list_for_each_safe(tmp, next, &res->lr_granted) {
2015 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2017 if (iter(lock, closure) == LDLM_ITER_STOP)
2018 GOTO(out, rc = LDLM_ITER_STOP);
2021 list_for_each_safe(tmp, next, &res->lr_converting) {
2022 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2024 if (iter(lock, closure) == LDLM_ITER_STOP)
2025 GOTO(out, rc = LDLM_ITER_STOP);
2028 list_for_each_safe(tmp, next, &res->lr_waiting) {
2029 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
2031 if (iter(lock, closure) == LDLM_ITER_STOP)
2032 GOTO(out, rc = LDLM_ITER_STOP);
2038 EXPORT_SYMBOL(ldlm_resource_foreach);
2040 struct iter_helper_data {
2041 ldlm_iterator_t iter;
2045 static int ldlm_iter_helper(struct ldlm_lock *lock, void *closure)
2047 struct iter_helper_data *helper = closure;
2048 return helper->iter(lock, helper->closure);
2051 static int ldlm_res_iter_helper(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2052 struct hlist_node *hnode, void *arg)
2055 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2057 return ldlm_resource_foreach(res, ldlm_iter_helper, arg) ==
2061 void ldlm_namespace_foreach(struct ldlm_namespace *ns,
2062 ldlm_iterator_t iter, void *closure)
2065 struct iter_helper_data helper = { .iter = iter, .closure = closure };
2067 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2068 ldlm_res_iter_helper, &helper);
2071 EXPORT_SYMBOL(ldlm_namespace_foreach);
2073 /* non-blocking function to manipulate a lock whose cb_data is being put away.
2074 * return 0: find no resource
2075 * > 0: must be LDLM_ITER_STOP/LDLM_ITER_CONTINUE.
2078 int ldlm_resource_iterate(struct ldlm_namespace *ns,
2079 const struct ldlm_res_id *res_id,
2080 ldlm_iterator_t iter, void *data)
2082 struct ldlm_resource *res;
2086 LASSERTF(ns != NULL, "must pass in namespace\n");
2088 res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
2092 LDLM_RESOURCE_ADDREF(res);
2093 rc = ldlm_resource_foreach(res, iter, data);
2094 LDLM_RESOURCE_DELREF(res);
2095 ldlm_resource_putref(res);
2098 EXPORT_SYMBOL(ldlm_resource_iterate);
2102 static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
2104 struct list_head *list = closure;
2106 /* we use l_pending_chain here, because it's unused on clients. */
2107 LASSERTF(list_empty(&lock->l_pending_chain),
2108 "lock %p next %p prev %p\n",
2109 lock, &lock->l_pending_chain.next,&lock->l_pending_chain.prev);
2110 /* bug 9573: don't replay locks left after eviction, or
2111 * bug 17614: locks being actively cancelled. Get a reference
2112 * on a lock so that it does not disapear under us (e.g. due to cancel)
2114 if (!(lock->l_flags & (LDLM_FL_FAILED|LDLM_FL_CANCELING))) {
2115 list_add(&lock->l_pending_chain, list);
2116 LDLM_LOCK_GET(lock);
2119 return LDLM_ITER_CONTINUE;
2122 static int replay_lock_interpret(const struct lu_env *env,
2123 struct ptlrpc_request *req,
2124 struct ldlm_async_args *aa, int rc)
2126 struct ldlm_lock *lock;
2127 struct ldlm_reply *reply;
2128 struct obd_export *exp;
2131 atomic_dec(&req->rq_import->imp_replay_inflight);
2135 reply = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2137 GOTO(out, rc = -EPROTO);
2139 lock = ldlm_handle2lock(&aa->lock_handle);
2141 CERROR("received replay ack for unknown local cookie "LPX64
2142 " remote cookie "LPX64 " from server %s id %s\n",
2143 aa->lock_handle.cookie, reply->lock_handle.cookie,
2144 req->rq_export->exp_client_uuid.uuid,
2145 libcfs_id2str(req->rq_peer));
2146 GOTO(out, rc = -ESTALE);
2149 /* Key change rehash lock in per-export hash with new key */
2150 exp = req->rq_export;
2151 if (exp && exp->exp_lock_hash) {
2152 /* In the function below, .hs_keycmp resolves to
2153 * ldlm_export_lock_keycmp() */
2154 /* coverity[overrun-buffer-val] */
2155 cfs_hash_rehash_key(exp->exp_lock_hash,
2156 &lock->l_remote_handle,
2157 &reply->lock_handle,
2160 lock->l_remote_handle = reply->lock_handle;
2163 LDLM_DEBUG(lock, "replayed lock:");
2164 ptlrpc_import_recovery_state_machine(req->rq_import);
2165 LDLM_LOCK_PUT(lock);
2168 ptlrpc_connect_import(req->rq_import);
2173 static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
2175 struct ptlrpc_request *req;
2176 struct ldlm_async_args *aa;
2177 struct ldlm_request *body;
2182 /* Bug 11974: Do not replay a lock which is actively being canceled */
2183 if (ldlm_is_canceling(lock)) {
2184 LDLM_DEBUG(lock, "Not replaying canceled lock:");
2188 /* If this is reply-less callback lock, we cannot replay it, since
2189 * server might have long dropped it, but notification of that event was
2190 * lost by network. (and server granted conflicting lock already) */
2191 if (ldlm_is_cancel_on_block(lock)) {
2192 LDLM_DEBUG(lock, "Not replaying reply-less lock:");
2193 ldlm_lock_cancel(lock);
2198 * If granted mode matches the requested mode, this lock is granted.
2200 * If they differ, but we have a granted mode, then we were granted
2201 * one mode and now want another: ergo, converting.
2203 * If we haven't been granted anything and are on a resource list,
2204 * then we're blocked/waiting.
2206 * If we haven't been granted anything and we're NOT on a resource list,
2207 * then we haven't got a reply yet and don't have a known disposition.
2208 * This happens whenever a lock enqueue is the request that triggers
2211 if (lock->l_granted_mode == lock->l_req_mode)
2212 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_GRANTED;
2213 else if (lock->l_granted_mode)
2214 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_CONV;
2215 else if (!list_empty(&lock->l_res_link))
2216 flags = LDLM_FL_REPLAY | LDLM_FL_BLOCK_WAIT;
2218 flags = LDLM_FL_REPLAY;
2220 req = ptlrpc_request_alloc_pack(imp, &RQF_LDLM_ENQUEUE,
2221 LUSTRE_DLM_VERSION, LDLM_ENQUEUE);
2225 /* We're part of recovery, so don't wait for it. */
2226 req->rq_send_state = LUSTRE_IMP_REPLAY_LOCKS;
2228 body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2229 ldlm_lock2desc(lock, &body->lock_desc);
2230 body->lock_flags = ldlm_flags_to_wire(flags);
2232 ldlm_lock2handle(lock, &body->lock_handle[0]);
2233 if (lock->l_lvb_len > 0)
2234 req_capsule_extend(&req->rq_pill, &RQF_LDLM_ENQUEUE_LVB);
2235 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2237 ptlrpc_request_set_replen(req);
2238 /* notify the server we've replayed all requests.
2239 * also, we mark the request to be put on a dedicated
2240 * queue to be processed after all request replayes.
2242 lustre_msg_set_flags(req->rq_reqmsg, MSG_REQ_REPLAY_DONE);
2244 LDLM_DEBUG(lock, "replaying lock:");
2246 atomic_inc(&req->rq_import->imp_replay_inflight);
2247 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2248 aa = ptlrpc_req_async_args(req);
2249 aa->lock_handle = body->lock_handle[0];
2250 req->rq_interpret_reply = (ptlrpc_interpterer_t)replay_lock_interpret;
2251 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2257 * Cancel as many unused locks as possible before replay. since we are
2258 * in recovery, we can't wait for any outstanding RPCs to send any RPC
2261 * Called only in recovery before replaying locks. there is no need to
2262 * replay locks that are unused. since the clients may hold thousands of
2263 * cached unused locks, dropping the unused locks can greatly reduce the
2264 * load on the servers at recovery time.
2266 static void ldlm_cancel_unused_locks_for_replay(struct ldlm_namespace *ns)
2269 struct list_head cancels = LIST_HEAD_INIT(cancels);
2271 CDEBUG(D_DLMTRACE, "Dropping as many unused locks as possible before"
2272 "replay for namespace %s (%d)\n",
2273 ldlm_ns_name(ns), ns->ns_nr_unused);
2275 /* We don't need to care whether or not LRU resize is enabled
2276 * because the LDLM_CANCEL_NO_WAIT policy doesn't use the
2277 * count parameter */
2278 canceled = ldlm_cancel_lru_local(ns, &cancels, ns->ns_nr_unused, 0,
2279 LCF_LOCAL, LDLM_CANCEL_NO_WAIT);
2281 CDEBUG(D_DLMTRACE, "Canceled %d unused locks from namespace %s\n",
2282 canceled, ldlm_ns_name(ns));
2285 int ldlm_replay_locks(struct obd_import *imp)
2287 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
2288 struct list_head list = LIST_HEAD_INIT(list);
2289 struct ldlm_lock *lock, *next;
2294 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
2296 /* don't replay locks if import failed recovery */
2297 if (imp->imp_vbr_failed)
2300 /* ensure this doesn't fall to 0 before all have been queued */
2301 atomic_inc(&imp->imp_replay_inflight);
2303 if (ldlm_cancel_unused_locks_before_replay)
2304 ldlm_cancel_unused_locks_for_replay(ns);
2306 ldlm_namespace_foreach(ns, ldlm_chain_lock_for_replay, &list);
2308 list_for_each_entry_safe(lock, next, &list, l_pending_chain) {
2309 list_del_init(&lock->l_pending_chain);
2311 LDLM_LOCK_RELEASE(lock);
2312 continue; /* or try to do the rest? */
2314 rc = replay_one_lock(imp, lock);
2315 LDLM_LOCK_RELEASE(lock);
2318 atomic_dec(&imp->imp_replay_inflight);
2322 EXPORT_SYMBOL(ldlm_replay_locks);