1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * underlying calls for lock creation
8 * Copyright (C) 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
28 #include <linux/module.h>
30 #include <linux/types.h>
31 #include <linux/slab.h>
32 #include <linux/highmem.h>
33 #include <linux/init.h>
34 #include <linux/sysctl.h>
35 #include <linux/random.h>
36 #include <linux/blkdev.h>
37 #include <linux/socket.h>
38 #include <linux/inet.h>
39 #include <linux/spinlock.h>
40 #include <linux/delay.h>
43 #include "cluster/heartbeat.h"
44 #include "cluster/nodemanager.h"
45 #include "cluster/tcp.h"
48 #include "dlmcommon.h"
50 #include "dlmconvert.h"
52 #define MLOG_MASK_PREFIX ML_DLM
53 #include "cluster/masklog.h"
55 static struct kmem_cache *dlm_lock_cache = NULL;
57 static DEFINE_SPINLOCK(dlm_cookie_lock);
58 static u64 dlm_next_cookie = 1;
60 static enum dlm_status dlm_send_remote_lock_request(struct dlm_ctxt *dlm,
61 struct dlm_lock_resource *res,
62 struct dlm_lock *lock, int flags);
63 static void dlm_init_lock(struct dlm_lock *newlock, int type,
65 static void dlm_lock_release(struct kref *kref);
66 static void dlm_lock_detach_lockres(struct dlm_lock *lock);
68 int dlm_init_lock_cache(void)
70 dlm_lock_cache = kmem_cache_create("o2dlm_lock",
71 sizeof(struct dlm_lock),
72 0, SLAB_HWCACHE_ALIGN, NULL);
73 if (dlm_lock_cache == NULL)
78 void dlm_destroy_lock_cache(void)
81 kmem_cache_destroy(dlm_lock_cache);
84 /* Tell us whether we can grant a new lock request.
86 * caller needs: res->spinlock
89 * returns: 1 if the lock can be granted, 0 otherwise.
91 static int dlm_can_grant_new_lock(struct dlm_lock_resource *res,
92 struct dlm_lock *lock)
94 struct list_head *iter;
95 struct dlm_lock *tmplock;
97 list_for_each(iter, &res->granted) {
98 tmplock = list_entry(iter, struct dlm_lock, list);
100 if (!dlm_lock_compatible(tmplock->ml.type, lock->ml.type))
104 list_for_each(iter, &res->converting) {
105 tmplock = list_entry(iter, struct dlm_lock, list);
107 if (!dlm_lock_compatible(tmplock->ml.type, lock->ml.type))
109 if (!dlm_lock_compatible(tmplock->ml.convert_type,
117 /* performs lock creation at the lockres master site
120 * taken: takes and drops res->spinlock
122 * returns: DLM_NORMAL, DLM_NOTQUEUED
124 static enum dlm_status dlmlock_master(struct dlm_ctxt *dlm,
125 struct dlm_lock_resource *res,
126 struct dlm_lock *lock, int flags)
128 int call_ast = 0, kick_thread = 0;
129 enum dlm_status status = DLM_NORMAL;
131 mlog(0, "type=%d\n", lock->ml.type);
133 spin_lock(&res->spinlock);
134 /* if called from dlm_create_lock_handler, need to
135 * ensure it will not sleep in dlm_wait_on_lockres */
136 status = __dlm_lockres_state_to_status(res);
137 if (status != DLM_NORMAL &&
138 lock->ml.node != dlm->node_num) {
139 /* erf. state changed after lock was dropped. */
140 spin_unlock(&res->spinlock);
144 __dlm_wait_on_lockres(res);
145 __dlm_lockres_reserve_ast(res);
147 if (dlm_can_grant_new_lock(res, lock)) {
148 mlog(0, "I can grant this lock right away\n");
149 /* got it right away */
150 lock->lksb->status = DLM_NORMAL;
153 list_add_tail(&lock->list, &res->granted);
155 /* for the recovery lock, we can't allow the ast
156 * to be queued since the dlmthread is already
157 * frozen. but the recovery lock is always locked
158 * with LKM_NOQUEUE so we do not need the ast in
159 * this special case */
160 if (!dlm_is_recovery_lock(res->lockname.name,
161 res->lockname.len)) {
165 mlog(0, "%s: returning DLM_NORMAL to "
166 "node %u for reco lock\n", dlm->name,
170 /* for NOQUEUE request, unless we get the
171 * lock right away, return DLM_NOTQUEUED */
172 if (flags & LKM_NOQUEUE) {
173 status = DLM_NOTQUEUED;
174 if (dlm_is_recovery_lock(res->lockname.name,
175 res->lockname.len)) {
176 mlog(0, "%s: returning NOTQUEUED to "
177 "node %u for reco lock\n", dlm->name,
182 list_add_tail(&lock->list, &res->blocked);
187 spin_unlock(&res->spinlock);
190 /* either queue the ast or release it */
192 dlm_queue_ast(dlm, lock);
194 dlm_lockres_release_ast(dlm, res);
196 dlm_lockres_calc_usage(dlm, res);
198 dlm_kick_thread(dlm, res);
203 void dlm_revert_pending_lock(struct dlm_lock_resource *res,
204 struct dlm_lock *lock)
206 /* remove from local queue if it failed */
207 list_del_init(&lock->list);
208 lock->lksb->flags &= ~DLM_LKSB_GET_LVB;
215 * taken: takes and drops res->spinlock
217 * returns: DLM_DENIED, DLM_RECOVERING, or net status
219 static enum dlm_status dlmlock_remote(struct dlm_ctxt *dlm,
220 struct dlm_lock_resource *res,
221 struct dlm_lock *lock, int flags)
223 enum dlm_status status = DLM_DENIED;
224 int lockres_changed = 1;
226 mlog(0, "type=%d, lockres %.*s, flags = 0x%x\n",
227 lock->ml.type, res->lockname.len,
228 res->lockname.name, flags);
230 spin_lock(&res->spinlock);
232 /* will exit this call with spinlock held */
233 __dlm_wait_on_lockres(res);
234 res->state |= DLM_LOCK_RES_IN_PROGRESS;
236 /* add lock to local (secondary) queue */
238 list_add_tail(&lock->list, &res->blocked);
239 lock->lock_pending = 1;
240 spin_unlock(&res->spinlock);
242 /* spec seems to say that you will get DLM_NORMAL when the lock
243 * has been queued, meaning we need to wait for a reply here. */
244 status = dlm_send_remote_lock_request(dlm, res, lock, flags);
246 spin_lock(&res->spinlock);
247 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
248 lock->lock_pending = 0;
249 if (status != DLM_NORMAL) {
250 if (status == DLM_RECOVERING &&
251 dlm_is_recovery_lock(res->lockname.name,
252 res->lockname.len)) {
253 /* recovery lock was mastered by dead node.
254 * we need to have calc_usage shoot down this
255 * lockres and completely remaster it. */
256 mlog(0, "%s: recovery lock was owned by "
257 "dead node %u, remaster it now.\n",
258 dlm->name, res->owner);
259 } else if (status != DLM_NOTQUEUED) {
261 * DO NOT call calc_usage, as this would unhash
262 * the remote lockres before we ever get to use
263 * it. treat as if we never made any change to
269 dlm_revert_pending_lock(res, lock);
271 } else if (dlm_is_recovery_lock(res->lockname.name,
272 res->lockname.len)) {
273 /* special case for the $RECOVERY lock.
274 * there will never be an AST delivered to put
275 * this lock on the proper secondary queue
276 * (granted), so do it manually. */
277 mlog(0, "%s: $RECOVERY lock for this node (%u) is "
278 "mastered by %u; got lock, manually granting (no ast)\n",
279 dlm->name, dlm->node_num, res->owner);
280 list_move_tail(&lock->list, &res->granted);
282 spin_unlock(&res->spinlock);
285 dlm_lockres_calc_usage(dlm, res);
292 /* for remote lock creation.
294 * caller needs: none, but need res->state & DLM_LOCK_RES_IN_PROGRESS
297 * returns: DLM_NOLOCKMGR, or net status
299 static enum dlm_status dlm_send_remote_lock_request(struct dlm_ctxt *dlm,
300 struct dlm_lock_resource *res,
301 struct dlm_lock *lock, int flags)
303 struct dlm_create_lock create;
304 int tmpret, status = 0;
307 memset(&create, 0, sizeof(create));
308 create.node_idx = dlm->node_num;
309 create.requested_type = lock->ml.type;
310 create.cookie = lock->ml.cookie;
311 create.namelen = res->lockname.len;
312 create.flags = cpu_to_be32(flags);
313 memcpy(create.name, res->lockname.name, create.namelen);
315 tmpret = o2net_send_message(DLM_CREATE_LOCK_MSG, dlm->key, &create,
316 sizeof(create), res->owner, &status);
319 if (ret == DLM_REJECTED) {
320 mlog(ML_ERROR, "%s: res %.*s, Stale lockres no longer "
321 "owned by node %u. That node is coming back up "
322 "currently.\n", dlm->name, create.namelen,
323 create.name, res->owner);
324 dlm_print_one_lock_resource(res);
328 mlog(ML_ERROR, "%s: res %.*s, Error %d send CREATE LOCK to "
329 "node %u\n", dlm->name, create.namelen, create.name,
331 if (dlm_is_host_down(tmpret))
332 ret = DLM_RECOVERING;
334 ret = dlm_err_to_dlm_status(tmpret);
340 void dlm_lock_get(struct dlm_lock *lock)
342 kref_get(&lock->lock_refs);
345 void dlm_lock_put(struct dlm_lock *lock)
347 kref_put(&lock->lock_refs, dlm_lock_release);
350 static void dlm_lock_release(struct kref *kref)
352 struct dlm_lock *lock;
354 lock = container_of(kref, struct dlm_lock, lock_refs);
356 BUG_ON(!list_empty(&lock->list));
357 BUG_ON(!list_empty(&lock->ast_list));
358 BUG_ON(!list_empty(&lock->bast_list));
359 BUG_ON(lock->ast_pending);
360 BUG_ON(lock->bast_pending);
362 dlm_lock_detach_lockres(lock);
364 if (lock->lksb_kernel_allocated) {
365 mlog(0, "freeing kernel-allocated lksb\n");
368 kmem_cache_free(dlm_lock_cache, lock);
371 /* associate a lock with it's lockres, getting a ref on the lockres */
372 void dlm_lock_attach_lockres(struct dlm_lock *lock,
373 struct dlm_lock_resource *res)
375 dlm_lockres_get(res);
379 /* drop ref on lockres, if there is still one associated with lock */
380 static void dlm_lock_detach_lockres(struct dlm_lock *lock)
382 struct dlm_lock_resource *res;
386 lock->lockres = NULL;
387 mlog(0, "removing lock's lockres reference\n");
388 dlm_lockres_put(res);
392 static void dlm_init_lock(struct dlm_lock *newlock, int type,
395 INIT_LIST_HEAD(&newlock->list);
396 INIT_LIST_HEAD(&newlock->ast_list);
397 INIT_LIST_HEAD(&newlock->bast_list);
398 spin_lock_init(&newlock->spinlock);
399 newlock->ml.type = type;
400 newlock->ml.convert_type = LKM_IVMODE;
401 newlock->ml.highest_blocked = LKM_IVMODE;
402 newlock->ml.node = node;
403 newlock->ml.pad1 = 0;
404 newlock->ml.list = 0;
405 newlock->ml.flags = 0;
407 newlock->bast = NULL;
408 newlock->astdata = NULL;
409 newlock->ml.cookie = cpu_to_be64(cookie);
410 newlock->ast_pending = 0;
411 newlock->bast_pending = 0;
412 newlock->convert_pending = 0;
413 newlock->lock_pending = 0;
414 newlock->unlock_pending = 0;
415 newlock->cancel_pending = 0;
416 newlock->lksb_kernel_allocated = 0;
418 kref_init(&newlock->lock_refs);
421 struct dlm_lock * dlm_new_lock(int type, u8 node, u64 cookie,
422 struct dlm_lockstatus *lksb)
424 struct dlm_lock *lock;
425 int kernel_allocated = 0;
427 lock = kmem_cache_zalloc(dlm_lock_cache, GFP_NOFS);
432 /* zero memory only if kernel-allocated */
433 lksb = kzalloc(sizeof(*lksb), GFP_NOFS);
438 kernel_allocated = 1;
441 dlm_init_lock(lock, type, node, cookie);
442 if (kernel_allocated)
443 lock->lksb_kernel_allocated = 1;
449 /* handler for lock creation net message
452 * taken: takes and drops res->spinlock
454 * returns: DLM_NORMAL, DLM_SYSERR, DLM_IVLOCKID, DLM_NOTQUEUED
456 int dlm_create_lock_handler(struct o2net_msg *msg, u32 len, void *data,
459 struct dlm_ctxt *dlm = data;
460 struct dlm_create_lock *create = (struct dlm_create_lock *)msg->buf;
461 struct dlm_lock_resource *res = NULL;
462 struct dlm_lock *newlock = NULL;
463 struct dlm_lockstatus *lksb = NULL;
464 enum dlm_status status = DLM_NORMAL;
466 unsigned int namelen;
474 namelen = create->namelen;
475 status = DLM_REJECTED;
476 if (!dlm_domain_fully_joined(dlm)) {
477 mlog(ML_ERROR, "Domain %s not fully joined, but node %u is "
478 "sending a create_lock message for lock %.*s!\n",
479 dlm->name, create->node_idx, namelen, name);
484 status = DLM_IVBUFLEN;
485 if (namelen > DLM_LOCKID_NAME_MAX) {
491 newlock = dlm_new_lock(create->requested_type,
493 be64_to_cpu(create->cookie), NULL);
499 lksb = newlock->lksb;
501 if (be32_to_cpu(create->flags) & LKM_GET_LVB) {
502 lksb->flags |= DLM_LKSB_GET_LVB;
503 mlog(0, "set DLM_LKSB_GET_LVB flag\n");
506 status = DLM_IVLOCKID;
507 res = dlm_lookup_lockres(dlm, name, namelen);
513 spin_lock(&res->spinlock);
514 status = __dlm_lockres_state_to_status(res);
515 spin_unlock(&res->spinlock);
517 if (status != DLM_NORMAL) {
518 mlog(0, "lockres recovering/migrating/in-progress\n");
522 dlm_lock_attach_lockres(newlock, res);
524 status = dlmlock_master(dlm, res, newlock, be32_to_cpu(create->flags));
526 if (status != DLM_NORMAL)
528 dlm_lock_put(newlock);
531 dlm_lockres_put(res);
539 /* fetch next node-local (u8 nodenum + u56 cookie) into u64 */
540 static inline void dlm_get_next_cookie(u8 node_num, u64 *cookie)
542 u64 tmpnode = node_num;
544 /* shift single byte of node num into top 8 bits */
547 spin_lock(&dlm_cookie_lock);
548 *cookie = (dlm_next_cookie | tmpnode);
549 if (++dlm_next_cookie & 0xff00000000000000ull) {
550 mlog(0, "This node's cookie will now wrap!\n");
553 spin_unlock(&dlm_cookie_lock);
556 enum dlm_status dlmlock(struct dlm_ctxt *dlm, int mode,
557 struct dlm_lockstatus *lksb, int flags,
558 const char *name, int namelen, dlm_astlockfunc_t *ast,
559 void *data, dlm_bastlockfunc_t *bast)
561 enum dlm_status status;
562 struct dlm_lock_resource *res = NULL;
563 struct dlm_lock *lock = NULL;
564 int convert = 0, recovery = 0;
566 /* yes this function is a mess.
567 * TODO: clean this up. lots of common code in the
568 * lock and convert paths, especially in the retry blocks */
570 dlm_error(DLM_BADARGS);
574 status = DLM_BADPARAM;
575 if (mode != LKM_EXMODE && mode != LKM_PRMODE && mode != LKM_NLMODE) {
580 if (flags & ~LKM_VALID_FLAGS) {
585 convert = (flags & LKM_CONVERT);
586 recovery = (flags & LKM_RECOVERY);
589 (!dlm_is_recovery_lock(name, namelen) || convert) ) {
593 if (convert && (flags & LKM_LOCAL)) {
594 mlog(ML_ERROR, "strange LOCAL convert request!\n");
599 /* CONVERT request */
601 /* if converting, must pass in a valid dlm_lock */
604 mlog(ML_ERROR, "NULL lock pointer in convert "
611 mlog(ML_ERROR, "NULL lockres pointer in convert "
615 dlm_lockres_get(res);
617 /* XXX: for ocfs2 purposes, the ast/bast/astdata/lksb are
618 * static after the original lock call. convert requests will
619 * ensure that everything is the same, or return DLM_BADARGS.
620 * this means that DLM_DENIED_NOASTS will never be returned.
622 if (lock->lksb != lksb || lock->ast != ast ||
623 lock->bast != bast || lock->astdata != data) {
624 status = DLM_BADARGS;
625 mlog(ML_ERROR, "new args: lksb=%p, ast=%p, bast=%p, "
626 "astdata=%p\n", lksb, ast, bast, data);
627 mlog(ML_ERROR, "orig args: lksb=%p, ast=%p, bast=%p, "
628 "astdata=%p\n", lock->lksb, lock->ast,
629 lock->bast, lock->astdata);
633 dlm_wait_for_recovery(dlm);
635 if (res->owner == dlm->node_num)
636 status = dlmconvert_master(dlm, res, lock, flags, mode);
638 status = dlmconvert_remote(dlm, res, lock, flags, mode);
639 if (status == DLM_RECOVERING || status == DLM_MIGRATING ||
640 status == DLM_FORWARD) {
641 /* for now, see how this works without sleeping
642 * and just retry right away. I suspect the reco
643 * or migration will complete fast enough that
644 * no waiting will be necessary */
645 mlog(0, "retrying convert with migration/recovery/"
654 status = DLM_BADARGS;
660 status = DLM_IVBUFLEN;
661 if (namelen > DLM_LOCKID_NAME_MAX || namelen < 1) {
666 dlm_get_next_cookie(dlm->node_num, &tmpcookie);
667 lock = dlm_new_lock(mode, dlm->node_num, tmpcookie, lksb);
674 dlm_wait_for_recovery(dlm);
676 /* find or create the lock resource */
677 res = dlm_get_lock_resource(dlm, name, namelen, flags);
679 status = DLM_IVLOCKID;
684 mlog(0, "type=%d, flags = 0x%x\n", mode, flags);
685 mlog(0, "creating lock: lock=%p res=%p\n", lock, res);
687 dlm_lock_attach_lockres(lock, res);
690 lock->astdata = data;
693 if (flags & LKM_VALBLK) {
694 mlog(0, "LKM_VALBLK passed by caller\n");
696 /* LVB requests for non PR, PW or EX locks are
698 if (mode < LKM_PRMODE)
699 flags &= ~LKM_VALBLK;
701 flags |= LKM_GET_LVB;
702 lock->lksb->flags |= DLM_LKSB_GET_LVB;
706 if (res->owner == dlm->node_num)
707 status = dlmlock_master(dlm, res, lock, flags);
709 status = dlmlock_remote(dlm, res, lock, flags);
711 if (status == DLM_RECOVERING || status == DLM_MIGRATING ||
712 status == DLM_FORWARD) {
713 mlog(0, "retrying lock with migration/"
714 "recovery/in progress\n");
716 /* no waiting for dlm_reco_thread */
718 if (status != DLM_RECOVERING)
721 mlog(0, "%s: got RECOVERING "
722 "for $RECOVERY lock, master "
723 "was %u\n", dlm->name,
725 /* wait to see the node go down, then
726 * drop down and allow the lockres to
727 * get cleaned up. need to remaster. */
728 dlm_wait_for_node_death(dlm, res->owner,
729 DLM_NODE_DEATH_WAIT_MAX);
731 dlm_wait_for_recovery(dlm);
736 /* Inflight taken in dlm_get_lock_resource() is dropped here */
737 spin_lock(&res->spinlock);
738 dlm_lockres_drop_inflight_ref(dlm, res);
739 spin_unlock(&res->spinlock);
741 dlm_lockres_calc_usage(dlm, res);
742 dlm_kick_thread(dlm, res);
744 if (status != DLM_NORMAL) {
745 lock->lksb->flags &= ~DLM_LKSB_GET_LVB;
746 if (status != DLM_NOTQUEUED)
753 if (status != DLM_NORMAL) {
754 if (lock && !convert)
756 // this is kind of unnecessary
757 lksb->status = status;
760 /* put lockres ref from the convert path
761 * or from dlm_get_lock_resource */
763 dlm_lockres_put(res);
767 EXPORT_SYMBOL_GPL(dlmlock);