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);
231 * Wait if resource is getting recovered, remastered, etc.
232 * If the resource was remastered and new owner is self, then exit.
234 spin_lock(&res->spinlock);
235 __dlm_wait_on_lockres(res);
236 if (res->owner == dlm->node_num) {
237 spin_unlock(&res->spinlock);
238 return DLM_RECOVERING;
240 res->state |= DLM_LOCK_RES_IN_PROGRESS;
242 /* add lock to local (secondary) queue */
244 list_add_tail(&lock->list, &res->blocked);
245 lock->lock_pending = 1;
246 spin_unlock(&res->spinlock);
248 /* spec seems to say that you will get DLM_NORMAL when the lock
249 * has been queued, meaning we need to wait for a reply here. */
250 status = dlm_send_remote_lock_request(dlm, res, lock, flags);
252 spin_lock(&res->spinlock);
253 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
254 lock->lock_pending = 0;
255 if (status != DLM_NORMAL) {
256 if (status == DLM_RECOVERING &&
257 dlm_is_recovery_lock(res->lockname.name,
258 res->lockname.len)) {
259 /* recovery lock was mastered by dead node.
260 * we need to have calc_usage shoot down this
261 * lockres and completely remaster it. */
262 mlog(0, "%s: recovery lock was owned by "
263 "dead node %u, remaster it now.\n",
264 dlm->name, res->owner);
265 } else if (status != DLM_NOTQUEUED) {
267 * DO NOT call calc_usage, as this would unhash
268 * the remote lockres before we ever get to use
269 * it. treat as if we never made any change to
275 dlm_revert_pending_lock(res, lock);
277 } else if (dlm_is_recovery_lock(res->lockname.name,
278 res->lockname.len)) {
279 /* special case for the $RECOVERY lock.
280 * there will never be an AST delivered to put
281 * this lock on the proper secondary queue
282 * (granted), so do it manually. */
283 mlog(0, "%s: $RECOVERY lock for this node (%u) is "
284 "mastered by %u; got lock, manually granting (no ast)\n",
285 dlm->name, dlm->node_num, res->owner);
286 list_move_tail(&lock->list, &res->granted);
288 spin_unlock(&res->spinlock);
291 dlm_lockres_calc_usage(dlm, res);
298 /* for remote lock creation.
300 * caller needs: none, but need res->state & DLM_LOCK_RES_IN_PROGRESS
303 * returns: DLM_NOLOCKMGR, or net status
305 static enum dlm_status dlm_send_remote_lock_request(struct dlm_ctxt *dlm,
306 struct dlm_lock_resource *res,
307 struct dlm_lock *lock, int flags)
309 struct dlm_create_lock create;
310 int tmpret, status = 0;
313 memset(&create, 0, sizeof(create));
314 create.node_idx = dlm->node_num;
315 create.requested_type = lock->ml.type;
316 create.cookie = lock->ml.cookie;
317 create.namelen = res->lockname.len;
318 create.flags = cpu_to_be32(flags);
319 memcpy(create.name, res->lockname.name, create.namelen);
321 tmpret = o2net_send_message(DLM_CREATE_LOCK_MSG, dlm->key, &create,
322 sizeof(create), res->owner, &status);
325 if (ret == DLM_REJECTED) {
326 mlog(ML_ERROR, "%s: res %.*s, Stale lockres no longer "
327 "owned by node %u. That node is coming back up "
328 "currently.\n", dlm->name, create.namelen,
329 create.name, res->owner);
330 dlm_print_one_lock_resource(res);
334 mlog(ML_ERROR, "%s: res %.*s, Error %d send CREATE LOCK to "
335 "node %u\n", dlm->name, create.namelen, create.name,
337 if (dlm_is_host_down(tmpret))
338 ret = DLM_RECOVERING;
340 ret = dlm_err_to_dlm_status(tmpret);
346 void dlm_lock_get(struct dlm_lock *lock)
348 kref_get(&lock->lock_refs);
351 void dlm_lock_put(struct dlm_lock *lock)
353 kref_put(&lock->lock_refs, dlm_lock_release);
356 static void dlm_lock_release(struct kref *kref)
358 struct dlm_lock *lock;
360 lock = container_of(kref, struct dlm_lock, lock_refs);
362 BUG_ON(!list_empty(&lock->list));
363 BUG_ON(!list_empty(&lock->ast_list));
364 BUG_ON(!list_empty(&lock->bast_list));
365 BUG_ON(lock->ast_pending);
366 BUG_ON(lock->bast_pending);
368 dlm_lock_detach_lockres(lock);
370 if (lock->lksb_kernel_allocated) {
371 mlog(0, "freeing kernel-allocated lksb\n");
374 kmem_cache_free(dlm_lock_cache, lock);
377 /* associate a lock with it's lockres, getting a ref on the lockres */
378 void dlm_lock_attach_lockres(struct dlm_lock *lock,
379 struct dlm_lock_resource *res)
381 dlm_lockres_get(res);
385 /* drop ref on lockres, if there is still one associated with lock */
386 static void dlm_lock_detach_lockres(struct dlm_lock *lock)
388 struct dlm_lock_resource *res;
392 lock->lockres = NULL;
393 mlog(0, "removing lock's lockres reference\n");
394 dlm_lockres_put(res);
398 static void dlm_init_lock(struct dlm_lock *newlock, int type,
401 INIT_LIST_HEAD(&newlock->list);
402 INIT_LIST_HEAD(&newlock->ast_list);
403 INIT_LIST_HEAD(&newlock->bast_list);
404 spin_lock_init(&newlock->spinlock);
405 newlock->ml.type = type;
406 newlock->ml.convert_type = LKM_IVMODE;
407 newlock->ml.highest_blocked = LKM_IVMODE;
408 newlock->ml.node = node;
409 newlock->ml.pad1 = 0;
410 newlock->ml.list = 0;
411 newlock->ml.flags = 0;
413 newlock->bast = NULL;
414 newlock->astdata = NULL;
415 newlock->ml.cookie = cpu_to_be64(cookie);
416 newlock->ast_pending = 0;
417 newlock->bast_pending = 0;
418 newlock->convert_pending = 0;
419 newlock->lock_pending = 0;
420 newlock->unlock_pending = 0;
421 newlock->cancel_pending = 0;
422 newlock->lksb_kernel_allocated = 0;
424 kref_init(&newlock->lock_refs);
427 struct dlm_lock * dlm_new_lock(int type, u8 node, u64 cookie,
428 struct dlm_lockstatus *lksb)
430 struct dlm_lock *lock;
431 int kernel_allocated = 0;
433 lock = kmem_cache_zalloc(dlm_lock_cache, GFP_NOFS);
438 /* zero memory only if kernel-allocated */
439 lksb = kzalloc(sizeof(*lksb), GFP_NOFS);
441 kmem_cache_free(dlm_lock_cache, lock);
444 kernel_allocated = 1;
447 dlm_init_lock(lock, type, node, cookie);
448 if (kernel_allocated)
449 lock->lksb_kernel_allocated = 1;
455 /* handler for lock creation net message
458 * taken: takes and drops res->spinlock
460 * returns: DLM_NORMAL, DLM_SYSERR, DLM_IVLOCKID, DLM_NOTQUEUED
462 int dlm_create_lock_handler(struct o2net_msg *msg, u32 len, void *data,
465 struct dlm_ctxt *dlm = data;
466 struct dlm_create_lock *create = (struct dlm_create_lock *)msg->buf;
467 struct dlm_lock_resource *res = NULL;
468 struct dlm_lock *newlock = NULL;
469 struct dlm_lockstatus *lksb = NULL;
470 enum dlm_status status = DLM_NORMAL;
472 unsigned int namelen;
480 namelen = create->namelen;
481 status = DLM_REJECTED;
482 if (!dlm_domain_fully_joined(dlm)) {
483 mlog(ML_ERROR, "Domain %s not fully joined, but node %u is "
484 "sending a create_lock message for lock %.*s!\n",
485 dlm->name, create->node_idx, namelen, name);
490 status = DLM_IVBUFLEN;
491 if (namelen > DLM_LOCKID_NAME_MAX) {
497 newlock = dlm_new_lock(create->requested_type,
499 be64_to_cpu(create->cookie), NULL);
505 lksb = newlock->lksb;
507 if (be32_to_cpu(create->flags) & LKM_GET_LVB) {
508 lksb->flags |= DLM_LKSB_GET_LVB;
509 mlog(0, "set DLM_LKSB_GET_LVB flag\n");
512 status = DLM_IVLOCKID;
513 res = dlm_lookup_lockres(dlm, name, namelen);
519 spin_lock(&res->spinlock);
520 status = __dlm_lockres_state_to_status(res);
521 spin_unlock(&res->spinlock);
523 if (status != DLM_NORMAL) {
524 mlog(0, "lockres recovering/migrating/in-progress\n");
528 dlm_lock_attach_lockres(newlock, res);
530 status = dlmlock_master(dlm, res, newlock, be32_to_cpu(create->flags));
532 if (status != DLM_NORMAL)
534 dlm_lock_put(newlock);
537 dlm_lockres_put(res);
545 /* fetch next node-local (u8 nodenum + u56 cookie) into u64 */
546 static inline void dlm_get_next_cookie(u8 node_num, u64 *cookie)
548 u64 tmpnode = node_num;
550 /* shift single byte of node num into top 8 bits */
553 spin_lock(&dlm_cookie_lock);
554 *cookie = (dlm_next_cookie | tmpnode);
555 if (++dlm_next_cookie & 0xff00000000000000ull) {
556 mlog(0, "This node's cookie will now wrap!\n");
559 spin_unlock(&dlm_cookie_lock);
562 enum dlm_status dlmlock(struct dlm_ctxt *dlm, int mode,
563 struct dlm_lockstatus *lksb, int flags,
564 const char *name, int namelen, dlm_astlockfunc_t *ast,
565 void *data, dlm_bastlockfunc_t *bast)
567 enum dlm_status status;
568 struct dlm_lock_resource *res = NULL;
569 struct dlm_lock *lock = NULL;
570 int convert = 0, recovery = 0;
572 /* yes this function is a mess.
573 * TODO: clean this up. lots of common code in the
574 * lock and convert paths, especially in the retry blocks */
576 dlm_error(DLM_BADARGS);
580 status = DLM_BADPARAM;
581 if (mode != LKM_EXMODE && mode != LKM_PRMODE && mode != LKM_NLMODE) {
586 if (flags & ~LKM_VALID_FLAGS) {
591 convert = (flags & LKM_CONVERT);
592 recovery = (flags & LKM_RECOVERY);
595 (!dlm_is_recovery_lock(name, namelen) || convert) ) {
599 if (convert && (flags & LKM_LOCAL)) {
600 mlog(ML_ERROR, "strange LOCAL convert request!\n");
605 /* CONVERT request */
607 /* if converting, must pass in a valid dlm_lock */
610 mlog(ML_ERROR, "NULL lock pointer in convert "
617 mlog(ML_ERROR, "NULL lockres pointer in convert "
621 dlm_lockres_get(res);
623 /* XXX: for ocfs2 purposes, the ast/bast/astdata/lksb are
624 * static after the original lock call. convert requests will
625 * ensure that everything is the same, or return DLM_BADARGS.
626 * this means that DLM_DENIED_NOASTS will never be returned.
628 if (lock->lksb != lksb || lock->ast != ast ||
629 lock->bast != bast || lock->astdata != data) {
630 status = DLM_BADARGS;
631 mlog(ML_ERROR, "new args: lksb=%p, ast=%p, bast=%p, "
632 "astdata=%p\n", lksb, ast, bast, data);
633 mlog(ML_ERROR, "orig args: lksb=%p, ast=%p, bast=%p, "
634 "astdata=%p\n", lock->lksb, lock->ast,
635 lock->bast, lock->astdata);
639 dlm_wait_for_recovery(dlm);
641 if (res->owner == dlm->node_num)
642 status = dlmconvert_master(dlm, res, lock, flags, mode);
644 status = dlmconvert_remote(dlm, res, lock, flags, mode);
645 if (status == DLM_RECOVERING || status == DLM_MIGRATING ||
646 status == DLM_FORWARD) {
647 /* for now, see how this works without sleeping
648 * and just retry right away. I suspect the reco
649 * or migration will complete fast enough that
650 * no waiting will be necessary */
651 mlog(0, "retrying convert with migration/recovery/"
660 status = DLM_BADARGS;
666 status = DLM_IVBUFLEN;
667 if (namelen > DLM_LOCKID_NAME_MAX || namelen < 1) {
672 dlm_get_next_cookie(dlm->node_num, &tmpcookie);
673 lock = dlm_new_lock(mode, dlm->node_num, tmpcookie, lksb);
680 dlm_wait_for_recovery(dlm);
682 /* find or create the lock resource */
683 res = dlm_get_lock_resource(dlm, name, namelen, flags);
685 status = DLM_IVLOCKID;
690 mlog(0, "type=%d, flags = 0x%x\n", mode, flags);
691 mlog(0, "creating lock: lock=%p res=%p\n", lock, res);
693 dlm_lock_attach_lockres(lock, res);
696 lock->astdata = data;
699 if (flags & LKM_VALBLK) {
700 mlog(0, "LKM_VALBLK passed by caller\n");
702 /* LVB requests for non PR, PW or EX locks are
704 if (mode < LKM_PRMODE)
705 flags &= ~LKM_VALBLK;
707 flags |= LKM_GET_LVB;
708 lock->lksb->flags |= DLM_LKSB_GET_LVB;
712 if (res->owner == dlm->node_num)
713 status = dlmlock_master(dlm, res, lock, flags);
715 status = dlmlock_remote(dlm, res, lock, flags);
717 if (status == DLM_RECOVERING || status == DLM_MIGRATING ||
718 status == DLM_FORWARD) {
721 if (status != DLM_RECOVERING)
723 /* wait to see the node go down, then
724 * drop down and allow the lockres to
725 * get cleaned up. need to remaster. */
726 dlm_wait_for_node_death(dlm, res->owner,
727 DLM_NODE_DEATH_WAIT_MAX);
729 dlm_wait_for_recovery(dlm);
734 /* Inflight taken in dlm_get_lock_resource() is dropped here */
735 spin_lock(&res->spinlock);
736 dlm_lockres_drop_inflight_ref(dlm, res);
737 spin_unlock(&res->spinlock);
739 dlm_lockres_calc_usage(dlm, res);
740 dlm_kick_thread(dlm, res);
742 if (status != DLM_NORMAL) {
743 lock->lksb->flags &= ~DLM_LKSB_GET_LVB;
744 if (status != DLM_NOTQUEUED)
751 if (status != DLM_NORMAL) {
752 if (lock && !convert)
754 // this is kind of unnecessary
755 lksb->status = status;
758 /* put lockres ref from the convert path
759 * or from dlm_get_lock_resource */
761 dlm_lockres_put(res);
765 EXPORT_SYMBOL_GPL(dlmlock);