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/utsname.h>
34 #include <linux/init.h>
35 #include <linux/sysctl.h>
36 #include <linux/random.h>
37 #include <linux/blkdev.h>
38 #include <linux/socket.h>
39 #include <linux/inet.h>
40 #include <linux/spinlock.h>
41 #include <linux/delay.h>
44 #include "cluster/heartbeat.h"
45 #include "cluster/nodemanager.h"
46 #include "cluster/tcp.h"
49 #include "dlmcommon.h"
51 #include "dlmconvert.h"
53 #define MLOG_MASK_PREFIX ML_DLM
54 #include "cluster/masklog.h"
56 static spinlock_t dlm_cookie_lock = SPIN_LOCK_UNLOCKED;
57 static u64 dlm_next_cookie = 1;
59 static enum dlm_status dlm_send_remote_lock_request(struct dlm_ctxt *dlm,
60 struct dlm_lock_resource *res,
61 struct dlm_lock *lock, int flags);
62 static void dlm_init_lock(struct dlm_lock *newlock, int type,
64 static void dlm_lock_release(struct kref *kref);
65 static void dlm_lock_detach_lockres(struct dlm_lock *lock);
67 /* Tell us whether we can grant a new lock request.
69 * caller needs: res->spinlock
72 * returns: 1 if the lock can be granted, 0 otherwise.
74 static int dlm_can_grant_new_lock(struct dlm_lock_resource *res,
75 struct dlm_lock *lock)
77 struct list_head *iter;
78 struct dlm_lock *tmplock;
80 list_for_each(iter, &res->granted) {
81 tmplock = list_entry(iter, struct dlm_lock, list);
83 if (!dlm_lock_compatible(tmplock->ml.type, lock->ml.type))
87 list_for_each(iter, &res->converting) {
88 tmplock = list_entry(iter, struct dlm_lock, list);
90 if (!dlm_lock_compatible(tmplock->ml.type, lock->ml.type))
97 /* performs lock creation at the lockres master site
100 * taken: takes and drops res->spinlock
102 * returns: DLM_NORMAL, DLM_NOTQUEUED
104 static enum dlm_status dlmlock_master(struct dlm_ctxt *dlm,
105 struct dlm_lock_resource *res,
106 struct dlm_lock *lock, int flags)
108 int call_ast = 0, kick_thread = 0;
109 enum dlm_status status = DLM_NORMAL;
111 mlog_entry("type=%d\n", lock->ml.type);
113 spin_lock(&res->spinlock);
114 /* if called from dlm_create_lock_handler, need to
115 * ensure it will not sleep in dlm_wait_on_lockres */
116 status = __dlm_lockres_state_to_status(res);
117 if (status != DLM_NORMAL &&
118 lock->ml.node != dlm->node_num) {
119 /* erf. state changed after lock was dropped. */
120 spin_unlock(&res->spinlock);
124 __dlm_wait_on_lockres(res);
125 __dlm_lockres_reserve_ast(res);
127 if (dlm_can_grant_new_lock(res, lock)) {
128 mlog(0, "I can grant this lock right away\n");
129 /* got it right away */
130 lock->lksb->status = DLM_NORMAL;
133 list_add_tail(&lock->list, &res->granted);
135 /* for the recovery lock, we can't allow the ast
136 * to be queued since the dlmthread is already
137 * frozen. but the recovery lock is always locked
138 * with LKM_NOQUEUE so we do not need the ast in
139 * this special case */
140 if (!dlm_is_recovery_lock(res->lockname.name,
141 res->lockname.len)) {
145 mlog(0, "%s: returning DLM_NORMAL to "
146 "node %u for reco lock\n", dlm->name,
150 /* for NOQUEUE request, unless we get the
151 * lock right away, return DLM_NOTQUEUED */
152 if (flags & LKM_NOQUEUE) {
153 status = DLM_NOTQUEUED;
154 if (dlm_is_recovery_lock(res->lockname.name,
155 res->lockname.len)) {
156 mlog(0, "%s: returning NOTQUEUED to "
157 "node %u for reco lock\n", dlm->name,
162 list_add_tail(&lock->list, &res->blocked);
167 spin_unlock(&res->spinlock);
170 /* either queue the ast or release it */
172 dlm_queue_ast(dlm, lock);
174 dlm_lockres_release_ast(dlm, res);
176 dlm_lockres_calc_usage(dlm, res);
178 dlm_kick_thread(dlm, res);
183 void dlm_revert_pending_lock(struct dlm_lock_resource *res,
184 struct dlm_lock *lock)
186 /* remove from local queue if it failed */
187 list_del_init(&lock->list);
188 lock->lksb->flags &= ~DLM_LKSB_GET_LVB;
195 * taken: takes and drops res->spinlock
197 * returns: DLM_DENIED, DLM_RECOVERING, or net status
199 static enum dlm_status dlmlock_remote(struct dlm_ctxt *dlm,
200 struct dlm_lock_resource *res,
201 struct dlm_lock *lock, int flags)
203 enum dlm_status status = DLM_DENIED;
204 int lockres_changed = 1;
206 mlog_entry("type=%d\n", lock->ml.type);
207 mlog(0, "lockres %.*s, flags = 0x%x\n", res->lockname.len,
208 res->lockname.name, flags);
210 spin_lock(&res->spinlock);
212 /* will exit this call with spinlock held */
213 __dlm_wait_on_lockres(res);
214 res->state |= DLM_LOCK_RES_IN_PROGRESS;
216 /* add lock to local (secondary) queue */
218 list_add_tail(&lock->list, &res->blocked);
219 lock->lock_pending = 1;
220 spin_unlock(&res->spinlock);
222 /* spec seems to say that you will get DLM_NORMAL when the lock
223 * has been queued, meaning we need to wait for a reply here. */
224 status = dlm_send_remote_lock_request(dlm, res, lock, flags);
226 spin_lock(&res->spinlock);
227 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
228 lock->lock_pending = 0;
229 if (status != DLM_NORMAL) {
230 if (status != DLM_NOTQUEUED) {
232 * DO NOT call calc_usage, as this would unhash
233 * the remote lockres before we ever get to use
234 * it. treat as if we never made any change to
240 dlm_revert_pending_lock(res, lock);
242 } else if (dlm_is_recovery_lock(res->lockname.name,
243 res->lockname.len)) {
244 /* special case for the $RECOVERY lock.
245 * there will never be an AST delivered to put
246 * this lock on the proper secondary queue
247 * (granted), so do it manually. */
248 mlog(0, "%s: $RECOVERY lock for this node (%u) is "
249 "mastered by %u; got lock, manually granting (no ast)\n",
250 dlm->name, dlm->node_num, res->owner);
251 list_move_tail(&lock->list, &res->granted);
253 spin_unlock(&res->spinlock);
256 dlm_lockres_calc_usage(dlm, res);
263 /* for remote lock creation.
265 * caller needs: none, but need res->state & DLM_LOCK_RES_IN_PROGRESS
268 * returns: DLM_NOLOCKMGR, or net status
270 static enum dlm_status dlm_send_remote_lock_request(struct dlm_ctxt *dlm,
271 struct dlm_lock_resource *res,
272 struct dlm_lock *lock, int flags)
274 struct dlm_create_lock create;
275 int tmpret, status = 0;
280 memset(&create, 0, sizeof(create));
281 create.node_idx = dlm->node_num;
282 create.requested_type = lock->ml.type;
283 create.cookie = lock->ml.cookie;
284 create.namelen = res->lockname.len;
285 create.flags = cpu_to_be32(flags);
286 memcpy(create.name, res->lockname.name, create.namelen);
288 tmpret = o2net_send_message(DLM_CREATE_LOCK_MSG, dlm->key, &create,
289 sizeof(create), res->owner, &status);
291 // successfully sent and received
292 ret = status; // this is already a dlm_status
293 if (ret == DLM_RECOVERING) {
294 mlog(ML_ERROR, "%s:%.*s: BUG. this is a stale lockres "
295 "no longer owned by %u. that node is coming back "
296 "up currently.\n", dlm->name, create.namelen,
297 create.name, res->owner);
298 dlm_print_one_lock_resource(res);
303 if (dlm_is_host_down(tmpret)) {
304 ret = DLM_RECOVERING;
305 mlog(0, "node %u died so returning DLM_RECOVERING "
306 "from lock message!\n", res->owner);
308 ret = dlm_err_to_dlm_status(tmpret);
315 void dlm_lock_get(struct dlm_lock *lock)
317 kref_get(&lock->lock_refs);
320 void dlm_lock_put(struct dlm_lock *lock)
322 kref_put(&lock->lock_refs, dlm_lock_release);
325 static void dlm_lock_release(struct kref *kref)
327 struct dlm_lock *lock;
329 lock = container_of(kref, struct dlm_lock, lock_refs);
331 BUG_ON(!list_empty(&lock->list));
332 BUG_ON(!list_empty(&lock->ast_list));
333 BUG_ON(!list_empty(&lock->bast_list));
334 BUG_ON(lock->ast_pending);
335 BUG_ON(lock->bast_pending);
337 dlm_lock_detach_lockres(lock);
339 if (lock->lksb_kernel_allocated) {
340 mlog(0, "freeing kernel-allocated lksb\n");
346 /* associate a lock with it's lockres, getting a ref on the lockres */
347 void dlm_lock_attach_lockres(struct dlm_lock *lock,
348 struct dlm_lock_resource *res)
350 dlm_lockres_get(res);
354 /* drop ref on lockres, if there is still one associated with lock */
355 static void dlm_lock_detach_lockres(struct dlm_lock *lock)
357 struct dlm_lock_resource *res;
361 lock->lockres = NULL;
362 mlog(0, "removing lock's lockres reference\n");
363 dlm_lockres_put(res);
367 static void dlm_init_lock(struct dlm_lock *newlock, int type,
370 INIT_LIST_HEAD(&newlock->list);
371 INIT_LIST_HEAD(&newlock->ast_list);
372 INIT_LIST_HEAD(&newlock->bast_list);
373 spin_lock_init(&newlock->spinlock);
374 newlock->ml.type = type;
375 newlock->ml.convert_type = LKM_IVMODE;
376 newlock->ml.highest_blocked = LKM_IVMODE;
377 newlock->ml.node = node;
378 newlock->ml.pad1 = 0;
379 newlock->ml.list = 0;
380 newlock->ml.flags = 0;
382 newlock->bast = NULL;
383 newlock->astdata = NULL;
384 newlock->ml.cookie = cpu_to_be64(cookie);
385 newlock->ast_pending = 0;
386 newlock->bast_pending = 0;
387 newlock->convert_pending = 0;
388 newlock->lock_pending = 0;
389 newlock->unlock_pending = 0;
390 newlock->cancel_pending = 0;
391 newlock->lksb_kernel_allocated = 0;
393 kref_init(&newlock->lock_refs);
396 struct dlm_lock * dlm_new_lock(int type, u8 node, u64 cookie,
397 struct dlm_lockstatus *lksb)
399 struct dlm_lock *lock;
400 int kernel_allocated = 0;
402 lock = kcalloc(1, sizeof(*lock), GFP_KERNEL);
407 /* zero memory only if kernel-allocated */
408 lksb = kcalloc(1, sizeof(*lksb), GFP_KERNEL);
413 kernel_allocated = 1;
416 dlm_init_lock(lock, type, node, cookie);
417 if (kernel_allocated)
418 lock->lksb_kernel_allocated = 1;
424 /* handler for lock creation net message
427 * taken: takes and drops res->spinlock
429 * returns: DLM_NORMAL, DLM_SYSERR, DLM_IVLOCKID, DLM_NOTQUEUED
431 int dlm_create_lock_handler(struct o2net_msg *msg, u32 len, void *data)
433 struct dlm_ctxt *dlm = data;
434 struct dlm_create_lock *create = (struct dlm_create_lock *)msg->buf;
435 struct dlm_lock_resource *res = NULL;
436 struct dlm_lock *newlock = NULL;
437 struct dlm_lockstatus *lksb = NULL;
438 enum dlm_status status = DLM_NORMAL;
440 unsigned int namelen;
450 namelen = create->namelen;
451 status = DLM_RECOVERING;
452 if (!dlm_domain_fully_joined(dlm)) {
453 mlog(ML_ERROR, "Domain %s not fully joined, but node %u is "
454 "sending a create_lock message for lock %.*s!\n",
455 dlm->name, create->node_idx, namelen, name);
460 status = DLM_IVBUFLEN;
461 if (namelen > DLM_LOCKID_NAME_MAX) {
467 newlock = dlm_new_lock(create->requested_type,
469 be64_to_cpu(create->cookie), NULL);
475 lksb = newlock->lksb;
477 if (be32_to_cpu(create->flags) & LKM_GET_LVB) {
478 lksb->flags |= DLM_LKSB_GET_LVB;
479 mlog(0, "set DLM_LKSB_GET_LVB flag\n");
482 status = DLM_IVLOCKID;
483 res = dlm_lookup_lockres(dlm, name, namelen);
489 spin_lock(&res->spinlock);
490 status = __dlm_lockres_state_to_status(res);
491 spin_unlock(&res->spinlock);
493 if (status != DLM_NORMAL) {
494 mlog(0, "lockres recovering/migrating/in-progress\n");
498 dlm_lock_attach_lockres(newlock, res);
500 status = dlmlock_master(dlm, res, newlock, be32_to_cpu(create->flags));
502 if (status != DLM_NORMAL)
504 dlm_lock_put(newlock);
507 dlm_lockres_put(res);
515 /* fetch next node-local (u8 nodenum + u56 cookie) into u64 */
516 static inline void dlm_get_next_cookie(u8 node_num, u64 *cookie)
518 u64 tmpnode = node_num;
520 /* shift single byte of node num into top 8 bits */
523 spin_lock(&dlm_cookie_lock);
524 *cookie = (dlm_next_cookie | tmpnode);
525 if (++dlm_next_cookie & 0xff00000000000000ull) {
526 mlog(0, "This node's cookie will now wrap!\n");
529 spin_unlock(&dlm_cookie_lock);
532 enum dlm_status dlmlock(struct dlm_ctxt *dlm, int mode,
533 struct dlm_lockstatus *lksb, int flags,
534 const char *name, dlm_astlockfunc_t *ast, void *data,
535 dlm_bastlockfunc_t *bast)
537 enum dlm_status status;
538 struct dlm_lock_resource *res = NULL;
539 struct dlm_lock *lock = NULL;
540 int convert = 0, recovery = 0;
542 /* yes this function is a mess.
543 * TODO: clean this up. lots of common code in the
544 * lock and convert paths, especially in the retry blocks */
546 dlm_error(DLM_BADARGS);
550 status = DLM_BADPARAM;
551 if (mode != LKM_EXMODE && mode != LKM_PRMODE && mode != LKM_NLMODE) {
556 if (flags & ~LKM_VALID_FLAGS) {
561 convert = (flags & LKM_CONVERT);
562 recovery = (flags & LKM_RECOVERY);
565 (!dlm_is_recovery_lock(name, strlen(name)) || convert) ) {
569 if (convert && (flags & LKM_LOCAL)) {
570 mlog(ML_ERROR, "strange LOCAL convert request!\n");
575 /* CONVERT request */
577 /* if converting, must pass in a valid dlm_lock */
580 mlog(ML_ERROR, "NULL lock pointer in convert "
587 mlog(ML_ERROR, "NULL lockres pointer in convert "
591 dlm_lockres_get(res);
593 /* XXX: for ocfs2 purposes, the ast/bast/astdata/lksb are
594 * static after the original lock call. convert requests will
595 * ensure that everything is the same, or return DLM_BADARGS.
596 * this means that DLM_DENIED_NOASTS will never be returned.
598 if (lock->lksb != lksb || lock->ast != ast ||
599 lock->bast != bast || lock->astdata != data) {
600 status = DLM_BADARGS;
601 mlog(ML_ERROR, "new args: lksb=%p, ast=%p, bast=%p, "
602 "astdata=%p\n", lksb, ast, bast, data);
603 mlog(ML_ERROR, "orig args: lksb=%p, ast=%p, bast=%p, "
604 "astdata=%p\n", lock->lksb, lock->ast,
605 lock->bast, lock->astdata);
609 dlm_wait_for_recovery(dlm);
611 if (res->owner == dlm->node_num)
612 status = dlmconvert_master(dlm, res, lock, flags, mode);
614 status = dlmconvert_remote(dlm, res, lock, flags, mode);
615 if (status == DLM_RECOVERING || status == DLM_MIGRATING ||
616 status == DLM_FORWARD) {
617 /* for now, see how this works without sleeping
618 * and just retry right away. I suspect the reco
619 * or migration will complete fast enough that
620 * no waiting will be necessary */
621 mlog(0, "retrying convert with migration/recovery/"
630 status = DLM_BADARGS;
636 status = DLM_IVBUFLEN;
637 if (strlen(name) > DLM_LOCKID_NAME_MAX || strlen(name) < 1) {
642 dlm_get_next_cookie(dlm->node_num, &tmpcookie);
643 lock = dlm_new_lock(mode, dlm->node_num, tmpcookie, lksb);
650 dlm_wait_for_recovery(dlm);
652 /* find or create the lock resource */
653 res = dlm_get_lock_resource(dlm, name, flags);
655 status = DLM_IVLOCKID;
660 mlog(0, "type=%d, flags = 0x%x\n", mode, flags);
661 mlog(0, "creating lock: lock=%p res=%p\n", lock, res);
663 dlm_lock_attach_lockres(lock, res);
666 lock->astdata = data;
669 if (flags & LKM_VALBLK) {
670 mlog(0, "LKM_VALBLK passed by caller\n");
672 /* LVB requests for non PR, PW or EX locks are
674 if (mode < LKM_PRMODE)
675 flags &= ~LKM_VALBLK;
677 flags |= LKM_GET_LVB;
678 lock->lksb->flags |= DLM_LKSB_GET_LVB;
682 if (res->owner == dlm->node_num)
683 status = dlmlock_master(dlm, res, lock, flags);
685 status = dlmlock_remote(dlm, res, lock, flags);
687 if (status == DLM_RECOVERING || status == DLM_MIGRATING ||
688 status == DLM_FORWARD) {
689 mlog(0, "retrying lock with migration/"
690 "recovery/in progress\n");
692 /* no waiting for dlm_reco_thread */
694 if (status == DLM_RECOVERING) {
695 mlog(0, "%s: got RECOVERING "
696 "for $REOCVERY lock, master "
697 "was %u\n", dlm->name,
699 dlm_wait_for_node_death(dlm, res->owner,
700 DLM_NODE_DEATH_WAIT_MAX);
703 dlm_wait_for_recovery(dlm);
708 if (status != DLM_NORMAL) {
709 lock->lksb->flags &= ~DLM_LKSB_GET_LVB;
710 if (status != DLM_NOTQUEUED)
717 if (status != DLM_NORMAL) {
718 if (lock && !convert)
720 // this is kind of unnecessary
721 lksb->status = status;
724 /* put lockres ref from the convert path
725 * or from dlm_get_lock_resource */
727 dlm_lockres_put(res);
731 EXPORT_SYMBOL_GPL(dlmlock);