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;
205 mlog_entry("type=%d\n", lock->ml.type);
206 mlog(0, "lockres %.*s, flags = 0x%x\n", res->lockname.len,
207 res->lockname.name, flags);
209 spin_lock(&res->spinlock);
211 /* will exit this call with spinlock held */
212 __dlm_wait_on_lockres(res);
213 res->state |= DLM_LOCK_RES_IN_PROGRESS;
215 /* add lock to local (secondary) queue */
217 list_add_tail(&lock->list, &res->blocked);
218 lock->lock_pending = 1;
219 spin_unlock(&res->spinlock);
221 /* spec seems to say that you will get DLM_NORMAL when the lock
222 * has been queued, meaning we need to wait for a reply here. */
223 status = dlm_send_remote_lock_request(dlm, res, lock, flags);
225 spin_lock(&res->spinlock);
226 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
227 lock->lock_pending = 0;
228 if (status != DLM_NORMAL) {
229 if (status != DLM_NOTQUEUED)
231 dlm_revert_pending_lock(res, lock);
233 } else if (dlm_is_recovery_lock(res->lockname.name,
234 res->lockname.len)) {
235 /* special case for the $RECOVERY lock.
236 * there will never be an AST delivered to put
237 * this lock on the proper secondary queue
238 * (granted), so do it manually. */
239 mlog(0, "%s: $RECOVERY lock for this node (%u) is "
240 "mastered by %u; got lock, manually granting (no ast)\n",
241 dlm->name, dlm->node_num, res->owner);
242 list_move_tail(&lock->list, &res->granted);
244 spin_unlock(&res->spinlock);
246 dlm_lockres_calc_usage(dlm, res);
253 /* for remote lock creation.
255 * caller needs: none, but need res->state & DLM_LOCK_RES_IN_PROGRESS
258 * returns: DLM_NOLOCKMGR, or net status
260 static enum dlm_status dlm_send_remote_lock_request(struct dlm_ctxt *dlm,
261 struct dlm_lock_resource *res,
262 struct dlm_lock *lock, int flags)
264 struct dlm_create_lock create;
265 int tmpret, status = 0;
270 memset(&create, 0, sizeof(create));
271 create.node_idx = dlm->node_num;
272 create.requested_type = lock->ml.type;
273 create.cookie = lock->ml.cookie;
274 create.namelen = res->lockname.len;
275 create.flags = cpu_to_be32(flags);
276 memcpy(create.name, res->lockname.name, create.namelen);
278 tmpret = o2net_send_message(DLM_CREATE_LOCK_MSG, dlm->key, &create,
279 sizeof(create), res->owner, &status);
281 // successfully sent and received
282 ret = status; // this is already a dlm_status
283 if (ret == DLM_RECOVERING) {
284 mlog(ML_ERROR, "%s:%.*s: BUG. this is a stale lockres "
285 "no longer owned by %u. that node is coming back "
286 "up currently.\n", dlm->name, create.namelen,
287 create.name, res->owner);
288 dlm_print_one_lock_resource(res);
293 if (dlm_is_host_down(tmpret)) {
294 ret = DLM_RECOVERING;
295 mlog(0, "node %u died so returning DLM_RECOVERING "
296 "from lock message!\n", res->owner);
298 ret = dlm_err_to_dlm_status(tmpret);
305 void dlm_lock_get(struct dlm_lock *lock)
307 kref_get(&lock->lock_refs);
310 void dlm_lock_put(struct dlm_lock *lock)
312 kref_put(&lock->lock_refs, dlm_lock_release);
315 static void dlm_lock_release(struct kref *kref)
317 struct dlm_lock *lock;
319 lock = container_of(kref, struct dlm_lock, lock_refs);
321 BUG_ON(!list_empty(&lock->list));
322 BUG_ON(!list_empty(&lock->ast_list));
323 BUG_ON(!list_empty(&lock->bast_list));
324 BUG_ON(lock->ast_pending);
325 BUG_ON(lock->bast_pending);
327 dlm_lock_detach_lockres(lock);
329 if (lock->lksb_kernel_allocated) {
330 mlog(0, "freeing kernel-allocated lksb\n");
336 /* associate a lock with it's lockres, getting a ref on the lockres */
337 void dlm_lock_attach_lockres(struct dlm_lock *lock,
338 struct dlm_lock_resource *res)
340 dlm_lockres_get(res);
344 /* drop ref on lockres, if there is still one associated with lock */
345 static void dlm_lock_detach_lockres(struct dlm_lock *lock)
347 struct dlm_lock_resource *res;
351 lock->lockres = NULL;
352 mlog(0, "removing lock's lockres reference\n");
353 dlm_lockres_put(res);
357 static void dlm_init_lock(struct dlm_lock *newlock, int type,
360 INIT_LIST_HEAD(&newlock->list);
361 INIT_LIST_HEAD(&newlock->ast_list);
362 INIT_LIST_HEAD(&newlock->bast_list);
363 spin_lock_init(&newlock->spinlock);
364 newlock->ml.type = type;
365 newlock->ml.convert_type = LKM_IVMODE;
366 newlock->ml.highest_blocked = LKM_IVMODE;
367 newlock->ml.node = node;
368 newlock->ml.pad1 = 0;
369 newlock->ml.list = 0;
370 newlock->ml.flags = 0;
372 newlock->bast = NULL;
373 newlock->astdata = NULL;
374 newlock->ml.cookie = cpu_to_be64(cookie);
375 newlock->ast_pending = 0;
376 newlock->bast_pending = 0;
377 newlock->convert_pending = 0;
378 newlock->lock_pending = 0;
379 newlock->unlock_pending = 0;
380 newlock->cancel_pending = 0;
381 newlock->lksb_kernel_allocated = 0;
383 kref_init(&newlock->lock_refs);
386 struct dlm_lock * dlm_new_lock(int type, u8 node, u64 cookie,
387 struct dlm_lockstatus *lksb)
389 struct dlm_lock *lock;
390 int kernel_allocated = 0;
392 lock = kcalloc(1, sizeof(*lock), GFP_KERNEL);
397 /* zero memory only if kernel-allocated */
398 lksb = kcalloc(1, sizeof(*lksb), GFP_KERNEL);
403 kernel_allocated = 1;
406 dlm_init_lock(lock, type, node, cookie);
407 if (kernel_allocated)
408 lock->lksb_kernel_allocated = 1;
414 /* handler for lock creation net message
417 * taken: takes and drops res->spinlock
419 * returns: DLM_NORMAL, DLM_SYSERR, DLM_IVLOCKID, DLM_NOTQUEUED
421 int dlm_create_lock_handler(struct o2net_msg *msg, u32 len, void *data)
423 struct dlm_ctxt *dlm = data;
424 struct dlm_create_lock *create = (struct dlm_create_lock *)msg->buf;
425 struct dlm_lock_resource *res = NULL;
426 struct dlm_lock *newlock = NULL;
427 struct dlm_lockstatus *lksb = NULL;
428 enum dlm_status status = DLM_NORMAL;
430 unsigned int namelen;
440 namelen = create->namelen;
441 status = DLM_RECOVERING;
442 if (!dlm_domain_fully_joined(dlm)) {
443 mlog(ML_ERROR, "Domain %s not fully joined, but node %u is "
444 "sending a create_lock message for lock %.*s!\n",
445 dlm->name, create->node_idx, namelen, name);
450 status = DLM_IVBUFLEN;
451 if (namelen > DLM_LOCKID_NAME_MAX) {
457 newlock = dlm_new_lock(create->requested_type,
459 be64_to_cpu(create->cookie), NULL);
465 lksb = newlock->lksb;
467 if (be32_to_cpu(create->flags) & LKM_GET_LVB) {
468 lksb->flags |= DLM_LKSB_GET_LVB;
469 mlog(0, "set DLM_LKSB_GET_LVB flag\n");
472 status = DLM_IVLOCKID;
473 res = dlm_lookup_lockres(dlm, name, namelen);
479 spin_lock(&res->spinlock);
480 status = __dlm_lockres_state_to_status(res);
481 spin_unlock(&res->spinlock);
483 if (status != DLM_NORMAL) {
484 mlog(0, "lockres recovering/migrating/in-progress\n");
488 dlm_lock_attach_lockres(newlock, res);
490 status = dlmlock_master(dlm, res, newlock, be32_to_cpu(create->flags));
492 if (status != DLM_NORMAL)
494 dlm_lock_put(newlock);
497 dlm_lockres_put(res);
505 /* fetch next node-local (u8 nodenum + u56 cookie) into u64 */
506 static inline void dlm_get_next_cookie(u8 node_num, u64 *cookie)
508 u64 tmpnode = node_num;
510 /* shift single byte of node num into top 8 bits */
513 spin_lock(&dlm_cookie_lock);
514 *cookie = (dlm_next_cookie | tmpnode);
515 if (++dlm_next_cookie & 0xff00000000000000ull) {
516 mlog(0, "This node's cookie will now wrap!\n");
519 spin_unlock(&dlm_cookie_lock);
522 enum dlm_status dlmlock(struct dlm_ctxt *dlm, int mode,
523 struct dlm_lockstatus *lksb, int flags,
524 const char *name, dlm_astlockfunc_t *ast, void *data,
525 dlm_bastlockfunc_t *bast)
527 enum dlm_status status;
528 struct dlm_lock_resource *res = NULL;
529 struct dlm_lock *lock = NULL;
530 int convert = 0, recovery = 0;
532 /* yes this function is a mess.
533 * TODO: clean this up. lots of common code in the
534 * lock and convert paths, especially in the retry blocks */
536 dlm_error(DLM_BADARGS);
540 status = DLM_BADPARAM;
541 if (mode != LKM_EXMODE && mode != LKM_PRMODE && mode != LKM_NLMODE) {
546 if (flags & ~LKM_VALID_FLAGS) {
551 convert = (flags & LKM_CONVERT);
552 recovery = (flags & LKM_RECOVERY);
555 (!dlm_is_recovery_lock(name, strlen(name)) || convert) ) {
559 if (convert && (flags & LKM_LOCAL)) {
560 mlog(ML_ERROR, "strange LOCAL convert request!\n");
565 /* CONVERT request */
567 /* if converting, must pass in a valid dlm_lock */
570 mlog(ML_ERROR, "NULL lock pointer in convert "
577 mlog(ML_ERROR, "NULL lockres pointer in convert "
581 dlm_lockres_get(res);
583 /* XXX: for ocfs2 purposes, the ast/bast/astdata/lksb are
584 * static after the original lock call. convert requests will
585 * ensure that everything is the same, or return DLM_BADARGS.
586 * this means that DLM_DENIED_NOASTS will never be returned.
588 if (lock->lksb != lksb || lock->ast != ast ||
589 lock->bast != bast || lock->astdata != data) {
590 status = DLM_BADARGS;
591 mlog(ML_ERROR, "new args: lksb=%p, ast=%p, bast=%p, "
592 "astdata=%p\n", lksb, ast, bast, data);
593 mlog(ML_ERROR, "orig args: lksb=%p, ast=%p, bast=%p, "
594 "astdata=%p\n", lock->lksb, lock->ast,
595 lock->bast, lock->astdata);
599 dlm_wait_for_recovery(dlm);
601 if (res->owner == dlm->node_num)
602 status = dlmconvert_master(dlm, res, lock, flags, mode);
604 status = dlmconvert_remote(dlm, res, lock, flags, mode);
605 if (status == DLM_RECOVERING || status == DLM_MIGRATING ||
606 status == DLM_FORWARD) {
607 /* for now, see how this works without sleeping
608 * and just retry right away. I suspect the reco
609 * or migration will complete fast enough that
610 * no waiting will be necessary */
611 mlog(0, "retrying convert with migration/recovery/"
620 status = DLM_BADARGS;
626 status = DLM_IVBUFLEN;
627 if (strlen(name) > DLM_LOCKID_NAME_MAX || strlen(name) < 1) {
632 dlm_get_next_cookie(dlm->node_num, &tmpcookie);
633 lock = dlm_new_lock(mode, dlm->node_num, tmpcookie, lksb);
640 dlm_wait_for_recovery(dlm);
642 /* find or create the lock resource */
643 res = dlm_get_lock_resource(dlm, name, flags);
645 status = DLM_IVLOCKID;
650 mlog(0, "type=%d, flags = 0x%x\n", mode, flags);
651 mlog(0, "creating lock: lock=%p res=%p\n", lock, res);
653 dlm_lock_attach_lockres(lock, res);
656 lock->astdata = data;
659 if (flags & LKM_VALBLK) {
660 mlog(0, "LKM_VALBLK passed by caller\n");
662 /* LVB requests for non PR, PW or EX locks are
664 if (mode < LKM_PRMODE)
665 flags &= ~LKM_VALBLK;
667 flags |= LKM_GET_LVB;
668 lock->lksb->flags |= DLM_LKSB_GET_LVB;
672 if (res->owner == dlm->node_num)
673 status = dlmlock_master(dlm, res, lock, flags);
675 status = dlmlock_remote(dlm, res, lock, flags);
677 if (status == DLM_RECOVERING || status == DLM_MIGRATING ||
678 status == DLM_FORWARD) {
679 mlog(0, "retrying lock with migration/"
680 "recovery/in progress\n");
682 /* no waiting for dlm_reco_thread */
684 if (status == DLM_RECOVERING) {
685 mlog(0, "%s: got RECOVERING "
686 "for $REOCVERY lock, master "
687 "was %u\n", dlm->name,
689 dlm_wait_for_node_death(dlm, res->owner,
690 DLM_NODE_DEATH_WAIT_MAX);
693 dlm_wait_for_recovery(dlm);
698 if (status != DLM_NORMAL) {
699 lock->lksb->flags &= ~DLM_LKSB_GET_LVB;
700 if (status != DLM_NOTQUEUED)
707 if (status != DLM_NORMAL) {
708 if (lock && !convert)
710 // this is kind of unnecessary
711 lksb->status = status;
714 /* put lockres ref from the convert path
715 * or from dlm_get_lock_resource */
717 dlm_lockres_put(res);
721 EXPORT_SYMBOL_GPL(dlmlock);