On Tue, Nov 06, 2007 at 02:33:53AM -0800, akpm@linux-foundation.org wrote:
[mmotm.git] / fs / ocfs2 / dlm / dlmconvert.c
blobca96bce50e18cb6629343f6a0436b8bec2aa20a6
1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
4 * dlmconvert.c
6 * underlying calls for lock conversion
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>
29 #include <linux/fs.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>
42 #include "cluster/heartbeat.h"
43 #include "cluster/nodemanager.h"
44 #include "cluster/tcp.h"
46 #include "dlmapi.h"
47 #include "dlmcommon.h"
49 #include "dlmconvert.h"
51 #define MLOG_MASK_PREFIX ML_DLM
52 #include "cluster/masklog.h"
54 /* NOTE: __dlmconvert_master is the only function in here that
55 * needs a spinlock held on entry (res->spinlock) and it is the
56 * only one that holds a lock on exit (res->spinlock).
57 * All other functions in here need no locks and drop all of
58 * the locks that they acquire. */
59 static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm,
60 struct dlm_lock_resource *res,
61 struct dlm_lock *lock, int flags,
62 int type, int *call_ast,
63 int *kick_thread);
64 static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm,
65 struct dlm_lock_resource *res,
66 struct dlm_lock *lock, int flags, int type);
69 * this is only called directly by dlmlock(), and only when the
70 * local node is the owner of the lockres
71 * locking:
72 * caller needs: none
73 * taken: takes and drops res->spinlock
74 * held on exit: none
75 * returns: see __dlmconvert_master
77 enum dlm_status dlmconvert_master(struct dlm_ctxt *dlm,
78 struct dlm_lock_resource *res,
79 struct dlm_lock *lock, int flags, int type)
81 int call_ast = 0, kick_thread = 0;
82 enum dlm_status status;
84 spin_lock(&res->spinlock);
85 /* we are not in a network handler, this is fine */
86 __dlm_wait_on_lockres(res);
87 __dlm_lockres_reserve_ast(res);
88 res->state |= DLM_LOCK_RES_IN_PROGRESS;
90 status = __dlmconvert_master(dlm, res, lock, flags, type,
91 &call_ast, &kick_thread);
93 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
94 spin_unlock(&res->spinlock);
95 wake_up(&res->wq);
96 if (status != DLM_NORMAL && status != DLM_NOTQUEUED)
97 dlm_error(status);
99 /* either queue the ast or release it */
100 if (call_ast)
101 dlm_queue_ast(dlm, lock);
102 else
103 dlm_lockres_release_ast(dlm, res);
105 if (kick_thread)
106 dlm_kick_thread(dlm, res);
108 return status;
111 /* performs lock conversion at the lockres master site
112 * locking:
113 * caller needs: res->spinlock
114 * taken: takes and drops lock->spinlock
115 * held on exit: res->spinlock
116 * returns: DLM_NORMAL, DLM_NOTQUEUED, DLM_DENIED
117 * call_ast: whether ast should be called for this lock
118 * kick_thread: whether dlm_kick_thread should be called
120 static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm,
121 struct dlm_lock_resource *res,
122 struct dlm_lock *lock, int flags,
123 int type, int *call_ast,
124 int *kick_thread)
126 enum dlm_status status = DLM_NORMAL;
127 struct list_head *iter;
128 struct dlm_lock *tmplock=NULL;
130 assert_spin_locked(&res->spinlock);
132 mlog_entry("type=%d, convert_type=%d, new convert_type=%d\n",
133 lock->ml.type, lock->ml.convert_type, type);
135 spin_lock(&lock->spinlock);
137 /* already converting? */
138 if (lock->ml.convert_type != LKM_IVMODE) {
139 mlog(ML_ERROR, "attempted to convert a lock with a lock "
140 "conversion pending\n");
141 status = DLM_DENIED;
142 goto unlock_exit;
145 /* must be on grant queue to convert */
146 if (!dlm_lock_on_list(&res->granted, lock)) {
147 mlog(ML_ERROR, "attempted to convert a lock not on grant "
148 "queue\n");
149 status = DLM_DENIED;
150 goto unlock_exit;
153 if (flags & LKM_VALBLK) {
154 switch (lock->ml.type) {
155 case LKM_EXMODE:
156 /* EX + LKM_VALBLK + convert == set lvb */
157 mlog(0, "will set lvb: converting %s->%s\n",
158 dlm_lock_mode_name(lock->ml.type),
159 dlm_lock_mode_name(type));
160 lock->lksb->flags |= DLM_LKSB_PUT_LVB;
161 break;
162 case LKM_PRMODE:
163 case LKM_NLMODE:
164 /* refetch if new level is not NL */
165 if (type > LKM_NLMODE) {
166 mlog(0, "will fetch new value into "
167 "lvb: converting %s->%s\n",
168 dlm_lock_mode_name(lock->ml.type),
169 dlm_lock_mode_name(type));
170 lock->lksb->flags |= DLM_LKSB_GET_LVB;
171 } else {
172 mlog(0, "will NOT fetch new value "
173 "into lvb: converting %s->%s\n",
174 dlm_lock_mode_name(lock->ml.type),
175 dlm_lock_mode_name(type));
176 flags &= ~(LKM_VALBLK);
178 break;
183 /* in-place downconvert? */
184 if (type <= lock->ml.type)
185 goto grant;
187 /* upconvert from here on */
188 status = DLM_NORMAL;
189 list_for_each(iter, &res->granted) {
190 tmplock = list_entry(iter, struct dlm_lock, list);
191 if (tmplock == lock)
192 continue;
193 if (!dlm_lock_compatible(tmplock->ml.type, type))
194 goto switch_queues;
197 list_for_each(iter, &res->converting) {
198 tmplock = list_entry(iter, struct dlm_lock, list);
199 if (!dlm_lock_compatible(tmplock->ml.type, type))
200 goto switch_queues;
201 /* existing conversion requests take precedence */
202 if (!dlm_lock_compatible(tmplock->ml.convert_type, type))
203 goto switch_queues;
206 /* fall thru to grant */
208 grant:
209 mlog(0, "res %.*s, granting %s lock\n", res->lockname.len,
210 res->lockname.name, dlm_lock_mode_name(type));
211 /* immediately grant the new lock type */
212 lock->lksb->status = DLM_NORMAL;
213 if (lock->ml.node == dlm->node_num)
214 mlog(0, "doing in-place convert for nonlocal lock\n");
215 lock->ml.type = type;
216 if (lock->lksb->flags & DLM_LKSB_PUT_LVB)
217 memcpy(res->lvb, lock->lksb->lvb, DLM_LVB_LEN);
219 status = DLM_NORMAL;
220 *call_ast = 1;
221 goto unlock_exit;
223 switch_queues:
224 if (flags & LKM_NOQUEUE) {
225 mlog(0, "failed to convert NOQUEUE lock %.*s from "
226 "%d to %d...\n", res->lockname.len, res->lockname.name,
227 lock->ml.type, type);
228 status = DLM_NOTQUEUED;
229 goto unlock_exit;
231 mlog(0, "res %.*s, queueing...\n", res->lockname.len,
232 res->lockname.name);
234 lock->ml.convert_type = type;
235 /* do not alter lock refcount. switching lists. */
236 list_move_tail(&lock->list, &res->converting);
238 unlock_exit:
239 spin_unlock(&lock->spinlock);
240 if (status == DLM_DENIED) {
241 __dlm_print_one_lock_resource(res);
243 if (status == DLM_NORMAL)
244 *kick_thread = 1;
245 return status;
248 void dlm_revert_pending_convert(struct dlm_lock_resource *res,
249 struct dlm_lock *lock)
251 /* do not alter lock refcount. switching lists. */
252 list_move_tail(&lock->list, &res->granted);
253 lock->ml.convert_type = LKM_IVMODE;
254 lock->lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB);
257 /* messages the master site to do lock conversion
258 * locking:
259 * caller needs: none
260 * taken: takes and drops res->spinlock, uses DLM_LOCK_RES_IN_PROGRESS
261 * held on exit: none
262 * returns: DLM_NORMAL, DLM_RECOVERING, status from remote node
264 enum dlm_status dlmconvert_remote(struct dlm_ctxt *dlm,
265 struct dlm_lock_resource *res,
266 struct dlm_lock *lock, int flags, int type)
268 enum dlm_status status;
270 mlog(0, "type=%d, convert_type=%d, busy=%d\n", lock->ml.type,
271 lock->ml.convert_type, res->state & DLM_LOCK_RES_IN_PROGRESS);
273 spin_lock(&res->spinlock);
274 if (res->state & DLM_LOCK_RES_RECOVERING) {
275 mlog(0, "bailing out early since res is RECOVERING "
276 "on secondary queue\n");
277 /* __dlm_print_one_lock_resource(res); */
278 status = DLM_RECOVERING;
279 goto bail;
281 /* will exit this call with spinlock held */
282 __dlm_wait_on_lockres(res);
284 if (lock->ml.convert_type != LKM_IVMODE) {
285 __dlm_print_one_lock_resource(res);
286 mlog(ML_ERROR, "converting a remote lock that is already "
287 "converting! (cookie=%u:%llu, conv=%d)\n",
288 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
289 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
290 lock->ml.convert_type);
291 status = DLM_DENIED;
292 goto bail;
294 res->state |= DLM_LOCK_RES_IN_PROGRESS;
295 /* move lock to local convert queue */
296 /* do not alter lock refcount. switching lists. */
297 list_move_tail(&lock->list, &res->converting);
298 lock->convert_pending = 1;
299 lock->ml.convert_type = type;
301 if (flags & LKM_VALBLK) {
302 if (lock->ml.type == LKM_EXMODE) {
303 flags |= LKM_PUT_LVB;
304 lock->lksb->flags |= DLM_LKSB_PUT_LVB;
305 } else {
306 if (lock->ml.convert_type == LKM_NLMODE)
307 flags &= ~LKM_VALBLK;
308 else {
309 flags |= LKM_GET_LVB;
310 lock->lksb->flags |= DLM_LKSB_GET_LVB;
314 spin_unlock(&res->spinlock);
316 /* no locks held here.
317 * need to wait for a reply as to whether it got queued or not. */
318 status = dlm_send_remote_convert_request(dlm, res, lock, flags, type);
320 spin_lock(&res->spinlock);
321 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
322 lock->convert_pending = 0;
323 /* if it failed, move it back to granted queue */
324 if (status != DLM_NORMAL) {
325 if (status != DLM_NOTQUEUED)
326 dlm_error(status);
327 dlm_revert_pending_convert(res, lock);
329 bail:
330 spin_unlock(&res->spinlock);
332 /* TODO: should this be a wake_one? */
333 /* wake up any IN_PROGRESS waiters */
334 wake_up(&res->wq);
336 return status;
339 /* sends DLM_CONVERT_LOCK_MSG to master site
340 * locking:
341 * caller needs: none
342 * taken: none
343 * held on exit: none
344 * returns: DLM_NOLOCKMGR, status from remote node
346 static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm,
347 struct dlm_lock_resource *res,
348 struct dlm_lock *lock, int flags, int type)
350 struct dlm_convert_lock convert;
351 int tmpret;
352 enum dlm_status ret;
353 int status = 0;
354 struct kvec vec[2];
355 size_t veclen = 1;
357 mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
359 memset(&convert, 0, sizeof(struct dlm_convert_lock));
360 convert.node_idx = dlm->node_num;
361 convert.requested_type = type;
362 convert.cookie = lock->ml.cookie;
363 convert.namelen = res->lockname.len;
364 convert.flags = cpu_to_be32(flags);
365 memcpy(convert.name, res->lockname.name, convert.namelen);
367 vec[0].iov_len = sizeof(struct dlm_convert_lock);
368 vec[0].iov_base = &convert;
370 if (flags & LKM_PUT_LVB) {
371 /* extra data to send if we are updating lvb */
372 vec[1].iov_len = DLM_LVB_LEN;
373 vec[1].iov_base = lock->lksb->lvb;
374 veclen++;
377 tmpret = o2net_send_message_vec(DLM_CONVERT_LOCK_MSG, dlm->key,
378 vec, veclen, res->owner, &status);
379 if (tmpret >= 0) {
380 // successfully sent and received
381 ret = status; // this is already a dlm_status
382 if (ret == DLM_RECOVERING) {
383 mlog(0, "node %u returned DLM_RECOVERING from convert "
384 "message!\n", res->owner);
385 } else if (ret == DLM_MIGRATING) {
386 mlog(0, "node %u returned DLM_MIGRATING from convert "
387 "message!\n", res->owner);
388 } else if (ret == DLM_FORWARD) {
389 mlog(0, "node %u returned DLM_FORWARD from convert "
390 "message!\n", res->owner);
391 } else if (ret != DLM_NORMAL && ret != DLM_NOTQUEUED)
392 dlm_error(ret);
393 } else {
394 mlog_errno(tmpret);
395 if (dlm_is_host_down(tmpret)) {
396 /* instead of logging the same network error over
397 * and over, sleep here and wait for the heartbeat
398 * to notice the node is dead. times out after 5s. */
399 dlm_wait_for_node_death(dlm, res->owner,
400 DLM_NODE_DEATH_WAIT_MAX);
401 ret = DLM_RECOVERING;
402 mlog(0, "node %u died so returning DLM_RECOVERING "
403 "from convert message!\n", res->owner);
404 } else {
405 ret = dlm_err_to_dlm_status(tmpret);
409 return ret;
412 /* handler for DLM_CONVERT_LOCK_MSG on master site
413 * locking:
414 * caller needs: none
415 * taken: takes and drop res->spinlock
416 * held on exit: none
417 * returns: DLM_NORMAL, DLM_IVLOCKID, DLM_BADARGS,
418 * status from __dlmconvert_master
420 int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data,
421 void **ret_data)
423 struct dlm_ctxt *dlm = data;
424 struct dlm_convert_lock *cnv = (struct dlm_convert_lock *)msg->buf;
425 struct dlm_lock_resource *res = NULL;
426 struct list_head *iter;
427 struct dlm_lock *lock = NULL;
428 struct dlm_lockstatus *lksb;
429 enum dlm_status status = DLM_NORMAL;
430 u32 flags;
431 int call_ast = 0, kick_thread = 0, ast_reserved = 0, wake = 0;
433 if (!dlm_grab(dlm)) {
434 dlm_error(DLM_REJECTED);
435 return DLM_REJECTED;
438 mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
439 "Domain %s not fully joined!\n", dlm->name);
441 if (cnv->namelen > DLM_LOCKID_NAME_MAX) {
442 status = DLM_IVBUFLEN;
443 dlm_error(status);
444 goto leave;
447 flags = be32_to_cpu(cnv->flags);
449 if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) ==
450 (LKM_PUT_LVB|LKM_GET_LVB)) {
451 mlog(ML_ERROR, "both PUT and GET lvb specified\n");
452 status = DLM_BADARGS;
453 goto leave;
456 mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" :
457 (flags & LKM_GET_LVB ? "get lvb" : "none"));
459 status = DLM_IVLOCKID;
460 res = dlm_lookup_lockres(dlm, cnv->name, cnv->namelen);
461 if (!res) {
462 dlm_error(status);
463 goto leave;
466 spin_lock(&res->spinlock);
467 status = __dlm_lockres_state_to_status(res);
468 if (status != DLM_NORMAL) {
469 spin_unlock(&res->spinlock);
470 dlm_error(status);
471 goto leave;
473 list_for_each(iter, &res->granted) {
474 lock = list_entry(iter, struct dlm_lock, list);
475 if (lock->ml.cookie == cnv->cookie &&
476 lock->ml.node == cnv->node_idx) {
477 dlm_lock_get(lock);
478 break;
480 lock = NULL;
482 spin_unlock(&res->spinlock);
483 if (!lock) {
484 status = DLM_IVLOCKID;
485 mlog(ML_ERROR, "did not find lock to convert on grant queue! "
486 "cookie=%u:%llu\n",
487 dlm_get_lock_cookie_node(be64_to_cpu(cnv->cookie)),
488 dlm_get_lock_cookie_seq(be64_to_cpu(cnv->cookie)));
489 dlm_print_one_lock_resource(res);
490 goto leave;
493 /* found the lock */
494 lksb = lock->lksb;
496 /* see if caller needed to get/put lvb */
497 if (flags & LKM_PUT_LVB) {
498 BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
499 lksb->flags |= DLM_LKSB_PUT_LVB;
500 memcpy(&lksb->lvb[0], &cnv->lvb[0], DLM_LVB_LEN);
501 } else if (flags & LKM_GET_LVB) {
502 BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
503 lksb->flags |= DLM_LKSB_GET_LVB;
506 spin_lock(&res->spinlock);
507 status = __dlm_lockres_state_to_status(res);
508 if (status == DLM_NORMAL) {
509 __dlm_lockres_reserve_ast(res);
510 ast_reserved = 1;
511 res->state |= DLM_LOCK_RES_IN_PROGRESS;
512 status = __dlmconvert_master(dlm, res, lock, flags,
513 cnv->requested_type,
514 &call_ast, &kick_thread);
515 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
516 wake = 1;
518 spin_unlock(&res->spinlock);
519 if (wake)
520 wake_up(&res->wq);
522 if (status != DLM_NORMAL) {
523 if (status != DLM_NOTQUEUED)
524 dlm_error(status);
525 lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB);
528 leave:
529 if (lock)
530 dlm_lock_put(lock);
532 /* either queue the ast or release it, if reserved */
533 if (call_ast)
534 dlm_queue_ast(dlm, lock);
535 else if (ast_reserved)
536 dlm_lockres_release_ast(dlm, res);
538 if (kick_thread)
539 dlm_kick_thread(dlm, res);
541 if (res)
542 dlm_lockres_put(res);
544 dlm_put(dlm);
546 return status;