1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
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>
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>
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 /* NOTE: __dlmconvert_master is the only function in here that
56 * needs a spinlock held on entry (res->spinlock) and it is the
57 * only one that holds a lock on exit (res->spinlock).
58 * All other functions in here need no locks and drop all of
59 * the locks that they acquire. */
60 static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm,
61 struct dlm_lock_resource *res,
62 struct dlm_lock *lock, int flags,
63 int type, int *call_ast,
65 static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm,
66 struct dlm_lock_resource *res,
67 struct dlm_lock *lock, int flags, int type);
70 * this is only called directly by dlmlock(), and only when the
71 * local node is the owner of the lockres
74 * taken: takes and drops res->spinlock
76 * returns: see __dlmconvert_master
78 enum dlm_status dlmconvert_master(struct dlm_ctxt *dlm,
79 struct dlm_lock_resource *res,
80 struct dlm_lock *lock, int flags, int type)
82 int call_ast = 0, kick_thread = 0;
83 enum dlm_status status;
85 spin_lock(&res->spinlock);
86 /* we are not in a network handler, this is fine */
87 __dlm_wait_on_lockres(res);
88 __dlm_lockres_reserve_ast(res);
89 res->state |= DLM_LOCK_RES_IN_PROGRESS;
91 status = __dlmconvert_master(dlm, res, lock, flags, type,
92 &call_ast, &kick_thread);
94 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
95 spin_unlock(&res->spinlock);
97 if (status != DLM_NORMAL && status != DLM_NOTQUEUED)
100 /* either queue the ast or release it */
102 dlm_queue_ast(dlm, lock);
104 dlm_lockres_release_ast(dlm, res);
107 dlm_kick_thread(dlm, res);
112 /* performs lock conversion at the lockres master site
114 * caller needs: res->spinlock
115 * taken: takes and drops lock->spinlock
116 * held on exit: res->spinlock
117 * returns: DLM_NORMAL, DLM_NOTQUEUED, DLM_DENIED
118 * call_ast: whether ast should be called for this lock
119 * kick_thread: whether dlm_kick_thread should be called
121 static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm,
122 struct dlm_lock_resource *res,
123 struct dlm_lock *lock, int flags,
124 int type, int *call_ast,
127 enum dlm_status status = DLM_NORMAL;
128 struct list_head *iter;
129 struct dlm_lock *tmplock=NULL;
131 assert_spin_locked(&res->spinlock);
133 mlog_entry("type=%d, convert_type=%d, new convert_type=%d\n",
134 lock->ml.type, lock->ml.convert_type, type);
136 spin_lock(&lock->spinlock);
138 /* already converting? */
139 if (lock->ml.convert_type != LKM_IVMODE) {
140 mlog(ML_ERROR, "attempted to convert a lock with a lock "
141 "conversion pending\n");
146 /* must be on grant queue to convert */
147 if (!dlm_lock_on_list(&res->granted, lock)) {
148 mlog(ML_ERROR, "attempted to convert a lock not on grant "
154 if (flags & LKM_VALBLK) {
155 switch (lock->ml.type) {
157 /* EX + LKM_VALBLK + convert == set lvb */
158 mlog(0, "will set lvb: converting %s->%s\n",
159 dlm_lock_mode_name(lock->ml.type),
160 dlm_lock_mode_name(type));
161 lock->lksb->flags |= DLM_LKSB_PUT_LVB;
165 /* refetch if new level is not NL */
166 if (type > LKM_NLMODE) {
167 mlog(0, "will fetch new value into "
168 "lvb: converting %s->%s\n",
169 dlm_lock_mode_name(lock->ml.type),
170 dlm_lock_mode_name(type));
171 lock->lksb->flags |= DLM_LKSB_GET_LVB;
173 mlog(0, "will NOT fetch new value "
174 "into lvb: converting %s->%s\n",
175 dlm_lock_mode_name(lock->ml.type),
176 dlm_lock_mode_name(type));
177 flags &= ~(LKM_VALBLK);
184 /* in-place downconvert? */
185 if (type <= lock->ml.type)
188 /* upconvert from here on */
190 list_for_each(iter, &res->granted) {
191 tmplock = list_entry(iter, struct dlm_lock, list);
194 if (!dlm_lock_compatible(tmplock->ml.type, type))
198 list_for_each(iter, &res->converting) {
199 tmplock = list_entry(iter, struct dlm_lock, list);
200 if (!dlm_lock_compatible(tmplock->ml.type, type))
202 /* existing conversion requests take precedence */
203 if (!dlm_lock_compatible(tmplock->ml.convert_type, type))
207 /* fall thru to grant */
210 mlog(0, "res %.*s, granting %s lock\n", res->lockname.len,
211 res->lockname.name, dlm_lock_mode_name(type));
212 /* immediately grant the new lock type */
213 lock->lksb->status = DLM_NORMAL;
214 if (lock->ml.node == dlm->node_num)
215 mlog(0, "doing in-place convert for nonlocal lock\n");
216 lock->ml.type = type;
222 if (flags & LKM_NOQUEUE) {
223 mlog(0, "failed to convert NOQUEUE lock %.*s from "
224 "%d to %d...\n", res->lockname.len, res->lockname.name,
225 lock->ml.type, type);
226 status = DLM_NOTQUEUED;
229 mlog(0, "res %.*s, queueing...\n", res->lockname.len,
232 lock->ml.convert_type = type;
233 /* do not alter lock refcount. switching lists. */
234 list_del_init(&lock->list);
235 list_add_tail(&lock->list, &res->converting);
238 spin_unlock(&lock->spinlock);
239 if (status == DLM_DENIED) {
240 __dlm_print_one_lock_resource(res);
242 if (status == DLM_NORMAL)
247 void dlm_revert_pending_convert(struct dlm_lock_resource *res,
248 struct dlm_lock *lock)
250 /* do not alter lock refcount. switching lists. */
251 list_del_init(&lock->list);
252 list_add_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
260 * taken: takes and drops res->spinlock, uses DLM_LOCK_RES_IN_PROGRESS
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;
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=%"MLFu64", conv=%d)\n",
288 lock->ml.cookie, lock->ml.convert_type);
292 res->state |= DLM_LOCK_RES_IN_PROGRESS;
293 /* move lock to local convert queue */
294 /* do not alter lock refcount. switching lists. */
295 list_del_init(&lock->list);
296 list_add_tail(&lock->list, &res->converting);
297 lock->convert_pending = 1;
298 lock->ml.convert_type = type;
300 if (flags & LKM_VALBLK) {
301 if (lock->ml.type == LKM_EXMODE) {
302 flags |= LKM_PUT_LVB;
303 lock->lksb->flags |= DLM_LKSB_PUT_LVB;
305 if (lock->ml.convert_type == LKM_NLMODE)
306 flags &= ~LKM_VALBLK;
308 flags |= LKM_GET_LVB;
309 lock->lksb->flags |= DLM_LKSB_GET_LVB;
313 spin_unlock(&res->spinlock);
315 /* no locks held here.
316 * need to wait for a reply as to whether it got queued or not. */
317 status = dlm_send_remote_convert_request(dlm, res, lock, flags, type);
319 spin_lock(&res->spinlock);
320 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
321 lock->convert_pending = 0;
322 /* if it failed, move it back to granted queue */
323 if (status != DLM_NORMAL) {
324 if (status != DLM_NOTQUEUED)
326 dlm_revert_pending_convert(res, lock);
329 spin_unlock(&res->spinlock);
331 /* TODO: should this be a wake_one? */
332 /* wake up any IN_PROGRESS waiters */
338 /* sends DLM_CONVERT_LOCK_MSG to master site
343 * returns: DLM_NOLOCKMGR, status from remote node
345 static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm,
346 struct dlm_lock_resource *res,
347 struct dlm_lock *lock, int flags, int type)
349 struct dlm_convert_lock convert;
356 mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
358 memset(&convert, 0, sizeof(struct dlm_convert_lock));
359 convert.node_idx = dlm->node_num;
360 convert.requested_type = type;
361 convert.cookie = lock->ml.cookie;
362 convert.namelen = res->lockname.len;
363 convert.flags = cpu_to_be32(flags);
364 memcpy(convert.name, res->lockname.name, convert.namelen);
366 vec[0].iov_len = sizeof(struct dlm_convert_lock);
367 vec[0].iov_base = &convert;
369 if (flags & LKM_PUT_LVB) {
370 /* extra data to send if we are updating lvb */
371 vec[1].iov_len = DLM_LVB_LEN;
372 vec[1].iov_base = lock->lksb->lvb;
376 tmpret = o2net_send_message_vec(DLM_CONVERT_LOCK_MSG, dlm->key,
377 vec, veclen, res->owner, &status);
379 // successfully sent and received
380 ret = status; // this is already a dlm_status
381 if (ret == DLM_RECOVERING) {
382 mlog(0, "node %u returned DLM_RECOVERING from convert "
383 "message!\n", res->owner);
384 } else if (ret == DLM_MIGRATING) {
385 mlog(0, "node %u returned DLM_MIGRATING from convert "
386 "message!\n", res->owner);
387 } else if (ret == DLM_FORWARD) {
388 mlog(0, "node %u returned DLM_FORWARD from convert "
389 "message!\n", res->owner);
390 } else if (ret != DLM_NORMAL && ret != DLM_NOTQUEUED)
394 if (dlm_is_host_down(tmpret)) {
395 /* instead of logging the same network error over
396 * and over, sleep here and wait for the heartbeat
397 * to notice the node is dead. times out after 5s. */
398 dlm_wait_for_node_death(dlm, res->owner,
399 DLM_NODE_DEATH_WAIT_MAX);
400 ret = DLM_RECOVERING;
401 mlog(0, "node %u died so returning DLM_RECOVERING "
402 "from convert message!\n", res->owner);
404 ret = dlm_err_to_dlm_status(tmpret);
411 /* handler for DLM_CONVERT_LOCK_MSG on master site
414 * taken: takes and drop res->spinlock
416 * returns: DLM_NORMAL, DLM_IVLOCKID, DLM_BADARGS,
417 * status from __dlmconvert_master
419 int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data)
421 struct dlm_ctxt *dlm = data;
422 struct dlm_convert_lock *cnv = (struct dlm_convert_lock *)msg->buf;
423 struct dlm_lock_resource *res = NULL;
424 struct list_head *iter;
425 struct dlm_lock *lock = NULL;
426 struct dlm_lockstatus *lksb;
427 enum dlm_status status = DLM_NORMAL;
429 int call_ast = 0, kick_thread = 0, ast_reserved = 0;
431 if (!dlm_grab(dlm)) {
432 dlm_error(DLM_REJECTED);
436 mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
437 "Domain %s not fully joined!\n", dlm->name);
439 if (cnv->namelen > DLM_LOCKID_NAME_MAX) {
440 status = DLM_IVBUFLEN;
445 flags = be32_to_cpu(cnv->flags);
447 if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) ==
448 (LKM_PUT_LVB|LKM_GET_LVB)) {
449 mlog(ML_ERROR, "both PUT and GET lvb specified\n");
450 status = DLM_BADARGS;
454 mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" :
455 (flags & LKM_GET_LVB ? "get lvb" : "none"));
457 status = DLM_IVLOCKID;
458 res = dlm_lookup_lockres(dlm, cnv->name, cnv->namelen);
464 spin_lock(&res->spinlock);
465 list_for_each(iter, &res->granted) {
466 lock = list_entry(iter, struct dlm_lock, list);
467 if (lock->ml.cookie == cnv->cookie &&
468 lock->ml.node == cnv->node_idx) {
474 spin_unlock(&res->spinlock);
476 status = DLM_IVLOCKID;
484 /* see if caller needed to get/put lvb */
485 if (flags & LKM_PUT_LVB) {
486 BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
487 lksb->flags |= DLM_LKSB_PUT_LVB;
488 memcpy(&lksb->lvb[0], &cnv->lvb[0], DLM_LVB_LEN);
489 } else if (flags & LKM_GET_LVB) {
490 BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
491 lksb->flags |= DLM_LKSB_GET_LVB;
494 spin_lock(&res->spinlock);
495 status = __dlm_lockres_state_to_status(res);
496 if (status == DLM_NORMAL) {
497 __dlm_lockres_reserve_ast(res);
499 res->state |= DLM_LOCK_RES_IN_PROGRESS;
500 status = __dlmconvert_master(dlm, res, lock, flags,
502 &call_ast, &kick_thread);
503 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
505 spin_unlock(&res->spinlock);
507 if (status != DLM_NORMAL) {
508 if (status != DLM_NOTQUEUED)
510 lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB);
515 mlog(ML_ERROR, "did not find lock to convert on grant queue! "
516 "cookie=%"MLFu64"\n",
521 /* either queue the ast or release it, if reserved */
523 dlm_queue_ast(dlm, lock);
524 else if (ast_reserved)
525 dlm_lockres_release_ast(dlm, res);
528 dlm_kick_thread(dlm, res);
531 dlm_lockres_put(res);