blob: 8d17d28ef91c127122b9724b7b22a5eaa2889593 [file] [log] [blame]
Kurt Hackel6714d8e2005-12-15 14:31:23 -08001/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmast.c
5 *
6 * AST and BAST functionality for local and remote nodes
7 *
8 * Copyright (C) 2004 Oracle. All rights reserved.
9 *
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.
14 *
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.
19 *
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.
24 *
25 */
26
27
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/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
42
43#include "cluster/heartbeat.h"
44#include "cluster/nodemanager.h"
45#include "cluster/tcp.h"
46#include "cluster/endian.h"
47
48#include "dlmapi.h"
49#include "dlmcommon.h"
50
51#define MLOG_MASK_PREFIX ML_DLM
52#include "cluster/masklog.h"
53
54static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
55 struct dlm_lock *lock);
56static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
57
58/* Should be called as an ast gets queued to see if the new
59 * lock level will obsolete a pending bast.
60 * For example, if dlm_thread queued a bast for an EX lock that
61 * was blocking another EX, but before sending the bast the
62 * lock owner downconverted to NL, the bast is now obsolete.
63 * Only the ast should be sent.
64 * This is needed because the lock and convert paths can queue
65 * asts out-of-band (not waiting for dlm_thread) in order to
66 * allow for LKM_NOQUEUE to get immediate responses. */
67static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
68{
69 assert_spin_locked(&dlm->ast_lock);
70 assert_spin_locked(&lock->spinlock);
71
72 if (lock->ml.highest_blocked == LKM_IVMODE)
73 return 0;
74 BUG_ON(lock->ml.highest_blocked == LKM_NLMODE);
75
76 if (lock->bast_pending &&
77 list_empty(&lock->bast_list))
78 /* old bast already sent, ok */
79 return 0;
80
81 if (lock->ml.type == LKM_EXMODE)
82 /* EX blocks anything left, any bast still valid */
83 return 0;
84 else if (lock->ml.type == LKM_NLMODE)
85 /* NL blocks nothing, no reason to send any bast, cancel it */
86 return 1;
87 else if (lock->ml.highest_blocked != LKM_EXMODE)
88 /* PR only blocks EX */
89 return 1;
90
91 return 0;
92}
93
94static void __dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
95{
96 mlog_entry_void();
97
98 BUG_ON(!dlm);
99 BUG_ON(!lock);
100
101 assert_spin_locked(&dlm->ast_lock);
102 if (!list_empty(&lock->ast_list)) {
103 mlog(ML_ERROR, "ast list not empty!! pending=%d, newlevel=%d\n",
104 lock->ast_pending, lock->ml.type);
105 BUG();
106 }
107 BUG_ON(!list_empty(&lock->ast_list));
108 if (lock->ast_pending)
109 mlog(0, "lock has an ast getting flushed right now\n");
110
111 /* putting lock on list, add a ref */
112 dlm_lock_get(lock);
113 spin_lock(&lock->spinlock);
114
115 /* check to see if this ast obsoletes the bast */
116 if (dlm_should_cancel_bast(dlm, lock)) {
117 struct dlm_lock_resource *res = lock->lockres;
118 mlog(0, "%s: cancelling bast for %.*s\n",
119 dlm->name, res->lockname.len, res->lockname.name);
120 lock->bast_pending = 0;
121 list_del_init(&lock->bast_list);
122 lock->ml.highest_blocked = LKM_IVMODE;
123 /* removing lock from list, remove a ref. guaranteed
124 * this won't be the last ref because of the get above,
125 * so res->spinlock will not be taken here */
126 dlm_lock_put(lock);
127 /* free up the reserved bast that we are cancelling.
128 * guaranteed that this will not be the last reserved
129 * ast because *both* an ast and a bast were reserved
130 * to get to this point. the res->spinlock will not be
131 * taken here */
132 dlm_lockres_release_ast(dlm, res);
133 }
134 list_add_tail(&lock->ast_list, &dlm->pending_asts);
135 lock->ast_pending = 1;
136 spin_unlock(&lock->spinlock);
137}
138
139void dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
140{
141 mlog_entry_void();
142
143 BUG_ON(!dlm);
144 BUG_ON(!lock);
145
146 spin_lock(&dlm->ast_lock);
147 __dlm_queue_ast(dlm, lock);
148 spin_unlock(&dlm->ast_lock);
149}
150
151
152static void __dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
153{
154 mlog_entry_void();
155
156 BUG_ON(!dlm);
157 BUG_ON(!lock);
158 assert_spin_locked(&dlm->ast_lock);
159
160 BUG_ON(!list_empty(&lock->bast_list));
161 if (lock->bast_pending)
162 mlog(0, "lock has a bast getting flushed right now\n");
163
164 /* putting lock on list, add a ref */
165 dlm_lock_get(lock);
166 spin_lock(&lock->spinlock);
167 list_add_tail(&lock->bast_list, &dlm->pending_basts);
168 lock->bast_pending = 1;
169 spin_unlock(&lock->spinlock);
170}
171
172void dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
173{
174 mlog_entry_void();
175
176 BUG_ON(!dlm);
177 BUG_ON(!lock);
178
179 spin_lock(&dlm->ast_lock);
180 __dlm_queue_bast(dlm, lock);
181 spin_unlock(&dlm->ast_lock);
182}
183
184static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
185 struct dlm_lock *lock)
186{
187 struct dlm_lockstatus *lksb = lock->lksb;
188 BUG_ON(!lksb);
189
190 /* only updates if this node masters the lockres */
191 if (res->owner == dlm->node_num) {
192
193 spin_lock(&res->spinlock);
194 /* check the lksb flags for the direction */
195 if (lksb->flags & DLM_LKSB_GET_LVB) {
196 mlog(0, "getting lvb from lockres for %s node\n",
197 lock->ml.node == dlm->node_num ? "master" :
198 "remote");
199 memcpy(lksb->lvb, res->lvb, DLM_LVB_LEN);
200 } else if (lksb->flags & DLM_LKSB_PUT_LVB) {
201 mlog(0, "setting lvb from lockres for %s node\n",
202 lock->ml.node == dlm->node_num ? "master" :
203 "remote");
204 memcpy(res->lvb, lksb->lvb, DLM_LVB_LEN);
205 }
206 spin_unlock(&res->spinlock);
207 }
208
209 /* reset any lvb flags on the lksb */
210 lksb->flags &= ~(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB);
211}
212
213void dlm_do_local_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
214 struct dlm_lock *lock)
215{
216 dlm_astlockfunc_t *fn;
217 struct dlm_lockstatus *lksb;
218
219 mlog_entry_void();
220
221 lksb = lock->lksb;
222 fn = lock->ast;
223 BUG_ON(lock->ml.node != dlm->node_num);
224
225 dlm_update_lvb(dlm, res, lock);
226 (*fn)(lock->astdata);
227}
228
229
230int dlm_do_remote_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
231 struct dlm_lock *lock)
232{
233 int ret;
234 struct dlm_lockstatus *lksb;
235 int lksbflags;
236
237 mlog_entry_void();
238
239 lksb = lock->lksb;
240 BUG_ON(lock->ml.node == dlm->node_num);
241
242 lksbflags = lksb->flags;
243 dlm_update_lvb(dlm, res, lock);
244
245 /* lock request came from another node
246 * go do the ast over there */
247 ret = dlm_send_proxy_ast(dlm, res, lock, lksbflags);
248 return ret;
249}
250
251void dlm_do_local_bast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
252 struct dlm_lock *lock, int blocked_type)
253{
254 dlm_bastlockfunc_t *fn = lock->bast;
255
256 mlog_entry_void();
257 BUG_ON(lock->ml.node != dlm->node_num);
258
259 (*fn)(lock->astdata, blocked_type);
260}
261
262
263
264int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data)
265{
266 int ret;
267 unsigned int locklen;
268 struct dlm_ctxt *dlm = data;
269 struct dlm_lock_resource *res = NULL;
270 struct dlm_lock *lock = NULL;
271 struct dlm_proxy_ast *past = (struct dlm_proxy_ast *) msg->buf;
272 char *name;
273 struct list_head *iter, *head=NULL;
274 u64 cookie;
275 u32 flags;
276
277 if (!dlm_grab(dlm)) {
278 dlm_error(DLM_REJECTED);
279 return DLM_REJECTED;
280 }
281
282 mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
283 "Domain %s not fully joined!\n", dlm->name);
284
285 name = past->name;
286 locklen = past->namelen;
287 cookie = be64_to_cpu(past->cookie);
288 flags = be32_to_cpu(past->flags);
289
290 if (locklen > DLM_LOCKID_NAME_MAX) {
291 ret = DLM_IVBUFLEN;
292 mlog(ML_ERROR, "Invalid name length in proxy ast handler!\n");
293 goto leave;
294 }
295
296 if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) ==
297 (LKM_PUT_LVB|LKM_GET_LVB)) {
298 mlog(ML_ERROR, "both PUT and GET lvb specified\n");
299 ret = DLM_BADARGS;
300 goto leave;
301 }
302
303 mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" :
304 (flags & LKM_GET_LVB ? "get lvb" : "none"));
305
306 mlog(0, "type=%d, blocked_type=%d\n", past->type, past->blocked_type);
307
308 if (past->type != DLM_AST &&
309 past->type != DLM_BAST) {
310 mlog(ML_ERROR, "Unknown ast type! %d, cookie=%"MLFu64", "
311 "name=%.*s\n", past->type, cookie, locklen, name);
312 ret = DLM_IVLOCKID;
313 goto leave;
314 }
315
316 res = dlm_lookup_lockres(dlm, name, locklen);
317 if (!res) {
318 mlog(ML_ERROR, "got %sast for unknown lockres! "
319 "cookie=%"MLFu64", name=%.*s, namelen=%u\n",
320 past->type == DLM_AST ? "" : "b",
321 cookie, locklen, name, locklen);
322 ret = DLM_IVLOCKID;
323 goto leave;
324 }
325
326 /* cannot get a proxy ast message if this node owns it */
327 BUG_ON(res->owner == dlm->node_num);
328
329 mlog(0, "lockres %.*s\n", res->lockname.len, res->lockname.name);
330
331 spin_lock(&res->spinlock);
332 if (res->state & DLM_LOCK_RES_RECOVERING) {
333 mlog(0, "responding with DLM_RECOVERING!\n");
334 ret = DLM_RECOVERING;
335 goto unlock_out;
336 }
337 if (res->state & DLM_LOCK_RES_MIGRATING) {
338 mlog(0, "responding with DLM_MIGRATING!\n");
339 ret = DLM_MIGRATING;
340 goto unlock_out;
341 }
342 /* try convert queue for both ast/bast */
343 head = &res->converting;
344 lock = NULL;
345 list_for_each(iter, head) {
346 lock = list_entry (iter, struct dlm_lock, list);
347 if (be64_to_cpu(lock->ml.cookie) == cookie)
348 goto do_ast;
349 }
350
351 /* if not on convert, try blocked for ast, granted for bast */
352 if (past->type == DLM_AST)
353 head = &res->blocked;
354 else
355 head = &res->granted;
356
357 list_for_each(iter, head) {
358 lock = list_entry (iter, struct dlm_lock, list);
359 if (be64_to_cpu(lock->ml.cookie) == cookie)
360 goto do_ast;
361 }
362
363 mlog(ML_ERROR, "got %sast for unknown lock! cookie=%"MLFu64", "
364 "name=%.*s, namelen=%u\n",
365 past->type == DLM_AST ? "" : "b", cookie, locklen, name, locklen);
366
367 ret = DLM_NORMAL;
368unlock_out:
369 spin_unlock(&res->spinlock);
370 goto leave;
371
372do_ast:
373 ret = DLM_NORMAL;
374 if (past->type == DLM_AST) {
375 /* do not alter lock refcount. switching lists. */
376 list_del_init(&lock->list);
377 list_add_tail(&lock->list, &res->granted);
378 mlog(0, "ast: adding to granted list... type=%d, "
379 "convert_type=%d\n", lock->ml.type, lock->ml.convert_type);
380 if (lock->ml.convert_type != LKM_IVMODE) {
381 lock->ml.type = lock->ml.convert_type;
382 lock->ml.convert_type = LKM_IVMODE;
383 } else {
384 // should already be there....
385 }
386
387 lock->lksb->status = DLM_NORMAL;
388
389 /* if we requested the lvb, fetch it into our lksb now */
390 if (flags & LKM_GET_LVB) {
391 BUG_ON(!(lock->lksb->flags & DLM_LKSB_GET_LVB));
392 memcpy(lock->lksb->lvb, past->lvb, DLM_LVB_LEN);
393 }
394 }
395 spin_unlock(&res->spinlock);
396
397 if (past->type == DLM_AST)
398 dlm_do_local_ast(dlm, res, lock);
399 else
400 dlm_do_local_bast(dlm, res, lock, past->blocked_type);
401
402leave:
403
404 if (res)
405 dlm_lockres_put(res);
406
407 dlm_put(dlm);
408 return ret;
409}
410
411
412
413int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
414 struct dlm_lock *lock, int msg_type,
415 int blocked_type, int flags)
416{
417 int ret = 0;
418 struct dlm_proxy_ast past;
419 struct kvec vec[2];
420 size_t veclen = 1;
421 int status;
422
423 mlog_entry("res %.*s, to=%u, type=%d, blocked_type=%d\n",
424 res->lockname.len, res->lockname.name, lock->ml.node,
425 msg_type, blocked_type);
426
427 memset(&past, 0, sizeof(struct dlm_proxy_ast));
428 past.node_idx = dlm->node_num;
429 past.type = msg_type;
430 past.blocked_type = blocked_type;
431 past.namelen = res->lockname.len;
432 memcpy(past.name, res->lockname.name, past.namelen);
433 past.cookie = lock->ml.cookie;
434
435 vec[0].iov_len = sizeof(struct dlm_proxy_ast);
436 vec[0].iov_base = &past;
437 if (flags & DLM_LKSB_GET_LVB) {
438 mlog(0, "returning requested LVB data\n");
439 be32_add_cpu(&past.flags, LKM_GET_LVB);
440 vec[1].iov_len = DLM_LVB_LEN;
441 vec[1].iov_base = lock->lksb->lvb;
442 veclen++;
443 }
444
445 ret = o2net_send_message_vec(DLM_PROXY_AST_MSG, dlm->key, vec, veclen,
446 lock->ml.node, &status);
447 if (ret < 0)
448 mlog_errno(ret);
449 else {
450 if (status == DLM_RECOVERING) {
451 mlog(ML_ERROR, "sent AST to node %u, it thinks this "
452 "node is dead!\n", lock->ml.node);
453 BUG();
454 } else if (status == DLM_MIGRATING) {
455 mlog(ML_ERROR, "sent AST to node %u, it returned "
456 "DLM_MIGRATING!\n", lock->ml.node);
457 BUG();
458 } else if (status != DLM_NORMAL) {
459 mlog(ML_ERROR, "AST to node %u returned %d!\n",
460 lock->ml.node, status);
461 /* ignore it */
462 }
463 ret = 0;
464 }
465 return ret;
466}