1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * AST and BAST functionality for local and remote nodes
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 #define MLOG_MASK_PREFIX ML_DLM
51 #include "cluster/masklog.h"
53 static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
54 struct dlm_lock *lock);
55 static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
57 /* Should be called as an ast gets queued to see if the new
58 * lock level will obsolete a pending bast.
59 * For example, if dlm_thread queued a bast for an EX lock that
60 * was blocking another EX, but before sending the bast the
61 * lock owner downconverted to NL, the bast is now obsolete.
62 * Only the ast should be sent.
63 * This is needed because the lock and convert paths can queue
64 * asts out-of-band (not waiting for dlm_thread) in order to
65 * allow for LKM_NOQUEUE to get immediate responses. */
66 static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
68 assert_spin_locked(&dlm->ast_lock);
69 assert_spin_locked(&lock->spinlock);
71 if (lock->ml.highest_blocked == LKM_IVMODE)
73 BUG_ON(lock->ml.highest_blocked == LKM_NLMODE);
75 if (lock->bast_pending &&
76 list_empty(&lock->bast_list))
77 /* old bast already sent, ok */
80 if (lock->ml.type == LKM_EXMODE)
81 /* EX blocks anything left, any bast still valid */
83 else if (lock->ml.type == LKM_NLMODE)
84 /* NL blocks nothing, no reason to send any bast, cancel it */
86 else if (lock->ml.highest_blocked != LKM_EXMODE)
87 /* PR only blocks EX */
93 static void __dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
100 assert_spin_locked(&dlm->ast_lock);
101 if (!list_empty(&lock->ast_list)) {
102 mlog(ML_ERROR, "ast list not empty!! pending=%d, newlevel=%d\n",
103 lock->ast_pending, lock->ml.type);
106 BUG_ON(!list_empty(&lock->ast_list));
107 if (lock->ast_pending)
108 mlog(0, "lock has an ast getting flushed right now\n");
110 /* putting lock on list, add a ref */
112 spin_lock(&lock->spinlock);
114 /* check to see if this ast obsoletes the bast */
115 if (dlm_should_cancel_bast(dlm, lock)) {
116 struct dlm_lock_resource *res = lock->lockres;
117 mlog(0, "%s: cancelling bast for %.*s\n",
118 dlm->name, res->lockname.len, res->lockname.name);
119 lock->bast_pending = 0;
120 list_del_init(&lock->bast_list);
121 lock->ml.highest_blocked = LKM_IVMODE;
122 /* removing lock from list, remove a ref. guaranteed
123 * this won't be the last ref because of the get above,
124 * so res->spinlock will not be taken here */
126 /* free up the reserved bast that we are cancelling.
127 * guaranteed that this will not be the last reserved
128 * ast because *both* an ast and a bast were reserved
129 * to get to this point. the res->spinlock will not be
131 dlm_lockres_release_ast(dlm, res);
133 list_add_tail(&lock->ast_list, &dlm->pending_asts);
134 lock->ast_pending = 1;
135 spin_unlock(&lock->spinlock);
138 void dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
145 spin_lock(&dlm->ast_lock);
146 __dlm_queue_ast(dlm, lock);
147 spin_unlock(&dlm->ast_lock);
151 static void __dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
157 assert_spin_locked(&dlm->ast_lock);
159 BUG_ON(!list_empty(&lock->bast_list));
160 if (lock->bast_pending)
161 mlog(0, "lock has a bast getting flushed right now\n");
163 /* putting lock on list, add a ref */
165 spin_lock(&lock->spinlock);
166 list_add_tail(&lock->bast_list, &dlm->pending_basts);
167 lock->bast_pending = 1;
168 spin_unlock(&lock->spinlock);
171 void dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
178 spin_lock(&dlm->ast_lock);
179 __dlm_queue_bast(dlm, lock);
180 spin_unlock(&dlm->ast_lock);
183 static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
184 struct dlm_lock *lock)
186 struct dlm_lockstatus *lksb = lock->lksb;
189 /* only updates if this node masters the lockres */
190 if (res->owner == dlm->node_num) {
192 spin_lock(&res->spinlock);
193 /* check the lksb flags for the direction */
194 if (lksb->flags & DLM_LKSB_GET_LVB) {
195 mlog(0, "getting lvb from lockres for %s node\n",
196 lock->ml.node == dlm->node_num ? "master" :
198 memcpy(lksb->lvb, res->lvb, DLM_LVB_LEN);
200 /* Do nothing for lvb put requests - they should be done in
201 * place when the lock is downconverted - otherwise we risk
202 * racing gets and puts which could result in old lvb data
203 * being propagated. We leave the put flag set and clear it
204 * here. In the future we might want to clear it at the time
205 * the put is actually done.
207 spin_unlock(&res->spinlock);
210 /* reset any lvb flags on the lksb */
211 lksb->flags &= ~(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB);
214 void dlm_do_local_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
215 struct dlm_lock *lock)
217 dlm_astlockfunc_t *fn;
218 struct dlm_lockstatus *lksb;
224 BUG_ON(lock->ml.node != dlm->node_num);
226 dlm_update_lvb(dlm, res, lock);
227 (*fn)(lock->astdata);
231 int dlm_do_remote_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
232 struct dlm_lock *lock)
235 struct dlm_lockstatus *lksb;
241 BUG_ON(lock->ml.node == dlm->node_num);
243 lksbflags = lksb->flags;
244 dlm_update_lvb(dlm, res, lock);
246 /* lock request came from another node
247 * go do the ast over there */
248 ret = dlm_send_proxy_ast(dlm, res, lock, lksbflags);
252 void dlm_do_local_bast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
253 struct dlm_lock *lock, int blocked_type)
255 dlm_bastlockfunc_t *fn = lock->bast;
258 BUG_ON(lock->ml.node != dlm->node_num);
260 (*fn)(lock->astdata, blocked_type);
265 int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data,
269 unsigned int locklen;
270 struct dlm_ctxt *dlm = data;
271 struct dlm_lock_resource *res = NULL;
272 struct dlm_lock *lock = NULL;
273 struct dlm_proxy_ast *past = (struct dlm_proxy_ast *) msg->buf;
275 struct list_head *iter, *head=NULL;
280 if (!dlm_grab(dlm)) {
281 dlm_error(DLM_REJECTED);
285 mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
286 "Domain %s not fully joined!\n", dlm->name);
289 locklen = past->namelen;
290 cookie = past->cookie;
291 flags = be32_to_cpu(past->flags);
292 node = past->node_idx;
294 if (locklen > DLM_LOCKID_NAME_MAX) {
296 mlog(ML_ERROR, "Invalid name length (%d) in proxy ast "
297 "handler!\n", locklen);
301 if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) ==
302 (LKM_PUT_LVB|LKM_GET_LVB)) {
303 mlog(ML_ERROR, "Both PUT and GET lvb specified, (0x%x)\n",
309 mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" :
310 (flags & LKM_GET_LVB ? "get lvb" : "none"));
312 mlog(0, "type=%d, blocked_type=%d\n", past->type, past->blocked_type);
314 if (past->type != DLM_AST &&
315 past->type != DLM_BAST) {
316 mlog(ML_ERROR, "Unknown ast type! %d, cookie=%u:%llu"
317 "name=%.*s, node=%u\n", past->type,
318 dlm_get_lock_cookie_node(be64_to_cpu(cookie)),
319 dlm_get_lock_cookie_seq(be64_to_cpu(cookie)),
320 locklen, name, node);
325 res = dlm_lookup_lockres(dlm, name, locklen);
327 mlog(0, "Got %sast for unknown lockres! cookie=%u:%llu, "
328 "name=%.*s, node=%u\n", (past->type == DLM_AST ? "" : "b"),
329 dlm_get_lock_cookie_node(be64_to_cpu(cookie)),
330 dlm_get_lock_cookie_seq(be64_to_cpu(cookie)),
331 locklen, name, node);
336 /* cannot get a proxy ast message if this node owns it */
337 BUG_ON(res->owner == dlm->node_num);
339 mlog(0, "lockres %.*s\n", res->lockname.len, res->lockname.name);
341 spin_lock(&res->spinlock);
342 if (res->state & DLM_LOCK_RES_RECOVERING) {
343 mlog(0, "Responding with DLM_RECOVERING!\n");
344 ret = DLM_RECOVERING;
347 if (res->state & DLM_LOCK_RES_MIGRATING) {
348 mlog(0, "Responding with DLM_MIGRATING!\n");
352 /* try convert queue for both ast/bast */
353 head = &res->converting;
355 list_for_each(iter, head) {
356 lock = list_entry (iter, struct dlm_lock, list);
357 if (lock->ml.cookie == cookie)
361 /* if not on convert, try blocked for ast, granted for bast */
362 if (past->type == DLM_AST)
363 head = &res->blocked;
365 head = &res->granted;
367 list_for_each(iter, head) {
368 lock = list_entry (iter, struct dlm_lock, list);
369 if (lock->ml.cookie == cookie)
373 mlog(0, "Got %sast for unknown lock! cookie=%u:%llu, name=%.*s, "
374 "node=%u\n", past->type == DLM_AST ? "" : "b",
375 dlm_get_lock_cookie_node(be64_to_cpu(cookie)),
376 dlm_get_lock_cookie_seq(be64_to_cpu(cookie)),
377 locklen, name, node);
381 spin_unlock(&res->spinlock);
386 if (past->type == DLM_AST) {
387 /* do not alter lock refcount. switching lists. */
388 list_move_tail(&lock->list, &res->granted);
389 mlog(0, "ast: Adding to granted list... type=%d, "
390 "convert_type=%d\n", lock->ml.type, lock->ml.convert_type);
391 if (lock->ml.convert_type != LKM_IVMODE) {
392 lock->ml.type = lock->ml.convert_type;
393 lock->ml.convert_type = LKM_IVMODE;
395 // should already be there....
398 lock->lksb->status = DLM_NORMAL;
400 /* if we requested the lvb, fetch it into our lksb now */
401 if (flags & LKM_GET_LVB) {
402 BUG_ON(!(lock->lksb->flags & DLM_LKSB_GET_LVB));
403 memcpy(lock->lksb->lvb, past->lvb, DLM_LVB_LEN);
406 spin_unlock(&res->spinlock);
408 if (past->type == DLM_AST)
409 dlm_do_local_ast(dlm, res, lock);
411 dlm_do_local_bast(dlm, res, lock, past->blocked_type);
415 dlm_lockres_put(res);
423 int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
424 struct dlm_lock *lock, int msg_type,
425 int blocked_type, int flags)
428 struct dlm_proxy_ast past;
433 mlog_entry("res %.*s, to=%u, type=%d, blocked_type=%d\n",
434 res->lockname.len, res->lockname.name, lock->ml.node,
435 msg_type, blocked_type);
437 memset(&past, 0, sizeof(struct dlm_proxy_ast));
438 past.node_idx = dlm->node_num;
439 past.type = msg_type;
440 past.blocked_type = blocked_type;
441 past.namelen = res->lockname.len;
442 memcpy(past.name, res->lockname.name, past.namelen);
443 past.cookie = lock->ml.cookie;
445 vec[0].iov_len = sizeof(struct dlm_proxy_ast);
446 vec[0].iov_base = &past;
447 if (flags & DLM_LKSB_GET_LVB) {
448 mlog(0, "returning requested LVB data\n");
449 be32_add_cpu(&past.flags, LKM_GET_LVB);
450 vec[1].iov_len = DLM_LVB_LEN;
451 vec[1].iov_base = lock->lksb->lvb;
455 ret = o2net_send_message_vec(DLM_PROXY_AST_MSG, dlm->key, vec, veclen,
456 lock->ml.node, &status);
460 if (status == DLM_RECOVERING) {
461 mlog(ML_ERROR, "sent AST to node %u, it thinks this "
462 "node is dead!\n", lock->ml.node);
464 } else if (status == DLM_MIGRATING) {
465 mlog(ML_ERROR, "sent AST to node %u, it returned "
466 "DLM_MIGRATING!\n", lock->ml.node);
468 } else if (status != DLM_NORMAL && status != DLM_IVLOCKID) {
469 mlog(ML_ERROR, "AST to node %u returned %d!\n",
470 lock->ml.node, status);