fs/ocfs2/dlm/dlmast.c

   1 /* -*- 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/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>
  40
  41
  42 #include "cluster/heartbeat.h"
  43 #include "cluster/nodemanager.h"
  44 #include "cluster/tcp.h"
  45
  46 #include "dlmapi.h"
  47 #include "dlmcommon.h"
  48
  49 #define MLOG_MASK_PREFIX ML_DLM
  50 #include "cluster/masklog.h"
  51
  52 static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
  53                            struct dlm_lock *lock);
  54 static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
  55
  56 /* Should be called as an ast gets queued to see if the new
  57  * lock level will obsolete a pending bast.
  58  * For example, if dlm_thread queued a bast for an EX lock that
  59  * was blocking another EX, but before sending the bast the
  60  * lock owner downconverted to NL, the bast is now obsolete.
  61  * Only the ast should be sent.
  62  * This is needed because the lock and convert paths can queue
  63  * asts out-of-band (not waiting for dlm_thread) in order to
  64  * allow for LKM_NOQUEUE to get immediate responses. */
  65 static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
  66 {
  67         assert_spin_locked(&dlm->ast_lock);
  68         assert_spin_locked(&lock->spinlock);
  69
  70         if (lock->ml.highest_blocked == LKM_IVMODE)
  71                 return 0;
  72         BUG_ON(lock->ml.highest_blocked == LKM_NLMODE);
  73
  74         if (lock->bast_pending &&
  75             list_empty(&lock->bast_list))
  76                 /* old bast already sent, ok */
  77                 return 0;
  78
  79         if (lock->ml.type == LKM_EXMODE)
  80                 /* EX blocks anything left, any bast still valid */
  81                 return 0;
  82         else if (lock->ml.type == LKM_NLMODE)
  83                 /* NL blocks nothing, no reason to send any bast, cancel it */
  84                 return 1;
  85         else if (lock->ml.highest_blocked != LKM_EXMODE)
  86                 /* PR only blocks EX */
  87                 return 1;
  88
  89         return 0;
  90 }
  91
  92 static void __dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
  93 {
  94         mlog_entry_void();
  95
  96         BUG_ON(!dlm);
  97         BUG_ON(!lock);
  98
  99         assert_spin_locked(&dlm->ast_lock);
 100         if (!list_empty(&lock->ast_list)) {
 101                 mlog(ML_ERROR, "ast list not empty!!  pending=%d, newlevel=%d\n",
 102                      lock->ast_pending, lock->ml.type);
 103                 BUG();
 104         }
 105         if (lock->ast_pending)
 106                 mlog(0, "lock has an ast getting flushed right now\n");
 107
 108         /* putting lock on list, add a ref */
 109         dlm_lock_get(lock);
 110         spin_lock(&lock->spinlock);
 111
 112         /* check to see if this ast obsoletes the bast */
 113         if (dlm_should_cancel_bast(dlm, lock)) {
 114                 struct dlm_lock_resource *res = lock->lockres;
 115                 mlog(0, "%s: cancelling bast for %.*s\n",
 116                      dlm->name, res->lockname.len, res->lockname.name);
 117                 lock->bast_pending = 0;
 118                 list_del_init(&lock->bast_list);
 119                 lock->ml.highest_blocked = LKM_IVMODE;
 120                 /* removing lock from list, remove a ref.  guaranteed
 121                  * this won't be the last ref because of the get above,
 122                  * so res->spinlock will not be taken here */
 123                 dlm_lock_put(lock);
 124                 /* free up the reserved bast that we are cancelling.
 125                  * guaranteed that this will not be the last reserved
 126                  * ast because *both* an ast and a bast were reserved
 127                  * to get to this point.  the res->spinlock will not be
 128                  * taken here */
 129                 dlm_lockres_release_ast(dlm, res);
 130         }
 131         list_add_tail(&lock->ast_list, &dlm->pending_asts);
 132         lock->ast_pending = 1;
 133         spin_unlock(&lock->spinlock);
 134 }
 135
 136 void dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
 137 {
 138         mlog_entry_void();
 139
 140         BUG_ON(!dlm);
 141         BUG_ON(!lock);
 142
 143         spin_lock(&dlm->ast_lock);
 144         __dlm_queue_ast(dlm, lock);
 145         spin_unlock(&dlm->ast_lock);
 146 }
 147
 148
 149 static void __dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
 150 {
 151         mlog_entry_void();
 152
 153         BUG_ON(!dlm);
 154         BUG_ON(!lock);
 155         assert_spin_locked(&dlm->ast_lock);
 156
 157         BUG_ON(!list_empty(&lock->bast_list));
 158         if (lock->bast_pending)
 159                 mlog(0, "lock has a bast getting flushed right now\n");
 160
 161         /* putting lock on list, add a ref */
 162         dlm_lock_get(lock);
 163         spin_lock(&lock->spinlock);
 164         list_add_tail(&lock->bast_list, &dlm->pending_basts);
 165         lock->bast_pending = 1;
 166         spin_unlock(&lock->spinlock);
 167 }
 168
 169 void dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
 170 {
 171         mlog_entry_void();
 172
 173         BUG_ON(!dlm);
 174         BUG_ON(!lock);
 175
 176         spin_lock(&dlm->ast_lock);
 177         __dlm_queue_bast(dlm, lock);
 178         spin_unlock(&dlm->ast_lock);
 179 }
 180
 181 static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 182                            struct dlm_lock *lock)
 183 {
 184         struct dlm_lockstatus *lksb = lock->lksb;
 185         BUG_ON(!lksb);
 186
 187         /* only updates if this node masters the lockres */
 188         if (res->owner == dlm->node_num) {
 189
 190                 spin_lock(&res->spinlock);
 191                 /* check the lksb flags for the direction */
 192                 if (lksb->flags & DLM_LKSB_GET_LVB) {
 193                         mlog(0, "getting lvb from lockres for %s node\n",
 194                                   lock->ml.node == dlm->node_num ? "master" :
 195                                   "remote");
 196                         memcpy(lksb->lvb, res->lvb, DLM_LVB_LEN);
 197                 }
 198                 /* Do nothing for lvb put requests - they should be done in
 199                  * place when the lock is downconverted - otherwise we risk
 200                  * racing gets and puts which could result in old lvb data
 201                  * being propagated. We leave the put flag set and clear it
 202                  * here. In the future we might want to clear it at the time
 203                  * the put is actually done.
 204                  */
 205                 spin_unlock(&res->spinlock);
 206         }
 207
 208         /* reset any lvb flags on the lksb */
 209         lksb->flags &= ~(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB);
 210 }
 211
 212 void dlm_do_local_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 213                       struct dlm_lock *lock)
 214 {
 215         dlm_astlockfunc_t *fn;
 216         struct dlm_lockstatus *lksb;
 217
 218         mlog_entry_void();
 219
 220         lksb = lock->lksb;
 221         fn = lock->ast;
 222         BUG_ON(lock->ml.node != dlm->node_num);
 223
 224         dlm_update_lvb(dlm, res, lock);
 225         (*fn)(lock->astdata);
 226 }
 227
 228
 229 int dlm_do_remote_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 230                       struct dlm_lock *lock)
 231 {
 232         int ret;
 233         struct dlm_lockstatus *lksb;
 234         int lksbflags;
 235
 236         mlog_entry_void();
 237
 238         lksb = lock->lksb;
 239         BUG_ON(lock->ml.node == dlm->node_num);
 240
 241         lksbflags = lksb->flags;
 242         dlm_update_lvb(dlm, res, lock);
 243
 244         /* lock request came from another node
 245          * go do the ast over there */
 246         ret = dlm_send_proxy_ast(dlm, res, lock, lksbflags);
 247         return ret;
 248 }
 249
 250 void dlm_do_local_bast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 251                        struct dlm_lock *lock, int blocked_type)
 252 {
 253         dlm_bastlockfunc_t *fn = lock->bast;
 254
 255         mlog_entry_void();
 256         BUG_ON(lock->ml.node != dlm->node_num);
 257
 258         (*fn)(lock->astdata, blocked_type);
 259 }
 260
 261
 262
 263 int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data,
 264                           void **ret_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         u8 node;
 277
 278         if (!dlm_grab(dlm)) {
 279                 dlm_error(DLM_REJECTED);
 280                 return DLM_REJECTED;
 281         }
 282
 283         mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
 284                         "Domain %s not fully joined!\n", dlm->name);
 285
 286         name = past->name;
 287         locklen = past->namelen;
 288         cookie = past->cookie;
 289         flags = be32_to_cpu(past->flags);
 290         node = past->node_idx;
 291
 292         if (locklen > DLM_LOCKID_NAME_MAX) {
 293                 ret = DLM_IVBUFLEN;
 294                 mlog(ML_ERROR, "Invalid name length (%d) in proxy ast "
 295                      "handler!\n", locklen);
 296                 goto leave;
 297         }
 298
 299         if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) ==
 300              (LKM_PUT_LVB|LKM_GET_LVB)) {
 301                 mlog(ML_ERROR, "Both PUT and GET lvb specified, (0x%x)\n",
 302                      flags);
 303                 ret = DLM_BADARGS;
 304                 goto leave;
 305         }
 306
 307         mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" :
 308                   (flags & LKM_GET_LVB ? "get lvb" : "none"));
 309
 310         mlog(0, "type=%d, blocked_type=%d\n", past->type, past->blocked_type);
 311
 312         if (past->type != DLM_AST &&
 313             past->type != DLM_BAST) {
 314                 mlog(ML_ERROR, "Unknown ast type! %d, cookie=%u:%llu"
 315                      "name=%.*s, node=%u\n", past->type,
 316                      dlm_get_lock_cookie_node(be64_to_cpu(cookie)),
 317                      dlm_get_lock_cookie_seq(be64_to_cpu(cookie)),
 318                      locklen, name, node);
 319                 ret = DLM_IVLOCKID;
 320                 goto leave;
 321         }
 322
 323         res = dlm_lookup_lockres(dlm, name, locklen);
 324         if (!res) {
 325                 mlog(0, "Got %sast for unknown lockres! cookie=%u:%llu, "
 326                      "name=%.*s, node=%u\n", (past->type == DLM_AST ? "" : "b"),
 327                      dlm_get_lock_cookie_node(be64_to_cpu(cookie)),
 328                      dlm_get_lock_cookie_seq(be64_to_cpu(cookie)),
 329                      locklen, name, node);
 330                 ret = DLM_IVLOCKID;
 331                 goto leave;
 332         }
 333
 334         /* cannot get a proxy ast message if this node owns it */
 335         BUG_ON(res->owner == dlm->node_num);
 336
 337         mlog(0, "lockres %.*s\n", res->lockname.len, res->lockname.name);
 338
 339         spin_lock(&res->spinlock);
 340         if (res->state & DLM_LOCK_RES_RECOVERING) {
 341                 mlog(0, "Responding with DLM_RECOVERING!\n");
 342                 ret = DLM_RECOVERING;
 343                 goto unlock_out;
 344         }
 345         if (res->state & DLM_LOCK_RES_MIGRATING) {
 346                 mlog(0, "Responding with DLM_MIGRATING!\n");
 347                 ret = DLM_MIGRATING;
 348                 goto unlock_out;
 349         }
 350         /* try convert queue for both ast/bast */
 351         head = &res->converting;
 352         lock = NULL;
 353         list_for_each(iter, head) {
 354                 lock = list_entry (iter, struct dlm_lock, list);
 355                 if (lock->ml.cookie == cookie)
 356                         goto do_ast;
 357         }
 358
 359         /* if not on convert, try blocked for ast, granted for bast */
 360         if (past->type == DLM_AST)
 361                 head = &res->blocked;
 362         else
 363                 head = &res->granted;
 364
 365         list_for_each(iter, head) {
 366                 lock = list_entry (iter, struct dlm_lock, list);
 367                 if (lock->ml.cookie == cookie)
 368                         goto do_ast;
 369         }
 370
 371         mlog(0, "Got %sast for unknown lock! cookie=%u:%llu, name=%.*s, "
 372              "node=%u\n", past->type == DLM_AST ? "" : "b",
 373              dlm_get_lock_cookie_node(be64_to_cpu(cookie)),
 374              dlm_get_lock_cookie_seq(be64_to_cpu(cookie)),
 375              locklen, name, node);
 376
 377         ret = DLM_NORMAL;
 378 unlock_out:
 379         spin_unlock(&res->spinlock);
 380         goto leave;
 381
 382 do_ast:
 383         ret = DLM_NORMAL;
 384         if (past->type == DLM_AST) {
 385                 /* do not alter lock refcount.  switching lists. */
 386                 list_move_tail(&lock->list, &res->granted);
 387                 mlog(0, "ast: Adding to granted list... type=%d, "
 388                      "convert_type=%d\n", lock->ml.type, lock->ml.convert_type);
 389                 if (lock->ml.convert_type != LKM_IVMODE) {
 390                         lock->ml.type = lock->ml.convert_type;
 391                         lock->ml.convert_type = LKM_IVMODE;
 392                 } else {
 393                         // should already be there....
 394                 }
 395
 396                 lock->lksb->status = DLM_NORMAL;
 397
 398                 /* if we requested the lvb, fetch it into our lksb now */
 399                 if (flags & LKM_GET_LVB) {
 400                         BUG_ON(!(lock->lksb->flags & DLM_LKSB_GET_LVB));
 401                         memcpy(lock->lksb->lvb, past->lvb, DLM_LVB_LEN);
 402                 }
 403         }
 404         spin_unlock(&res->spinlock);
 405
 406         if (past->type == DLM_AST)
 407                 dlm_do_local_ast(dlm, res, lock);
 408         else
 409                 dlm_do_local_bast(dlm, res, lock, past->blocked_type);
 410
 411 leave:
 412         if (res)
 413                 dlm_lockres_put(res);
 414
 415         dlm_put(dlm);
 416         return ret;
 417 }
 418
 419
 420
 421 int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 422                            struct dlm_lock *lock, int msg_type,
 423                            int blocked_type, int flags)
 424 {
 425         int ret = 0;
 426         struct dlm_proxy_ast past;
 427         struct kvec vec[2];
 428         size_t veclen = 1;
 429         int status;
 430
 431         mlog_entry("res %.*s, to=%u, type=%d, blocked_type=%d\n",
 432                    res->lockname.len, res->lockname.name, lock->ml.node,
 433                    msg_type, blocked_type);
 434
 435         memset(&past, 0, sizeof(struct dlm_proxy_ast));
 436         past.node_idx = dlm->node_num;
 437         past.type = msg_type;
 438         past.blocked_type = blocked_type;
 439         past.namelen = res->lockname.len;
 440         memcpy(past.name, res->lockname.name, past.namelen);
 441         past.cookie = lock->ml.cookie;
 442
 443         vec[0].iov_len = sizeof(struct dlm_proxy_ast);
 444         vec[0].iov_base = &past;
 445         if (flags & DLM_LKSB_GET_LVB) {
 446                 mlog(0, "returning requested LVB data\n");
 447                 be32_add_cpu(&past.flags, LKM_GET_LVB);
 448                 vec[1].iov_len = DLM_LVB_LEN;
 449                 vec[1].iov_base = lock->lksb->lvb;
 450                 veclen++;
 451         }
 452
 453         ret = o2net_send_message_vec(DLM_PROXY_AST_MSG, dlm->key, vec, veclen,
 454                                      lock->ml.node, &status);
 455         if (ret < 0)
 456                 mlog_errno(ret);
 457         else {
 458                 if (status == DLM_RECOVERING) {
 459                         mlog(ML_ERROR, "sent AST to node %u, it thinks this "
 460                              "node is dead!\n", lock->ml.node);
 461                         BUG();
 462                 } else if (status == DLM_MIGRATING) {
 463                         mlog(ML_ERROR, "sent AST to node %u, it returned "
 464                              "DLM_MIGRATING!\n", lock->ml.node);
 465                         BUG();
 466                 } else if (status != DLM_NORMAL && status != DLM_IVLOCKID) {
 467                         mlog(ML_ERROR, "AST to node %u returned %d!\n",
 468                              lock->ml.node, status);
 469                         /* ignore it */
 470                 }
 471                 ret = 0;
 472         }
 473         return ret;
 474 }