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/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
  54 static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
  55                            struct dlm_lock *lock);
  56 static 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. */
  67 static 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
  94 static 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
 139 void 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
 152 static 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
 172 void 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
 184 static 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                 }
 201                 /* Do nothing for lvb put requests - they should be done in
 202                  * place when the lock is downconverted - otherwise we risk
 203                  * racing gets and puts which could result in old lvb data
 204                  * being propagated. We leave the put flag set and clear it
 205                  * here. In the future we might want to clear it at the time
 206                  * the put is actually done.
 207                  */
 208                 spin_unlock(&res->spinlock);
 209         }
 210
 211         /* reset any lvb flags on the lksb */
 212         lksb->flags &= ~(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB);
 213 }
 214
 215 void dlm_do_local_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 216                       struct dlm_lock *lock)
 217 {
 218         dlm_astlockfunc_t *fn;
 219         struct dlm_lockstatus *lksb;
 220
 221         mlog_entry_void();
 222
 223         lksb = lock->lksb;
 224         fn = lock->ast;
 225         BUG_ON(lock->ml.node != dlm->node_num);
 226
 227         dlm_update_lvb(dlm, res, lock);
 228         (*fn)(lock->astdata);
 229 }
 230
 231
 232 int dlm_do_remote_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 233                       struct dlm_lock *lock)
 234 {
 235         int ret;
 236         struct dlm_lockstatus *lksb;
 237         int lksbflags;
 238
 239         mlog_entry_void();
 240
 241         lksb = lock->lksb;
 242         BUG_ON(lock->ml.node == dlm->node_num);
 243
 244         lksbflags = lksb->flags;
 245         dlm_update_lvb(dlm, res, lock);
 246
 247         /* lock request came from another node
 248          * go do the ast over there */
 249         ret = dlm_send_proxy_ast(dlm, res, lock, lksbflags);
 250         return ret;
 251 }
 252
 253 void dlm_do_local_bast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 254                        struct dlm_lock *lock, int blocked_type)
 255 {
 256         dlm_bastlockfunc_t *fn = lock->bast;
 257
 258         mlog_entry_void();
 259         BUG_ON(lock->ml.node != dlm->node_num);
 260
 261         (*fn)(lock->astdata, blocked_type);
 262 }
 263
 264
 265
 266 int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data,
 267                           void **ret_data)
 268 {
 269         int ret;
 270         unsigned int locklen;
 271         struct dlm_ctxt *dlm = data;
 272         struct dlm_lock_resource *res = NULL;
 273         struct dlm_lock *lock = NULL;
 274         struct dlm_proxy_ast *past = (struct dlm_proxy_ast *) msg->buf;
 275         char *name;
 276         struct list_head *iter, *head=NULL;
 277         u64 cookie;
 278         u32 flags;
 279
 280         if (!dlm_grab(dlm)) {
 281                 dlm_error(DLM_REJECTED);
 282                 return DLM_REJECTED;
 283         }
 284
 285         mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
 286                         "Domain %s not fully joined!\n", dlm->name);
 287
 288         name = past->name;
 289         locklen = past->namelen;
 290         cookie = be64_to_cpu(past->cookie);
 291         flags = be32_to_cpu(past->flags);
 292
 293         if (locklen > DLM_LOCKID_NAME_MAX) {
 294                 ret = DLM_IVBUFLEN;
 295                 mlog(ML_ERROR, "Invalid name length in proxy ast handler!\n");
 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\n");
 302                 ret = DLM_BADARGS;
 303                 goto leave;
 304         }
 305
 306         mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" :
 307                   (flags & LKM_GET_LVB ? "get lvb" : "none"));
 308
 309         mlog(0, "type=%d, blocked_type=%d\n", past->type, past->blocked_type);
 310
 311         if (past->type != DLM_AST &&
 312             past->type != DLM_BAST) {
 313                 mlog(ML_ERROR, "Unknown ast type! %d, cookie=%u:%llu"
 314                      "name=%.*s\n", past->type,
 315                      dlm_get_lock_cookie_node(be64_to_cpu(cookie)),
 316                      dlm_get_lock_cookie_seq(be64_to_cpu(cookie)),
 317                      locklen, name);
 318                 ret = DLM_IVLOCKID;
 319                 goto leave;
 320         }
 321
 322         res = dlm_lookup_lockres(dlm, name, locklen);
 323         if (!res) {
 324                 mlog(0, "got %sast for unknown lockres! "
 325                      "cookie=%u:%llu, name=%.*s, namelen=%u\n",
 326                      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, locklen);
 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 (be64_to_cpu(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 (be64_to_cpu(lock->ml.cookie) == cookie)
 368                         goto do_ast;
 369         }
 370
 371         mlog(0, "got %sast for unknown lock!  cookie=%u:%llu, "
 372              "name=%.*s, namelen=%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, locklen);
 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
 413         if (res)
 414                 dlm_lockres_put(res);
 415
 416         dlm_put(dlm);
 417         return ret;
 418 }
 419
 420
 421
 422 int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 423                            struct dlm_lock *lock, int msg_type,
 424                            int blocked_type, int flags)
 425 {
 426         int ret = 0;
 427         struct dlm_proxy_ast past;
 428         struct kvec vec[2];
 429         size_t veclen = 1;
 430         int status;
 431
 432         mlog_entry("res %.*s, to=%u, type=%d, blocked_type=%d\n",
 433                    res->lockname.len, res->lockname.name, lock->ml.node,
 434                    msg_type, blocked_type);
 435
 436         memset(&past, 0, sizeof(struct dlm_proxy_ast));
 437         past.node_idx = dlm->node_num;
 438         past.type = msg_type;
 439         past.blocked_type = blocked_type;
 440         past.namelen = res->lockname.len;
 441         memcpy(past.name, res->lockname.name, past.namelen);
 442         past.cookie = lock->ml.cookie;
 443
 444         vec[0].iov_len = sizeof(struct dlm_proxy_ast);
 445         vec[0].iov_base = &past;
 446         if (flags & DLM_LKSB_GET_LVB) {
 447                 mlog(0, "returning requested LVB data\n");
 448                 be32_add_cpu(&past.flags, LKM_GET_LVB);
 449                 vec[1].iov_len = DLM_LVB_LEN;
 450                 vec[1].iov_base = lock->lksb->lvb;
 451                 veclen++;
 452         }
 453
 454         ret = o2net_send_message_vec(DLM_PROXY_AST_MSG, dlm->key, vec, veclen,
 455                                      lock->ml.node, &status);
 456         if (ret < 0)
 457                 mlog_errno(ret);
 458         else {
 459                 if (status == DLM_RECOVERING) {
 460                         mlog(ML_ERROR, "sent AST to node %u, it thinks this "
 461                              "node is dead!\n", lock->ml.node);
 462                         BUG();
 463                 } else if (status == DLM_MIGRATING) {
 464                         mlog(ML_ERROR, "sent AST to node %u, it returned "
 465                              "DLM_MIGRATING!\n", lock->ml.node);
 466                         BUG();
 467                 } else if (status != DLM_NORMAL && status != DLM_IVLOCKID) {
 468                         mlog(ML_ERROR, "AST to node %u returned %d!\n",
 469                              lock->ml.node, status);
 470                         /* ignore it */
 471                 }
 472                 ret = 0;
 473         }
 474         return ret;
 475 }