SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/dlm/rcom.c

   1 /******************************************************************************
   2 *******************************************************************************
   3 **
   4 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
   5 **  Copyright (C) 2005 Red Hat, Inc.  All rights reserved.
   6 **
   7 **  This copyrighted material is made available to anyone wishing to use,
   8 **  modify, copy, or redistribute it subject to the terms and conditions
   9 **  of the GNU General Public License v.2.
  10 **
  11 *******************************************************************************
  12 ******************************************************************************/
  13
  14 #include "dlm_internal.h"
  15 #include "lockspace.h"
  16 #include "member.h"
  17 #include "lowcomms.h"
  18 #include "midcomms.h"
  19 #include "rcom.h"
  20 #include "recover.h"
  21 #include "dir.h"
  22 #include "config.h"
  23 #include "memory.h"
  24 #include "lock.h"
  25 #include "util.h"
  26
  27
  28 static int rcom_response(struct dlm_ls *ls)
  29 {
  30         return test_bit(LSFL_RCOM_READY, &ls->ls_flags);
  31 }
  32
  33 static int create_rcom(struct dlm_ls *ls, int to_nodeid, int type, int len,
  34                        struct dlm_rcom **rc_ret, struct dlm_mhandle **mh_ret)
  35 {
  36         struct dlm_rcom *rc;
  37         struct dlm_mhandle *mh;
  38         char *mb;
  39         int mb_len = sizeof(struct dlm_rcom) + len;
  40
  41         mh = dlm_lowcomms_get_buffer(to_nodeid, mb_len, GFP_KERNEL, &mb);
  42         if (!mh) {
  43                 log_print("create_rcom to %d type %d len %d ENOBUFS",
  44                           to_nodeid, type, len);
  45                 return -ENOBUFS;
  46         }
  47         memset(mb, 0, mb_len);
  48
  49         rc = (struct dlm_rcom *) mb;
  50
  51         rc->rc_header.h_version = (DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
  52         rc->rc_header.h_lockspace = ls->ls_global_id;
  53         rc->rc_header.h_nodeid = dlm_our_nodeid();
  54         rc->rc_header.h_length = mb_len;
  55         rc->rc_header.h_cmd = DLM_RCOM;
  56
  57         rc->rc_type = type;
  58
  59         *mh_ret = mh;
  60         *rc_ret = rc;
  61         return 0;
  62 }
  63
  64 static void send_rcom(struct dlm_ls *ls, struct dlm_mhandle *mh,
  65                       struct dlm_rcom *rc)
  66 {
  67         dlm_rcom_out(rc);
  68         dlm_lowcomms_commit_buffer(mh);
  69 }
  70
  71 /* When replying to a status request, a node also sends back its
  72    configuration values.  The requesting node then checks that the remote
  73    node is configured the same way as itself. */
  74
  75 static void make_config(struct dlm_ls *ls, struct rcom_config *rf)
  76 {
  77         rf->rf_lvblen = ls->ls_lvblen;
  78         rf->rf_lsflags = ls->ls_exflags;
  79 }
  80
  81 static int check_config(struct dlm_ls *ls, struct rcom_config *rf, int nodeid)
  82 {
  83         if (rf->rf_lvblen != ls->ls_lvblen ||
  84             rf->rf_lsflags != ls->ls_exflags) {
  85                 log_error(ls, "config mismatch: %d,%x nodeid %d: %d,%x",
  86                           ls->ls_lvblen, ls->ls_exflags,
  87                           nodeid, rf->rf_lvblen, rf->rf_lsflags);
  88                 return -EINVAL;
  89         }
  90         return 0;
  91 }
  92
  93 int dlm_rcom_status(struct dlm_ls *ls, int nodeid)
  94 {
  95         struct dlm_rcom *rc;
  96         struct dlm_mhandle *mh;
  97         int error = 0;
  98
  99         memset(ls->ls_recover_buf, 0, dlm_config.buffer_size);
 100
 101         if (nodeid == dlm_our_nodeid()) {
 102                 rc = (struct dlm_rcom *) ls->ls_recover_buf;
 103                 rc->rc_result = dlm_recover_status(ls);
 104                 goto out;
 105         }
 106
 107         error = create_rcom(ls, nodeid, DLM_RCOM_STATUS, 0, &rc, &mh);
 108         if (error)
 109                 goto out;
 110
 111         send_rcom(ls, mh, rc);
 112
 113         error = dlm_wait_function(ls, &rcom_response);
 114         clear_bit(LSFL_RCOM_READY, &ls->ls_flags);
 115         if (error)
 116                 goto out;
 117
 118         rc = (struct dlm_rcom *) ls->ls_recover_buf;
 119
 120         if (rc->rc_result == -ESRCH) {
 121                 /* we pretend the remote lockspace exists with 0 status */
 122                 log_debug(ls, "remote node %d not ready", nodeid);
 123                 rc->rc_result = 0;
 124         } else
 125                 error = check_config(ls, (struct rcom_config *) rc->rc_buf,
 126                                      nodeid);
 127         /* the caller looks at rc_result for the remote recovery status */
 128  out:
 129         return error;
 130 }
 131
 132 static void receive_rcom_status(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 133 {
 134         struct dlm_rcom *rc;
 135         struct dlm_mhandle *mh;
 136         int error, nodeid = rc_in->rc_header.h_nodeid;
 137
 138         error = create_rcom(ls, nodeid, DLM_RCOM_STATUS_REPLY,
 139                             sizeof(struct rcom_config), &rc, &mh);
 140         if (error)
 141                 return;
 142         rc->rc_result = dlm_recover_status(ls);
 143         make_config(ls, (struct rcom_config *) rc->rc_buf);
 144
 145         send_rcom(ls, mh, rc);
 146 }
 147
 148 static void receive_rcom_status_reply(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 149 {
 150         memcpy(ls->ls_recover_buf, rc_in, rc_in->rc_header.h_length);
 151         set_bit(LSFL_RCOM_READY, &ls->ls_flags);
 152         wake_up(&ls->ls_wait_general);
 153 }
 154
 155 int dlm_rcom_names(struct dlm_ls *ls, int nodeid, char *last_name, int last_len)
 156 {
 157         struct dlm_rcom *rc;
 158         struct dlm_mhandle *mh;
 159         int error = 0, len = sizeof(struct dlm_rcom);
 160
 161         memset(ls->ls_recover_buf, 0, dlm_config.buffer_size);
 162
 163         if (nodeid == dlm_our_nodeid()) {
 164                 dlm_copy_master_names(ls, last_name, last_len,
 165                                       ls->ls_recover_buf + len,
 166                                       dlm_config.buffer_size - len, nodeid);
 167                 goto out;
 168         }
 169
 170         error = create_rcom(ls, nodeid, DLM_RCOM_NAMES, last_len, &rc, &mh);
 171         if (error)
 172                 goto out;
 173         memcpy(rc->rc_buf, last_name, last_len);
 174
 175         send_rcom(ls, mh, rc);
 176
 177         error = dlm_wait_function(ls, &rcom_response);
 178         clear_bit(LSFL_RCOM_READY, &ls->ls_flags);
 179  out:
 180         return error;
 181 }
 182
 183 static void receive_rcom_names(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 184 {
 185         struct dlm_rcom *rc;
 186         struct dlm_mhandle *mh;
 187         int error, inlen, outlen;
 188         int nodeid = rc_in->rc_header.h_nodeid;
 189         uint32_t status = dlm_recover_status(ls);
 190
 191         /*
 192          * We can't run dlm_dir_rebuild_send (which uses ls_nodes) while
 193          * dlm_recoverd is running ls_nodes_reconfig (which changes ls_nodes).
 194          * It could only happen in rare cases where we get a late NAMES
 195          * message from a previous instance of recovery.
 196          */
 197
 198         if (!(status & DLM_RS_NODES)) {
 199                 log_debug(ls, "ignoring RCOM_NAMES from %u", nodeid);
 200                 return;
 201         }
 202
 203         nodeid = rc_in->rc_header.h_nodeid;
 204         inlen = rc_in->rc_header.h_length - sizeof(struct dlm_rcom);
 205         outlen = dlm_config.buffer_size - sizeof(struct dlm_rcom);
 206
 207         error = create_rcom(ls, nodeid, DLM_RCOM_NAMES_REPLY, outlen, &rc, &mh);
 208         if (error)
 209                 return;
 210
 211         dlm_copy_master_names(ls, rc_in->rc_buf, inlen, rc->rc_buf, outlen,
 212                               nodeid);
 213         send_rcom(ls, mh, rc);
 214 }
 215
 216 static void receive_rcom_names_reply(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 217 {
 218         memcpy(ls->ls_recover_buf, rc_in, rc_in->rc_header.h_length);
 219         set_bit(LSFL_RCOM_READY, &ls->ls_flags);
 220         wake_up(&ls->ls_wait_general);
 221 }
 222
 223 int dlm_send_rcom_lookup(struct dlm_rsb *r, int dir_nodeid)
 224 {
 225         struct dlm_rcom *rc;
 226         struct dlm_mhandle *mh;
 227         struct dlm_ls *ls = r->res_ls;
 228         int error;
 229
 230         error = create_rcom(ls, dir_nodeid, DLM_RCOM_LOOKUP, r->res_length,
 231                             &rc, &mh);
 232         if (error)
 233                 goto out;
 234         memcpy(rc->rc_buf, r->res_name, r->res_length);
 235         rc->rc_id = (unsigned long) r;
 236
 237         send_rcom(ls, mh, rc);
 238  out:
 239         return error;
 240 }
 241
 242 static void receive_rcom_lookup(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 243 {
 244         struct dlm_rcom *rc;
 245         struct dlm_mhandle *mh;
 246         int error, ret_nodeid, nodeid = rc_in->rc_header.h_nodeid;
 247         int len = rc_in->rc_header.h_length - sizeof(struct dlm_rcom);
 248
 249         error = create_rcom(ls, nodeid, DLM_RCOM_LOOKUP_REPLY, 0, &rc, &mh);
 250         if (error)
 251                 return;
 252
 253         error = dlm_dir_lookup(ls, nodeid, rc_in->rc_buf, len, &ret_nodeid);
 254         if (error)
 255                 ret_nodeid = error;
 256         rc->rc_result = ret_nodeid;
 257         rc->rc_id = rc_in->rc_id;
 258
 259         send_rcom(ls, mh, rc);
 260 }
 261
 262 static void receive_rcom_lookup_reply(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 263 {
 264         dlm_recover_master_reply(ls, rc_in);
 265 }
 266
 267 static void pack_rcom_lock(struct dlm_rsb *r, struct dlm_lkb *lkb,
 268                            struct rcom_lock *rl)
 269 {
 270         memset(rl, 0, sizeof(*rl));
 271
 272         rl->rl_ownpid = lkb->lkb_ownpid;
 273         rl->rl_lkid = lkb->lkb_id;
 274         rl->rl_exflags = lkb->lkb_exflags;
 275         rl->rl_flags = lkb->lkb_flags;
 276         rl->rl_lvbseq = lkb->lkb_lvbseq;
 277         rl->rl_rqmode = lkb->lkb_rqmode;
 278         rl->rl_grmode = lkb->lkb_grmode;
 279         rl->rl_status = lkb->lkb_status;
 280         rl->rl_wait_type = lkb->lkb_wait_type;
 281
 282         if (lkb->lkb_bastaddr)
 283                 rl->rl_asts |= AST_BAST;
 284         if (lkb->lkb_astaddr)
 285                 rl->rl_asts |= AST_COMP;
 286
 287         rl->rl_namelen = r->res_length;
 288         memcpy(rl->rl_name, r->res_name, r->res_length);
 289
 290         /* FIXME: might we have an lvb without DLM_LKF_VALBLK set ?
 291            If so, receive_rcom_lock_args() won't take this copy. */
 292
 293         if (lkb->lkb_lvbptr)
 294                 memcpy(rl->rl_lvb, lkb->lkb_lvbptr, r->res_ls->ls_lvblen);
 295 }
 296
 297 int dlm_send_rcom_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
 298 {
 299         struct dlm_ls *ls = r->res_ls;
 300         struct dlm_rcom *rc;
 301         struct dlm_mhandle *mh;
 302         struct rcom_lock *rl;
 303         int error, len = sizeof(struct rcom_lock);
 304
 305         if (lkb->lkb_lvbptr)
 306                 len += ls->ls_lvblen;
 307
 308         error = create_rcom(ls, r->res_nodeid, DLM_RCOM_LOCK, len, &rc, &mh);
 309         if (error)
 310                 goto out;
 311
 312         rl = (struct rcom_lock *) rc->rc_buf;
 313         pack_rcom_lock(r, lkb, rl);
 314         rc->rc_id = (unsigned long) r;
 315
 316         send_rcom(ls, mh, rc);
 317  out:
 318         return error;
 319 }
 320
 321 static void receive_rcom_lock(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 322 {
 323         struct dlm_rcom *rc;
 324         struct dlm_mhandle *mh;
 325         int error, nodeid = rc_in->rc_header.h_nodeid;
 326
 327         dlm_recover_master_copy(ls, rc_in);
 328
 329         error = create_rcom(ls, nodeid, DLM_RCOM_LOCK_REPLY,
 330                             sizeof(struct rcom_lock), &rc, &mh);
 331         if (error)
 332                 return;
 333
 334         /* We send back the same rcom_lock struct we received, but
 335            dlm_recover_master_copy() has filled in rl_remid and rl_result */
 336
 337         memcpy(rc->rc_buf, rc_in->rc_buf, sizeof(struct rcom_lock));
 338         rc->rc_id = rc_in->rc_id;
 339
 340         send_rcom(ls, mh, rc);
 341 }
 342
 343 static void receive_rcom_lock_reply(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 344 {
 345         uint32_t status = dlm_recover_status(ls);
 346
 347         if (!(status & DLM_RS_DIR)) {
 348                 log_debug(ls, "ignoring RCOM_LOCK_REPLY from %u",
 349                           rc_in->rc_header.h_nodeid);
 350                 return;
 351         }
 352
 353         dlm_recover_process_copy(ls, rc_in);
 354 }
 355
 356 static int send_ls_not_ready(int nodeid, struct dlm_rcom *rc_in)
 357 {
 358         struct dlm_rcom *rc;
 359         struct dlm_mhandle *mh;
 360         char *mb;
 361         int mb_len = sizeof(struct dlm_rcom);
 362
 363         mh = dlm_lowcomms_get_buffer(nodeid, mb_len, GFP_KERNEL, &mb);
 364         if (!mh)
 365                 return -ENOBUFS;
 366         memset(mb, 0, mb_len);
 367
 368         rc = (struct dlm_rcom *) mb;
 369
 370         rc->rc_header.h_version = (DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
 371         rc->rc_header.h_lockspace = rc_in->rc_header.h_lockspace;
 372         rc->rc_header.h_nodeid = dlm_our_nodeid();
 373         rc->rc_header.h_length = mb_len;
 374         rc->rc_header.h_cmd = DLM_RCOM;
 375
 376         rc->rc_type = DLM_RCOM_STATUS_REPLY;
 377         rc->rc_result = -ESRCH;
 378
 379         dlm_rcom_out(rc);
 380         dlm_lowcomms_commit_buffer(mh);
 381
 382         return 0;
 383 }
 384
 385 /* Called by dlm_recvd; corresponds to dlm_receive_message() but special
 386    recovery-only comms are sent through here. */
 387
 388 void dlm_receive_rcom(struct dlm_header *hd, int nodeid)
 389 {
 390         struct dlm_rcom *rc = (struct dlm_rcom *) hd;
 391         struct dlm_ls *ls;
 392
 393         dlm_rcom_in(rc);
 394
 395         /* If the lockspace doesn't exist then still send a status message
 396            back; it's possible that it just doesn't have its global_id yet. */
 397
 398         ls = dlm_find_lockspace_global(hd->h_lockspace);
 399         if (!ls) {
 400                 log_print("lockspace %x from %d not found",
 401                           hd->h_lockspace, nodeid);
 402                 send_ls_not_ready(nodeid, rc);
 403                 return;
 404         }
 405
 406         if (dlm_recovery_stopped(ls) && (rc->rc_type != DLM_RCOM_STATUS)) {
 407                 log_error(ls, "ignoring recovery message %x from %d",
 408                           rc->rc_type, nodeid);
 409                 goto out;
 410         }
 411
 412         if (nodeid != rc->rc_header.h_nodeid) {
 413                 log_error(ls, "bad rcom nodeid %d from %d",
 414                           rc->rc_header.h_nodeid, nodeid);
 415                 goto out;
 416         }
 417
 418         switch (rc->rc_type) {
 419         case DLM_RCOM_STATUS:
 420                 receive_rcom_status(ls, rc);
 421                 break;
 422
 423         case DLM_RCOM_NAMES:
 424                 receive_rcom_names(ls, rc);
 425                 break;
 426
 427         case DLM_RCOM_LOOKUP:
 428                 receive_rcom_lookup(ls, rc);
 429                 break;
 430
 431         case DLM_RCOM_LOCK:
 432                 receive_rcom_lock(ls, rc);
 433                 break;
 434
 435         case DLM_RCOM_STATUS_REPLY:
 436                 receive_rcom_status_reply(ls, rc);
 437                 break;
 438
 439         case DLM_RCOM_NAMES_REPLY:
 440                 receive_rcom_names_reply(ls, rc);
 441                 break;
 442
 443         case DLM_RCOM_LOOKUP_REPLY:
 444                 receive_rcom_lookup_reply(ls, rc);
 445                 break;
 446
 447         case DLM_RCOM_LOCK_REPLY:
 448                 receive_rcom_lock_reply(ls, rc);
 449                 break;
 450
 451         default:
 452                 DLM_ASSERT(0, printk("rc_type=%x\n", rc->rc_type););
 453         }
 454  out:
 455         dlm_put_lockspace(ls);
 456 }
 457