lustre/obdclass/llog_cat.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see [sun.com URL with a
  20  * copy of GPLv2].
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdclass/llog_cat.c
  37  *
  38  * OST<->MDS recovery logging infrastructure.
  39  *
  40  * Invariants in implementation:
  41  * - we do not share logs among different OST<->MDS connections, so that
  42  *   if an OST or MDS fails it need only look at log(s) relevant to itself
  43  *
  44  * Author: Andreas Dilger <adilger@clusterfs.com>
  45  */
  46
  47 #define DEBUG_SUBSYSTEM S_LOG
  48
  49 #ifndef EXPORT_SYMTAB
  50 #define EXPORT_SYMTAB
  51 #endif
  52
  53 #ifndef __KERNEL__
  54 #include <liblustre.h>
  55 #endif
  56
  57 #include <obd_class.h>
  58 #include <lustre_log.h>
  59 #include <libcfs/list.h>
  60
  61 /* Create a new log handle and add it to the open list.
  62  * This log handle will be closed when all of the records in it are removed.
  63  *
  64  * Assumes caller has already pushed us into the kernel context and is locking.
  65  */
  66 static struct llog_handle *llog_cat_new_log(struct llog_handle *cathandle)
  67 {
  68         struct llog_handle *loghandle;
  69         struct llog_log_hdr *llh;
  70         struct llog_logid_rec rec = { { 0 }, };
  71         int rc, index, bitmap_size;
  72         ENTRY;
  73
  74         llh = cathandle->lgh_hdr;
  75         bitmap_size = LLOG_BITMAP_SIZE(llh);
  76
  77         index = (cathandle->lgh_last_idx + 1) % bitmap_size;
  78
  79         /* maximum number of available slots in catlog is bitmap_size - 2 */
  80         if (llh->llh_cat_idx == index) {
  81                 CERROR("no free catalog slots for log...\n");
  82                 RETURN(ERR_PTR(-ENOSPC));
  83         }
  84
  85         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED))
  86                 RETURN(ERR_PTR(-ENOSPC));
  87
  88         rc = llog_create(cathandle->lgh_ctxt, &loghandle, NULL, NULL);
  89         if (rc)
  90                 RETURN(ERR_PTR(rc));
  91
  92         rc = llog_init_handle(loghandle,
  93                               LLOG_F_IS_PLAIN | LLOG_F_ZAP_WHEN_EMPTY,
  94                               &cathandle->lgh_hdr->llh_tgtuuid);
  95         if (rc)
  96                 GOTO(out_destroy, rc);
  97
  98         if (index == 0)
  99                 index = 1;
 100         if (ext2_set_bit(index, llh->llh_bitmap)) {
 101                 CERROR("argh, index %u already set in log bitmap?\n",
 102                        index);
 103                 LBUG(); /* should never happen */
 104         }
 105         cathandle->lgh_last_idx = index;
 106         llh->llh_count++;
 107         llh->llh_tail.lrt_index = index;
 108
 109         CDEBUG(D_RPCTRACE,"new recovery log "LPX64":%x for index %u of catalog "
 110                LPX64"\n", loghandle->lgh_id.lgl_oid, loghandle->lgh_id.lgl_ogen,
 111                index, cathandle->lgh_id.lgl_oid);
 112         /* build the record for this log in the catalog */
 113         rec.lid_hdr.lrh_len = sizeof(rec);
 114         rec.lid_hdr.lrh_index = index;
 115         rec.lid_hdr.lrh_type = LLOG_LOGID_MAGIC;
 116         rec.lid_id = loghandle->lgh_id;
 117         rec.lid_tail.lrt_len = sizeof(rec);
 118         rec.lid_tail.lrt_index = index;
 119
 120         /* update the catalog: header and record */
 121         rc = llog_write_rec(cathandle, &rec.lid_hdr,
 122                             &loghandle->u.phd.phd_cookie, 1, NULL, index);
 123         if (rc < 0) {
 124                 GOTO(out_destroy, rc);
 125         }
 126
 127         loghandle->lgh_hdr->llh_cat_idx = index;
 128         cathandle->u.chd.chd_current_log = loghandle;
 129         LASSERT(list_empty(&loghandle->u.phd.phd_entry));
 130         list_add_tail(&loghandle->u.phd.phd_entry, &cathandle->u.chd.chd_head);
 131
 132  out_destroy:
 133         if (rc < 0)
 134                 llog_destroy(loghandle);
 135
 136         RETURN(loghandle);
 137 }
 138 EXPORT_SYMBOL(llog_cat_new_log);
 139
 140 /* Open an existent log handle and add it to the open list.
 141  * This log handle will be closed when all of the records in it are removed.
 142  *
 143  * Assumes caller has already pushed us into the kernel context and is locking.
 144  * We return a lock on the handle to ensure nobody yanks it from us.
 145  */
 146 int llog_cat_id2handle(struct llog_handle *cathandle, struct llog_handle **res,
 147                        struct llog_logid *logid)
 148 {
 149         struct llog_handle *loghandle;
 150         int rc = 0;
 151         ENTRY;
 152
 153         if (cathandle == NULL)
 154                 RETURN(-EBADF);
 155
 156         list_for_each_entry(loghandle, &cathandle->u.chd.chd_head,
 157                             u.phd.phd_entry) {
 158                 struct llog_logid *cgl = &loghandle->lgh_id;
 159                 if (cgl->lgl_oid == logid->lgl_oid) {
 160                         if (cgl->lgl_ogen != logid->lgl_ogen) {
 161                                 CERROR("log "LPX64" generation %x != %x\n",
 162                                        logid->lgl_oid, cgl->lgl_ogen,
 163                                        logid->lgl_ogen);
 164                                 continue;
 165                         }
 166                         loghandle->u.phd.phd_cat_handle = cathandle;
 167                         GOTO(out, rc = 0);
 168                 }
 169         }
 170
 171         rc = llog_create(cathandle->lgh_ctxt, &loghandle, logid, NULL);
 172         if (rc) {
 173                 CERROR("error opening log id "LPX64":%x: rc %d\n",
 174                        logid->lgl_oid, logid->lgl_ogen, rc);
 175         } else {
 176                 rc = llog_init_handle(loghandle, LLOG_F_IS_PLAIN, NULL);
 177                 if (!rc) {
 178                         list_add(&loghandle->u.phd.phd_entry,
 179                                  &cathandle->u.chd.chd_head);
 180                 }
 181         }
 182         if (!rc) {
 183                 loghandle->u.phd.phd_cat_handle = cathandle;
 184                 loghandle->u.phd.phd_cookie.lgc_lgl = cathandle->lgh_id;
 185                 loghandle->u.phd.phd_cookie.lgc_index =
 186                         loghandle->lgh_hdr->llh_cat_idx;
 187         }
 188
 189 out:
 190         *res = loghandle;
 191         RETURN(rc);
 192 }
 193
 194 int llog_cat_put(struct llog_handle *cathandle)
 195 {
 196         struct llog_handle *loghandle, *n;
 197         int rc;
 198         ENTRY;
 199
 200         list_for_each_entry_safe(loghandle, n, &cathandle->u.chd.chd_head,
 201                                  u.phd.phd_entry) {
 202                 int err = llog_close(loghandle);
 203                 if (err)
 204                         CERROR("error closing loghandle\n");
 205         }
 206         rc = llog_close(cathandle);
 207         RETURN(rc);
 208 }
 209 EXPORT_SYMBOL(llog_cat_put);
 210
 211 /* Return the currently active log handle.  If the current log handle doesn't
 212  * have enough space left for the current record, start a new one.
 213  *
 214  * If reclen is 0, we only want to know what the currently active log is,
 215  * otherwise we get a lock on this log so nobody can steal our space.
 216  *
 217  * Assumes caller has already pushed us into the kernel context and is locking.
 218  *
 219  * NOTE: loghandle is write-locked upon successful return
 220  */
 221 static struct llog_handle *llog_cat_current_log(struct llog_handle *cathandle,
 222                                                 int create)
 223 {
 224         struct llog_handle *loghandle = NULL;
 225         ENTRY;
 226
 227         down_read(&cathandle->lgh_lock);
 228         loghandle = cathandle->u.chd.chd_current_log;
 229         if (loghandle) {
 230                 struct llog_log_hdr *llh = loghandle->lgh_hdr;
 231                 down_write(&loghandle->lgh_lock);
 232                 if (loghandle->lgh_last_idx < LLOG_BITMAP_SIZE(llh) - 1) {
 233                         up_read(&cathandle->lgh_lock);
 234                         RETURN(loghandle);
 235                 } else {
 236                         up_write(&loghandle->lgh_lock);
 237                 }
 238         }
 239         if (!create) {
 240                 if (loghandle)
 241                         down_write(&loghandle->lgh_lock);
 242                 up_read(&cathandle->lgh_lock);
 243                 RETURN(loghandle);
 244         }
 245         up_read(&cathandle->lgh_lock);
 246
 247         /* time to create new log */
 248
 249         /* first, we have to make sure the state hasn't changed */
 250         down_write(&cathandle->lgh_lock);
 251         loghandle = cathandle->u.chd.chd_current_log;
 252         if (loghandle) {
 253                 struct llog_log_hdr *llh = loghandle->lgh_hdr;
 254                 down_write(&loghandle->lgh_lock);
 255                 if (loghandle->lgh_last_idx < LLOG_BITMAP_SIZE(llh) - 1) {
 256                         up_write(&cathandle->lgh_lock);
 257                         RETURN(loghandle);
 258                 } else {
 259                         up_write(&loghandle->lgh_lock);
 260                 }
 261         }
 262
 263         CDEBUG(D_INODE, "creating new log\n");
 264         loghandle = llog_cat_new_log(cathandle);
 265         if (!IS_ERR(loghandle))
 266                 down_write(&loghandle->lgh_lock);
 267         up_write(&cathandle->lgh_lock);
 268         RETURN(loghandle);
 269 }
 270
 271 /* Add a single record to the recovery log(s) using a catalog
 272  * Returns as llog_write_record
 273  *
 274  * Assumes caller has already pushed us into the kernel context.
 275  */
 276 int llog_cat_add_rec(struct llog_handle *cathandle, struct llog_rec_hdr *rec,
 277                      struct llog_cookie *reccookie, void *buf)
 278 {
 279         struct llog_handle *loghandle;
 280         int rc;
 281         ENTRY;
 282
 283         LASSERT(rec->lrh_len <= LLOG_CHUNK_SIZE);
 284         loghandle = llog_cat_current_log(cathandle, 1);
 285         if (IS_ERR(loghandle))
 286                 RETURN(PTR_ERR(loghandle));
 287         /* loghandle is already locked by llog_cat_current_log() for us */
 288         rc = llog_write_rec(loghandle, rec, reccookie, 1, buf, -1);
 289         up_write(&loghandle->lgh_lock);
 290         if (rc == -ENOSPC) {
 291                 /* to create a new plain log */
 292                 loghandle = llog_cat_current_log(cathandle, 1);
 293                 if (IS_ERR(loghandle))
 294                         RETURN(PTR_ERR(loghandle));
 295                 rc = llog_write_rec(loghandle, rec, reccookie, 1, buf, -1);
 296                 up_write(&loghandle->lgh_lock);
 297         }
 298
 299         RETURN(rc);
 300 }
 301 EXPORT_SYMBOL(llog_cat_add_rec);
 302
 303 /* For each cookie in the cookie array, we clear the log in-use bit and either:
 304  * - the log is empty, so mark it free in the catalog header and delete it
 305  * - the log is not empty, just write out the log header
 306  *
 307  * The cookies may be in different log files, so we need to get new logs
 308  * each time.
 309  *
 310  * Assumes caller has already pushed us into the kernel context.
 311  */
 312 int llog_cat_cancel_records(struct llog_handle *cathandle, int count,
 313                         struct llog_cookie *cookies)
 314 {
 315         int i, index, rc = 0;
 316         ENTRY;
 317
 318         down_write(&cathandle->lgh_lock);
 319         for (i = 0; i < count; i++, cookies++) {
 320                 struct llog_handle *loghandle;
 321                 struct llog_logid *lgl = &cookies->lgc_lgl;
 322
 323                 rc = llog_cat_id2handle(cathandle, &loghandle, lgl);
 324                 if (rc) {
 325                         CERROR("Cannot find log "LPX64"\n", lgl->lgl_oid);
 326                         break;
 327                 }
 328
 329                 down_write(&loghandle->lgh_lock);
 330                 rc = llog_cancel_rec(loghandle, cookies->lgc_index);
 331                 up_write(&loghandle->lgh_lock);
 332
 333                 if (rc == 1) {          /* log has been destroyed */
 334                         index = loghandle->u.phd.phd_cookie.lgc_index;
 335                         if (cathandle->u.chd.chd_current_log == loghandle)
 336                                 cathandle->u.chd.chd_current_log = NULL;
 337                         llog_free_handle(loghandle);
 338
 339                         LASSERT(index);
 340                         llog_cat_set_first_idx(cathandle, index);
 341                         rc = llog_cancel_rec(cathandle, index);
 342                         if (rc == 0)
 343                                 CDEBUG(D_RPCTRACE,"cancel plain log at index %u"
 344                                        " of catalog "LPX64"\n",
 345                                        index, cathandle->lgh_id.lgl_oid);
 346                 }
 347         }
 348         up_write(&cathandle->lgh_lock);
 349
 350         RETURN(rc);
 351 }
 352 EXPORT_SYMBOL(llog_cat_cancel_records);
 353
 354 int llog_cat_process_cb(struct llog_handle *cat_llh, struct llog_rec_hdr *rec,
 355                         void *data)
 356 {
 357         struct llog_process_data *d = data;
 358         struct llog_logid_rec *lir = (struct llog_logid_rec *)rec;
 359         struct llog_handle *llh;
 360         int rc;
 361
 362         ENTRY;
 363         if (rec->lrh_type != LLOG_LOGID_MAGIC) {
 364                 CERROR("invalid record in catalog\n");
 365                 RETURN(-EINVAL);
 366         }
 367         CDEBUG(D_HA, "processing log "LPX64":%x at index %u of catalog "
 368                LPX64"\n", lir->lid_id.lgl_oid, lir->lid_id.lgl_ogen,
 369                rec->lrh_index, cat_llh->lgh_id.lgl_oid);
 370
 371         rc = llog_cat_id2handle(cat_llh, &llh, &lir->lid_id);
 372         if (rc) {
 373                 CERROR("Cannot find handle for log "LPX64"\n",
 374                        lir->lid_id.lgl_oid);
 375                 RETURN(rc);
 376         }
 377
 378         rc = llog_process(llh, d->lpd_cb, d->lpd_data, NULL);
 379         RETURN(rc);
 380 }
 381
 382 int llog_cat_process(struct llog_handle *cat_llh, llog_cb_t cb, void *data)
 383 {
 384         struct llog_process_data d;
 385         struct llog_process_cat_data cd;
 386         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
 387         int rc;
 388         ENTRY;
 389
 390         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
 391         d.lpd_data = data;
 392         d.lpd_cb = cb;
 393
 394         if (llh->llh_cat_idx > cat_llh->lgh_last_idx) {
 395                 CWARN("catlog "LPX64" crosses index zero\n",
 396                       cat_llh->lgh_id.lgl_oid);
 397
 398                 cd.first_idx = llh->llh_cat_idx;
 399                 cd.last_idx = 0;
 400                 rc = llog_process(cat_llh, llog_cat_process_cb, &d, &cd);
 401                 if (rc != 0)
 402                         RETURN(rc);
 403
 404                 cd.first_idx = 0;
 405                 cd.last_idx = cat_llh->lgh_last_idx;
 406                 rc = llog_process(cat_llh, llog_cat_process_cb, &d, &cd);
 407         } else {
 408                 rc = llog_process(cat_llh, llog_cat_process_cb, &d, NULL);
 409         }
 410
 411         RETURN(rc);
 412 }
 413 EXPORT_SYMBOL(llog_cat_process);
 414
 415 static int llog_cat_reverse_process_cb(struct llog_handle *cat_llh,
 416                                        struct llog_rec_hdr *rec, void *data)
 417 {
 418         struct llog_process_data *d = data;
 419         struct llog_logid_rec *lir = (struct llog_logid_rec *)rec;
 420         struct llog_handle *llh;
 421         int rc;
 422
 423         if (le32_to_cpu(rec->lrh_type) != LLOG_LOGID_MAGIC) {
 424                 CERROR("invalid record in catalog\n");
 425                 RETURN(-EINVAL);
 426         }
 427         CDEBUG(D_HA, "processing log "LPX64":%x at index %u of catalog "
 428                LPX64"\n", lir->lid_id.lgl_oid, lir->lid_id.lgl_ogen,
 429                le32_to_cpu(rec->lrh_index), cat_llh->lgh_id.lgl_oid);
 430
 431         rc = llog_cat_id2handle(cat_llh, &llh, &lir->lid_id);
 432         if (rc) {
 433                 CERROR("Cannot find handle for log "LPX64"\n",
 434                        lir->lid_id.lgl_oid);
 435                 RETURN(rc);
 436         }
 437
 438         rc = llog_reverse_process(llh, d->lpd_cb, d->lpd_data, NULL);
 439         RETURN(rc);
 440 }
 441
 442 int llog_cat_reverse_process(struct llog_handle *cat_llh,
 443                              llog_cb_t cb, void *data)
 444 {
 445         struct llog_process_data d;
 446         struct llog_process_cat_data cd;
 447         struct llog_log_hdr *llh = cat_llh->lgh_hdr;
 448         int rc;
 449         ENTRY;
 450
 451         LASSERT(llh->llh_flags & LLOG_F_IS_CAT);
 452         d.lpd_data = data;
 453         d.lpd_cb = cb;
 454
 455         if (llh->llh_cat_idx > cat_llh->lgh_last_idx) {
 456                 CWARN("catalog "LPX64" crosses index zero\n",
 457                       cat_llh->lgh_id.lgl_oid);
 458
 459                 cd.first_idx = 0;
 460                 cd.last_idx = cat_llh->lgh_last_idx;
 461                 rc = llog_reverse_process(cat_llh, llog_cat_reverse_process_cb,
 462                                           &d, &cd);
 463                 if (rc != 0)
 464                         RETURN(rc);
 465
 466                 cd.first_idx = le32_to_cpu(llh->llh_cat_idx);
 467                 cd.last_idx = 0;
 468                 rc = llog_reverse_process(cat_llh, llog_cat_reverse_process_cb,
 469                                           &d, &cd);
 470         } else {
 471                 rc = llog_reverse_process(cat_llh, llog_cat_reverse_process_cb,
 472                                           &d, NULL);
 473         }
 474
 475         RETURN(rc);
 476 }
 477 EXPORT_SYMBOL(llog_cat_reverse_process);
 478
 479 int llog_cat_set_first_idx(struct llog_handle *cathandle, int index)
 480 {
 481         struct llog_log_hdr *llh = cathandle->lgh_hdr;
 482         int i, bitmap_size, idx;
 483         ENTRY;
 484
 485         bitmap_size = LLOG_BITMAP_SIZE(llh);
 486         if (llh->llh_cat_idx == (index - 1)) {
 487                 idx = llh->llh_cat_idx + 1;
 488                 llh->llh_cat_idx = idx;
 489                 if (idx == cathandle->lgh_last_idx)
 490                         goto out;
 491                 for (i = (index + 1) % bitmap_size;
 492                      i != cathandle->lgh_last_idx;
 493                      i = (i + 1) % bitmap_size) {
 494                         if (!ext2_test_bit(i, llh->llh_bitmap)) {
 495                                 idx = llh->llh_cat_idx + 1;
 496                                 llh->llh_cat_idx = idx;
 497                         } else if (i == 0) {
 498                                 llh->llh_cat_idx = 0;
 499                         } else {
 500                                 break;
 501                         }
 502                 }
 503 out:
 504                 CDEBUG(D_RPCTRACE, "set catlog "LPX64" first idx %u\n",
 505                        cathandle->lgh_id.lgl_oid, llh->llh_cat_idx);
 506         }
 507
 508         RETURN(0);
 509 }
 510
 511 #if 0
 512 /* Assumes caller has already pushed us into the kernel context. */
 513 int llog_cat_init(struct llog_handle *cathandle, struct obd_uuid *tgtuuid)
 514 {
 515         struct llog_log_hdr *llh;
 516         loff_t offset = 0;
 517         int rc = 0;
 518         ENTRY;
 519
 520         LASSERT(sizeof(*llh) == LLOG_CHUNK_SIZE);
 521
 522         down(&cathandle->lgh_lock);
 523         llh = cathandle->lgh_hdr;
 524
 525         if (i_size_read(cathandle->lgh_file->f_dentry->d_inode) == 0) {
 526                 llog_write_rec(cathandle, &llh->llh_hdr, NULL, 0, NULL, 0);
 527
 528 write_hdr:
 529                 rc = lustre_fwrite(cathandle->lgh_file, llh, LLOG_CHUNK_SIZE,
 530                                    &offset);
 531                 if (rc != LLOG_CHUNK_SIZE) {
 532                         CERROR("error writing catalog header: rc %d\n", rc);
 533                         OBD_FREE(llh, sizeof(*llh));
 534                         if (rc >= 0)
 535                                 rc = -ENOSPC;
 536                 } else
 537                         rc = 0;
 538         } else {
 539                 rc = lustre_fread(cathandle->lgh_file, llh, LLOG_CHUNK_SIZE,
 540                                   &offset);
 541                 if (rc != LLOG_CHUNK_SIZE) {
 542                         CERROR("error reading catalog header: rc %d\n", rc);
 543                         /* Can we do much else if the header is bad? */
 544                         goto write_hdr;
 545                 } else
 546                         rc = 0;
 547         }
 548
 549         cathandle->lgh_tgtuuid = &llh->llh_tgtuuid;
 550         up(&cathandle->lgh_lock);
 551         RETURN(rc);
 552 }
 553 EXPORT_SYMBOL(llog_cat_init);
 554
 555 #endif