fs/gfs2/util.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   4  * Copyright (C) 2004-2006 Red Hat, Inc.  All rights reserved.
   5  */
   6
   7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   8
   9 #include <linux/spinlock.h>
  10 #include <linux/completion.h>
  11 #include <linux/buffer_head.h>
  12 #include <linux/crc32.h>
  13 #include <linux/gfs2_ondisk.h>
  14 #include <linux/delay.h>
  15 #include <linux/uaccess.h>
  16
  17 #include "gfs2.h"
  18 #include "incore.h"
  19 #include "glock.h"
  20 #include "glops.h"
  21 #include "log.h"
  22 #include "lops.h"
  23 #include "recovery.h"
  24 #include "rgrp.h"
  25 #include "super.h"
  26 #include "util.h"
  27
  28 struct kmem_cache *gfs2_glock_cachep __read_mostly;
  29 struct kmem_cache *gfs2_glock_aspace_cachep __read_mostly;
  30 struct kmem_cache *gfs2_inode_cachep __read_mostly;
  31 struct kmem_cache *gfs2_bufdata_cachep __read_mostly;
  32 struct kmem_cache *gfs2_rgrpd_cachep __read_mostly;
  33 struct kmem_cache *gfs2_quotad_cachep __read_mostly;
  34 struct kmem_cache *gfs2_qadata_cachep __read_mostly;
  35 struct kmem_cache *gfs2_trans_cachep __read_mostly;
  36 mempool_t *gfs2_page_pool __read_mostly;
  37
  38 void gfs2_assert_i(struct gfs2_sbd *sdp)
  39 {
  40         fs_emerg(sdp, "fatal assertion failed\n");
  41 }
  42
  43 /**
  44  * check_journal_clean - Make sure a journal is clean for a spectator mount
  45  * @sdp: The GFS2 superblock
  46  * @jd: The journal descriptor
  47  * @verbose: Show more prints in the log
  48  *
  49  * Returns: 0 if the journal is clean or locked, else an error
  50  */
  51 int check_journal_clean(struct gfs2_sbd *sdp, struct gfs2_jdesc *jd,
  52                         bool verbose)
  53 {
  54         int error;
  55         struct gfs2_holder j_gh;
  56         struct gfs2_log_header_host head;
  57         struct gfs2_inode *ip;
  58
  59         ip = GFS2_I(jd->jd_inode);
  60         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_NOEXP |
  61                                    GL_EXACT | GL_NOCACHE, &j_gh);
  62         if (error) {
  63                 if (verbose)
  64                         fs_err(sdp, "Error %d locking journal for spectator "
  65                                "mount.\n", error);
  66                 return -EPERM;
  67         }
  68         error = gfs2_jdesc_check(jd);
  69         if (error) {
  70                 if (verbose)
  71                         fs_err(sdp, "Error checking journal for spectator "
  72                                "mount.\n");
  73                 goto out_unlock;
  74         }
  75         error = gfs2_find_jhead(jd, &head, false);
  76         if (error) {
  77                 if (verbose)
  78                         fs_err(sdp, "Error parsing journal for spectator "
  79                                "mount.\n");
  80                 goto out_unlock;
  81         }
  82         if (!(head.lh_flags & GFS2_LOG_HEAD_UNMOUNT)) {
  83                 error = -EPERM;
  84                 if (verbose)
  85                         fs_err(sdp, "jid=%u: Journal is dirty, so the first "
  86                                "mounter must not be a spectator.\n",
  87                                jd->jd_jid);
  88         }
  89
  90 out_unlock:
  91         gfs2_glock_dq_uninit(&j_gh);
  92         return error;
  93 }
  94
  95 /**
  96  * gfs2_freeze_lock - hold the freeze glock
  97  * @sdp: the superblock
  98  * @freeze_gh: pointer to the requested holder
  99  * @caller_flags: any additional flags needed by the caller
 100  */
 101 int gfs2_freeze_lock(struct gfs2_sbd *sdp, struct gfs2_holder *freeze_gh,
 102                      int caller_flags)
 103 {
 104         int flags = LM_FLAG_NOEXP | GL_EXACT | caller_flags;
 105         int error;
 106
 107         error = gfs2_glock_nq_init(sdp->sd_freeze_gl, LM_ST_SHARED, flags,
 108                                    freeze_gh);
 109         if (error && error != GLR_TRYFAILED)
 110                 fs_err(sdp, "can't lock the freeze lock: %d\n", error);
 111         return error;
 112 }
 113
 114 void gfs2_freeze_unlock(struct gfs2_holder *freeze_gh)
 115 {
 116         if (gfs2_holder_initialized(freeze_gh))
 117                 gfs2_glock_dq_uninit(freeze_gh);
 118 }
 119
 120 static void signal_our_withdraw(struct gfs2_sbd *sdp)
 121 {
 122         struct gfs2_glock *live_gl = sdp->sd_live_gh.gh_gl;
 123         struct inode *inode;
 124         struct gfs2_inode *ip;
 125         struct gfs2_glock *i_gl;
 126         u64 no_formal_ino;
 127         int log_write_allowed = test_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags);
 128         int ret = 0;
 129         int tries;
 130
 131         if (test_bit(SDF_NORECOVERY, &sdp->sd_flags) || !sdp->sd_jdesc)
 132                 return;
 133
 134         gfs2_ail_drain(sdp); /* frees all transactions */
 135         inode = sdp->sd_jdesc->jd_inode;
 136         ip = GFS2_I(inode);
 137         i_gl = ip->i_gl;
 138         no_formal_ino = ip->i_no_formal_ino;
 139
 140         /* Prevent any glock dq until withdraw recovery is complete */
 141         set_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags);
 142         /*
 143          * Don't tell dlm we're bailing until we have no more buffers in the
 144          * wind. If journal had an IO error, the log code should just purge
 145          * the outstanding buffers rather than submitting new IO. Making the
 146          * file system read-only will flush the journal, etc.
 147          *
 148          * During a normal unmount, gfs2_make_fs_ro calls gfs2_log_shutdown
 149          * which clears SDF_JOURNAL_LIVE. In a withdraw, we must not write
 150          * any UNMOUNT log header, so we can't call gfs2_log_shutdown, and
 151          * therefore we need to clear SDF_JOURNAL_LIVE manually.
 152          */
 153         clear_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags);
 154         if (!sb_rdonly(sdp->sd_vfs)) {
 155                 struct gfs2_holder freeze_gh;
 156
 157                 gfs2_holder_mark_uninitialized(&freeze_gh);
 158                 if (sdp->sd_freeze_gl &&
 159                     !gfs2_glock_is_locked_by_me(sdp->sd_freeze_gl)) {
 160                         ret = gfs2_freeze_lock(sdp, &freeze_gh,
 161                                        log_write_allowed ? 0 : LM_FLAG_TRY);
 162                         if (ret == GLR_TRYFAILED)
 163                                 ret = 0;
 164                 }
 165                 if (!ret)
 166                         gfs2_make_fs_ro(sdp);
 167                 gfs2_freeze_unlock(&freeze_gh);
 168         }
 169
 170         if (sdp->sd_lockstruct.ls_ops->lm_lock == NULL) { /* lock_nolock */
 171                 if (!ret)
 172                         ret = -EIO;
 173                 clear_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags);
 174                 goto skip_recovery;
 175         }
 176         /*
 177          * Drop the glock for our journal so another node can recover it.
 178          */
 179         if (gfs2_holder_initialized(&sdp->sd_journal_gh)) {
 180                 gfs2_glock_dq_wait(&sdp->sd_journal_gh);
 181                 gfs2_holder_uninit(&sdp->sd_journal_gh);
 182         }
 183         sdp->sd_jinode_gh.gh_flags |= GL_NOCACHE;
 184         gfs2_glock_dq(&sdp->sd_jinode_gh);
 185         if (test_bit(SDF_FS_FROZEN, &sdp->sd_flags)) {
 186                 /* Make sure gfs2_unfreeze works if partially-frozen */
 187                 flush_work(&sdp->sd_freeze_work);
 188                 atomic_set(&sdp->sd_freeze_state, SFS_FROZEN);
 189                 thaw_super(sdp->sd_vfs);
 190         } else {
 191                 wait_on_bit(&i_gl->gl_flags, GLF_DEMOTE,
 192                             TASK_UNINTERRUPTIBLE);
 193         }
 194
 195         /*
 196          * holder_uninit to force glock_put, to force dlm to let go
 197          */
 198         gfs2_holder_uninit(&sdp->sd_jinode_gh);
 199
 200         /*
 201          * Note: We need to be careful here:
 202          * Our iput of jd_inode will evict it. The evict will dequeue its
 203          * glock, but the glock dq will wait for the withdraw unless we have
 204          * exception code in glock_dq.
 205          */
 206         iput(inode);
 207         /*
 208          * Wait until the journal inode's glock is freed. This allows try locks
 209          * on other nodes to be successful, otherwise we remain the owner of
 210          * the glock as far as dlm is concerned.
 211          */
 212         if (i_gl->gl_ops->go_free) {
 213                 set_bit(GLF_FREEING, &i_gl->gl_flags);
 214                 wait_on_bit(&i_gl->gl_flags, GLF_FREEING, TASK_UNINTERRUPTIBLE);
 215         }
 216
 217         /*
 218          * Dequeue the "live" glock, but keep a reference so it's never freed.
 219          */
 220         gfs2_glock_hold(live_gl);
 221         gfs2_glock_dq_wait(&sdp->sd_live_gh);
 222         /*
 223          * We enqueue the "live" glock in EX so that all other nodes
 224          * get a demote request and act on it. We don't really want the
 225          * lock in EX, so we send a "try" lock with 1CB to produce a callback.
 226          */
 227         fs_warn(sdp, "Requesting recovery of jid %d.\n",
 228                 sdp->sd_lockstruct.ls_jid);
 229         gfs2_holder_reinit(LM_ST_EXCLUSIVE, LM_FLAG_TRY_1CB | LM_FLAG_NOEXP,
 230                            &sdp->sd_live_gh);
 231         msleep(GL_GLOCK_MAX_HOLD);
 232         /*
 233          * This will likely fail in a cluster, but succeed standalone:
 234          */
 235         ret = gfs2_glock_nq(&sdp->sd_live_gh);
 236
 237         /*
 238          * If we actually got the "live" lock in EX mode, there are no other
 239          * nodes available to replay our journal. So we try to replay it
 240          * ourselves. We hold the "live" glock to prevent other mounters
 241          * during recovery, then just dequeue it and reacquire it in our
 242          * normal SH mode. Just in case the problem that caused us to
 243          * withdraw prevents us from recovering our journal (e.g. io errors
 244          * and such) we still check if the journal is clean before proceeding
 245          * but we may wait forever until another mounter does the recovery.
 246          */
 247         if (ret == 0) {
 248                 fs_warn(sdp, "No other mounters found. Trying to recover our "
 249                         "own journal jid %d.\n", sdp->sd_lockstruct.ls_jid);
 250                 if (gfs2_recover_journal(sdp->sd_jdesc, 1))
 251                         fs_warn(sdp, "Unable to recover our journal jid %d.\n",
 252                                 sdp->sd_lockstruct.ls_jid);
 253                 gfs2_glock_dq_wait(&sdp->sd_live_gh);
 254                 gfs2_holder_reinit(LM_ST_SHARED, LM_FLAG_NOEXP | GL_EXACT,
 255                                    &sdp->sd_live_gh);
 256                 gfs2_glock_nq(&sdp->sd_live_gh);
 257         }
 258
 259         gfs2_glock_queue_put(live_gl); /* drop extra reference we acquired */
 260         clear_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags);
 261
 262         /*
 263          * At this point our journal is evicted, so we need to get a new inode
 264          * for it. Once done, we need to call gfs2_find_jhead which
 265          * calls gfs2_map_journal_extents to map it for us again.
 266          *
 267          * Note that we don't really want it to look up a FREE block. The
 268          * GFS2_BLKST_FREE simply overrides a block check in gfs2_inode_lookup
 269          * which would otherwise fail because it requires grabbing an rgrp
 270          * glock, which would fail with -EIO because we're withdrawing.
 271          */
 272         inode = gfs2_inode_lookup(sdp->sd_vfs, DT_UNKNOWN,
 273                                   sdp->sd_jdesc->jd_no_addr, no_formal_ino,
 274                                   GFS2_BLKST_FREE);
 275         if (IS_ERR(inode)) {
 276                 fs_warn(sdp, "Reprocessing of jid %d failed with %ld.\n",
 277                         sdp->sd_lockstruct.ls_jid, PTR_ERR(inode));
 278                 goto skip_recovery;
 279         }
 280         sdp->sd_jdesc->jd_inode = inode;
 281         d_mark_dontcache(inode);
 282
 283         /*
 284          * Now wait until recovery is complete.
 285          */
 286         for (tries = 0; tries < 10; tries++) {
 287                 ret = check_journal_clean(sdp, sdp->sd_jdesc, false);
 288                 if (!ret)
 289                         break;
 290                 msleep(HZ);
 291                 fs_warn(sdp, "Waiting for journal recovery jid %d.\n",
 292                         sdp->sd_lockstruct.ls_jid);
 293         }
 294 skip_recovery:
 295         if (!ret)
 296                 fs_warn(sdp, "Journal recovery complete for jid %d.\n",
 297                         sdp->sd_lockstruct.ls_jid);
 298         else
 299                 fs_warn(sdp, "Journal recovery skipped for jid %d until next "
 300                         "mount.\n", sdp->sd_lockstruct.ls_jid);
 301         fs_warn(sdp, "Glock dequeues delayed: %lu\n", sdp->sd_glock_dqs_held);
 302         sdp->sd_glock_dqs_held = 0;
 303         wake_up_bit(&sdp->sd_flags, SDF_WITHDRAW_RECOVERY);
 304 }
 305
 306 void gfs2_lm(struct gfs2_sbd *sdp, const char *fmt, ...)
 307 {
 308         struct va_format vaf;
 309         va_list args;
 310
 311         if (sdp->sd_args.ar_errors == GFS2_ERRORS_WITHDRAW &&
 312             test_bit(SDF_WITHDRAWN, &sdp->sd_flags))
 313                 return;
 314
 315         va_start(args, fmt);
 316         vaf.fmt = fmt;
 317         vaf.va = &args;
 318         fs_err(sdp, "%pV", &vaf);
 319         va_end(args);
 320 }
 321
 322 int gfs2_withdraw(struct gfs2_sbd *sdp)
 323 {
 324         struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 325         const struct lm_lockops *lm = ls->ls_ops;
 326
 327         if (sdp->sd_args.ar_errors == GFS2_ERRORS_WITHDRAW &&
 328             test_and_set_bit(SDF_WITHDRAWN, &sdp->sd_flags)) {
 329                 if (!test_bit(SDF_WITHDRAW_IN_PROG, &sdp->sd_flags))
 330                         return -1;
 331
 332                 wait_on_bit(&sdp->sd_flags, SDF_WITHDRAW_IN_PROG,
 333                             TASK_UNINTERRUPTIBLE);
 334                 return -1;
 335         }
 336
 337         set_bit(SDF_WITHDRAW_IN_PROG, &sdp->sd_flags);
 338
 339         if (sdp->sd_args.ar_errors == GFS2_ERRORS_WITHDRAW) {
 340                 fs_err(sdp, "about to withdraw this file system\n");
 341                 BUG_ON(sdp->sd_args.ar_debug);
 342
 343                 signal_our_withdraw(sdp);
 344
 345                 kobject_uevent(&sdp->sd_kobj, KOBJ_OFFLINE);
 346
 347                 if (!strcmp(sdp->sd_lockstruct.ls_ops->lm_proto_name, "lock_dlm"))
 348                         wait_for_completion(&sdp->sd_wdack);
 349
 350                 if (lm->lm_unmount) {
 351                         fs_err(sdp, "telling LM to unmount\n");
 352                         lm->lm_unmount(sdp);
 353                 }
 354                 set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
 355                 fs_err(sdp, "File system withdrawn\n");
 356                 dump_stack();
 357                 clear_bit(SDF_WITHDRAW_IN_PROG, &sdp->sd_flags);
 358                 smp_mb__after_atomic();
 359                 wake_up_bit(&sdp->sd_flags, SDF_WITHDRAW_IN_PROG);
 360         }
 361
 362         if (sdp->sd_args.ar_errors == GFS2_ERRORS_PANIC)
 363                 panic("GFS2: fsid=%s: panic requested\n", sdp->sd_fsname);
 364
 365         return -1;
 366 }
 367
 368 /*
 369  * gfs2_assert_withdraw_i - Cause the machine to withdraw if @assertion is false
 370  */
 371
 372 void gfs2_assert_withdraw_i(struct gfs2_sbd *sdp, char *assertion,
 373                             const char *function, char *file, unsigned int line,
 374                             bool delayed)
 375 {
 376         if (gfs2_withdrawn(sdp))
 377                 return;
 378
 379         fs_err(sdp,
 380                "fatal: assertion \"%s\" failed\n"
 381                "   function = %s, file = %s, line = %u\n",
 382                assertion, function, file, line);
 383
 384         /*
 385          * If errors=panic was specified on mount, it won't help to delay the
 386          * withdraw.
 387          */
 388         if (sdp->sd_args.ar_errors == GFS2_ERRORS_PANIC)
 389                 delayed = false;
 390
 391         if (delayed)
 392                 gfs2_withdraw_delayed(sdp);
 393         else
 394                 gfs2_withdraw(sdp);
 395         dump_stack();
 396 }
 397
 398 /*
 399  * gfs2_assert_warn_i - Print a message to the console if @assertion is false
 400  */
 401
 402 void gfs2_assert_warn_i(struct gfs2_sbd *sdp, char *assertion,
 403                         const char *function, char *file, unsigned int line)
 404 {
 405         if (time_before(jiffies,
 406                         sdp->sd_last_warning +
 407                         gfs2_tune_get(sdp, gt_complain_secs) * HZ))
 408                 return;
 409
 410         if (sdp->sd_args.ar_errors == GFS2_ERRORS_WITHDRAW)
 411                 fs_warn(sdp, "warning: assertion \"%s\" failed at function = %s, file = %s, line = %u\n",
 412                         assertion, function, file, line);
 413
 414         if (sdp->sd_args.ar_debug)
 415                 BUG();
 416         else
 417                 dump_stack();
 418
 419         if (sdp->sd_args.ar_errors == GFS2_ERRORS_PANIC)
 420                 panic("GFS2: fsid=%s: warning: assertion \"%s\" failed\n"
 421                       "GFS2: fsid=%s:   function = %s, file = %s, line = %u\n",
 422                       sdp->sd_fsname, assertion,
 423                       sdp->sd_fsname, function, file, line);
 424
 425         sdp->sd_last_warning = jiffies;
 426 }
 427
 428 /*
 429  * gfs2_consist_i - Flag a filesystem consistency error and withdraw
 430  */
 431
 432 void gfs2_consist_i(struct gfs2_sbd *sdp, const char *function,
 433                     char *file, unsigned int line)
 434 {
 435         gfs2_lm(sdp,
 436                 "fatal: filesystem consistency error - function = %s, file = %s, line = %u\n",
 437                 function, file, line);
 438         gfs2_withdraw(sdp);
 439 }
 440
 441 /*
 442  * gfs2_consist_inode_i - Flag an inode consistency error and withdraw
 443  */
 444
 445 void gfs2_consist_inode_i(struct gfs2_inode *ip,
 446                           const char *function, char *file, unsigned int line)
 447 {
 448         struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
 449
 450         gfs2_lm(sdp,
 451                 "fatal: filesystem consistency error\n"
 452                 "  inode = %llu %llu\n"
 453                 "  function = %s, file = %s, line = %u\n",
 454                 (unsigned long long)ip->i_no_formal_ino,
 455                 (unsigned long long)ip->i_no_addr,
 456                 function, file, line);
 457         gfs2_withdraw(sdp);
 458 }
 459
 460 /*
 461  * gfs2_consist_rgrpd_i - Flag a RG consistency error and withdraw
 462  */
 463
 464 void gfs2_consist_rgrpd_i(struct gfs2_rgrpd *rgd,
 465                           const char *function, char *file, unsigned int line)
 466 {
 467         struct gfs2_sbd *sdp = rgd->rd_sbd;
 468         char fs_id_buf[sizeof(sdp->sd_fsname) + 7];
 469
 470         sprintf(fs_id_buf, "fsid=%s: ", sdp->sd_fsname);
 471         gfs2_rgrp_dump(NULL, rgd, fs_id_buf);
 472         gfs2_lm(sdp,
 473                 "fatal: filesystem consistency error\n"
 474                 "  RG = %llu\n"
 475                 "  function = %s, file = %s, line = %u\n",
 476                 (unsigned long long)rgd->rd_addr,
 477                 function, file, line);
 478         gfs2_withdraw(sdp);
 479 }
 480
 481 /*
 482  * gfs2_meta_check_ii - Flag a magic number consistency error and withdraw
 483  * Returns: -1 if this call withdrew the machine,
 484  *          -2 if it was already withdrawn
 485  */
 486
 487 int gfs2_meta_check_ii(struct gfs2_sbd *sdp, struct buffer_head *bh,
 488                        const char *type, const char *function, char *file,
 489                        unsigned int line)
 490 {
 491         int me;
 492
 493         gfs2_lm(sdp,
 494                 "fatal: invalid metadata block\n"
 495                 "  bh = %llu (%s)\n"
 496                 "  function = %s, file = %s, line = %u\n",
 497                 (unsigned long long)bh->b_blocknr, type,
 498                 function, file, line);
 499         me = gfs2_withdraw(sdp);
 500         return (me) ? -1 : -2;
 501 }
 502
 503 /*
 504  * gfs2_metatype_check_ii - Flag a metadata type consistency error and withdraw
 505  * Returns: -1 if this call withdrew the machine,
 506  *          -2 if it was already withdrawn
 507  */
 508
 509 int gfs2_metatype_check_ii(struct gfs2_sbd *sdp, struct buffer_head *bh,
 510                            u16 type, u16 t, const char *function,
 511                            char *file, unsigned int line)
 512 {
 513         int me;
 514
 515         gfs2_lm(sdp,
 516                 "fatal: invalid metadata block\n"
 517                 "  bh = %llu (type: exp=%u, found=%u)\n"
 518                 "  function = %s, file = %s, line = %u\n",
 519                 (unsigned long long)bh->b_blocknr, type, t,
 520                 function, file, line);
 521         me = gfs2_withdraw(sdp);
 522         return (me) ? -1 : -2;
 523 }
 524
 525 /*
 526  * gfs2_io_error_i - Flag an I/O error and withdraw
 527  * Returns: -1 if this call withdrew the machine,
 528  *          0 if it was already withdrawn
 529  */
 530
 531 int gfs2_io_error_i(struct gfs2_sbd *sdp, const char *function, char *file,
 532                     unsigned int line)
 533 {
 534         gfs2_lm(sdp,
 535                 "fatal: I/O error\n"
 536                 "  function = %s, file = %s, line = %u\n",
 537                 function, file, line);
 538         return gfs2_withdraw(sdp);
 539 }
 540
 541 /*
 542  * gfs2_io_error_bh_i - Flag a buffer I/O error
 543  * @withdraw: withdraw the filesystem
 544  */
 545
 546 void gfs2_io_error_bh_i(struct gfs2_sbd *sdp, struct buffer_head *bh,
 547                         const char *function, char *file, unsigned int line,
 548                         bool withdraw)
 549 {
 550         if (gfs2_withdrawn(sdp))
 551                 return;
 552
 553         fs_err(sdp, "fatal: I/O error\n"
 554                "  block = %llu\n"
 555                "  function = %s, file = %s, line = %u\n",
 556                (unsigned long long)bh->b_blocknr, function, file, line);
 557         if (withdraw)
 558                 gfs2_withdraw(sdp);
 559 }
 560