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_shared - hold the freeze glock
  97  * @sdp: the superblock
  98  */
  99 int gfs2_freeze_lock_shared(struct gfs2_sbd *sdp)
 100 {
 101         int error;
 102
 103         error = gfs2_glock_nq_init(sdp->sd_freeze_gl, LM_ST_SHARED,
 104                                    LM_FLAG_NOEXP | GL_EXACT,
 105                                    &sdp->sd_freeze_gh);
 106         if (error)
 107                 fs_err(sdp, "can't lock the freeze glock: %d\n", error);
 108         return error;
 109 }
 110
 111 void gfs2_freeze_unlock(struct gfs2_holder *freeze_gh)
 112 {
 113         if (gfs2_holder_initialized(freeze_gh))
 114                 gfs2_glock_dq_uninit(freeze_gh);
 115 }
 116
 117 static void signal_our_withdraw(struct gfs2_sbd *sdp)
 118 {
 119         struct gfs2_glock *live_gl = sdp->sd_live_gh.gh_gl;
 120         struct inode *inode;
 121         struct gfs2_inode *ip;
 122         struct gfs2_glock *i_gl;
 123         u64 no_formal_ino;
 124         int ret = 0;
 125         int tries;
 126
 127         if (test_bit(SDF_NORECOVERY, &sdp->sd_flags) || !sdp->sd_jdesc)
 128                 return;
 129
 130         gfs2_ail_drain(sdp); /* frees all transactions */
 131         inode = sdp->sd_jdesc->jd_inode;
 132         ip = GFS2_I(inode);
 133         i_gl = ip->i_gl;
 134         no_formal_ino = ip->i_no_formal_ino;
 135
 136         /* Prevent any glock dq until withdraw recovery is complete */
 137         set_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags);
 138         /*
 139          * Don't tell dlm we're bailing until we have no more buffers in the
 140          * wind. If journal had an IO error, the log code should just purge
 141          * the outstanding buffers rather than submitting new IO. Making the
 142          * file system read-only will flush the journal, etc.
 143          *
 144          * During a normal unmount, gfs2_make_fs_ro calls gfs2_log_shutdown
 145          * which clears SDF_JOURNAL_LIVE. In a withdraw, we must not write
 146          * any UNMOUNT log header, so we can't call gfs2_log_shutdown, and
 147          * therefore we need to clear SDF_JOURNAL_LIVE manually.
 148          */
 149         clear_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags);
 150         if (!sb_rdonly(sdp->sd_vfs)) {
 151                 bool locked = mutex_trylock(&sdp->sd_freeze_mutex);
 152
 153                 gfs2_make_fs_ro(sdp);
 154
 155                 if (locked)
 156                         mutex_unlock(&sdp->sd_freeze_mutex);
 157
 158                 /*
 159                  * Dequeue any pending non-system glock holders that can no
 160                  * longer be granted because the file system is withdrawn.
 161                  */
 162                 gfs2_gl_dq_holders(sdp);
 163         }
 164
 165         if (sdp->sd_lockstruct.ls_ops->lm_lock == NULL) { /* lock_nolock */
 166                 if (!ret)
 167                         ret = -EIO;
 168                 clear_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags);
 169                 goto skip_recovery;
 170         }
 171         /*
 172          * Drop the glock for our journal so another node can recover it.
 173          */
 174         if (gfs2_holder_initialized(&sdp->sd_journal_gh)) {
 175                 gfs2_glock_dq_wait(&sdp->sd_journal_gh);
 176                 gfs2_holder_uninit(&sdp->sd_journal_gh);
 177         }
 178         sdp->sd_jinode_gh.gh_flags |= GL_NOCACHE;
 179         gfs2_glock_dq(&sdp->sd_jinode_gh);
 180         gfs2_thaw_freeze_initiator(sdp->sd_vfs);
 181         wait_on_bit(&i_gl->gl_flags, GLF_DEMOTE, TASK_UNINTERRUPTIBLE);
 182
 183         /*
 184          * holder_uninit to force glock_put, to force dlm to let go
 185          */
 186         gfs2_holder_uninit(&sdp->sd_jinode_gh);
 187
 188         /*
 189          * Note: We need to be careful here:
 190          * Our iput of jd_inode will evict it. The evict will dequeue its
 191          * glock, but the glock dq will wait for the withdraw unless we have
 192          * exception code in glock_dq.
 193          */
 194         iput(inode);
 195         sdp->sd_jdesc->jd_inode = NULL;
 196         /*
 197          * Wait until the journal inode's glock is freed. This allows try locks
 198          * on other nodes to be successful, otherwise we remain the owner of
 199          * the glock as far as dlm is concerned.
 200          */
 201         if (i_gl->gl_ops->go_free) {
 202                 set_bit(GLF_FREEING, &i_gl->gl_flags);
 203                 wait_on_bit(&i_gl->gl_flags, GLF_FREEING, TASK_UNINTERRUPTIBLE);
 204         }
 205
 206         /*
 207          * Dequeue the "live" glock, but keep a reference so it's never freed.
 208          */
 209         gfs2_glock_hold(live_gl);
 210         gfs2_glock_dq_wait(&sdp->sd_live_gh);
 211         /*
 212          * We enqueue the "live" glock in EX so that all other nodes
 213          * get a demote request and act on it. We don't really want the
 214          * lock in EX, so we send a "try" lock with 1CB to produce a callback.
 215          */
 216         fs_warn(sdp, "Requesting recovery of jid %d.\n",
 217                 sdp->sd_lockstruct.ls_jid);
 218         gfs2_holder_reinit(LM_ST_EXCLUSIVE,
 219                            LM_FLAG_TRY_1CB | LM_FLAG_NOEXP | GL_NOPID,
 220                            &sdp->sd_live_gh);
 221         msleep(GL_GLOCK_MAX_HOLD);
 222         /*
 223          * This will likely fail in a cluster, but succeed standalone:
 224          */
 225         ret = gfs2_glock_nq(&sdp->sd_live_gh);
 226
 227         /*
 228          * If we actually got the "live" lock in EX mode, there are no other
 229          * nodes available to replay our journal. So we try to replay it
 230          * ourselves. We hold the "live" glock to prevent other mounters
 231          * during recovery, then just dequeue it and reacquire it in our
 232          * normal SH mode. Just in case the problem that caused us to
 233          * withdraw prevents us from recovering our journal (e.g. io errors
 234          * and such) we still check if the journal is clean before proceeding
 235          * but we may wait forever until another mounter does the recovery.
 236          */
 237         if (ret == 0) {
 238                 fs_warn(sdp, "No other mounters found. Trying to recover our "
 239                         "own journal jid %d.\n", sdp->sd_lockstruct.ls_jid);
 240                 if (gfs2_recover_journal(sdp->sd_jdesc, 1))
 241                         fs_warn(sdp, "Unable to recover our journal jid %d.\n",
 242                                 sdp->sd_lockstruct.ls_jid);
 243                 gfs2_glock_dq_wait(&sdp->sd_live_gh);
 244                 gfs2_holder_reinit(LM_ST_SHARED,
 245                                    LM_FLAG_NOEXP | GL_EXACT | GL_NOPID,
 246                                    &sdp->sd_live_gh);
 247                 gfs2_glock_nq(&sdp->sd_live_gh);
 248         }
 249
 250         gfs2_glock_queue_put(live_gl); /* drop extra reference we acquired */
 251         clear_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags);
 252
 253         /*
 254          * At this point our journal is evicted, so we need to get a new inode
 255          * for it. Once done, we need to call gfs2_find_jhead which
 256          * calls gfs2_map_journal_extents to map it for us again.
 257          *
 258          * Note that we don't really want it to look up a FREE block. The
 259          * GFS2_BLKST_FREE simply overrides a block check in gfs2_inode_lookup
 260          * which would otherwise fail because it requires grabbing an rgrp
 261          * glock, which would fail with -EIO because we're withdrawing.
 262          */
 263         inode = gfs2_inode_lookup(sdp->sd_vfs, DT_UNKNOWN,
 264                                   sdp->sd_jdesc->jd_no_addr, no_formal_ino,
 265                                   GFS2_BLKST_FREE);
 266         if (IS_ERR(inode)) {
 267                 fs_warn(sdp, "Reprocessing of jid %d failed with %ld.\n",
 268                         sdp->sd_lockstruct.ls_jid, PTR_ERR(inode));
 269                 goto skip_recovery;
 270         }
 271         sdp->sd_jdesc->jd_inode = inode;
 272         d_mark_dontcache(inode);
 273
 274         /*
 275          * Now wait until recovery is complete.
 276          */
 277         for (tries = 0; tries < 10; tries++) {
 278                 ret = check_journal_clean(sdp, sdp->sd_jdesc, false);
 279                 if (!ret)
 280                         break;
 281                 msleep(HZ);
 282                 fs_warn(sdp, "Waiting for journal recovery jid %d.\n",
 283                         sdp->sd_lockstruct.ls_jid);
 284         }
 285 skip_recovery:
 286         if (!ret)
 287                 fs_warn(sdp, "Journal recovery complete for jid %d.\n",
 288                         sdp->sd_lockstruct.ls_jid);
 289         else
 290                 fs_warn(sdp, "Journal recovery skipped for jid %d until next "
 291                         "mount.\n", sdp->sd_lockstruct.ls_jid);
 292         fs_warn(sdp, "Glock dequeues delayed: %lu\n", sdp->sd_glock_dqs_held);
 293         sdp->sd_glock_dqs_held = 0;
 294         wake_up_bit(&sdp->sd_flags, SDF_WITHDRAW_RECOVERY);
 295 }
 296
 297 void gfs2_lm(struct gfs2_sbd *sdp, const char *fmt, ...)
 298 {
 299         struct va_format vaf;
 300         va_list args;
 301
 302         if (sdp->sd_args.ar_errors == GFS2_ERRORS_WITHDRAW &&
 303             test_bit(SDF_WITHDRAWN, &sdp->sd_flags))
 304                 return;
 305
 306         va_start(args, fmt);
 307         vaf.fmt = fmt;
 308         vaf.va = &args;
 309         fs_err(sdp, "%pV", &vaf);
 310         va_end(args);
 311 }
 312
 313 int gfs2_withdraw(struct gfs2_sbd *sdp)
 314 {
 315         struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 316         const struct lm_lockops *lm = ls->ls_ops;
 317
 318         if (sdp->sd_args.ar_errors == GFS2_ERRORS_WITHDRAW &&
 319             test_and_set_bit(SDF_WITHDRAWN, &sdp->sd_flags)) {
 320                 if (!test_bit(SDF_WITHDRAW_IN_PROG, &sdp->sd_flags))
 321                         return -1;
 322
 323                 wait_on_bit(&sdp->sd_flags, SDF_WITHDRAW_IN_PROG,
 324                             TASK_UNINTERRUPTIBLE);
 325                 return -1;
 326         }
 327
 328         set_bit(SDF_WITHDRAW_IN_PROG, &sdp->sd_flags);
 329
 330         if (sdp->sd_args.ar_errors == GFS2_ERRORS_WITHDRAW) {
 331                 fs_err(sdp, "about to withdraw this file system\n");
 332                 BUG_ON(sdp->sd_args.ar_debug);
 333
 334                 signal_our_withdraw(sdp);
 335
 336                 kobject_uevent(&sdp->sd_kobj, KOBJ_OFFLINE);
 337
 338                 if (!strcmp(sdp->sd_lockstruct.ls_ops->lm_proto_name, "lock_dlm"))
 339                         wait_for_completion(&sdp->sd_wdack);
 340
 341                 if (lm->lm_unmount) {
 342                         fs_err(sdp, "telling LM to unmount\n");
 343                         lm->lm_unmount(sdp);
 344                 }
 345                 set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
 346                 fs_err(sdp, "File system withdrawn\n");
 347                 dump_stack();
 348                 clear_bit(SDF_WITHDRAW_IN_PROG, &sdp->sd_flags);
 349                 smp_mb__after_atomic();
 350                 wake_up_bit(&sdp->sd_flags, SDF_WITHDRAW_IN_PROG);
 351         }
 352
 353         if (sdp->sd_args.ar_errors == GFS2_ERRORS_PANIC)
 354                 panic("GFS2: fsid=%s: panic requested\n", sdp->sd_fsname);
 355
 356         return -1;
 357 }
 358
 359 /*
 360  * gfs2_assert_withdraw_i - Cause the machine to withdraw if @assertion is false
 361  */
 362
 363 void gfs2_assert_withdraw_i(struct gfs2_sbd *sdp, char *assertion,
 364                             const char *function, char *file, unsigned int line,
 365                             bool delayed)
 366 {
 367         if (gfs2_withdrawn(sdp))
 368                 return;
 369
 370         fs_err(sdp,
 371                "fatal: assertion \"%s\" failed\n"
 372                "   function = %s, file = %s, line = %u\n",
 373                assertion, function, file, line);
 374
 375         /*
 376          * If errors=panic was specified on mount, it won't help to delay the
 377          * withdraw.
 378          */
 379         if (sdp->sd_args.ar_errors == GFS2_ERRORS_PANIC)
 380                 delayed = false;
 381
 382         if (delayed)
 383                 gfs2_withdraw_delayed(sdp);
 384         else
 385                 gfs2_withdraw(sdp);
 386         dump_stack();
 387 }
 388
 389 /*
 390  * gfs2_assert_warn_i - Print a message to the console if @assertion is false
 391  */
 392
 393 void gfs2_assert_warn_i(struct gfs2_sbd *sdp, char *assertion,
 394                         const char *function, char *file, unsigned int line)
 395 {
 396         if (time_before(jiffies,
 397                         sdp->sd_last_warning +
 398                         gfs2_tune_get(sdp, gt_complain_secs) * HZ))
 399                 return;
 400
 401         if (sdp->sd_args.ar_errors == GFS2_ERRORS_WITHDRAW)
 402                 fs_warn(sdp, "warning: assertion \"%s\" failed at function = %s, file = %s, line = %u\n",
 403                         assertion, function, file, line);
 404
 405         if (sdp->sd_args.ar_debug)
 406                 BUG();
 407         else
 408                 dump_stack();
 409
 410         if (sdp->sd_args.ar_errors == GFS2_ERRORS_PANIC)
 411                 panic("GFS2: fsid=%s: warning: assertion \"%s\" failed\n"
 412                       "GFS2: fsid=%s:   function = %s, file = %s, line = %u\n",
 413                       sdp->sd_fsname, assertion,
 414                       sdp->sd_fsname, function, file, line);
 415
 416         sdp->sd_last_warning = jiffies;
 417 }
 418
 419 /*
 420  * gfs2_consist_i - Flag a filesystem consistency error and withdraw
 421  */
 422
 423 void gfs2_consist_i(struct gfs2_sbd *sdp, const char *function,
 424                     char *file, unsigned int line)
 425 {
 426         gfs2_lm(sdp,
 427                 "fatal: filesystem consistency error - function = %s, file = %s, line = %u\n",
 428                 function, file, line);
 429         gfs2_withdraw(sdp);
 430 }
 431
 432 /*
 433  * gfs2_consist_inode_i - Flag an inode consistency error and withdraw
 434  */
 435
 436 void gfs2_consist_inode_i(struct gfs2_inode *ip,
 437                           const char *function, char *file, unsigned int line)
 438 {
 439         struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
 440
 441         gfs2_lm(sdp,
 442                 "fatal: filesystem consistency error\n"
 443                 "  inode = %llu %llu\n"
 444                 "  function = %s, file = %s, line = %u\n",
 445                 (unsigned long long)ip->i_no_formal_ino,
 446                 (unsigned long long)ip->i_no_addr,
 447                 function, file, line);
 448         gfs2_dump_glock(NULL, ip->i_gl, 1);
 449         gfs2_withdraw(sdp);
 450 }
 451
 452 /*
 453  * gfs2_consist_rgrpd_i - Flag a RG consistency error and withdraw
 454  */
 455
 456 void gfs2_consist_rgrpd_i(struct gfs2_rgrpd *rgd,
 457                           const char *function, char *file, unsigned int line)
 458 {
 459         struct gfs2_sbd *sdp = rgd->rd_sbd;
 460         char fs_id_buf[sizeof(sdp->sd_fsname) + 7];
 461
 462         sprintf(fs_id_buf, "fsid=%s: ", sdp->sd_fsname);
 463         gfs2_rgrp_dump(NULL, rgd, fs_id_buf);
 464         gfs2_lm(sdp,
 465                 "fatal: filesystem consistency error\n"
 466                 "  RG = %llu\n"
 467                 "  function = %s, file = %s, line = %u\n",
 468                 (unsigned long long)rgd->rd_addr,
 469                 function, file, line);
 470         gfs2_dump_glock(NULL, rgd->rd_gl, 1);
 471         gfs2_withdraw(sdp);
 472 }
 473
 474 /*
 475  * gfs2_meta_check_ii - Flag a magic number consistency error and withdraw
 476  * Returns: -1 if this call withdrew the machine,
 477  *          -2 if it was already withdrawn
 478  */
 479
 480 int gfs2_meta_check_ii(struct gfs2_sbd *sdp, struct buffer_head *bh,
 481                        const char *type, const char *function, char *file,
 482                        unsigned int line)
 483 {
 484         int me;
 485
 486         gfs2_lm(sdp,
 487                 "fatal: invalid metadata block\n"
 488                 "  bh = %llu (%s)\n"
 489                 "  function = %s, file = %s, line = %u\n",
 490                 (unsigned long long)bh->b_blocknr, type,
 491                 function, file, line);
 492         me = gfs2_withdraw(sdp);
 493         return (me) ? -1 : -2;
 494 }
 495
 496 /*
 497  * gfs2_metatype_check_ii - Flag a metadata type consistency error and withdraw
 498  * Returns: -1 if this call withdrew the machine,
 499  *          -2 if it was already withdrawn
 500  */
 501
 502 int gfs2_metatype_check_ii(struct gfs2_sbd *sdp, struct buffer_head *bh,
 503                            u16 type, u16 t, const char *function,
 504                            char *file, unsigned int line)
 505 {
 506         int me;
 507
 508         gfs2_lm(sdp,
 509                 "fatal: invalid metadata block\n"
 510                 "  bh = %llu (type: exp=%u, found=%u)\n"
 511                 "  function = %s, file = %s, line = %u\n",
 512                 (unsigned long long)bh->b_blocknr, type, t,
 513                 function, file, line);
 514         me = gfs2_withdraw(sdp);
 515         return (me) ? -1 : -2;
 516 }
 517
 518 /*
 519  * gfs2_io_error_i - Flag an I/O error and withdraw
 520  * Returns: -1 if this call withdrew the machine,
 521  *          0 if it was already withdrawn
 522  */
 523
 524 int gfs2_io_error_i(struct gfs2_sbd *sdp, const char *function, char *file,
 525                     unsigned int line)
 526 {
 527         gfs2_lm(sdp,
 528                 "fatal: I/O error\n"
 529                 "  function = %s, file = %s, line = %u\n",
 530                 function, file, line);
 531         return gfs2_withdraw(sdp);
 532 }
 533
 534 /*
 535  * gfs2_io_error_bh_i - Flag a buffer I/O error
 536  * @withdraw: withdraw the filesystem
 537  */
 538
 539 void gfs2_io_error_bh_i(struct gfs2_sbd *sdp, struct buffer_head *bh,
 540                         const char *function, char *file, unsigned int line,
 541                         bool withdraw)
 542 {
 543         if (gfs2_withdrawn(sdp))
 544                 return;
 545
 546         fs_err(sdp, "fatal: I/O error\n"
 547                "  block = %llu\n"
 548                "  function = %s, file = %s, line = %u\n",
 549                (unsigned long long)bh->b_blocknr, function, file, line);
 550         if (withdraw)
 551                 gfs2_withdraw(sdp);
 552 }
 553