fs/xfs/xfs_mount.c

   1 /*
   2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_types.h"
  21 #include "xfs_bit.h"
  22 #include "xfs_log.h"
  23 #include "xfs_inum.h"
  24 #include "xfs_trans.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_ag.h"
  27 #include "xfs_dir2.h"
  28 #include "xfs_dmapi.h"
  29 #include "xfs_mount.h"
  30 #include "xfs_bmap_btree.h"
  31 #include "xfs_alloc_btree.h"
  32 #include "xfs_ialloc_btree.h"
  33 #include "xfs_dir2_sf.h"
  34 #include "xfs_attr_sf.h"
  35 #include "xfs_dinode.h"
  36 #include "xfs_inode.h"
  37 #include "xfs_btree.h"
  38 #include "xfs_ialloc.h"
  39 #include "xfs_alloc.h"
  40 #include "xfs_rtalloc.h"
  41 #include "xfs_bmap.h"
  42 #include "xfs_error.h"
  43 #include "xfs_rw.h"
  44 #include "xfs_quota.h"
  45 #include "xfs_fsops.h"
  46
  47 STATIC void     xfs_mount_log_sbunit(xfs_mount_t *, __int64_t);
  48 STATIC int      xfs_uuid_mount(xfs_mount_t *);
  49 STATIC void     xfs_uuid_unmount(xfs_mount_t *mp);
  50 STATIC void     xfs_unmountfs_wait(xfs_mount_t *);
  51
  52
  53 #ifdef HAVE_PERCPU_SB
  54 STATIC void     xfs_icsb_destroy_counters(xfs_mount_t *);
  55 STATIC void     xfs_icsb_balance_counter(xfs_mount_t *, xfs_sb_field_t, int);
  56 STATIC void     xfs_icsb_sync_counters(xfs_mount_t *);
  57 STATIC int      xfs_icsb_modify_counters(xfs_mount_t *, xfs_sb_field_t,
  58                                                 int, int);
  59 STATIC int      xfs_icsb_modify_counters_locked(xfs_mount_t *, xfs_sb_field_t,
  60                                                 int, int);
  61 STATIC int      xfs_icsb_disable_counter(xfs_mount_t *, xfs_sb_field_t);
  62
  63 #else
  64
  65 #define xfs_icsb_destroy_counters(mp)                   do { } while (0)
  66 #define xfs_icsb_balance_counter(mp, a, b)              do { } while (0)
  67 #define xfs_icsb_sync_counters(mp)                      do { } while (0)
  68 #define xfs_icsb_modify_counters(mp, a, b, c)           do { } while (0)
  69 #define xfs_icsb_modify_counters_locked(mp, a, b, c)    do { } while (0)
  70
  71 #endif
  72
  73 static const struct {
  74         short offset;
  75         short type;     /* 0 = integer
  76                          * 1 = binary / string (no translation)
  77                          */
  78 } xfs_sb_info[] = {
  79     { offsetof(xfs_sb_t, sb_magicnum),   0 },
  80     { offsetof(xfs_sb_t, sb_blocksize),  0 },
  81     { offsetof(xfs_sb_t, sb_dblocks),    0 },
  82     { offsetof(xfs_sb_t, sb_rblocks),    0 },
  83     { offsetof(xfs_sb_t, sb_rextents),   0 },
  84     { offsetof(xfs_sb_t, sb_uuid),       1 },
  85     { offsetof(xfs_sb_t, sb_logstart),   0 },
  86     { offsetof(xfs_sb_t, sb_rootino),    0 },
  87     { offsetof(xfs_sb_t, sb_rbmino),     0 },
  88     { offsetof(xfs_sb_t, sb_rsumino),    0 },
  89     { offsetof(xfs_sb_t, sb_rextsize),   0 },
  90     { offsetof(xfs_sb_t, sb_agblocks),   0 },
  91     { offsetof(xfs_sb_t, sb_agcount),    0 },
  92     { offsetof(xfs_sb_t, sb_rbmblocks),  0 },
  93     { offsetof(xfs_sb_t, sb_logblocks),  0 },
  94     { offsetof(xfs_sb_t, sb_versionnum), 0 },
  95     { offsetof(xfs_sb_t, sb_sectsize),   0 },
  96     { offsetof(xfs_sb_t, sb_inodesize),  0 },
  97     { offsetof(xfs_sb_t, sb_inopblock),  0 },
  98     { offsetof(xfs_sb_t, sb_fname[0]),   1 },
  99     { offsetof(xfs_sb_t, sb_blocklog),   0 },
 100     { offsetof(xfs_sb_t, sb_sectlog),    0 },
 101     { offsetof(xfs_sb_t, sb_inodelog),   0 },
 102     { offsetof(xfs_sb_t, sb_inopblog),   0 },
 103     { offsetof(xfs_sb_t, sb_agblklog),   0 },
 104     { offsetof(xfs_sb_t, sb_rextslog),   0 },
 105     { offsetof(xfs_sb_t, sb_inprogress), 0 },
 106     { offsetof(xfs_sb_t, sb_imax_pct),   0 },
 107     { offsetof(xfs_sb_t, sb_icount),     0 },
 108     { offsetof(xfs_sb_t, sb_ifree),      0 },
 109     { offsetof(xfs_sb_t, sb_fdblocks),   0 },
 110     { offsetof(xfs_sb_t, sb_frextents),  0 },
 111     { offsetof(xfs_sb_t, sb_uquotino),   0 },
 112     { offsetof(xfs_sb_t, sb_gquotino),   0 },
 113     { offsetof(xfs_sb_t, sb_qflags),     0 },
 114     { offsetof(xfs_sb_t, sb_flags),      0 },
 115     { offsetof(xfs_sb_t, sb_shared_vn),  0 },
 116     { offsetof(xfs_sb_t, sb_inoalignmt), 0 },
 117     { offsetof(xfs_sb_t, sb_unit),       0 },
 118     { offsetof(xfs_sb_t, sb_width),      0 },
 119     { offsetof(xfs_sb_t, sb_dirblklog),  0 },
 120     { offsetof(xfs_sb_t, sb_logsectlog), 0 },
 121     { offsetof(xfs_sb_t, sb_logsectsize),0 },
 122     { offsetof(xfs_sb_t, sb_logsunit),   0 },
 123     { offsetof(xfs_sb_t, sb_features2),  0 },
 124     { sizeof(xfs_sb_t),                  0 }
 125 };
 126
 127 /*
 128  * Return a pointer to an initialized xfs_mount structure.
 129  */
 130 xfs_mount_t *
 131 xfs_mount_init(void)
 132 {
 133         xfs_mount_t *mp;
 134
 135         mp = kmem_zalloc(sizeof(xfs_mount_t), KM_SLEEP);
 136
 137         if (xfs_icsb_init_counters(mp)) {
 138                 mp->m_flags |= XFS_MOUNT_NO_PERCPU_SB;
 139         }
 140
 141         AIL_LOCKINIT(&mp->m_ail_lock, "xfs_ail");
 142         spinlock_init(&mp->m_sb_lock, "xfs_sb");
 143         mutex_init(&mp->m_ilock);
 144         initnsema(&mp->m_growlock, 1, "xfs_grow");
 145         /*
 146          * Initialize the AIL.
 147          */
 148         xfs_trans_ail_init(mp);
 149
 150         atomic_set(&mp->m_active_trans, 0);
 151
 152         return mp;
 153 }
 154
 155 /*
 156  * Free up the resources associated with a mount structure.  Assume that
 157  * the structure was initially zeroed, so we can tell which fields got
 158  * initialized.
 159  */
 160 void
 161 xfs_mount_free(
 162         xfs_mount_t     *mp,
 163         int             remove_bhv)
 164 {
 165         if (mp->m_ihash)
 166                 xfs_ihash_free(mp);
 167         if (mp->m_chash)
 168                 xfs_chash_free(mp);
 169
 170         if (mp->m_perag) {
 171                 int     agno;
 172
 173                 for (agno = 0; agno < mp->m_maxagi; agno++)
 174                         if (mp->m_perag[agno].pagb_list)
 175                                 kmem_free(mp->m_perag[agno].pagb_list,
 176                                                 sizeof(xfs_perag_busy_t) *
 177                                                         XFS_PAGB_NUM_SLOTS);
 178                 kmem_free(mp->m_perag,
 179                           sizeof(xfs_perag_t) * mp->m_sb.sb_agcount);
 180         }
 181
 182         AIL_LOCK_DESTROY(&mp->m_ail_lock);
 183         spinlock_destroy(&mp->m_sb_lock);
 184         mutex_destroy(&mp->m_ilock);
 185         freesema(&mp->m_growlock);
 186         if (mp->m_quotainfo)
 187                 XFS_QM_DONE(mp);
 188
 189         if (mp->m_fsname != NULL)
 190                 kmem_free(mp->m_fsname, mp->m_fsname_len);
 191         if (mp->m_rtname != NULL)
 192                 kmem_free(mp->m_rtname, strlen(mp->m_rtname) + 1);
 193         if (mp->m_logname != NULL)
 194                 kmem_free(mp->m_logname, strlen(mp->m_logname) + 1);
 195
 196         if (remove_bhv) {
 197                 struct bhv_vfs  *vfsp = XFS_MTOVFS(mp);
 198
 199                 bhv_remove_all_vfsops(vfsp, 0);
 200                 VFS_REMOVEBHV(vfsp, &mp->m_bhv);
 201         }
 202
 203         xfs_icsb_destroy_counters(mp);
 204         kmem_free(mp, sizeof(xfs_mount_t));
 205 }
 206
 207
 208 /*
 209  * Check the validity of the SB found.
 210  */
 211 STATIC int
 212 xfs_mount_validate_sb(
 213         xfs_mount_t     *mp,
 214         xfs_sb_t        *sbp,
 215         int             flags)
 216 {
 217         /*
 218          * If the log device and data device have the
 219          * same device number, the log is internal.
 220          * Consequently, the sb_logstart should be non-zero.  If
 221          * we have a zero sb_logstart in this case, we may be trying to mount
 222          * a volume filesystem in a non-volume manner.
 223          */
 224         if (sbp->sb_magicnum != XFS_SB_MAGIC) {
 225                 xfs_fs_mount_cmn_err(flags, "bad magic number");
 226                 return XFS_ERROR(EWRONGFS);
 227         }
 228
 229         if (!XFS_SB_GOOD_VERSION(sbp)) {
 230                 xfs_fs_mount_cmn_err(flags, "bad version");
 231                 return XFS_ERROR(EWRONGFS);
 232         }
 233
 234         if (unlikely(
 235             sbp->sb_logstart == 0 && mp->m_logdev_targp == mp->m_ddev_targp)) {
 236                 xfs_fs_mount_cmn_err(flags,
 237                         "filesystem is marked as having an external log; "
 238                         "specify logdev on the\nmount command line.");
 239                 return XFS_ERROR(EINVAL);
 240         }
 241
 242         if (unlikely(
 243             sbp->sb_logstart != 0 && mp->m_logdev_targp != mp->m_ddev_targp)) {
 244                 xfs_fs_mount_cmn_err(flags,
 245                         "filesystem is marked as having an internal log; "
 246                         "do not specify logdev on\nthe mount command line.");
 247                 return XFS_ERROR(EINVAL);
 248         }
 249
 250         /*
 251          * More sanity checking. These were stolen directly from
 252          * xfs_repair.
 253          */
 254         if (unlikely(
 255             sbp->sb_agcount <= 0                                        ||
 256             sbp->sb_sectsize < XFS_MIN_SECTORSIZE                       ||
 257             sbp->sb_sectsize > XFS_MAX_SECTORSIZE                       ||
 258             sbp->sb_sectlog < XFS_MIN_SECTORSIZE_LOG                    ||
 259             sbp->sb_sectlog > XFS_MAX_SECTORSIZE_LOG                    ||
 260             sbp->sb_blocksize < XFS_MIN_BLOCKSIZE                       ||
 261             sbp->sb_blocksize > XFS_MAX_BLOCKSIZE                       ||
 262             sbp->sb_blocklog < XFS_MIN_BLOCKSIZE_LOG                    ||
 263             sbp->sb_blocklog > XFS_MAX_BLOCKSIZE_LOG                    ||
 264             sbp->sb_inodesize < XFS_DINODE_MIN_SIZE                     ||
 265             sbp->sb_inodesize > XFS_DINODE_MAX_SIZE                     ||
 266             sbp->sb_inodelog < XFS_DINODE_MIN_LOG                       ||
 267             sbp->sb_inodelog > XFS_DINODE_MAX_LOG                       ||
 268             (sbp->sb_blocklog - sbp->sb_inodelog != sbp->sb_inopblog)   ||
 269             (sbp->sb_rextsize * sbp->sb_blocksize > XFS_MAX_RTEXTSIZE)  ||
 270             (sbp->sb_rextsize * sbp->sb_blocksize < XFS_MIN_RTEXTSIZE)  ||
 271             (sbp->sb_imax_pct > 100 /* zero sb_imax_pct is valid */))) {
 272                 xfs_fs_mount_cmn_err(flags, "SB sanity check 1 failed");
 273                 return XFS_ERROR(EFSCORRUPTED);
 274         }
 275
 276         /*
 277          * Sanity check AG count, size fields against data size field
 278          */
 279         if (unlikely(
 280             sbp->sb_dblocks == 0 ||
 281             sbp->sb_dblocks >
 282              (xfs_drfsbno_t)sbp->sb_agcount * sbp->sb_agblocks ||
 283             sbp->sb_dblocks < (xfs_drfsbno_t)(sbp->sb_agcount - 1) *
 284                               sbp->sb_agblocks + XFS_MIN_AG_BLOCKS)) {
 285                 xfs_fs_mount_cmn_err(flags, "SB sanity check 2 failed");
 286                 return XFS_ERROR(EFSCORRUPTED);
 287         }
 288
 289         ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
 290         ASSERT(sbp->sb_blocklog >= BBSHIFT);
 291
 292 #if XFS_BIG_BLKNOS     /* Limited by ULONG_MAX of page cache index */
 293         if (unlikely(
 294             (sbp->sb_dblocks >> (PAGE_SHIFT - sbp->sb_blocklog)) > ULONG_MAX ||
 295             (sbp->sb_rblocks >> (PAGE_SHIFT - sbp->sb_blocklog)) > ULONG_MAX)) {
 296 #else                  /* Limited by UINT_MAX of sectors */
 297         if (unlikely(
 298             (sbp->sb_dblocks << (sbp->sb_blocklog - BBSHIFT)) > UINT_MAX ||
 299             (sbp->sb_rblocks << (sbp->sb_blocklog - BBSHIFT)) > UINT_MAX)) {
 300 #endif
 301                 xfs_fs_mount_cmn_err(flags,
 302                         "file system too large to be mounted on this system.");
 303                 return XFS_ERROR(E2BIG);
 304         }
 305
 306         if (unlikely(sbp->sb_inprogress)) {
 307                 xfs_fs_mount_cmn_err(flags, "file system busy");
 308                 return XFS_ERROR(EFSCORRUPTED);
 309         }
 310
 311         /*
 312          * Version 1 directory format has never worked on Linux.
 313          */
 314         if (unlikely(!XFS_SB_VERSION_HASDIRV2(sbp))) {
 315                 xfs_fs_mount_cmn_err(flags,
 316                         "file system using version 1 directory format");
 317                 return XFS_ERROR(ENOSYS);
 318         }
 319
 320         /*
 321          * Until this is fixed only page-sized or smaller data blocks work.
 322          */
 323         if (unlikely(sbp->sb_blocksize > PAGE_SIZE)) {
 324                 xfs_fs_mount_cmn_err(flags,
 325                         "file system with blocksize %d bytes",
 326                         sbp->sb_blocksize);
 327                 xfs_fs_mount_cmn_err(flags,
 328                         "only pagesize (%ld) or less will currently work.",
 329                         PAGE_SIZE);
 330                 return XFS_ERROR(ENOSYS);
 331         }
 332
 333         return 0;
 334 }
 335
 336 xfs_agnumber_t
 337 xfs_initialize_perag(
 338         bhv_vfs_t       *vfs,
 339         xfs_mount_t     *mp,
 340         xfs_agnumber_t  agcount)
 341 {
 342         xfs_agnumber_t  index, max_metadata;
 343         xfs_perag_t     *pag;
 344         xfs_agino_t     agino;
 345         xfs_ino_t       ino;
 346         xfs_sb_t        *sbp = &mp->m_sb;
 347         xfs_ino_t       max_inum = XFS_MAXINUMBER_32;
 348
 349         /* Check to see if the filesystem can overflow 32 bit inodes */
 350         agino = XFS_OFFBNO_TO_AGINO(mp, sbp->sb_agblocks - 1, 0);
 351         ino = XFS_AGINO_TO_INO(mp, agcount - 1, agino);
 352
 353         /* Clear the mount flag if no inode can overflow 32 bits
 354          * on this filesystem, or if specifically requested..
 355          */
 356         if ((vfs->vfs_flag & VFS_32BITINODES) && ino > max_inum) {
 357                 mp->m_flags |= XFS_MOUNT_32BITINODES;
 358         } else {
 359                 mp->m_flags &= ~XFS_MOUNT_32BITINODES;
 360         }
 361
 362         /* If we can overflow then setup the ag headers accordingly */
 363         if (mp->m_flags & XFS_MOUNT_32BITINODES) {
 364                 /* Calculate how much should be reserved for inodes to
 365                  * meet the max inode percentage.
 366                  */
 367                 if (mp->m_maxicount) {
 368                         __uint64_t      icount;
 369
 370                         icount = sbp->sb_dblocks * sbp->sb_imax_pct;
 371                         do_div(icount, 100);
 372                         icount += sbp->sb_agblocks - 1;
 373                         do_div(icount, sbp->sb_agblocks);
 374                         max_metadata = icount;
 375                 } else {
 376                         max_metadata = agcount;
 377                 }
 378                 for (index = 0; index < agcount; index++) {
 379                         ino = XFS_AGINO_TO_INO(mp, index, agino);
 380                         if (ino > max_inum) {
 381                                 index++;
 382                                 break;
 383                         }
 384
 385                         /* This ag is preferred for inodes */
 386                         pag = &mp->m_perag[index];
 387                         pag->pagi_inodeok = 1;
 388                         if (index < max_metadata)
 389                                 pag->pagf_metadata = 1;
 390                 }
 391         } else {
 392                 /* Setup default behavior for smaller filesystems */
 393                 for (index = 0; index < agcount; index++) {
 394                         pag = &mp->m_perag[index];
 395                         pag->pagi_inodeok = 1;
 396                 }
 397         }
 398         return index;
 399 }
 400
 401 /*
 402  * xfs_xlatesb
 403  *
 404  *     data       - on disk version of sb
 405  *     sb         - a superblock
 406  *     dir        - conversion direction: <0 - convert sb to buf
 407  *                                        >0 - convert buf to sb
 408  *     fields     - which fields to copy (bitmask)
 409  */
 410 void
 411 xfs_xlatesb(
 412         void            *data,
 413         xfs_sb_t        *sb,
 414         int             dir,
 415         __int64_t       fields)
 416 {
 417         xfs_caddr_t     buf_ptr;
 418         xfs_caddr_t     mem_ptr;
 419         xfs_sb_field_t  f;
 420         int             first;
 421         int             size;
 422
 423         ASSERT(dir);
 424         ASSERT(fields);
 425
 426         if (!fields)
 427                 return;
 428
 429         buf_ptr = (xfs_caddr_t)data;
 430         mem_ptr = (xfs_caddr_t)sb;
 431
 432         while (fields) {
 433                 f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
 434                 first = xfs_sb_info[f].offset;
 435                 size = xfs_sb_info[f + 1].offset - first;
 436
 437                 ASSERT(xfs_sb_info[f].type == 0 || xfs_sb_info[f].type == 1);
 438
 439                 if (size == 1 || xfs_sb_info[f].type == 1) {
 440                         if (dir > 0) {
 441                                 memcpy(mem_ptr + first, buf_ptr + first, size);
 442                         } else {
 443                                 memcpy(buf_ptr + first, mem_ptr + first, size);
 444                         }
 445                 } else {
 446                         switch (size) {
 447                         case 2:
 448                                 INT_XLATE(*(__uint16_t*)(buf_ptr+first),
 449                                           *(__uint16_t*)(mem_ptr+first),
 450                                           dir, ARCH_CONVERT);
 451                                 break;
 452                         case 4:
 453                                 INT_XLATE(*(__uint32_t*)(buf_ptr+first),
 454                                           *(__uint32_t*)(mem_ptr+first),
 455                                           dir, ARCH_CONVERT);
 456                                 break;
 457                         case 8:
 458                                 INT_XLATE(*(__uint64_t*)(buf_ptr+first),
 459                                           *(__uint64_t*)(mem_ptr+first), dir, ARCH_CONVERT);
 460                                 break;
 461                         default:
 462                                 ASSERT(0);
 463                         }
 464                 }
 465
 466                 fields &= ~(1LL << f);
 467         }
 468 }
 469
 470 /*
 471  * xfs_readsb
 472  *
 473  * Does the initial read of the superblock.
 474  */
 475 int
 476 xfs_readsb(xfs_mount_t *mp, int flags)
 477 {
 478         unsigned int    sector_size;
 479         unsigned int    extra_flags;
 480         xfs_buf_t       *bp;
 481         xfs_sb_t        *sbp;
 482         int             error;
 483
 484         ASSERT(mp->m_sb_bp == NULL);
 485         ASSERT(mp->m_ddev_targp != NULL);
 486
 487         /*
 488          * Allocate a (locked) buffer to hold the superblock.
 489          * This will be kept around at all times to optimize
 490          * access to the superblock.
 491          */
 492         sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
 493         extra_flags = XFS_BUF_LOCK | XFS_BUF_MANAGE | XFS_BUF_MAPPED;
 494
 495         bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
 496                                 BTOBB(sector_size), extra_flags);
 497         if (!bp || XFS_BUF_ISERROR(bp)) {
 498                 xfs_fs_mount_cmn_err(flags, "SB read failed");
 499                 error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
 500                 goto fail;
 501         }
 502         ASSERT(XFS_BUF_ISBUSY(bp));
 503         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 504
 505         /*
 506          * Initialize the mount structure from the superblock.
 507          * But first do some basic consistency checking.
 508          */
 509         sbp = XFS_BUF_TO_SBP(bp);
 510         xfs_xlatesb(XFS_BUF_PTR(bp), &(mp->m_sb), 1, XFS_SB_ALL_BITS);
 511
 512         error = xfs_mount_validate_sb(mp, &(mp->m_sb), flags);
 513         if (error) {
 514                 xfs_fs_mount_cmn_err(flags, "SB validate failed");
 515                 goto fail;
 516         }
 517
 518         /*
 519          * We must be able to do sector-sized and sector-aligned IO.
 520          */
 521         if (sector_size > mp->m_sb.sb_sectsize) {
 522                 xfs_fs_mount_cmn_err(flags,
 523                         "device supports only %u byte sectors (not %u)",
 524                         sector_size, mp->m_sb.sb_sectsize);
 525                 error = ENOSYS;
 526                 goto fail;
 527         }
 528
 529         /*
 530          * If device sector size is smaller than the superblock size,
 531          * re-read the superblock so the buffer is correctly sized.
 532          */
 533         if (sector_size < mp->m_sb.sb_sectsize) {
 534                 XFS_BUF_UNMANAGE(bp);
 535                 xfs_buf_relse(bp);
 536                 sector_size = mp->m_sb.sb_sectsize;
 537                 bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
 538                                         BTOBB(sector_size), extra_flags);
 539                 if (!bp || XFS_BUF_ISERROR(bp)) {
 540                         xfs_fs_mount_cmn_err(flags, "SB re-read failed");
 541                         error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
 542                         goto fail;
 543                 }
 544                 ASSERT(XFS_BUF_ISBUSY(bp));
 545                 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 546         }
 547
 548         xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
 549         xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
 550         xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
 551
 552         mp->m_sb_bp = bp;
 553         xfs_buf_relse(bp);
 554         ASSERT(XFS_BUF_VALUSEMA(bp) > 0);
 555         return 0;
 556
 557  fail:
 558         if (bp) {
 559                 XFS_BUF_UNMANAGE(bp);
 560                 xfs_buf_relse(bp);
 561         }
 562         return error;
 563 }
 564
 565
 566 /*
 567  * xfs_mount_common
 568  *
 569  * Mount initialization code establishing various mount
 570  * fields from the superblock associated with the given
 571  * mount structure
 572  */
 573 STATIC void
 574 xfs_mount_common(xfs_mount_t *mp, xfs_sb_t *sbp)
 575 {
 576         int     i;
 577
 578         mp->m_agfrotor = mp->m_agirotor = 0;
 579         spinlock_init(&mp->m_agirotor_lock, "m_agirotor_lock");
 580         mp->m_maxagi = mp->m_sb.sb_agcount;
 581         mp->m_blkbit_log = sbp->sb_blocklog + XFS_NBBYLOG;
 582         mp->m_blkbb_log = sbp->sb_blocklog - BBSHIFT;
 583         mp->m_sectbb_log = sbp->sb_sectlog - BBSHIFT;
 584         mp->m_agno_log = xfs_highbit32(sbp->sb_agcount - 1) + 1;
 585         mp->m_agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
 586         mp->m_litino = sbp->sb_inodesize -
 587                 ((uint)sizeof(xfs_dinode_core_t) + (uint)sizeof(xfs_agino_t));
 588         mp->m_blockmask = sbp->sb_blocksize - 1;
 589         mp->m_blockwsize = sbp->sb_blocksize >> XFS_WORDLOG;
 590         mp->m_blockwmask = mp->m_blockwsize - 1;
 591         INIT_LIST_HEAD(&mp->m_del_inodes);
 592
 593         /*
 594          * Setup for attributes, in case they get created.
 595          * This value is for inodes getting attributes for the first time,
 596          * the per-inode value is for old attribute values.
 597          */
 598         ASSERT(sbp->sb_inodesize >= 256 && sbp->sb_inodesize <= 2048);
 599         switch (sbp->sb_inodesize) {
 600         case 256:
 601                 mp->m_attroffset = XFS_LITINO(mp) -
 602                                    XFS_BMDR_SPACE_CALC(MINABTPTRS);
 603                 break;
 604         case 512:
 605         case 1024:
 606         case 2048:
 607                 mp->m_attroffset = XFS_BMDR_SPACE_CALC(6 * MINABTPTRS);
 608                 break;
 609         default:
 610                 ASSERT(0);
 611         }
 612         ASSERT(mp->m_attroffset < XFS_LITINO(mp));
 613
 614         for (i = 0; i < 2; i++) {
 615                 mp->m_alloc_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 616                         xfs_alloc, i == 0);
 617                 mp->m_alloc_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 618                         xfs_alloc, i == 0);
 619         }
 620         for (i = 0; i < 2; i++) {
 621                 mp->m_bmap_dmxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 622                         xfs_bmbt, i == 0);
 623                 mp->m_bmap_dmnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 624                         xfs_bmbt, i == 0);
 625         }
 626         for (i = 0; i < 2; i++) {
 627                 mp->m_inobt_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 628                         xfs_inobt, i == 0);
 629                 mp->m_inobt_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 630                         xfs_inobt, i == 0);
 631         }
 632
 633         mp->m_bsize = XFS_FSB_TO_BB(mp, 1);
 634         mp->m_ialloc_inos = (int)MAX((__uint16_t)XFS_INODES_PER_CHUNK,
 635                                         sbp->sb_inopblock);
 636         mp->m_ialloc_blks = mp->m_ialloc_inos >> sbp->sb_inopblog;
 637 }
 638 /*
 639  * xfs_mountfs
 640  *
 641  * This function does the following on an initial mount of a file system:
 642  *      - reads the superblock from disk and init the mount struct
 643  *      - if we're a 32-bit kernel, do a size check on the superblock
 644  *              so we don't mount terabyte filesystems
 645  *      - init mount struct realtime fields
 646  *      - allocate inode hash table for fs
 647  *      - init directory manager
 648  *      - perform recovery and init the log manager
 649  */
 650 int
 651 xfs_mountfs(
 652         bhv_vfs_t       *vfsp,
 653         xfs_mount_t     *mp,
 654         int             mfsi_flags)
 655 {
 656         xfs_buf_t       *bp;
 657         xfs_sb_t        *sbp = &(mp->m_sb);
 658         xfs_inode_t     *rip;
 659         bhv_vnode_t     *rvp = NULL;
 660         int             readio_log, writeio_log;
 661         xfs_daddr_t     d;
 662         __uint64_t      ret64;
 663         __int64_t       update_flags;
 664         uint            quotamount, quotaflags;
 665         int             agno;
 666         int             uuid_mounted = 0;
 667         int             error = 0;
 668
 669         if (mp->m_sb_bp == NULL) {
 670                 if ((error = xfs_readsb(mp, mfsi_flags))) {
 671                         return error;
 672                 }
 673         }
 674         xfs_mount_common(mp, sbp);
 675
 676         /*
 677          * Check if sb_agblocks is aligned at stripe boundary
 678          * If sb_agblocks is NOT aligned turn off m_dalign since
 679          * allocator alignment is within an ag, therefore ag has
 680          * to be aligned at stripe boundary.
 681          */
 682         update_flags = 0LL;
 683         if (mp->m_dalign && !(mfsi_flags & XFS_MFSI_SECOND)) {
 684                 /*
 685                  * If stripe unit and stripe width are not multiples
 686                  * of the fs blocksize turn off alignment.
 687                  */
 688                 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
 689                     (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
 690                         if (mp->m_flags & XFS_MOUNT_RETERR) {
 691                                 cmn_err(CE_WARN,
 692                                         "XFS: alignment check 1 failed");
 693                                 error = XFS_ERROR(EINVAL);
 694                                 goto error1;
 695                         }
 696                         mp->m_dalign = mp->m_swidth = 0;
 697                 } else {
 698                         /*
 699                          * Convert the stripe unit and width to FSBs.
 700                          */
 701                         mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
 702                         if (mp->m_dalign && (sbp->sb_agblocks % mp->m_dalign)) {
 703                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
 704                                         error = XFS_ERROR(EINVAL);
 705                                         goto error1;
 706                                 }
 707                                 xfs_fs_cmn_err(CE_WARN, mp,
 708 "stripe alignment turned off: sunit(%d)/swidth(%d) incompatible with agsize(%d)",
 709                                         mp->m_dalign, mp->m_swidth,
 710                                         sbp->sb_agblocks);
 711
 712                                 mp->m_dalign = 0;
 713                                 mp->m_swidth = 0;
 714                         } else if (mp->m_dalign) {
 715                                 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
 716                         } else {
 717                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
 718                                         xfs_fs_cmn_err(CE_WARN, mp,
 719 "stripe alignment turned off: sunit(%d) less than bsize(%d)",
 720                                                 mp->m_dalign,
 721                                                 mp->m_blockmask +1);
 722                                         error = XFS_ERROR(EINVAL);
 723                                         goto error1;
 724                                 }
 725                                 mp->m_swidth = 0;
 726                         }
 727                 }
 728
 729                 /*
 730                  * Update superblock with new values
 731                  * and log changes
 732                  */
 733                 if (XFS_SB_VERSION_HASDALIGN(sbp)) {
 734                         if (sbp->sb_unit != mp->m_dalign) {
 735                                 sbp->sb_unit = mp->m_dalign;
 736                                 update_flags |= XFS_SB_UNIT;
 737                         }
 738                         if (sbp->sb_width != mp->m_swidth) {
 739                                 sbp->sb_width = mp->m_swidth;
 740                                 update_flags |= XFS_SB_WIDTH;
 741                         }
 742                 }
 743         } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
 744                     XFS_SB_VERSION_HASDALIGN(&mp->m_sb)) {
 745                         mp->m_dalign = sbp->sb_unit;
 746                         mp->m_swidth = sbp->sb_width;
 747         }
 748
 749         xfs_alloc_compute_maxlevels(mp);
 750         xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
 751         xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
 752         xfs_ialloc_compute_maxlevels(mp);
 753
 754         if (sbp->sb_imax_pct) {
 755                 __uint64_t      icount;
 756
 757                 /* Make sure the maximum inode count is a multiple of the
 758                  * units we allocate inodes in.
 759                  */
 760
 761                 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
 762                 do_div(icount, 100);
 763                 do_div(icount, mp->m_ialloc_blks);
 764                 mp->m_maxicount = (icount * mp->m_ialloc_blks)  <<
 765                                    sbp->sb_inopblog;
 766         } else
 767                 mp->m_maxicount = 0;
 768
 769         mp->m_maxioffset = xfs_max_file_offset(sbp->sb_blocklog);
 770
 771         /*
 772          * XFS uses the uuid from the superblock as the unique
 773          * identifier for fsid.  We can not use the uuid from the volume
 774          * since a single partition filesystem is identical to a single
 775          * partition volume/filesystem.
 776          */
 777         if ((mfsi_flags & XFS_MFSI_SECOND) == 0 &&
 778             (mp->m_flags & XFS_MOUNT_NOUUID) == 0) {
 779                 if (xfs_uuid_mount(mp)) {
 780                         error = XFS_ERROR(EINVAL);
 781                         goto error1;
 782                 }
 783                 uuid_mounted=1;
 784                 ret64 = uuid_hash64(&sbp->sb_uuid);
 785                 memcpy(&vfsp->vfs_fsid, &ret64, sizeof(ret64));
 786         }
 787
 788         /*
 789          * Set the default minimum read and write sizes unless
 790          * already specified in a mount option.
 791          * We use smaller I/O sizes when the file system
 792          * is being used for NFS service (wsync mount option).
 793          */
 794         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
 795                 if (mp->m_flags & XFS_MOUNT_WSYNC) {
 796                         readio_log = XFS_WSYNC_READIO_LOG;
 797                         writeio_log = XFS_WSYNC_WRITEIO_LOG;
 798                 } else {
 799                         readio_log = XFS_READIO_LOG_LARGE;
 800                         writeio_log = XFS_WRITEIO_LOG_LARGE;
 801                 }
 802         } else {
 803                 readio_log = mp->m_readio_log;
 804                 writeio_log = mp->m_writeio_log;
 805         }
 806
 807         /*
 808          * Set the number of readahead buffers to use based on
 809          * physical memory size.
 810          */
 811         if (xfs_physmem <= 4096)                /* <= 16MB */
 812                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_16MB;
 813         else if (xfs_physmem <= 8192)   /* <= 32MB */
 814                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_32MB;
 815         else
 816                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_K32;
 817         if (sbp->sb_blocklog > readio_log) {
 818                 mp->m_readio_log = sbp->sb_blocklog;
 819         } else {
 820                 mp->m_readio_log = readio_log;
 821         }
 822         mp->m_readio_blocks = 1 << (mp->m_readio_log - sbp->sb_blocklog);
 823         if (sbp->sb_blocklog > writeio_log) {
 824                 mp->m_writeio_log = sbp->sb_blocklog;
 825         } else {
 826                 mp->m_writeio_log = writeio_log;
 827         }
 828         mp->m_writeio_blocks = 1 << (mp->m_writeio_log - sbp->sb_blocklog);
 829
 830         /*
 831          * Set the inode cluster size based on the physical memory
 832          * size.  This may still be overridden by the file system
 833          * block size if it is larger than the chosen cluster size.
 834          */
 835         if (xfs_physmem <= btoc(32 * 1024 * 1024)) { /* <= 32 MB */
 836                 mp->m_inode_cluster_size = XFS_INODE_SMALL_CLUSTER_SIZE;
 837         } else {
 838                 mp->m_inode_cluster_size = XFS_INODE_BIG_CLUSTER_SIZE;
 839         }
 840         /*
 841          * Set whether we're using inode alignment.
 842          */
 843         if (XFS_SB_VERSION_HASALIGN(&mp->m_sb) &&
 844             mp->m_sb.sb_inoalignmt >=
 845             XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size))
 846                 mp->m_inoalign_mask = mp->m_sb.sb_inoalignmt - 1;
 847         else
 848                 mp->m_inoalign_mask = 0;
 849         /*
 850          * If we are using stripe alignment, check whether
 851          * the stripe unit is a multiple of the inode alignment
 852          */
 853         if (mp->m_dalign && mp->m_inoalign_mask &&
 854             !(mp->m_dalign & mp->m_inoalign_mask))
 855                 mp->m_sinoalign = mp->m_dalign;
 856         else
 857                 mp->m_sinoalign = 0;
 858         /*
 859          * Check that the data (and log if separate) are an ok size.
 860          */
 861         d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
 862         if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
 863                 cmn_err(CE_WARN, "XFS: size check 1 failed");
 864                 error = XFS_ERROR(E2BIG);
 865                 goto error1;
 866         }
 867         error = xfs_read_buf(mp, mp->m_ddev_targp,
 868                              d - XFS_FSS_TO_BB(mp, 1),
 869                              XFS_FSS_TO_BB(mp, 1), 0, &bp);
 870         if (!error) {
 871                 xfs_buf_relse(bp);
 872         } else {
 873                 cmn_err(CE_WARN, "XFS: size check 2 failed");
 874                 if (error == ENOSPC) {
 875                         error = XFS_ERROR(E2BIG);
 876                 }
 877                 goto error1;
 878         }
 879
 880         if (((mfsi_flags & XFS_MFSI_CLIENT) == 0) &&
 881             mp->m_logdev_targp != mp->m_ddev_targp) {
 882                 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
 883                 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
 884                         cmn_err(CE_WARN, "XFS: size check 3 failed");
 885                         error = XFS_ERROR(E2BIG);
 886                         goto error1;
 887                 }
 888                 error = xfs_read_buf(mp, mp->m_logdev_targp,
 889                                      d - XFS_FSB_TO_BB(mp, 1),
 890                                      XFS_FSB_TO_BB(mp, 1), 0, &bp);
 891                 if (!error) {
 892                         xfs_buf_relse(bp);
 893                 } else {
 894                         cmn_err(CE_WARN, "XFS: size check 3 failed");
 895                         if (error == ENOSPC) {
 896                                 error = XFS_ERROR(E2BIG);
 897                         }
 898                         goto error1;
 899                 }
 900         }
 901
 902         /*
 903          * Initialize realtime fields in the mount structure
 904          */
 905         if ((error = xfs_rtmount_init(mp))) {
 906                 cmn_err(CE_WARN, "XFS: RT mount failed");
 907                 goto error1;
 908         }
 909
 910         /*
 911          * For client case we are done now
 912          */
 913         if (mfsi_flags & XFS_MFSI_CLIENT) {
 914                 return 0;
 915         }
 916
 917         /*
 918          *  Copies the low order bits of the timestamp and the randomly
 919          *  set "sequence" number out of a UUID.
 920          */
 921         uuid_getnodeuniq(&sbp->sb_uuid, mp->m_fixedfsid);
 922
 923         /*
 924          *  The vfs structure needs to have a file system independent
 925          *  way of checking for the invariant file system ID.  Since it
 926          *  can't look at mount structures it has a pointer to the data
 927          *  in the mount structure.
 928          *
 929          *  File systems that don't support user level file handles (i.e.
 930          *  all of them except for XFS) will leave vfs_altfsid as NULL.
 931          */
 932         vfsp->vfs_altfsid = (xfs_fsid_t *)mp->m_fixedfsid;
 933         mp->m_dmevmask = 0;     /* not persistent; set after each mount */
 934
 935         xfs_dir_mount(mp);
 936
 937         /*
 938          * Initialize the attribute manager's entries.
 939          */
 940         mp->m_attr_magicpct = (mp->m_sb.sb_blocksize * 37) / 100;
 941
 942         /*
 943          * Initialize the precomputed transaction reservations values.
 944          */
 945         xfs_trans_init(mp);
 946
 947         /*
 948          * Allocate and initialize the inode hash table for this
 949          * file system.
 950          */
 951         xfs_ihash_init(mp);
 952         xfs_chash_init(mp);
 953
 954         /*
 955          * Allocate and initialize the per-ag data.
 956          */
 957         init_rwsem(&mp->m_peraglock);
 958         mp->m_perag =
 959                 kmem_zalloc(sbp->sb_agcount * sizeof(xfs_perag_t), KM_SLEEP);
 960
 961         mp->m_maxagi = xfs_initialize_perag(vfsp, mp, sbp->sb_agcount);
 962
 963         /*
 964          * log's mount-time initialization. Perform 1st part recovery if needed
 965          */
 966         if (likely(sbp->sb_logblocks > 0)) {    /* check for volume case */
 967                 error = xfs_log_mount(mp, mp->m_logdev_targp,
 968                                       XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
 969                                       XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
 970                 if (error) {
 971                         cmn_err(CE_WARN, "XFS: log mount failed");
 972                         goto error2;
 973                 }
 974         } else {        /* No log has been defined */
 975                 cmn_err(CE_WARN, "XFS: no log defined");
 976                 XFS_ERROR_REPORT("xfs_mountfs_int(1)", XFS_ERRLEVEL_LOW, mp);
 977                 error = XFS_ERROR(EFSCORRUPTED);
 978                 goto error2;
 979         }
 980
 981         /*
 982          * Get and sanity-check the root inode.
 983          * Save the pointer to it in the mount structure.
 984          */
 985         error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip, 0);
 986         if (error) {
 987                 cmn_err(CE_WARN, "XFS: failed to read root inode");
 988                 goto error3;
 989         }
 990
 991         ASSERT(rip != NULL);
 992         rvp = XFS_ITOV(rip);
 993
 994         if (unlikely((rip->i_d.di_mode & S_IFMT) != S_IFDIR)) {
 995                 cmn_err(CE_WARN, "XFS: corrupted root inode");
 996                 cmn_err(CE_WARN, "Device %s - root %llu is not a directory",
 997                         XFS_BUFTARG_NAME(mp->m_ddev_targp),
 998                         (unsigned long long)rip->i_ino);
 999                 xfs_iunlock(rip, XFS_ILOCK_EXCL);
1000                 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
1001                                  mp);
1002                 error = XFS_ERROR(EFSCORRUPTED);
1003                 goto error4;
1004         }
1005         mp->m_rootip = rip;     /* save it */
1006
1007         xfs_iunlock(rip, XFS_ILOCK_EXCL);
1008
1009         /*
1010          * Initialize realtime inode pointers in the mount structure
1011          */
1012         if ((error = xfs_rtmount_inodes(mp))) {
1013                 /*
1014                  * Free up the root inode.
1015                  */
1016                 cmn_err(CE_WARN, "XFS: failed to read RT inodes");
1017                 goto error4;
1018         }
1019
1020         /*
1021          * If fs is not mounted readonly, then update the superblock
1022          * unit and width changes.
1023          */
1024         if (update_flags && !(vfsp->vfs_flag & VFS_RDONLY))
1025                 xfs_mount_log_sbunit(mp, update_flags);
1026
1027         /*
1028          * Initialise the XFS quota management subsystem for this mount
1029          */
1030         if ((error = XFS_QM_INIT(mp, &quotamount, &quotaflags)))
1031                 goto error4;
1032
1033         /*
1034          * Finish recovering the file system.  This part needed to be
1035          * delayed until after the root and real-time bitmap inodes
1036          * were consistently read in.
1037          */
1038         error = xfs_log_mount_finish(mp, mfsi_flags);
1039         if (error) {
1040                 cmn_err(CE_WARN, "XFS: log mount finish failed");
1041                 goto error4;
1042         }
1043
1044         /*
1045          * Complete the quota initialisation, post-log-replay component.
1046          */
1047         if ((error = XFS_QM_MOUNT(mp, quotamount, quotaflags, mfsi_flags)))
1048                 goto error4;
1049
1050         return 0;
1051
1052  error4:
1053         /*
1054          * Free up the root inode.
1055          */
1056         VN_RELE(rvp);
1057  error3:
1058         xfs_log_unmount_dealloc(mp);
1059  error2:
1060         xfs_ihash_free(mp);
1061         xfs_chash_free(mp);
1062         for (agno = 0; agno < sbp->sb_agcount; agno++)
1063                 if (mp->m_perag[agno].pagb_list)
1064                         kmem_free(mp->m_perag[agno].pagb_list,
1065                           sizeof(xfs_perag_busy_t) * XFS_PAGB_NUM_SLOTS);
1066         kmem_free(mp->m_perag, sbp->sb_agcount * sizeof(xfs_perag_t));
1067         mp->m_perag = NULL;
1068         /* FALLTHROUGH */
1069  error1:
1070         if (uuid_mounted)
1071                 xfs_uuid_unmount(mp);
1072         xfs_freesb(mp);
1073         return error;
1074 }
1075
1076 /*
1077  * xfs_unmountfs
1078  *
1079  * This flushes out the inodes,dquots and the superblock, unmounts the
1080  * log and makes sure that incore structures are freed.
1081  */
1082 int
1083 xfs_unmountfs(xfs_mount_t *mp, struct cred *cr)
1084 {
1085         struct bhv_vfs  *vfsp = XFS_MTOVFS(mp);
1086 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1087         int64_t         fsid;
1088 #endif
1089
1090         xfs_iflush_all(mp);
1091
1092         XFS_QM_DQPURGEALL(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_UMOUNTING);
1093
1094         /*
1095          * Flush out the log synchronously so that we know for sure
1096          * that nothing is pinned.  This is important because bflush()
1097          * will skip pinned buffers.
1098          */
1099         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1100
1101         xfs_binval(mp->m_ddev_targp);
1102         if (mp->m_rtdev_targp) {
1103                 xfs_binval(mp->m_rtdev_targp);
1104         }
1105
1106         xfs_unmountfs_writesb(mp);
1107
1108         xfs_unmountfs_wait(mp);                 /* wait for async bufs */
1109
1110         xfs_log_unmount(mp);                    /* Done! No more fs ops. */
1111
1112         xfs_freesb(mp);
1113
1114         /*
1115          * All inodes from this mount point should be freed.
1116          */
1117         ASSERT(mp->m_inodes == NULL);
1118
1119         xfs_unmountfs_close(mp, cr);
1120         if ((mp->m_flags & XFS_MOUNT_NOUUID) == 0)
1121                 xfs_uuid_unmount(mp);
1122
1123 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1124         /*
1125          * clear all error tags on this filesystem
1126          */
1127         memcpy(&fsid, &vfsp->vfs_fsid, sizeof(int64_t));
1128         xfs_errortag_clearall_umount(fsid, mp->m_fsname, 0);
1129 #endif
1130         XFS_IODONE(vfsp);
1131         xfs_mount_free(mp, 1);
1132         return 0;
1133 }
1134
1135 void
1136 xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
1137 {
1138         if (mp->m_logdev_targp != mp->m_ddev_targp)
1139                 xfs_free_buftarg(mp->m_logdev_targp, 1);
1140         if (mp->m_rtdev_targp)
1141                 xfs_free_buftarg(mp->m_rtdev_targp, 1);
1142         xfs_free_buftarg(mp->m_ddev_targp, 0);
1143 }
1144
1145 STATIC void
1146 xfs_unmountfs_wait(xfs_mount_t *mp)
1147 {
1148         if (mp->m_logdev_targp != mp->m_ddev_targp)
1149                 xfs_wait_buftarg(mp->m_logdev_targp);
1150         if (mp->m_rtdev_targp)
1151                 xfs_wait_buftarg(mp->m_rtdev_targp);
1152         xfs_wait_buftarg(mp->m_ddev_targp);
1153 }
1154
1155 int
1156 xfs_unmountfs_writesb(xfs_mount_t *mp)
1157 {
1158         xfs_buf_t       *sbp;
1159         xfs_sb_t        *sb;
1160         int             error = 0;
1161
1162         /*
1163          * skip superblock write if fs is read-only, or
1164          * if we are doing a forced umount.
1165          */
1166         sbp = xfs_getsb(mp, 0);
1167         if (!(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY ||
1168                 XFS_FORCED_SHUTDOWN(mp))) {
1169
1170                 xfs_icsb_sync_counters(mp);
1171
1172                 /*
1173                  * mark shared-readonly if desired
1174                  */
1175                 sb = XFS_BUF_TO_SBP(sbp);
1176                 if (mp->m_mk_sharedro) {
1177                         if (!(sb->sb_flags & XFS_SBF_READONLY))
1178                                 sb->sb_flags |= XFS_SBF_READONLY;
1179                         if (!XFS_SB_VERSION_HASSHARED(sb))
1180                                 XFS_SB_VERSION_ADDSHARED(sb);
1181                         xfs_fs_cmn_err(CE_NOTE, mp,
1182                                 "Unmounting, marking shared read-only");
1183                 }
1184                 XFS_BUF_UNDONE(sbp);
1185                 XFS_BUF_UNREAD(sbp);
1186                 XFS_BUF_UNDELAYWRITE(sbp);
1187                 XFS_BUF_WRITE(sbp);
1188                 XFS_BUF_UNASYNC(sbp);
1189                 ASSERT(XFS_BUF_TARGET(sbp) == mp->m_ddev_targp);
1190                 xfsbdstrat(mp, sbp);
1191                 /* Nevermind errors we might get here. */
1192                 error = xfs_iowait(sbp);
1193                 if (error)
1194                         xfs_ioerror_alert("xfs_unmountfs_writesb",
1195                                           mp, sbp, XFS_BUF_ADDR(sbp));
1196                 if (error && mp->m_mk_sharedro)
1197                         xfs_fs_cmn_err(CE_ALERT, mp, "Superblock write error detected while unmounting.  Filesystem may not be marked shared readonly");
1198         }
1199         xfs_buf_relse(sbp);
1200         return error;
1201 }
1202
1203 /*
1204  * xfs_mod_sb() can be used to copy arbitrary changes to the
1205  * in-core superblock into the superblock buffer to be logged.
1206  * It does not provide the higher level of locking that is
1207  * needed to protect the in-core superblock from concurrent
1208  * access.
1209  */
1210 void
1211 xfs_mod_sb(xfs_trans_t *tp, __int64_t fields)
1212 {
1213         xfs_buf_t       *bp;
1214         int             first;
1215         int             last;
1216         xfs_mount_t     *mp;
1217         xfs_sb_t        *sbp;
1218         xfs_sb_field_t  f;
1219
1220         ASSERT(fields);
1221         if (!fields)
1222                 return;
1223         mp = tp->t_mountp;
1224         bp = xfs_trans_getsb(tp, mp, 0);
1225         sbp = XFS_BUF_TO_SBP(bp);
1226         first = sizeof(xfs_sb_t);
1227         last = 0;
1228
1229         /* translate/copy */
1230
1231         xfs_xlatesb(XFS_BUF_PTR(bp), &(mp->m_sb), -1, fields);
1232
1233         /* find modified range */
1234
1235         f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
1236         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1237         first = xfs_sb_info[f].offset;
1238
1239         f = (xfs_sb_field_t)xfs_highbit64((__uint64_t)fields);
1240         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1241         last = xfs_sb_info[f + 1].offset - 1;
1242
1243         xfs_trans_log_buf(tp, bp, first, last);
1244 }
1245
1246 /*
1247  * In order to avoid ENOSPC-related deadlock caused by
1248  * out-of-order locking of AGF buffer (PV 947395), we place
1249  * constraints on the relationship among actual allocations for
1250  * data blocks, freelist blocks, and potential file data bmap
1251  * btree blocks. However, these restrictions may result in no
1252  * actual space allocated for a delayed extent, for example, a data
1253  * block in a certain AG is allocated but there is no additional
1254  * block for the additional bmap btree block due to a split of the
1255  * bmap btree of the file. The result of this may lead to an
1256  * infinite loop in xfssyncd when the file gets flushed to disk and
1257  * all delayed extents need to be actually allocated. To get around
1258  * this, we explicitly set aside a few blocks which will not be
1259  * reserved in delayed allocation. Considering the minimum number of
1260  * needed freelist blocks is 4 fsbs, a potential split of file's bmap
1261  * btree requires 1 fsb, so we set the number of set-aside blocks to 8.
1262 */
1263 #define SET_ASIDE_BLOCKS 8
1264
1265 /*
1266  * xfs_mod_incore_sb_unlocked() is a utility routine common used to apply
1267  * a delta to a specified field in the in-core superblock.  Simply
1268  * switch on the field indicated and apply the delta to that field.
1269  * Fields are not allowed to dip below zero, so if the delta would
1270  * do this do not apply it and return EINVAL.
1271  *
1272  * The SB_LOCK must be held when this routine is called.
1273  */
1274 int
1275 xfs_mod_incore_sb_unlocked(xfs_mount_t *mp, xfs_sb_field_t field,
1276                         int delta, int rsvd)
1277 {
1278         int             scounter;       /* short counter for 32 bit fields */
1279         long long       lcounter;       /* long counter for 64 bit fields */
1280         long long       res_used, rem;
1281
1282         /*
1283          * With the in-core superblock spin lock held, switch
1284          * on the indicated field.  Apply the delta to the
1285          * proper field.  If the fields value would dip below
1286          * 0, then do not apply the delta and return EINVAL.
1287          */
1288         switch (field) {
1289         case XFS_SBS_ICOUNT:
1290                 lcounter = (long long)mp->m_sb.sb_icount;
1291                 lcounter += delta;
1292                 if (lcounter < 0) {
1293                         ASSERT(0);
1294                         return XFS_ERROR(EINVAL);
1295                 }
1296                 mp->m_sb.sb_icount = lcounter;
1297                 return 0;
1298         case XFS_SBS_IFREE:
1299                 lcounter = (long long)mp->m_sb.sb_ifree;
1300                 lcounter += delta;
1301                 if (lcounter < 0) {
1302                         ASSERT(0);
1303                         return XFS_ERROR(EINVAL);
1304                 }
1305                 mp->m_sb.sb_ifree = lcounter;
1306                 return 0;
1307         case XFS_SBS_FDBLOCKS:
1308
1309                 lcounter = (long long)mp->m_sb.sb_fdblocks - SET_ASIDE_BLOCKS;
1310                 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1311
1312                 if (delta > 0) {                /* Putting blocks back */
1313                         if (res_used > delta) {
1314                                 mp->m_resblks_avail += delta;
1315                         } else {
1316                                 rem = delta - res_used;
1317                                 mp->m_resblks_avail = mp->m_resblks;
1318                                 lcounter += rem;
1319                         }
1320                 } else {                                /* Taking blocks away */
1321
1322                         lcounter += delta;
1323
1324                 /*
1325                  * If were out of blocks, use any available reserved blocks if
1326                  * were allowed to.
1327                  */
1328
1329                         if (lcounter < 0) {
1330                                 if (rsvd) {
1331                                         lcounter = (long long)mp->m_resblks_avail + delta;
1332                                         if (lcounter < 0) {
1333                                                 return XFS_ERROR(ENOSPC);
1334                                         }
1335                                         mp->m_resblks_avail = lcounter;
1336                                         return 0;
1337                                 } else {        /* not reserved */
1338                                         return XFS_ERROR(ENOSPC);
1339                                 }
1340                         }
1341                 }
1342
1343                 mp->m_sb.sb_fdblocks = lcounter + SET_ASIDE_BLOCKS;
1344                 return 0;
1345         case XFS_SBS_FREXTENTS:
1346                 lcounter = (long long)mp->m_sb.sb_frextents;
1347                 lcounter += delta;
1348                 if (lcounter < 0) {
1349                         return XFS_ERROR(ENOSPC);
1350                 }
1351                 mp->m_sb.sb_frextents = lcounter;
1352                 return 0;
1353         case XFS_SBS_DBLOCKS:
1354                 lcounter = (long long)mp->m_sb.sb_dblocks;
1355                 lcounter += delta;
1356                 if (lcounter < 0) {
1357                         ASSERT(0);
1358                         return XFS_ERROR(EINVAL);
1359                 }
1360                 mp->m_sb.sb_dblocks = lcounter;
1361                 return 0;
1362         case XFS_SBS_AGCOUNT:
1363                 scounter = mp->m_sb.sb_agcount;
1364                 scounter += delta;
1365                 if (scounter < 0) {
1366                         ASSERT(0);
1367                         return XFS_ERROR(EINVAL);
1368                 }
1369                 mp->m_sb.sb_agcount = scounter;
1370                 return 0;
1371         case XFS_SBS_IMAX_PCT:
1372                 scounter = mp->m_sb.sb_imax_pct;
1373                 scounter += delta;
1374                 if (scounter < 0) {
1375                         ASSERT(0);
1376                         return XFS_ERROR(EINVAL);
1377                 }
1378                 mp->m_sb.sb_imax_pct = scounter;
1379                 return 0;
1380         case XFS_SBS_REXTSIZE:
1381                 scounter = mp->m_sb.sb_rextsize;
1382                 scounter += delta;
1383                 if (scounter < 0) {
1384                         ASSERT(0);
1385                         return XFS_ERROR(EINVAL);
1386                 }
1387                 mp->m_sb.sb_rextsize = scounter;
1388                 return 0;
1389         case XFS_SBS_RBMBLOCKS:
1390                 scounter = mp->m_sb.sb_rbmblocks;
1391                 scounter += delta;
1392                 if (scounter < 0) {
1393                         ASSERT(0);
1394                         return XFS_ERROR(EINVAL);
1395                 }
1396                 mp->m_sb.sb_rbmblocks = scounter;
1397                 return 0;
1398         case XFS_SBS_RBLOCKS:
1399                 lcounter = (long long)mp->m_sb.sb_rblocks;
1400                 lcounter += delta;
1401                 if (lcounter < 0) {
1402                         ASSERT(0);
1403                         return XFS_ERROR(EINVAL);
1404                 }
1405                 mp->m_sb.sb_rblocks = lcounter;
1406                 return 0;
1407         case XFS_SBS_REXTENTS:
1408                 lcounter = (long long)mp->m_sb.sb_rextents;
1409                 lcounter += delta;
1410                 if (lcounter < 0) {
1411                         ASSERT(0);
1412                         return XFS_ERROR(EINVAL);
1413                 }
1414                 mp->m_sb.sb_rextents = lcounter;
1415                 return 0;
1416         case XFS_SBS_REXTSLOG:
1417                 scounter = mp->m_sb.sb_rextslog;
1418                 scounter += delta;
1419                 if (scounter < 0) {
1420                         ASSERT(0);
1421                         return XFS_ERROR(EINVAL);
1422                 }
1423                 mp->m_sb.sb_rextslog = scounter;
1424                 return 0;
1425         default:
1426                 ASSERT(0);
1427                 return XFS_ERROR(EINVAL);
1428         }
1429 }
1430
1431 /*
1432  * xfs_mod_incore_sb() is used to change a field in the in-core
1433  * superblock structure by the specified delta.  This modification
1434  * is protected by the SB_LOCK.  Just use the xfs_mod_incore_sb_unlocked()
1435  * routine to do the work.
1436  */
1437 int
1438 xfs_mod_incore_sb(xfs_mount_t *mp, xfs_sb_field_t field, int delta, int rsvd)
1439 {
1440         unsigned long   s;
1441         int     status;
1442
1443         /* check for per-cpu counters */
1444         switch (field) {
1445 #ifdef HAVE_PERCPU_SB
1446         case XFS_SBS_ICOUNT:
1447         case XFS_SBS_IFREE:
1448         case XFS_SBS_FDBLOCKS:
1449                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1450                         status = xfs_icsb_modify_counters(mp, field,
1451                                                         delta, rsvd);
1452                         break;
1453                 }
1454                 /* FALLTHROUGH */
1455 #endif
1456         default:
1457                 s = XFS_SB_LOCK(mp);
1458                 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1459                 XFS_SB_UNLOCK(mp, s);
1460                 break;
1461         }
1462
1463         return status;
1464 }
1465
1466 /*
1467  * xfs_mod_incore_sb_batch() is used to change more than one field
1468  * in the in-core superblock structure at a time.  This modification
1469  * is protected by a lock internal to this module.  The fields and
1470  * changes to those fields are specified in the array of xfs_mod_sb
1471  * structures passed in.
1472  *
1473  * Either all of the specified deltas will be applied or none of
1474  * them will.  If any modified field dips below 0, then all modifications
1475  * will be backed out and EINVAL will be returned.
1476  */
1477 int
1478 xfs_mod_incore_sb_batch(xfs_mount_t *mp, xfs_mod_sb_t *msb, uint nmsb, int rsvd)
1479 {
1480         unsigned long   s;
1481         int             status=0;
1482         xfs_mod_sb_t    *msbp;
1483
1484         /*
1485          * Loop through the array of mod structures and apply each
1486          * individually.  If any fail, then back out all those
1487          * which have already been applied.  Do all of this within
1488          * the scope of the SB_LOCK so that all of the changes will
1489          * be atomic.
1490          */
1491         s = XFS_SB_LOCK(mp);
1492         msbp = &msb[0];
1493         for (msbp = &msbp[0]; msbp < (msb + nmsb); msbp++) {
1494                 /*
1495                  * Apply the delta at index n.  If it fails, break
1496                  * from the loop so we'll fall into the undo loop
1497                  * below.
1498                  */
1499                 switch (msbp->msb_field) {
1500 #ifdef HAVE_PERCPU_SB
1501                 case XFS_SBS_ICOUNT:
1502                 case XFS_SBS_IFREE:
1503                 case XFS_SBS_FDBLOCKS:
1504                         if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1505                                 status = xfs_icsb_modify_counters_locked(mp,
1506                                                         msbp->msb_field,
1507                                                         msbp->msb_delta, rsvd);
1508                                 break;
1509                         }
1510                         /* FALLTHROUGH */
1511 #endif
1512                 default:
1513                         status = xfs_mod_incore_sb_unlocked(mp,
1514                                                 msbp->msb_field,
1515                                                 msbp->msb_delta, rsvd);
1516                         break;
1517                 }
1518
1519                 if (status != 0) {
1520                         break;
1521                 }
1522         }
1523
1524         /*
1525          * If we didn't complete the loop above, then back out
1526          * any changes made to the superblock.  If you add code
1527          * between the loop above and here, make sure that you
1528          * preserve the value of status. Loop back until
1529          * we step below the beginning of the array.  Make sure
1530          * we don't touch anything back there.
1531          */
1532         if (status != 0) {
1533                 msbp--;
1534                 while (msbp >= msb) {
1535                         switch (msbp->msb_field) {
1536 #ifdef HAVE_PERCPU_SB
1537                         case XFS_SBS_ICOUNT:
1538                         case XFS_SBS_IFREE:
1539                         case XFS_SBS_FDBLOCKS:
1540                                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1541                                         status =
1542                                             xfs_icsb_modify_counters_locked(mp,
1543                                                         msbp->msb_field,
1544                                                         -(msbp->msb_delta),
1545                                                         rsvd);
1546                                         break;
1547                                 }
1548                                 /* FALLTHROUGH */
1549 #endif
1550                         default:
1551                                 status = xfs_mod_incore_sb_unlocked(mp,
1552                                                         msbp->msb_field,
1553                                                         -(msbp->msb_delta),
1554                                                         rsvd);
1555                                 break;
1556                         }
1557                         ASSERT(status == 0);
1558                         msbp--;
1559                 }
1560         }
1561         XFS_SB_UNLOCK(mp, s);
1562         return status;
1563 }
1564
1565 /*
1566  * xfs_getsb() is called to obtain the buffer for the superblock.
1567  * The buffer is returned locked and read in from disk.
1568  * The buffer should be released with a call to xfs_brelse().
1569  *
1570  * If the flags parameter is BUF_TRYLOCK, then we'll only return
1571  * the superblock buffer if it can be locked without sleeping.
1572  * If it can't then we'll return NULL.
1573  */
1574 xfs_buf_t *
1575 xfs_getsb(
1576         xfs_mount_t     *mp,
1577         int             flags)
1578 {
1579         xfs_buf_t       *bp;
1580
1581         ASSERT(mp->m_sb_bp != NULL);
1582         bp = mp->m_sb_bp;
1583         if (flags & XFS_BUF_TRYLOCK) {
1584                 if (!XFS_BUF_CPSEMA(bp)) {
1585                         return NULL;
1586                 }
1587         } else {
1588                 XFS_BUF_PSEMA(bp, PRIBIO);
1589         }
1590         XFS_BUF_HOLD(bp);
1591         ASSERT(XFS_BUF_ISDONE(bp));
1592         return bp;
1593 }
1594
1595 /*
1596  * Used to free the superblock along various error paths.
1597  */
1598 void
1599 xfs_freesb(
1600         xfs_mount_t     *mp)
1601 {
1602         xfs_buf_t       *bp;
1603
1604         /*
1605          * Use xfs_getsb() so that the buffer will be locked
1606          * when we call xfs_buf_relse().
1607          */
1608         bp = xfs_getsb(mp, 0);
1609         XFS_BUF_UNMANAGE(bp);
1610         xfs_buf_relse(bp);
1611         mp->m_sb_bp = NULL;
1612 }
1613
1614 /*
1615  * See if the UUID is unique among mounted XFS filesystems.
1616  * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
1617  */
1618 STATIC int
1619 xfs_uuid_mount(
1620         xfs_mount_t     *mp)
1621 {
1622         if (uuid_is_nil(&mp->m_sb.sb_uuid)) {
1623                 cmn_err(CE_WARN,
1624                         "XFS: Filesystem %s has nil UUID - can't mount",
1625                         mp->m_fsname);
1626                 return -1;
1627         }
1628         if (!uuid_table_insert(&mp->m_sb.sb_uuid)) {
1629                 cmn_err(CE_WARN,
1630                         "XFS: Filesystem %s has duplicate UUID - can't mount",
1631                         mp->m_fsname);
1632                 return -1;
1633         }
1634         return 0;
1635 }
1636
1637 /*
1638  * Remove filesystem from the UUID table.
1639  */
1640 STATIC void
1641 xfs_uuid_unmount(
1642         xfs_mount_t     *mp)
1643 {
1644         uuid_table_remove(&mp->m_sb.sb_uuid);
1645 }
1646
1647 /*
1648  * Used to log changes to the superblock unit and width fields which could
1649  * be altered by the mount options. Only the first superblock is updated.
1650  */
1651 STATIC void
1652 xfs_mount_log_sbunit(
1653         xfs_mount_t     *mp,
1654         __int64_t       fields)
1655 {
1656         xfs_trans_t     *tp;
1657
1658         ASSERT(fields & (XFS_SB_UNIT|XFS_SB_WIDTH|XFS_SB_UUID));
1659
1660         tp = xfs_trans_alloc(mp, XFS_TRANS_SB_UNIT);
1661         if (xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1662                                 XFS_DEFAULT_LOG_COUNT)) {
1663                 xfs_trans_cancel(tp, 0);
1664                 return;
1665         }
1666         xfs_mod_sb(tp, fields);
1667         xfs_trans_commit(tp, 0, NULL);
1668 }
1669
1670
1671 #ifdef HAVE_PERCPU_SB
1672 /*
1673  * Per-cpu incore superblock counters
1674  *
1675  * Simple concept, difficult implementation
1676  *
1677  * Basically, replace the incore superblock counters with a distributed per cpu
1678  * counter for contended fields (e.g.  free block count).
1679  *
1680  * Difficulties arise in that the incore sb is used for ENOSPC checking, and
1681  * hence needs to be accurately read when we are running low on space. Hence
1682  * there is a method to enable and disable the per-cpu counters based on how
1683  * much "stuff" is available in them.
1684  *
1685  * Basically, a counter is enabled if there is enough free resource to justify
1686  * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
1687  * ENOSPC), then we disable the counters to synchronise all callers and
1688  * re-distribute the available resources.
1689  *
1690  * If, once we redistributed the available resources, we still get a failure,
1691  * we disable the per-cpu counter and go through the slow path.
1692  *
1693  * The slow path is the current xfs_mod_incore_sb() function.  This means that
1694  * when we disable a per-cpu counter, we need to drain it's resources back to
1695  * the global superblock. We do this after disabling the counter to prevent
1696  * more threads from queueing up on the counter.
1697  *
1698  * Essentially, this means that we still need a lock in the fast path to enable
1699  * synchronisation between the global counters and the per-cpu counters. This
1700  * is not a problem because the lock will be local to a CPU almost all the time
1701  * and have little contention except when we get to ENOSPC conditions.
1702  *
1703  * Basically, this lock becomes a barrier that enables us to lock out the fast
1704  * path while we do things like enabling and disabling counters and
1705  * synchronising the counters.
1706  *
1707  * Locking rules:
1708  *
1709  *      1. XFS_SB_LOCK() before picking up per-cpu locks
1710  *      2. per-cpu locks always picked up via for_each_online_cpu() order
1711  *      3. accurate counter sync requires XFS_SB_LOCK + per cpu locks
1712  *      4. modifying per-cpu counters requires holding per-cpu lock
1713  *      5. modifying global counters requires holding XFS_SB_LOCK
1714  *      6. enabling or disabling a counter requires holding the XFS_SB_LOCK
1715  *         and _none_ of the per-cpu locks.
1716  *
1717  * Disabled counters are only ever re-enabled by a balance operation
1718  * that results in more free resources per CPU than a given threshold.
1719  * To ensure counters don't remain disabled, they are rebalanced when
1720  * the global resource goes above a higher threshold (i.e. some hysteresis
1721  * is present to prevent thrashing).
1722  */
1723
1724 #ifdef CONFIG_HOTPLUG_CPU
1725 /*
1726  * hot-plug CPU notifier support.
1727  *
1728  * We need a notifier per filesystem as we need to be able to identify
1729  * the filesystem to balance the counters out. This is achieved by
1730  * having a notifier block embedded in the xfs_mount_t and doing pointer
1731  * magic to get the mount pointer from the notifier block address.
1732  */
1733 STATIC int
1734 xfs_icsb_cpu_notify(
1735         struct notifier_block *nfb,
1736         unsigned long action,
1737         void *hcpu)
1738 {
1739         xfs_icsb_cnts_t *cntp;
1740         xfs_mount_t     *mp;
1741         int             s;
1742
1743         mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
1744         cntp = (xfs_icsb_cnts_t *)
1745                         per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
1746         switch (action) {
1747         case CPU_UP_PREPARE:
1748                 /* Easy Case - initialize the area and locks, and
1749                  * then rebalance when online does everything else for us. */
1750                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1751                 break;
1752         case CPU_ONLINE:
1753                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
1754                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
1755                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
1756                 break;
1757         case CPU_DEAD:
1758                 /* Disable all the counters, then fold the dead cpu's
1759                  * count into the total on the global superblock and
1760                  * re-enable the counters. */
1761                 s = XFS_SB_LOCK(mp);
1762                 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
1763                 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
1764                 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
1765
1766                 mp->m_sb.sb_icount += cntp->icsb_icount;
1767                 mp->m_sb.sb_ifree += cntp->icsb_ifree;
1768                 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
1769
1770                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1771
1772                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, XFS_ICSB_SB_LOCKED);
1773                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, XFS_ICSB_SB_LOCKED);
1774                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, XFS_ICSB_SB_LOCKED);
1775                 XFS_SB_UNLOCK(mp, s);
1776                 break;
1777         }
1778
1779         return NOTIFY_OK;
1780 }
1781 #endif /* CONFIG_HOTPLUG_CPU */
1782
1783 int
1784 xfs_icsb_init_counters(
1785         xfs_mount_t     *mp)
1786 {
1787         xfs_icsb_cnts_t *cntp;
1788         int             i;
1789
1790         mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
1791         if (mp->m_sb_cnts == NULL)
1792                 return -ENOMEM;
1793
1794 #ifdef CONFIG_HOTPLUG_CPU
1795         mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
1796         mp->m_icsb_notifier.priority = 0;
1797         register_hotcpu_notifier(&mp->m_icsb_notifier);
1798 #endif /* CONFIG_HOTPLUG_CPU */
1799
1800         for_each_online_cpu(i) {
1801                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1802                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1803         }
1804         /*
1805          * start with all counters disabled so that the
1806          * initial balance kicks us off correctly
1807          */
1808         mp->m_icsb_counters = -1;
1809         return 0;
1810 }
1811
1812 STATIC void
1813 xfs_icsb_destroy_counters(
1814         xfs_mount_t     *mp)
1815 {
1816         if (mp->m_sb_cnts) {
1817                 unregister_hotcpu_notifier(&mp->m_icsb_notifier);
1818                 free_percpu(mp->m_sb_cnts);
1819         }
1820 }
1821
1822 STATIC inline void
1823 xfs_icsb_lock_cntr(
1824         xfs_icsb_cnts_t *icsbp)
1825 {
1826         while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
1827                 ndelay(1000);
1828         }
1829 }
1830
1831 STATIC inline void
1832 xfs_icsb_unlock_cntr(
1833         xfs_icsb_cnts_t *icsbp)
1834 {
1835         clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
1836 }
1837
1838
1839 STATIC inline void
1840 xfs_icsb_lock_all_counters(
1841         xfs_mount_t     *mp)
1842 {
1843         xfs_icsb_cnts_t *cntp;
1844         int             i;
1845
1846         for_each_online_cpu(i) {
1847                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1848                 xfs_icsb_lock_cntr(cntp);
1849         }
1850 }
1851
1852 STATIC inline void
1853 xfs_icsb_unlock_all_counters(
1854         xfs_mount_t     *mp)
1855 {
1856         xfs_icsb_cnts_t *cntp;
1857         int             i;
1858
1859         for_each_online_cpu(i) {
1860                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1861                 xfs_icsb_unlock_cntr(cntp);
1862         }
1863 }
1864
1865 STATIC void
1866 xfs_icsb_count(
1867         xfs_mount_t     *mp,
1868         xfs_icsb_cnts_t *cnt,
1869         int             flags)
1870 {
1871         xfs_icsb_cnts_t *cntp;
1872         int             i;
1873
1874         memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
1875
1876         if (!(flags & XFS_ICSB_LAZY_COUNT))
1877                 xfs_icsb_lock_all_counters(mp);
1878
1879         for_each_online_cpu(i) {
1880                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1881                 cnt->icsb_icount += cntp->icsb_icount;
1882                 cnt->icsb_ifree += cntp->icsb_ifree;
1883                 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
1884         }
1885
1886         if (!(flags & XFS_ICSB_LAZY_COUNT))
1887                 xfs_icsb_unlock_all_counters(mp);
1888 }
1889
1890 STATIC int
1891 xfs_icsb_counter_disabled(
1892         xfs_mount_t     *mp,
1893         xfs_sb_field_t  field)
1894 {
1895         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1896         return test_bit(field, &mp->m_icsb_counters);
1897 }
1898
1899 STATIC int
1900 xfs_icsb_disable_counter(
1901         xfs_mount_t     *mp,
1902         xfs_sb_field_t  field)
1903 {
1904         xfs_icsb_cnts_t cnt;
1905
1906         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1907
1908         xfs_icsb_lock_all_counters(mp);
1909         if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
1910                 /* drain back to superblock */
1911
1912                 xfs_icsb_count(mp, &cnt, XFS_ICSB_SB_LOCKED|XFS_ICSB_LAZY_COUNT);
1913                 switch(field) {
1914                 case XFS_SBS_ICOUNT:
1915                         mp->m_sb.sb_icount = cnt.icsb_icount;
1916                         break;
1917                 case XFS_SBS_IFREE:
1918                         mp->m_sb.sb_ifree = cnt.icsb_ifree;
1919                         break;
1920                 case XFS_SBS_FDBLOCKS:
1921                         mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1922                         break;
1923                 default:
1924                         BUG();
1925                 }
1926         }
1927
1928         xfs_icsb_unlock_all_counters(mp);
1929
1930         return 0;
1931 }
1932
1933 STATIC void
1934 xfs_icsb_enable_counter(
1935         xfs_mount_t     *mp,
1936         xfs_sb_field_t  field,
1937         uint64_t        count,
1938         uint64_t        resid)
1939 {
1940         xfs_icsb_cnts_t *cntp;
1941         int             i;
1942
1943         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1944
1945         xfs_icsb_lock_all_counters(mp);
1946         for_each_online_cpu(i) {
1947                 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
1948                 switch (field) {
1949                 case XFS_SBS_ICOUNT:
1950                         cntp->icsb_icount = count + resid;
1951                         break;
1952                 case XFS_SBS_IFREE:
1953                         cntp->icsb_ifree = count + resid;
1954                         break;
1955                 case XFS_SBS_FDBLOCKS:
1956                         cntp->icsb_fdblocks = count + resid;
1957                         break;
1958                 default:
1959                         BUG();
1960                         break;
1961                 }
1962                 resid = 0;
1963         }
1964         clear_bit(field, &mp->m_icsb_counters);
1965         xfs_icsb_unlock_all_counters(mp);
1966 }
1967
1968 STATIC void
1969 xfs_icsb_sync_counters_int(
1970         xfs_mount_t     *mp,
1971         int             flags)
1972 {
1973         xfs_icsb_cnts_t cnt;
1974         int             s;
1975
1976         /* Pass 1: lock all counters */
1977         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
1978                 s = XFS_SB_LOCK(mp);
1979
1980         xfs_icsb_count(mp, &cnt, flags);
1981
1982         /* Step 3: update mp->m_sb fields */
1983         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
1984                 mp->m_sb.sb_icount = cnt.icsb_icount;
1985         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
1986                 mp->m_sb.sb_ifree = cnt.icsb_ifree;
1987         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
1988                 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1989
1990         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
1991                 XFS_SB_UNLOCK(mp, s);
1992 }
1993
1994 /*
1995  * Accurate update of per-cpu counters to incore superblock
1996  */
1997 STATIC void
1998 xfs_icsb_sync_counters(
1999         xfs_mount_t     *mp)
2000 {
2001         xfs_icsb_sync_counters_int(mp, 0);
2002 }
2003
2004 /*
2005  * lazy addition used for things like df, background sb syncs, etc
2006  */
2007 void
2008 xfs_icsb_sync_counters_lazy(
2009         xfs_mount_t     *mp)
2010 {
2011         xfs_icsb_sync_counters_int(mp, XFS_ICSB_LAZY_COUNT);
2012 }
2013
2014 /*
2015  * Balance and enable/disable counters as necessary.
2016  *
2017  * Thresholds for re-enabling counters are somewhat magic.
2018  * inode counts are chosen to be the same number as single
2019  * on disk allocation chunk per CPU, and free blocks is
2020  * something far enough zero that we aren't going thrash
2021  * when we get near ENOSPC.
2022  */
2023 #define XFS_ICSB_INO_CNTR_REENABLE      64
2024 #define XFS_ICSB_FDBLK_CNTR_REENABLE    512
2025 STATIC void
2026 xfs_icsb_balance_counter(
2027         xfs_mount_t     *mp,
2028         xfs_sb_field_t  field,
2029         int             flags)
2030 {
2031         uint64_t        count, resid;
2032         int             weight = num_online_cpus();
2033         int             s;
2034
2035         if (!(flags & XFS_ICSB_SB_LOCKED))
2036                 s = XFS_SB_LOCK(mp);
2037
2038         /* disable counter and sync counter */
2039         xfs_icsb_disable_counter(mp, field);
2040
2041         /* update counters  - first CPU gets residual*/
2042         switch (field) {
2043         case XFS_SBS_ICOUNT:
2044                 count = mp->m_sb.sb_icount;
2045                 resid = do_div(count, weight);
2046                 if (count < XFS_ICSB_INO_CNTR_REENABLE)
2047                         goto out;
2048                 break;
2049         case XFS_SBS_IFREE:
2050                 count = mp->m_sb.sb_ifree;
2051                 resid = do_div(count, weight);
2052                 if (count < XFS_ICSB_INO_CNTR_REENABLE)
2053                         goto out;
2054                 break;
2055         case XFS_SBS_FDBLOCKS:
2056                 count = mp->m_sb.sb_fdblocks;
2057                 resid = do_div(count, weight);
2058                 if (count < XFS_ICSB_FDBLK_CNTR_REENABLE)
2059                         goto out;
2060                 break;
2061         default:
2062                 BUG();
2063                 count = resid = 0;      /* quiet, gcc */
2064                 break;
2065         }
2066
2067         xfs_icsb_enable_counter(mp, field, count, resid);
2068 out:
2069         if (!(flags & XFS_ICSB_SB_LOCKED))
2070                 XFS_SB_UNLOCK(mp, s);
2071 }
2072
2073 STATIC int
2074 xfs_icsb_modify_counters_int(
2075         xfs_mount_t     *mp,
2076         xfs_sb_field_t  field,
2077         int             delta,
2078         int             rsvd,
2079         int             flags)
2080 {
2081         xfs_icsb_cnts_t *icsbp;
2082         long long       lcounter;       /* long counter for 64 bit fields */
2083         int             cpu, s, locked = 0;
2084         int             ret = 0, balance_done = 0;
2085
2086 again:
2087         cpu = get_cpu();
2088         icsbp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, cpu),
2089         xfs_icsb_lock_cntr(icsbp);
2090         if (unlikely(xfs_icsb_counter_disabled(mp, field)))
2091                 goto slow_path;
2092
2093         switch (field) {
2094         case XFS_SBS_ICOUNT:
2095                 lcounter = icsbp->icsb_icount;
2096                 lcounter += delta;
2097                 if (unlikely(lcounter < 0))
2098                         goto slow_path;
2099                 icsbp->icsb_icount = lcounter;
2100                 break;
2101
2102         case XFS_SBS_IFREE:
2103                 lcounter = icsbp->icsb_ifree;
2104                 lcounter += delta;
2105                 if (unlikely(lcounter < 0))
2106                         goto slow_path;
2107                 icsbp->icsb_ifree = lcounter;
2108                 break;
2109
2110         case XFS_SBS_FDBLOCKS:
2111                 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
2112
2113                 lcounter = icsbp->icsb_fdblocks;
2114                 lcounter += delta;
2115                 if (unlikely(lcounter < 0))
2116                         goto slow_path;
2117                 icsbp->icsb_fdblocks = lcounter;
2118                 break;
2119         default:
2120                 BUG();
2121                 break;
2122         }
2123         xfs_icsb_unlock_cntr(icsbp);
2124         put_cpu();
2125         if (locked)
2126                 XFS_SB_UNLOCK(mp, s);
2127         return 0;
2128
2129         /*
2130          * The slow path needs to be run with the SBLOCK
2131          * held so that we prevent other threads from
2132          * attempting to run this path at the same time.
2133          * this provides exclusion for the balancing code,
2134          * and exclusive fallback if the balance does not
2135          * provide enough resources to continue in an unlocked
2136          * manner.
2137          */
2138 slow_path:
2139         xfs_icsb_unlock_cntr(icsbp);
2140         put_cpu();
2141
2142         /* need to hold superblock incase we need
2143          * to disable a counter */
2144         if (!(flags & XFS_ICSB_SB_LOCKED)) {
2145                 s = XFS_SB_LOCK(mp);
2146                 locked = 1;
2147                 flags |= XFS_ICSB_SB_LOCKED;
2148         }
2149         if (!balance_done) {
2150                 xfs_icsb_balance_counter(mp, field, flags);
2151                 balance_done = 1;
2152                 goto again;
2153         } else {
2154                 /*
2155                  * we might not have enough on this local
2156                  * cpu to allocate for a bulk request.
2157                  * We need to drain this field from all CPUs
2158                  * and disable the counter fastpath
2159                  */
2160                 xfs_icsb_disable_counter(mp, field);
2161         }
2162
2163         ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
2164
2165         if (locked)
2166                 XFS_SB_UNLOCK(mp, s);
2167         return ret;
2168 }
2169
2170 STATIC int
2171 xfs_icsb_modify_counters(
2172         xfs_mount_t     *mp,
2173         xfs_sb_field_t  field,
2174         int             delta,
2175         int             rsvd)
2176 {
2177         return xfs_icsb_modify_counters_int(mp, field, delta, rsvd, 0);
2178 }
2179
2180 /*
2181  * Called when superblock is already locked
2182  */
2183 STATIC int
2184 xfs_icsb_modify_counters_locked(
2185         xfs_mount_t     *mp,
2186         xfs_sb_field_t  field,
2187         int             delta,
2188         int             rsvd)
2189 {
2190         return xfs_icsb_modify_counters_int(mp, field, delta,
2191                                                 rsvd, XFS_ICSB_SB_LOCKED);
2192 }
2193 #endif