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,
  56                                                 int, int);
  57 STATIC void     xfs_icsb_sync_counters(xfs_mount_t *);
  58 STATIC int      xfs_icsb_modify_counters(xfs_mount_t *, xfs_sb_field_t,
  59                                                 int64_t, int);
  60 STATIC int      xfs_icsb_disable_counter(xfs_mount_t *, xfs_sb_field_t);
  61
  62 #else
  63
  64 #define xfs_icsb_destroy_counters(mp)                   do { } while (0)
  65 #define xfs_icsb_balance_counter(mp, a, b, c)           do { } while (0)
  66 #define xfs_icsb_sync_counters(mp)                      do { } while (0)
  67 #define xfs_icsb_modify_counters(mp, a, b, c)           do { } while (0)
  68
  69 #endif
  70
  71 static const struct {
  72         short offset;
  73         short type;     /* 0 = integer
  74                          * 1 = binary / string (no translation)
  75                          */
  76 } xfs_sb_info[] = {
  77     { offsetof(xfs_sb_t, sb_magicnum),   0 },
  78     { offsetof(xfs_sb_t, sb_blocksize),  0 },
  79     { offsetof(xfs_sb_t, sb_dblocks),    0 },
  80     { offsetof(xfs_sb_t, sb_rblocks),    0 },
  81     { offsetof(xfs_sb_t, sb_rextents),   0 },
  82     { offsetof(xfs_sb_t, sb_uuid),       1 },
  83     { offsetof(xfs_sb_t, sb_logstart),   0 },
  84     { offsetof(xfs_sb_t, sb_rootino),    0 },
  85     { offsetof(xfs_sb_t, sb_rbmino),     0 },
  86     { offsetof(xfs_sb_t, sb_rsumino),    0 },
  87     { offsetof(xfs_sb_t, sb_rextsize),   0 },
  88     { offsetof(xfs_sb_t, sb_agblocks),   0 },
  89     { offsetof(xfs_sb_t, sb_agcount),    0 },
  90     { offsetof(xfs_sb_t, sb_rbmblocks),  0 },
  91     { offsetof(xfs_sb_t, sb_logblocks),  0 },
  92     { offsetof(xfs_sb_t, sb_versionnum), 0 },
  93     { offsetof(xfs_sb_t, sb_sectsize),   0 },
  94     { offsetof(xfs_sb_t, sb_inodesize),  0 },
  95     { offsetof(xfs_sb_t, sb_inopblock),  0 },
  96     { offsetof(xfs_sb_t, sb_fname[0]),   1 },
  97     { offsetof(xfs_sb_t, sb_blocklog),   0 },
  98     { offsetof(xfs_sb_t, sb_sectlog),    0 },
  99     { offsetof(xfs_sb_t, sb_inodelog),   0 },
 100     { offsetof(xfs_sb_t, sb_inopblog),   0 },
 101     { offsetof(xfs_sb_t, sb_agblklog),   0 },
 102     { offsetof(xfs_sb_t, sb_rextslog),   0 },
 103     { offsetof(xfs_sb_t, sb_inprogress), 0 },
 104     { offsetof(xfs_sb_t, sb_imax_pct),   0 },
 105     { offsetof(xfs_sb_t, sb_icount),     0 },
 106     { offsetof(xfs_sb_t, sb_ifree),      0 },
 107     { offsetof(xfs_sb_t, sb_fdblocks),   0 },
 108     { offsetof(xfs_sb_t, sb_frextents),  0 },
 109     { offsetof(xfs_sb_t, sb_uquotino),   0 },
 110     { offsetof(xfs_sb_t, sb_gquotino),   0 },
 111     { offsetof(xfs_sb_t, sb_qflags),     0 },
 112     { offsetof(xfs_sb_t, sb_flags),      0 },
 113     { offsetof(xfs_sb_t, sb_shared_vn),  0 },
 114     { offsetof(xfs_sb_t, sb_inoalignmt), 0 },
 115     { offsetof(xfs_sb_t, sb_unit),       0 },
 116     { offsetof(xfs_sb_t, sb_width),      0 },
 117     { offsetof(xfs_sb_t, sb_dirblklog),  0 },
 118     { offsetof(xfs_sb_t, sb_logsectlog), 0 },
 119     { offsetof(xfs_sb_t, sb_logsectsize),0 },
 120     { offsetof(xfs_sb_t, sb_logsunit),   0 },
 121     { offsetof(xfs_sb_t, sb_features2),  0 },
 122     { sizeof(xfs_sb_t),                  0 }
 123 };
 124
 125 /*
 126  * Return a pointer to an initialized xfs_mount structure.
 127  */
 128 xfs_mount_t *
 129 xfs_mount_init(void)
 130 {
 131         xfs_mount_t *mp;
 132
 133         mp = kmem_zalloc(sizeof(xfs_mount_t), KM_SLEEP);
 134
 135         if (xfs_icsb_init_counters(mp)) {
 136                 mp->m_flags |= XFS_MOUNT_NO_PERCPU_SB;
 137         }
 138
 139         AIL_LOCKINIT(&mp->m_ail_lock, "xfs_ail");
 140         spinlock_init(&mp->m_sb_lock, "xfs_sb");
 141         mutex_init(&mp->m_ilock);
 142         initnsema(&mp->m_growlock, 1, "xfs_grow");
 143         /*
 144          * Initialize the AIL.
 145          */
 146         xfs_trans_ail_init(mp);
 147
 148         atomic_set(&mp->m_active_trans, 0);
 149
 150         return mp;
 151 }
 152
 153 /*
 154  * Free up the resources associated with a mount structure.  Assume that
 155  * the structure was initially zeroed, so we can tell which fields got
 156  * initialized.
 157  */
 158 void
 159 xfs_mount_free(
 160         xfs_mount_t     *mp,
 161         int             remove_bhv)
 162 {
 163         if (mp->m_perag) {
 164                 int     agno;
 165
 166                 for (agno = 0; agno < mp->m_maxagi; agno++)
 167                         if (mp->m_perag[agno].pagb_list)
 168                                 kmem_free(mp->m_perag[agno].pagb_list,
 169                                                 sizeof(xfs_perag_busy_t) *
 170                                                         XFS_PAGB_NUM_SLOTS);
 171                 kmem_free(mp->m_perag,
 172                           sizeof(xfs_perag_t) * mp->m_sb.sb_agcount);
 173         }
 174
 175         AIL_LOCK_DESTROY(&mp->m_ail_lock);
 176         spinlock_destroy(&mp->m_sb_lock);
 177         mutex_destroy(&mp->m_ilock);
 178         freesema(&mp->m_growlock);
 179         if (mp->m_quotainfo)
 180                 XFS_QM_DONE(mp);
 181
 182         if (mp->m_fsname != NULL)
 183                 kmem_free(mp->m_fsname, mp->m_fsname_len);
 184         if (mp->m_rtname != NULL)
 185                 kmem_free(mp->m_rtname, strlen(mp->m_rtname) + 1);
 186         if (mp->m_logname != NULL)
 187                 kmem_free(mp->m_logname, strlen(mp->m_logname) + 1);
 188
 189         if (remove_bhv) {
 190                 struct bhv_vfs  *vfsp = XFS_MTOVFS(mp);
 191
 192                 bhv_remove_all_vfsops(vfsp, 0);
 193                 VFS_REMOVEBHV(vfsp, &mp->m_bhv);
 194         }
 195
 196         xfs_icsb_destroy_counters(mp);
 197         kmem_free(mp, sizeof(xfs_mount_t));
 198 }
 199
 200 /*
 201  * Check size of device based on the (data/realtime) block count.
 202  * Note: this check is used by the growfs code as well as mount.
 203  */
 204 int
 205 xfs_sb_validate_fsb_count(
 206         xfs_sb_t        *sbp,
 207         __uint64_t      nblocks)
 208 {
 209         ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
 210         ASSERT(sbp->sb_blocklog >= BBSHIFT);
 211
 212 #if XFS_BIG_BLKNOS     /* Limited by ULONG_MAX of page cache index */
 213         if (nblocks >> (PAGE_CACHE_SHIFT - sbp->sb_blocklog) > ULONG_MAX)
 214                 return E2BIG;
 215 #else                  /* Limited by UINT_MAX of sectors */
 216         if (nblocks << (sbp->sb_blocklog - BBSHIFT) > UINT_MAX)
 217                 return E2BIG;
 218 #endif
 219         return 0;
 220 }
 221
 222 /*
 223  * Check the validity of the SB found.
 224  */
 225 STATIC int
 226 xfs_mount_validate_sb(
 227         xfs_mount_t     *mp,
 228         xfs_sb_t        *sbp,
 229         int             flags)
 230 {
 231         /*
 232          * If the log device and data device have the
 233          * same device number, the log is internal.
 234          * Consequently, the sb_logstart should be non-zero.  If
 235          * we have a zero sb_logstart in this case, we may be trying to mount
 236          * a volume filesystem in a non-volume manner.
 237          */
 238         if (sbp->sb_magicnum != XFS_SB_MAGIC) {
 239                 xfs_fs_mount_cmn_err(flags, "bad magic number");
 240                 return XFS_ERROR(EWRONGFS);
 241         }
 242
 243         if (!XFS_SB_GOOD_VERSION(sbp)) {
 244                 xfs_fs_mount_cmn_err(flags, "bad version");
 245                 return XFS_ERROR(EWRONGFS);
 246         }
 247
 248         if (unlikely(
 249             sbp->sb_logstart == 0 && mp->m_logdev_targp == mp->m_ddev_targp)) {
 250                 xfs_fs_mount_cmn_err(flags,
 251                         "filesystem is marked as having an external log; "
 252                         "specify logdev on the\nmount command line.");
 253                 return XFS_ERROR(EINVAL);
 254         }
 255
 256         if (unlikely(
 257             sbp->sb_logstart != 0 && mp->m_logdev_targp != mp->m_ddev_targp)) {
 258                 xfs_fs_mount_cmn_err(flags,
 259                         "filesystem is marked as having an internal log; "
 260                         "do not specify logdev on\nthe mount command line.");
 261                 return XFS_ERROR(EINVAL);
 262         }
 263
 264         /*
 265          * More sanity checking. These were stolen directly from
 266          * xfs_repair.
 267          */
 268         if (unlikely(
 269             sbp->sb_agcount <= 0                                        ||
 270             sbp->sb_sectsize < XFS_MIN_SECTORSIZE                       ||
 271             sbp->sb_sectsize > XFS_MAX_SECTORSIZE                       ||
 272             sbp->sb_sectlog < XFS_MIN_SECTORSIZE_LOG                    ||
 273             sbp->sb_sectlog > XFS_MAX_SECTORSIZE_LOG                    ||
 274             sbp->sb_blocksize < XFS_MIN_BLOCKSIZE                       ||
 275             sbp->sb_blocksize > XFS_MAX_BLOCKSIZE                       ||
 276             sbp->sb_blocklog < XFS_MIN_BLOCKSIZE_LOG                    ||
 277             sbp->sb_blocklog > XFS_MAX_BLOCKSIZE_LOG                    ||
 278             sbp->sb_inodesize < XFS_DINODE_MIN_SIZE                     ||
 279             sbp->sb_inodesize > XFS_DINODE_MAX_SIZE                     ||
 280             sbp->sb_inodelog < XFS_DINODE_MIN_LOG                       ||
 281             sbp->sb_inodelog > XFS_DINODE_MAX_LOG                       ||
 282             (sbp->sb_blocklog - sbp->sb_inodelog != sbp->sb_inopblog)   ||
 283             (sbp->sb_rextsize * sbp->sb_blocksize > XFS_MAX_RTEXTSIZE)  ||
 284             (sbp->sb_rextsize * sbp->sb_blocksize < XFS_MIN_RTEXTSIZE)  ||
 285             (sbp->sb_imax_pct > 100 /* zero sb_imax_pct is valid */))) {
 286                 xfs_fs_mount_cmn_err(flags, "SB sanity check 1 failed");
 287                 return XFS_ERROR(EFSCORRUPTED);
 288         }
 289
 290         /*
 291          * Sanity check AG count, size fields against data size field
 292          */
 293         if (unlikely(
 294             sbp->sb_dblocks == 0 ||
 295             sbp->sb_dblocks >
 296              (xfs_drfsbno_t)sbp->sb_agcount * sbp->sb_agblocks ||
 297             sbp->sb_dblocks < (xfs_drfsbno_t)(sbp->sb_agcount - 1) *
 298                               sbp->sb_agblocks + XFS_MIN_AG_BLOCKS)) {
 299                 xfs_fs_mount_cmn_err(flags, "SB sanity check 2 failed");
 300                 return XFS_ERROR(EFSCORRUPTED);
 301         }
 302
 303         if (xfs_sb_validate_fsb_count(sbp, sbp->sb_dblocks) ||
 304             xfs_sb_validate_fsb_count(sbp, sbp->sb_rblocks)) {
 305                 xfs_fs_mount_cmn_err(flags,
 306                         "file system too large to be mounted on this system.");
 307                 return XFS_ERROR(E2BIG);
 308         }
 309
 310         if (unlikely(sbp->sb_inprogress)) {
 311                 xfs_fs_mount_cmn_err(flags, "file system busy");
 312                 return XFS_ERROR(EFSCORRUPTED);
 313         }
 314
 315         /*
 316          * Version 1 directory format has never worked on Linux.
 317          */
 318         if (unlikely(!XFS_SB_VERSION_HASDIRV2(sbp))) {
 319                 xfs_fs_mount_cmn_err(flags,
 320                         "file system using version 1 directory format");
 321                 return XFS_ERROR(ENOSYS);
 322         }
 323
 324         /*
 325          * Until this is fixed only page-sized or smaller data blocks work.
 326          */
 327         if (unlikely(sbp->sb_blocksize > PAGE_SIZE)) {
 328                 xfs_fs_mount_cmn_err(flags,
 329                         "file system with blocksize %d bytes",
 330                         sbp->sb_blocksize);
 331                 xfs_fs_mount_cmn_err(flags,
 332                         "only pagesize (%ld) or less will currently work.",
 333                         PAGE_SIZE);
 334                 return XFS_ERROR(ENOSYS);
 335         }
 336
 337         return 0;
 338 }
 339
 340 STATIC void
 341 xfs_initialize_perag_icache(
 342         xfs_perag_t     *pag)
 343 {
 344         if (!pag->pag_ici_init) {
 345                 rwlock_init(&pag->pag_ici_lock);
 346                 INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
 347                 pag->pag_ici_init = 1;
 348         }
 349 }
 350
 351 xfs_agnumber_t
 352 xfs_initialize_perag(
 353         bhv_vfs_t       *vfs,
 354         xfs_mount_t     *mp,
 355         xfs_agnumber_t  agcount)
 356 {
 357         xfs_agnumber_t  index, max_metadata;
 358         xfs_perag_t     *pag;
 359         xfs_agino_t     agino;
 360         xfs_ino_t       ino;
 361         xfs_sb_t        *sbp = &mp->m_sb;
 362         xfs_ino_t       max_inum = XFS_MAXINUMBER_32;
 363
 364         /* Check to see if the filesystem can overflow 32 bit inodes */
 365         agino = XFS_OFFBNO_TO_AGINO(mp, sbp->sb_agblocks - 1, 0);
 366         ino = XFS_AGINO_TO_INO(mp, agcount - 1, agino);
 367
 368         /* Clear the mount flag if no inode can overflow 32 bits
 369          * on this filesystem, or if specifically requested..
 370          */
 371         if ((vfs->vfs_flag & VFS_32BITINODES) && ino > max_inum) {
 372                 mp->m_flags |= XFS_MOUNT_32BITINODES;
 373         } else {
 374                 mp->m_flags &= ~XFS_MOUNT_32BITINODES;
 375         }
 376
 377         /* If we can overflow then setup the ag headers accordingly */
 378         if (mp->m_flags & XFS_MOUNT_32BITINODES) {
 379                 /* Calculate how much should be reserved for inodes to
 380                  * meet the max inode percentage.
 381                  */
 382                 if (mp->m_maxicount) {
 383                         __uint64_t      icount;
 384
 385                         icount = sbp->sb_dblocks * sbp->sb_imax_pct;
 386                         do_div(icount, 100);
 387                         icount += sbp->sb_agblocks - 1;
 388                         do_div(icount, sbp->sb_agblocks);
 389                         max_metadata = icount;
 390                 } else {
 391                         max_metadata = agcount;
 392                 }
 393                 for (index = 0; index < agcount; index++) {
 394                         ino = XFS_AGINO_TO_INO(mp, index, agino);
 395                         if (ino > max_inum) {
 396                                 index++;
 397                                 break;
 398                         }
 399
 400                         /* This ag is preferred for inodes */
 401                         pag = &mp->m_perag[index];
 402                         pag->pagi_inodeok = 1;
 403                         if (index < max_metadata)
 404                                 pag->pagf_metadata = 1;
 405                         xfs_initialize_perag_icache(pag);
 406                 }
 407         } else {
 408                 /* Setup default behavior for smaller filesystems */
 409                 for (index = 0; index < agcount; index++) {
 410                         pag = &mp->m_perag[index];
 411                         pag->pagi_inodeok = 1;
 412                         xfs_initialize_perag_icache(pag);
 413                 }
 414         }
 415         return index;
 416 }
 417
 418 void
 419 xfs_sb_from_disk(
 420         xfs_sb_t        *to,
 421         xfs_dsb_t       *from)
 422 {
 423         to->sb_magicnum = be32_to_cpu(from->sb_magicnum);
 424         to->sb_blocksize = be32_to_cpu(from->sb_blocksize);
 425         to->sb_dblocks = be64_to_cpu(from->sb_dblocks);
 426         to->sb_rblocks = be64_to_cpu(from->sb_rblocks);
 427         to->sb_rextents = be64_to_cpu(from->sb_rextents);
 428         memcpy(&to->sb_uuid, &from->sb_uuid, sizeof(to->sb_uuid));
 429         to->sb_logstart = be64_to_cpu(from->sb_logstart);
 430         to->sb_rootino = be64_to_cpu(from->sb_rootino);
 431         to->sb_rbmino = be64_to_cpu(from->sb_rbmino);
 432         to->sb_rsumino = be64_to_cpu(from->sb_rsumino);
 433         to->sb_rextsize = be32_to_cpu(from->sb_rextsize);
 434         to->sb_agblocks = be32_to_cpu(from->sb_agblocks);
 435         to->sb_agcount = be32_to_cpu(from->sb_agcount);
 436         to->sb_rbmblocks = be32_to_cpu(from->sb_rbmblocks);
 437         to->sb_logblocks = be32_to_cpu(from->sb_logblocks);
 438         to->sb_versionnum = be16_to_cpu(from->sb_versionnum);
 439         to->sb_sectsize = be16_to_cpu(from->sb_sectsize);
 440         to->sb_inodesize = be16_to_cpu(from->sb_inodesize);
 441         to->sb_inopblock = be16_to_cpu(from->sb_inopblock);
 442         memcpy(&to->sb_fname, &from->sb_fname, sizeof(to->sb_fname));
 443         to->sb_blocklog = from->sb_blocklog;
 444         to->sb_sectlog = from->sb_sectlog;
 445         to->sb_inodelog = from->sb_inodelog;
 446         to->sb_inopblog = from->sb_inopblog;
 447         to->sb_agblklog = from->sb_agblklog;
 448         to->sb_rextslog = from->sb_rextslog;
 449         to->sb_inprogress = from->sb_inprogress;
 450         to->sb_imax_pct = from->sb_imax_pct;
 451         to->sb_icount = be64_to_cpu(from->sb_icount);
 452         to->sb_ifree = be64_to_cpu(from->sb_ifree);
 453         to->sb_fdblocks = be64_to_cpu(from->sb_fdblocks);
 454         to->sb_frextents = be64_to_cpu(from->sb_frextents);
 455         to->sb_uquotino = be64_to_cpu(from->sb_uquotino);
 456         to->sb_gquotino = be64_to_cpu(from->sb_gquotino);
 457         to->sb_qflags = be16_to_cpu(from->sb_qflags);
 458         to->sb_flags = from->sb_flags;
 459         to->sb_shared_vn = from->sb_shared_vn;
 460         to->sb_inoalignmt = be32_to_cpu(from->sb_inoalignmt);
 461         to->sb_unit = be32_to_cpu(from->sb_unit);
 462         to->sb_width = be32_to_cpu(from->sb_width);
 463         to->sb_dirblklog = from->sb_dirblklog;
 464         to->sb_logsectlog = from->sb_logsectlog;
 465         to->sb_logsectsize = be16_to_cpu(from->sb_logsectsize);
 466         to->sb_logsunit = be32_to_cpu(from->sb_logsunit);
 467         to->sb_features2 = be32_to_cpu(from->sb_features2);
 468 }
 469
 470 /*
 471  * Copy in core superblock to ondisk one.
 472  *
 473  * The fields argument is mask of superblock fields to copy.
 474  */
 475 void
 476 xfs_sb_to_disk(
 477         xfs_dsb_t       *to,
 478         xfs_sb_t        *from,
 479         __int64_t       fields)
 480 {
 481         xfs_caddr_t     to_ptr = (xfs_caddr_t)to;
 482         xfs_caddr_t     from_ptr = (xfs_caddr_t)from;
 483         xfs_sb_field_t  f;
 484         int             first;
 485         int             size;
 486
 487         ASSERT(fields);
 488         if (!fields)
 489                 return;
 490
 491         while (fields) {
 492                 f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
 493                 first = xfs_sb_info[f].offset;
 494                 size = xfs_sb_info[f + 1].offset - first;
 495
 496                 ASSERT(xfs_sb_info[f].type == 0 || xfs_sb_info[f].type == 1);
 497
 498                 if (size == 1 || xfs_sb_info[f].type == 1) {
 499                         memcpy(to_ptr + first, from_ptr + first, size);
 500                 } else {
 501                         switch (size) {
 502                         case 2:
 503                                 *(__be16 *)(to_ptr + first) =
 504                                         cpu_to_be16(*(__u16 *)(from_ptr + first));
 505                                 break;
 506                         case 4:
 507                                 *(__be32 *)(to_ptr + first) =
 508                                         cpu_to_be32(*(__u32 *)(from_ptr + first));
 509                                 break;
 510                         case 8:
 511                                 *(__be64 *)(to_ptr + first) =
 512                                         cpu_to_be64(*(__u64 *)(from_ptr + first));
 513                                 break;
 514                         default:
 515                                 ASSERT(0);
 516                         }
 517                 }
 518
 519                 fields &= ~(1LL << f);
 520         }
 521 }
 522
 523 /*
 524  * xfs_readsb
 525  *
 526  * Does the initial read of the superblock.
 527  */
 528 int
 529 xfs_readsb(xfs_mount_t *mp, int flags)
 530 {
 531         unsigned int    sector_size;
 532         unsigned int    extra_flags;
 533         xfs_buf_t       *bp;
 534         int             error;
 535
 536         ASSERT(mp->m_sb_bp == NULL);
 537         ASSERT(mp->m_ddev_targp != NULL);
 538
 539         /*
 540          * Allocate a (locked) buffer to hold the superblock.
 541          * This will be kept around at all times to optimize
 542          * access to the superblock.
 543          */
 544         sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
 545         extra_flags = XFS_BUF_LOCK | XFS_BUF_MANAGE | XFS_BUF_MAPPED;
 546
 547         bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
 548                                 BTOBB(sector_size), extra_flags);
 549         if (!bp || XFS_BUF_ISERROR(bp)) {
 550                 xfs_fs_mount_cmn_err(flags, "SB read failed");
 551                 error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
 552                 goto fail;
 553         }
 554         ASSERT(XFS_BUF_ISBUSY(bp));
 555         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 556
 557         /*
 558          * Initialize the mount structure from the superblock.
 559          * But first do some basic consistency checking.
 560          */
 561         xfs_sb_from_disk(&mp->m_sb, XFS_BUF_TO_SBP(bp));
 562
 563         error = xfs_mount_validate_sb(mp, &(mp->m_sb), flags);
 564         if (error) {
 565                 xfs_fs_mount_cmn_err(flags, "SB validate failed");
 566                 goto fail;
 567         }
 568
 569         /*
 570          * We must be able to do sector-sized and sector-aligned IO.
 571          */
 572         if (sector_size > mp->m_sb.sb_sectsize) {
 573                 xfs_fs_mount_cmn_err(flags,
 574                         "device supports only %u byte sectors (not %u)",
 575                         sector_size, mp->m_sb.sb_sectsize);
 576                 error = ENOSYS;
 577                 goto fail;
 578         }
 579
 580         /*
 581          * If device sector size is smaller than the superblock size,
 582          * re-read the superblock so the buffer is correctly sized.
 583          */
 584         if (sector_size < mp->m_sb.sb_sectsize) {
 585                 XFS_BUF_UNMANAGE(bp);
 586                 xfs_buf_relse(bp);
 587                 sector_size = mp->m_sb.sb_sectsize;
 588                 bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
 589                                         BTOBB(sector_size), extra_flags);
 590                 if (!bp || XFS_BUF_ISERROR(bp)) {
 591                         xfs_fs_mount_cmn_err(flags, "SB re-read failed");
 592                         error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
 593                         goto fail;
 594                 }
 595                 ASSERT(XFS_BUF_ISBUSY(bp));
 596                 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 597         }
 598
 599         /* Initialize per-cpu counters */
 600         xfs_icsb_reinit_counters(mp);
 601
 602         mp->m_sb_bp = bp;
 603         xfs_buf_relse(bp);
 604         ASSERT(XFS_BUF_VALUSEMA(bp) > 0);
 605         return 0;
 606
 607  fail:
 608         if (bp) {
 609                 XFS_BUF_UNMANAGE(bp);
 610                 xfs_buf_relse(bp);
 611         }
 612         return error;
 613 }
 614
 615
 616 /*
 617  * xfs_mount_common
 618  *
 619  * Mount initialization code establishing various mount
 620  * fields from the superblock associated with the given
 621  * mount structure
 622  */
 623 STATIC void
 624 xfs_mount_common(xfs_mount_t *mp, xfs_sb_t *sbp)
 625 {
 626         int     i;
 627
 628         mp->m_agfrotor = mp->m_agirotor = 0;
 629         spinlock_init(&mp->m_agirotor_lock, "m_agirotor_lock");
 630         mp->m_maxagi = mp->m_sb.sb_agcount;
 631         mp->m_blkbit_log = sbp->sb_blocklog + XFS_NBBYLOG;
 632         mp->m_blkbb_log = sbp->sb_blocklog - BBSHIFT;
 633         mp->m_sectbb_log = sbp->sb_sectlog - BBSHIFT;
 634         mp->m_agno_log = xfs_highbit32(sbp->sb_agcount - 1) + 1;
 635         mp->m_agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
 636         mp->m_litino = sbp->sb_inodesize -
 637                 ((uint)sizeof(xfs_dinode_core_t) + (uint)sizeof(xfs_agino_t));
 638         mp->m_blockmask = sbp->sb_blocksize - 1;
 639         mp->m_blockwsize = sbp->sb_blocksize >> XFS_WORDLOG;
 640         mp->m_blockwmask = mp->m_blockwsize - 1;
 641         INIT_LIST_HEAD(&mp->m_del_inodes);
 642
 643         /*
 644          * Setup for attributes, in case they get created.
 645          * This value is for inodes getting attributes for the first time,
 646          * the per-inode value is for old attribute values.
 647          */
 648         ASSERT(sbp->sb_inodesize >= 256 && sbp->sb_inodesize <= 2048);
 649         switch (sbp->sb_inodesize) {
 650         case 256:
 651                 mp->m_attroffset = XFS_LITINO(mp) -
 652                                    XFS_BMDR_SPACE_CALC(MINABTPTRS);
 653                 break;
 654         case 512:
 655         case 1024:
 656         case 2048:
 657                 mp->m_attroffset = XFS_BMDR_SPACE_CALC(6 * MINABTPTRS);
 658                 break;
 659         default:
 660                 ASSERT(0);
 661         }
 662         ASSERT(mp->m_attroffset < XFS_LITINO(mp));
 663
 664         for (i = 0; i < 2; i++) {
 665                 mp->m_alloc_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 666                         xfs_alloc, i == 0);
 667                 mp->m_alloc_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 668                         xfs_alloc, i == 0);
 669         }
 670         for (i = 0; i < 2; i++) {
 671                 mp->m_bmap_dmxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 672                         xfs_bmbt, i == 0);
 673                 mp->m_bmap_dmnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 674                         xfs_bmbt, i == 0);
 675         }
 676         for (i = 0; i < 2; i++) {
 677                 mp->m_inobt_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 678                         xfs_inobt, i == 0);
 679                 mp->m_inobt_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 680                         xfs_inobt, i == 0);
 681         }
 682
 683         mp->m_bsize = XFS_FSB_TO_BB(mp, 1);
 684         mp->m_ialloc_inos = (int)MAX((__uint16_t)XFS_INODES_PER_CHUNK,
 685                                         sbp->sb_inopblock);
 686         mp->m_ialloc_blks = mp->m_ialloc_inos >> sbp->sb_inopblog;
 687 }
 688
 689 /*
 690  * xfs_initialize_perag_data
 691  *
 692  * Read in each per-ag structure so we can count up the number of
 693  * allocated inodes, free inodes and used filesystem blocks as this
 694  * information is no longer persistent in the superblock. Once we have
 695  * this information, write it into the in-core superblock structure.
 696  */
 697 STATIC int
 698 xfs_initialize_perag_data(xfs_mount_t *mp, xfs_agnumber_t agcount)
 699 {
 700         xfs_agnumber_t  index;
 701         xfs_perag_t     *pag;
 702         xfs_sb_t        *sbp = &mp->m_sb;
 703         uint64_t        ifree = 0;
 704         uint64_t        ialloc = 0;
 705         uint64_t        bfree = 0;
 706         uint64_t        bfreelst = 0;
 707         uint64_t        btree = 0;
 708         int             error;
 709         int             s;
 710
 711         for (index = 0; index < agcount; index++) {
 712                 /*
 713                  * read the agf, then the agi. This gets us
 714                  * all the inforamtion we need and populates the
 715                  * per-ag structures for us.
 716                  */
 717                 error = xfs_alloc_pagf_init(mp, NULL, index, 0);
 718                 if (error)
 719                         return error;
 720
 721                 error = xfs_ialloc_pagi_init(mp, NULL, index);
 722                 if (error)
 723                         return error;
 724                 pag = &mp->m_perag[index];
 725                 ifree += pag->pagi_freecount;
 726                 ialloc += pag->pagi_count;
 727                 bfree += pag->pagf_freeblks;
 728                 bfreelst += pag->pagf_flcount;
 729                 btree += pag->pagf_btreeblks;
 730         }
 731         /*
 732          * Overwrite incore superblock counters with just-read data
 733          */
 734         s = XFS_SB_LOCK(mp);
 735         sbp->sb_ifree = ifree;
 736         sbp->sb_icount = ialloc;
 737         sbp->sb_fdblocks = bfree + bfreelst + btree;
 738         XFS_SB_UNLOCK(mp, s);
 739
 740         /* Fixup the per-cpu counters as well. */
 741         xfs_icsb_reinit_counters(mp);
 742
 743         return 0;
 744 }
 745
 746 /*
 747  * xfs_mountfs
 748  *
 749  * This function does the following on an initial mount of a file system:
 750  *      - reads the superblock from disk and init the mount struct
 751  *      - if we're a 32-bit kernel, do a size check on the superblock
 752  *              so we don't mount terabyte filesystems
 753  *      - init mount struct realtime fields
 754  *      - allocate inode hash table for fs
 755  *      - init directory manager
 756  *      - perform recovery and init the log manager
 757  */
 758 int
 759 xfs_mountfs(
 760         bhv_vfs_t       *vfsp,
 761         xfs_mount_t     *mp,
 762         int             mfsi_flags)
 763 {
 764         xfs_buf_t       *bp;
 765         xfs_sb_t        *sbp = &(mp->m_sb);
 766         xfs_inode_t     *rip;
 767         bhv_vnode_t     *rvp = NULL;
 768         int             readio_log, writeio_log;
 769         xfs_daddr_t     d;
 770         __uint64_t      resblks;
 771         __int64_t       update_flags;
 772         uint            quotamount, quotaflags;
 773         int             agno;
 774         int             uuid_mounted = 0;
 775         int             error = 0;
 776
 777         if (mp->m_sb_bp == NULL) {
 778                 if ((error = xfs_readsb(mp, mfsi_flags))) {
 779                         return error;
 780                 }
 781         }
 782         xfs_mount_common(mp, sbp);
 783
 784         /*
 785          * Check if sb_agblocks is aligned at stripe boundary
 786          * If sb_agblocks is NOT aligned turn off m_dalign since
 787          * allocator alignment is within an ag, therefore ag has
 788          * to be aligned at stripe boundary.
 789          */
 790         update_flags = 0LL;
 791         if (mp->m_dalign && !(mfsi_flags & XFS_MFSI_SECOND)) {
 792                 /*
 793                  * If stripe unit and stripe width are not multiples
 794                  * of the fs blocksize turn off alignment.
 795                  */
 796                 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
 797                     (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
 798                         if (mp->m_flags & XFS_MOUNT_RETERR) {
 799                                 cmn_err(CE_WARN,
 800                                         "XFS: alignment check 1 failed");
 801                                 error = XFS_ERROR(EINVAL);
 802                                 goto error1;
 803                         }
 804                         mp->m_dalign = mp->m_swidth = 0;
 805                 } else {
 806                         /*
 807                          * Convert the stripe unit and width to FSBs.
 808                          */
 809                         mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
 810                         if (mp->m_dalign && (sbp->sb_agblocks % mp->m_dalign)) {
 811                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
 812                                         error = XFS_ERROR(EINVAL);
 813                                         goto error1;
 814                                 }
 815                                 xfs_fs_cmn_err(CE_WARN, mp,
 816 "stripe alignment turned off: sunit(%d)/swidth(%d) incompatible with agsize(%d)",
 817                                         mp->m_dalign, mp->m_swidth,
 818                                         sbp->sb_agblocks);
 819
 820                                 mp->m_dalign = 0;
 821                                 mp->m_swidth = 0;
 822                         } else if (mp->m_dalign) {
 823                                 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
 824                         } else {
 825                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
 826                                         xfs_fs_cmn_err(CE_WARN, mp,
 827 "stripe alignment turned off: sunit(%d) less than bsize(%d)",
 828                                                 mp->m_dalign,
 829                                                 mp->m_blockmask +1);
 830                                         error = XFS_ERROR(EINVAL);
 831                                         goto error1;
 832                                 }
 833                                 mp->m_swidth = 0;
 834                         }
 835                 }
 836
 837                 /*
 838                  * Update superblock with new values
 839                  * and log changes
 840                  */
 841                 if (XFS_SB_VERSION_HASDALIGN(sbp)) {
 842                         if (sbp->sb_unit != mp->m_dalign) {
 843                                 sbp->sb_unit = mp->m_dalign;
 844                                 update_flags |= XFS_SB_UNIT;
 845                         }
 846                         if (sbp->sb_width != mp->m_swidth) {
 847                                 sbp->sb_width = mp->m_swidth;
 848                                 update_flags |= XFS_SB_WIDTH;
 849                         }
 850                 }
 851         } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
 852                     XFS_SB_VERSION_HASDALIGN(&mp->m_sb)) {
 853                         mp->m_dalign = sbp->sb_unit;
 854                         mp->m_swidth = sbp->sb_width;
 855         }
 856
 857         xfs_alloc_compute_maxlevels(mp);
 858         xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
 859         xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
 860         xfs_ialloc_compute_maxlevels(mp);
 861
 862         if (sbp->sb_imax_pct) {
 863                 __uint64_t      icount;
 864
 865                 /* Make sure the maximum inode count is a multiple of the
 866                  * units we allocate inodes in.
 867                  */
 868
 869                 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
 870                 do_div(icount, 100);
 871                 do_div(icount, mp->m_ialloc_blks);
 872                 mp->m_maxicount = (icount * mp->m_ialloc_blks)  <<
 873                                    sbp->sb_inopblog;
 874         } else
 875                 mp->m_maxicount = 0;
 876
 877         mp->m_maxioffset = xfs_max_file_offset(sbp->sb_blocklog);
 878
 879         /*
 880          * XFS uses the uuid from the superblock as the unique
 881          * identifier for fsid.  We can not use the uuid from the volume
 882          * since a single partition filesystem is identical to a single
 883          * partition volume/filesystem.
 884          */
 885         if ((mfsi_flags & XFS_MFSI_SECOND) == 0 &&
 886             (mp->m_flags & XFS_MOUNT_NOUUID) == 0) {
 887                 __uint64_t      ret64;
 888                 if (xfs_uuid_mount(mp)) {
 889                         error = XFS_ERROR(EINVAL);
 890                         goto error1;
 891                 }
 892                 uuid_mounted=1;
 893                 ret64 = uuid_hash64(&sbp->sb_uuid);
 894                 memcpy(&vfsp->vfs_fsid, &ret64, sizeof(ret64));
 895         }
 896
 897         /*
 898          * Set the default minimum read and write sizes unless
 899          * already specified in a mount option.
 900          * We use smaller I/O sizes when the file system
 901          * is being used for NFS service (wsync mount option).
 902          */
 903         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
 904                 if (mp->m_flags & XFS_MOUNT_WSYNC) {
 905                         readio_log = XFS_WSYNC_READIO_LOG;
 906                         writeio_log = XFS_WSYNC_WRITEIO_LOG;
 907                 } else {
 908                         readio_log = XFS_READIO_LOG_LARGE;
 909                         writeio_log = XFS_WRITEIO_LOG_LARGE;
 910                 }
 911         } else {
 912                 readio_log = mp->m_readio_log;
 913                 writeio_log = mp->m_writeio_log;
 914         }
 915
 916         if (sbp->sb_blocklog > readio_log) {
 917                 mp->m_readio_log = sbp->sb_blocklog;
 918         } else {
 919                 mp->m_readio_log = readio_log;
 920         }
 921         mp->m_readio_blocks = 1 << (mp->m_readio_log - sbp->sb_blocklog);
 922         if (sbp->sb_blocklog > writeio_log) {
 923                 mp->m_writeio_log = sbp->sb_blocklog;
 924         } else {
 925                 mp->m_writeio_log = writeio_log;
 926         }
 927         mp->m_writeio_blocks = 1 << (mp->m_writeio_log - sbp->sb_blocklog);
 928
 929         /*
 930          * Set the inode cluster size.
 931          * This may still be overridden by the file system
 932          * block size if it is larger than the chosen cluster size.
 933          */
 934         mp->m_inode_cluster_size = XFS_INODE_BIG_CLUSTER_SIZE;
 935
 936         /*
 937          * Set whether we're using inode alignment.
 938          */
 939         if (XFS_SB_VERSION_HASALIGN(&mp->m_sb) &&
 940             mp->m_sb.sb_inoalignmt >=
 941             XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size))
 942                 mp->m_inoalign_mask = mp->m_sb.sb_inoalignmt - 1;
 943         else
 944                 mp->m_inoalign_mask = 0;
 945         /*
 946          * If we are using stripe alignment, check whether
 947          * the stripe unit is a multiple of the inode alignment
 948          */
 949         if (mp->m_dalign && mp->m_inoalign_mask &&
 950             !(mp->m_dalign & mp->m_inoalign_mask))
 951                 mp->m_sinoalign = mp->m_dalign;
 952         else
 953                 mp->m_sinoalign = 0;
 954         /*
 955          * Check that the data (and log if separate) are an ok size.
 956          */
 957         d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
 958         if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
 959                 cmn_err(CE_WARN, "XFS: size check 1 failed");
 960                 error = XFS_ERROR(E2BIG);
 961                 goto error1;
 962         }
 963         error = xfs_read_buf(mp, mp->m_ddev_targp,
 964                              d - XFS_FSS_TO_BB(mp, 1),
 965                              XFS_FSS_TO_BB(mp, 1), 0, &bp);
 966         if (!error) {
 967                 xfs_buf_relse(bp);
 968         } else {
 969                 cmn_err(CE_WARN, "XFS: size check 2 failed");
 970                 if (error == ENOSPC) {
 971                         error = XFS_ERROR(E2BIG);
 972                 }
 973                 goto error1;
 974         }
 975
 976         if (((mfsi_flags & XFS_MFSI_CLIENT) == 0) &&
 977             mp->m_logdev_targp != mp->m_ddev_targp) {
 978                 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
 979                 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
 980                         cmn_err(CE_WARN, "XFS: size check 3 failed");
 981                         error = XFS_ERROR(E2BIG);
 982                         goto error1;
 983                 }
 984                 error = xfs_read_buf(mp, mp->m_logdev_targp,
 985                                      d - XFS_FSB_TO_BB(mp, 1),
 986                                      XFS_FSB_TO_BB(mp, 1), 0, &bp);
 987                 if (!error) {
 988                         xfs_buf_relse(bp);
 989                 } else {
 990                         cmn_err(CE_WARN, "XFS: size check 3 failed");
 991                         if (error == ENOSPC) {
 992                                 error = XFS_ERROR(E2BIG);
 993                         }
 994                         goto error1;
 995                 }
 996         }
 997
 998         /*
 999          * Initialize realtime fields in the mount structure
1000          */
1001         if ((error = xfs_rtmount_init(mp))) {
1002                 cmn_err(CE_WARN, "XFS: RT mount failed");
1003                 goto error1;
1004         }
1005
1006         /*
1007          * For client case we are done now
1008          */
1009         if (mfsi_flags & XFS_MFSI_CLIENT) {
1010                 return 0;
1011         }
1012
1013         /*
1014          *  Copies the low order bits of the timestamp and the randomly
1015          *  set "sequence" number out of a UUID.
1016          */
1017         uuid_getnodeuniq(&sbp->sb_uuid, mp->m_fixedfsid);
1018
1019         /*
1020          *  The vfs structure needs to have a file system independent
1021          *  way of checking for the invariant file system ID.  Since it
1022          *  can't look at mount structures it has a pointer to the data
1023          *  in the mount structure.
1024          *
1025          *  File systems that don't support user level file handles (i.e.
1026          *  all of them except for XFS) will leave vfs_altfsid as NULL.
1027          */
1028         vfsp->vfs_altfsid = (xfs_fsid_t *)mp->m_fixedfsid;
1029         mp->m_dmevmask = 0;     /* not persistent; set after each mount */
1030
1031         xfs_dir_mount(mp);
1032
1033         /*
1034          * Initialize the attribute manager's entries.
1035          */
1036         mp->m_attr_magicpct = (mp->m_sb.sb_blocksize * 37) / 100;
1037
1038         /*
1039          * Initialize the precomputed transaction reservations values.
1040          */
1041         xfs_trans_init(mp);
1042
1043         /*
1044          * Allocate and initialize the per-ag data.
1045          */
1046         init_rwsem(&mp->m_peraglock);
1047         mp->m_perag =
1048                 kmem_zalloc(sbp->sb_agcount * sizeof(xfs_perag_t), KM_SLEEP);
1049
1050         mp->m_maxagi = xfs_initialize_perag(vfsp, mp, sbp->sb_agcount);
1051
1052         /*
1053          * log's mount-time initialization. Perform 1st part recovery if needed
1054          */
1055         if (likely(sbp->sb_logblocks > 0)) {    /* check for volume case */
1056                 error = xfs_log_mount(mp, mp->m_logdev_targp,
1057                                       XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
1058                                       XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
1059                 if (error) {
1060                         cmn_err(CE_WARN, "XFS: log mount failed");
1061                         goto error2;
1062                 }
1063         } else {        /* No log has been defined */
1064                 cmn_err(CE_WARN, "XFS: no log defined");
1065                 XFS_ERROR_REPORT("xfs_mountfs_int(1)", XFS_ERRLEVEL_LOW, mp);
1066                 error = XFS_ERROR(EFSCORRUPTED);
1067                 goto error2;
1068         }
1069
1070         /*
1071          * Now the log is mounted, we know if it was an unclean shutdown or
1072          * not. If it was, with the first phase of recovery has completed, we
1073          * have consistent AG blocks on disk. We have not recovered EFIs yet,
1074          * but they are recovered transactionally in the second recovery phase
1075          * later.
1076          *
1077          * Hence we can safely re-initialise incore superblock counters from
1078          * the per-ag data. These may not be correct if the filesystem was not
1079          * cleanly unmounted, so we need to wait for recovery to finish before
1080          * doing this.
1081          *
1082          * If the filesystem was cleanly unmounted, then we can trust the
1083          * values in the superblock to be correct and we don't need to do
1084          * anything here.
1085          *
1086          * If we are currently making the filesystem, the initialisation will
1087          * fail as the perag data is in an undefined state.
1088          */
1089
1090         if (xfs_sb_version_haslazysbcount(&mp->m_sb) &&
1091             !XFS_LAST_UNMOUNT_WAS_CLEAN(mp) &&
1092              !mp->m_sb.sb_inprogress) {
1093                 error = xfs_initialize_perag_data(mp, sbp->sb_agcount);
1094                 if (error) {
1095                         goto error2;
1096                 }
1097         }
1098         /*
1099          * Get and sanity-check the root inode.
1100          * Save the pointer to it in the mount structure.
1101          */
1102         error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip, 0);
1103         if (error) {
1104                 cmn_err(CE_WARN, "XFS: failed to read root inode");
1105                 goto error3;
1106         }
1107
1108         ASSERT(rip != NULL);
1109         rvp = XFS_ITOV(rip);
1110
1111         if (unlikely((rip->i_d.di_mode & S_IFMT) != S_IFDIR)) {
1112                 cmn_err(CE_WARN, "XFS: corrupted root inode");
1113                 cmn_err(CE_WARN, "Device %s - root %llu is not a directory",
1114                         XFS_BUFTARG_NAME(mp->m_ddev_targp),
1115                         (unsigned long long)rip->i_ino);
1116                 xfs_iunlock(rip, XFS_ILOCK_EXCL);
1117                 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
1118                                  mp);
1119                 error = XFS_ERROR(EFSCORRUPTED);
1120                 goto error4;
1121         }
1122         mp->m_rootip = rip;     /* save it */
1123
1124         xfs_iunlock(rip, XFS_ILOCK_EXCL);
1125
1126         /*
1127          * Initialize realtime inode pointers in the mount structure
1128          */
1129         if ((error = xfs_rtmount_inodes(mp))) {
1130                 /*
1131                  * Free up the root inode.
1132                  */
1133                 cmn_err(CE_WARN, "XFS: failed to read RT inodes");
1134                 goto error4;
1135         }
1136
1137         /*
1138          * If fs is not mounted readonly, then update the superblock
1139          * unit and width changes.
1140          */
1141         if (update_flags && !(vfsp->vfs_flag & VFS_RDONLY))
1142                 xfs_mount_log_sbunit(mp, update_flags);
1143
1144         /*
1145          * Initialise the XFS quota management subsystem for this mount
1146          */
1147         if ((error = XFS_QM_INIT(mp, &quotamount, &quotaflags)))
1148                 goto error4;
1149
1150         /*
1151          * Finish recovering the file system.  This part needed to be
1152          * delayed until after the root and real-time bitmap inodes
1153          * were consistently read in.
1154          */
1155         error = xfs_log_mount_finish(mp, mfsi_flags);
1156         if (error) {
1157                 cmn_err(CE_WARN, "XFS: log mount finish failed");
1158                 goto error4;
1159         }
1160
1161         /*
1162          * Complete the quota initialisation, post-log-replay component.
1163          */
1164         if ((error = XFS_QM_MOUNT(mp, quotamount, quotaflags, mfsi_flags)))
1165                 goto error4;
1166
1167         /*
1168          * Now we are mounted, reserve a small amount of unused space for
1169          * privileged transactions. This is needed so that transaction
1170          * space required for critical operations can dip into this pool
1171          * when at ENOSPC. This is needed for operations like create with
1172          * attr, unwritten extent conversion at ENOSPC, etc. Data allocations
1173          * are not allowed to use this reserved space.
1174          *
1175          * We default to 5% or 1024 fsbs of space reserved, whichever is smaller.
1176          * This may drive us straight to ENOSPC on mount, but that implies
1177          * we were already there on the last unmount.
1178          */
1179         resblks = mp->m_sb.sb_dblocks;
1180         do_div(resblks, 20);
1181         resblks = min_t(__uint64_t, resblks, 1024);
1182         xfs_reserve_blocks(mp, &resblks, NULL);
1183
1184         return 0;
1185
1186  error4:
1187         /*
1188          * Free up the root inode.
1189          */
1190         VN_RELE(rvp);
1191  error3:
1192         xfs_log_unmount_dealloc(mp);
1193  error2:
1194         for (agno = 0; agno < sbp->sb_agcount; agno++)
1195                 if (mp->m_perag[agno].pagb_list)
1196                         kmem_free(mp->m_perag[agno].pagb_list,
1197                           sizeof(xfs_perag_busy_t) * XFS_PAGB_NUM_SLOTS);
1198         kmem_free(mp->m_perag, sbp->sb_agcount * sizeof(xfs_perag_t));
1199         mp->m_perag = NULL;
1200         /* FALLTHROUGH */
1201  error1:
1202         if (uuid_mounted)
1203                 xfs_uuid_unmount(mp);
1204         xfs_freesb(mp);
1205         return error;
1206 }
1207
1208 /*
1209  * xfs_unmountfs
1210  *
1211  * This flushes out the inodes,dquots and the superblock, unmounts the
1212  * log and makes sure that incore structures are freed.
1213  */
1214 int
1215 xfs_unmountfs(xfs_mount_t *mp, struct cred *cr)
1216 {
1217         struct bhv_vfs  *vfsp = XFS_MTOVFS(mp);
1218 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1219         int64_t         fsid;
1220 #endif
1221         __uint64_t      resblks;
1222
1223         /*
1224          * We can potentially deadlock here if we have an inode cluster
1225          * that has been freed has it's buffer still pinned in memory because
1226          * the transaction is still sitting in a iclog. The stale inodes
1227          * on that buffer will have their flush locks held until the
1228          * transaction hits the disk and the callbacks run. the inode
1229          * flush takes the flush lock unconditionally and with nothing to
1230          * push out the iclog we will never get that unlocked. hence we
1231          * need to force the log first.
1232          */
1233         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1234         xfs_iflush_all(mp);
1235
1236         XFS_QM_DQPURGEALL(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_UMOUNTING);
1237
1238         /*
1239          * Flush out the log synchronously so that we know for sure
1240          * that nothing is pinned.  This is important because bflush()
1241          * will skip pinned buffers.
1242          */
1243         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1244
1245         xfs_binval(mp->m_ddev_targp);
1246         if (mp->m_rtdev_targp) {
1247                 xfs_binval(mp->m_rtdev_targp);
1248         }
1249
1250         /*
1251          * Unreserve any blocks we have so that when we unmount we don't account
1252          * the reserved free space as used. This is really only necessary for
1253          * lazy superblock counting because it trusts the incore superblock
1254          * counters to be aboslutely correct on clean unmount.
1255          *
1256          * We don't bother correcting this elsewhere for lazy superblock
1257          * counting because on mount of an unclean filesystem we reconstruct the
1258          * correct counter value and this is irrelevant.
1259          *
1260          * For non-lazy counter filesystems, this doesn't matter at all because
1261          * we only every apply deltas to the superblock and hence the incore
1262          * value does not matter....
1263          */
1264         resblks = 0;
1265         xfs_reserve_blocks(mp, &resblks, NULL);
1266
1267         xfs_log_sbcount(mp, 1);
1268         xfs_unmountfs_writesb(mp);
1269         xfs_unmountfs_wait(mp);                 /* wait for async bufs */
1270         xfs_log_unmount(mp);                    /* Done! No more fs ops. */
1271
1272         xfs_freesb(mp);
1273
1274         /*
1275          * All inodes from this mount point should be freed.
1276          */
1277         ASSERT(mp->m_inodes == NULL);
1278
1279         xfs_unmountfs_close(mp, cr);
1280         if ((mp->m_flags & XFS_MOUNT_NOUUID) == 0)
1281                 xfs_uuid_unmount(mp);
1282
1283 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1284         /*
1285          * clear all error tags on this filesystem
1286          */
1287         memcpy(&fsid, &vfsp->vfs_fsid, sizeof(int64_t));
1288         xfs_errortag_clearall_umount(fsid, mp->m_fsname, 0);
1289 #endif
1290         XFS_IODONE(vfsp);
1291         xfs_mount_free(mp, 1);
1292         return 0;
1293 }
1294
1295 void
1296 xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
1297 {
1298         if (mp->m_logdev_targp && mp->m_logdev_targp != mp->m_ddev_targp)
1299                 xfs_free_buftarg(mp->m_logdev_targp, 1);
1300         if (mp->m_rtdev_targp)
1301                 xfs_free_buftarg(mp->m_rtdev_targp, 1);
1302         xfs_free_buftarg(mp->m_ddev_targp, 0);
1303 }
1304
1305 STATIC void
1306 xfs_unmountfs_wait(xfs_mount_t *mp)
1307 {
1308         if (mp->m_logdev_targp != mp->m_ddev_targp)
1309                 xfs_wait_buftarg(mp->m_logdev_targp);
1310         if (mp->m_rtdev_targp)
1311                 xfs_wait_buftarg(mp->m_rtdev_targp);
1312         xfs_wait_buftarg(mp->m_ddev_targp);
1313 }
1314
1315 int
1316 xfs_fs_writable(xfs_mount_t *mp)
1317 {
1318         bhv_vfs_t       *vfsp = XFS_MTOVFS(mp);
1319
1320         return !(vfs_test_for_freeze(vfsp) || XFS_FORCED_SHUTDOWN(mp) ||
1321                 (vfsp->vfs_flag & VFS_RDONLY));
1322 }
1323
1324 /*
1325  * xfs_log_sbcount
1326  *
1327  * Called either periodically to keep the on disk superblock values
1328  * roughly up to date or from unmount to make sure the values are
1329  * correct on a clean unmount.
1330  *
1331  * Note this code can be called during the process of freezing, so
1332  * we may need to use the transaction allocator which does not not
1333  * block when the transaction subsystem is in its frozen state.
1334  */
1335 int
1336 xfs_log_sbcount(
1337         xfs_mount_t     *mp,
1338         uint            sync)
1339 {
1340         xfs_trans_t     *tp;
1341         int             error;
1342
1343         if (!xfs_fs_writable(mp))
1344                 return 0;
1345
1346         xfs_icsb_sync_counters(mp);
1347
1348         /*
1349          * we don't need to do this if we are updating the superblock
1350          * counters on every modification.
1351          */
1352         if (!xfs_sb_version_haslazysbcount(&mp->m_sb))
1353                 return 0;
1354
1355         tp = _xfs_trans_alloc(mp, XFS_TRANS_SB_COUNT);
1356         error = xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1357                                         XFS_DEFAULT_LOG_COUNT);
1358         if (error) {
1359                 xfs_trans_cancel(tp, 0);
1360                 return error;
1361         }
1362
1363         xfs_mod_sb(tp, XFS_SB_IFREE | XFS_SB_ICOUNT | XFS_SB_FDBLOCKS);
1364         if (sync)
1365                 xfs_trans_set_sync(tp);
1366         xfs_trans_commit(tp, 0);
1367
1368         return 0;
1369 }
1370
1371 STATIC void
1372 xfs_mark_shared_ro(
1373         xfs_mount_t     *mp,
1374         xfs_buf_t       *bp)
1375 {
1376         xfs_dsb_t       *sb = XFS_BUF_TO_SBP(bp);
1377         __uint16_t      version;
1378
1379         if (!(sb->sb_flags & XFS_SBF_READONLY))
1380                 sb->sb_flags |= XFS_SBF_READONLY;
1381
1382         version = be16_to_cpu(sb->sb_versionnum);
1383         if ((version & XFS_SB_VERSION_NUMBITS) != XFS_SB_VERSION_4 ||
1384             !(version & XFS_SB_VERSION_SHAREDBIT))
1385                 version |= XFS_SB_VERSION_SHAREDBIT;
1386         sb->sb_versionnum = cpu_to_be16(version);
1387 }
1388
1389 int
1390 xfs_unmountfs_writesb(xfs_mount_t *mp)
1391 {
1392         xfs_buf_t       *sbp;
1393         int             error = 0;
1394
1395         /*
1396          * skip superblock write if fs is read-only, or
1397          * if we are doing a forced umount.
1398          */
1399         if (!(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY ||
1400                 XFS_FORCED_SHUTDOWN(mp))) {
1401
1402                 sbp = xfs_getsb(mp, 0);
1403
1404                 /*
1405                  * mark shared-readonly if desired
1406                  */
1407                 if (mp->m_mk_sharedro)
1408                         xfs_mark_shared_ro(mp, sbp);
1409
1410                 XFS_BUF_UNDONE(sbp);
1411                 XFS_BUF_UNREAD(sbp);
1412                 XFS_BUF_UNDELAYWRITE(sbp);
1413                 XFS_BUF_WRITE(sbp);
1414                 XFS_BUF_UNASYNC(sbp);
1415                 ASSERT(XFS_BUF_TARGET(sbp) == mp->m_ddev_targp);
1416                 xfsbdstrat(mp, sbp);
1417                 /* Nevermind errors we might get here. */
1418                 error = xfs_iowait(sbp);
1419                 if (error)
1420                         xfs_ioerror_alert("xfs_unmountfs_writesb",
1421                                           mp, sbp, XFS_BUF_ADDR(sbp));
1422                 if (error && mp->m_mk_sharedro)
1423                         xfs_fs_cmn_err(CE_ALERT, mp, "Superblock write error detected while unmounting.  Filesystem may not be marked shared readonly");
1424                 xfs_buf_relse(sbp);
1425         }
1426         return error;
1427 }
1428
1429 /*
1430  * xfs_mod_sb() can be used to copy arbitrary changes to the
1431  * in-core superblock into the superblock buffer to be logged.
1432  * It does not provide the higher level of locking that is
1433  * needed to protect the in-core superblock from concurrent
1434  * access.
1435  */
1436 void
1437 xfs_mod_sb(xfs_trans_t *tp, __int64_t fields)
1438 {
1439         xfs_buf_t       *bp;
1440         int             first;
1441         int             last;
1442         xfs_mount_t     *mp;
1443         xfs_sb_field_t  f;
1444
1445         ASSERT(fields);
1446         if (!fields)
1447                 return;
1448         mp = tp->t_mountp;
1449         bp = xfs_trans_getsb(tp, mp, 0);
1450         first = sizeof(xfs_sb_t);
1451         last = 0;
1452
1453         /* translate/copy */
1454
1455         xfs_sb_to_disk(XFS_BUF_TO_SBP(bp), &mp->m_sb, fields);
1456
1457         /* find modified range */
1458
1459         f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
1460         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1461         first = xfs_sb_info[f].offset;
1462
1463         f = (xfs_sb_field_t)xfs_highbit64((__uint64_t)fields);
1464         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1465         last = xfs_sb_info[f + 1].offset - 1;
1466
1467         xfs_trans_log_buf(tp, bp, first, last);
1468 }
1469
1470
1471 /*
1472  * xfs_mod_incore_sb_unlocked() is a utility routine common used to apply
1473  * a delta to a specified field in the in-core superblock.  Simply
1474  * switch on the field indicated and apply the delta to that field.
1475  * Fields are not allowed to dip below zero, so if the delta would
1476  * do this do not apply it and return EINVAL.
1477  *
1478  * The SB_LOCK must be held when this routine is called.
1479  */
1480 int
1481 xfs_mod_incore_sb_unlocked(
1482         xfs_mount_t     *mp,
1483         xfs_sb_field_t  field,
1484         int64_t         delta,
1485         int             rsvd)
1486 {
1487         int             scounter;       /* short counter for 32 bit fields */
1488         long long       lcounter;       /* long counter for 64 bit fields */
1489         long long       res_used, rem;
1490
1491         /*
1492          * With the in-core superblock spin lock held, switch
1493          * on the indicated field.  Apply the delta to the
1494          * proper field.  If the fields value would dip below
1495          * 0, then do not apply the delta and return EINVAL.
1496          */
1497         switch (field) {
1498         case XFS_SBS_ICOUNT:
1499                 lcounter = (long long)mp->m_sb.sb_icount;
1500                 lcounter += delta;
1501                 if (lcounter < 0) {
1502                         ASSERT(0);
1503                         return XFS_ERROR(EINVAL);
1504                 }
1505                 mp->m_sb.sb_icount = lcounter;
1506                 return 0;
1507         case XFS_SBS_IFREE:
1508                 lcounter = (long long)mp->m_sb.sb_ifree;
1509                 lcounter += delta;
1510                 if (lcounter < 0) {
1511                         ASSERT(0);
1512                         return XFS_ERROR(EINVAL);
1513                 }
1514                 mp->m_sb.sb_ifree = lcounter;
1515                 return 0;
1516         case XFS_SBS_FDBLOCKS:
1517                 lcounter = (long long)
1518                         mp->m_sb.sb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
1519                 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1520
1521                 if (delta > 0) {                /* Putting blocks back */
1522                         if (res_used > delta) {
1523                                 mp->m_resblks_avail += delta;
1524                         } else {
1525                                 rem = delta - res_used;
1526                                 mp->m_resblks_avail = mp->m_resblks;
1527                                 lcounter += rem;
1528                         }
1529                 } else {                                /* Taking blocks away */
1530
1531                         lcounter += delta;
1532
1533                 /*
1534                  * If were out of blocks, use any available reserved blocks if
1535                  * were allowed to.
1536                  */
1537
1538                         if (lcounter < 0) {
1539                                 if (rsvd) {
1540                                         lcounter = (long long)mp->m_resblks_avail + delta;
1541                                         if (lcounter < 0) {
1542                                                 return XFS_ERROR(ENOSPC);
1543                                         }
1544                                         mp->m_resblks_avail = lcounter;
1545                                         return 0;
1546                                 } else {        /* not reserved */
1547                                         return XFS_ERROR(ENOSPC);
1548                                 }
1549                         }
1550                 }
1551
1552                 mp->m_sb.sb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
1553                 return 0;
1554         case XFS_SBS_FREXTENTS:
1555                 lcounter = (long long)mp->m_sb.sb_frextents;
1556                 lcounter += delta;
1557                 if (lcounter < 0) {
1558                         return XFS_ERROR(ENOSPC);
1559                 }
1560                 mp->m_sb.sb_frextents = lcounter;
1561                 return 0;
1562         case XFS_SBS_DBLOCKS:
1563                 lcounter = (long long)mp->m_sb.sb_dblocks;
1564                 lcounter += delta;
1565                 if (lcounter < 0) {
1566                         ASSERT(0);
1567                         return XFS_ERROR(EINVAL);
1568                 }
1569                 mp->m_sb.sb_dblocks = lcounter;
1570                 return 0;
1571         case XFS_SBS_AGCOUNT:
1572                 scounter = mp->m_sb.sb_agcount;
1573                 scounter += delta;
1574                 if (scounter < 0) {
1575                         ASSERT(0);
1576                         return XFS_ERROR(EINVAL);
1577                 }
1578                 mp->m_sb.sb_agcount = scounter;
1579                 return 0;
1580         case XFS_SBS_IMAX_PCT:
1581                 scounter = mp->m_sb.sb_imax_pct;
1582                 scounter += delta;
1583                 if (scounter < 0) {
1584                         ASSERT(0);
1585                         return XFS_ERROR(EINVAL);
1586                 }
1587                 mp->m_sb.sb_imax_pct = scounter;
1588                 return 0;
1589         case XFS_SBS_REXTSIZE:
1590                 scounter = mp->m_sb.sb_rextsize;
1591                 scounter += delta;
1592                 if (scounter < 0) {
1593                         ASSERT(0);
1594                         return XFS_ERROR(EINVAL);
1595                 }
1596                 mp->m_sb.sb_rextsize = scounter;
1597                 return 0;
1598         case XFS_SBS_RBMBLOCKS:
1599                 scounter = mp->m_sb.sb_rbmblocks;
1600                 scounter += delta;
1601                 if (scounter < 0) {
1602                         ASSERT(0);
1603                         return XFS_ERROR(EINVAL);
1604                 }
1605                 mp->m_sb.sb_rbmblocks = scounter;
1606                 return 0;
1607         case XFS_SBS_RBLOCKS:
1608                 lcounter = (long long)mp->m_sb.sb_rblocks;
1609                 lcounter += delta;
1610                 if (lcounter < 0) {
1611                         ASSERT(0);
1612                         return XFS_ERROR(EINVAL);
1613                 }
1614                 mp->m_sb.sb_rblocks = lcounter;
1615                 return 0;
1616         case XFS_SBS_REXTENTS:
1617                 lcounter = (long long)mp->m_sb.sb_rextents;
1618                 lcounter += delta;
1619                 if (lcounter < 0) {
1620                         ASSERT(0);
1621                         return XFS_ERROR(EINVAL);
1622                 }
1623                 mp->m_sb.sb_rextents = lcounter;
1624                 return 0;
1625         case XFS_SBS_REXTSLOG:
1626                 scounter = mp->m_sb.sb_rextslog;
1627                 scounter += delta;
1628                 if (scounter < 0) {
1629                         ASSERT(0);
1630                         return XFS_ERROR(EINVAL);
1631                 }
1632                 mp->m_sb.sb_rextslog = scounter;
1633                 return 0;
1634         default:
1635                 ASSERT(0);
1636                 return XFS_ERROR(EINVAL);
1637         }
1638 }
1639
1640 /*
1641  * xfs_mod_incore_sb() is used to change a field in the in-core
1642  * superblock structure by the specified delta.  This modification
1643  * is protected by the SB_LOCK.  Just use the xfs_mod_incore_sb_unlocked()
1644  * routine to do the work.
1645  */
1646 int
1647 xfs_mod_incore_sb(
1648         xfs_mount_t     *mp,
1649         xfs_sb_field_t  field,
1650         int64_t         delta,
1651         int             rsvd)
1652 {
1653         unsigned long   s;
1654         int     status;
1655
1656         /* check for per-cpu counters */
1657         switch (field) {
1658 #ifdef HAVE_PERCPU_SB
1659         case XFS_SBS_ICOUNT:
1660         case XFS_SBS_IFREE:
1661         case XFS_SBS_FDBLOCKS:
1662                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1663                         status = xfs_icsb_modify_counters(mp, field,
1664                                                         delta, rsvd);
1665                         break;
1666                 }
1667                 /* FALLTHROUGH */
1668 #endif
1669         default:
1670                 s = XFS_SB_LOCK(mp);
1671                 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1672                 XFS_SB_UNLOCK(mp, s);
1673                 break;
1674         }
1675
1676         return status;
1677 }
1678
1679 /*
1680  * xfs_mod_incore_sb_batch() is used to change more than one field
1681  * in the in-core superblock structure at a time.  This modification
1682  * is protected by a lock internal to this module.  The fields and
1683  * changes to those fields are specified in the array of xfs_mod_sb
1684  * structures passed in.
1685  *
1686  * Either all of the specified deltas will be applied or none of
1687  * them will.  If any modified field dips below 0, then all modifications
1688  * will be backed out and EINVAL will be returned.
1689  */
1690 int
1691 xfs_mod_incore_sb_batch(xfs_mount_t *mp, xfs_mod_sb_t *msb, uint nmsb, int rsvd)
1692 {
1693         unsigned long   s;
1694         int             status=0;
1695         xfs_mod_sb_t    *msbp;
1696
1697         /*
1698          * Loop through the array of mod structures and apply each
1699          * individually.  If any fail, then back out all those
1700          * which have already been applied.  Do all of this within
1701          * the scope of the SB_LOCK so that all of the changes will
1702          * be atomic.
1703          */
1704         s = XFS_SB_LOCK(mp);
1705         msbp = &msb[0];
1706         for (msbp = &msbp[0]; msbp < (msb + nmsb); msbp++) {
1707                 /*
1708                  * Apply the delta at index n.  If it fails, break
1709                  * from the loop so we'll fall into the undo loop
1710                  * below.
1711                  */
1712                 switch (msbp->msb_field) {
1713 #ifdef HAVE_PERCPU_SB
1714                 case XFS_SBS_ICOUNT:
1715                 case XFS_SBS_IFREE:
1716                 case XFS_SBS_FDBLOCKS:
1717                         if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1718                                 XFS_SB_UNLOCK(mp, s);
1719                                 status = xfs_icsb_modify_counters(mp,
1720                                                         msbp->msb_field,
1721                                                         msbp->msb_delta, rsvd);
1722                                 s = XFS_SB_LOCK(mp);
1723                                 break;
1724                         }
1725                         /* FALLTHROUGH */
1726 #endif
1727                 default:
1728                         status = xfs_mod_incore_sb_unlocked(mp,
1729                                                 msbp->msb_field,
1730                                                 msbp->msb_delta, rsvd);
1731                         break;
1732                 }
1733
1734                 if (status != 0) {
1735                         break;
1736                 }
1737         }
1738
1739         /*
1740          * If we didn't complete the loop above, then back out
1741          * any changes made to the superblock.  If you add code
1742          * between the loop above and here, make sure that you
1743          * preserve the value of status. Loop back until
1744          * we step below the beginning of the array.  Make sure
1745          * we don't touch anything back there.
1746          */
1747         if (status != 0) {
1748                 msbp--;
1749                 while (msbp >= msb) {
1750                         switch (msbp->msb_field) {
1751 #ifdef HAVE_PERCPU_SB
1752                         case XFS_SBS_ICOUNT:
1753                         case XFS_SBS_IFREE:
1754                         case XFS_SBS_FDBLOCKS:
1755                                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1756                                         XFS_SB_UNLOCK(mp, s);
1757                                         status = xfs_icsb_modify_counters(mp,
1758                                                         msbp->msb_field,
1759                                                         -(msbp->msb_delta),
1760                                                         rsvd);
1761                                         s = XFS_SB_LOCK(mp);
1762                                         break;
1763                                 }
1764                                 /* FALLTHROUGH */
1765 #endif
1766                         default:
1767                                 status = xfs_mod_incore_sb_unlocked(mp,
1768                                                         msbp->msb_field,
1769                                                         -(msbp->msb_delta),
1770                                                         rsvd);
1771                                 break;
1772                         }
1773                         ASSERT(status == 0);
1774                         msbp--;
1775                 }
1776         }
1777         XFS_SB_UNLOCK(mp, s);
1778         return status;
1779 }
1780
1781 /*
1782  * xfs_getsb() is called to obtain the buffer for the superblock.
1783  * The buffer is returned locked and read in from disk.
1784  * The buffer should be released with a call to xfs_brelse().
1785  *
1786  * If the flags parameter is BUF_TRYLOCK, then we'll only return
1787  * the superblock buffer if it can be locked without sleeping.
1788  * If it can't then we'll return NULL.
1789  */
1790 xfs_buf_t *
1791 xfs_getsb(
1792         xfs_mount_t     *mp,
1793         int             flags)
1794 {
1795         xfs_buf_t       *bp;
1796
1797         ASSERT(mp->m_sb_bp != NULL);
1798         bp = mp->m_sb_bp;
1799         if (flags & XFS_BUF_TRYLOCK) {
1800                 if (!XFS_BUF_CPSEMA(bp)) {
1801                         return NULL;
1802                 }
1803         } else {
1804                 XFS_BUF_PSEMA(bp, PRIBIO);
1805         }
1806         XFS_BUF_HOLD(bp);
1807         ASSERT(XFS_BUF_ISDONE(bp));
1808         return bp;
1809 }
1810
1811 /*
1812  * Used to free the superblock along various error paths.
1813  */
1814 void
1815 xfs_freesb(
1816         xfs_mount_t     *mp)
1817 {
1818         xfs_buf_t       *bp;
1819
1820         /*
1821          * Use xfs_getsb() so that the buffer will be locked
1822          * when we call xfs_buf_relse().
1823          */
1824         bp = xfs_getsb(mp, 0);
1825         XFS_BUF_UNMANAGE(bp);
1826         xfs_buf_relse(bp);
1827         mp->m_sb_bp = NULL;
1828 }
1829
1830 /*
1831  * See if the UUID is unique among mounted XFS filesystems.
1832  * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
1833  */
1834 STATIC int
1835 xfs_uuid_mount(
1836         xfs_mount_t     *mp)
1837 {
1838         if (uuid_is_nil(&mp->m_sb.sb_uuid)) {
1839                 cmn_err(CE_WARN,
1840                         "XFS: Filesystem %s has nil UUID - can't mount",
1841                         mp->m_fsname);
1842                 return -1;
1843         }
1844         if (!uuid_table_insert(&mp->m_sb.sb_uuid)) {
1845                 cmn_err(CE_WARN,
1846                         "XFS: Filesystem %s has duplicate UUID - can't mount",
1847                         mp->m_fsname);
1848                 return -1;
1849         }
1850         return 0;
1851 }
1852
1853 /*
1854  * Remove filesystem from the UUID table.
1855  */
1856 STATIC void
1857 xfs_uuid_unmount(
1858         xfs_mount_t     *mp)
1859 {
1860         uuid_table_remove(&mp->m_sb.sb_uuid);
1861 }
1862
1863 /*
1864  * Used to log changes to the superblock unit and width fields which could
1865  * be altered by the mount options. Only the first superblock is updated.
1866  */
1867 STATIC void
1868 xfs_mount_log_sbunit(
1869         xfs_mount_t     *mp,
1870         __int64_t       fields)
1871 {
1872         xfs_trans_t     *tp;
1873
1874         ASSERT(fields & (XFS_SB_UNIT|XFS_SB_WIDTH|XFS_SB_UUID));
1875
1876         tp = xfs_trans_alloc(mp, XFS_TRANS_SB_UNIT);
1877         if (xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1878                                 XFS_DEFAULT_LOG_COUNT)) {
1879                 xfs_trans_cancel(tp, 0);
1880                 return;
1881         }
1882         xfs_mod_sb(tp, fields);
1883         xfs_trans_commit(tp, 0);
1884 }
1885
1886
1887 #ifdef HAVE_PERCPU_SB
1888 /*
1889  * Per-cpu incore superblock counters
1890  *
1891  * Simple concept, difficult implementation
1892  *
1893  * Basically, replace the incore superblock counters with a distributed per cpu
1894  * counter for contended fields (e.g.  free block count).
1895  *
1896  * Difficulties arise in that the incore sb is used for ENOSPC checking, and
1897  * hence needs to be accurately read when we are running low on space. Hence
1898  * there is a method to enable and disable the per-cpu counters based on how
1899  * much "stuff" is available in them.
1900  *
1901  * Basically, a counter is enabled if there is enough free resource to justify
1902  * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
1903  * ENOSPC), then we disable the counters to synchronise all callers and
1904  * re-distribute the available resources.
1905  *
1906  * If, once we redistributed the available resources, we still get a failure,
1907  * we disable the per-cpu counter and go through the slow path.
1908  *
1909  * The slow path is the current xfs_mod_incore_sb() function.  This means that
1910  * when we disable a per-cpu counter, we need to drain it's resources back to
1911  * the global superblock. We do this after disabling the counter to prevent
1912  * more threads from queueing up on the counter.
1913  *
1914  * Essentially, this means that we still need a lock in the fast path to enable
1915  * synchronisation between the global counters and the per-cpu counters. This
1916  * is not a problem because the lock will be local to a CPU almost all the time
1917  * and have little contention except when we get to ENOSPC conditions.
1918  *
1919  * Basically, this lock becomes a barrier that enables us to lock out the fast
1920  * path while we do things like enabling and disabling counters and
1921  * synchronising the counters.
1922  *
1923  * Locking rules:
1924  *
1925  *      1. XFS_SB_LOCK() before picking up per-cpu locks
1926  *      2. per-cpu locks always picked up via for_each_online_cpu() order
1927  *      3. accurate counter sync requires XFS_SB_LOCK + per cpu locks
1928  *      4. modifying per-cpu counters requires holding per-cpu lock
1929  *      5. modifying global counters requires holding XFS_SB_LOCK
1930  *      6. enabling or disabling a counter requires holding the XFS_SB_LOCK
1931  *         and _none_ of the per-cpu locks.
1932  *
1933  * Disabled counters are only ever re-enabled by a balance operation
1934  * that results in more free resources per CPU than a given threshold.
1935  * To ensure counters don't remain disabled, they are rebalanced when
1936  * the global resource goes above a higher threshold (i.e. some hysteresis
1937  * is present to prevent thrashing).
1938  */
1939
1940 #ifdef CONFIG_HOTPLUG_CPU
1941 /*
1942  * hot-plug CPU notifier support.
1943  *
1944  * We need a notifier per filesystem as we need to be able to identify
1945  * the filesystem to balance the counters out. This is achieved by
1946  * having a notifier block embedded in the xfs_mount_t and doing pointer
1947  * magic to get the mount pointer from the notifier block address.
1948  */
1949 STATIC int
1950 xfs_icsb_cpu_notify(
1951         struct notifier_block *nfb,
1952         unsigned long action,
1953         void *hcpu)
1954 {
1955         xfs_icsb_cnts_t *cntp;
1956         xfs_mount_t     *mp;
1957         int             s;
1958
1959         mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
1960         cntp = (xfs_icsb_cnts_t *)
1961                         per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
1962         switch (action) {
1963         case CPU_UP_PREPARE:
1964         case CPU_UP_PREPARE_FROZEN:
1965                 /* Easy Case - initialize the area and locks, and
1966                  * then rebalance when online does everything else for us. */
1967                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1968                 break;
1969         case CPU_ONLINE:
1970         case CPU_ONLINE_FROZEN:
1971                 xfs_icsb_lock(mp);
1972                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0, 0);
1973                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0, 0);
1974                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0, 0);
1975                 xfs_icsb_unlock(mp);
1976                 break;
1977         case CPU_DEAD:
1978         case CPU_DEAD_FROZEN:
1979                 /* Disable all the counters, then fold the dead cpu's
1980                  * count into the total on the global superblock and
1981                  * re-enable the counters. */
1982                 xfs_icsb_lock(mp);
1983                 s = XFS_SB_LOCK(mp);
1984                 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
1985                 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
1986                 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
1987
1988                 mp->m_sb.sb_icount += cntp->icsb_icount;
1989                 mp->m_sb.sb_ifree += cntp->icsb_ifree;
1990                 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
1991
1992                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1993
1994                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT,
1995                                          XFS_ICSB_SB_LOCKED, 0);
1996                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE,
1997                                          XFS_ICSB_SB_LOCKED, 0);
1998                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS,
1999                                          XFS_ICSB_SB_LOCKED, 0);
2000                 XFS_SB_UNLOCK(mp, s);
2001                 xfs_icsb_unlock(mp);
2002                 break;
2003         }
2004
2005         return NOTIFY_OK;
2006 }
2007 #endif /* CONFIG_HOTPLUG_CPU */
2008
2009 int
2010 xfs_icsb_init_counters(
2011         xfs_mount_t     *mp)
2012 {
2013         xfs_icsb_cnts_t *cntp;
2014         int             i;
2015
2016         mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
2017         if (mp->m_sb_cnts == NULL)
2018                 return -ENOMEM;
2019
2020 #ifdef CONFIG_HOTPLUG_CPU
2021         mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
2022         mp->m_icsb_notifier.priority = 0;
2023         register_hotcpu_notifier(&mp->m_icsb_notifier);
2024 #endif /* CONFIG_HOTPLUG_CPU */
2025
2026         for_each_online_cpu(i) {
2027                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2028                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
2029         }
2030
2031         mutex_init(&mp->m_icsb_mutex);
2032
2033         /*
2034          * start with all counters disabled so that the
2035          * initial balance kicks us off correctly
2036          */
2037         mp->m_icsb_counters = -1;
2038         return 0;
2039 }
2040
2041 void
2042 xfs_icsb_reinit_counters(
2043         xfs_mount_t     *mp)
2044 {
2045         xfs_icsb_lock(mp);
2046         /*
2047          * start with all counters disabled so that the
2048          * initial balance kicks us off correctly
2049          */
2050         mp->m_icsb_counters = -1;
2051         xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0, 0);
2052         xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0, 0);
2053         xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0, 0);
2054         xfs_icsb_unlock(mp);
2055 }
2056
2057 STATIC void
2058 xfs_icsb_destroy_counters(
2059         xfs_mount_t     *mp)
2060 {
2061         if (mp->m_sb_cnts) {
2062                 unregister_hotcpu_notifier(&mp->m_icsb_notifier);
2063                 free_percpu(mp->m_sb_cnts);
2064         }
2065         mutex_destroy(&mp->m_icsb_mutex);
2066 }
2067
2068 STATIC_INLINE void
2069 xfs_icsb_lock_cntr(
2070         xfs_icsb_cnts_t *icsbp)
2071 {
2072         while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
2073                 ndelay(1000);
2074         }
2075 }
2076
2077 STATIC_INLINE void
2078 xfs_icsb_unlock_cntr(
2079         xfs_icsb_cnts_t *icsbp)
2080 {
2081         clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
2082 }
2083
2084
2085 STATIC_INLINE void
2086 xfs_icsb_lock_all_counters(
2087         xfs_mount_t     *mp)
2088 {
2089         xfs_icsb_cnts_t *cntp;
2090         int             i;
2091
2092         for_each_online_cpu(i) {
2093                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2094                 xfs_icsb_lock_cntr(cntp);
2095         }
2096 }
2097
2098 STATIC_INLINE void
2099 xfs_icsb_unlock_all_counters(
2100         xfs_mount_t     *mp)
2101 {
2102         xfs_icsb_cnts_t *cntp;
2103         int             i;
2104
2105         for_each_online_cpu(i) {
2106                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2107                 xfs_icsb_unlock_cntr(cntp);
2108         }
2109 }
2110
2111 STATIC void
2112 xfs_icsb_count(
2113         xfs_mount_t     *mp,
2114         xfs_icsb_cnts_t *cnt,
2115         int             flags)
2116 {
2117         xfs_icsb_cnts_t *cntp;
2118         int             i;
2119
2120         memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
2121
2122         if (!(flags & XFS_ICSB_LAZY_COUNT))
2123                 xfs_icsb_lock_all_counters(mp);
2124
2125         for_each_online_cpu(i) {
2126                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2127                 cnt->icsb_icount += cntp->icsb_icount;
2128                 cnt->icsb_ifree += cntp->icsb_ifree;
2129                 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
2130         }
2131
2132         if (!(flags & XFS_ICSB_LAZY_COUNT))
2133                 xfs_icsb_unlock_all_counters(mp);
2134 }
2135
2136 STATIC int
2137 xfs_icsb_counter_disabled(
2138         xfs_mount_t     *mp,
2139         xfs_sb_field_t  field)
2140 {
2141         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2142         return test_bit(field, &mp->m_icsb_counters);
2143 }
2144
2145 STATIC int
2146 xfs_icsb_disable_counter(
2147         xfs_mount_t     *mp,
2148         xfs_sb_field_t  field)
2149 {
2150         xfs_icsb_cnts_t cnt;
2151
2152         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2153
2154         /*
2155          * If we are already disabled, then there is nothing to do
2156          * here. We check before locking all the counters to avoid
2157          * the expensive lock operation when being called in the
2158          * slow path and the counter is already disabled. This is
2159          * safe because the only time we set or clear this state is under
2160          * the m_icsb_mutex.
2161          */
2162         if (xfs_icsb_counter_disabled(mp, field))
2163                 return 0;
2164
2165         xfs_icsb_lock_all_counters(mp);
2166         if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
2167                 /* drain back to superblock */
2168
2169                 xfs_icsb_count(mp, &cnt, XFS_ICSB_SB_LOCKED|XFS_ICSB_LAZY_COUNT);
2170                 switch(field) {
2171                 case XFS_SBS_ICOUNT:
2172                         mp->m_sb.sb_icount = cnt.icsb_icount;
2173                         break;
2174                 case XFS_SBS_IFREE:
2175                         mp->m_sb.sb_ifree = cnt.icsb_ifree;
2176                         break;
2177                 case XFS_SBS_FDBLOCKS:
2178                         mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2179                         break;
2180                 default:
2181                         BUG();
2182                 }
2183         }
2184
2185         xfs_icsb_unlock_all_counters(mp);
2186
2187         return 0;
2188 }
2189
2190 STATIC void
2191 xfs_icsb_enable_counter(
2192         xfs_mount_t     *mp,
2193         xfs_sb_field_t  field,
2194         uint64_t        count,
2195         uint64_t        resid)
2196 {
2197         xfs_icsb_cnts_t *cntp;
2198         int             i;
2199
2200         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2201
2202         xfs_icsb_lock_all_counters(mp);
2203         for_each_online_cpu(i) {
2204                 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
2205                 switch (field) {
2206                 case XFS_SBS_ICOUNT:
2207                         cntp->icsb_icount = count + resid;
2208                         break;
2209                 case XFS_SBS_IFREE:
2210                         cntp->icsb_ifree = count + resid;
2211                         break;
2212                 case XFS_SBS_FDBLOCKS:
2213                         cntp->icsb_fdblocks = count + resid;
2214                         break;
2215                 default:
2216                         BUG();
2217                         break;
2218                 }
2219                 resid = 0;
2220         }
2221         clear_bit(field, &mp->m_icsb_counters);
2222         xfs_icsb_unlock_all_counters(mp);
2223 }
2224
2225 void
2226 xfs_icsb_sync_counters_flags(
2227         xfs_mount_t     *mp,
2228         int             flags)
2229 {
2230         xfs_icsb_cnts_t cnt;
2231         int             s;
2232
2233         /* Pass 1: lock all counters */
2234         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
2235                 s = XFS_SB_LOCK(mp);
2236
2237         xfs_icsb_count(mp, &cnt, flags);
2238
2239         /* Step 3: update mp->m_sb fields */
2240         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
2241                 mp->m_sb.sb_icount = cnt.icsb_icount;
2242         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
2243                 mp->m_sb.sb_ifree = cnt.icsb_ifree;
2244         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
2245                 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2246
2247         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
2248                 XFS_SB_UNLOCK(mp, s);
2249 }
2250
2251 /*
2252  * Accurate update of per-cpu counters to incore superblock
2253  */
2254 STATIC void
2255 xfs_icsb_sync_counters(
2256         xfs_mount_t     *mp)
2257 {
2258         xfs_icsb_sync_counters_flags(mp, 0);
2259 }
2260
2261 /*
2262  * Balance and enable/disable counters as necessary.
2263  *
2264  * Thresholds for re-enabling counters are somewhat magic.  inode counts are
2265  * chosen to be the same number as single on disk allocation chunk per CPU, and
2266  * free blocks is something far enough zero that we aren't going thrash when we
2267  * get near ENOSPC. We also need to supply a minimum we require per cpu to
2268  * prevent looping endlessly when xfs_alloc_space asks for more than will
2269  * be distributed to a single CPU but each CPU has enough blocks to be
2270  * reenabled.
2271  *
2272  * Note that we can be called when counters are already disabled.
2273  * xfs_icsb_disable_counter() optimises the counter locking in this case to
2274  * prevent locking every per-cpu counter needlessly.
2275  */
2276
2277 #define XFS_ICSB_INO_CNTR_REENABLE      (uint64_t)64
2278 #define XFS_ICSB_FDBLK_CNTR_REENABLE(mp) \
2279                 (uint64_t)(512 + XFS_ALLOC_SET_ASIDE(mp))
2280 STATIC void
2281 xfs_icsb_balance_counter(
2282         xfs_mount_t     *mp,
2283         xfs_sb_field_t  field,
2284         int             flags,
2285         int             min_per_cpu)
2286 {
2287         uint64_t        count, resid;
2288         int             weight = num_online_cpus();
2289         int             s;
2290         uint64_t        min = (uint64_t)min_per_cpu;
2291
2292         if (!(flags & XFS_ICSB_SB_LOCKED))
2293                 s = XFS_SB_LOCK(mp);
2294
2295         /* disable counter and sync counter */
2296         xfs_icsb_disable_counter(mp, field);
2297
2298         /* update counters  - first CPU gets residual*/
2299         switch (field) {
2300         case XFS_SBS_ICOUNT:
2301                 count = mp->m_sb.sb_icount;
2302                 resid = do_div(count, weight);
2303                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2304                         goto out;
2305                 break;
2306         case XFS_SBS_IFREE:
2307                 count = mp->m_sb.sb_ifree;
2308                 resid = do_div(count, weight);
2309                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2310                         goto out;
2311                 break;
2312         case XFS_SBS_FDBLOCKS:
2313                 count = mp->m_sb.sb_fdblocks;
2314                 resid = do_div(count, weight);
2315                 if (count < max(min, XFS_ICSB_FDBLK_CNTR_REENABLE(mp)))
2316                         goto out;
2317                 break;
2318         default:
2319                 BUG();
2320                 count = resid = 0;      /* quiet, gcc */
2321                 break;
2322         }
2323
2324         xfs_icsb_enable_counter(mp, field, count, resid);
2325 out:
2326         if (!(flags & XFS_ICSB_SB_LOCKED))
2327                 XFS_SB_UNLOCK(mp, s);
2328 }
2329
2330 int
2331 xfs_icsb_modify_counters(
2332         xfs_mount_t     *mp,
2333         xfs_sb_field_t  field,
2334         int64_t         delta,
2335         int             rsvd)
2336 {
2337         xfs_icsb_cnts_t *icsbp;
2338         long long       lcounter;       /* long counter for 64 bit fields */
2339         int             cpu, ret = 0, s;
2340
2341         might_sleep();
2342 again:
2343         cpu = get_cpu();
2344         icsbp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, cpu);
2345
2346         /*
2347          * if the counter is disabled, go to slow path
2348          */
2349         if (unlikely(xfs_icsb_counter_disabled(mp, field)))
2350                 goto slow_path;
2351         xfs_icsb_lock_cntr(icsbp);
2352         if (unlikely(xfs_icsb_counter_disabled(mp, field))) {
2353                 xfs_icsb_unlock_cntr(icsbp);
2354                 goto slow_path;
2355         }
2356
2357         switch (field) {
2358         case XFS_SBS_ICOUNT:
2359                 lcounter = icsbp->icsb_icount;
2360                 lcounter += delta;
2361                 if (unlikely(lcounter < 0))
2362                         goto balance_counter;
2363                 icsbp->icsb_icount = lcounter;
2364                 break;
2365
2366         case XFS_SBS_IFREE:
2367                 lcounter = icsbp->icsb_ifree;
2368                 lcounter += delta;
2369                 if (unlikely(lcounter < 0))
2370                         goto balance_counter;
2371                 icsbp->icsb_ifree = lcounter;
2372                 break;
2373
2374         case XFS_SBS_FDBLOCKS:
2375                 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
2376
2377                 lcounter = icsbp->icsb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
2378                 lcounter += delta;
2379                 if (unlikely(lcounter < 0))
2380                         goto balance_counter;
2381                 icsbp->icsb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
2382                 break;
2383         default:
2384                 BUG();
2385                 break;
2386         }
2387         xfs_icsb_unlock_cntr(icsbp);
2388         put_cpu();
2389         return 0;
2390
2391 slow_path:
2392         put_cpu();
2393
2394         /*
2395          * serialise with a mutex so we don't burn lots of cpu on
2396          * the superblock lock. We still need to hold the superblock
2397          * lock, however, when we modify the global structures.
2398          */
2399         xfs_icsb_lock(mp);
2400
2401         /*
2402          * Now running atomically.
2403          *
2404          * If the counter is enabled, someone has beaten us to rebalancing.
2405          * Drop the lock and try again in the fast path....
2406          */
2407         if (!(xfs_icsb_counter_disabled(mp, field))) {
2408                 xfs_icsb_unlock(mp);
2409                 goto again;
2410         }
2411
2412         /*
2413          * The counter is currently disabled. Because we are
2414          * running atomically here, we know a rebalance cannot
2415          * be in progress. Hence we can go straight to operating
2416          * on the global superblock. We do not call xfs_mod_incore_sb()
2417          * here even though we need to get the SB_LOCK. Doing so
2418          * will cause us to re-enter this function and deadlock.
2419          * Hence we get the SB_LOCK ourselves and then call
2420          * xfs_mod_incore_sb_unlocked() as the unlocked path operates
2421          * directly on the global counters.
2422          */
2423         s = XFS_SB_LOCK(mp);
2424         ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
2425         XFS_SB_UNLOCK(mp, s);
2426
2427         /*
2428          * Now that we've modified the global superblock, we
2429          * may be able to re-enable the distributed counters
2430          * (e.g. lots of space just got freed). After that
2431          * we are done.
2432          */
2433         if (ret != ENOSPC)
2434                 xfs_icsb_balance_counter(mp, field, 0, 0);
2435         xfs_icsb_unlock(mp);
2436         return ret;
2437
2438 balance_counter:
2439         xfs_icsb_unlock_cntr(icsbp);
2440         put_cpu();
2441
2442         /*
2443          * We may have multiple threads here if multiple per-cpu
2444          * counters run dry at the same time. This will mean we can
2445          * do more balances than strictly necessary but it is not
2446          * the common slowpath case.
2447          */
2448         xfs_icsb_lock(mp);
2449
2450         /*
2451          * running atomically.
2452          *
2453          * This will leave the counter in the correct state for future
2454          * accesses. After the rebalance, we simply try again and our retry
2455          * will either succeed through the fast path or slow path without
2456          * another balance operation being required.
2457          */
2458         xfs_icsb_balance_counter(mp, field, 0, delta);
2459         xfs_icsb_unlock(mp);
2460         goto again;
2461 }
2462
2463 #endif