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         spin_lock_init(&mp->m_ail_lock);
 140         spin_lock_init(&mp->m_sb_lock);
 141         mutex_init(&mp->m_ilock);
 142         mutex_init(&mp->m_growlock);
 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 {
 162         if (mp->m_perag) {
 163                 int     agno;
 164
 165                 for (agno = 0; agno < mp->m_maxagi; agno++)
 166                         if (mp->m_perag[agno].pagb_list)
 167                                 kmem_free(mp->m_perag[agno].pagb_list,
 168                                                 sizeof(xfs_perag_busy_t) *
 169                                                         XFS_PAGB_NUM_SLOTS);
 170                 kmem_free(mp->m_perag,
 171                           sizeof(xfs_perag_t) * mp->m_sb.sb_agcount);
 172         }
 173
 174         spinlock_destroy(&mp->m_ail_lock);
 175         spinlock_destroy(&mp->m_sb_lock);
 176         mutex_destroy(&mp->m_ilock);
 177         mutex_destroy(&mp->m_growlock);
 178         if (mp->m_quotainfo)
 179                 XFS_QM_DONE(mp);
 180
 181         if (mp->m_fsname != NULL)
 182                 kmem_free(mp->m_fsname, mp->m_fsname_len);
 183         if (mp->m_rtname != NULL)
 184                 kmem_free(mp->m_rtname, strlen(mp->m_rtname) + 1);
 185         if (mp->m_logname != NULL)
 186                 kmem_free(mp->m_logname, strlen(mp->m_logname) + 1);
 187
 188         xfs_icsb_destroy_counters(mp);
 189 }
 190
 191 /*
 192  * Check size of device based on the (data/realtime) block count.
 193  * Note: this check is used by the growfs code as well as mount.
 194  */
 195 int
 196 xfs_sb_validate_fsb_count(
 197         xfs_sb_t        *sbp,
 198         __uint64_t      nblocks)
 199 {
 200         ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
 201         ASSERT(sbp->sb_blocklog >= BBSHIFT);
 202
 203 #if XFS_BIG_BLKNOS     /* Limited by ULONG_MAX of page cache index */
 204         if (nblocks >> (PAGE_CACHE_SHIFT - sbp->sb_blocklog) > ULONG_MAX)
 205                 return E2BIG;
 206 #else                  /* Limited by UINT_MAX of sectors */
 207         if (nblocks << (sbp->sb_blocklog - BBSHIFT) > UINT_MAX)
 208                 return E2BIG;
 209 #endif
 210         return 0;
 211 }
 212
 213 /*
 214  * Check the validity of the SB found.
 215  */
 216 STATIC int
 217 xfs_mount_validate_sb(
 218         xfs_mount_t     *mp,
 219         xfs_sb_t        *sbp,
 220         int             flags)
 221 {
 222         /*
 223          * If the log device and data device have the
 224          * same device number, the log is internal.
 225          * Consequently, the sb_logstart should be non-zero.  If
 226          * we have a zero sb_logstart in this case, we may be trying to mount
 227          * a volume filesystem in a non-volume manner.
 228          */
 229         if (sbp->sb_magicnum != XFS_SB_MAGIC) {
 230                 xfs_fs_mount_cmn_err(flags, "bad magic number");
 231                 return XFS_ERROR(EWRONGFS);
 232         }
 233
 234         if (!XFS_SB_GOOD_VERSION(sbp)) {
 235                 xfs_fs_mount_cmn_err(flags, "bad version");
 236                 return XFS_ERROR(EWRONGFS);
 237         }
 238
 239         if (unlikely(
 240             sbp->sb_logstart == 0 && mp->m_logdev_targp == mp->m_ddev_targp)) {
 241                 xfs_fs_mount_cmn_err(flags,
 242                         "filesystem is marked as having an external log; "
 243                         "specify logdev on the\nmount command line.");
 244                 return XFS_ERROR(EINVAL);
 245         }
 246
 247         if (unlikely(
 248             sbp->sb_logstart != 0 && mp->m_logdev_targp != mp->m_ddev_targp)) {
 249                 xfs_fs_mount_cmn_err(flags,
 250                         "filesystem is marked as having an internal log; "
 251                         "do not specify logdev on\nthe mount command line.");
 252                 return XFS_ERROR(EINVAL);
 253         }
 254
 255         /*
 256          * More sanity checking. These were stolen directly from
 257          * xfs_repair.
 258          */
 259         if (unlikely(
 260             sbp->sb_agcount <= 0                                        ||
 261             sbp->sb_sectsize < XFS_MIN_SECTORSIZE                       ||
 262             sbp->sb_sectsize > XFS_MAX_SECTORSIZE                       ||
 263             sbp->sb_sectlog < XFS_MIN_SECTORSIZE_LOG                    ||
 264             sbp->sb_sectlog > XFS_MAX_SECTORSIZE_LOG                    ||
 265             sbp->sb_blocksize < XFS_MIN_BLOCKSIZE                       ||
 266             sbp->sb_blocksize > XFS_MAX_BLOCKSIZE                       ||
 267             sbp->sb_blocklog < XFS_MIN_BLOCKSIZE_LOG                    ||
 268             sbp->sb_blocklog > XFS_MAX_BLOCKSIZE_LOG                    ||
 269             sbp->sb_inodesize < XFS_DINODE_MIN_SIZE                     ||
 270             sbp->sb_inodesize > XFS_DINODE_MAX_SIZE                     ||
 271             sbp->sb_inodelog < XFS_DINODE_MIN_LOG                       ||
 272             sbp->sb_inodelog > XFS_DINODE_MAX_LOG                       ||
 273             (sbp->sb_blocklog - sbp->sb_inodelog != sbp->sb_inopblog)   ||
 274             (sbp->sb_rextsize * sbp->sb_blocksize > XFS_MAX_RTEXTSIZE)  ||
 275             (sbp->sb_rextsize * sbp->sb_blocksize < XFS_MIN_RTEXTSIZE)  ||
 276             (sbp->sb_imax_pct > 100 /* zero sb_imax_pct is valid */))) {
 277                 xfs_fs_mount_cmn_err(flags, "SB sanity check 1 failed");
 278                 return XFS_ERROR(EFSCORRUPTED);
 279         }
 280
 281         /*
 282          * Sanity check AG count, size fields against data size field
 283          */
 284         if (unlikely(
 285             sbp->sb_dblocks == 0 ||
 286             sbp->sb_dblocks >
 287              (xfs_drfsbno_t)sbp->sb_agcount * sbp->sb_agblocks ||
 288             sbp->sb_dblocks < (xfs_drfsbno_t)(sbp->sb_agcount - 1) *
 289                               sbp->sb_agblocks + XFS_MIN_AG_BLOCKS)) {
 290                 xfs_fs_mount_cmn_err(flags, "SB sanity check 2 failed");
 291                 return XFS_ERROR(EFSCORRUPTED);
 292         }
 293
 294         if (xfs_sb_validate_fsb_count(sbp, sbp->sb_dblocks) ||
 295             xfs_sb_validate_fsb_count(sbp, sbp->sb_rblocks)) {
 296                 xfs_fs_mount_cmn_err(flags,
 297                         "file system too large to be mounted on this system.");
 298                 return XFS_ERROR(E2BIG);
 299         }
 300
 301         if (unlikely(sbp->sb_inprogress)) {
 302                 xfs_fs_mount_cmn_err(flags, "file system busy");
 303                 return XFS_ERROR(EFSCORRUPTED);
 304         }
 305
 306         /*
 307          * Version 1 directory format has never worked on Linux.
 308          */
 309         if (unlikely(!XFS_SB_VERSION_HASDIRV2(sbp))) {
 310                 xfs_fs_mount_cmn_err(flags,
 311                         "file system using version 1 directory format");
 312                 return XFS_ERROR(ENOSYS);
 313         }
 314
 315         /*
 316          * Until this is fixed only page-sized or smaller data blocks work.
 317          */
 318         if (unlikely(sbp->sb_blocksize > PAGE_SIZE)) {
 319                 xfs_fs_mount_cmn_err(flags,
 320                         "file system with blocksize %d bytes",
 321                         sbp->sb_blocksize);
 322                 xfs_fs_mount_cmn_err(flags,
 323                         "only pagesize (%ld) or less will currently work.",
 324                         PAGE_SIZE);
 325                 return XFS_ERROR(ENOSYS);
 326         }
 327
 328         return 0;
 329 }
 330
 331 STATIC void
 332 xfs_initialize_perag_icache(
 333         xfs_perag_t     *pag)
 334 {
 335         if (!pag->pag_ici_init) {
 336                 rwlock_init(&pag->pag_ici_lock);
 337                 INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
 338                 pag->pag_ici_init = 1;
 339         }
 340 }
 341
 342 xfs_agnumber_t
 343 xfs_initialize_perag(
 344         xfs_mount_t     *mp,
 345         xfs_agnumber_t  agcount)
 346 {
 347         xfs_agnumber_t  index, max_metadata;
 348         xfs_perag_t     *pag;
 349         xfs_agino_t     agino;
 350         xfs_ino_t       ino;
 351         xfs_sb_t        *sbp = &mp->m_sb;
 352         xfs_ino_t       max_inum = XFS_MAXINUMBER_32;
 353
 354         /* Check to see if the filesystem can overflow 32 bit inodes */
 355         agino = XFS_OFFBNO_TO_AGINO(mp, sbp->sb_agblocks - 1, 0);
 356         ino = XFS_AGINO_TO_INO(mp, agcount - 1, agino);
 357
 358         /* Clear the mount flag if no inode can overflow 32 bits
 359          * on this filesystem, or if specifically requested..
 360          */
 361         if ((mp->m_flags & XFS_MOUNT_SMALL_INUMS) && ino > max_inum) {
 362                 mp->m_flags |= XFS_MOUNT_32BITINODES;
 363         } else {
 364                 mp->m_flags &= ~XFS_MOUNT_32BITINODES;
 365         }
 366
 367         /* If we can overflow then setup the ag headers accordingly */
 368         if (mp->m_flags & XFS_MOUNT_32BITINODES) {
 369                 /* Calculate how much should be reserved for inodes to
 370                  * meet the max inode percentage.
 371                  */
 372                 if (mp->m_maxicount) {
 373                         __uint64_t      icount;
 374
 375                         icount = sbp->sb_dblocks * sbp->sb_imax_pct;
 376                         do_div(icount, 100);
 377                         icount += sbp->sb_agblocks - 1;
 378                         do_div(icount, sbp->sb_agblocks);
 379                         max_metadata = icount;
 380                 } else {
 381                         max_metadata = agcount;
 382                 }
 383                 for (index = 0; index < agcount; index++) {
 384                         ino = XFS_AGINO_TO_INO(mp, index, agino);
 385                         if (ino > max_inum) {
 386                                 index++;
 387                                 break;
 388                         }
 389
 390                         /* This ag is preferred for inodes */
 391                         pag = &mp->m_perag[index];
 392                         pag->pagi_inodeok = 1;
 393                         if (index < max_metadata)
 394                                 pag->pagf_metadata = 1;
 395                         xfs_initialize_perag_icache(pag);
 396                 }
 397         } else {
 398                 /* Setup default behavior for smaller filesystems */
 399                 for (index = 0; index < agcount; index++) {
 400                         pag = &mp->m_perag[index];
 401                         pag->pagi_inodeok = 1;
 402                         xfs_initialize_perag_icache(pag);
 403                 }
 404         }
 405         return index;
 406 }
 407
 408 void
 409 xfs_sb_from_disk(
 410         xfs_sb_t        *to,
 411         xfs_dsb_t       *from)
 412 {
 413         to->sb_magicnum = be32_to_cpu(from->sb_magicnum);
 414         to->sb_blocksize = be32_to_cpu(from->sb_blocksize);
 415         to->sb_dblocks = be64_to_cpu(from->sb_dblocks);
 416         to->sb_rblocks = be64_to_cpu(from->sb_rblocks);
 417         to->sb_rextents = be64_to_cpu(from->sb_rextents);
 418         memcpy(&to->sb_uuid, &from->sb_uuid, sizeof(to->sb_uuid));
 419         to->sb_logstart = be64_to_cpu(from->sb_logstart);
 420         to->sb_rootino = be64_to_cpu(from->sb_rootino);
 421         to->sb_rbmino = be64_to_cpu(from->sb_rbmino);
 422         to->sb_rsumino = be64_to_cpu(from->sb_rsumino);
 423         to->sb_rextsize = be32_to_cpu(from->sb_rextsize);
 424         to->sb_agblocks = be32_to_cpu(from->sb_agblocks);
 425         to->sb_agcount = be32_to_cpu(from->sb_agcount);
 426         to->sb_rbmblocks = be32_to_cpu(from->sb_rbmblocks);
 427         to->sb_logblocks = be32_to_cpu(from->sb_logblocks);
 428         to->sb_versionnum = be16_to_cpu(from->sb_versionnum);
 429         to->sb_sectsize = be16_to_cpu(from->sb_sectsize);
 430         to->sb_inodesize = be16_to_cpu(from->sb_inodesize);
 431         to->sb_inopblock = be16_to_cpu(from->sb_inopblock);
 432         memcpy(&to->sb_fname, &from->sb_fname, sizeof(to->sb_fname));
 433         to->sb_blocklog = from->sb_blocklog;
 434         to->sb_sectlog = from->sb_sectlog;
 435         to->sb_inodelog = from->sb_inodelog;
 436         to->sb_inopblog = from->sb_inopblog;
 437         to->sb_agblklog = from->sb_agblklog;
 438         to->sb_rextslog = from->sb_rextslog;
 439         to->sb_inprogress = from->sb_inprogress;
 440         to->sb_imax_pct = from->sb_imax_pct;
 441         to->sb_icount = be64_to_cpu(from->sb_icount);
 442         to->sb_ifree = be64_to_cpu(from->sb_ifree);
 443         to->sb_fdblocks = be64_to_cpu(from->sb_fdblocks);
 444         to->sb_frextents = be64_to_cpu(from->sb_frextents);
 445         to->sb_uquotino = be64_to_cpu(from->sb_uquotino);
 446         to->sb_gquotino = be64_to_cpu(from->sb_gquotino);
 447         to->sb_qflags = be16_to_cpu(from->sb_qflags);
 448         to->sb_flags = from->sb_flags;
 449         to->sb_shared_vn = from->sb_shared_vn;
 450         to->sb_inoalignmt = be32_to_cpu(from->sb_inoalignmt);
 451         to->sb_unit = be32_to_cpu(from->sb_unit);
 452         to->sb_width = be32_to_cpu(from->sb_width);
 453         to->sb_dirblklog = from->sb_dirblklog;
 454         to->sb_logsectlog = from->sb_logsectlog;
 455         to->sb_logsectsize = be16_to_cpu(from->sb_logsectsize);
 456         to->sb_logsunit = be32_to_cpu(from->sb_logsunit);
 457         to->sb_features2 = be32_to_cpu(from->sb_features2);
 458 }
 459
 460 /*
 461  * Copy in core superblock to ondisk one.
 462  *
 463  * The fields argument is mask of superblock fields to copy.
 464  */
 465 void
 466 xfs_sb_to_disk(
 467         xfs_dsb_t       *to,
 468         xfs_sb_t        *from,
 469         __int64_t       fields)
 470 {
 471         xfs_caddr_t     to_ptr = (xfs_caddr_t)to;
 472         xfs_caddr_t     from_ptr = (xfs_caddr_t)from;
 473         xfs_sb_field_t  f;
 474         int             first;
 475         int             size;
 476
 477         ASSERT(fields);
 478         if (!fields)
 479                 return;
 480
 481         while (fields) {
 482                 f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
 483                 first = xfs_sb_info[f].offset;
 484                 size = xfs_sb_info[f + 1].offset - first;
 485
 486                 ASSERT(xfs_sb_info[f].type == 0 || xfs_sb_info[f].type == 1);
 487
 488                 if (size == 1 || xfs_sb_info[f].type == 1) {
 489                         memcpy(to_ptr + first, from_ptr + first, size);
 490                 } else {
 491                         switch (size) {
 492                         case 2:
 493                                 *(__be16 *)(to_ptr + first) =
 494                                         cpu_to_be16(*(__u16 *)(from_ptr + first));
 495                                 break;
 496                         case 4:
 497                                 *(__be32 *)(to_ptr + first) =
 498                                         cpu_to_be32(*(__u32 *)(from_ptr + first));
 499                                 break;
 500                         case 8:
 501                                 *(__be64 *)(to_ptr + first) =
 502                                         cpu_to_be64(*(__u64 *)(from_ptr + first));
 503                                 break;
 504                         default:
 505                                 ASSERT(0);
 506                         }
 507                 }
 508
 509                 fields &= ~(1LL << f);
 510         }
 511 }
 512
 513 /*
 514  * xfs_readsb
 515  *
 516  * Does the initial read of the superblock.
 517  */
 518 int
 519 xfs_readsb(xfs_mount_t *mp, int flags)
 520 {
 521         unsigned int    sector_size;
 522         unsigned int    extra_flags;
 523         xfs_buf_t       *bp;
 524         int             error;
 525
 526         ASSERT(mp->m_sb_bp == NULL);
 527         ASSERT(mp->m_ddev_targp != NULL);
 528
 529         /*
 530          * Allocate a (locked) buffer to hold the superblock.
 531          * This will be kept around at all times to optimize
 532          * access to the superblock.
 533          */
 534         sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
 535         extra_flags = XFS_BUF_LOCK | XFS_BUF_MANAGE | XFS_BUF_MAPPED;
 536
 537         bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
 538                                 BTOBB(sector_size), extra_flags);
 539         if (!bp || XFS_BUF_ISERROR(bp)) {
 540                 xfs_fs_mount_cmn_err(flags, "SB read failed");
 541                 error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
 542                 goto fail;
 543         }
 544         ASSERT(XFS_BUF_ISBUSY(bp));
 545         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 546
 547         /*
 548          * Initialize the mount structure from the superblock.
 549          * But first do some basic consistency checking.
 550          */
 551         xfs_sb_from_disk(&mp->m_sb, XFS_BUF_TO_SBP(bp));
 552
 553         error = xfs_mount_validate_sb(mp, &(mp->m_sb), flags);
 554         if (error) {
 555                 xfs_fs_mount_cmn_err(flags, "SB validate failed");
 556                 goto fail;
 557         }
 558
 559         /*
 560          * We must be able to do sector-sized and sector-aligned IO.
 561          */
 562         if (sector_size > mp->m_sb.sb_sectsize) {
 563                 xfs_fs_mount_cmn_err(flags,
 564                         "device supports only %u byte sectors (not %u)",
 565                         sector_size, mp->m_sb.sb_sectsize);
 566                 error = ENOSYS;
 567                 goto fail;
 568         }
 569
 570         /*
 571          * If device sector size is smaller than the superblock size,
 572          * re-read the superblock so the buffer is correctly sized.
 573          */
 574         if (sector_size < mp->m_sb.sb_sectsize) {
 575                 XFS_BUF_UNMANAGE(bp);
 576                 xfs_buf_relse(bp);
 577                 sector_size = mp->m_sb.sb_sectsize;
 578                 bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
 579                                         BTOBB(sector_size), extra_flags);
 580                 if (!bp || XFS_BUF_ISERROR(bp)) {
 581                         xfs_fs_mount_cmn_err(flags, "SB re-read failed");
 582                         error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
 583                         goto fail;
 584                 }
 585                 ASSERT(XFS_BUF_ISBUSY(bp));
 586                 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 587         }
 588
 589         /* Initialize per-cpu counters */
 590         xfs_icsb_reinit_counters(mp);
 591
 592         mp->m_sb_bp = bp;
 593         xfs_buf_relse(bp);
 594         ASSERT(XFS_BUF_VALUSEMA(bp) > 0);
 595         return 0;
 596
 597  fail:
 598         if (bp) {
 599                 XFS_BUF_UNMANAGE(bp);
 600                 xfs_buf_relse(bp);
 601         }
 602         return error;
 603 }
 604
 605
 606 /*
 607  * xfs_mount_common
 608  *
 609  * Mount initialization code establishing various mount
 610  * fields from the superblock associated with the given
 611  * mount structure
 612  */
 613 STATIC void
 614 xfs_mount_common(xfs_mount_t *mp, xfs_sb_t *sbp)
 615 {
 616         int     i;
 617
 618         mp->m_agfrotor = mp->m_agirotor = 0;
 619         spin_lock_init(&mp->m_agirotor_lock);
 620         mp->m_maxagi = mp->m_sb.sb_agcount;
 621         mp->m_blkbit_log = sbp->sb_blocklog + XFS_NBBYLOG;
 622         mp->m_blkbb_log = sbp->sb_blocklog - BBSHIFT;
 623         mp->m_sectbb_log = sbp->sb_sectlog - BBSHIFT;
 624         mp->m_agno_log = xfs_highbit32(sbp->sb_agcount - 1) + 1;
 625         mp->m_agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
 626         mp->m_litino = sbp->sb_inodesize -
 627                 ((uint)sizeof(xfs_dinode_core_t) + (uint)sizeof(xfs_agino_t));
 628         mp->m_blockmask = sbp->sb_blocksize - 1;
 629         mp->m_blockwsize = sbp->sb_blocksize >> XFS_WORDLOG;
 630         mp->m_blockwmask = mp->m_blockwsize - 1;
 631         INIT_LIST_HEAD(&mp->m_del_inodes);
 632
 633         /*
 634          * Setup for attributes, in case they get created.
 635          * This value is for inodes getting attributes for the first time,
 636          * the per-inode value is for old attribute values.
 637          */
 638         ASSERT(sbp->sb_inodesize >= 256 && sbp->sb_inodesize <= 2048);
 639         switch (sbp->sb_inodesize) {
 640         case 256:
 641                 mp->m_attroffset = XFS_LITINO(mp) -
 642                                    XFS_BMDR_SPACE_CALC(MINABTPTRS);
 643                 break;
 644         case 512:
 645         case 1024:
 646         case 2048:
 647                 mp->m_attroffset = XFS_BMDR_SPACE_CALC(6 * MINABTPTRS);
 648                 break;
 649         default:
 650                 ASSERT(0);
 651         }
 652         ASSERT(mp->m_attroffset < XFS_LITINO(mp));
 653
 654         for (i = 0; i < 2; i++) {
 655                 mp->m_alloc_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 656                         xfs_alloc, i == 0);
 657                 mp->m_alloc_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 658                         xfs_alloc, i == 0);
 659         }
 660         for (i = 0; i < 2; i++) {
 661                 mp->m_bmap_dmxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 662                         xfs_bmbt, i == 0);
 663                 mp->m_bmap_dmnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 664                         xfs_bmbt, i == 0);
 665         }
 666         for (i = 0; i < 2; i++) {
 667                 mp->m_inobt_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 668                         xfs_inobt, i == 0);
 669                 mp->m_inobt_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 670                         xfs_inobt, i == 0);
 671         }
 672
 673         mp->m_bsize = XFS_FSB_TO_BB(mp, 1);
 674         mp->m_ialloc_inos = (int)MAX((__uint16_t)XFS_INODES_PER_CHUNK,
 675                                         sbp->sb_inopblock);
 676         mp->m_ialloc_blks = mp->m_ialloc_inos >> sbp->sb_inopblog;
 677 }
 678
 679 /*
 680  * xfs_initialize_perag_data
 681  *
 682  * Read in each per-ag structure so we can count up the number of
 683  * allocated inodes, free inodes and used filesystem blocks as this
 684  * information is no longer persistent in the superblock. Once we have
 685  * this information, write it into the in-core superblock structure.
 686  */
 687 STATIC int
 688 xfs_initialize_perag_data(xfs_mount_t *mp, xfs_agnumber_t agcount)
 689 {
 690         xfs_agnumber_t  index;
 691         xfs_perag_t     *pag;
 692         xfs_sb_t        *sbp = &mp->m_sb;
 693         uint64_t        ifree = 0;
 694         uint64_t        ialloc = 0;
 695         uint64_t        bfree = 0;
 696         uint64_t        bfreelst = 0;
 697         uint64_t        btree = 0;
 698         int             error;
 699
 700         for (index = 0; index < agcount; index++) {
 701                 /*
 702                  * read the agf, then the agi. This gets us
 703                  * all the inforamtion we need and populates the
 704                  * per-ag structures for us.
 705                  */
 706                 error = xfs_alloc_pagf_init(mp, NULL, index, 0);
 707                 if (error)
 708                         return error;
 709
 710                 error = xfs_ialloc_pagi_init(mp, NULL, index);
 711                 if (error)
 712                         return error;
 713                 pag = &mp->m_perag[index];
 714                 ifree += pag->pagi_freecount;
 715                 ialloc += pag->pagi_count;
 716                 bfree += pag->pagf_freeblks;
 717                 bfreelst += pag->pagf_flcount;
 718                 btree += pag->pagf_btreeblks;
 719         }
 720         /*
 721          * Overwrite incore superblock counters with just-read data
 722          */
 723         spin_lock(&mp->m_sb_lock);
 724         sbp->sb_ifree = ifree;
 725         sbp->sb_icount = ialloc;
 726         sbp->sb_fdblocks = bfree + bfreelst + btree;
 727         spin_unlock(&mp->m_sb_lock);
 728
 729         /* Fixup the per-cpu counters as well. */
 730         xfs_icsb_reinit_counters(mp);
 731
 732         return 0;
 733 }
 734
 735 /*
 736  * Update alignment values based on mount options and sb values
 737  */
 738 STATIC int
 739 xfs_update_alignment(xfs_mount_t *mp, int mfsi_flags, __uint64_t *update_flags)
 740 {
 741         xfs_sb_t        *sbp = &(mp->m_sb);
 742
 743         if (mp->m_dalign && !(mfsi_flags & XFS_MFSI_SECOND)) {
 744                 /*
 745                  * If stripe unit and stripe width are not multiples
 746                  * of the fs blocksize turn off alignment.
 747                  */
 748                 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
 749                     (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
 750                         if (mp->m_flags & XFS_MOUNT_RETERR) {
 751                                 cmn_err(CE_WARN,
 752                                         "XFS: alignment check 1 failed");
 753                                 return XFS_ERROR(EINVAL);
 754                         }
 755                         mp->m_dalign = mp->m_swidth = 0;
 756                 } else {
 757                         /*
 758                          * Convert the stripe unit and width to FSBs.
 759                          */
 760                         mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
 761                         if (mp->m_dalign && (sbp->sb_agblocks % mp->m_dalign)) {
 762                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
 763                                         return XFS_ERROR(EINVAL);
 764                                 }
 765                                 xfs_fs_cmn_err(CE_WARN, mp,
 766 "stripe alignment turned off: sunit(%d)/swidth(%d) incompatible with agsize(%d)",
 767                                         mp->m_dalign, mp->m_swidth,
 768                                         sbp->sb_agblocks);
 769
 770                                 mp->m_dalign = 0;
 771                                 mp->m_swidth = 0;
 772                         } else if (mp->m_dalign) {
 773                                 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
 774                         } else {
 775                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
 776                                         xfs_fs_cmn_err(CE_WARN, mp,
 777 "stripe alignment turned off: sunit(%d) less than bsize(%d)",
 778                                                 mp->m_dalign,
 779                                                 mp->m_blockmask +1);
 780                                         return XFS_ERROR(EINVAL);
 781                                 }
 782                                 mp->m_swidth = 0;
 783                         }
 784                 }
 785
 786                 /*
 787                  * Update superblock with new values
 788                  * and log changes
 789                  */
 790                 if (XFS_SB_VERSION_HASDALIGN(sbp)) {
 791                         if (sbp->sb_unit != mp->m_dalign) {
 792                                 sbp->sb_unit = mp->m_dalign;
 793                                 *update_flags |= XFS_SB_UNIT;
 794                         }
 795                         if (sbp->sb_width != mp->m_swidth) {
 796                                 sbp->sb_width = mp->m_swidth;
 797                                 *update_flags |= XFS_SB_WIDTH;
 798                         }
 799                 }
 800         } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
 801                     XFS_SB_VERSION_HASDALIGN(&mp->m_sb)) {
 802                         mp->m_dalign = sbp->sb_unit;
 803                         mp->m_swidth = sbp->sb_width;
 804         }
 805
 806         return 0;
 807 }
 808
 809 /*
 810  * Set the maximum inode count for this filesystem
 811  */
 812 STATIC void
 813 xfs_set_maxicount(xfs_mount_t *mp)
 814 {
 815         xfs_sb_t        *sbp = &(mp->m_sb);
 816         __uint64_t      icount;
 817
 818         if (sbp->sb_imax_pct) {
 819                 /*
 820                  * Make sure the maximum inode count is a multiple
 821                  * of the units we allocate inodes in.
 822                  */
 823                 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
 824                 do_div(icount, 100);
 825                 do_div(icount, mp->m_ialloc_blks);
 826                 mp->m_maxicount = (icount * mp->m_ialloc_blks)  <<
 827                                    sbp->sb_inopblog;
 828         } else {
 829                 mp->m_maxicount = 0;
 830         }
 831 }
 832
 833 /*
 834  * Set the default minimum read and write sizes unless
 835  * already specified in a mount option.
 836  * We use smaller I/O sizes when the file system
 837  * is being used for NFS service (wsync mount option).
 838  */
 839 STATIC void
 840 xfs_set_rw_sizes(xfs_mount_t *mp)
 841 {
 842         xfs_sb_t        *sbp = &(mp->m_sb);
 843         int             readio_log, writeio_log;
 844
 845         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
 846                 if (mp->m_flags & XFS_MOUNT_WSYNC) {
 847                         readio_log = XFS_WSYNC_READIO_LOG;
 848                         writeio_log = XFS_WSYNC_WRITEIO_LOG;
 849                 } else {
 850                         readio_log = XFS_READIO_LOG_LARGE;
 851                         writeio_log = XFS_WRITEIO_LOG_LARGE;
 852                 }
 853         } else {
 854                 readio_log = mp->m_readio_log;
 855                 writeio_log = mp->m_writeio_log;
 856         }
 857
 858         if (sbp->sb_blocklog > readio_log) {
 859                 mp->m_readio_log = sbp->sb_blocklog;
 860         } else {
 861                 mp->m_readio_log = readio_log;
 862         }
 863         mp->m_readio_blocks = 1 << (mp->m_readio_log - sbp->sb_blocklog);
 864         if (sbp->sb_blocklog > writeio_log) {
 865                 mp->m_writeio_log = sbp->sb_blocklog;
 866         } else {
 867                 mp->m_writeio_log = writeio_log;
 868         }
 869         mp->m_writeio_blocks = 1 << (mp->m_writeio_log - sbp->sb_blocklog);
 870 }
 871
 872 /*
 873  * Set whether we're using inode alignment.
 874  */
 875 STATIC void
 876 xfs_set_inoalignment(xfs_mount_t *mp)
 877 {
 878         if (XFS_SB_VERSION_HASALIGN(&mp->m_sb) &&
 879             mp->m_sb.sb_inoalignmt >=
 880             XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size))
 881                 mp->m_inoalign_mask = mp->m_sb.sb_inoalignmt - 1;
 882         else
 883                 mp->m_inoalign_mask = 0;
 884         /*
 885          * If we are using stripe alignment, check whether
 886          * the stripe unit is a multiple of the inode alignment
 887          */
 888         if (mp->m_dalign && mp->m_inoalign_mask &&
 889             !(mp->m_dalign & mp->m_inoalign_mask))
 890                 mp->m_sinoalign = mp->m_dalign;
 891         else
 892                 mp->m_sinoalign = 0;
 893 }
 894
 895 /*
 896  * Check that the data (and log if separate) are an ok size.
 897  */
 898 STATIC int
 899 xfs_check_sizes(xfs_mount_t *mp, int mfsi_flags)
 900 {
 901         xfs_buf_t       *bp;
 902         xfs_daddr_t     d;
 903         int             error;
 904
 905         d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
 906         if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
 907                 cmn_err(CE_WARN, "XFS: size check 1 failed");
 908                 return XFS_ERROR(E2BIG);
 909         }
 910         error = xfs_read_buf(mp, mp->m_ddev_targp,
 911                              d - XFS_FSS_TO_BB(mp, 1),
 912                              XFS_FSS_TO_BB(mp, 1), 0, &bp);
 913         if (!error) {
 914                 xfs_buf_relse(bp);
 915         } else {
 916                 cmn_err(CE_WARN, "XFS: size check 2 failed");
 917                 if (error == ENOSPC)
 918                         error = XFS_ERROR(E2BIG);
 919                 return error;
 920         }
 921
 922         if (((mfsi_flags & XFS_MFSI_CLIENT) == 0) &&
 923             mp->m_logdev_targp != mp->m_ddev_targp) {
 924                 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
 925                 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
 926                         cmn_err(CE_WARN, "XFS: size check 3 failed");
 927                         return XFS_ERROR(E2BIG);
 928                 }
 929                 error = xfs_read_buf(mp, mp->m_logdev_targp,
 930                                      d - XFS_FSB_TO_BB(mp, 1),
 931                                      XFS_FSB_TO_BB(mp, 1), 0, &bp);
 932                 if (!error) {
 933                         xfs_buf_relse(bp);
 934                 } else {
 935                         cmn_err(CE_WARN, "XFS: size check 3 failed");
 936                         if (error == ENOSPC)
 937                                 error = XFS_ERROR(E2BIG);
 938                         return error;
 939                 }
 940         }
 941         return 0;
 942 }
 943
 944 /*
 945  * xfs_mountfs
 946  *
 947  * This function does the following on an initial mount of a file system:
 948  *      - reads the superblock from disk and init the mount struct
 949  *      - if we're a 32-bit kernel, do a size check on the superblock
 950  *              so we don't mount terabyte filesystems
 951  *      - init mount struct realtime fields
 952  *      - allocate inode hash table for fs
 953  *      - init directory manager
 954  *      - perform recovery and init the log manager
 955  */
 956 int
 957 xfs_mountfs(
 958         xfs_mount_t     *mp,
 959         int             mfsi_flags)
 960 {
 961         xfs_sb_t        *sbp = &(mp->m_sb);
 962         xfs_inode_t     *rip;
 963         bhv_vnode_t     *rvp = NULL;
 964         __uint64_t      resblks;
 965         __int64_t       update_flags = 0LL;
 966         uint            quotamount, quotaflags;
 967         int             agno;
 968         int             uuid_mounted = 0;
 969         int             error = 0;
 970
 971         if (mp->m_sb_bp == NULL) {
 972                 error = xfs_readsb(mp, mfsi_flags);
 973                 if (error)
 974                         return error;
 975         }
 976         xfs_mount_common(mp, sbp);
 977
 978         /*
 979          * Check if sb_agblocks is aligned at stripe boundary
 980          * If sb_agblocks is NOT aligned turn off m_dalign since
 981          * allocator alignment is within an ag, therefore ag has
 982          * to be aligned at stripe boundary.
 983          */
 984         error = xfs_update_alignment(mp, mfsi_flags, &update_flags);
 985         if (error)
 986                 goto error1;
 987
 988         xfs_alloc_compute_maxlevels(mp);
 989         xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
 990         xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
 991         xfs_ialloc_compute_maxlevels(mp);
 992
 993         xfs_set_maxicount(mp);
 994
 995         mp->m_maxioffset = xfs_max_file_offset(sbp->sb_blocklog);
 996
 997         /*
 998          * XFS uses the uuid from the superblock as the unique
 999          * identifier for fsid.  We can not use the uuid from the volume
1000          * since a single partition filesystem is identical to a single
1001          * partition volume/filesystem.
1002          */
1003         if ((mfsi_flags & XFS_MFSI_SECOND) == 0 &&
1004             (mp->m_flags & XFS_MOUNT_NOUUID) == 0) {
1005                 if (xfs_uuid_mount(mp)) {
1006                         error = XFS_ERROR(EINVAL);
1007                         goto error1;
1008                 }
1009                 uuid_mounted=1;
1010         }
1011
1012         /*
1013          * Set the minimum read and write sizes
1014          */
1015         xfs_set_rw_sizes(mp);
1016
1017         /*
1018          * Set the inode cluster size.
1019          * This may still be overridden by the file system
1020          * block size if it is larger than the chosen cluster size.
1021          */
1022         mp->m_inode_cluster_size = XFS_INODE_BIG_CLUSTER_SIZE;
1023
1024         /*
1025          * Set inode alignment fields
1026          */
1027         xfs_set_inoalignment(mp);
1028
1029         /*
1030          * Check that the data (and log if separate) are an ok size.
1031          */
1032         error = xfs_check_sizes(mp, mfsi_flags);
1033         if (error)
1034                 goto error1;
1035
1036         /*
1037          * Initialize realtime fields in the mount structure
1038          */
1039         error = xfs_rtmount_init(mp);
1040         if (error) {
1041                 cmn_err(CE_WARN, "XFS: RT mount failed");
1042                 goto error1;
1043         }
1044
1045         /*
1046          * For client case we are done now
1047          */
1048         if (mfsi_flags & XFS_MFSI_CLIENT) {
1049                 return 0;
1050         }
1051
1052         /*
1053          *  Copies the low order bits of the timestamp and the randomly
1054          *  set "sequence" number out of a UUID.
1055          */
1056         uuid_getnodeuniq(&sbp->sb_uuid, mp->m_fixedfsid);
1057
1058         mp->m_dmevmask = 0;     /* not persistent; set after each mount */
1059
1060         xfs_dir_mount(mp);
1061
1062         /*
1063          * Initialize the attribute manager's entries.
1064          */
1065         mp->m_attr_magicpct = (mp->m_sb.sb_blocksize * 37) / 100;
1066
1067         /*
1068          * Initialize the precomputed transaction reservations values.
1069          */
1070         xfs_trans_init(mp);
1071
1072         /*
1073          * Allocate and initialize the per-ag data.
1074          */
1075         init_rwsem(&mp->m_peraglock);
1076         mp->m_perag =
1077                 kmem_zalloc(sbp->sb_agcount * sizeof(xfs_perag_t), KM_SLEEP);
1078
1079         mp->m_maxagi = xfs_initialize_perag(mp, sbp->sb_agcount);
1080
1081         /*
1082          * log's mount-time initialization. Perform 1st part recovery if needed
1083          */
1084         if (likely(sbp->sb_logblocks > 0)) {    /* check for volume case */
1085                 error = xfs_log_mount(mp, mp->m_logdev_targp,
1086                                       XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
1087                                       XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
1088                 if (error) {
1089                         cmn_err(CE_WARN, "XFS: log mount failed");
1090                         goto error2;
1091                 }
1092         } else {        /* No log has been defined */
1093                 cmn_err(CE_WARN, "XFS: no log defined");
1094                 XFS_ERROR_REPORT("xfs_mountfs_int(1)", XFS_ERRLEVEL_LOW, mp);
1095                 error = XFS_ERROR(EFSCORRUPTED);
1096                 goto error2;
1097         }
1098
1099         /*
1100          * Now the log is mounted, we know if it was an unclean shutdown or
1101          * not. If it was, with the first phase of recovery has completed, we
1102          * have consistent AG blocks on disk. We have not recovered EFIs yet,
1103          * but they are recovered transactionally in the second recovery phase
1104          * later.
1105          *
1106          * Hence we can safely re-initialise incore superblock counters from
1107          * the per-ag data. These may not be correct if the filesystem was not
1108          * cleanly unmounted, so we need to wait for recovery to finish before
1109          * doing this.
1110          *
1111          * If the filesystem was cleanly unmounted, then we can trust the
1112          * values in the superblock to be correct and we don't need to do
1113          * anything here.
1114          *
1115          * If we are currently making the filesystem, the initialisation will
1116          * fail as the perag data is in an undefined state.
1117          */
1118
1119         if (xfs_sb_version_haslazysbcount(&mp->m_sb) &&
1120             !XFS_LAST_UNMOUNT_WAS_CLEAN(mp) &&
1121              !mp->m_sb.sb_inprogress) {
1122                 error = xfs_initialize_perag_data(mp, sbp->sb_agcount);
1123                 if (error) {
1124                         goto error2;
1125                 }
1126         }
1127         /*
1128          * Get and sanity-check the root inode.
1129          * Save the pointer to it in the mount structure.
1130          */
1131         error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip, 0);
1132         if (error) {
1133                 cmn_err(CE_WARN, "XFS: failed to read root inode");
1134                 goto error3;
1135         }
1136
1137         ASSERT(rip != NULL);
1138         rvp = XFS_ITOV(rip);
1139
1140         if (unlikely((rip->i_d.di_mode & S_IFMT) != S_IFDIR)) {
1141                 cmn_err(CE_WARN, "XFS: corrupted root inode");
1142                 cmn_err(CE_WARN, "Device %s - root %llu is not a directory",
1143                         XFS_BUFTARG_NAME(mp->m_ddev_targp),
1144                         (unsigned long long)rip->i_ino);
1145                 xfs_iunlock(rip, XFS_ILOCK_EXCL);
1146                 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
1147                                  mp);
1148                 error = XFS_ERROR(EFSCORRUPTED);
1149                 goto error4;
1150         }
1151         mp->m_rootip = rip;     /* save it */
1152
1153         xfs_iunlock(rip, XFS_ILOCK_EXCL);
1154
1155         /*
1156          * Initialize realtime inode pointers in the mount structure
1157          */
1158         error = xfs_rtmount_inodes(mp);
1159         if (error) {
1160                 /*
1161                  * Free up the root inode.
1162                  */
1163                 cmn_err(CE_WARN, "XFS: failed to read RT inodes");
1164                 goto error4;
1165         }
1166
1167         /*
1168          * If fs is not mounted readonly, then update the superblock
1169          * unit and width changes.
1170          */
1171         if (update_flags && !(mp->m_flags & XFS_MOUNT_RDONLY))
1172                 xfs_mount_log_sbunit(mp, update_flags);
1173
1174         /*
1175          * Initialise the XFS quota management subsystem for this mount
1176          */
1177         error = XFS_QM_INIT(mp, &quotamount, &quotaflags);
1178         if (error)
1179                 goto error4;
1180
1181         /*
1182          * Finish recovering the file system.  This part needed to be
1183          * delayed until after the root and real-time bitmap inodes
1184          * were consistently read in.
1185          */
1186         error = xfs_log_mount_finish(mp, mfsi_flags);
1187         if (error) {
1188                 cmn_err(CE_WARN, "XFS: log mount finish failed");
1189                 goto error4;
1190         }
1191
1192         /*
1193          * Complete the quota initialisation, post-log-replay component.
1194          */
1195         error = XFS_QM_MOUNT(mp, quotamount, quotaflags, mfsi_flags);
1196         if (error)
1197                 goto error4;
1198
1199         /*
1200          * Now we are mounted, reserve a small amount of unused space for
1201          * privileged transactions. This is needed so that transaction
1202          * space required for critical operations can dip into this pool
1203          * when at ENOSPC. This is needed for operations like create with
1204          * attr, unwritten extent conversion at ENOSPC, etc. Data allocations
1205          * are not allowed to use this reserved space.
1206          *
1207          * We default to 5% or 1024 fsbs of space reserved, whichever is smaller.
1208          * This may drive us straight to ENOSPC on mount, but that implies
1209          * we were already there on the last unmount.
1210          */
1211         resblks = mp->m_sb.sb_dblocks;
1212         do_div(resblks, 20);
1213         resblks = min_t(__uint64_t, resblks, 1024);
1214         xfs_reserve_blocks(mp, &resblks, NULL);
1215
1216         return 0;
1217
1218  error4:
1219         /*
1220          * Free up the root inode.
1221          */
1222         VN_RELE(rvp);
1223  error3:
1224         xfs_log_unmount_dealloc(mp);
1225  error2:
1226         for (agno = 0; agno < sbp->sb_agcount; agno++)
1227                 if (mp->m_perag[agno].pagb_list)
1228                         kmem_free(mp->m_perag[agno].pagb_list,
1229                           sizeof(xfs_perag_busy_t) * XFS_PAGB_NUM_SLOTS);
1230         kmem_free(mp->m_perag, sbp->sb_agcount * sizeof(xfs_perag_t));
1231         mp->m_perag = NULL;
1232         /* FALLTHROUGH */
1233  error1:
1234         if (uuid_mounted)
1235                 xfs_uuid_unmount(mp);
1236         xfs_freesb(mp);
1237         return error;
1238 }
1239
1240 /*
1241  * xfs_unmountfs
1242  *
1243  * This flushes out the inodes,dquots and the superblock, unmounts the
1244  * log and makes sure that incore structures are freed.
1245  */
1246 int
1247 xfs_unmountfs(xfs_mount_t *mp, struct cred *cr)
1248 {
1249         __uint64_t      resblks;
1250
1251         /*
1252          * We can potentially deadlock here if we have an inode cluster
1253          * that has been freed has it's buffer still pinned in memory because
1254          * the transaction is still sitting in a iclog. The stale inodes
1255          * on that buffer will have their flush locks held until the
1256          * transaction hits the disk and the callbacks run. the inode
1257          * flush takes the flush lock unconditionally and with nothing to
1258          * push out the iclog we will never get that unlocked. hence we
1259          * need to force the log first.
1260          */
1261         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1262         xfs_iflush_all(mp);
1263
1264         XFS_QM_DQPURGEALL(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_UMOUNTING);
1265
1266         /*
1267          * Flush out the log synchronously so that we know for sure
1268          * that nothing is pinned.  This is important because bflush()
1269          * will skip pinned buffers.
1270          */
1271         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1272
1273         xfs_binval(mp->m_ddev_targp);
1274         if (mp->m_rtdev_targp) {
1275                 xfs_binval(mp->m_rtdev_targp);
1276         }
1277
1278         /*
1279          * Unreserve any blocks we have so that when we unmount we don't account
1280          * the reserved free space as used. This is really only necessary for
1281          * lazy superblock counting because it trusts the incore superblock
1282          * counters to be aboslutely correct on clean unmount.
1283          *
1284          * We don't bother correcting this elsewhere for lazy superblock
1285          * counting because on mount of an unclean filesystem we reconstruct the
1286          * correct counter value and this is irrelevant.
1287          *
1288          * For non-lazy counter filesystems, this doesn't matter at all because
1289          * we only every apply deltas to the superblock and hence the incore
1290          * value does not matter....
1291          */
1292         resblks = 0;
1293         xfs_reserve_blocks(mp, &resblks, NULL);
1294
1295         xfs_log_sbcount(mp, 1);
1296         xfs_unmountfs_writesb(mp);
1297         xfs_unmountfs_wait(mp);                 /* wait for async bufs */
1298         xfs_log_unmount(mp);                    /* Done! No more fs ops. */
1299
1300         xfs_freesb(mp);
1301
1302         /*
1303          * All inodes from this mount point should be freed.
1304          */
1305         ASSERT(mp->m_inodes == NULL);
1306
1307         xfs_unmountfs_close(mp, cr);
1308         if ((mp->m_flags & XFS_MOUNT_NOUUID) == 0)
1309                 xfs_uuid_unmount(mp);
1310
1311 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1312         xfs_errortag_clearall(mp, 0);
1313 #endif
1314         xfs_mount_free(mp);
1315         return 0;
1316 }
1317
1318 void
1319 xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
1320 {
1321         if (mp->m_logdev_targp && mp->m_logdev_targp != mp->m_ddev_targp)
1322                 xfs_free_buftarg(mp->m_logdev_targp, 1);
1323         if (mp->m_rtdev_targp)
1324                 xfs_free_buftarg(mp->m_rtdev_targp, 1);
1325         xfs_free_buftarg(mp->m_ddev_targp, 0);
1326 }
1327
1328 STATIC void
1329 xfs_unmountfs_wait(xfs_mount_t *mp)
1330 {
1331         if (mp->m_logdev_targp != mp->m_ddev_targp)
1332                 xfs_wait_buftarg(mp->m_logdev_targp);
1333         if (mp->m_rtdev_targp)
1334                 xfs_wait_buftarg(mp->m_rtdev_targp);
1335         xfs_wait_buftarg(mp->m_ddev_targp);
1336 }
1337
1338 int
1339 xfs_fs_writable(xfs_mount_t *mp)
1340 {
1341         return !(xfs_test_for_freeze(mp) || XFS_FORCED_SHUTDOWN(mp) ||
1342                 (mp->m_flags & XFS_MOUNT_RDONLY));
1343 }
1344
1345 /*
1346  * xfs_log_sbcount
1347  *
1348  * Called either periodically to keep the on disk superblock values
1349  * roughly up to date or from unmount to make sure the values are
1350  * correct on a clean unmount.
1351  *
1352  * Note this code can be called during the process of freezing, so
1353  * we may need to use the transaction allocator which does not not
1354  * block when the transaction subsystem is in its frozen state.
1355  */
1356 int
1357 xfs_log_sbcount(
1358         xfs_mount_t     *mp,
1359         uint            sync)
1360 {
1361         xfs_trans_t     *tp;
1362         int             error;
1363
1364         if (!xfs_fs_writable(mp))
1365                 return 0;
1366
1367         xfs_icsb_sync_counters(mp);
1368
1369         /*
1370          * we don't need to do this if we are updating the superblock
1371          * counters on every modification.
1372          */
1373         if (!xfs_sb_version_haslazysbcount(&mp->m_sb))
1374                 return 0;
1375
1376         tp = _xfs_trans_alloc(mp, XFS_TRANS_SB_COUNT);
1377         error = xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1378                                         XFS_DEFAULT_LOG_COUNT);
1379         if (error) {
1380                 xfs_trans_cancel(tp, 0);
1381                 return error;
1382         }
1383
1384         xfs_mod_sb(tp, XFS_SB_IFREE | XFS_SB_ICOUNT | XFS_SB_FDBLOCKS);
1385         if (sync)
1386                 xfs_trans_set_sync(tp);
1387         xfs_trans_commit(tp, 0);
1388
1389         return 0;
1390 }
1391
1392 STATIC void
1393 xfs_mark_shared_ro(
1394         xfs_mount_t     *mp,
1395         xfs_buf_t       *bp)
1396 {
1397         xfs_dsb_t       *sb = XFS_BUF_TO_SBP(bp);
1398         __uint16_t      version;
1399
1400         if (!(sb->sb_flags & XFS_SBF_READONLY))
1401                 sb->sb_flags |= XFS_SBF_READONLY;
1402
1403         version = be16_to_cpu(sb->sb_versionnum);
1404         if ((version & XFS_SB_VERSION_NUMBITS) != XFS_SB_VERSION_4 ||
1405             !(version & XFS_SB_VERSION_SHAREDBIT))
1406                 version |= XFS_SB_VERSION_SHAREDBIT;
1407         sb->sb_versionnum = cpu_to_be16(version);
1408 }
1409
1410 int
1411 xfs_unmountfs_writesb(xfs_mount_t *mp)
1412 {
1413         xfs_buf_t       *sbp;
1414         int             error = 0;
1415
1416         /*
1417          * skip superblock write if fs is read-only, or
1418          * if we are doing a forced umount.
1419          */
1420         if (!((mp->m_flags & XFS_MOUNT_RDONLY) ||
1421                 XFS_FORCED_SHUTDOWN(mp))) {
1422
1423                 sbp = xfs_getsb(mp, 0);
1424
1425                 /*
1426                  * mark shared-readonly if desired
1427                  */
1428                 if (mp->m_mk_sharedro)
1429                         xfs_mark_shared_ro(mp, sbp);
1430
1431                 XFS_BUF_UNDONE(sbp);
1432                 XFS_BUF_UNREAD(sbp);
1433                 XFS_BUF_UNDELAYWRITE(sbp);
1434                 XFS_BUF_WRITE(sbp);
1435                 XFS_BUF_UNASYNC(sbp);
1436                 ASSERT(XFS_BUF_TARGET(sbp) == mp->m_ddev_targp);
1437                 xfsbdstrat(mp, sbp);
1438                 /* Nevermind errors we might get here. */
1439                 error = xfs_iowait(sbp);
1440                 if (error)
1441                         xfs_ioerror_alert("xfs_unmountfs_writesb",
1442                                           mp, sbp, XFS_BUF_ADDR(sbp));
1443                 if (error && mp->m_mk_sharedro)
1444                         xfs_fs_cmn_err(CE_ALERT, mp, "Superblock write error detected while unmounting.  Filesystem may not be marked shared readonly");
1445                 xfs_buf_relse(sbp);
1446         }
1447         return error;
1448 }
1449
1450 /*
1451  * xfs_mod_sb() can be used to copy arbitrary changes to the
1452  * in-core superblock into the superblock buffer to be logged.
1453  * It does not provide the higher level of locking that is
1454  * needed to protect the in-core superblock from concurrent
1455  * access.
1456  */
1457 void
1458 xfs_mod_sb(xfs_trans_t *tp, __int64_t fields)
1459 {
1460         xfs_buf_t       *bp;
1461         int             first;
1462         int             last;
1463         xfs_mount_t     *mp;
1464         xfs_sb_field_t  f;
1465
1466         ASSERT(fields);
1467         if (!fields)
1468                 return;
1469         mp = tp->t_mountp;
1470         bp = xfs_trans_getsb(tp, mp, 0);
1471         first = sizeof(xfs_sb_t);
1472         last = 0;
1473
1474         /* translate/copy */
1475
1476         xfs_sb_to_disk(XFS_BUF_TO_SBP(bp), &mp->m_sb, fields);
1477
1478         /* find modified range */
1479
1480         f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
1481         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1482         first = xfs_sb_info[f].offset;
1483
1484         f = (xfs_sb_field_t)xfs_highbit64((__uint64_t)fields);
1485         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1486         last = xfs_sb_info[f + 1].offset - 1;
1487
1488         xfs_trans_log_buf(tp, bp, first, last);
1489 }
1490
1491
1492 /*
1493  * xfs_mod_incore_sb_unlocked() is a utility routine common used to apply
1494  * a delta to a specified field in the in-core superblock.  Simply
1495  * switch on the field indicated and apply the delta to that field.
1496  * Fields are not allowed to dip below zero, so if the delta would
1497  * do this do not apply it and return EINVAL.
1498  *
1499  * The m_sb_lock must be held when this routine is called.
1500  */
1501 int
1502 xfs_mod_incore_sb_unlocked(
1503         xfs_mount_t     *mp,
1504         xfs_sb_field_t  field,
1505         int64_t         delta,
1506         int             rsvd)
1507 {
1508         int             scounter;       /* short counter for 32 bit fields */
1509         long long       lcounter;       /* long counter for 64 bit fields */
1510         long long       res_used, rem;
1511
1512         /*
1513          * With the in-core superblock spin lock held, switch
1514          * on the indicated field.  Apply the delta to the
1515          * proper field.  If the fields value would dip below
1516          * 0, then do not apply the delta and return EINVAL.
1517          */
1518         switch (field) {
1519         case XFS_SBS_ICOUNT:
1520                 lcounter = (long long)mp->m_sb.sb_icount;
1521                 lcounter += delta;
1522                 if (lcounter < 0) {
1523                         ASSERT(0);
1524                         return XFS_ERROR(EINVAL);
1525                 }
1526                 mp->m_sb.sb_icount = lcounter;
1527                 return 0;
1528         case XFS_SBS_IFREE:
1529                 lcounter = (long long)mp->m_sb.sb_ifree;
1530                 lcounter += delta;
1531                 if (lcounter < 0) {
1532                         ASSERT(0);
1533                         return XFS_ERROR(EINVAL);
1534                 }
1535                 mp->m_sb.sb_ifree = lcounter;
1536                 return 0;
1537         case XFS_SBS_FDBLOCKS:
1538                 lcounter = (long long)
1539                         mp->m_sb.sb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
1540                 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1541
1542                 if (delta > 0) {                /* Putting blocks back */
1543                         if (res_used > delta) {
1544                                 mp->m_resblks_avail += delta;
1545                         } else {
1546                                 rem = delta - res_used;
1547                                 mp->m_resblks_avail = mp->m_resblks;
1548                                 lcounter += rem;
1549                         }
1550                 } else {                                /* Taking blocks away */
1551
1552                         lcounter += delta;
1553
1554                 /*
1555                  * If were out of blocks, use any available reserved blocks if
1556                  * were allowed to.
1557                  */
1558
1559                         if (lcounter < 0) {
1560                                 if (rsvd) {
1561                                         lcounter = (long long)mp->m_resblks_avail + delta;
1562                                         if (lcounter < 0) {
1563                                                 return XFS_ERROR(ENOSPC);
1564                                         }
1565                                         mp->m_resblks_avail = lcounter;
1566                                         return 0;
1567                                 } else {        /* not reserved */
1568                                         return XFS_ERROR(ENOSPC);
1569                                 }
1570                         }
1571                 }
1572
1573                 mp->m_sb.sb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
1574                 return 0;
1575         case XFS_SBS_FREXTENTS:
1576                 lcounter = (long long)mp->m_sb.sb_frextents;
1577                 lcounter += delta;
1578                 if (lcounter < 0) {
1579                         return XFS_ERROR(ENOSPC);
1580                 }
1581                 mp->m_sb.sb_frextents = lcounter;
1582                 return 0;
1583         case XFS_SBS_DBLOCKS:
1584                 lcounter = (long long)mp->m_sb.sb_dblocks;
1585                 lcounter += delta;
1586                 if (lcounter < 0) {
1587                         ASSERT(0);
1588                         return XFS_ERROR(EINVAL);
1589                 }
1590                 mp->m_sb.sb_dblocks = lcounter;
1591                 return 0;
1592         case XFS_SBS_AGCOUNT:
1593                 scounter = mp->m_sb.sb_agcount;
1594                 scounter += delta;
1595                 if (scounter < 0) {
1596                         ASSERT(0);
1597                         return XFS_ERROR(EINVAL);
1598                 }
1599                 mp->m_sb.sb_agcount = scounter;
1600                 return 0;
1601         case XFS_SBS_IMAX_PCT:
1602                 scounter = mp->m_sb.sb_imax_pct;
1603                 scounter += delta;
1604                 if (scounter < 0) {
1605                         ASSERT(0);
1606                         return XFS_ERROR(EINVAL);
1607                 }
1608                 mp->m_sb.sb_imax_pct = scounter;
1609                 return 0;
1610         case XFS_SBS_REXTSIZE:
1611                 scounter = mp->m_sb.sb_rextsize;
1612                 scounter += delta;
1613                 if (scounter < 0) {
1614                         ASSERT(0);
1615                         return XFS_ERROR(EINVAL);
1616                 }
1617                 mp->m_sb.sb_rextsize = scounter;
1618                 return 0;
1619         case XFS_SBS_RBMBLOCKS:
1620                 scounter = mp->m_sb.sb_rbmblocks;
1621                 scounter += delta;
1622                 if (scounter < 0) {
1623                         ASSERT(0);
1624                         return XFS_ERROR(EINVAL);
1625                 }
1626                 mp->m_sb.sb_rbmblocks = scounter;
1627                 return 0;
1628         case XFS_SBS_RBLOCKS:
1629                 lcounter = (long long)mp->m_sb.sb_rblocks;
1630                 lcounter += delta;
1631                 if (lcounter < 0) {
1632                         ASSERT(0);
1633                         return XFS_ERROR(EINVAL);
1634                 }
1635                 mp->m_sb.sb_rblocks = lcounter;
1636                 return 0;
1637         case XFS_SBS_REXTENTS:
1638                 lcounter = (long long)mp->m_sb.sb_rextents;
1639                 lcounter += delta;
1640                 if (lcounter < 0) {
1641                         ASSERT(0);
1642                         return XFS_ERROR(EINVAL);
1643                 }
1644                 mp->m_sb.sb_rextents = lcounter;
1645                 return 0;
1646         case XFS_SBS_REXTSLOG:
1647                 scounter = mp->m_sb.sb_rextslog;
1648                 scounter += delta;
1649                 if (scounter < 0) {
1650                         ASSERT(0);
1651                         return XFS_ERROR(EINVAL);
1652                 }
1653                 mp->m_sb.sb_rextslog = scounter;
1654                 return 0;
1655         default:
1656                 ASSERT(0);
1657                 return XFS_ERROR(EINVAL);
1658         }
1659 }
1660
1661 /*
1662  * xfs_mod_incore_sb() is used to change a field in the in-core
1663  * superblock structure by the specified delta.  This modification
1664  * is protected by the m_sb_lock.  Just use the xfs_mod_incore_sb_unlocked()
1665  * routine to do the work.
1666  */
1667 int
1668 xfs_mod_incore_sb(
1669         xfs_mount_t     *mp,
1670         xfs_sb_field_t  field,
1671         int64_t         delta,
1672         int             rsvd)
1673 {
1674         int     status;
1675
1676         /* check for per-cpu counters */
1677         switch (field) {
1678 #ifdef HAVE_PERCPU_SB
1679         case XFS_SBS_ICOUNT:
1680         case XFS_SBS_IFREE:
1681         case XFS_SBS_FDBLOCKS:
1682                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1683                         status = xfs_icsb_modify_counters(mp, field,
1684                                                         delta, rsvd);
1685                         break;
1686                 }
1687                 /* FALLTHROUGH */
1688 #endif
1689         default:
1690                 spin_lock(&mp->m_sb_lock);
1691                 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1692                 spin_unlock(&mp->m_sb_lock);
1693                 break;
1694         }
1695
1696         return status;
1697 }
1698
1699 /*
1700  * xfs_mod_incore_sb_batch() is used to change more than one field
1701  * in the in-core superblock structure at a time.  This modification
1702  * is protected by a lock internal to this module.  The fields and
1703  * changes to those fields are specified in the array of xfs_mod_sb
1704  * structures passed in.
1705  *
1706  * Either all of the specified deltas will be applied or none of
1707  * them will.  If any modified field dips below 0, then all modifications
1708  * will be backed out and EINVAL will be returned.
1709  */
1710 int
1711 xfs_mod_incore_sb_batch(xfs_mount_t *mp, xfs_mod_sb_t *msb, uint nmsb, int rsvd)
1712 {
1713         int             status=0;
1714         xfs_mod_sb_t    *msbp;
1715
1716         /*
1717          * Loop through the array of mod structures and apply each
1718          * individually.  If any fail, then back out all those
1719          * which have already been applied.  Do all of this within
1720          * the scope of the m_sb_lock so that all of the changes will
1721          * be atomic.
1722          */
1723         spin_lock(&mp->m_sb_lock);
1724         msbp = &msb[0];
1725         for (msbp = &msbp[0]; msbp < (msb + nmsb); msbp++) {
1726                 /*
1727                  * Apply the delta at index n.  If it fails, break
1728                  * from the loop so we'll fall into the undo loop
1729                  * below.
1730                  */
1731                 switch (msbp->msb_field) {
1732 #ifdef HAVE_PERCPU_SB
1733                 case XFS_SBS_ICOUNT:
1734                 case XFS_SBS_IFREE:
1735                 case XFS_SBS_FDBLOCKS:
1736                         if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1737                                 spin_unlock(&mp->m_sb_lock);
1738                                 status = xfs_icsb_modify_counters(mp,
1739                                                         msbp->msb_field,
1740                                                         msbp->msb_delta, rsvd);
1741                                 spin_lock(&mp->m_sb_lock);
1742                                 break;
1743                         }
1744                         /* FALLTHROUGH */
1745 #endif
1746                 default:
1747                         status = xfs_mod_incore_sb_unlocked(mp,
1748                                                 msbp->msb_field,
1749                                                 msbp->msb_delta, rsvd);
1750                         break;
1751                 }
1752
1753                 if (status != 0) {
1754                         break;
1755                 }
1756         }
1757
1758         /*
1759          * If we didn't complete the loop above, then back out
1760          * any changes made to the superblock.  If you add code
1761          * between the loop above and here, make sure that you
1762          * preserve the value of status. Loop back until
1763          * we step below the beginning of the array.  Make sure
1764          * we don't touch anything back there.
1765          */
1766         if (status != 0) {
1767                 msbp--;
1768                 while (msbp >= msb) {
1769                         switch (msbp->msb_field) {
1770 #ifdef HAVE_PERCPU_SB
1771                         case XFS_SBS_ICOUNT:
1772                         case XFS_SBS_IFREE:
1773                         case XFS_SBS_FDBLOCKS:
1774                                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1775                                         spin_unlock(&mp->m_sb_lock);
1776                                         status = xfs_icsb_modify_counters(mp,
1777                                                         msbp->msb_field,
1778                                                         -(msbp->msb_delta),
1779                                                         rsvd);
1780                                         spin_lock(&mp->m_sb_lock);
1781                                         break;
1782                                 }
1783                                 /* FALLTHROUGH */
1784 #endif
1785                         default:
1786                                 status = xfs_mod_incore_sb_unlocked(mp,
1787                                                         msbp->msb_field,
1788                                                         -(msbp->msb_delta),
1789                                                         rsvd);
1790                                 break;
1791                         }
1792                         ASSERT(status == 0);
1793                         msbp--;
1794                 }
1795         }
1796         spin_unlock(&mp->m_sb_lock);
1797         return status;
1798 }
1799
1800 /*
1801  * xfs_getsb() is called to obtain the buffer for the superblock.
1802  * The buffer is returned locked and read in from disk.
1803  * The buffer should be released with a call to xfs_brelse().
1804  *
1805  * If the flags parameter is BUF_TRYLOCK, then we'll only return
1806  * the superblock buffer if it can be locked without sleeping.
1807  * If it can't then we'll return NULL.
1808  */
1809 xfs_buf_t *
1810 xfs_getsb(
1811         xfs_mount_t     *mp,
1812         int             flags)
1813 {
1814         xfs_buf_t       *bp;
1815
1816         ASSERT(mp->m_sb_bp != NULL);
1817         bp = mp->m_sb_bp;
1818         if (flags & XFS_BUF_TRYLOCK) {
1819                 if (!XFS_BUF_CPSEMA(bp)) {
1820                         return NULL;
1821                 }
1822         } else {
1823                 XFS_BUF_PSEMA(bp, PRIBIO);
1824         }
1825         XFS_BUF_HOLD(bp);
1826         ASSERT(XFS_BUF_ISDONE(bp));
1827         return bp;
1828 }
1829
1830 /*
1831  * Used to free the superblock along various error paths.
1832  */
1833 void
1834 xfs_freesb(
1835         xfs_mount_t     *mp)
1836 {
1837         xfs_buf_t       *bp;
1838
1839         /*
1840          * Use xfs_getsb() so that the buffer will be locked
1841          * when we call xfs_buf_relse().
1842          */
1843         bp = xfs_getsb(mp, 0);
1844         XFS_BUF_UNMANAGE(bp);
1845         xfs_buf_relse(bp);
1846         mp->m_sb_bp = NULL;
1847 }
1848
1849 /*
1850  * See if the UUID is unique among mounted XFS filesystems.
1851  * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
1852  */
1853 STATIC int
1854 xfs_uuid_mount(
1855         xfs_mount_t     *mp)
1856 {
1857         if (uuid_is_nil(&mp->m_sb.sb_uuid)) {
1858                 cmn_err(CE_WARN,
1859                         "XFS: Filesystem %s has nil UUID - can't mount",
1860                         mp->m_fsname);
1861                 return -1;
1862         }
1863         if (!uuid_table_insert(&mp->m_sb.sb_uuid)) {
1864                 cmn_err(CE_WARN,
1865                         "XFS: Filesystem %s has duplicate UUID - can't mount",
1866                         mp->m_fsname);
1867                 return -1;
1868         }
1869         return 0;
1870 }
1871
1872 /*
1873  * Remove filesystem from the UUID table.
1874  */
1875 STATIC void
1876 xfs_uuid_unmount(
1877         xfs_mount_t     *mp)
1878 {
1879         uuid_table_remove(&mp->m_sb.sb_uuid);
1880 }
1881
1882 /*
1883  * Used to log changes to the superblock unit and width fields which could
1884  * be altered by the mount options. Only the first superblock is updated.
1885  */
1886 STATIC void
1887 xfs_mount_log_sbunit(
1888         xfs_mount_t     *mp,
1889         __int64_t       fields)
1890 {
1891         xfs_trans_t     *tp;
1892
1893         ASSERT(fields & (XFS_SB_UNIT|XFS_SB_WIDTH|XFS_SB_UUID));
1894
1895         tp = xfs_trans_alloc(mp, XFS_TRANS_SB_UNIT);
1896         if (xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1897                                 XFS_DEFAULT_LOG_COUNT)) {
1898                 xfs_trans_cancel(tp, 0);
1899                 return;
1900         }
1901         xfs_mod_sb(tp, fields);
1902         xfs_trans_commit(tp, 0);
1903 }
1904
1905
1906 #ifdef HAVE_PERCPU_SB
1907 /*
1908  * Per-cpu incore superblock counters
1909  *
1910  * Simple concept, difficult implementation
1911  *
1912  * Basically, replace the incore superblock counters with a distributed per cpu
1913  * counter for contended fields (e.g.  free block count).
1914  *
1915  * Difficulties arise in that the incore sb is used for ENOSPC checking, and
1916  * hence needs to be accurately read when we are running low on space. Hence
1917  * there is a method to enable and disable the per-cpu counters based on how
1918  * much "stuff" is available in them.
1919  *
1920  * Basically, a counter is enabled if there is enough free resource to justify
1921  * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
1922  * ENOSPC), then we disable the counters to synchronise all callers and
1923  * re-distribute the available resources.
1924  *
1925  * If, once we redistributed the available resources, we still get a failure,
1926  * we disable the per-cpu counter and go through the slow path.
1927  *
1928  * The slow path is the current xfs_mod_incore_sb() function.  This means that
1929  * when we disable a per-cpu counter, we need to drain it's resources back to
1930  * the global superblock. We do this after disabling the counter to prevent
1931  * more threads from queueing up on the counter.
1932  *
1933  * Essentially, this means that we still need a lock in the fast path to enable
1934  * synchronisation between the global counters and the per-cpu counters. This
1935  * is not a problem because the lock will be local to a CPU almost all the time
1936  * and have little contention except when we get to ENOSPC conditions.
1937  *
1938  * Basically, this lock becomes a barrier that enables us to lock out the fast
1939  * path while we do things like enabling and disabling counters and
1940  * synchronising the counters.
1941  *
1942  * Locking rules:
1943  *
1944  *      1. m_sb_lock before picking up per-cpu locks
1945  *      2. per-cpu locks always picked up via for_each_online_cpu() order
1946  *      3. accurate counter sync requires m_sb_lock + per cpu locks
1947  *      4. modifying per-cpu counters requires holding per-cpu lock
1948  *      5. modifying global counters requires holding m_sb_lock
1949  *      6. enabling or disabling a counter requires holding the m_sb_lock
1950  *         and _none_ of the per-cpu locks.
1951  *
1952  * Disabled counters are only ever re-enabled by a balance operation
1953  * that results in more free resources per CPU than a given threshold.
1954  * To ensure counters don't remain disabled, they are rebalanced when
1955  * the global resource goes above a higher threshold (i.e. some hysteresis
1956  * is present to prevent thrashing).
1957  */
1958
1959 #ifdef CONFIG_HOTPLUG_CPU
1960 /*
1961  * hot-plug CPU notifier support.
1962  *
1963  * We need a notifier per filesystem as we need to be able to identify
1964  * the filesystem to balance the counters out. This is achieved by
1965  * having a notifier block embedded in the xfs_mount_t and doing pointer
1966  * magic to get the mount pointer from the notifier block address.
1967  */
1968 STATIC int
1969 xfs_icsb_cpu_notify(
1970         struct notifier_block *nfb,
1971         unsigned long action,
1972         void *hcpu)
1973 {
1974         xfs_icsb_cnts_t *cntp;
1975         xfs_mount_t     *mp;
1976
1977         mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
1978         cntp = (xfs_icsb_cnts_t *)
1979                         per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
1980         switch (action) {
1981         case CPU_UP_PREPARE:
1982         case CPU_UP_PREPARE_FROZEN:
1983                 /* Easy Case - initialize the area and locks, and
1984                  * then rebalance when online does everything else for us. */
1985                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1986                 break;
1987         case CPU_ONLINE:
1988         case CPU_ONLINE_FROZEN:
1989                 xfs_icsb_lock(mp);
1990                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0, 0);
1991                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0, 0);
1992                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0, 0);
1993                 xfs_icsb_unlock(mp);
1994                 break;
1995         case CPU_DEAD:
1996         case CPU_DEAD_FROZEN:
1997                 /* Disable all the counters, then fold the dead cpu's
1998                  * count into the total on the global superblock and
1999                  * re-enable the counters. */
2000                 xfs_icsb_lock(mp);
2001                 spin_lock(&mp->m_sb_lock);
2002                 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
2003                 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
2004                 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
2005
2006                 mp->m_sb.sb_icount += cntp->icsb_icount;
2007                 mp->m_sb.sb_ifree += cntp->icsb_ifree;
2008                 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
2009
2010                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
2011
2012                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT,
2013                                          XFS_ICSB_SB_LOCKED, 0);
2014                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE,
2015                                          XFS_ICSB_SB_LOCKED, 0);
2016                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS,
2017                                          XFS_ICSB_SB_LOCKED, 0);
2018                 spin_unlock(&mp->m_sb_lock);
2019                 xfs_icsb_unlock(mp);
2020                 break;
2021         }
2022
2023         return NOTIFY_OK;
2024 }
2025 #endif /* CONFIG_HOTPLUG_CPU */
2026
2027 int
2028 xfs_icsb_init_counters(
2029         xfs_mount_t     *mp)
2030 {
2031         xfs_icsb_cnts_t *cntp;
2032         int             i;
2033
2034         mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
2035         if (mp->m_sb_cnts == NULL)
2036                 return -ENOMEM;
2037
2038 #ifdef CONFIG_HOTPLUG_CPU
2039         mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
2040         mp->m_icsb_notifier.priority = 0;
2041         register_hotcpu_notifier(&mp->m_icsb_notifier);
2042 #endif /* CONFIG_HOTPLUG_CPU */
2043
2044         for_each_online_cpu(i) {
2045                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2046                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
2047         }
2048
2049         mutex_init(&mp->m_icsb_mutex);
2050
2051         /*
2052          * start with all counters disabled so that the
2053          * initial balance kicks us off correctly
2054          */
2055         mp->m_icsb_counters = -1;
2056         return 0;
2057 }
2058
2059 void
2060 xfs_icsb_reinit_counters(
2061         xfs_mount_t     *mp)
2062 {
2063         xfs_icsb_lock(mp);
2064         /*
2065          * start with all counters disabled so that the
2066          * initial balance kicks us off correctly
2067          */
2068         mp->m_icsb_counters = -1;
2069         xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0, 0);
2070         xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0, 0);
2071         xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0, 0);
2072         xfs_icsb_unlock(mp);
2073 }
2074
2075 STATIC void
2076 xfs_icsb_destroy_counters(
2077         xfs_mount_t     *mp)
2078 {
2079         if (mp->m_sb_cnts) {
2080                 unregister_hotcpu_notifier(&mp->m_icsb_notifier);
2081                 free_percpu(mp->m_sb_cnts);
2082         }
2083         mutex_destroy(&mp->m_icsb_mutex);
2084 }
2085
2086 STATIC_INLINE void
2087 xfs_icsb_lock_cntr(
2088         xfs_icsb_cnts_t *icsbp)
2089 {
2090         while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
2091                 ndelay(1000);
2092         }
2093 }
2094
2095 STATIC_INLINE void
2096 xfs_icsb_unlock_cntr(
2097         xfs_icsb_cnts_t *icsbp)
2098 {
2099         clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
2100 }
2101
2102
2103 STATIC_INLINE void
2104 xfs_icsb_lock_all_counters(
2105         xfs_mount_t     *mp)
2106 {
2107         xfs_icsb_cnts_t *cntp;
2108         int             i;
2109
2110         for_each_online_cpu(i) {
2111                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2112                 xfs_icsb_lock_cntr(cntp);
2113         }
2114 }
2115
2116 STATIC_INLINE void
2117 xfs_icsb_unlock_all_counters(
2118         xfs_mount_t     *mp)
2119 {
2120         xfs_icsb_cnts_t *cntp;
2121         int             i;
2122
2123         for_each_online_cpu(i) {
2124                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2125                 xfs_icsb_unlock_cntr(cntp);
2126         }
2127 }
2128
2129 STATIC void
2130 xfs_icsb_count(
2131         xfs_mount_t     *mp,
2132         xfs_icsb_cnts_t *cnt,
2133         int             flags)
2134 {
2135         xfs_icsb_cnts_t *cntp;
2136         int             i;
2137
2138         memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
2139
2140         if (!(flags & XFS_ICSB_LAZY_COUNT))
2141                 xfs_icsb_lock_all_counters(mp);
2142
2143         for_each_online_cpu(i) {
2144                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
2145                 cnt->icsb_icount += cntp->icsb_icount;
2146                 cnt->icsb_ifree += cntp->icsb_ifree;
2147                 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
2148         }
2149
2150         if (!(flags & XFS_ICSB_LAZY_COUNT))
2151                 xfs_icsb_unlock_all_counters(mp);
2152 }
2153
2154 STATIC int
2155 xfs_icsb_counter_disabled(
2156         xfs_mount_t     *mp,
2157         xfs_sb_field_t  field)
2158 {
2159         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2160         return test_bit(field, &mp->m_icsb_counters);
2161 }
2162
2163 STATIC int
2164 xfs_icsb_disable_counter(
2165         xfs_mount_t     *mp,
2166         xfs_sb_field_t  field)
2167 {
2168         xfs_icsb_cnts_t cnt;
2169
2170         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2171
2172         /*
2173          * If we are already disabled, then there is nothing to do
2174          * here. We check before locking all the counters to avoid
2175          * the expensive lock operation when being called in the
2176          * slow path and the counter is already disabled. This is
2177          * safe because the only time we set or clear this state is under
2178          * the m_icsb_mutex.
2179          */
2180         if (xfs_icsb_counter_disabled(mp, field))
2181                 return 0;
2182
2183         xfs_icsb_lock_all_counters(mp);
2184         if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
2185                 /* drain back to superblock */
2186
2187                 xfs_icsb_count(mp, &cnt, XFS_ICSB_SB_LOCKED|XFS_ICSB_LAZY_COUNT);
2188                 switch(field) {
2189                 case XFS_SBS_ICOUNT:
2190                         mp->m_sb.sb_icount = cnt.icsb_icount;
2191                         break;
2192                 case XFS_SBS_IFREE:
2193                         mp->m_sb.sb_ifree = cnt.icsb_ifree;
2194                         break;
2195                 case XFS_SBS_FDBLOCKS:
2196                         mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2197                         break;
2198                 default:
2199                         BUG();
2200                 }
2201         }
2202
2203         xfs_icsb_unlock_all_counters(mp);
2204
2205         return 0;
2206 }
2207
2208 STATIC void
2209 xfs_icsb_enable_counter(
2210         xfs_mount_t     *mp,
2211         xfs_sb_field_t  field,
2212         uint64_t        count,
2213         uint64_t        resid)
2214 {
2215         xfs_icsb_cnts_t *cntp;
2216         int             i;
2217
2218         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
2219
2220         xfs_icsb_lock_all_counters(mp);
2221         for_each_online_cpu(i) {
2222                 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
2223                 switch (field) {
2224                 case XFS_SBS_ICOUNT:
2225                         cntp->icsb_icount = count + resid;
2226                         break;
2227                 case XFS_SBS_IFREE:
2228                         cntp->icsb_ifree = count + resid;
2229                         break;
2230                 case XFS_SBS_FDBLOCKS:
2231                         cntp->icsb_fdblocks = count + resid;
2232                         break;
2233                 default:
2234                         BUG();
2235                         break;
2236                 }
2237                 resid = 0;
2238         }
2239         clear_bit(field, &mp->m_icsb_counters);
2240         xfs_icsb_unlock_all_counters(mp);
2241 }
2242
2243 void
2244 xfs_icsb_sync_counters_flags(
2245         xfs_mount_t     *mp,
2246         int             flags)
2247 {
2248         xfs_icsb_cnts_t cnt;
2249
2250         /* Pass 1: lock all counters */
2251         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
2252                 spin_lock(&mp->m_sb_lock);
2253
2254         xfs_icsb_count(mp, &cnt, flags);
2255
2256         /* Step 3: update mp->m_sb fields */
2257         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
2258                 mp->m_sb.sb_icount = cnt.icsb_icount;
2259         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
2260                 mp->m_sb.sb_ifree = cnt.icsb_ifree;
2261         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
2262                 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2263
2264         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
2265                 spin_unlock(&mp->m_sb_lock);
2266 }
2267
2268 /*
2269  * Accurate update of per-cpu counters to incore superblock
2270  */
2271 STATIC void
2272 xfs_icsb_sync_counters(
2273         xfs_mount_t     *mp)
2274 {
2275         xfs_icsb_sync_counters_flags(mp, 0);
2276 }
2277
2278 /*
2279  * Balance and enable/disable counters as necessary.
2280  *
2281  * Thresholds for re-enabling counters are somewhat magic.  inode counts are
2282  * chosen to be the same number as single on disk allocation chunk per CPU, and
2283  * free blocks is something far enough zero that we aren't going thrash when we
2284  * get near ENOSPC. We also need to supply a minimum we require per cpu to
2285  * prevent looping endlessly when xfs_alloc_space asks for more than will
2286  * be distributed to a single CPU but each CPU has enough blocks to be
2287  * reenabled.
2288  *
2289  * Note that we can be called when counters are already disabled.
2290  * xfs_icsb_disable_counter() optimises the counter locking in this case to
2291  * prevent locking every per-cpu counter needlessly.
2292  */
2293
2294 #define XFS_ICSB_INO_CNTR_REENABLE      (uint64_t)64
2295 #define XFS_ICSB_FDBLK_CNTR_REENABLE(mp) \
2296                 (uint64_t)(512 + XFS_ALLOC_SET_ASIDE(mp))
2297 STATIC void
2298 xfs_icsb_balance_counter(
2299         xfs_mount_t     *mp,
2300         xfs_sb_field_t  field,
2301         int             flags,
2302         int             min_per_cpu)
2303 {
2304         uint64_t        count, resid;
2305         int             weight = num_online_cpus();
2306         uint64_t        min = (uint64_t)min_per_cpu;
2307
2308         if (!(flags & XFS_ICSB_SB_LOCKED))
2309                 spin_lock(&mp->m_sb_lock);
2310
2311         /* disable counter and sync counter */
2312         xfs_icsb_disable_counter(mp, field);
2313
2314         /* update counters  - first CPU gets residual*/
2315         switch (field) {
2316         case XFS_SBS_ICOUNT:
2317                 count = mp->m_sb.sb_icount;
2318                 resid = do_div(count, weight);
2319                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2320                         goto out;
2321                 break;
2322         case XFS_SBS_IFREE:
2323                 count = mp->m_sb.sb_ifree;
2324                 resid = do_div(count, weight);
2325                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2326                         goto out;
2327                 break;
2328         case XFS_SBS_FDBLOCKS:
2329                 count = mp->m_sb.sb_fdblocks;
2330                 resid = do_div(count, weight);
2331                 if (count < max(min, XFS_ICSB_FDBLK_CNTR_REENABLE(mp)))
2332                         goto out;
2333                 break;
2334         default:
2335                 BUG();
2336                 count = resid = 0;      /* quiet, gcc */
2337                 break;
2338         }
2339
2340         xfs_icsb_enable_counter(mp, field, count, resid);
2341 out:
2342         if (!(flags & XFS_ICSB_SB_LOCKED))
2343                 spin_unlock(&mp->m_sb_lock);
2344 }
2345
2346 STATIC int
2347 xfs_icsb_modify_counters(
2348         xfs_mount_t     *mp,
2349         xfs_sb_field_t  field,
2350         int64_t         delta,
2351         int             rsvd)
2352 {
2353         xfs_icsb_cnts_t *icsbp;
2354         long long       lcounter;       /* long counter for 64 bit fields */
2355         int             cpu, ret = 0;
2356
2357         might_sleep();
2358 again:
2359         cpu = get_cpu();
2360         icsbp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, cpu);
2361
2362         /*
2363          * if the counter is disabled, go to slow path
2364          */
2365         if (unlikely(xfs_icsb_counter_disabled(mp, field)))
2366                 goto slow_path;
2367         xfs_icsb_lock_cntr(icsbp);
2368         if (unlikely(xfs_icsb_counter_disabled(mp, field))) {
2369                 xfs_icsb_unlock_cntr(icsbp);
2370                 goto slow_path;
2371         }
2372
2373         switch (field) {
2374         case XFS_SBS_ICOUNT:
2375                 lcounter = icsbp->icsb_icount;
2376                 lcounter += delta;
2377                 if (unlikely(lcounter < 0))
2378                         goto balance_counter;
2379                 icsbp->icsb_icount = lcounter;
2380                 break;
2381
2382         case XFS_SBS_IFREE:
2383                 lcounter = icsbp->icsb_ifree;
2384                 lcounter += delta;
2385                 if (unlikely(lcounter < 0))
2386                         goto balance_counter;
2387                 icsbp->icsb_ifree = lcounter;
2388                 break;
2389
2390         case XFS_SBS_FDBLOCKS:
2391                 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
2392
2393                 lcounter = icsbp->icsb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
2394                 lcounter += delta;
2395                 if (unlikely(lcounter < 0))
2396                         goto balance_counter;
2397                 icsbp->icsb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
2398                 break;
2399         default:
2400                 BUG();
2401                 break;
2402         }
2403         xfs_icsb_unlock_cntr(icsbp);
2404         put_cpu();
2405         return 0;
2406
2407 slow_path:
2408         put_cpu();
2409
2410         /*
2411          * serialise with a mutex so we don't burn lots of cpu on
2412          * the superblock lock. We still need to hold the superblock
2413          * lock, however, when we modify the global structures.
2414          */
2415         xfs_icsb_lock(mp);
2416
2417         /*
2418          * Now running atomically.
2419          *
2420          * If the counter is enabled, someone has beaten us to rebalancing.
2421          * Drop the lock and try again in the fast path....
2422          */
2423         if (!(xfs_icsb_counter_disabled(mp, field))) {
2424                 xfs_icsb_unlock(mp);
2425                 goto again;
2426         }
2427
2428         /*
2429          * The counter is currently disabled. Because we are
2430          * running atomically here, we know a rebalance cannot
2431          * be in progress. Hence we can go straight to operating
2432          * on the global superblock. We do not call xfs_mod_incore_sb()
2433          * here even though we need to get the m_sb_lock. Doing so
2434          * will cause us to re-enter this function and deadlock.
2435          * Hence we get the m_sb_lock ourselves and then call
2436          * xfs_mod_incore_sb_unlocked() as the unlocked path operates
2437          * directly on the global counters.
2438          */
2439         spin_lock(&mp->m_sb_lock);
2440         ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
2441         spin_unlock(&mp->m_sb_lock);
2442
2443         /*
2444          * Now that we've modified the global superblock, we
2445          * may be able to re-enable the distributed counters
2446          * (e.g. lots of space just got freed). After that
2447          * we are done.
2448          */
2449         if (ret != ENOSPC)
2450                 xfs_icsb_balance_counter(mp, field, 0, 0);
2451         xfs_icsb_unlock(mp);
2452         return ret;
2453
2454 balance_counter:
2455         xfs_icsb_unlock_cntr(icsbp);
2456         put_cpu();
2457
2458         /*
2459          * We may have multiple threads here if multiple per-cpu
2460          * counters run dry at the same time. This will mean we can
2461          * do more balances than strictly necessary but it is not
2462          * the common slowpath case.
2463          */
2464         xfs_icsb_lock(mp);
2465
2466         /*
2467          * running atomically.
2468          *
2469          * This will leave the counter in the correct state for future
2470          * accesses. After the rebalance, we simply try again and our retry
2471          * will either succeed through the fast path or slow path without
2472          * another balance operation being required.
2473          */
2474         xfs_icsb_balance_counter(mp, field, 0, delta);
2475         xfs_icsb_unlock(mp);
2476         goto again;
2477 }
2478
2479 #endif