err.no Git - linux-2.6/blob - fs/xfs/xfs_mount.c

   1 /*
   2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_types.h"
  21 #include "xfs_bit.h"
  22 #include "xfs_log.h"
  23 #include "xfs_inum.h"
  24 #include "xfs_trans.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_ag.h"
  27 #include "xfs_dir2.h"
  28 #include "xfs_dmapi.h"
  29 #include "xfs_mount.h"
  30 #include "xfs_bmap_btree.h"
  31 #include "xfs_alloc_btree.h"
  32 #include "xfs_ialloc_btree.h"
  33 #include "xfs_dir2_sf.h"
  34 #include "xfs_attr_sf.h"
  35 #include "xfs_dinode.h"
  36 #include "xfs_inode.h"
  37 #include "xfs_btree.h"
  38 #include "xfs_ialloc.h"
  39 #include "xfs_alloc.h"
  40 #include "xfs_rtalloc.h"
  41 #include "xfs_bmap.h"
  42 #include "xfs_error.h"
  43 #include "xfs_rw.h"
  44 #include "xfs_quota.h"
  45 #include "xfs_fsops.h"
  46
  47 STATIC void     xfs_mount_log_sbunit(xfs_mount_t *, __int64_t);
  48 STATIC int      xfs_uuid_mount(xfs_mount_t *);
  49 STATIC void     xfs_uuid_unmount(xfs_mount_t *mp);
  50 STATIC void     xfs_unmountfs_wait(xfs_mount_t *);
  51
  52
  53 #ifdef HAVE_PERCPU_SB
  54 STATIC void     xfs_icsb_destroy_counters(xfs_mount_t *);
  55 STATIC void     xfs_icsb_balance_counter(xfs_mount_t *, xfs_sb_field_t,
  56                                                 int, int);
  57 STATIC void     xfs_icsb_sync_counters(xfs_mount_t *);
  58 STATIC int      xfs_icsb_modify_counters(xfs_mount_t *, xfs_sb_field_t,
  59                                                 int64_t, int);
  60 STATIC int      xfs_icsb_disable_counter(xfs_mount_t *, xfs_sb_field_t);
  61
  62 #else
  63
  64 #define xfs_icsb_destroy_counters(mp)                   do { } while (0)
  65 #define xfs_icsb_balance_counter(mp, a, b, c)           do { } while (0)
  66 #define xfs_icsb_sync_counters(mp)                      do { } while (0)
  67 #define xfs_icsb_modify_counters(mp, a, b, c)           do { } while (0)
  68
  69 #endif
  70
  71 static const struct {
  72         short offset;
  73         short type;     /* 0 = integer
  74                          * 1 = binary / string (no translation)
  75                          */
  76 } xfs_sb_info[] = {
  77     { offsetof(xfs_sb_t, sb_magicnum),   0 },
  78     { offsetof(xfs_sb_t, sb_blocksize),  0 },
  79     { offsetof(xfs_sb_t, sb_dblocks),    0 },
  80     { offsetof(xfs_sb_t, sb_rblocks),    0 },
  81     { offsetof(xfs_sb_t, sb_rextents),   0 },
  82     { offsetof(xfs_sb_t, sb_uuid),       1 },
  83     { offsetof(xfs_sb_t, sb_logstart),   0 },
  84     { offsetof(xfs_sb_t, sb_rootino),    0 },
  85     { offsetof(xfs_sb_t, sb_rbmino),     0 },
  86     { offsetof(xfs_sb_t, sb_rsumino),    0 },
  87     { offsetof(xfs_sb_t, sb_rextsize),   0 },
  88     { offsetof(xfs_sb_t, sb_agblocks),   0 },
  89     { offsetof(xfs_sb_t, sb_agcount),    0 },
  90     { offsetof(xfs_sb_t, sb_rbmblocks),  0 },
  91     { offsetof(xfs_sb_t, sb_logblocks),  0 },
  92     { offsetof(xfs_sb_t, sb_versionnum), 0 },
  93     { offsetof(xfs_sb_t, sb_sectsize),   0 },
  94     { offsetof(xfs_sb_t, sb_inodesize),  0 },
  95     { offsetof(xfs_sb_t, sb_inopblock),  0 },
  96     { offsetof(xfs_sb_t, sb_fname[0]),   1 },
  97     { offsetof(xfs_sb_t, sb_blocklog),   0 },
  98     { offsetof(xfs_sb_t, sb_sectlog),    0 },
  99     { offsetof(xfs_sb_t, sb_inodelog),   0 },
 100     { offsetof(xfs_sb_t, sb_inopblog),   0 },
 101     { offsetof(xfs_sb_t, sb_agblklog),   0 },
 102     { offsetof(xfs_sb_t, sb_rextslog),   0 },
 103     { offsetof(xfs_sb_t, sb_inprogress), 0 },
 104     { offsetof(xfs_sb_t, sb_imax_pct),   0 },
 105     { offsetof(xfs_sb_t, sb_icount),     0 },
 106     { offsetof(xfs_sb_t, sb_ifree),      0 },
 107     { offsetof(xfs_sb_t, sb_fdblocks),   0 },
 108     { offsetof(xfs_sb_t, sb_frextents),  0 },
 109     { offsetof(xfs_sb_t, sb_uquotino),   0 },
 110     { offsetof(xfs_sb_t, sb_gquotino),   0 },
 111     { offsetof(xfs_sb_t, sb_qflags),     0 },
 112     { offsetof(xfs_sb_t, sb_flags),      0 },
 113     { offsetof(xfs_sb_t, sb_shared_vn),  0 },
 114     { offsetof(xfs_sb_t, sb_inoalignmt), 0 },
 115     { offsetof(xfs_sb_t, sb_unit),       0 },
 116     { offsetof(xfs_sb_t, sb_width),      0 },
 117     { offsetof(xfs_sb_t, sb_dirblklog),  0 },
 118     { offsetof(xfs_sb_t, sb_logsectlog), 0 },
 119     { offsetof(xfs_sb_t, sb_logsectsize),0 },
 120     { offsetof(xfs_sb_t, sb_logsunit),   0 },
 121     { offsetof(xfs_sb_t, sb_features2),  0 },
 122     { sizeof(xfs_sb_t),                  0 }
 123 };
 124
 125 /*
 126  * Return a pointer to an initialized xfs_mount structure.
 127  */
 128 xfs_mount_t *
 129 xfs_mount_init(void)
 130 {
 131         xfs_mount_t *mp;
 132
 133         mp = kmem_zalloc(sizeof(xfs_mount_t), KM_SLEEP);
 134
 135         if (xfs_icsb_init_counters(mp)) {
 136                 mp->m_flags |= XFS_MOUNT_NO_PERCPU_SB;
 137         }
 138
 139         AIL_LOCKINIT(&mp->m_ail_lock, "xfs_ail");
 140         spinlock_init(&mp->m_sb_lock, "xfs_sb");
 141         mutex_init(&mp->m_ilock);
 142         initnsema(&mp->m_growlock, 1, "xfs_grow");
 143         /*
 144          * Initialize the AIL.
 145          */
 146         xfs_trans_ail_init(mp);
 147
 148         atomic_set(&mp->m_active_trans, 0);
 149
 150         return mp;
 151 }
 152
 153 /*
 154  * Free up the resources associated with a mount structure.  Assume that
 155  * the structure was initially zeroed, so we can tell which fields got
 156  * initialized.
 157  */
 158 void
 159 xfs_mount_free(
 160         xfs_mount_t     *mp,
 161         int             remove_bhv)
 162 {
 163         if (mp->m_ihash)
 164                 xfs_ihash_free(mp);
 165         if (mp->m_chash)
 166                 xfs_chash_free(mp);
 167
 168         if (mp->m_perag) {
 169                 int     agno;
 170
 171                 for (agno = 0; agno < mp->m_maxagi; agno++)
 172                         if (mp->m_perag[agno].pagb_list)
 173                                 kmem_free(mp->m_perag[agno].pagb_list,
 174                                                 sizeof(xfs_perag_busy_t) *
 175                                                         XFS_PAGB_NUM_SLOTS);
 176                 kmem_free(mp->m_perag,
 177                           sizeof(xfs_perag_t) * mp->m_sb.sb_agcount);
 178         }
 179
 180         AIL_LOCK_DESTROY(&mp->m_ail_lock);
 181         spinlock_destroy(&mp->m_sb_lock);
 182         mutex_destroy(&mp->m_ilock);
 183         freesema(&mp->m_growlock);
 184         if (mp->m_quotainfo)
 185                 XFS_QM_DONE(mp);
 186
 187         if (mp->m_fsname != NULL)
 188                 kmem_free(mp->m_fsname, mp->m_fsname_len);
 189         if (mp->m_rtname != NULL)
 190                 kmem_free(mp->m_rtname, strlen(mp->m_rtname) + 1);
 191         if (mp->m_logname != NULL)
 192                 kmem_free(mp->m_logname, strlen(mp->m_logname) + 1);
 193
 194         if (remove_bhv) {
 195                 struct bhv_vfs  *vfsp = XFS_MTOVFS(mp);
 196
 197                 bhv_remove_all_vfsops(vfsp, 0);
 198                 VFS_REMOVEBHV(vfsp, &mp->m_bhv);
 199         }
 200
 201         xfs_icsb_destroy_counters(mp);
 202         kmem_free(mp, sizeof(xfs_mount_t));
 203 }
 204
 205 /*
 206  * Check size of device based on the (data/realtime) block count.
 207  * Note: this check is used by the growfs code as well as mount.
 208  */
 209 int
 210 xfs_sb_validate_fsb_count(
 211         xfs_sb_t        *sbp,
 212         __uint64_t      nblocks)
 213 {
 214         ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
 215         ASSERT(sbp->sb_blocklog >= BBSHIFT);
 216
 217 #if XFS_BIG_BLKNOS     /* Limited by ULONG_MAX of page cache index */
 218         if (nblocks >> (PAGE_CACHE_SHIFT - sbp->sb_blocklog) > ULONG_MAX)
 219                 return E2BIG;
 220 #else                  /* Limited by UINT_MAX of sectors */
 221         if (nblocks << (sbp->sb_blocklog - BBSHIFT) > UINT_MAX)
 222                 return E2BIG;
 223 #endif
 224         return 0;
 225 }
 226
 227 /*
 228  * Check the validity of the SB found.
 229  */
 230 STATIC int
 231 xfs_mount_validate_sb(
 232         xfs_mount_t     *mp,
 233         xfs_sb_t        *sbp,
 234         int             flags)
 235 {
 236         /*
 237          * If the log device and data device have the
 238          * same device number, the log is internal.
 239          * Consequently, the sb_logstart should be non-zero.  If
 240          * we have a zero sb_logstart in this case, we may be trying to mount
 241          * a volume filesystem in a non-volume manner.
 242          */
 243         if (sbp->sb_magicnum != XFS_SB_MAGIC) {
 244                 xfs_fs_mount_cmn_err(flags, "bad magic number");
 245                 return XFS_ERROR(EWRONGFS);
 246         }
 247
 248         if (!XFS_SB_GOOD_VERSION(sbp)) {
 249                 xfs_fs_mount_cmn_err(flags, "bad version");
 250                 return XFS_ERROR(EWRONGFS);
 251         }
 252
 253         if (unlikely(
 254             sbp->sb_logstart == 0 && mp->m_logdev_targp == mp->m_ddev_targp)) {
 255                 xfs_fs_mount_cmn_err(flags,
 256                         "filesystem is marked as having an external log; "
 257                         "specify logdev on the\nmount command line.");
 258                 return XFS_ERROR(EINVAL);
 259         }
 260
 261         if (unlikely(
 262             sbp->sb_logstart != 0 && mp->m_logdev_targp != mp->m_ddev_targp)) {
 263                 xfs_fs_mount_cmn_err(flags,
 264                         "filesystem is marked as having an internal log; "
 265                         "do not specify logdev on\nthe mount command line.");
 266                 return XFS_ERROR(EINVAL);
 267         }
 268
 269         /*
 270          * More sanity checking. These were stolen directly from
 271          * xfs_repair.
 272          */
 273         if (unlikely(
 274             sbp->sb_agcount <= 0                                        ||
 275             sbp->sb_sectsize < XFS_MIN_SECTORSIZE                       ||
 276             sbp->sb_sectsize > XFS_MAX_SECTORSIZE                       ||
 277             sbp->sb_sectlog < XFS_MIN_SECTORSIZE_LOG                    ||
 278             sbp->sb_sectlog > XFS_MAX_SECTORSIZE_LOG                    ||
 279             sbp->sb_blocksize < XFS_MIN_BLOCKSIZE                       ||
 280             sbp->sb_blocksize > XFS_MAX_BLOCKSIZE                       ||
 281             sbp->sb_blocklog < XFS_MIN_BLOCKSIZE_LOG                    ||
 282             sbp->sb_blocklog > XFS_MAX_BLOCKSIZE_LOG                    ||
 283             sbp->sb_inodesize < XFS_DINODE_MIN_SIZE                     ||
 284             sbp->sb_inodesize > XFS_DINODE_MAX_SIZE                     ||
 285             sbp->sb_inodelog < XFS_DINODE_MIN_LOG                       ||
 286             sbp->sb_inodelog > XFS_DINODE_MAX_LOG                       ||
 287             (sbp->sb_blocklog - sbp->sb_inodelog != sbp->sb_inopblog)   ||
 288             (sbp->sb_rextsize * sbp->sb_blocksize > XFS_MAX_RTEXTSIZE)  ||
 289             (sbp->sb_rextsize * sbp->sb_blocksize < XFS_MIN_RTEXTSIZE)  ||
 290             (sbp->sb_imax_pct > 100 /* zero sb_imax_pct is valid */))) {
 291                 xfs_fs_mount_cmn_err(flags, "SB sanity check 1 failed");
 292                 return XFS_ERROR(EFSCORRUPTED);
 293         }
 294
 295         /*
 296          * Sanity check AG count, size fields against data size field
 297          */
 298         if (unlikely(
 299             sbp->sb_dblocks == 0 ||
 300             sbp->sb_dblocks >
 301              (xfs_drfsbno_t)sbp->sb_agcount * sbp->sb_agblocks ||
 302             sbp->sb_dblocks < (xfs_drfsbno_t)(sbp->sb_agcount - 1) *
 303                               sbp->sb_agblocks + XFS_MIN_AG_BLOCKS)) {
 304                 xfs_fs_mount_cmn_err(flags, "SB sanity check 2 failed");
 305                 return XFS_ERROR(EFSCORRUPTED);
 306         }
 307
 308         if (xfs_sb_validate_fsb_count(sbp, sbp->sb_dblocks) ||
 309             xfs_sb_validate_fsb_count(sbp, sbp->sb_rblocks)) {
 310                 xfs_fs_mount_cmn_err(flags,
 311                         "file system too large to be mounted on this system.");
 312                 return XFS_ERROR(E2BIG);
 313         }
 314
 315         if (unlikely(sbp->sb_inprogress)) {
 316                 xfs_fs_mount_cmn_err(flags, "file system busy");
 317                 return XFS_ERROR(EFSCORRUPTED);
 318         }
 319
 320         /*
 321          * Version 1 directory format has never worked on Linux.
 322          */
 323         if (unlikely(!XFS_SB_VERSION_HASDIRV2(sbp))) {
 324                 xfs_fs_mount_cmn_err(flags,
 325                         "file system using version 1 directory format");
 326                 return XFS_ERROR(ENOSYS);
 327         }
 328
 329         /*
 330          * Until this is fixed only page-sized or smaller data blocks work.
 331          */
 332         if (unlikely(sbp->sb_blocksize > PAGE_SIZE)) {
 333                 xfs_fs_mount_cmn_err(flags,
 334                         "file system with blocksize %d bytes",
 335                         sbp->sb_blocksize);
 336                 xfs_fs_mount_cmn_err(flags,
 337                         "only pagesize (%ld) or less will currently work.",
 338                         PAGE_SIZE);
 339                 return XFS_ERROR(ENOSYS);
 340         }
 341
 342         return 0;
 343 }
 344
 345 xfs_agnumber_t
 346 xfs_initialize_perag(
 347         bhv_vfs_t       *vfs,
 348         xfs_mount_t     *mp,
 349         xfs_agnumber_t  agcount)
 350 {
 351         xfs_agnumber_t  index, max_metadata;
 352         xfs_perag_t     *pag;
 353         xfs_agino_t     agino;
 354         xfs_ino_t       ino;
 355         xfs_sb_t        *sbp = &mp->m_sb;
 356         xfs_ino_t       max_inum = XFS_MAXINUMBER_32;
 357
 358         /* Check to see if the filesystem can overflow 32 bit inodes */
 359         agino = XFS_OFFBNO_TO_AGINO(mp, sbp->sb_agblocks - 1, 0);
 360         ino = XFS_AGINO_TO_INO(mp, agcount - 1, agino);
 361
 362         /* Clear the mount flag if no inode can overflow 32 bits
 363          * on this filesystem, or if specifically requested..
 364          */
 365         if ((vfs->vfs_flag & VFS_32BITINODES) && ino > max_inum) {
 366                 mp->m_flags |= XFS_MOUNT_32BITINODES;
 367         } else {
 368                 mp->m_flags &= ~XFS_MOUNT_32BITINODES;
 369         }
 370
 371         /* If we can overflow then setup the ag headers accordingly */
 372         if (mp->m_flags & XFS_MOUNT_32BITINODES) {
 373                 /* Calculate how much should be reserved for inodes to
 374                  * meet the max inode percentage.
 375                  */
 376                 if (mp->m_maxicount) {
 377                         __uint64_t      icount;
 378
 379                         icount = sbp->sb_dblocks * sbp->sb_imax_pct;
 380                         do_div(icount, 100);
 381                         icount += sbp->sb_agblocks - 1;
 382                         do_div(icount, sbp->sb_agblocks);
 383                         max_metadata = icount;
 384                 } else {
 385                         max_metadata = agcount;
 386                 }
 387                 for (index = 0; index < agcount; index++) {
 388                         ino = XFS_AGINO_TO_INO(mp, index, agino);
 389                         if (ino > max_inum) {
 390                                 index++;
 391                                 break;
 392                         }
 393
 394                         /* This ag is preferred for inodes */
 395                         pag = &mp->m_perag[index];
 396                         pag->pagi_inodeok = 1;
 397                         if (index < max_metadata)
 398                                 pag->pagf_metadata = 1;
 399                 }
 400         } else {
 401                 /* Setup default behavior for smaller filesystems */
 402                 for (index = 0; index < agcount; index++) {
 403                         pag = &mp->m_perag[index];
 404                         pag->pagi_inodeok = 1;
 405                 }
 406         }
 407         return index;
 408 }
 409
 410 /*
 411  * xfs_xlatesb
 412  *
 413  *     data       - on disk version of sb
 414  *     sb         - a superblock
 415  *     dir        - conversion direction: <0 - convert sb to buf
 416  *                                        >0 - convert buf to sb
 417  *     fields     - which fields to copy (bitmask)
 418  */
 419 void
 420 xfs_xlatesb(
 421         void            *data,
 422         xfs_sb_t        *sb,
 423         int             dir,
 424         __int64_t       fields)
 425 {
 426         xfs_caddr_t     buf_ptr;
 427         xfs_caddr_t     mem_ptr;
 428         xfs_sb_field_t  f;
 429         int             first;
 430         int             size;
 431
 432         ASSERT(dir);
 433         ASSERT(fields);
 434
 435         if (!fields)
 436                 return;
 437
 438         buf_ptr = (xfs_caddr_t)data;
 439         mem_ptr = (xfs_caddr_t)sb;
 440
 441         while (fields) {
 442                 f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
 443                 first = xfs_sb_info[f].offset;
 444                 size = xfs_sb_info[f + 1].offset - first;
 445
 446                 ASSERT(xfs_sb_info[f].type == 0 || xfs_sb_info[f].type == 1);
 447
 448                 if (size == 1 || xfs_sb_info[f].type == 1) {
 449                         if (dir > 0) {
 450                                 memcpy(mem_ptr + first, buf_ptr + first, size);
 451                         } else {
 452                                 memcpy(buf_ptr + first, mem_ptr + first, size);
 453                         }
 454                 } else {
 455                         switch (size) {
 456                         case 2:
 457                                 INT_XLATE(*(__uint16_t*)(buf_ptr+first),
 458                                           *(__uint16_t*)(mem_ptr+first),
 459                                           dir, ARCH_CONVERT);
 460                                 break;
 461                         case 4:
 462                                 INT_XLATE(*(__uint32_t*)(buf_ptr+first),
 463                                           *(__uint32_t*)(mem_ptr+first),
 464                                           dir, ARCH_CONVERT);
 465                                 break;
 466                         case 8:
 467                                 INT_XLATE(*(__uint64_t*)(buf_ptr+first),
 468                                           *(__uint64_t*)(mem_ptr+first), dir, ARCH_CONVERT);
 469                                 break;
 470                         default:
 471                                 ASSERT(0);
 472                         }
 473                 }
 474
 475                 fields &= ~(1LL << f);
 476         }
 477 }
 478
 479 /*
 480  * xfs_readsb
 481  *
 482  * Does the initial read of the superblock.
 483  */
 484 int
 485 xfs_readsb(xfs_mount_t *mp, int flags)
 486 {
 487         unsigned int    sector_size;
 488         unsigned int    extra_flags;
 489         xfs_buf_t       *bp;
 490         xfs_sb_t        *sbp;
 491         int             error;
 492
 493         ASSERT(mp->m_sb_bp == NULL);
 494         ASSERT(mp->m_ddev_targp != NULL);
 495
 496         /*
 497          * Allocate a (locked) buffer to hold the superblock.
 498          * This will be kept around at all times to optimize
 499          * access to the superblock.
 500          */
 501         sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
 502         extra_flags = XFS_BUF_LOCK | XFS_BUF_MANAGE | XFS_BUF_MAPPED;
 503
 504         bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
 505                                 BTOBB(sector_size), extra_flags);
 506         if (!bp || XFS_BUF_ISERROR(bp)) {
 507                 xfs_fs_mount_cmn_err(flags, "SB read failed");
 508                 error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
 509                 goto fail;
 510         }
 511         ASSERT(XFS_BUF_ISBUSY(bp));
 512         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 513
 514         /*
 515          * Initialize the mount structure from the superblock.
 516          * But first do some basic consistency checking.
 517          */
 518         sbp = XFS_BUF_TO_SBP(bp);
 519         xfs_xlatesb(XFS_BUF_PTR(bp), &(mp->m_sb), 1, XFS_SB_ALL_BITS);
 520
 521         error = xfs_mount_validate_sb(mp, &(mp->m_sb), flags);
 522         if (error) {
 523                 xfs_fs_mount_cmn_err(flags, "SB validate failed");
 524                 goto fail;
 525         }
 526
 527         /*
 528          * We must be able to do sector-sized and sector-aligned IO.
 529          */
 530         if (sector_size > mp->m_sb.sb_sectsize) {
 531                 xfs_fs_mount_cmn_err(flags,
 532                         "device supports only %u byte sectors (not %u)",
 533                         sector_size, mp->m_sb.sb_sectsize);
 534                 error = ENOSYS;
 535                 goto fail;
 536         }
 537
 538         /*
 539          * If device sector size is smaller than the superblock size,
 540          * re-read the superblock so the buffer is correctly sized.
 541          */
 542         if (sector_size < mp->m_sb.sb_sectsize) {
 543                 XFS_BUF_UNMANAGE(bp);
 544                 xfs_buf_relse(bp);
 545                 sector_size = mp->m_sb.sb_sectsize;
 546                 bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
 547                                         BTOBB(sector_size), extra_flags);
 548                 if (!bp || XFS_BUF_ISERROR(bp)) {
 549                         xfs_fs_mount_cmn_err(flags, "SB re-read failed");
 550                         error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
 551                         goto fail;
 552                 }
 553                 ASSERT(XFS_BUF_ISBUSY(bp));
 554                 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 555         }
 556
 557         /* Initialize per-cpu counters */
 558         xfs_icsb_reinit_counters(mp);
 559
 560         mp->m_sb_bp = bp;
 561         xfs_buf_relse(bp);
 562         ASSERT(XFS_BUF_VALUSEMA(bp) > 0);
 563         return 0;
 564
 565  fail:
 566         if (bp) {
 567                 XFS_BUF_UNMANAGE(bp);
 568                 xfs_buf_relse(bp);
 569         }
 570         return error;
 571 }
 572
 573
 574 /*
 575  * xfs_mount_common
 576  *
 577  * Mount initialization code establishing various mount
 578  * fields from the superblock associated with the given
 579  * mount structure
 580  */
 581 STATIC void
 582 xfs_mount_common(xfs_mount_t *mp, xfs_sb_t *sbp)
 583 {
 584         int     i;
 585
 586         mp->m_agfrotor = mp->m_agirotor = 0;
 587         spinlock_init(&mp->m_agirotor_lock, "m_agirotor_lock");
 588         mp->m_maxagi = mp->m_sb.sb_agcount;
 589         mp->m_blkbit_log = sbp->sb_blocklog + XFS_NBBYLOG;
 590         mp->m_blkbb_log = sbp->sb_blocklog - BBSHIFT;
 591         mp->m_sectbb_log = sbp->sb_sectlog - BBSHIFT;
 592         mp->m_agno_log = xfs_highbit32(sbp->sb_agcount - 1) + 1;
 593         mp->m_agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
 594         mp->m_litino = sbp->sb_inodesize -
 595                 ((uint)sizeof(xfs_dinode_core_t) + (uint)sizeof(xfs_agino_t));
 596         mp->m_blockmask = sbp->sb_blocksize - 1;
 597         mp->m_blockwsize = sbp->sb_blocksize >> XFS_WORDLOG;
 598         mp->m_blockwmask = mp->m_blockwsize - 1;
 599         INIT_LIST_HEAD(&mp->m_del_inodes);
 600
 601         /*
 602          * Setup for attributes, in case they get created.
 603          * This value is for inodes getting attributes for the first time,
 604          * the per-inode value is for old attribute values.
 605          */
 606         ASSERT(sbp->sb_inodesize >= 256 && sbp->sb_inodesize <= 2048);
 607         switch (sbp->sb_inodesize) {
 608         case 256:
 609                 mp->m_attroffset = XFS_LITINO(mp) -
 610                                    XFS_BMDR_SPACE_CALC(MINABTPTRS);
 611                 break;
 612         case 512:
 613         case 1024:
 614         case 2048:
 615                 mp->m_attroffset = XFS_BMDR_SPACE_CALC(6 * MINABTPTRS);
 616                 break;
 617         default:
 618                 ASSERT(0);
 619         }
 620         ASSERT(mp->m_attroffset < XFS_LITINO(mp));
 621
 622         for (i = 0; i < 2; i++) {
 623                 mp->m_alloc_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 624                         xfs_alloc, i == 0);
 625                 mp->m_alloc_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 626                         xfs_alloc, i == 0);
 627         }
 628         for (i = 0; i < 2; i++) {
 629                 mp->m_bmap_dmxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 630                         xfs_bmbt, i == 0);
 631                 mp->m_bmap_dmnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 632                         xfs_bmbt, i == 0);
 633         }
 634         for (i = 0; i < 2; i++) {
 635                 mp->m_inobt_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 636                         xfs_inobt, i == 0);
 637                 mp->m_inobt_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 638                         xfs_inobt, i == 0);
 639         }
 640
 641         mp->m_bsize = XFS_FSB_TO_BB(mp, 1);
 642         mp->m_ialloc_inos = (int)MAX((__uint16_t)XFS_INODES_PER_CHUNK,
 643                                         sbp->sb_inopblock);
 644         mp->m_ialloc_blks = mp->m_ialloc_inos >> sbp->sb_inopblog;
 645 }
 646 /*
 647  * xfs_mountfs
 648  *
 649  * This function does the following on an initial mount of a file system:
 650  *      - reads the superblock from disk and init the mount struct
 651  *      - if we're a 32-bit kernel, do a size check on the superblock
 652  *              so we don't mount terabyte filesystems
 653  *      - init mount struct realtime fields
 654  *      - allocate inode hash table for fs
 655  *      - init directory manager
 656  *      - perform recovery and init the log manager
 657  */
 658 int
 659 xfs_mountfs(
 660         bhv_vfs_t       *vfsp,
 661         xfs_mount_t     *mp,
 662         int             mfsi_flags)
 663 {
 664         xfs_buf_t       *bp;
 665         xfs_sb_t        *sbp = &(mp->m_sb);
 666         xfs_inode_t     *rip;
 667         bhv_vnode_t     *rvp = NULL;
 668         int             readio_log, writeio_log;
 669         xfs_daddr_t     d;
 670         __uint64_t      ret64;
 671         __int64_t       update_flags;
 672         uint            quotamount, quotaflags;
 673         int             agno;
 674         int             uuid_mounted = 0;
 675         int             error = 0;
 676
 677         if (mp->m_sb_bp == NULL) {
 678                 if ((error = xfs_readsb(mp, mfsi_flags))) {
 679                         return error;
 680                 }
 681         }
 682         xfs_mount_common(mp, sbp);
 683
 684         /*
 685          * Check if sb_agblocks is aligned at stripe boundary
 686          * If sb_agblocks is NOT aligned turn off m_dalign since
 687          * allocator alignment is within an ag, therefore ag has
 688          * to be aligned at stripe boundary.
 689          */
 690         update_flags = 0LL;
 691         if (mp->m_dalign && !(mfsi_flags & XFS_MFSI_SECOND)) {
 692                 /*
 693                  * If stripe unit and stripe width are not multiples
 694                  * of the fs blocksize turn off alignment.
 695                  */
 696                 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
 697                     (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
 698                         if (mp->m_flags & XFS_MOUNT_RETERR) {
 699                                 cmn_err(CE_WARN,
 700                                         "XFS: alignment check 1 failed");
 701                                 error = XFS_ERROR(EINVAL);
 702                                 goto error1;
 703                         }
 704                         mp->m_dalign = mp->m_swidth = 0;
 705                 } else {
 706                         /*
 707                          * Convert the stripe unit and width to FSBs.
 708                          */
 709                         mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
 710                         if (mp->m_dalign && (sbp->sb_agblocks % mp->m_dalign)) {
 711                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
 712                                         error = XFS_ERROR(EINVAL);
 713                                         goto error1;
 714                                 }
 715                                 xfs_fs_cmn_err(CE_WARN, mp,
 716 "stripe alignment turned off: sunit(%d)/swidth(%d) incompatible with agsize(%d)",
 717                                         mp->m_dalign, mp->m_swidth,
 718                                         sbp->sb_agblocks);
 719
 720                                 mp->m_dalign = 0;
 721                                 mp->m_swidth = 0;
 722                         } else if (mp->m_dalign) {
 723                                 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
 724                         } else {
 725                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
 726                                         xfs_fs_cmn_err(CE_WARN, mp,
 727 "stripe alignment turned off: sunit(%d) less than bsize(%d)",
 728                                                 mp->m_dalign,
 729                                                 mp->m_blockmask +1);
 730                                         error = XFS_ERROR(EINVAL);
 731                                         goto error1;
 732                                 }
 733                                 mp->m_swidth = 0;
 734                         }
 735                 }
 736
 737                 /*
 738                  * Update superblock with new values
 739                  * and log changes
 740                  */
 741                 if (XFS_SB_VERSION_HASDALIGN(sbp)) {
 742                         if (sbp->sb_unit != mp->m_dalign) {
 743                                 sbp->sb_unit = mp->m_dalign;
 744                                 update_flags |= XFS_SB_UNIT;
 745                         }
 746                         if (sbp->sb_width != mp->m_swidth) {
 747                                 sbp->sb_width = mp->m_swidth;
 748                                 update_flags |= XFS_SB_WIDTH;
 749                         }
 750                 }
 751         } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
 752                     XFS_SB_VERSION_HASDALIGN(&mp->m_sb)) {
 753                         mp->m_dalign = sbp->sb_unit;
 754                         mp->m_swidth = sbp->sb_width;
 755         }
 756
 757         xfs_alloc_compute_maxlevels(mp);
 758         xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
 759         xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
 760         xfs_ialloc_compute_maxlevels(mp);
 761
 762         if (sbp->sb_imax_pct) {
 763                 __uint64_t      icount;
 764
 765                 /* Make sure the maximum inode count is a multiple of the
 766                  * units we allocate inodes in.
 767                  */
 768
 769                 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
 770                 do_div(icount, 100);
 771                 do_div(icount, mp->m_ialloc_blks);
 772                 mp->m_maxicount = (icount * mp->m_ialloc_blks)  <<
 773                                    sbp->sb_inopblog;
 774         } else
 775                 mp->m_maxicount = 0;
 776
 777         mp->m_maxioffset = xfs_max_file_offset(sbp->sb_blocklog);
 778
 779         /*
 780          * XFS uses the uuid from the superblock as the unique
 781          * identifier for fsid.  We can not use the uuid from the volume
 782          * since a single partition filesystem is identical to a single
 783          * partition volume/filesystem.
 784          */
 785         if ((mfsi_flags & XFS_MFSI_SECOND) == 0 &&
 786             (mp->m_flags & XFS_MOUNT_NOUUID) == 0) {
 787                 if (xfs_uuid_mount(mp)) {
 788                         error = XFS_ERROR(EINVAL);
 789                         goto error1;
 790                 }
 791                 uuid_mounted=1;
 792                 ret64 = uuid_hash64(&sbp->sb_uuid);
 793                 memcpy(&vfsp->vfs_fsid, &ret64, sizeof(ret64));
 794         }
 795
 796         /*
 797          * Set the default minimum read and write sizes unless
 798          * already specified in a mount option.
 799          * We use smaller I/O sizes when the file system
 800          * is being used for NFS service (wsync mount option).
 801          */
 802         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
 803                 if (mp->m_flags & XFS_MOUNT_WSYNC) {
 804                         readio_log = XFS_WSYNC_READIO_LOG;
 805                         writeio_log = XFS_WSYNC_WRITEIO_LOG;
 806                 } else {
 807                         readio_log = XFS_READIO_LOG_LARGE;
 808                         writeio_log = XFS_WRITEIO_LOG_LARGE;
 809                 }
 810         } else {
 811                 readio_log = mp->m_readio_log;
 812                 writeio_log = mp->m_writeio_log;
 813         }
 814
 815         /*
 816          * Set the number of readahead buffers to use based on
 817          * physical memory size.
 818          */
 819         if (xfs_physmem <= 4096)                /* <= 16MB */
 820                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_16MB;
 821         else if (xfs_physmem <= 8192)   /* <= 32MB */
 822                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_32MB;
 823         else
 824                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_K32;
 825         if (sbp->sb_blocklog > readio_log) {
 826                 mp->m_readio_log = sbp->sb_blocklog;
 827         } else {
 828                 mp->m_readio_log = readio_log;
 829         }
 830         mp->m_readio_blocks = 1 << (mp->m_readio_log - sbp->sb_blocklog);
 831         if (sbp->sb_blocklog > writeio_log) {
 832                 mp->m_writeio_log = sbp->sb_blocklog;
 833         } else {
 834                 mp->m_writeio_log = writeio_log;
 835         }
 836         mp->m_writeio_blocks = 1 << (mp->m_writeio_log - sbp->sb_blocklog);
 837
 838         /*
 839          * Set the inode cluster size based on the physical memory
 840          * size.  This may still be overridden by the file system
 841          * block size if it is larger than the chosen cluster size.
 842          */
 843         if (xfs_physmem <= btoc(32 * 1024 * 1024)) { /* <= 32 MB */
 844                 mp->m_inode_cluster_size = XFS_INODE_SMALL_CLUSTER_SIZE;
 845         } else {
 846                 mp->m_inode_cluster_size = XFS_INODE_BIG_CLUSTER_SIZE;
 847         }
 848         /*
 849          * Set whether we're using inode alignment.
 850          */
 851         if (XFS_SB_VERSION_HASALIGN(&mp->m_sb) &&
 852             mp->m_sb.sb_inoalignmt >=
 853             XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size))
 854                 mp->m_inoalign_mask = mp->m_sb.sb_inoalignmt - 1;
 855         else
 856                 mp->m_inoalign_mask = 0;
 857         /*
 858          * If we are using stripe alignment, check whether
 859          * the stripe unit is a multiple of the inode alignment
 860          */
 861         if (mp->m_dalign && mp->m_inoalign_mask &&
 862             !(mp->m_dalign & mp->m_inoalign_mask))
 863                 mp->m_sinoalign = mp->m_dalign;
 864         else
 865                 mp->m_sinoalign = 0;
 866         /*
 867          * Check that the data (and log if separate) are an ok size.
 868          */
 869         d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
 870         if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
 871                 cmn_err(CE_WARN, "XFS: size check 1 failed");
 872                 error = XFS_ERROR(E2BIG);
 873                 goto error1;
 874         }
 875         error = xfs_read_buf(mp, mp->m_ddev_targp,
 876                              d - XFS_FSS_TO_BB(mp, 1),
 877                              XFS_FSS_TO_BB(mp, 1), 0, &bp);
 878         if (!error) {
 879                 xfs_buf_relse(bp);
 880         } else {
 881                 cmn_err(CE_WARN, "XFS: size check 2 failed");
 882                 if (error == ENOSPC) {
 883                         error = XFS_ERROR(E2BIG);
 884                 }
 885                 goto error1;
 886         }
 887
 888         if (((mfsi_flags & XFS_MFSI_CLIENT) == 0) &&
 889             mp->m_logdev_targp != mp->m_ddev_targp) {
 890                 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
 891                 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
 892                         cmn_err(CE_WARN, "XFS: size check 3 failed");
 893                         error = XFS_ERROR(E2BIG);
 894                         goto error1;
 895                 }
 896                 error = xfs_read_buf(mp, mp->m_logdev_targp,
 897                                      d - XFS_FSB_TO_BB(mp, 1),
 898                                      XFS_FSB_TO_BB(mp, 1), 0, &bp);
 899                 if (!error) {
 900                         xfs_buf_relse(bp);
 901                 } else {
 902                         cmn_err(CE_WARN, "XFS: size check 3 failed");
 903                         if (error == ENOSPC) {
 904                                 error = XFS_ERROR(E2BIG);
 905                         }
 906                         goto error1;
 907                 }
 908         }
 909
 910         /*
 911          * Initialize realtime fields in the mount structure
 912          */
 913         if ((error = xfs_rtmount_init(mp))) {
 914                 cmn_err(CE_WARN, "XFS: RT mount failed");
 915                 goto error1;
 916         }
 917
 918         /*
 919          * For client case we are done now
 920          */
 921         if (mfsi_flags & XFS_MFSI_CLIENT) {
 922                 return 0;
 923         }
 924
 925         /*
 926          *  Copies the low order bits of the timestamp and the randomly
 927          *  set "sequence" number out of a UUID.
 928          */
 929         uuid_getnodeuniq(&sbp->sb_uuid, mp->m_fixedfsid);
 930
 931         /*
 932          *  The vfs structure needs to have a file system independent
 933          *  way of checking for the invariant file system ID.  Since it
 934          *  can't look at mount structures it has a pointer to the data
 935          *  in the mount structure.
 936          *
 937          *  File systems that don't support user level file handles (i.e.
 938          *  all of them except for XFS) will leave vfs_altfsid as NULL.
 939          */
 940         vfsp->vfs_altfsid = (xfs_fsid_t *)mp->m_fixedfsid;
 941         mp->m_dmevmask = 0;     /* not persistent; set after each mount */
 942
 943         xfs_dir_mount(mp);
 944
 945         /*
 946          * Initialize the attribute manager's entries.
 947          */
 948         mp->m_attr_magicpct = (mp->m_sb.sb_blocksize * 37) / 100;
 949
 950         /*
 951          * Initialize the precomputed transaction reservations values.
 952          */
 953         xfs_trans_init(mp);
 954
 955         /*
 956          * Allocate and initialize the inode hash table for this
 957          * file system.
 958          */
 959         xfs_ihash_init(mp);
 960         xfs_chash_init(mp);
 961
 962         /*
 963          * Allocate and initialize the per-ag data.
 964          */
 965         init_rwsem(&mp->m_peraglock);
 966         mp->m_perag =
 967                 kmem_zalloc(sbp->sb_agcount * sizeof(xfs_perag_t), KM_SLEEP);
 968
 969         mp->m_maxagi = xfs_initialize_perag(vfsp, mp, sbp->sb_agcount);
 970
 971         /*
 972          * log's mount-time initialization. Perform 1st part recovery if needed
 973          */
 974         if (likely(sbp->sb_logblocks > 0)) {    /* check for volume case */
 975                 error = xfs_log_mount(mp, mp->m_logdev_targp,
 976                                       XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
 977                                       XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
 978                 if (error) {
 979                         cmn_err(CE_WARN, "XFS: log mount failed");
 980                         goto error2;
 981                 }
 982         } else {        /* No log has been defined */
 983                 cmn_err(CE_WARN, "XFS: no log defined");
 984                 XFS_ERROR_REPORT("xfs_mountfs_int(1)", XFS_ERRLEVEL_LOW, mp);
 985                 error = XFS_ERROR(EFSCORRUPTED);
 986                 goto error2;
 987         }
 988
 989         /*
 990          * Get and sanity-check the root inode.
 991          * Save the pointer to it in the mount structure.
 992          */
 993         error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip, 0);
 994         if (error) {
 995                 cmn_err(CE_WARN, "XFS: failed to read root inode");
 996                 goto error3;
 997         }
 998
 999         ASSERT(rip != NULL);
1000         rvp = XFS_ITOV(rip);
1001
1002         if (unlikely((rip->i_d.di_mode & S_IFMT) != S_IFDIR)) {
1003                 cmn_err(CE_WARN, "XFS: corrupted root inode");
1004                 cmn_err(CE_WARN, "Device %s - root %llu is not a directory",
1005                         XFS_BUFTARG_NAME(mp->m_ddev_targp),
1006                         (unsigned long long)rip->i_ino);
1007                 xfs_iunlock(rip, XFS_ILOCK_EXCL);
1008                 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
1009                                  mp);
1010                 error = XFS_ERROR(EFSCORRUPTED);
1011                 goto error4;
1012         }
1013         mp->m_rootip = rip;     /* save it */
1014
1015         xfs_iunlock(rip, XFS_ILOCK_EXCL);
1016
1017         /*
1018          * Initialize realtime inode pointers in the mount structure
1019          */
1020         if ((error = xfs_rtmount_inodes(mp))) {
1021                 /*
1022                  * Free up the root inode.
1023                  */
1024                 cmn_err(CE_WARN, "XFS: failed to read RT inodes");
1025                 goto error4;
1026         }
1027
1028         /*
1029          * If fs is not mounted readonly, then update the superblock
1030          * unit and width changes.
1031          */
1032         if (update_flags && !(vfsp->vfs_flag & VFS_RDONLY))
1033                 xfs_mount_log_sbunit(mp, update_flags);
1034
1035         /*
1036          * Initialise the XFS quota management subsystem for this mount
1037          */
1038         if ((error = XFS_QM_INIT(mp, &quotamount, &quotaflags)))
1039                 goto error4;
1040
1041         /*
1042          * Finish recovering the file system.  This part needed to be
1043          * delayed until after the root and real-time bitmap inodes
1044          * were consistently read in.
1045          */
1046         error = xfs_log_mount_finish(mp, mfsi_flags);
1047         if (error) {
1048                 cmn_err(CE_WARN, "XFS: log mount finish failed");
1049                 goto error4;
1050         }
1051
1052         /*
1053          * Complete the quota initialisation, post-log-replay component.
1054          */
1055         if ((error = XFS_QM_MOUNT(mp, quotamount, quotaflags, mfsi_flags)))
1056                 goto error4;
1057
1058         return 0;
1059
1060  error4:
1061         /*
1062          * Free up the root inode.
1063          */
1064         VN_RELE(rvp);
1065  error3:
1066         xfs_log_unmount_dealloc(mp);
1067  error2:
1068         xfs_ihash_free(mp);
1069         xfs_chash_free(mp);
1070         for (agno = 0; agno < sbp->sb_agcount; agno++)
1071                 if (mp->m_perag[agno].pagb_list)
1072                         kmem_free(mp->m_perag[agno].pagb_list,
1073                           sizeof(xfs_perag_busy_t) * XFS_PAGB_NUM_SLOTS);
1074         kmem_free(mp->m_perag, sbp->sb_agcount * sizeof(xfs_perag_t));
1075         mp->m_perag = NULL;
1076         /* FALLTHROUGH */
1077  error1:
1078         if (uuid_mounted)
1079                 xfs_uuid_unmount(mp);
1080         xfs_freesb(mp);
1081         return error;
1082 }
1083
1084 /*
1085  * xfs_unmountfs
1086  *
1087  * This flushes out the inodes,dquots and the superblock, unmounts the
1088  * log and makes sure that incore structures are freed.
1089  */
1090 int
1091 xfs_unmountfs(xfs_mount_t *mp, struct cred *cr)
1092 {
1093         struct bhv_vfs  *vfsp = XFS_MTOVFS(mp);
1094 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1095         int64_t         fsid;
1096 #endif
1097
1098         xfs_iflush_all(mp);
1099
1100         XFS_QM_DQPURGEALL(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_UMOUNTING);
1101
1102         /*
1103          * Flush out the log synchronously so that we know for sure
1104          * that nothing is pinned.  This is important because bflush()
1105          * will skip pinned buffers.
1106          */
1107         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1108
1109         xfs_binval(mp->m_ddev_targp);
1110         if (mp->m_rtdev_targp) {
1111                 xfs_binval(mp->m_rtdev_targp);
1112         }
1113
1114         xfs_unmountfs_writesb(mp);
1115
1116         xfs_unmountfs_wait(mp);                 /* wait for async bufs */
1117
1118         xfs_log_unmount(mp);                    /* Done! No more fs ops. */
1119
1120         xfs_freesb(mp);
1121
1122         /*
1123          * All inodes from this mount point should be freed.
1124          */
1125         ASSERT(mp->m_inodes == NULL);
1126
1127         xfs_unmountfs_close(mp, cr);
1128         if ((mp->m_flags & XFS_MOUNT_NOUUID) == 0)
1129                 xfs_uuid_unmount(mp);
1130
1131 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1132         /*
1133          * clear all error tags on this filesystem
1134          */
1135         memcpy(&fsid, &vfsp->vfs_fsid, sizeof(int64_t));
1136         xfs_errortag_clearall_umount(fsid, mp->m_fsname, 0);
1137 #endif
1138         XFS_IODONE(vfsp);
1139         xfs_mount_free(mp, 1);
1140         return 0;
1141 }
1142
1143 void
1144 xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
1145 {
1146         if (mp->m_logdev_targp != mp->m_ddev_targp)
1147                 xfs_free_buftarg(mp->m_logdev_targp, 1);
1148         if (mp->m_rtdev_targp)
1149                 xfs_free_buftarg(mp->m_rtdev_targp, 1);
1150         xfs_free_buftarg(mp->m_ddev_targp, 0);
1151 }
1152
1153 STATIC void
1154 xfs_unmountfs_wait(xfs_mount_t *mp)
1155 {
1156         if (mp->m_logdev_targp != mp->m_ddev_targp)
1157                 xfs_wait_buftarg(mp->m_logdev_targp);
1158         if (mp->m_rtdev_targp)
1159                 xfs_wait_buftarg(mp->m_rtdev_targp);
1160         xfs_wait_buftarg(mp->m_ddev_targp);
1161 }
1162
1163 int
1164 xfs_unmountfs_writesb(xfs_mount_t *mp)
1165 {
1166         xfs_buf_t       *sbp;
1167         xfs_sb_t        *sb;
1168         int             error = 0;
1169
1170         /*
1171          * skip superblock write if fs is read-only, or
1172          * if we are doing a forced umount.
1173          */
1174         sbp = xfs_getsb(mp, 0);
1175         if (!(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY ||
1176                 XFS_FORCED_SHUTDOWN(mp))) {
1177
1178                 xfs_icsb_sync_counters(mp);
1179
1180                 /*
1181                  * mark shared-readonly if desired
1182                  */
1183                 sb = XFS_BUF_TO_SBP(sbp);
1184                 if (mp->m_mk_sharedro) {
1185                         if (!(sb->sb_flags & XFS_SBF_READONLY))
1186                                 sb->sb_flags |= XFS_SBF_READONLY;
1187                         if (!XFS_SB_VERSION_HASSHARED(sb))
1188                                 XFS_SB_VERSION_ADDSHARED(sb);
1189                         xfs_fs_cmn_err(CE_NOTE, mp,
1190                                 "Unmounting, marking shared read-only");
1191                 }
1192                 XFS_BUF_UNDONE(sbp);
1193                 XFS_BUF_UNREAD(sbp);
1194                 XFS_BUF_UNDELAYWRITE(sbp);
1195                 XFS_BUF_WRITE(sbp);
1196                 XFS_BUF_UNASYNC(sbp);
1197                 ASSERT(XFS_BUF_TARGET(sbp) == mp->m_ddev_targp);
1198                 xfsbdstrat(mp, sbp);
1199                 /* Nevermind errors we might get here. */
1200                 error = xfs_iowait(sbp);
1201                 if (error)
1202                         xfs_ioerror_alert("xfs_unmountfs_writesb",
1203                                           mp, sbp, XFS_BUF_ADDR(sbp));
1204                 if (error && mp->m_mk_sharedro)
1205                         xfs_fs_cmn_err(CE_ALERT, mp, "Superblock write error detected while unmounting.  Filesystem may not be marked shared readonly");
1206         }
1207         xfs_buf_relse(sbp);
1208         return error;
1209 }
1210
1211 /*
1212  * xfs_mod_sb() can be used to copy arbitrary changes to the
1213  * in-core superblock into the superblock buffer to be logged.
1214  * It does not provide the higher level of locking that is
1215  * needed to protect the in-core superblock from concurrent
1216  * access.
1217  */
1218 void
1219 xfs_mod_sb(xfs_trans_t *tp, __int64_t fields)
1220 {
1221         xfs_buf_t       *bp;
1222         int             first;
1223         int             last;
1224         xfs_mount_t     *mp;
1225         xfs_sb_t        *sbp;
1226         xfs_sb_field_t  f;
1227
1228         ASSERT(fields);
1229         if (!fields)
1230                 return;
1231         mp = tp->t_mountp;
1232         bp = xfs_trans_getsb(tp, mp, 0);
1233         sbp = XFS_BUF_TO_SBP(bp);
1234         first = sizeof(xfs_sb_t);
1235         last = 0;
1236
1237         /* translate/copy */
1238
1239         xfs_xlatesb(XFS_BUF_PTR(bp), &(mp->m_sb), -1, fields);
1240
1241         /* find modified range */
1242
1243         f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
1244         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1245         first = xfs_sb_info[f].offset;
1246
1247         f = (xfs_sb_field_t)xfs_highbit64((__uint64_t)fields);
1248         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1249         last = xfs_sb_info[f + 1].offset - 1;
1250
1251         xfs_trans_log_buf(tp, bp, first, last);
1252 }
1253
1254
1255 /*
1256  * xfs_mod_incore_sb_unlocked() is a utility routine common used to apply
1257  * a delta to a specified field in the in-core superblock.  Simply
1258  * switch on the field indicated and apply the delta to that field.
1259  * Fields are not allowed to dip below zero, so if the delta would
1260  * do this do not apply it and return EINVAL.
1261  *
1262  * The SB_LOCK must be held when this routine is called.
1263  */
1264 int
1265 xfs_mod_incore_sb_unlocked(
1266         xfs_mount_t     *mp,
1267         xfs_sb_field_t  field,
1268         int64_t         delta,
1269         int             rsvd)
1270 {
1271         int             scounter;       /* short counter for 32 bit fields */
1272         long long       lcounter;       /* long counter for 64 bit fields */
1273         long long       res_used, rem;
1274
1275         /*
1276          * With the in-core superblock spin lock held, switch
1277          * on the indicated field.  Apply the delta to the
1278          * proper field.  If the fields value would dip below
1279          * 0, then do not apply the delta and return EINVAL.
1280          */
1281         switch (field) {
1282         case XFS_SBS_ICOUNT:
1283                 lcounter = (long long)mp->m_sb.sb_icount;
1284                 lcounter += delta;
1285                 if (lcounter < 0) {
1286                         ASSERT(0);
1287                         return XFS_ERROR(EINVAL);
1288                 }
1289                 mp->m_sb.sb_icount = lcounter;
1290                 return 0;
1291         case XFS_SBS_IFREE:
1292                 lcounter = (long long)mp->m_sb.sb_ifree;
1293                 lcounter += delta;
1294                 if (lcounter < 0) {
1295                         ASSERT(0);
1296                         return XFS_ERROR(EINVAL);
1297                 }
1298                 mp->m_sb.sb_ifree = lcounter;
1299                 return 0;
1300         case XFS_SBS_FDBLOCKS:
1301                 lcounter = (long long)
1302                         mp->m_sb.sb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
1303                 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1304
1305                 if (delta > 0) {                /* Putting blocks back */
1306                         if (res_used > delta) {
1307                                 mp->m_resblks_avail += delta;
1308                         } else {
1309                                 rem = delta - res_used;
1310                                 mp->m_resblks_avail = mp->m_resblks;
1311                                 lcounter += rem;
1312                         }
1313                 } else {                                /* Taking blocks away */
1314
1315                         lcounter += delta;
1316
1317                 /*
1318                  * If were out of blocks, use any available reserved blocks if
1319                  * were allowed to.
1320                  */
1321
1322                         if (lcounter < 0) {
1323                                 if (rsvd) {
1324                                         lcounter = (long long)mp->m_resblks_avail + delta;
1325                                         if (lcounter < 0) {
1326                                                 return XFS_ERROR(ENOSPC);
1327                                         }
1328                                         mp->m_resblks_avail = lcounter;
1329                                         return 0;
1330                                 } else {        /* not reserved */
1331                                         return XFS_ERROR(ENOSPC);
1332                                 }
1333                         }
1334                 }
1335
1336                 mp->m_sb.sb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
1337                 return 0;
1338         case XFS_SBS_FREXTENTS:
1339                 lcounter = (long long)mp->m_sb.sb_frextents;
1340                 lcounter += delta;
1341                 if (lcounter < 0) {
1342                         return XFS_ERROR(ENOSPC);
1343                 }
1344                 mp->m_sb.sb_frextents = lcounter;
1345                 return 0;
1346         case XFS_SBS_DBLOCKS:
1347                 lcounter = (long long)mp->m_sb.sb_dblocks;
1348                 lcounter += delta;
1349                 if (lcounter < 0) {
1350                         ASSERT(0);
1351                         return XFS_ERROR(EINVAL);
1352                 }
1353                 mp->m_sb.sb_dblocks = lcounter;
1354                 return 0;
1355         case XFS_SBS_AGCOUNT:
1356                 scounter = mp->m_sb.sb_agcount;
1357                 scounter += delta;
1358                 if (scounter < 0) {
1359                         ASSERT(0);
1360                         return XFS_ERROR(EINVAL);
1361                 }
1362                 mp->m_sb.sb_agcount = scounter;
1363                 return 0;
1364         case XFS_SBS_IMAX_PCT:
1365                 scounter = mp->m_sb.sb_imax_pct;
1366                 scounter += delta;
1367                 if (scounter < 0) {
1368                         ASSERT(0);
1369                         return XFS_ERROR(EINVAL);
1370                 }
1371                 mp->m_sb.sb_imax_pct = scounter;
1372                 return 0;
1373         case XFS_SBS_REXTSIZE:
1374                 scounter = mp->m_sb.sb_rextsize;
1375                 scounter += delta;
1376                 if (scounter < 0) {
1377                         ASSERT(0);
1378                         return XFS_ERROR(EINVAL);
1379                 }
1380                 mp->m_sb.sb_rextsize = scounter;
1381                 return 0;
1382         case XFS_SBS_RBMBLOCKS:
1383                 scounter = mp->m_sb.sb_rbmblocks;
1384                 scounter += delta;
1385                 if (scounter < 0) {
1386                         ASSERT(0);
1387                         return XFS_ERROR(EINVAL);
1388                 }
1389                 mp->m_sb.sb_rbmblocks = scounter;
1390                 return 0;
1391         case XFS_SBS_RBLOCKS:
1392                 lcounter = (long long)mp->m_sb.sb_rblocks;
1393                 lcounter += delta;
1394                 if (lcounter < 0) {
1395                         ASSERT(0);
1396                         return XFS_ERROR(EINVAL);
1397                 }
1398                 mp->m_sb.sb_rblocks = lcounter;
1399                 return 0;
1400         case XFS_SBS_REXTENTS:
1401                 lcounter = (long long)mp->m_sb.sb_rextents;
1402                 lcounter += delta;
1403                 if (lcounter < 0) {
1404                         ASSERT(0);
1405                         return XFS_ERROR(EINVAL);
1406                 }
1407                 mp->m_sb.sb_rextents = lcounter;
1408                 return 0;
1409         case XFS_SBS_REXTSLOG:
1410                 scounter = mp->m_sb.sb_rextslog;
1411                 scounter += delta;
1412                 if (scounter < 0) {
1413                         ASSERT(0);
1414                         return XFS_ERROR(EINVAL);
1415                 }
1416                 mp->m_sb.sb_rextslog = scounter;
1417                 return 0;
1418         default:
1419                 ASSERT(0);
1420                 return XFS_ERROR(EINVAL);
1421         }
1422 }
1423
1424 /*
1425  * xfs_mod_incore_sb() is used to change a field in the in-core
1426  * superblock structure by the specified delta.  This modification
1427  * is protected by the SB_LOCK.  Just use the xfs_mod_incore_sb_unlocked()
1428  * routine to do the work.
1429  */
1430 int
1431 xfs_mod_incore_sb(
1432         xfs_mount_t     *mp,
1433         xfs_sb_field_t  field,
1434         int64_t         delta,
1435         int             rsvd)
1436 {
1437         unsigned long   s;
1438         int     status;
1439
1440         /* check for per-cpu counters */
1441         switch (field) {
1442 #ifdef HAVE_PERCPU_SB
1443         case XFS_SBS_ICOUNT:
1444         case XFS_SBS_IFREE:
1445         case XFS_SBS_FDBLOCKS:
1446                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1447                         status = xfs_icsb_modify_counters(mp, field,
1448                                                         delta, rsvd);
1449                         break;
1450                 }
1451                 /* FALLTHROUGH */
1452 #endif
1453         default:
1454                 s = XFS_SB_LOCK(mp);
1455                 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1456                 XFS_SB_UNLOCK(mp, s);
1457                 break;
1458         }
1459
1460         return status;
1461 }
1462
1463 /*
1464  * xfs_mod_incore_sb_batch() is used to change more than one field
1465  * in the in-core superblock structure at a time.  This modification
1466  * is protected by a lock internal to this module.  The fields and
1467  * changes to those fields are specified in the array of xfs_mod_sb
1468  * structures passed in.
1469  *
1470  * Either all of the specified deltas will be applied or none of
1471  * them will.  If any modified field dips below 0, then all modifications
1472  * will be backed out and EINVAL will be returned.
1473  */
1474 int
1475 xfs_mod_incore_sb_batch(xfs_mount_t *mp, xfs_mod_sb_t *msb, uint nmsb, int rsvd)
1476 {
1477         unsigned long   s;
1478         int             status=0;
1479         xfs_mod_sb_t    *msbp;
1480
1481         /*
1482          * Loop through the array of mod structures and apply each
1483          * individually.  If any fail, then back out all those
1484          * which have already been applied.  Do all of this within
1485          * the scope of the SB_LOCK so that all of the changes will
1486          * be atomic.
1487          */
1488         s = XFS_SB_LOCK(mp);
1489         msbp = &msb[0];
1490         for (msbp = &msbp[0]; msbp < (msb + nmsb); msbp++) {
1491                 /*
1492                  * Apply the delta at index n.  If it fails, break
1493                  * from the loop so we'll fall into the undo loop
1494                  * below.
1495                  */
1496                 switch (msbp->msb_field) {
1497 #ifdef HAVE_PERCPU_SB
1498                 case XFS_SBS_ICOUNT:
1499                 case XFS_SBS_IFREE:
1500                 case XFS_SBS_FDBLOCKS:
1501                         if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1502                                 XFS_SB_UNLOCK(mp, s);
1503                                 status = xfs_icsb_modify_counters(mp,
1504                                                         msbp->msb_field,
1505                                                         msbp->msb_delta, rsvd);
1506                                 s = XFS_SB_LOCK(mp);
1507                                 break;
1508                         }
1509                         /* FALLTHROUGH */
1510 #endif
1511                 default:
1512                         status = xfs_mod_incore_sb_unlocked(mp,
1513                                                 msbp->msb_field,
1514                                                 msbp->msb_delta, rsvd);
1515                         break;
1516                 }
1517
1518                 if (status != 0) {
1519                         break;
1520                 }
1521         }
1522
1523         /*
1524          * If we didn't complete the loop above, then back out
1525          * any changes made to the superblock.  If you add code
1526          * between the loop above and here, make sure that you
1527          * preserve the value of status. Loop back until
1528          * we step below the beginning of the array.  Make sure
1529          * we don't touch anything back there.
1530          */
1531         if (status != 0) {
1532                 msbp--;
1533                 while (msbp >= msb) {
1534                         switch (msbp->msb_field) {
1535 #ifdef HAVE_PERCPU_SB
1536                         case XFS_SBS_ICOUNT:
1537                         case XFS_SBS_IFREE:
1538                         case XFS_SBS_FDBLOCKS:
1539                                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1540                                         XFS_SB_UNLOCK(mp, s);
1541                                         status = xfs_icsb_modify_counters(mp,
1542                                                         msbp->msb_field,
1543                                                         -(msbp->msb_delta),
1544                                                         rsvd);
1545                                         s = XFS_SB_LOCK(mp);
1546                                         break;
1547                                 }
1548                                 /* FALLTHROUGH */
1549 #endif
1550                         default:
1551                                 status = xfs_mod_incore_sb_unlocked(mp,
1552                                                         msbp->msb_field,
1553                                                         -(msbp->msb_delta),
1554                                                         rsvd);
1555                                 break;
1556                         }
1557                         ASSERT(status == 0);
1558                         msbp--;
1559                 }
1560         }
1561         XFS_SB_UNLOCK(mp, s);
1562         return status;
1563 }
1564
1565 /*
1566  * xfs_getsb() is called to obtain the buffer for the superblock.
1567  * The buffer is returned locked and read in from disk.
1568  * The buffer should be released with a call to xfs_brelse().
1569  *
1570  * If the flags parameter is BUF_TRYLOCK, then we'll only return
1571  * the superblock buffer if it can be locked without sleeping.
1572  * If it can't then we'll return NULL.
1573  */
1574 xfs_buf_t *
1575 xfs_getsb(
1576         xfs_mount_t     *mp,
1577         int             flags)
1578 {
1579         xfs_buf_t       *bp;
1580
1581         ASSERT(mp->m_sb_bp != NULL);
1582         bp = mp->m_sb_bp;
1583         if (flags & XFS_BUF_TRYLOCK) {
1584                 if (!XFS_BUF_CPSEMA(bp)) {
1585                         return NULL;
1586                 }
1587         } else {
1588                 XFS_BUF_PSEMA(bp, PRIBIO);
1589         }
1590         XFS_BUF_HOLD(bp);
1591         ASSERT(XFS_BUF_ISDONE(bp));
1592         return bp;
1593 }
1594
1595 /*
1596  * Used to free the superblock along various error paths.
1597  */
1598 void
1599 xfs_freesb(
1600         xfs_mount_t     *mp)
1601 {
1602         xfs_buf_t       *bp;
1603
1604         /*
1605          * Use xfs_getsb() so that the buffer will be locked
1606          * when we call xfs_buf_relse().
1607          */
1608         bp = xfs_getsb(mp, 0);
1609         XFS_BUF_UNMANAGE(bp);
1610         xfs_buf_relse(bp);
1611         mp->m_sb_bp = NULL;
1612 }
1613
1614 /*
1615  * See if the UUID is unique among mounted XFS filesystems.
1616  * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
1617  */
1618 STATIC int
1619 xfs_uuid_mount(
1620         xfs_mount_t     *mp)
1621 {
1622         if (uuid_is_nil(&mp->m_sb.sb_uuid)) {
1623                 cmn_err(CE_WARN,
1624                         "XFS: Filesystem %s has nil UUID - can't mount",
1625                         mp->m_fsname);
1626                 return -1;
1627         }
1628         if (!uuid_table_insert(&mp->m_sb.sb_uuid)) {
1629                 cmn_err(CE_WARN,
1630                         "XFS: Filesystem %s has duplicate UUID - can't mount",
1631                         mp->m_fsname);
1632                 return -1;
1633         }
1634         return 0;
1635 }
1636
1637 /*
1638  * Remove filesystem from the UUID table.
1639  */
1640 STATIC void
1641 xfs_uuid_unmount(
1642         xfs_mount_t     *mp)
1643 {
1644         uuid_table_remove(&mp->m_sb.sb_uuid);
1645 }
1646
1647 /*
1648  * Used to log changes to the superblock unit and width fields which could
1649  * be altered by the mount options. Only the first superblock is updated.
1650  */
1651 STATIC void
1652 xfs_mount_log_sbunit(
1653         xfs_mount_t     *mp,
1654         __int64_t       fields)
1655 {
1656         xfs_trans_t     *tp;
1657
1658         ASSERT(fields & (XFS_SB_UNIT|XFS_SB_WIDTH|XFS_SB_UUID));
1659
1660         tp = xfs_trans_alloc(mp, XFS_TRANS_SB_UNIT);
1661         if (xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1662                                 XFS_DEFAULT_LOG_COUNT)) {
1663                 xfs_trans_cancel(tp, 0);
1664                 return;
1665         }
1666         xfs_mod_sb(tp, fields);
1667         xfs_trans_commit(tp, 0);
1668 }
1669
1670
1671 #ifdef HAVE_PERCPU_SB
1672 /*
1673  * Per-cpu incore superblock counters
1674  *
1675  * Simple concept, difficult implementation
1676  *
1677  * Basically, replace the incore superblock counters with a distributed per cpu
1678  * counter for contended fields (e.g.  free block count).
1679  *
1680  * Difficulties arise in that the incore sb is used for ENOSPC checking, and
1681  * hence needs to be accurately read when we are running low on space. Hence
1682  * there is a method to enable and disable the per-cpu counters based on how
1683  * much "stuff" is available in them.
1684  *
1685  * Basically, a counter is enabled if there is enough free resource to justify
1686  * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
1687  * ENOSPC), then we disable the counters to synchronise all callers and
1688  * re-distribute the available resources.
1689  *
1690  * If, once we redistributed the available resources, we still get a failure,
1691  * we disable the per-cpu counter and go through the slow path.
1692  *
1693  * The slow path is the current xfs_mod_incore_sb() function.  This means that
1694  * when we disable a per-cpu counter, we need to drain it's resources back to
1695  * the global superblock. We do this after disabling the counter to prevent
1696  * more threads from queueing up on the counter.
1697  *
1698  * Essentially, this means that we still need a lock in the fast path to enable
1699  * synchronisation between the global counters and the per-cpu counters. This
1700  * is not a problem because the lock will be local to a CPU almost all the time
1701  * and have little contention except when we get to ENOSPC conditions.
1702  *
1703  * Basically, this lock becomes a barrier that enables us to lock out the fast
1704  * path while we do things like enabling and disabling counters and
1705  * synchronising the counters.
1706  *
1707  * Locking rules:
1708  *
1709  *      1. XFS_SB_LOCK() before picking up per-cpu locks
1710  *      2. per-cpu locks always picked up via for_each_online_cpu() order
1711  *      3. accurate counter sync requires XFS_SB_LOCK + per cpu locks
1712  *      4. modifying per-cpu counters requires holding per-cpu lock
1713  *      5. modifying global counters requires holding XFS_SB_LOCK
1714  *      6. enabling or disabling a counter requires holding the XFS_SB_LOCK
1715  *         and _none_ of the per-cpu locks.
1716  *
1717  * Disabled counters are only ever re-enabled by a balance operation
1718  * that results in more free resources per CPU than a given threshold.
1719  * To ensure counters don't remain disabled, they are rebalanced when
1720  * the global resource goes above a higher threshold (i.e. some hysteresis
1721  * is present to prevent thrashing).
1722  */
1723
1724 #ifdef CONFIG_HOTPLUG_CPU
1725 /*
1726  * hot-plug CPU notifier support.
1727  *
1728  * We need a notifier per filesystem as we need to be able to identify
1729  * the filesystem to balance the counters out. This is achieved by
1730  * having a notifier block embedded in the xfs_mount_t and doing pointer
1731  * magic to get the mount pointer from the notifier block address.
1732  */
1733 STATIC int
1734 xfs_icsb_cpu_notify(
1735         struct notifier_block *nfb,
1736         unsigned long action,
1737         void *hcpu)
1738 {
1739         xfs_icsb_cnts_t *cntp;
1740         xfs_mount_t     *mp;
1741         int             s;
1742
1743         mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
1744         cntp = (xfs_icsb_cnts_t *)
1745                         per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
1746         switch (action) {
1747         case CPU_UP_PREPARE:
1748         case CPU_UP_PREPARE_FROZEN:
1749                 /* Easy Case - initialize the area and locks, and
1750                  * then rebalance when online does everything else for us. */
1751                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1752                 break;
1753         case CPU_ONLINE:
1754         case CPU_ONLINE_FROZEN:
1755                 xfs_icsb_lock(mp);
1756                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0, 0);
1757                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0, 0);
1758                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0, 0);
1759                 xfs_icsb_unlock(mp);
1760                 break;
1761         case CPU_DEAD:
1762         case CPU_DEAD_FROZEN:
1763                 /* Disable all the counters, then fold the dead cpu's
1764                  * count into the total on the global superblock and
1765                  * re-enable the counters. */
1766                 xfs_icsb_lock(mp);
1767                 s = XFS_SB_LOCK(mp);
1768                 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
1769                 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
1770                 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
1771
1772                 mp->m_sb.sb_icount += cntp->icsb_icount;
1773                 mp->m_sb.sb_ifree += cntp->icsb_ifree;
1774                 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
1775
1776                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1777
1778                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT,
1779                                          XFS_ICSB_SB_LOCKED, 0);
1780                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE,
1781                                          XFS_ICSB_SB_LOCKED, 0);
1782                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS,
1783                                          XFS_ICSB_SB_LOCKED, 0);
1784                 XFS_SB_UNLOCK(mp, s);
1785                 xfs_icsb_unlock(mp);
1786                 break;
1787         }
1788
1789         return NOTIFY_OK;
1790 }
1791 #endif /* CONFIG_HOTPLUG_CPU */
1792
1793 int
1794 xfs_icsb_init_counters(
1795         xfs_mount_t     *mp)
1796 {
1797         xfs_icsb_cnts_t *cntp;
1798         int             i;
1799
1800         mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
1801         if (mp->m_sb_cnts == NULL)
1802                 return -ENOMEM;
1803
1804 #ifdef CONFIG_HOTPLUG_CPU
1805         mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
1806         mp->m_icsb_notifier.priority = 0;
1807         register_hotcpu_notifier(&mp->m_icsb_notifier);
1808 #endif /* CONFIG_HOTPLUG_CPU */
1809
1810         for_each_online_cpu(i) {
1811                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1812                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1813         }
1814
1815         mutex_init(&mp->m_icsb_mutex);
1816
1817         /*
1818          * start with all counters disabled so that the
1819          * initial balance kicks us off correctly
1820          */
1821         mp->m_icsb_counters = -1;
1822         return 0;
1823 }
1824
1825 void
1826 xfs_icsb_reinit_counters(
1827         xfs_mount_t     *mp)
1828 {
1829         xfs_icsb_lock(mp);
1830         /*
1831          * start with all counters disabled so that the
1832          * initial balance kicks us off correctly
1833          */
1834         mp->m_icsb_counters = -1;
1835         xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0, 0);
1836         xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0, 0);
1837         xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0, 0);
1838         xfs_icsb_unlock(mp);
1839 }
1840
1841 STATIC void
1842 xfs_icsb_destroy_counters(
1843         xfs_mount_t     *mp)
1844 {
1845         if (mp->m_sb_cnts) {
1846                 unregister_hotcpu_notifier(&mp->m_icsb_notifier);
1847                 free_percpu(mp->m_sb_cnts);
1848         }
1849         mutex_destroy(&mp->m_icsb_mutex);
1850 }
1851
1852 STATIC_INLINE void
1853 xfs_icsb_lock_cntr(
1854         xfs_icsb_cnts_t *icsbp)
1855 {
1856         while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
1857                 ndelay(1000);
1858         }
1859 }
1860
1861 STATIC_INLINE void
1862 xfs_icsb_unlock_cntr(
1863         xfs_icsb_cnts_t *icsbp)
1864 {
1865         clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
1866 }
1867
1868
1869 STATIC_INLINE void
1870 xfs_icsb_lock_all_counters(
1871         xfs_mount_t     *mp)
1872 {
1873         xfs_icsb_cnts_t *cntp;
1874         int             i;
1875
1876         for_each_online_cpu(i) {
1877                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1878                 xfs_icsb_lock_cntr(cntp);
1879         }
1880 }
1881
1882 STATIC_INLINE void
1883 xfs_icsb_unlock_all_counters(
1884         xfs_mount_t     *mp)
1885 {
1886         xfs_icsb_cnts_t *cntp;
1887         int             i;
1888
1889         for_each_online_cpu(i) {
1890                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1891                 xfs_icsb_unlock_cntr(cntp);
1892         }
1893 }
1894
1895 STATIC void
1896 xfs_icsb_count(
1897         xfs_mount_t     *mp,
1898         xfs_icsb_cnts_t *cnt,
1899         int             flags)
1900 {
1901         xfs_icsb_cnts_t *cntp;
1902         int             i;
1903
1904         memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
1905
1906         if (!(flags & XFS_ICSB_LAZY_COUNT))
1907                 xfs_icsb_lock_all_counters(mp);
1908
1909         for_each_online_cpu(i) {
1910                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1911                 cnt->icsb_icount += cntp->icsb_icount;
1912                 cnt->icsb_ifree += cntp->icsb_ifree;
1913                 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
1914         }
1915
1916         if (!(flags & XFS_ICSB_LAZY_COUNT))
1917                 xfs_icsb_unlock_all_counters(mp);
1918 }
1919
1920 STATIC int
1921 xfs_icsb_counter_disabled(
1922         xfs_mount_t     *mp,
1923         xfs_sb_field_t  field)
1924 {
1925         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1926         return test_bit(field, &mp->m_icsb_counters);
1927 }
1928
1929 STATIC int
1930 xfs_icsb_disable_counter(
1931         xfs_mount_t     *mp,
1932         xfs_sb_field_t  field)
1933 {
1934         xfs_icsb_cnts_t cnt;
1935
1936         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1937
1938         /*
1939          * If we are already disabled, then there is nothing to do
1940          * here. We check before locking all the counters to avoid
1941          * the expensive lock operation when being called in the
1942          * slow path and the counter is already disabled. This is
1943          * safe because the only time we set or clear this state is under
1944          * the m_icsb_mutex.
1945          */
1946         if (xfs_icsb_counter_disabled(mp, field))
1947                 return 0;
1948
1949         xfs_icsb_lock_all_counters(mp);
1950         if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
1951                 /* drain back to superblock */
1952
1953                 xfs_icsb_count(mp, &cnt, XFS_ICSB_SB_LOCKED|XFS_ICSB_LAZY_COUNT);
1954                 switch(field) {
1955                 case XFS_SBS_ICOUNT:
1956                         mp->m_sb.sb_icount = cnt.icsb_icount;
1957                         break;
1958                 case XFS_SBS_IFREE:
1959                         mp->m_sb.sb_ifree = cnt.icsb_ifree;
1960                         break;
1961                 case XFS_SBS_FDBLOCKS:
1962                         mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1963                         break;
1964                 default:
1965                         BUG();
1966                 }
1967         }
1968
1969         xfs_icsb_unlock_all_counters(mp);
1970
1971         return 0;
1972 }
1973
1974 STATIC void
1975 xfs_icsb_enable_counter(
1976         xfs_mount_t     *mp,
1977         xfs_sb_field_t  field,
1978         uint64_t        count,
1979         uint64_t        resid)
1980 {
1981         xfs_icsb_cnts_t *cntp;
1982         int             i;
1983
1984         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1985
1986         xfs_icsb_lock_all_counters(mp);
1987         for_each_online_cpu(i) {
1988                 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
1989                 switch (field) {
1990                 case XFS_SBS_ICOUNT:
1991                         cntp->icsb_icount = count + resid;
1992                         break;
1993                 case XFS_SBS_IFREE:
1994                         cntp->icsb_ifree = count + resid;
1995                         break;
1996                 case XFS_SBS_FDBLOCKS:
1997                         cntp->icsb_fdblocks = count + resid;
1998                         break;
1999                 default:
2000                         BUG();
2001                         break;
2002                 }
2003                 resid = 0;
2004         }
2005         clear_bit(field, &mp->m_icsb_counters);
2006         xfs_icsb_unlock_all_counters(mp);
2007 }
2008
2009 void
2010 xfs_icsb_sync_counters_flags(
2011         xfs_mount_t     *mp,
2012         int             flags)
2013 {
2014         xfs_icsb_cnts_t cnt;
2015         int             s;
2016
2017         /* Pass 1: lock all counters */
2018         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
2019                 s = XFS_SB_LOCK(mp);
2020
2021         xfs_icsb_count(mp, &cnt, flags);
2022
2023         /* Step 3: update mp->m_sb fields */
2024         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
2025                 mp->m_sb.sb_icount = cnt.icsb_icount;
2026         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
2027                 mp->m_sb.sb_ifree = cnt.icsb_ifree;
2028         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
2029                 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
2030
2031         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
2032                 XFS_SB_UNLOCK(mp, s);
2033 }
2034
2035 /*
2036  * Accurate update of per-cpu counters to incore superblock
2037  */
2038 STATIC void
2039 xfs_icsb_sync_counters(
2040         xfs_mount_t     *mp)
2041 {
2042         xfs_icsb_sync_counters_flags(mp, 0);
2043 }
2044
2045 /*
2046  * Balance and enable/disable counters as necessary.
2047  *
2048  * Thresholds for re-enabling counters are somewhat magic.  inode counts are
2049  * chosen to be the same number as single on disk allocation chunk per CPU, and
2050  * free blocks is something far enough zero that we aren't going thrash when we
2051  * get near ENOSPC. We also need to supply a minimum we require per cpu to
2052  * prevent looping endlessly when xfs_alloc_space asks for more than will
2053  * be distributed to a single CPU but each CPU has enough blocks to be
2054  * reenabled.
2055  *
2056  * Note that we can be called when counters are already disabled.
2057  * xfs_icsb_disable_counter() optimises the counter locking in this case to
2058  * prevent locking every per-cpu counter needlessly.
2059  */
2060
2061 #define XFS_ICSB_INO_CNTR_REENABLE      (uint64_t)64
2062 #define XFS_ICSB_FDBLK_CNTR_REENABLE(mp) \
2063                 (uint64_t)(512 + XFS_ALLOC_SET_ASIDE(mp))
2064 STATIC void
2065 xfs_icsb_balance_counter(
2066         xfs_mount_t     *mp,
2067         xfs_sb_field_t  field,
2068         int             flags,
2069         int             min_per_cpu)
2070 {
2071         uint64_t        count, resid;
2072         int             weight = num_online_cpus();
2073         int             s;
2074         uint64_t        min = (uint64_t)min_per_cpu;
2075
2076         if (!(flags & XFS_ICSB_SB_LOCKED))
2077                 s = XFS_SB_LOCK(mp);
2078
2079         /* disable counter and sync counter */
2080         xfs_icsb_disable_counter(mp, field);
2081
2082         /* update counters  - first CPU gets residual*/
2083         switch (field) {
2084         case XFS_SBS_ICOUNT:
2085                 count = mp->m_sb.sb_icount;
2086                 resid = do_div(count, weight);
2087                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2088                         goto out;
2089                 break;
2090         case XFS_SBS_IFREE:
2091                 count = mp->m_sb.sb_ifree;
2092                 resid = do_div(count, weight);
2093                 if (count < max(min, XFS_ICSB_INO_CNTR_REENABLE))
2094                         goto out;
2095                 break;
2096         case XFS_SBS_FDBLOCKS:
2097                 count = mp->m_sb.sb_fdblocks;
2098                 resid = do_div(count, weight);
2099                 if (count < max(min, XFS_ICSB_FDBLK_CNTR_REENABLE(mp)))
2100                         goto out;
2101                 break;
2102         default:
2103                 BUG();
2104                 count = resid = 0;      /* quiet, gcc */
2105                 break;
2106         }
2107
2108         xfs_icsb_enable_counter(mp, field, count, resid);
2109 out:
2110         if (!(flags & XFS_ICSB_SB_LOCKED))
2111                 XFS_SB_UNLOCK(mp, s);
2112 }
2113
2114 int
2115 xfs_icsb_modify_counters(
2116         xfs_mount_t     *mp,
2117         xfs_sb_field_t  field,
2118         int64_t         delta,
2119         int             rsvd)
2120 {
2121         xfs_icsb_cnts_t *icsbp;
2122         long long       lcounter;       /* long counter for 64 bit fields */
2123         int             cpu, ret = 0, s;
2124
2125         might_sleep();
2126 again:
2127         cpu = get_cpu();
2128         icsbp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, cpu);
2129
2130         /*
2131          * if the counter is disabled, go to slow path
2132          */
2133         if (unlikely(xfs_icsb_counter_disabled(mp, field)))
2134                 goto slow_path;
2135         xfs_icsb_lock_cntr(icsbp);
2136         if (unlikely(xfs_icsb_counter_disabled(mp, field))) {
2137                 xfs_icsb_unlock_cntr(icsbp);
2138                 goto slow_path;
2139         }
2140
2141         switch (field) {
2142         case XFS_SBS_ICOUNT:
2143                 lcounter = icsbp->icsb_icount;
2144                 lcounter += delta;
2145                 if (unlikely(lcounter < 0))
2146                         goto balance_counter;
2147                 icsbp->icsb_icount = lcounter;
2148                 break;
2149
2150         case XFS_SBS_IFREE:
2151                 lcounter = icsbp->icsb_ifree;
2152                 lcounter += delta;
2153                 if (unlikely(lcounter < 0))
2154                         goto balance_counter;
2155                 icsbp->icsb_ifree = lcounter;
2156                 break;
2157
2158         case XFS_SBS_FDBLOCKS:
2159                 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
2160
2161                 lcounter = icsbp->icsb_fdblocks - XFS_ALLOC_SET_ASIDE(mp);
2162                 lcounter += delta;
2163                 if (unlikely(lcounter < 0))
2164                         goto balance_counter;
2165                 icsbp->icsb_fdblocks = lcounter + XFS_ALLOC_SET_ASIDE(mp);
2166                 break;
2167         default:
2168                 BUG();
2169                 break;
2170         }
2171         xfs_icsb_unlock_cntr(icsbp);
2172         put_cpu();
2173         return 0;
2174
2175 slow_path:
2176         put_cpu();
2177
2178         /*
2179          * serialise with a mutex so we don't burn lots of cpu on
2180          * the superblock lock. We still need to hold the superblock
2181          * lock, however, when we modify the global structures.
2182          */
2183         xfs_icsb_lock(mp);
2184
2185         /*
2186          * Now running atomically.
2187          *
2188          * If the counter is enabled, someone has beaten us to rebalancing.
2189          * Drop the lock and try again in the fast path....
2190          */
2191         if (!(xfs_icsb_counter_disabled(mp, field))) {
2192                 xfs_icsb_unlock(mp);
2193                 goto again;
2194         }
2195
2196         /*
2197          * The counter is currently disabled. Because we are
2198          * running atomically here, we know a rebalance cannot
2199          * be in progress. Hence we can go straight to operating
2200          * on the global superblock. We do not call xfs_mod_incore_sb()
2201          * here even though we need to get the SB_LOCK. Doing so
2202          * will cause us to re-enter this function and deadlock.
2203          * Hence we get the SB_LOCK ourselves and then call
2204          * xfs_mod_incore_sb_unlocked() as the unlocked path operates
2205          * directly on the global counters.
2206          */
2207         s = XFS_SB_LOCK(mp);
2208         ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
2209         XFS_SB_UNLOCK(mp, s);
2210
2211         /*
2212          * Now that we've modified the global superblock, we
2213          * may be able to re-enable the distributed counters
2214          * (e.g. lots of space just got freed). After that
2215          * we are done.
2216          */
2217         if (ret != ENOSPC)
2218                 xfs_icsb_balance_counter(mp, field, 0, 0);
2219         xfs_icsb_unlock(mp);
2220         return ret;
2221
2222 balance_counter:
2223         xfs_icsb_unlock_cntr(icsbp);
2224         put_cpu();
2225
2226         /*
2227          * We may have multiple threads here if multiple per-cpu
2228          * counters run dry at the same time. This will mean we can
2229          * do more balances than strictly necessary but it is not
2230          * the common slowpath case.
2231          */
2232         xfs_icsb_lock(mp);
2233
2234         /*
2235          * running atomically.
2236          *
2237          * This will leave the counter in the correct state for future
2238          * accesses. After the rebalance, we simply try again and our retry
2239          * will either succeed through the fast path or slow path without
2240          * another balance operation being required.
2241          */
2242         xfs_icsb_balance_counter(mp, field, 0, delta);
2243         xfs_icsb_unlock(mp);
2244         goto again;
2245 }
2246
2247 #endif