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_dir.h"
  28 #include "xfs_dir2.h"
  29 #include "xfs_dmapi.h"
  30 #include "xfs_mount.h"
  31 #include "xfs_bmap_btree.h"
  32 #include "xfs_alloc_btree.h"
  33 #include "xfs_ialloc_btree.h"
  34 #include "xfs_dir_sf.h"
  35 #include "xfs_dir2_sf.h"
  36 #include "xfs_attr_sf.h"
  37 #include "xfs_dinode.h"
  38 #include "xfs_inode.h"
  39 #include "xfs_btree.h"
  40 #include "xfs_ialloc.h"
  41 #include "xfs_alloc.h"
  42 #include "xfs_rtalloc.h"
  43 #include "xfs_bmap.h"
  44 #include "xfs_error.h"
  45 #include "xfs_rw.h"
  46 #include "xfs_quota.h"
  47 #include "xfs_fsops.h"
  48
  49 STATIC void     xfs_mount_log_sbunit(xfs_mount_t *, __int64_t);
  50 STATIC int      xfs_uuid_mount(xfs_mount_t *);
  51 STATIC void     xfs_uuid_unmount(xfs_mount_t *mp);
  52 STATIC void     xfs_unmountfs_wait(xfs_mount_t *);
  53
  54
  55 #ifdef HAVE_PERCPU_SB
  56 STATIC void     xfs_icsb_destroy_counters(xfs_mount_t *);
  57 STATIC void     xfs_icsb_balance_counter(xfs_mount_t *, xfs_sb_field_t, int);
  58 STATIC void     xfs_icsb_sync_counters(xfs_mount_t *);
  59 STATIC int      xfs_icsb_modify_counters(xfs_mount_t *, xfs_sb_field_t,
  60                                                 int, int);
  61 STATIC int      xfs_icsb_modify_counters_locked(xfs_mount_t *, xfs_sb_field_t,
  62                                                 int, int);
  63 STATIC int      xfs_icsb_disable_counter(xfs_mount_t *, xfs_sb_field_t);
  64
  65 #else
  66
  67 #define xfs_icsb_destroy_counters(mp)                   do { } while (0)
  68 #define xfs_icsb_balance_counter(mp, a, b)              do { } while (0)
  69 #define xfs_icsb_sync_counters(mp)                      do { } while (0)
  70 #define xfs_icsb_modify_counters(mp, a, b, c)           do { } while (0)
  71 #define xfs_icsb_modify_counters_locked(mp, a, b, c)    do { } while (0)
  72
  73 #endif
  74
  75 static const struct {
  76         short offset;
  77         short type;     /* 0 = integer
  78                          * 1 = binary / string (no translation)
  79                          */
  80 } xfs_sb_info[] = {
  81     { offsetof(xfs_sb_t, sb_magicnum),   0 },
  82     { offsetof(xfs_sb_t, sb_blocksize),  0 },
  83     { offsetof(xfs_sb_t, sb_dblocks),    0 },
  84     { offsetof(xfs_sb_t, sb_rblocks),    0 },
  85     { offsetof(xfs_sb_t, sb_rextents),   0 },
  86     { offsetof(xfs_sb_t, sb_uuid),       1 },
  87     { offsetof(xfs_sb_t, sb_logstart),   0 },
  88     { offsetof(xfs_sb_t, sb_rootino),    0 },
  89     { offsetof(xfs_sb_t, sb_rbmino),     0 },
  90     { offsetof(xfs_sb_t, sb_rsumino),    0 },
  91     { offsetof(xfs_sb_t, sb_rextsize),   0 },
  92     { offsetof(xfs_sb_t, sb_agblocks),   0 },
  93     { offsetof(xfs_sb_t, sb_agcount),    0 },
  94     { offsetof(xfs_sb_t, sb_rbmblocks),  0 },
  95     { offsetof(xfs_sb_t, sb_logblocks),  0 },
  96     { offsetof(xfs_sb_t, sb_versionnum), 0 },
  97     { offsetof(xfs_sb_t, sb_sectsize),   0 },
  98     { offsetof(xfs_sb_t, sb_inodesize),  0 },
  99     { offsetof(xfs_sb_t, sb_inopblock),  0 },
 100     { offsetof(xfs_sb_t, sb_fname[0]),   1 },
 101     { offsetof(xfs_sb_t, sb_blocklog),   0 },
 102     { offsetof(xfs_sb_t, sb_sectlog),    0 },
 103     { offsetof(xfs_sb_t, sb_inodelog),   0 },
 104     { offsetof(xfs_sb_t, sb_inopblog),   0 },
 105     { offsetof(xfs_sb_t, sb_agblklog),   0 },
 106     { offsetof(xfs_sb_t, sb_rextslog),   0 },
 107     { offsetof(xfs_sb_t, sb_inprogress), 0 },
 108     { offsetof(xfs_sb_t, sb_imax_pct),   0 },
 109     { offsetof(xfs_sb_t, sb_icount),     0 },
 110     { offsetof(xfs_sb_t, sb_ifree),      0 },
 111     { offsetof(xfs_sb_t, sb_fdblocks),   0 },
 112     { offsetof(xfs_sb_t, sb_frextents),  0 },
 113     { offsetof(xfs_sb_t, sb_uquotino),   0 },
 114     { offsetof(xfs_sb_t, sb_gquotino),   0 },
 115     { offsetof(xfs_sb_t, sb_qflags),     0 },
 116     { offsetof(xfs_sb_t, sb_flags),      0 },
 117     { offsetof(xfs_sb_t, sb_shared_vn),  0 },
 118     { offsetof(xfs_sb_t, sb_inoalignmt), 0 },
 119     { offsetof(xfs_sb_t, sb_unit),       0 },
 120     { offsetof(xfs_sb_t, sb_width),      0 },
 121     { offsetof(xfs_sb_t, sb_dirblklog),  0 },
 122     { offsetof(xfs_sb_t, sb_logsectlog), 0 },
 123     { offsetof(xfs_sb_t, sb_logsectsize),0 },
 124     { offsetof(xfs_sb_t, sb_logsunit),   0 },
 125     { offsetof(xfs_sb_t, sb_features2),  0 },
 126     { sizeof(xfs_sb_t),                  0 }
 127 };
 128
 129 /*
 130  * Return a pointer to an initialized xfs_mount structure.
 131  */
 132 xfs_mount_t *
 133 xfs_mount_init(void)
 134 {
 135         xfs_mount_t *mp;
 136
 137         mp = kmem_zalloc(sizeof(xfs_mount_t), KM_SLEEP);
 138
 139         if (xfs_icsb_init_counters(mp)) {
 140                 mp->m_flags |= XFS_MOUNT_NO_PERCPU_SB;
 141         }
 142
 143         AIL_LOCKINIT(&mp->m_ail_lock, "xfs_ail");
 144         spinlock_init(&mp->m_sb_lock, "xfs_sb");
 145         mutex_init(&mp->m_ilock);
 146         initnsema(&mp->m_growlock, 1, "xfs_grow");
 147         /*
 148          * Initialize the AIL.
 149          */
 150         xfs_trans_ail_init(mp);
 151
 152         atomic_set(&mp->m_active_trans, 0);
 153
 154         return mp;
 155 }
 156
 157 /*
 158  * Free up the resources associated with a mount structure.  Assume that
 159  * the structure was initially zeroed, so we can tell which fields got
 160  * initialized.
 161  */
 162 void
 163 xfs_mount_free(
 164         xfs_mount_t     *mp,
 165         int             remove_bhv)
 166 {
 167         if (mp->m_ihash)
 168                 xfs_ihash_free(mp);
 169         if (mp->m_chash)
 170                 xfs_chash_free(mp);
 171
 172         if (mp->m_perag) {
 173                 int     agno;
 174
 175                 for (agno = 0; agno < mp->m_maxagi; agno++)
 176                         if (mp->m_perag[agno].pagb_list)
 177                                 kmem_free(mp->m_perag[agno].pagb_list,
 178                                                 sizeof(xfs_perag_busy_t) *
 179                                                         XFS_PAGB_NUM_SLOTS);
 180                 kmem_free(mp->m_perag,
 181                           sizeof(xfs_perag_t) * mp->m_sb.sb_agcount);
 182         }
 183
 184         AIL_LOCK_DESTROY(&mp->m_ail_lock);
 185         spinlock_destroy(&mp->m_sb_lock);
 186         mutex_destroy(&mp->m_ilock);
 187         freesema(&mp->m_growlock);
 188         if (mp->m_quotainfo)
 189                 XFS_QM_DONE(mp);
 190
 191         if (mp->m_fsname != NULL)
 192                 kmem_free(mp->m_fsname, mp->m_fsname_len);
 193         if (mp->m_rtname != NULL)
 194                 kmem_free(mp->m_rtname, strlen(mp->m_rtname) + 1);
 195         if (mp->m_logname != NULL)
 196                 kmem_free(mp->m_logname, strlen(mp->m_logname) + 1);
 197
 198         if (remove_bhv) {
 199                 struct vfs      *vfsp = XFS_MTOVFS(mp);
 200
 201                 bhv_remove_all_vfsops(vfsp, 0);
 202                 VFS_REMOVEBHV(vfsp, &mp->m_bhv);
 203         }
 204
 205         xfs_icsb_destroy_counters(mp);
 206         kmem_free(mp, sizeof(xfs_mount_t));
 207 }
 208
 209
 210 /*
 211  * Check the validity of the SB found.
 212  */
 213 STATIC int
 214 xfs_mount_validate_sb(
 215         xfs_mount_t     *mp,
 216         xfs_sb_t        *sbp)
 217 {
 218         /*
 219          * If the log device and data device have the
 220          * same device number, the log is internal.
 221          * Consequently, the sb_logstart should be non-zero.  If
 222          * we have a zero sb_logstart in this case, we may be trying to mount
 223          * a volume filesystem in a non-volume manner.
 224          */
 225         if (sbp->sb_magicnum != XFS_SB_MAGIC) {
 226                 cmn_err(CE_WARN, "XFS: bad magic number");
 227                 return XFS_ERROR(EWRONGFS);
 228         }
 229
 230         if (!XFS_SB_GOOD_VERSION(sbp)) {
 231                 cmn_err(CE_WARN, "XFS: bad version");
 232                 return XFS_ERROR(EWRONGFS);
 233         }
 234
 235         if (unlikely(
 236             sbp->sb_logstart == 0 && mp->m_logdev_targp == mp->m_ddev_targp)) {
 237                 cmn_err(CE_WARN,
 238         "XFS: filesystem is marked as having an external log; "
 239         "specify logdev on the\nmount command line.");
 240                 XFS_CORRUPTION_ERROR("xfs_mount_validate_sb(1)",
 241                                      XFS_ERRLEVEL_HIGH, mp, sbp);
 242                 return XFS_ERROR(EFSCORRUPTED);
 243         }
 244
 245         if (unlikely(
 246             sbp->sb_logstart != 0 && mp->m_logdev_targp != mp->m_ddev_targp)) {
 247                 cmn_err(CE_WARN,
 248         "XFS: filesystem is marked as having an internal log; "
 249         "don't specify logdev on\nthe mount command line.");
 250                 XFS_CORRUPTION_ERROR("xfs_mount_validate_sb(2)",
 251                                      XFS_ERRLEVEL_HIGH, mp, sbp);
 252                 return XFS_ERROR(EFSCORRUPTED);
 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 || sbp->sb_imax_pct < 1))) {
 277                 cmn_err(CE_WARN, "XFS: SB sanity check 1 failed");
 278                 XFS_CORRUPTION_ERROR("xfs_mount_validate_sb(3)",
 279                                      XFS_ERRLEVEL_LOW, mp, sbp);
 280                 return XFS_ERROR(EFSCORRUPTED);
 281         }
 282
 283         /*
 284          * Sanity check AG count, size fields against data size field
 285          */
 286         if (unlikely(
 287             sbp->sb_dblocks == 0 ||
 288             sbp->sb_dblocks >
 289              (xfs_drfsbno_t)sbp->sb_agcount * sbp->sb_agblocks ||
 290             sbp->sb_dblocks < (xfs_drfsbno_t)(sbp->sb_agcount - 1) *
 291                               sbp->sb_agblocks + XFS_MIN_AG_BLOCKS)) {
 292                 cmn_err(CE_WARN, "XFS: SB sanity check 2 failed");
 293                 XFS_ERROR_REPORT("xfs_mount_validate_sb(4)",
 294                                  XFS_ERRLEVEL_LOW, mp);
 295                 return XFS_ERROR(EFSCORRUPTED);
 296         }
 297
 298         ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
 299         ASSERT(sbp->sb_blocklog >= BBSHIFT);
 300
 301 #if XFS_BIG_BLKNOS     /* Limited by ULONG_MAX of page cache index */
 302         if (unlikely(
 303             (sbp->sb_dblocks >> (PAGE_SHIFT - sbp->sb_blocklog)) > ULONG_MAX ||
 304             (sbp->sb_rblocks >> (PAGE_SHIFT - sbp->sb_blocklog)) > ULONG_MAX)) {
 305 #else                  /* Limited by UINT_MAX of sectors */
 306         if (unlikely(
 307             (sbp->sb_dblocks << (sbp->sb_blocklog - BBSHIFT)) > UINT_MAX ||
 308             (sbp->sb_rblocks << (sbp->sb_blocklog - BBSHIFT)) > UINT_MAX)) {
 309 #endif
 310                 cmn_err(CE_WARN,
 311         "XFS: File system is too large to be mounted on this system.");
 312                 return XFS_ERROR(E2BIG);
 313         }
 314
 315         if (unlikely(sbp->sb_inprogress)) {
 316                 cmn_err(CE_WARN, "XFS: file system busy");
 317                 XFS_ERROR_REPORT("xfs_mount_validate_sb(5)",
 318                                  XFS_ERRLEVEL_LOW, mp);
 319                 return XFS_ERROR(EFSCORRUPTED);
 320         }
 321
 322         /*
 323          * Version 1 directory format has never worked on Linux.
 324          */
 325         if (unlikely(!XFS_SB_VERSION_HASDIRV2(sbp))) {
 326                 cmn_err(CE_WARN,
 327         "XFS: Attempted to mount file system using version 1 directory format");
 328                 return XFS_ERROR(ENOSYS);
 329         }
 330
 331         /*
 332          * Until this is fixed only page-sized or smaller data blocks work.
 333          */
 334         if (unlikely(sbp->sb_blocksize > PAGE_SIZE)) {
 335                 cmn_err(CE_WARN,
 336                 "XFS: Attempted to mount file system with blocksize %d bytes",
 337                         sbp->sb_blocksize);
 338                 cmn_err(CE_WARN,
 339                 "XFS: Only page-sized (%ld) or less blocksizes currently work.",
 340                         PAGE_SIZE);
 341                 return XFS_ERROR(ENOSYS);
 342         }
 343
 344         return 0;
 345 }
 346
 347 xfs_agnumber_t
 348 xfs_initialize_perag(
 349         struct vfs      *vfs,
 350         xfs_mount_t     *mp,
 351         xfs_agnumber_t  agcount)
 352 {
 353         xfs_agnumber_t  index, max_metadata;
 354         xfs_perag_t     *pag;
 355         xfs_agino_t     agino;
 356         xfs_ino_t       ino;
 357         xfs_sb_t        *sbp = &mp->m_sb;
 358         xfs_ino_t       max_inum = XFS_MAXINUMBER_32;
 359
 360         /* Check to see if the filesystem can overflow 32 bit inodes */
 361         agino = XFS_OFFBNO_TO_AGINO(mp, sbp->sb_agblocks - 1, 0);
 362         ino = XFS_AGINO_TO_INO(mp, agcount - 1, agino);
 363
 364         /* Clear the mount flag if no inode can overflow 32 bits
 365          * on this filesystem, or if specifically requested..
 366          */
 367         if ((vfs->vfs_flag & VFS_32BITINODES) && ino > max_inum) {
 368                 mp->m_flags |= XFS_MOUNT_32BITINODES;
 369         } else {
 370                 mp->m_flags &= ~XFS_MOUNT_32BITINODES;
 371         }
 372
 373         /* If we can overflow then setup the ag headers accordingly */
 374         if (mp->m_flags & XFS_MOUNT_32BITINODES) {
 375                 /* Calculate how much should be reserved for inodes to
 376                  * meet the max inode percentage.
 377                  */
 378                 if (mp->m_maxicount) {
 379                         __uint64_t      icount;
 380
 381                         icount = sbp->sb_dblocks * sbp->sb_imax_pct;
 382                         do_div(icount, 100);
 383                         icount += sbp->sb_agblocks - 1;
 384                         do_div(icount, sbp->sb_agblocks);
 385                         max_metadata = icount;
 386                 } else {
 387                         max_metadata = agcount;
 388                 }
 389                 for (index = 0; index < agcount; index++) {
 390                         ino = XFS_AGINO_TO_INO(mp, index, agino);
 391                         if (ino > max_inum) {
 392                                 index++;
 393                                 break;
 394                         }
 395
 396                         /* This ag is preferred for inodes */
 397                         pag = &mp->m_perag[index];
 398                         pag->pagi_inodeok = 1;
 399                         if (index < max_metadata)
 400                                 pag->pagf_metadata = 1;
 401                 }
 402         } else {
 403                 /* Setup default behavior for smaller filesystems */
 404                 for (index = 0; index < agcount; index++) {
 405                         pag = &mp->m_perag[index];
 406                         pag->pagi_inodeok = 1;
 407                 }
 408         }
 409         return index;
 410 }
 411
 412 /*
 413  * xfs_xlatesb
 414  *
 415  *     data       - on disk version of sb
 416  *     sb         - a superblock
 417  *     dir        - conversion direction: <0 - convert sb to buf
 418  *                                        >0 - convert buf to sb
 419  *     fields     - which fields to copy (bitmask)
 420  */
 421 void
 422 xfs_xlatesb(
 423         void            *data,
 424         xfs_sb_t        *sb,
 425         int             dir,
 426         __int64_t       fields)
 427 {
 428         xfs_caddr_t     buf_ptr;
 429         xfs_caddr_t     mem_ptr;
 430         xfs_sb_field_t  f;
 431         int             first;
 432         int             size;
 433
 434         ASSERT(dir);
 435         ASSERT(fields);
 436
 437         if (!fields)
 438                 return;
 439
 440         buf_ptr = (xfs_caddr_t)data;
 441         mem_ptr = (xfs_caddr_t)sb;
 442
 443         while (fields) {
 444                 f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
 445                 first = xfs_sb_info[f].offset;
 446                 size = xfs_sb_info[f + 1].offset - first;
 447
 448                 ASSERT(xfs_sb_info[f].type == 0 || xfs_sb_info[f].type == 1);
 449
 450                 if (size == 1 || xfs_sb_info[f].type == 1) {
 451                         if (dir > 0) {
 452                                 memcpy(mem_ptr + first, buf_ptr + first, size);
 453                         } else {
 454                                 memcpy(buf_ptr + first, mem_ptr + first, size);
 455                         }
 456                 } else {
 457                         switch (size) {
 458                         case 2:
 459                                 INT_XLATE(*(__uint16_t*)(buf_ptr+first),
 460                                           *(__uint16_t*)(mem_ptr+first),
 461                                           dir, ARCH_CONVERT);
 462                                 break;
 463                         case 4:
 464                                 INT_XLATE(*(__uint32_t*)(buf_ptr+first),
 465                                           *(__uint32_t*)(mem_ptr+first),
 466                                           dir, ARCH_CONVERT);
 467                                 break;
 468                         case 8:
 469                                 INT_XLATE(*(__uint64_t*)(buf_ptr+first),
 470                                           *(__uint64_t*)(mem_ptr+first), dir, ARCH_CONVERT);
 471                                 break;
 472                         default:
 473                                 ASSERT(0);
 474                         }
 475                 }
 476
 477                 fields &= ~(1LL << f);
 478         }
 479 }
 480
 481 /*
 482  * xfs_readsb
 483  *
 484  * Does the initial read of the superblock.
 485  */
 486 int
 487 xfs_readsb(xfs_mount_t *mp)
 488 {
 489         unsigned int    sector_size;
 490         unsigned int    extra_flags;
 491         xfs_buf_t       *bp;
 492         xfs_sb_t        *sbp;
 493         int             error;
 494
 495         ASSERT(mp->m_sb_bp == NULL);
 496         ASSERT(mp->m_ddev_targp != NULL);
 497
 498         /*
 499          * Allocate a (locked) buffer to hold the superblock.
 500          * This will be kept around at all times to optimize
 501          * access to the superblock.
 502          */
 503         sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
 504         extra_flags = XFS_BUF_LOCK | XFS_BUF_MANAGE | XFS_BUF_MAPPED;
 505
 506         bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
 507                                 BTOBB(sector_size), extra_flags);
 508         if (!bp || XFS_BUF_ISERROR(bp)) {
 509                 cmn_err(CE_WARN, "XFS: SB read failed");
 510                 error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
 511                 goto fail;
 512         }
 513         ASSERT(XFS_BUF_ISBUSY(bp));
 514         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 515
 516         /*
 517          * Initialize the mount structure from the superblock.
 518          * But first do some basic consistency checking.
 519          */
 520         sbp = XFS_BUF_TO_SBP(bp);
 521         xfs_xlatesb(XFS_BUF_PTR(bp), &(mp->m_sb), 1, XFS_SB_ALL_BITS);
 522
 523         error = xfs_mount_validate_sb(mp, &(mp->m_sb));
 524         if (error) {
 525                 cmn_err(CE_WARN, "XFS: SB validate failed");
 526                 goto fail;
 527         }
 528
 529         /*
 530          * We must be able to do sector-sized and sector-aligned IO.
 531          */
 532         if (sector_size > mp->m_sb.sb_sectsize) {
 533                 cmn_err(CE_WARN,
 534                         "XFS: device supports only %u byte sectors (not %u)",
 535                         sector_size, mp->m_sb.sb_sectsize);
 536                 error = ENOSYS;
 537                 goto fail;
 538         }
 539
 540         /*
 541          * If device sector size is smaller than the superblock size,
 542          * re-read the superblock so the buffer is correctly sized.
 543          */
 544         if (sector_size < mp->m_sb.sb_sectsize) {
 545                 XFS_BUF_UNMANAGE(bp);
 546                 xfs_buf_relse(bp);
 547                 sector_size = mp->m_sb.sb_sectsize;
 548                 bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
 549                                         BTOBB(sector_size), extra_flags);
 550                 if (!bp || XFS_BUF_ISERROR(bp)) {
 551                         cmn_err(CE_WARN, "XFS: SB re-read failed");
 552                         error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
 553                         goto fail;
 554                 }
 555                 ASSERT(XFS_BUF_ISBUSY(bp));
 556                 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 557         }
 558
 559         xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
 560         xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
 561         xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
 562
 563         mp->m_sb_bp = bp;
 564         xfs_buf_relse(bp);
 565         ASSERT(XFS_BUF_VALUSEMA(bp) > 0);
 566         return 0;
 567
 568  fail:
 569         if (bp) {
 570                 XFS_BUF_UNMANAGE(bp);
 571                 xfs_buf_relse(bp);
 572         }
 573         return error;
 574 }
 575
 576
 577 /*
 578  * xfs_mount_common
 579  *
 580  * Mount initialization code establishing various mount
 581  * fields from the superblock associated with the given
 582  * mount structure
 583  */
 584 STATIC void
 585 xfs_mount_common(xfs_mount_t *mp, xfs_sb_t *sbp)
 586 {
 587         int     i;
 588
 589         mp->m_agfrotor = mp->m_agirotor = 0;
 590         spinlock_init(&mp->m_agirotor_lock, "m_agirotor_lock");
 591         mp->m_maxagi = mp->m_sb.sb_agcount;
 592         mp->m_blkbit_log = sbp->sb_blocklog + XFS_NBBYLOG;
 593         mp->m_blkbb_log = sbp->sb_blocklog - BBSHIFT;
 594         mp->m_sectbb_log = sbp->sb_sectlog - BBSHIFT;
 595         mp->m_agno_log = xfs_highbit32(sbp->sb_agcount - 1) + 1;
 596         mp->m_agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
 597         mp->m_litino = sbp->sb_inodesize -
 598                 ((uint)sizeof(xfs_dinode_core_t) + (uint)sizeof(xfs_agino_t));
 599         mp->m_blockmask = sbp->sb_blocksize - 1;
 600         mp->m_blockwsize = sbp->sb_blocksize >> XFS_WORDLOG;
 601         mp->m_blockwmask = mp->m_blockwsize - 1;
 602         INIT_LIST_HEAD(&mp->m_del_inodes);
 603
 604         /*
 605          * Setup for attributes, in case they get created.
 606          * This value is for inodes getting attributes for the first time,
 607          * the per-inode value is for old attribute values.
 608          */
 609         ASSERT(sbp->sb_inodesize >= 256 && sbp->sb_inodesize <= 2048);
 610         switch (sbp->sb_inodesize) {
 611         case 256:
 612                 mp->m_attroffset = XFS_LITINO(mp) -
 613                                    XFS_BMDR_SPACE_CALC(MINABTPTRS);
 614                 break;
 615         case 512:
 616         case 1024:
 617         case 2048:
 618                 mp->m_attroffset = XFS_BMDR_SPACE_CALC(6 * MINABTPTRS);
 619                 break;
 620         default:
 621                 ASSERT(0);
 622         }
 623         ASSERT(mp->m_attroffset < XFS_LITINO(mp));
 624
 625         for (i = 0; i < 2; i++) {
 626                 mp->m_alloc_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 627                         xfs_alloc, i == 0);
 628                 mp->m_alloc_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 629                         xfs_alloc, i == 0);
 630         }
 631         for (i = 0; i < 2; i++) {
 632                 mp->m_bmap_dmxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 633                         xfs_bmbt, i == 0);
 634                 mp->m_bmap_dmnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 635                         xfs_bmbt, i == 0);
 636         }
 637         for (i = 0; i < 2; i++) {
 638                 mp->m_inobt_mxr[i] = XFS_BTREE_BLOCK_MAXRECS(sbp->sb_blocksize,
 639                         xfs_inobt, i == 0);
 640                 mp->m_inobt_mnr[i] = XFS_BTREE_BLOCK_MINRECS(sbp->sb_blocksize,
 641                         xfs_inobt, i == 0);
 642         }
 643
 644         mp->m_bsize = XFS_FSB_TO_BB(mp, 1);
 645         mp->m_ialloc_inos = (int)MAX((__uint16_t)XFS_INODES_PER_CHUNK,
 646                                         sbp->sb_inopblock);
 647         mp->m_ialloc_blks = mp->m_ialloc_inos >> sbp->sb_inopblog;
 648 }
 649 /*
 650  * xfs_mountfs
 651  *
 652  * This function does the following on an initial mount of a file system:
 653  *      - reads the superblock from disk and init the mount struct
 654  *      - if we're a 32-bit kernel, do a size check on the superblock
 655  *              so we don't mount terabyte filesystems
 656  *      - init mount struct realtime fields
 657  *      - allocate inode hash table for fs
 658  *      - init directory manager
 659  *      - perform recovery and init the log manager
 660  */
 661 int
 662 xfs_mountfs(
 663         vfs_t           *vfsp,
 664         xfs_mount_t     *mp,
 665         int             mfsi_flags)
 666 {
 667         xfs_buf_t       *bp;
 668         xfs_sb_t        *sbp = &(mp->m_sb);
 669         xfs_inode_t     *rip;
 670         vnode_t         *rvp = NULL;
 671         int             readio_log, writeio_log;
 672         xfs_daddr_t     d;
 673         __uint64_t      ret64;
 674         __int64_t       update_flags;
 675         uint            quotamount, quotaflags;
 676         int             agno;
 677         int             uuid_mounted = 0;
 678         int             error = 0;
 679
 680         if (mp->m_sb_bp == NULL) {
 681                 if ((error = xfs_readsb(mp))) {
 682                         return error;
 683                 }
 684         }
 685         xfs_mount_common(mp, sbp);
 686
 687         /*
 688          * Check if sb_agblocks is aligned at stripe boundary
 689          * If sb_agblocks is NOT aligned turn off m_dalign since
 690          * allocator alignment is within an ag, therefore ag has
 691          * to be aligned at stripe boundary.
 692          */
 693         update_flags = 0LL;
 694         if (mp->m_dalign && !(mfsi_flags & XFS_MFSI_SECOND)) {
 695                 /*
 696                  * If stripe unit and stripe width are not multiples
 697                  * of the fs blocksize turn off alignment.
 698                  */
 699                 if ((BBTOB(mp->m_dalign) & mp->m_blockmask) ||
 700                     (BBTOB(mp->m_swidth) & mp->m_blockmask)) {
 701                         if (mp->m_flags & XFS_MOUNT_RETERR) {
 702                                 cmn_err(CE_WARN,
 703                                         "XFS: alignment check 1 failed");
 704                                 error = XFS_ERROR(EINVAL);
 705                                 goto error1;
 706                         }
 707                         mp->m_dalign = mp->m_swidth = 0;
 708                 } else {
 709                         /*
 710                          * Convert the stripe unit and width to FSBs.
 711                          */
 712                         mp->m_dalign = XFS_BB_TO_FSBT(mp, mp->m_dalign);
 713                         if (mp->m_dalign && (sbp->sb_agblocks % mp->m_dalign)) {
 714                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
 715                                         error = XFS_ERROR(EINVAL);
 716                                         goto error1;
 717                                 }
 718                                 xfs_fs_cmn_err(CE_WARN, mp,
 719 "stripe alignment turned off: sunit(%d)/swidth(%d) incompatible with agsize(%d)",
 720                                         mp->m_dalign, mp->m_swidth,
 721                                         sbp->sb_agblocks);
 722
 723                                 mp->m_dalign = 0;
 724                                 mp->m_swidth = 0;
 725                         } else if (mp->m_dalign) {
 726                                 mp->m_swidth = XFS_BB_TO_FSBT(mp, mp->m_swidth);
 727                         } else {
 728                                 if (mp->m_flags & XFS_MOUNT_RETERR) {
 729                                         xfs_fs_cmn_err(CE_WARN, mp,
 730 "stripe alignment turned off: sunit(%d) less than bsize(%d)",
 731                                                 mp->m_dalign,
 732                                                 mp->m_blockmask +1);
 733                                         error = XFS_ERROR(EINVAL);
 734                                         goto error1;
 735                                 }
 736                                 mp->m_swidth = 0;
 737                         }
 738                 }
 739
 740                 /*
 741                  * Update superblock with new values
 742                  * and log changes
 743                  */
 744                 if (XFS_SB_VERSION_HASDALIGN(sbp)) {
 745                         if (sbp->sb_unit != mp->m_dalign) {
 746                                 sbp->sb_unit = mp->m_dalign;
 747                                 update_flags |= XFS_SB_UNIT;
 748                         }
 749                         if (sbp->sb_width != mp->m_swidth) {
 750                                 sbp->sb_width = mp->m_swidth;
 751                                 update_flags |= XFS_SB_WIDTH;
 752                         }
 753                 }
 754         } else if ((mp->m_flags & XFS_MOUNT_NOALIGN) != XFS_MOUNT_NOALIGN &&
 755                     XFS_SB_VERSION_HASDALIGN(&mp->m_sb)) {
 756                         mp->m_dalign = sbp->sb_unit;
 757                         mp->m_swidth = sbp->sb_width;
 758         }
 759
 760         xfs_alloc_compute_maxlevels(mp);
 761         xfs_bmap_compute_maxlevels(mp, XFS_DATA_FORK);
 762         xfs_bmap_compute_maxlevels(mp, XFS_ATTR_FORK);
 763         xfs_ialloc_compute_maxlevels(mp);
 764
 765         if (sbp->sb_imax_pct) {
 766                 __uint64_t      icount;
 767
 768                 /* Make sure the maximum inode count is a multiple of the
 769                  * units we allocate inodes in.
 770                  */
 771
 772                 icount = sbp->sb_dblocks * sbp->sb_imax_pct;
 773                 do_div(icount, 100);
 774                 do_div(icount, mp->m_ialloc_blks);
 775                 mp->m_maxicount = (icount * mp->m_ialloc_blks)  <<
 776                                    sbp->sb_inopblog;
 777         } else
 778                 mp->m_maxicount = 0;
 779
 780         mp->m_maxioffset = xfs_max_file_offset(sbp->sb_blocklog);
 781
 782         /*
 783          * XFS uses the uuid from the superblock as the unique
 784          * identifier for fsid.  We can not use the uuid from the volume
 785          * since a single partition filesystem is identical to a single
 786          * partition volume/filesystem.
 787          */
 788         if ((mfsi_flags & XFS_MFSI_SECOND) == 0 &&
 789             (mp->m_flags & XFS_MOUNT_NOUUID) == 0) {
 790                 if (xfs_uuid_mount(mp)) {
 791                         error = XFS_ERROR(EINVAL);
 792                         goto error1;
 793                 }
 794                 uuid_mounted=1;
 795                 ret64 = uuid_hash64(&sbp->sb_uuid);
 796                 memcpy(&vfsp->vfs_fsid, &ret64, sizeof(ret64));
 797         }
 798
 799         /*
 800          * Set the default minimum read and write sizes unless
 801          * already specified in a mount option.
 802          * We use smaller I/O sizes when the file system
 803          * is being used for NFS service (wsync mount option).
 804          */
 805         if (!(mp->m_flags & XFS_MOUNT_DFLT_IOSIZE)) {
 806                 if (mp->m_flags & XFS_MOUNT_WSYNC) {
 807                         readio_log = XFS_WSYNC_READIO_LOG;
 808                         writeio_log = XFS_WSYNC_WRITEIO_LOG;
 809                 } else {
 810                         readio_log = XFS_READIO_LOG_LARGE;
 811                         writeio_log = XFS_WRITEIO_LOG_LARGE;
 812                 }
 813         } else {
 814                 readio_log = mp->m_readio_log;
 815                 writeio_log = mp->m_writeio_log;
 816         }
 817
 818         /*
 819          * Set the number of readahead buffers to use based on
 820          * physical memory size.
 821          */
 822         if (xfs_physmem <= 4096)                /* <= 16MB */
 823                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_16MB;
 824         else if (xfs_physmem <= 8192)   /* <= 32MB */
 825                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_32MB;
 826         else
 827                 mp->m_nreadaheads = XFS_RW_NREADAHEAD_K32;
 828         if (sbp->sb_blocklog > readio_log) {
 829                 mp->m_readio_log = sbp->sb_blocklog;
 830         } else {
 831                 mp->m_readio_log = readio_log;
 832         }
 833         mp->m_readio_blocks = 1 << (mp->m_readio_log - sbp->sb_blocklog);
 834         if (sbp->sb_blocklog > writeio_log) {
 835                 mp->m_writeio_log = sbp->sb_blocklog;
 836         } else {
 837                 mp->m_writeio_log = writeio_log;
 838         }
 839         mp->m_writeio_blocks = 1 << (mp->m_writeio_log - sbp->sb_blocklog);
 840
 841         /*
 842          * Set the inode cluster size based on the physical memory
 843          * size.  This may still be overridden by the file system
 844          * block size if it is larger than the chosen cluster size.
 845          */
 846         if (xfs_physmem <= btoc(32 * 1024 * 1024)) { /* <= 32 MB */
 847                 mp->m_inode_cluster_size = XFS_INODE_SMALL_CLUSTER_SIZE;
 848         } else {
 849                 mp->m_inode_cluster_size = XFS_INODE_BIG_CLUSTER_SIZE;
 850         }
 851         /*
 852          * Set whether we're using inode alignment.
 853          */
 854         if (XFS_SB_VERSION_HASALIGN(&mp->m_sb) &&
 855             mp->m_sb.sb_inoalignmt >=
 856             XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size))
 857                 mp->m_inoalign_mask = mp->m_sb.sb_inoalignmt - 1;
 858         else
 859                 mp->m_inoalign_mask = 0;
 860         /*
 861          * If we are using stripe alignment, check whether
 862          * the stripe unit is a multiple of the inode alignment
 863          */
 864         if (mp->m_dalign && mp->m_inoalign_mask &&
 865             !(mp->m_dalign & mp->m_inoalign_mask))
 866                 mp->m_sinoalign = mp->m_dalign;
 867         else
 868                 mp->m_sinoalign = 0;
 869         /*
 870          * Check that the data (and log if separate) are an ok size.
 871          */
 872         d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);
 873         if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_dblocks) {
 874                 cmn_err(CE_WARN, "XFS: size check 1 failed");
 875                 error = XFS_ERROR(E2BIG);
 876                 goto error1;
 877         }
 878         error = xfs_read_buf(mp, mp->m_ddev_targp,
 879                              d - XFS_FSS_TO_BB(mp, 1),
 880                              XFS_FSS_TO_BB(mp, 1), 0, &bp);
 881         if (!error) {
 882                 xfs_buf_relse(bp);
 883         } else {
 884                 cmn_err(CE_WARN, "XFS: size check 2 failed");
 885                 if (error == ENOSPC) {
 886                         error = XFS_ERROR(E2BIG);
 887                 }
 888                 goto error1;
 889         }
 890
 891         if (((mfsi_flags & XFS_MFSI_CLIENT) == 0) &&
 892             mp->m_logdev_targp != mp->m_ddev_targp) {
 893                 d = (xfs_daddr_t)XFS_FSB_TO_BB(mp, mp->m_sb.sb_logblocks);
 894                 if (XFS_BB_TO_FSB(mp, d) != mp->m_sb.sb_logblocks) {
 895                         cmn_err(CE_WARN, "XFS: size check 3 failed");
 896                         error = XFS_ERROR(E2BIG);
 897                         goto error1;
 898                 }
 899                 error = xfs_read_buf(mp, mp->m_logdev_targp,
 900                                      d - XFS_FSB_TO_BB(mp, 1),
 901                                      XFS_FSB_TO_BB(mp, 1), 0, &bp);
 902                 if (!error) {
 903                         xfs_buf_relse(bp);
 904                 } else {
 905                         cmn_err(CE_WARN, "XFS: size check 3 failed");
 906                         if (error == ENOSPC) {
 907                                 error = XFS_ERROR(E2BIG);
 908                         }
 909                         goto error1;
 910                 }
 911         }
 912
 913         /*
 914          * Initialize realtime fields in the mount structure
 915          */
 916         if ((error = xfs_rtmount_init(mp))) {
 917                 cmn_err(CE_WARN, "XFS: RT mount failed");
 918                 goto error1;
 919         }
 920
 921         /*
 922          * For client case we are done now
 923          */
 924         if (mfsi_flags & XFS_MFSI_CLIENT) {
 925                 return 0;
 926         }
 927
 928         /*
 929          *  Copies the low order bits of the timestamp and the randomly
 930          *  set "sequence" number out of a UUID.
 931          */
 932         uuid_getnodeuniq(&sbp->sb_uuid, mp->m_fixedfsid);
 933
 934         /*
 935          *  The vfs structure needs to have a file system independent
 936          *  way of checking for the invariant file system ID.  Since it
 937          *  can't look at mount structures it has a pointer to the data
 938          *  in the mount structure.
 939          *
 940          *  File systems that don't support user level file handles (i.e.
 941          *  all of them except for XFS) will leave vfs_altfsid as NULL.
 942          */
 943         vfsp->vfs_altfsid = (xfs_fsid_t *)mp->m_fixedfsid;
 944         mp->m_dmevmask = 0;     /* not persistent; set after each mount */
 945
 946         /*
 947          * Select the right directory manager.
 948          */
 949         mp->m_dirops =
 950                 XFS_SB_VERSION_HASDIRV2(&mp->m_sb) ?
 951                         xfsv2_dirops :
 952                         xfsv1_dirops;
 953
 954         /*
 955          * Initialize directory manager's entries.
 956          */
 957         XFS_DIR_MOUNT(mp);
 958
 959         /*
 960          * Initialize the attribute manager's entries.
 961          */
 962         mp->m_attr_magicpct = (mp->m_sb.sb_blocksize * 37) / 100;
 963
 964         /*
 965          * Initialize the precomputed transaction reservations values.
 966          */
 967         xfs_trans_init(mp);
 968
 969         /*
 970          * Allocate and initialize the inode hash table for this
 971          * file system.
 972          */
 973         xfs_ihash_init(mp);
 974         xfs_chash_init(mp);
 975
 976         /*
 977          * Allocate and initialize the per-ag data.
 978          */
 979         init_rwsem(&mp->m_peraglock);
 980         mp->m_perag =
 981                 kmem_zalloc(sbp->sb_agcount * sizeof(xfs_perag_t), KM_SLEEP);
 982
 983         mp->m_maxagi = xfs_initialize_perag(vfsp, mp, sbp->sb_agcount);
 984
 985         /*
 986          * log's mount-time initialization. Perform 1st part recovery if needed
 987          */
 988         if (likely(sbp->sb_logblocks > 0)) {    /* check for volume case */
 989                 error = xfs_log_mount(mp, mp->m_logdev_targp,
 990                                       XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
 991                                       XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
 992                 if (error) {
 993                         cmn_err(CE_WARN, "XFS: log mount failed");
 994                         goto error2;
 995                 }
 996         } else {        /* No log has been defined */
 997                 cmn_err(CE_WARN, "XFS: no log defined");
 998                 XFS_ERROR_REPORT("xfs_mountfs_int(1)", XFS_ERRLEVEL_LOW, mp);
 999                 error = XFS_ERROR(EFSCORRUPTED);
1000                 goto error2;
1001         }
1002
1003         /*
1004          * Get and sanity-check the root inode.
1005          * Save the pointer to it in the mount structure.
1006          */
1007         error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip, 0);
1008         if (error) {
1009                 cmn_err(CE_WARN, "XFS: failed to read root inode");
1010                 goto error3;
1011         }
1012
1013         ASSERT(rip != NULL);
1014         rvp = XFS_ITOV(rip);
1015
1016         if (unlikely((rip->i_d.di_mode & S_IFMT) != S_IFDIR)) {
1017                 cmn_err(CE_WARN, "XFS: corrupted root inode");
1018                 prdev("Root inode %llu is not a directory",
1019                       mp->m_ddev_targp, (unsigned long long)rip->i_ino);
1020                 xfs_iunlock(rip, XFS_ILOCK_EXCL);
1021                 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
1022                                  mp);
1023                 error = XFS_ERROR(EFSCORRUPTED);
1024                 goto error4;
1025         }
1026         mp->m_rootip = rip;     /* save it */
1027
1028         xfs_iunlock(rip, XFS_ILOCK_EXCL);
1029
1030         /*
1031          * Initialize realtime inode pointers in the mount structure
1032          */
1033         if ((error = xfs_rtmount_inodes(mp))) {
1034                 /*
1035                  * Free up the root inode.
1036                  */
1037                 cmn_err(CE_WARN, "XFS: failed to read RT inodes");
1038                 goto error4;
1039         }
1040
1041         /*
1042          * If fs is not mounted readonly, then update the superblock
1043          * unit and width changes.
1044          */
1045         if (update_flags && !(vfsp->vfs_flag & VFS_RDONLY))
1046                 xfs_mount_log_sbunit(mp, update_flags);
1047
1048         /*
1049          * Initialise the XFS quota management subsystem for this mount
1050          */
1051         if ((error = XFS_QM_INIT(mp, &quotamount, &quotaflags)))
1052                 goto error4;
1053
1054         /*
1055          * Finish recovering the file system.  This part needed to be
1056          * delayed until after the root and real-time bitmap inodes
1057          * were consistently read in.
1058          */
1059         error = xfs_log_mount_finish(mp, mfsi_flags);
1060         if (error) {
1061                 cmn_err(CE_WARN, "XFS: log mount finish failed");
1062                 goto error4;
1063         }
1064
1065         /*
1066          * Complete the quota initialisation, post-log-replay component.
1067          */
1068         if ((error = XFS_QM_MOUNT(mp, quotamount, quotaflags, mfsi_flags)))
1069                 goto error4;
1070
1071         return 0;
1072
1073  error4:
1074         /*
1075          * Free up the root inode.
1076          */
1077         VN_RELE(rvp);
1078  error3:
1079         xfs_log_unmount_dealloc(mp);
1080  error2:
1081         xfs_ihash_free(mp);
1082         xfs_chash_free(mp);
1083         for (agno = 0; agno < sbp->sb_agcount; agno++)
1084                 if (mp->m_perag[agno].pagb_list)
1085                         kmem_free(mp->m_perag[agno].pagb_list,
1086                           sizeof(xfs_perag_busy_t) * XFS_PAGB_NUM_SLOTS);
1087         kmem_free(mp->m_perag, sbp->sb_agcount * sizeof(xfs_perag_t));
1088         mp->m_perag = NULL;
1089         /* FALLTHROUGH */
1090  error1:
1091         if (uuid_mounted)
1092                 xfs_uuid_unmount(mp);
1093         xfs_freesb(mp);
1094         return error;
1095 }
1096
1097 /*
1098  * xfs_unmountfs
1099  *
1100  * This flushes out the inodes,dquots and the superblock, unmounts the
1101  * log and makes sure that incore structures are freed.
1102  */
1103 int
1104 xfs_unmountfs(xfs_mount_t *mp, struct cred *cr)
1105 {
1106         struct vfs      *vfsp = XFS_MTOVFS(mp);
1107 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1108         int64_t         fsid;
1109 #endif
1110
1111         xfs_iflush_all(mp);
1112
1113         XFS_QM_DQPURGEALL(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_UMOUNTING);
1114
1115         /*
1116          * Flush out the log synchronously so that we know for sure
1117          * that nothing is pinned.  This is important because bflush()
1118          * will skip pinned buffers.
1119          */
1120         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1121
1122         xfs_binval(mp->m_ddev_targp);
1123         if (mp->m_rtdev_targp) {
1124                 xfs_binval(mp->m_rtdev_targp);
1125         }
1126
1127         xfs_unmountfs_writesb(mp);
1128
1129         xfs_unmountfs_wait(mp);                 /* wait for async bufs */
1130
1131         xfs_log_unmount(mp);                    /* Done! No more fs ops. */
1132
1133         xfs_freesb(mp);
1134
1135         /*
1136          * All inodes from this mount point should be freed.
1137          */
1138         ASSERT(mp->m_inodes == NULL);
1139
1140         xfs_unmountfs_close(mp, cr);
1141         if ((mp->m_flags & XFS_MOUNT_NOUUID) == 0)
1142                 xfs_uuid_unmount(mp);
1143
1144 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1145         /*
1146          * clear all error tags on this filesystem
1147          */
1148         memcpy(&fsid, &vfsp->vfs_fsid, sizeof(int64_t));
1149         xfs_errortag_clearall_umount(fsid, mp->m_fsname, 0);
1150 #endif
1151         XFS_IODONE(vfsp);
1152         xfs_mount_free(mp, 1);
1153         return 0;
1154 }
1155
1156 void
1157 xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
1158 {
1159         if (mp->m_logdev_targp != mp->m_ddev_targp)
1160                 xfs_free_buftarg(mp->m_logdev_targp, 1);
1161         if (mp->m_rtdev_targp)
1162                 xfs_free_buftarg(mp->m_rtdev_targp, 1);
1163         xfs_free_buftarg(mp->m_ddev_targp, 0);
1164 }
1165
1166 STATIC void
1167 xfs_unmountfs_wait(xfs_mount_t *mp)
1168 {
1169         if (mp->m_logdev_targp != mp->m_ddev_targp)
1170                 xfs_wait_buftarg(mp->m_logdev_targp);
1171         if (mp->m_rtdev_targp)
1172                 xfs_wait_buftarg(mp->m_rtdev_targp);
1173         xfs_wait_buftarg(mp->m_ddev_targp);
1174 }
1175
1176 int
1177 xfs_unmountfs_writesb(xfs_mount_t *mp)
1178 {
1179         xfs_buf_t       *sbp;
1180         xfs_sb_t        *sb;
1181         int             error = 0;
1182
1183         /*
1184          * skip superblock write if fs is read-only, or
1185          * if we are doing a forced umount.
1186          */
1187         sbp = xfs_getsb(mp, 0);
1188         if (!(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY ||
1189                 XFS_FORCED_SHUTDOWN(mp))) {
1190
1191                 xfs_icsb_sync_counters(mp);
1192
1193                 /*
1194                  * mark shared-readonly if desired
1195                  */
1196                 sb = XFS_BUF_TO_SBP(sbp);
1197                 if (mp->m_mk_sharedro) {
1198                         if (!(sb->sb_flags & XFS_SBF_READONLY))
1199                                 sb->sb_flags |= XFS_SBF_READONLY;
1200                         if (!XFS_SB_VERSION_HASSHARED(sb))
1201                                 XFS_SB_VERSION_ADDSHARED(sb);
1202                         xfs_fs_cmn_err(CE_NOTE, mp,
1203                                 "Unmounting, marking shared read-only");
1204                 }
1205                 XFS_BUF_UNDONE(sbp);
1206                 XFS_BUF_UNREAD(sbp);
1207                 XFS_BUF_UNDELAYWRITE(sbp);
1208                 XFS_BUF_WRITE(sbp);
1209                 XFS_BUF_UNASYNC(sbp);
1210                 ASSERT(XFS_BUF_TARGET(sbp) == mp->m_ddev_targp);
1211                 xfsbdstrat(mp, sbp);
1212                 /* Nevermind errors we might get here. */
1213                 error = xfs_iowait(sbp);
1214                 if (error)
1215                         xfs_ioerror_alert("xfs_unmountfs_writesb",
1216                                           mp, sbp, XFS_BUF_ADDR(sbp));
1217                 if (error && mp->m_mk_sharedro)
1218                         xfs_fs_cmn_err(CE_ALERT, mp, "Superblock write error detected while unmounting.  Filesystem may not be marked shared readonly");
1219         }
1220         xfs_buf_relse(sbp);
1221         return error;
1222 }
1223
1224 /*
1225  * xfs_mod_sb() can be used to copy arbitrary changes to the
1226  * in-core superblock into the superblock buffer to be logged.
1227  * It does not provide the higher level of locking that is
1228  * needed to protect the in-core superblock from concurrent
1229  * access.
1230  */
1231 void
1232 xfs_mod_sb(xfs_trans_t *tp, __int64_t fields)
1233 {
1234         xfs_buf_t       *bp;
1235         int             first;
1236         int             last;
1237         xfs_mount_t     *mp;
1238         xfs_sb_t        *sbp;
1239         xfs_sb_field_t  f;
1240
1241         ASSERT(fields);
1242         if (!fields)
1243                 return;
1244         mp = tp->t_mountp;
1245         bp = xfs_trans_getsb(tp, mp, 0);
1246         sbp = XFS_BUF_TO_SBP(bp);
1247         first = sizeof(xfs_sb_t);
1248         last = 0;
1249
1250         /* translate/copy */
1251
1252         xfs_xlatesb(XFS_BUF_PTR(bp), &(mp->m_sb), -1, fields);
1253
1254         /* find modified range */
1255
1256         f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
1257         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1258         first = xfs_sb_info[f].offset;
1259
1260         f = (xfs_sb_field_t)xfs_highbit64((__uint64_t)fields);
1261         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1262         last = xfs_sb_info[f + 1].offset - 1;
1263
1264         xfs_trans_log_buf(tp, bp, first, last);
1265 }
1266 /*
1267  * xfs_mod_incore_sb_unlocked() is a utility routine common used to apply
1268  * a delta to a specified field in the in-core superblock.  Simply
1269  * switch on the field indicated and apply the delta to that field.
1270  * Fields are not allowed to dip below zero, so if the delta would
1271  * do this do not apply it and return EINVAL.
1272  *
1273  * The SB_LOCK must be held when this routine is called.
1274  */
1275 int
1276 xfs_mod_incore_sb_unlocked(xfs_mount_t *mp, xfs_sb_field_t field,
1277                         int delta, int rsvd)
1278 {
1279         int             scounter;       /* short counter for 32 bit fields */
1280         long long       lcounter;       /* long counter for 64 bit fields */
1281         long long       res_used, rem;
1282
1283         /*
1284          * With the in-core superblock spin lock held, switch
1285          * on the indicated field.  Apply the delta to the
1286          * proper field.  If the fields value would dip below
1287          * 0, then do not apply the delta and return EINVAL.
1288          */
1289         switch (field) {
1290         case XFS_SBS_ICOUNT:
1291                 lcounter = (long long)mp->m_sb.sb_icount;
1292                 lcounter += delta;
1293                 if (lcounter < 0) {
1294                         ASSERT(0);
1295                         return XFS_ERROR(EINVAL);
1296                 }
1297                 mp->m_sb.sb_icount = lcounter;
1298                 return 0;
1299         case XFS_SBS_IFREE:
1300                 lcounter = (long long)mp->m_sb.sb_ifree;
1301                 lcounter += delta;
1302                 if (lcounter < 0) {
1303                         ASSERT(0);
1304                         return XFS_ERROR(EINVAL);
1305                 }
1306                 mp->m_sb.sb_ifree = lcounter;
1307                 return 0;
1308         case XFS_SBS_FDBLOCKS:
1309
1310                 lcounter = (long long)mp->m_sb.sb_fdblocks;
1311                 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1312
1313                 if (delta > 0) {                /* Putting blocks back */
1314                         if (res_used > delta) {
1315                                 mp->m_resblks_avail += delta;
1316                         } else {
1317                                 rem = delta - res_used;
1318                                 mp->m_resblks_avail = mp->m_resblks;
1319                                 lcounter += rem;
1320                         }
1321                 } else {                                /* Taking blocks away */
1322
1323                         lcounter += delta;
1324
1325                 /*
1326                  * If were out of blocks, use any available reserved blocks if
1327                  * were allowed to.
1328                  */
1329
1330                         if (lcounter < 0) {
1331                                 if (rsvd) {
1332                                         lcounter = (long long)mp->m_resblks_avail + delta;
1333                                         if (lcounter < 0) {
1334                                                 return XFS_ERROR(ENOSPC);
1335                                         }
1336                                         mp->m_resblks_avail = lcounter;
1337                                         return 0;
1338                                 } else {        /* not reserved */
1339                                         return XFS_ERROR(ENOSPC);
1340                                 }
1341                         }
1342                 }
1343
1344                 mp->m_sb.sb_fdblocks = lcounter;
1345                 return 0;
1346         case XFS_SBS_FREXTENTS:
1347                 lcounter = (long long)mp->m_sb.sb_frextents;
1348                 lcounter += delta;
1349                 if (lcounter < 0) {
1350                         return XFS_ERROR(ENOSPC);
1351                 }
1352                 mp->m_sb.sb_frextents = lcounter;
1353                 return 0;
1354         case XFS_SBS_DBLOCKS:
1355                 lcounter = (long long)mp->m_sb.sb_dblocks;
1356                 lcounter += delta;
1357                 if (lcounter < 0) {
1358                         ASSERT(0);
1359                         return XFS_ERROR(EINVAL);
1360                 }
1361                 mp->m_sb.sb_dblocks = lcounter;
1362                 return 0;
1363         case XFS_SBS_AGCOUNT:
1364                 scounter = mp->m_sb.sb_agcount;
1365                 scounter += delta;
1366                 if (scounter < 0) {
1367                         ASSERT(0);
1368                         return XFS_ERROR(EINVAL);
1369                 }
1370                 mp->m_sb.sb_agcount = scounter;
1371                 return 0;
1372         case XFS_SBS_IMAX_PCT:
1373                 scounter = mp->m_sb.sb_imax_pct;
1374                 scounter += delta;
1375                 if (scounter < 0) {
1376                         ASSERT(0);
1377                         return XFS_ERROR(EINVAL);
1378                 }
1379                 mp->m_sb.sb_imax_pct = scounter;
1380                 return 0;
1381         case XFS_SBS_REXTSIZE:
1382                 scounter = mp->m_sb.sb_rextsize;
1383                 scounter += delta;
1384                 if (scounter < 0) {
1385                         ASSERT(0);
1386                         return XFS_ERROR(EINVAL);
1387                 }
1388                 mp->m_sb.sb_rextsize = scounter;
1389                 return 0;
1390         case XFS_SBS_RBMBLOCKS:
1391                 scounter = mp->m_sb.sb_rbmblocks;
1392                 scounter += delta;
1393                 if (scounter < 0) {
1394                         ASSERT(0);
1395                         return XFS_ERROR(EINVAL);
1396                 }
1397                 mp->m_sb.sb_rbmblocks = scounter;
1398                 return 0;
1399         case XFS_SBS_RBLOCKS:
1400                 lcounter = (long long)mp->m_sb.sb_rblocks;
1401                 lcounter += delta;
1402                 if (lcounter < 0) {
1403                         ASSERT(0);
1404                         return XFS_ERROR(EINVAL);
1405                 }
1406                 mp->m_sb.sb_rblocks = lcounter;
1407                 return 0;
1408         case XFS_SBS_REXTENTS:
1409                 lcounter = (long long)mp->m_sb.sb_rextents;
1410                 lcounter += delta;
1411                 if (lcounter < 0) {
1412                         ASSERT(0);
1413                         return XFS_ERROR(EINVAL);
1414                 }
1415                 mp->m_sb.sb_rextents = lcounter;
1416                 return 0;
1417         case XFS_SBS_REXTSLOG:
1418                 scounter = mp->m_sb.sb_rextslog;
1419                 scounter += delta;
1420                 if (scounter < 0) {
1421                         ASSERT(0);
1422                         return XFS_ERROR(EINVAL);
1423                 }
1424                 mp->m_sb.sb_rextslog = scounter;
1425                 return 0;
1426         default:
1427                 ASSERT(0);
1428                 return XFS_ERROR(EINVAL);
1429         }
1430 }
1431
1432 /*
1433  * xfs_mod_incore_sb() is used to change a field in the in-core
1434  * superblock structure by the specified delta.  This modification
1435  * is protected by the SB_LOCK.  Just use the xfs_mod_incore_sb_unlocked()
1436  * routine to do the work.
1437  */
1438 int
1439 xfs_mod_incore_sb(xfs_mount_t *mp, xfs_sb_field_t field, int delta, int rsvd)
1440 {
1441         unsigned long   s;
1442         int     status;
1443
1444         /* check for per-cpu counters */
1445         switch (field) {
1446 #ifdef HAVE_PERCPU_SB
1447         case XFS_SBS_ICOUNT:
1448         case XFS_SBS_IFREE:
1449         case XFS_SBS_FDBLOCKS:
1450                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1451                         status = xfs_icsb_modify_counters(mp, field,
1452                                                         delta, rsvd);
1453                         break;
1454                 }
1455                 /* FALLTHROUGH */
1456 #endif
1457         default:
1458                 s = XFS_SB_LOCK(mp);
1459                 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1460                 XFS_SB_UNLOCK(mp, s);
1461                 break;
1462         }
1463
1464         return status;
1465 }
1466
1467 /*
1468  * xfs_mod_incore_sb_batch() is used to change more than one field
1469  * in the in-core superblock structure at a time.  This modification
1470  * is protected by a lock internal to this module.  The fields and
1471  * changes to those fields are specified in the array of xfs_mod_sb
1472  * structures passed in.
1473  *
1474  * Either all of the specified deltas will be applied or none of
1475  * them will.  If any modified field dips below 0, then all modifications
1476  * will be backed out and EINVAL will be returned.
1477  */
1478 int
1479 xfs_mod_incore_sb_batch(xfs_mount_t *mp, xfs_mod_sb_t *msb, uint nmsb, int rsvd)
1480 {
1481         unsigned long   s;
1482         int             status=0;
1483         xfs_mod_sb_t    *msbp;
1484
1485         /*
1486          * Loop through the array of mod structures and apply each
1487          * individually.  If any fail, then back out all those
1488          * which have already been applied.  Do all of this within
1489          * the scope of the SB_LOCK so that all of the changes will
1490          * be atomic.
1491          */
1492         s = XFS_SB_LOCK(mp);
1493         msbp = &msb[0];
1494         for (msbp = &msbp[0]; msbp < (msb + nmsb); msbp++) {
1495                 /*
1496                  * Apply the delta at index n.  If it fails, break
1497                  * from the loop so we'll fall into the undo loop
1498                  * below.
1499                  */
1500                 switch (msbp->msb_field) {
1501 #ifdef HAVE_PERCPU_SB
1502                 case XFS_SBS_ICOUNT:
1503                 case XFS_SBS_IFREE:
1504                 case XFS_SBS_FDBLOCKS:
1505                         if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1506                                 status = xfs_icsb_modify_counters_locked(mp,
1507                                                         msbp->msb_field,
1508                                                         msbp->msb_delta, rsvd);
1509                                 break;
1510                         }
1511                         /* FALLTHROUGH */
1512 #endif
1513                 default:
1514                         status = xfs_mod_incore_sb_unlocked(mp,
1515                                                 msbp->msb_field,
1516                                                 msbp->msb_delta, rsvd);
1517                         break;
1518                 }
1519
1520                 if (status != 0) {
1521                         break;
1522                 }
1523         }
1524
1525         /*
1526          * If we didn't complete the loop above, then back out
1527          * any changes made to the superblock.  If you add code
1528          * between the loop above and here, make sure that you
1529          * preserve the value of status. Loop back until
1530          * we step below the beginning of the array.  Make sure
1531          * we don't touch anything back there.
1532          */
1533         if (status != 0) {
1534                 msbp--;
1535                 while (msbp >= msb) {
1536                         switch (msbp->msb_field) {
1537 #ifdef HAVE_PERCPU_SB
1538                         case XFS_SBS_ICOUNT:
1539                         case XFS_SBS_IFREE:
1540                         case XFS_SBS_FDBLOCKS:
1541                                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1542                                         status =
1543                                             xfs_icsb_modify_counters_locked(mp,
1544                                                         msbp->msb_field,
1545                                                         -(msbp->msb_delta),
1546                                                         rsvd);
1547                                         break;
1548                                 }
1549                                 /* FALLTHROUGH */
1550 #endif
1551                         default:
1552                                 status = xfs_mod_incore_sb_unlocked(mp,
1553                                                         msbp->msb_field,
1554                                                         -(msbp->msb_delta),
1555                                                         rsvd);
1556                                 break;
1557                         }
1558                         ASSERT(status == 0);
1559                         msbp--;
1560                 }
1561         }
1562         XFS_SB_UNLOCK(mp, s);
1563         return status;
1564 }
1565
1566 /*
1567  * xfs_getsb() is called to obtain the buffer for the superblock.
1568  * The buffer is returned locked and read in from disk.
1569  * The buffer should be released with a call to xfs_brelse().
1570  *
1571  * If the flags parameter is BUF_TRYLOCK, then we'll only return
1572  * the superblock buffer if it can be locked without sleeping.
1573  * If it can't then we'll return NULL.
1574  */
1575 xfs_buf_t *
1576 xfs_getsb(
1577         xfs_mount_t     *mp,
1578         int             flags)
1579 {
1580         xfs_buf_t       *bp;
1581
1582         ASSERT(mp->m_sb_bp != NULL);
1583         bp = mp->m_sb_bp;
1584         if (flags & XFS_BUF_TRYLOCK) {
1585                 if (!XFS_BUF_CPSEMA(bp)) {
1586                         return NULL;
1587                 }
1588         } else {
1589                 XFS_BUF_PSEMA(bp, PRIBIO);
1590         }
1591         XFS_BUF_HOLD(bp);
1592         ASSERT(XFS_BUF_ISDONE(bp));
1593         return bp;
1594 }
1595
1596 /*
1597  * Used to free the superblock along various error paths.
1598  */
1599 void
1600 xfs_freesb(
1601         xfs_mount_t     *mp)
1602 {
1603         xfs_buf_t       *bp;
1604
1605         /*
1606          * Use xfs_getsb() so that the buffer will be locked
1607          * when we call xfs_buf_relse().
1608          */
1609         bp = xfs_getsb(mp, 0);
1610         XFS_BUF_UNMANAGE(bp);
1611         xfs_buf_relse(bp);
1612         mp->m_sb_bp = NULL;
1613 }
1614
1615 /*
1616  * See if the UUID is unique among mounted XFS filesystems.
1617  * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
1618  */
1619 STATIC int
1620 xfs_uuid_mount(
1621         xfs_mount_t     *mp)
1622 {
1623         if (uuid_is_nil(&mp->m_sb.sb_uuid)) {
1624                 cmn_err(CE_WARN,
1625                         "XFS: Filesystem %s has nil UUID - can't mount",
1626                         mp->m_fsname);
1627                 return -1;
1628         }
1629         if (!uuid_table_insert(&mp->m_sb.sb_uuid)) {
1630                 cmn_err(CE_WARN,
1631                         "XFS: Filesystem %s has duplicate UUID - can't mount",
1632                         mp->m_fsname);
1633                 return -1;
1634         }
1635         return 0;
1636 }
1637
1638 /*
1639  * Remove filesystem from the UUID table.
1640  */
1641 STATIC void
1642 xfs_uuid_unmount(
1643         xfs_mount_t     *mp)
1644 {
1645         uuid_table_remove(&mp->m_sb.sb_uuid);
1646 }
1647
1648 /*
1649  * Used to log changes to the superblock unit and width fields which could
1650  * be altered by the mount options. Only the first superblock is updated.
1651  */
1652 STATIC void
1653 xfs_mount_log_sbunit(
1654         xfs_mount_t     *mp,
1655         __int64_t       fields)
1656 {
1657         xfs_trans_t     *tp;
1658
1659         ASSERT(fields & (XFS_SB_UNIT|XFS_SB_WIDTH|XFS_SB_UUID));
1660
1661         tp = xfs_trans_alloc(mp, XFS_TRANS_SB_UNIT);
1662         if (xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1663                                 XFS_DEFAULT_LOG_COUNT)) {
1664                 xfs_trans_cancel(tp, 0);
1665                 return;
1666         }
1667         xfs_mod_sb(tp, fields);
1668         xfs_trans_commit(tp, 0, NULL);
1669 }
1670
1671
1672 #ifdef HAVE_PERCPU_SB
1673 /*
1674  * Per-cpu incore superblock counters
1675  *
1676  * Simple concept, difficult implementation
1677  *
1678  * Basically, replace the incore superblock counters with a distributed per cpu
1679  * counter for contended fields (e.g.  free block count).
1680  *
1681  * Difficulties arise in that the incore sb is used for ENOSPC checking, and
1682  * hence needs to be accurately read when we are running low on space. Hence
1683  * there is a method to enable and disable the per-cpu counters based on how
1684  * much "stuff" is available in them.
1685  *
1686  * Basically, a counter is enabled if there is enough free resource to justify
1687  * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
1688  * ENOSPC), then we disable the counters to synchronise all callers and
1689  * re-distribute the available resources.
1690  *
1691  * If, once we redistributed the available resources, we still get a failure,
1692  * we disable the per-cpu counter and go through the slow path.
1693  *
1694  * The slow path is the current xfs_mod_incore_sb() function.  This means that
1695  * when we disable a per-cpu counter, we need to drain it's resources back to
1696  * the global superblock. We do this after disabling the counter to prevent
1697  * more threads from queueing up on the counter.
1698  *
1699  * Essentially, this means that we still need a lock in the fast path to enable
1700  * synchronisation between the global counters and the per-cpu counters. This
1701  * is not a problem because the lock will be local to a CPU almost all the time
1702  * and have little contention except when we get to ENOSPC conditions.
1703  *
1704  * Basically, this lock becomes a barrier that enables us to lock out the fast
1705  * path while we do things like enabling and disabling counters and
1706  * synchronising the counters.
1707  *
1708  * Locking rules:
1709  *
1710  *      1. XFS_SB_LOCK() before picking up per-cpu locks
1711  *      2. per-cpu locks always picked up via for_each_online_cpu() order
1712  *      3. accurate counter sync requires XFS_SB_LOCK + per cpu locks
1713  *      4. modifying per-cpu counters requires holding per-cpu lock
1714  *      5. modifying global counters requires holding XFS_SB_LOCK
1715  *      6. enabling or disabling a counter requires holding the XFS_SB_LOCK
1716  *         and _none_ of the per-cpu locks.
1717  *
1718  * Disabled counters are only ever re-enabled by a balance operation
1719  * that results in more free resources per CPU than a given threshold.
1720  * To ensure counters don't remain disabled, they are rebalanced when
1721  * the global resource goes above a higher threshold (i.e. some hysteresis
1722  * is present to prevent thrashing).
1723  */
1724
1725 /*
1726  * hot-plug CPU notifier support.
1727  *
1728  * We cannot use the hotcpu_register() function because it does
1729  * not allow notifier instances. We need a notifier per filesystem
1730  * as we need to be able to identify the filesystem to balance
1731  * the counters out. This is achieved by having a notifier block
1732  * embedded in the xfs_mount_t and doing pointer magic to get the
1733  * mount pointer from the notifier block address.
1734  */
1735 STATIC int
1736 xfs_icsb_cpu_notify(
1737         struct notifier_block *nfb,
1738         unsigned long action,
1739         void *hcpu)
1740 {
1741         xfs_icsb_cnts_t *cntp;
1742         xfs_mount_t     *mp;
1743         int             s;
1744
1745         mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
1746         cntp = (xfs_icsb_cnts_t *)
1747                         per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
1748         switch (action) {
1749         case CPU_UP_PREPARE:
1750                 /* Easy Case - initialize the area and locks, and
1751                  * then rebalance when online does everything else for us. */
1752                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1753                 break;
1754         case CPU_ONLINE:
1755                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
1756                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
1757                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
1758                 break;
1759         case CPU_DEAD:
1760                 /* Disable all the counters, then fold the dead cpu's
1761                  * count into the total on the global superblock and
1762                  * re-enable the counters. */
1763                 s = XFS_SB_LOCK(mp);
1764                 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
1765                 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
1766                 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
1767
1768                 mp->m_sb.sb_icount += cntp->icsb_icount;
1769                 mp->m_sb.sb_ifree += cntp->icsb_ifree;
1770                 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
1771
1772                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1773
1774                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, XFS_ICSB_SB_LOCKED);
1775                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, XFS_ICSB_SB_LOCKED);
1776                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, XFS_ICSB_SB_LOCKED);
1777                 XFS_SB_UNLOCK(mp, s);
1778                 break;
1779         }
1780
1781         return NOTIFY_OK;
1782 }
1783
1784 int
1785 xfs_icsb_init_counters(
1786         xfs_mount_t     *mp)
1787 {
1788         xfs_icsb_cnts_t *cntp;
1789         int             i;
1790
1791         mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
1792         if (mp->m_sb_cnts == NULL)
1793                 return -ENOMEM;
1794
1795         mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
1796         mp->m_icsb_notifier.priority = 0;
1797         register_cpu_notifier(&mp->m_icsb_notifier);
1798
1799         for_each_online_cpu(i) {
1800                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1801                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1802         }
1803         /*
1804          * start with all counters disabled so that the
1805          * initial balance kicks us off correctly
1806          */
1807         mp->m_icsb_counters = -1;
1808         return 0;
1809 }
1810
1811 STATIC void
1812 xfs_icsb_destroy_counters(
1813         xfs_mount_t     *mp)
1814 {
1815         if (mp->m_sb_cnts) {
1816                 unregister_cpu_notifier(&mp->m_icsb_notifier);
1817                 free_percpu(mp->m_sb_cnts);
1818         }
1819 }
1820
1821 STATIC inline void
1822 xfs_icsb_lock_cntr(
1823         xfs_icsb_cnts_t *icsbp)
1824 {
1825         while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
1826                 ndelay(1000);
1827         }
1828 }
1829
1830 STATIC inline void
1831 xfs_icsb_unlock_cntr(
1832         xfs_icsb_cnts_t *icsbp)
1833 {
1834         clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
1835 }
1836
1837
1838 STATIC inline void
1839 xfs_icsb_lock_all_counters(
1840         xfs_mount_t     *mp)
1841 {
1842         xfs_icsb_cnts_t *cntp;
1843         int             i;
1844
1845         for_each_online_cpu(i) {
1846                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1847                 xfs_icsb_lock_cntr(cntp);
1848         }
1849 }
1850
1851 STATIC inline void
1852 xfs_icsb_unlock_all_counters(
1853         xfs_mount_t     *mp)
1854 {
1855         xfs_icsb_cnts_t *cntp;
1856         int             i;
1857
1858         for_each_online_cpu(i) {
1859                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1860                 xfs_icsb_unlock_cntr(cntp);
1861         }
1862 }
1863
1864 STATIC void
1865 xfs_icsb_count(
1866         xfs_mount_t     *mp,
1867         xfs_icsb_cnts_t *cnt,
1868         int             flags)
1869 {
1870         xfs_icsb_cnts_t *cntp;
1871         int             i;
1872
1873         memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
1874
1875         if (!(flags & XFS_ICSB_LAZY_COUNT))
1876                 xfs_icsb_lock_all_counters(mp);
1877
1878         for_each_online_cpu(i) {
1879                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1880                 cnt->icsb_icount += cntp->icsb_icount;
1881                 cnt->icsb_ifree += cntp->icsb_ifree;
1882                 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
1883         }
1884
1885         if (!(flags & XFS_ICSB_LAZY_COUNT))
1886                 xfs_icsb_unlock_all_counters(mp);
1887 }
1888
1889 STATIC int
1890 xfs_icsb_counter_disabled(
1891         xfs_mount_t     *mp,
1892         xfs_sb_field_t  field)
1893 {
1894         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1895         return test_bit(field, &mp->m_icsb_counters);
1896 }
1897
1898 STATIC int
1899 xfs_icsb_disable_counter(
1900         xfs_mount_t     *mp,
1901         xfs_sb_field_t  field)
1902 {
1903         xfs_icsb_cnts_t cnt;
1904
1905         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1906
1907         xfs_icsb_lock_all_counters(mp);
1908         if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
1909                 /* drain back to superblock */
1910
1911                 xfs_icsb_count(mp, &cnt, XFS_ICSB_SB_LOCKED|XFS_ICSB_LAZY_COUNT);
1912                 switch(field) {
1913                 case XFS_SBS_ICOUNT:
1914                         mp->m_sb.sb_icount = cnt.icsb_icount;
1915                         break;
1916                 case XFS_SBS_IFREE:
1917                         mp->m_sb.sb_ifree = cnt.icsb_ifree;
1918                         break;
1919                 case XFS_SBS_FDBLOCKS:
1920                         mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1921                         break;
1922                 default:
1923                         BUG();
1924                 }
1925         }
1926
1927         xfs_icsb_unlock_all_counters(mp);
1928
1929         return 0;
1930 }
1931
1932 STATIC void
1933 xfs_icsb_enable_counter(
1934         xfs_mount_t     *mp,
1935         xfs_sb_field_t  field,
1936         uint64_t        count,
1937         uint64_t        resid)
1938 {
1939         xfs_icsb_cnts_t *cntp;
1940         int             i;
1941
1942         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1943
1944         xfs_icsb_lock_all_counters(mp);
1945         for_each_online_cpu(i) {
1946                 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
1947                 switch (field) {
1948                 case XFS_SBS_ICOUNT:
1949                         cntp->icsb_icount = count + resid;
1950                         break;
1951                 case XFS_SBS_IFREE:
1952                         cntp->icsb_ifree = count + resid;
1953                         break;
1954                 case XFS_SBS_FDBLOCKS:
1955                         cntp->icsb_fdblocks = count + resid;
1956                         break;
1957                 default:
1958                         BUG();
1959                         break;
1960                 }
1961                 resid = 0;
1962         }
1963         clear_bit(field, &mp->m_icsb_counters);
1964         xfs_icsb_unlock_all_counters(mp);
1965 }
1966
1967 STATIC void
1968 xfs_icsb_sync_counters_int(
1969         xfs_mount_t     *mp,
1970         int             flags)
1971 {
1972         xfs_icsb_cnts_t cnt;
1973         int             s;
1974
1975         /* Pass 1: lock all counters */
1976         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
1977                 s = XFS_SB_LOCK(mp);
1978
1979         xfs_icsb_count(mp, &cnt, flags);
1980
1981         /* Step 3: update mp->m_sb fields */
1982         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
1983                 mp->m_sb.sb_icount = cnt.icsb_icount;
1984         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
1985                 mp->m_sb.sb_ifree = cnt.icsb_ifree;
1986         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
1987                 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1988
1989         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
1990                 XFS_SB_UNLOCK(mp, s);
1991 }
1992
1993 /*
1994  * Accurate update of per-cpu counters to incore superblock
1995  */
1996 STATIC void
1997 xfs_icsb_sync_counters(
1998         xfs_mount_t     *mp)
1999 {
2000         xfs_icsb_sync_counters_int(mp, 0);
2001 }
2002
2003 /*
2004  * lazy addition used for things like df, background sb syncs, etc
2005  */
2006 void
2007 xfs_icsb_sync_counters_lazy(
2008         xfs_mount_t     *mp)
2009 {
2010         xfs_icsb_sync_counters_int(mp, XFS_ICSB_LAZY_COUNT);
2011 }
2012
2013 /*
2014  * Balance and enable/disable counters as necessary.
2015  *
2016  * Thresholds for re-enabling counters are somewhat magic.
2017  * inode counts are chosen to be the same number as single
2018  * on disk allocation chunk per CPU, and free blocks is
2019  * something far enough zero that we aren't going thrash
2020  * when we get near ENOSPC.
2021  */
2022 #define XFS_ICSB_INO_CNTR_REENABLE      64
2023 #define XFS_ICSB_FDBLK_CNTR_REENABLE    512
2024 STATIC void
2025 xfs_icsb_balance_counter(
2026         xfs_mount_t     *mp,
2027         xfs_sb_field_t  field,
2028         int             flags)
2029 {
2030         uint64_t        count, resid = 0;
2031         int             weight = num_online_cpus();
2032         int             s;
2033
2034         if (!(flags & XFS_ICSB_SB_LOCKED))
2035                 s = XFS_SB_LOCK(mp);
2036
2037         /* disable counter and sync counter */
2038         xfs_icsb_disable_counter(mp, field);
2039
2040         /* update counters  - first CPU gets residual*/
2041         switch (field) {
2042         case XFS_SBS_ICOUNT:
2043                 count = mp->m_sb.sb_icount;
2044                 resid = do_div(count, weight);
2045                 if (count < XFS_ICSB_INO_CNTR_REENABLE)
2046                         goto out;
2047                 break;
2048         case XFS_SBS_IFREE:
2049                 count = mp->m_sb.sb_ifree;
2050                 resid = do_div(count, weight);
2051                 if (count < XFS_ICSB_INO_CNTR_REENABLE)
2052                         goto out;
2053                 break;
2054         case XFS_SBS_FDBLOCKS:
2055                 count = mp->m_sb.sb_fdblocks;
2056                 resid = do_div(count, weight);
2057                 if (count < XFS_ICSB_FDBLK_CNTR_REENABLE)
2058                         goto out;
2059                 break;
2060         default:
2061                 BUG();
2062                 break;
2063         }
2064
2065         xfs_icsb_enable_counter(mp, field, count, resid);
2066 out:
2067         if (!(flags & XFS_ICSB_SB_LOCKED))
2068                 XFS_SB_UNLOCK(mp, s);
2069 }
2070
2071 STATIC int
2072 xfs_icsb_modify_counters_int(
2073         xfs_mount_t     *mp,
2074         xfs_sb_field_t  field,
2075         int             delta,
2076         int             rsvd,
2077         int             flags)
2078 {
2079         xfs_icsb_cnts_t *icsbp;
2080         long long       lcounter;       /* long counter for 64 bit fields */
2081         int             cpu, s, locked = 0;
2082         int             ret = 0, balance_done = 0;
2083
2084 again:
2085         cpu = get_cpu();
2086         icsbp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, cpu),
2087         xfs_icsb_lock_cntr(icsbp);
2088         if (unlikely(xfs_icsb_counter_disabled(mp, field)))
2089                 goto slow_path;
2090
2091         switch (field) {
2092         case XFS_SBS_ICOUNT:
2093                 lcounter = icsbp->icsb_icount;
2094                 lcounter += delta;
2095                 if (unlikely(lcounter < 0))
2096                         goto slow_path;
2097                 icsbp->icsb_icount = lcounter;
2098                 break;
2099
2100         case XFS_SBS_IFREE:
2101                 lcounter = icsbp->icsb_ifree;
2102                 lcounter += delta;
2103                 if (unlikely(lcounter < 0))
2104                         goto slow_path;
2105                 icsbp->icsb_ifree = lcounter;
2106                 break;
2107
2108         case XFS_SBS_FDBLOCKS:
2109                 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
2110
2111                 lcounter = icsbp->icsb_fdblocks;
2112                 lcounter += delta;
2113                 if (unlikely(lcounter < 0))
2114                         goto slow_path;
2115                 icsbp->icsb_fdblocks = lcounter;
2116                 break;
2117         default:
2118                 BUG();
2119                 break;
2120         }
2121         xfs_icsb_unlock_cntr(icsbp);
2122         put_cpu();
2123         if (locked)
2124                 XFS_SB_UNLOCK(mp, s);
2125         return 0;
2126
2127         /*
2128          * The slow path needs to be run with the SBLOCK
2129          * held so that we prevent other threads from
2130          * attempting to run this path at the same time.
2131          * this provides exclusion for the balancing code,
2132          * and exclusive fallback if the balance does not
2133          * provide enough resources to continue in an unlocked
2134          * manner.
2135          */
2136 slow_path:
2137         xfs_icsb_unlock_cntr(icsbp);
2138         put_cpu();
2139
2140         /* need to hold superblock incase we need
2141          * to disable a counter */
2142         if (!(flags & XFS_ICSB_SB_LOCKED)) {
2143                 s = XFS_SB_LOCK(mp);
2144                 locked = 1;
2145                 flags |= XFS_ICSB_SB_LOCKED;
2146         }
2147         if (!balance_done) {
2148                 xfs_icsb_balance_counter(mp, field, flags);
2149                 balance_done = 1;
2150                 goto again;
2151         } else {
2152                 /*
2153                  * we might not have enough on this local
2154                  * cpu to allocate for a bulk request.
2155                  * We need to drain this field from all CPUs
2156                  * and disable the counter fastpath
2157                  */
2158                 xfs_icsb_disable_counter(mp, field);
2159         }
2160
2161         ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
2162
2163         if (locked)
2164                 XFS_SB_UNLOCK(mp, s);
2165         return ret;
2166 }
2167
2168 STATIC int
2169 xfs_icsb_modify_counters(
2170         xfs_mount_t     *mp,
2171         xfs_sb_field_t  field,
2172         int             delta,
2173         int             rsvd)
2174 {
2175         return xfs_icsb_modify_counters_int(mp, field, delta, rsvd, 0);
2176 }
2177
2178 /*
2179  * Called when superblock is already locked
2180  */
2181 STATIC int
2182 xfs_icsb_modify_counters_locked(
2183         xfs_mount_t     *mp,
2184         xfs_sb_field_t  field,
2185         int             delta,
2186         int             rsvd)
2187 {
2188         return xfs_icsb_modify_counters_int(mp, field, delta,
2189                                                 rsvd, XFS_ICSB_SB_LOCKED);
2190 }
2191 #endif