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_rextsize * sbp->sb_blocksize > XFS_MAX_RTEXTSIZE)  ||
 272             (sbp->sb_rextsize * sbp->sb_blocksize < XFS_MIN_RTEXTSIZE)  ||
 273             sbp->sb_imax_pct > 100)) {
 274                 cmn_err(CE_WARN, "XFS: SB sanity check 1 failed");
 275                 XFS_CORRUPTION_ERROR("xfs_mount_validate_sb(3)",
 276                                      XFS_ERRLEVEL_LOW, mp, sbp);
 277                 return XFS_ERROR(EFSCORRUPTED);
 278         }
 279
 280         /*
 281          * Sanity check AG count, size fields against data size field
 282          */
 283         if (unlikely(
 284             sbp->sb_dblocks == 0 ||
 285             sbp->sb_dblocks >
 286              (xfs_drfsbno_t)sbp->sb_agcount * sbp->sb_agblocks ||
 287             sbp->sb_dblocks < (xfs_drfsbno_t)(sbp->sb_agcount - 1) *
 288                               sbp->sb_agblocks + XFS_MIN_AG_BLOCKS)) {
 289                 cmn_err(CE_WARN, "XFS: SB sanity check 2 failed");
 290                 XFS_ERROR_REPORT("xfs_mount_validate_sb(4)",
 291                                  XFS_ERRLEVEL_LOW, mp);
 292                 return XFS_ERROR(EFSCORRUPTED);
 293         }
 294
 295         ASSERT(PAGE_SHIFT >= sbp->sb_blocklog);
 296         ASSERT(sbp->sb_blocklog >= BBSHIFT);
 297
 298 #if XFS_BIG_BLKNOS     /* Limited by ULONG_MAX of page cache index */
 299         if (unlikely(
 300             (sbp->sb_dblocks >> (PAGE_SHIFT - sbp->sb_blocklog)) > ULONG_MAX ||
 301             (sbp->sb_rblocks >> (PAGE_SHIFT - sbp->sb_blocklog)) > ULONG_MAX)) {
 302 #else                  /* Limited by UINT_MAX of sectors */
 303         if (unlikely(
 304             (sbp->sb_dblocks << (sbp->sb_blocklog - BBSHIFT)) > UINT_MAX ||
 305             (sbp->sb_rblocks << (sbp->sb_blocklog - BBSHIFT)) > UINT_MAX)) {
 306 #endif
 307                 cmn_err(CE_WARN,
 308         "XFS: File system is too large to be mounted on this system.");
 309                 return XFS_ERROR(E2BIG);
 310         }
 311
 312         if (unlikely(sbp->sb_inprogress)) {
 313                 cmn_err(CE_WARN, "XFS: file system busy");
 314                 XFS_ERROR_REPORT("xfs_mount_validate_sb(5)",
 315                                  XFS_ERRLEVEL_LOW, mp);
 316                 return XFS_ERROR(EFSCORRUPTED);
 317         }
 318
 319         /*
 320          * Version 1 directory format has never worked on Linux.
 321          */
 322         if (unlikely(!XFS_SB_VERSION_HASDIRV2(sbp))) {
 323                 cmn_err(CE_WARN,
 324         "XFS: Attempted to mount file system using version 1 directory format");
 325                 return XFS_ERROR(ENOSYS);
 326         }
 327
 328         /*
 329          * Until this is fixed only page-sized or smaller data blocks work.
 330          */
 331         if (unlikely(sbp->sb_blocksize > PAGE_SIZE)) {
 332                 cmn_err(CE_WARN,
 333                 "XFS: Attempted to mount file system with blocksize %d bytes",
 334                         sbp->sb_blocksize);
 335                 cmn_err(CE_WARN,
 336                 "XFS: Only page-sized (%ld) or less blocksizes currently work.",
 337                         PAGE_SIZE);
 338                 return XFS_ERROR(ENOSYS);
 339         }
 340
 341         return 0;
 342 }
 343
 344 xfs_agnumber_t
 345 xfs_initialize_perag(
 346         struct vfs      *vfs,
 347         xfs_mount_t     *mp,
 348         xfs_agnumber_t  agcount)
 349 {
 350         xfs_agnumber_t  index, max_metadata;
 351         xfs_perag_t     *pag;
 352         xfs_agino_t     agino;
 353         xfs_ino_t       ino;
 354         xfs_sb_t        *sbp = &mp->m_sb;
 355         xfs_ino_t       max_inum = XFS_MAXINUMBER_32;
 356
 357         /* Check to see if the filesystem can overflow 32 bit inodes */
 358         agino = XFS_OFFBNO_TO_AGINO(mp, sbp->sb_agblocks - 1, 0);
 359         ino = XFS_AGINO_TO_INO(mp, agcount - 1, agino);
 360
 361         /* Clear the mount flag if no inode can overflow 32 bits
 362          * on this filesystem, or if specifically requested..
 363          */
 364         if ((vfs->vfs_flag & VFS_32BITINODES) && ino > max_inum) {
 365                 mp->m_flags |= XFS_MOUNT_32BITINODES;
 366         } else {
 367                 mp->m_flags &= ~XFS_MOUNT_32BITINODES;
 368         }
 369
 370         /* If we can overflow then setup the ag headers accordingly */
 371         if (mp->m_flags & XFS_MOUNT_32BITINODES) {
 372                 /* Calculate how much should be reserved for inodes to
 373                  * meet the max inode percentage.
 374                  */
 375                 if (mp->m_maxicount) {
 376                         __uint64_t      icount;
 377
 378                         icount = sbp->sb_dblocks * sbp->sb_imax_pct;
 379                         do_div(icount, 100);
 380                         icount += sbp->sb_agblocks - 1;
 381                         do_div(icount, sbp->sb_agblocks);
 382                         max_metadata = icount;
 383                 } else {
 384                         max_metadata = agcount;
 385                 }
 386                 for (index = 0; index < agcount; index++) {
 387                         ino = XFS_AGINO_TO_INO(mp, index, agino);
 388                         if (ino > max_inum) {
 389                                 index++;
 390                                 break;
 391                         }
 392
 393                         /* This ag is prefered for inodes */
 394                         pag = &mp->m_perag[index];
 395                         pag->pagi_inodeok = 1;
 396                         if (index < max_metadata)
 397                                 pag->pagf_metadata = 1;
 398                 }
 399         } else {
 400                 /* Setup default behavior for smaller filesystems */
 401                 for (index = 0; index < agcount; index++) {
 402                         pag = &mp->m_perag[index];
 403                         pag->pagi_inodeok = 1;
 404                 }
 405         }
 406         return index;
 407 }
 408
 409 /*
 410  * xfs_xlatesb
 411  *
 412  *     data       - on disk version of sb
 413  *     sb         - a superblock
 414  *     dir        - conversion direction: <0 - convert sb to buf
 415  *                                        >0 - convert buf to sb
 416  *     fields     - which fields to copy (bitmask)
 417  */
 418 void
 419 xfs_xlatesb(
 420         void            *data,
 421         xfs_sb_t        *sb,
 422         int             dir,
 423         __int64_t       fields)
 424 {
 425         xfs_caddr_t     buf_ptr;
 426         xfs_caddr_t     mem_ptr;
 427         xfs_sb_field_t  f;
 428         int             first;
 429         int             size;
 430
 431         ASSERT(dir);
 432         ASSERT(fields);
 433
 434         if (!fields)
 435                 return;
 436
 437         buf_ptr = (xfs_caddr_t)data;
 438         mem_ptr = (xfs_caddr_t)sb;
 439
 440         while (fields) {
 441                 f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
 442                 first = xfs_sb_info[f].offset;
 443                 size = xfs_sb_info[f + 1].offset - first;
 444
 445                 ASSERT(xfs_sb_info[f].type == 0 || xfs_sb_info[f].type == 1);
 446
 447                 if (size == 1 || xfs_sb_info[f].type == 1) {
 448                         if (dir > 0) {
 449                                 memcpy(mem_ptr + first, buf_ptr + first, size);
 450                         } else {
 451                                 memcpy(buf_ptr + first, mem_ptr + first, size);
 452                         }
 453                 } else {
 454                         switch (size) {
 455                         case 2:
 456                                 INT_XLATE(*(__uint16_t*)(buf_ptr+first),
 457                                           *(__uint16_t*)(mem_ptr+first),
 458                                           dir, ARCH_CONVERT);
 459                                 break;
 460                         case 4:
 461                                 INT_XLATE(*(__uint32_t*)(buf_ptr+first),
 462                                           *(__uint32_t*)(mem_ptr+first),
 463                                           dir, ARCH_CONVERT);
 464                                 break;
 465                         case 8:
 466                                 INT_XLATE(*(__uint64_t*)(buf_ptr+first),
 467                                           *(__uint64_t*)(mem_ptr+first), dir, ARCH_CONVERT);
 468                                 break;
 469                         default:
 470                                 ASSERT(0);
 471                         }
 472                 }
 473
 474                 fields &= ~(1LL << f);
 475         }
 476 }
 477
 478 /*
 479  * xfs_readsb
 480  *
 481  * Does the initial read of the superblock.
 482  */
 483 int
 484 xfs_readsb(xfs_mount_t *mp)
 485 {
 486         unsigned int    sector_size;
 487         unsigned int    extra_flags;
 488         xfs_buf_t       *bp;
 489         xfs_sb_t        *sbp;
 490         int             error;
 491
 492         ASSERT(mp->m_sb_bp == NULL);
 493         ASSERT(mp->m_ddev_targp != NULL);
 494
 495         /*
 496          * Allocate a (locked) buffer to hold the superblock.
 497          * This will be kept around at all times to optimize
 498          * access to the superblock.
 499          */
 500         sector_size = xfs_getsize_buftarg(mp->m_ddev_targp);
 501         extra_flags = XFS_BUF_LOCK | XFS_BUF_MANAGE | XFS_BUF_MAPPED;
 502
 503         bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
 504                                 BTOBB(sector_size), extra_flags);
 505         if (!bp || XFS_BUF_ISERROR(bp)) {
 506                 cmn_err(CE_WARN, "XFS: SB read failed");
 507                 error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
 508                 goto fail;
 509         }
 510         ASSERT(XFS_BUF_ISBUSY(bp));
 511         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 512
 513         /*
 514          * Initialize the mount structure from the superblock.
 515          * But first do some basic consistency checking.
 516          */
 517         sbp = XFS_BUF_TO_SBP(bp);
 518         xfs_xlatesb(XFS_BUF_PTR(bp), &(mp->m_sb), 1, XFS_SB_ALL_BITS);
 519
 520         error = xfs_mount_validate_sb(mp, &(mp->m_sb));
 521         if (error) {
 522                 cmn_err(CE_WARN, "XFS: SB validate failed");
 523                 goto fail;
 524         }
 525
 526         /*
 527          * We must be able to do sector-sized and sector-aligned IO.
 528          */
 529         if (sector_size > mp->m_sb.sb_sectsize) {
 530                 cmn_err(CE_WARN,
 531                         "XFS: device supports only %u byte sectors (not %u)",
 532                         sector_size, mp->m_sb.sb_sectsize);
 533                 error = ENOSYS;
 534                 goto fail;
 535         }
 536
 537         /*
 538          * If device sector size is smaller than the superblock size,
 539          * re-read the superblock so the buffer is correctly sized.
 540          */
 541         if (sector_size < mp->m_sb.sb_sectsize) {
 542                 XFS_BUF_UNMANAGE(bp);
 543                 xfs_buf_relse(bp);
 544                 sector_size = mp->m_sb.sb_sectsize;
 545                 bp = xfs_buf_read_flags(mp->m_ddev_targp, XFS_SB_DADDR,
 546                                         BTOBB(sector_size), extra_flags);
 547                 if (!bp || XFS_BUF_ISERROR(bp)) {
 548                         cmn_err(CE_WARN, "XFS: SB re-read failed");
 549                         error = bp ? XFS_BUF_GETERROR(bp) : ENOMEM;
 550                         goto fail;
 551                 }
 552                 ASSERT(XFS_BUF_ISBUSY(bp));
 553                 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 554         }
 555
 556         xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
 557         xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
 558         xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
 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         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         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))) {
 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         /*
 944          * Select the right directory manager.
 945          */
 946         mp->m_dirops =
 947                 XFS_SB_VERSION_HASDIRV2(&mp->m_sb) ?
 948                         xfsv2_dirops :
 949                         xfsv1_dirops;
 950
 951         /*
 952          * Initialize directory manager's entries.
 953          */
 954         XFS_DIR_MOUNT(mp);
 955
 956         /*
 957          * Initialize the attribute manager's entries.
 958          */
 959         mp->m_attr_magicpct = (mp->m_sb.sb_blocksize * 37) / 100;
 960
 961         /*
 962          * Initialize the precomputed transaction reservations values.
 963          */
 964         xfs_trans_init(mp);
 965
 966         /*
 967          * Allocate and initialize the inode hash table for this
 968          * file system.
 969          */
 970         xfs_ihash_init(mp);
 971         xfs_chash_init(mp);
 972
 973         /*
 974          * Allocate and initialize the per-ag data.
 975          */
 976         init_rwsem(&mp->m_peraglock);
 977         mp->m_perag =
 978                 kmem_zalloc(sbp->sb_agcount * sizeof(xfs_perag_t), KM_SLEEP);
 979
 980         mp->m_maxagi = xfs_initialize_perag(vfsp, mp, sbp->sb_agcount);
 981
 982         /*
 983          * log's mount-time initialization. Perform 1st part recovery if needed
 984          */
 985         if (likely(sbp->sb_logblocks > 0)) {    /* check for volume case */
 986                 error = xfs_log_mount(mp, mp->m_logdev_targp,
 987                                       XFS_FSB_TO_DADDR(mp, sbp->sb_logstart),
 988                                       XFS_FSB_TO_BB(mp, sbp->sb_logblocks));
 989                 if (error) {
 990                         cmn_err(CE_WARN, "XFS: log mount failed");
 991                         goto error2;
 992                 }
 993         } else {        /* No log has been defined */
 994                 cmn_err(CE_WARN, "XFS: no log defined");
 995                 XFS_ERROR_REPORT("xfs_mountfs_int(1)", XFS_ERRLEVEL_LOW, mp);
 996                 error = XFS_ERROR(EFSCORRUPTED);
 997                 goto error2;
 998         }
 999
1000         /*
1001          * Get and sanity-check the root inode.
1002          * Save the pointer to it in the mount structure.
1003          */
1004         error = xfs_iget(mp, NULL, sbp->sb_rootino, 0, XFS_ILOCK_EXCL, &rip, 0);
1005         if (error) {
1006                 cmn_err(CE_WARN, "XFS: failed to read root inode");
1007                 goto error3;
1008         }
1009
1010         ASSERT(rip != NULL);
1011         rvp = XFS_ITOV(rip);
1012
1013         if (unlikely((rip->i_d.di_mode & S_IFMT) != S_IFDIR)) {
1014                 cmn_err(CE_WARN, "XFS: corrupted root inode");
1015                 prdev("Root inode %llu is not a directory",
1016                       mp->m_ddev_targp, (unsigned long long)rip->i_ino);
1017                 xfs_iunlock(rip, XFS_ILOCK_EXCL);
1018                 XFS_ERROR_REPORT("xfs_mountfs_int(2)", XFS_ERRLEVEL_LOW,
1019                                  mp);
1020                 error = XFS_ERROR(EFSCORRUPTED);
1021                 goto error4;
1022         }
1023         mp->m_rootip = rip;     /* save it */
1024
1025         xfs_iunlock(rip, XFS_ILOCK_EXCL);
1026
1027         /*
1028          * Initialize realtime inode pointers in the mount structure
1029          */
1030         if ((error = xfs_rtmount_inodes(mp))) {
1031                 /*
1032                  * Free up the root inode.
1033                  */
1034                 cmn_err(CE_WARN, "XFS: failed to read RT inodes");
1035                 goto error4;
1036         }
1037
1038         /*
1039          * If fs is not mounted readonly, then update the superblock
1040          * unit and width changes.
1041          */
1042         if (update_flags && !(vfsp->vfs_flag & VFS_RDONLY))
1043                 xfs_mount_log_sbunit(mp, update_flags);
1044
1045         /*
1046          * Initialise the XFS quota management subsystem for this mount
1047          */
1048         if ((error = XFS_QM_INIT(mp, &quotamount, &quotaflags)))
1049                 goto error4;
1050
1051         /*
1052          * Finish recovering the file system.  This part needed to be
1053          * delayed until after the root and real-time bitmap inodes
1054          * were consistently read in.
1055          */
1056         error = xfs_log_mount_finish(mp, mfsi_flags);
1057         if (error) {
1058                 cmn_err(CE_WARN, "XFS: log mount finish failed");
1059                 goto error4;
1060         }
1061
1062         /*
1063          * Complete the quota initialisation, post-log-replay component.
1064          */
1065         if ((error = XFS_QM_MOUNT(mp, quotamount, quotaflags, mfsi_flags)))
1066                 goto error4;
1067
1068         return 0;
1069
1070  error4:
1071         /*
1072          * Free up the root inode.
1073          */
1074         VN_RELE(rvp);
1075  error3:
1076         xfs_log_unmount_dealloc(mp);
1077  error2:
1078         xfs_ihash_free(mp);
1079         xfs_chash_free(mp);
1080         for (agno = 0; agno < sbp->sb_agcount; agno++)
1081                 if (mp->m_perag[agno].pagb_list)
1082                         kmem_free(mp->m_perag[agno].pagb_list,
1083                           sizeof(xfs_perag_busy_t) * XFS_PAGB_NUM_SLOTS);
1084         kmem_free(mp->m_perag, sbp->sb_agcount * sizeof(xfs_perag_t));
1085         mp->m_perag = NULL;
1086         /* FALLTHROUGH */
1087  error1:
1088         if (uuid_mounted)
1089                 xfs_uuid_unmount(mp);
1090         xfs_freesb(mp);
1091         return error;
1092 }
1093
1094 /*
1095  * xfs_unmountfs
1096  *
1097  * This flushes out the inodes,dquots and the superblock, unmounts the
1098  * log and makes sure that incore structures are freed.
1099  */
1100 int
1101 xfs_unmountfs(xfs_mount_t *mp, struct cred *cr)
1102 {
1103         struct vfs      *vfsp = XFS_MTOVFS(mp);
1104 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1105         int64_t         fsid;
1106 #endif
1107
1108         xfs_iflush_all(mp);
1109
1110         XFS_QM_DQPURGEALL(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_UMOUNTING);
1111
1112         /*
1113          * Flush out the log synchronously so that we know for sure
1114          * that nothing is pinned.  This is important because bflush()
1115          * will skip pinned buffers.
1116          */
1117         xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE | XFS_LOG_SYNC);
1118
1119         xfs_binval(mp->m_ddev_targp);
1120         if (mp->m_rtdev_targp) {
1121                 xfs_binval(mp->m_rtdev_targp);
1122         }
1123
1124         xfs_unmountfs_writesb(mp);
1125
1126         xfs_unmountfs_wait(mp);                 /* wait for async bufs */
1127
1128         xfs_log_unmount(mp);                    /* Done! No more fs ops. */
1129
1130         xfs_freesb(mp);
1131
1132         /*
1133          * All inodes from this mount point should be freed.
1134          */
1135         ASSERT(mp->m_inodes == NULL);
1136
1137         xfs_unmountfs_close(mp, cr);
1138         if ((mp->m_flags & XFS_MOUNT_NOUUID) == 0)
1139                 xfs_uuid_unmount(mp);
1140
1141 #if defined(DEBUG) || defined(INDUCE_IO_ERROR)
1142         /*
1143          * clear all error tags on this filesystem
1144          */
1145         memcpy(&fsid, &vfsp->vfs_fsid, sizeof(int64_t));
1146         xfs_errortag_clearall_umount(fsid, mp->m_fsname, 0);
1147 #endif
1148         XFS_IODONE(vfsp);
1149         xfs_mount_free(mp, 1);
1150         return 0;
1151 }
1152
1153 void
1154 xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
1155 {
1156         if (mp->m_logdev_targp != mp->m_ddev_targp)
1157                 xfs_free_buftarg(mp->m_logdev_targp, 1);
1158         if (mp->m_rtdev_targp)
1159                 xfs_free_buftarg(mp->m_rtdev_targp, 1);
1160         xfs_free_buftarg(mp->m_ddev_targp, 0);
1161 }
1162
1163 STATIC void
1164 xfs_unmountfs_wait(xfs_mount_t *mp)
1165 {
1166         if (mp->m_logdev_targp != mp->m_ddev_targp)
1167                 xfs_wait_buftarg(mp->m_logdev_targp);
1168         if (mp->m_rtdev_targp)
1169                 xfs_wait_buftarg(mp->m_rtdev_targp);
1170         xfs_wait_buftarg(mp->m_ddev_targp);
1171 }
1172
1173 int
1174 xfs_unmountfs_writesb(xfs_mount_t *mp)
1175 {
1176         xfs_buf_t       *sbp;
1177         xfs_sb_t        *sb;
1178         int             error = 0;
1179
1180         /*
1181          * skip superblock write if fs is read-only, or
1182          * if we are doing a forced umount.
1183          */
1184         sbp = xfs_getsb(mp, 0);
1185         if (!(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY ||
1186                 XFS_FORCED_SHUTDOWN(mp))) {
1187
1188                 xfs_icsb_sync_counters(mp);
1189
1190                 /*
1191                  * mark shared-readonly if desired
1192                  */
1193                 sb = XFS_BUF_TO_SBP(sbp);
1194                 if (mp->m_mk_sharedro) {
1195                         if (!(sb->sb_flags & XFS_SBF_READONLY))
1196                                 sb->sb_flags |= XFS_SBF_READONLY;
1197                         if (!XFS_SB_VERSION_HASSHARED(sb))
1198                                 XFS_SB_VERSION_ADDSHARED(sb);
1199                         xfs_fs_cmn_err(CE_NOTE, mp,
1200                                 "Unmounting, marking shared read-only");
1201                 }
1202                 XFS_BUF_UNDONE(sbp);
1203                 XFS_BUF_UNREAD(sbp);
1204                 XFS_BUF_UNDELAYWRITE(sbp);
1205                 XFS_BUF_WRITE(sbp);
1206                 XFS_BUF_UNASYNC(sbp);
1207                 ASSERT(XFS_BUF_TARGET(sbp) == mp->m_ddev_targp);
1208                 xfsbdstrat(mp, sbp);
1209                 /* Nevermind errors we might get here. */
1210                 error = xfs_iowait(sbp);
1211                 if (error)
1212                         xfs_ioerror_alert("xfs_unmountfs_writesb",
1213                                           mp, sbp, XFS_BUF_ADDR(sbp));
1214                 if (error && mp->m_mk_sharedro)
1215                         xfs_fs_cmn_err(CE_ALERT, mp, "Superblock write error detected while unmounting.  Filesystem may not be marked shared readonly");
1216         }
1217         xfs_buf_relse(sbp);
1218         return error;
1219 }
1220
1221 /*
1222  * xfs_mod_sb() can be used to copy arbitrary changes to the
1223  * in-core superblock into the superblock buffer to be logged.
1224  * It does not provide the higher level of locking that is
1225  * needed to protect the in-core superblock from concurrent
1226  * access.
1227  */
1228 void
1229 xfs_mod_sb(xfs_trans_t *tp, __int64_t fields)
1230 {
1231         xfs_buf_t       *bp;
1232         int             first;
1233         int             last;
1234         xfs_mount_t     *mp;
1235         xfs_sb_t        *sbp;
1236         xfs_sb_field_t  f;
1237
1238         ASSERT(fields);
1239         if (!fields)
1240                 return;
1241         mp = tp->t_mountp;
1242         bp = xfs_trans_getsb(tp, mp, 0);
1243         sbp = XFS_BUF_TO_SBP(bp);
1244         first = sizeof(xfs_sb_t);
1245         last = 0;
1246
1247         /* translate/copy */
1248
1249         xfs_xlatesb(XFS_BUF_PTR(bp), &(mp->m_sb), -1, fields);
1250
1251         /* find modified range */
1252
1253         f = (xfs_sb_field_t)xfs_lowbit64((__uint64_t)fields);
1254         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1255         first = xfs_sb_info[f].offset;
1256
1257         f = (xfs_sb_field_t)xfs_highbit64((__uint64_t)fields);
1258         ASSERT((1LL << f) & XFS_SB_MOD_BITS);
1259         last = xfs_sb_info[f + 1].offset - 1;
1260
1261         xfs_trans_log_buf(tp, bp, first, last);
1262 }
1263 /*
1264  * xfs_mod_incore_sb_unlocked() is a utility routine common used to apply
1265  * a delta to a specified field in the in-core superblock.  Simply
1266  * switch on the field indicated and apply the delta to that field.
1267  * Fields are not allowed to dip below zero, so if the delta would
1268  * do this do not apply it and return EINVAL.
1269  *
1270  * The SB_LOCK must be held when this routine is called.
1271  */
1272 int
1273 xfs_mod_incore_sb_unlocked(xfs_mount_t *mp, xfs_sb_field_t field,
1274                         int delta, int rsvd)
1275 {
1276         int             scounter;       /* short counter for 32 bit fields */
1277         long long       lcounter;       /* long counter for 64 bit fields */
1278         long long       res_used, rem;
1279
1280         /*
1281          * With the in-core superblock spin lock held, switch
1282          * on the indicated field.  Apply the delta to the
1283          * proper field.  If the fields value would dip below
1284          * 0, then do not apply the delta and return EINVAL.
1285          */
1286         switch (field) {
1287         case XFS_SBS_ICOUNT:
1288                 lcounter = (long long)mp->m_sb.sb_icount;
1289                 lcounter += delta;
1290                 if (lcounter < 0) {
1291                         ASSERT(0);
1292                         return XFS_ERROR(EINVAL);
1293                 }
1294                 mp->m_sb.sb_icount = lcounter;
1295                 return 0;
1296         case XFS_SBS_IFREE:
1297                 lcounter = (long long)mp->m_sb.sb_ifree;
1298                 lcounter += delta;
1299                 if (lcounter < 0) {
1300                         ASSERT(0);
1301                         return XFS_ERROR(EINVAL);
1302                 }
1303                 mp->m_sb.sb_ifree = lcounter;
1304                 return 0;
1305         case XFS_SBS_FDBLOCKS:
1306
1307                 lcounter = (long long)mp->m_sb.sb_fdblocks;
1308                 res_used = (long long)(mp->m_resblks - mp->m_resblks_avail);
1309
1310                 if (delta > 0) {                /* Putting blocks back */
1311                         if (res_used > delta) {
1312                                 mp->m_resblks_avail += delta;
1313                         } else {
1314                                 rem = delta - res_used;
1315                                 mp->m_resblks_avail = mp->m_resblks;
1316                                 lcounter += rem;
1317                         }
1318                 } else {                                /* Taking blocks away */
1319
1320                         lcounter += delta;
1321
1322                 /*
1323                  * If were out of blocks, use any available reserved blocks if
1324                  * were allowed to.
1325                  */
1326
1327                         if (lcounter < 0) {
1328                                 if (rsvd) {
1329                                         lcounter = (long long)mp->m_resblks_avail + delta;
1330                                         if (lcounter < 0) {
1331                                                 return XFS_ERROR(ENOSPC);
1332                                         }
1333                                         mp->m_resblks_avail = lcounter;
1334                                         return 0;
1335                                 } else {        /* not reserved */
1336                                         return XFS_ERROR(ENOSPC);
1337                                 }
1338                         }
1339                 }
1340
1341                 mp->m_sb.sb_fdblocks = lcounter;
1342                 return 0;
1343         case XFS_SBS_FREXTENTS:
1344                 lcounter = (long long)mp->m_sb.sb_frextents;
1345                 lcounter += delta;
1346                 if (lcounter < 0) {
1347                         return XFS_ERROR(ENOSPC);
1348                 }
1349                 mp->m_sb.sb_frextents = lcounter;
1350                 return 0;
1351         case XFS_SBS_DBLOCKS:
1352                 lcounter = (long long)mp->m_sb.sb_dblocks;
1353                 lcounter += delta;
1354                 if (lcounter < 0) {
1355                         ASSERT(0);
1356                         return XFS_ERROR(EINVAL);
1357                 }
1358                 mp->m_sb.sb_dblocks = lcounter;
1359                 return 0;
1360         case XFS_SBS_AGCOUNT:
1361                 scounter = mp->m_sb.sb_agcount;
1362                 scounter += delta;
1363                 if (scounter < 0) {
1364                         ASSERT(0);
1365                         return XFS_ERROR(EINVAL);
1366                 }
1367                 mp->m_sb.sb_agcount = scounter;
1368                 return 0;
1369         case XFS_SBS_IMAX_PCT:
1370                 scounter = mp->m_sb.sb_imax_pct;
1371                 scounter += delta;
1372                 if (scounter < 0) {
1373                         ASSERT(0);
1374                         return XFS_ERROR(EINVAL);
1375                 }
1376                 mp->m_sb.sb_imax_pct = scounter;
1377                 return 0;
1378         case XFS_SBS_REXTSIZE:
1379                 scounter = mp->m_sb.sb_rextsize;
1380                 scounter += delta;
1381                 if (scounter < 0) {
1382                         ASSERT(0);
1383                         return XFS_ERROR(EINVAL);
1384                 }
1385                 mp->m_sb.sb_rextsize = scounter;
1386                 return 0;
1387         case XFS_SBS_RBMBLOCKS:
1388                 scounter = mp->m_sb.sb_rbmblocks;
1389                 scounter += delta;
1390                 if (scounter < 0) {
1391                         ASSERT(0);
1392                         return XFS_ERROR(EINVAL);
1393                 }
1394                 mp->m_sb.sb_rbmblocks = scounter;
1395                 return 0;
1396         case XFS_SBS_RBLOCKS:
1397                 lcounter = (long long)mp->m_sb.sb_rblocks;
1398                 lcounter += delta;
1399                 if (lcounter < 0) {
1400                         ASSERT(0);
1401                         return XFS_ERROR(EINVAL);
1402                 }
1403                 mp->m_sb.sb_rblocks = lcounter;
1404                 return 0;
1405         case XFS_SBS_REXTENTS:
1406                 lcounter = (long long)mp->m_sb.sb_rextents;
1407                 lcounter += delta;
1408                 if (lcounter < 0) {
1409                         ASSERT(0);
1410                         return XFS_ERROR(EINVAL);
1411                 }
1412                 mp->m_sb.sb_rextents = lcounter;
1413                 return 0;
1414         case XFS_SBS_REXTSLOG:
1415                 scounter = mp->m_sb.sb_rextslog;
1416                 scounter += delta;
1417                 if (scounter < 0) {
1418                         ASSERT(0);
1419                         return XFS_ERROR(EINVAL);
1420                 }
1421                 mp->m_sb.sb_rextslog = scounter;
1422                 return 0;
1423         default:
1424                 ASSERT(0);
1425                 return XFS_ERROR(EINVAL);
1426         }
1427 }
1428
1429 /*
1430  * xfs_mod_incore_sb() is used to change a field in the in-core
1431  * superblock structure by the specified delta.  This modification
1432  * is protected by the SB_LOCK.  Just use the xfs_mod_incore_sb_unlocked()
1433  * routine to do the work.
1434  */
1435 int
1436 xfs_mod_incore_sb(xfs_mount_t *mp, xfs_sb_field_t field, int delta, int rsvd)
1437 {
1438         unsigned long   s;
1439         int     status;
1440
1441         /* check for per-cpu counters */
1442         switch (field) {
1443 #ifdef HAVE_PERCPU_SB
1444         case XFS_SBS_ICOUNT:
1445         case XFS_SBS_IFREE:
1446         case XFS_SBS_FDBLOCKS:
1447                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1448                         status = xfs_icsb_modify_counters(mp, field,
1449                                                         delta, rsvd);
1450                         break;
1451                 }
1452                 /* FALLTHROUGH */
1453 #endif
1454         default:
1455                 s = XFS_SB_LOCK(mp);
1456                 status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
1457                 XFS_SB_UNLOCK(mp, s);
1458                 break;
1459         }
1460
1461         return status;
1462 }
1463
1464 /*
1465  * xfs_mod_incore_sb_batch() is used to change more than one field
1466  * in the in-core superblock structure at a time.  This modification
1467  * is protected by a lock internal to this module.  The fields and
1468  * changes to those fields are specified in the array of xfs_mod_sb
1469  * structures passed in.
1470  *
1471  * Either all of the specified deltas will be applied or none of
1472  * them will.  If any modified field dips below 0, then all modifications
1473  * will be backed out and EINVAL will be returned.
1474  */
1475 int
1476 xfs_mod_incore_sb_batch(xfs_mount_t *mp, xfs_mod_sb_t *msb, uint nmsb, int rsvd)
1477 {
1478         unsigned long   s;
1479         int             status=0;
1480         xfs_mod_sb_t    *msbp;
1481
1482         /*
1483          * Loop through the array of mod structures and apply each
1484          * individually.  If any fail, then back out all those
1485          * which have already been applied.  Do all of this within
1486          * the scope of the SB_LOCK so that all of the changes will
1487          * be atomic.
1488          */
1489         s = XFS_SB_LOCK(mp);
1490         msbp = &msb[0];
1491         for (msbp = &msbp[0]; msbp < (msb + nmsb); msbp++) {
1492                 /*
1493                  * Apply the delta at index n.  If it fails, break
1494                  * from the loop so we'll fall into the undo loop
1495                  * below.
1496                  */
1497                 switch (msbp->msb_field) {
1498 #ifdef HAVE_PERCPU_SB
1499                 case XFS_SBS_ICOUNT:
1500                 case XFS_SBS_IFREE:
1501                 case XFS_SBS_FDBLOCKS:
1502                         if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1503                                 status = xfs_icsb_modify_counters_locked(mp,
1504                                                         msbp->msb_field,
1505                                                         msbp->msb_delta, rsvd);
1506                                 break;
1507                         }
1508                         /* FALLTHROUGH */
1509 #endif
1510                 default:
1511                         status = xfs_mod_incore_sb_unlocked(mp,
1512                                                 msbp->msb_field,
1513                                                 msbp->msb_delta, rsvd);
1514                         break;
1515                 }
1516
1517                 if (status != 0) {
1518                         break;
1519                 }
1520         }
1521
1522         /*
1523          * If we didn't complete the loop above, then back out
1524          * any changes made to the superblock.  If you add code
1525          * between the loop above and here, make sure that you
1526          * preserve the value of status. Loop back until
1527          * we step below the beginning of the array.  Make sure
1528          * we don't touch anything back there.
1529          */
1530         if (status != 0) {
1531                 msbp--;
1532                 while (msbp >= msb) {
1533                         switch (msbp->msb_field) {
1534 #ifdef HAVE_PERCPU_SB
1535                         case XFS_SBS_ICOUNT:
1536                         case XFS_SBS_IFREE:
1537                         case XFS_SBS_FDBLOCKS:
1538                                 if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
1539                                         status =
1540                                             xfs_icsb_modify_counters_locked(mp,
1541                                                         msbp->msb_field,
1542                                                         -(msbp->msb_delta),
1543                                                         rsvd);
1544                                         break;
1545                                 }
1546                                 /* FALLTHROUGH */
1547 #endif
1548                         default:
1549                                 status = xfs_mod_incore_sb_unlocked(mp,
1550                                                         msbp->msb_field,
1551                                                         -(msbp->msb_delta),
1552                                                         rsvd);
1553                                 break;
1554                         }
1555                         ASSERT(status == 0);
1556                         msbp--;
1557                 }
1558         }
1559         XFS_SB_UNLOCK(mp, s);
1560         return status;
1561 }
1562
1563 /*
1564  * xfs_getsb() is called to obtain the buffer for the superblock.
1565  * The buffer is returned locked and read in from disk.
1566  * The buffer should be released with a call to xfs_brelse().
1567  *
1568  * If the flags parameter is BUF_TRYLOCK, then we'll only return
1569  * the superblock buffer if it can be locked without sleeping.
1570  * If it can't then we'll return NULL.
1571  */
1572 xfs_buf_t *
1573 xfs_getsb(
1574         xfs_mount_t     *mp,
1575         int             flags)
1576 {
1577         xfs_buf_t       *bp;
1578
1579         ASSERT(mp->m_sb_bp != NULL);
1580         bp = mp->m_sb_bp;
1581         if (flags & XFS_BUF_TRYLOCK) {
1582                 if (!XFS_BUF_CPSEMA(bp)) {
1583                         return NULL;
1584                 }
1585         } else {
1586                 XFS_BUF_PSEMA(bp, PRIBIO);
1587         }
1588         XFS_BUF_HOLD(bp);
1589         ASSERT(XFS_BUF_ISDONE(bp));
1590         return bp;
1591 }
1592
1593 /*
1594  * Used to free the superblock along various error paths.
1595  */
1596 void
1597 xfs_freesb(
1598         xfs_mount_t     *mp)
1599 {
1600         xfs_buf_t       *bp;
1601
1602         /*
1603          * Use xfs_getsb() so that the buffer will be locked
1604          * when we call xfs_buf_relse().
1605          */
1606         bp = xfs_getsb(mp, 0);
1607         XFS_BUF_UNMANAGE(bp);
1608         xfs_buf_relse(bp);
1609         mp->m_sb_bp = NULL;
1610 }
1611
1612 /*
1613  * See if the UUID is unique among mounted XFS filesystems.
1614  * Mount fails if UUID is nil or a FS with the same UUID is already mounted.
1615  */
1616 STATIC int
1617 xfs_uuid_mount(
1618         xfs_mount_t     *mp)
1619 {
1620         if (uuid_is_nil(&mp->m_sb.sb_uuid)) {
1621                 cmn_err(CE_WARN,
1622                         "XFS: Filesystem %s has nil UUID - can't mount",
1623                         mp->m_fsname);
1624                 return -1;
1625         }
1626         if (!uuid_table_insert(&mp->m_sb.sb_uuid)) {
1627                 cmn_err(CE_WARN,
1628                         "XFS: Filesystem %s has duplicate UUID - can't mount",
1629                         mp->m_fsname);
1630                 return -1;
1631         }
1632         return 0;
1633 }
1634
1635 /*
1636  * Remove filesystem from the UUID table.
1637  */
1638 STATIC void
1639 xfs_uuid_unmount(
1640         xfs_mount_t     *mp)
1641 {
1642         uuid_table_remove(&mp->m_sb.sb_uuid);
1643 }
1644
1645 /*
1646  * Used to log changes to the superblock unit and width fields which could
1647  * be altered by the mount options. Only the first superblock is updated.
1648  */
1649 STATIC void
1650 xfs_mount_log_sbunit(
1651         xfs_mount_t     *mp,
1652         __int64_t       fields)
1653 {
1654         xfs_trans_t     *tp;
1655
1656         ASSERT(fields & (XFS_SB_UNIT|XFS_SB_WIDTH|XFS_SB_UUID));
1657
1658         tp = xfs_trans_alloc(mp, XFS_TRANS_SB_UNIT);
1659         if (xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
1660                                 XFS_DEFAULT_LOG_COUNT)) {
1661                 xfs_trans_cancel(tp, 0);
1662                 return;
1663         }
1664         xfs_mod_sb(tp, fields);
1665         xfs_trans_commit(tp, 0, NULL);
1666 }
1667
1668
1669 #ifdef HAVE_PERCPU_SB
1670 /*
1671  * Per-cpu incore superblock counters
1672  *
1673  * Simple concept, difficult implementation
1674  *
1675  * Basically, replace the incore superblock counters with a distributed per cpu
1676  * counter for contended fields (e.g.  free block count).
1677  *
1678  * Difficulties arise in that the incore sb is used for ENOSPC checking, and
1679  * hence needs to be accurately read when we are running low on space. Hence
1680  * there is a method to enable and disable the per-cpu counters based on how
1681  * much "stuff" is available in them.
1682  *
1683  * Basically, a counter is enabled if there is enough free resource to justify
1684  * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
1685  * ENOSPC), then we disable the counters to synchronise all callers and
1686  * re-distribute the available resources.
1687  *
1688  * If, once we redistributed the available resources, we still get a failure,
1689  * we disable the per-cpu counter and go through the slow path.
1690  *
1691  * The slow path is the current xfs_mod_incore_sb() function.  This means that
1692  * when we disable a per-cpu counter, we need to drain it's resources back to
1693  * the global superblock. We do this after disabling the counter to prevent
1694  * more threads from queueing up on the counter.
1695  *
1696  * Essentially, this means that we still need a lock in the fast path to enable
1697  * synchronisation between the global counters and the per-cpu counters. This
1698  * is not a problem because the lock will be local to a CPU almost all the time
1699  * and have little contention except when we get to ENOSPC conditions.
1700  *
1701  * Basically, this lock becomes a barrier that enables us to lock out the fast
1702  * path while we do things like enabling and disabling counters and
1703  * synchronising the counters.
1704  *
1705  * Locking rules:
1706  *
1707  *      1. XFS_SB_LOCK() before picking up per-cpu locks
1708  *      2. per-cpu locks always picked up via for_each_online_cpu() order
1709  *      3. accurate counter sync requires XFS_SB_LOCK + per cpu locks
1710  *      4. modifying per-cpu counters requires holding per-cpu lock
1711  *      5. modifying global counters requires holding XFS_SB_LOCK
1712  *      6. enabling or disabling a counter requires holding the XFS_SB_LOCK
1713  *         and _none_ of the per-cpu locks.
1714  *
1715  * Disabled counters are only ever re-enabled by a balance operation
1716  * that results in more free resources per CPU than a given threshold.
1717  * To ensure counters don't remain disabled, they are rebalanced when
1718  * the global resource goes above a higher threshold (i.e. some hysteresis
1719  * is present to prevent thrashing).
1720  */
1721
1722 /*
1723  * hot-plug CPU notifier support.
1724  *
1725  * We cannot use the hotcpu_register() function because it does
1726  * not allow notifier instances. We need a notifier per filesystem
1727  * as we need to be able to identify the filesystem to balance
1728  * the counters out. This is acheived by having a notifier block
1729  * embedded in the xfs_mount_t and doing pointer magic to get the
1730  * mount pointer from the notifier block address.
1731  */
1732 STATIC int
1733 xfs_icsb_cpu_notify(
1734         struct notifier_block *nfb,
1735         unsigned long action,
1736         void *hcpu)
1737 {
1738         xfs_icsb_cnts_t *cntp;
1739         xfs_mount_t     *mp;
1740         int             s;
1741
1742         mp = (xfs_mount_t *)container_of(nfb, xfs_mount_t, m_icsb_notifier);
1743         cntp = (xfs_icsb_cnts_t *)
1744                         per_cpu_ptr(mp->m_sb_cnts, (unsigned long)hcpu);
1745         switch (action) {
1746         case CPU_UP_PREPARE:
1747                 /* Easy Case - initialize the area and locks, and
1748                  * then rebalance when online does everything else for us. */
1749                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1750                 break;
1751         case CPU_ONLINE:
1752                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
1753                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
1754                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
1755                 break;
1756         case CPU_DEAD:
1757                 /* Disable all the counters, then fold the dead cpu's
1758                  * count into the total on the global superblock and
1759                  * re-enable the counters. */
1760                 s = XFS_SB_LOCK(mp);
1761                 xfs_icsb_disable_counter(mp, XFS_SBS_ICOUNT);
1762                 xfs_icsb_disable_counter(mp, XFS_SBS_IFREE);
1763                 xfs_icsb_disable_counter(mp, XFS_SBS_FDBLOCKS);
1764
1765                 mp->m_sb.sb_icount += cntp->icsb_icount;
1766                 mp->m_sb.sb_ifree += cntp->icsb_ifree;
1767                 mp->m_sb.sb_fdblocks += cntp->icsb_fdblocks;
1768
1769                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1770
1771                 xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, XFS_ICSB_SB_LOCKED);
1772                 xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, XFS_ICSB_SB_LOCKED);
1773                 xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, XFS_ICSB_SB_LOCKED);
1774                 XFS_SB_UNLOCK(mp, s);
1775                 break;
1776         }
1777
1778         return NOTIFY_OK;
1779 }
1780
1781 int
1782 xfs_icsb_init_counters(
1783         xfs_mount_t     *mp)
1784 {
1785         xfs_icsb_cnts_t *cntp;
1786         int             i;
1787
1788         mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
1789         if (mp->m_sb_cnts == NULL)
1790                 return -ENOMEM;
1791
1792         mp->m_icsb_notifier.notifier_call = xfs_icsb_cpu_notify;
1793         mp->m_icsb_notifier.priority = 0;
1794         register_cpu_notifier(&mp->m_icsb_notifier);
1795
1796         for_each_online_cpu(i) {
1797                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1798                 memset(cntp, 0, sizeof(xfs_icsb_cnts_t));
1799         }
1800         /*
1801          * start with all counters disabled so that the
1802          * initial balance kicks us off correctly
1803          */
1804         mp->m_icsb_counters = -1;
1805         return 0;
1806 }
1807
1808 STATIC void
1809 xfs_icsb_destroy_counters(
1810         xfs_mount_t     *mp)
1811 {
1812         if (mp->m_sb_cnts) {
1813                 unregister_cpu_notifier(&mp->m_icsb_notifier);
1814                 free_percpu(mp->m_sb_cnts);
1815         }
1816 }
1817
1818 STATIC inline void
1819 xfs_icsb_lock_cntr(
1820         xfs_icsb_cnts_t *icsbp)
1821 {
1822         while (test_and_set_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags)) {
1823                 ndelay(1000);
1824         }
1825 }
1826
1827 STATIC inline void
1828 xfs_icsb_unlock_cntr(
1829         xfs_icsb_cnts_t *icsbp)
1830 {
1831         clear_bit(XFS_ICSB_FLAG_LOCK, &icsbp->icsb_flags);
1832 }
1833
1834
1835 STATIC inline void
1836 xfs_icsb_lock_all_counters(
1837         xfs_mount_t     *mp)
1838 {
1839         xfs_icsb_cnts_t *cntp;
1840         int             i;
1841
1842         for_each_online_cpu(i) {
1843                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1844                 xfs_icsb_lock_cntr(cntp);
1845         }
1846 }
1847
1848 STATIC inline void
1849 xfs_icsb_unlock_all_counters(
1850         xfs_mount_t     *mp)
1851 {
1852         xfs_icsb_cnts_t *cntp;
1853         int             i;
1854
1855         for_each_online_cpu(i) {
1856                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1857                 xfs_icsb_unlock_cntr(cntp);
1858         }
1859 }
1860
1861 STATIC void
1862 xfs_icsb_count(
1863         xfs_mount_t     *mp,
1864         xfs_icsb_cnts_t *cnt,
1865         int             flags)
1866 {
1867         xfs_icsb_cnts_t *cntp;
1868         int             i;
1869
1870         memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
1871
1872         if (!(flags & XFS_ICSB_LAZY_COUNT))
1873                 xfs_icsb_lock_all_counters(mp);
1874
1875         for_each_online_cpu(i) {
1876                 cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
1877                 cnt->icsb_icount += cntp->icsb_icount;
1878                 cnt->icsb_ifree += cntp->icsb_ifree;
1879                 cnt->icsb_fdblocks += cntp->icsb_fdblocks;
1880         }
1881
1882         if (!(flags & XFS_ICSB_LAZY_COUNT))
1883                 xfs_icsb_unlock_all_counters(mp);
1884 }
1885
1886 STATIC int
1887 xfs_icsb_counter_disabled(
1888         xfs_mount_t     *mp,
1889         xfs_sb_field_t  field)
1890 {
1891         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1892         return test_bit(field, &mp->m_icsb_counters);
1893 }
1894
1895 STATIC int
1896 xfs_icsb_disable_counter(
1897         xfs_mount_t     *mp,
1898         xfs_sb_field_t  field)
1899 {
1900         xfs_icsb_cnts_t cnt;
1901
1902         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1903
1904         xfs_icsb_lock_all_counters(mp);
1905         if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
1906                 /* drain back to superblock */
1907
1908                 xfs_icsb_count(mp, &cnt, XFS_ICSB_SB_LOCKED|XFS_ICSB_LAZY_COUNT);
1909                 switch(field) {
1910                 case XFS_SBS_ICOUNT:
1911                         mp->m_sb.sb_icount = cnt.icsb_icount;
1912                         break;
1913                 case XFS_SBS_IFREE:
1914                         mp->m_sb.sb_ifree = cnt.icsb_ifree;
1915                         break;
1916                 case XFS_SBS_FDBLOCKS:
1917                         mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1918                         break;
1919                 default:
1920                         BUG();
1921                 }
1922         }
1923
1924         xfs_icsb_unlock_all_counters(mp);
1925
1926         return 0;
1927 }
1928
1929 STATIC void
1930 xfs_icsb_enable_counter(
1931         xfs_mount_t     *mp,
1932         xfs_sb_field_t  field,
1933         uint64_t        count,
1934         uint64_t        resid)
1935 {
1936         xfs_icsb_cnts_t *cntp;
1937         int             i;
1938
1939         ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
1940
1941         xfs_icsb_lock_all_counters(mp);
1942         for_each_online_cpu(i) {
1943                 cntp = per_cpu_ptr(mp->m_sb_cnts, i);
1944                 switch (field) {
1945                 case XFS_SBS_ICOUNT:
1946                         cntp->icsb_icount = count + resid;
1947                         break;
1948                 case XFS_SBS_IFREE:
1949                         cntp->icsb_ifree = count + resid;
1950                         break;
1951                 case XFS_SBS_FDBLOCKS:
1952                         cntp->icsb_fdblocks = count + resid;
1953                         break;
1954                 default:
1955                         BUG();
1956                         break;
1957                 }
1958                 resid = 0;
1959         }
1960         clear_bit(field, &mp->m_icsb_counters);
1961         xfs_icsb_unlock_all_counters(mp);
1962 }
1963
1964 STATIC void
1965 xfs_icsb_sync_counters_int(
1966         xfs_mount_t     *mp,
1967         int             flags)
1968 {
1969         xfs_icsb_cnts_t cnt;
1970         int             s;
1971
1972         /* Pass 1: lock all counters */
1973         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
1974                 s = XFS_SB_LOCK(mp);
1975
1976         xfs_icsb_count(mp, &cnt, flags);
1977
1978         /* Step 3: update mp->m_sb fields */
1979         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
1980                 mp->m_sb.sb_icount = cnt.icsb_icount;
1981         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
1982                 mp->m_sb.sb_ifree = cnt.icsb_ifree;
1983         if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
1984                 mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
1985
1986         if ((flags & XFS_ICSB_SB_LOCKED) == 0)
1987                 XFS_SB_UNLOCK(mp, s);
1988 }
1989
1990 /*
1991  * Accurate update of per-cpu counters to incore superblock
1992  */
1993 STATIC void
1994 xfs_icsb_sync_counters(
1995         xfs_mount_t     *mp)
1996 {
1997         xfs_icsb_sync_counters_int(mp, 0);
1998 }
1999
2000 /*
2001  * lazy addition used for things like df, background sb syncs, etc
2002  */
2003 void
2004 xfs_icsb_sync_counters_lazy(
2005         xfs_mount_t     *mp)
2006 {
2007         xfs_icsb_sync_counters_int(mp, XFS_ICSB_LAZY_COUNT);
2008 }
2009
2010 /*
2011  * Balance and enable/disable counters as necessary.
2012  *
2013  * Thresholds for re-enabling counters are somewhat magic.
2014  * inode counts are chosen to be the same number as single
2015  * on disk allocation chunk per CPU, and free blocks is
2016  * something far enough zero that we aren't going thrash
2017  * when we get near ENOSPC.
2018  */
2019 #define XFS_ICSB_INO_CNTR_REENABLE      64
2020 #define XFS_ICSB_FDBLK_CNTR_REENABLE    512
2021 STATIC void
2022 xfs_icsb_balance_counter(
2023         xfs_mount_t     *mp,
2024         xfs_sb_field_t  field,
2025         int             flags)
2026 {
2027         uint64_t        count, resid = 0;
2028         int             weight = num_online_cpus();
2029         int             s;
2030
2031         if (!(flags & XFS_ICSB_SB_LOCKED))
2032                 s = XFS_SB_LOCK(mp);
2033
2034         /* disable counter and sync counter */
2035         xfs_icsb_disable_counter(mp, field);
2036
2037         /* update counters  - first CPU gets residual*/
2038         switch (field) {
2039         case XFS_SBS_ICOUNT:
2040                 count = mp->m_sb.sb_icount;
2041                 resid = do_div(count, weight);
2042                 if (count < XFS_ICSB_INO_CNTR_REENABLE)
2043                         goto out;
2044                 break;
2045         case XFS_SBS_IFREE:
2046                 count = mp->m_sb.sb_ifree;
2047                 resid = do_div(count, weight);
2048                 if (count < XFS_ICSB_INO_CNTR_REENABLE)
2049                         goto out;
2050                 break;
2051         case XFS_SBS_FDBLOCKS:
2052                 count = mp->m_sb.sb_fdblocks;
2053                 resid = do_div(count, weight);
2054                 if (count < XFS_ICSB_FDBLK_CNTR_REENABLE)
2055                         goto out;
2056                 break;
2057         default:
2058                 BUG();
2059                 break;
2060         }
2061
2062         xfs_icsb_enable_counter(mp, field, count, resid);
2063 out:
2064         if (!(flags & XFS_ICSB_SB_LOCKED))
2065                 XFS_SB_UNLOCK(mp, s);
2066 }
2067
2068 STATIC int
2069 xfs_icsb_modify_counters_int(
2070         xfs_mount_t     *mp,
2071         xfs_sb_field_t  field,
2072         int             delta,
2073         int             rsvd,
2074         int             flags)
2075 {
2076         xfs_icsb_cnts_t *icsbp;
2077         long long       lcounter;       /* long counter for 64 bit fields */
2078         int             cpu, s, locked = 0;
2079         int             ret = 0, balance_done = 0;
2080
2081 again:
2082         cpu = get_cpu();
2083         icsbp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, cpu),
2084         xfs_icsb_lock_cntr(icsbp);
2085         if (unlikely(xfs_icsb_counter_disabled(mp, field)))
2086                 goto slow_path;
2087
2088         switch (field) {
2089         case XFS_SBS_ICOUNT:
2090                 lcounter = icsbp->icsb_icount;
2091                 lcounter += delta;
2092                 if (unlikely(lcounter < 0))
2093                         goto slow_path;
2094                 icsbp->icsb_icount = lcounter;
2095                 break;
2096
2097         case XFS_SBS_IFREE:
2098                 lcounter = icsbp->icsb_ifree;
2099                 lcounter += delta;
2100                 if (unlikely(lcounter < 0))
2101                         goto slow_path;
2102                 icsbp->icsb_ifree = lcounter;
2103                 break;
2104
2105         case XFS_SBS_FDBLOCKS:
2106                 BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
2107
2108                 lcounter = icsbp->icsb_fdblocks;
2109                 lcounter += delta;
2110                 if (unlikely(lcounter < 0))
2111                         goto slow_path;
2112                 icsbp->icsb_fdblocks = lcounter;
2113                 break;
2114         default:
2115                 BUG();
2116                 break;
2117         }
2118         xfs_icsb_unlock_cntr(icsbp);
2119         put_cpu();
2120         if (locked)
2121                 XFS_SB_UNLOCK(mp, s);
2122         return 0;
2123
2124         /*
2125          * The slow path needs to be run with the SBLOCK
2126          * held so that we prevent other threads from
2127          * attempting to run this path at the same time.
2128          * this provides exclusion for the balancing code,
2129          * and exclusive fallback if the balance does not
2130          * provide enough resources to continue in an unlocked
2131          * manner.
2132          */
2133 slow_path:
2134         xfs_icsb_unlock_cntr(icsbp);
2135         put_cpu();
2136
2137         /* need to hold superblock incase we need
2138          * to disable a counter */
2139         if (!(flags & XFS_ICSB_SB_LOCKED)) {
2140                 s = XFS_SB_LOCK(mp);
2141                 locked = 1;
2142                 flags |= XFS_ICSB_SB_LOCKED;
2143         }
2144         if (!balance_done) {
2145                 xfs_icsb_balance_counter(mp, field, flags);
2146                 balance_done = 1;
2147                 goto again;
2148         } else {
2149                 /*
2150                  * we might not have enough on this local
2151                  * cpu to allocate for a bulk request.
2152                  * We need to drain this field from all CPUs
2153                  * and disable the counter fastpath
2154                  */
2155                 xfs_icsb_disable_counter(mp, field);
2156         }
2157
2158         ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
2159
2160         if (locked)
2161                 XFS_SB_UNLOCK(mp, s);
2162         return ret;
2163 }
2164
2165 STATIC int
2166 xfs_icsb_modify_counters(
2167         xfs_mount_t     *mp,
2168         xfs_sb_field_t  field,
2169         int             delta,
2170         int             rsvd)
2171 {
2172         return xfs_icsb_modify_counters_int(mp, field, delta, rsvd, 0);
2173 }
2174
2175 /*
2176  * Called when superblock is already locked
2177  */
2178 STATIC int
2179 xfs_icsb_modify_counters_locked(
2180         xfs_mount_t     *mp,
2181         xfs_sb_field_t  field,
2182         int             delta,
2183         int             rsvd)
2184 {
2185         return xfs_icsb_modify_counters_int(mp, field, delta,
2186                                                 rsvd, XFS_ICSB_SB_LOCKED);
2187 }
2188 #endif