fs/xfs/xfs_inode_item.c

   1 /*
   2  * Copyright (c) 2000-2002,2005 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_mount.h"
  28 #include "xfs_trans_priv.h"
  29 #include "xfs_bmap_btree.h"
  30 #include "xfs_dinode.h"
  31 #include "xfs_inode.h"
  32 #include "xfs_inode_item.h"
  33 #include "xfs_error.h"
  34 #include "xfs_trace.h"
  35
  36
  37 kmem_zone_t     *xfs_ili_zone;          /* inode log item zone */
  38
  39 /*
  40  * This returns the number of iovecs needed to log the given inode item.
  41  *
  42  * We need one iovec for the inode log format structure, one for the
  43  * inode core, and possibly one for the inode data/extents/b-tree root
  44  * and one for the inode attribute data/extents/b-tree root.
  45  */
  46 STATIC uint
  47 xfs_inode_item_size(
  48         xfs_inode_log_item_t    *iip)
  49 {
  50         uint            nvecs;
  51         xfs_inode_t     *ip;
  52
  53         ip = iip->ili_inode;
  54         nvecs = 2;
  55
  56         /*
  57          * Only log the data/extents/b-tree root if there is something
  58          * left to log.
  59          */
  60         iip->ili_format.ilf_fields |= XFS_ILOG_CORE;
  61
  62         switch (ip->i_d.di_format) {
  63         case XFS_DINODE_FMT_EXTENTS:
  64                 iip->ili_format.ilf_fields &=
  65                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
  66                           XFS_ILOG_DEV | XFS_ILOG_UUID);
  67                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DEXT) &&
  68                     (ip->i_d.di_nextents > 0) &&
  69                     (ip->i_df.if_bytes > 0)) {
  70                         ASSERT(ip->i_df.if_u1.if_extents != NULL);
  71                         nvecs++;
  72                 } else {
  73                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DEXT;
  74                 }
  75                 break;
  76
  77         case XFS_DINODE_FMT_BTREE:
  78                 ASSERT(ip->i_df.if_ext_max ==
  79                        XFS_IFORK_DSIZE(ip) / (uint)sizeof(xfs_bmbt_rec_t));
  80                 iip->ili_format.ilf_fields &=
  81                         ~(XFS_ILOG_DDATA | XFS_ILOG_DEXT |
  82                           XFS_ILOG_DEV | XFS_ILOG_UUID);
  83                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DBROOT) &&
  84                     (ip->i_df.if_broot_bytes > 0)) {
  85                         ASSERT(ip->i_df.if_broot != NULL);
  86                         nvecs++;
  87                 } else {
  88                         ASSERT(!(iip->ili_format.ilf_fields &
  89                                  XFS_ILOG_DBROOT));
  90 #ifdef XFS_TRANS_DEBUG
  91                         if (iip->ili_root_size > 0) {
  92                                 ASSERT(iip->ili_root_size ==
  93                                        ip->i_df.if_broot_bytes);
  94                                 ASSERT(memcmp(iip->ili_orig_root,
  95                                             ip->i_df.if_broot,
  96                                             iip->ili_root_size) == 0);
  97                         } else {
  98                                 ASSERT(ip->i_df.if_broot_bytes == 0);
  99                         }
 100 #endif
 101                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DBROOT;
 102                 }
 103                 break;
 104
 105         case XFS_DINODE_FMT_LOCAL:
 106                 iip->ili_format.ilf_fields &=
 107                         ~(XFS_ILOG_DEXT | XFS_ILOG_DBROOT |
 108                           XFS_ILOG_DEV | XFS_ILOG_UUID);
 109                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DDATA) &&
 110                     (ip->i_df.if_bytes > 0)) {
 111                         ASSERT(ip->i_df.if_u1.if_data != NULL);
 112                         ASSERT(ip->i_d.di_size > 0);
 113                         nvecs++;
 114                 } else {
 115                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DDATA;
 116                 }
 117                 break;
 118
 119         case XFS_DINODE_FMT_DEV:
 120                 iip->ili_format.ilf_fields &=
 121                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 122                           XFS_ILOG_DEXT | XFS_ILOG_UUID);
 123                 break;
 124
 125         case XFS_DINODE_FMT_UUID:
 126                 iip->ili_format.ilf_fields &=
 127                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 128                           XFS_ILOG_DEXT | XFS_ILOG_DEV);
 129                 break;
 130
 131         default:
 132                 ASSERT(0);
 133                 break;
 134         }
 135
 136         /*
 137          * If there are no attributes associated with this file,
 138          * then there cannot be anything more to log.
 139          * Clear all attribute-related log flags.
 140          */
 141         if (!XFS_IFORK_Q(ip)) {
 142                 iip->ili_format.ilf_fields &=
 143                         ~(XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT);
 144                 return nvecs;
 145         }
 146
 147         /*
 148          * Log any necessary attribute data.
 149          */
 150         switch (ip->i_d.di_aformat) {
 151         case XFS_DINODE_FMT_EXTENTS:
 152                 iip->ili_format.ilf_fields &=
 153                         ~(XFS_ILOG_ADATA | XFS_ILOG_ABROOT);
 154                 if ((iip->ili_format.ilf_fields & XFS_ILOG_AEXT) &&
 155                     (ip->i_d.di_anextents > 0) &&
 156                     (ip->i_afp->if_bytes > 0)) {
 157                         ASSERT(ip->i_afp->if_u1.if_extents != NULL);
 158                         nvecs++;
 159                 } else {
 160                         iip->ili_format.ilf_fields &= ~XFS_ILOG_AEXT;
 161                 }
 162                 break;
 163
 164         case XFS_DINODE_FMT_BTREE:
 165                 iip->ili_format.ilf_fields &=
 166                         ~(XFS_ILOG_ADATA | XFS_ILOG_AEXT);
 167                 if ((iip->ili_format.ilf_fields & XFS_ILOG_ABROOT) &&
 168                     (ip->i_afp->if_broot_bytes > 0)) {
 169                         ASSERT(ip->i_afp->if_broot != NULL);
 170                         nvecs++;
 171                 } else {
 172                         iip->ili_format.ilf_fields &= ~XFS_ILOG_ABROOT;
 173                 }
 174                 break;
 175
 176         case XFS_DINODE_FMT_LOCAL:
 177                 iip->ili_format.ilf_fields &=
 178                         ~(XFS_ILOG_AEXT | XFS_ILOG_ABROOT);
 179                 if ((iip->ili_format.ilf_fields & XFS_ILOG_ADATA) &&
 180                     (ip->i_afp->if_bytes > 0)) {
 181                         ASSERT(ip->i_afp->if_u1.if_data != NULL);
 182                         nvecs++;
 183                 } else {
 184                         iip->ili_format.ilf_fields &= ~XFS_ILOG_ADATA;
 185                 }
 186                 break;
 187
 188         default:
 189                 ASSERT(0);
 190                 break;
 191         }
 192
 193         return nvecs;
 194 }
 195
 196 /*
 197  * This is called to fill in the vector of log iovecs for the
 198  * given inode log item.  It fills the first item with an inode
 199  * log format structure, the second with the on-disk inode structure,
 200  * and a possible third and/or fourth with the inode data/extents/b-tree
 201  * root and inode attributes data/extents/b-tree root.
 202  */
 203 STATIC void
 204 xfs_inode_item_format(
 205         xfs_inode_log_item_t    *iip,
 206         xfs_log_iovec_t         *log_vector)
 207 {
 208         uint                    nvecs;
 209         xfs_log_iovec_t         *vecp;
 210         xfs_inode_t             *ip;
 211         size_t                  data_bytes;
 212         xfs_bmbt_rec_t          *ext_buffer;
 213         int                     nrecs;
 214         xfs_mount_t             *mp;
 215
 216         ip = iip->ili_inode;
 217         vecp = log_vector;
 218
 219         vecp->i_addr = (xfs_caddr_t)&iip->ili_format;
 220         vecp->i_len  = sizeof(xfs_inode_log_format_t);
 221         vecp->i_type = XLOG_REG_TYPE_IFORMAT;
 222         vecp++;
 223         nvecs        = 1;
 224
 225         /*
 226          * Make sure the linux inode is dirty. We do this before
 227          * clearing i_update_core as the VFS will call back into
 228          * XFS here and set i_update_core, so we need to dirty the
 229          * inode first so that the ordering of i_update_core and
 230          * unlogged modifications still works as described below.
 231          */
 232         xfs_mark_inode_dirty_sync(ip);
 233
 234         /*
 235          * Clear i_update_core if the timestamps (or any other
 236          * non-transactional modification) need flushing/logging
 237          * and we're about to log them with the rest of the core.
 238          *
 239          * This is the same logic as xfs_iflush() but this code can't
 240          * run at the same time as xfs_iflush because we're in commit
 241          * processing here and so we have the inode lock held in
 242          * exclusive mode.  Although it doesn't really matter
 243          * for the timestamps if both routines were to grab the
 244          * timestamps or not.  That would be ok.
 245          *
 246          * We clear i_update_core before copying out the data.
 247          * This is for coordination with our timestamp updates
 248          * that don't hold the inode lock. They will always
 249          * update the timestamps BEFORE setting i_update_core,
 250          * so if we clear i_update_core after they set it we
 251          * are guaranteed to see their updates to the timestamps
 252          * either here.  Likewise, if they set it after we clear it
 253          * here, we'll see it either on the next commit of this
 254          * inode or the next time the inode gets flushed via
 255          * xfs_iflush().  This depends on strongly ordered memory
 256          * semantics, but we have that.  We use the SYNCHRONIZE
 257          * macro to make sure that the compiler does not reorder
 258          * the i_update_core access below the data copy below.
 259          */
 260         if (ip->i_update_core)  {
 261                 ip->i_update_core = 0;
 262                 SYNCHRONIZE();
 263         }
 264
 265         /*
 266          * Make sure to get the latest timestamps from the Linux inode.
 267          */
 268         xfs_synchronize_times(ip);
 269
 270         vecp->i_addr = (xfs_caddr_t)&ip->i_d;
 271         vecp->i_len  = sizeof(struct xfs_icdinode);
 272         vecp->i_type = XLOG_REG_TYPE_ICORE;
 273         vecp++;
 274         nvecs++;
 275         iip->ili_format.ilf_fields |= XFS_ILOG_CORE;
 276
 277         /*
 278          * If this is really an old format inode, then we need to
 279          * log it as such.  This means that we have to copy the link
 280          * count from the new field to the old.  We don't have to worry
 281          * about the new fields, because nothing trusts them as long as
 282          * the old inode version number is there.  If the superblock already
 283          * has a new version number, then we don't bother converting back.
 284          */
 285         mp = ip->i_mount;
 286         ASSERT(ip->i_d.di_version == 1 || xfs_sb_version_hasnlink(&mp->m_sb));
 287         if (ip->i_d.di_version == 1) {
 288                 if (!xfs_sb_version_hasnlink(&mp->m_sb)) {
 289                         /*
 290                          * Convert it back.
 291                          */
 292                         ASSERT(ip->i_d.di_nlink <= XFS_MAXLINK_1);
 293                         ip->i_d.di_onlink = ip->i_d.di_nlink;
 294                 } else {
 295                         /*
 296                          * The superblock version has already been bumped,
 297                          * so just make the conversion to the new inode
 298                          * format permanent.
 299                          */
 300                         ip->i_d.di_version = 2;
 301                         ip->i_d.di_onlink = 0;
 302                         memset(&(ip->i_d.di_pad[0]), 0, sizeof(ip->i_d.di_pad));
 303                 }
 304         }
 305
 306         switch (ip->i_d.di_format) {
 307         case XFS_DINODE_FMT_EXTENTS:
 308                 ASSERT(!(iip->ili_format.ilf_fields &
 309                          (XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 310                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 311                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEXT) {
 312                         ASSERT(ip->i_df.if_bytes > 0);
 313                         ASSERT(ip->i_df.if_u1.if_extents != NULL);
 314                         ASSERT(ip->i_d.di_nextents > 0);
 315                         ASSERT(iip->ili_extents_buf == NULL);
 316                         nrecs = ip->i_df.if_bytes /
 317                                 (uint)sizeof(xfs_bmbt_rec_t);
 318                         ASSERT(nrecs > 0);
 319 #ifdef XFS_NATIVE_HOST
 320                         if (nrecs == ip->i_d.di_nextents) {
 321                                 /*
 322                                  * There are no delayed allocation
 323                                  * extents, so just point to the
 324                                  * real extents array.
 325                                  */
 326                                 vecp->i_addr =
 327                                         (char *)(ip->i_df.if_u1.if_extents);
 328                                 vecp->i_len = ip->i_df.if_bytes;
 329                                 vecp->i_type = XLOG_REG_TYPE_IEXT;
 330                         } else
 331 #endif
 332                         {
 333                                 /*
 334                                  * There are delayed allocation extents
 335                                  * in the inode, or we need to convert
 336                                  * the extents to on disk format.
 337                                  * Use xfs_iextents_copy()
 338                                  * to copy only the real extents into
 339                                  * a separate buffer.  We'll free the
 340                                  * buffer in the unlock routine.
 341                                  */
 342                                 ext_buffer = kmem_alloc(ip->i_df.if_bytes,
 343                                         KM_SLEEP);
 344                                 iip->ili_extents_buf = ext_buffer;
 345                                 vecp->i_addr = (xfs_caddr_t)ext_buffer;
 346                                 vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
 347                                                 XFS_DATA_FORK);
 348                                 vecp->i_type = XLOG_REG_TYPE_IEXT;
 349                         }
 350                         ASSERT(vecp->i_len <= ip->i_df.if_bytes);
 351                         iip->ili_format.ilf_dsize = vecp->i_len;
 352                         vecp++;
 353                         nvecs++;
 354                 }
 355                 break;
 356
 357         case XFS_DINODE_FMT_BTREE:
 358                 ASSERT(!(iip->ili_format.ilf_fields &
 359                          (XFS_ILOG_DDATA | XFS_ILOG_DEXT |
 360                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 361                 if (iip->ili_format.ilf_fields & XFS_ILOG_DBROOT) {
 362                         ASSERT(ip->i_df.if_broot_bytes > 0);
 363                         ASSERT(ip->i_df.if_broot != NULL);
 364                         vecp->i_addr = (xfs_caddr_t)ip->i_df.if_broot;
 365                         vecp->i_len = ip->i_df.if_broot_bytes;
 366                         vecp->i_type = XLOG_REG_TYPE_IBROOT;
 367                         vecp++;
 368                         nvecs++;
 369                         iip->ili_format.ilf_dsize = ip->i_df.if_broot_bytes;
 370                 }
 371                 break;
 372
 373         case XFS_DINODE_FMT_LOCAL:
 374                 ASSERT(!(iip->ili_format.ilf_fields &
 375                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 376                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 377                 if (iip->ili_format.ilf_fields & XFS_ILOG_DDATA) {
 378                         ASSERT(ip->i_df.if_bytes > 0);
 379                         ASSERT(ip->i_df.if_u1.if_data != NULL);
 380                         ASSERT(ip->i_d.di_size > 0);
 381
 382                         vecp->i_addr = (xfs_caddr_t)ip->i_df.if_u1.if_data;
 383                         /*
 384                          * Round i_bytes up to a word boundary.
 385                          * The underlying memory is guaranteed to
 386                          * to be there by xfs_idata_realloc().
 387                          */
 388                         data_bytes = roundup(ip->i_df.if_bytes, 4);
 389                         ASSERT((ip->i_df.if_real_bytes == 0) ||
 390                                (ip->i_df.if_real_bytes == data_bytes));
 391                         vecp->i_len = (int)data_bytes;
 392                         vecp->i_type = XLOG_REG_TYPE_ILOCAL;
 393                         vecp++;
 394                         nvecs++;
 395                         iip->ili_format.ilf_dsize = (unsigned)data_bytes;
 396                 }
 397                 break;
 398
 399         case XFS_DINODE_FMT_DEV:
 400                 ASSERT(!(iip->ili_format.ilf_fields &
 401                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 402                           XFS_ILOG_DDATA | XFS_ILOG_UUID)));
 403                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEV) {
 404                         iip->ili_format.ilf_u.ilfu_rdev =
 405                                 ip->i_df.if_u2.if_rdev;
 406                 }
 407                 break;
 408
 409         case XFS_DINODE_FMT_UUID:
 410                 ASSERT(!(iip->ili_format.ilf_fields &
 411                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 412                           XFS_ILOG_DDATA | XFS_ILOG_DEV)));
 413                 if (iip->ili_format.ilf_fields & XFS_ILOG_UUID) {
 414                         iip->ili_format.ilf_u.ilfu_uuid =
 415                                 ip->i_df.if_u2.if_uuid;
 416                 }
 417                 break;
 418
 419         default:
 420                 ASSERT(0);
 421                 break;
 422         }
 423
 424         /*
 425          * If there are no attributes associated with the file,
 426          * then we're done.
 427          * Assert that no attribute-related log flags are set.
 428          */
 429         if (!XFS_IFORK_Q(ip)) {
 430                 ASSERT(nvecs == iip->ili_item.li_desc->lid_size);
 431                 iip->ili_format.ilf_size = nvecs;
 432                 ASSERT(!(iip->ili_format.ilf_fields &
 433                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
 434                 return;
 435         }
 436
 437         switch (ip->i_d.di_aformat) {
 438         case XFS_DINODE_FMT_EXTENTS:
 439                 ASSERT(!(iip->ili_format.ilf_fields &
 440                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT)));
 441                 if (iip->ili_format.ilf_fields & XFS_ILOG_AEXT) {
 442                         ASSERT(ip->i_afp->if_bytes > 0);
 443                         ASSERT(ip->i_afp->if_u1.if_extents != NULL);
 444                         ASSERT(ip->i_d.di_anextents > 0);
 445 #ifdef DEBUG
 446                         nrecs = ip->i_afp->if_bytes /
 447                                 (uint)sizeof(xfs_bmbt_rec_t);
 448 #endif
 449                         ASSERT(nrecs > 0);
 450                         ASSERT(nrecs == ip->i_d.di_anextents);
 451 #ifdef XFS_NATIVE_HOST
 452                         /*
 453                          * There are not delayed allocation extents
 454                          * for attributes, so just point at the array.
 455                          */
 456                         vecp->i_addr = (char *)(ip->i_afp->if_u1.if_extents);
 457                         vecp->i_len = ip->i_afp->if_bytes;
 458 #else
 459                         ASSERT(iip->ili_aextents_buf == NULL);
 460                         /*
 461                          * Need to endian flip before logging
 462                          */
 463                         ext_buffer = kmem_alloc(ip->i_afp->if_bytes,
 464                                 KM_SLEEP);
 465                         iip->ili_aextents_buf = ext_buffer;
 466                         vecp->i_addr = (xfs_caddr_t)ext_buffer;
 467                         vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
 468                                         XFS_ATTR_FORK);
 469 #endif
 470                         vecp->i_type = XLOG_REG_TYPE_IATTR_EXT;
 471                         iip->ili_format.ilf_asize = vecp->i_len;
 472                         vecp++;
 473                         nvecs++;
 474                 }
 475                 break;
 476
 477         case XFS_DINODE_FMT_BTREE:
 478                 ASSERT(!(iip->ili_format.ilf_fields &
 479                          (XFS_ILOG_ADATA | XFS_ILOG_AEXT)));
 480                 if (iip->ili_format.ilf_fields & XFS_ILOG_ABROOT) {
 481                         ASSERT(ip->i_afp->if_broot_bytes > 0);
 482                         ASSERT(ip->i_afp->if_broot != NULL);
 483                         vecp->i_addr = (xfs_caddr_t)ip->i_afp->if_broot;
 484                         vecp->i_len = ip->i_afp->if_broot_bytes;
 485                         vecp->i_type = XLOG_REG_TYPE_IATTR_BROOT;
 486                         vecp++;
 487                         nvecs++;
 488                         iip->ili_format.ilf_asize = ip->i_afp->if_broot_bytes;
 489                 }
 490                 break;
 491
 492         case XFS_DINODE_FMT_LOCAL:
 493                 ASSERT(!(iip->ili_format.ilf_fields &
 494                          (XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
 495                 if (iip->ili_format.ilf_fields & XFS_ILOG_ADATA) {
 496                         ASSERT(ip->i_afp->if_bytes > 0);
 497                         ASSERT(ip->i_afp->if_u1.if_data != NULL);
 498
 499                         vecp->i_addr = (xfs_caddr_t)ip->i_afp->if_u1.if_data;
 500                         /*
 501                          * Round i_bytes up to a word boundary.
 502                          * The underlying memory is guaranteed to
 503                          * to be there by xfs_idata_realloc().
 504                          */
 505                         data_bytes = roundup(ip->i_afp->if_bytes, 4);
 506                         ASSERT((ip->i_afp->if_real_bytes == 0) ||
 507                                (ip->i_afp->if_real_bytes == data_bytes));
 508                         vecp->i_len = (int)data_bytes;
 509                         vecp->i_type = XLOG_REG_TYPE_IATTR_LOCAL;
 510                         vecp++;
 511                         nvecs++;
 512                         iip->ili_format.ilf_asize = (unsigned)data_bytes;
 513                 }
 514                 break;
 515
 516         default:
 517                 ASSERT(0);
 518                 break;
 519         }
 520
 521         ASSERT(nvecs == iip->ili_item.li_desc->lid_size);
 522         iip->ili_format.ilf_size = nvecs;
 523 }
 524
 525
 526 /*
 527  * This is called to pin the inode associated with the inode log
 528  * item in memory so it cannot be written out.
 529  */
 530 STATIC void
 531 xfs_inode_item_pin(
 532         xfs_inode_log_item_t    *iip)
 533 {
 534         ASSERT(xfs_isilocked(iip->ili_inode, XFS_ILOCK_EXCL));
 535
 536         trace_xfs_inode_pin(iip->ili_inode, _RET_IP_);
 537         atomic_inc(&iip->ili_inode->i_pincount);
 538 }
 539
 540
 541 /*
 542  * This is called to unpin the inode associated with the inode log
 543  * item which was previously pinned with a call to xfs_inode_item_pin().
 544  *
 545  * Also wake up anyone in xfs_iunpin_wait() if the count goes to 0.
 546  */
 547 STATIC void
 548 xfs_inode_item_unpin(
 549         xfs_inode_log_item_t    *iip,
 550         int                     remove)
 551 {
 552         struct xfs_inode        *ip = iip->ili_inode;
 553
 554         trace_xfs_inode_unpin(ip, _RET_IP_);
 555         ASSERT(atomic_read(&ip->i_pincount) > 0);
 556         if (atomic_dec_and_test(&ip->i_pincount))
 557                 wake_up(&ip->i_ipin_wait);
 558 }
 559
 560 /*
 561  * This is called to attempt to lock the inode associated with this
 562  * inode log item, in preparation for the push routine which does the actual
 563  * iflush.  Don't sleep on the inode lock or the flush lock.
 564  *
 565  * If the flush lock is already held, indicating that the inode has
 566  * been or is in the process of being flushed, then (ideally) we'd like to
 567  * see if the inode's buffer is still incore, and if so give it a nudge.
 568  * We delay doing so until the pushbuf routine, though, to avoid holding
 569  * the AIL lock across a call to the blackhole which is the buffer cache.
 570  * Also we don't want to sleep in any device strategy routines, which can happen
 571  * if we do the subsequent bawrite in here.
 572  */
 573 STATIC uint
 574 xfs_inode_item_trylock(
 575         xfs_inode_log_item_t    *iip)
 576 {
 577         register xfs_inode_t    *ip;
 578
 579         ip = iip->ili_inode;
 580
 581         if (xfs_ipincount(ip) > 0) {
 582                 return XFS_ITEM_PINNED;
 583         }
 584
 585         if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
 586                 return XFS_ITEM_LOCKED;
 587         }
 588
 589         if (!xfs_iflock_nowait(ip)) {
 590                 /*
 591                  * inode has already been flushed to the backing buffer,
 592                  * leave it locked in shared mode, pushbuf routine will
 593                  * unlock it.
 594                  */
 595                 return XFS_ITEM_PUSHBUF;
 596         }
 597
 598         /* Stale items should force out the iclog */
 599         if (ip->i_flags & XFS_ISTALE) {
 600                 xfs_ifunlock(ip);
 601                 /*
 602                  * we hold the AIL lock - notify the unlock routine of this
 603                  * so it doesn't try to get the lock again.
 604                  */
 605                 xfs_iunlock(ip, XFS_ILOCK_SHARED|XFS_IUNLOCK_NONOTIFY);
 606                 return XFS_ITEM_PINNED;
 607         }
 608
 609 #ifdef DEBUG
 610         if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
 611                 ASSERT(iip->ili_format.ilf_fields != 0);
 612                 ASSERT(iip->ili_logged == 0);
 613                 ASSERT(iip->ili_item.li_flags & XFS_LI_IN_AIL);
 614         }
 615 #endif
 616         return XFS_ITEM_SUCCESS;
 617 }
 618
 619 /*
 620  * Unlock the inode associated with the inode log item.
 621  * Clear the fields of the inode and inode log item that
 622  * are specific to the current transaction.  If the
 623  * hold flags is set, do not unlock the inode.
 624  */
 625 STATIC void
 626 xfs_inode_item_unlock(
 627         xfs_inode_log_item_t    *iip)
 628 {
 629         uint            hold;
 630         uint            iolocked;
 631         uint            lock_flags;
 632         xfs_inode_t     *ip;
 633
 634         ASSERT(iip != NULL);
 635         ASSERT(iip->ili_inode->i_itemp != NULL);
 636         ASSERT(xfs_isilocked(iip->ili_inode, XFS_ILOCK_EXCL));
 637         ASSERT((!(iip->ili_inode->i_itemp->ili_flags &
 638                   XFS_ILI_IOLOCKED_EXCL)) ||
 639                xfs_isilocked(iip->ili_inode, XFS_IOLOCK_EXCL));
 640         ASSERT((!(iip->ili_inode->i_itemp->ili_flags &
 641                   XFS_ILI_IOLOCKED_SHARED)) ||
 642                xfs_isilocked(iip->ili_inode, XFS_IOLOCK_SHARED));
 643         /*
 644          * Clear the transaction pointer in the inode.
 645          */
 646         ip = iip->ili_inode;
 647         ip->i_transp = NULL;
 648
 649         /*
 650          * If the inode needed a separate buffer with which to log
 651          * its extents, then free it now.
 652          */
 653         if (iip->ili_extents_buf != NULL) {
 654                 ASSERT(ip->i_d.di_format == XFS_DINODE_FMT_EXTENTS);
 655                 ASSERT(ip->i_d.di_nextents > 0);
 656                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_DEXT);
 657                 ASSERT(ip->i_df.if_bytes > 0);
 658                 kmem_free(iip->ili_extents_buf);
 659                 iip->ili_extents_buf = NULL;
 660         }
 661         if (iip->ili_aextents_buf != NULL) {
 662                 ASSERT(ip->i_d.di_aformat == XFS_DINODE_FMT_EXTENTS);
 663                 ASSERT(ip->i_d.di_anextents > 0);
 664                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_AEXT);
 665                 ASSERT(ip->i_afp->if_bytes > 0);
 666                 kmem_free(iip->ili_aextents_buf);
 667                 iip->ili_aextents_buf = NULL;
 668         }
 669
 670         /*
 671          * Figure out if we should unlock the inode or not.
 672          */
 673         hold = iip->ili_flags & XFS_ILI_HOLD;
 674
 675         /*
 676          * Before clearing out the flags, remember whether we
 677          * are holding the inode's IO lock.
 678          */
 679         iolocked = iip->ili_flags & XFS_ILI_IOLOCKED_ANY;
 680
 681         /*
 682          * Clear out the fields of the inode log item particular
 683          * to the current transaction.
 684          */
 685         iip->ili_flags = 0;
 686
 687         /*
 688          * Unlock the inode if XFS_ILI_HOLD was not set.
 689          */
 690         if (!hold) {
 691                 lock_flags = XFS_ILOCK_EXCL;
 692                 if (iolocked & XFS_ILI_IOLOCKED_EXCL) {
 693                         lock_flags |= XFS_IOLOCK_EXCL;
 694                 } else if (iolocked & XFS_ILI_IOLOCKED_SHARED) {
 695                         lock_flags |= XFS_IOLOCK_SHARED;
 696                 }
 697                 xfs_iput(iip->ili_inode, lock_flags);
 698         }
 699 }
 700
 701 /*
 702  * This is called to find out where the oldest active copy of the
 703  * inode log item in the on disk log resides now that the last log
 704  * write of it completed at the given lsn.  Since we always re-log
 705  * all dirty data in an inode, the latest copy in the on disk log
 706  * is the only one that matters.  Therefore, simply return the
 707  * given lsn.
 708  */
 709 /*ARGSUSED*/
 710 STATIC xfs_lsn_t
 711 xfs_inode_item_committed(
 712         xfs_inode_log_item_t    *iip,
 713         xfs_lsn_t               lsn)
 714 {
 715         return (lsn);
 716 }
 717
 718 /*
 719  * This gets called by xfs_trans_push_ail(), when IOP_TRYLOCK
 720  * failed to get the inode flush lock but did get the inode locked SHARED.
 721  * Here we're trying to see if the inode buffer is incore, and if so whether it's
 722  * marked delayed write. If that's the case, we'll promote it and that will
 723  * allow the caller to write the buffer by triggering the xfsbufd to run.
 724  */
 725 STATIC void
 726 xfs_inode_item_pushbuf(
 727         xfs_inode_log_item_t    *iip)
 728 {
 729         xfs_inode_t     *ip;
 730         xfs_mount_t     *mp;
 731         xfs_buf_t       *bp;
 732
 733         ip = iip->ili_inode;
 734         ASSERT(xfs_isilocked(ip, XFS_ILOCK_SHARED));
 735
 736         /*
 737          * If a flush is not in progress anymore, chances are that the
 738          * inode was taken off the AIL. So, just get out.
 739          */
 740         if (completion_done(&ip->i_flush) ||
 741             ((iip->ili_item.li_flags & XFS_LI_IN_AIL) == 0)) {
 742                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
 743                 return;
 744         }
 745
 746         mp = ip->i_mount;
 747         bp = xfs_incore(mp->m_ddev_targp, iip->ili_format.ilf_blkno,
 748                     iip->ili_format.ilf_len, XBF_TRYLOCK);
 749
 750         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 751         if (!bp)
 752                 return;
 753         if (XFS_BUF_ISDELAYWRITE(bp))
 754                 xfs_buf_delwri_promote(bp);
 755         xfs_buf_relse(bp);
 756         return;
 757 }
 758
 759
 760 /*
 761  * This is called to asynchronously write the inode associated with this
 762  * inode log item out to disk. The inode will already have been locked by
 763  * a successful call to xfs_inode_item_trylock().
 764  */
 765 STATIC void
 766 xfs_inode_item_push(
 767         xfs_inode_log_item_t    *iip)
 768 {
 769         xfs_inode_t     *ip;
 770
 771         ip = iip->ili_inode;
 772
 773         ASSERT(xfs_isilocked(ip, XFS_ILOCK_SHARED));
 774         ASSERT(!completion_done(&ip->i_flush));
 775         /*
 776          * Since we were able to lock the inode's flush lock and
 777          * we found it on the AIL, the inode must be dirty.  This
 778          * is because the inode is removed from the AIL while still
 779          * holding the flush lock in xfs_iflush_done().  Thus, if
 780          * we found it in the AIL and were able to obtain the flush
 781          * lock without sleeping, then there must not have been
 782          * anyone in the process of flushing the inode.
 783          */
 784         ASSERT(XFS_FORCED_SHUTDOWN(ip->i_mount) ||
 785                iip->ili_format.ilf_fields != 0);
 786
 787         /*
 788          * Push the inode to it's backing buffer. This will not remove the
 789          * inode from the AIL - a further push will be required to trigger a
 790          * buffer push. However, this allows all the dirty inodes to be pushed
 791          * to the buffer before it is pushed to disk. THe buffer IO completion
 792          * will pull th einode from the AIL, mark it clean and unlock the flush
 793          * lock.
 794          */
 795         (void) xfs_iflush(ip, 0);
 796         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 797
 798         return;
 799 }
 800
 801 /*
 802  * XXX rcc - this one really has to do something.  Probably needs
 803  * to stamp in a new field in the incore inode.
 804  */
 805 /* ARGSUSED */
 806 STATIC void
 807 xfs_inode_item_committing(
 808         xfs_inode_log_item_t    *iip,
 809         xfs_lsn_t               lsn)
 810 {
 811         iip->ili_last_lsn = lsn;
 812         return;
 813 }
 814
 815 /*
 816  * This is the ops vector shared by all buf log items.
 817  */
 818 static struct xfs_item_ops xfs_inode_item_ops = {
 819         .iop_size       = (uint(*)(xfs_log_item_t*))xfs_inode_item_size,
 820         .iop_format     = (void(*)(xfs_log_item_t*, xfs_log_iovec_t*))
 821                                         xfs_inode_item_format,
 822         .iop_pin        = (void(*)(xfs_log_item_t*))xfs_inode_item_pin,
 823         .iop_unpin      = (void(*)(xfs_log_item_t*, int))xfs_inode_item_unpin,
 824         .iop_trylock    = (uint(*)(xfs_log_item_t*))xfs_inode_item_trylock,
 825         .iop_unlock     = (void(*)(xfs_log_item_t*))xfs_inode_item_unlock,
 826         .iop_committed  = (xfs_lsn_t(*)(xfs_log_item_t*, xfs_lsn_t))
 827                                         xfs_inode_item_committed,
 828         .iop_push       = (void(*)(xfs_log_item_t*))xfs_inode_item_push,
 829         .iop_pushbuf    = (void(*)(xfs_log_item_t*))xfs_inode_item_pushbuf,
 830         .iop_committing = (void(*)(xfs_log_item_t*, xfs_lsn_t))
 831                                         xfs_inode_item_committing
 832 };
 833
 834
 835 /*
 836  * Initialize the inode log item for a newly allocated (in-core) inode.
 837  */
 838 void
 839 xfs_inode_item_init(
 840         xfs_inode_t     *ip,
 841         xfs_mount_t     *mp)
 842 {
 843         xfs_inode_log_item_t    *iip;
 844
 845         ASSERT(ip->i_itemp == NULL);
 846         iip = ip->i_itemp = kmem_zone_zalloc(xfs_ili_zone, KM_SLEEP);
 847
 848         iip->ili_inode = ip;
 849         xfs_log_item_init(mp, &iip->ili_item, XFS_LI_INODE,
 850                                                 &xfs_inode_item_ops);
 851         iip->ili_format.ilf_type = XFS_LI_INODE;
 852         iip->ili_format.ilf_ino = ip->i_ino;
 853         iip->ili_format.ilf_blkno = ip->i_imap.im_blkno;
 854         iip->ili_format.ilf_len = ip->i_imap.im_len;
 855         iip->ili_format.ilf_boffset = ip->i_imap.im_boffset;
 856 }
 857
 858 /*
 859  * Free the inode log item and any memory hanging off of it.
 860  */
 861 void
 862 xfs_inode_item_destroy(
 863         xfs_inode_t     *ip)
 864 {
 865 #ifdef XFS_TRANS_DEBUG
 866         if (ip->i_itemp->ili_root_size != 0) {
 867                 kmem_free(ip->i_itemp->ili_orig_root);
 868         }
 869 #endif
 870         kmem_zone_free(xfs_ili_zone, ip->i_itemp);
 871 }
 872
 873
 874 /*
 875  * This is the inode flushing I/O completion routine.  It is called
 876  * from interrupt level when the buffer containing the inode is
 877  * flushed to disk.  It is responsible for removing the inode item
 878  * from the AIL if it has not been re-logged, and unlocking the inode's
 879  * flush lock.
 880  */
 881 /*ARGSUSED*/
 882 void
 883 xfs_iflush_done(
 884         xfs_buf_t               *bp,
 885         xfs_inode_log_item_t    *iip)
 886 {
 887         xfs_inode_t             *ip = iip->ili_inode;
 888         struct xfs_ail          *ailp = iip->ili_item.li_ailp;
 889
 890         /*
 891          * We only want to pull the item from the AIL if it is
 892          * actually there and its location in the log has not
 893          * changed since we started the flush.  Thus, we only bother
 894          * if the ili_logged flag is set and the inode's lsn has not
 895          * changed.  First we check the lsn outside
 896          * the lock since it's cheaper, and then we recheck while
 897          * holding the lock before removing the inode from the AIL.
 898          */
 899         if (iip->ili_logged &&
 900             (iip->ili_item.li_lsn == iip->ili_flush_lsn)) {
 901                 spin_lock(&ailp->xa_lock);
 902                 if (iip->ili_item.li_lsn == iip->ili_flush_lsn) {
 903                         /* xfs_trans_ail_delete() drops the AIL lock. */
 904                         xfs_trans_ail_delete(ailp, (xfs_log_item_t*)iip);
 905                 } else {
 906                         spin_unlock(&ailp->xa_lock);
 907                 }
 908         }
 909
 910         iip->ili_logged = 0;
 911
 912         /*
 913          * Clear the ili_last_fields bits now that we know that the
 914          * data corresponding to them is safely on disk.
 915          */
 916         iip->ili_last_fields = 0;
 917
 918         /*
 919          * Release the inode's flush lock since we're done with it.
 920          */
 921         xfs_ifunlock(ip);
 922
 923         return;
 924 }
 925
 926 /*
 927  * This is the inode flushing abort routine.  It is called
 928  * from xfs_iflush when the filesystem is shutting down to clean
 929  * up the inode state.
 930  * It is responsible for removing the inode item
 931  * from the AIL if it has not been re-logged, and unlocking the inode's
 932  * flush lock.
 933  */
 934 void
 935 xfs_iflush_abort(
 936         xfs_inode_t             *ip)
 937 {
 938         xfs_inode_log_item_t    *iip = ip->i_itemp;
 939         xfs_mount_t             *mp;
 940
 941         iip = ip->i_itemp;
 942         mp = ip->i_mount;
 943         if (iip) {
 944                 struct xfs_ail  *ailp = iip->ili_item.li_ailp;
 945                 if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
 946                         spin_lock(&ailp->xa_lock);
 947                         if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
 948                                 /* xfs_trans_ail_delete() drops the AIL lock. */
 949                                 xfs_trans_ail_delete(ailp, (xfs_log_item_t *)iip);
 950                         } else
 951                                 spin_unlock(&ailp->xa_lock);
 952                 }
 953                 iip->ili_logged = 0;
 954                 /*
 955                  * Clear the ili_last_fields bits now that we know that the
 956                  * data corresponding to them is safely on disk.
 957                  */
 958                 iip->ili_last_fields = 0;
 959                 /*
 960                  * Clear the inode logging fields so no more flushes are
 961                  * attempted.
 962                  */
 963                 iip->ili_format.ilf_fields = 0;
 964         }
 965         /*
 966          * Release the inode's flush lock since we're done with it.
 967          */
 968         xfs_ifunlock(ip);
 969 }
 970
 971 void
 972 xfs_istale_done(
 973         xfs_buf_t               *bp,
 974         xfs_inode_log_item_t    *iip)
 975 {
 976         xfs_iflush_abort(iip->ili_inode);
 977 }
 978
 979 /*
 980  * convert an xfs_inode_log_format struct from either 32 or 64 bit versions
 981  * (which can have different field alignments) to the native version
 982  */
 983 int
 984 xfs_inode_item_format_convert(
 985         xfs_log_iovec_t         *buf,
 986         xfs_inode_log_format_t  *in_f)
 987 {
 988         if (buf->i_len == sizeof(xfs_inode_log_format_32_t)) {
 989                 xfs_inode_log_format_32_t *in_f32;
 990
 991                 in_f32 = (xfs_inode_log_format_32_t *)buf->i_addr;
 992                 in_f->ilf_type = in_f32->ilf_type;
 993                 in_f->ilf_size = in_f32->ilf_size;
 994                 in_f->ilf_fields = in_f32->ilf_fields;
 995                 in_f->ilf_asize = in_f32->ilf_asize;
 996                 in_f->ilf_dsize = in_f32->ilf_dsize;
 997                 in_f->ilf_ino = in_f32->ilf_ino;
 998                 /* copy biggest field of ilf_u */
 999                 memcpy(in_f->ilf_u.ilfu_uuid.__u_bits,
1000                        in_f32->ilf_u.ilfu_uuid.__u_bits,
1001                        sizeof(uuid_t));
1002                 in_f->ilf_blkno = in_f32->ilf_blkno;
1003                 in_f->ilf_len = in_f32->ilf_len;
1004                 in_f->ilf_boffset = in_f32->ilf_boffset;
1005                 return 0;
1006         } else if (buf->i_len == sizeof(xfs_inode_log_format_64_t)){
1007                 xfs_inode_log_format_64_t *in_f64;
1008
1009                 in_f64 = (xfs_inode_log_format_64_t *)buf->i_addr;
1010                 in_f->ilf_type = in_f64->ilf_type;
1011                 in_f->ilf_size = in_f64->ilf_size;
1012                 in_f->ilf_fields = in_f64->ilf_fields;
1013                 in_f->ilf_asize = in_f64->ilf_asize;
1014                 in_f->ilf_dsize = in_f64->ilf_dsize;
1015                 in_f->ilf_ino = in_f64->ilf_ino;
1016                 /* copy biggest field of ilf_u */
1017                 memcpy(in_f->ilf_u.ilfu_uuid.__u_bits,
1018                        in_f64->ilf_u.ilfu_uuid.__u_bits,
1019                        sizeof(uuid_t));
1020                 in_f->ilf_blkno = in_f64->ilf_blkno;
1021                 in_f->ilf_len = in_f64->ilf_len;
1022                 in_f->ilf_boffset = in_f64->ilf_boffset;
1023                 return 0;
1024         }
1025         return EFSCORRUPTED;
1026 }