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 static inline struct xfs_inode_log_item *INODE_ITEM(struct xfs_log_item *lip)
  40 {
  41         return container_of(lip, struct xfs_inode_log_item, ili_item);
  42 }
  43
  44
  45 /*
  46  * This returns the number of iovecs needed to log the given inode item.
  47  *
  48  * We need one iovec for the inode log format structure, one for the
  49  * inode core, and possibly one for the inode data/extents/b-tree root
  50  * and one for the inode attribute data/extents/b-tree root.
  51  */
  52 STATIC uint
  53 xfs_inode_item_size(
  54         struct xfs_log_item     *lip)
  55 {
  56         struct xfs_inode_log_item *iip = INODE_ITEM(lip);
  57         struct xfs_inode        *ip = iip->ili_inode;
  58         uint                    nvecs = 2;
  59
  60         /*
  61          * Only log the data/extents/b-tree root if there is something
  62          * left to log.
  63          */
  64         iip->ili_format.ilf_fields |= XFS_ILOG_CORE;
  65
  66         switch (ip->i_d.di_format) {
  67         case XFS_DINODE_FMT_EXTENTS:
  68                 iip->ili_format.ilf_fields &=
  69                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
  70                           XFS_ILOG_DEV | XFS_ILOG_UUID);
  71                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DEXT) &&
  72                     (ip->i_d.di_nextents > 0) &&
  73                     (ip->i_df.if_bytes > 0)) {
  74                         ASSERT(ip->i_df.if_u1.if_extents != NULL);
  75                         nvecs++;
  76                 } else {
  77                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DEXT;
  78                 }
  79                 break;
  80
  81         case XFS_DINODE_FMT_BTREE:
  82                 ASSERT(ip->i_df.if_ext_max ==
  83                        XFS_IFORK_DSIZE(ip) / (uint)sizeof(xfs_bmbt_rec_t));
  84                 iip->ili_format.ilf_fields &=
  85                         ~(XFS_ILOG_DDATA | XFS_ILOG_DEXT |
  86                           XFS_ILOG_DEV | XFS_ILOG_UUID);
  87                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DBROOT) &&
  88                     (ip->i_df.if_broot_bytes > 0)) {
  89                         ASSERT(ip->i_df.if_broot != NULL);
  90                         nvecs++;
  91                 } else {
  92                         ASSERT(!(iip->ili_format.ilf_fields &
  93                                  XFS_ILOG_DBROOT));
  94 #ifdef XFS_TRANS_DEBUG
  95                         if (iip->ili_root_size > 0) {
  96                                 ASSERT(iip->ili_root_size ==
  97                                        ip->i_df.if_broot_bytes);
  98                                 ASSERT(memcmp(iip->ili_orig_root,
  99                                             ip->i_df.if_broot,
 100                                             iip->ili_root_size) == 0);
 101                         } else {
 102                                 ASSERT(ip->i_df.if_broot_bytes == 0);
 103                         }
 104 #endif
 105                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DBROOT;
 106                 }
 107                 break;
 108
 109         case XFS_DINODE_FMT_LOCAL:
 110                 iip->ili_format.ilf_fields &=
 111                         ~(XFS_ILOG_DEXT | XFS_ILOG_DBROOT |
 112                           XFS_ILOG_DEV | XFS_ILOG_UUID);
 113                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DDATA) &&
 114                     (ip->i_df.if_bytes > 0)) {
 115                         ASSERT(ip->i_df.if_u1.if_data != NULL);
 116                         ASSERT(ip->i_d.di_size > 0);
 117                         nvecs++;
 118                 } else {
 119                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DDATA;
 120                 }
 121                 break;
 122
 123         case XFS_DINODE_FMT_DEV:
 124                 iip->ili_format.ilf_fields &=
 125                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 126                           XFS_ILOG_DEXT | XFS_ILOG_UUID);
 127                 break;
 128
 129         case XFS_DINODE_FMT_UUID:
 130                 iip->ili_format.ilf_fields &=
 131                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 132                           XFS_ILOG_DEXT | XFS_ILOG_DEV);
 133                 break;
 134
 135         default:
 136                 ASSERT(0);
 137                 break;
 138         }
 139
 140         /*
 141          * If there are no attributes associated with this file,
 142          * then there cannot be anything more to log.
 143          * Clear all attribute-related log flags.
 144          */
 145         if (!XFS_IFORK_Q(ip)) {
 146                 iip->ili_format.ilf_fields &=
 147                         ~(XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT);
 148                 return nvecs;
 149         }
 150
 151         /*
 152          * Log any necessary attribute data.
 153          */
 154         switch (ip->i_d.di_aformat) {
 155         case XFS_DINODE_FMT_EXTENTS:
 156                 iip->ili_format.ilf_fields &=
 157                         ~(XFS_ILOG_ADATA | XFS_ILOG_ABROOT);
 158                 if ((iip->ili_format.ilf_fields & XFS_ILOG_AEXT) &&
 159                     (ip->i_d.di_anextents > 0) &&
 160                     (ip->i_afp->if_bytes > 0)) {
 161                         ASSERT(ip->i_afp->if_u1.if_extents != NULL);
 162                         nvecs++;
 163                 } else {
 164                         iip->ili_format.ilf_fields &= ~XFS_ILOG_AEXT;
 165                 }
 166                 break;
 167
 168         case XFS_DINODE_FMT_BTREE:
 169                 iip->ili_format.ilf_fields &=
 170                         ~(XFS_ILOG_ADATA | XFS_ILOG_AEXT);
 171                 if ((iip->ili_format.ilf_fields & XFS_ILOG_ABROOT) &&
 172                     (ip->i_afp->if_broot_bytes > 0)) {
 173                         ASSERT(ip->i_afp->if_broot != NULL);
 174                         nvecs++;
 175                 } else {
 176                         iip->ili_format.ilf_fields &= ~XFS_ILOG_ABROOT;
 177                 }
 178                 break;
 179
 180         case XFS_DINODE_FMT_LOCAL:
 181                 iip->ili_format.ilf_fields &=
 182                         ~(XFS_ILOG_AEXT | XFS_ILOG_ABROOT);
 183                 if ((iip->ili_format.ilf_fields & XFS_ILOG_ADATA) &&
 184                     (ip->i_afp->if_bytes > 0)) {
 185                         ASSERT(ip->i_afp->if_u1.if_data != NULL);
 186                         nvecs++;
 187                 } else {
 188                         iip->ili_format.ilf_fields &= ~XFS_ILOG_ADATA;
 189                 }
 190                 break;
 191
 192         default:
 193                 ASSERT(0);
 194                 break;
 195         }
 196
 197         return nvecs;
 198 }
 199
 200 /*
 201  * This is called to fill in the vector of log iovecs for the
 202  * given inode log item.  It fills the first item with an inode
 203  * log format structure, the second with the on-disk inode structure,
 204  * and a possible third and/or fourth with the inode data/extents/b-tree
 205  * root and inode attributes data/extents/b-tree root.
 206  */
 207 STATIC void
 208 xfs_inode_item_format(
 209         struct xfs_log_item     *lip,
 210         struct xfs_log_iovec    *vecp)
 211 {
 212         struct xfs_inode_log_item *iip = INODE_ITEM(lip);
 213         struct xfs_inode        *ip = iip->ili_inode;
 214         uint                    nvecs;
 215         size_t                  data_bytes;
 216         xfs_bmbt_rec_t          *ext_buffer;
 217         int                     nrecs;
 218         xfs_mount_t             *mp;
 219
 220         vecp->i_addr = &iip->ili_format;
 221         vecp->i_len  = sizeof(xfs_inode_log_format_t);
 222         vecp->i_type = XLOG_REG_TYPE_IFORMAT;
 223         vecp++;
 224         nvecs        = 1;
 225
 226         /*
 227          * Make sure the linux inode is dirty. We do this before
 228          * clearing i_update_core as the VFS will call back into
 229          * XFS here and set i_update_core, so we need to dirty the
 230          * inode first so that the ordering of i_update_core and
 231          * unlogged modifications still works as described below.
 232          */
 233         xfs_mark_inode_dirty_sync(ip);
 234
 235         /*
 236          * Clear i_update_core if the timestamps (or any other
 237          * non-transactional modification) need flushing/logging
 238          * and we're about to log them with the rest of the core.
 239          *
 240          * This is the same logic as xfs_iflush() but this code can't
 241          * run at the same time as xfs_iflush because we're in commit
 242          * processing here and so we have the inode lock held in
 243          * exclusive mode.  Although it doesn't really matter
 244          * for the timestamps if both routines were to grab the
 245          * timestamps or not.  That would be ok.
 246          *
 247          * We clear i_update_core before copying out the data.
 248          * This is for coordination with our timestamp updates
 249          * that don't hold the inode lock. They will always
 250          * update the timestamps BEFORE setting i_update_core,
 251          * so if we clear i_update_core after they set it we
 252          * are guaranteed to see their updates to the timestamps
 253          * either here.  Likewise, if they set it after we clear it
 254          * here, we'll see it either on the next commit of this
 255          * inode or the next time the inode gets flushed via
 256          * xfs_iflush().  This depends on strongly ordered memory
 257          * semantics, but we have that.  We use the SYNCHRONIZE
 258          * macro to make sure that the compiler does not reorder
 259          * the i_update_core access below the data copy below.
 260          */
 261         if (ip->i_update_core)  {
 262                 ip->i_update_core = 0;
 263                 SYNCHRONIZE();
 264         }
 265
 266         /*
 267          * Make sure to get the latest timestamps from the Linux inode.
 268          */
 269         xfs_synchronize_times(ip);
 270
 271         vecp->i_addr = &ip->i_d;
 272         vecp->i_len  = sizeof(struct xfs_icdinode);
 273         vecp->i_type = XLOG_REG_TYPE_ICORE;
 274         vecp++;
 275         nvecs++;
 276         iip->ili_format.ilf_fields |= XFS_ILOG_CORE;
 277
 278         /*
 279          * If this is really an old format inode, then we need to
 280          * log it as such.  This means that we have to copy the link
 281          * count from the new field to the old.  We don't have to worry
 282          * about the new fields, because nothing trusts them as long as
 283          * the old inode version number is there.  If the superblock already
 284          * has a new version number, then we don't bother converting back.
 285          */
 286         mp = ip->i_mount;
 287         ASSERT(ip->i_d.di_version == 1 || xfs_sb_version_hasnlink(&mp->m_sb));
 288         if (ip->i_d.di_version == 1) {
 289                 if (!xfs_sb_version_hasnlink(&mp->m_sb)) {
 290                         /*
 291                          * Convert it back.
 292                          */
 293                         ASSERT(ip->i_d.di_nlink <= XFS_MAXLINK_1);
 294                         ip->i_d.di_onlink = ip->i_d.di_nlink;
 295                 } else {
 296                         /*
 297                          * The superblock version has already been bumped,
 298                          * so just make the conversion to the new inode
 299                          * format permanent.
 300                          */
 301                         ip->i_d.di_version = 2;
 302                         ip->i_d.di_onlink = 0;
 303                         memset(&(ip->i_d.di_pad[0]), 0, sizeof(ip->i_d.di_pad));
 304                 }
 305         }
 306
 307         switch (ip->i_d.di_format) {
 308         case XFS_DINODE_FMT_EXTENTS:
 309                 ASSERT(!(iip->ili_format.ilf_fields &
 310                          (XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 311                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 312                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEXT) {
 313                         ASSERT(ip->i_df.if_bytes > 0);
 314                         ASSERT(ip->i_df.if_u1.if_extents != NULL);
 315                         ASSERT(ip->i_d.di_nextents > 0);
 316                         ASSERT(iip->ili_extents_buf == NULL);
 317                         nrecs = ip->i_df.if_bytes /
 318                                 (uint)sizeof(xfs_bmbt_rec_t);
 319                         ASSERT(nrecs > 0);
 320 #ifdef XFS_NATIVE_HOST
 321                         if (nrecs == ip->i_d.di_nextents) {
 322                                 /*
 323                                  * There are no delayed allocation
 324                                  * extents, so just point to the
 325                                  * real extents array.
 326                                  */
 327                                 vecp->i_addr = 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 = 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 = 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 = 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 == lip->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 = 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 = 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 = 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 = 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 == lip->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         struct xfs_log_item     *lip)
 533 {
 534         struct xfs_inode        *ip = INODE_ITEM(lip)->ili_inode;
 535
 536         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 537
 538         trace_xfs_inode_pin(ip, _RET_IP_);
 539         atomic_inc(&ip->i_pincount);
 540 }
 541
 542
 543 /*
 544  * This is called to unpin the inode associated with the inode log
 545  * item which was previously pinned with a call to xfs_inode_item_pin().
 546  *
 547  * Also wake up anyone in xfs_iunpin_wait() if the count goes to 0.
 548  */
 549 STATIC void
 550 xfs_inode_item_unpin(
 551         struct xfs_log_item     *lip,
 552         int                     remove)
 553 {
 554         struct xfs_inode        *ip = INODE_ITEM(lip)->ili_inode;
 555
 556         trace_xfs_inode_unpin(ip, _RET_IP_);
 557         ASSERT(atomic_read(&ip->i_pincount) > 0);
 558         if (atomic_dec_and_test(&ip->i_pincount))
 559                 wake_up(&ip->i_ipin_wait);
 560 }
 561
 562 /*
 563  * This is called to attempt to lock the inode associated with this
 564  * inode log item, in preparation for the push routine which does the actual
 565  * iflush.  Don't sleep on the inode lock or the flush lock.
 566  *
 567  * If the flush lock is already held, indicating that the inode has
 568  * been or is in the process of being flushed, then (ideally) we'd like to
 569  * see if the inode's buffer is still incore, and if so give it a nudge.
 570  * We delay doing so until the pushbuf routine, though, to avoid holding
 571  * the AIL lock across a call to the blackhole which is the buffer cache.
 572  * Also we don't want to sleep in any device strategy routines, which can happen
 573  * if we do the subsequent bawrite in here.
 574  */
 575 STATIC uint
 576 xfs_inode_item_trylock(
 577         struct xfs_log_item     *lip)
 578 {
 579         struct xfs_inode_log_item *iip = INODE_ITEM(lip);
 580         struct xfs_inode        *ip = iip->ili_inode;
 581
 582         if (xfs_ipincount(ip) > 0)
 583                 return XFS_ITEM_PINNED;
 584
 585         if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED))
 586                 return XFS_ITEM_LOCKED;
 587
 588         if (!xfs_iflock_nowait(ip)) {
 589                 /*
 590                  * inode has already been flushed to the backing buffer,
 591                  * leave it locked in shared mode, pushbuf routine will
 592                  * unlock it.
 593                  */
 594                 return XFS_ITEM_PUSHBUF;
 595         }
 596
 597         /* Stale items should force out the iclog */
 598         if (ip->i_flags & XFS_ISTALE) {
 599                 xfs_ifunlock(ip);
 600                 /*
 601                  * we hold the AIL lock - notify the unlock routine of this
 602                  * so it doesn't try to get the lock again.
 603                  */
 604                 xfs_iunlock(ip, XFS_ILOCK_SHARED|XFS_IUNLOCK_NONOTIFY);
 605                 return XFS_ITEM_PINNED;
 606         }
 607
 608 #ifdef DEBUG
 609         if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
 610                 ASSERT(iip->ili_format.ilf_fields != 0);
 611                 ASSERT(iip->ili_logged == 0);
 612                 ASSERT(lip->li_flags & XFS_LI_IN_AIL);
 613         }
 614 #endif
 615         return XFS_ITEM_SUCCESS;
 616 }
 617
 618 /*
 619  * Unlock the inode associated with the inode log item.
 620  * Clear the fields of the inode and inode log item that
 621  * are specific to the current transaction.  If the
 622  * hold flags is set, do not unlock the inode.
 623  */
 624 STATIC void
 625 xfs_inode_item_unlock(
 626         struct xfs_log_item     *lip)
 627 {
 628         struct xfs_inode_log_item *iip = INODE_ITEM(lip);
 629         struct xfs_inode        *ip = iip->ili_inode;
 630         unsigned short          lock_flags;
 631
 632         ASSERT(iip->ili_inode->i_itemp != NULL);
 633         ASSERT(xfs_isilocked(iip->ili_inode, XFS_ILOCK_EXCL));
 634
 635         /*
 636          * Clear the transaction pointer in the inode.
 637          */
 638         ip->i_transp = NULL;
 639
 640         /*
 641          * If the inode needed a separate buffer with which to log
 642          * its extents, then free it now.
 643          */
 644         if (iip->ili_extents_buf != NULL) {
 645                 ASSERT(ip->i_d.di_format == XFS_DINODE_FMT_EXTENTS);
 646                 ASSERT(ip->i_d.di_nextents > 0);
 647                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_DEXT);
 648                 ASSERT(ip->i_df.if_bytes > 0);
 649                 kmem_free(iip->ili_extents_buf);
 650                 iip->ili_extents_buf = NULL;
 651         }
 652         if (iip->ili_aextents_buf != NULL) {
 653                 ASSERT(ip->i_d.di_aformat == XFS_DINODE_FMT_EXTENTS);
 654                 ASSERT(ip->i_d.di_anextents > 0);
 655                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_AEXT);
 656                 ASSERT(ip->i_afp->if_bytes > 0);
 657                 kmem_free(iip->ili_aextents_buf);
 658                 iip->ili_aextents_buf = NULL;
 659         }
 660
 661         lock_flags = iip->ili_lock_flags;
 662         iip->ili_lock_flags = 0;
 663         if (lock_flags) {
 664                 xfs_iunlock(iip->ili_inode, lock_flags);
 665                 IRELE(iip->ili_inode);
 666         }
 667 }
 668
 669 /*
 670  * This is called to find out where the oldest active copy of the
 671  * inode log item in the on disk log resides now that the last log
 672  * write of it completed at the given lsn.  Since we always re-log
 673  * all dirty data in an inode, the latest copy in the on disk log
 674  * is the only one that matters.  Therefore, simply return the
 675  * given lsn.
 676  */
 677 STATIC xfs_lsn_t
 678 xfs_inode_item_committed(
 679         struct xfs_log_item     *lip,
 680         xfs_lsn_t               lsn)
 681 {
 682         return lsn;
 683 }
 684
 685 /*
 686  * This gets called by xfs_trans_push_ail(), when IOP_TRYLOCK
 687  * failed to get the inode flush lock but did get the inode locked SHARED.
 688  * Here we're trying to see if the inode buffer is incore, and if so whether it's
 689  * marked delayed write. If that's the case, we'll promote it and that will
 690  * allow the caller to write the buffer by triggering the xfsbufd to run.
 691  */
 692 STATIC void
 693 xfs_inode_item_pushbuf(
 694         struct xfs_log_item     *lip)
 695 {
 696         struct xfs_inode_log_item *iip = INODE_ITEM(lip);
 697         struct xfs_inode        *ip = iip->ili_inode;
 698         struct xfs_buf          *bp;
 699
 700         ASSERT(xfs_isilocked(ip, XFS_ILOCK_SHARED));
 701
 702         /*
 703          * If a flush is not in progress anymore, chances are that the
 704          * inode was taken off the AIL. So, just get out.
 705          */
 706         if (completion_done(&ip->i_flush) ||
 707             !(lip->li_flags & XFS_LI_IN_AIL)) {
 708                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
 709                 return;
 710         }
 711
 712         bp = xfs_incore(ip->i_mount->m_ddev_targp, iip->ili_format.ilf_blkno,
 713                         iip->ili_format.ilf_len, XBF_TRYLOCK);
 714
 715         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 716         if (!bp)
 717                 return;
 718         if (XFS_BUF_ISDELAYWRITE(bp))
 719                 xfs_buf_delwri_promote(bp);
 720         xfs_buf_relse(bp);
 721 }
 722
 723 /*
 724  * This is called to asynchronously write the inode associated with this
 725  * inode log item out to disk. The inode will already have been locked by
 726  * a successful call to xfs_inode_item_trylock().
 727  */
 728 STATIC void
 729 xfs_inode_item_push(
 730         struct xfs_log_item     *lip)
 731 {
 732         struct xfs_inode_log_item *iip = INODE_ITEM(lip);
 733         struct xfs_inode        *ip = iip->ili_inode;
 734
 735         ASSERT(xfs_isilocked(ip, XFS_ILOCK_SHARED));
 736         ASSERT(!completion_done(&ip->i_flush));
 737
 738         /*
 739          * Since we were able to lock the inode's flush lock and
 740          * we found it on the AIL, the inode must be dirty.  This
 741          * is because the inode is removed from the AIL while still
 742          * holding the flush lock in xfs_iflush_done().  Thus, if
 743          * we found it in the AIL and were able to obtain the flush
 744          * lock without sleeping, then there must not have been
 745          * anyone in the process of flushing the inode.
 746          */
 747         ASSERT(XFS_FORCED_SHUTDOWN(ip->i_mount) ||
 748                iip->ili_format.ilf_fields != 0);
 749
 750         /*
 751          * Push the inode to it's backing buffer. This will not remove the
 752          * inode from the AIL - a further push will be required to trigger a
 753          * buffer push. However, this allows all the dirty inodes to be pushed
 754          * to the buffer before it is pushed to disk. THe buffer IO completion
 755          * will pull th einode from the AIL, mark it clean and unlock the flush
 756          * lock.
 757          */
 758         (void) xfs_iflush(ip, 0);
 759         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 760 }
 761
 762 /*
 763  * XXX rcc - this one really has to do something.  Probably needs
 764  * to stamp in a new field in the incore inode.
 765  */
 766 STATIC void
 767 xfs_inode_item_committing(
 768         struct xfs_log_item     *lip,
 769         xfs_lsn_t               lsn)
 770 {
 771         INODE_ITEM(lip)->ili_last_lsn = lsn;
 772 }
 773
 774 /*
 775  * This is the ops vector shared by all buf log items.
 776  */
 777 static struct xfs_item_ops xfs_inode_item_ops = {
 778         .iop_size       = xfs_inode_item_size,
 779         .iop_format     = xfs_inode_item_format,
 780         .iop_pin        = xfs_inode_item_pin,
 781         .iop_unpin      = xfs_inode_item_unpin,
 782         .iop_trylock    = xfs_inode_item_trylock,
 783         .iop_unlock     = xfs_inode_item_unlock,
 784         .iop_committed  = xfs_inode_item_committed,
 785         .iop_push       = xfs_inode_item_push,
 786         .iop_pushbuf    = xfs_inode_item_pushbuf,
 787         .iop_committing = xfs_inode_item_committing
 788 };
 789
 790
 791 /*
 792  * Initialize the inode log item for a newly allocated (in-core) inode.
 793  */
 794 void
 795 xfs_inode_item_init(
 796         struct xfs_inode        *ip,
 797         struct xfs_mount        *mp)
 798 {
 799         struct xfs_inode_log_item *iip;
 800
 801         ASSERT(ip->i_itemp == NULL);
 802         iip = ip->i_itemp = kmem_zone_zalloc(xfs_ili_zone, KM_SLEEP);
 803
 804         iip->ili_inode = ip;
 805         xfs_log_item_init(mp, &iip->ili_item, XFS_LI_INODE,
 806                                                 &xfs_inode_item_ops);
 807         iip->ili_format.ilf_type = XFS_LI_INODE;
 808         iip->ili_format.ilf_ino = ip->i_ino;
 809         iip->ili_format.ilf_blkno = ip->i_imap.im_blkno;
 810         iip->ili_format.ilf_len = ip->i_imap.im_len;
 811         iip->ili_format.ilf_boffset = ip->i_imap.im_boffset;
 812 }
 813
 814 /*
 815  * Free the inode log item and any memory hanging off of it.
 816  */
 817 void
 818 xfs_inode_item_destroy(
 819         xfs_inode_t     *ip)
 820 {
 821 #ifdef XFS_TRANS_DEBUG
 822         if (ip->i_itemp->ili_root_size != 0) {
 823                 kmem_free(ip->i_itemp->ili_orig_root);
 824         }
 825 #endif
 826         kmem_zone_free(xfs_ili_zone, ip->i_itemp);
 827 }
 828
 829
 830 /*
 831  * This is the inode flushing I/O completion routine.  It is called
 832  * from interrupt level when the buffer containing the inode is
 833  * flushed to disk.  It is responsible for removing the inode item
 834  * from the AIL if it has not been re-logged, and unlocking the inode's
 835  * flush lock.
 836  */
 837 void
 838 xfs_iflush_done(
 839         struct xfs_buf          *bp,
 840         struct xfs_log_item     *lip)
 841 {
 842         struct xfs_inode_log_item *iip = INODE_ITEM(lip);
 843         xfs_inode_t             *ip = iip->ili_inode;
 844         struct xfs_ail          *ailp = lip->li_ailp;
 845
 846         /*
 847          * We only want to pull the item from the AIL if it is
 848          * actually there and its location in the log has not
 849          * changed since we started the flush.  Thus, we only bother
 850          * if the ili_logged flag is set and the inode's lsn has not
 851          * changed.  First we check the lsn outside
 852          * the lock since it's cheaper, and then we recheck while
 853          * holding the lock before removing the inode from the AIL.
 854          */
 855         if (iip->ili_logged && lip->li_lsn == iip->ili_flush_lsn) {
 856                 spin_lock(&ailp->xa_lock);
 857                 if (lip->li_lsn == iip->ili_flush_lsn) {
 858                         /* xfs_trans_ail_delete() drops the AIL lock. */
 859                         xfs_trans_ail_delete(ailp, lip);
 860                 } else {
 861                         spin_unlock(&ailp->xa_lock);
 862                 }
 863         }
 864
 865         iip->ili_logged = 0;
 866
 867         /*
 868          * Clear the ili_last_fields bits now that we know that the
 869          * data corresponding to them is safely on disk.
 870          */
 871         iip->ili_last_fields = 0;
 872
 873         /*
 874          * Release the inode's flush lock since we're done with it.
 875          */
 876         xfs_ifunlock(ip);
 877 }
 878
 879 /*
 880  * This is the inode flushing abort routine.  It is called
 881  * from xfs_iflush when the filesystem is shutting down to clean
 882  * up the inode state.
 883  * It is responsible for removing the inode item
 884  * from the AIL if it has not been re-logged, and unlocking the inode's
 885  * flush lock.
 886  */
 887 void
 888 xfs_iflush_abort(
 889         xfs_inode_t             *ip)
 890 {
 891         xfs_inode_log_item_t    *iip = ip->i_itemp;
 892         xfs_mount_t             *mp;
 893
 894         iip = ip->i_itemp;
 895         mp = ip->i_mount;
 896         if (iip) {
 897                 struct xfs_ail  *ailp = iip->ili_item.li_ailp;
 898                 if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
 899                         spin_lock(&ailp->xa_lock);
 900                         if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
 901                                 /* xfs_trans_ail_delete() drops the AIL lock. */
 902                                 xfs_trans_ail_delete(ailp, (xfs_log_item_t *)iip);
 903                         } else
 904                                 spin_unlock(&ailp->xa_lock);
 905                 }
 906                 iip->ili_logged = 0;
 907                 /*
 908                  * Clear the ili_last_fields bits now that we know that the
 909                  * data corresponding to them is safely on disk.
 910                  */
 911                 iip->ili_last_fields = 0;
 912                 /*
 913                  * Clear the inode logging fields so no more flushes are
 914                  * attempted.
 915                  */
 916                 iip->ili_format.ilf_fields = 0;
 917         }
 918         /*
 919          * Release the inode's flush lock since we're done with it.
 920          */
 921         xfs_ifunlock(ip);
 922 }
 923
 924 void
 925 xfs_istale_done(
 926         struct xfs_buf          *bp,
 927         struct xfs_log_item     *lip)
 928 {
 929         xfs_iflush_abort(INODE_ITEM(lip)->ili_inode);
 930 }
 931
 932 /*
 933  * convert an xfs_inode_log_format struct from either 32 or 64 bit versions
 934  * (which can have different field alignments) to the native version
 935  */
 936 int
 937 xfs_inode_item_format_convert(
 938         xfs_log_iovec_t         *buf,
 939         xfs_inode_log_format_t  *in_f)
 940 {
 941         if (buf->i_len == sizeof(xfs_inode_log_format_32_t)) {
 942                 xfs_inode_log_format_32_t *in_f32 = buf->i_addr;
 943
 944                 in_f->ilf_type = in_f32->ilf_type;
 945                 in_f->ilf_size = in_f32->ilf_size;
 946                 in_f->ilf_fields = in_f32->ilf_fields;
 947                 in_f->ilf_asize = in_f32->ilf_asize;
 948                 in_f->ilf_dsize = in_f32->ilf_dsize;
 949                 in_f->ilf_ino = in_f32->ilf_ino;
 950                 /* copy biggest field of ilf_u */
 951                 memcpy(in_f->ilf_u.ilfu_uuid.__u_bits,
 952                        in_f32->ilf_u.ilfu_uuid.__u_bits,
 953                        sizeof(uuid_t));
 954                 in_f->ilf_blkno = in_f32->ilf_blkno;
 955                 in_f->ilf_len = in_f32->ilf_len;
 956                 in_f->ilf_boffset = in_f32->ilf_boffset;
 957                 return 0;
 958         } else if (buf->i_len == sizeof(xfs_inode_log_format_64_t)){
 959                 xfs_inode_log_format_64_t *in_f64 = buf->i_addr;
 960
 961                 in_f->ilf_type = in_f64->ilf_type;
 962                 in_f->ilf_size = in_f64->ilf_size;
 963                 in_f->ilf_fields = in_f64->ilf_fields;
 964                 in_f->ilf_asize = in_f64->ilf_asize;
 965                 in_f->ilf_dsize = in_f64->ilf_dsize;
 966                 in_f->ilf_ino = in_f64->ilf_ino;
 967                 /* copy biggest field of ilf_u */
 968                 memcpy(in_f->ilf_u.ilfu_uuid.__u_bits,
 969                        in_f64->ilf_u.ilfu_uuid.__u_bits,
 970                        sizeof(uuid_t));
 971                 in_f->ilf_blkno = in_f64->ilf_blkno;
 972                 in_f->ilf_len = in_f64->ilf_len;
 973                 in_f->ilf_boffset = in_f64->ilf_boffset;
 974                 return 0;
 975         }
 976         return EFSCORRUPTED;
 977 }