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         xfs_mount_t             *mp;
 218
 219         vecp->i_addr = &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 = &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                         ASSERT((ip->i_df.if_bytes /
 317                                 (uint)sizeof(xfs_bmbt_rec_t)) > 0);
 318 #ifdef XFS_NATIVE_HOST
 319                        if (ip->i_d.di_nextents == ip->i_df.if_bytes /
 320                                                (uint)sizeof(xfs_bmbt_rec_t)) {
 321                                 /*
 322                                  * There are no delayed allocation
 323                                  * extents, so just point to the
 324                                  * real extents array.
 325                                  */
 326                                 vecp->i_addr = ip->i_df.if_u1.if_extents;
 327                                 vecp->i_len = ip->i_df.if_bytes;
 328                                 vecp->i_type = XLOG_REG_TYPE_IEXT;
 329                         } else
 330 #endif
 331                         {
 332                                 /*
 333                                  * There are delayed allocation extents
 334                                  * in the inode, or we need to convert
 335                                  * the extents to on disk format.
 336                                  * Use xfs_iextents_copy()
 337                                  * to copy only the real extents into
 338                                  * a separate buffer.  We'll free the
 339                                  * buffer in the unlock routine.
 340                                  */
 341                                 ext_buffer = kmem_alloc(ip->i_df.if_bytes,
 342                                         KM_SLEEP);
 343                                 iip->ili_extents_buf = ext_buffer;
 344                                 vecp->i_addr = ext_buffer;
 345                                 vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
 346                                                 XFS_DATA_FORK);
 347                                 vecp->i_type = XLOG_REG_TYPE_IEXT;
 348                         }
 349                         ASSERT(vecp->i_len <= ip->i_df.if_bytes);
 350                         iip->ili_format.ilf_dsize = vecp->i_len;
 351                         vecp++;
 352                         nvecs++;
 353                 }
 354                 break;
 355
 356         case XFS_DINODE_FMT_BTREE:
 357                 ASSERT(!(iip->ili_format.ilf_fields &
 358                          (XFS_ILOG_DDATA | XFS_ILOG_DEXT |
 359                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 360                 if (iip->ili_format.ilf_fields & XFS_ILOG_DBROOT) {
 361                         ASSERT(ip->i_df.if_broot_bytes > 0);
 362                         ASSERT(ip->i_df.if_broot != NULL);
 363                         vecp->i_addr = ip->i_df.if_broot;
 364                         vecp->i_len = ip->i_df.if_broot_bytes;
 365                         vecp->i_type = XLOG_REG_TYPE_IBROOT;
 366                         vecp++;
 367                         nvecs++;
 368                         iip->ili_format.ilf_dsize = ip->i_df.if_broot_bytes;
 369                 }
 370                 break;
 371
 372         case XFS_DINODE_FMT_LOCAL:
 373                 ASSERT(!(iip->ili_format.ilf_fields &
 374                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 375                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 376                 if (iip->ili_format.ilf_fields & XFS_ILOG_DDATA) {
 377                         ASSERT(ip->i_df.if_bytes > 0);
 378                         ASSERT(ip->i_df.if_u1.if_data != NULL);
 379                         ASSERT(ip->i_d.di_size > 0);
 380
 381                         vecp->i_addr = ip->i_df.if_u1.if_data;
 382                         /*
 383                          * Round i_bytes up to a word boundary.
 384                          * The underlying memory is guaranteed to
 385                          * to be there by xfs_idata_realloc().
 386                          */
 387                         data_bytes = roundup(ip->i_df.if_bytes, 4);
 388                         ASSERT((ip->i_df.if_real_bytes == 0) ||
 389                                (ip->i_df.if_real_bytes == data_bytes));
 390                         vecp->i_len = (int)data_bytes;
 391                         vecp->i_type = XLOG_REG_TYPE_ILOCAL;
 392                         vecp++;
 393                         nvecs++;
 394                         iip->ili_format.ilf_dsize = (unsigned)data_bytes;
 395                 }
 396                 break;
 397
 398         case XFS_DINODE_FMT_DEV:
 399                 ASSERT(!(iip->ili_format.ilf_fields &
 400                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 401                           XFS_ILOG_DDATA | XFS_ILOG_UUID)));
 402                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEV) {
 403                         iip->ili_format.ilf_u.ilfu_rdev =
 404                                 ip->i_df.if_u2.if_rdev;
 405                 }
 406                 break;
 407
 408         case XFS_DINODE_FMT_UUID:
 409                 ASSERT(!(iip->ili_format.ilf_fields &
 410                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 411                           XFS_ILOG_DDATA | XFS_ILOG_DEV)));
 412                 if (iip->ili_format.ilf_fields & XFS_ILOG_UUID) {
 413                         iip->ili_format.ilf_u.ilfu_uuid =
 414                                 ip->i_df.if_u2.if_uuid;
 415                 }
 416                 break;
 417
 418         default:
 419                 ASSERT(0);
 420                 break;
 421         }
 422
 423         /*
 424          * If there are no attributes associated with the file,
 425          * then we're done.
 426          * Assert that no attribute-related log flags are set.
 427          */
 428         if (!XFS_IFORK_Q(ip)) {
 429                 ASSERT(nvecs == lip->li_desc->lid_size);
 430                 iip->ili_format.ilf_size = nvecs;
 431                 ASSERT(!(iip->ili_format.ilf_fields &
 432                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
 433                 return;
 434         }
 435
 436         switch (ip->i_d.di_aformat) {
 437         case XFS_DINODE_FMT_EXTENTS:
 438                 ASSERT(!(iip->ili_format.ilf_fields &
 439                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT)));
 440                 if (iip->ili_format.ilf_fields & XFS_ILOG_AEXT) {
 441 #ifdef DEBUG
 442                         int nrecs = ip->i_afp->if_bytes /
 443                                 (uint)sizeof(xfs_bmbt_rec_t);
 444                         ASSERT(nrecs > 0);
 445                         ASSERT(nrecs == ip->i_d.di_anextents);
 446                         ASSERT(ip->i_afp->if_bytes > 0);
 447                         ASSERT(ip->i_afp->if_u1.if_extents != NULL);
 448                         ASSERT(ip->i_d.di_anextents > 0);
 449 #endif
 450 #ifdef XFS_NATIVE_HOST
 451                         /*
 452                          * There are not delayed allocation extents
 453                          * for attributes, so just point at the array.
 454                          */
 455                         vecp->i_addr = ip->i_afp->if_u1.if_extents;
 456                         vecp->i_len = ip->i_afp->if_bytes;
 457 #else
 458                         ASSERT(iip->ili_aextents_buf == NULL);
 459                         /*
 460                          * Need to endian flip before logging
 461                          */
 462                         ext_buffer = kmem_alloc(ip->i_afp->if_bytes,
 463                                 KM_SLEEP);
 464                         iip->ili_aextents_buf = ext_buffer;
 465                         vecp->i_addr = ext_buffer;
 466                         vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
 467                                         XFS_ATTR_FORK);
 468 #endif
 469                         vecp->i_type = XLOG_REG_TYPE_IATTR_EXT;
 470                         iip->ili_format.ilf_asize = vecp->i_len;
 471                         vecp++;
 472                         nvecs++;
 473                 }
 474                 break;
 475
 476         case XFS_DINODE_FMT_BTREE:
 477                 ASSERT(!(iip->ili_format.ilf_fields &
 478                          (XFS_ILOG_ADATA | XFS_ILOG_AEXT)));
 479                 if (iip->ili_format.ilf_fields & XFS_ILOG_ABROOT) {
 480                         ASSERT(ip->i_afp->if_broot_bytes > 0);
 481                         ASSERT(ip->i_afp->if_broot != NULL);
 482                         vecp->i_addr = ip->i_afp->if_broot;
 483                         vecp->i_len = ip->i_afp->if_broot_bytes;
 484                         vecp->i_type = XLOG_REG_TYPE_IATTR_BROOT;
 485                         vecp++;
 486                         nvecs++;
 487                         iip->ili_format.ilf_asize = ip->i_afp->if_broot_bytes;
 488                 }
 489                 break;
 490
 491         case XFS_DINODE_FMT_LOCAL:
 492                 ASSERT(!(iip->ili_format.ilf_fields &
 493                          (XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
 494                 if (iip->ili_format.ilf_fields & XFS_ILOG_ADATA) {
 495                         ASSERT(ip->i_afp->if_bytes > 0);
 496                         ASSERT(ip->i_afp->if_u1.if_data != NULL);
 497
 498                         vecp->i_addr = ip->i_afp->if_u1.if_data;
 499                         /*
 500                          * Round i_bytes up to a word boundary.
 501                          * The underlying memory is guaranteed to
 502                          * to be there by xfs_idata_realloc().
 503                          */
 504                         data_bytes = roundup(ip->i_afp->if_bytes, 4);
 505                         ASSERT((ip->i_afp->if_real_bytes == 0) ||
 506                                (ip->i_afp->if_real_bytes == data_bytes));
 507                         vecp->i_len = (int)data_bytes;
 508                         vecp->i_type = XLOG_REG_TYPE_IATTR_LOCAL;
 509                         vecp++;
 510                         nvecs++;
 511                         iip->ili_format.ilf_asize = (unsigned)data_bytes;
 512                 }
 513                 break;
 514
 515         default:
 516                 ASSERT(0);
 517                 break;
 518         }
 519
 520         ASSERT(nvecs == lip->li_desc->lid_size);
 521         iip->ili_format.ilf_size = nvecs;
 522 }
 523
 524
 525 /*
 526  * This is called to pin the inode associated with the inode log
 527  * item in memory so it cannot be written out.
 528  */
 529 STATIC void
 530 xfs_inode_item_pin(
 531         struct xfs_log_item     *lip)
 532 {
 533         struct xfs_inode        *ip = INODE_ITEM(lip)->ili_inode;
 534
 535         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 536
 537         trace_xfs_inode_pin(ip, _RET_IP_);
 538         atomic_inc(&ip->i_pincount);
 539 }
 540
 541
 542 /*
 543  * This is called to unpin the inode associated with the inode log
 544  * item which was previously pinned with a call to xfs_inode_item_pin().
 545  *
 546  * Also wake up anyone in xfs_iunpin_wait() if the count goes to 0.
 547  */
 548 STATIC void
 549 xfs_inode_item_unpin(
 550         struct xfs_log_item     *lip,
 551         int                     remove)
 552 {
 553         struct xfs_inode        *ip = INODE_ITEM(lip)->ili_inode;
 554
 555         trace_xfs_inode_unpin(ip, _RET_IP_);
 556         ASSERT(atomic_read(&ip->i_pincount) > 0);
 557         if (atomic_dec_and_test(&ip->i_pincount))
 558                 wake_up(&ip->i_ipin_wait);
 559 }
 560
 561 /*
 562  * This is called to attempt to lock the inode associated with this
 563  * inode log item, in preparation for the push routine which does the actual
 564  * iflush.  Don't sleep on the inode lock or the flush lock.
 565  *
 566  * If the flush lock is already held, indicating that the inode has
 567  * been or is in the process of being flushed, then (ideally) we'd like to
 568  * see if the inode's buffer is still incore, and if so give it a nudge.
 569  * We delay doing so until the pushbuf routine, though, to avoid holding
 570  * the AIL lock across a call to the blackhole which is the buffer cache.
 571  * Also we don't want to sleep in any device strategy routines, which can happen
 572  * if we do the subsequent bawrite in here.
 573  */
 574 STATIC uint
 575 xfs_inode_item_trylock(
 576         struct xfs_log_item     *lip)
 577 {
 578         struct xfs_inode_log_item *iip = INODE_ITEM(lip);
 579         struct xfs_inode        *ip = iip->ili_inode;
 580
 581         if (xfs_ipincount(ip) > 0)
 582                 return XFS_ITEM_PINNED;
 583
 584         if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED))
 585                 return XFS_ITEM_LOCKED;
 586
 587         if (!xfs_iflock_nowait(ip)) {
 588                 /*
 589                  * inode has already been flushed to the backing buffer,
 590                  * leave it locked in shared mode, pushbuf routine will
 591                  * unlock it.
 592                  */
 593                 return XFS_ITEM_PUSHBUF;
 594         }
 595
 596         /* Stale items should force out the iclog */
 597         if (ip->i_flags & XFS_ISTALE) {
 598                 xfs_ifunlock(ip);
 599                 /*
 600                  * we hold the AIL lock - notify the unlock routine of this
 601                  * so it doesn't try to get the lock again.
 602                  */
 603                 xfs_iunlock(ip, XFS_ILOCK_SHARED|XFS_IUNLOCK_NONOTIFY);
 604                 return XFS_ITEM_PINNED;
 605         }
 606
 607 #ifdef DEBUG
 608         if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
 609                 ASSERT(iip->ili_format.ilf_fields != 0);
 610                 ASSERT(iip->ili_logged == 0);
 611                 ASSERT(lip->li_flags & XFS_LI_IN_AIL);
 612         }
 613 #endif
 614         return XFS_ITEM_SUCCESS;
 615 }
 616
 617 /*
 618  * Unlock the inode associated with the inode log item.
 619  * Clear the fields of the inode and inode log item that
 620  * are specific to the current transaction.  If the
 621  * hold flags is set, do not unlock the inode.
 622  */
 623 STATIC void
 624 xfs_inode_item_unlock(
 625         struct xfs_log_item     *lip)
 626 {
 627         struct xfs_inode_log_item *iip = INODE_ITEM(lip);
 628         struct xfs_inode        *ip = iip->ili_inode;
 629         unsigned short          lock_flags;
 630
 631         ASSERT(iip->ili_inode->i_itemp != NULL);
 632         ASSERT(xfs_isilocked(iip->ili_inode, XFS_ILOCK_EXCL));
 633
 634         /*
 635          * Clear the transaction pointer in the inode.
 636          */
 637         ip->i_transp = NULL;
 638
 639         /*
 640          * If the inode needed a separate buffer with which to log
 641          * its extents, then free it now.
 642          */
 643         if (iip->ili_extents_buf != NULL) {
 644                 ASSERT(ip->i_d.di_format == XFS_DINODE_FMT_EXTENTS);
 645                 ASSERT(ip->i_d.di_nextents > 0);
 646                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_DEXT);
 647                 ASSERT(ip->i_df.if_bytes > 0);
 648                 kmem_free(iip->ili_extents_buf);
 649                 iip->ili_extents_buf = NULL;
 650         }
 651         if (iip->ili_aextents_buf != NULL) {
 652                 ASSERT(ip->i_d.di_aformat == XFS_DINODE_FMT_EXTENTS);
 653                 ASSERT(ip->i_d.di_anextents > 0);
 654                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_AEXT);
 655                 ASSERT(ip->i_afp->if_bytes > 0);
 656                 kmem_free(iip->ili_aextents_buf);
 657                 iip->ili_aextents_buf = NULL;
 658         }
 659
 660         lock_flags = iip->ili_lock_flags;
 661         iip->ili_lock_flags = 0;
 662         if (lock_flags) {
 663                 xfs_iunlock(iip->ili_inode, lock_flags);
 664                 IRELE(iip->ili_inode);
 665         }
 666 }
 667
 668 /*
 669  * This is called to find out where the oldest active copy of the
 670  * inode log item in the on disk log resides now that the last log
 671  * write of it completed at the given lsn.  Since we always re-log
 672  * all dirty data in an inode, the latest copy in the on disk log
 673  * is the only one that matters.  Therefore, simply return the
 674  * given lsn.
 675  */
 676 STATIC xfs_lsn_t
 677 xfs_inode_item_committed(
 678         struct xfs_log_item     *lip,
 679         xfs_lsn_t               lsn)
 680 {
 681         return lsn;
 682 }
 683
 684 /*
 685  * This gets called by xfs_trans_push_ail(), when IOP_TRYLOCK
 686  * failed to get the inode flush lock but did get the inode locked SHARED.
 687  * Here we're trying to see if the inode buffer is incore, and if so whether it's
 688  * marked delayed write. If that's the case, we'll promote it and that will
 689  * allow the caller to write the buffer by triggering the xfsbufd to run.
 690  */
 691 STATIC void
 692 xfs_inode_item_pushbuf(
 693         struct xfs_log_item     *lip)
 694 {
 695         struct xfs_inode_log_item *iip = INODE_ITEM(lip);
 696         struct xfs_inode        *ip = iip->ili_inode;
 697         struct xfs_buf          *bp;
 698
 699         ASSERT(xfs_isilocked(ip, XFS_ILOCK_SHARED));
 700
 701         /*
 702          * If a flush is not in progress anymore, chances are that the
 703          * inode was taken off the AIL. So, just get out.
 704          */
 705         if (completion_done(&ip->i_flush) ||
 706             !(lip->li_flags & XFS_LI_IN_AIL)) {
 707                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
 708                 return;
 709         }
 710
 711         bp = xfs_incore(ip->i_mount->m_ddev_targp, iip->ili_format.ilf_blkno,
 712                         iip->ili_format.ilf_len, XBF_TRYLOCK);
 713
 714         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 715         if (!bp)
 716                 return;
 717         if (XFS_BUF_ISDELAYWRITE(bp))
 718                 xfs_buf_delwri_promote(bp);
 719         xfs_buf_relse(bp);
 720 }
 721
 722 /*
 723  * This is called to asynchronously write the inode associated with this
 724  * inode log item out to disk. The inode will already have been locked by
 725  * a successful call to xfs_inode_item_trylock().
 726  */
 727 STATIC void
 728 xfs_inode_item_push(
 729         struct xfs_log_item     *lip)
 730 {
 731         struct xfs_inode_log_item *iip = INODE_ITEM(lip);
 732         struct xfs_inode        *ip = iip->ili_inode;
 733
 734         ASSERT(xfs_isilocked(ip, XFS_ILOCK_SHARED));
 735         ASSERT(!completion_done(&ip->i_flush));
 736
 737         /*
 738          * Since we were able to lock the inode's flush lock and
 739          * we found it on the AIL, the inode must be dirty.  This
 740          * is because the inode is removed from the AIL while still
 741          * holding the flush lock in xfs_iflush_done().  Thus, if
 742          * we found it in the AIL and were able to obtain the flush
 743          * lock without sleeping, then there must not have been
 744          * anyone in the process of flushing the inode.
 745          */
 746         ASSERT(XFS_FORCED_SHUTDOWN(ip->i_mount) ||
 747                iip->ili_format.ilf_fields != 0);
 748
 749         /*
 750          * Push the inode to it's backing buffer. This will not remove the
 751          * inode from the AIL - a further push will be required to trigger a
 752          * buffer push. However, this allows all the dirty inodes to be pushed
 753          * to the buffer before it is pushed to disk. THe buffer IO completion
 754          * will pull th einode from the AIL, mark it clean and unlock the flush
 755          * lock.
 756          */
 757         (void) xfs_iflush(ip, 0);
 758         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 759 }
 760
 761 /*
 762  * XXX rcc - this one really has to do something.  Probably needs
 763  * to stamp in a new field in the incore inode.
 764  */
 765 STATIC void
 766 xfs_inode_item_committing(
 767         struct xfs_log_item     *lip,
 768         xfs_lsn_t               lsn)
 769 {
 770         INODE_ITEM(lip)->ili_last_lsn = lsn;
 771 }
 772
 773 /*
 774  * This is the ops vector shared by all buf log items.
 775  */
 776 static struct xfs_item_ops xfs_inode_item_ops = {
 777         .iop_size       = xfs_inode_item_size,
 778         .iop_format     = xfs_inode_item_format,
 779         .iop_pin        = xfs_inode_item_pin,
 780         .iop_unpin      = xfs_inode_item_unpin,
 781         .iop_trylock    = xfs_inode_item_trylock,
 782         .iop_unlock     = xfs_inode_item_unlock,
 783         .iop_committed  = xfs_inode_item_committed,
 784         .iop_push       = xfs_inode_item_push,
 785         .iop_pushbuf    = xfs_inode_item_pushbuf,
 786         .iop_committing = xfs_inode_item_committing
 787 };
 788
 789
 790 /*
 791  * Initialize the inode log item for a newly allocated (in-core) inode.
 792  */
 793 void
 794 xfs_inode_item_init(
 795         struct xfs_inode        *ip,
 796         struct xfs_mount        *mp)
 797 {
 798         struct xfs_inode_log_item *iip;
 799
 800         ASSERT(ip->i_itemp == NULL);
 801         iip = ip->i_itemp = kmem_zone_zalloc(xfs_ili_zone, KM_SLEEP);
 802
 803         iip->ili_inode = ip;
 804         xfs_log_item_init(mp, &iip->ili_item, XFS_LI_INODE,
 805                                                 &xfs_inode_item_ops);
 806         iip->ili_format.ilf_type = XFS_LI_INODE;
 807         iip->ili_format.ilf_ino = ip->i_ino;
 808         iip->ili_format.ilf_blkno = ip->i_imap.im_blkno;
 809         iip->ili_format.ilf_len = ip->i_imap.im_len;
 810         iip->ili_format.ilf_boffset = ip->i_imap.im_boffset;
 811 }
 812
 813 /*
 814  * Free the inode log item and any memory hanging off of it.
 815  */
 816 void
 817 xfs_inode_item_destroy(
 818         xfs_inode_t     *ip)
 819 {
 820 #ifdef XFS_TRANS_DEBUG
 821         if (ip->i_itemp->ili_root_size != 0) {
 822                 kmem_free(ip->i_itemp->ili_orig_root);
 823         }
 824 #endif
 825         kmem_zone_free(xfs_ili_zone, ip->i_itemp);
 826 }
 827
 828
 829 /*
 830  * This is the inode flushing I/O completion routine.  It is called
 831  * from interrupt level when the buffer containing the inode is
 832  * flushed to disk.  It is responsible for removing the inode item
 833  * from the AIL if it has not been re-logged, and unlocking the inode's
 834  * flush lock.
 835  */
 836 void
 837 xfs_iflush_done(
 838         struct xfs_buf          *bp,
 839         struct xfs_log_item     *lip)
 840 {
 841         struct xfs_inode_log_item *iip = INODE_ITEM(lip);
 842         xfs_inode_t             *ip = iip->ili_inode;
 843         struct xfs_ail          *ailp = lip->li_ailp;
 844
 845         /*
 846          * We only want to pull the item from the AIL if it is
 847          * actually there and its location in the log has not
 848          * changed since we started the flush.  Thus, we only bother
 849          * if the ili_logged flag is set and the inode's lsn has not
 850          * changed.  First we check the lsn outside
 851          * the lock since it's cheaper, and then we recheck while
 852          * holding the lock before removing the inode from the AIL.
 853          */
 854         if (iip->ili_logged && lip->li_lsn == iip->ili_flush_lsn) {
 855                 spin_lock(&ailp->xa_lock);
 856                 if (lip->li_lsn == iip->ili_flush_lsn) {
 857                         /* xfs_trans_ail_delete() drops the AIL lock. */
 858                         xfs_trans_ail_delete(ailp, lip);
 859                 } else {
 860                         spin_unlock(&ailp->xa_lock);
 861                 }
 862         }
 863
 864         iip->ili_logged = 0;
 865
 866         /*
 867          * Clear the ili_last_fields bits now that we know that the
 868          * data corresponding to them is safely on disk.
 869          */
 870         iip->ili_last_fields = 0;
 871
 872         /*
 873          * Release the inode's flush lock since we're done with it.
 874          */
 875         xfs_ifunlock(ip);
 876 }
 877
 878 /*
 879  * This is the inode flushing abort routine.  It is called
 880  * from xfs_iflush when the filesystem is shutting down to clean
 881  * up the inode state.
 882  * It is responsible for removing the inode item
 883  * from the AIL if it has not been re-logged, and unlocking the inode's
 884  * flush lock.
 885  */
 886 void
 887 xfs_iflush_abort(
 888         xfs_inode_t             *ip)
 889 {
 890         xfs_inode_log_item_t    *iip = ip->i_itemp;
 891
 892         iip = ip->i_itemp;
 893         if (iip) {
 894                 struct xfs_ail  *ailp = iip->ili_item.li_ailp;
 895                 if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
 896                         spin_lock(&ailp->xa_lock);
 897                         if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
 898                                 /* xfs_trans_ail_delete() drops the AIL lock. */
 899                                 xfs_trans_ail_delete(ailp, (xfs_log_item_t *)iip);
 900                         } else
 901                                 spin_unlock(&ailp->xa_lock);
 902                 }
 903                 iip->ili_logged = 0;
 904                 /*
 905                  * Clear the ili_last_fields bits now that we know that the
 906                  * data corresponding to them is safely on disk.
 907                  */
 908                 iip->ili_last_fields = 0;
 909                 /*
 910                  * Clear the inode logging fields so no more flushes are
 911                  * attempted.
 912                  */
 913                 iip->ili_format.ilf_fields = 0;
 914         }
 915         /*
 916          * Release the inode's flush lock since we're done with it.
 917          */
 918         xfs_ifunlock(ip);
 919 }
 920
 921 void
 922 xfs_istale_done(
 923         struct xfs_buf          *bp,
 924         struct xfs_log_item     *lip)
 925 {
 926         xfs_iflush_abort(INODE_ITEM(lip)->ili_inode);
 927 }
 928
 929 /*
 930  * convert an xfs_inode_log_format struct from either 32 or 64 bit versions
 931  * (which can have different field alignments) to the native version
 932  */
 933 int
 934 xfs_inode_item_format_convert(
 935         xfs_log_iovec_t         *buf,
 936         xfs_inode_log_format_t  *in_f)
 937 {
 938         if (buf->i_len == sizeof(xfs_inode_log_format_32_t)) {
 939                 xfs_inode_log_format_32_t *in_f32 = buf->i_addr;
 940
 941                 in_f->ilf_type = in_f32->ilf_type;
 942                 in_f->ilf_size = in_f32->ilf_size;
 943                 in_f->ilf_fields = in_f32->ilf_fields;
 944                 in_f->ilf_asize = in_f32->ilf_asize;
 945                 in_f->ilf_dsize = in_f32->ilf_dsize;
 946                 in_f->ilf_ino = in_f32->ilf_ino;
 947                 /* copy biggest field of ilf_u */
 948                 memcpy(in_f->ilf_u.ilfu_uuid.__u_bits,
 949                        in_f32->ilf_u.ilfu_uuid.__u_bits,
 950                        sizeof(uuid_t));
 951                 in_f->ilf_blkno = in_f32->ilf_blkno;
 952                 in_f->ilf_len = in_f32->ilf_len;
 953                 in_f->ilf_boffset = in_f32->ilf_boffset;
 954                 return 0;
 955         } else if (buf->i_len == sizeof(xfs_inode_log_format_64_t)){
 956                 xfs_inode_log_format_64_t *in_f64 = buf->i_addr;
 957
 958                 in_f->ilf_type = in_f64->ilf_type;
 959                 in_f->ilf_size = in_f64->ilf_size;
 960                 in_f->ilf_fields = in_f64->ilf_fields;
 961                 in_f->ilf_asize = in_f64->ilf_asize;
 962                 in_f->ilf_dsize = in_f64->ilf_dsize;
 963                 in_f->ilf_ino = in_f64->ilf_ino;
 964                 /* copy biggest field of ilf_u */
 965                 memcpy(in_f->ilf_u.ilfu_uuid.__u_bits,
 966                        in_f64->ilf_u.ilfu_uuid.__u_bits,
 967                        sizeof(uuid_t));
 968                 in_f->ilf_blkno = in_f64->ilf_blkno;
 969                 in_f->ilf_len = in_f64->ilf_len;
 970                 in_f->ilf_boffset = in_f64->ilf_boffset;
 971                 return 0;
 972         }
 973         return EFSCORRUPTED;
 974 }