git.oblomov.eu Git - linux-2.6/blob - 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_buf_item.h"
  26 #include "xfs_sb.h"
  27 #include "xfs_ag.h"
  28 #include "xfs_dir.h"
  29 #include "xfs_dir2.h"
  30 #include "xfs_dmapi.h"
  31 #include "xfs_mount.h"
  32 #include "xfs_trans_priv.h"
  33 #include "xfs_bmap_btree.h"
  34 #include "xfs_alloc_btree.h"
  35 #include "xfs_ialloc_btree.h"
  36 #include "xfs_dir_sf.h"
  37 #include "xfs_dir2_sf.h"
  38 #include "xfs_attr_sf.h"
  39 #include "xfs_dinode.h"
  40 #include "xfs_inode.h"
  41 #include "xfs_inode_item.h"
  42 #include "xfs_btree.h"
  43 #include "xfs_ialloc.h"
  44 #include "xfs_rw.h"
  45
  46
  47 kmem_zone_t     *xfs_ili_zone;          /* inode log item zone */
  48
  49 /*
  50  * This returns the number of iovecs needed to log the given inode item.
  51  *
  52  * We need one iovec for the inode log format structure, one for the
  53  * inode core, and possibly one for the inode data/extents/b-tree root
  54  * and one for the inode attribute data/extents/b-tree root.
  55  */
  56 STATIC uint
  57 xfs_inode_item_size(
  58         xfs_inode_log_item_t    *iip)
  59 {
  60         uint            nvecs;
  61         xfs_inode_t     *ip;
  62
  63         ip = iip->ili_inode;
  64         nvecs = 2;
  65
  66         /*
  67          * Only log the data/extents/b-tree root if there is something
  68          * left to log.
  69          */
  70         iip->ili_format.ilf_fields |= XFS_ILOG_CORE;
  71
  72         switch (ip->i_d.di_format) {
  73         case XFS_DINODE_FMT_EXTENTS:
  74                 iip->ili_format.ilf_fields &=
  75                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
  76                           XFS_ILOG_DEV | XFS_ILOG_UUID);
  77                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DEXT) &&
  78                     (ip->i_d.di_nextents > 0) &&
  79                     (ip->i_df.if_bytes > 0)) {
  80                         ASSERT(ip->i_df.if_u1.if_extents != NULL);
  81                         nvecs++;
  82                 } else {
  83                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DEXT;
  84                 }
  85                 break;
  86
  87         case XFS_DINODE_FMT_BTREE:
  88                 ASSERT(ip->i_df.if_ext_max ==
  89                        XFS_IFORK_DSIZE(ip) / (uint)sizeof(xfs_bmbt_rec_t));
  90                 iip->ili_format.ilf_fields &=
  91                         ~(XFS_ILOG_DDATA | XFS_ILOG_DEXT |
  92                           XFS_ILOG_DEV | XFS_ILOG_UUID);
  93                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DBROOT) &&
  94                     (ip->i_df.if_broot_bytes > 0)) {
  95                         ASSERT(ip->i_df.if_broot != NULL);
  96                         nvecs++;
  97                 } else {
  98                         ASSERT(!(iip->ili_format.ilf_fields &
  99                                  XFS_ILOG_DBROOT));
 100 #ifdef XFS_TRANS_DEBUG
 101                         if (iip->ili_root_size > 0) {
 102                                 ASSERT(iip->ili_root_size ==
 103                                        ip->i_df.if_broot_bytes);
 104                                 ASSERT(memcmp(iip->ili_orig_root,
 105                                             ip->i_df.if_broot,
 106                                             iip->ili_root_size) == 0);
 107                         } else {
 108                                 ASSERT(ip->i_df.if_broot_bytes == 0);
 109                         }
 110 #endif
 111                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DBROOT;
 112                 }
 113                 break;
 114
 115         case XFS_DINODE_FMT_LOCAL:
 116                 iip->ili_format.ilf_fields &=
 117                         ~(XFS_ILOG_DEXT | XFS_ILOG_DBROOT |
 118                           XFS_ILOG_DEV | XFS_ILOG_UUID);
 119                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DDATA) &&
 120                     (ip->i_df.if_bytes > 0)) {
 121                         ASSERT(ip->i_df.if_u1.if_data != NULL);
 122                         ASSERT(ip->i_d.di_size > 0);
 123                         nvecs++;
 124                 } else {
 125                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DDATA;
 126                 }
 127                 break;
 128
 129         case XFS_DINODE_FMT_DEV:
 130                 iip->ili_format.ilf_fields &=
 131                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 132                           XFS_ILOG_DEXT | XFS_ILOG_UUID);
 133                 break;
 134
 135         case XFS_DINODE_FMT_UUID:
 136                 iip->ili_format.ilf_fields &=
 137                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 138                           XFS_ILOG_DEXT | XFS_ILOG_DEV);
 139                 break;
 140
 141         default:
 142                 ASSERT(0);
 143                 break;
 144         }
 145
 146         /*
 147          * If there are no attributes associated with this file,
 148          * then there cannot be anything more to log.
 149          * Clear all attribute-related log flags.
 150          */
 151         if (!XFS_IFORK_Q(ip)) {
 152                 iip->ili_format.ilf_fields &=
 153                         ~(XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT);
 154                 return nvecs;
 155         }
 156
 157         /*
 158          * Log any necessary attribute data.
 159          */
 160         switch (ip->i_d.di_aformat) {
 161         case XFS_DINODE_FMT_EXTENTS:
 162                 iip->ili_format.ilf_fields &=
 163                         ~(XFS_ILOG_ADATA | XFS_ILOG_ABROOT);
 164                 if ((iip->ili_format.ilf_fields & XFS_ILOG_AEXT) &&
 165                     (ip->i_d.di_anextents > 0) &&
 166                     (ip->i_afp->if_bytes > 0)) {
 167                         ASSERT(ip->i_afp->if_u1.if_extents != NULL);
 168                         nvecs++;
 169                 } else {
 170                         iip->ili_format.ilf_fields &= ~XFS_ILOG_AEXT;
 171                 }
 172                 break;
 173
 174         case XFS_DINODE_FMT_BTREE:
 175                 iip->ili_format.ilf_fields &=
 176                         ~(XFS_ILOG_ADATA | XFS_ILOG_AEXT);
 177                 if ((iip->ili_format.ilf_fields & XFS_ILOG_ABROOT) &&
 178                     (ip->i_afp->if_broot_bytes > 0)) {
 179                         ASSERT(ip->i_afp->if_broot != NULL);
 180                         nvecs++;
 181                 } else {
 182                         iip->ili_format.ilf_fields &= ~XFS_ILOG_ABROOT;
 183                 }
 184                 break;
 185
 186         case XFS_DINODE_FMT_LOCAL:
 187                 iip->ili_format.ilf_fields &=
 188                         ~(XFS_ILOG_AEXT | XFS_ILOG_ABROOT);
 189                 if ((iip->ili_format.ilf_fields & XFS_ILOG_ADATA) &&
 190                     (ip->i_afp->if_bytes > 0)) {
 191                         ASSERT(ip->i_afp->if_u1.if_data != NULL);
 192                         nvecs++;
 193                 } else {
 194                         iip->ili_format.ilf_fields &= ~XFS_ILOG_ADATA;
 195                 }
 196                 break;
 197
 198         default:
 199                 ASSERT(0);
 200                 break;
 201         }
 202
 203         return nvecs;
 204 }
 205
 206 /*
 207  * This is called to fill in the vector of log iovecs for the
 208  * given inode log item.  It fills the first item with an inode
 209  * log format structure, the second with the on-disk inode structure,
 210  * and a possible third and/or fourth with the inode data/extents/b-tree
 211  * root and inode attributes data/extents/b-tree root.
 212  */
 213 STATIC void
 214 xfs_inode_item_format(
 215         xfs_inode_log_item_t    *iip,
 216         xfs_log_iovec_t         *log_vector)
 217 {
 218         uint                    nvecs;
 219         xfs_log_iovec_t         *vecp;
 220         xfs_inode_t             *ip;
 221         size_t                  data_bytes;
 222         xfs_bmbt_rec_t          *ext_buffer;
 223         int                     nrecs;
 224         xfs_mount_t             *mp;
 225
 226         ip = iip->ili_inode;
 227         vecp = log_vector;
 228
 229         vecp->i_addr = (xfs_caddr_t)&iip->ili_format;
 230         vecp->i_len  = sizeof(xfs_inode_log_format_t);
 231         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IFORMAT);
 232         vecp++;
 233         nvecs        = 1;
 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          * We don't have to worry about re-ordering here because
 268          * the update_size field is protected by the inode lock
 269          * and we have that held in exclusive mode.
 270          */
 271         if (ip->i_update_size)
 272                 ip->i_update_size = 0;
 273
 274         vecp->i_addr = (xfs_caddr_t)&ip->i_d;
 275         vecp->i_len  = sizeof(xfs_dinode_core_t);
 276         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_ICORE);
 277         vecp++;
 278         nvecs++;
 279         iip->ili_format.ilf_fields |= XFS_ILOG_CORE;
 280
 281         /*
 282          * If this is really an old format inode, then we need to
 283          * log it as such.  This means that we have to copy the link
 284          * count from the new field to the old.  We don't have to worry
 285          * about the new fields, because nothing trusts them as long as
 286          * the old inode version number is there.  If the superblock already
 287          * has a new version number, then we don't bother converting back.
 288          */
 289         mp = ip->i_mount;
 290         ASSERT(ip->i_d.di_version == XFS_DINODE_VERSION_1 ||
 291                XFS_SB_VERSION_HASNLINK(&mp->m_sb));
 292         if (ip->i_d.di_version == XFS_DINODE_VERSION_1) {
 293                 if (!XFS_SB_VERSION_HASNLINK(&mp->m_sb)) {
 294                         /*
 295                          * Convert it back.
 296                          */
 297                         ASSERT(ip->i_d.di_nlink <= XFS_MAXLINK_1);
 298                         ip->i_d.di_onlink = ip->i_d.di_nlink;
 299                 } else {
 300                         /*
 301                          * The superblock version has already been bumped,
 302                          * so just make the conversion to the new inode
 303                          * format permanent.
 304                          */
 305                         ip->i_d.di_version = XFS_DINODE_VERSION_2;
 306                         ip->i_d.di_onlink = 0;
 307                         memset(&(ip->i_d.di_pad[0]), 0, sizeof(ip->i_d.di_pad));
 308                 }
 309         }
 310
 311         switch (ip->i_d.di_format) {
 312         case XFS_DINODE_FMT_EXTENTS:
 313                 ASSERT(!(iip->ili_format.ilf_fields &
 314                          (XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 315                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 316                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEXT) {
 317                         ASSERT(ip->i_df.if_bytes > 0);
 318                         ASSERT(ip->i_df.if_u1.if_extents != NULL);
 319                         ASSERT(ip->i_d.di_nextents > 0);
 320                         ASSERT(iip->ili_extents_buf == NULL);
 321                         nrecs = ip->i_df.if_bytes /
 322                                 (uint)sizeof(xfs_bmbt_rec_t);
 323                         ASSERT(nrecs > 0);
 324 #ifdef XFS_NATIVE_HOST
 325                         if (nrecs == ip->i_d.di_nextents) {
 326                                 /*
 327                                  * There are no delayed allocation
 328                                  * extents, so just point to the
 329                                  * real extents array.
 330                                  */
 331                                 vecp->i_addr =
 332                                         (char *)(ip->i_df.if_u1.if_extents);
 333                                 vecp->i_len = ip->i_df.if_bytes;
 334                                 XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IEXT);
 335                         } else
 336 #endif
 337                         {
 338                                 /*
 339                                  * There are delayed allocation extents
 340                                  * in the inode, or we need to convert
 341                                  * the extents to on disk format.
 342                                  * Use xfs_iextents_copy()
 343                                  * to copy only the real extents into
 344                                  * a separate buffer.  We'll free the
 345                                  * buffer in the unlock routine.
 346                                  */
 347                                 ext_buffer = kmem_alloc(ip->i_df.if_bytes,
 348                                         KM_SLEEP);
 349                                 iip->ili_extents_buf = ext_buffer;
 350                                 vecp->i_addr = (xfs_caddr_t)ext_buffer;
 351                                 vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
 352                                                 XFS_DATA_FORK);
 353                                 XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IEXT);
 354                         }
 355                         ASSERT(vecp->i_len <= ip->i_df.if_bytes);
 356                         iip->ili_format.ilf_dsize = vecp->i_len;
 357                         vecp++;
 358                         nvecs++;
 359                 }
 360                 break;
 361
 362         case XFS_DINODE_FMT_BTREE:
 363                 ASSERT(!(iip->ili_format.ilf_fields &
 364                          (XFS_ILOG_DDATA | XFS_ILOG_DEXT |
 365                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 366                 if (iip->ili_format.ilf_fields & XFS_ILOG_DBROOT) {
 367                         ASSERT(ip->i_df.if_broot_bytes > 0);
 368                         ASSERT(ip->i_df.if_broot != NULL);
 369                         vecp->i_addr = (xfs_caddr_t)ip->i_df.if_broot;
 370                         vecp->i_len = ip->i_df.if_broot_bytes;
 371                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IBROOT);
 372                         vecp++;
 373                         nvecs++;
 374                         iip->ili_format.ilf_dsize = ip->i_df.if_broot_bytes;
 375                 }
 376                 break;
 377
 378         case XFS_DINODE_FMT_LOCAL:
 379                 ASSERT(!(iip->ili_format.ilf_fields &
 380                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 381                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 382                 if (iip->ili_format.ilf_fields & XFS_ILOG_DDATA) {
 383                         ASSERT(ip->i_df.if_bytes > 0);
 384                         ASSERT(ip->i_df.if_u1.if_data != NULL);
 385                         ASSERT(ip->i_d.di_size > 0);
 386
 387                         vecp->i_addr = (xfs_caddr_t)ip->i_df.if_u1.if_data;
 388                         /*
 389                          * Round i_bytes up to a word boundary.
 390                          * The underlying memory is guaranteed to
 391                          * to be there by xfs_idata_realloc().
 392                          */
 393                         data_bytes = roundup(ip->i_df.if_bytes, 4);
 394                         ASSERT((ip->i_df.if_real_bytes == 0) ||
 395                                (ip->i_df.if_real_bytes == data_bytes));
 396                         vecp->i_len = (int)data_bytes;
 397                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_ILOCAL);
 398                         vecp++;
 399                         nvecs++;
 400                         iip->ili_format.ilf_dsize = (unsigned)data_bytes;
 401                 }
 402                 break;
 403
 404         case XFS_DINODE_FMT_DEV:
 405                 ASSERT(!(iip->ili_format.ilf_fields &
 406                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 407                           XFS_ILOG_DDATA | XFS_ILOG_UUID)));
 408                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEV) {
 409                         iip->ili_format.ilf_u.ilfu_rdev =
 410                                 ip->i_df.if_u2.if_rdev;
 411                 }
 412                 break;
 413
 414         case XFS_DINODE_FMT_UUID:
 415                 ASSERT(!(iip->ili_format.ilf_fields &
 416                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 417                           XFS_ILOG_DDATA | XFS_ILOG_DEV)));
 418                 if (iip->ili_format.ilf_fields & XFS_ILOG_UUID) {
 419                         iip->ili_format.ilf_u.ilfu_uuid =
 420                                 ip->i_df.if_u2.if_uuid;
 421                 }
 422                 break;
 423
 424         default:
 425                 ASSERT(0);
 426                 break;
 427         }
 428
 429         /*
 430          * If there are no attributes associated with the file,
 431          * then we're done.
 432          * Assert that no attribute-related log flags are set.
 433          */
 434         if (!XFS_IFORK_Q(ip)) {
 435                 ASSERT(nvecs == iip->ili_item.li_desc->lid_size);
 436                 iip->ili_format.ilf_size = nvecs;
 437                 ASSERT(!(iip->ili_format.ilf_fields &
 438                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
 439                 return;
 440         }
 441
 442         switch (ip->i_d.di_aformat) {
 443         case XFS_DINODE_FMT_EXTENTS:
 444                 ASSERT(!(iip->ili_format.ilf_fields &
 445                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT)));
 446                 if (iip->ili_format.ilf_fields & XFS_ILOG_AEXT) {
 447                         ASSERT(ip->i_afp->if_bytes > 0);
 448                         ASSERT(ip->i_afp->if_u1.if_extents != NULL);
 449                         ASSERT(ip->i_d.di_anextents > 0);
 450 #ifdef DEBUG
 451                         nrecs = ip->i_afp->if_bytes /
 452                                 (uint)sizeof(xfs_bmbt_rec_t);
 453 #endif
 454                         ASSERT(nrecs > 0);
 455                         ASSERT(nrecs == ip->i_d.di_anextents);
 456 #ifdef XFS_NATIVE_HOST
 457                         /*
 458                          * There are not delayed allocation extents
 459                          * for attributes, so just point at the array.
 460                          */
 461                         vecp->i_addr = (char *)(ip->i_afp->if_u1.if_extents);
 462                         vecp->i_len = ip->i_afp->if_bytes;
 463 #else
 464                         ASSERT(iip->ili_aextents_buf == NULL);
 465                         /*
 466                          * Need to endian flip before logging
 467                          */
 468                         ext_buffer = kmem_alloc(ip->i_afp->if_bytes,
 469                                 KM_SLEEP);
 470                         iip->ili_aextents_buf = ext_buffer;
 471                         vecp->i_addr = (xfs_caddr_t)ext_buffer;
 472                         vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
 473                                         XFS_ATTR_FORK);
 474 #endif
 475                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IATTR_EXT);
 476                         iip->ili_format.ilf_asize = vecp->i_len;
 477                         vecp++;
 478                         nvecs++;
 479                 }
 480                 break;
 481
 482         case XFS_DINODE_FMT_BTREE:
 483                 ASSERT(!(iip->ili_format.ilf_fields &
 484                          (XFS_ILOG_ADATA | XFS_ILOG_AEXT)));
 485                 if (iip->ili_format.ilf_fields & XFS_ILOG_ABROOT) {
 486                         ASSERT(ip->i_afp->if_broot_bytes > 0);
 487                         ASSERT(ip->i_afp->if_broot != NULL);
 488                         vecp->i_addr = (xfs_caddr_t)ip->i_afp->if_broot;
 489                         vecp->i_len = ip->i_afp->if_broot_bytes;
 490                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IATTR_BROOT);
 491                         vecp++;
 492                         nvecs++;
 493                         iip->ili_format.ilf_asize = ip->i_afp->if_broot_bytes;
 494                 }
 495                 break;
 496
 497         case XFS_DINODE_FMT_LOCAL:
 498                 ASSERT(!(iip->ili_format.ilf_fields &
 499                          (XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
 500                 if (iip->ili_format.ilf_fields & XFS_ILOG_ADATA) {
 501                         ASSERT(ip->i_afp->if_bytes > 0);
 502                         ASSERT(ip->i_afp->if_u1.if_data != NULL);
 503
 504                         vecp->i_addr = (xfs_caddr_t)ip->i_afp->if_u1.if_data;
 505                         /*
 506                          * Round i_bytes up to a word boundary.
 507                          * The underlying memory is guaranteed to
 508                          * to be there by xfs_idata_realloc().
 509                          */
 510                         data_bytes = roundup(ip->i_afp->if_bytes, 4);
 511                         ASSERT((ip->i_afp->if_real_bytes == 0) ||
 512                                (ip->i_afp->if_real_bytes == data_bytes));
 513                         vecp->i_len = (int)data_bytes;
 514                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IATTR_LOCAL);
 515                         vecp++;
 516                         nvecs++;
 517                         iip->ili_format.ilf_asize = (unsigned)data_bytes;
 518                 }
 519                 break;
 520
 521         default:
 522                 ASSERT(0);
 523                 break;
 524         }
 525
 526         ASSERT(nvecs == iip->ili_item.li_desc->lid_size);
 527         iip->ili_format.ilf_size = nvecs;
 528 }
 529
 530
 531 /*
 532  * This is called to pin the inode associated with the inode log
 533  * item in memory so it cannot be written out.  Do this by calling
 534  * xfs_ipin() to bump the pin count in the inode while holding the
 535  * inode pin lock.
 536  */
 537 STATIC void
 538 xfs_inode_item_pin(
 539         xfs_inode_log_item_t    *iip)
 540 {
 541         ASSERT(ismrlocked(&(iip->ili_inode->i_lock), MR_UPDATE));
 542         xfs_ipin(iip->ili_inode);
 543 }
 544
 545
 546 /*
 547  * This is called to unpin the inode associated with the inode log
 548  * item which was previously pinned with a call to xfs_inode_item_pin().
 549  * Just call xfs_iunpin() on the inode to do this.
 550  */
 551 /* ARGSUSED */
 552 STATIC void
 553 xfs_inode_item_unpin(
 554         xfs_inode_log_item_t    *iip,
 555         int                     stale)
 556 {
 557         xfs_iunpin(iip->ili_inode);
 558 }
 559
 560 /* ARGSUSED */
 561 STATIC void
 562 xfs_inode_item_unpin_remove(
 563         xfs_inode_log_item_t    *iip,
 564         xfs_trans_t             *tp)
 565 {
 566         xfs_iunpin(iip->ili_inode);
 567 }
 568
 569 /*
 570  * This is called to attempt to lock the inode associated with this
 571  * inode log item, in preparation for the push routine which does the actual
 572  * iflush.  Don't sleep on the inode lock or the flush lock.
 573  *
 574  * If the flush lock is already held, indicating that the inode has
 575  * been or is in the process of being flushed, then (ideally) we'd like to
 576  * see if the inode's buffer is still incore, and if so give it a nudge.
 577  * We delay doing so until the pushbuf routine, though, to avoid holding
 578  * the AIL lock across a call to the blackhole which is the buffercache.
 579  * Also we don't want to sleep in any device strategy routines, which can happen
 580  * if we do the subsequent bawrite in here.
 581  */
 582 STATIC uint
 583 xfs_inode_item_trylock(
 584         xfs_inode_log_item_t    *iip)
 585 {
 586         register xfs_inode_t    *ip;
 587
 588         ip = iip->ili_inode;
 589
 590         if (xfs_ipincount(ip) > 0) {
 591                 return XFS_ITEM_PINNED;
 592         }
 593
 594         if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
 595                 return XFS_ITEM_LOCKED;
 596         }
 597
 598         if (!xfs_iflock_nowait(ip)) {
 599                 /*
 600                  * If someone else isn't already trying to push the inode
 601                  * buffer, we get to do it.
 602                  */
 603                 if (iip->ili_pushbuf_flag == 0) {
 604                         iip->ili_pushbuf_flag = 1;
 605 #ifdef DEBUG
 606                         iip->ili_push_owner = get_thread_id();
 607 #endif
 608                         /*
 609                          * Inode is left locked in shared mode.
 610                          * Pushbuf routine gets to unlock it.
 611                          */
 612                         return XFS_ITEM_PUSHBUF;
 613                 } else {
 614                         /*
 615                          * We hold the AIL_LOCK, so we must specify the
 616                          * NONOTIFY flag so that we won't double trip.
 617                          */
 618                         xfs_iunlock(ip, XFS_ILOCK_SHARED|XFS_IUNLOCK_NONOTIFY);
 619                         return XFS_ITEM_FLUSHING;
 620                 }
 621                 /* NOTREACHED */
 622         }
 623
 624         /* Stale items should force out the iclog */
 625         if (ip->i_flags & XFS_ISTALE) {
 626                 xfs_ifunlock(ip);
 627                 xfs_iunlock(ip, XFS_ILOCK_SHARED|XFS_IUNLOCK_NONOTIFY);
 628                 return XFS_ITEM_PINNED;
 629         }
 630
 631 #ifdef DEBUG
 632         if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
 633                 ASSERT(iip->ili_format.ilf_fields != 0);
 634                 ASSERT(iip->ili_logged == 0);
 635                 ASSERT(iip->ili_item.li_flags & XFS_LI_IN_AIL);
 636         }
 637 #endif
 638         return XFS_ITEM_SUCCESS;
 639 }
 640
 641 /*
 642  * Unlock the inode associated with the inode log item.
 643  * Clear the fields of the inode and inode log item that
 644  * are specific to the current transaction.  If the
 645  * hold flags is set, do not unlock the inode.
 646  */
 647 STATIC void
 648 xfs_inode_item_unlock(
 649         xfs_inode_log_item_t    *iip)
 650 {
 651         uint            hold;
 652         uint            iolocked;
 653         uint            lock_flags;
 654         xfs_inode_t     *ip;
 655
 656         ASSERT(iip != NULL);
 657         ASSERT(iip->ili_inode->i_itemp != NULL);
 658         ASSERT(ismrlocked(&(iip->ili_inode->i_lock), MR_UPDATE));
 659         ASSERT((!(iip->ili_inode->i_itemp->ili_flags &
 660                   XFS_ILI_IOLOCKED_EXCL)) ||
 661                ismrlocked(&(iip->ili_inode->i_iolock), MR_UPDATE));
 662         ASSERT((!(iip->ili_inode->i_itemp->ili_flags &
 663                   XFS_ILI_IOLOCKED_SHARED)) ||
 664                ismrlocked(&(iip->ili_inode->i_iolock), MR_ACCESS));
 665         /*
 666          * Clear the transaction pointer in the inode.
 667          */
 668         ip = iip->ili_inode;
 669         ip->i_transp = NULL;
 670
 671         /*
 672          * If the inode needed a separate buffer with which to log
 673          * its extents, then free it now.
 674          */
 675         if (iip->ili_extents_buf != NULL) {
 676                 ASSERT(ip->i_d.di_format == XFS_DINODE_FMT_EXTENTS);
 677                 ASSERT(ip->i_d.di_nextents > 0);
 678                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_DEXT);
 679                 ASSERT(ip->i_df.if_bytes > 0);
 680                 kmem_free(iip->ili_extents_buf, ip->i_df.if_bytes);
 681                 iip->ili_extents_buf = NULL;
 682         }
 683         if (iip->ili_aextents_buf != NULL) {
 684                 ASSERT(ip->i_d.di_aformat == XFS_DINODE_FMT_EXTENTS);
 685                 ASSERT(ip->i_d.di_anextents > 0);
 686                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_AEXT);
 687                 ASSERT(ip->i_afp->if_bytes > 0);
 688                 kmem_free(iip->ili_aextents_buf, ip->i_afp->if_bytes);
 689                 iip->ili_aextents_buf = NULL;
 690         }
 691
 692         /*
 693          * Figure out if we should unlock the inode or not.
 694          */
 695         hold = iip->ili_flags & XFS_ILI_HOLD;
 696
 697         /*
 698          * Before clearing out the flags, remember whether we
 699          * are holding the inode's IO lock.
 700          */
 701         iolocked = iip->ili_flags & XFS_ILI_IOLOCKED_ANY;
 702
 703         /*
 704          * Clear out the fields of the inode log item particular
 705          * to the current transaction.
 706          */
 707         iip->ili_ilock_recur = 0;
 708         iip->ili_iolock_recur = 0;
 709         iip->ili_flags = 0;
 710
 711         /*
 712          * Unlock the inode if XFS_ILI_HOLD was not set.
 713          */
 714         if (!hold) {
 715                 lock_flags = XFS_ILOCK_EXCL;
 716                 if (iolocked & XFS_ILI_IOLOCKED_EXCL) {
 717                         lock_flags |= XFS_IOLOCK_EXCL;
 718                 } else if (iolocked & XFS_ILI_IOLOCKED_SHARED) {
 719                         lock_flags |= XFS_IOLOCK_SHARED;
 720                 }
 721                 xfs_iput(iip->ili_inode, lock_flags);
 722         }
 723 }
 724
 725 /*
 726  * This is called to find out where the oldest active copy of the
 727  * inode log item in the on disk log resides now that the last log
 728  * write of it completed at the given lsn.  Since we always re-log
 729  * all dirty data in an inode, the latest copy in the on disk log
 730  * is the only one that matters.  Therefore, simply return the
 731  * given lsn.
 732  */
 733 /*ARGSUSED*/
 734 STATIC xfs_lsn_t
 735 xfs_inode_item_committed(
 736         xfs_inode_log_item_t    *iip,
 737         xfs_lsn_t               lsn)
 738 {
 739         return (lsn);
 740 }
 741
 742 /*
 743  * The transaction with the inode locked has aborted.  The inode
 744  * must not be dirty within the transaction (unless we're forcibly
 745  * shutting down).  We simply unlock just as if the transaction
 746  * had been cancelled.
 747  */
 748 STATIC void
 749 xfs_inode_item_abort(
 750         xfs_inode_log_item_t    *iip)
 751 {
 752         xfs_inode_item_unlock(iip);
 753         return;
 754 }
 755
 756
 757 /*
 758  * This gets called by xfs_trans_push_ail(), when IOP_TRYLOCK
 759  * failed to get the inode flush lock but did get the inode locked SHARED.
 760  * Here we're trying to see if the inode buffer is incore, and if so whether it's
 761  * marked delayed write. If that's the case, we'll initiate a bawrite on that
 762  * buffer to expedite the process.
 763  *
 764  * We aren't holding the AIL_LOCK (or the flush lock) when this gets called,
 765  * so it is inherently race-y.
 766  */
 767 STATIC void
 768 xfs_inode_item_pushbuf(
 769         xfs_inode_log_item_t    *iip)
 770 {
 771         xfs_inode_t     *ip;
 772         xfs_mount_t     *mp;
 773         xfs_buf_t       *bp;
 774         uint            dopush;
 775
 776         ip = iip->ili_inode;
 777
 778         ASSERT(ismrlocked(&(ip->i_lock), MR_ACCESS));
 779
 780         /*
 781          * The ili_pushbuf_flag keeps others from
 782          * trying to duplicate our effort.
 783          */
 784         ASSERT(iip->ili_pushbuf_flag != 0);
 785         ASSERT(iip->ili_push_owner == get_thread_id());
 786
 787         /*
 788          * If flushlock isn't locked anymore, chances are that the
 789          * inode flush completed and the inode was taken off the AIL.
 790          * So, just get out.
 791          */
 792         if ((valusema(&(ip->i_flock)) > 0)  ||
 793             ((iip->ili_item.li_flags & XFS_LI_IN_AIL) == 0)) {
 794                 iip->ili_pushbuf_flag = 0;
 795                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
 796                 return;
 797         }
 798
 799         mp = ip->i_mount;
 800         bp = xfs_incore(mp->m_ddev_targp, iip->ili_format.ilf_blkno,
 801                     iip->ili_format.ilf_len, XFS_INCORE_TRYLOCK);
 802
 803         if (bp != NULL) {
 804                 if (XFS_BUF_ISDELAYWRITE(bp)) {
 805                         /*
 806                          * We were racing with iflush because we don't hold
 807                          * the AIL_LOCK or the flush lock. However, at this point,
 808                          * we have the buffer, and we know that it's dirty.
 809                          * So, it's possible that iflush raced with us, and
 810                          * this item is already taken off the AIL.
 811                          * If not, we can flush it async.
 812                          */
 813                         dopush = ((iip->ili_item.li_flags & XFS_LI_IN_AIL) &&
 814                                   (valusema(&(ip->i_flock)) <= 0));
 815                         iip->ili_pushbuf_flag = 0;
 816                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 817                         xfs_buftrace("INODE ITEM PUSH", bp);
 818                         if (XFS_BUF_ISPINNED(bp)) {
 819                                 xfs_log_force(mp, (xfs_lsn_t)0,
 820                                               XFS_LOG_FORCE);
 821                         }
 822                         if (dopush) {
 823                                 xfs_bawrite(mp, bp);
 824                         } else {
 825                                 xfs_buf_relse(bp);
 826                         }
 827                 } else {
 828                         iip->ili_pushbuf_flag = 0;
 829                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 830                         xfs_buf_relse(bp);
 831                 }
 832                 return;
 833         }
 834         /*
 835          * We have to be careful about resetting pushbuf flag too early (above).
 836          * Even though in theory we can do it as soon as we have the buflock,
 837          * we don't want others to be doing work needlessly. They'll come to
 838          * this function thinking that pushing the buffer is their
 839          * responsibility only to find that the buffer is still locked by
 840          * another doing the same thing
 841          */
 842         iip->ili_pushbuf_flag = 0;
 843         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 844         return;
 845 }
 846
 847
 848 /*
 849  * This is called to asynchronously write the inode associated with this
 850  * inode log item out to disk. The inode will already have been locked by
 851  * a successful call to xfs_inode_item_trylock().
 852  */
 853 STATIC void
 854 xfs_inode_item_push(
 855         xfs_inode_log_item_t    *iip)
 856 {
 857         xfs_inode_t     *ip;
 858
 859         ip = iip->ili_inode;
 860
 861         ASSERT(ismrlocked(&(ip->i_lock), MR_ACCESS));
 862         ASSERT(valusema(&(ip->i_flock)) <= 0);
 863         /*
 864          * Since we were able to lock the inode's flush lock and
 865          * we found it on the AIL, the inode must be dirty.  This
 866          * is because the inode is removed from the AIL while still
 867          * holding the flush lock in xfs_iflush_done().  Thus, if
 868          * we found it in the AIL and were able to obtain the flush
 869          * lock without sleeping, then there must not have been
 870          * anyone in the process of flushing the inode.
 871          */
 872         ASSERT(XFS_FORCED_SHUTDOWN(ip->i_mount) ||
 873                iip->ili_format.ilf_fields != 0);
 874
 875         /*
 876          * Write out the inode.  The completion routine ('iflush_done') will
 877          * pull it from the AIL, mark it clean, unlock the flush lock.
 878          */
 879         (void) xfs_iflush(ip, XFS_IFLUSH_ASYNC);
 880         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 881
 882         return;
 883 }
 884
 885 /*
 886  * XXX rcc - this one really has to do something.  Probably needs
 887  * to stamp in a new field in the incore inode.
 888  */
 889 /* ARGSUSED */
 890 STATIC void
 891 xfs_inode_item_committing(
 892         xfs_inode_log_item_t    *iip,
 893         xfs_lsn_t               lsn)
 894 {
 895         iip->ili_last_lsn = lsn;
 896         return;
 897 }
 898
 899 /*
 900  * This is the ops vector shared by all buf log items.
 901  */
 902 STATIC struct xfs_item_ops xfs_inode_item_ops = {
 903         .iop_size       = (uint(*)(xfs_log_item_t*))xfs_inode_item_size,
 904         .iop_format     = (void(*)(xfs_log_item_t*, xfs_log_iovec_t*))
 905                                         xfs_inode_item_format,
 906         .iop_pin        = (void(*)(xfs_log_item_t*))xfs_inode_item_pin,
 907         .iop_unpin      = (void(*)(xfs_log_item_t*, int))xfs_inode_item_unpin,
 908         .iop_unpin_remove = (void(*)(xfs_log_item_t*, xfs_trans_t*))
 909                                         xfs_inode_item_unpin_remove,
 910         .iop_trylock    = (uint(*)(xfs_log_item_t*))xfs_inode_item_trylock,
 911         .iop_unlock     = (void(*)(xfs_log_item_t*))xfs_inode_item_unlock,
 912         .iop_committed  = (xfs_lsn_t(*)(xfs_log_item_t*, xfs_lsn_t))
 913                                         xfs_inode_item_committed,
 914         .iop_push       = (void(*)(xfs_log_item_t*))xfs_inode_item_push,
 915         .iop_abort      = (void(*)(xfs_log_item_t*))xfs_inode_item_abort,
 916         .iop_pushbuf    = (void(*)(xfs_log_item_t*))xfs_inode_item_pushbuf,
 917         .iop_committing = (void(*)(xfs_log_item_t*, xfs_lsn_t))
 918                                         xfs_inode_item_committing
 919 };
 920
 921
 922 /*
 923  * Initialize the inode log item for a newly allocated (in-core) inode.
 924  */
 925 void
 926 xfs_inode_item_init(
 927         xfs_inode_t     *ip,
 928         xfs_mount_t     *mp)
 929 {
 930         xfs_inode_log_item_t    *iip;
 931
 932         ASSERT(ip->i_itemp == NULL);
 933         iip = ip->i_itemp = kmem_zone_zalloc(xfs_ili_zone, KM_SLEEP);
 934
 935         iip->ili_item.li_type = XFS_LI_INODE;
 936         iip->ili_item.li_ops = &xfs_inode_item_ops;
 937         iip->ili_item.li_mountp = mp;
 938         iip->ili_inode = ip;
 939
 940         /*
 941            We have zeroed memory. No need ...
 942            iip->ili_extents_buf = NULL;
 943            iip->ili_pushbuf_flag = 0;
 944          */
 945
 946         iip->ili_format.ilf_type = XFS_LI_INODE;
 947         iip->ili_format.ilf_ino = ip->i_ino;
 948         iip->ili_format.ilf_blkno = ip->i_blkno;
 949         iip->ili_format.ilf_len = ip->i_len;
 950         iip->ili_format.ilf_boffset = ip->i_boffset;
 951 }
 952
 953 /*
 954  * Free the inode log item and any memory hanging off of it.
 955  */
 956 void
 957 xfs_inode_item_destroy(
 958         xfs_inode_t     *ip)
 959 {
 960 #ifdef XFS_TRANS_DEBUG
 961         if (ip->i_itemp->ili_root_size != 0) {
 962                 kmem_free(ip->i_itemp->ili_orig_root,
 963                           ip->i_itemp->ili_root_size);
 964         }
 965 #endif
 966         kmem_zone_free(xfs_ili_zone, ip->i_itemp);
 967 }
 968
 969
 970 /*
 971  * This is the inode flushing I/O completion routine.  It is called
 972  * from interrupt level when the buffer containing the inode is
 973  * flushed to disk.  It is responsible for removing the inode item
 974  * from the AIL if it has not been re-logged, and unlocking the inode's
 975  * flush lock.
 976  */
 977 /*ARGSUSED*/
 978 void
 979 xfs_iflush_done(
 980         xfs_buf_t               *bp,
 981         xfs_inode_log_item_t    *iip)
 982 {
 983         xfs_inode_t     *ip;
 984         SPLDECL(s);
 985
 986         ip = iip->ili_inode;
 987
 988         /*
 989          * We only want to pull the item from the AIL if it is
 990          * actually there and its location in the log has not
 991          * changed since we started the flush.  Thus, we only bother
 992          * if the ili_logged flag is set and the inode's lsn has not
 993          * changed.  First we check the lsn outside
 994          * the lock since it's cheaper, and then we recheck while
 995          * holding the lock before removing the inode from the AIL.
 996          */
 997         if (iip->ili_logged &&
 998             (iip->ili_item.li_lsn == iip->ili_flush_lsn)) {
 999                 AIL_LOCK(ip->i_mount, s);
1000                 if (iip->ili_item.li_lsn == iip->ili_flush_lsn) {
1001                         /*
1002                          * xfs_trans_delete_ail() drops the AIL lock.
1003                          */
1004                         xfs_trans_delete_ail(ip->i_mount,
1005                                              (xfs_log_item_t*)iip, s);
1006                 } else {
1007                         AIL_UNLOCK(ip->i_mount, s);
1008                 }
1009         }
1010
1011         iip->ili_logged = 0;
1012
1013         /*
1014          * Clear the ili_last_fields bits now that we know that the
1015          * data corresponding to them is safely on disk.
1016          */
1017         iip->ili_last_fields = 0;
1018
1019         /*
1020          * Release the inode's flush lock since we're done with it.
1021          */
1022         xfs_ifunlock(ip);
1023
1024         return;
1025 }
1026
1027 /*
1028  * This is the inode flushing abort routine.  It is called
1029  * from xfs_iflush when the filesystem is shutting down to clean
1030  * up the inode state.
1031  * It is responsible for removing the inode item
1032  * from the AIL if it has not been re-logged, and unlocking the inode's
1033  * flush lock.
1034  */
1035 void
1036 xfs_iflush_abort(
1037         xfs_inode_t             *ip)
1038 {
1039         xfs_inode_log_item_t    *iip;
1040         xfs_mount_t             *mp;
1041         SPLDECL(s);
1042
1043         iip = ip->i_itemp;
1044         mp = ip->i_mount;
1045         if (iip) {
1046                 if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
1047                         AIL_LOCK(mp, s);
1048                         if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
1049                                 /*
1050                                  * xfs_trans_delete_ail() drops the AIL lock.
1051                                  */
1052                                 xfs_trans_delete_ail(mp, (xfs_log_item_t *)iip,
1053                                         s);
1054                         } else
1055                                 AIL_UNLOCK(mp, s);
1056                 }
1057                 iip->ili_logged = 0;
1058                 /*
1059                  * Clear the ili_last_fields bits now that we know that the
1060                  * data corresponding to them is safely on disk.
1061                  */
1062                 iip->ili_last_fields = 0;
1063                 /*
1064                  * Clear the inode logging fields so no more flushes are
1065                  * attempted.
1066                  */
1067                 iip->ili_format.ilf_fields = 0;
1068         }
1069         /*
1070          * Release the inode's flush lock since we're done with it.
1071          */
1072         xfs_ifunlock(ip);
1073 }
1074
1075 void
1076 xfs_istale_done(
1077         xfs_buf_t               *bp,
1078         xfs_inode_log_item_t    *iip)
1079 {
1080         xfs_iflush_abort(iip->ili_inode);
1081 }