SAFE public projects git trees. - safe/jmp/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_dir2.h"
  29 #include "xfs_dmapi.h"
  30 #include "xfs_mount.h"
  31 #include "xfs_trans_priv.h"
  32 #include "xfs_bmap_btree.h"
  33 #include "xfs_alloc_btree.h"
  34 #include "xfs_ialloc_btree.h"
  35 #include "xfs_dir2_sf.h"
  36 #include "xfs_attr_sf.h"
  37 #include "xfs_dinode.h"
  38 #include "xfs_inode.h"
  39 #include "xfs_inode_item.h"
  40 #include "xfs_btree.h"
  41 #include "xfs_ialloc.h"
  42 #include "xfs_rw.h"
  43 #include "xfs_error.h"
  44
  45
  46 kmem_zone_t     *xfs_ili_zone;          /* inode log item zone */
  47
  48 /*
  49  * This returns the number of iovecs needed to log the given inode item.
  50  *
  51  * We need one iovec for the inode log format structure, one for the
  52  * inode core, and possibly one for the inode data/extents/b-tree root
  53  * and one for the inode attribute data/extents/b-tree root.
  54  */
  55 STATIC uint
  56 xfs_inode_item_size(
  57         xfs_inode_log_item_t    *iip)
  58 {
  59         uint            nvecs;
  60         xfs_inode_t     *ip;
  61
  62         ip = iip->ili_inode;
  63         nvecs = 2;
  64
  65         /*
  66          * Only log the data/extents/b-tree root if there is something
  67          * left to log.
  68          */
  69         iip->ili_format.ilf_fields |= XFS_ILOG_CORE;
  70
  71         switch (ip->i_d.di_format) {
  72         case XFS_DINODE_FMT_EXTENTS:
  73                 iip->ili_format.ilf_fields &=
  74                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
  75                           XFS_ILOG_DEV | XFS_ILOG_UUID);
  76                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DEXT) &&
  77                     (ip->i_d.di_nextents > 0) &&
  78                     (ip->i_df.if_bytes > 0)) {
  79                         ASSERT(ip->i_df.if_u1.if_extents != NULL);
  80                         nvecs++;
  81                 } else {
  82                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DEXT;
  83                 }
  84                 break;
  85
  86         case XFS_DINODE_FMT_BTREE:
  87                 ASSERT(ip->i_df.if_ext_max ==
  88                        XFS_IFORK_DSIZE(ip) / (uint)sizeof(xfs_bmbt_rec_t));
  89                 iip->ili_format.ilf_fields &=
  90                         ~(XFS_ILOG_DDATA | XFS_ILOG_DEXT |
  91                           XFS_ILOG_DEV | XFS_ILOG_UUID);
  92                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DBROOT) &&
  93                     (ip->i_df.if_broot_bytes > 0)) {
  94                         ASSERT(ip->i_df.if_broot != NULL);
  95                         nvecs++;
  96                 } else {
  97                         ASSERT(!(iip->ili_format.ilf_fields &
  98                                  XFS_ILOG_DBROOT));
  99 #ifdef XFS_TRANS_DEBUG
 100                         if (iip->ili_root_size > 0) {
 101                                 ASSERT(iip->ili_root_size ==
 102                                        ip->i_df.if_broot_bytes);
 103                                 ASSERT(memcmp(iip->ili_orig_root,
 104                                             ip->i_df.if_broot,
 105                                             iip->ili_root_size) == 0);
 106                         } else {
 107                                 ASSERT(ip->i_df.if_broot_bytes == 0);
 108                         }
 109 #endif
 110                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DBROOT;
 111                 }
 112                 break;
 113
 114         case XFS_DINODE_FMT_LOCAL:
 115                 iip->ili_format.ilf_fields &=
 116                         ~(XFS_ILOG_DEXT | XFS_ILOG_DBROOT |
 117                           XFS_ILOG_DEV | XFS_ILOG_UUID);
 118                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DDATA) &&
 119                     (ip->i_df.if_bytes > 0)) {
 120                         ASSERT(ip->i_df.if_u1.if_data != NULL);
 121                         ASSERT(ip->i_d.di_size > 0);
 122                         nvecs++;
 123                 } else {
 124                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DDATA;
 125                 }
 126                 break;
 127
 128         case XFS_DINODE_FMT_DEV:
 129                 iip->ili_format.ilf_fields &=
 130                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 131                           XFS_ILOG_DEXT | XFS_ILOG_UUID);
 132                 break;
 133
 134         case XFS_DINODE_FMT_UUID:
 135                 iip->ili_format.ilf_fields &=
 136                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 137                           XFS_ILOG_DEXT | XFS_ILOG_DEV);
 138                 break;
 139
 140         default:
 141                 ASSERT(0);
 142                 break;
 143         }
 144
 145         /*
 146          * If there are no attributes associated with this file,
 147          * then there cannot be anything more to log.
 148          * Clear all attribute-related log flags.
 149          */
 150         if (!XFS_IFORK_Q(ip)) {
 151                 iip->ili_format.ilf_fields &=
 152                         ~(XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT);
 153                 return nvecs;
 154         }
 155
 156         /*
 157          * Log any necessary attribute data.
 158          */
 159         switch (ip->i_d.di_aformat) {
 160         case XFS_DINODE_FMT_EXTENTS:
 161                 iip->ili_format.ilf_fields &=
 162                         ~(XFS_ILOG_ADATA | XFS_ILOG_ABROOT);
 163                 if ((iip->ili_format.ilf_fields & XFS_ILOG_AEXT) &&
 164                     (ip->i_d.di_anextents > 0) &&
 165                     (ip->i_afp->if_bytes > 0)) {
 166                         ASSERT(ip->i_afp->if_u1.if_extents != NULL);
 167                         nvecs++;
 168                 } else {
 169                         iip->ili_format.ilf_fields &= ~XFS_ILOG_AEXT;
 170                 }
 171                 break;
 172
 173         case XFS_DINODE_FMT_BTREE:
 174                 iip->ili_format.ilf_fields &=
 175                         ~(XFS_ILOG_ADATA | XFS_ILOG_AEXT);
 176                 if ((iip->ili_format.ilf_fields & XFS_ILOG_ABROOT) &&
 177                     (ip->i_afp->if_broot_bytes > 0)) {
 178                         ASSERT(ip->i_afp->if_broot != NULL);
 179                         nvecs++;
 180                 } else {
 181                         iip->ili_format.ilf_fields &= ~XFS_ILOG_ABROOT;
 182                 }
 183                 break;
 184
 185         case XFS_DINODE_FMT_LOCAL:
 186                 iip->ili_format.ilf_fields &=
 187                         ~(XFS_ILOG_AEXT | XFS_ILOG_ABROOT);
 188                 if ((iip->ili_format.ilf_fields & XFS_ILOG_ADATA) &&
 189                     (ip->i_afp->if_bytes > 0)) {
 190                         ASSERT(ip->i_afp->if_u1.if_data != NULL);
 191                         nvecs++;
 192                 } else {
 193                         iip->ili_format.ilf_fields &= ~XFS_ILOG_ADATA;
 194                 }
 195                 break;
 196
 197         default:
 198                 ASSERT(0);
 199                 break;
 200         }
 201
 202         return nvecs;
 203 }
 204
 205 /*
 206  * This is called to fill in the vector of log iovecs for the
 207  * given inode log item.  It fills the first item with an inode
 208  * log format structure, the second with the on-disk inode structure,
 209  * and a possible third and/or fourth with the inode data/extents/b-tree
 210  * root and inode attributes data/extents/b-tree root.
 211  */
 212 STATIC void
 213 xfs_inode_item_format(
 214         xfs_inode_log_item_t    *iip,
 215         xfs_log_iovec_t         *log_vector)
 216 {
 217         uint                    nvecs;
 218         xfs_log_iovec_t         *vecp;
 219         xfs_inode_t             *ip;
 220         size_t                  data_bytes;
 221         xfs_bmbt_rec_t          *ext_buffer;
 222         int                     nrecs;
 223         xfs_mount_t             *mp;
 224
 225         ip = iip->ili_inode;
 226         vecp = log_vector;
 227
 228         vecp->i_addr = (xfs_caddr_t)&iip->ili_format;
 229         vecp->i_len  = sizeof(xfs_inode_log_format_t);
 230         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IFORMAT);
 231         vecp++;
 232         nvecs        = 1;
 233
 234         /*
 235          * Make sure the linux inode is dirty. We do this before
 236          * clearing i_update_core as the VFS will call back into
 237          * XFS here and set i_update_core, so we need to dirty the
 238          * inode first so that the ordering of i_update_core and
 239          * unlogged modifications still works as described below.
 240          */
 241         xfs_mark_inode_dirty_sync(ip);
 242
 243         /*
 244          * Clear i_update_core if the timestamps (or any other
 245          * non-transactional modification) need flushing/logging
 246          * and we're about to log them with the rest of the core.
 247          *
 248          * This is the same logic as xfs_iflush() but this code can't
 249          * run at the same time as xfs_iflush because we're in commit
 250          * processing here and so we have the inode lock held in
 251          * exclusive mode.  Although it doesn't really matter
 252          * for the timestamps if both routines were to grab the
 253          * timestamps or not.  That would be ok.
 254          *
 255          * We clear i_update_core before copying out the data.
 256          * This is for coordination with our timestamp updates
 257          * that don't hold the inode lock. They will always
 258          * update the timestamps BEFORE setting i_update_core,
 259          * so if we clear i_update_core after they set it we
 260          * are guaranteed to see their updates to the timestamps
 261          * either here.  Likewise, if they set it after we clear it
 262          * here, we'll see it either on the next commit of this
 263          * inode or the next time the inode gets flushed via
 264          * xfs_iflush().  This depends on strongly ordered memory
 265          * semantics, but we have that.  We use the SYNCHRONIZE
 266          * macro to make sure that the compiler does not reorder
 267          * the i_update_core access below the data copy below.
 268          */
 269         if (ip->i_update_core)  {
 270                 ip->i_update_core = 0;
 271                 SYNCHRONIZE();
 272         }
 273
 274         /*
 275          * Make sure to get the latest timestamps from the Linux inode.
 276          */
 277         xfs_synchronize_times(ip);
 278
 279         vecp->i_addr = (xfs_caddr_t)&ip->i_d;
 280         vecp->i_len  = sizeof(struct xfs_icdinode);
 281         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_ICORE);
 282         vecp++;
 283         nvecs++;
 284         iip->ili_format.ilf_fields |= XFS_ILOG_CORE;
 285
 286         /*
 287          * If this is really an old format inode, then we need to
 288          * log it as such.  This means that we have to copy the link
 289          * count from the new field to the old.  We don't have to worry
 290          * about the new fields, because nothing trusts them as long as
 291          * the old inode version number is there.  If the superblock already
 292          * has a new version number, then we don't bother converting back.
 293          */
 294         mp = ip->i_mount;
 295         ASSERT(ip->i_d.di_version == 1 || xfs_sb_version_hasnlink(&mp->m_sb));
 296         if (ip->i_d.di_version == 1) {
 297                 if (!xfs_sb_version_hasnlink(&mp->m_sb)) {
 298                         /*
 299                          * Convert it back.
 300                          */
 301                         ASSERT(ip->i_d.di_nlink <= XFS_MAXLINK_1);
 302                         ip->i_d.di_onlink = ip->i_d.di_nlink;
 303                 } else {
 304                         /*
 305                          * The superblock version has already been bumped,
 306                          * so just make the conversion to the new inode
 307                          * format permanent.
 308                          */
 309                         ip->i_d.di_version = 2;
 310                         ip->i_d.di_onlink = 0;
 311                         memset(&(ip->i_d.di_pad[0]), 0, sizeof(ip->i_d.di_pad));
 312                 }
 313         }
 314
 315         switch (ip->i_d.di_format) {
 316         case XFS_DINODE_FMT_EXTENTS:
 317                 ASSERT(!(iip->ili_format.ilf_fields &
 318                          (XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 319                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 320                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEXT) {
 321                         ASSERT(ip->i_df.if_bytes > 0);
 322                         ASSERT(ip->i_df.if_u1.if_extents != NULL);
 323                         ASSERT(ip->i_d.di_nextents > 0);
 324                         ASSERT(iip->ili_extents_buf == NULL);
 325                         nrecs = ip->i_df.if_bytes /
 326                                 (uint)sizeof(xfs_bmbt_rec_t);
 327                         ASSERT(nrecs > 0);
 328 #ifdef XFS_NATIVE_HOST
 329                         if (nrecs == ip->i_d.di_nextents) {
 330                                 /*
 331                                  * There are no delayed allocation
 332                                  * extents, so just point to the
 333                                  * real extents array.
 334                                  */
 335                                 vecp->i_addr =
 336                                         (char *)(ip->i_df.if_u1.if_extents);
 337                                 vecp->i_len = ip->i_df.if_bytes;
 338                                 XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IEXT);
 339                         } else
 340 #endif
 341                         {
 342                                 /*
 343                                  * There are delayed allocation extents
 344                                  * in the inode, or we need to convert
 345                                  * the extents to on disk format.
 346                                  * Use xfs_iextents_copy()
 347                                  * to copy only the real extents into
 348                                  * a separate buffer.  We'll free the
 349                                  * buffer in the unlock routine.
 350                                  */
 351                                 ext_buffer = kmem_alloc(ip->i_df.if_bytes,
 352                                         KM_SLEEP);
 353                                 iip->ili_extents_buf = ext_buffer;
 354                                 vecp->i_addr = (xfs_caddr_t)ext_buffer;
 355                                 vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
 356                                                 XFS_DATA_FORK);
 357                                 XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IEXT);
 358                         }
 359                         ASSERT(vecp->i_len <= ip->i_df.if_bytes);
 360                         iip->ili_format.ilf_dsize = vecp->i_len;
 361                         vecp++;
 362                         nvecs++;
 363                 }
 364                 break;
 365
 366         case XFS_DINODE_FMT_BTREE:
 367                 ASSERT(!(iip->ili_format.ilf_fields &
 368                          (XFS_ILOG_DDATA | XFS_ILOG_DEXT |
 369                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 370                 if (iip->ili_format.ilf_fields & XFS_ILOG_DBROOT) {
 371                         ASSERT(ip->i_df.if_broot_bytes > 0);
 372                         ASSERT(ip->i_df.if_broot != NULL);
 373                         vecp->i_addr = (xfs_caddr_t)ip->i_df.if_broot;
 374                         vecp->i_len = ip->i_df.if_broot_bytes;
 375                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IBROOT);
 376                         vecp++;
 377                         nvecs++;
 378                         iip->ili_format.ilf_dsize = ip->i_df.if_broot_bytes;
 379                 }
 380                 break;
 381
 382         case XFS_DINODE_FMT_LOCAL:
 383                 ASSERT(!(iip->ili_format.ilf_fields &
 384                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 385                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 386                 if (iip->ili_format.ilf_fields & XFS_ILOG_DDATA) {
 387                         ASSERT(ip->i_df.if_bytes > 0);
 388                         ASSERT(ip->i_df.if_u1.if_data != NULL);
 389                         ASSERT(ip->i_d.di_size > 0);
 390
 391                         vecp->i_addr = (xfs_caddr_t)ip->i_df.if_u1.if_data;
 392                         /*
 393                          * Round i_bytes up to a word boundary.
 394                          * The underlying memory is guaranteed to
 395                          * to be there by xfs_idata_realloc().
 396                          */
 397                         data_bytes = roundup(ip->i_df.if_bytes, 4);
 398                         ASSERT((ip->i_df.if_real_bytes == 0) ||
 399                                (ip->i_df.if_real_bytes == data_bytes));
 400                         vecp->i_len = (int)data_bytes;
 401                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_ILOCAL);
 402                         vecp++;
 403                         nvecs++;
 404                         iip->ili_format.ilf_dsize = (unsigned)data_bytes;
 405                 }
 406                 break;
 407
 408         case XFS_DINODE_FMT_DEV:
 409                 ASSERT(!(iip->ili_format.ilf_fields &
 410                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 411                           XFS_ILOG_DDATA | XFS_ILOG_UUID)));
 412                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEV) {
 413                         iip->ili_format.ilf_u.ilfu_rdev =
 414                                 ip->i_df.if_u2.if_rdev;
 415                 }
 416                 break;
 417
 418         case XFS_DINODE_FMT_UUID:
 419                 ASSERT(!(iip->ili_format.ilf_fields &
 420                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 421                           XFS_ILOG_DDATA | XFS_ILOG_DEV)));
 422                 if (iip->ili_format.ilf_fields & XFS_ILOG_UUID) {
 423                         iip->ili_format.ilf_u.ilfu_uuid =
 424                                 ip->i_df.if_u2.if_uuid;
 425                 }
 426                 break;
 427
 428         default:
 429                 ASSERT(0);
 430                 break;
 431         }
 432
 433         /*
 434          * If there are no attributes associated with the file,
 435          * then we're done.
 436          * Assert that no attribute-related log flags are set.
 437          */
 438         if (!XFS_IFORK_Q(ip)) {
 439                 ASSERT(nvecs == iip->ili_item.li_desc->lid_size);
 440                 iip->ili_format.ilf_size = nvecs;
 441                 ASSERT(!(iip->ili_format.ilf_fields &
 442                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
 443                 return;
 444         }
 445
 446         switch (ip->i_d.di_aformat) {
 447         case XFS_DINODE_FMT_EXTENTS:
 448                 ASSERT(!(iip->ili_format.ilf_fields &
 449                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT)));
 450                 if (iip->ili_format.ilf_fields & XFS_ILOG_AEXT) {
 451                         ASSERT(ip->i_afp->if_bytes > 0);
 452                         ASSERT(ip->i_afp->if_u1.if_extents != NULL);
 453                         ASSERT(ip->i_d.di_anextents > 0);
 454 #ifdef DEBUG
 455                         nrecs = ip->i_afp->if_bytes /
 456                                 (uint)sizeof(xfs_bmbt_rec_t);
 457 #endif
 458                         ASSERT(nrecs > 0);
 459                         ASSERT(nrecs == ip->i_d.di_anextents);
 460 #ifdef XFS_NATIVE_HOST
 461                         /*
 462                          * There are not delayed allocation extents
 463                          * for attributes, so just point at the array.
 464                          */
 465                         vecp->i_addr = (char *)(ip->i_afp->if_u1.if_extents);
 466                         vecp->i_len = ip->i_afp->if_bytes;
 467 #else
 468                         ASSERT(iip->ili_aextents_buf == NULL);
 469                         /*
 470                          * Need to endian flip before logging
 471                          */
 472                         ext_buffer = kmem_alloc(ip->i_afp->if_bytes,
 473                                 KM_SLEEP);
 474                         iip->ili_aextents_buf = ext_buffer;
 475                         vecp->i_addr = (xfs_caddr_t)ext_buffer;
 476                         vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
 477                                         XFS_ATTR_FORK);
 478 #endif
 479                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IATTR_EXT);
 480                         iip->ili_format.ilf_asize = vecp->i_len;
 481                         vecp++;
 482                         nvecs++;
 483                 }
 484                 break;
 485
 486         case XFS_DINODE_FMT_BTREE:
 487                 ASSERT(!(iip->ili_format.ilf_fields &
 488                          (XFS_ILOG_ADATA | XFS_ILOG_AEXT)));
 489                 if (iip->ili_format.ilf_fields & XFS_ILOG_ABROOT) {
 490                         ASSERT(ip->i_afp->if_broot_bytes > 0);
 491                         ASSERT(ip->i_afp->if_broot != NULL);
 492                         vecp->i_addr = (xfs_caddr_t)ip->i_afp->if_broot;
 493                         vecp->i_len = ip->i_afp->if_broot_bytes;
 494                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IATTR_BROOT);
 495                         vecp++;
 496                         nvecs++;
 497                         iip->ili_format.ilf_asize = ip->i_afp->if_broot_bytes;
 498                 }
 499                 break;
 500
 501         case XFS_DINODE_FMT_LOCAL:
 502                 ASSERT(!(iip->ili_format.ilf_fields &
 503                          (XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
 504                 if (iip->ili_format.ilf_fields & XFS_ILOG_ADATA) {
 505                         ASSERT(ip->i_afp->if_bytes > 0);
 506                         ASSERT(ip->i_afp->if_u1.if_data != NULL);
 507
 508                         vecp->i_addr = (xfs_caddr_t)ip->i_afp->if_u1.if_data;
 509                         /*
 510                          * Round i_bytes up to a word boundary.
 511                          * The underlying memory is guaranteed to
 512                          * to be there by xfs_idata_realloc().
 513                          */
 514                         data_bytes = roundup(ip->i_afp->if_bytes, 4);
 515                         ASSERT((ip->i_afp->if_real_bytes == 0) ||
 516                                (ip->i_afp->if_real_bytes == data_bytes));
 517                         vecp->i_len = (int)data_bytes;
 518                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IATTR_LOCAL);
 519                         vecp++;
 520                         nvecs++;
 521                         iip->ili_format.ilf_asize = (unsigned)data_bytes;
 522                 }
 523                 break;
 524
 525         default:
 526                 ASSERT(0);
 527                 break;
 528         }
 529
 530         ASSERT(nvecs == iip->ili_item.li_desc->lid_size);
 531         iip->ili_format.ilf_size = nvecs;
 532 }
 533
 534
 535 /*
 536  * This is called to pin the inode associated with the inode log
 537  * item in memory so it cannot be written out.  Do this by calling
 538  * xfs_ipin() to bump the pin count in the inode while holding the
 539  * inode pin lock.
 540  */
 541 STATIC void
 542 xfs_inode_item_pin(
 543         xfs_inode_log_item_t    *iip)
 544 {
 545         ASSERT(xfs_isilocked(iip->ili_inode, XFS_ILOCK_EXCL));
 546         xfs_ipin(iip->ili_inode);
 547 }
 548
 549
 550 /*
 551  * This is called to unpin the inode associated with the inode log
 552  * item which was previously pinned with a call to xfs_inode_item_pin().
 553  * Just call xfs_iunpin() on the inode to do this.
 554  */
 555 /* ARGSUSED */
 556 STATIC void
 557 xfs_inode_item_unpin(
 558         xfs_inode_log_item_t    *iip,
 559         int                     stale)
 560 {
 561         xfs_iunpin(iip->ili_inode);
 562 }
 563
 564 /* ARGSUSED */
 565 STATIC void
 566 xfs_inode_item_unpin_remove(
 567         xfs_inode_log_item_t    *iip,
 568         xfs_trans_t             *tp)
 569 {
 570         xfs_iunpin(iip->ili_inode);
 571 }
 572
 573 /*
 574  * This is called to attempt to lock the inode associated with this
 575  * inode log item, in preparation for the push routine which does the actual
 576  * iflush.  Don't sleep on the inode lock or the flush lock.
 577  *
 578  * If the flush lock is already held, indicating that the inode has
 579  * been or is in the process of being flushed, then (ideally) we'd like to
 580  * see if the inode's buffer is still incore, and if so give it a nudge.
 581  * We delay doing so until the pushbuf routine, though, to avoid holding
 582  * the AIL lock across a call to the blackhole which is the buffer cache.
 583  * Also we don't want to sleep in any device strategy routines, which can happen
 584  * if we do the subsequent bawrite in here.
 585  */
 586 STATIC uint
 587 xfs_inode_item_trylock(
 588         xfs_inode_log_item_t    *iip)
 589 {
 590         register xfs_inode_t    *ip;
 591
 592         ip = iip->ili_inode;
 593
 594         if (xfs_ipincount(ip) > 0) {
 595                 return XFS_ITEM_PINNED;
 596         }
 597
 598         if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
 599                 return XFS_ITEM_LOCKED;
 600         }
 601
 602         if (!xfs_iflock_nowait(ip)) {
 603                 /*
 604                  * If someone else isn't already trying to push the inode
 605                  * buffer, we get to do it.
 606                  */
 607                 if (iip->ili_pushbuf_flag == 0) {
 608                         iip->ili_pushbuf_flag = 1;
 609 #ifdef DEBUG
 610                         iip->ili_push_owner = current_pid();
 611 #endif
 612                         /*
 613                          * Inode is left locked in shared mode.
 614                          * Pushbuf routine gets to unlock it.
 615                          */
 616                         return XFS_ITEM_PUSHBUF;
 617                 } else {
 618                         /*
 619                          * We hold the AIL lock, so we must specify the
 620                          * NONOTIFY flag so that we won't double trip.
 621                          */
 622                         xfs_iunlock(ip, XFS_ILOCK_SHARED|XFS_IUNLOCK_NONOTIFY);
 623                         return XFS_ITEM_FLUSHING;
 624                 }
 625                 /* NOTREACHED */
 626         }
 627
 628         /* Stale items should force out the iclog */
 629         if (ip->i_flags & XFS_ISTALE) {
 630                 xfs_ifunlock(ip);
 631                 xfs_iunlock(ip, XFS_ILOCK_SHARED|XFS_IUNLOCK_NONOTIFY);
 632                 return XFS_ITEM_PINNED;
 633         }
 634
 635 #ifdef DEBUG
 636         if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
 637                 ASSERT(iip->ili_format.ilf_fields != 0);
 638                 ASSERT(iip->ili_logged == 0);
 639                 ASSERT(iip->ili_item.li_flags & XFS_LI_IN_AIL);
 640         }
 641 #endif
 642         return XFS_ITEM_SUCCESS;
 643 }
 644
 645 /*
 646  * Unlock the inode associated with the inode log item.
 647  * Clear the fields of the inode and inode log item that
 648  * are specific to the current transaction.  If the
 649  * hold flags is set, do not unlock the inode.
 650  */
 651 STATIC void
 652 xfs_inode_item_unlock(
 653         xfs_inode_log_item_t    *iip)
 654 {
 655         uint            hold;
 656         uint            iolocked;
 657         uint            lock_flags;
 658         xfs_inode_t     *ip;
 659
 660         ASSERT(iip != NULL);
 661         ASSERT(iip->ili_inode->i_itemp != NULL);
 662         ASSERT(xfs_isilocked(iip->ili_inode, XFS_ILOCK_EXCL));
 663         ASSERT((!(iip->ili_inode->i_itemp->ili_flags &
 664                   XFS_ILI_IOLOCKED_EXCL)) ||
 665                xfs_isilocked(iip->ili_inode, XFS_IOLOCK_EXCL));
 666         ASSERT((!(iip->ili_inode->i_itemp->ili_flags &
 667                   XFS_ILI_IOLOCKED_SHARED)) ||
 668                xfs_isilocked(iip->ili_inode, XFS_IOLOCK_SHARED));
 669         /*
 670          * Clear the transaction pointer in the inode.
 671          */
 672         ip = iip->ili_inode;
 673         ip->i_transp = NULL;
 674
 675         /*
 676          * If the inode needed a separate buffer with which to log
 677          * its extents, then free it now.
 678          */
 679         if (iip->ili_extents_buf != NULL) {
 680                 ASSERT(ip->i_d.di_format == XFS_DINODE_FMT_EXTENTS);
 681                 ASSERT(ip->i_d.di_nextents > 0);
 682                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_DEXT);
 683                 ASSERT(ip->i_df.if_bytes > 0);
 684                 kmem_free(iip->ili_extents_buf);
 685                 iip->ili_extents_buf = NULL;
 686         }
 687         if (iip->ili_aextents_buf != NULL) {
 688                 ASSERT(ip->i_d.di_aformat == XFS_DINODE_FMT_EXTENTS);
 689                 ASSERT(ip->i_d.di_anextents > 0);
 690                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_AEXT);
 691                 ASSERT(ip->i_afp->if_bytes > 0);
 692                 kmem_free(iip->ili_aextents_buf);
 693                 iip->ili_aextents_buf = NULL;
 694         }
 695
 696         /*
 697          * Figure out if we should unlock the inode or not.
 698          */
 699         hold = iip->ili_flags & XFS_ILI_HOLD;
 700
 701         /*
 702          * Before clearing out the flags, remember whether we
 703          * are holding the inode's IO lock.
 704          */
 705         iolocked = iip->ili_flags & XFS_ILI_IOLOCKED_ANY;
 706
 707         /*
 708          * Clear out the fields of the inode log item particular
 709          * to the current transaction.
 710          */
 711         iip->ili_flags = 0;
 712
 713         /*
 714          * Unlock the inode if XFS_ILI_HOLD was not set.
 715          */
 716         if (!hold) {
 717                 lock_flags = XFS_ILOCK_EXCL;
 718                 if (iolocked & XFS_ILI_IOLOCKED_EXCL) {
 719                         lock_flags |= XFS_IOLOCK_EXCL;
 720                 } else if (iolocked & XFS_ILI_IOLOCKED_SHARED) {
 721                         lock_flags |= XFS_IOLOCK_SHARED;
 722                 }
 723                 xfs_iput(iip->ili_inode, lock_flags);
 724         }
 725 }
 726
 727 /*
 728  * This is called to find out where the oldest active copy of the
 729  * inode log item in the on disk log resides now that the last log
 730  * write of it completed at the given lsn.  Since we always re-log
 731  * all dirty data in an inode, the latest copy in the on disk log
 732  * is the only one that matters.  Therefore, simply return the
 733  * given lsn.
 734  */
 735 /*ARGSUSED*/
 736 STATIC xfs_lsn_t
 737 xfs_inode_item_committed(
 738         xfs_inode_log_item_t    *iip,
 739         xfs_lsn_t               lsn)
 740 {
 741         return (lsn);
 742 }
 743
 744 /*
 745  * This gets called by xfs_trans_push_ail(), when IOP_TRYLOCK
 746  * failed to get the inode flush lock but did get the inode locked SHARED.
 747  * Here we're trying to see if the inode buffer is incore, and if so whether it's
 748  * marked delayed write. If that's the case, we'll initiate a bawrite on that
 749  * buffer to expedite the process.
 750  *
 751  * We aren't holding the AIL lock (or the flush lock) when this gets called,
 752  * so it is inherently race-y.
 753  */
 754 STATIC void
 755 xfs_inode_item_pushbuf(
 756         xfs_inode_log_item_t    *iip)
 757 {
 758         xfs_inode_t     *ip;
 759         xfs_mount_t     *mp;
 760         xfs_buf_t       *bp;
 761         uint            dopush;
 762
 763         ip = iip->ili_inode;
 764
 765         ASSERT(xfs_isilocked(ip, XFS_ILOCK_SHARED));
 766
 767         /*
 768          * The ili_pushbuf_flag keeps others from
 769          * trying to duplicate our effort.
 770          */
 771         ASSERT(iip->ili_pushbuf_flag != 0);
 772         ASSERT(iip->ili_push_owner == current_pid());
 773
 774         /*
 775          * If a flush is not in progress anymore, chances are that the
 776          * inode was taken off the AIL. So, just get out.
 777          */
 778         if (completion_done(&ip->i_flush) ||
 779             ((iip->ili_item.li_flags & XFS_LI_IN_AIL) == 0)) {
 780                 iip->ili_pushbuf_flag = 0;
 781                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
 782                 return;
 783         }
 784
 785         mp = ip->i_mount;
 786         bp = xfs_incore(mp->m_ddev_targp, iip->ili_format.ilf_blkno,
 787                     iip->ili_format.ilf_len, XFS_INCORE_TRYLOCK);
 788
 789         if (bp != NULL) {
 790                 if (XFS_BUF_ISDELAYWRITE(bp)) {
 791                         /*
 792                          * We were racing with iflush because we don't hold
 793                          * the AIL lock or the flush lock. However, at this point,
 794                          * we have the buffer, and we know that it's dirty.
 795                          * So, it's possible that iflush raced with us, and
 796                          * this item is already taken off the AIL.
 797                          * If not, we can flush it async.
 798                          */
 799                         dopush = ((iip->ili_item.li_flags & XFS_LI_IN_AIL) &&
 800                                   !completion_done(&ip->i_flush));
 801                         iip->ili_pushbuf_flag = 0;
 802                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 803                         xfs_buftrace("INODE ITEM PUSH", bp);
 804                         if (XFS_BUF_ISPINNED(bp)) {
 805                                 xfs_log_force(mp, (xfs_lsn_t)0,
 806                                               XFS_LOG_FORCE);
 807                         }
 808                         if (dopush) {
 809                                 int     error;
 810                                 error = xfs_bawrite(mp, bp);
 811                                 if (error)
 812                                         xfs_fs_cmn_err(CE_WARN, mp,
 813                 "xfs_inode_item_pushbuf: pushbuf error %d on iip %p, bp %p",
 814                                                         error, iip, bp);
 815                         } else {
 816                                 xfs_buf_relse(bp);
 817                         }
 818                 } else {
 819                         iip->ili_pushbuf_flag = 0;
 820                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 821                         xfs_buf_relse(bp);
 822                 }
 823                 return;
 824         }
 825         /*
 826          * We have to be careful about resetting pushbuf flag too early (above).
 827          * Even though in theory we can do it as soon as we have the buflock,
 828          * we don't want others to be doing work needlessly. They'll come to
 829          * this function thinking that pushing the buffer is their
 830          * responsibility only to find that the buffer is still locked by
 831          * another doing the same thing
 832          */
 833         iip->ili_pushbuf_flag = 0;
 834         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 835         return;
 836 }
 837
 838
 839 /*
 840  * This is called to asynchronously write the inode associated with this
 841  * inode log item out to disk. The inode will already have been locked by
 842  * a successful call to xfs_inode_item_trylock().
 843  */
 844 STATIC void
 845 xfs_inode_item_push(
 846         xfs_inode_log_item_t    *iip)
 847 {
 848         xfs_inode_t     *ip;
 849
 850         ip = iip->ili_inode;
 851
 852         ASSERT(xfs_isilocked(ip, XFS_ILOCK_SHARED));
 853         ASSERT(!completion_done(&ip->i_flush));
 854         /*
 855          * Since we were able to lock the inode's flush lock and
 856          * we found it on the AIL, the inode must be dirty.  This
 857          * is because the inode is removed from the AIL while still
 858          * holding the flush lock in xfs_iflush_done().  Thus, if
 859          * we found it in the AIL and were able to obtain the flush
 860          * lock without sleeping, then there must not have been
 861          * anyone in the process of flushing the inode.
 862          */
 863         ASSERT(XFS_FORCED_SHUTDOWN(ip->i_mount) ||
 864                iip->ili_format.ilf_fields != 0);
 865
 866         /*
 867          * Write out the inode.  The completion routine ('iflush_done') will
 868          * pull it from the AIL, mark it clean, unlock the flush lock.
 869          */
 870         (void) xfs_iflush(ip, XFS_IFLUSH_ASYNC);
 871         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 872
 873         return;
 874 }
 875
 876 /*
 877  * XXX rcc - this one really has to do something.  Probably needs
 878  * to stamp in a new field in the incore inode.
 879  */
 880 /* ARGSUSED */
 881 STATIC void
 882 xfs_inode_item_committing(
 883         xfs_inode_log_item_t    *iip,
 884         xfs_lsn_t               lsn)
 885 {
 886         iip->ili_last_lsn = lsn;
 887         return;
 888 }
 889
 890 /*
 891  * This is the ops vector shared by all buf log items.
 892  */
 893 static struct xfs_item_ops xfs_inode_item_ops = {
 894         .iop_size       = (uint(*)(xfs_log_item_t*))xfs_inode_item_size,
 895         .iop_format     = (void(*)(xfs_log_item_t*, xfs_log_iovec_t*))
 896                                         xfs_inode_item_format,
 897         .iop_pin        = (void(*)(xfs_log_item_t*))xfs_inode_item_pin,
 898         .iop_unpin      = (void(*)(xfs_log_item_t*, int))xfs_inode_item_unpin,
 899         .iop_unpin_remove = (void(*)(xfs_log_item_t*, xfs_trans_t*))
 900                                         xfs_inode_item_unpin_remove,
 901         .iop_trylock    = (uint(*)(xfs_log_item_t*))xfs_inode_item_trylock,
 902         .iop_unlock     = (void(*)(xfs_log_item_t*))xfs_inode_item_unlock,
 903         .iop_committed  = (xfs_lsn_t(*)(xfs_log_item_t*, xfs_lsn_t))
 904                                         xfs_inode_item_committed,
 905         .iop_push       = (void(*)(xfs_log_item_t*))xfs_inode_item_push,
 906         .iop_pushbuf    = (void(*)(xfs_log_item_t*))xfs_inode_item_pushbuf,
 907         .iop_committing = (void(*)(xfs_log_item_t*, xfs_lsn_t))
 908                                         xfs_inode_item_committing
 909 };
 910
 911
 912 /*
 913  * Initialize the inode log item for a newly allocated (in-core) inode.
 914  */
 915 void
 916 xfs_inode_item_init(
 917         xfs_inode_t     *ip,
 918         xfs_mount_t     *mp)
 919 {
 920         xfs_inode_log_item_t    *iip;
 921
 922         ASSERT(ip->i_itemp == NULL);
 923         iip = ip->i_itemp = kmem_zone_zalloc(xfs_ili_zone, KM_SLEEP);
 924
 925         iip->ili_item.li_type = XFS_LI_INODE;
 926         iip->ili_item.li_ops = &xfs_inode_item_ops;
 927         iip->ili_item.li_mountp = mp;
 928         iip->ili_item.li_ailp = mp->m_ail;
 929         iip->ili_inode = ip;
 930
 931         /*
 932            We have zeroed memory. No need ...
 933            iip->ili_extents_buf = NULL;
 934            iip->ili_pushbuf_flag = 0;
 935          */
 936
 937         iip->ili_format.ilf_type = XFS_LI_INODE;
 938         iip->ili_format.ilf_ino = ip->i_ino;
 939         iip->ili_format.ilf_blkno = ip->i_imap.im_blkno;
 940         iip->ili_format.ilf_len = ip->i_imap.im_len;
 941         iip->ili_format.ilf_boffset = ip->i_imap.im_boffset;
 942 }
 943
 944 /*
 945  * Free the inode log item and any memory hanging off of it.
 946  */
 947 void
 948 xfs_inode_item_destroy(
 949         xfs_inode_t     *ip)
 950 {
 951 #ifdef XFS_TRANS_DEBUG
 952         if (ip->i_itemp->ili_root_size != 0) {
 953                 kmem_free(ip->i_itemp->ili_orig_root);
 954         }
 955 #endif
 956         kmem_zone_free(xfs_ili_zone, ip->i_itemp);
 957 }
 958
 959
 960 /*
 961  * This is the inode flushing I/O completion routine.  It is called
 962  * from interrupt level when the buffer containing the inode is
 963  * flushed to disk.  It is responsible for removing the inode item
 964  * from the AIL if it has not been re-logged, and unlocking the inode's
 965  * flush lock.
 966  */
 967 /*ARGSUSED*/
 968 void
 969 xfs_iflush_done(
 970         xfs_buf_t               *bp,
 971         xfs_inode_log_item_t    *iip)
 972 {
 973         xfs_inode_t             *ip = iip->ili_inode;
 974         struct xfs_ail          *ailp = iip->ili_item.li_ailp;
 975
 976         /*
 977          * We only want to pull the item from the AIL if it is
 978          * actually there and its location in the log has not
 979          * changed since we started the flush.  Thus, we only bother
 980          * if the ili_logged flag is set and the inode's lsn has not
 981          * changed.  First we check the lsn outside
 982          * the lock since it's cheaper, and then we recheck while
 983          * holding the lock before removing the inode from the AIL.
 984          */
 985         if (iip->ili_logged &&
 986             (iip->ili_item.li_lsn == iip->ili_flush_lsn)) {
 987                 spin_lock(&ailp->xa_lock);
 988                 if (iip->ili_item.li_lsn == iip->ili_flush_lsn) {
 989                         /* xfs_trans_ail_delete() drops the AIL lock. */
 990                         xfs_trans_ail_delete(ailp, (xfs_log_item_t*)iip);
 991                 } else {
 992                         spin_unlock(&ailp->xa_lock);
 993                 }
 994         }
 995
 996         iip->ili_logged = 0;
 997
 998         /*
 999          * Clear the ili_last_fields bits now that we know that the
1000          * data corresponding to them is safely on disk.
1001          */
1002         iip->ili_last_fields = 0;
1003
1004         /*
1005          * Release the inode's flush lock since we're done with it.
1006          */
1007         xfs_ifunlock(ip);
1008
1009         return;
1010 }
1011
1012 /*
1013  * This is the inode flushing abort routine.  It is called
1014  * from xfs_iflush when the filesystem is shutting down to clean
1015  * up the inode state.
1016  * It is responsible for removing the inode item
1017  * from the AIL if it has not been re-logged, and unlocking the inode's
1018  * flush lock.
1019  */
1020 void
1021 xfs_iflush_abort(
1022         xfs_inode_t             *ip)
1023 {
1024         xfs_inode_log_item_t    *iip = ip->i_itemp;
1025         xfs_mount_t             *mp;
1026
1027         iip = ip->i_itemp;
1028         mp = ip->i_mount;
1029         if (iip) {
1030                 struct xfs_ail  *ailp = iip->ili_item.li_ailp;
1031                 if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
1032                         spin_lock(&ailp->xa_lock);
1033                         if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
1034                                 /* xfs_trans_ail_delete() drops the AIL lock. */
1035                                 xfs_trans_ail_delete(ailp, (xfs_log_item_t *)iip);
1036                         } else
1037                                 spin_unlock(&ailp->xa_lock);
1038                 }
1039                 iip->ili_logged = 0;
1040                 /*
1041                  * Clear the ili_last_fields bits now that we know that the
1042                  * data corresponding to them is safely on disk.
1043                  */
1044                 iip->ili_last_fields = 0;
1045                 /*
1046                  * Clear the inode logging fields so no more flushes are
1047                  * attempted.
1048                  */
1049                 iip->ili_format.ilf_fields = 0;
1050         }
1051         /*
1052          * Release the inode's flush lock since we're done with it.
1053          */
1054         xfs_ifunlock(ip);
1055 }
1056
1057 void
1058 xfs_istale_done(
1059         xfs_buf_t               *bp,
1060         xfs_inode_log_item_t    *iip)
1061 {
1062         xfs_iflush_abort(iip->ili_inode);
1063 }
1064
1065 /*
1066  * convert an xfs_inode_log_format struct from either 32 or 64 bit versions
1067  * (which can have different field alignments) to the native version
1068  */
1069 int
1070 xfs_inode_item_format_convert(
1071         xfs_log_iovec_t         *buf,
1072         xfs_inode_log_format_t  *in_f)
1073 {
1074         if (buf->i_len == sizeof(xfs_inode_log_format_32_t)) {
1075                 xfs_inode_log_format_32_t *in_f32;
1076
1077                 in_f32 = (xfs_inode_log_format_32_t *)buf->i_addr;
1078                 in_f->ilf_type = in_f32->ilf_type;
1079                 in_f->ilf_size = in_f32->ilf_size;
1080                 in_f->ilf_fields = in_f32->ilf_fields;
1081                 in_f->ilf_asize = in_f32->ilf_asize;
1082                 in_f->ilf_dsize = in_f32->ilf_dsize;
1083                 in_f->ilf_ino = in_f32->ilf_ino;
1084                 /* copy biggest field of ilf_u */
1085                 memcpy(in_f->ilf_u.ilfu_uuid.__u_bits,
1086                        in_f32->ilf_u.ilfu_uuid.__u_bits,
1087                        sizeof(uuid_t));
1088                 in_f->ilf_blkno = in_f32->ilf_blkno;
1089                 in_f->ilf_len = in_f32->ilf_len;
1090                 in_f->ilf_boffset = in_f32->ilf_boffset;
1091                 return 0;
1092         } else if (buf->i_len == sizeof(xfs_inode_log_format_64_t)){
1093                 xfs_inode_log_format_64_t *in_f64;
1094
1095                 in_f64 = (xfs_inode_log_format_64_t *)buf->i_addr;
1096                 in_f->ilf_type = in_f64->ilf_type;
1097                 in_f->ilf_size = in_f64->ilf_size;
1098                 in_f->ilf_fields = in_f64->ilf_fields;
1099                 in_f->ilf_asize = in_f64->ilf_asize;
1100                 in_f->ilf_dsize = in_f64->ilf_dsize;
1101                 in_f->ilf_ino = in_f64->ilf_ino;
1102                 /* copy biggest field of ilf_u */
1103                 memcpy(in_f->ilf_u.ilfu_uuid.__u_bits,
1104                        in_f64->ilf_u.ilfu_uuid.__u_bits,
1105                        sizeof(uuid_t));
1106                 in_f->ilf_blkno = in_f64->ilf_blkno;
1107                 in_f->ilf_len = in_f64->ilf_len;
1108                 in_f->ilf_boffset = in_f64->ilf_boffset;
1109                 return 0;
1110         }
1111         return EFSCORRUPTED;
1112 }