fs/xfs/xfs_inode_item.c

   1 /*
   2  * Copyright (c) 2000-2002 Silicon Graphics, Inc.  All Rights Reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms of version 2 of the GNU General Public License as
   6  * published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it would be useful, but
   9  * WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  11  *
  12  * Further, this software is distributed without any warranty that it is
  13  * free of the rightful claim of any third person regarding infringement
  14  * or the like.  Any license provided herein, whether implied or
  15  * otherwise, applies only to this software file.  Patent licenses, if
  16  * any, provided herein do not apply to combinations of this program with
  17  * other software, or any other product whatsoever.
  18  *
  19  * You should have received a copy of the GNU General Public License along
  20  * with this program; if not, write the Free Software Foundation, Inc., 59
  21  * Temple Place - Suite 330, Boston MA 02111-1307, USA.
  22  *
  23  * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy,
  24  * Mountain View, CA  94043, or:
  25  *
  26  * http://www.sgi.com
  27  *
  28  * For further information regarding this notice, see:
  29  *
  30  * http://oss.sgi.com/projects/GenInfo/SGIGPLNoticeExplan/
  31  */
  32 #include "xfs.h"
  33 #include "xfs_fs.h"
  34 #include "xfs_types.h"
  35 #include "xfs_bit.h"
  36 #include "xfs_log.h"
  37 #include "xfs_inum.h"
  38 #include "xfs_trans.h"
  39 #include "xfs_buf_item.h"
  40 #include "xfs_sb.h"
  41 #include "xfs_ag.h"
  42 #include "xfs_dir.h"
  43 #include "xfs_dir2.h"
  44 #include "xfs_dmapi.h"
  45 #include "xfs_mount.h"
  46 #include "xfs_trans_priv.h"
  47 #include "xfs_bmap_btree.h"
  48 #include "xfs_alloc_btree.h"
  49 #include "xfs_ialloc_btree.h"
  50 #include "xfs_dir_sf.h"
  51 #include "xfs_dir2_sf.h"
  52 #include "xfs_attr_sf.h"
  53 #include "xfs_dinode.h"
  54 #include "xfs_inode.h"
  55 #include "xfs_inode_item.h"
  56 #include "xfs_btree.h"
  57 #include "xfs_ialloc.h"
  58 #include "xfs_rw.h"
  59
  60
  61 kmem_zone_t     *xfs_ili_zone;          /* inode log item zone */
  62
  63 /*
  64  * This returns the number of iovecs needed to log the given inode item.
  65  *
  66  * We need one iovec for the inode log format structure, one for the
  67  * inode core, and possibly one for the inode data/extents/b-tree root
  68  * and one for the inode attribute data/extents/b-tree root.
  69  */
  70 STATIC uint
  71 xfs_inode_item_size(
  72         xfs_inode_log_item_t    *iip)
  73 {
  74         uint            nvecs;
  75         xfs_inode_t     *ip;
  76
  77         ip = iip->ili_inode;
  78         nvecs = 2;
  79
  80         /*
  81          * Only log the data/extents/b-tree root if there is something
  82          * left to log.
  83          */
  84         iip->ili_format.ilf_fields |= XFS_ILOG_CORE;
  85
  86         switch (ip->i_d.di_format) {
  87         case XFS_DINODE_FMT_EXTENTS:
  88                 iip->ili_format.ilf_fields &=
  89                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
  90                           XFS_ILOG_DEV | XFS_ILOG_UUID);
  91                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DEXT) &&
  92                     (ip->i_d.di_nextents > 0) &&
  93                     (ip->i_df.if_bytes > 0)) {
  94                         ASSERT(ip->i_df.if_u1.if_extents != NULL);
  95                         nvecs++;
  96                 } else {
  97                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DEXT;
  98                 }
  99                 break;
 100
 101         case XFS_DINODE_FMT_BTREE:
 102                 ASSERT(ip->i_df.if_ext_max ==
 103                        XFS_IFORK_DSIZE(ip) / (uint)sizeof(xfs_bmbt_rec_t));
 104                 iip->ili_format.ilf_fields &=
 105                         ~(XFS_ILOG_DDATA | XFS_ILOG_DEXT |
 106                           XFS_ILOG_DEV | XFS_ILOG_UUID);
 107                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DBROOT) &&
 108                     (ip->i_df.if_broot_bytes > 0)) {
 109                         ASSERT(ip->i_df.if_broot != NULL);
 110                         nvecs++;
 111                 } else {
 112                         ASSERT(!(iip->ili_format.ilf_fields &
 113                                  XFS_ILOG_DBROOT));
 114 #ifdef XFS_TRANS_DEBUG
 115                         if (iip->ili_root_size > 0) {
 116                                 ASSERT(iip->ili_root_size ==
 117                                        ip->i_df.if_broot_bytes);
 118                                 ASSERT(memcmp(iip->ili_orig_root,
 119                                             ip->i_df.if_broot,
 120                                             iip->ili_root_size) == 0);
 121                         } else {
 122                                 ASSERT(ip->i_df.if_broot_bytes == 0);
 123                         }
 124 #endif
 125                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DBROOT;
 126                 }
 127                 break;
 128
 129         case XFS_DINODE_FMT_LOCAL:
 130                 iip->ili_format.ilf_fields &=
 131                         ~(XFS_ILOG_DEXT | XFS_ILOG_DBROOT |
 132                           XFS_ILOG_DEV | XFS_ILOG_UUID);
 133                 if ((iip->ili_format.ilf_fields & XFS_ILOG_DDATA) &&
 134                     (ip->i_df.if_bytes > 0)) {
 135                         ASSERT(ip->i_df.if_u1.if_data != NULL);
 136                         ASSERT(ip->i_d.di_size > 0);
 137                         nvecs++;
 138                 } else {
 139                         iip->ili_format.ilf_fields &= ~XFS_ILOG_DDATA;
 140                 }
 141                 break;
 142
 143         case XFS_DINODE_FMT_DEV:
 144                 iip->ili_format.ilf_fields &=
 145                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 146                           XFS_ILOG_DEXT | XFS_ILOG_UUID);
 147                 break;
 148
 149         case XFS_DINODE_FMT_UUID:
 150                 iip->ili_format.ilf_fields &=
 151                         ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 152                           XFS_ILOG_DEXT | XFS_ILOG_DEV);
 153                 break;
 154
 155         default:
 156                 ASSERT(0);
 157                 break;
 158         }
 159
 160         /*
 161          * If there are no attributes associated with this file,
 162          * then there cannot be anything more to log.
 163          * Clear all attribute-related log flags.
 164          */
 165         if (!XFS_IFORK_Q(ip)) {
 166                 iip->ili_format.ilf_fields &=
 167                         ~(XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT);
 168                 return nvecs;
 169         }
 170
 171         /*
 172          * Log any necessary attribute data.
 173          */
 174         switch (ip->i_d.di_aformat) {
 175         case XFS_DINODE_FMT_EXTENTS:
 176                 iip->ili_format.ilf_fields &=
 177                         ~(XFS_ILOG_ADATA | XFS_ILOG_ABROOT);
 178                 if ((iip->ili_format.ilf_fields & XFS_ILOG_AEXT) &&
 179                     (ip->i_d.di_anextents > 0) &&
 180                     (ip->i_afp->if_bytes > 0)) {
 181                         ASSERT(ip->i_afp->if_u1.if_extents != NULL);
 182                         nvecs++;
 183                 } else {
 184                         iip->ili_format.ilf_fields &= ~XFS_ILOG_AEXT;
 185                 }
 186                 break;
 187
 188         case XFS_DINODE_FMT_BTREE:
 189                 iip->ili_format.ilf_fields &=
 190                         ~(XFS_ILOG_ADATA | XFS_ILOG_AEXT);
 191                 if ((iip->ili_format.ilf_fields & XFS_ILOG_ABROOT) &&
 192                     (ip->i_afp->if_broot_bytes > 0)) {
 193                         ASSERT(ip->i_afp->if_broot != NULL);
 194                         nvecs++;
 195                 } else {
 196                         iip->ili_format.ilf_fields &= ~XFS_ILOG_ABROOT;
 197                 }
 198                 break;
 199
 200         case XFS_DINODE_FMT_LOCAL:
 201                 iip->ili_format.ilf_fields &=
 202                         ~(XFS_ILOG_AEXT | XFS_ILOG_ABROOT);
 203                 if ((iip->ili_format.ilf_fields & XFS_ILOG_ADATA) &&
 204                     (ip->i_afp->if_bytes > 0)) {
 205                         ASSERT(ip->i_afp->if_u1.if_data != NULL);
 206                         nvecs++;
 207                 } else {
 208                         iip->ili_format.ilf_fields &= ~XFS_ILOG_ADATA;
 209                 }
 210                 break;
 211
 212         default:
 213                 ASSERT(0);
 214                 break;
 215         }
 216
 217         return nvecs;
 218 }
 219
 220 /*
 221  * This is called to fill in the vector of log iovecs for the
 222  * given inode log item.  It fills the first item with an inode
 223  * log format structure, the second with the on-disk inode structure,
 224  * and a possible third and/or fourth with the inode data/extents/b-tree
 225  * root and inode attributes data/extents/b-tree root.
 226  */
 227 STATIC void
 228 xfs_inode_item_format(
 229         xfs_inode_log_item_t    *iip,
 230         xfs_log_iovec_t         *log_vector)
 231 {
 232         uint                    nvecs;
 233         xfs_log_iovec_t         *vecp;
 234         xfs_inode_t             *ip;
 235         size_t                  data_bytes;
 236         xfs_bmbt_rec_t          *ext_buffer;
 237         int                     nrecs;
 238         xfs_mount_t             *mp;
 239
 240         ip = iip->ili_inode;
 241         vecp = log_vector;
 242
 243         vecp->i_addr = (xfs_caddr_t)&iip->ili_format;
 244         vecp->i_len  = sizeof(xfs_inode_log_format_t);
 245         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IFORMAT);
 246         vecp++;
 247         nvecs        = 1;
 248
 249         /*
 250          * Clear i_update_core if the timestamps (or any other
 251          * non-transactional modification) need flushing/logging
 252          * and we're about to log them with the rest of the core.
 253          *
 254          * This is the same logic as xfs_iflush() but this code can't
 255          * run at the same time as xfs_iflush because we're in commit
 256          * processing here and so we have the inode lock held in
 257          * exclusive mode.  Although it doesn't really matter
 258          * for the timestamps if both routines were to grab the
 259          * timestamps or not.  That would be ok.
 260          *
 261          * We clear i_update_core before copying out the data.
 262          * This is for coordination with our timestamp updates
 263          * that don't hold the inode lock. They will always
 264          * update the timestamps BEFORE setting i_update_core,
 265          * so if we clear i_update_core after they set it we
 266          * are guaranteed to see their updates to the timestamps
 267          * either here.  Likewise, if they set it after we clear it
 268          * here, we'll see it either on the next commit of this
 269          * inode or the next time the inode gets flushed via
 270          * xfs_iflush().  This depends on strongly ordered memory
 271          * semantics, but we have that.  We use the SYNCHRONIZE
 272          * macro to make sure that the compiler does not reorder
 273          * the i_update_core access below the data copy below.
 274          */
 275         if (ip->i_update_core)  {
 276                 ip->i_update_core = 0;
 277                 SYNCHRONIZE();
 278         }
 279
 280         /*
 281          * We don't have to worry about re-ordering here because
 282          * the update_size field is protected by the inode lock
 283          * and we have that held in exclusive mode.
 284          */
 285         if (ip->i_update_size)
 286                 ip->i_update_size = 0;
 287
 288         vecp->i_addr = (xfs_caddr_t)&ip->i_d;
 289         vecp->i_len  = sizeof(xfs_dinode_core_t);
 290         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_ICORE);
 291         vecp++;
 292         nvecs++;
 293         iip->ili_format.ilf_fields |= XFS_ILOG_CORE;
 294
 295         /*
 296          * If this is really an old format inode, then we need to
 297          * log it as such.  This means that we have to copy the link
 298          * count from the new field to the old.  We don't have to worry
 299          * about the new fields, because nothing trusts them as long as
 300          * the old inode version number is there.  If the superblock already
 301          * has a new version number, then we don't bother converting back.
 302          */
 303         mp = ip->i_mount;
 304         ASSERT(ip->i_d.di_version == XFS_DINODE_VERSION_1 ||
 305                XFS_SB_VERSION_HASNLINK(&mp->m_sb));
 306         if (ip->i_d.di_version == XFS_DINODE_VERSION_1) {
 307                 if (!XFS_SB_VERSION_HASNLINK(&mp->m_sb)) {
 308                         /*
 309                          * Convert it back.
 310                          */
 311                         ASSERT(ip->i_d.di_nlink <= XFS_MAXLINK_1);
 312                         ip->i_d.di_onlink = ip->i_d.di_nlink;
 313                 } else {
 314                         /*
 315                          * The superblock version has already been bumped,
 316                          * so just make the conversion to the new inode
 317                          * format permanent.
 318                          */
 319                         ip->i_d.di_version = XFS_DINODE_VERSION_2;
 320                         ip->i_d.di_onlink = 0;
 321                         memset(&(ip->i_d.di_pad[0]), 0, sizeof(ip->i_d.di_pad));
 322                 }
 323         }
 324
 325         switch (ip->i_d.di_format) {
 326         case XFS_DINODE_FMT_EXTENTS:
 327                 ASSERT(!(iip->ili_format.ilf_fields &
 328                          (XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 329                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 330                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEXT) {
 331                         ASSERT(ip->i_df.if_bytes > 0);
 332                         ASSERT(ip->i_df.if_u1.if_extents != NULL);
 333                         ASSERT(ip->i_d.di_nextents > 0);
 334                         ASSERT(iip->ili_extents_buf == NULL);
 335                         nrecs = ip->i_df.if_bytes /
 336                                 (uint)sizeof(xfs_bmbt_rec_t);
 337                         ASSERT(nrecs > 0);
 338 #ifdef XFS_NATIVE_HOST
 339                         if (nrecs == ip->i_d.di_nextents) {
 340                                 /*
 341                                  * There are no delayed allocation
 342                                  * extents, so just point to the
 343                                  * real extents array.
 344                                  */
 345                                 vecp->i_addr =
 346                                         (char *)(ip->i_df.if_u1.if_extents);
 347                                 vecp->i_len = ip->i_df.if_bytes;
 348                                 XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IEXT);
 349                         } else
 350 #endif
 351                         {
 352                                 /*
 353                                  * There are delayed allocation extents
 354                                  * in the inode, or we need to convert
 355                                  * the extents to on disk format.
 356                                  * Use xfs_iextents_copy()
 357                                  * to copy only the real extents into
 358                                  * a separate buffer.  We'll free the
 359                                  * buffer in the unlock routine.
 360                                  */
 361                                 ext_buffer = kmem_alloc(ip->i_df.if_bytes,
 362                                         KM_SLEEP);
 363                                 iip->ili_extents_buf = ext_buffer;
 364                                 vecp->i_addr = (xfs_caddr_t)ext_buffer;
 365                                 vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
 366                                                 XFS_DATA_FORK);
 367                                 XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IEXT);
 368                         }
 369                         ASSERT(vecp->i_len <= ip->i_df.if_bytes);
 370                         iip->ili_format.ilf_dsize = vecp->i_len;
 371                         vecp++;
 372                         nvecs++;
 373                 }
 374                 break;
 375
 376         case XFS_DINODE_FMT_BTREE:
 377                 ASSERT(!(iip->ili_format.ilf_fields &
 378                          (XFS_ILOG_DDATA | XFS_ILOG_DEXT |
 379                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 380                 if (iip->ili_format.ilf_fields & XFS_ILOG_DBROOT) {
 381                         ASSERT(ip->i_df.if_broot_bytes > 0);
 382                         ASSERT(ip->i_df.if_broot != NULL);
 383                         vecp->i_addr = (xfs_caddr_t)ip->i_df.if_broot;
 384                         vecp->i_len = ip->i_df.if_broot_bytes;
 385                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IBROOT);
 386                         vecp++;
 387                         nvecs++;
 388                         iip->ili_format.ilf_dsize = ip->i_df.if_broot_bytes;
 389                 }
 390                 break;
 391
 392         case XFS_DINODE_FMT_LOCAL:
 393                 ASSERT(!(iip->ili_format.ilf_fields &
 394                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 395                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 396                 if (iip->ili_format.ilf_fields & XFS_ILOG_DDATA) {
 397                         ASSERT(ip->i_df.if_bytes > 0);
 398                         ASSERT(ip->i_df.if_u1.if_data != NULL);
 399                         ASSERT(ip->i_d.di_size > 0);
 400
 401                         vecp->i_addr = (xfs_caddr_t)ip->i_df.if_u1.if_data;
 402                         /*
 403                          * Round i_bytes up to a word boundary.
 404                          * The underlying memory is guaranteed to
 405                          * to be there by xfs_idata_realloc().
 406                          */
 407                         data_bytes = roundup(ip->i_df.if_bytes, 4);
 408                         ASSERT((ip->i_df.if_real_bytes == 0) ||
 409                                (ip->i_df.if_real_bytes == data_bytes));
 410                         vecp->i_len = (int)data_bytes;
 411                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_ILOCAL);
 412                         vecp++;
 413                         nvecs++;
 414                         iip->ili_format.ilf_dsize = (unsigned)data_bytes;
 415                 }
 416                 break;
 417
 418         case XFS_DINODE_FMT_DEV:
 419                 ASSERT(!(iip->ili_format.ilf_fields &
 420                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 421                           XFS_ILOG_DDATA | XFS_ILOG_UUID)));
 422                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEV) {
 423                         iip->ili_format.ilf_u.ilfu_rdev =
 424                                 ip->i_df.if_u2.if_rdev;
 425                 }
 426                 break;
 427
 428         case XFS_DINODE_FMT_UUID:
 429                 ASSERT(!(iip->ili_format.ilf_fields &
 430                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 431                           XFS_ILOG_DDATA | XFS_ILOG_DEV)));
 432                 if (iip->ili_format.ilf_fields & XFS_ILOG_UUID) {
 433                         iip->ili_format.ilf_u.ilfu_uuid =
 434                                 ip->i_df.if_u2.if_uuid;
 435                 }
 436                 break;
 437
 438         default:
 439                 ASSERT(0);
 440                 break;
 441         }
 442
 443         /*
 444          * If there are no attributes associated with the file,
 445          * then we're done.
 446          * Assert that no attribute-related log flags are set.
 447          */
 448         if (!XFS_IFORK_Q(ip)) {
 449                 ASSERT(nvecs == iip->ili_item.li_desc->lid_size);
 450                 iip->ili_format.ilf_size = nvecs;
 451                 ASSERT(!(iip->ili_format.ilf_fields &
 452                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
 453                 return;
 454         }
 455
 456         switch (ip->i_d.di_aformat) {
 457         case XFS_DINODE_FMT_EXTENTS:
 458                 ASSERT(!(iip->ili_format.ilf_fields &
 459                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT)));
 460                 if (iip->ili_format.ilf_fields & XFS_ILOG_AEXT) {
 461                         ASSERT(ip->i_afp->if_bytes > 0);
 462                         ASSERT(ip->i_afp->if_u1.if_extents != NULL);
 463                         ASSERT(ip->i_d.di_anextents > 0);
 464 #ifdef DEBUG
 465                         nrecs = ip->i_afp->if_bytes /
 466                                 (uint)sizeof(xfs_bmbt_rec_t);
 467 #endif
 468                         ASSERT(nrecs > 0);
 469                         ASSERT(nrecs == ip->i_d.di_anextents);
 470 #ifdef XFS_NATIVE_HOST
 471                         /*
 472                          * There are not delayed allocation extents
 473                          * for attributes, so just point at the array.
 474                          */
 475                         vecp->i_addr = (char *)(ip->i_afp->if_u1.if_extents);
 476                         vecp->i_len = ip->i_afp->if_bytes;
 477 #else
 478                         ASSERT(iip->ili_aextents_buf == NULL);
 479                         /*
 480                          * Need to endian flip before logging
 481                          */
 482                         ext_buffer = kmem_alloc(ip->i_afp->if_bytes,
 483                                 KM_SLEEP);
 484                         iip->ili_aextents_buf = ext_buffer;
 485                         vecp->i_addr = (xfs_caddr_t)ext_buffer;
 486                         vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
 487                                         XFS_ATTR_FORK);
 488 #endif
 489                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IATTR_EXT);
 490                         iip->ili_format.ilf_asize = vecp->i_len;
 491                         vecp++;
 492                         nvecs++;
 493                 }
 494                 break;
 495
 496         case XFS_DINODE_FMT_BTREE:
 497                 ASSERT(!(iip->ili_format.ilf_fields &
 498                          (XFS_ILOG_ADATA | XFS_ILOG_AEXT)));
 499                 if (iip->ili_format.ilf_fields & XFS_ILOG_ABROOT) {
 500                         ASSERT(ip->i_afp->if_broot_bytes > 0);
 501                         ASSERT(ip->i_afp->if_broot != NULL);
 502                         vecp->i_addr = (xfs_caddr_t)ip->i_afp->if_broot;
 503                         vecp->i_len = ip->i_afp->if_broot_bytes;
 504                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IATTR_BROOT);
 505                         vecp++;
 506                         nvecs++;
 507                         iip->ili_format.ilf_asize = ip->i_afp->if_broot_bytes;
 508                 }
 509                 break;
 510
 511         case XFS_DINODE_FMT_LOCAL:
 512                 ASSERT(!(iip->ili_format.ilf_fields &
 513                          (XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
 514                 if (iip->ili_format.ilf_fields & XFS_ILOG_ADATA) {
 515                         ASSERT(ip->i_afp->if_bytes > 0);
 516                         ASSERT(ip->i_afp->if_u1.if_data != NULL);
 517
 518                         vecp->i_addr = (xfs_caddr_t)ip->i_afp->if_u1.if_data;
 519                         /*
 520                          * Round i_bytes up to a word boundary.
 521                          * The underlying memory is guaranteed to
 522                          * to be there by xfs_idata_realloc().
 523                          */
 524                         data_bytes = roundup(ip->i_afp->if_bytes, 4);
 525                         ASSERT((ip->i_afp->if_real_bytes == 0) ||
 526                                (ip->i_afp->if_real_bytes == data_bytes));
 527                         vecp->i_len = (int)data_bytes;
 528                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IATTR_LOCAL);
 529                         vecp++;
 530                         nvecs++;
 531                         iip->ili_format.ilf_asize = (unsigned)data_bytes;
 532                 }
 533                 break;
 534
 535         default:
 536                 ASSERT(0);
 537                 break;
 538         }
 539
 540         ASSERT(nvecs == iip->ili_item.li_desc->lid_size);
 541         iip->ili_format.ilf_size = nvecs;
 542 }
 543
 544
 545 /*
 546  * This is called to pin the inode associated with the inode log
 547  * item in memory so it cannot be written out.  Do this by calling
 548  * xfs_ipin() to bump the pin count in the inode while holding the
 549  * inode pin lock.
 550  */
 551 STATIC void
 552 xfs_inode_item_pin(
 553         xfs_inode_log_item_t    *iip)
 554 {
 555         ASSERT(ismrlocked(&(iip->ili_inode->i_lock), MR_UPDATE));
 556         xfs_ipin(iip->ili_inode);
 557 }
 558
 559
 560 /*
 561  * This is called to unpin the inode associated with the inode log
 562  * item which was previously pinned with a call to xfs_inode_item_pin().
 563  * Just call xfs_iunpin() on the inode to do this.
 564  */
 565 /* ARGSUSED */
 566 STATIC void
 567 xfs_inode_item_unpin(
 568         xfs_inode_log_item_t    *iip,
 569         int                     stale)
 570 {
 571         xfs_iunpin(iip->ili_inode);
 572 }
 573
 574 /* ARGSUSED */
 575 STATIC void
 576 xfs_inode_item_unpin_remove(
 577         xfs_inode_log_item_t    *iip,
 578         xfs_trans_t             *tp)
 579 {
 580         xfs_iunpin(iip->ili_inode);
 581 }
 582
 583 /*
 584  * This is called to attempt to lock the inode associated with this
 585  * inode log item, in preparation for the push routine which does the actual
 586  * iflush.  Don't sleep on the inode lock or the flush lock.
 587  *
 588  * If the flush lock is already held, indicating that the inode has
 589  * been or is in the process of being flushed, then (ideally) we'd like to
 590  * see if the inode's buffer is still incore, and if so give it a nudge.
 591  * We delay doing so until the pushbuf routine, though, to avoid holding
 592  * the AIL lock across a call to the blackhole which is the buffercache.
 593  * Also we don't want to sleep in any device strategy routines, which can happen
 594  * if we do the subsequent bawrite in here.
 595  */
 596 STATIC uint
 597 xfs_inode_item_trylock(
 598         xfs_inode_log_item_t    *iip)
 599 {
 600         register xfs_inode_t    *ip;
 601
 602         ip = iip->ili_inode;
 603
 604         if (xfs_ipincount(ip) > 0) {
 605                 return XFS_ITEM_PINNED;
 606         }
 607
 608         if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
 609                 return XFS_ITEM_LOCKED;
 610         }
 611
 612         if (!xfs_iflock_nowait(ip)) {
 613                 /*
 614                  * If someone else isn't already trying to push the inode
 615                  * buffer, we get to do it.
 616                  */
 617                 if (iip->ili_pushbuf_flag == 0) {
 618                         iip->ili_pushbuf_flag = 1;
 619 #ifdef DEBUG
 620                         iip->ili_push_owner = get_thread_id();
 621 #endif
 622                         /*
 623                          * Inode is left locked in shared mode.
 624                          * Pushbuf routine gets to unlock it.
 625                          */
 626                         return XFS_ITEM_PUSHBUF;
 627                 } else {
 628                         /*
 629                          * We hold the AIL_LOCK, so we must specify the
 630                          * NONOTIFY flag so that we won't double trip.
 631                          */
 632                         xfs_iunlock(ip, XFS_ILOCK_SHARED|XFS_IUNLOCK_NONOTIFY);
 633                         return XFS_ITEM_FLUSHING;
 634                 }
 635                 /* NOTREACHED */
 636         }
 637
 638         /* Stale items should force out the iclog */
 639         if (ip->i_flags & XFS_ISTALE) {
 640                 xfs_ifunlock(ip);
 641                 xfs_iunlock(ip, XFS_ILOCK_SHARED|XFS_IUNLOCK_NONOTIFY);
 642                 return XFS_ITEM_PINNED;
 643         }
 644
 645 #ifdef DEBUG
 646         if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
 647                 ASSERT(iip->ili_format.ilf_fields != 0);
 648                 ASSERT(iip->ili_logged == 0);
 649                 ASSERT(iip->ili_item.li_flags & XFS_LI_IN_AIL);
 650         }
 651 #endif
 652         return XFS_ITEM_SUCCESS;
 653 }
 654
 655 /*
 656  * Unlock the inode associated with the inode log item.
 657  * Clear the fields of the inode and inode log item that
 658  * are specific to the current transaction.  If the
 659  * hold flags is set, do not unlock the inode.
 660  */
 661 STATIC void
 662 xfs_inode_item_unlock(
 663         xfs_inode_log_item_t    *iip)
 664 {
 665         uint            hold;
 666         uint            iolocked;
 667         uint            lock_flags;
 668         xfs_inode_t     *ip;
 669
 670         ASSERT(iip != NULL);
 671         ASSERT(iip->ili_inode->i_itemp != NULL);
 672         ASSERT(ismrlocked(&(iip->ili_inode->i_lock), MR_UPDATE));
 673         ASSERT((!(iip->ili_inode->i_itemp->ili_flags &
 674                   XFS_ILI_IOLOCKED_EXCL)) ||
 675                ismrlocked(&(iip->ili_inode->i_iolock), MR_UPDATE));
 676         ASSERT((!(iip->ili_inode->i_itemp->ili_flags &
 677                   XFS_ILI_IOLOCKED_SHARED)) ||
 678                ismrlocked(&(iip->ili_inode->i_iolock), MR_ACCESS));
 679         /*
 680          * Clear the transaction pointer in the inode.
 681          */
 682         ip = iip->ili_inode;
 683         ip->i_transp = NULL;
 684
 685         /*
 686          * If the inode needed a separate buffer with which to log
 687          * its extents, then free it now.
 688          */
 689         if (iip->ili_extents_buf != NULL) {
 690                 ASSERT(ip->i_d.di_format == XFS_DINODE_FMT_EXTENTS);
 691                 ASSERT(ip->i_d.di_nextents > 0);
 692                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_DEXT);
 693                 ASSERT(ip->i_df.if_bytes > 0);
 694                 kmem_free(iip->ili_extents_buf, ip->i_df.if_bytes);
 695                 iip->ili_extents_buf = NULL;
 696         }
 697         if (iip->ili_aextents_buf != NULL) {
 698                 ASSERT(ip->i_d.di_aformat == XFS_DINODE_FMT_EXTENTS);
 699                 ASSERT(ip->i_d.di_anextents > 0);
 700                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_AEXT);
 701                 ASSERT(ip->i_afp->if_bytes > 0);
 702                 kmem_free(iip->ili_aextents_buf, ip->i_afp->if_bytes);
 703                 iip->ili_aextents_buf = NULL;
 704         }
 705
 706         /*
 707          * Figure out if we should unlock the inode or not.
 708          */
 709         hold = iip->ili_flags & XFS_ILI_HOLD;
 710
 711         /*
 712          * Before clearing out the flags, remember whether we
 713          * are holding the inode's IO lock.
 714          */
 715         iolocked = iip->ili_flags & XFS_ILI_IOLOCKED_ANY;
 716
 717         /*
 718          * Clear out the fields of the inode log item particular
 719          * to the current transaction.
 720          */
 721         iip->ili_ilock_recur = 0;
 722         iip->ili_iolock_recur = 0;
 723         iip->ili_flags = 0;
 724
 725         /*
 726          * Unlock the inode if XFS_ILI_HOLD was not set.
 727          */
 728         if (!hold) {
 729                 lock_flags = XFS_ILOCK_EXCL;
 730                 if (iolocked & XFS_ILI_IOLOCKED_EXCL) {
 731                         lock_flags |= XFS_IOLOCK_EXCL;
 732                 } else if (iolocked & XFS_ILI_IOLOCKED_SHARED) {
 733                         lock_flags |= XFS_IOLOCK_SHARED;
 734                 }
 735                 xfs_iput(iip->ili_inode, lock_flags);
 736         }
 737 }
 738
 739 /*
 740  * This is called to find out where the oldest active copy of the
 741  * inode log item in the on disk log resides now that the last log
 742  * write of it completed at the given lsn.  Since we always re-log
 743  * all dirty data in an inode, the latest copy in the on disk log
 744  * is the only one that matters.  Therefore, simply return the
 745  * given lsn.
 746  */
 747 /*ARGSUSED*/
 748 STATIC xfs_lsn_t
 749 xfs_inode_item_committed(
 750         xfs_inode_log_item_t    *iip,
 751         xfs_lsn_t               lsn)
 752 {
 753         return (lsn);
 754 }
 755
 756 /*
 757  * The transaction with the inode locked has aborted.  The inode
 758  * must not be dirty within the transaction (unless we're forcibly
 759  * shutting down).  We simply unlock just as if the transaction
 760  * had been cancelled.
 761  */
 762 STATIC void
 763 xfs_inode_item_abort(
 764         xfs_inode_log_item_t    *iip)
 765 {
 766         xfs_inode_item_unlock(iip);
 767         return;
 768 }
 769
 770
 771 /*
 772  * This gets called by xfs_trans_push_ail(), when IOP_TRYLOCK
 773  * failed to get the inode flush lock but did get the inode locked SHARED.
 774  * Here we're trying to see if the inode buffer is incore, and if so whether it's
 775  * marked delayed write. If that's the case, we'll initiate a bawrite on that
 776  * buffer to expedite the process.
 777  *
 778  * We aren't holding the AIL_LOCK (or the flush lock) when this gets called,
 779  * so it is inherently race-y.
 780  */
 781 STATIC void
 782 xfs_inode_item_pushbuf(
 783         xfs_inode_log_item_t    *iip)
 784 {
 785         xfs_inode_t     *ip;
 786         xfs_mount_t     *mp;
 787         xfs_buf_t       *bp;
 788         uint            dopush;
 789
 790         ip = iip->ili_inode;
 791
 792         ASSERT(ismrlocked(&(ip->i_lock), MR_ACCESS));
 793
 794         /*
 795          * The ili_pushbuf_flag keeps others from
 796          * trying to duplicate our effort.
 797          */
 798         ASSERT(iip->ili_pushbuf_flag != 0);
 799         ASSERT(iip->ili_push_owner == get_thread_id());
 800
 801         /*
 802          * If flushlock isn't locked anymore, chances are that the
 803          * inode flush completed and the inode was taken off the AIL.
 804          * So, just get out.
 805          */
 806         if ((valusema(&(ip->i_flock)) > 0)  ||
 807             ((iip->ili_item.li_flags & XFS_LI_IN_AIL) == 0)) {
 808                 iip->ili_pushbuf_flag = 0;
 809                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
 810                 return;
 811         }
 812
 813         mp = ip->i_mount;
 814         bp = xfs_incore(mp->m_ddev_targp, iip->ili_format.ilf_blkno,
 815                     iip->ili_format.ilf_len, XFS_INCORE_TRYLOCK);
 816
 817         if (bp != NULL) {
 818                 if (XFS_BUF_ISDELAYWRITE(bp)) {
 819                         /*
 820                          * We were racing with iflush because we don't hold
 821                          * the AIL_LOCK or the flush lock. However, at this point,
 822                          * we have the buffer, and we know that it's dirty.
 823                          * So, it's possible that iflush raced with us, and
 824                          * this item is already taken off the AIL.
 825                          * If not, we can flush it async.
 826                          */
 827                         dopush = ((iip->ili_item.li_flags & XFS_LI_IN_AIL) &&
 828                                   (valusema(&(ip->i_flock)) <= 0));
 829                         iip->ili_pushbuf_flag = 0;
 830                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 831                         xfs_buftrace("INODE ITEM PUSH", bp);
 832                         if (XFS_BUF_ISPINNED(bp)) {
 833                                 xfs_log_force(mp, (xfs_lsn_t)0,
 834                                               XFS_LOG_FORCE);
 835                         }
 836                         if (dopush) {
 837                                 xfs_bawrite(mp, bp);
 838                         } else {
 839                                 xfs_buf_relse(bp);
 840                         }
 841                 } else {
 842                         iip->ili_pushbuf_flag = 0;
 843                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 844                         xfs_buf_relse(bp);
 845                 }
 846                 return;
 847         }
 848         /*
 849          * We have to be careful about resetting pushbuf flag too early (above).
 850          * Even though in theory we can do it as soon as we have the buflock,
 851          * we don't want others to be doing work needlessly. They'll come to
 852          * this function thinking that pushing the buffer is their
 853          * responsibility only to find that the buffer is still locked by
 854          * another doing the same thing
 855          */
 856         iip->ili_pushbuf_flag = 0;
 857         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 858         return;
 859 }
 860
 861
 862 /*
 863  * This is called to asynchronously write the inode associated with this
 864  * inode log item out to disk. The inode will already have been locked by
 865  * a successful call to xfs_inode_item_trylock().
 866  */
 867 STATIC void
 868 xfs_inode_item_push(
 869         xfs_inode_log_item_t    *iip)
 870 {
 871         xfs_inode_t     *ip;
 872
 873         ip = iip->ili_inode;
 874
 875         ASSERT(ismrlocked(&(ip->i_lock), MR_ACCESS));
 876         ASSERT(valusema(&(ip->i_flock)) <= 0);
 877         /*
 878          * Since we were able to lock the inode's flush lock and
 879          * we found it on the AIL, the inode must be dirty.  This
 880          * is because the inode is removed from the AIL while still
 881          * holding the flush lock in xfs_iflush_done().  Thus, if
 882          * we found it in the AIL and were able to obtain the flush
 883          * lock without sleeping, then there must not have been
 884          * anyone in the process of flushing the inode.
 885          */
 886         ASSERT(XFS_FORCED_SHUTDOWN(ip->i_mount) ||
 887                iip->ili_format.ilf_fields != 0);
 888
 889         /*
 890          * Write out the inode.  The completion routine ('iflush_done') will
 891          * pull it from the AIL, mark it clean, unlock the flush lock.
 892          */
 893         (void) xfs_iflush(ip, XFS_IFLUSH_ASYNC);
 894         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 895
 896         return;
 897 }
 898
 899 /*
 900  * XXX rcc - this one really has to do something.  Probably needs
 901  * to stamp in a new field in the incore inode.
 902  */
 903 /* ARGSUSED */
 904 STATIC void
 905 xfs_inode_item_committing(
 906         xfs_inode_log_item_t    *iip,
 907         xfs_lsn_t               lsn)
 908 {
 909         iip->ili_last_lsn = lsn;
 910         return;
 911 }
 912
 913 /*
 914  * This is the ops vector shared by all buf log items.
 915  */
 916 STATIC struct xfs_item_ops xfs_inode_item_ops = {
 917         .iop_size       = (uint(*)(xfs_log_item_t*))xfs_inode_item_size,
 918         .iop_format     = (void(*)(xfs_log_item_t*, xfs_log_iovec_t*))
 919                                         xfs_inode_item_format,
 920         .iop_pin        = (void(*)(xfs_log_item_t*))xfs_inode_item_pin,
 921         .iop_unpin      = (void(*)(xfs_log_item_t*, int))xfs_inode_item_unpin,
 922         .iop_unpin_remove = (void(*)(xfs_log_item_t*, xfs_trans_t*))
 923                                         xfs_inode_item_unpin_remove,
 924         .iop_trylock    = (uint(*)(xfs_log_item_t*))xfs_inode_item_trylock,
 925         .iop_unlock     = (void(*)(xfs_log_item_t*))xfs_inode_item_unlock,
 926         .iop_committed  = (xfs_lsn_t(*)(xfs_log_item_t*, xfs_lsn_t))
 927                                         xfs_inode_item_committed,
 928         .iop_push       = (void(*)(xfs_log_item_t*))xfs_inode_item_push,
 929         .iop_abort      = (void(*)(xfs_log_item_t*))xfs_inode_item_abort,
 930         .iop_pushbuf    = (void(*)(xfs_log_item_t*))xfs_inode_item_pushbuf,
 931         .iop_committing = (void(*)(xfs_log_item_t*, xfs_lsn_t))
 932                                         xfs_inode_item_committing
 933 };
 934
 935
 936 /*
 937  * Initialize the inode log item for a newly allocated (in-core) inode.
 938  */
 939 void
 940 xfs_inode_item_init(
 941         xfs_inode_t     *ip,
 942         xfs_mount_t     *mp)
 943 {
 944         xfs_inode_log_item_t    *iip;
 945
 946         ASSERT(ip->i_itemp == NULL);
 947         iip = ip->i_itemp = kmem_zone_zalloc(xfs_ili_zone, KM_SLEEP);
 948
 949         iip->ili_item.li_type = XFS_LI_INODE;
 950         iip->ili_item.li_ops = &xfs_inode_item_ops;
 951         iip->ili_item.li_mountp = mp;
 952         iip->ili_inode = ip;
 953
 954         /*
 955            We have zeroed memory. No need ...
 956            iip->ili_extents_buf = NULL;
 957            iip->ili_pushbuf_flag = 0;
 958          */
 959
 960         iip->ili_format.ilf_type = XFS_LI_INODE;
 961         iip->ili_format.ilf_ino = ip->i_ino;
 962         iip->ili_format.ilf_blkno = ip->i_blkno;
 963         iip->ili_format.ilf_len = ip->i_len;
 964         iip->ili_format.ilf_boffset = ip->i_boffset;
 965 }
 966
 967 /*
 968  * Free the inode log item and any memory hanging off of it.
 969  */
 970 void
 971 xfs_inode_item_destroy(
 972         xfs_inode_t     *ip)
 973 {
 974 #ifdef XFS_TRANS_DEBUG
 975         if (ip->i_itemp->ili_root_size != 0) {
 976                 kmem_free(ip->i_itemp->ili_orig_root,
 977                           ip->i_itemp->ili_root_size);
 978         }
 979 #endif
 980         kmem_zone_free(xfs_ili_zone, ip->i_itemp);
 981 }
 982
 983
 984 /*
 985  * This is the inode flushing I/O completion routine.  It is called
 986  * from interrupt level when the buffer containing the inode is
 987  * flushed to disk.  It is responsible for removing the inode item
 988  * from the AIL if it has not been re-logged, and unlocking the inode's
 989  * flush lock.
 990  */
 991 /*ARGSUSED*/
 992 void
 993 xfs_iflush_done(
 994         xfs_buf_t               *bp,
 995         xfs_inode_log_item_t    *iip)
 996 {
 997         xfs_inode_t     *ip;
 998         SPLDECL(s);
 999
1000         ip = iip->ili_inode;
1001
1002         /*
1003          * We only want to pull the item from the AIL if it is
1004          * actually there and its location in the log has not
1005          * changed since we started the flush.  Thus, we only bother
1006          * if the ili_logged flag is set and the inode's lsn has not
1007          * changed.  First we check the lsn outside
1008          * the lock since it's cheaper, and then we recheck while
1009          * holding the lock before removing the inode from the AIL.
1010          */
1011         if (iip->ili_logged &&
1012             (iip->ili_item.li_lsn == iip->ili_flush_lsn)) {
1013                 AIL_LOCK(ip->i_mount, s);
1014                 if (iip->ili_item.li_lsn == iip->ili_flush_lsn) {
1015                         /*
1016                          * xfs_trans_delete_ail() drops the AIL lock.
1017                          */
1018                         xfs_trans_delete_ail(ip->i_mount,
1019                                              (xfs_log_item_t*)iip, s);
1020                 } else {
1021                         AIL_UNLOCK(ip->i_mount, s);
1022                 }
1023         }
1024
1025         iip->ili_logged = 0;
1026
1027         /*
1028          * Clear the ili_last_fields bits now that we know that the
1029          * data corresponding to them is safely on disk.
1030          */
1031         iip->ili_last_fields = 0;
1032
1033         /*
1034          * Release the inode's flush lock since we're done with it.
1035          */
1036         xfs_ifunlock(ip);
1037
1038         return;
1039 }
1040
1041 /*
1042  * This is the inode flushing abort routine.  It is called
1043  * from xfs_iflush when the filesystem is shutting down to clean
1044  * up the inode state.
1045  * It is responsible for removing the inode item
1046  * from the AIL if it has not been re-logged, and unlocking the inode's
1047  * flush lock.
1048  */
1049 void
1050 xfs_iflush_abort(
1051         xfs_inode_t             *ip)
1052 {
1053         xfs_inode_log_item_t    *iip;
1054         xfs_mount_t             *mp;
1055         SPLDECL(s);
1056
1057         iip = ip->i_itemp;
1058         mp = ip->i_mount;
1059         if (iip) {
1060                 if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
1061                         AIL_LOCK(mp, s);
1062                         if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
1063                                 /*
1064                                  * xfs_trans_delete_ail() drops the AIL lock.
1065                                  */
1066                                 xfs_trans_delete_ail(mp, (xfs_log_item_t *)iip,
1067                                         s);
1068                         } else
1069                                 AIL_UNLOCK(mp, s);
1070                 }
1071                 iip->ili_logged = 0;
1072                 /*
1073                  * Clear the ili_last_fields bits now that we know that the
1074                  * data corresponding to them is safely on disk.
1075                  */
1076                 iip->ili_last_fields = 0;
1077                 /*
1078                  * Clear the inode logging fields so no more flushes are
1079                  * attempted.
1080                  */
1081                 iip->ili_format.ilf_fields = 0;
1082         }
1083         /*
1084          * Release the inode's flush lock since we're done with it.
1085          */
1086         xfs_ifunlock(ip);
1087 }
1088
1089 void
1090 xfs_istale_done(
1091         xfs_buf_t               *bp,
1092         xfs_inode_log_item_t    *iip)
1093 {
1094         xfs_iflush_abort(iip->ili_inode);
1095 }