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          * Clear i_update_core if the timestamps (or any other
 236          * non-transactional modification) need flushing/logging
 237          * and we're about to log them with the rest of the core.
 238          *
 239          * This is the same logic as xfs_iflush() but this code can't
 240          * run at the same time as xfs_iflush because we're in commit
 241          * processing here and so we have the inode lock held in
 242          * exclusive mode.  Although it doesn't really matter
 243          * for the timestamps if both routines were to grab the
 244          * timestamps or not.  That would be ok.
 245          *
 246          * We clear i_update_core before copying out the data.
 247          * This is for coordination with our timestamp updates
 248          * that don't hold the inode lock. They will always
 249          * update the timestamps BEFORE setting i_update_core,
 250          * so if we clear i_update_core after they set it we
 251          * are guaranteed to see their updates to the timestamps
 252          * either here.  Likewise, if they set it after we clear it
 253          * here, we'll see it either on the next commit of this
 254          * inode or the next time the inode gets flushed via
 255          * xfs_iflush().  This depends on strongly ordered memory
 256          * semantics, but we have that.  We use the SYNCHRONIZE
 257          * macro to make sure that the compiler does not reorder
 258          * the i_update_core access below the data copy below.
 259          */
 260         if (ip->i_update_core)  {
 261                 ip->i_update_core = 0;
 262                 SYNCHRONIZE();
 263         }
 264
 265         /*
 266          * We don't have to worry about re-ordering here because
 267          * the update_size field is protected by the inode lock
 268          * and we have that held in exclusive mode.
 269          */
 270         if (ip->i_update_size)
 271                 ip->i_update_size = 0;
 272
 273         /*
 274          * Make sure to get the latest atime from the Linux inode.
 275          */
 276         xfs_synchronize_atime(ip);
 277
 278         /*
 279          * make sure the linux inode is dirty
 280          */
 281         xfs_mark_inode_dirty_sync(ip);
 282
 283         vecp->i_addr = (xfs_caddr_t)&ip->i_d;
 284         vecp->i_len  = sizeof(struct xfs_icdinode);
 285         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_ICORE);
 286         vecp++;
 287         nvecs++;
 288         iip->ili_format.ilf_fields |= XFS_ILOG_CORE;
 289
 290         /*
 291          * If this is really an old format inode, then we need to
 292          * log it as such.  This means that we have to copy the link
 293          * count from the new field to the old.  We don't have to worry
 294          * about the new fields, because nothing trusts them as long as
 295          * the old inode version number is there.  If the superblock already
 296          * has a new version number, then we don't bother converting back.
 297          */
 298         mp = ip->i_mount;
 299         ASSERT(ip->i_d.di_version == 1 || xfs_sb_version_hasnlink(&mp->m_sb));
 300         if (ip->i_d.di_version == 1) {
 301                 if (!xfs_sb_version_hasnlink(&mp->m_sb)) {
 302                         /*
 303                          * Convert it back.
 304                          */
 305                         ASSERT(ip->i_d.di_nlink <= XFS_MAXLINK_1);
 306                         ip->i_d.di_onlink = ip->i_d.di_nlink;
 307                 } else {
 308                         /*
 309                          * The superblock version has already been bumped,
 310                          * so just make the conversion to the new inode
 311                          * format permanent.
 312                          */
 313                         ip->i_d.di_version = 2;
 314                         ip->i_d.di_onlink = 0;
 315                         memset(&(ip->i_d.di_pad[0]), 0, sizeof(ip->i_d.di_pad));
 316                 }
 317         }
 318
 319         switch (ip->i_d.di_format) {
 320         case XFS_DINODE_FMT_EXTENTS:
 321                 ASSERT(!(iip->ili_format.ilf_fields &
 322                          (XFS_ILOG_DDATA | XFS_ILOG_DBROOT |
 323                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 324                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEXT) {
 325                         ASSERT(ip->i_df.if_bytes > 0);
 326                         ASSERT(ip->i_df.if_u1.if_extents != NULL);
 327                         ASSERT(ip->i_d.di_nextents > 0);
 328                         ASSERT(iip->ili_extents_buf == NULL);
 329                         nrecs = ip->i_df.if_bytes /
 330                                 (uint)sizeof(xfs_bmbt_rec_t);
 331                         ASSERT(nrecs > 0);
 332 #ifdef XFS_NATIVE_HOST
 333                         if (nrecs == ip->i_d.di_nextents) {
 334                                 /*
 335                                  * There are no delayed allocation
 336                                  * extents, so just point to the
 337                                  * real extents array.
 338                                  */
 339                                 vecp->i_addr =
 340                                         (char *)(ip->i_df.if_u1.if_extents);
 341                                 vecp->i_len = ip->i_df.if_bytes;
 342                                 XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IEXT);
 343                         } else
 344 #endif
 345                         {
 346                                 /*
 347                                  * There are delayed allocation extents
 348                                  * in the inode, or we need to convert
 349                                  * the extents to on disk format.
 350                                  * Use xfs_iextents_copy()
 351                                  * to copy only the real extents into
 352                                  * a separate buffer.  We'll free the
 353                                  * buffer in the unlock routine.
 354                                  */
 355                                 ext_buffer = kmem_alloc(ip->i_df.if_bytes,
 356                                         KM_SLEEP);
 357                                 iip->ili_extents_buf = ext_buffer;
 358                                 vecp->i_addr = (xfs_caddr_t)ext_buffer;
 359                                 vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
 360                                                 XFS_DATA_FORK);
 361                                 XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IEXT);
 362                         }
 363                         ASSERT(vecp->i_len <= ip->i_df.if_bytes);
 364                         iip->ili_format.ilf_dsize = vecp->i_len;
 365                         vecp++;
 366                         nvecs++;
 367                 }
 368                 break;
 369
 370         case XFS_DINODE_FMT_BTREE:
 371                 ASSERT(!(iip->ili_format.ilf_fields &
 372                          (XFS_ILOG_DDATA | XFS_ILOG_DEXT |
 373                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 374                 if (iip->ili_format.ilf_fields & XFS_ILOG_DBROOT) {
 375                         ASSERT(ip->i_df.if_broot_bytes > 0);
 376                         ASSERT(ip->i_df.if_broot != NULL);
 377                         vecp->i_addr = (xfs_caddr_t)ip->i_df.if_broot;
 378                         vecp->i_len = ip->i_df.if_broot_bytes;
 379                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IBROOT);
 380                         vecp++;
 381                         nvecs++;
 382                         iip->ili_format.ilf_dsize = ip->i_df.if_broot_bytes;
 383                 }
 384                 break;
 385
 386         case XFS_DINODE_FMT_LOCAL:
 387                 ASSERT(!(iip->ili_format.ilf_fields &
 388                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 389                           XFS_ILOG_DEV | XFS_ILOG_UUID)));
 390                 if (iip->ili_format.ilf_fields & XFS_ILOG_DDATA) {
 391                         ASSERT(ip->i_df.if_bytes > 0);
 392                         ASSERT(ip->i_df.if_u1.if_data != NULL);
 393                         ASSERT(ip->i_d.di_size > 0);
 394
 395                         vecp->i_addr = (xfs_caddr_t)ip->i_df.if_u1.if_data;
 396                         /*
 397                          * Round i_bytes up to a word boundary.
 398                          * The underlying memory is guaranteed to
 399                          * to be there by xfs_idata_realloc().
 400                          */
 401                         data_bytes = roundup(ip->i_df.if_bytes, 4);
 402                         ASSERT((ip->i_df.if_real_bytes == 0) ||
 403                                (ip->i_df.if_real_bytes == data_bytes));
 404                         vecp->i_len = (int)data_bytes;
 405                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_ILOCAL);
 406                         vecp++;
 407                         nvecs++;
 408                         iip->ili_format.ilf_dsize = (unsigned)data_bytes;
 409                 }
 410                 break;
 411
 412         case XFS_DINODE_FMT_DEV:
 413                 ASSERT(!(iip->ili_format.ilf_fields &
 414                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 415                           XFS_ILOG_DDATA | XFS_ILOG_UUID)));
 416                 if (iip->ili_format.ilf_fields & XFS_ILOG_DEV) {
 417                         iip->ili_format.ilf_u.ilfu_rdev =
 418                                 ip->i_df.if_u2.if_rdev;
 419                 }
 420                 break;
 421
 422         case XFS_DINODE_FMT_UUID:
 423                 ASSERT(!(iip->ili_format.ilf_fields &
 424                          (XFS_ILOG_DBROOT | XFS_ILOG_DEXT |
 425                           XFS_ILOG_DDATA | XFS_ILOG_DEV)));
 426                 if (iip->ili_format.ilf_fields & XFS_ILOG_UUID) {
 427                         iip->ili_format.ilf_u.ilfu_uuid =
 428                                 ip->i_df.if_u2.if_uuid;
 429                 }
 430                 break;
 431
 432         default:
 433                 ASSERT(0);
 434                 break;
 435         }
 436
 437         /*
 438          * If there are no attributes associated with the file,
 439          * then we're done.
 440          * Assert that no attribute-related log flags are set.
 441          */
 442         if (!XFS_IFORK_Q(ip)) {
 443                 ASSERT(nvecs == iip->ili_item.li_desc->lid_size);
 444                 iip->ili_format.ilf_size = nvecs;
 445                 ASSERT(!(iip->ili_format.ilf_fields &
 446                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
 447                 return;
 448         }
 449
 450         switch (ip->i_d.di_aformat) {
 451         case XFS_DINODE_FMT_EXTENTS:
 452                 ASSERT(!(iip->ili_format.ilf_fields &
 453                          (XFS_ILOG_ADATA | XFS_ILOG_ABROOT)));
 454                 if (iip->ili_format.ilf_fields & XFS_ILOG_AEXT) {
 455                         ASSERT(ip->i_afp->if_bytes > 0);
 456                         ASSERT(ip->i_afp->if_u1.if_extents != NULL);
 457                         ASSERT(ip->i_d.di_anextents > 0);
 458 #ifdef DEBUG
 459                         nrecs = ip->i_afp->if_bytes /
 460                                 (uint)sizeof(xfs_bmbt_rec_t);
 461 #endif
 462                         ASSERT(nrecs > 0);
 463                         ASSERT(nrecs == ip->i_d.di_anextents);
 464 #ifdef XFS_NATIVE_HOST
 465                         /*
 466                          * There are not delayed allocation extents
 467                          * for attributes, so just point at the array.
 468                          */
 469                         vecp->i_addr = (char *)(ip->i_afp->if_u1.if_extents);
 470                         vecp->i_len = ip->i_afp->if_bytes;
 471 #else
 472                         ASSERT(iip->ili_aextents_buf == NULL);
 473                         /*
 474                          * Need to endian flip before logging
 475                          */
 476                         ext_buffer = kmem_alloc(ip->i_afp->if_bytes,
 477                                 KM_SLEEP);
 478                         iip->ili_aextents_buf = ext_buffer;
 479                         vecp->i_addr = (xfs_caddr_t)ext_buffer;
 480                         vecp->i_len = xfs_iextents_copy(ip, ext_buffer,
 481                                         XFS_ATTR_FORK);
 482 #endif
 483                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IATTR_EXT);
 484                         iip->ili_format.ilf_asize = vecp->i_len;
 485                         vecp++;
 486                         nvecs++;
 487                 }
 488                 break;
 489
 490         case XFS_DINODE_FMT_BTREE:
 491                 ASSERT(!(iip->ili_format.ilf_fields &
 492                          (XFS_ILOG_ADATA | XFS_ILOG_AEXT)));
 493                 if (iip->ili_format.ilf_fields & XFS_ILOG_ABROOT) {
 494                         ASSERT(ip->i_afp->if_broot_bytes > 0);
 495                         ASSERT(ip->i_afp->if_broot != NULL);
 496                         vecp->i_addr = (xfs_caddr_t)ip->i_afp->if_broot;
 497                         vecp->i_len = ip->i_afp->if_broot_bytes;
 498                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IATTR_BROOT);
 499                         vecp++;
 500                         nvecs++;
 501                         iip->ili_format.ilf_asize = ip->i_afp->if_broot_bytes;
 502                 }
 503                 break;
 504
 505         case XFS_DINODE_FMT_LOCAL:
 506                 ASSERT(!(iip->ili_format.ilf_fields &
 507                          (XFS_ILOG_ABROOT | XFS_ILOG_AEXT)));
 508                 if (iip->ili_format.ilf_fields & XFS_ILOG_ADATA) {
 509                         ASSERT(ip->i_afp->if_bytes > 0);
 510                         ASSERT(ip->i_afp->if_u1.if_data != NULL);
 511
 512                         vecp->i_addr = (xfs_caddr_t)ip->i_afp->if_u1.if_data;
 513                         /*
 514                          * Round i_bytes up to a word boundary.
 515                          * The underlying memory is guaranteed to
 516                          * to be there by xfs_idata_realloc().
 517                          */
 518                         data_bytes = roundup(ip->i_afp->if_bytes, 4);
 519                         ASSERT((ip->i_afp->if_real_bytes == 0) ||
 520                                (ip->i_afp->if_real_bytes == data_bytes));
 521                         vecp->i_len = (int)data_bytes;
 522                         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_IATTR_LOCAL);
 523                         vecp++;
 524                         nvecs++;
 525                         iip->ili_format.ilf_asize = (unsigned)data_bytes;
 526                 }
 527                 break;
 528
 529         default:
 530                 ASSERT(0);
 531                 break;
 532         }
 533
 534         ASSERT(nvecs == iip->ili_item.li_desc->lid_size);
 535         iip->ili_format.ilf_size = nvecs;
 536 }
 537
 538
 539 /*
 540  * This is called to pin the inode associated with the inode log
 541  * item in memory so it cannot be written out.  Do this by calling
 542  * xfs_ipin() to bump the pin count in the inode while holding the
 543  * inode pin lock.
 544  */
 545 STATIC void
 546 xfs_inode_item_pin(
 547         xfs_inode_log_item_t    *iip)
 548 {
 549         ASSERT(xfs_isilocked(iip->ili_inode, XFS_ILOCK_EXCL));
 550         xfs_ipin(iip->ili_inode);
 551 }
 552
 553
 554 /*
 555  * This is called to unpin the inode associated with the inode log
 556  * item which was previously pinned with a call to xfs_inode_item_pin().
 557  * Just call xfs_iunpin() on the inode to do this.
 558  */
 559 /* ARGSUSED */
 560 STATIC void
 561 xfs_inode_item_unpin(
 562         xfs_inode_log_item_t    *iip,
 563         int                     stale)
 564 {
 565         xfs_iunpin(iip->ili_inode);
 566 }
 567
 568 /* ARGSUSED */
 569 STATIC void
 570 xfs_inode_item_unpin_remove(
 571         xfs_inode_log_item_t    *iip,
 572         xfs_trans_t             *tp)
 573 {
 574         xfs_iunpin(iip->ili_inode);
 575 }
 576
 577 /*
 578  * This is called to attempt to lock the inode associated with this
 579  * inode log item, in preparation for the push routine which does the actual
 580  * iflush.  Don't sleep on the inode lock or the flush lock.
 581  *
 582  * If the flush lock is already held, indicating that the inode has
 583  * been or is in the process of being flushed, then (ideally) we'd like to
 584  * see if the inode's buffer is still incore, and if so give it a nudge.
 585  * We delay doing so until the pushbuf routine, though, to avoid holding
 586  * the AIL lock across a call to the blackhole which is the buffer cache.
 587  * Also we don't want to sleep in any device strategy routines, which can happen
 588  * if we do the subsequent bawrite in here.
 589  */
 590 STATIC uint
 591 xfs_inode_item_trylock(
 592         xfs_inode_log_item_t    *iip)
 593 {
 594         register xfs_inode_t    *ip;
 595
 596         ip = iip->ili_inode;
 597
 598         if (xfs_ipincount(ip) > 0) {
 599                 return XFS_ITEM_PINNED;
 600         }
 601
 602         if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
 603                 return XFS_ITEM_LOCKED;
 604         }
 605
 606         if (!xfs_iflock_nowait(ip)) {
 607                 /*
 608                  * If someone else isn't already trying to push the inode
 609                  * buffer, we get to do it.
 610                  */
 611                 if (iip->ili_pushbuf_flag == 0) {
 612                         iip->ili_pushbuf_flag = 1;
 613 #ifdef DEBUG
 614                         iip->ili_push_owner = current_pid();
 615 #endif
 616                         /*
 617                          * Inode is left locked in shared mode.
 618                          * Pushbuf routine gets to unlock it.
 619                          */
 620                         return XFS_ITEM_PUSHBUF;
 621                 } else {
 622                         /*
 623                          * We hold the AIL lock, so we must specify the
 624                          * NONOTIFY flag so that we won't double trip.
 625                          */
 626                         xfs_iunlock(ip, XFS_ILOCK_SHARED|XFS_IUNLOCK_NONOTIFY);
 627                         return XFS_ITEM_FLUSHING;
 628                 }
 629                 /* NOTREACHED */
 630         }
 631
 632         /* Stale items should force out the iclog */
 633         if (ip->i_flags & XFS_ISTALE) {
 634                 xfs_ifunlock(ip);
 635                 xfs_iunlock(ip, XFS_ILOCK_SHARED|XFS_IUNLOCK_NONOTIFY);
 636                 return XFS_ITEM_PINNED;
 637         }
 638
 639 #ifdef DEBUG
 640         if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
 641                 ASSERT(iip->ili_format.ilf_fields != 0);
 642                 ASSERT(iip->ili_logged == 0);
 643                 ASSERT(iip->ili_item.li_flags & XFS_LI_IN_AIL);
 644         }
 645 #endif
 646         return XFS_ITEM_SUCCESS;
 647 }
 648
 649 /*
 650  * Unlock the inode associated with the inode log item.
 651  * Clear the fields of the inode and inode log item that
 652  * are specific to the current transaction.  If the
 653  * hold flags is set, do not unlock the inode.
 654  */
 655 STATIC void
 656 xfs_inode_item_unlock(
 657         xfs_inode_log_item_t    *iip)
 658 {
 659         uint            hold;
 660         uint            iolocked;
 661         uint            lock_flags;
 662         xfs_inode_t     *ip;
 663
 664         ASSERT(iip != NULL);
 665         ASSERT(iip->ili_inode->i_itemp != NULL);
 666         ASSERT(xfs_isilocked(iip->ili_inode, XFS_ILOCK_EXCL));
 667         ASSERT((!(iip->ili_inode->i_itemp->ili_flags &
 668                   XFS_ILI_IOLOCKED_EXCL)) ||
 669                xfs_isilocked(iip->ili_inode, XFS_IOLOCK_EXCL));
 670         ASSERT((!(iip->ili_inode->i_itemp->ili_flags &
 671                   XFS_ILI_IOLOCKED_SHARED)) ||
 672                xfs_isilocked(iip->ili_inode, XFS_IOLOCK_SHARED));
 673         /*
 674          * Clear the transaction pointer in the inode.
 675          */
 676         ip = iip->ili_inode;
 677         ip->i_transp = NULL;
 678
 679         /*
 680          * If the inode needed a separate buffer with which to log
 681          * its extents, then free it now.
 682          */
 683         if (iip->ili_extents_buf != NULL) {
 684                 ASSERT(ip->i_d.di_format == XFS_DINODE_FMT_EXTENTS);
 685                 ASSERT(ip->i_d.di_nextents > 0);
 686                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_DEXT);
 687                 ASSERT(ip->i_df.if_bytes > 0);
 688                 kmem_free(iip->ili_extents_buf);
 689                 iip->ili_extents_buf = NULL;
 690         }
 691         if (iip->ili_aextents_buf != NULL) {
 692                 ASSERT(ip->i_d.di_aformat == XFS_DINODE_FMT_EXTENTS);
 693                 ASSERT(ip->i_d.di_anextents > 0);
 694                 ASSERT(iip->ili_format.ilf_fields & XFS_ILOG_AEXT);
 695                 ASSERT(ip->i_afp->if_bytes > 0);
 696                 kmem_free(iip->ili_aextents_buf);
 697                 iip->ili_aextents_buf = NULL;
 698         }
 699
 700         /*
 701          * Figure out if we should unlock the inode or not.
 702          */
 703         hold = iip->ili_flags & XFS_ILI_HOLD;
 704
 705         /*
 706          * Before clearing out the flags, remember whether we
 707          * are holding the inode's IO lock.
 708          */
 709         iolocked = iip->ili_flags & XFS_ILI_IOLOCKED_ANY;
 710
 711         /*
 712          * Clear out the fields of the inode log item particular
 713          * to the current transaction.
 714          */
 715         iip->ili_ilock_recur = 0;
 716         iip->ili_iolock_recur = 0;
 717         iip->ili_flags = 0;
 718
 719         /*
 720          * Unlock the inode if XFS_ILI_HOLD was not set.
 721          */
 722         if (!hold) {
 723                 lock_flags = XFS_ILOCK_EXCL;
 724                 if (iolocked & XFS_ILI_IOLOCKED_EXCL) {
 725                         lock_flags |= XFS_IOLOCK_EXCL;
 726                 } else if (iolocked & XFS_ILI_IOLOCKED_SHARED) {
 727                         lock_flags |= XFS_IOLOCK_SHARED;
 728                 }
 729                 xfs_iput(iip->ili_inode, lock_flags);
 730         }
 731 }
 732
 733 /*
 734  * This is called to find out where the oldest active copy of the
 735  * inode log item in the on disk log resides now that the last log
 736  * write of it completed at the given lsn.  Since we always re-log
 737  * all dirty data in an inode, the latest copy in the on disk log
 738  * is the only one that matters.  Therefore, simply return the
 739  * given lsn.
 740  */
 741 /*ARGSUSED*/
 742 STATIC xfs_lsn_t
 743 xfs_inode_item_committed(
 744         xfs_inode_log_item_t    *iip,
 745         xfs_lsn_t               lsn)
 746 {
 747         return (lsn);
 748 }
 749
 750 /*
 751  * This gets called by xfs_trans_push_ail(), when IOP_TRYLOCK
 752  * failed to get the inode flush lock but did get the inode locked SHARED.
 753  * Here we're trying to see if the inode buffer is incore, and if so whether it's
 754  * marked delayed write. If that's the case, we'll initiate a bawrite on that
 755  * buffer to expedite the process.
 756  *
 757  * We aren't holding the AIL lock (or the flush lock) when this gets called,
 758  * so it is inherently race-y.
 759  */
 760 STATIC void
 761 xfs_inode_item_pushbuf(
 762         xfs_inode_log_item_t    *iip)
 763 {
 764         xfs_inode_t     *ip;
 765         xfs_mount_t     *mp;
 766         xfs_buf_t       *bp;
 767         uint            dopush;
 768
 769         ip = iip->ili_inode;
 770
 771         ASSERT(xfs_isilocked(ip, XFS_ILOCK_SHARED));
 772
 773         /*
 774          * The ili_pushbuf_flag keeps others from
 775          * trying to duplicate our effort.
 776          */
 777         ASSERT(iip->ili_pushbuf_flag != 0);
 778         ASSERT(iip->ili_push_owner == current_pid());
 779
 780         /*
 781          * If a flush is not in progress anymore, chances are that the
 782          * inode was taken off the AIL. So, just get out.
 783          */
 784         if (completion_done(&ip->i_flush) ||
 785             ((iip->ili_item.li_flags & XFS_LI_IN_AIL) == 0)) {
 786                 iip->ili_pushbuf_flag = 0;
 787                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
 788                 return;
 789         }
 790
 791         mp = ip->i_mount;
 792         bp = xfs_incore(mp->m_ddev_targp, iip->ili_format.ilf_blkno,
 793                     iip->ili_format.ilf_len, XFS_INCORE_TRYLOCK);
 794
 795         if (bp != NULL) {
 796                 if (XFS_BUF_ISDELAYWRITE(bp)) {
 797                         /*
 798                          * We were racing with iflush because we don't hold
 799                          * the AIL lock or the flush lock. However, at this point,
 800                          * we have the buffer, and we know that it's dirty.
 801                          * So, it's possible that iflush raced with us, and
 802                          * this item is already taken off the AIL.
 803                          * If not, we can flush it async.
 804                          */
 805                         dopush = ((iip->ili_item.li_flags & XFS_LI_IN_AIL) &&
 806                                   !completion_done(&ip->i_flush));
 807                         iip->ili_pushbuf_flag = 0;
 808                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 809                         xfs_buftrace("INODE ITEM PUSH", bp);
 810                         if (XFS_BUF_ISPINNED(bp)) {
 811                                 xfs_log_force(mp, (xfs_lsn_t)0,
 812                                               XFS_LOG_FORCE);
 813                         }
 814                         if (dopush) {
 815                                 int     error;
 816                                 error = xfs_bawrite(mp, bp);
 817                                 if (error)
 818                                         xfs_fs_cmn_err(CE_WARN, mp,
 819                 "xfs_inode_item_pushbuf: pushbuf error %d on iip %p, bp %p",
 820                                                         error, iip, bp);
 821                         } else {
 822                                 xfs_buf_relse(bp);
 823                         }
 824                 } else {
 825                         iip->ili_pushbuf_flag = 0;
 826                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 827                         xfs_buf_relse(bp);
 828                 }
 829                 return;
 830         }
 831         /*
 832          * We have to be careful about resetting pushbuf flag too early (above).
 833          * Even though in theory we can do it as soon as we have the buflock,
 834          * we don't want others to be doing work needlessly. They'll come to
 835          * this function thinking that pushing the buffer is their
 836          * responsibility only to find that the buffer is still locked by
 837          * another doing the same thing
 838          */
 839         iip->ili_pushbuf_flag = 0;
 840         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 841         return;
 842 }
 843
 844
 845 /*
 846  * This is called to asynchronously write the inode associated with this
 847  * inode log item out to disk. The inode will already have been locked by
 848  * a successful call to xfs_inode_item_trylock().
 849  */
 850 STATIC void
 851 xfs_inode_item_push(
 852         xfs_inode_log_item_t    *iip)
 853 {
 854         xfs_inode_t     *ip;
 855
 856         ip = iip->ili_inode;
 857
 858         ASSERT(xfs_isilocked(ip, XFS_ILOCK_SHARED));
 859         ASSERT(!completion_done(&ip->i_flush));
 860         /*
 861          * Since we were able to lock the inode's flush lock and
 862          * we found it on the AIL, the inode must be dirty.  This
 863          * is because the inode is removed from the AIL while still
 864          * holding the flush lock in xfs_iflush_done().  Thus, if
 865          * we found it in the AIL and were able to obtain the flush
 866          * lock without sleeping, then there must not have been
 867          * anyone in the process of flushing the inode.
 868          */
 869         ASSERT(XFS_FORCED_SHUTDOWN(ip->i_mount) ||
 870                iip->ili_format.ilf_fields != 0);
 871
 872         /*
 873          * Write out the inode.  The completion routine ('iflush_done') will
 874          * pull it from the AIL, mark it clean, unlock the flush lock.
 875          */
 876         (void) xfs_iflush(ip, XFS_IFLUSH_ASYNC);
 877         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 878
 879         return;
 880 }
 881
 882 /*
 883  * XXX rcc - this one really has to do something.  Probably needs
 884  * to stamp in a new field in the incore inode.
 885  */
 886 /* ARGSUSED */
 887 STATIC void
 888 xfs_inode_item_committing(
 889         xfs_inode_log_item_t    *iip,
 890         xfs_lsn_t               lsn)
 891 {
 892         iip->ili_last_lsn = lsn;
 893         return;
 894 }
 895
 896 /*
 897  * This is the ops vector shared by all buf log items.
 898  */
 899 static struct xfs_item_ops xfs_inode_item_ops = {
 900         .iop_size       = (uint(*)(xfs_log_item_t*))xfs_inode_item_size,
 901         .iop_format     = (void(*)(xfs_log_item_t*, xfs_log_iovec_t*))
 902                                         xfs_inode_item_format,
 903         .iop_pin        = (void(*)(xfs_log_item_t*))xfs_inode_item_pin,
 904         .iop_unpin      = (void(*)(xfs_log_item_t*, int))xfs_inode_item_unpin,
 905         .iop_unpin_remove = (void(*)(xfs_log_item_t*, xfs_trans_t*))
 906                                         xfs_inode_item_unpin_remove,
 907         .iop_trylock    = (uint(*)(xfs_log_item_t*))xfs_inode_item_trylock,
 908         .iop_unlock     = (void(*)(xfs_log_item_t*))xfs_inode_item_unlock,
 909         .iop_committed  = (xfs_lsn_t(*)(xfs_log_item_t*, xfs_lsn_t))
 910                                         xfs_inode_item_committed,
 911         .iop_push       = (void(*)(xfs_log_item_t*))xfs_inode_item_push,
 912         .iop_pushbuf    = (void(*)(xfs_log_item_t*))xfs_inode_item_pushbuf,
 913         .iop_committing = (void(*)(xfs_log_item_t*, xfs_lsn_t))
 914                                         xfs_inode_item_committing
 915 };
 916
 917
 918 /*
 919  * Initialize the inode log item for a newly allocated (in-core) inode.
 920  */
 921 void
 922 xfs_inode_item_init(
 923         xfs_inode_t     *ip,
 924         xfs_mount_t     *mp)
 925 {
 926         xfs_inode_log_item_t    *iip;
 927
 928         ASSERT(ip->i_itemp == NULL);
 929         iip = ip->i_itemp = kmem_zone_zalloc(xfs_ili_zone, KM_SLEEP);
 930
 931         iip->ili_item.li_type = XFS_LI_INODE;
 932         iip->ili_item.li_ops = &xfs_inode_item_ops;
 933         iip->ili_item.li_mountp = mp;
 934         iip->ili_item.li_ailp = mp->m_ail;
 935         iip->ili_inode = ip;
 936
 937         /*
 938            We have zeroed memory. No need ...
 939            iip->ili_extents_buf = NULL;
 940            iip->ili_pushbuf_flag = 0;
 941          */
 942
 943         iip->ili_format.ilf_type = XFS_LI_INODE;
 944         iip->ili_format.ilf_ino = ip->i_ino;
 945         iip->ili_format.ilf_blkno = ip->i_imap.im_blkno;
 946         iip->ili_format.ilf_len = ip->i_imap.im_len;
 947         iip->ili_format.ilf_boffset = ip->i_imap.im_boffset;
 948 }
 949
 950 /*
 951  * Free the inode log item and any memory hanging off of it.
 952  */
 953 void
 954 xfs_inode_item_destroy(
 955         xfs_inode_t     *ip)
 956 {
 957 #ifdef XFS_TRANS_DEBUG
 958         if (ip->i_itemp->ili_root_size != 0) {
 959                 kmem_free(ip->i_itemp->ili_orig_root);
 960         }
 961 #endif
 962         kmem_zone_free(xfs_ili_zone, ip->i_itemp);
 963 }
 964
 965
 966 /*
 967  * This is the inode flushing I/O completion routine.  It is called
 968  * from interrupt level when the buffer containing the inode is
 969  * flushed to disk.  It is responsible for removing the inode item
 970  * from the AIL if it has not been re-logged, and unlocking the inode's
 971  * flush lock.
 972  */
 973 /*ARGSUSED*/
 974 void
 975 xfs_iflush_done(
 976         xfs_buf_t               *bp,
 977         xfs_inode_log_item_t    *iip)
 978 {
 979         xfs_inode_t             *ip = iip->ili_inode;
 980         struct xfs_ail          *ailp = iip->ili_item.li_ailp;
 981
 982         /*
 983          * We only want to pull the item from the AIL if it is
 984          * actually there and its location in the log has not
 985          * changed since we started the flush.  Thus, we only bother
 986          * if the ili_logged flag is set and the inode's lsn has not
 987          * changed.  First we check the lsn outside
 988          * the lock since it's cheaper, and then we recheck while
 989          * holding the lock before removing the inode from the AIL.
 990          */
 991         if (iip->ili_logged &&
 992             (iip->ili_item.li_lsn == iip->ili_flush_lsn)) {
 993                 spin_lock(&ailp->xa_lock);
 994                 if (iip->ili_item.li_lsn == iip->ili_flush_lsn) {
 995                         /* xfs_trans_ail_delete() drops the AIL lock. */
 996                         xfs_trans_ail_delete(ailp, (xfs_log_item_t*)iip);
 997                 } else {
 998                         spin_unlock(&ailp->xa_lock);
 999                 }
1000         }
1001
1002         iip->ili_logged = 0;
1003
1004         /*
1005          * Clear the ili_last_fields bits now that we know that the
1006          * data corresponding to them is safely on disk.
1007          */
1008         iip->ili_last_fields = 0;
1009
1010         /*
1011          * Release the inode's flush lock since we're done with it.
1012          */
1013         xfs_ifunlock(ip);
1014
1015         return;
1016 }
1017
1018 /*
1019  * This is the inode flushing abort routine.  It is called
1020  * from xfs_iflush when the filesystem is shutting down to clean
1021  * up the inode state.
1022  * It is responsible for removing the inode item
1023  * from the AIL if it has not been re-logged, and unlocking the inode's
1024  * flush lock.
1025  */
1026 void
1027 xfs_iflush_abort(
1028         xfs_inode_t             *ip)
1029 {
1030         xfs_inode_log_item_t    *iip = ip->i_itemp;
1031         xfs_mount_t             *mp;
1032
1033         iip = ip->i_itemp;
1034         mp = ip->i_mount;
1035         if (iip) {
1036                 struct xfs_ail  *ailp = iip->ili_item.li_ailp;
1037                 if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
1038                         spin_lock(&ailp->xa_lock);
1039                         if (iip->ili_item.li_flags & XFS_LI_IN_AIL) {
1040                                 /* xfs_trans_ail_delete() drops the AIL lock. */
1041                                 xfs_trans_ail_delete(ailp, (xfs_log_item_t *)iip);
1042                         } else
1043                                 spin_unlock(&ailp->xa_lock);
1044                 }
1045                 iip->ili_logged = 0;
1046                 /*
1047                  * Clear the ili_last_fields bits now that we know that the
1048                  * data corresponding to them is safely on disk.
1049                  */
1050                 iip->ili_last_fields = 0;
1051                 /*
1052                  * Clear the inode logging fields so no more flushes are
1053                  * attempted.
1054                  */
1055                 iip->ili_format.ilf_fields = 0;
1056         }
1057         /*
1058          * Release the inode's flush lock since we're done with it.
1059          */
1060         xfs_ifunlock(ip);
1061 }
1062
1063 void
1064 xfs_istale_done(
1065         xfs_buf_t               *bp,
1066         xfs_inode_log_item_t    *iip)
1067 {
1068         xfs_iflush_abort(iip->ili_inode);
1069 }
1070
1071 /*
1072  * convert an xfs_inode_log_format struct from either 32 or 64 bit versions
1073  * (which can have different field alignments) to the native version
1074  */
1075 int
1076 xfs_inode_item_format_convert(
1077         xfs_log_iovec_t         *buf,
1078         xfs_inode_log_format_t  *in_f)
1079 {
1080         if (buf->i_len == sizeof(xfs_inode_log_format_32_t)) {
1081                 xfs_inode_log_format_32_t *in_f32;
1082
1083                 in_f32 = (xfs_inode_log_format_32_t *)buf->i_addr;
1084                 in_f->ilf_type = in_f32->ilf_type;
1085                 in_f->ilf_size = in_f32->ilf_size;
1086                 in_f->ilf_fields = in_f32->ilf_fields;
1087                 in_f->ilf_asize = in_f32->ilf_asize;
1088                 in_f->ilf_dsize = in_f32->ilf_dsize;
1089                 in_f->ilf_ino = in_f32->ilf_ino;
1090                 /* copy biggest field of ilf_u */
1091                 memcpy(in_f->ilf_u.ilfu_uuid.__u_bits,
1092                        in_f32->ilf_u.ilfu_uuid.__u_bits,
1093                        sizeof(uuid_t));
1094                 in_f->ilf_blkno = in_f32->ilf_blkno;
1095                 in_f->ilf_len = in_f32->ilf_len;
1096                 in_f->ilf_boffset = in_f32->ilf_boffset;
1097                 return 0;
1098         } else if (buf->i_len == sizeof(xfs_inode_log_format_64_t)){
1099                 xfs_inode_log_format_64_t *in_f64;
1100
1101                 in_f64 = (xfs_inode_log_format_64_t *)buf->i_addr;
1102                 in_f->ilf_type = in_f64->ilf_type;
1103                 in_f->ilf_size = in_f64->ilf_size;
1104                 in_f->ilf_fields = in_f64->ilf_fields;
1105                 in_f->ilf_asize = in_f64->ilf_asize;
1106                 in_f->ilf_dsize = in_f64->ilf_dsize;
1107                 in_f->ilf_ino = in_f64->ilf_ino;
1108                 /* copy biggest field of ilf_u */
1109                 memcpy(in_f->ilf_u.ilfu_uuid.__u_bits,
1110                        in_f64->ilf_u.ilfu_uuid.__u_bits,
1111                        sizeof(uuid_t));
1112                 in_f->ilf_blkno = in_f64->ilf_blkno;
1113                 in_f->ilf_len = in_f64->ilf_len;
1114                 in_f->ilf_boffset = in_f64->ilf_boffset;
1115                 return 0;
1116         }
1117         return EFSCORRUPTED;
1118 }