Pileus Git - ~andy/linux/blob - fs/xfs/xfs_da_btree.c

   1 /*
   2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   3  * Copyright (c) 2013 Red Hat, Inc.
   4  * All Rights Reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License as
   8  * published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it would be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write the Free Software Foundation,
  17  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  18  */
  19 #include "xfs.h"
  20 #include "xfs_fs.h"
  21 #include "xfs_types.h"
  22 #include "xfs_bit.h"
  23 #include "xfs_log.h"
  24 #include "xfs_trans.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_ag.h"
  27 #include "xfs_mount.h"
  28 #include "xfs_da_btree.h"
  29 #include "xfs_bmap_btree.h"
  30 #include "xfs_dir2.h"
  31 #include "xfs_dir2_format.h"
  32 #include "xfs_dir2_priv.h"
  33 #include "xfs_dinode.h"
  34 #include "xfs_inode.h"
  35 #include "xfs_inode_item.h"
  36 #include "xfs_alloc.h"
  37 #include "xfs_bmap.h"
  38 #include "xfs_attr.h"
  39 #include "xfs_attr_leaf.h"
  40 #include "xfs_error.h"
  41 #include "xfs_trace.h"
  42 #include "xfs_cksum.h"
  43 #include "xfs_buf_item.h"
  44
  45 /*
  46  * xfs_da_btree.c
  47  *
  48  * Routines to implement directories as Btrees of hashed names.
  49  */
  50
  51 /*========================================================================
  52  * Function prototypes for the kernel.
  53  *========================================================================*/
  54
  55 /*
  56  * Routines used for growing the Btree.
  57  */
  58 STATIC int xfs_da3_root_split(xfs_da_state_t *state,
  59                                             xfs_da_state_blk_t *existing_root,
  60                                             xfs_da_state_blk_t *new_child);
  61 STATIC int xfs_da3_node_split(xfs_da_state_t *state,
  62                                             xfs_da_state_blk_t *existing_blk,
  63                                             xfs_da_state_blk_t *split_blk,
  64                                             xfs_da_state_blk_t *blk_to_add,
  65                                             int treelevel,
  66                                             int *result);
  67 STATIC void xfs_da3_node_rebalance(xfs_da_state_t *state,
  68                                          xfs_da_state_blk_t *node_blk_1,
  69                                          xfs_da_state_blk_t *node_blk_2);
  70 STATIC void xfs_da3_node_add(xfs_da_state_t *state,
  71                                    xfs_da_state_blk_t *old_node_blk,
  72                                    xfs_da_state_blk_t *new_node_blk);
  73
  74 /*
  75  * Routines used for shrinking the Btree.
  76  */
  77 STATIC int xfs_da3_root_join(xfs_da_state_t *state,
  78                                            xfs_da_state_blk_t *root_blk);
  79 STATIC int xfs_da3_node_toosmall(xfs_da_state_t *state, int *retval);
  80 STATIC void xfs_da3_node_remove(xfs_da_state_t *state,
  81                                               xfs_da_state_blk_t *drop_blk);
  82 STATIC void xfs_da3_node_unbalance(xfs_da_state_t *state,
  83                                          xfs_da_state_blk_t *src_node_blk,
  84                                          xfs_da_state_blk_t *dst_node_blk);
  85
  86 /*
  87  * Utility routines.
  88  */
  89 STATIC int      xfs_da3_blk_unlink(xfs_da_state_t *state,
  90                                   xfs_da_state_blk_t *drop_blk,
  91                                   xfs_da_state_blk_t *save_blk);
  92
  93
  94 kmem_zone_t *xfs_da_state_zone; /* anchor for state struct zone */
  95
  96 /*
  97  * Allocate a dir-state structure.
  98  * We don't put them on the stack since they're large.
  99  */
 100 xfs_da_state_t *
 101 xfs_da_state_alloc(void)
 102 {
 103         return kmem_zone_zalloc(xfs_da_state_zone, KM_NOFS);
 104 }
 105
 106 /*
 107  * Kill the altpath contents of a da-state structure.
 108  */
 109 STATIC void
 110 xfs_da_state_kill_altpath(xfs_da_state_t *state)
 111 {
 112         int     i;
 113
 114         for (i = 0; i < state->altpath.active; i++)
 115                 state->altpath.blk[i].bp = NULL;
 116         state->altpath.active = 0;
 117 }
 118
 119 /*
 120  * Free a da-state structure.
 121  */
 122 void
 123 xfs_da_state_free(xfs_da_state_t *state)
 124 {
 125         xfs_da_state_kill_altpath(state);
 126 #ifdef DEBUG
 127         memset((char *)state, 0, sizeof(*state));
 128 #endif /* DEBUG */
 129         kmem_zone_free(xfs_da_state_zone, state);
 130 }
 131
 132 void
 133 xfs_da3_node_hdr_from_disk(
 134         struct xfs_da3_icnode_hdr       *to,
 135         struct xfs_da_intnode           *from)
 136 {
 137         ASSERT(from->hdr.info.magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
 138                from->hdr.info.magic == cpu_to_be16(XFS_DA3_NODE_MAGIC));
 139
 140         if (from->hdr.info.magic == cpu_to_be16(XFS_DA3_NODE_MAGIC)) {
 141                 struct xfs_da3_node_hdr *hdr3 = (struct xfs_da3_node_hdr *)from;
 142
 143                 to->forw = be32_to_cpu(hdr3->info.hdr.forw);
 144                 to->back = be32_to_cpu(hdr3->info.hdr.back);
 145                 to->magic = be16_to_cpu(hdr3->info.hdr.magic);
 146                 to->count = be16_to_cpu(hdr3->__count);
 147                 to->level = be16_to_cpu(hdr3->__level);
 148                 return;
 149         }
 150         to->forw = be32_to_cpu(from->hdr.info.forw);
 151         to->back = be32_to_cpu(from->hdr.info.back);
 152         to->magic = be16_to_cpu(from->hdr.info.magic);
 153         to->count = be16_to_cpu(from->hdr.__count);
 154         to->level = be16_to_cpu(from->hdr.__level);
 155 }
 156
 157 void
 158 xfs_da3_node_hdr_to_disk(
 159         struct xfs_da_intnode           *to,
 160         struct xfs_da3_icnode_hdr       *from)
 161 {
 162         ASSERT(from->magic == XFS_DA_NODE_MAGIC ||
 163                from->magic == XFS_DA3_NODE_MAGIC);
 164
 165         if (from->magic == XFS_DA3_NODE_MAGIC) {
 166                 struct xfs_da3_node_hdr *hdr3 = (struct xfs_da3_node_hdr *)to;
 167
 168                 hdr3->info.hdr.forw = cpu_to_be32(from->forw);
 169                 hdr3->info.hdr.back = cpu_to_be32(from->back);
 170                 hdr3->info.hdr.magic = cpu_to_be16(from->magic);
 171                 hdr3->__count = cpu_to_be16(from->count);
 172                 hdr3->__level = cpu_to_be16(from->level);
 173                 return;
 174         }
 175         to->hdr.info.forw = cpu_to_be32(from->forw);
 176         to->hdr.info.back = cpu_to_be32(from->back);
 177         to->hdr.info.magic = cpu_to_be16(from->magic);
 178         to->hdr.__count = cpu_to_be16(from->count);
 179         to->hdr.__level = cpu_to_be16(from->level);
 180 }
 181
 182 static bool
 183 xfs_da3_node_verify(
 184         struct xfs_buf          *bp)
 185 {
 186         struct xfs_mount        *mp = bp->b_target->bt_mount;
 187         struct xfs_da_intnode   *hdr = bp->b_addr;
 188         struct xfs_da3_icnode_hdr ichdr;
 189
 190         xfs_da3_node_hdr_from_disk(&ichdr, hdr);
 191
 192         if (xfs_sb_version_hascrc(&mp->m_sb)) {
 193                 struct xfs_da3_node_hdr *hdr3 = bp->b_addr;
 194
 195                 if (ichdr.magic != XFS_DA3_NODE_MAGIC)
 196                         return false;
 197
 198                 if (!uuid_equal(&hdr3->info.uuid, &mp->m_sb.sb_uuid))
 199                         return false;
 200                 if (be64_to_cpu(hdr3->info.blkno) != bp->b_bn)
 201                         return false;
 202         } else {
 203                 if (ichdr.magic != XFS_DA_NODE_MAGIC)
 204                         return false;
 205         }
 206         if (ichdr.level == 0)
 207                 return false;
 208         if (ichdr.level > XFS_DA_NODE_MAXDEPTH)
 209                 return false;
 210         if (ichdr.count == 0)
 211                 return false;
 212
 213         /*
 214          * we don't know if the node is for and attribute or directory tree,
 215          * so only fail if the count is outside both bounds
 216          */
 217         if (ichdr.count > mp->m_dir_node_ents &&
 218             ichdr.count > mp->m_attr_node_ents)
 219                 return false;
 220
 221         /* XXX: hash order check? */
 222
 223         return true;
 224 }
 225
 226 static void
 227 xfs_da3_node_write_verify(
 228         struct xfs_buf  *bp)
 229 {
 230         struct xfs_mount        *mp = bp->b_target->bt_mount;
 231         struct xfs_buf_log_item *bip = bp->b_fspriv;
 232         struct xfs_da3_node_hdr *hdr3 = bp->b_addr;
 233
 234         if (!xfs_da3_node_verify(bp)) {
 235                 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp, bp->b_addr);
 236                 xfs_buf_ioerror(bp, EFSCORRUPTED);
 237                 return;
 238         }
 239
 240         if (!xfs_sb_version_hascrc(&mp->m_sb))
 241                 return;
 242
 243         if (bip)
 244                 hdr3->info.lsn = cpu_to_be64(bip->bli_item.li_lsn);
 245
 246         xfs_update_cksum(bp->b_addr, BBTOB(bp->b_length), XFS_DA3_NODE_CRC_OFF);
 247 }
 248
 249 /*
 250  * leaf/node format detection on trees is sketchy, so a node read can be done on
 251  * leaf level blocks when detection identifies the tree as a node format tree
 252  * incorrectly. In this case, we need to swap the verifier to match the correct
 253  * format of the block being read.
 254  */
 255 static void
 256 xfs_da3_node_read_verify(
 257         struct xfs_buf          *bp)
 258 {
 259         struct xfs_mount        *mp = bp->b_target->bt_mount;
 260         struct xfs_da_blkinfo   *info = bp->b_addr;
 261
 262         switch (be16_to_cpu(info->magic)) {
 263                 case XFS_DA3_NODE_MAGIC:
 264                         if (!xfs_verify_cksum(bp->b_addr, BBTOB(bp->b_length),
 265                                               XFS_DA3_NODE_CRC_OFF))
 266                                 break;
 267                         /* fall through */
 268                 case XFS_DA_NODE_MAGIC:
 269                         if (!xfs_da3_node_verify(bp))
 270                                 break;
 271                         return;
 272                 case XFS_ATTR_LEAF_MAGIC:
 273                 case XFS_ATTR3_LEAF_MAGIC:
 274                         bp->b_ops = &xfs_attr3_leaf_buf_ops;
 275                         bp->b_ops->verify_read(bp);
 276                         return;
 277                 case XFS_DIR2_LEAFN_MAGIC:
 278                 case XFS_DIR3_LEAFN_MAGIC:
 279                         bp->b_ops = &xfs_dir3_leafn_buf_ops;
 280                         bp->b_ops->verify_read(bp);
 281                         return;
 282                 default:
 283                         break;
 284         }
 285
 286         /* corrupt block */
 287         XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp, bp->b_addr);
 288         xfs_buf_ioerror(bp, EFSCORRUPTED);
 289 }
 290
 291 const struct xfs_buf_ops xfs_da3_node_buf_ops = {
 292         .verify_read = xfs_da3_node_read_verify,
 293         .verify_write = xfs_da3_node_write_verify,
 294 };
 295
 296 int
 297 xfs_da3_node_read(
 298         struct xfs_trans        *tp,
 299         struct xfs_inode        *dp,
 300         xfs_dablk_t             bno,
 301         xfs_daddr_t             mappedbno,
 302         struct xfs_buf          **bpp,
 303         int                     which_fork)
 304 {
 305         int                     err;
 306
 307         err = xfs_da_read_buf(tp, dp, bno, mappedbno, bpp,
 308                                         which_fork, &xfs_da3_node_buf_ops);
 309         if (!err && tp) {
 310                 struct xfs_da_blkinfo   *info = (*bpp)->b_addr;
 311                 int                     type;
 312
 313                 switch (be16_to_cpu(info->magic)) {
 314                 case XFS_DA_NODE_MAGIC:
 315                 case XFS_DA3_NODE_MAGIC:
 316                         type = XFS_BLFT_DA_NODE_BUF;
 317                         break;
 318                 case XFS_ATTR_LEAF_MAGIC:
 319                 case XFS_ATTR3_LEAF_MAGIC:
 320                         type = XFS_BLFT_ATTR_LEAF_BUF;
 321                         break;
 322                 case XFS_DIR2_LEAFN_MAGIC:
 323                 case XFS_DIR3_LEAFN_MAGIC:
 324                         type = XFS_BLFT_DIR_LEAFN_BUF;
 325                         break;
 326                 default:
 327                         type = 0;
 328                         ASSERT(0);
 329                         break;
 330                 }
 331                 xfs_trans_buf_set_type(tp, *bpp, type);
 332         }
 333         return err;
 334 }
 335
 336 /*========================================================================
 337  * Routines used for growing the Btree.
 338  *========================================================================*/
 339
 340 /*
 341  * Create the initial contents of an intermediate node.
 342  */
 343 int
 344 xfs_da3_node_create(
 345         struct xfs_da_args      *args,
 346         xfs_dablk_t             blkno,
 347         int                     level,
 348         struct xfs_buf          **bpp,
 349         int                     whichfork)
 350 {
 351         struct xfs_da_intnode   *node;
 352         struct xfs_trans        *tp = args->trans;
 353         struct xfs_mount        *mp = tp->t_mountp;
 354         struct xfs_da3_icnode_hdr ichdr = {0};
 355         struct xfs_buf          *bp;
 356         int                     error;
 357
 358         trace_xfs_da_node_create(args);
 359         ASSERT(level <= XFS_DA_NODE_MAXDEPTH);
 360
 361         error = xfs_da_get_buf(tp, args->dp, blkno, -1, &bp, whichfork);
 362         if (error)
 363                 return(error);
 364         bp->b_ops = &xfs_da3_node_buf_ops;
 365         xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DA_NODE_BUF);
 366         node = bp->b_addr;
 367
 368         if (xfs_sb_version_hascrc(&mp->m_sb)) {
 369                 struct xfs_da3_node_hdr *hdr3 = bp->b_addr;
 370
 371                 ichdr.magic = XFS_DA3_NODE_MAGIC;
 372                 hdr3->info.blkno = cpu_to_be64(bp->b_bn);
 373                 hdr3->info.owner = cpu_to_be64(args->dp->i_ino);
 374                 uuid_copy(&hdr3->info.uuid, &mp->m_sb.sb_uuid);
 375         } else {
 376                 ichdr.magic = XFS_DA_NODE_MAGIC;
 377         }
 378         ichdr.level = level;
 379
 380         xfs_da3_node_hdr_to_disk(node, &ichdr);
 381         xfs_trans_log_buf(tp, bp,
 382                 XFS_DA_LOGRANGE(node, &node->hdr, xfs_da3_node_hdr_size(node)));
 383
 384         *bpp = bp;
 385         return(0);
 386 }
 387
 388 /*
 389  * Split a leaf node, rebalance, then possibly split
 390  * intermediate nodes, rebalance, etc.
 391  */
 392 int                                                     /* error */
 393 xfs_da3_split(
 394         struct xfs_da_state     *state)
 395 {
 396         struct xfs_da_state_blk *oldblk;
 397         struct xfs_da_state_blk *newblk;
 398         struct xfs_da_state_blk *addblk;
 399         struct xfs_da_intnode   *node;
 400         struct xfs_buf          *bp;
 401         int                     max;
 402         int                     action;
 403         int                     error;
 404         int                     i;
 405
 406         trace_xfs_da_split(state->args);
 407
 408         /*
 409          * Walk back up the tree splitting/inserting/adjusting as necessary.
 410          * If we need to insert and there isn't room, split the node, then
 411          * decide which fragment to insert the new block from below into.
 412          * Note that we may split the root this way, but we need more fixup.
 413          */
 414         max = state->path.active - 1;
 415         ASSERT((max >= 0) && (max < XFS_DA_NODE_MAXDEPTH));
 416         ASSERT(state->path.blk[max].magic == XFS_ATTR_LEAF_MAGIC ||
 417                state->path.blk[max].magic == XFS_DIR2_LEAFN_MAGIC);
 418
 419         addblk = &state->path.blk[max];         /* initial dummy value */
 420         for (i = max; (i >= 0) && addblk; state->path.active--, i--) {
 421                 oldblk = &state->path.blk[i];
 422                 newblk = &state->altpath.blk[i];
 423
 424                 /*
 425                  * If a leaf node then
 426                  *     Allocate a new leaf node, then rebalance across them.
 427                  * else if an intermediate node then
 428                  *     We split on the last layer, must we split the node?
 429                  */
 430                 switch (oldblk->magic) {
 431                 case XFS_ATTR_LEAF_MAGIC:
 432                         error = xfs_attr3_leaf_split(state, oldblk, newblk);
 433                         if ((error != 0) && (error != ENOSPC)) {
 434                                 return(error);  /* GROT: attr is inconsistent */
 435                         }
 436                         if (!error) {
 437                                 addblk = newblk;
 438                                 break;
 439                         }
 440                         /*
 441                          * Entry wouldn't fit, split the leaf again.
 442                          */
 443                         state->extravalid = 1;
 444                         if (state->inleaf) {
 445                                 state->extraafter = 0;  /* before newblk */
 446                                 trace_xfs_attr_leaf_split_before(state->args);
 447                                 error = xfs_attr3_leaf_split(state, oldblk,
 448                                                             &state->extrablk);
 449                         } else {
 450                                 state->extraafter = 1;  /* after newblk */
 451                                 trace_xfs_attr_leaf_split_after(state->args);
 452                                 error = xfs_attr3_leaf_split(state, newblk,
 453                                                             &state->extrablk);
 454                         }
 455                         if (error)
 456                                 return(error);  /* GROT: attr inconsistent */
 457                         addblk = newblk;
 458                         break;
 459                 case XFS_DIR2_LEAFN_MAGIC:
 460                         error = xfs_dir2_leafn_split(state, oldblk, newblk);
 461                         if (error)
 462                                 return error;
 463                         addblk = newblk;
 464                         break;
 465                 case XFS_DA_NODE_MAGIC:
 466                         error = xfs_da3_node_split(state, oldblk, newblk, addblk,
 467                                                          max - i, &action);
 468                         addblk->bp = NULL;
 469                         if (error)
 470                                 return(error);  /* GROT: dir is inconsistent */
 471                         /*
 472                          * Record the newly split block for the next time thru?
 473                          */
 474                         if (action)
 475                                 addblk = newblk;
 476                         else
 477                                 addblk = NULL;
 478                         break;
 479                 }
 480
 481                 /*
 482                  * Update the btree to show the new hashval for this child.
 483                  */
 484                 xfs_da3_fixhashpath(state, &state->path);
 485         }
 486         if (!addblk)
 487                 return(0);
 488
 489         /*
 490          * Split the root node.
 491          */
 492         ASSERT(state->path.active == 0);
 493         oldblk = &state->path.blk[0];
 494         error = xfs_da3_root_split(state, oldblk, addblk);
 495         if (error) {
 496                 addblk->bp = NULL;
 497                 return(error);  /* GROT: dir is inconsistent */
 498         }
 499
 500         /*
 501          * Update pointers to the node which used to be block 0 and
 502          * just got bumped because of the addition of a new root node.
 503          * There might be three blocks involved if a double split occurred,
 504          * and the original block 0 could be at any position in the list.
 505          *
 506          * Note: the magic numbers and sibling pointers are in the same
 507          * physical place for both v2 and v3 headers (by design). Hence it
 508          * doesn't matter which version of the xfs_da_intnode structure we use
 509          * here as the result will be the same using either structure.
 510          */
 511         node = oldblk->bp->b_addr;
 512         if (node->hdr.info.forw) {
 513                 if (be32_to_cpu(node->hdr.info.forw) == addblk->blkno) {
 514                         bp = addblk->bp;
 515                 } else {
 516                         ASSERT(state->extravalid);
 517                         bp = state->extrablk.bp;
 518                 }
 519                 node = bp->b_addr;
 520                 node->hdr.info.back = cpu_to_be32(oldblk->blkno);
 521                 xfs_trans_log_buf(state->args->trans, bp,
 522                     XFS_DA_LOGRANGE(node, &node->hdr.info,
 523                     sizeof(node->hdr.info)));
 524         }
 525         node = oldblk->bp->b_addr;
 526         if (node->hdr.info.back) {
 527                 if (be32_to_cpu(node->hdr.info.back) == addblk->blkno) {
 528                         bp = addblk->bp;
 529                 } else {
 530                         ASSERT(state->extravalid);
 531                         bp = state->extrablk.bp;
 532                 }
 533                 node = bp->b_addr;
 534                 node->hdr.info.forw = cpu_to_be32(oldblk->blkno);
 535                 xfs_trans_log_buf(state->args->trans, bp,
 536                     XFS_DA_LOGRANGE(node, &node->hdr.info,
 537                     sizeof(node->hdr.info)));
 538         }
 539         addblk->bp = NULL;
 540         return(0);
 541 }
 542
 543 /*
 544  * Split the root.  We have to create a new root and point to the two
 545  * parts (the split old root) that we just created.  Copy block zero to
 546  * the EOF, extending the inode in process.
 547  */
 548 STATIC int                                              /* error */
 549 xfs_da3_root_split(
 550         struct xfs_da_state     *state,
 551         struct xfs_da_state_blk *blk1,
 552         struct xfs_da_state_blk *blk2)
 553 {
 554         struct xfs_da_intnode   *node;
 555         struct xfs_da_intnode   *oldroot;
 556         struct xfs_da_node_entry *btree;
 557         struct xfs_da3_icnode_hdr nodehdr;
 558         struct xfs_da_args      *args;
 559         struct xfs_buf          *bp;
 560         struct xfs_inode        *dp;
 561         struct xfs_trans        *tp;
 562         struct xfs_mount        *mp;
 563         struct xfs_dir2_leaf    *leaf;
 564         xfs_dablk_t             blkno;
 565         int                     level;
 566         int                     error;
 567         int                     size;
 568
 569         trace_xfs_da_root_split(state->args);
 570
 571         /*
 572          * Copy the existing (incorrect) block from the root node position
 573          * to a free space somewhere.
 574          */
 575         args = state->args;
 576         error = xfs_da_grow_inode(args, &blkno);
 577         if (error)
 578                 return error;
 579
 580         dp = args->dp;
 581         tp = args->trans;
 582         mp = state->mp;
 583         error = xfs_da_get_buf(tp, dp, blkno, -1, &bp, args->whichfork);
 584         if (error)
 585                 return error;
 586         node = bp->b_addr;
 587         oldroot = blk1->bp->b_addr;
 588         if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
 589             oldroot->hdr.info.magic == cpu_to_be16(XFS_DA3_NODE_MAGIC)) {
 590                 struct xfs_da3_icnode_hdr nodehdr;
 591
 592                 xfs_da3_node_hdr_from_disk(&nodehdr, oldroot);
 593                 btree = xfs_da3_node_tree_p(oldroot);
 594                 size = (int)((char *)&btree[nodehdr.count] - (char *)oldroot);
 595                 level = nodehdr.level;
 596
 597                 /*
 598                  * we are about to copy oldroot to bp, so set up the type
 599                  * of bp while we know exactly what it will be.
 600                  */
 601                 xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DA_NODE_BUF);
 602         } else {
 603                 struct xfs_dir3_icleaf_hdr leafhdr;
 604                 struct xfs_dir2_leaf_entry *ents;
 605
 606                 leaf = (xfs_dir2_leaf_t *)oldroot;
 607                 xfs_dir3_leaf_hdr_from_disk(&leafhdr, leaf);
 608                 ents = xfs_dir3_leaf_ents_p(leaf);
 609
 610                 ASSERT(leafhdr.magic == XFS_DIR2_LEAFN_MAGIC ||
 611                        leafhdr.magic == XFS_DIR3_LEAFN_MAGIC);
 612                 size = (int)((char *)&ents[leafhdr.count] - (char *)leaf);
 613                 level = 0;
 614
 615                 /*
 616                  * we are about to copy oldroot to bp, so set up the type
 617                  * of bp while we know exactly what it will be.
 618                  */
 619                 xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DIR_LEAFN_BUF);
 620         }
 621
 622         /*
 623          * we can copy most of the information in the node from one block to
 624          * another, but for CRC enabled headers we have to make sure that the
 625          * block specific identifiers are kept intact. We update the buffer
 626          * directly for this.
 627          */
 628         memcpy(node, oldroot, size);
 629         if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DA3_NODE_MAGIC) ||
 630             oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
 631                 struct xfs_da3_intnode *node3 = (struct xfs_da3_intnode *)node;
 632
 633                 node3->hdr.info.blkno = cpu_to_be64(bp->b_bn);
 634         }
 635         xfs_trans_log_buf(tp, bp, 0, size - 1);
 636
 637         bp->b_ops = blk1->bp->b_ops;
 638         blk1->bp = bp;
 639         blk1->blkno = blkno;
 640
 641         /*
 642          * Set up the new root node.
 643          */
 644         error = xfs_da3_node_create(args,
 645                 (args->whichfork == XFS_DATA_FORK) ? mp->m_dirleafblk : 0,
 646                 level + 1, &bp, args->whichfork);
 647         if (error)
 648                 return error;
 649
 650         node = bp->b_addr;
 651         xfs_da3_node_hdr_from_disk(&nodehdr, node);
 652         btree = xfs_da3_node_tree_p(node);
 653         btree[0].hashval = cpu_to_be32(blk1->hashval);
 654         btree[0].before = cpu_to_be32(blk1->blkno);
 655         btree[1].hashval = cpu_to_be32(blk2->hashval);
 656         btree[1].before = cpu_to_be32(blk2->blkno);
 657         nodehdr.count = 2;
 658         xfs_da3_node_hdr_to_disk(node, &nodehdr);
 659
 660 #ifdef DEBUG
 661         if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
 662             oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
 663                 ASSERT(blk1->blkno >= mp->m_dirleafblk &&
 664                        blk1->blkno < mp->m_dirfreeblk);
 665                 ASSERT(blk2->blkno >= mp->m_dirleafblk &&
 666                        blk2->blkno < mp->m_dirfreeblk);
 667         }
 668 #endif
 669
 670         /* Header is already logged by xfs_da_node_create */
 671         xfs_trans_log_buf(tp, bp,
 672                 XFS_DA_LOGRANGE(node, btree, sizeof(xfs_da_node_entry_t) * 2));
 673
 674         return 0;
 675 }
 676
 677 /*
 678  * Split the node, rebalance, then add the new entry.
 679  */
 680 STATIC int                                              /* error */
 681 xfs_da3_node_split(
 682         struct xfs_da_state     *state,
 683         struct xfs_da_state_blk *oldblk,
 684         struct xfs_da_state_blk *newblk,
 685         struct xfs_da_state_blk *addblk,
 686         int                     treelevel,
 687         int                     *result)
 688 {
 689         struct xfs_da_intnode   *node;
 690         struct xfs_da3_icnode_hdr nodehdr;
 691         xfs_dablk_t             blkno;
 692         int                     newcount;
 693         int                     error;
 694         int                     useextra;
 695
 696         trace_xfs_da_node_split(state->args);
 697
 698         node = oldblk->bp->b_addr;
 699         xfs_da3_node_hdr_from_disk(&nodehdr, node);
 700
 701         /*
 702          * With V2 dirs the extra block is data or freespace.
 703          */
 704         useextra = state->extravalid && state->args->whichfork == XFS_ATTR_FORK;
 705         newcount = 1 + useextra;
 706         /*
 707          * Do we have to split the node?
 708          */
 709         if (nodehdr.count + newcount > state->node_ents) {
 710                 /*
 711                  * Allocate a new node, add to the doubly linked chain of
 712                  * nodes, then move some of our excess entries into it.
 713                  */
 714                 error = xfs_da_grow_inode(state->args, &blkno);
 715                 if (error)
 716                         return(error);  /* GROT: dir is inconsistent */
 717
 718                 error = xfs_da3_node_create(state->args, blkno, treelevel,
 719                                            &newblk->bp, state->args->whichfork);
 720                 if (error)
 721                         return(error);  /* GROT: dir is inconsistent */
 722                 newblk->blkno = blkno;
 723                 newblk->magic = XFS_DA_NODE_MAGIC;
 724                 xfs_da3_node_rebalance(state, oldblk, newblk);
 725                 error = xfs_da3_blk_link(state, oldblk, newblk);
 726                 if (error)
 727                         return(error);
 728                 *result = 1;
 729         } else {
 730                 *result = 0;
 731         }
 732
 733         /*
 734          * Insert the new entry(s) into the correct block
 735          * (updating last hashval in the process).
 736          *
 737          * xfs_da3_node_add() inserts BEFORE the given index,
 738          * and as a result of using node_lookup_int() we always
 739          * point to a valid entry (not after one), but a split
 740          * operation always results in a new block whose hashvals
 741          * FOLLOW the current block.
 742          *
 743          * If we had double-split op below us, then add the extra block too.
 744          */
 745         node = oldblk->bp->b_addr;
 746         xfs_da3_node_hdr_from_disk(&nodehdr, node);
 747         if (oldblk->index <= nodehdr.count) {
 748                 oldblk->index++;
 749                 xfs_da3_node_add(state, oldblk, addblk);
 750                 if (useextra) {
 751                         if (state->extraafter)
 752                                 oldblk->index++;
 753                         xfs_da3_node_add(state, oldblk, &state->extrablk);
 754                         state->extravalid = 0;
 755                 }
 756         } else {
 757                 newblk->index++;
 758                 xfs_da3_node_add(state, newblk, addblk);
 759                 if (useextra) {
 760                         if (state->extraafter)
 761                                 newblk->index++;
 762                         xfs_da3_node_add(state, newblk, &state->extrablk);
 763                         state->extravalid = 0;
 764                 }
 765         }
 766
 767         return(0);
 768 }
 769
 770 /*
 771  * Balance the btree elements between two intermediate nodes,
 772  * usually one full and one empty.
 773  *
 774  * NOTE: if blk2 is empty, then it will get the upper half of blk1.
 775  */
 776 STATIC void
 777 xfs_da3_node_rebalance(
 778         struct xfs_da_state     *state,
 779         struct xfs_da_state_blk *blk1,
 780         struct xfs_da_state_blk *blk2)
 781 {
 782         struct xfs_da_intnode   *node1;
 783         struct xfs_da_intnode   *node2;
 784         struct xfs_da_intnode   *tmpnode;
 785         struct xfs_da_node_entry *btree1;
 786         struct xfs_da_node_entry *btree2;
 787         struct xfs_da_node_entry *btree_s;
 788         struct xfs_da_node_entry *btree_d;
 789         struct xfs_da3_icnode_hdr nodehdr1;
 790         struct xfs_da3_icnode_hdr nodehdr2;
 791         struct xfs_trans        *tp;
 792         int                     count;
 793         int                     tmp;
 794         int                     swap = 0;
 795
 796         trace_xfs_da_node_rebalance(state->args);
 797
 798         node1 = blk1->bp->b_addr;
 799         node2 = blk2->bp->b_addr;
 800         xfs_da3_node_hdr_from_disk(&nodehdr1, node1);
 801         xfs_da3_node_hdr_from_disk(&nodehdr2, node2);
 802         btree1 = xfs_da3_node_tree_p(node1);
 803         btree2 = xfs_da3_node_tree_p(node2);
 804
 805         /*
 806          * Figure out how many entries need to move, and in which direction.
 807          * Swap the nodes around if that makes it simpler.
 808          */
 809         if (nodehdr1.count > 0 && nodehdr2.count > 0 &&
 810             ((be32_to_cpu(btree2[0].hashval) < be32_to_cpu(btree1[0].hashval)) ||
 811              (be32_to_cpu(btree2[nodehdr2.count - 1].hashval) <
 812                         be32_to_cpu(btree1[nodehdr1.count - 1].hashval)))) {
 813                 tmpnode = node1;
 814                 node1 = node2;
 815                 node2 = tmpnode;
 816                 xfs_da3_node_hdr_from_disk(&nodehdr1, node1);
 817                 xfs_da3_node_hdr_from_disk(&nodehdr2, node2);
 818                 btree1 = xfs_da3_node_tree_p(node1);
 819                 btree2 = xfs_da3_node_tree_p(node2);
 820                 swap = 1;
 821         }
 822
 823         count = (nodehdr1.count - nodehdr2.count) / 2;
 824         if (count == 0)
 825                 return;
 826         tp = state->args->trans;
 827         /*
 828          * Two cases: high-to-low and low-to-high.
 829          */
 830         if (count > 0) {
 831                 /*
 832                  * Move elements in node2 up to make a hole.
 833                  */
 834                 tmp = nodehdr2.count;
 835                 if (tmp > 0) {
 836                         tmp *= (uint)sizeof(xfs_da_node_entry_t);
 837                         btree_s = &btree2[0];
 838                         btree_d = &btree2[count];
 839                         memmove(btree_d, btree_s, tmp);
 840                 }
 841
 842                 /*
 843                  * Move the req'd B-tree elements from high in node1 to
 844                  * low in node2.
 845                  */
 846                 nodehdr2.count += count;
 847                 tmp = count * (uint)sizeof(xfs_da_node_entry_t);
 848                 btree_s = &btree1[nodehdr1.count - count];
 849                 btree_d = &btree2[0];
 850                 memcpy(btree_d, btree_s, tmp);
 851                 nodehdr1.count -= count;
 852         } else {
 853                 /*
 854                  * Move the req'd B-tree elements from low in node2 to
 855                  * high in node1.
 856                  */
 857                 count = -count;
 858                 tmp = count * (uint)sizeof(xfs_da_node_entry_t);
 859                 btree_s = &btree2[0];
 860                 btree_d = &btree1[nodehdr1.count];
 861                 memcpy(btree_d, btree_s, tmp);
 862                 nodehdr1.count += count;
 863
 864                 xfs_trans_log_buf(tp, blk1->bp,
 865                         XFS_DA_LOGRANGE(node1, btree_d, tmp));
 866
 867                 /*
 868                  * Move elements in node2 down to fill the hole.
 869                  */
 870                 tmp  = nodehdr2.count - count;
 871                 tmp *= (uint)sizeof(xfs_da_node_entry_t);
 872                 btree_s = &btree2[count];
 873                 btree_d = &btree2[0];
 874                 memmove(btree_d, btree_s, tmp);
 875                 nodehdr2.count -= count;
 876         }
 877
 878         /*
 879          * Log header of node 1 and all current bits of node 2.
 880          */
 881         xfs_da3_node_hdr_to_disk(node1, &nodehdr1);
 882         xfs_trans_log_buf(tp, blk1->bp,
 883                 XFS_DA_LOGRANGE(node1, &node1->hdr,
 884                                 xfs_da3_node_hdr_size(node1)));
 885
 886         xfs_da3_node_hdr_to_disk(node2, &nodehdr2);
 887         xfs_trans_log_buf(tp, blk2->bp,
 888                 XFS_DA_LOGRANGE(node2, &node2->hdr,
 889                                 xfs_da3_node_hdr_size(node2) +
 890                                 (sizeof(btree2[0]) * nodehdr2.count)));
 891
 892         /*
 893          * Record the last hashval from each block for upward propagation.
 894          * (note: don't use the swapped node pointers)
 895          */
 896         if (swap) {
 897                 node1 = blk1->bp->b_addr;
 898                 node2 = blk2->bp->b_addr;
 899                 xfs_da3_node_hdr_from_disk(&nodehdr1, node1);
 900                 xfs_da3_node_hdr_from_disk(&nodehdr2, node2);
 901                 btree1 = xfs_da3_node_tree_p(node1);
 902                 btree2 = xfs_da3_node_tree_p(node2);
 903         }
 904         blk1->hashval = be32_to_cpu(btree1[nodehdr1.count - 1].hashval);
 905         blk2->hashval = be32_to_cpu(btree2[nodehdr2.count - 1].hashval);
 906
 907         /*
 908          * Adjust the expected index for insertion.
 909          */
 910         if (blk1->index >= nodehdr1.count) {
 911                 blk2->index = blk1->index - nodehdr1.count;
 912                 blk1->index = nodehdr1.count + 1;       /* make it invalid */
 913         }
 914 }
 915
 916 /*
 917  * Add a new entry to an intermediate node.
 918  */
 919 STATIC void
 920 xfs_da3_node_add(
 921         struct xfs_da_state     *state,
 922         struct xfs_da_state_blk *oldblk,
 923         struct xfs_da_state_blk *newblk)
 924 {
 925         struct xfs_da_intnode   *node;
 926         struct xfs_da3_icnode_hdr nodehdr;
 927         struct xfs_da_node_entry *btree;
 928         int                     tmp;
 929
 930         trace_xfs_da_node_add(state->args);
 931
 932         node = oldblk->bp->b_addr;
 933         xfs_da3_node_hdr_from_disk(&nodehdr, node);
 934         btree = xfs_da3_node_tree_p(node);
 935
 936         ASSERT(oldblk->index >= 0 && oldblk->index <= nodehdr.count);
 937         ASSERT(newblk->blkno != 0);
 938         if (state->args->whichfork == XFS_DATA_FORK)
 939                 ASSERT(newblk->blkno >= state->mp->m_dirleafblk &&
 940                        newblk->blkno < state->mp->m_dirfreeblk);
 941
 942         /*
 943          * We may need to make some room before we insert the new node.
 944          */
 945         tmp = 0;
 946         if (oldblk->index < nodehdr.count) {
 947                 tmp = (nodehdr.count - oldblk->index) * (uint)sizeof(*btree);
 948                 memmove(&btree[oldblk->index + 1], &btree[oldblk->index], tmp);
 949         }
 950         btree[oldblk->index].hashval = cpu_to_be32(newblk->hashval);
 951         btree[oldblk->index].before = cpu_to_be32(newblk->blkno);
 952         xfs_trans_log_buf(state->args->trans, oldblk->bp,
 953                 XFS_DA_LOGRANGE(node, &btree[oldblk->index],
 954                                 tmp + sizeof(*btree)));
 955
 956         nodehdr.count += 1;
 957         xfs_da3_node_hdr_to_disk(node, &nodehdr);
 958         xfs_trans_log_buf(state->args->trans, oldblk->bp,
 959                 XFS_DA_LOGRANGE(node, &node->hdr, xfs_da3_node_hdr_size(node)));
 960
 961         /*
 962          * Copy the last hash value from the oldblk to propagate upwards.
 963          */
 964         oldblk->hashval = be32_to_cpu(btree[nodehdr.count - 1].hashval);
 965 }
 966
 967 /*========================================================================
 968  * Routines used for shrinking the Btree.
 969  *========================================================================*/
 970
 971 /*
 972  * Deallocate an empty leaf node, remove it from its parent,
 973  * possibly deallocating that block, etc...
 974  */
 975 int
 976 xfs_da3_join(
 977         struct xfs_da_state     *state)
 978 {
 979         struct xfs_da_state_blk *drop_blk;
 980         struct xfs_da_state_blk *save_blk;
 981         int                     action = 0;
 982         int                     error;
 983
 984         trace_xfs_da_join(state->args);
 985
 986         drop_blk = &state->path.blk[ state->path.active-1 ];
 987         save_blk = &state->altpath.blk[ state->path.active-1 ];
 988         ASSERT(state->path.blk[0].magic == XFS_DA_NODE_MAGIC);
 989         ASSERT(drop_blk->magic == XFS_ATTR_LEAF_MAGIC ||
 990                drop_blk->magic == XFS_DIR2_LEAFN_MAGIC);
 991
 992         /*
 993          * Walk back up the tree joining/deallocating as necessary.
 994          * When we stop dropping blocks, break out.
 995          */
 996         for (  ; state->path.active >= 2; drop_blk--, save_blk--,
 997                  state->path.active--) {
 998                 /*
 999                  * See if we can combine the block with a neighbor.
1000                  *   (action == 0) => no options, just leave
1001                  *   (action == 1) => coalesce, then unlink
1002                  *   (action == 2) => block empty, unlink it
1003                  */
1004                 switch (drop_blk->magic) {
1005                 case XFS_ATTR_LEAF_MAGIC:
1006                         error = xfs_attr3_leaf_toosmall(state, &action);
1007                         if (error)
1008                                 return(error);
1009                         if (action == 0)
1010                                 return(0);
1011                         xfs_attr3_leaf_unbalance(state, drop_blk, save_blk);
1012                         break;
1013                 case XFS_DIR2_LEAFN_MAGIC:
1014                         error = xfs_dir2_leafn_toosmall(state, &action);
1015                         if (error)
1016                                 return error;
1017                         if (action == 0)
1018                                 return 0;
1019                         xfs_dir2_leafn_unbalance(state, drop_blk, save_blk);
1020                         break;
1021                 case XFS_DA_NODE_MAGIC:
1022                         /*
1023                          * Remove the offending node, fixup hashvals,
1024                          * check for a toosmall neighbor.
1025                          */
1026                         xfs_da3_node_remove(state, drop_blk);
1027                         xfs_da3_fixhashpath(state, &state->path);
1028                         error = xfs_da3_node_toosmall(state, &action);
1029                         if (error)
1030                                 return(error);
1031                         if (action == 0)
1032                                 return 0;
1033                         xfs_da3_node_unbalance(state, drop_blk, save_blk);
1034                         break;
1035                 }
1036                 xfs_da3_fixhashpath(state, &state->altpath);
1037                 error = xfs_da3_blk_unlink(state, drop_blk, save_blk);
1038                 xfs_da_state_kill_altpath(state);
1039                 if (error)
1040                         return(error);
1041                 error = xfs_da_shrink_inode(state->args, drop_blk->blkno,
1042                                                          drop_blk->bp);
1043                 drop_blk->bp = NULL;
1044                 if (error)
1045                         return(error);
1046         }
1047         /*
1048          * We joined all the way to the top.  If it turns out that
1049          * we only have one entry in the root, make the child block
1050          * the new root.
1051          */
1052         xfs_da3_node_remove(state, drop_blk);
1053         xfs_da3_fixhashpath(state, &state->path);
1054         error = xfs_da3_root_join(state, &state->path.blk[0]);
1055         return(error);
1056 }
1057
1058 #ifdef  DEBUG
1059 static void
1060 xfs_da_blkinfo_onlychild_validate(struct xfs_da_blkinfo *blkinfo, __u16 level)
1061 {
1062         __be16  magic = blkinfo->magic;
1063
1064         if (level == 1) {
1065                 ASSERT(magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
1066                        magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC) ||
1067                        magic == cpu_to_be16(XFS_ATTR_LEAF_MAGIC) ||
1068                        magic == cpu_to_be16(XFS_ATTR3_LEAF_MAGIC));
1069         } else {
1070                 ASSERT(magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
1071                        magic == cpu_to_be16(XFS_DA3_NODE_MAGIC));
1072         }
1073         ASSERT(!blkinfo->forw);
1074         ASSERT(!blkinfo->back);
1075 }
1076 #else   /* !DEBUG */
1077 #define xfs_da_blkinfo_onlychild_validate(blkinfo, level)
1078 #endif  /* !DEBUG */
1079
1080 /*
1081  * We have only one entry in the root.  Copy the only remaining child of
1082  * the old root to block 0 as the new root node.
1083  */
1084 STATIC int
1085 xfs_da3_root_join(
1086         struct xfs_da_state     *state,
1087         struct xfs_da_state_blk *root_blk)
1088 {
1089         struct xfs_da_intnode   *oldroot;
1090         struct xfs_da_args      *args;
1091         xfs_dablk_t             child;
1092         struct xfs_buf          *bp;
1093         struct xfs_da3_icnode_hdr oldroothdr;
1094         struct xfs_da_node_entry *btree;
1095         int                     error;
1096
1097         trace_xfs_da_root_join(state->args);
1098
1099         ASSERT(root_blk->magic == XFS_DA_NODE_MAGIC);
1100
1101         args = state->args;
1102         oldroot = root_blk->bp->b_addr;
1103         xfs_da3_node_hdr_from_disk(&oldroothdr, oldroot);
1104         ASSERT(oldroothdr.forw == 0);
1105         ASSERT(oldroothdr.back == 0);
1106
1107         /*
1108          * If the root has more than one child, then don't do anything.
1109          */
1110         if (oldroothdr.count > 1)
1111                 return 0;
1112
1113         /*
1114          * Read in the (only) child block, then copy those bytes into
1115          * the root block's buffer and free the original child block.
1116          */
1117         btree = xfs_da3_node_tree_p(oldroot);
1118         child = be32_to_cpu(btree[0].before);
1119         ASSERT(child != 0);
1120         error = xfs_da3_node_read(args->trans, args->dp, child, -1, &bp,
1121                                              args->whichfork);
1122         if (error)
1123                 return error;
1124         xfs_da_blkinfo_onlychild_validate(bp->b_addr, oldroothdr.level);
1125
1126         /*
1127          * This could be copying a leaf back into the root block in the case of
1128          * there only being a single leaf block left in the tree. Hence we have
1129          * to update the b_ops pointer as well to match the buffer type change
1130          * that could occur. For dir3 blocks we also need to update the block
1131          * number in the buffer header.
1132          */
1133         memcpy(root_blk->bp->b_addr, bp->b_addr, state->blocksize);
1134         root_blk->bp->b_ops = bp->b_ops;
1135         xfs_trans_buf_copy_type(root_blk->bp, bp);
1136         if (oldroothdr.magic == XFS_DA3_NODE_MAGIC) {
1137                 struct xfs_da3_blkinfo *da3 = root_blk->bp->b_addr;
1138                 da3->blkno = cpu_to_be64(root_blk->bp->b_bn);
1139         }
1140         xfs_trans_log_buf(args->trans, root_blk->bp, 0, state->blocksize - 1);
1141         error = xfs_da_shrink_inode(args, child, bp);
1142         return(error);
1143 }
1144
1145 /*
1146  * Check a node block and its neighbors to see if the block should be
1147  * collapsed into one or the other neighbor.  Always keep the block
1148  * with the smaller block number.
1149  * If the current block is over 50% full, don't try to join it, return 0.
1150  * If the block is empty, fill in the state structure and return 2.
1151  * If it can be collapsed, fill in the state structure and return 1.
1152  * If nothing can be done, return 0.
1153  */
1154 STATIC int
1155 xfs_da3_node_toosmall(
1156         struct xfs_da_state     *state,
1157         int                     *action)
1158 {
1159         struct xfs_da_intnode   *node;
1160         struct xfs_da_state_blk *blk;
1161         struct xfs_da_blkinfo   *info;
1162         xfs_dablk_t             blkno;
1163         struct xfs_buf          *bp;
1164         struct xfs_da3_icnode_hdr nodehdr;
1165         int                     count;
1166         int                     forward;
1167         int                     error;
1168         int                     retval;
1169         int                     i;
1170
1171         trace_xfs_da_node_toosmall(state->args);
1172
1173         /*
1174          * Check for the degenerate case of the block being over 50% full.
1175          * If so, it's not worth even looking to see if we might be able
1176          * to coalesce with a sibling.
1177          */
1178         blk = &state->path.blk[ state->path.active-1 ];
1179         info = blk->bp->b_addr;
1180         node = (xfs_da_intnode_t *)info;
1181         xfs_da3_node_hdr_from_disk(&nodehdr, node);
1182         if (nodehdr.count > (state->node_ents >> 1)) {
1183                 *action = 0;    /* blk over 50%, don't try to join */
1184                 return(0);      /* blk over 50%, don't try to join */
1185         }
1186
1187         /*
1188          * Check for the degenerate case of the block being empty.
1189          * If the block is empty, we'll simply delete it, no need to
1190          * coalesce it with a sibling block.  We choose (arbitrarily)
1191          * to merge with the forward block unless it is NULL.
1192          */
1193         if (nodehdr.count == 0) {
1194                 /*
1195                  * Make altpath point to the block we want to keep and
1196                  * path point to the block we want to drop (this one).
1197                  */
1198                 forward = (info->forw != 0);
1199                 memcpy(&state->altpath, &state->path, sizeof(state->path));
1200                 error = xfs_da3_path_shift(state, &state->altpath, forward,
1201                                                  0, &retval);
1202                 if (error)
1203                         return(error);
1204                 if (retval) {
1205                         *action = 0;
1206                 } else {
1207                         *action = 2;
1208                 }
1209                 return(0);
1210         }
1211
1212         /*
1213          * Examine each sibling block to see if we can coalesce with
1214          * at least 25% free space to spare.  We need to figure out
1215          * whether to merge with the forward or the backward block.
1216          * We prefer coalescing with the lower numbered sibling so as
1217          * to shrink a directory over time.
1218          */
1219         count  = state->node_ents;
1220         count -= state->node_ents >> 2;
1221         count -= nodehdr.count;
1222
1223         /* start with smaller blk num */
1224         forward = nodehdr.forw < nodehdr.back;
1225         for (i = 0; i < 2; forward = !forward, i++) {
1226                 if (forward)
1227                         blkno = nodehdr.forw;
1228                 else
1229                         blkno = nodehdr.back;
1230                 if (blkno == 0)
1231                         continue;
1232                 error = xfs_da3_node_read(state->args->trans, state->args->dp,
1233                                         blkno, -1, &bp, state->args->whichfork);
1234                 if (error)
1235                         return(error);
1236
1237                 node = bp->b_addr;
1238                 xfs_da3_node_hdr_from_disk(&nodehdr, node);
1239                 xfs_trans_brelse(state->args->trans, bp);
1240
1241                 if (count - nodehdr.count >= 0)
1242                         break;  /* fits with at least 25% to spare */
1243         }
1244         if (i >= 2) {
1245                 *action = 0;
1246                 return 0;
1247         }
1248
1249         /*
1250          * Make altpath point to the block we want to keep (the lower
1251          * numbered block) and path point to the block we want to drop.
1252          */
1253         memcpy(&state->altpath, &state->path, sizeof(state->path));
1254         if (blkno < blk->blkno) {
1255                 error = xfs_da3_path_shift(state, &state->altpath, forward,
1256                                                  0, &retval);
1257         } else {
1258                 error = xfs_da3_path_shift(state, &state->path, forward,
1259                                                  0, &retval);
1260         }
1261         if (error)
1262                 return error;
1263         if (retval) {
1264                 *action = 0;
1265                 return 0;
1266         }
1267         *action = 1;
1268         return 0;
1269 }
1270
1271 /*
1272  * Pick up the last hashvalue from an intermediate node.
1273  */
1274 STATIC uint
1275 xfs_da3_node_lasthash(
1276         struct xfs_buf          *bp,
1277         int                     *count)
1278 {
1279         struct xfs_da_intnode    *node;
1280         struct xfs_da_node_entry *btree;
1281         struct xfs_da3_icnode_hdr nodehdr;
1282
1283         node = bp->b_addr;
1284         xfs_da3_node_hdr_from_disk(&nodehdr, node);
1285         if (count)
1286                 *count = nodehdr.count;
1287         if (!nodehdr.count)
1288                 return 0;
1289         btree = xfs_da3_node_tree_p(node);
1290         return be32_to_cpu(btree[nodehdr.count - 1].hashval);
1291 }
1292
1293 /*
1294  * Walk back up the tree adjusting hash values as necessary,
1295  * when we stop making changes, return.
1296  */
1297 void
1298 xfs_da3_fixhashpath(
1299         struct xfs_da_state     *state,
1300         struct xfs_da_state_path *path)
1301 {
1302         struct xfs_da_state_blk *blk;
1303         struct xfs_da_intnode   *node;
1304         struct xfs_da_node_entry *btree;
1305         xfs_dahash_t            lasthash=0;
1306         int                     level;
1307         int                     count;
1308
1309         trace_xfs_da_fixhashpath(state->args);
1310
1311         level = path->active-1;
1312         blk = &path->blk[ level ];
1313         switch (blk->magic) {
1314         case XFS_ATTR_LEAF_MAGIC:
1315                 lasthash = xfs_attr_leaf_lasthash(blk->bp, &count);
1316                 if (count == 0)
1317                         return;
1318                 break;
1319         case XFS_DIR2_LEAFN_MAGIC:
1320                 lasthash = xfs_dir2_leafn_lasthash(blk->bp, &count);
1321                 if (count == 0)
1322                         return;
1323                 break;
1324         case XFS_DA_NODE_MAGIC:
1325                 lasthash = xfs_da3_node_lasthash(blk->bp, &count);
1326                 if (count == 0)
1327                         return;
1328                 break;
1329         }
1330         for (blk--, level--; level >= 0; blk--, level--) {
1331                 struct xfs_da3_icnode_hdr nodehdr;
1332
1333                 node = blk->bp->b_addr;
1334                 xfs_da3_node_hdr_from_disk(&nodehdr, node);
1335                 btree = xfs_da3_node_tree_p(node);
1336                 if (be32_to_cpu(btree->hashval) == lasthash)
1337                         break;
1338                 blk->hashval = lasthash;
1339                 btree[blk->index].hashval = cpu_to_be32(lasthash);
1340                 xfs_trans_log_buf(state->args->trans, blk->bp,
1341                                   XFS_DA_LOGRANGE(node, &btree[blk->index],
1342                                                   sizeof(*btree)));
1343
1344                 lasthash = be32_to_cpu(btree[nodehdr.count - 1].hashval);
1345         }
1346 }
1347
1348 /*
1349  * Remove an entry from an intermediate node.
1350  */
1351 STATIC void
1352 xfs_da3_node_remove(
1353         struct xfs_da_state     *state,
1354         struct xfs_da_state_blk *drop_blk)
1355 {
1356         struct xfs_da_intnode   *node;
1357         struct xfs_da3_icnode_hdr nodehdr;
1358         struct xfs_da_node_entry *btree;
1359         int                     index;
1360         int                     tmp;
1361
1362         trace_xfs_da_node_remove(state->args);
1363
1364         node = drop_blk->bp->b_addr;
1365         xfs_da3_node_hdr_from_disk(&nodehdr, node);
1366         ASSERT(drop_blk->index < nodehdr.count);
1367         ASSERT(drop_blk->index >= 0);
1368
1369         /*
1370          * Copy over the offending entry, or just zero it out.
1371          */
1372         index = drop_blk->index;
1373         btree = xfs_da3_node_tree_p(node);
1374         if (index < nodehdr.count - 1) {
1375                 tmp  = nodehdr.count - index - 1;
1376                 tmp *= (uint)sizeof(xfs_da_node_entry_t);
1377                 memmove(&btree[index], &btree[index + 1], tmp);
1378                 xfs_trans_log_buf(state->args->trans, drop_blk->bp,
1379                     XFS_DA_LOGRANGE(node, &btree[index], tmp));
1380                 index = nodehdr.count - 1;
1381         }
1382         memset(&btree[index], 0, sizeof(xfs_da_node_entry_t));
1383         xfs_trans_log_buf(state->args->trans, drop_blk->bp,
1384             XFS_DA_LOGRANGE(node, &btree[index], sizeof(btree[index])));
1385         nodehdr.count -= 1;
1386         xfs_da3_node_hdr_to_disk(node, &nodehdr);
1387         xfs_trans_log_buf(state->args->trans, drop_blk->bp,
1388             XFS_DA_LOGRANGE(node, &node->hdr, xfs_da3_node_hdr_size(node)));
1389
1390         /*
1391          * Copy the last hash value from the block to propagate upwards.
1392          */
1393         drop_blk->hashval = be32_to_cpu(btree[index - 1].hashval);
1394 }
1395
1396 /*
1397  * Unbalance the elements between two intermediate nodes,
1398  * move all Btree elements from one node into another.
1399  */
1400 STATIC void
1401 xfs_da3_node_unbalance(
1402         struct xfs_da_state     *state,
1403         struct xfs_da_state_blk *drop_blk,
1404         struct xfs_da_state_blk *save_blk)
1405 {
1406         struct xfs_da_intnode   *drop_node;
1407         struct xfs_da_intnode   *save_node;
1408         struct xfs_da_node_entry *drop_btree;
1409         struct xfs_da_node_entry *save_btree;
1410         struct xfs_da3_icnode_hdr drop_hdr;
1411         struct xfs_da3_icnode_hdr save_hdr;
1412         struct xfs_trans        *tp;
1413         int                     sindex;
1414         int                     tmp;
1415
1416         trace_xfs_da_node_unbalance(state->args);
1417
1418         drop_node = drop_blk->bp->b_addr;
1419         save_node = save_blk->bp->b_addr;
1420         xfs_da3_node_hdr_from_disk(&drop_hdr, drop_node);
1421         xfs_da3_node_hdr_from_disk(&save_hdr, save_node);
1422         drop_btree = xfs_da3_node_tree_p(drop_node);
1423         save_btree = xfs_da3_node_tree_p(save_node);
1424         tp = state->args->trans;
1425
1426         /*
1427          * If the dying block has lower hashvals, then move all the
1428          * elements in the remaining block up to make a hole.
1429          */
1430         if ((be32_to_cpu(drop_btree[0].hashval) <
1431                         be32_to_cpu(save_btree[0].hashval)) ||
1432             (be32_to_cpu(drop_btree[drop_hdr.count - 1].hashval) <
1433                         be32_to_cpu(save_btree[save_hdr.count - 1].hashval))) {
1434                 /* XXX: check this - is memmove dst correct? */
1435                 tmp = save_hdr.count * sizeof(xfs_da_node_entry_t);
1436                 memmove(&save_btree[drop_hdr.count], &save_btree[0], tmp);
1437
1438                 sindex = 0;
1439                 xfs_trans_log_buf(tp, save_blk->bp,
1440                         XFS_DA_LOGRANGE(save_node, &save_btree[0],
1441                                 (save_hdr.count + drop_hdr.count) *
1442                                                 sizeof(xfs_da_node_entry_t)));
1443         } else {
1444                 sindex = save_hdr.count;
1445                 xfs_trans_log_buf(tp, save_blk->bp,
1446                         XFS_DA_LOGRANGE(save_node, &save_btree[sindex],
1447                                 drop_hdr.count * sizeof(xfs_da_node_entry_t)));
1448         }
1449
1450         /*
1451          * Move all the B-tree elements from drop_blk to save_blk.
1452          */
1453         tmp = drop_hdr.count * (uint)sizeof(xfs_da_node_entry_t);
1454         memcpy(&save_btree[sindex], &drop_btree[0], tmp);
1455         save_hdr.count += drop_hdr.count;
1456
1457         xfs_da3_node_hdr_to_disk(save_node, &save_hdr);
1458         xfs_trans_log_buf(tp, save_blk->bp,
1459                 XFS_DA_LOGRANGE(save_node, &save_node->hdr,
1460                                 xfs_da3_node_hdr_size(save_node)));
1461
1462         /*
1463          * Save the last hashval in the remaining block for upward propagation.
1464          */
1465         save_blk->hashval = be32_to_cpu(save_btree[save_hdr.count - 1].hashval);
1466 }
1467
1468 /*========================================================================
1469  * Routines used for finding things in the Btree.
1470  *========================================================================*/
1471
1472 /*
1473  * Walk down the Btree looking for a particular filename, filling
1474  * in the state structure as we go.
1475  *
1476  * We will set the state structure to point to each of the elements
1477  * in each of the nodes where either the hashval is or should be.
1478  *
1479  * We support duplicate hashval's so for each entry in the current
1480  * node that could contain the desired hashval, descend.  This is a
1481  * pruned depth-first tree search.
1482  */
1483 int                                                     /* error */
1484 xfs_da3_node_lookup_int(
1485         struct xfs_da_state     *state,
1486         int                     *result)
1487 {
1488         struct xfs_da_state_blk *blk;
1489         struct xfs_da_blkinfo   *curr;
1490         struct xfs_da_intnode   *node;
1491         struct xfs_da_node_entry *btree;
1492         struct xfs_da3_icnode_hdr nodehdr;
1493         struct xfs_da_args      *args;
1494         xfs_dablk_t             blkno;
1495         xfs_dahash_t            hashval;
1496         xfs_dahash_t            btreehashval;
1497         int                     probe;
1498         int                     span;
1499         int                     max;
1500         int                     error;
1501         int                     retval;
1502
1503         args = state->args;
1504
1505         /*
1506          * Descend thru the B-tree searching each level for the right
1507          * node to use, until the right hashval is found.
1508          */
1509         blkno = (args->whichfork == XFS_DATA_FORK)? state->mp->m_dirleafblk : 0;
1510         for (blk = &state->path.blk[0], state->path.active = 1;
1511                          state->path.active <= XFS_DA_NODE_MAXDEPTH;
1512                          blk++, state->path.active++) {
1513                 /*
1514                  * Read the next node down in the tree.
1515                  */
1516                 blk->blkno = blkno;
1517                 error = xfs_da3_node_read(args->trans, args->dp, blkno,
1518                                         -1, &blk->bp, args->whichfork);
1519                 if (error) {
1520                         blk->blkno = 0;
1521                         state->path.active--;
1522                         return(error);
1523                 }
1524                 curr = blk->bp->b_addr;
1525                 blk->magic = be16_to_cpu(curr->magic);
1526
1527                 if (blk->magic == XFS_ATTR_LEAF_MAGIC ||
1528                     blk->magic == XFS_ATTR3_LEAF_MAGIC) {
1529                         blk->magic = XFS_ATTR_LEAF_MAGIC;
1530                         blk->hashval = xfs_attr_leaf_lasthash(blk->bp, NULL);
1531                         break;
1532                 }
1533
1534                 if (blk->magic == XFS_DIR2_LEAFN_MAGIC ||
1535                     blk->magic == XFS_DIR3_LEAFN_MAGIC) {
1536                         blk->magic = XFS_DIR2_LEAFN_MAGIC;
1537                         blk->hashval = xfs_dir2_leafn_lasthash(blk->bp, NULL);
1538                         break;
1539                 }
1540
1541                 blk->magic = XFS_DA_NODE_MAGIC;
1542
1543
1544                 /*
1545                  * Search an intermediate node for a match.
1546                  */
1547                 node = blk->bp->b_addr;
1548                 xfs_da3_node_hdr_from_disk(&nodehdr, node);
1549                 btree = xfs_da3_node_tree_p(node);
1550
1551                 max = nodehdr.count;
1552                 blk->hashval = be32_to_cpu(btree[max - 1].hashval);
1553
1554                 /*
1555                  * Binary search.  (note: small blocks will skip loop)
1556                  */
1557                 probe = span = max / 2;
1558                 hashval = args->hashval;
1559                 while (span > 4) {
1560                         span /= 2;
1561                         btreehashval = be32_to_cpu(btree[probe].hashval);
1562                         if (btreehashval < hashval)
1563                                 probe += span;
1564                         else if (btreehashval > hashval)
1565                                 probe -= span;
1566                         else
1567                                 break;
1568                 }
1569                 ASSERT((probe >= 0) && (probe < max));
1570                 ASSERT((span <= 4) ||
1571                         (be32_to_cpu(btree[probe].hashval) == hashval));
1572
1573                 /*
1574                  * Since we may have duplicate hashval's, find the first
1575                  * matching hashval in the node.
1576                  */
1577                 while (probe > 0 &&
1578                        be32_to_cpu(btree[probe].hashval) >= hashval) {
1579                         probe--;
1580                 }
1581                 while (probe < max &&
1582                        be32_to_cpu(btree[probe].hashval) < hashval) {
1583                         probe++;
1584                 }
1585
1586                 /*
1587                  * Pick the right block to descend on.
1588                  */
1589                 if (probe == max) {
1590                         blk->index = max - 1;
1591                         blkno = be32_to_cpu(btree[max - 1].before);
1592                 } else {
1593                         blk->index = probe;
1594                         blkno = be32_to_cpu(btree[probe].before);
1595                 }
1596         }
1597
1598         /*
1599          * A leaf block that ends in the hashval that we are interested in
1600          * (final hashval == search hashval) means that the next block may
1601          * contain more entries with the same hashval, shift upward to the
1602          * next leaf and keep searching.
1603          */
1604         for (;;) {
1605                 if (blk->magic == XFS_DIR2_LEAFN_MAGIC) {
1606                         retval = xfs_dir2_leafn_lookup_int(blk->bp, args,
1607                                                         &blk->index, state);
1608                 } else if (blk->magic == XFS_ATTR_LEAF_MAGIC) {
1609                         retval = xfs_attr3_leaf_lookup_int(blk->bp, args);
1610                         blk->index = args->index;
1611                         args->blkno = blk->blkno;
1612                 } else {
1613                         ASSERT(0);
1614                         return XFS_ERROR(EFSCORRUPTED);
1615                 }
1616                 if (((retval == ENOENT) || (retval == ENOATTR)) &&
1617                     (blk->hashval == args->hashval)) {
1618                         error = xfs_da3_path_shift(state, &state->path, 1, 1,
1619                                                          &retval);
1620                         if (error)
1621                                 return(error);
1622                         if (retval == 0) {
1623                                 continue;
1624                         } else if (blk->magic == XFS_ATTR_LEAF_MAGIC) {
1625                                 /* path_shift() gives ENOENT */
1626                                 retval = XFS_ERROR(ENOATTR);
1627                         }
1628                 }
1629                 break;
1630         }
1631         *result = retval;
1632         return(0);
1633 }
1634
1635 /*========================================================================
1636  * Utility routines.
1637  *========================================================================*/
1638
1639 /*
1640  * Compare two intermediate nodes for "order".
1641  */
1642 STATIC int
1643 xfs_da3_node_order(
1644         struct xfs_buf  *node1_bp,
1645         struct xfs_buf  *node2_bp)
1646 {
1647         struct xfs_da_intnode   *node1;
1648         struct xfs_da_intnode   *node2;
1649         struct xfs_da_node_entry *btree1;
1650         struct xfs_da_node_entry *btree2;
1651         struct xfs_da3_icnode_hdr node1hdr;
1652         struct xfs_da3_icnode_hdr node2hdr;
1653
1654         node1 = node1_bp->b_addr;
1655         node2 = node2_bp->b_addr;
1656         xfs_da3_node_hdr_from_disk(&node1hdr, node1);
1657         xfs_da3_node_hdr_from_disk(&node2hdr, node2);
1658         btree1 = xfs_da3_node_tree_p(node1);
1659         btree2 = xfs_da3_node_tree_p(node2);
1660
1661         if (node1hdr.count > 0 && node2hdr.count > 0 &&
1662             ((be32_to_cpu(btree2[0].hashval) < be32_to_cpu(btree1[0].hashval)) ||
1663              (be32_to_cpu(btree2[node2hdr.count - 1].hashval) <
1664               be32_to_cpu(btree1[node1hdr.count - 1].hashval)))) {
1665                 return 1;
1666         }
1667         return 0;
1668 }
1669
1670 /*
1671  * Link a new block into a doubly linked list of blocks (of whatever type).
1672  */
1673 int                                                     /* error */
1674 xfs_da3_blk_link(
1675         struct xfs_da_state     *state,
1676         struct xfs_da_state_blk *old_blk,
1677         struct xfs_da_state_blk *new_blk)
1678 {
1679         struct xfs_da_blkinfo   *old_info;
1680         struct xfs_da_blkinfo   *new_info;
1681         struct xfs_da_blkinfo   *tmp_info;
1682         struct xfs_da_args      *args;
1683         struct xfs_buf          *bp;
1684         int                     before = 0;
1685         int                     error;
1686
1687         /*
1688          * Set up environment.
1689          */
1690         args = state->args;
1691         ASSERT(args != NULL);
1692         old_info = old_blk->bp->b_addr;
1693         new_info = new_blk->bp->b_addr;
1694         ASSERT(old_blk->magic == XFS_DA_NODE_MAGIC ||
1695                old_blk->magic == XFS_DIR2_LEAFN_MAGIC ||
1696                old_blk->magic == XFS_ATTR_LEAF_MAGIC);
1697
1698         switch (old_blk->magic) {
1699         case XFS_ATTR_LEAF_MAGIC:
1700                 before = xfs_attr_leaf_order(old_blk->bp, new_blk->bp);
1701                 break;
1702         case XFS_DIR2_LEAFN_MAGIC:
1703                 before = xfs_dir2_leafn_order(old_blk->bp, new_blk->bp);
1704                 break;
1705         case XFS_DA_NODE_MAGIC:
1706                 before = xfs_da3_node_order(old_blk->bp, new_blk->bp);
1707                 break;
1708         }
1709
1710         /*
1711          * Link blocks in appropriate order.
1712          */
1713         if (before) {
1714                 /*
1715                  * Link new block in before existing block.
1716                  */
1717                 trace_xfs_da_link_before(args);
1718                 new_info->forw = cpu_to_be32(old_blk->blkno);
1719                 new_info->back = old_info->back;
1720                 if (old_info->back) {
1721                         error = xfs_da3_node_read(args->trans, args->dp,
1722                                                 be32_to_cpu(old_info->back),
1723                                                 -1, &bp, args->whichfork);
1724                         if (error)
1725                                 return(error);
1726                         ASSERT(bp != NULL);
1727                         tmp_info = bp->b_addr;
1728                         ASSERT(tmp_info->magic == old_info->magic);
1729                         ASSERT(be32_to_cpu(tmp_info->forw) == old_blk->blkno);
1730                         tmp_info->forw = cpu_to_be32(new_blk->blkno);
1731                         xfs_trans_log_buf(args->trans, bp, 0, sizeof(*tmp_info)-1);
1732                 }
1733                 old_info->back = cpu_to_be32(new_blk->blkno);
1734         } else {
1735                 /*
1736                  * Link new block in after existing block.
1737                  */
1738                 trace_xfs_da_link_after(args);
1739                 new_info->forw = old_info->forw;
1740                 new_info->back = cpu_to_be32(old_blk->blkno);
1741                 if (old_info->forw) {
1742                         error = xfs_da3_node_read(args->trans, args->dp,
1743                                                 be32_to_cpu(old_info->forw),
1744                                                 -1, &bp, args->whichfork);
1745                         if (error)
1746                                 return(error);
1747                         ASSERT(bp != NULL);
1748                         tmp_info = bp->b_addr;
1749                         ASSERT(tmp_info->magic == old_info->magic);
1750                         ASSERT(be32_to_cpu(tmp_info->back) == old_blk->blkno);
1751                         tmp_info->back = cpu_to_be32(new_blk->blkno);
1752                         xfs_trans_log_buf(args->trans, bp, 0, sizeof(*tmp_info)-1);
1753                 }
1754                 old_info->forw = cpu_to_be32(new_blk->blkno);
1755         }
1756
1757         xfs_trans_log_buf(args->trans, old_blk->bp, 0, sizeof(*tmp_info) - 1);
1758         xfs_trans_log_buf(args->trans, new_blk->bp, 0, sizeof(*tmp_info) - 1);
1759         return(0);
1760 }
1761
1762 /*
1763  * Unlink a block from a doubly linked list of blocks.
1764  */
1765 STATIC int                                              /* error */
1766 xfs_da3_blk_unlink(
1767         struct xfs_da_state     *state,
1768         struct xfs_da_state_blk *drop_blk,
1769         struct xfs_da_state_blk *save_blk)
1770 {
1771         struct xfs_da_blkinfo   *drop_info;
1772         struct xfs_da_blkinfo   *save_info;
1773         struct xfs_da_blkinfo   *tmp_info;
1774         struct xfs_da_args      *args;
1775         struct xfs_buf          *bp;
1776         int                     error;
1777
1778         /*
1779          * Set up environment.
1780          */
1781         args = state->args;
1782         ASSERT(args != NULL);
1783         save_info = save_blk->bp->b_addr;
1784         drop_info = drop_blk->bp->b_addr;
1785         ASSERT(save_blk->magic == XFS_DA_NODE_MAGIC ||
1786                save_blk->magic == XFS_DIR2_LEAFN_MAGIC ||
1787                save_blk->magic == XFS_ATTR_LEAF_MAGIC);
1788         ASSERT(save_blk->magic == drop_blk->magic);
1789         ASSERT((be32_to_cpu(save_info->forw) == drop_blk->blkno) ||
1790                (be32_to_cpu(save_info->back) == drop_blk->blkno));
1791         ASSERT((be32_to_cpu(drop_info->forw) == save_blk->blkno) ||
1792                (be32_to_cpu(drop_info->back) == save_blk->blkno));
1793
1794         /*
1795          * Unlink the leaf block from the doubly linked chain of leaves.
1796          */
1797         if (be32_to_cpu(save_info->back) == drop_blk->blkno) {
1798                 trace_xfs_da_unlink_back(args);
1799                 save_info->back = drop_info->back;
1800                 if (drop_info->back) {
1801                         error = xfs_da3_node_read(args->trans, args->dp,
1802                                                 be32_to_cpu(drop_info->back),
1803                                                 -1, &bp, args->whichfork);
1804                         if (error)
1805                                 return(error);
1806                         ASSERT(bp != NULL);
1807                         tmp_info = bp->b_addr;
1808                         ASSERT(tmp_info->magic == save_info->magic);
1809                         ASSERT(be32_to_cpu(tmp_info->forw) == drop_blk->blkno);
1810                         tmp_info->forw = cpu_to_be32(save_blk->blkno);
1811                         xfs_trans_log_buf(args->trans, bp, 0,
1812                                                     sizeof(*tmp_info) - 1);
1813                 }
1814         } else {
1815                 trace_xfs_da_unlink_forward(args);
1816                 save_info->forw = drop_info->forw;
1817                 if (drop_info->forw) {
1818                         error = xfs_da3_node_read(args->trans, args->dp,
1819                                                 be32_to_cpu(drop_info->forw),
1820                                                 -1, &bp, args->whichfork);
1821                         if (error)
1822                                 return(error);
1823                         ASSERT(bp != NULL);
1824                         tmp_info = bp->b_addr;
1825                         ASSERT(tmp_info->magic == save_info->magic);
1826                         ASSERT(be32_to_cpu(tmp_info->back) == drop_blk->blkno);
1827                         tmp_info->back = cpu_to_be32(save_blk->blkno);
1828                         xfs_trans_log_buf(args->trans, bp, 0,
1829                                                     sizeof(*tmp_info) - 1);
1830                 }
1831         }
1832
1833         xfs_trans_log_buf(args->trans, save_blk->bp, 0, sizeof(*save_info) - 1);
1834         return(0);
1835 }
1836
1837 /*
1838  * Move a path "forward" or "!forward" one block at the current level.
1839  *
1840  * This routine will adjust a "path" to point to the next block
1841  * "forward" (higher hashvalues) or "!forward" (lower hashvals) in the
1842  * Btree, including updating pointers to the intermediate nodes between
1843  * the new bottom and the root.
1844  */
1845 int                                                     /* error */
1846 xfs_da3_path_shift(
1847         struct xfs_da_state     *state,
1848         struct xfs_da_state_path *path,
1849         int                     forward,
1850         int                     release,
1851         int                     *result)
1852 {
1853         struct xfs_da_state_blk *blk;
1854         struct xfs_da_blkinfo   *info;
1855         struct xfs_da_intnode   *node;
1856         struct xfs_da_args      *args;
1857         struct xfs_da_node_entry *btree;
1858         struct xfs_da3_icnode_hdr nodehdr;
1859         xfs_dablk_t             blkno = 0;
1860         int                     level;
1861         int                     error;
1862
1863         trace_xfs_da_path_shift(state->args);
1864
1865         /*
1866          * Roll up the Btree looking for the first block where our
1867          * current index is not at the edge of the block.  Note that
1868          * we skip the bottom layer because we want the sibling block.
1869          */
1870         args = state->args;
1871         ASSERT(args != NULL);
1872         ASSERT(path != NULL);
1873         ASSERT((path->active > 0) && (path->active < XFS_DA_NODE_MAXDEPTH));
1874         level = (path->active-1) - 1;   /* skip bottom layer in path */
1875         for (blk = &path->blk[level]; level >= 0; blk--, level--) {
1876                 node = blk->bp->b_addr;
1877                 xfs_da3_node_hdr_from_disk(&nodehdr, node);
1878                 btree = xfs_da3_node_tree_p(node);
1879
1880                 if (forward && (blk->index < nodehdr.count - 1)) {
1881                         blk->index++;
1882                         blkno = be32_to_cpu(btree[blk->index].before);
1883                         break;
1884                 } else if (!forward && (blk->index > 0)) {
1885                         blk->index--;
1886                         blkno = be32_to_cpu(btree[blk->index].before);
1887                         break;
1888                 }
1889         }
1890         if (level < 0) {
1891                 *result = XFS_ERROR(ENOENT);    /* we're out of our tree */
1892                 ASSERT(args->op_flags & XFS_DA_OP_OKNOENT);
1893                 return(0);
1894         }
1895
1896         /*
1897          * Roll down the edge of the subtree until we reach the
1898          * same depth we were at originally.
1899          */
1900         for (blk++, level++; level < path->active; blk++, level++) {
1901                 /*
1902                  * Release the old block.
1903                  * (if it's dirty, trans won't actually let go)
1904                  */
1905                 if (release)
1906                         xfs_trans_brelse(args->trans, blk->bp);
1907
1908                 /*
1909                  * Read the next child block.
1910                  */
1911                 blk->blkno = blkno;
1912                 error = xfs_da3_node_read(args->trans, args->dp, blkno, -1,
1913                                         &blk->bp, args->whichfork);
1914                 if (error)
1915                         return(error);
1916                 info = blk->bp->b_addr;
1917                 ASSERT(info->magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
1918                        info->magic == cpu_to_be16(XFS_DA3_NODE_MAGIC) ||
1919                        info->magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
1920                        info->magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC) ||
1921                        info->magic == cpu_to_be16(XFS_ATTR_LEAF_MAGIC) ||
1922                        info->magic == cpu_to_be16(XFS_ATTR3_LEAF_MAGIC));
1923
1924
1925                 /*
1926                  * Note: we flatten the magic number to a single type so we
1927                  * don't have to compare against crc/non-crc types elsewhere.
1928                  */
1929                 switch (be16_to_cpu(info->magic)) {
1930                 case XFS_DA_NODE_MAGIC:
1931                 case XFS_DA3_NODE_MAGIC:
1932                         blk->magic = XFS_DA_NODE_MAGIC;
1933                         node = (xfs_da_intnode_t *)info;
1934                         xfs_da3_node_hdr_from_disk(&nodehdr, node);
1935                         btree = xfs_da3_node_tree_p(node);
1936                         blk->hashval = be32_to_cpu(btree[nodehdr.count - 1].hashval);
1937                         if (forward)
1938                                 blk->index = 0;
1939                         else
1940                                 blk->index = nodehdr.count - 1;
1941                         blkno = be32_to_cpu(btree[blk->index].before);
1942                         break;
1943                 case XFS_ATTR_LEAF_MAGIC:
1944                 case XFS_ATTR3_LEAF_MAGIC:
1945                         blk->magic = XFS_ATTR_LEAF_MAGIC;
1946                         ASSERT(level == path->active-1);
1947                         blk->index = 0;
1948                         blk->hashval = xfs_attr_leaf_lasthash(blk->bp,
1949                                                               NULL);
1950                         break;
1951                 case XFS_DIR2_LEAFN_MAGIC:
1952                 case XFS_DIR3_LEAFN_MAGIC:
1953                         blk->magic = XFS_DIR2_LEAFN_MAGIC;
1954                         ASSERT(level == path->active-1);
1955                         blk->index = 0;
1956                         blk->hashval = xfs_dir2_leafn_lasthash(blk->bp,
1957                                                                NULL);
1958                         break;
1959                 default:
1960                         ASSERT(0);
1961                         break;
1962                 }
1963         }
1964         *result = 0;
1965         return 0;
1966 }
1967
1968
1969 /*========================================================================
1970  * Utility routines.
1971  *========================================================================*/
1972
1973 /*
1974  * Implement a simple hash on a character string.
1975  * Rotate the hash value by 7 bits, then XOR each character in.
1976  * This is implemented with some source-level loop unrolling.
1977  */
1978 xfs_dahash_t
1979 xfs_da_hashname(const __uint8_t *name, int namelen)
1980 {
1981         xfs_dahash_t hash;
1982
1983         /*
1984          * Do four characters at a time as long as we can.
1985          */
1986         for (hash = 0; namelen >= 4; namelen -= 4, name += 4)
1987                 hash = (name[0] << 21) ^ (name[1] << 14) ^ (name[2] << 7) ^
1988                        (name[3] << 0) ^ rol32(hash, 7 * 4);
1989
1990         /*
1991          * Now do the rest of the characters.
1992          */
1993         switch (namelen) {
1994         case 3:
1995                 return (name[0] << 14) ^ (name[1] << 7) ^ (name[2] << 0) ^
1996                        rol32(hash, 7 * 3);
1997         case 2:
1998                 return (name[0] << 7) ^ (name[1] << 0) ^ rol32(hash, 7 * 2);
1999         case 1:
2000                 return (name[0] << 0) ^ rol32(hash, 7 * 1);
2001         default: /* case 0: */
2002                 return hash;
2003         }
2004 }
2005
2006 enum xfs_dacmp
2007 xfs_da_compname(
2008         struct xfs_da_args *args,
2009         const unsigned char *name,
2010         int             len)
2011 {
2012         return (args->namelen == len && memcmp(args->name, name, len) == 0) ?
2013                                         XFS_CMP_EXACT : XFS_CMP_DIFFERENT;
2014 }
2015
2016 static xfs_dahash_t
2017 xfs_default_hashname(
2018         struct xfs_name *name)
2019 {
2020         return xfs_da_hashname(name->name, name->len);
2021 }
2022
2023 const struct xfs_nameops xfs_default_nameops = {
2024         .hashname       = xfs_default_hashname,
2025         .compname       = xfs_da_compname
2026 };
2027
2028 int
2029 xfs_da_grow_inode_int(
2030         struct xfs_da_args      *args,
2031         xfs_fileoff_t           *bno,
2032         int                     count)
2033 {
2034         struct xfs_trans        *tp = args->trans;
2035         struct xfs_inode        *dp = args->dp;
2036         int                     w = args->whichfork;
2037         xfs_drfsbno_t           nblks = dp->i_d.di_nblocks;
2038         struct xfs_bmbt_irec    map, *mapp;
2039         int                     nmap, error, got, i, mapi;
2040
2041         /*
2042          * Find a spot in the file space to put the new block.
2043          */
2044         error = xfs_bmap_first_unused(tp, dp, count, bno, w);
2045         if (error)
2046                 return error;
2047
2048         /*
2049          * Try mapping it in one filesystem block.
2050          */
2051         nmap = 1;
2052         ASSERT(args->firstblock != NULL);
2053         error = xfs_bmapi_write(tp, dp, *bno, count,
2054                         xfs_bmapi_aflag(w)|XFS_BMAPI_METADATA|XFS_BMAPI_CONTIG,
2055                         args->firstblock, args->total, &map, &nmap,
2056                         args->flist);
2057         if (error)
2058                 return error;
2059
2060         ASSERT(nmap <= 1);
2061         if (nmap == 1) {
2062                 mapp = &map;
2063                 mapi = 1;
2064         } else if (nmap == 0 && count > 1) {
2065                 xfs_fileoff_t           b;
2066                 int                     c;
2067
2068                 /*
2069                  * If we didn't get it and the block might work if fragmented,
2070                  * try without the CONTIG flag.  Loop until we get it all.
2071                  */
2072                 mapp = kmem_alloc(sizeof(*mapp) * count, KM_SLEEP);
2073                 for (b = *bno, mapi = 0; b < *bno + count; ) {
2074                         nmap = MIN(XFS_BMAP_MAX_NMAP, count);
2075                         c = (int)(*bno + count - b);
2076                         error = xfs_bmapi_write(tp, dp, b, c,
2077                                         xfs_bmapi_aflag(w)|XFS_BMAPI_METADATA,
2078                                         args->firstblock, args->total,
2079                                         &mapp[mapi], &nmap, args->flist);
2080                         if (error)
2081                                 goto out_free_map;
2082                         if (nmap < 1)
2083                                 break;
2084                         mapi += nmap;
2085                         b = mapp[mapi - 1].br_startoff +
2086                             mapp[mapi - 1].br_blockcount;
2087                 }
2088         } else {
2089                 mapi = 0;
2090                 mapp = NULL;
2091         }
2092
2093         /*
2094          * Count the blocks we got, make sure it matches the total.
2095          */
2096         for (i = 0, got = 0; i < mapi; i++)
2097                 got += mapp[i].br_blockcount;
2098         if (got != count || mapp[0].br_startoff != *bno ||
2099             mapp[mapi - 1].br_startoff + mapp[mapi - 1].br_blockcount !=
2100             *bno + count) {
2101                 error = XFS_ERROR(ENOSPC);
2102                 goto out_free_map;
2103         }
2104
2105         /* account for newly allocated blocks in reserved blocks total */
2106         args->total -= dp->i_d.di_nblocks - nblks;
2107
2108 out_free_map:
2109         if (mapp != &map)
2110                 kmem_free(mapp);
2111         return error;
2112 }
2113
2114 /*
2115  * Add a block to the btree ahead of the file.
2116  * Return the new block number to the caller.
2117  */
2118 int
2119 xfs_da_grow_inode(
2120         struct xfs_da_args      *args,
2121         xfs_dablk_t             *new_blkno)
2122 {
2123         xfs_fileoff_t           bno;
2124         int                     count;
2125         int                     error;
2126
2127         trace_xfs_da_grow_inode(args);
2128
2129         if (args->whichfork == XFS_DATA_FORK) {
2130                 bno = args->dp->i_mount->m_dirleafblk;
2131                 count = args->dp->i_mount->m_dirblkfsbs;
2132         } else {
2133                 bno = 0;
2134                 count = 1;
2135         }
2136
2137         error = xfs_da_grow_inode_int(args, &bno, count);
2138         if (!error)
2139                 *new_blkno = (xfs_dablk_t)bno;
2140         return error;
2141 }
2142
2143 /*
2144  * Ick.  We need to always be able to remove a btree block, even
2145  * if there's no space reservation because the filesystem is full.
2146  * This is called if xfs_bunmapi on a btree block fails due to ENOSPC.
2147  * It swaps the target block with the last block in the file.  The
2148  * last block in the file can always be removed since it can't cause
2149  * a bmap btree split to do that.
2150  */
2151 STATIC int
2152 xfs_da3_swap_lastblock(
2153         struct xfs_da_args      *args,
2154         xfs_dablk_t             *dead_blknop,
2155         struct xfs_buf          **dead_bufp)
2156 {
2157         struct xfs_da_blkinfo   *dead_info;
2158         struct xfs_da_blkinfo   *sib_info;
2159         struct xfs_da_intnode   *par_node;
2160         struct xfs_da_intnode   *dead_node;
2161         struct xfs_dir2_leaf    *dead_leaf2;
2162         struct xfs_da_node_entry *btree;
2163         struct xfs_da3_icnode_hdr par_hdr;
2164         struct xfs_inode        *ip;
2165         struct xfs_trans        *tp;
2166         struct xfs_mount        *mp;
2167         struct xfs_buf          *dead_buf;
2168         struct xfs_buf          *last_buf;
2169         struct xfs_buf          *sib_buf;
2170         struct xfs_buf          *par_buf;
2171         xfs_dahash_t            dead_hash;
2172         xfs_fileoff_t           lastoff;
2173         xfs_dablk_t             dead_blkno;
2174         xfs_dablk_t             last_blkno;
2175         xfs_dablk_t             sib_blkno;
2176         xfs_dablk_t             par_blkno;
2177         int                     error;
2178         int                     w;
2179         int                     entno;
2180         int                     level;
2181         int                     dead_level;
2182
2183         trace_xfs_da_swap_lastblock(args);
2184
2185         dead_buf = *dead_bufp;
2186         dead_blkno = *dead_blknop;
2187         tp = args->trans;
2188         ip = args->dp;
2189         w = args->whichfork;
2190         ASSERT(w == XFS_DATA_FORK);
2191         mp = ip->i_mount;
2192         lastoff = mp->m_dirfreeblk;
2193         error = xfs_bmap_last_before(tp, ip, &lastoff, w);
2194         if (error)
2195                 return error;
2196         if (unlikely(lastoff == 0)) {
2197                 XFS_ERROR_REPORT("xfs_da_swap_lastblock(1)", XFS_ERRLEVEL_LOW,
2198                                  mp);
2199                 return XFS_ERROR(EFSCORRUPTED);
2200         }
2201         /*
2202          * Read the last block in the btree space.
2203          */
2204         last_blkno = (xfs_dablk_t)lastoff - mp->m_dirblkfsbs;
2205         error = xfs_da3_node_read(tp, ip, last_blkno, -1, &last_buf, w);
2206         if (error)
2207                 return error;
2208         /*
2209          * Copy the last block into the dead buffer and log it.
2210          */
2211         memcpy(dead_buf->b_addr, last_buf->b_addr, mp->m_dirblksize);
2212         xfs_trans_log_buf(tp, dead_buf, 0, mp->m_dirblksize - 1);
2213         dead_info = dead_buf->b_addr;
2214         /*
2215          * Get values from the moved block.
2216          */
2217         if (dead_info->magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
2218             dead_info->magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
2219                 struct xfs_dir3_icleaf_hdr leafhdr;
2220                 struct xfs_dir2_leaf_entry *ents;
2221
2222                 dead_leaf2 = (xfs_dir2_leaf_t *)dead_info;
2223                 xfs_dir3_leaf_hdr_from_disk(&leafhdr, dead_leaf2);
2224                 ents = xfs_dir3_leaf_ents_p(dead_leaf2);
2225                 dead_level = 0;
2226                 dead_hash = be32_to_cpu(ents[leafhdr.count - 1].hashval);
2227         } else {
2228                 struct xfs_da3_icnode_hdr deadhdr;
2229
2230                 dead_node = (xfs_da_intnode_t *)dead_info;
2231                 xfs_da3_node_hdr_from_disk(&deadhdr, dead_node);
2232                 btree = xfs_da3_node_tree_p(dead_node);
2233                 dead_level = deadhdr.level;
2234                 dead_hash = be32_to_cpu(btree[deadhdr.count - 1].hashval);
2235         }
2236         sib_buf = par_buf = NULL;
2237         /*
2238          * If the moved block has a left sibling, fix up the pointers.
2239          */
2240         if ((sib_blkno = be32_to_cpu(dead_info->back))) {
2241                 error = xfs_da3_node_read(tp, ip, sib_blkno, -1, &sib_buf, w);
2242                 if (error)
2243                         goto done;
2244                 sib_info = sib_buf->b_addr;
2245                 if (unlikely(
2246                     be32_to_cpu(sib_info->forw) != last_blkno ||
2247                     sib_info->magic != dead_info->magic)) {
2248                         XFS_ERROR_REPORT("xfs_da_swap_lastblock(2)",
2249                                          XFS_ERRLEVEL_LOW, mp);
2250                         error = XFS_ERROR(EFSCORRUPTED);
2251                         goto done;
2252                 }
2253                 sib_info->forw = cpu_to_be32(dead_blkno);
2254                 xfs_trans_log_buf(tp, sib_buf,
2255                         XFS_DA_LOGRANGE(sib_info, &sib_info->forw,
2256                                         sizeof(sib_info->forw)));
2257                 sib_buf = NULL;
2258         }
2259         /*
2260          * If the moved block has a right sibling, fix up the pointers.
2261          */
2262         if ((sib_blkno = be32_to_cpu(dead_info->forw))) {
2263                 error = xfs_da3_node_read(tp, ip, sib_blkno, -1, &sib_buf, w);
2264                 if (error)
2265                         goto done;
2266                 sib_info = sib_buf->b_addr;
2267                 if (unlikely(
2268                        be32_to_cpu(sib_info->back) != last_blkno ||
2269                        sib_info->magic != dead_info->magic)) {
2270                         XFS_ERROR_REPORT("xfs_da_swap_lastblock(3)",
2271                                          XFS_ERRLEVEL_LOW, mp);
2272                         error = XFS_ERROR(EFSCORRUPTED);
2273                         goto done;
2274                 }
2275                 sib_info->back = cpu_to_be32(dead_blkno);
2276                 xfs_trans_log_buf(tp, sib_buf,
2277                         XFS_DA_LOGRANGE(sib_info, &sib_info->back,
2278                                         sizeof(sib_info->back)));
2279                 sib_buf = NULL;
2280         }
2281         par_blkno = mp->m_dirleafblk;
2282         level = -1;
2283         /*
2284          * Walk down the tree looking for the parent of the moved block.
2285          */
2286         for (;;) {
2287                 error = xfs_da3_node_read(tp, ip, par_blkno, -1, &par_buf, w);
2288                 if (error)
2289                         goto done;
2290                 par_node = par_buf->b_addr;
2291                 xfs_da3_node_hdr_from_disk(&par_hdr, par_node);
2292                 if (level >= 0 && level != par_hdr.level + 1) {
2293                         XFS_ERROR_REPORT("xfs_da_swap_lastblock(4)",
2294                                          XFS_ERRLEVEL_LOW, mp);
2295                         error = XFS_ERROR(EFSCORRUPTED);
2296                         goto done;
2297                 }
2298                 level = par_hdr.level;
2299                 btree = xfs_da3_node_tree_p(par_node);
2300                 for (entno = 0;
2301                      entno < par_hdr.count &&
2302                      be32_to_cpu(btree[entno].hashval) < dead_hash;
2303                      entno++)
2304                         continue;
2305                 if (entno == par_hdr.count) {
2306                         XFS_ERROR_REPORT("xfs_da_swap_lastblock(5)",
2307                                          XFS_ERRLEVEL_LOW, mp);
2308                         error = XFS_ERROR(EFSCORRUPTED);
2309                         goto done;
2310                 }
2311                 par_blkno = be32_to_cpu(btree[entno].before);
2312                 if (level == dead_level + 1)
2313                         break;
2314                 xfs_trans_brelse(tp, par_buf);
2315                 par_buf = NULL;
2316         }
2317         /*
2318          * We're in the right parent block.
2319          * Look for the right entry.
2320          */
2321         for (;;) {
2322                 for (;
2323                      entno < par_hdr.count &&
2324                      be32_to_cpu(btree[entno].before) != last_blkno;
2325                      entno++)
2326                         continue;
2327                 if (entno < par_hdr.count)
2328                         break;
2329                 par_blkno = par_hdr.forw;
2330                 xfs_trans_brelse(tp, par_buf);
2331                 par_buf = NULL;
2332                 if (unlikely(par_blkno == 0)) {
2333                         XFS_ERROR_REPORT("xfs_da_swap_lastblock(6)",
2334                                          XFS_ERRLEVEL_LOW, mp);
2335                         error = XFS_ERROR(EFSCORRUPTED);
2336                         goto done;
2337                 }
2338                 error = xfs_da3_node_read(tp, ip, par_blkno, -1, &par_buf, w);
2339                 if (error)
2340                         goto done;
2341                 par_node = par_buf->b_addr;
2342                 xfs_da3_node_hdr_from_disk(&par_hdr, par_node);
2343                 if (par_hdr.level != level) {
2344                         XFS_ERROR_REPORT("xfs_da_swap_lastblock(7)",
2345                                          XFS_ERRLEVEL_LOW, mp);
2346                         error = XFS_ERROR(EFSCORRUPTED);
2347                         goto done;
2348                 }
2349                 btree = xfs_da3_node_tree_p(par_node);
2350                 entno = 0;
2351         }
2352         /*
2353          * Update the parent entry pointing to the moved block.
2354          */
2355         btree[entno].before = cpu_to_be32(dead_blkno);
2356         xfs_trans_log_buf(tp, par_buf,
2357                 XFS_DA_LOGRANGE(par_node, &btree[entno].before,
2358                                 sizeof(btree[entno].before)));
2359         *dead_blknop = last_blkno;
2360         *dead_bufp = last_buf;
2361         return 0;
2362 done:
2363         if (par_buf)
2364                 xfs_trans_brelse(tp, par_buf);
2365         if (sib_buf)
2366                 xfs_trans_brelse(tp, sib_buf);
2367         xfs_trans_brelse(tp, last_buf);
2368         return error;
2369 }
2370
2371 /*
2372  * Remove a btree block from a directory or attribute.
2373  */
2374 int
2375 xfs_da_shrink_inode(
2376         xfs_da_args_t   *args,
2377         xfs_dablk_t     dead_blkno,
2378         struct xfs_buf  *dead_buf)
2379 {
2380         xfs_inode_t *dp;
2381         int done, error, w, count;
2382         xfs_trans_t *tp;
2383         xfs_mount_t *mp;
2384
2385         trace_xfs_da_shrink_inode(args);
2386
2387         dp = args->dp;
2388         w = args->whichfork;
2389         tp = args->trans;
2390         mp = dp->i_mount;
2391         if (w == XFS_DATA_FORK)
2392                 count = mp->m_dirblkfsbs;
2393         else
2394                 count = 1;
2395         for (;;) {
2396                 /*
2397                  * Remove extents.  If we get ENOSPC for a dir we have to move
2398                  * the last block to the place we want to kill.
2399                  */
2400                 error = xfs_bunmapi(tp, dp, dead_blkno, count,
2401                                     xfs_bmapi_aflag(w)|XFS_BMAPI_METADATA,
2402                                     0, args->firstblock, args->flist, &done);
2403                 if (error == ENOSPC) {
2404                         if (w != XFS_DATA_FORK)
2405                                 break;
2406                         error = xfs_da3_swap_lastblock(args, &dead_blkno,
2407                                                       &dead_buf);
2408                         if (error)
2409                                 break;
2410                 } else {
2411                         break;
2412                 }
2413         }
2414         xfs_trans_binval(tp, dead_buf);
2415         return error;
2416 }
2417
2418 /*
2419  * See if the mapping(s) for this btree block are valid, i.e.
2420  * don't contain holes, are logically contiguous, and cover the whole range.
2421  */
2422 STATIC int
2423 xfs_da_map_covers_blocks(
2424         int             nmap,
2425         xfs_bmbt_irec_t *mapp,
2426         xfs_dablk_t     bno,
2427         int             count)
2428 {
2429         int             i;
2430         xfs_fileoff_t   off;
2431
2432         for (i = 0, off = bno; i < nmap; i++) {
2433                 if (mapp[i].br_startblock == HOLESTARTBLOCK ||
2434                     mapp[i].br_startblock == DELAYSTARTBLOCK) {
2435                         return 0;
2436                 }
2437                 if (off != mapp[i].br_startoff) {
2438                         return 0;
2439                 }
2440                 off += mapp[i].br_blockcount;
2441         }
2442         return off == bno + count;
2443 }
2444
2445 /*
2446  * Convert a struct xfs_bmbt_irec to a struct xfs_buf_map.
2447  *
2448  * For the single map case, it is assumed that the caller has provided a pointer
2449  * to a valid xfs_buf_map.  For the multiple map case, this function will
2450  * allocate the xfs_buf_map to hold all the maps and replace the caller's single
2451  * map pointer with the allocated map.
2452  */
2453 static int
2454 xfs_buf_map_from_irec(
2455         struct xfs_mount        *mp,
2456         struct xfs_buf_map      **mapp,
2457         unsigned int            *nmaps,
2458         struct xfs_bmbt_irec    *irecs,
2459         unsigned int            nirecs)
2460 {
2461         struct xfs_buf_map      *map;
2462         int                     i;
2463
2464         ASSERT(*nmaps == 1);
2465         ASSERT(nirecs >= 1);
2466
2467         if (nirecs > 1) {
2468                 map = kmem_zalloc(nirecs * sizeof(struct xfs_buf_map),
2469                                   KM_SLEEP | KM_NOFS);
2470                 if (!map)
2471                         return ENOMEM;
2472                 *mapp = map;
2473         }
2474
2475         *nmaps = nirecs;
2476         map = *mapp;
2477         for (i = 0; i < *nmaps; i++) {
2478                 ASSERT(irecs[i].br_startblock != DELAYSTARTBLOCK &&
2479                        irecs[i].br_startblock != HOLESTARTBLOCK);
2480                 map[i].bm_bn = XFS_FSB_TO_DADDR(mp, irecs[i].br_startblock);
2481                 map[i].bm_len = XFS_FSB_TO_BB(mp, irecs[i].br_blockcount);
2482         }
2483         return 0;
2484 }
2485
2486 /*
2487  * Map the block we are given ready for reading. There are three possible return
2488  * values:
2489  *      -1 - will be returned if we land in a hole and mappedbno == -2 so the
2490  *           caller knows not to execute a subsequent read.
2491  *       0 - if we mapped the block successfully
2492  *      >0 - positive error number if there was an error.
2493  */
2494 static int
2495 xfs_dabuf_map(
2496         struct xfs_trans        *trans,
2497         struct xfs_inode        *dp,
2498         xfs_dablk_t             bno,
2499         xfs_daddr_t             mappedbno,
2500         int                     whichfork,
2501         struct xfs_buf_map      **map,
2502         int                     *nmaps)
2503 {
2504         struct xfs_mount        *mp = dp->i_mount;
2505         int                     nfsb;
2506         int                     error = 0;
2507         struct xfs_bmbt_irec    irec;
2508         struct xfs_bmbt_irec    *irecs = &irec;
2509         int                     nirecs;
2510
2511         ASSERT(map && *map);
2512         ASSERT(*nmaps == 1);
2513
2514         nfsb = (whichfork == XFS_DATA_FORK) ? mp->m_dirblkfsbs : 1;
2515
2516         /*
2517          * Caller doesn't have a mapping.  -2 means don't complain
2518          * if we land in a hole.
2519          */
2520         if (mappedbno == -1 || mappedbno == -2) {
2521                 /*
2522                  * Optimize the one-block case.
2523                  */
2524                 if (nfsb != 1)
2525                         irecs = kmem_zalloc(sizeof(irec) * nfsb,
2526                                             KM_SLEEP | KM_NOFS);
2527
2528                 nirecs = nfsb;
2529                 error = xfs_bmapi_read(dp, (xfs_fileoff_t)bno, nfsb, irecs,
2530                                        &nirecs, xfs_bmapi_aflag(whichfork));
2531                 if (error)
2532                         goto out;
2533         } else {
2534                 irecs->br_startblock = XFS_DADDR_TO_FSB(mp, mappedbno);
2535                 irecs->br_startoff = (xfs_fileoff_t)bno;
2536                 irecs->br_blockcount = nfsb;
2537                 irecs->br_state = 0;
2538                 nirecs = 1;
2539         }
2540
2541         if (!xfs_da_map_covers_blocks(nirecs, irecs, bno, nfsb)) {
2542                 error = mappedbno == -2 ? -1 : XFS_ERROR(EFSCORRUPTED);
2543                 if (unlikely(error == EFSCORRUPTED)) {
2544                         if (xfs_error_level >= XFS_ERRLEVEL_LOW) {
2545                                 int i;
2546                                 xfs_alert(mp, "%s: bno %lld dir: inode %lld",
2547                                         __func__, (long long)bno,
2548                                         (long long)dp->i_ino);
2549                                 for (i = 0; i < *nmaps; i++) {
2550                                         xfs_alert(mp,
2551 "[%02d] br_startoff %lld br_startblock %lld br_blockcount %lld br_state %d",
2552                                                 i,
2553                                                 (long long)irecs[i].br_startoff,
2554                                                 (long long)irecs[i].br_startblock,
2555                                                 (long long)irecs[i].br_blockcount,
2556                                                 irecs[i].br_state);
2557                                 }
2558                         }
2559                         XFS_ERROR_REPORT("xfs_da_do_buf(1)",
2560                                          XFS_ERRLEVEL_LOW, mp);
2561                 }
2562                 goto out;
2563         }
2564         error = xfs_buf_map_from_irec(mp, map, nmaps, irecs, nirecs);
2565 out:
2566         if (irecs != &irec)
2567                 kmem_free(irecs);
2568         return error;
2569 }
2570
2571 /*
2572  * Get a buffer for the dir/attr block.
2573  */
2574 int
2575 xfs_da_get_buf(
2576         struct xfs_trans        *trans,
2577         struct xfs_inode        *dp,
2578         xfs_dablk_t             bno,
2579         xfs_daddr_t             mappedbno,
2580         struct xfs_buf          **bpp,
2581         int                     whichfork)
2582 {
2583         struct xfs_buf          *bp;
2584         struct xfs_buf_map      map;
2585         struct xfs_buf_map      *mapp;
2586         int                     nmap;
2587         int                     error;
2588
2589         *bpp = NULL;
2590         mapp = &map;
2591         nmap = 1;
2592         error = xfs_dabuf_map(trans, dp, bno, mappedbno, whichfork,
2593                                 &mapp, &nmap);
2594         if (error) {
2595                 /* mapping a hole is not an error, but we don't continue */
2596                 if (error == -1)
2597                         error = 0;
2598                 goto out_free;
2599         }
2600
2601         bp = xfs_trans_get_buf_map(trans, dp->i_mount->m_ddev_targp,
2602                                     mapp, nmap, 0);
2603         error = bp ? bp->b_error : XFS_ERROR(EIO);
2604         if (error) {
2605                 xfs_trans_brelse(trans, bp);
2606                 goto out_free;
2607         }
2608
2609         *bpp = bp;
2610
2611 out_free:
2612         if (mapp != &map)
2613                 kmem_free(mapp);
2614
2615         return error;
2616 }
2617
2618 /*
2619  * Get a buffer for the dir/attr block, fill in the contents.
2620  */
2621 int
2622 xfs_da_read_buf(
2623         struct xfs_trans        *trans,
2624         struct xfs_inode        *dp,
2625         xfs_dablk_t             bno,
2626         xfs_daddr_t             mappedbno,
2627         struct xfs_buf          **bpp,
2628         int                     whichfork,
2629         const struct xfs_buf_ops *ops)
2630 {
2631         struct xfs_buf          *bp;
2632         struct xfs_buf_map      map;
2633         struct xfs_buf_map      *mapp;
2634         int                     nmap;
2635         int                     error;
2636
2637         *bpp = NULL;
2638         mapp = &map;
2639         nmap = 1;
2640         error = xfs_dabuf_map(trans, dp, bno, mappedbno, whichfork,
2641                                 &mapp, &nmap);
2642         if (error) {
2643                 /* mapping a hole is not an error, but we don't continue */
2644                 if (error == -1)
2645                         error = 0;
2646                 goto out_free;
2647         }
2648
2649         error = xfs_trans_read_buf_map(dp->i_mount, trans,
2650                                         dp->i_mount->m_ddev_targp,
2651                                         mapp, nmap, 0, &bp, ops);
2652         if (error)
2653                 goto out_free;
2654
2655         if (whichfork == XFS_ATTR_FORK)
2656                 xfs_buf_set_ref(bp, XFS_ATTR_BTREE_REF);
2657         else
2658                 xfs_buf_set_ref(bp, XFS_DIR_BTREE_REF);
2659
2660         /*
2661          * This verification code will be moved to a CRC verification callback
2662          * function so just leave it here unchanged until then.
2663          */
2664         {
2665                 xfs_dir2_data_hdr_t     *hdr = bp->b_addr;
2666                 xfs_dir2_free_t         *free = bp->b_addr;
2667                 xfs_da_blkinfo_t        *info = bp->b_addr;
2668                 uint                    magic, magic1;
2669                 struct xfs_mount        *mp = dp->i_mount;
2670
2671                 magic = be16_to_cpu(info->magic);
2672                 magic1 = be32_to_cpu(hdr->magic);
2673                 if (unlikely(
2674                     XFS_TEST_ERROR((magic != XFS_DA_NODE_MAGIC) &&
2675                                    (magic != XFS_DA3_NODE_MAGIC) &&
2676                                    (magic != XFS_ATTR_LEAF_MAGIC) &&
2677                                    (magic != XFS_ATTR3_LEAF_MAGIC) &&
2678                                    (magic != XFS_DIR2_LEAF1_MAGIC) &&
2679                                    (magic != XFS_DIR3_LEAF1_MAGIC) &&
2680                                    (magic != XFS_DIR2_LEAFN_MAGIC) &&
2681                                    (magic != XFS_DIR3_LEAFN_MAGIC) &&
2682                                    (magic1 != XFS_DIR2_BLOCK_MAGIC) &&
2683                                    (magic1 != XFS_DIR3_BLOCK_MAGIC) &&
2684                                    (magic1 != XFS_DIR2_DATA_MAGIC) &&
2685                                    (magic1 != XFS_DIR3_DATA_MAGIC) &&
2686                                    (free->hdr.magic !=
2687                                         cpu_to_be32(XFS_DIR2_FREE_MAGIC)) &&
2688                                    (free->hdr.magic !=
2689                                         cpu_to_be32(XFS_DIR3_FREE_MAGIC)),
2690                                 mp, XFS_ERRTAG_DA_READ_BUF,
2691                                 XFS_RANDOM_DA_READ_BUF))) {
2692                         trace_xfs_da_btree_corrupt(bp, _RET_IP_);
2693                         XFS_CORRUPTION_ERROR("xfs_da_do_buf(2)",
2694                                              XFS_ERRLEVEL_LOW, mp, info);
2695                         error = XFS_ERROR(EFSCORRUPTED);
2696                         xfs_trans_brelse(trans, bp);
2697                         goto out_free;
2698                 }
2699         }
2700         *bpp = bp;
2701 out_free:
2702         if (mapp != &map)
2703                 kmem_free(mapp);
2704
2705         return error;
2706 }
2707
2708 /*
2709  * Readahead the dir/attr block.
2710  */
2711 xfs_daddr_t
2712 xfs_da_reada_buf(
2713         struct xfs_trans        *trans,
2714         struct xfs_inode        *dp,
2715         xfs_dablk_t             bno,
2716         xfs_daddr_t             mappedbno,
2717         int                     whichfork,
2718         const struct xfs_buf_ops *ops)
2719 {
2720         struct xfs_buf_map      map;
2721         struct xfs_buf_map      *mapp;
2722         int                     nmap;
2723         int                     error;
2724
2725         mapp = &map;
2726         nmap = 1;
2727         error = xfs_dabuf_map(trans, dp, bno, mappedbno, whichfork,
2728                                 &mapp, &nmap);
2729         if (error) {
2730                 /* mapping a hole is not an error, but we don't continue */
2731                 if (error == -1)
2732                         error = 0;
2733                 goto out_free;
2734         }
2735
2736         mappedbno = mapp[0].bm_bn;
2737         xfs_buf_readahead_map(dp->i_mount->m_ddev_targp, mapp, nmap, ops);
2738
2739 out_free:
2740         if (mapp != &map)
2741                 kmem_free(mapp);
2742
2743         if (error)
2744                 return -1;
2745         return mappedbno;
2746 }