fs/xfs/libxfs/xfs_da_btree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   4  * Copyright (c) 2013 Red Hat, Inc.
   5  * All Rights Reserved.
   6  */
   7 #include "xfs.h"
   8 #include "xfs_fs.h"
   9 #include "xfs_shared.h"
  10 #include "xfs_format.h"
  11 #include "xfs_log_format.h"
  12 #include "xfs_trans_resv.h"
  13 #include "xfs_bit.h"
  14 #include "xfs_mount.h"
  15 #include "xfs_dir2.h"
  16 #include "xfs_dir2_priv.h"
  17 #include "xfs_inode.h"
  18 #include "xfs_trans.h"
  19 #include "xfs_bmap.h"
  20 #include "xfs_attr_leaf.h"
  21 #include "xfs_error.h"
  22 #include "xfs_trace.h"
  23 #include "xfs_buf_item.h"
  24 #include "xfs_log.h"
  25
  26 /*
  27  * xfs_da_btree.c
  28  *
  29  * Routines to implement directories as Btrees of hashed names.
  30  */
  31
  32 /*========================================================================
  33  * Function prototypes for the kernel.
  34  *========================================================================*/
  35
  36 /*
  37  * Routines used for growing the Btree.
  38  */
  39 STATIC int xfs_da3_root_split(xfs_da_state_t *state,
  40                                             xfs_da_state_blk_t *existing_root,
  41                                             xfs_da_state_blk_t *new_child);
  42 STATIC int xfs_da3_node_split(xfs_da_state_t *state,
  43                                             xfs_da_state_blk_t *existing_blk,
  44                                             xfs_da_state_blk_t *split_blk,
  45                                             xfs_da_state_blk_t *blk_to_add,
  46                                             int treelevel,
  47                                             int *result);
  48 STATIC void xfs_da3_node_rebalance(xfs_da_state_t *state,
  49                                          xfs_da_state_blk_t *node_blk_1,
  50                                          xfs_da_state_blk_t *node_blk_2);
  51 STATIC void xfs_da3_node_add(xfs_da_state_t *state,
  52                                    xfs_da_state_blk_t *old_node_blk,
  53                                    xfs_da_state_blk_t *new_node_blk);
  54
  55 /*
  56  * Routines used for shrinking the Btree.
  57  */
  58 STATIC int xfs_da3_root_join(xfs_da_state_t *state,
  59                                            xfs_da_state_blk_t *root_blk);
  60 STATIC int xfs_da3_node_toosmall(xfs_da_state_t *state, int *retval);
  61 STATIC void xfs_da3_node_remove(xfs_da_state_t *state,
  62                                               xfs_da_state_blk_t *drop_blk);
  63 STATIC void xfs_da3_node_unbalance(xfs_da_state_t *state,
  64                                          xfs_da_state_blk_t *src_node_blk,
  65                                          xfs_da_state_blk_t *dst_node_blk);
  66
  67 /*
  68  * Utility routines.
  69  */
  70 STATIC int      xfs_da3_blk_unlink(xfs_da_state_t *state,
  71                                   xfs_da_state_blk_t *drop_blk,
  72                                   xfs_da_state_blk_t *save_blk);
  73
  74
  75 kmem_zone_t *xfs_da_state_zone; /* anchor for state struct zone */
  76
  77 /*
  78  * Allocate a dir-state structure.
  79  * We don't put them on the stack since they're large.
  80  */
  81 xfs_da_state_t *
  82 xfs_da_state_alloc(void)
  83 {
  84         return kmem_zone_zalloc(xfs_da_state_zone, KM_NOFS);
  85 }
  86
  87 /*
  88  * Kill the altpath contents of a da-state structure.
  89  */
  90 STATIC void
  91 xfs_da_state_kill_altpath(xfs_da_state_t *state)
  92 {
  93         int     i;
  94
  95         for (i = 0; i < state->altpath.active; i++)
  96                 state->altpath.blk[i].bp = NULL;
  97         state->altpath.active = 0;
  98 }
  99
 100 /*
 101  * Free a da-state structure.
 102  */
 103 void
 104 xfs_da_state_free(xfs_da_state_t *state)
 105 {
 106         xfs_da_state_kill_altpath(state);
 107 #ifdef DEBUG
 108         memset((char *)state, 0, sizeof(*state));
 109 #endif /* DEBUG */
 110         kmem_zone_free(xfs_da_state_zone, state);
 111 }
 112
 113 void
 114 xfs_da3_node_hdr_from_disk(
 115         struct xfs_mount                *mp,
 116         struct xfs_da3_icnode_hdr       *to,
 117         struct xfs_da_intnode           *from)
 118 {
 119         if (xfs_sb_version_hascrc(&mp->m_sb)) {
 120                 struct xfs_da3_intnode  *from3 = (struct xfs_da3_intnode *)from;
 121
 122                 to->forw = be32_to_cpu(from3->hdr.info.hdr.forw);
 123                 to->back = be32_to_cpu(from3->hdr.info.hdr.back);
 124                 to->magic = be16_to_cpu(from3->hdr.info.hdr.magic);
 125                 to->count = be16_to_cpu(from3->hdr.__count);
 126                 to->level = be16_to_cpu(from3->hdr.__level);
 127                 to->btree = from3->__btree;
 128                 ASSERT(to->magic == XFS_DA3_NODE_MAGIC);
 129         } else {
 130                 to->forw = be32_to_cpu(from->hdr.info.forw);
 131                 to->back = be32_to_cpu(from->hdr.info.back);
 132                 to->magic = be16_to_cpu(from->hdr.info.magic);
 133                 to->count = be16_to_cpu(from->hdr.__count);
 134                 to->level = be16_to_cpu(from->hdr.__level);
 135                 to->btree = from->__btree;
 136                 ASSERT(to->magic == XFS_DA_NODE_MAGIC);
 137         }
 138 }
 139
 140 void
 141 xfs_da3_node_hdr_to_disk(
 142         struct xfs_mount                *mp,
 143         struct xfs_da_intnode           *to,
 144         struct xfs_da3_icnode_hdr       *from)
 145 {
 146         if (xfs_sb_version_hascrc(&mp->m_sb)) {
 147                 struct xfs_da3_intnode  *to3 = (struct xfs_da3_intnode *)to;
 148
 149                 ASSERT(from->magic == XFS_DA3_NODE_MAGIC);
 150                 to3->hdr.info.hdr.forw = cpu_to_be32(from->forw);
 151                 to3->hdr.info.hdr.back = cpu_to_be32(from->back);
 152                 to3->hdr.info.hdr.magic = cpu_to_be16(from->magic);
 153                 to3->hdr.__count = cpu_to_be16(from->count);
 154                 to3->hdr.__level = cpu_to_be16(from->level);
 155         } else {
 156                 ASSERT(from->magic == XFS_DA_NODE_MAGIC);
 157                 to->hdr.info.forw = cpu_to_be32(from->forw);
 158                 to->hdr.info.back = cpu_to_be32(from->back);
 159                 to->hdr.info.magic = cpu_to_be16(from->magic);
 160                 to->hdr.__count = cpu_to_be16(from->count);
 161                 to->hdr.__level = cpu_to_be16(from->level);
 162         }
 163 }
 164
 165 /*
 166  * Verify an xfs_da3_blkinfo structure. Note that the da3 fields are only
 167  * accessible on v5 filesystems. This header format is common across da node,
 168  * attr leaf and dir leaf blocks.
 169  */
 170 xfs_failaddr_t
 171 xfs_da3_blkinfo_verify(
 172         struct xfs_buf          *bp,
 173         struct xfs_da3_blkinfo  *hdr3)
 174 {
 175         struct xfs_mount        *mp = bp->b_mount;
 176         struct xfs_da_blkinfo   *hdr = &hdr3->hdr;
 177
 178         if (!xfs_verify_magic16(bp, hdr->magic))
 179                 return __this_address;
 180
 181         if (xfs_sb_version_hascrc(&mp->m_sb)) {
 182                 if (!uuid_equal(&hdr3->uuid, &mp->m_sb.sb_meta_uuid))
 183                         return __this_address;
 184                 if (be64_to_cpu(hdr3->blkno) != bp->b_bn)
 185                         return __this_address;
 186                 if (!xfs_log_check_lsn(mp, be64_to_cpu(hdr3->lsn)))
 187                         return __this_address;
 188         }
 189
 190         return NULL;
 191 }
 192
 193 static xfs_failaddr_t
 194 xfs_da3_node_verify(
 195         struct xfs_buf          *bp)
 196 {
 197         struct xfs_mount        *mp = bp->b_mount;
 198         struct xfs_da_intnode   *hdr = bp->b_addr;
 199         struct xfs_da3_icnode_hdr ichdr;
 200         xfs_failaddr_t          fa;
 201
 202         xfs_da3_node_hdr_from_disk(mp, &ichdr, hdr);
 203
 204         fa = xfs_da3_blkinfo_verify(bp, bp->b_addr);
 205         if (fa)
 206                 return fa;
 207
 208         if (ichdr.level == 0)
 209                 return __this_address;
 210         if (ichdr.level > XFS_DA_NODE_MAXDEPTH)
 211                 return __this_address;
 212         if (ichdr.count == 0)
 213                 return __this_address;
 214
 215         /*
 216          * we don't know if the node is for and attribute or directory tree,
 217          * so only fail if the count is outside both bounds
 218          */
 219         if (ichdr.count > mp->m_dir_geo->node_ents &&
 220             ichdr.count > mp->m_attr_geo->node_ents)
 221                 return __this_address;
 222
 223         /* XXX: hash order check? */
 224
 225         return NULL;
 226 }
 227
 228 static void
 229 xfs_da3_node_write_verify(
 230         struct xfs_buf  *bp)
 231 {
 232         struct xfs_mount        *mp = bp->b_mount;
 233         struct xfs_buf_log_item *bip = bp->b_log_item;
 234         struct xfs_da3_node_hdr *hdr3 = bp->b_addr;
 235         xfs_failaddr_t          fa;
 236
 237         fa = xfs_da3_node_verify(bp);
 238         if (fa) {
 239                 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
 240                 return;
 241         }
 242
 243         if (!xfs_sb_version_hascrc(&mp->m_sb))
 244                 return;
 245
 246         if (bip)
 247                 hdr3->info.lsn = cpu_to_be64(bip->bli_item.li_lsn);
 248
 249         xfs_buf_update_cksum(bp, XFS_DA3_NODE_CRC_OFF);
 250 }
 251
 252 /*
 253  * leaf/node format detection on trees is sketchy, so a node read can be done on
 254  * leaf level blocks when detection identifies the tree as a node format tree
 255  * incorrectly. In this case, we need to swap the verifier to match the correct
 256  * format of the block being read.
 257  */
 258 static void
 259 xfs_da3_node_read_verify(
 260         struct xfs_buf          *bp)
 261 {
 262         struct xfs_da_blkinfo   *info = bp->b_addr;
 263         xfs_failaddr_t          fa;
 264
 265         switch (be16_to_cpu(info->magic)) {
 266                 case XFS_DA3_NODE_MAGIC:
 267                         if (!xfs_buf_verify_cksum(bp, XFS_DA3_NODE_CRC_OFF)) {
 268                                 xfs_verifier_error(bp, -EFSBADCRC,
 269                                                 __this_address);
 270                                 break;
 271                         }
 272                         /* fall through */
 273                 case XFS_DA_NODE_MAGIC:
 274                         fa = xfs_da3_node_verify(bp);
 275                         if (fa)
 276                                 xfs_verifier_error(bp, -EFSCORRUPTED, fa);
 277                         return;
 278                 case XFS_ATTR_LEAF_MAGIC:
 279                 case XFS_ATTR3_LEAF_MAGIC:
 280                         bp->b_ops = &xfs_attr3_leaf_buf_ops;
 281                         bp->b_ops->verify_read(bp);
 282                         return;
 283                 case XFS_DIR2_LEAFN_MAGIC:
 284                 case XFS_DIR3_LEAFN_MAGIC:
 285                         bp->b_ops = &xfs_dir3_leafn_buf_ops;
 286                         bp->b_ops->verify_read(bp);
 287                         return;
 288                 default:
 289                         xfs_verifier_error(bp, -EFSCORRUPTED, __this_address);
 290                         break;
 291         }
 292 }
 293
 294 /* Verify the structure of a da3 block. */
 295 static xfs_failaddr_t
 296 xfs_da3_node_verify_struct(
 297         struct xfs_buf          *bp)
 298 {
 299         struct xfs_da_blkinfo   *info = bp->b_addr;
 300
 301         switch (be16_to_cpu(info->magic)) {
 302         case XFS_DA3_NODE_MAGIC:
 303         case XFS_DA_NODE_MAGIC:
 304                 return xfs_da3_node_verify(bp);
 305         case XFS_ATTR_LEAF_MAGIC:
 306         case XFS_ATTR3_LEAF_MAGIC:
 307                 bp->b_ops = &xfs_attr3_leaf_buf_ops;
 308                 return bp->b_ops->verify_struct(bp);
 309         case XFS_DIR2_LEAFN_MAGIC:
 310         case XFS_DIR3_LEAFN_MAGIC:
 311                 bp->b_ops = &xfs_dir3_leafn_buf_ops;
 312                 return bp->b_ops->verify_struct(bp);
 313         default:
 314                 return __this_address;
 315         }
 316 }
 317
 318 const struct xfs_buf_ops xfs_da3_node_buf_ops = {
 319         .name = "xfs_da3_node",
 320         .magic16 = { cpu_to_be16(XFS_DA_NODE_MAGIC),
 321                      cpu_to_be16(XFS_DA3_NODE_MAGIC) },
 322         .verify_read = xfs_da3_node_read_verify,
 323         .verify_write = xfs_da3_node_write_verify,
 324         .verify_struct = xfs_da3_node_verify_struct,
 325 };
 326
 327 int
 328 xfs_da3_node_read(
 329         struct xfs_trans        *tp,
 330         struct xfs_inode        *dp,
 331         xfs_dablk_t             bno,
 332         xfs_daddr_t             mappedbno,
 333         struct xfs_buf          **bpp,
 334         int                     which_fork)
 335 {
 336         int                     err;
 337
 338         err = xfs_da_read_buf(tp, dp, bno, mappedbno, bpp,
 339                                         which_fork, &xfs_da3_node_buf_ops);
 340         if (!err && tp && *bpp) {
 341                 struct xfs_da_blkinfo   *info = (*bpp)->b_addr;
 342                 int                     type;
 343
 344                 switch (be16_to_cpu(info->magic)) {
 345                 case XFS_DA_NODE_MAGIC:
 346                 case XFS_DA3_NODE_MAGIC:
 347                         type = XFS_BLFT_DA_NODE_BUF;
 348                         break;
 349                 case XFS_ATTR_LEAF_MAGIC:
 350                 case XFS_ATTR3_LEAF_MAGIC:
 351                         type = XFS_BLFT_ATTR_LEAF_BUF;
 352                         break;
 353                 case XFS_DIR2_LEAFN_MAGIC:
 354                 case XFS_DIR3_LEAFN_MAGIC:
 355                         type = XFS_BLFT_DIR_LEAFN_BUF;
 356                         break;
 357                 default:
 358                         XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW,
 359                                         tp->t_mountp, info, sizeof(*info));
 360                         xfs_trans_brelse(tp, *bpp);
 361                         *bpp = NULL;
 362                         return -EFSCORRUPTED;
 363                 }
 364                 xfs_trans_buf_set_type(tp, *bpp, type);
 365         }
 366         return err;
 367 }
 368
 369 /*========================================================================
 370  * Routines used for growing the Btree.
 371  *========================================================================*/
 372
 373 /*
 374  * Create the initial contents of an intermediate node.
 375  */
 376 int
 377 xfs_da3_node_create(
 378         struct xfs_da_args      *args,
 379         xfs_dablk_t             blkno,
 380         int                     level,
 381         struct xfs_buf          **bpp,
 382         int                     whichfork)
 383 {
 384         struct xfs_da_intnode   *node;
 385         struct xfs_trans        *tp = args->trans;
 386         struct xfs_mount        *mp = tp->t_mountp;
 387         struct xfs_da3_icnode_hdr ichdr = {0};
 388         struct xfs_buf          *bp;
 389         int                     error;
 390         struct xfs_inode        *dp = args->dp;
 391
 392         trace_xfs_da_node_create(args);
 393         ASSERT(level <= XFS_DA_NODE_MAXDEPTH);
 394
 395         error = xfs_da_get_buf(tp, dp, blkno, -1, &bp, whichfork);
 396         if (error)
 397                 return error;
 398         bp->b_ops = &xfs_da3_node_buf_ops;
 399         xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DA_NODE_BUF);
 400         node = bp->b_addr;
 401
 402         if (xfs_sb_version_hascrc(&mp->m_sb)) {
 403                 struct xfs_da3_node_hdr *hdr3 = bp->b_addr;
 404
 405                 memset(hdr3, 0, sizeof(struct xfs_da3_node_hdr));
 406                 ichdr.magic = XFS_DA3_NODE_MAGIC;
 407                 hdr3->info.blkno = cpu_to_be64(bp->b_bn);
 408                 hdr3->info.owner = cpu_to_be64(args->dp->i_ino);
 409                 uuid_copy(&hdr3->info.uuid, &mp->m_sb.sb_meta_uuid);
 410         } else {
 411                 ichdr.magic = XFS_DA_NODE_MAGIC;
 412         }
 413         ichdr.level = level;
 414
 415         xfs_da3_node_hdr_to_disk(dp->i_mount, node, &ichdr);
 416         xfs_trans_log_buf(tp, bp,
 417                 XFS_DA_LOGRANGE(node, &node->hdr, args->geo->node_hdr_size));
 418
 419         *bpp = bp;
 420         return 0;
 421 }
 422
 423 /*
 424  * Split a leaf node, rebalance, then possibly split
 425  * intermediate nodes, rebalance, etc.
 426  */
 427 int                                                     /* error */
 428 xfs_da3_split(
 429         struct xfs_da_state     *state)
 430 {
 431         struct xfs_da_state_blk *oldblk;
 432         struct xfs_da_state_blk *newblk;
 433         struct xfs_da_state_blk *addblk;
 434         struct xfs_da_intnode   *node;
 435         int                     max;
 436         int                     action = 0;
 437         int                     error;
 438         int                     i;
 439
 440         trace_xfs_da_split(state->args);
 441
 442         /*
 443          * Walk back up the tree splitting/inserting/adjusting as necessary.
 444          * If we need to insert and there isn't room, split the node, then
 445          * decide which fragment to insert the new block from below into.
 446          * Note that we may split the root this way, but we need more fixup.
 447          */
 448         max = state->path.active - 1;
 449         ASSERT((max >= 0) && (max < XFS_DA_NODE_MAXDEPTH));
 450         ASSERT(state->path.blk[max].magic == XFS_ATTR_LEAF_MAGIC ||
 451                state->path.blk[max].magic == XFS_DIR2_LEAFN_MAGIC);
 452
 453         addblk = &state->path.blk[max];         /* initial dummy value */
 454         for (i = max; (i >= 0) && addblk; state->path.active--, i--) {
 455                 oldblk = &state->path.blk[i];
 456                 newblk = &state->altpath.blk[i];
 457
 458                 /*
 459                  * If a leaf node then
 460                  *     Allocate a new leaf node, then rebalance across them.
 461                  * else if an intermediate node then
 462                  *     We split on the last layer, must we split the node?
 463                  */
 464                 switch (oldblk->magic) {
 465                 case XFS_ATTR_LEAF_MAGIC:
 466                         error = xfs_attr3_leaf_split(state, oldblk, newblk);
 467                         if ((error != 0) && (error != -ENOSPC)) {
 468                                 return error;   /* GROT: attr is inconsistent */
 469                         }
 470                         if (!error) {
 471                                 addblk = newblk;
 472                                 break;
 473                         }
 474                         /*
 475                          * Entry wouldn't fit, split the leaf again. The new
 476                          * extrablk will be consumed by xfs_da3_node_split if
 477                          * the node is split.
 478                          */
 479                         state->extravalid = 1;
 480                         if (state->inleaf) {
 481                                 state->extraafter = 0;  /* before newblk */
 482                                 trace_xfs_attr_leaf_split_before(state->args);
 483                                 error = xfs_attr3_leaf_split(state, oldblk,
 484                                                             &state->extrablk);
 485                         } else {
 486                                 state->extraafter = 1;  /* after newblk */
 487                                 trace_xfs_attr_leaf_split_after(state->args);
 488                                 error = xfs_attr3_leaf_split(state, newblk,
 489                                                             &state->extrablk);
 490                         }
 491                         if (error)
 492                                 return error;   /* GROT: attr inconsistent */
 493                         addblk = newblk;
 494                         break;
 495                 case XFS_DIR2_LEAFN_MAGIC:
 496                         error = xfs_dir2_leafn_split(state, oldblk, newblk);
 497                         if (error)
 498                                 return error;
 499                         addblk = newblk;
 500                         break;
 501                 case XFS_DA_NODE_MAGIC:
 502                         error = xfs_da3_node_split(state, oldblk, newblk, addblk,
 503                                                          max - i, &action);
 504                         addblk->bp = NULL;
 505                         if (error)
 506                                 return error;   /* GROT: dir is inconsistent */
 507                         /*
 508                          * Record the newly split block for the next time thru?
 509                          */
 510                         if (action)
 511                                 addblk = newblk;
 512                         else
 513                                 addblk = NULL;
 514                         break;
 515                 }
 516
 517                 /*
 518                  * Update the btree to show the new hashval for this child.
 519                  */
 520                 xfs_da3_fixhashpath(state, &state->path);
 521         }
 522         if (!addblk)
 523                 return 0;
 524
 525         /*
 526          * xfs_da3_node_split() should have consumed any extra blocks we added
 527          * during a double leaf split in the attr fork. This is guaranteed as
 528          * we can't be here if the attr fork only has a single leaf block.
 529          */
 530         ASSERT(state->extravalid == 0 ||
 531                state->path.blk[max].magic == XFS_DIR2_LEAFN_MAGIC);
 532
 533         /*
 534          * Split the root node.
 535          */
 536         ASSERT(state->path.active == 0);
 537         oldblk = &state->path.blk[0];
 538         error = xfs_da3_root_split(state, oldblk, addblk);
 539         if (error)
 540                 goto out;
 541
 542         /*
 543          * Update pointers to the node which used to be block 0 and just got
 544          * bumped because of the addition of a new root node.  Note that the
 545          * original block 0 could be at any position in the list of blocks in
 546          * the tree.
 547          *
 548          * Note: the magic numbers and sibling pointers are in the same physical
 549          * place for both v2 and v3 headers (by design). Hence it doesn't matter
 550          * which version of the xfs_da_intnode structure we use here as the
 551          * result will be the same using either structure.
 552          */
 553         node = oldblk->bp->b_addr;
 554         if (node->hdr.info.forw) {
 555                 if (be32_to_cpu(node->hdr.info.forw) != addblk->blkno) {
 556                         xfs_buf_corruption_error(oldblk->bp);
 557                         error = -EFSCORRUPTED;
 558                         goto out;
 559                 }
 560                 node = addblk->bp->b_addr;
 561                 node->hdr.info.back = cpu_to_be32(oldblk->blkno);
 562                 xfs_trans_log_buf(state->args->trans, addblk->bp,
 563                                   XFS_DA_LOGRANGE(node, &node->hdr.info,
 564                                   sizeof(node->hdr.info)));
 565         }
 566         node = oldblk->bp->b_addr;
 567         if (node->hdr.info.back) {
 568                 if (be32_to_cpu(node->hdr.info.back) != addblk->blkno) {
 569                         xfs_buf_corruption_error(oldblk->bp);
 570                         error = -EFSCORRUPTED;
 571                         goto out;
 572                 }
 573                 node = addblk->bp->b_addr;
 574                 node->hdr.info.forw = cpu_to_be32(oldblk->blkno);
 575                 xfs_trans_log_buf(state->args->trans, addblk->bp,
 576                                   XFS_DA_LOGRANGE(node, &node->hdr.info,
 577                                   sizeof(node->hdr.info)));
 578         }
 579 out:
 580         addblk->bp = NULL;
 581         return error;
 582 }
 583
 584 /*
 585  * Split the root.  We have to create a new root and point to the two
 586  * parts (the split old root) that we just created.  Copy block zero to
 587  * the EOF, extending the inode in process.
 588  */
 589 STATIC int                                              /* error */
 590 xfs_da3_root_split(
 591         struct xfs_da_state     *state,
 592         struct xfs_da_state_blk *blk1,
 593         struct xfs_da_state_blk *blk2)
 594 {
 595         struct xfs_da_intnode   *node;
 596         struct xfs_da_intnode   *oldroot;
 597         struct xfs_da_node_entry *btree;
 598         struct xfs_da3_icnode_hdr nodehdr;
 599         struct xfs_da_args      *args;
 600         struct xfs_buf          *bp;
 601         struct xfs_inode        *dp;
 602         struct xfs_trans        *tp;
 603         struct xfs_dir2_leaf    *leaf;
 604         xfs_dablk_t             blkno;
 605         int                     level;
 606         int                     error;
 607         int                     size;
 608
 609         trace_xfs_da_root_split(state->args);
 610
 611         /*
 612          * Copy the existing (incorrect) block from the root node position
 613          * to a free space somewhere.
 614          */
 615         args = state->args;
 616         error = xfs_da_grow_inode(args, &blkno);
 617         if (error)
 618                 return error;
 619
 620         dp = args->dp;
 621         tp = args->trans;
 622         error = xfs_da_get_buf(tp, dp, blkno, -1, &bp, args->whichfork);
 623         if (error)
 624                 return error;
 625         node = bp->b_addr;
 626         oldroot = blk1->bp->b_addr;
 627         if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
 628             oldroot->hdr.info.magic == cpu_to_be16(XFS_DA3_NODE_MAGIC)) {
 629                 struct xfs_da3_icnode_hdr icnodehdr;
 630
 631                 xfs_da3_node_hdr_from_disk(dp->i_mount, &icnodehdr, oldroot);
 632                 btree = icnodehdr.btree;
 633                 size = (int)((char *)&btree[icnodehdr.count] - (char *)oldroot);
 634                 level = icnodehdr.level;
 635
 636                 /*
 637                  * we are about to copy oldroot to bp, so set up the type
 638                  * of bp while we know exactly what it will be.
 639                  */
 640                 xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DA_NODE_BUF);
 641         } else {
 642                 struct xfs_dir3_icleaf_hdr leafhdr;
 643
 644                 leaf = (xfs_dir2_leaf_t *)oldroot;
 645                 xfs_dir2_leaf_hdr_from_disk(dp->i_mount, &leafhdr, leaf);
 646
 647                 ASSERT(leafhdr.magic == XFS_DIR2_LEAFN_MAGIC ||
 648                        leafhdr.magic == XFS_DIR3_LEAFN_MAGIC);
 649                 size = (int)((char *)&leafhdr.ents[leafhdr.count] -
 650                         (char *)leaf);
 651                 level = 0;
 652
 653                 /*
 654                  * we are about to copy oldroot to bp, so set up the type
 655                  * of bp while we know exactly what it will be.
 656                  */
 657                 xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DIR_LEAFN_BUF);
 658         }
 659
 660         /*
 661          * we can copy most of the information in the node from one block to
 662          * another, but for CRC enabled headers we have to make sure that the
 663          * block specific identifiers are kept intact. We update the buffer
 664          * directly for this.
 665          */
 666         memcpy(node, oldroot, size);
 667         if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DA3_NODE_MAGIC) ||
 668             oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
 669                 struct xfs_da3_intnode *node3 = (struct xfs_da3_intnode *)node;
 670
 671                 node3->hdr.info.blkno = cpu_to_be64(bp->b_bn);
 672         }
 673         xfs_trans_log_buf(tp, bp, 0, size - 1);
 674
 675         bp->b_ops = blk1->bp->b_ops;
 676         xfs_trans_buf_copy_type(bp, blk1->bp);
 677         blk1->bp = bp;
 678         blk1->blkno = blkno;
 679
 680         /*
 681          * Set up the new root node.
 682          */
 683         error = xfs_da3_node_create(args,
 684                 (args->whichfork == XFS_DATA_FORK) ? args->geo->leafblk : 0,
 685                 level + 1, &bp, args->whichfork);
 686         if (error)
 687                 return error;
 688
 689         node = bp->b_addr;
 690         xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
 691         btree = nodehdr.btree;
 692         btree[0].hashval = cpu_to_be32(blk1->hashval);
 693         btree[0].before = cpu_to_be32(blk1->blkno);
 694         btree[1].hashval = cpu_to_be32(blk2->hashval);
 695         btree[1].before = cpu_to_be32(blk2->blkno);
 696         nodehdr.count = 2;
 697         xfs_da3_node_hdr_to_disk(dp->i_mount, node, &nodehdr);
 698
 699 #ifdef DEBUG
 700         if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
 701             oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
 702                 ASSERT(blk1->blkno >= args->geo->leafblk &&
 703                        blk1->blkno < args->geo->freeblk);
 704                 ASSERT(blk2->blkno >= args->geo->leafblk &&
 705                        blk2->blkno < args->geo->freeblk);
 706         }
 707 #endif
 708
 709         /* Header is already logged by xfs_da_node_create */
 710         xfs_trans_log_buf(tp, bp,
 711                 XFS_DA_LOGRANGE(node, btree, sizeof(xfs_da_node_entry_t) * 2));
 712
 713         return 0;
 714 }
 715
 716 /*
 717  * Split the node, rebalance, then add the new entry.
 718  */
 719 STATIC int                                              /* error */
 720 xfs_da3_node_split(
 721         struct xfs_da_state     *state,
 722         struct xfs_da_state_blk *oldblk,
 723         struct xfs_da_state_blk *newblk,
 724         struct xfs_da_state_blk *addblk,
 725         int                     treelevel,
 726         int                     *result)
 727 {
 728         struct xfs_da_intnode   *node;
 729         struct xfs_da3_icnode_hdr nodehdr;
 730         xfs_dablk_t             blkno;
 731         int                     newcount;
 732         int                     error;
 733         int                     useextra;
 734         struct xfs_inode        *dp = state->args->dp;
 735
 736         trace_xfs_da_node_split(state->args);
 737
 738         node = oldblk->bp->b_addr;
 739         xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
 740
 741         /*
 742          * With V2 dirs the extra block is data or freespace.
 743          */
 744         useextra = state->extravalid && state->args->whichfork == XFS_ATTR_FORK;
 745         newcount = 1 + useextra;
 746         /*
 747          * Do we have to split the node?
 748          */
 749         if (nodehdr.count + newcount > state->args->geo->node_ents) {
 750                 /*
 751                  * Allocate a new node, add to the doubly linked chain of
 752                  * nodes, then move some of our excess entries into it.
 753                  */
 754                 error = xfs_da_grow_inode(state->args, &blkno);
 755                 if (error)
 756                         return error;   /* GROT: dir is inconsistent */
 757
 758                 error = xfs_da3_node_create(state->args, blkno, treelevel,
 759                                            &newblk->bp, state->args->whichfork);
 760                 if (error)
 761                         return error;   /* GROT: dir is inconsistent */
 762                 newblk->blkno = blkno;
 763                 newblk->magic = XFS_DA_NODE_MAGIC;
 764                 xfs_da3_node_rebalance(state, oldblk, newblk);
 765                 error = xfs_da3_blk_link(state, oldblk, newblk);
 766                 if (error)
 767                         return error;
 768                 *result = 1;
 769         } else {
 770                 *result = 0;
 771         }
 772
 773         /*
 774          * Insert the new entry(s) into the correct block
 775          * (updating last hashval in the process).
 776          *
 777          * xfs_da3_node_add() inserts BEFORE the given index,
 778          * and as a result of using node_lookup_int() we always
 779          * point to a valid entry (not after one), but a split
 780          * operation always results in a new block whose hashvals
 781          * FOLLOW the current block.
 782          *
 783          * If we had double-split op below us, then add the extra block too.
 784          */
 785         node = oldblk->bp->b_addr;
 786         xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
 787         if (oldblk->index <= nodehdr.count) {
 788                 oldblk->index++;
 789                 xfs_da3_node_add(state, oldblk, addblk);
 790                 if (useextra) {
 791                         if (state->extraafter)
 792                                 oldblk->index++;
 793                         xfs_da3_node_add(state, oldblk, &state->extrablk);
 794                         state->extravalid = 0;
 795                 }
 796         } else {
 797                 newblk->index++;
 798                 xfs_da3_node_add(state, newblk, addblk);
 799                 if (useextra) {
 800                         if (state->extraafter)
 801                                 newblk->index++;
 802                         xfs_da3_node_add(state, newblk, &state->extrablk);
 803                         state->extravalid = 0;
 804                 }
 805         }
 806
 807         return 0;
 808 }
 809
 810 /*
 811  * Balance the btree elements between two intermediate nodes,
 812  * usually one full and one empty.
 813  *
 814  * NOTE: if blk2 is empty, then it will get the upper half of blk1.
 815  */
 816 STATIC void
 817 xfs_da3_node_rebalance(
 818         struct xfs_da_state     *state,
 819         struct xfs_da_state_blk *blk1,
 820         struct xfs_da_state_blk *blk2)
 821 {
 822         struct xfs_da_intnode   *node1;
 823         struct xfs_da_intnode   *node2;
 824         struct xfs_da_intnode   *tmpnode;
 825         struct xfs_da_node_entry *btree1;
 826         struct xfs_da_node_entry *btree2;
 827         struct xfs_da_node_entry *btree_s;
 828         struct xfs_da_node_entry *btree_d;
 829         struct xfs_da3_icnode_hdr nodehdr1;
 830         struct xfs_da3_icnode_hdr nodehdr2;
 831         struct xfs_trans        *tp;
 832         int                     count;
 833         int                     tmp;
 834         int                     swap = 0;
 835         struct xfs_inode        *dp = state->args->dp;
 836
 837         trace_xfs_da_node_rebalance(state->args);
 838
 839         node1 = blk1->bp->b_addr;
 840         node2 = blk2->bp->b_addr;
 841         xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr1, node1);
 842         xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr2, node2);
 843         btree1 = nodehdr1.btree;
 844         btree2 = nodehdr2.btree;
 845
 846         /*
 847          * Figure out how many entries need to move, and in which direction.
 848          * Swap the nodes around if that makes it simpler.
 849          */
 850         if (nodehdr1.count > 0 && nodehdr2.count > 0 &&
 851             ((be32_to_cpu(btree2[0].hashval) < be32_to_cpu(btree1[0].hashval)) ||
 852              (be32_to_cpu(btree2[nodehdr2.count - 1].hashval) <
 853                         be32_to_cpu(btree1[nodehdr1.count - 1].hashval)))) {
 854                 tmpnode = node1;
 855                 node1 = node2;
 856                 node2 = tmpnode;
 857                 xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr1, node1);
 858                 xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr2, node2);
 859                 btree1 = nodehdr1.btree;
 860                 btree2 = nodehdr2.btree;
 861                 swap = 1;
 862         }
 863
 864         count = (nodehdr1.count - nodehdr2.count) / 2;
 865         if (count == 0)
 866                 return;
 867         tp = state->args->trans;
 868         /*
 869          * Two cases: high-to-low and low-to-high.
 870          */
 871         if (count > 0) {
 872                 /*
 873                  * Move elements in node2 up to make a hole.
 874                  */
 875                 tmp = nodehdr2.count;
 876                 if (tmp > 0) {
 877                         tmp *= (uint)sizeof(xfs_da_node_entry_t);
 878                         btree_s = &btree2[0];
 879                         btree_d = &btree2[count];
 880                         memmove(btree_d, btree_s, tmp);
 881                 }
 882
 883                 /*
 884                  * Move the req'd B-tree elements from high in node1 to
 885                  * low in node2.
 886                  */
 887                 nodehdr2.count += count;
 888                 tmp = count * (uint)sizeof(xfs_da_node_entry_t);
 889                 btree_s = &btree1[nodehdr1.count - count];
 890                 btree_d = &btree2[0];
 891                 memcpy(btree_d, btree_s, tmp);
 892                 nodehdr1.count -= count;
 893         } else {
 894                 /*
 895                  * Move the req'd B-tree elements from low in node2 to
 896                  * high in node1.
 897                  */
 898                 count = -count;
 899                 tmp = count * (uint)sizeof(xfs_da_node_entry_t);
 900                 btree_s = &btree2[0];
 901                 btree_d = &btree1[nodehdr1.count];
 902                 memcpy(btree_d, btree_s, tmp);
 903                 nodehdr1.count += count;
 904
 905                 xfs_trans_log_buf(tp, blk1->bp,
 906                         XFS_DA_LOGRANGE(node1, btree_d, tmp));
 907
 908                 /*
 909                  * Move elements in node2 down to fill the hole.
 910                  */
 911                 tmp  = nodehdr2.count - count;
 912                 tmp *= (uint)sizeof(xfs_da_node_entry_t);
 913                 btree_s = &btree2[count];
 914                 btree_d = &btree2[0];
 915                 memmove(btree_d, btree_s, tmp);
 916                 nodehdr2.count -= count;
 917         }
 918
 919         /*
 920          * Log header of node 1 and all current bits of node 2.
 921          */
 922         xfs_da3_node_hdr_to_disk(dp->i_mount, node1, &nodehdr1);
 923         xfs_trans_log_buf(tp, blk1->bp,
 924                 XFS_DA_LOGRANGE(node1, &node1->hdr,
 925                                 state->args->geo->node_hdr_size));
 926
 927         xfs_da3_node_hdr_to_disk(dp->i_mount, node2, &nodehdr2);
 928         xfs_trans_log_buf(tp, blk2->bp,
 929                 XFS_DA_LOGRANGE(node2, &node2->hdr,
 930                                 state->args->geo->node_hdr_size +
 931                                 (sizeof(btree2[0]) * nodehdr2.count)));
 932
 933         /*
 934          * Record the last hashval from each block for upward propagation.
 935          * (note: don't use the swapped node pointers)
 936          */
 937         if (swap) {
 938                 node1 = blk1->bp->b_addr;
 939                 node2 = blk2->bp->b_addr;
 940                 xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr1, node1);
 941                 xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr2, node2);
 942                 btree1 = nodehdr1.btree;
 943                 btree2 = nodehdr2.btree;
 944         }
 945         blk1->hashval = be32_to_cpu(btree1[nodehdr1.count - 1].hashval);
 946         blk2->hashval = be32_to_cpu(btree2[nodehdr2.count - 1].hashval);
 947
 948         /*
 949          * Adjust the expected index for insertion.
 950          */
 951         if (blk1->index >= nodehdr1.count) {
 952                 blk2->index = blk1->index - nodehdr1.count;
 953                 blk1->index = nodehdr1.count + 1;       /* make it invalid */
 954         }
 955 }
 956
 957 /*
 958  * Add a new entry to an intermediate node.
 959  */
 960 STATIC void
 961 xfs_da3_node_add(
 962         struct xfs_da_state     *state,
 963         struct xfs_da_state_blk *oldblk,
 964         struct xfs_da_state_blk *newblk)
 965 {
 966         struct xfs_da_intnode   *node;
 967         struct xfs_da3_icnode_hdr nodehdr;
 968         struct xfs_da_node_entry *btree;
 969         int                     tmp;
 970         struct xfs_inode        *dp = state->args->dp;
 971
 972         trace_xfs_da_node_add(state->args);
 973
 974         node = oldblk->bp->b_addr;
 975         xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
 976         btree = nodehdr.btree;
 977
 978         ASSERT(oldblk->index >= 0 && oldblk->index <= nodehdr.count);
 979         ASSERT(newblk->blkno != 0);
 980         if (state->args->whichfork == XFS_DATA_FORK)
 981                 ASSERT(newblk->blkno >= state->args->geo->leafblk &&
 982                        newblk->blkno < state->args->geo->freeblk);
 983
 984         /*
 985          * We may need to make some room before we insert the new node.
 986          */
 987         tmp = 0;
 988         if (oldblk->index < nodehdr.count) {
 989                 tmp = (nodehdr.count - oldblk->index) * (uint)sizeof(*btree);
 990                 memmove(&btree[oldblk->index + 1], &btree[oldblk->index], tmp);
 991         }
 992         btree[oldblk->index].hashval = cpu_to_be32(newblk->hashval);
 993         btree[oldblk->index].before = cpu_to_be32(newblk->blkno);
 994         xfs_trans_log_buf(state->args->trans, oldblk->bp,
 995                 XFS_DA_LOGRANGE(node, &btree[oldblk->index],
 996                                 tmp + sizeof(*btree)));
 997
 998         nodehdr.count += 1;
 999         xfs_da3_node_hdr_to_disk(dp->i_mount, node, &nodehdr);
1000         xfs_trans_log_buf(state->args->trans, oldblk->bp,
1001                 XFS_DA_LOGRANGE(node, &node->hdr,
1002                                 state->args->geo->node_hdr_size));
1003
1004         /*
1005          * Copy the last hash value from the oldblk to propagate upwards.
1006          */
1007         oldblk->hashval = be32_to_cpu(btree[nodehdr.count - 1].hashval);
1008 }
1009
1010 /*========================================================================
1011  * Routines used for shrinking the Btree.
1012  *========================================================================*/
1013
1014 /*
1015  * Deallocate an empty leaf node, remove it from its parent,
1016  * possibly deallocating that block, etc...
1017  */
1018 int
1019 xfs_da3_join(
1020         struct xfs_da_state     *state)
1021 {
1022         struct xfs_da_state_blk *drop_blk;
1023         struct xfs_da_state_blk *save_blk;
1024         int                     action = 0;
1025         int                     error;
1026
1027         trace_xfs_da_join(state->args);
1028
1029         drop_blk = &state->path.blk[ state->path.active-1 ];
1030         save_blk = &state->altpath.blk[ state->path.active-1 ];
1031         ASSERT(state->path.blk[0].magic == XFS_DA_NODE_MAGIC);
1032         ASSERT(drop_blk->magic == XFS_ATTR_LEAF_MAGIC ||
1033                drop_blk->magic == XFS_DIR2_LEAFN_MAGIC);
1034
1035         /*
1036          * Walk back up the tree joining/deallocating as necessary.
1037          * When we stop dropping blocks, break out.
1038          */
1039         for (  ; state->path.active >= 2; drop_blk--, save_blk--,
1040                  state->path.active--) {
1041                 /*
1042                  * See if we can combine the block with a neighbor.
1043                  *   (action == 0) => no options, just leave
1044                  *   (action == 1) => coalesce, then unlink
1045                  *   (action == 2) => block empty, unlink it
1046                  */
1047                 switch (drop_blk->magic) {
1048                 case XFS_ATTR_LEAF_MAGIC:
1049                         error = xfs_attr3_leaf_toosmall(state, &action);
1050                         if (error)
1051                                 return error;
1052                         if (action == 0)
1053                                 return 0;
1054                         xfs_attr3_leaf_unbalance(state, drop_blk, save_blk);
1055                         break;
1056                 case XFS_DIR2_LEAFN_MAGIC:
1057                         error = xfs_dir2_leafn_toosmall(state, &action);
1058                         if (error)
1059                                 return error;
1060                         if (action == 0)
1061                                 return 0;
1062                         xfs_dir2_leafn_unbalance(state, drop_blk, save_blk);
1063                         break;
1064                 case XFS_DA_NODE_MAGIC:
1065                         /*
1066                          * Remove the offending node, fixup hashvals,
1067                          * check for a toosmall neighbor.
1068                          */
1069                         xfs_da3_node_remove(state, drop_blk);
1070                         xfs_da3_fixhashpath(state, &state->path);
1071                         error = xfs_da3_node_toosmall(state, &action);
1072                         if (error)
1073                                 return error;
1074                         if (action == 0)
1075                                 return 0;
1076                         xfs_da3_node_unbalance(state, drop_blk, save_blk);
1077                         break;
1078                 }
1079                 xfs_da3_fixhashpath(state, &state->altpath);
1080                 error = xfs_da3_blk_unlink(state, drop_blk, save_blk);
1081                 xfs_da_state_kill_altpath(state);
1082                 if (error)
1083                         return error;
1084                 error = xfs_da_shrink_inode(state->args, drop_blk->blkno,
1085                                                          drop_blk->bp);
1086                 drop_blk->bp = NULL;
1087                 if (error)
1088                         return error;
1089         }
1090         /*
1091          * We joined all the way to the top.  If it turns out that
1092          * we only have one entry in the root, make the child block
1093          * the new root.
1094          */
1095         xfs_da3_node_remove(state, drop_blk);
1096         xfs_da3_fixhashpath(state, &state->path);
1097         error = xfs_da3_root_join(state, &state->path.blk[0]);
1098         return error;
1099 }
1100
1101 #ifdef  DEBUG
1102 static void
1103 xfs_da_blkinfo_onlychild_validate(struct xfs_da_blkinfo *blkinfo, __u16 level)
1104 {
1105         __be16  magic = blkinfo->magic;
1106
1107         if (level == 1) {
1108                 ASSERT(magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
1109                        magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC) ||
1110                        magic == cpu_to_be16(XFS_ATTR_LEAF_MAGIC) ||
1111                        magic == cpu_to_be16(XFS_ATTR3_LEAF_MAGIC));
1112         } else {
1113                 ASSERT(magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
1114                        magic == cpu_to_be16(XFS_DA3_NODE_MAGIC));
1115         }
1116         ASSERT(!blkinfo->forw);
1117         ASSERT(!blkinfo->back);
1118 }
1119 #else   /* !DEBUG */
1120 #define xfs_da_blkinfo_onlychild_validate(blkinfo, level)
1121 #endif  /* !DEBUG */
1122
1123 /*
1124  * We have only one entry in the root.  Copy the only remaining child of
1125  * the old root to block 0 as the new root node.
1126  */
1127 STATIC int
1128 xfs_da3_root_join(
1129         struct xfs_da_state     *state,
1130         struct xfs_da_state_blk *root_blk)
1131 {
1132         struct xfs_da_intnode   *oldroot;
1133         struct xfs_da_args      *args;
1134         xfs_dablk_t             child;
1135         struct xfs_buf          *bp;
1136         struct xfs_da3_icnode_hdr oldroothdr;
1137         int                     error;
1138         struct xfs_inode        *dp = state->args->dp;
1139
1140         trace_xfs_da_root_join(state->args);
1141
1142         ASSERT(root_blk->magic == XFS_DA_NODE_MAGIC);
1143
1144         args = state->args;
1145         oldroot = root_blk->bp->b_addr;
1146         xfs_da3_node_hdr_from_disk(dp->i_mount, &oldroothdr, oldroot);
1147         ASSERT(oldroothdr.forw == 0);
1148         ASSERT(oldroothdr.back == 0);
1149
1150         /*
1151          * If the root has more than one child, then don't do anything.
1152          */
1153         if (oldroothdr.count > 1)
1154                 return 0;
1155
1156         /*
1157          * Read in the (only) child block, then copy those bytes into
1158          * the root block's buffer and free the original child block.
1159          */
1160         child = be32_to_cpu(oldroothdr.btree[0].before);
1161         ASSERT(child != 0);
1162         error = xfs_da3_node_read(args->trans, dp, child, -1, &bp,
1163                                              args->whichfork);
1164         if (error)
1165                 return error;
1166         xfs_da_blkinfo_onlychild_validate(bp->b_addr, oldroothdr.level);
1167
1168         /*
1169          * This could be copying a leaf back into the root block in the case of
1170          * there only being a single leaf block left in the tree. Hence we have
1171          * to update the b_ops pointer as well to match the buffer type change
1172          * that could occur. For dir3 blocks we also need to update the block
1173          * number in the buffer header.
1174          */
1175         memcpy(root_blk->bp->b_addr, bp->b_addr, args->geo->blksize);
1176         root_blk->bp->b_ops = bp->b_ops;
1177         xfs_trans_buf_copy_type(root_blk->bp, bp);
1178         if (oldroothdr.magic == XFS_DA3_NODE_MAGIC) {
1179                 struct xfs_da3_blkinfo *da3 = root_blk->bp->b_addr;
1180                 da3->blkno = cpu_to_be64(root_blk->bp->b_bn);
1181         }
1182         xfs_trans_log_buf(args->trans, root_blk->bp, 0,
1183                           args->geo->blksize - 1);
1184         error = xfs_da_shrink_inode(args, child, bp);
1185         return error;
1186 }
1187
1188 /*
1189  * Check a node block and its neighbors to see if the block should be
1190  * collapsed into one or the other neighbor.  Always keep the block
1191  * with the smaller block number.
1192  * If the current block is over 50% full, don't try to join it, return 0.
1193  * If the block is empty, fill in the state structure and return 2.
1194  * If it can be collapsed, fill in the state structure and return 1.
1195  * If nothing can be done, return 0.
1196  */
1197 STATIC int
1198 xfs_da3_node_toosmall(
1199         struct xfs_da_state     *state,
1200         int                     *action)
1201 {
1202         struct xfs_da_intnode   *node;
1203         struct xfs_da_state_blk *blk;
1204         struct xfs_da_blkinfo   *info;
1205         xfs_dablk_t             blkno;
1206         struct xfs_buf          *bp;
1207         struct xfs_da3_icnode_hdr nodehdr;
1208         int                     count;
1209         int                     forward;
1210         int                     error;
1211         int                     retval;
1212         int                     i;
1213         struct xfs_inode        *dp = state->args->dp;
1214
1215         trace_xfs_da_node_toosmall(state->args);
1216
1217         /*
1218          * Check for the degenerate case of the block being over 50% full.
1219          * If so, it's not worth even looking to see if we might be able
1220          * to coalesce with a sibling.
1221          */
1222         blk = &state->path.blk[ state->path.active-1 ];
1223         info = blk->bp->b_addr;
1224         node = (xfs_da_intnode_t *)info;
1225         xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
1226         if (nodehdr.count > (state->args->geo->node_ents >> 1)) {
1227                 *action = 0;    /* blk over 50%, don't try to join */
1228                 return 0;       /* blk over 50%, don't try to join */
1229         }
1230
1231         /*
1232          * Check for the degenerate case of the block being empty.
1233          * If the block is empty, we'll simply delete it, no need to
1234          * coalesce it with a sibling block.  We choose (arbitrarily)
1235          * to merge with the forward block unless it is NULL.
1236          */
1237         if (nodehdr.count == 0) {
1238                 /*
1239                  * Make altpath point to the block we want to keep and
1240                  * path point to the block we want to drop (this one).
1241                  */
1242                 forward = (info->forw != 0);
1243                 memcpy(&state->altpath, &state->path, sizeof(state->path));
1244                 error = xfs_da3_path_shift(state, &state->altpath, forward,
1245                                                  0, &retval);
1246                 if (error)
1247                         return error;
1248                 if (retval) {
1249                         *action = 0;
1250                 } else {
1251                         *action = 2;
1252                 }
1253                 return 0;
1254         }
1255
1256         /*
1257          * Examine each sibling block to see if we can coalesce with
1258          * at least 25% free space to spare.  We need to figure out
1259          * whether to merge with the forward or the backward block.
1260          * We prefer coalescing with the lower numbered sibling so as
1261          * to shrink a directory over time.
1262          */
1263         count  = state->args->geo->node_ents;
1264         count -= state->args->geo->node_ents >> 2;
1265         count -= nodehdr.count;
1266
1267         /* start with smaller blk num */
1268         forward = nodehdr.forw < nodehdr.back;
1269         for (i = 0; i < 2; forward = !forward, i++) {
1270                 struct xfs_da3_icnode_hdr thdr;
1271                 if (forward)
1272                         blkno = nodehdr.forw;
1273                 else
1274                         blkno = nodehdr.back;
1275                 if (blkno == 0)
1276                         continue;
1277                 error = xfs_da3_node_read(state->args->trans, dp,
1278                                         blkno, -1, &bp, state->args->whichfork);
1279                 if (error)
1280                         return error;
1281
1282                 node = bp->b_addr;
1283                 xfs_da3_node_hdr_from_disk(dp->i_mount, &thdr, node);
1284                 xfs_trans_brelse(state->args->trans, bp);
1285
1286                 if (count - thdr.count >= 0)
1287                         break;  /* fits with at least 25% to spare */
1288         }
1289         if (i >= 2) {
1290                 *action = 0;
1291                 return 0;
1292         }
1293
1294         /*
1295          * Make altpath point to the block we want to keep (the lower
1296          * numbered block) and path point to the block we want to drop.
1297          */
1298         memcpy(&state->altpath, &state->path, sizeof(state->path));
1299         if (blkno < blk->blkno) {
1300                 error = xfs_da3_path_shift(state, &state->altpath, forward,
1301                                                  0, &retval);
1302         } else {
1303                 error = xfs_da3_path_shift(state, &state->path, forward,
1304                                                  0, &retval);
1305         }
1306         if (error)
1307                 return error;
1308         if (retval) {
1309                 *action = 0;
1310                 return 0;
1311         }
1312         *action = 1;
1313         return 0;
1314 }
1315
1316 /*
1317  * Pick up the last hashvalue from an intermediate node.
1318  */
1319 STATIC uint
1320 xfs_da3_node_lasthash(
1321         struct xfs_inode        *dp,
1322         struct xfs_buf          *bp,
1323         int                     *count)
1324 {
1325         struct xfs_da3_icnode_hdr nodehdr;
1326
1327         xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, bp->b_addr);
1328         if (count)
1329                 *count = nodehdr.count;
1330         if (!nodehdr.count)
1331                 return 0;
1332         return be32_to_cpu(nodehdr.btree[nodehdr.count - 1].hashval);
1333 }
1334
1335 /*
1336  * Walk back up the tree adjusting hash values as necessary,
1337  * when we stop making changes, return.
1338  */
1339 void
1340 xfs_da3_fixhashpath(
1341         struct xfs_da_state     *state,
1342         struct xfs_da_state_path *path)
1343 {
1344         struct xfs_da_state_blk *blk;
1345         struct xfs_da_intnode   *node;
1346         struct xfs_da_node_entry *btree;
1347         xfs_dahash_t            lasthash=0;
1348         int                     level;
1349         int                     count;
1350         struct xfs_inode        *dp = state->args->dp;
1351
1352         trace_xfs_da_fixhashpath(state->args);
1353
1354         level = path->active-1;
1355         blk = &path->blk[ level ];
1356         switch (blk->magic) {
1357         case XFS_ATTR_LEAF_MAGIC:
1358                 lasthash = xfs_attr_leaf_lasthash(blk->bp, &count);
1359                 if (count == 0)
1360                         return;
1361                 break;
1362         case XFS_DIR2_LEAFN_MAGIC:
1363                 lasthash = xfs_dir2_leaf_lasthash(dp, blk->bp, &count);
1364                 if (count == 0)
1365                         return;
1366                 break;
1367         case XFS_DA_NODE_MAGIC:
1368                 lasthash = xfs_da3_node_lasthash(dp, blk->bp, &count);
1369                 if (count == 0)
1370                         return;
1371                 break;
1372         }
1373         for (blk--, level--; level >= 0; blk--, level--) {
1374                 struct xfs_da3_icnode_hdr nodehdr;
1375
1376                 node = blk->bp->b_addr;
1377                 xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
1378                 btree = nodehdr.btree;
1379                 if (be32_to_cpu(btree[blk->index].hashval) == lasthash)
1380                         break;
1381                 blk->hashval = lasthash;
1382                 btree[blk->index].hashval = cpu_to_be32(lasthash);
1383                 xfs_trans_log_buf(state->args->trans, blk->bp,
1384                                   XFS_DA_LOGRANGE(node, &btree[blk->index],
1385                                                   sizeof(*btree)));
1386
1387                 lasthash = be32_to_cpu(btree[nodehdr.count - 1].hashval);
1388         }
1389 }
1390
1391 /*
1392  * Remove an entry from an intermediate node.
1393  */
1394 STATIC void
1395 xfs_da3_node_remove(
1396         struct xfs_da_state     *state,
1397         struct xfs_da_state_blk *drop_blk)
1398 {
1399         struct xfs_da_intnode   *node;
1400         struct xfs_da3_icnode_hdr nodehdr;
1401         struct xfs_da_node_entry *btree;
1402         int                     index;
1403         int                     tmp;
1404         struct xfs_inode        *dp = state->args->dp;
1405
1406         trace_xfs_da_node_remove(state->args);
1407
1408         node = drop_blk->bp->b_addr;
1409         xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
1410         ASSERT(drop_blk->index < nodehdr.count);
1411         ASSERT(drop_blk->index >= 0);
1412
1413         /*
1414          * Copy over the offending entry, or just zero it out.
1415          */
1416         index = drop_blk->index;
1417         btree = nodehdr.btree;
1418         if (index < nodehdr.count - 1) {
1419                 tmp  = nodehdr.count - index - 1;
1420                 tmp *= (uint)sizeof(xfs_da_node_entry_t);
1421                 memmove(&btree[index], &btree[index + 1], tmp);
1422                 xfs_trans_log_buf(state->args->trans, drop_blk->bp,
1423                     XFS_DA_LOGRANGE(node, &btree[index], tmp));
1424                 index = nodehdr.count - 1;
1425         }
1426         memset(&btree[index], 0, sizeof(xfs_da_node_entry_t));
1427         xfs_trans_log_buf(state->args->trans, drop_blk->bp,
1428             XFS_DA_LOGRANGE(node, &btree[index], sizeof(btree[index])));
1429         nodehdr.count -= 1;
1430         xfs_da3_node_hdr_to_disk(dp->i_mount, node, &nodehdr);
1431         xfs_trans_log_buf(state->args->trans, drop_blk->bp,
1432             XFS_DA_LOGRANGE(node, &node->hdr, state->args->geo->node_hdr_size));
1433
1434         /*
1435          * Copy the last hash value from the block to propagate upwards.
1436          */
1437         drop_blk->hashval = be32_to_cpu(btree[index - 1].hashval);
1438 }
1439
1440 /*
1441  * Unbalance the elements between two intermediate nodes,
1442  * move all Btree elements from one node into another.
1443  */
1444 STATIC void
1445 xfs_da3_node_unbalance(
1446         struct xfs_da_state     *state,
1447         struct xfs_da_state_blk *drop_blk,
1448         struct xfs_da_state_blk *save_blk)
1449 {
1450         struct xfs_da_intnode   *drop_node;
1451         struct xfs_da_intnode   *save_node;
1452         struct xfs_da_node_entry *drop_btree;
1453         struct xfs_da_node_entry *save_btree;
1454         struct xfs_da3_icnode_hdr drop_hdr;
1455         struct xfs_da3_icnode_hdr save_hdr;
1456         struct xfs_trans        *tp;
1457         int                     sindex;
1458         int                     tmp;
1459         struct xfs_inode        *dp = state->args->dp;
1460
1461         trace_xfs_da_node_unbalance(state->args);
1462
1463         drop_node = drop_blk->bp->b_addr;
1464         save_node = save_blk->bp->b_addr;
1465         xfs_da3_node_hdr_from_disk(dp->i_mount, &drop_hdr, drop_node);
1466         xfs_da3_node_hdr_from_disk(dp->i_mount, &save_hdr, save_node);
1467         drop_btree = drop_hdr.btree;
1468         save_btree = save_hdr.btree;
1469         tp = state->args->trans;
1470
1471         /*
1472          * If the dying block has lower hashvals, then move all the
1473          * elements in the remaining block up to make a hole.
1474          */
1475         if ((be32_to_cpu(drop_btree[0].hashval) <
1476                         be32_to_cpu(save_btree[0].hashval)) ||
1477             (be32_to_cpu(drop_btree[drop_hdr.count - 1].hashval) <
1478                         be32_to_cpu(save_btree[save_hdr.count - 1].hashval))) {
1479                 /* XXX: check this - is memmove dst correct? */
1480                 tmp = save_hdr.count * sizeof(xfs_da_node_entry_t);
1481                 memmove(&save_btree[drop_hdr.count], &save_btree[0], tmp);
1482
1483                 sindex = 0;
1484                 xfs_trans_log_buf(tp, save_blk->bp,
1485                         XFS_DA_LOGRANGE(save_node, &save_btree[0],
1486                                 (save_hdr.count + drop_hdr.count) *
1487                                                 sizeof(xfs_da_node_entry_t)));
1488         } else {
1489                 sindex = save_hdr.count;
1490                 xfs_trans_log_buf(tp, save_blk->bp,
1491                         XFS_DA_LOGRANGE(save_node, &save_btree[sindex],
1492                                 drop_hdr.count * sizeof(xfs_da_node_entry_t)));
1493         }
1494
1495         /*
1496          * Move all the B-tree elements from drop_blk to save_blk.
1497          */
1498         tmp = drop_hdr.count * (uint)sizeof(xfs_da_node_entry_t);
1499         memcpy(&save_btree[sindex], &drop_btree[0], tmp);
1500         save_hdr.count += drop_hdr.count;
1501
1502         xfs_da3_node_hdr_to_disk(dp->i_mount, save_node, &save_hdr);
1503         xfs_trans_log_buf(tp, save_blk->bp,
1504                 XFS_DA_LOGRANGE(save_node, &save_node->hdr,
1505                                 state->args->geo->node_hdr_size));
1506
1507         /*
1508          * Save the last hashval in the remaining block for upward propagation.
1509          */
1510         save_blk->hashval = be32_to_cpu(save_btree[save_hdr.count - 1].hashval);
1511 }
1512
1513 /*========================================================================
1514  * Routines used for finding things in the Btree.
1515  *========================================================================*/
1516
1517 /*
1518  * Walk down the Btree looking for a particular filename, filling
1519  * in the state structure as we go.
1520  *
1521  * We will set the state structure to point to each of the elements
1522  * in each of the nodes where either the hashval is or should be.
1523  *
1524  * We support duplicate hashval's so for each entry in the current
1525  * node that could contain the desired hashval, descend.  This is a
1526  * pruned depth-first tree search.
1527  */
1528 int                                                     /* error */
1529 xfs_da3_node_lookup_int(
1530         struct xfs_da_state     *state,
1531         int                     *result)
1532 {
1533         struct xfs_da_state_blk *blk;
1534         struct xfs_da_blkinfo   *curr;
1535         struct xfs_da_intnode   *node;
1536         struct xfs_da_node_entry *btree;
1537         struct xfs_da3_icnode_hdr nodehdr;
1538         struct xfs_da_args      *args;
1539         xfs_dablk_t             blkno;
1540         xfs_dahash_t            hashval;
1541         xfs_dahash_t            btreehashval;
1542         int                     probe;
1543         int                     span;
1544         int                     max;
1545         int                     error;
1546         int                     retval;
1547         unsigned int            expected_level = 0;
1548         uint16_t                magic;
1549         struct xfs_inode        *dp = state->args->dp;
1550
1551         args = state->args;
1552
1553         /*
1554          * Descend thru the B-tree searching each level for the right
1555          * node to use, until the right hashval is found.
1556          */
1557         blkno = args->geo->leafblk;
1558         for (blk = &state->path.blk[0], state->path.active = 1;
1559                          state->path.active <= XFS_DA_NODE_MAXDEPTH;
1560                          blk++, state->path.active++) {
1561                 /*
1562                  * Read the next node down in the tree.
1563                  */
1564                 blk->blkno = blkno;
1565                 error = xfs_da3_node_read(args->trans, args->dp, blkno,
1566                                         -1, &blk->bp, args->whichfork);
1567                 if (error) {
1568                         blk->blkno = 0;
1569                         state->path.active--;
1570                         return error;
1571                 }
1572                 curr = blk->bp->b_addr;
1573                 magic = be16_to_cpu(curr->magic);
1574
1575                 if (magic == XFS_ATTR_LEAF_MAGIC ||
1576                     magic == XFS_ATTR3_LEAF_MAGIC) {
1577                         blk->magic = XFS_ATTR_LEAF_MAGIC;
1578                         blk->hashval = xfs_attr_leaf_lasthash(blk->bp, NULL);
1579                         break;
1580                 }
1581
1582                 if (magic == XFS_DIR2_LEAFN_MAGIC ||
1583                     magic == XFS_DIR3_LEAFN_MAGIC) {
1584                         blk->magic = XFS_DIR2_LEAFN_MAGIC;
1585                         blk->hashval = xfs_dir2_leaf_lasthash(args->dp,
1586                                                               blk->bp, NULL);
1587                         break;
1588                 }
1589
1590                 if (magic != XFS_DA_NODE_MAGIC && magic != XFS_DA3_NODE_MAGIC) {
1591                         xfs_buf_corruption_error(blk->bp);
1592                         return -EFSCORRUPTED;
1593                 }
1594
1595                 blk->magic = XFS_DA_NODE_MAGIC;
1596
1597                 /*
1598                  * Search an intermediate node for a match.
1599                  */
1600                 node = blk->bp->b_addr;
1601                 xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
1602                 btree = nodehdr.btree;
1603
1604                 /* Tree taller than we can handle; bail out! */
1605                 if (nodehdr.level >= XFS_DA_NODE_MAXDEPTH) {
1606                         xfs_buf_corruption_error(blk->bp);
1607                         return -EFSCORRUPTED;
1608                 }
1609
1610                 /* Check the level from the root. */
1611                 if (blkno == args->geo->leafblk)
1612                         expected_level = nodehdr.level - 1;
1613                 else if (expected_level != nodehdr.level) {
1614                         xfs_buf_corruption_error(blk->bp);
1615                         return -EFSCORRUPTED;
1616                 } else
1617                         expected_level--;
1618
1619                 max = nodehdr.count;
1620                 blk->hashval = be32_to_cpu(btree[max - 1].hashval);
1621
1622                 /*
1623                  * Binary search.  (note: small blocks will skip loop)
1624                  */
1625                 probe = span = max / 2;
1626                 hashval = args->hashval;
1627                 while (span > 4) {
1628                         span /= 2;
1629                         btreehashval = be32_to_cpu(btree[probe].hashval);
1630                         if (btreehashval < hashval)
1631                                 probe += span;
1632                         else if (btreehashval > hashval)
1633                                 probe -= span;
1634                         else
1635                                 break;
1636                 }
1637                 ASSERT((probe >= 0) && (probe < max));
1638                 ASSERT((span <= 4) ||
1639                         (be32_to_cpu(btree[probe].hashval) == hashval));
1640
1641                 /*
1642                  * Since we may have duplicate hashval's, find the first
1643                  * matching hashval in the node.
1644                  */
1645                 while (probe > 0 &&
1646                        be32_to_cpu(btree[probe].hashval) >= hashval) {
1647                         probe--;
1648                 }
1649                 while (probe < max &&
1650                        be32_to_cpu(btree[probe].hashval) < hashval) {
1651                         probe++;
1652                 }
1653
1654                 /*
1655                  * Pick the right block to descend on.
1656                  */
1657                 if (probe == max) {
1658                         blk->index = max - 1;
1659                         blkno = be32_to_cpu(btree[max - 1].before);
1660                 } else {
1661                         blk->index = probe;
1662                         blkno = be32_to_cpu(btree[probe].before);
1663                 }
1664
1665                 /* We can't point back to the root. */
1666                 if (XFS_IS_CORRUPT(dp->i_mount, blkno == args->geo->leafblk))
1667                         return -EFSCORRUPTED;
1668         }
1669
1670         if (XFS_IS_CORRUPT(dp->i_mount, expected_level != 0))
1671                 return -EFSCORRUPTED;
1672
1673         /*
1674          * A leaf block that ends in the hashval that we are interested in
1675          * (final hashval == search hashval) means that the next block may
1676          * contain more entries with the same hashval, shift upward to the
1677          * next leaf and keep searching.
1678          */
1679         for (;;) {
1680                 if (blk->magic == XFS_DIR2_LEAFN_MAGIC) {
1681                         retval = xfs_dir2_leafn_lookup_int(blk->bp, args,
1682                                                         &blk->index, state);
1683                 } else if (blk->magic == XFS_ATTR_LEAF_MAGIC) {
1684                         retval = xfs_attr3_leaf_lookup_int(blk->bp, args);
1685                         blk->index = args->index;
1686                         args->blkno = blk->blkno;
1687                 } else {
1688                         ASSERT(0);
1689                         return -EFSCORRUPTED;
1690                 }
1691                 if (((retval == -ENOENT) || (retval == -ENOATTR)) &&
1692                     (blk->hashval == args->hashval)) {
1693                         error = xfs_da3_path_shift(state, &state->path, 1, 1,
1694                                                          &retval);
1695                         if (error)
1696                                 return error;
1697                         if (retval == 0) {
1698                                 continue;
1699                         } else if (blk->magic == XFS_ATTR_LEAF_MAGIC) {
1700                                 /* path_shift() gives ENOENT */
1701                                 retval = -ENOATTR;
1702                         }
1703                 }
1704                 break;
1705         }
1706         *result = retval;
1707         return 0;
1708 }
1709
1710 /*========================================================================
1711  * Utility routines.
1712  *========================================================================*/
1713
1714 /*
1715  * Compare two intermediate nodes for "order".
1716  */
1717 STATIC int
1718 xfs_da3_node_order(
1719         struct xfs_inode *dp,
1720         struct xfs_buf  *node1_bp,
1721         struct xfs_buf  *node2_bp)
1722 {
1723         struct xfs_da_intnode   *node1;
1724         struct xfs_da_intnode   *node2;
1725         struct xfs_da_node_entry *btree1;
1726         struct xfs_da_node_entry *btree2;
1727         struct xfs_da3_icnode_hdr node1hdr;
1728         struct xfs_da3_icnode_hdr node2hdr;
1729
1730         node1 = node1_bp->b_addr;
1731         node2 = node2_bp->b_addr;
1732         xfs_da3_node_hdr_from_disk(dp->i_mount, &node1hdr, node1);
1733         xfs_da3_node_hdr_from_disk(dp->i_mount, &node2hdr, node2);
1734         btree1 = node1hdr.btree;
1735         btree2 = node2hdr.btree;
1736
1737         if (node1hdr.count > 0 && node2hdr.count > 0 &&
1738             ((be32_to_cpu(btree2[0].hashval) < be32_to_cpu(btree1[0].hashval)) ||
1739              (be32_to_cpu(btree2[node2hdr.count - 1].hashval) <
1740               be32_to_cpu(btree1[node1hdr.count - 1].hashval)))) {
1741                 return 1;
1742         }
1743         return 0;
1744 }
1745
1746 /*
1747  * Link a new block into a doubly linked list of blocks (of whatever type).
1748  */
1749 int                                                     /* error */
1750 xfs_da3_blk_link(
1751         struct xfs_da_state     *state,
1752         struct xfs_da_state_blk *old_blk,
1753         struct xfs_da_state_blk *new_blk)
1754 {
1755         struct xfs_da_blkinfo   *old_info;
1756         struct xfs_da_blkinfo   *new_info;
1757         struct xfs_da_blkinfo   *tmp_info;
1758         struct xfs_da_args      *args;
1759         struct xfs_buf          *bp;
1760         int                     before = 0;
1761         int                     error;
1762         struct xfs_inode        *dp = state->args->dp;
1763
1764         /*
1765          * Set up environment.
1766          */
1767         args = state->args;
1768         ASSERT(args != NULL);
1769         old_info = old_blk->bp->b_addr;
1770         new_info = new_blk->bp->b_addr;
1771         ASSERT(old_blk->magic == XFS_DA_NODE_MAGIC ||
1772                old_blk->magic == XFS_DIR2_LEAFN_MAGIC ||
1773                old_blk->magic == XFS_ATTR_LEAF_MAGIC);
1774
1775         switch (old_blk->magic) {
1776         case XFS_ATTR_LEAF_MAGIC:
1777                 before = xfs_attr_leaf_order(old_blk->bp, new_blk->bp);
1778                 break;
1779         case XFS_DIR2_LEAFN_MAGIC:
1780                 before = xfs_dir2_leafn_order(dp, old_blk->bp, new_blk->bp);
1781                 break;
1782         case XFS_DA_NODE_MAGIC:
1783                 before = xfs_da3_node_order(dp, old_blk->bp, new_blk->bp);
1784                 break;
1785         }
1786
1787         /*
1788          * Link blocks in appropriate order.
1789          */
1790         if (before) {
1791                 /*
1792                  * Link new block in before existing block.
1793                  */
1794                 trace_xfs_da_link_before(args);
1795                 new_info->forw = cpu_to_be32(old_blk->blkno);
1796                 new_info->back = old_info->back;
1797                 if (old_info->back) {
1798                         error = xfs_da3_node_read(args->trans, dp,
1799                                                 be32_to_cpu(old_info->back),
1800                                                 -1, &bp, args->whichfork);
1801                         if (error)
1802                                 return error;
1803                         ASSERT(bp != NULL);
1804                         tmp_info = bp->b_addr;
1805                         ASSERT(tmp_info->magic == old_info->magic);
1806                         ASSERT(be32_to_cpu(tmp_info->forw) == old_blk->blkno);
1807                         tmp_info->forw = cpu_to_be32(new_blk->blkno);
1808                         xfs_trans_log_buf(args->trans, bp, 0, sizeof(*tmp_info)-1);
1809                 }
1810                 old_info->back = cpu_to_be32(new_blk->blkno);
1811         } else {
1812                 /*
1813                  * Link new block in after existing block.
1814                  */
1815                 trace_xfs_da_link_after(args);
1816                 new_info->forw = old_info->forw;
1817                 new_info->back = cpu_to_be32(old_blk->blkno);
1818                 if (old_info->forw) {
1819                         error = xfs_da3_node_read(args->trans, dp,
1820                                                 be32_to_cpu(old_info->forw),
1821                                                 -1, &bp, args->whichfork);
1822                         if (error)
1823                                 return error;
1824                         ASSERT(bp != NULL);
1825                         tmp_info = bp->b_addr;
1826                         ASSERT(tmp_info->magic == old_info->magic);
1827                         ASSERT(be32_to_cpu(tmp_info->back) == old_blk->blkno);
1828                         tmp_info->back = cpu_to_be32(new_blk->blkno);
1829                         xfs_trans_log_buf(args->trans, bp, 0, sizeof(*tmp_info)-1);
1830                 }
1831                 old_info->forw = cpu_to_be32(new_blk->blkno);
1832         }
1833
1834         xfs_trans_log_buf(args->trans, old_blk->bp, 0, sizeof(*tmp_info) - 1);
1835         xfs_trans_log_buf(args->trans, new_blk->bp, 0, sizeof(*tmp_info) - 1);
1836         return 0;
1837 }
1838
1839 /*
1840  * Unlink a block from a doubly linked list of blocks.
1841  */
1842 STATIC int                                              /* error */
1843 xfs_da3_blk_unlink(
1844         struct xfs_da_state     *state,
1845         struct xfs_da_state_blk *drop_blk,
1846         struct xfs_da_state_blk *save_blk)
1847 {
1848         struct xfs_da_blkinfo   *drop_info;
1849         struct xfs_da_blkinfo   *save_info;
1850         struct xfs_da_blkinfo   *tmp_info;
1851         struct xfs_da_args      *args;
1852         struct xfs_buf          *bp;
1853         int                     error;
1854
1855         /*
1856          * Set up environment.
1857          */
1858         args = state->args;
1859         ASSERT(args != NULL);
1860         save_info = save_blk->bp->b_addr;
1861         drop_info = drop_blk->bp->b_addr;
1862         ASSERT(save_blk->magic == XFS_DA_NODE_MAGIC ||
1863                save_blk->magic == XFS_DIR2_LEAFN_MAGIC ||
1864                save_blk->magic == XFS_ATTR_LEAF_MAGIC);
1865         ASSERT(save_blk->magic == drop_blk->magic);
1866         ASSERT((be32_to_cpu(save_info->forw) == drop_blk->blkno) ||
1867                (be32_to_cpu(save_info->back) == drop_blk->blkno));
1868         ASSERT((be32_to_cpu(drop_info->forw) == save_blk->blkno) ||
1869                (be32_to_cpu(drop_info->back) == save_blk->blkno));
1870
1871         /*
1872          * Unlink the leaf block from the doubly linked chain of leaves.
1873          */
1874         if (be32_to_cpu(save_info->back) == drop_blk->blkno) {
1875                 trace_xfs_da_unlink_back(args);
1876                 save_info->back = drop_info->back;
1877                 if (drop_info->back) {
1878                         error = xfs_da3_node_read(args->trans, args->dp,
1879                                                 be32_to_cpu(drop_info->back),
1880                                                 -1, &bp, args->whichfork);
1881                         if (error)
1882                                 return error;
1883                         ASSERT(bp != NULL);
1884                         tmp_info = bp->b_addr;
1885                         ASSERT(tmp_info->magic == save_info->magic);
1886                         ASSERT(be32_to_cpu(tmp_info->forw) == drop_blk->blkno);
1887                         tmp_info->forw = cpu_to_be32(save_blk->blkno);
1888                         xfs_trans_log_buf(args->trans, bp, 0,
1889                                                     sizeof(*tmp_info) - 1);
1890                 }
1891         } else {
1892                 trace_xfs_da_unlink_forward(args);
1893                 save_info->forw = drop_info->forw;
1894                 if (drop_info->forw) {
1895                         error = xfs_da3_node_read(args->trans, args->dp,
1896                                                 be32_to_cpu(drop_info->forw),
1897                                                 -1, &bp, args->whichfork);
1898                         if (error)
1899                                 return error;
1900                         ASSERT(bp != NULL);
1901                         tmp_info = bp->b_addr;
1902                         ASSERT(tmp_info->magic == save_info->magic);
1903                         ASSERT(be32_to_cpu(tmp_info->back) == drop_blk->blkno);
1904                         tmp_info->back = cpu_to_be32(save_blk->blkno);
1905                         xfs_trans_log_buf(args->trans, bp, 0,
1906                                                     sizeof(*tmp_info) - 1);
1907                 }
1908         }
1909
1910         xfs_trans_log_buf(args->trans, save_blk->bp, 0, sizeof(*save_info) - 1);
1911         return 0;
1912 }
1913
1914 /*
1915  * Move a path "forward" or "!forward" one block at the current level.
1916  *
1917  * This routine will adjust a "path" to point to the next block
1918  * "forward" (higher hashvalues) or "!forward" (lower hashvals) in the
1919  * Btree, including updating pointers to the intermediate nodes between
1920  * the new bottom and the root.
1921  */
1922 int                                                     /* error */
1923 xfs_da3_path_shift(
1924         struct xfs_da_state     *state,
1925         struct xfs_da_state_path *path,
1926         int                     forward,
1927         int                     release,
1928         int                     *result)
1929 {
1930         struct xfs_da_state_blk *blk;
1931         struct xfs_da_blkinfo   *info;
1932         struct xfs_da_args      *args;
1933         struct xfs_da_node_entry *btree;
1934         struct xfs_da3_icnode_hdr nodehdr;
1935         struct xfs_buf          *bp;
1936         xfs_dablk_t             blkno = 0;
1937         int                     level;
1938         int                     error;
1939         struct xfs_inode        *dp = state->args->dp;
1940
1941         trace_xfs_da_path_shift(state->args);
1942
1943         /*
1944          * Roll up the Btree looking for the first block where our
1945          * current index is not at the edge of the block.  Note that
1946          * we skip the bottom layer because we want the sibling block.
1947          */
1948         args = state->args;
1949         ASSERT(args != NULL);
1950         ASSERT(path != NULL);
1951         ASSERT((path->active > 0) && (path->active < XFS_DA_NODE_MAXDEPTH));
1952         level = (path->active-1) - 1;   /* skip bottom layer in path */
1953         for (blk = &path->blk[level]; level >= 0; blk--, level--) {
1954                 xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr,
1955                                            blk->bp->b_addr);
1956
1957                 if (forward && (blk->index < nodehdr.count - 1)) {
1958                         blk->index++;
1959                         blkno = be32_to_cpu(nodehdr.btree[blk->index].before);
1960                         break;
1961                 } else if (!forward && (blk->index > 0)) {
1962                         blk->index--;
1963                         blkno = be32_to_cpu(nodehdr.btree[blk->index].before);
1964                         break;
1965                 }
1966         }
1967         if (level < 0) {
1968                 *result = -ENOENT;      /* we're out of our tree */
1969                 ASSERT(args->op_flags & XFS_DA_OP_OKNOENT);
1970                 return 0;
1971         }
1972
1973         /*
1974          * Roll down the edge of the subtree until we reach the
1975          * same depth we were at originally.
1976          */
1977         for (blk++, level++; level < path->active; blk++, level++) {
1978                 /*
1979                  * Read the next child block into a local buffer.
1980                  */
1981                 error = xfs_da3_node_read(args->trans, dp, blkno, -1, &bp,
1982                                           args->whichfork);
1983                 if (error)
1984                         return error;
1985
1986                 /*
1987                  * Release the old block (if it's dirty, the trans doesn't
1988                  * actually let go) and swap the local buffer into the path
1989                  * structure. This ensures failure of the above read doesn't set
1990                  * a NULL buffer in an active slot in the path.
1991                  */
1992                 if (release)
1993                         xfs_trans_brelse(args->trans, blk->bp);
1994                 blk->blkno = blkno;
1995                 blk->bp = bp;
1996
1997                 info = blk->bp->b_addr;
1998                 ASSERT(info->magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
1999                        info->magic == cpu_to_be16(XFS_DA3_NODE_MAGIC) ||
2000                        info->magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
2001                        info->magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC) ||
2002                        info->magic == cpu_to_be16(XFS_ATTR_LEAF_MAGIC) ||
2003                        info->magic == cpu_to_be16(XFS_ATTR3_LEAF_MAGIC));
2004
2005
2006                 /*
2007                  * Note: we flatten the magic number to a single type so we
2008                  * don't have to compare against crc/non-crc types elsewhere.
2009                  */
2010                 switch (be16_to_cpu(info->magic)) {
2011                 case XFS_DA_NODE_MAGIC:
2012                 case XFS_DA3_NODE_MAGIC:
2013                         blk->magic = XFS_DA_NODE_MAGIC;
2014                         xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr,
2015                                                    bp->b_addr);
2016                         btree = nodehdr.btree;
2017                         blk->hashval = be32_to_cpu(btree[nodehdr.count - 1].hashval);
2018                         if (forward)
2019                                 blk->index = 0;
2020                         else
2021                                 blk->index = nodehdr.count - 1;
2022                         blkno = be32_to_cpu(btree[blk->index].before);
2023                         break;
2024                 case XFS_ATTR_LEAF_MAGIC:
2025                 case XFS_ATTR3_LEAF_MAGIC:
2026                         blk->magic = XFS_ATTR_LEAF_MAGIC;
2027                         ASSERT(level == path->active-1);
2028                         blk->index = 0;
2029                         blk->hashval = xfs_attr_leaf_lasthash(blk->bp, NULL);
2030                         break;
2031                 case XFS_DIR2_LEAFN_MAGIC:
2032                 case XFS_DIR3_LEAFN_MAGIC:
2033                         blk->magic = XFS_DIR2_LEAFN_MAGIC;
2034                         ASSERT(level == path->active-1);
2035                         blk->index = 0;
2036                         blk->hashval = xfs_dir2_leaf_lasthash(args->dp,
2037                                                               blk->bp, NULL);
2038                         break;
2039                 default:
2040                         ASSERT(0);
2041                         break;
2042                 }
2043         }
2044         *result = 0;
2045         return 0;
2046 }
2047
2048
2049 /*========================================================================
2050  * Utility routines.
2051  *========================================================================*/
2052
2053 /*
2054  * Implement a simple hash on a character string.
2055  * Rotate the hash value by 7 bits, then XOR each character in.
2056  * This is implemented with some source-level loop unrolling.
2057  */
2058 xfs_dahash_t
2059 xfs_da_hashname(const uint8_t *name, int namelen)
2060 {
2061         xfs_dahash_t hash;
2062
2063         /*
2064          * Do four characters at a time as long as we can.
2065          */
2066         for (hash = 0; namelen >= 4; namelen -= 4, name += 4)
2067                 hash = (name[0] << 21) ^ (name[1] << 14) ^ (name[2] << 7) ^
2068                        (name[3] << 0) ^ rol32(hash, 7 * 4);
2069
2070         /*
2071          * Now do the rest of the characters.
2072          */
2073         switch (namelen) {
2074         case 3:
2075                 return (name[0] << 14) ^ (name[1] << 7) ^ (name[2] << 0) ^
2076                        rol32(hash, 7 * 3);
2077         case 2:
2078                 return (name[0] << 7) ^ (name[1] << 0) ^ rol32(hash, 7 * 2);
2079         case 1:
2080                 return (name[0] << 0) ^ rol32(hash, 7 * 1);
2081         default: /* case 0: */
2082                 return hash;
2083         }
2084 }
2085
2086 enum xfs_dacmp
2087 xfs_da_compname(
2088         struct xfs_da_args *args,
2089         const unsigned char *name,
2090         int             len)
2091 {
2092         return (args->namelen == len && memcmp(args->name, name, len) == 0) ?
2093                                         XFS_CMP_EXACT : XFS_CMP_DIFFERENT;
2094 }
2095
2096 static xfs_dahash_t
2097 xfs_default_hashname(
2098         struct xfs_name *name)
2099 {
2100         return xfs_da_hashname(name->name, name->len);
2101 }
2102
2103 const struct xfs_nameops xfs_default_nameops = {
2104         .hashname       = xfs_default_hashname,
2105         .compname       = xfs_da_compname
2106 };
2107
2108 int
2109 xfs_da_grow_inode_int(
2110         struct xfs_da_args      *args,
2111         xfs_fileoff_t           *bno,
2112         int                     count)
2113 {
2114         struct xfs_trans        *tp = args->trans;
2115         struct xfs_inode        *dp = args->dp;
2116         int                     w = args->whichfork;
2117         xfs_rfsblock_t          nblks = dp->i_d.di_nblocks;
2118         struct xfs_bmbt_irec    map, *mapp;
2119         int                     nmap, error, got, i, mapi;
2120
2121         /*
2122          * Find a spot in the file space to put the new block.
2123          */
2124         error = xfs_bmap_first_unused(tp, dp, count, bno, w);
2125         if (error)
2126                 return error;
2127
2128         /*
2129          * Try mapping it in one filesystem block.
2130          */
2131         nmap = 1;
2132         error = xfs_bmapi_write(tp, dp, *bno, count,
2133                         xfs_bmapi_aflag(w)|XFS_BMAPI_METADATA|XFS_BMAPI_CONTIG,
2134                         args->total, &map, &nmap);
2135         if (error)
2136                 return error;
2137
2138         ASSERT(nmap <= 1);
2139         if (nmap == 1) {
2140                 mapp = &map;
2141                 mapi = 1;
2142         } else if (nmap == 0 && count > 1) {
2143                 xfs_fileoff_t           b;
2144                 int                     c;
2145
2146                 /*
2147                  * If we didn't get it and the block might work if fragmented,
2148                  * try without the CONTIG flag.  Loop until we get it all.
2149                  */
2150                 mapp = kmem_alloc(sizeof(*mapp) * count, 0);
2151                 for (b = *bno, mapi = 0; b < *bno + count; ) {
2152                         nmap = min(XFS_BMAP_MAX_NMAP, count);
2153                         c = (int)(*bno + count - b);
2154                         error = xfs_bmapi_write(tp, dp, b, c,
2155                                         xfs_bmapi_aflag(w)|XFS_BMAPI_METADATA,
2156                                         args->total, &mapp[mapi], &nmap);
2157                         if (error)
2158                                 goto out_free_map;
2159                         if (nmap < 1)
2160                                 break;
2161                         mapi += nmap;
2162                         b = mapp[mapi - 1].br_startoff +
2163                             mapp[mapi - 1].br_blockcount;
2164                 }
2165         } else {
2166                 mapi = 0;
2167                 mapp = NULL;
2168         }
2169
2170         /*
2171          * Count the blocks we got, make sure it matches the total.
2172          */
2173         for (i = 0, got = 0; i < mapi; i++)
2174                 got += mapp[i].br_blockcount;
2175         if (got != count || mapp[0].br_startoff != *bno ||
2176             mapp[mapi - 1].br_startoff + mapp[mapi - 1].br_blockcount !=
2177             *bno + count) {
2178                 error = -ENOSPC;
2179                 goto out_free_map;
2180         }
2181
2182         /* account for newly allocated blocks in reserved blocks total */
2183         args->total -= dp->i_d.di_nblocks - nblks;
2184
2185 out_free_map:
2186         if (mapp != &map)
2187                 kmem_free(mapp);
2188         return error;
2189 }
2190
2191 /*
2192  * Add a block to the btree ahead of the file.
2193  * Return the new block number to the caller.
2194  */
2195 int
2196 xfs_da_grow_inode(
2197         struct xfs_da_args      *args,
2198         xfs_dablk_t             *new_blkno)
2199 {
2200         xfs_fileoff_t           bno;
2201         int                     error;
2202
2203         trace_xfs_da_grow_inode(args);
2204
2205         bno = args->geo->leafblk;
2206         error = xfs_da_grow_inode_int(args, &bno, args->geo->fsbcount);
2207         if (!error)
2208                 *new_blkno = (xfs_dablk_t)bno;
2209         return error;
2210 }
2211
2212 /*
2213  * Ick.  We need to always be able to remove a btree block, even
2214  * if there's no space reservation because the filesystem is full.
2215  * This is called if xfs_bunmapi on a btree block fails due to ENOSPC.
2216  * It swaps the target block with the last block in the file.  The
2217  * last block in the file can always be removed since it can't cause
2218  * a bmap btree split to do that.
2219  */
2220 STATIC int
2221 xfs_da3_swap_lastblock(
2222         struct xfs_da_args      *args,
2223         xfs_dablk_t             *dead_blknop,
2224         struct xfs_buf          **dead_bufp)
2225 {
2226         struct xfs_da_blkinfo   *dead_info;
2227         struct xfs_da_blkinfo   *sib_info;
2228         struct xfs_da_intnode   *par_node;
2229         struct xfs_da_intnode   *dead_node;
2230         struct xfs_dir2_leaf    *dead_leaf2;
2231         struct xfs_da_node_entry *btree;
2232         struct xfs_da3_icnode_hdr par_hdr;
2233         struct xfs_inode        *dp;
2234         struct xfs_trans        *tp;
2235         struct xfs_mount        *mp;
2236         struct xfs_buf          *dead_buf;
2237         struct xfs_buf          *last_buf;
2238         struct xfs_buf          *sib_buf;
2239         struct xfs_buf          *par_buf;
2240         xfs_dahash_t            dead_hash;
2241         xfs_fileoff_t           lastoff;
2242         xfs_dablk_t             dead_blkno;
2243         xfs_dablk_t             last_blkno;
2244         xfs_dablk_t             sib_blkno;
2245         xfs_dablk_t             par_blkno;
2246         int                     error;
2247         int                     w;
2248         int                     entno;
2249         int                     level;
2250         int                     dead_level;
2251
2252         trace_xfs_da_swap_lastblock(args);
2253
2254         dead_buf = *dead_bufp;
2255         dead_blkno = *dead_blknop;
2256         tp = args->trans;
2257         dp = args->dp;
2258         w = args->whichfork;
2259         ASSERT(w == XFS_DATA_FORK);
2260         mp = dp->i_mount;
2261         lastoff = args->geo->freeblk;
2262         error = xfs_bmap_last_before(tp, dp, &lastoff, w);
2263         if (error)
2264                 return error;
2265         if (XFS_IS_CORRUPT(mp, lastoff == 0))
2266                 return -EFSCORRUPTED;
2267         /*
2268          * Read the last block in the btree space.
2269          */
2270         last_blkno = (xfs_dablk_t)lastoff - args->geo->fsbcount;
2271         error = xfs_da3_node_read(tp, dp, last_blkno, -1, &last_buf, w);
2272         if (error)
2273                 return error;
2274         /*
2275          * Copy the last block into the dead buffer and log it.
2276          */
2277         memcpy(dead_buf->b_addr, last_buf->b_addr, args->geo->blksize);
2278         xfs_trans_log_buf(tp, dead_buf, 0, args->geo->blksize - 1);
2279         dead_info = dead_buf->b_addr;
2280         /*
2281          * Get values from the moved block.
2282          */
2283         if (dead_info->magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
2284             dead_info->magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
2285                 struct xfs_dir3_icleaf_hdr leafhdr;
2286                 struct xfs_dir2_leaf_entry *ents;
2287
2288                 dead_leaf2 = (xfs_dir2_leaf_t *)dead_info;
2289                 xfs_dir2_leaf_hdr_from_disk(dp->i_mount, &leafhdr,
2290                                             dead_leaf2);
2291                 ents = leafhdr.ents;
2292                 dead_level = 0;
2293                 dead_hash = be32_to_cpu(ents[leafhdr.count - 1].hashval);
2294         } else {
2295                 struct xfs_da3_icnode_hdr deadhdr;
2296
2297                 dead_node = (xfs_da_intnode_t *)dead_info;
2298                 xfs_da3_node_hdr_from_disk(dp->i_mount, &deadhdr, dead_node);
2299                 btree = deadhdr.btree;
2300                 dead_level = deadhdr.level;
2301                 dead_hash = be32_to_cpu(btree[deadhdr.count - 1].hashval);
2302         }
2303         sib_buf = par_buf = NULL;
2304         /*
2305          * If the moved block has a left sibling, fix up the pointers.
2306          */
2307         if ((sib_blkno = be32_to_cpu(dead_info->back))) {
2308                 error = xfs_da3_node_read(tp, dp, sib_blkno, -1, &sib_buf, w);
2309                 if (error)
2310                         goto done;
2311                 sib_info = sib_buf->b_addr;
2312                 if (XFS_IS_CORRUPT(mp,
2313                                    be32_to_cpu(sib_info->forw) != last_blkno ||
2314                                    sib_info->magic != dead_info->magic)) {
2315                         error = -EFSCORRUPTED;
2316                         goto done;
2317                 }
2318                 sib_info->forw = cpu_to_be32(dead_blkno);
2319                 xfs_trans_log_buf(tp, sib_buf,
2320                         XFS_DA_LOGRANGE(sib_info, &sib_info->forw,
2321                                         sizeof(sib_info->forw)));
2322                 sib_buf = NULL;
2323         }
2324         /*
2325          * If the moved block has a right sibling, fix up the pointers.
2326          */
2327         if ((sib_blkno = be32_to_cpu(dead_info->forw))) {
2328                 error = xfs_da3_node_read(tp, dp, sib_blkno, -1, &sib_buf, w);
2329                 if (error)
2330                         goto done;
2331                 sib_info = sib_buf->b_addr;
2332                 if (XFS_IS_CORRUPT(mp,
2333                                    be32_to_cpu(sib_info->back) != last_blkno ||
2334                                    sib_info->magic != dead_info->magic)) {
2335                         error = -EFSCORRUPTED;
2336                         goto done;
2337                 }
2338                 sib_info->back = cpu_to_be32(dead_blkno);
2339                 xfs_trans_log_buf(tp, sib_buf,
2340                         XFS_DA_LOGRANGE(sib_info, &sib_info->back,
2341                                         sizeof(sib_info->back)));
2342                 sib_buf = NULL;
2343         }
2344         par_blkno = args->geo->leafblk;
2345         level = -1;
2346         /*
2347          * Walk down the tree looking for the parent of the moved block.
2348          */
2349         for (;;) {
2350                 error = xfs_da3_node_read(tp, dp, par_blkno, -1, &par_buf, w);
2351                 if (error)
2352                         goto done;
2353                 par_node = par_buf->b_addr;
2354                 xfs_da3_node_hdr_from_disk(dp->i_mount, &par_hdr, par_node);
2355                 if (XFS_IS_CORRUPT(mp,
2356                                    level >= 0 && level != par_hdr.level + 1)) {
2357                         error = -EFSCORRUPTED;
2358                         goto done;
2359                 }
2360                 level = par_hdr.level;
2361                 btree = par_hdr.btree;
2362                 for (entno = 0;
2363                      entno < par_hdr.count &&
2364                      be32_to_cpu(btree[entno].hashval) < dead_hash;
2365                      entno++)
2366                         continue;
2367                 if (XFS_IS_CORRUPT(mp, entno == par_hdr.count)) {
2368                         error = -EFSCORRUPTED;
2369                         goto done;
2370                 }
2371                 par_blkno = be32_to_cpu(btree[entno].before);
2372                 if (level == dead_level + 1)
2373                         break;
2374                 xfs_trans_brelse(tp, par_buf);
2375                 par_buf = NULL;
2376         }
2377         /*
2378          * We're in the right parent block.
2379          * Look for the right entry.
2380          */
2381         for (;;) {
2382                 for (;
2383                      entno < par_hdr.count &&
2384                      be32_to_cpu(btree[entno].before) != last_blkno;
2385                      entno++)
2386                         continue;
2387                 if (entno < par_hdr.count)
2388                         break;
2389                 par_blkno = par_hdr.forw;
2390                 xfs_trans_brelse(tp, par_buf);
2391                 par_buf = NULL;
2392                 if (XFS_IS_CORRUPT(mp, par_blkno == 0)) {
2393                         error = -EFSCORRUPTED;
2394                         goto done;
2395                 }
2396                 error = xfs_da3_node_read(tp, dp, par_blkno, -1, &par_buf, w);
2397                 if (error)
2398                         goto done;
2399                 par_node = par_buf->b_addr;
2400                 xfs_da3_node_hdr_from_disk(dp->i_mount, &par_hdr, par_node);
2401                 if (XFS_IS_CORRUPT(mp, par_hdr.level != level)) {
2402                         error = -EFSCORRUPTED;
2403                         goto done;
2404                 }
2405                 btree = par_hdr.btree;
2406                 entno = 0;
2407         }
2408         /*
2409          * Update the parent entry pointing to the moved block.
2410          */
2411         btree[entno].before = cpu_to_be32(dead_blkno);
2412         xfs_trans_log_buf(tp, par_buf,
2413                 XFS_DA_LOGRANGE(par_node, &btree[entno].before,
2414                                 sizeof(btree[entno].before)));
2415         *dead_blknop = last_blkno;
2416         *dead_bufp = last_buf;
2417         return 0;
2418 done:
2419         if (par_buf)
2420                 xfs_trans_brelse(tp, par_buf);
2421         if (sib_buf)
2422                 xfs_trans_brelse(tp, sib_buf);
2423         xfs_trans_brelse(tp, last_buf);
2424         return error;
2425 }
2426
2427 /*
2428  * Remove a btree block from a directory or attribute.
2429  */
2430 int
2431 xfs_da_shrink_inode(
2432         struct xfs_da_args      *args,
2433         xfs_dablk_t             dead_blkno,
2434         struct xfs_buf          *dead_buf)
2435 {
2436         struct xfs_inode        *dp;
2437         int                     done, error, w, count;
2438         struct xfs_trans        *tp;
2439
2440         trace_xfs_da_shrink_inode(args);
2441
2442         dp = args->dp;
2443         w = args->whichfork;
2444         tp = args->trans;
2445         count = args->geo->fsbcount;
2446         for (;;) {
2447                 /*
2448                  * Remove extents.  If we get ENOSPC for a dir we have to move
2449                  * the last block to the place we want to kill.
2450                  */
2451                 error = xfs_bunmapi(tp, dp, dead_blkno, count,
2452                                     xfs_bmapi_aflag(w), 0, &done);
2453                 if (error == -ENOSPC) {
2454                         if (w != XFS_DATA_FORK)
2455                                 break;
2456                         error = xfs_da3_swap_lastblock(args, &dead_blkno,
2457                                                       &dead_buf);
2458                         if (error)
2459                                 break;
2460                 } else {
2461                         break;
2462                 }
2463         }
2464         xfs_trans_binval(tp, dead_buf);
2465         return error;
2466 }
2467
2468 /*
2469  * See if the mapping(s) for this btree block are valid, i.e.
2470  * don't contain holes, are logically contiguous, and cover the whole range.
2471  */
2472 STATIC int
2473 xfs_da_map_covers_blocks(
2474         int             nmap,
2475         xfs_bmbt_irec_t *mapp,
2476         xfs_dablk_t     bno,
2477         int             count)
2478 {
2479         int             i;
2480         xfs_fileoff_t   off;
2481
2482         for (i = 0, off = bno; i < nmap; i++) {
2483                 if (mapp[i].br_startblock == HOLESTARTBLOCK ||
2484                     mapp[i].br_startblock == DELAYSTARTBLOCK) {
2485                         return 0;
2486                 }
2487                 if (off != mapp[i].br_startoff) {
2488                         return 0;
2489                 }
2490                 off += mapp[i].br_blockcount;
2491         }
2492         return off == bno + count;
2493 }
2494
2495 /*
2496  * Convert a struct xfs_bmbt_irec to a struct xfs_buf_map.
2497  *
2498  * For the single map case, it is assumed that the caller has provided a pointer
2499  * to a valid xfs_buf_map.  For the multiple map case, this function will
2500  * allocate the xfs_buf_map to hold all the maps and replace the caller's single
2501  * map pointer with the allocated map.
2502  */
2503 static int
2504 xfs_buf_map_from_irec(
2505         struct xfs_mount        *mp,
2506         struct xfs_buf_map      **mapp,
2507         int                     *nmaps,
2508         struct xfs_bmbt_irec    *irecs,
2509         int                     nirecs)
2510 {
2511         struct xfs_buf_map      *map;
2512         int                     i;
2513
2514         ASSERT(*nmaps == 1);
2515         ASSERT(nirecs >= 1);
2516
2517         if (nirecs > 1) {
2518                 map = kmem_zalloc(nirecs * sizeof(struct xfs_buf_map),
2519                                   KM_NOFS);
2520                 if (!map)
2521                         return -ENOMEM;
2522                 *mapp = map;
2523         }
2524
2525         *nmaps = nirecs;
2526         map = *mapp;
2527         for (i = 0; i < *nmaps; i++) {
2528                 ASSERT(irecs[i].br_startblock != DELAYSTARTBLOCK &&
2529                        irecs[i].br_startblock != HOLESTARTBLOCK);
2530                 map[i].bm_bn = XFS_FSB_TO_DADDR(mp, irecs[i].br_startblock);
2531                 map[i].bm_len = XFS_FSB_TO_BB(mp, irecs[i].br_blockcount);
2532         }
2533         return 0;
2534 }
2535
2536 /*
2537  * Map the block we are given ready for reading. There are three possible return
2538  * values:
2539  *      -1 - will be returned if we land in a hole and mappedbno == -2 so the
2540  *           caller knows not to execute a subsequent read.
2541  *       0 - if we mapped the block successfully
2542  *      >0 - positive error number if there was an error.
2543  */
2544 static int
2545 xfs_dabuf_map(
2546         struct xfs_inode        *dp,
2547         xfs_dablk_t             bno,
2548         xfs_daddr_t             mappedbno,
2549         int                     whichfork,
2550         struct xfs_buf_map      **map,
2551         int                     *nmaps)
2552 {
2553         struct xfs_mount        *mp = dp->i_mount;
2554         int                     nfsb;
2555         int                     error = 0;
2556         struct xfs_bmbt_irec    irec;
2557         struct xfs_bmbt_irec    *irecs = &irec;
2558         int                     nirecs;
2559
2560         ASSERT(map && *map);
2561         ASSERT(*nmaps == 1);
2562
2563         if (whichfork == XFS_DATA_FORK)
2564                 nfsb = mp->m_dir_geo->fsbcount;
2565         else
2566                 nfsb = mp->m_attr_geo->fsbcount;
2567
2568         /*
2569          * Caller doesn't have a mapping.  -2 means don't complain
2570          * if we land in a hole.
2571          */
2572         if (mappedbno == -1 || mappedbno == -2) {
2573                 /*
2574                  * Optimize the one-block case.
2575                  */
2576                 if (nfsb != 1)
2577                         irecs = kmem_zalloc(sizeof(irec) * nfsb,
2578                                             KM_NOFS);
2579
2580                 nirecs = nfsb;
2581                 error = xfs_bmapi_read(dp, (xfs_fileoff_t)bno, nfsb, irecs,
2582                                        &nirecs, xfs_bmapi_aflag(whichfork));
2583                 if (error)
2584                         goto out;
2585         } else {
2586                 irecs->br_startblock = XFS_DADDR_TO_FSB(mp, mappedbno);
2587                 irecs->br_startoff = (xfs_fileoff_t)bno;
2588                 irecs->br_blockcount = nfsb;
2589                 irecs->br_state = 0;
2590                 nirecs = 1;
2591         }
2592
2593         if (!xfs_da_map_covers_blocks(nirecs, irecs, bno, nfsb)) {
2594                 /* Caller ok with no mapping. */
2595                 if (!XFS_IS_CORRUPT(mp, mappedbno != -2)) {
2596                         error = -1;
2597                         goto out;
2598                 }
2599
2600                 /* Caller expected a mapping, so abort. */
2601                 if (xfs_error_level >= XFS_ERRLEVEL_LOW) {
2602                         int i;
2603
2604                         xfs_alert(mp, "%s: bno %lld dir: inode %lld", __func__,
2605                                         (long long)bno, (long long)dp->i_ino);
2606                         for (i = 0; i < *nmaps; i++) {
2607                                 xfs_alert(mp,
2608 "[%02d] br_startoff %lld br_startblock %lld br_blockcount %lld br_state %d",
2609                                         i,
2610                                         (long long)irecs[i].br_startoff,
2611                                         (long long)irecs[i].br_startblock,
2612                                         (long long)irecs[i].br_blockcount,
2613                                         irecs[i].br_state);
2614                         }
2615                 }
2616                 error = -EFSCORRUPTED;
2617                 goto out;
2618         }
2619         error = xfs_buf_map_from_irec(mp, map, nmaps, irecs, nirecs);
2620 out:
2621         if (irecs != &irec)
2622                 kmem_free(irecs);
2623         return error;
2624 }
2625
2626 /*
2627  * Get a buffer for the dir/attr block.
2628  */
2629 int
2630 xfs_da_get_buf(
2631         struct xfs_trans        *trans,
2632         struct xfs_inode        *dp,
2633         xfs_dablk_t             bno,
2634         xfs_daddr_t             mappedbno,
2635         struct xfs_buf          **bpp,
2636         int                     whichfork)
2637 {
2638         struct xfs_buf          *bp;
2639         struct xfs_buf_map      map;
2640         struct xfs_buf_map      *mapp;
2641         int                     nmap;
2642         int                     error;
2643
2644         *bpp = NULL;
2645         mapp = &map;
2646         nmap = 1;
2647         error = xfs_dabuf_map(dp, bno, mappedbno, whichfork,
2648                                 &mapp, &nmap);
2649         if (error) {
2650                 /* mapping a hole is not an error, but we don't continue */
2651                 if (error == -1)
2652                         error = 0;
2653                 goto out_free;
2654         }
2655
2656         bp = xfs_trans_get_buf_map(trans, dp->i_mount->m_ddev_targp,
2657                                     mapp, nmap, 0);
2658         error = bp ? bp->b_error : -EIO;
2659         if (error) {
2660                 if (bp)
2661                         xfs_trans_brelse(trans, bp);
2662                 goto out_free;
2663         }
2664
2665         *bpp = bp;
2666
2667 out_free:
2668         if (mapp != &map)
2669                 kmem_free(mapp);
2670
2671         return error;
2672 }
2673
2674 /*
2675  * Get a buffer for the dir/attr block, fill in the contents.
2676  */
2677 int
2678 xfs_da_read_buf(
2679         struct xfs_trans        *trans,
2680         struct xfs_inode        *dp,
2681         xfs_dablk_t             bno,
2682         xfs_daddr_t             mappedbno,
2683         struct xfs_buf          **bpp,
2684         int                     whichfork,
2685         const struct xfs_buf_ops *ops)
2686 {
2687         struct xfs_buf          *bp;
2688         struct xfs_buf_map      map;
2689         struct xfs_buf_map      *mapp;
2690         int                     nmap;
2691         int                     error;
2692
2693         *bpp = NULL;
2694         mapp = &map;
2695         nmap = 1;
2696         error = xfs_dabuf_map(dp, bno, mappedbno, whichfork,
2697                                 &mapp, &nmap);
2698         if (error) {
2699                 /* mapping a hole is not an error, but we don't continue */
2700                 if (error == -1)
2701                         error = 0;
2702                 goto out_free;
2703         }
2704
2705         error = xfs_trans_read_buf_map(dp->i_mount, trans,
2706                                         dp->i_mount->m_ddev_targp,
2707                                         mapp, nmap, 0, &bp, ops);
2708         if (error)
2709                 goto out_free;
2710
2711         if (whichfork == XFS_ATTR_FORK)
2712                 xfs_buf_set_ref(bp, XFS_ATTR_BTREE_REF);
2713         else
2714                 xfs_buf_set_ref(bp, XFS_DIR_BTREE_REF);
2715         *bpp = bp;
2716 out_free:
2717         if (mapp != &map)
2718                 kmem_free(mapp);
2719
2720         return error;
2721 }
2722
2723 /*
2724  * Readahead the dir/attr block.
2725  */
2726 int
2727 xfs_da_reada_buf(
2728         struct xfs_inode        *dp,
2729         xfs_dablk_t             bno,
2730         xfs_daddr_t             mappedbno,
2731         int                     whichfork,
2732         const struct xfs_buf_ops *ops)
2733 {
2734         struct xfs_buf_map      map;
2735         struct xfs_buf_map      *mapp;
2736         int                     nmap;
2737         int                     error;
2738
2739         mapp = &map;
2740         nmap = 1;
2741         error = xfs_dabuf_map(dp, bno, mappedbno, whichfork,
2742                                 &mapp, &nmap);
2743         if (error) {
2744                 /* mapping a hole is not an error, but we don't continue */
2745                 if (error == -1)
2746                         error = 0;
2747                 goto out_free;
2748         }
2749
2750         mappedbno = mapp[0].bm_bn;
2751         xfs_buf_readahead_map(dp->i_mount->m_ddev_targp, mapp, nmap, ops);
2752
2753 out_free:
2754         if (mapp != &map)
2755                 kmem_free(mapp);
2756
2757         return error;
2758 }