fs/xfs/xfs_da_btree.c

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