fs/xfs/xfs_da_btree.c

   1 /*
   2  * Copyright (c) 2000-2004 Silicon Graphics, Inc.  All Rights Reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms of version 2 of the GNU General Public License as
   6  * published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it would be useful, but
   9  * WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  11  *
  12  * Further, this software is distributed without any warranty that it is
  13  * free of the rightful claim of any third person regarding infringement
  14  * or the like.  Any license provided herein, whether implied or
  15  * otherwise, applies only to this software file.  Patent licenses, if
  16  * any, provided herein do not apply to combinations of this program with
  17  * other software, or any other product whatsoever.
  18  *
  19  * You should have received a copy of the GNU General Public License along
  20  * with this program; if not, write the Free Software Foundation, Inc., 59
  21  * Temple Place - Suite 330, Boston MA 02111-1307, USA.
  22  *
  23  * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy,
  24  * Mountain View, CA  94043, or:
  25  *
  26  * http://www.sgi.com
  27  *
  28  * For further information regarding this notice, see:
  29  *
  30  * http://oss.sgi.com/projects/GenInfo/SGIGPLNoticeExplan/
  31  */
  32
  33 #include "xfs.h"
  34
  35 #include "xfs_macros.h"
  36 #include "xfs_types.h"
  37 #include "xfs_inum.h"
  38 #include "xfs_log.h"
  39 #include "xfs_trans.h"
  40 #include "xfs_sb.h"
  41 #include "xfs_ag.h"
  42 #include "xfs_dir.h"
  43 #include "xfs_dir2.h"
  44 #include "xfs_dmapi.h"
  45 #include "xfs_mount.h"
  46 #include "xfs_alloc_btree.h"
  47 #include "xfs_bmap_btree.h"
  48 #include "xfs_ialloc_btree.h"
  49 #include "xfs_alloc.h"
  50 #include "xfs_btree.h"
  51 #include "xfs_attr_sf.h"
  52 #include "xfs_dir_sf.h"
  53 #include "xfs_dir2_sf.h"
  54 #include "xfs_dinode.h"
  55 #include "xfs_inode_item.h"
  56 #include "xfs_inode.h"
  57 #include "xfs_bmap.h"
  58 #include "xfs_da_btree.h"
  59 #include "xfs_attr.h"
  60 #include "xfs_attr_leaf.h"
  61 #include "xfs_dir_leaf.h"
  62 #include "xfs_dir2_data.h"
  63 #include "xfs_dir2_leaf.h"
  64 #include "xfs_dir2_block.h"
  65 #include "xfs_dir2_node.h"
  66 #include "xfs_error.h"
  67 #include "xfs_bit.h"
  68
  69 /*
  70  * xfs_da_btree.c
  71  *
  72  * Routines to implement directories as Btrees of hashed names.
  73  */
  74
  75 /*========================================================================
  76  * Function prototypes for the kernel.
  77  *========================================================================*/
  78
  79 /*
  80  * Routines used for growing the Btree.
  81  */
  82 STATIC int xfs_da_root_split(xfs_da_state_t *state,
  83                                             xfs_da_state_blk_t *existing_root,
  84                                             xfs_da_state_blk_t *new_child);
  85 STATIC int xfs_da_node_split(xfs_da_state_t *state,
  86                                             xfs_da_state_blk_t *existing_blk,
  87                                             xfs_da_state_blk_t *split_blk,
  88                                             xfs_da_state_blk_t *blk_to_add,
  89                                             int treelevel,
  90                                             int *result);
  91 STATIC void xfs_da_node_rebalance(xfs_da_state_t *state,
  92                                          xfs_da_state_blk_t *node_blk_1,
  93                                          xfs_da_state_blk_t *node_blk_2);
  94 STATIC void xfs_da_node_add(xfs_da_state_t *state,
  95                                    xfs_da_state_blk_t *old_node_blk,
  96                                    xfs_da_state_blk_t *new_node_blk);
  97
  98 /*
  99  * Routines used for shrinking the Btree.
 100  */
 101 STATIC int xfs_da_root_join(xfs_da_state_t *state,
 102                                            xfs_da_state_blk_t *root_blk);
 103 STATIC int xfs_da_node_toosmall(xfs_da_state_t *state, int *retval);
 104 STATIC void xfs_da_node_remove(xfs_da_state_t *state,
 105                                               xfs_da_state_blk_t *drop_blk);
 106 STATIC void xfs_da_node_unbalance(xfs_da_state_t *state,
 107                                          xfs_da_state_blk_t *src_node_blk,
 108                                          xfs_da_state_blk_t *dst_node_blk);
 109
 110 /*
 111  * Utility routines.
 112  */
 113 STATIC uint     xfs_da_node_lasthash(xfs_dabuf_t *bp, int *count);
 114 STATIC int      xfs_da_node_order(xfs_dabuf_t *node1_bp, xfs_dabuf_t *node2_bp);
 115 STATIC xfs_dabuf_t *xfs_da_buf_make(int nbuf, xfs_buf_t **bps, inst_t *ra);
 116 STATIC int      xfs_da_blk_unlink(xfs_da_state_t *state,
 117                                   xfs_da_state_blk_t *drop_blk,
 118                                   xfs_da_state_blk_t *save_blk);
 119 STATIC void     xfs_da_state_kill_altpath(xfs_da_state_t *state);
 120
 121 /*========================================================================
 122  * Routines used for growing the Btree.
 123  *========================================================================*/
 124
 125 /*
 126  * Create the initial contents of an intermediate node.
 127  */
 128 int
 129 xfs_da_node_create(xfs_da_args_t *args, xfs_dablk_t blkno, int level,
 130                                  xfs_dabuf_t **bpp, int whichfork)
 131 {
 132         xfs_da_intnode_t *node;
 133         xfs_dabuf_t *bp;
 134         int error;
 135         xfs_trans_t *tp;
 136
 137         tp = args->trans;
 138         error = xfs_da_get_buf(tp, args->dp, blkno, -1, &bp, whichfork);
 139         if (error)
 140                 return(error);
 141         ASSERT(bp != NULL);
 142         node = bp->data;
 143         node->hdr.info.forw = 0;
 144         node->hdr.info.back = 0;
 145         INT_SET(node->hdr.info.magic, ARCH_CONVERT, XFS_DA_NODE_MAGIC);
 146         node->hdr.info.pad = 0;
 147         node->hdr.count = 0;
 148         INT_SET(node->hdr.level, ARCH_CONVERT, level);
 149
 150         xfs_da_log_buf(tp, bp,
 151                 XFS_DA_LOGRANGE(node, &node->hdr, sizeof(node->hdr)));
 152
 153         *bpp = bp;
 154         return(0);
 155 }
 156
 157 /*
 158  * Split a leaf node, rebalance, then possibly split
 159  * intermediate nodes, rebalance, etc.
 160  */
 161 int                                                     /* error */
 162 xfs_da_split(xfs_da_state_t *state)
 163 {
 164         xfs_da_state_blk_t *oldblk, *newblk, *addblk;
 165         xfs_da_intnode_t *node;
 166         xfs_dabuf_t *bp;
 167         int max, action, error, i;
 168
 169         /*
 170          * Walk back up the tree splitting/inserting/adjusting as necessary.
 171          * If we need to insert and there isn't room, split the node, then
 172          * decide which fragment to insert the new block from below into.
 173          * Note that we may split the root this way, but we need more fixup.
 174          */
 175         max = state->path.active - 1;
 176         ASSERT((max >= 0) && (max < XFS_DA_NODE_MAXDEPTH));
 177         ASSERT(state->path.blk[max].magic == XFS_ATTR_LEAF_MAGIC ||
 178                state->path.blk[max].magic == XFS_DIRX_LEAF_MAGIC(state->mp));
 179
 180         addblk = &state->path.blk[max];         /* initial dummy value */
 181         for (i = max; (i >= 0) && addblk; state->path.active--, i--) {
 182                 oldblk = &state->path.blk[i];
 183                 newblk = &state->altpath.blk[i];
 184
 185                 /*
 186                  * If a leaf node then
 187                  *     Allocate a new leaf node, then rebalance across them.
 188                  * else if an intermediate node then
 189                  *     We split on the last layer, must we split the node?
 190                  */
 191                 switch (oldblk->magic) {
 192                 case XFS_ATTR_LEAF_MAGIC:
 193 #ifndef __KERNEL__
 194                         return(ENOTTY);
 195 #else
 196                         error = xfs_attr_leaf_split(state, oldblk, newblk);
 197                         if ((error != 0) && (error != ENOSPC)) {
 198                                 return(error);  /* GROT: attr is inconsistent */
 199                         }
 200                         if (!error) {
 201                                 addblk = newblk;
 202                                 break;
 203                         }
 204                         /*
 205                          * Entry wouldn't fit, split the leaf again.
 206                          */
 207                         state->extravalid = 1;
 208                         if (state->inleaf) {
 209                                 state->extraafter = 0;  /* before newblk */
 210                                 error = xfs_attr_leaf_split(state, oldblk,
 211                                                             &state->extrablk);
 212                         } else {
 213                                 state->extraafter = 1;  /* after newblk */
 214                                 error = xfs_attr_leaf_split(state, newblk,
 215                                                             &state->extrablk);
 216                         }
 217                         if (error)
 218                                 return(error);  /* GROT: attr inconsistent */
 219                         addblk = newblk;
 220                         break;
 221 #endif
 222                 case XFS_DIR_LEAF_MAGIC:
 223                         ASSERT(XFS_DIR_IS_V1(state->mp));
 224                         error = xfs_dir_leaf_split(state, oldblk, newblk);
 225                         if ((error != 0) && (error != ENOSPC)) {
 226                                 return(error);  /* GROT: dir is inconsistent */
 227                         }
 228                         if (!error) {
 229                                 addblk = newblk;
 230                                 break;
 231                         }
 232                         /*
 233                          * Entry wouldn't fit, split the leaf again.
 234                          */
 235                         state->extravalid = 1;
 236                         if (state->inleaf) {
 237                                 state->extraafter = 0;  /* before newblk */
 238                                 error = xfs_dir_leaf_split(state, oldblk,
 239                                                            &state->extrablk);
 240                                 if (error)
 241                                         return(error);  /* GROT: dir incon. */
 242                                 addblk = newblk;
 243                         } else {
 244                                 state->extraafter = 1;  /* after newblk */
 245                                 error = xfs_dir_leaf_split(state, newblk,
 246                                                            &state->extrablk);
 247                                 if (error)
 248                                         return(error);  /* GROT: dir incon. */
 249                                 addblk = newblk;
 250                         }
 251                         break;
 252                 case XFS_DIR2_LEAFN_MAGIC:
 253                         ASSERT(XFS_DIR_IS_V2(state->mp));
 254                         error = xfs_dir2_leafn_split(state, oldblk, newblk);
 255                         if (error)
 256                                 return error;
 257                         addblk = newblk;
 258                         break;
 259                 case XFS_DA_NODE_MAGIC:
 260                         error = xfs_da_node_split(state, oldblk, newblk, addblk,
 261                                                          max - i, &action);
 262                         xfs_da_buf_done(addblk->bp);
 263                         addblk->bp = NULL;
 264                         if (error)
 265                                 return(error);  /* GROT: dir is inconsistent */
 266                         /*
 267                          * Record the newly split block for the next time thru?
 268                          */
 269                         if (action)
 270                                 addblk = newblk;
 271                         else
 272                                 addblk = NULL;
 273                         break;
 274                 }
 275
 276                 /*
 277                  * Update the btree to show the new hashval for this child.
 278                  */
 279                 xfs_da_fixhashpath(state, &state->path);
 280                 /*
 281                  * If we won't need this block again, it's getting dropped
 282                  * from the active path by the loop control, so we need
 283                  * to mark it done now.
 284                  */
 285                 if (i > 0 || !addblk)
 286                         xfs_da_buf_done(oldblk->bp);
 287         }
 288         if (!addblk)
 289                 return(0);
 290
 291         /*
 292          * Split the root node.
 293          */
 294         ASSERT(state->path.active == 0);
 295         oldblk = &state->path.blk[0];
 296         error = xfs_da_root_split(state, oldblk, addblk);
 297         if (error) {
 298                 xfs_da_buf_done(oldblk->bp);
 299                 xfs_da_buf_done(addblk->bp);
 300                 addblk->bp = NULL;
 301                 return(error);  /* GROT: dir is inconsistent */
 302         }
 303
 304         /*
 305          * Update pointers to the node which used to be block 0 and
 306          * just got bumped because of the addition of a new root node.
 307          * There might be three blocks involved if a double split occurred,
 308          * and the original block 0 could be at any position in the list.
 309          */
 310
 311         node = oldblk->bp->data;
 312         if (node->hdr.info.forw) {
 313                 if (INT_GET(node->hdr.info.forw, ARCH_CONVERT) == addblk->blkno) {
 314                         bp = addblk->bp;
 315                 } else {
 316                         ASSERT(state->extravalid);
 317                         bp = state->extrablk.bp;
 318                 }
 319                 node = bp->data;
 320                 INT_SET(node->hdr.info.back, ARCH_CONVERT, oldblk->blkno);
 321                 xfs_da_log_buf(state->args->trans, bp,
 322                     XFS_DA_LOGRANGE(node, &node->hdr.info,
 323                     sizeof(node->hdr.info)));
 324         }
 325         node = oldblk->bp->data;
 326         if (INT_GET(node->hdr.info.back, ARCH_CONVERT)) {
 327                 if (INT_GET(node->hdr.info.back, ARCH_CONVERT) == addblk->blkno) {
 328                         bp = addblk->bp;
 329                 } else {
 330                         ASSERT(state->extravalid);
 331                         bp = state->extrablk.bp;
 332                 }
 333                 node = bp->data;
 334                 INT_SET(node->hdr.info.forw, ARCH_CONVERT, oldblk->blkno);
 335                 xfs_da_log_buf(state->args->trans, bp,
 336                     XFS_DA_LOGRANGE(node, &node->hdr.info,
 337                     sizeof(node->hdr.info)));
 338         }
 339         xfs_da_buf_done(oldblk->bp);
 340         xfs_da_buf_done(addblk->bp);
 341         addblk->bp = NULL;
 342         return(0);
 343 }
 344
 345 /*
 346  * Split the root.  We have to create a new root and point to the two
 347  * parts (the split old root) that we just created.  Copy block zero to
 348  * the EOF, extending the inode in process.
 349  */
 350 STATIC int                                              /* error */
 351 xfs_da_root_split(xfs_da_state_t *state, xfs_da_state_blk_t *blk1,
 352                                  xfs_da_state_blk_t *blk2)
 353 {
 354         xfs_da_intnode_t *node, *oldroot;
 355         xfs_da_args_t *args;
 356         xfs_dablk_t blkno;
 357         xfs_dabuf_t *bp;
 358         int error, size;
 359         xfs_inode_t *dp;
 360         xfs_trans_t *tp;
 361         xfs_mount_t *mp;
 362         xfs_dir2_leaf_t *leaf;
 363
 364         /*
 365          * Copy the existing (incorrect) block from the root node position
 366          * to a free space somewhere.
 367          */
 368         args = state->args;
 369         ASSERT(args != NULL);
 370         error = xfs_da_grow_inode(args, &blkno);
 371         if (error)
 372                 return(error);
 373         dp = args->dp;
 374         tp = args->trans;
 375         mp = state->mp;
 376         error = xfs_da_get_buf(tp, dp, blkno, -1, &bp, args->whichfork);
 377         if (error)
 378                 return(error);
 379         ASSERT(bp != NULL);
 380         node = bp->data;
 381         oldroot = blk1->bp->data;
 382         if (INT_GET(oldroot->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC) {
 383                 size = (int)((char *)&oldroot->btree[INT_GET(oldroot->hdr.count, ARCH_CONVERT)] -
 384                              (char *)oldroot);
 385         } else {
 386                 ASSERT(XFS_DIR_IS_V2(mp));
 387                 ASSERT(INT_GET(oldroot->hdr.info.magic, ARCH_CONVERT) == XFS_DIR2_LEAFN_MAGIC);
 388                 leaf = (xfs_dir2_leaf_t *)oldroot;
 389                 size = (int)((char *)&leaf->ents[INT_GET(leaf->hdr.count, ARCH_CONVERT)] -
 390                              (char *)leaf);
 391         }
 392         memcpy(node, oldroot, size);
 393         xfs_da_log_buf(tp, bp, 0, size - 1);
 394         xfs_da_buf_done(blk1->bp);
 395         blk1->bp = bp;
 396         blk1->blkno = blkno;
 397
 398         /*
 399          * Set up the new root node.
 400          */
 401         error = xfs_da_node_create(args,
 402                 args->whichfork == XFS_DATA_FORK &&
 403                 XFS_DIR_IS_V2(mp) ? mp->m_dirleafblk : 0,
 404                 INT_GET(node->hdr.level, ARCH_CONVERT) + 1, &bp, args->whichfork);
 405         if (error)
 406                 return(error);
 407         node = bp->data;
 408         INT_SET(node->btree[0].hashval, ARCH_CONVERT, blk1->hashval);
 409         INT_SET(node->btree[0].before, ARCH_CONVERT, blk1->blkno);
 410         INT_SET(node->btree[1].hashval, ARCH_CONVERT, blk2->hashval);
 411         INT_SET(node->btree[1].before, ARCH_CONVERT, blk2->blkno);
 412         INT_SET(node->hdr.count, ARCH_CONVERT, 2);
 413
 414 #ifdef DEBUG
 415         if (INT_GET(oldroot->hdr.info.magic, ARCH_CONVERT) == XFS_DIR2_LEAFN_MAGIC) {
 416                 ASSERT(blk1->blkno >= mp->m_dirleafblk &&
 417                        blk1->blkno < mp->m_dirfreeblk);
 418                 ASSERT(blk2->blkno >= mp->m_dirleafblk &&
 419                        blk2->blkno < mp->m_dirfreeblk);
 420         }
 421 #endif
 422
 423         /* Header is already logged by xfs_da_node_create */
 424         xfs_da_log_buf(tp, bp,
 425                 XFS_DA_LOGRANGE(node, node->btree,
 426                         sizeof(xfs_da_node_entry_t) * 2));
 427         xfs_da_buf_done(bp);
 428
 429         return(0);
 430 }
 431
 432 /*
 433  * Split the node, rebalance, then add the new entry.
 434  */
 435 STATIC int                                              /* error */
 436 xfs_da_node_split(xfs_da_state_t *state, xfs_da_state_blk_t *oldblk,
 437                                  xfs_da_state_blk_t *newblk,
 438                                  xfs_da_state_blk_t *addblk,
 439                                  int treelevel, int *result)
 440 {
 441         xfs_da_intnode_t *node;
 442         xfs_dablk_t blkno;
 443         int newcount, error;
 444         int useextra;
 445
 446         node = oldblk->bp->data;
 447         ASSERT(INT_GET(node->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
 448
 449         /*
 450          * With V2 the extra block is data or freespace.
 451          */
 452         useextra = state->extravalid && XFS_DIR_IS_V1(state->mp);
 453         newcount = 1 + useextra;
 454         /*
 455          * Do we have to split the node?
 456          */
 457         if ((INT_GET(node->hdr.count, ARCH_CONVERT) + newcount) > state->node_ents) {
 458                 /*
 459                  * Allocate a new node, add to the doubly linked chain of
 460                  * nodes, then move some of our excess entries into it.
 461                  */
 462                 error = xfs_da_grow_inode(state->args, &blkno);
 463                 if (error)
 464                         return(error);  /* GROT: dir is inconsistent */
 465
 466                 error = xfs_da_node_create(state->args, blkno, treelevel,
 467                                            &newblk->bp, state->args->whichfork);
 468                 if (error)
 469                         return(error);  /* GROT: dir is inconsistent */
 470                 newblk->blkno = blkno;
 471                 newblk->magic = XFS_DA_NODE_MAGIC;
 472                 xfs_da_node_rebalance(state, oldblk, newblk);
 473                 error = xfs_da_blk_link(state, oldblk, newblk);
 474                 if (error)
 475                         return(error);
 476                 *result = 1;
 477         } else {
 478                 *result = 0;
 479         }
 480
 481         /*
 482          * Insert the new entry(s) into the correct block
 483          * (updating last hashval in the process).
 484          *
 485          * xfs_da_node_add() inserts BEFORE the given index,
 486          * and as a result of using node_lookup_int() we always
 487          * point to a valid entry (not after one), but a split
 488          * operation always results in a new block whose hashvals
 489          * FOLLOW the current block.
 490          *
 491          * If we had double-split op below us, then add the extra block too.
 492          */
 493         node = oldblk->bp->data;
 494         if (oldblk->index <= INT_GET(node->hdr.count, ARCH_CONVERT)) {
 495                 oldblk->index++;
 496                 xfs_da_node_add(state, oldblk, addblk);
 497                 if (useextra) {
 498                         if (state->extraafter)
 499                                 oldblk->index++;
 500                         xfs_da_node_add(state, oldblk, &state->extrablk);
 501                         state->extravalid = 0;
 502                 }
 503         } else {
 504                 newblk->index++;
 505                 xfs_da_node_add(state, newblk, addblk);
 506                 if (useextra) {
 507                         if (state->extraafter)
 508                                 newblk->index++;
 509                         xfs_da_node_add(state, newblk, &state->extrablk);
 510                         state->extravalid = 0;
 511                 }
 512         }
 513
 514         return(0);
 515 }
 516
 517 /*
 518  * Balance the btree elements between two intermediate nodes,
 519  * usually one full and one empty.
 520  *
 521  * NOTE: if blk2 is empty, then it will get the upper half of blk1.
 522  */
 523 STATIC void
 524 xfs_da_node_rebalance(xfs_da_state_t *state, xfs_da_state_blk_t *blk1,
 525                                      xfs_da_state_blk_t *blk2)
 526 {
 527         xfs_da_intnode_t *node1, *node2, *tmpnode;
 528         xfs_da_node_entry_t *btree_s, *btree_d;
 529         int count, tmp;
 530         xfs_trans_t *tp;
 531
 532         node1 = blk1->bp->data;
 533         node2 = blk2->bp->data;
 534         /*
 535          * Figure out how many entries need to move, and in which direction.
 536          * Swap the nodes around if that makes it simpler.
 537          */
 538         if ((INT_GET(node1->hdr.count, ARCH_CONVERT) > 0) && (INT_GET(node2->hdr.count, ARCH_CONVERT) > 0) &&
 539             ((INT_GET(node2->btree[ 0 ].hashval, ARCH_CONVERT) < INT_GET(node1->btree[ 0 ].hashval, ARCH_CONVERT)) ||
 540              (INT_GET(node2->btree[ INT_GET(node2->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT) <
 541               INT_GET(node1->btree[ INT_GET(node1->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT)))) {
 542                 tmpnode = node1;
 543                 node1 = node2;
 544                 node2 = tmpnode;
 545         }
 546         ASSERT(INT_GET(node1->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
 547         ASSERT(INT_GET(node2->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
 548         count = (INT_GET(node1->hdr.count, ARCH_CONVERT) - INT_GET(node2->hdr.count, ARCH_CONVERT)) / 2;
 549         if (count == 0)
 550                 return;
 551         tp = state->args->trans;
 552         /*
 553          * Two cases: high-to-low and low-to-high.
 554          */
 555         if (count > 0) {
 556                 /*
 557                  * Move elements in node2 up to make a hole.
 558                  */
 559                 if ((tmp = INT_GET(node2->hdr.count, ARCH_CONVERT)) > 0) {
 560                         tmp *= (uint)sizeof(xfs_da_node_entry_t);
 561                         btree_s = &node2->btree[0];
 562                         btree_d = &node2->btree[count];
 563                         memmove(btree_d, btree_s, tmp);
 564                 }
 565
 566                 /*
 567                  * Move the req'd B-tree elements from high in node1 to
 568                  * low in node2.
 569                  */
 570                 INT_MOD(node2->hdr.count, ARCH_CONVERT, count);
 571                 tmp = count * (uint)sizeof(xfs_da_node_entry_t);
 572                 btree_s = &node1->btree[INT_GET(node1->hdr.count, ARCH_CONVERT) - count];
 573                 btree_d = &node2->btree[0];
 574                 memcpy(btree_d, btree_s, tmp);
 575                 INT_MOD(node1->hdr.count, ARCH_CONVERT, -(count));
 576
 577         } else {
 578                 /*
 579                  * Move the req'd B-tree elements from low in node2 to
 580                  * high in node1.
 581                  */
 582                 count = -count;
 583                 tmp = count * (uint)sizeof(xfs_da_node_entry_t);
 584                 btree_s = &node2->btree[0];
 585                 btree_d = &node1->btree[INT_GET(node1->hdr.count, ARCH_CONVERT)];
 586                 memcpy(btree_d, btree_s, tmp);
 587                 INT_MOD(node1->hdr.count, ARCH_CONVERT, count);
 588                 xfs_da_log_buf(tp, blk1->bp,
 589                         XFS_DA_LOGRANGE(node1, btree_d, tmp));
 590
 591                 /*
 592                  * Move elements in node2 down to fill the hole.
 593                  */
 594                 tmp  = INT_GET(node2->hdr.count, ARCH_CONVERT) - count;
 595                 tmp *= (uint)sizeof(xfs_da_node_entry_t);
 596                 btree_s = &node2->btree[count];
 597                 btree_d = &node2->btree[0];
 598                 memmove(btree_d, btree_s, tmp);
 599                 INT_MOD(node2->hdr.count, ARCH_CONVERT, -(count));
 600         }
 601
 602         /*
 603          * Log header of node 1 and all current bits of node 2.
 604          */
 605         xfs_da_log_buf(tp, blk1->bp,
 606                 XFS_DA_LOGRANGE(node1, &node1->hdr, sizeof(node1->hdr)));
 607         xfs_da_log_buf(tp, blk2->bp,
 608                 XFS_DA_LOGRANGE(node2, &node2->hdr,
 609                         sizeof(node2->hdr) +
 610                         sizeof(node2->btree[0]) * INT_GET(node2->hdr.count, ARCH_CONVERT)));
 611
 612         /*
 613          * Record the last hashval from each block for upward propagation.
 614          * (note: don't use the swapped node pointers)
 615          */
 616         node1 = blk1->bp->data;
 617         node2 = blk2->bp->data;
 618         blk1->hashval = INT_GET(node1->btree[ INT_GET(node1->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT);
 619         blk2->hashval = INT_GET(node2->btree[ INT_GET(node2->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT);
 620
 621         /*
 622          * Adjust the expected index for insertion.
 623          */
 624         if (blk1->index >= INT_GET(node1->hdr.count, ARCH_CONVERT)) {
 625                 blk2->index = blk1->index - INT_GET(node1->hdr.count, ARCH_CONVERT);
 626                 blk1->index = INT_GET(node1->hdr.count, ARCH_CONVERT) + 1;      /* make it invalid */
 627         }
 628 }
 629
 630 /*
 631  * Add a new entry to an intermediate node.
 632  */
 633 STATIC void
 634 xfs_da_node_add(xfs_da_state_t *state, xfs_da_state_blk_t *oldblk,
 635                                xfs_da_state_blk_t *newblk)
 636 {
 637         xfs_da_intnode_t *node;
 638         xfs_da_node_entry_t *btree;
 639         int tmp;
 640         xfs_mount_t *mp;
 641
 642         node = oldblk->bp->data;
 643         mp = state->mp;
 644         ASSERT(INT_GET(node->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
 645         ASSERT((oldblk->index >= 0) && (oldblk->index <= INT_GET(node->hdr.count, ARCH_CONVERT)));
 646         ASSERT(newblk->blkno != 0);
 647         if (state->args->whichfork == XFS_DATA_FORK && XFS_DIR_IS_V2(mp))
 648                 ASSERT(newblk->blkno >= mp->m_dirleafblk &&
 649                        newblk->blkno < mp->m_dirfreeblk);
 650
 651         /*
 652          * We may need to make some room before we insert the new node.
 653          */
 654         tmp = 0;
 655         btree = &node->btree[ oldblk->index ];
 656         if (oldblk->index < INT_GET(node->hdr.count, ARCH_CONVERT)) {
 657                 tmp = (INT_GET(node->hdr.count, ARCH_CONVERT) - oldblk->index) * (uint)sizeof(*btree);
 658                 memmove(btree + 1, btree, tmp);
 659         }
 660         INT_SET(btree->hashval, ARCH_CONVERT, newblk->hashval);
 661         INT_SET(btree->before, ARCH_CONVERT, newblk->blkno);
 662         xfs_da_log_buf(state->args->trans, oldblk->bp,
 663                 XFS_DA_LOGRANGE(node, btree, tmp + sizeof(*btree)));
 664         INT_MOD(node->hdr.count, ARCH_CONVERT, +1);
 665         xfs_da_log_buf(state->args->trans, oldblk->bp,
 666                 XFS_DA_LOGRANGE(node, &node->hdr, sizeof(node->hdr)));
 667
 668         /*
 669          * Copy the last hash value from the oldblk to propagate upwards.
 670          */
 671         oldblk->hashval = INT_GET(node->btree[ INT_GET(node->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT);
 672 }
 673
 674 /*========================================================================
 675  * Routines used for shrinking the Btree.
 676  *========================================================================*/
 677
 678 /*
 679  * Deallocate an empty leaf node, remove it from its parent,
 680  * possibly deallocating that block, etc...
 681  */
 682 int
 683 xfs_da_join(xfs_da_state_t *state)
 684 {
 685         xfs_da_state_blk_t *drop_blk, *save_blk;
 686         int action, error;
 687
 688         action = 0;
 689         drop_blk = &state->path.blk[ state->path.active-1 ];
 690         save_blk = &state->altpath.blk[ state->path.active-1 ];
 691         ASSERT(state->path.blk[0].magic == XFS_DA_NODE_MAGIC);
 692         ASSERT(drop_blk->magic == XFS_ATTR_LEAF_MAGIC ||
 693                drop_blk->magic == XFS_DIRX_LEAF_MAGIC(state->mp));
 694
 695         /*
 696          * Walk back up the tree joining/deallocating as necessary.
 697          * When we stop dropping blocks, break out.
 698          */
 699         for (  ; state->path.active >= 2; drop_blk--, save_blk--,
 700                  state->path.active--) {
 701                 /*
 702                  * See if we can combine the block with a neighbor.
 703                  *   (action == 0) => no options, just leave
 704                  *   (action == 1) => coalesce, then unlink
 705                  *   (action == 2) => block empty, unlink it
 706                  */
 707                 switch (drop_blk->magic) {
 708                 case XFS_ATTR_LEAF_MAGIC:
 709 #ifndef __KERNEL__
 710                         error = ENOTTY;
 711 #else
 712                         error = xfs_attr_leaf_toosmall(state, &action);
 713 #endif
 714                         if (error)
 715                                 return(error);
 716                         if (action == 0)
 717                                 return(0);
 718 #ifdef __KERNEL__
 719                         xfs_attr_leaf_unbalance(state, drop_blk, save_blk);
 720 #endif
 721                         break;
 722                 case XFS_DIR_LEAF_MAGIC:
 723                         ASSERT(XFS_DIR_IS_V1(state->mp));
 724                         error = xfs_dir_leaf_toosmall(state, &action);
 725                         if (error)
 726                                 return(error);
 727                         if (action == 0)
 728                                 return(0);
 729                         xfs_dir_leaf_unbalance(state, drop_blk, save_blk);
 730                         break;
 731                 case XFS_DIR2_LEAFN_MAGIC:
 732                         ASSERT(XFS_DIR_IS_V2(state->mp));
 733                         error = xfs_dir2_leafn_toosmall(state, &action);
 734                         if (error)
 735                                 return error;
 736                         if (action == 0)
 737                                 return 0;
 738                         xfs_dir2_leafn_unbalance(state, drop_blk, save_blk);
 739                         break;
 740                 case XFS_DA_NODE_MAGIC:
 741                         /*
 742                          * Remove the offending node, fixup hashvals,
 743                          * check for a toosmall neighbor.
 744                          */
 745                         xfs_da_node_remove(state, drop_blk);
 746                         xfs_da_fixhashpath(state, &state->path);
 747                         error = xfs_da_node_toosmall(state, &action);
 748                         if (error)
 749                                 return(error);
 750                         if (action == 0)
 751                                 return 0;
 752                         xfs_da_node_unbalance(state, drop_blk, save_blk);
 753                         break;
 754                 }
 755                 xfs_da_fixhashpath(state, &state->altpath);
 756                 error = xfs_da_blk_unlink(state, drop_blk, save_blk);
 757                 xfs_da_state_kill_altpath(state);
 758                 if (error)
 759                         return(error);
 760                 error = xfs_da_shrink_inode(state->args, drop_blk->blkno,
 761                                                          drop_blk->bp);
 762                 drop_blk->bp = NULL;
 763                 if (error)
 764                         return(error);
 765         }
 766         /*
 767          * We joined all the way to the top.  If it turns out that
 768          * we only have one entry in the root, make the child block
 769          * the new root.
 770          */
 771         xfs_da_node_remove(state, drop_blk);
 772         xfs_da_fixhashpath(state, &state->path);
 773         error = xfs_da_root_join(state, &state->path.blk[0]);
 774         return(error);
 775 }
 776
 777 /*
 778  * We have only one entry in the root.  Copy the only remaining child of
 779  * the old root to block 0 as the new root node.
 780  */
 781 STATIC int
 782 xfs_da_root_join(xfs_da_state_t *state, xfs_da_state_blk_t *root_blk)
 783 {
 784         xfs_da_intnode_t *oldroot;
 785         /* REFERENCED */
 786         xfs_da_blkinfo_t *blkinfo;
 787         xfs_da_args_t *args;
 788         xfs_dablk_t child;
 789         xfs_dabuf_t *bp;
 790         int error;
 791
 792         args = state->args;
 793         ASSERT(args != NULL);
 794         ASSERT(root_blk->magic == XFS_DA_NODE_MAGIC);
 795         oldroot = root_blk->bp->data;
 796         ASSERT(INT_GET(oldroot->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
 797         ASSERT(!oldroot->hdr.info.forw);
 798         ASSERT(!oldroot->hdr.info.back);
 799
 800         /*
 801          * If the root has more than one child, then don't do anything.
 802          */
 803         if (INT_GET(oldroot->hdr.count, ARCH_CONVERT) > 1)
 804                 return(0);
 805
 806         /*
 807          * Read in the (only) child block, then copy those bytes into
 808          * the root block's buffer and free the original child block.
 809          */
 810         child = INT_GET(oldroot->btree[ 0 ].before, ARCH_CONVERT);
 811         ASSERT(child != 0);
 812         error = xfs_da_read_buf(args->trans, args->dp, child, -1, &bp,
 813                                              args->whichfork);
 814         if (error)
 815                 return(error);
 816         ASSERT(bp != NULL);
 817         blkinfo = bp->data;
 818         if (INT_GET(oldroot->hdr.level, ARCH_CONVERT) == 1) {
 819                 ASSERT(INT_GET(blkinfo->magic, ARCH_CONVERT) == XFS_DIRX_LEAF_MAGIC(state->mp) ||
 820                        INT_GET(blkinfo->magic, ARCH_CONVERT) == XFS_ATTR_LEAF_MAGIC);
 821         } else {
 822                 ASSERT(INT_GET(blkinfo->magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
 823         }
 824         ASSERT(!blkinfo->forw);
 825         ASSERT(!blkinfo->back);
 826         memcpy(root_blk->bp->data, bp->data, state->blocksize);
 827         xfs_da_log_buf(args->trans, root_blk->bp, 0, state->blocksize - 1);
 828         error = xfs_da_shrink_inode(args, child, bp);
 829         return(error);
 830 }
 831
 832 /*
 833  * Check a node block and its neighbors to see if the block should be
 834  * collapsed into one or the other neighbor.  Always keep the block
 835  * with the smaller block number.
 836  * If the current block is over 50% full, don't try to join it, return 0.
 837  * If the block is empty, fill in the state structure and return 2.
 838  * If it can be collapsed, fill in the state structure and return 1.
 839  * If nothing can be done, return 0.
 840  */
 841 STATIC int
 842 xfs_da_node_toosmall(xfs_da_state_t *state, int *action)
 843 {
 844         xfs_da_intnode_t *node;
 845         xfs_da_state_blk_t *blk;
 846         xfs_da_blkinfo_t *info;
 847         int count, forward, error, retval, i;
 848         xfs_dablk_t blkno;
 849         xfs_dabuf_t *bp;
 850
 851         /*
 852          * Check for the degenerate case of the block being over 50% full.
 853          * If so, it's not worth even looking to see if we might be able
 854          * to coalesce with a sibling.
 855          */
 856         blk = &state->path.blk[ state->path.active-1 ];
 857         info = blk->bp->data;
 858         ASSERT(INT_GET(info->magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
 859         node = (xfs_da_intnode_t *)info;
 860         count = INT_GET(node->hdr.count, ARCH_CONVERT);
 861         if (count > (state->node_ents >> 1)) {
 862                 *action = 0;    /* blk over 50%, don't try to join */
 863                 return(0);      /* blk over 50%, don't try to join */
 864         }
 865
 866         /*
 867          * Check for the degenerate case of the block being empty.
 868          * If the block is empty, we'll simply delete it, no need to
 869          * coalesce it with a sibling block.  We choose (aribtrarily)
 870          * to merge with the forward block unless it is NULL.
 871          */
 872         if (count == 0) {
 873                 /*
 874                  * Make altpath point to the block we want to keep and
 875                  * path point to the block we want to drop (this one).
 876                  */
 877                 forward = info->forw;
 878                 memcpy(&state->altpath, &state->path, sizeof(state->path));
 879                 error = xfs_da_path_shift(state, &state->altpath, forward,
 880                                                  0, &retval);
 881                 if (error)
 882                         return(error);
 883                 if (retval) {
 884                         *action = 0;
 885                 } else {
 886                         *action = 2;
 887                 }
 888                 return(0);
 889         }
 890
 891         /*
 892          * Examine each sibling block to see if we can coalesce with
 893          * at least 25% free space to spare.  We need to figure out
 894          * whether to merge with the forward or the backward block.
 895          * We prefer coalescing with the lower numbered sibling so as
 896          * to shrink a directory over time.
 897          */
 898         /* start with smaller blk num */
 899         forward = (INT_GET(info->forw, ARCH_CONVERT)
 900                                 < INT_GET(info->back, ARCH_CONVERT));
 901         for (i = 0; i < 2; forward = !forward, i++) {
 902                 if (forward)
 903                         blkno = INT_GET(info->forw, ARCH_CONVERT);
 904                 else
 905                         blkno = INT_GET(info->back, ARCH_CONVERT);
 906                 if (blkno == 0)
 907                         continue;
 908                 error = xfs_da_read_buf(state->args->trans, state->args->dp,
 909                                         blkno, -1, &bp, state->args->whichfork);
 910                 if (error)
 911                         return(error);
 912                 ASSERT(bp != NULL);
 913
 914                 node = (xfs_da_intnode_t *)info;
 915                 count  = state->node_ents;
 916                 count -= state->node_ents >> 2;
 917                 count -= INT_GET(node->hdr.count, ARCH_CONVERT);
 918                 node = bp->data;
 919                 ASSERT(INT_GET(node->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
 920                 count -= INT_GET(node->hdr.count, ARCH_CONVERT);
 921                 xfs_da_brelse(state->args->trans, bp);
 922                 if (count >= 0)
 923                         break;  /* fits with at least 25% to spare */
 924         }
 925         if (i >= 2) {
 926                 *action = 0;
 927                 return(0);
 928         }
 929
 930         /*
 931          * Make altpath point to the block we want to keep (the lower
 932          * numbered block) and path point to the block we want to drop.
 933          */
 934         memcpy(&state->altpath, &state->path, sizeof(state->path));
 935         if (blkno < blk->blkno) {
 936                 error = xfs_da_path_shift(state, &state->altpath, forward,
 937                                                  0, &retval);
 938                 if (error) {
 939                         return(error);
 940                 }
 941                 if (retval) {
 942                         *action = 0;
 943                         return(0);
 944                 }
 945         } else {
 946                 error = xfs_da_path_shift(state, &state->path, forward,
 947                                                  0, &retval);
 948                 if (error) {
 949                         return(error);
 950                 }
 951                 if (retval) {
 952                         *action = 0;
 953                         return(0);
 954                 }
 955         }
 956         *action = 1;
 957         return(0);
 958 }
 959
 960 /*
 961  * Walk back up the tree adjusting hash values as necessary,
 962  * when we stop making changes, return.
 963  */
 964 void
 965 xfs_da_fixhashpath(xfs_da_state_t *state, xfs_da_state_path_t *path)
 966 {
 967         xfs_da_state_blk_t *blk;
 968         xfs_da_intnode_t *node;
 969         xfs_da_node_entry_t *btree;
 970         xfs_dahash_t lasthash=0;
 971         int level, count;
 972
 973         level = path->active-1;
 974         blk = &path->blk[ level ];
 975         switch (blk->magic) {
 976 #ifdef __KERNEL__
 977         case XFS_ATTR_LEAF_MAGIC:
 978                 lasthash = xfs_attr_leaf_lasthash(blk->bp, &count);
 979                 if (count == 0)
 980                         return;
 981                 break;
 982 #endif
 983         case XFS_DIR_LEAF_MAGIC:
 984                 ASSERT(XFS_DIR_IS_V1(state->mp));
 985                 lasthash = xfs_dir_leaf_lasthash(blk->bp, &count);
 986                 if (count == 0)
 987                         return;
 988                 break;
 989         case XFS_DIR2_LEAFN_MAGIC:
 990                 ASSERT(XFS_DIR_IS_V2(state->mp));
 991                 lasthash = xfs_dir2_leafn_lasthash(blk->bp, &count);
 992                 if (count == 0)
 993                         return;
 994                 break;
 995         case XFS_DA_NODE_MAGIC:
 996                 lasthash = xfs_da_node_lasthash(blk->bp, &count);
 997                 if (count == 0)
 998                         return;
 999                 break;
1000         }
1001         for (blk--, level--; level >= 0; blk--, level--) {
1002                 node = blk->bp->data;
1003                 ASSERT(INT_GET(node->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
1004                 btree = &node->btree[ blk->index ];
1005                 if (INT_GET(btree->hashval, ARCH_CONVERT) == lasthash)
1006                         break;
1007                 blk->hashval = lasthash;
1008                 INT_SET(btree->hashval, ARCH_CONVERT, lasthash);
1009                 xfs_da_log_buf(state->args->trans, blk->bp,
1010                                   XFS_DA_LOGRANGE(node, btree, sizeof(*btree)));
1011
1012                 lasthash = INT_GET(node->btree[ INT_GET(node->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT);
1013         }
1014 }
1015
1016 /*
1017  * Remove an entry from an intermediate node.
1018  */
1019 STATIC void
1020 xfs_da_node_remove(xfs_da_state_t *state, xfs_da_state_blk_t *drop_blk)
1021 {
1022         xfs_da_intnode_t *node;
1023         xfs_da_node_entry_t *btree;
1024         int tmp;
1025
1026         node = drop_blk->bp->data;
1027         ASSERT(drop_blk->index < INT_GET(node->hdr.count, ARCH_CONVERT));
1028         ASSERT(drop_blk->index >= 0);
1029
1030         /*
1031          * Copy over the offending entry, or just zero it out.
1032          */
1033         btree = &node->btree[drop_blk->index];
1034         if (drop_blk->index < (INT_GET(node->hdr.count, ARCH_CONVERT)-1)) {
1035                 tmp  = INT_GET(node->hdr.count, ARCH_CONVERT) - drop_blk->index - 1;
1036                 tmp *= (uint)sizeof(xfs_da_node_entry_t);
1037                 memmove(btree, btree + 1, tmp);
1038                 xfs_da_log_buf(state->args->trans, drop_blk->bp,
1039                     XFS_DA_LOGRANGE(node, btree, tmp));
1040                 btree = &node->btree[ INT_GET(node->hdr.count, ARCH_CONVERT)-1 ];
1041         }
1042         memset((char *)btree, 0, sizeof(xfs_da_node_entry_t));
1043         xfs_da_log_buf(state->args->trans, drop_blk->bp,
1044             XFS_DA_LOGRANGE(node, btree, sizeof(*btree)));
1045         INT_MOD(node->hdr.count, ARCH_CONVERT, -1);
1046         xfs_da_log_buf(state->args->trans, drop_blk->bp,
1047             XFS_DA_LOGRANGE(node, &node->hdr, sizeof(node->hdr)));
1048
1049         /*
1050          * Copy the last hash value from the block to propagate upwards.
1051          */
1052         btree--;
1053         drop_blk->hashval = INT_GET(btree->hashval, ARCH_CONVERT);
1054 }
1055
1056 /*
1057  * Unbalance the btree elements between two intermediate nodes,
1058  * move all Btree elements from one node into another.
1059  */
1060 STATIC void
1061 xfs_da_node_unbalance(xfs_da_state_t *state, xfs_da_state_blk_t *drop_blk,
1062                                      xfs_da_state_blk_t *save_blk)
1063 {
1064         xfs_da_intnode_t *drop_node, *save_node;
1065         xfs_da_node_entry_t *btree;
1066         int tmp;
1067         xfs_trans_t *tp;
1068
1069         drop_node = drop_blk->bp->data;
1070         save_node = save_blk->bp->data;
1071         ASSERT(INT_GET(drop_node->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
1072         ASSERT(INT_GET(save_node->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
1073         tp = state->args->trans;
1074
1075         /*
1076          * If the dying block has lower hashvals, then move all the
1077          * elements in the remaining block up to make a hole.
1078          */
1079         if ((INT_GET(drop_node->btree[ 0 ].hashval, ARCH_CONVERT) < INT_GET(save_node->btree[ 0 ].hashval, ARCH_CONVERT)) ||
1080             (INT_GET(drop_node->btree[ INT_GET(drop_node->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT) <
1081              INT_GET(save_node->btree[ INT_GET(save_node->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT)))
1082         {
1083                 btree = &save_node->btree[ INT_GET(drop_node->hdr.count, ARCH_CONVERT) ];
1084                 tmp = INT_GET(save_node->hdr.count, ARCH_CONVERT) * (uint)sizeof(xfs_da_node_entry_t);
1085                 memmove(btree, &save_node->btree[0], tmp);
1086                 btree = &save_node->btree[0];
1087                 xfs_da_log_buf(tp, save_blk->bp,
1088                         XFS_DA_LOGRANGE(save_node, btree,
1089                                 (INT_GET(save_node->hdr.count, ARCH_CONVERT) + INT_GET(drop_node->hdr.count, ARCH_CONVERT)) *
1090                                 sizeof(xfs_da_node_entry_t)));
1091         } else {
1092                 btree = &save_node->btree[ INT_GET(save_node->hdr.count, ARCH_CONVERT) ];
1093                 xfs_da_log_buf(tp, save_blk->bp,
1094                         XFS_DA_LOGRANGE(save_node, btree,
1095                                 INT_GET(drop_node->hdr.count, ARCH_CONVERT) *
1096                                 sizeof(xfs_da_node_entry_t)));
1097         }
1098
1099         /*
1100          * Move all the B-tree elements from drop_blk to save_blk.
1101          */
1102         tmp = INT_GET(drop_node->hdr.count, ARCH_CONVERT) * (uint)sizeof(xfs_da_node_entry_t);
1103         memcpy(btree, &drop_node->btree[0], tmp);
1104         INT_MOD(save_node->hdr.count, ARCH_CONVERT, INT_GET(drop_node->hdr.count, ARCH_CONVERT));
1105
1106         xfs_da_log_buf(tp, save_blk->bp,
1107                 XFS_DA_LOGRANGE(save_node, &save_node->hdr,
1108                         sizeof(save_node->hdr)));
1109
1110         /*
1111          * Save the last hashval in the remaining block for upward propagation.
1112          */
1113         save_blk->hashval = INT_GET(save_node->btree[ INT_GET(save_node->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT);
1114 }
1115
1116 /*========================================================================
1117  * Routines used for finding things in the Btree.
1118  *========================================================================*/
1119
1120 /*
1121  * Walk down the Btree looking for a particular filename, filling
1122  * in the state structure as we go.
1123  *
1124  * We will set the state structure to point to each of the elements
1125  * in each of the nodes where either the hashval is or should be.
1126  *
1127  * We support duplicate hashval's so for each entry in the current
1128  * node that could contain the desired hashval, descend.  This is a
1129  * pruned depth-first tree search.
1130  */
1131 int                                                     /* error */
1132 xfs_da_node_lookup_int(xfs_da_state_t *state, int *result)
1133 {
1134         xfs_da_state_blk_t *blk;
1135         xfs_da_blkinfo_t *curr;
1136         xfs_da_intnode_t *node;
1137         xfs_da_node_entry_t *btree;
1138         xfs_dablk_t blkno;
1139         int probe, span, max, error, retval;
1140         xfs_dahash_t hashval;
1141         xfs_da_args_t *args;
1142
1143         args = state->args;
1144
1145         /*
1146          * Descend thru the B-tree searching each level for the right
1147          * node to use, until the right hashval is found.
1148          */
1149         if (args->whichfork == XFS_DATA_FORK && XFS_DIR_IS_V2(state->mp))
1150                 blkno = state->mp->m_dirleafblk;
1151         else
1152                 blkno = 0;
1153         for (blk = &state->path.blk[0], state->path.active = 1;
1154                          state->path.active <= XFS_DA_NODE_MAXDEPTH;
1155                          blk++, state->path.active++) {
1156                 /*
1157                  * Read the next node down in the tree.
1158                  */
1159                 blk->blkno = blkno;
1160                 error = xfs_da_read_buf(args->trans, args->dp, blkno,
1161                                         -1, &blk->bp, args->whichfork);
1162                 if (error) {
1163                         blk->blkno = 0;
1164                         state->path.active--;
1165                         return(error);
1166                 }
1167                 curr = blk->bp->data;
1168                 ASSERT(INT_GET(curr->magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC ||
1169                        INT_GET(curr->magic, ARCH_CONVERT) == XFS_DIRX_LEAF_MAGIC(state->mp) ||
1170                        INT_GET(curr->magic, ARCH_CONVERT) == XFS_ATTR_LEAF_MAGIC);
1171
1172                 /*
1173                  * Search an intermediate node for a match.
1174                  */
1175                 blk->magic = INT_GET(curr->magic, ARCH_CONVERT);
1176                 if (INT_GET(curr->magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC) {
1177                         node = blk->bp->data;
1178                         blk->hashval = INT_GET(node->btree[ INT_GET(node->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT);
1179
1180                         /*
1181                          * Binary search.  (note: small blocks will skip loop)
1182                          */
1183                         max = INT_GET(node->hdr.count, ARCH_CONVERT);
1184                         probe = span = max / 2;
1185                         hashval = args->hashval;
1186                         for (btree = &node->btree[probe]; span > 4;
1187                                    btree = &node->btree[probe]) {
1188                                 span /= 2;
1189                                 if (INT_GET(btree->hashval, ARCH_CONVERT) < hashval)
1190                                         probe += span;
1191                                 else if (INT_GET(btree->hashval, ARCH_CONVERT) > hashval)
1192                                         probe -= span;
1193                                 else
1194                                         break;
1195                         }
1196                         ASSERT((probe >= 0) && (probe < max));
1197                         ASSERT((span <= 4) || (INT_GET(btree->hashval, ARCH_CONVERT) == hashval));
1198
1199                         /*
1200                          * Since we may have duplicate hashval's, find the first
1201                          * matching hashval in the node.
1202                          */
1203                         while ((probe > 0) && (INT_GET(btree->hashval, ARCH_CONVERT) >= hashval)) {
1204                                 btree--;
1205                                 probe--;
1206                         }
1207                         while ((probe < max) && (INT_GET(btree->hashval, ARCH_CONVERT) < hashval)) {
1208                                 btree++;
1209                                 probe++;
1210                         }
1211
1212                         /*
1213                          * Pick the right block to descend on.
1214                          */
1215                         if (probe == max) {
1216                                 blk->index = max-1;
1217                                 blkno = INT_GET(node->btree[ max-1 ].before, ARCH_CONVERT);
1218                         } else {
1219                                 blk->index = probe;
1220                                 blkno = INT_GET(btree->before, ARCH_CONVERT);
1221                         }
1222                 }
1223 #ifdef __KERNEL__
1224                 else if (INT_GET(curr->magic, ARCH_CONVERT) == XFS_ATTR_LEAF_MAGIC) {
1225                         blk->hashval = xfs_attr_leaf_lasthash(blk->bp, NULL);
1226                         break;
1227                 }
1228 #endif
1229                 else if (INT_GET(curr->magic, ARCH_CONVERT) == XFS_DIR_LEAF_MAGIC) {
1230                         blk->hashval = xfs_dir_leaf_lasthash(blk->bp, NULL);
1231                         break;
1232                 }
1233                 else if (INT_GET(curr->magic, ARCH_CONVERT) == XFS_DIR2_LEAFN_MAGIC) {
1234                         blk->hashval = xfs_dir2_leafn_lasthash(blk->bp, NULL);
1235                         break;
1236                 }
1237         }
1238
1239         /*
1240          * A leaf block that ends in the hashval that we are interested in
1241          * (final hashval == search hashval) means that the next block may
1242          * contain more entries with the same hashval, shift upward to the
1243          * next leaf and keep searching.
1244          */
1245         for (;;) {
1246                 if (blk->magic == XFS_DIR_LEAF_MAGIC) {
1247                         ASSERT(XFS_DIR_IS_V1(state->mp));
1248                         retval = xfs_dir_leaf_lookup_int(blk->bp, args,
1249                                                                   &blk->index);
1250                 } else if (blk->magic == XFS_DIR2_LEAFN_MAGIC) {
1251                         ASSERT(XFS_DIR_IS_V2(state->mp));
1252                         retval = xfs_dir2_leafn_lookup_int(blk->bp, args,
1253                                                         &blk->index, state);
1254                 }
1255 #ifdef __KERNEL__
1256                 else if (blk->magic == XFS_ATTR_LEAF_MAGIC) {
1257                         retval = xfs_attr_leaf_lookup_int(blk->bp, args);
1258                         blk->index = args->index;
1259                         args->blkno = blk->blkno;
1260                 }
1261 #endif
1262                 if (((retval == ENOENT) || (retval == ENOATTR)) &&
1263                     (blk->hashval == args->hashval)) {
1264                         error = xfs_da_path_shift(state, &state->path, 1, 1,
1265                                                          &retval);
1266                         if (error)
1267                                 return(error);
1268                         if (retval == 0) {
1269                                 continue;
1270                         }
1271 #ifdef __KERNEL__
1272                         else if (blk->magic == XFS_ATTR_LEAF_MAGIC) {
1273                                 /* path_shift() gives ENOENT */
1274                                 retval = XFS_ERROR(ENOATTR);
1275                         }
1276 #endif
1277                 }
1278                 break;
1279         }
1280         *result = retval;
1281         return(0);
1282 }
1283
1284 /*========================================================================
1285  * Utility routines.
1286  *========================================================================*/
1287
1288 /*
1289  * Link a new block into a doubly linked list of blocks (of whatever type).
1290  */
1291 int                                                     /* error */
1292 xfs_da_blk_link(xfs_da_state_t *state, xfs_da_state_blk_t *old_blk,
1293                                xfs_da_state_blk_t *new_blk)
1294 {
1295         xfs_da_blkinfo_t *old_info, *new_info, *tmp_info;
1296         xfs_da_args_t *args;
1297         int before=0, error;
1298         xfs_dabuf_t *bp;
1299
1300         /*
1301          * Set up environment.
1302          */
1303         args = state->args;
1304         ASSERT(args != NULL);
1305         old_info = old_blk->bp->data;
1306         new_info = new_blk->bp->data;
1307         ASSERT(old_blk->magic == XFS_DA_NODE_MAGIC ||
1308                old_blk->magic == XFS_DIRX_LEAF_MAGIC(state->mp) ||
1309                old_blk->magic == XFS_ATTR_LEAF_MAGIC);
1310         ASSERT(old_blk->magic == INT_GET(old_info->magic, ARCH_CONVERT));
1311         ASSERT(new_blk->magic == INT_GET(new_info->magic, ARCH_CONVERT));
1312         ASSERT(old_blk->magic == new_blk->magic);
1313
1314         switch (old_blk->magic) {
1315 #ifdef __KERNEL__
1316         case XFS_ATTR_LEAF_MAGIC:
1317                 before = xfs_attr_leaf_order(old_blk->bp, new_blk->bp);
1318                 break;
1319 #endif
1320         case XFS_DIR_LEAF_MAGIC:
1321                 ASSERT(XFS_DIR_IS_V1(state->mp));
1322                 before = xfs_dir_leaf_order(old_blk->bp, new_blk->bp);
1323                 break;
1324         case XFS_DIR2_LEAFN_MAGIC:
1325                 ASSERT(XFS_DIR_IS_V2(state->mp));
1326                 before = xfs_dir2_leafn_order(old_blk->bp, new_blk->bp);
1327                 break;
1328         case XFS_DA_NODE_MAGIC:
1329                 before = xfs_da_node_order(old_blk->bp, new_blk->bp);
1330                 break;
1331         }
1332
1333         /*
1334          * Link blocks in appropriate order.
1335          */
1336         if (before) {
1337                 /*
1338                  * Link new block in before existing block.
1339                  */
1340                 INT_SET(new_info->forw, ARCH_CONVERT, old_blk->blkno);
1341                 new_info->back = old_info->back; /* INT_: direct copy */
1342                 if (INT_GET(old_info->back, ARCH_CONVERT)) {
1343                         error = xfs_da_read_buf(args->trans, args->dp,
1344                                                 INT_GET(old_info->back,
1345                                                         ARCH_CONVERT), -1, &bp,
1346                                                 args->whichfork);
1347                         if (error)
1348                                 return(error);
1349                         ASSERT(bp != NULL);
1350                         tmp_info = bp->data;
1351                         ASSERT(INT_GET(tmp_info->magic, ARCH_CONVERT) == INT_GET(old_info->magic, ARCH_CONVERT));
1352                         ASSERT(INT_GET(tmp_info->forw, ARCH_CONVERT) == old_blk->blkno);
1353                         INT_SET(tmp_info->forw, ARCH_CONVERT, new_blk->blkno);
1354                         xfs_da_log_buf(args->trans, bp, 0, sizeof(*tmp_info)-1);
1355                         xfs_da_buf_done(bp);
1356                 }
1357                 INT_SET(old_info->back, ARCH_CONVERT, new_blk->blkno);
1358         } else {
1359                 /*
1360                  * Link new block in after existing block.
1361                  */
1362                 new_info->forw = old_info->forw; /* INT_: direct copy */
1363                 INT_SET(new_info->back, ARCH_CONVERT, old_blk->blkno);
1364                 if (INT_GET(old_info->forw, ARCH_CONVERT)) {
1365                         error = xfs_da_read_buf(args->trans, args->dp,
1366                                                 INT_GET(old_info->forw, ARCH_CONVERT), -1, &bp,
1367                                                 args->whichfork);
1368                         if (error)
1369                                 return(error);
1370                         ASSERT(bp != NULL);
1371                         tmp_info = bp->data;
1372                         ASSERT(INT_GET(tmp_info->magic, ARCH_CONVERT)
1373                                     == INT_GET(old_info->magic, ARCH_CONVERT));
1374                         ASSERT(INT_GET(tmp_info->back, ARCH_CONVERT)
1375                                     == old_blk->blkno);
1376                         INT_SET(tmp_info->back, ARCH_CONVERT, new_blk->blkno);
1377                         xfs_da_log_buf(args->trans, bp, 0, sizeof(*tmp_info)-1);
1378                         xfs_da_buf_done(bp);
1379                 }
1380                 INT_SET(old_info->forw, ARCH_CONVERT, new_blk->blkno);
1381         }
1382
1383         xfs_da_log_buf(args->trans, old_blk->bp, 0, sizeof(*tmp_info) - 1);
1384         xfs_da_log_buf(args->trans, new_blk->bp, 0, sizeof(*tmp_info) - 1);
1385         return(0);
1386 }
1387
1388 /*
1389  * Compare two intermediate nodes for "order".
1390  */
1391 STATIC int
1392 xfs_da_node_order(xfs_dabuf_t *node1_bp, xfs_dabuf_t *node2_bp)
1393 {
1394         xfs_da_intnode_t *node1, *node2;
1395
1396         node1 = node1_bp->data;
1397         node2 = node2_bp->data;
1398         ASSERT((INT_GET(node1->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC) &&
1399                (INT_GET(node2->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC));
1400         if ((INT_GET(node1->hdr.count, ARCH_CONVERT) > 0) && (INT_GET(node2->hdr.count, ARCH_CONVERT) > 0) &&
1401             ((INT_GET(node2->btree[ 0 ].hashval, ARCH_CONVERT) <
1402               INT_GET(node1->btree[ 0 ].hashval, ARCH_CONVERT)) ||
1403              (INT_GET(node2->btree[ INT_GET(node2->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT) <
1404               INT_GET(node1->btree[ INT_GET(node1->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT)))) {
1405                 return(1);
1406         }
1407         return(0);
1408 }
1409
1410 /*
1411  * Pick up the last hashvalue from an intermediate node.
1412  */
1413 STATIC uint
1414 xfs_da_node_lasthash(xfs_dabuf_t *bp, int *count)
1415 {
1416         xfs_da_intnode_t *node;
1417
1418         node = bp->data;
1419         ASSERT(INT_GET(node->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
1420         if (count)
1421                 *count = INT_GET(node->hdr.count, ARCH_CONVERT);
1422         if (!node->hdr.count)
1423                 return(0);
1424         return(INT_GET(node->btree[ INT_GET(node->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT));
1425 }
1426
1427 /*
1428  * Unlink a block from a doubly linked list of blocks.
1429  */
1430 STATIC int                                              /* error */
1431 xfs_da_blk_unlink(xfs_da_state_t *state, xfs_da_state_blk_t *drop_blk,
1432                                  xfs_da_state_blk_t *save_blk)
1433 {
1434         xfs_da_blkinfo_t *drop_info, *save_info, *tmp_info;
1435         xfs_da_args_t *args;
1436         xfs_dabuf_t *bp;
1437         int error;
1438
1439         /*
1440          * Set up environment.
1441          */
1442         args = state->args;
1443         ASSERT(args != NULL);
1444         save_info = save_blk->bp->data;
1445         drop_info = drop_blk->bp->data;
1446         ASSERT(save_blk->magic == XFS_DA_NODE_MAGIC ||
1447                save_blk->magic == XFS_DIRX_LEAF_MAGIC(state->mp) ||
1448                save_blk->magic == XFS_ATTR_LEAF_MAGIC);
1449         ASSERT(save_blk->magic == INT_GET(save_info->magic, ARCH_CONVERT));
1450         ASSERT(drop_blk->magic == INT_GET(drop_info->magic, ARCH_CONVERT));
1451         ASSERT(save_blk->magic == drop_blk->magic);
1452         ASSERT((INT_GET(save_info->forw, ARCH_CONVERT) == drop_blk->blkno) ||
1453                (INT_GET(save_info->back, ARCH_CONVERT) == drop_blk->blkno));
1454         ASSERT((INT_GET(drop_info->forw, ARCH_CONVERT) == save_blk->blkno) ||
1455                (INT_GET(drop_info->back, ARCH_CONVERT) == save_blk->blkno));
1456
1457         /*
1458          * Unlink the leaf block from the doubly linked chain of leaves.
1459          */
1460         if (INT_GET(save_info->back, ARCH_CONVERT) == drop_blk->blkno) {
1461                 save_info->back = drop_info->back; /* INT_: direct copy */
1462                 if (INT_GET(drop_info->back, ARCH_CONVERT)) {
1463                         error = xfs_da_read_buf(args->trans, args->dp,
1464                                                 INT_GET(drop_info->back,
1465                                                         ARCH_CONVERT), -1, &bp,
1466                                                 args->whichfork);
1467                         if (error)
1468                                 return(error);
1469                         ASSERT(bp != NULL);
1470                         tmp_info = bp->data;
1471                         ASSERT(INT_GET(tmp_info->magic, ARCH_CONVERT) == INT_GET(save_info->magic, ARCH_CONVERT));
1472                         ASSERT(INT_GET(tmp_info->forw, ARCH_CONVERT) == drop_blk->blkno);
1473                         INT_SET(tmp_info->forw, ARCH_CONVERT, save_blk->blkno);
1474                         xfs_da_log_buf(args->trans, bp, 0,
1475                                                     sizeof(*tmp_info) - 1);
1476                         xfs_da_buf_done(bp);
1477                 }
1478         } else {
1479                 save_info->forw = drop_info->forw; /* INT_: direct copy */
1480                 if (INT_GET(drop_info->forw, ARCH_CONVERT)) {
1481                         error = xfs_da_read_buf(args->trans, args->dp,
1482                                                 INT_GET(drop_info->forw, ARCH_CONVERT), -1, &bp,
1483                                                 args->whichfork);
1484                         if (error)
1485                                 return(error);
1486                         ASSERT(bp != NULL);
1487                         tmp_info = bp->data;
1488                         ASSERT(INT_GET(tmp_info->magic, ARCH_CONVERT)
1489                                     == INT_GET(save_info->magic, ARCH_CONVERT));
1490                         ASSERT(INT_GET(tmp_info->back, ARCH_CONVERT)
1491                                     == drop_blk->blkno);
1492                         INT_SET(tmp_info->back, ARCH_CONVERT, save_blk->blkno);
1493                         xfs_da_log_buf(args->trans, bp, 0,
1494                                                     sizeof(*tmp_info) - 1);
1495                         xfs_da_buf_done(bp);
1496                 }
1497         }
1498
1499         xfs_da_log_buf(args->trans, save_blk->bp, 0, sizeof(*save_info) - 1);
1500         return(0);
1501 }
1502
1503 /*
1504  * Move a path "forward" or "!forward" one block at the current level.
1505  *
1506  * This routine will adjust a "path" to point to the next block
1507  * "forward" (higher hashvalues) or "!forward" (lower hashvals) in the
1508  * Btree, including updating pointers to the intermediate nodes between
1509  * the new bottom and the root.
1510  */
1511 int                                                     /* error */
1512 xfs_da_path_shift(xfs_da_state_t *state, xfs_da_state_path_t *path,
1513                                  int forward, int release, int *result)
1514 {
1515         xfs_da_state_blk_t *blk;
1516         xfs_da_blkinfo_t *info;
1517         xfs_da_intnode_t *node;
1518         xfs_da_args_t *args;
1519         xfs_dablk_t blkno=0;
1520         int level, error;
1521
1522         /*
1523          * Roll up the Btree looking for the first block where our
1524          * current index is not at the edge of the block.  Note that
1525          * we skip the bottom layer because we want the sibling block.
1526          */
1527         args = state->args;
1528         ASSERT(args != NULL);
1529         ASSERT(path != NULL);
1530         ASSERT((path->active > 0) && (path->active < XFS_DA_NODE_MAXDEPTH));
1531         level = (path->active-1) - 1;   /* skip bottom layer in path */
1532         for (blk = &path->blk[level]; level >= 0; blk--, level--) {
1533                 ASSERT(blk->bp != NULL);
1534                 node = blk->bp->data;
1535                 ASSERT(INT_GET(node->hdr.info.magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
1536                 if (forward && (blk->index < INT_GET(node->hdr.count, ARCH_CONVERT)-1)) {
1537                         blk->index++;
1538                         blkno = INT_GET(node->btree[ blk->index ].before, ARCH_CONVERT);
1539                         break;
1540                 } else if (!forward && (blk->index > 0)) {
1541                         blk->index--;
1542                         blkno = INT_GET(node->btree[ blk->index ].before, ARCH_CONVERT);
1543                         break;
1544                 }
1545         }
1546         if (level < 0) {
1547                 *result = XFS_ERROR(ENOENT);    /* we're out of our tree */
1548                 ASSERT(args->oknoent);
1549                 return(0);
1550         }
1551
1552         /*
1553          * Roll down the edge of the subtree until we reach the
1554          * same depth we were at originally.
1555          */
1556         for (blk++, level++; level < path->active; blk++, level++) {
1557                 /*
1558                  * Release the old block.
1559                  * (if it's dirty, trans won't actually let go)
1560                  */
1561                 if (release)
1562                         xfs_da_brelse(args->trans, blk->bp);
1563
1564                 /*
1565                  * Read the next child block.
1566                  */
1567                 blk->blkno = blkno;
1568                 error = xfs_da_read_buf(args->trans, args->dp, blkno, -1,
1569                                                      &blk->bp, args->whichfork);
1570                 if (error)
1571                         return(error);
1572                 ASSERT(blk->bp != NULL);
1573                 info = blk->bp->data;
1574                 ASSERT(INT_GET(info->magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC ||
1575                        INT_GET(info->magic, ARCH_CONVERT) == XFS_DIRX_LEAF_MAGIC(state->mp) ||
1576                        INT_GET(info->magic, ARCH_CONVERT) == XFS_ATTR_LEAF_MAGIC);
1577                 blk->magic = INT_GET(info->magic, ARCH_CONVERT);
1578                 if (INT_GET(info->magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC) {
1579                         node = (xfs_da_intnode_t *)info;
1580                         blk->hashval = INT_GET(node->btree[ INT_GET(node->hdr.count, ARCH_CONVERT)-1 ].hashval, ARCH_CONVERT);
1581                         if (forward)
1582                                 blk->index = 0;
1583                         else
1584                                 blk->index = INT_GET(node->hdr.count, ARCH_CONVERT)-1;
1585                         blkno = INT_GET(node->btree[ blk->index ].before, ARCH_CONVERT);
1586                 } else {
1587                         ASSERT(level == path->active-1);
1588                         blk->index = 0;
1589                         switch(blk->magic) {
1590 #ifdef __KERNEL__
1591                         case XFS_ATTR_LEAF_MAGIC:
1592                                 blk->hashval = xfs_attr_leaf_lasthash(blk->bp,
1593                                                                       NULL);
1594                                 break;
1595 #endif
1596                         case XFS_DIR_LEAF_MAGIC:
1597                                 ASSERT(XFS_DIR_IS_V1(state->mp));
1598                                 blk->hashval = xfs_dir_leaf_lasthash(blk->bp,
1599                                                                      NULL);
1600                                 break;
1601                         case XFS_DIR2_LEAFN_MAGIC:
1602                                 ASSERT(XFS_DIR_IS_V2(state->mp));
1603                                 blk->hashval = xfs_dir2_leafn_lasthash(blk->bp,
1604                                                                        NULL);
1605                                 break;
1606                         default:
1607                                 ASSERT(blk->magic == XFS_ATTR_LEAF_MAGIC ||
1608                                        blk->magic ==
1609                                        XFS_DIRX_LEAF_MAGIC(state->mp));
1610                                 break;
1611                         }
1612                 }
1613         }
1614         *result = 0;
1615         return(0);
1616 }
1617
1618
1619 /*========================================================================
1620  * Utility routines.
1621  *========================================================================*/
1622
1623 /*
1624  * Implement a simple hash on a character string.
1625  * Rotate the hash value by 7 bits, then XOR each character in.
1626  * This is implemented with some source-level loop unrolling.
1627  */
1628 xfs_dahash_t
1629 xfs_da_hashname(uchar_t *name, int namelen)
1630 {
1631         xfs_dahash_t hash;
1632
1633 #ifdef SLOWVERSION
1634         /*
1635          * This is the old one-byte-at-a-time version.
1636          */
1637         for (hash = 0; namelen > 0; namelen--)
1638                 hash = *name++ ^ rol32(hash, 7);
1639
1640         return(hash);
1641 #else
1642         /*
1643          * Do four characters at a time as long as we can.
1644          */
1645         for (hash = 0; namelen >= 4; namelen -= 4, name += 4)
1646                 hash = (name[0] << 21) ^ (name[1] << 14) ^ (name[2] << 7) ^
1647                        (name[3] << 0) ^ rol32(hash, 7 * 4);
1648
1649         /*
1650          * Now do the rest of the characters.
1651          */
1652         switch (namelen) {
1653         case 3:
1654                 return (name[0] << 14) ^ (name[1] << 7) ^ (name[2] << 0) ^
1655                        rol32(hash, 7 * 3);
1656         case 2:
1657                 return (name[0] << 7) ^ (name[1] << 0) ^ rol32(hash, 7 * 2);
1658         case 1:
1659                 return (name[0] << 0) ^ rol32(hash, 7 * 1);
1660         case 0:
1661                 return hash;
1662         }
1663         /* NOTREACHED */
1664 #endif
1665         return 0; /* keep gcc happy */
1666 }
1667
1668 /*
1669  * Add a block to the btree ahead of the file.
1670  * Return the new block number to the caller.
1671  */
1672 int
1673 xfs_da_grow_inode(xfs_da_args_t *args, xfs_dablk_t *new_blkno)
1674 {
1675         xfs_fileoff_t bno, b;
1676         xfs_bmbt_irec_t map;
1677         xfs_bmbt_irec_t *mapp;
1678         xfs_inode_t *dp;
1679         int nmap, error, w, count, c, got, i, mapi;
1680         xfs_fsize_t size;
1681         xfs_trans_t *tp;
1682         xfs_mount_t *mp;
1683
1684         dp = args->dp;
1685         mp = dp->i_mount;
1686         w = args->whichfork;
1687         tp = args->trans;
1688         /*
1689          * For new directories adjust the file offset and block count.
1690          */
1691         if (w == XFS_DATA_FORK && XFS_DIR_IS_V2(mp)) {
1692                 bno = mp->m_dirleafblk;
1693                 count = mp->m_dirblkfsbs;
1694         } else {
1695                 bno = 0;
1696                 count = 1;
1697         }
1698         /*
1699          * Find a spot in the file space to put the new block.
1700          */
1701         if ((error = xfs_bmap_first_unused(tp, dp, count, &bno, w))) {
1702                 return error;
1703         }
1704         if (w == XFS_DATA_FORK && XFS_DIR_IS_V2(mp))
1705                 ASSERT(bno >= mp->m_dirleafblk && bno < mp->m_dirfreeblk);
1706         /*
1707          * Try mapping it in one filesystem block.
1708          */
1709         nmap = 1;
1710         ASSERT(args->firstblock != NULL);
1711         if ((error = xfs_bmapi(tp, dp, bno, count,
1712                         XFS_BMAPI_AFLAG(w)|XFS_BMAPI_WRITE|XFS_BMAPI_METADATA|
1713                         XFS_BMAPI_CONTIG,
1714                         args->firstblock, args->total, &map, &nmap,
1715                         args->flist))) {
1716                 return error;
1717         }
1718         ASSERT(nmap <= 1);
1719         if (nmap == 1) {
1720                 mapp = &map;
1721                 mapi = 1;
1722         }
1723         /*
1724          * If we didn't get it and the block might work if fragmented,
1725          * try without the CONTIG flag.  Loop until we get it all.
1726          */
1727         else if (nmap == 0 && count > 1) {
1728                 mapp = kmem_alloc(sizeof(*mapp) * count, KM_SLEEP);
1729                 for (b = bno, mapi = 0; b < bno + count; ) {
1730                         nmap = MIN(XFS_BMAP_MAX_NMAP, count);
1731                         c = (int)(bno + count - b);
1732                         if ((error = xfs_bmapi(tp, dp, b, c,
1733                                         XFS_BMAPI_AFLAG(w)|XFS_BMAPI_WRITE|
1734                                         XFS_BMAPI_METADATA,
1735                                         args->firstblock, args->total,
1736                                         &mapp[mapi], &nmap, args->flist))) {
1737                                 kmem_free(mapp, sizeof(*mapp) * count);
1738                                 return error;
1739                         }
1740                         if (nmap < 1)
1741                                 break;
1742                         mapi += nmap;
1743                         b = mapp[mapi - 1].br_startoff +
1744                             mapp[mapi - 1].br_blockcount;
1745                 }
1746         } else {
1747                 mapi = 0;
1748                 mapp = NULL;
1749         }
1750         /*
1751          * Count the blocks we got, make sure it matches the total.
1752          */
1753         for (i = 0, got = 0; i < mapi; i++)
1754                 got += mapp[i].br_blockcount;
1755         if (got != count || mapp[0].br_startoff != bno ||
1756             mapp[mapi - 1].br_startoff + mapp[mapi - 1].br_blockcount !=
1757             bno + count) {
1758                 if (mapp != &map)
1759                         kmem_free(mapp, sizeof(*mapp) * count);
1760                 return XFS_ERROR(ENOSPC);
1761         }
1762         if (mapp != &map)
1763                 kmem_free(mapp, sizeof(*mapp) * count);
1764         *new_blkno = (xfs_dablk_t)bno;
1765         /*
1766          * For version 1 directories, adjust the file size if it changed.
1767          */
1768         if (w == XFS_DATA_FORK && XFS_DIR_IS_V1(mp)) {
1769                 ASSERT(mapi == 1);
1770                 if ((error = xfs_bmap_last_offset(tp, dp, &bno, w)))
1771                         return error;
1772                 size = XFS_FSB_TO_B(mp, bno);
1773                 if (size != dp->i_d.di_size) {
1774                         dp->i_d.di_size = size;
1775                         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1776                 }
1777         }
1778         return 0;
1779 }
1780
1781 /*
1782  * Ick.  We need to always be able to remove a btree block, even
1783  * if there's no space reservation because the filesystem is full.
1784  * This is called if xfs_bunmapi on a btree block fails due to ENOSPC.
1785  * It swaps the target block with the last block in the file.  The
1786  * last block in the file can always be removed since it can't cause
1787  * a bmap btree split to do that.
1788  */
1789 STATIC int
1790 xfs_da_swap_lastblock(xfs_da_args_t *args, xfs_dablk_t *dead_blknop,
1791                       xfs_dabuf_t **dead_bufp)
1792 {
1793         xfs_dablk_t dead_blkno, last_blkno, sib_blkno, par_blkno;
1794         xfs_dabuf_t *dead_buf, *last_buf, *sib_buf, *par_buf;
1795         xfs_fileoff_t lastoff;
1796         xfs_inode_t *ip;
1797         xfs_trans_t *tp;
1798         xfs_mount_t *mp;
1799         int error, w, entno, level, dead_level;
1800         xfs_da_blkinfo_t *dead_info, *sib_info;
1801         xfs_da_intnode_t *par_node, *dead_node;
1802         xfs_dir_leafblock_t *dead_leaf;
1803         xfs_dir2_leaf_t *dead_leaf2;
1804         xfs_dahash_t dead_hash;
1805
1806         dead_buf = *dead_bufp;
1807         dead_blkno = *dead_blknop;
1808         tp = args->trans;
1809         ip = args->dp;
1810         w = args->whichfork;
1811         ASSERT(w == XFS_DATA_FORK);
1812         mp = ip->i_mount;
1813         if (XFS_DIR_IS_V2(mp)) {
1814                 lastoff = mp->m_dirfreeblk;
1815                 error = xfs_bmap_last_before(tp, ip, &lastoff, w);
1816         } else
1817                 error = xfs_bmap_last_offset(tp, ip, &lastoff, w);
1818         if (error)
1819                 return error;
1820         if (unlikely(lastoff == 0)) {
1821                 XFS_ERROR_REPORT("xfs_da_swap_lastblock(1)", XFS_ERRLEVEL_LOW,
1822                                  mp);
1823                 return XFS_ERROR(EFSCORRUPTED);
1824         }
1825         /*
1826          * Read the last block in the btree space.
1827          */
1828         last_blkno = (xfs_dablk_t)lastoff - mp->m_dirblkfsbs;
1829         if ((error = xfs_da_read_buf(tp, ip, last_blkno, -1, &last_buf, w)))
1830                 return error;
1831         /*
1832          * Copy the last block into the dead buffer and log it.
1833          */
1834         memcpy(dead_buf->data, last_buf->data, mp->m_dirblksize);
1835         xfs_da_log_buf(tp, dead_buf, 0, mp->m_dirblksize - 1);
1836         dead_info = dead_buf->data;
1837         /*
1838          * Get values from the moved block.
1839          */
1840         if (INT_GET(dead_info->magic, ARCH_CONVERT) == XFS_DIR_LEAF_MAGIC) {
1841                 ASSERT(XFS_DIR_IS_V1(mp));
1842                 dead_leaf = (xfs_dir_leafblock_t *)dead_info;
1843                 dead_level = 0;
1844                 dead_hash =
1845                         INT_GET(dead_leaf->entries[INT_GET(dead_leaf->hdr.count, ARCH_CONVERT) - 1].hashval, ARCH_CONVERT);
1846         } else if (INT_GET(dead_info->magic, ARCH_CONVERT) == XFS_DIR2_LEAFN_MAGIC) {
1847                 ASSERT(XFS_DIR_IS_V2(mp));
1848                 dead_leaf2 = (xfs_dir2_leaf_t *)dead_info;
1849                 dead_level = 0;
1850                 dead_hash = INT_GET(dead_leaf2->ents[INT_GET(dead_leaf2->hdr.count, ARCH_CONVERT) - 1].hashval, ARCH_CONVERT);
1851         } else {
1852                 ASSERT(INT_GET(dead_info->magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC);
1853                 dead_node = (xfs_da_intnode_t *)dead_info;
1854                 dead_level = INT_GET(dead_node->hdr.level, ARCH_CONVERT);
1855                 dead_hash = INT_GET(dead_node->btree[INT_GET(dead_node->hdr.count, ARCH_CONVERT) - 1].hashval, ARCH_CONVERT);
1856         }
1857         sib_buf = par_buf = NULL;
1858         /*
1859          * If the moved block has a left sibling, fix up the pointers.
1860          */
1861         if ((sib_blkno = INT_GET(dead_info->back, ARCH_CONVERT))) {
1862                 if ((error = xfs_da_read_buf(tp, ip, sib_blkno, -1, &sib_buf, w)))
1863                         goto done;
1864                 sib_info = sib_buf->data;
1865                 if (unlikely(
1866                     INT_GET(sib_info->forw, ARCH_CONVERT) != last_blkno ||
1867                     INT_GET(sib_info->magic, ARCH_CONVERT) != INT_GET(dead_info->magic, ARCH_CONVERT))) {
1868                         XFS_ERROR_REPORT("xfs_da_swap_lastblock(2)",
1869                                          XFS_ERRLEVEL_LOW, mp);
1870                         error = XFS_ERROR(EFSCORRUPTED);
1871                         goto done;
1872                 }
1873                 INT_SET(sib_info->forw, ARCH_CONVERT, dead_blkno);
1874                 xfs_da_log_buf(tp, sib_buf,
1875                         XFS_DA_LOGRANGE(sib_info, &sib_info->forw,
1876                                         sizeof(sib_info->forw)));
1877                 xfs_da_buf_done(sib_buf);
1878                 sib_buf = NULL;
1879         }
1880         /*
1881          * If the moved block has a right sibling, fix up the pointers.
1882          */
1883         if ((sib_blkno = INT_GET(dead_info->forw, ARCH_CONVERT))) {
1884                 if ((error = xfs_da_read_buf(tp, ip, sib_blkno, -1, &sib_buf, w)))
1885                         goto done;
1886                 sib_info = sib_buf->data;
1887                 if (unlikely(
1888                        INT_GET(sib_info->back, ARCH_CONVERT) != last_blkno
1889                     || INT_GET(sib_info->magic, ARCH_CONVERT)
1890                                 != INT_GET(dead_info->magic, ARCH_CONVERT))) {
1891                         XFS_ERROR_REPORT("xfs_da_swap_lastblock(3)",
1892                                          XFS_ERRLEVEL_LOW, mp);
1893                         error = XFS_ERROR(EFSCORRUPTED);
1894                         goto done;
1895                 }
1896                 INT_SET(sib_info->back, ARCH_CONVERT, dead_blkno);
1897                 xfs_da_log_buf(tp, sib_buf,
1898                         XFS_DA_LOGRANGE(sib_info, &sib_info->back,
1899                                         sizeof(sib_info->back)));
1900                 xfs_da_buf_done(sib_buf);
1901                 sib_buf = NULL;
1902         }
1903         par_blkno = XFS_DIR_IS_V1(mp) ? 0 : mp->m_dirleafblk;
1904         level = -1;
1905         /*
1906          * Walk down the tree looking for the parent of the moved block.
1907          */
1908         for (;;) {
1909                 if ((error = xfs_da_read_buf(tp, ip, par_blkno, -1, &par_buf, w)))
1910                         goto done;
1911                 par_node = par_buf->data;
1912                 if (unlikely(
1913                     INT_GET(par_node->hdr.info.magic, ARCH_CONVERT) != XFS_DA_NODE_MAGIC ||
1914                     (level >= 0 && level != INT_GET(par_node->hdr.level, ARCH_CONVERT) + 1))) {
1915                         XFS_ERROR_REPORT("xfs_da_swap_lastblock(4)",
1916                                          XFS_ERRLEVEL_LOW, mp);
1917                         error = XFS_ERROR(EFSCORRUPTED);
1918                         goto done;
1919                 }
1920                 level = INT_GET(par_node->hdr.level, ARCH_CONVERT);
1921                 for (entno = 0;
1922                      entno < INT_GET(par_node->hdr.count, ARCH_CONVERT) &&
1923                      INT_GET(par_node->btree[entno].hashval, ARCH_CONVERT) < dead_hash;
1924                      entno++)
1925                         continue;
1926                 if (unlikely(entno == INT_GET(par_node->hdr.count, ARCH_CONVERT))) {
1927                         XFS_ERROR_REPORT("xfs_da_swap_lastblock(5)",
1928                                          XFS_ERRLEVEL_LOW, mp);
1929                         error = XFS_ERROR(EFSCORRUPTED);
1930                         goto done;
1931                 }
1932                 par_blkno = INT_GET(par_node->btree[entno].before, ARCH_CONVERT);
1933                 if (level == dead_level + 1)
1934                         break;
1935                 xfs_da_brelse(tp, par_buf);
1936                 par_buf = NULL;
1937         }
1938         /*
1939          * We're in the right parent block.
1940          * Look for the right entry.
1941          */
1942         for (;;) {
1943                 for (;
1944                      entno < INT_GET(par_node->hdr.count, ARCH_CONVERT) &&
1945                      INT_GET(par_node->btree[entno].before, ARCH_CONVERT) != last_blkno;
1946                      entno++)
1947                         continue;
1948                 if (entno < INT_GET(par_node->hdr.count, ARCH_CONVERT))
1949                         break;
1950                 par_blkno = INT_GET(par_node->hdr.info.forw, ARCH_CONVERT);
1951                 xfs_da_brelse(tp, par_buf);
1952                 par_buf = NULL;
1953                 if (unlikely(par_blkno == 0)) {
1954                         XFS_ERROR_REPORT("xfs_da_swap_lastblock(6)",
1955                                          XFS_ERRLEVEL_LOW, mp);
1956                         error = XFS_ERROR(EFSCORRUPTED);
1957                         goto done;
1958                 }
1959                 if ((error = xfs_da_read_buf(tp, ip, par_blkno, -1, &par_buf, w)))
1960                         goto done;
1961                 par_node = par_buf->data;
1962                 if (unlikely(
1963                     INT_GET(par_node->hdr.level, ARCH_CONVERT) != level ||
1964                     INT_GET(par_node->hdr.info.magic, ARCH_CONVERT) != XFS_DA_NODE_MAGIC)) {
1965                         XFS_ERROR_REPORT("xfs_da_swap_lastblock(7)",
1966                                          XFS_ERRLEVEL_LOW, mp);
1967                         error = XFS_ERROR(EFSCORRUPTED);
1968                         goto done;
1969                 }
1970                 entno = 0;
1971         }
1972         /*
1973          * Update the parent entry pointing to the moved block.
1974          */
1975         INT_SET(par_node->btree[entno].before, ARCH_CONVERT, dead_blkno);
1976         xfs_da_log_buf(tp, par_buf,
1977                 XFS_DA_LOGRANGE(par_node, &par_node->btree[entno].before,
1978                                 sizeof(par_node->btree[entno].before)));
1979         xfs_da_buf_done(par_buf);
1980         xfs_da_buf_done(dead_buf);
1981         *dead_blknop = last_blkno;
1982         *dead_bufp = last_buf;
1983         return 0;
1984 done:
1985         if (par_buf)
1986                 xfs_da_brelse(tp, par_buf);
1987         if (sib_buf)
1988                 xfs_da_brelse(tp, sib_buf);
1989         xfs_da_brelse(tp, last_buf);
1990         return error;
1991 }
1992
1993 /*
1994  * Remove a btree block from a directory or attribute.
1995  */
1996 int
1997 xfs_da_shrink_inode(xfs_da_args_t *args, xfs_dablk_t dead_blkno,
1998                     xfs_dabuf_t *dead_buf)
1999 {
2000         xfs_inode_t *dp;
2001         int done, error, w, count;
2002         xfs_fileoff_t bno;
2003         xfs_fsize_t size;
2004         xfs_trans_t *tp;
2005         xfs_mount_t *mp;
2006
2007         dp = args->dp;
2008         w = args->whichfork;
2009         tp = args->trans;
2010         mp = dp->i_mount;
2011         if (w == XFS_DATA_FORK && XFS_DIR_IS_V2(mp))
2012                 count = mp->m_dirblkfsbs;
2013         else
2014                 count = 1;
2015         for (;;) {
2016                 /*
2017                  * Remove extents.  If we get ENOSPC for a dir we have to move
2018                  * the last block to the place we want to kill.
2019                  */
2020                 if ((error = xfs_bunmapi(tp, dp, dead_blkno, count,
2021                                 XFS_BMAPI_AFLAG(w)|XFS_BMAPI_METADATA,
2022                                 0, args->firstblock, args->flist,
2023                                 &done)) == ENOSPC) {
2024                         if (w != XFS_DATA_FORK)
2025                                 goto done;
2026                         if ((error = xfs_da_swap_lastblock(args, &dead_blkno,
2027                                         &dead_buf)))
2028                                 goto done;
2029                 } else if (error)
2030                         goto done;
2031                 else
2032                         break;
2033         }
2034         ASSERT(done);
2035         xfs_da_binval(tp, dead_buf);
2036         /*
2037          * Adjust the directory size for version 1.
2038          */
2039         if (w == XFS_DATA_FORK && XFS_DIR_IS_V1(mp)) {
2040                 if ((error = xfs_bmap_last_offset(tp, dp, &bno, w)))
2041                         return error;
2042                 size = XFS_FSB_TO_B(dp->i_mount, bno);
2043                 if (size != dp->i_d.di_size) {
2044                         dp->i_d.di_size = size;
2045                         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2046                 }
2047         }
2048         return 0;
2049 done:
2050         xfs_da_binval(tp, dead_buf);
2051         return error;
2052 }
2053
2054 /*
2055  * See if the mapping(s) for this btree block are valid, i.e.
2056  * don't contain holes, are logically contiguous, and cover the whole range.
2057  */
2058 STATIC int
2059 xfs_da_map_covers_blocks(
2060         int             nmap,
2061         xfs_bmbt_irec_t *mapp,
2062         xfs_dablk_t     bno,
2063         int             count)
2064 {
2065         int             i;
2066         xfs_fileoff_t   off;
2067
2068         for (i = 0, off = bno; i < nmap; i++) {
2069                 if (mapp[i].br_startblock == HOLESTARTBLOCK ||
2070                     mapp[i].br_startblock == DELAYSTARTBLOCK) {
2071                         return 0;
2072                 }
2073                 if (off != mapp[i].br_startoff) {
2074                         return 0;
2075                 }
2076                 off += mapp[i].br_blockcount;
2077         }
2078         return off == bno + count;
2079 }
2080
2081 /*
2082  * Make a dabuf.
2083  * Used for get_buf, read_buf, read_bufr, and reada_buf.
2084  */
2085 STATIC int
2086 xfs_da_do_buf(
2087         xfs_trans_t     *trans,
2088         xfs_inode_t     *dp,
2089         xfs_dablk_t     bno,
2090         xfs_daddr_t     *mappedbnop,
2091         xfs_dabuf_t     **bpp,
2092         int             whichfork,
2093         int             caller,
2094         inst_t          *ra)
2095 {
2096         xfs_buf_t       *bp = NULL;
2097         xfs_buf_t       **bplist;
2098         int             error=0;
2099         int             i;
2100         xfs_bmbt_irec_t map;
2101         xfs_bmbt_irec_t *mapp;
2102         xfs_daddr_t     mappedbno;
2103         xfs_mount_t     *mp;
2104         int             nbplist=0;
2105         int             nfsb;
2106         int             nmap;
2107         xfs_dabuf_t     *rbp;
2108
2109         mp = dp->i_mount;
2110         if (whichfork == XFS_DATA_FORK && XFS_DIR_IS_V2(mp))
2111                 nfsb = mp->m_dirblkfsbs;
2112         else
2113                 nfsb = 1;
2114         mappedbno = *mappedbnop;
2115         /*
2116          * Caller doesn't have a mapping.  -2 means don't complain
2117          * if we land in a hole.
2118          */
2119         if (mappedbno == -1 || mappedbno == -2) {
2120                 /*
2121                  * Optimize the one-block case.
2122                  */
2123                 if (nfsb == 1) {
2124                         xfs_fsblock_t   fsb;
2125
2126                         if ((error =
2127                             xfs_bmapi_single(trans, dp, whichfork, &fsb,
2128                                     (xfs_fileoff_t)bno))) {
2129                                 return error;
2130                         }
2131                         mapp = &map;
2132                         if (fsb == NULLFSBLOCK) {
2133                                 nmap = 0;
2134                         } else {
2135                                 map.br_startblock = fsb;
2136                                 map.br_startoff = (xfs_fileoff_t)bno;
2137                                 map.br_blockcount = 1;
2138                                 nmap = 1;
2139                         }
2140                 } else {
2141                         mapp = kmem_alloc(sizeof(*mapp) * nfsb, KM_SLEEP);
2142                         nmap = nfsb;
2143                         if ((error = xfs_bmapi(trans, dp, (xfs_fileoff_t)bno,
2144                                         nfsb,
2145                                         XFS_BMAPI_METADATA |
2146                                                 XFS_BMAPI_AFLAG(whichfork),
2147                                         NULL, 0, mapp, &nmap, NULL)))
2148                                 goto exit0;
2149                 }
2150         } else {
2151                 map.br_startblock = XFS_DADDR_TO_FSB(mp, mappedbno);
2152                 map.br_startoff = (xfs_fileoff_t)bno;
2153                 map.br_blockcount = nfsb;
2154                 mapp = &map;
2155                 nmap = 1;
2156         }
2157         if (!xfs_da_map_covers_blocks(nmap, mapp, bno, nfsb)) {
2158                 error = mappedbno == -2 ? 0 : XFS_ERROR(EFSCORRUPTED);
2159                 if (unlikely(error == EFSCORRUPTED)) {
2160                         if (xfs_error_level >= XFS_ERRLEVEL_LOW) {
2161                                 int     i;
2162                                 cmn_err(CE_ALERT, "xfs_da_do_buf: bno %lld\n",
2163                                         (long long)bno);
2164                                 cmn_err(CE_ALERT, "dir: inode %lld\n",
2165                                         (long long)dp->i_ino);
2166                                 for (i = 0; i < nmap; i++) {
2167                                         cmn_err(CE_ALERT,
2168                                                 "[%02d] br_startoff %lld br_startblock %lld br_blockcount %lld br_state %d\n",
2169                                                 i,
2170                                                 (long long)mapp[i].br_startoff,
2171                                                 (long long)mapp[i].br_startblock,
2172                                                 (long long)mapp[i].br_blockcount,
2173                                                 mapp[i].br_state);
2174                                 }
2175                         }
2176                         XFS_ERROR_REPORT("xfs_da_do_buf(1)",
2177                                          XFS_ERRLEVEL_LOW, mp);
2178                 }
2179                 goto exit0;
2180         }
2181         if (caller != 3 && nmap > 1) {
2182                 bplist = kmem_alloc(sizeof(*bplist) * nmap, KM_SLEEP);
2183                 nbplist = 0;
2184         } else
2185                 bplist = NULL;
2186         /*
2187          * Turn the mapping(s) into buffer(s).
2188          */
2189         for (i = 0; i < nmap; i++) {
2190                 int     nmapped;
2191
2192                 mappedbno = XFS_FSB_TO_DADDR(mp, mapp[i].br_startblock);
2193                 if (i == 0)
2194                         *mappedbnop = mappedbno;
2195                 nmapped = (int)XFS_FSB_TO_BB(mp, mapp[i].br_blockcount);
2196                 switch (caller) {
2197                 case 0:
2198                         bp = xfs_trans_get_buf(trans, mp->m_ddev_targp,
2199                                 mappedbno, nmapped, 0);
2200                         error = bp ? XFS_BUF_GETERROR(bp) : XFS_ERROR(EIO);
2201                         break;
2202                 case 1:
2203 #ifndef __KERNEL__
2204                 case 2:
2205 #endif
2206                         bp = NULL;
2207                         error = xfs_trans_read_buf(mp, trans, mp->m_ddev_targp,
2208                                 mappedbno, nmapped, 0, &bp);
2209                         break;
2210 #ifdef __KERNEL__
2211                 case 3:
2212                         xfs_baread(mp->m_ddev_targp, mappedbno, nmapped);
2213                         error = 0;
2214                         bp = NULL;
2215                         break;
2216 #endif
2217                 }
2218                 if (error) {
2219                         if (bp)
2220                                 xfs_trans_brelse(trans, bp);
2221                         goto exit1;
2222                 }
2223                 if (!bp)
2224                         continue;
2225                 if (caller == 1) {
2226                         if (whichfork == XFS_ATTR_FORK) {
2227                                 XFS_BUF_SET_VTYPE_REF(bp, B_FS_ATTR_BTREE,
2228                                                 XFS_ATTR_BTREE_REF);
2229                         } else {
2230                                 XFS_BUF_SET_VTYPE_REF(bp, B_FS_DIR_BTREE,
2231                                                 XFS_DIR_BTREE_REF);
2232                         }
2233                 }
2234                 if (bplist) {
2235                         bplist[nbplist++] = bp;
2236                 }
2237         }
2238         /*
2239          * Build a dabuf structure.
2240          */
2241         if (bplist) {
2242                 rbp = xfs_da_buf_make(nbplist, bplist, ra);
2243         } else if (bp)
2244                 rbp = xfs_da_buf_make(1, &bp, ra);
2245         else
2246                 rbp = NULL;
2247         /*
2248          * For read_buf, check the magic number.
2249          */
2250         if (caller == 1) {
2251                 xfs_dir2_data_t         *data;
2252                 xfs_dir2_free_t         *free;
2253                 xfs_da_blkinfo_t        *info;
2254                 uint                    magic, magic1;
2255
2256                 info = rbp->data;
2257                 data = rbp->data;
2258                 free = rbp->data;
2259                 magic = INT_GET(info->magic, ARCH_CONVERT);
2260                 magic1 = INT_GET(data->hdr.magic, ARCH_CONVERT);
2261                 if (unlikely(
2262                     XFS_TEST_ERROR((magic != XFS_DA_NODE_MAGIC) &&
2263                                    (magic != XFS_DIR_LEAF_MAGIC) &&
2264                                    (magic != XFS_ATTR_LEAF_MAGIC) &&
2265                                    (magic != XFS_DIR2_LEAF1_MAGIC) &&
2266                                    (magic != XFS_DIR2_LEAFN_MAGIC) &&
2267                                    (magic1 != XFS_DIR2_BLOCK_MAGIC) &&
2268                                    (magic1 != XFS_DIR2_DATA_MAGIC) &&
2269                                    (INT_GET(free->hdr.magic, ARCH_CONVERT) != XFS_DIR2_FREE_MAGIC),
2270                                 mp, XFS_ERRTAG_DA_READ_BUF,
2271                                 XFS_RANDOM_DA_READ_BUF))) {
2272                         xfs_buftrace("DA READ ERROR", rbp->bps[0]);
2273                         XFS_CORRUPTION_ERROR("xfs_da_do_buf(2)",
2274                                              XFS_ERRLEVEL_LOW, mp, info);
2275                         error = XFS_ERROR(EFSCORRUPTED);
2276                         xfs_da_brelse(trans, rbp);
2277                         nbplist = 0;
2278                         goto exit1;
2279                 }
2280         }
2281         if (bplist) {
2282                 kmem_free(bplist, sizeof(*bplist) * nmap);
2283         }
2284         if (mapp != &map) {
2285                 kmem_free(mapp, sizeof(*mapp) * nfsb);
2286         }
2287         if (bpp)
2288                 *bpp = rbp;
2289         return 0;
2290 exit1:
2291         if (bplist) {
2292                 for (i = 0; i < nbplist; i++)
2293                         xfs_trans_brelse(trans, bplist[i]);
2294                 kmem_free(bplist, sizeof(*bplist) * nmap);
2295         }
2296 exit0:
2297         if (mapp != &map)
2298                 kmem_free(mapp, sizeof(*mapp) * nfsb);
2299         if (bpp)
2300                 *bpp = NULL;
2301         return error;
2302 }
2303
2304 /*
2305  * Get a buffer for the dir/attr block.
2306  */
2307 int
2308 xfs_da_get_buf(
2309         xfs_trans_t     *trans,
2310         xfs_inode_t     *dp,
2311         xfs_dablk_t     bno,
2312         xfs_daddr_t             mappedbno,
2313         xfs_dabuf_t     **bpp,
2314         int             whichfork)
2315 {
2316         return xfs_da_do_buf(trans, dp, bno, &mappedbno, bpp, whichfork, 0,
2317                                                  (inst_t *)__return_address);
2318 }
2319
2320 /*
2321  * Get a buffer for the dir/attr block, fill in the contents.
2322  */
2323 int
2324 xfs_da_read_buf(
2325         xfs_trans_t     *trans,
2326         xfs_inode_t     *dp,
2327         xfs_dablk_t     bno,
2328         xfs_daddr_t             mappedbno,
2329         xfs_dabuf_t     **bpp,
2330         int             whichfork)
2331 {
2332         return xfs_da_do_buf(trans, dp, bno, &mappedbno, bpp, whichfork, 1,
2333                 (inst_t *)__return_address);
2334 }
2335
2336 /*
2337  * Readahead the dir/attr block.
2338  */
2339 xfs_daddr_t
2340 xfs_da_reada_buf(
2341         xfs_trans_t     *trans,
2342         xfs_inode_t     *dp,
2343         xfs_dablk_t     bno,
2344         int             whichfork)
2345 {
2346         xfs_daddr_t             rval;
2347
2348         rval = -1;
2349         if (xfs_da_do_buf(trans, dp, bno, &rval, NULL, whichfork, 3,
2350                         (inst_t *)__return_address))
2351                 return -1;
2352         else
2353                 return rval;
2354 }
2355
2356 /*
2357  * Calculate the number of bits needed to hold i different values.
2358  */
2359 uint
2360 xfs_da_log2_roundup(uint i)
2361 {
2362         uint rval;
2363
2364         for (rval = 0; rval < NBBY * sizeof(i); rval++) {
2365                 if ((1 << rval) >= i)
2366                         break;
2367         }
2368         return(rval);
2369 }
2370
2371 kmem_zone_t *xfs_da_state_zone; /* anchor for state struct zone */
2372 kmem_zone_t *xfs_dabuf_zone;            /* dabuf zone */
2373
2374 /*
2375  * Allocate a dir-state structure.
2376  * We don't put them on the stack since they're large.
2377  */
2378 xfs_da_state_t *
2379 xfs_da_state_alloc(void)
2380 {
2381         return kmem_zone_zalloc(xfs_da_state_zone, KM_SLEEP);
2382 }
2383
2384 /*
2385  * Kill the altpath contents of a da-state structure.
2386  */
2387 STATIC void
2388 xfs_da_state_kill_altpath(xfs_da_state_t *state)
2389 {
2390         int     i;
2391
2392         for (i = 0; i < state->altpath.active; i++) {
2393                 if (state->altpath.blk[i].bp) {
2394                         if (state->altpath.blk[i].bp != state->path.blk[i].bp)
2395                                 xfs_da_buf_done(state->altpath.blk[i].bp);
2396                         state->altpath.blk[i].bp = NULL;
2397                 }
2398         }
2399         state->altpath.active = 0;
2400 }
2401
2402 /*
2403  * Free a da-state structure.
2404  */
2405 void
2406 xfs_da_state_free(xfs_da_state_t *state)
2407 {
2408         int     i;
2409
2410         xfs_da_state_kill_altpath(state);
2411         for (i = 0; i < state->path.active; i++) {
2412                 if (state->path.blk[i].bp)
2413                         xfs_da_buf_done(state->path.blk[i].bp);
2414         }
2415         if (state->extravalid && state->extrablk.bp)
2416                 xfs_da_buf_done(state->extrablk.bp);
2417 #ifdef DEBUG
2418         memset((char *)state, 0, sizeof(*state));
2419 #endif /* DEBUG */
2420         kmem_zone_free(xfs_da_state_zone, state);
2421 }
2422
2423 #ifdef XFS_DABUF_DEBUG
2424 xfs_dabuf_t     *xfs_dabuf_global_list;
2425 lock_t          xfs_dabuf_global_lock;
2426 #endif
2427
2428 /*
2429  * Create a dabuf.
2430  */
2431 /* ARGSUSED */
2432 STATIC xfs_dabuf_t *
2433 xfs_da_buf_make(int nbuf, xfs_buf_t **bps, inst_t *ra)
2434 {
2435         xfs_buf_t       *bp;
2436         xfs_dabuf_t     *dabuf;
2437         int             i;
2438         int             off;
2439
2440         if (nbuf == 1)
2441                 dabuf = kmem_zone_alloc(xfs_dabuf_zone, KM_SLEEP);
2442         else
2443                 dabuf = kmem_alloc(XFS_DA_BUF_SIZE(nbuf), KM_SLEEP);
2444         dabuf->dirty = 0;
2445 #ifdef XFS_DABUF_DEBUG
2446         dabuf->ra = ra;
2447         dabuf->target = XFS_BUF_TARGET(bps[0]);
2448         dabuf->blkno = XFS_BUF_ADDR(bps[0]);
2449 #endif
2450         if (nbuf == 1) {
2451                 dabuf->nbuf = 1;
2452                 bp = bps[0];
2453                 dabuf->bbcount = (short)BTOBB(XFS_BUF_COUNT(bp));
2454                 dabuf->data = XFS_BUF_PTR(bp);
2455                 dabuf->bps[0] = bp;
2456         } else {
2457                 dabuf->nbuf = nbuf;
2458                 for (i = 0, dabuf->bbcount = 0; i < nbuf; i++) {
2459                         dabuf->bps[i] = bp = bps[i];
2460                         dabuf->bbcount += BTOBB(XFS_BUF_COUNT(bp));
2461                 }
2462                 dabuf->data = kmem_alloc(BBTOB(dabuf->bbcount), KM_SLEEP);
2463                 for (i = off = 0; i < nbuf; i++, off += XFS_BUF_COUNT(bp)) {
2464                         bp = bps[i];
2465                         memcpy((char *)dabuf->data + off, XFS_BUF_PTR(bp),
2466                                 XFS_BUF_COUNT(bp));
2467                 }
2468         }
2469 #ifdef XFS_DABUF_DEBUG
2470         {
2471                 SPLDECL(s);
2472                 xfs_dabuf_t     *p;
2473
2474                 s = mutex_spinlock(&xfs_dabuf_global_lock);
2475                 for (p = xfs_dabuf_global_list; p; p = p->next) {
2476                         ASSERT(p->blkno != dabuf->blkno ||
2477                                p->target != dabuf->target);
2478                 }
2479                 dabuf->prev = NULL;
2480                 if (xfs_dabuf_global_list)
2481                         xfs_dabuf_global_list->prev = dabuf;
2482                 dabuf->next = xfs_dabuf_global_list;
2483                 xfs_dabuf_global_list = dabuf;
2484                 mutex_spinunlock(&xfs_dabuf_global_lock, s);
2485         }
2486 #endif
2487         return dabuf;
2488 }
2489
2490 /*
2491  * Un-dirty a dabuf.
2492  */
2493 STATIC void
2494 xfs_da_buf_clean(xfs_dabuf_t *dabuf)
2495 {
2496         xfs_buf_t       *bp;
2497         int             i;
2498         int             off;
2499
2500         if (dabuf->dirty) {
2501                 ASSERT(dabuf->nbuf > 1);
2502                 dabuf->dirty = 0;
2503                 for (i = off = 0; i < dabuf->nbuf;
2504                                 i++, off += XFS_BUF_COUNT(bp)) {
2505                         bp = dabuf->bps[i];
2506                         memcpy(XFS_BUF_PTR(bp), (char *)dabuf->data + off,
2507                                 XFS_BUF_COUNT(bp));
2508                 }
2509         }
2510 }
2511
2512 /*
2513  * Release a dabuf.
2514  */
2515 void
2516 xfs_da_buf_done(xfs_dabuf_t *dabuf)
2517 {
2518         ASSERT(dabuf);
2519         ASSERT(dabuf->nbuf && dabuf->data && dabuf->bbcount && dabuf->bps[0]);
2520         if (dabuf->dirty)
2521                 xfs_da_buf_clean(dabuf);
2522         if (dabuf->nbuf > 1)
2523                 kmem_free(dabuf->data, BBTOB(dabuf->bbcount));
2524 #ifdef XFS_DABUF_DEBUG
2525         {
2526                 SPLDECL(s);
2527
2528                 s = mutex_spinlock(&xfs_dabuf_global_lock);
2529                 if (dabuf->prev)
2530                         dabuf->prev->next = dabuf->next;
2531                 else
2532                         xfs_dabuf_global_list = dabuf->next;
2533                 if (dabuf->next)
2534                         dabuf->next->prev = dabuf->prev;
2535                 mutex_spinunlock(&xfs_dabuf_global_lock, s);
2536         }
2537         memset(dabuf, 0, XFS_DA_BUF_SIZE(dabuf->nbuf));
2538 #endif
2539         if (dabuf->nbuf == 1)
2540                 kmem_zone_free(xfs_dabuf_zone, dabuf);
2541         else
2542                 kmem_free(dabuf, XFS_DA_BUF_SIZE(dabuf->nbuf));
2543 }
2544
2545 /*
2546  * Log transaction from a dabuf.
2547  */
2548 void
2549 xfs_da_log_buf(xfs_trans_t *tp, xfs_dabuf_t *dabuf, uint first, uint last)
2550 {
2551         xfs_buf_t       *bp;
2552         uint            f;
2553         int             i;
2554         uint            l;
2555         int             off;
2556
2557         ASSERT(dabuf->nbuf && dabuf->data && dabuf->bbcount && dabuf->bps[0]);
2558         if (dabuf->nbuf == 1) {
2559                 ASSERT(dabuf->data == (void *)XFS_BUF_PTR(dabuf->bps[0]));
2560                 xfs_trans_log_buf(tp, dabuf->bps[0], first, last);
2561                 return;
2562         }
2563         dabuf->dirty = 1;
2564         ASSERT(first <= last);
2565         for (i = off = 0; i < dabuf->nbuf; i++, off += XFS_BUF_COUNT(bp)) {
2566                 bp = dabuf->bps[i];
2567                 f = off;
2568                 l = f + XFS_BUF_COUNT(bp) - 1;
2569                 if (f < first)
2570                         f = first;
2571                 if (l > last)
2572                         l = last;
2573                 if (f <= l)
2574                         xfs_trans_log_buf(tp, bp, f - off, l - off);
2575                 /*
2576                  * B_DONE is set by xfs_trans_log buf.
2577                  * If we don't set it on a new buffer (get not read)
2578                  * then if we don't put anything in the buffer it won't
2579                  * be set, and at commit it it released into the cache,
2580                  * and then a read will fail.
2581                  */
2582                 else if (!(XFS_BUF_ISDONE(bp)))
2583                   XFS_BUF_DONE(bp);
2584         }
2585         ASSERT(last < off);
2586 }
2587
2588 /*
2589  * Release dabuf from a transaction.
2590  * Have to free up the dabuf before the buffers are released,
2591  * since the synchronization on the dabuf is really the lock on the buffer.
2592  */
2593 void
2594 xfs_da_brelse(xfs_trans_t *tp, xfs_dabuf_t *dabuf)
2595 {
2596         xfs_buf_t       *bp;
2597         xfs_buf_t       **bplist;
2598         int             i;
2599         int             nbuf;
2600
2601         ASSERT(dabuf->nbuf && dabuf->data && dabuf->bbcount && dabuf->bps[0]);
2602         if ((nbuf = dabuf->nbuf) == 1) {
2603                 bplist = &bp;
2604                 bp = dabuf->bps[0];
2605         } else {
2606                 bplist = kmem_alloc(nbuf * sizeof(*bplist), KM_SLEEP);
2607                 memcpy(bplist, dabuf->bps, nbuf * sizeof(*bplist));
2608         }
2609         xfs_da_buf_done(dabuf);
2610         for (i = 0; i < nbuf; i++)
2611                 xfs_trans_brelse(tp, bplist[i]);
2612         if (bplist != &bp)
2613                 kmem_free(bplist, nbuf * sizeof(*bplist));
2614 }
2615
2616 /*
2617  * Invalidate dabuf from a transaction.
2618  */
2619 void
2620 xfs_da_binval(xfs_trans_t *tp, xfs_dabuf_t *dabuf)
2621 {
2622         xfs_buf_t       *bp;
2623         xfs_buf_t       **bplist;
2624         int             i;
2625         int             nbuf;
2626
2627         ASSERT(dabuf->nbuf && dabuf->data && dabuf->bbcount && dabuf->bps[0]);
2628         if ((nbuf = dabuf->nbuf) == 1) {
2629                 bplist = &bp;
2630                 bp = dabuf->bps[0];
2631         } else {
2632                 bplist = kmem_alloc(nbuf * sizeof(*bplist), KM_SLEEP);
2633                 memcpy(bplist, dabuf->bps, nbuf * sizeof(*bplist));
2634         }
2635         xfs_da_buf_done(dabuf);
2636         for (i = 0; i < nbuf; i++)
2637                 xfs_trans_binval(tp, bplist[i]);
2638         if (bplist != &bp)
2639                 kmem_free(bplist, nbuf * sizeof(*bplist));
2640 }
2641
2642 /*
2643  * Get the first daddr from a dabuf.
2644  */
2645 xfs_daddr_t
2646 xfs_da_blkno(xfs_dabuf_t *dabuf)
2647 {
2648         ASSERT(dabuf->nbuf);
2649         ASSERT(dabuf->data);
2650         return XFS_BUF_ADDR(dabuf->bps[0]);
2651 }