xfs_da_btree.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   4 * Copyright (c) 2013 Red Hat, Inc.
   5 * All Rights Reserved.
   6 */
   7#include "xfs.h"
   8#include "xfs_fs.h"
   9#include "xfs_shared.h"
  10#include "xfs_format.h"
  11#include "xfs_log_format.h"
  12#include "xfs_trans_resv.h"
  13#include "xfs_bit.h"
  14#include "xfs_mount.h"
  15#include "xfs_inode.h"
  16#include "xfs_dir2.h"
  17#include "xfs_dir2_priv.h"
  18#include "xfs_trans.h"
  19#include "xfs_bmap.h"
  20#include "xfs_attr_leaf.h"
  21#include "xfs_error.h"
  22#include "xfs_trace.h"
  23#include "xfs_buf_item.h"
  24#include "xfs_log.h"
  25#include "xfs_errortag.h"
  26#include "xfs_health.h"
  27
  28/*
  29 * xfs_da_btree.c
  30 *
  31 * Routines to implement directories as Btrees of hashed names.
  32 */
  33
  34/*========================================================================
  35 * Function prototypes for the kernel.
  36 *========================================================================*/
  37
  38/*
  39 * Routines used for growing the Btree.
  40 */
  41STATIC int xfs_da3_root_split(xfs_da_state_t *state,
  42					    xfs_da_state_blk_t *existing_root,
  43					    xfs_da_state_blk_t *new_child);
  44STATIC int xfs_da3_node_split(xfs_da_state_t *state,
  45					    xfs_da_state_blk_t *existing_blk,
  46					    xfs_da_state_blk_t *split_blk,
  47					    xfs_da_state_blk_t *blk_to_add,
  48					    int treelevel,
  49					    int *result);
  50STATIC void xfs_da3_node_rebalance(xfs_da_state_t *state,
  51					 xfs_da_state_blk_t *node_blk_1,
  52					 xfs_da_state_blk_t *node_blk_2);
  53STATIC void xfs_da3_node_add(xfs_da_state_t *state,
  54				   xfs_da_state_blk_t *old_node_blk,
  55				   xfs_da_state_blk_t *new_node_blk);
  56
  57/*
  58 * Routines used for shrinking the Btree.
  59 */
  60STATIC int xfs_da3_root_join(xfs_da_state_t *state,
  61					   xfs_da_state_blk_t *root_blk);
  62STATIC int xfs_da3_node_toosmall(xfs_da_state_t *state, int *retval);
  63STATIC void xfs_da3_node_remove(xfs_da_state_t *state,
  64					      xfs_da_state_blk_t *drop_blk);
  65STATIC void xfs_da3_node_unbalance(xfs_da_state_t *state,
  66					 xfs_da_state_blk_t *src_node_blk,
  67					 xfs_da_state_blk_t *dst_node_blk);
  68
  69/*
  70 * Utility routines.
  71 */
  72STATIC int	xfs_da3_blk_unlink(xfs_da_state_t *state,
  73				  xfs_da_state_blk_t *drop_blk,
  74				  xfs_da_state_blk_t *save_blk);
  75
  76
  77struct kmem_cache	*xfs_da_state_cache;	/* anchor for dir/attr state */
  78
  79/*
  80 * Allocate a dir-state structure.
  81 * We don't put them on the stack since they're large.
  82 */
  83struct xfs_da_state *
  84xfs_da_state_alloc(
  85	struct xfs_da_args	*args)
  86{
  87	struct xfs_da_state	*state;
  88
  89	state = kmem_cache_zalloc(xfs_da_state_cache,
  90			GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOFAIL);
  91	state->args = args;
  92	state->mp = args->dp->i_mount;
  93	return state;
  94}
  95
  96/*
  97 * Kill the altpath contents of a da-state structure.
  98 */
  99STATIC void
 100xfs_da_state_kill_altpath(xfs_da_state_t *state)
 101{
 102	int	i;
 103
 104	for (i = 0; i < state->altpath.active; i++)
 105		state->altpath.blk[i].bp = NULL;
 106	state->altpath.active = 0;
 107}
 108
 109/*
 110 * Free a da-state structure.
 111 */
 112void
 113xfs_da_state_free(xfs_da_state_t *state)
 114{
 115	xfs_da_state_kill_altpath(state);
 116#ifdef DEBUG
 117	memset((char *)state, 0, sizeof(*state));
 118#endif /* DEBUG */
 119	kmem_cache_free(xfs_da_state_cache, state);
 120}
 121
 122void
 123xfs_da_state_reset(
 124	struct xfs_da_state	*state,
 125	struct xfs_da_args	*args)
 126{
 127	xfs_da_state_kill_altpath(state);
 128	memset(state, 0, sizeof(struct xfs_da_state));
 129	state->args = args;
 130	state->mp = state->args->dp->i_mount;
 131}
 132
 133static inline int xfs_dabuf_nfsb(struct xfs_mount *mp, int whichfork)
 134{
 135	if (whichfork == XFS_DATA_FORK)
 136		return mp->m_dir_geo->fsbcount;
 137	return mp->m_attr_geo->fsbcount;
 138}
 139
 140void
 141xfs_da3_node_hdr_from_disk(
 142	struct xfs_mount		*mp,
 143	struct xfs_da3_icnode_hdr	*to,
 144	struct xfs_da_intnode		*from)
 145{
 146	if (xfs_has_crc(mp)) {
 147		struct xfs_da3_intnode	*from3 = (struct xfs_da3_intnode *)from;
 148
 149		to->forw = be32_to_cpu(from3->hdr.info.hdr.forw);
 150		to->back = be32_to_cpu(from3->hdr.info.hdr.back);
 151		to->magic = be16_to_cpu(from3->hdr.info.hdr.magic);
 152		to->count = be16_to_cpu(from3->hdr.__count);
 153		to->level = be16_to_cpu(from3->hdr.__level);
 154		to->btree = from3->__btree;
 155		ASSERT(to->magic == XFS_DA3_NODE_MAGIC);
 156	} else {
 157		to->forw = be32_to_cpu(from->hdr.info.forw);
 158		to->back = be32_to_cpu(from->hdr.info.back);
 159		to->magic = be16_to_cpu(from->hdr.info.magic);
 160		to->count = be16_to_cpu(from->hdr.__count);
 161		to->level = be16_to_cpu(from->hdr.__level);
 162		to->btree = from->__btree;
 163		ASSERT(to->magic == XFS_DA_NODE_MAGIC);
 164	}
 165}
 166
 167void
 168xfs_da3_node_hdr_to_disk(
 169	struct xfs_mount		*mp,
 170	struct xfs_da_intnode		*to,
 171	struct xfs_da3_icnode_hdr	*from)
 172{
 173	if (xfs_has_crc(mp)) {
 174		struct xfs_da3_intnode	*to3 = (struct xfs_da3_intnode *)to;
 175
 176		ASSERT(from->magic == XFS_DA3_NODE_MAGIC);
 177		to3->hdr.info.hdr.forw = cpu_to_be32(from->forw);
 178		to3->hdr.info.hdr.back = cpu_to_be32(from->back);
 179		to3->hdr.info.hdr.magic = cpu_to_be16(from->magic);
 180		to3->hdr.__count = cpu_to_be16(from->count);
 181		to3->hdr.__level = cpu_to_be16(from->level);
 182	} else {
 183		ASSERT(from->magic == XFS_DA_NODE_MAGIC);
 184		to->hdr.info.forw = cpu_to_be32(from->forw);
 185		to->hdr.info.back = cpu_to_be32(from->back);
 186		to->hdr.info.magic = cpu_to_be16(from->magic);
 187		to->hdr.__count = cpu_to_be16(from->count);
 188		to->hdr.__level = cpu_to_be16(from->level);
 189	}
 190}
 191
 192/*
 193 * Verify an xfs_da3_blkinfo structure. Note that the da3 fields are only
 194 * accessible on v5 filesystems. This header format is common across da node,
 195 * attr leaf and dir leaf blocks.
 196 */
 197xfs_failaddr_t
 198xfs_da3_blkinfo_verify(
 199	struct xfs_buf		*bp,
 200	struct xfs_da3_blkinfo	*hdr3)
 201{
 202	struct xfs_mount	*mp = bp->b_mount;
 203	struct xfs_da_blkinfo	*hdr = &hdr3->hdr;
 204
 205	if (!xfs_verify_magic16(bp, hdr->magic))
 206		return __this_address;
 207
 208	if (xfs_has_crc(mp)) {
 209		if (!uuid_equal(&hdr3->uuid, &mp->m_sb.sb_meta_uuid))
 210			return __this_address;
 211		if (be64_to_cpu(hdr3->blkno) != xfs_buf_daddr(bp))
 212			return __this_address;
 213		if (!xfs_log_check_lsn(mp, be64_to_cpu(hdr3->lsn)))
 214			return __this_address;
 215	}
 216
 217	return NULL;
 218}
 219
 220static xfs_failaddr_t
 221xfs_da3_node_verify(
 222	struct xfs_buf		*bp)
 223{
 224	struct xfs_mount	*mp = bp->b_mount;
 225	struct xfs_da_intnode	*hdr = bp->b_addr;
 226	struct xfs_da3_icnode_hdr ichdr;
 227	xfs_failaddr_t		fa;
 228
 229	xfs_da3_node_hdr_from_disk(mp, &ichdr, hdr);
 230
 231	fa = xfs_da3_blkinfo_verify(bp, bp->b_addr);
 232	if (fa)
 233		return fa;
 234
 235	if (ichdr.level == 0)
 236		return __this_address;
 237	if (ichdr.level > XFS_DA_NODE_MAXDEPTH)
 238		return __this_address;
 239	if (ichdr.count == 0)
 240		return __this_address;
 241
 242	/*
 243	 * we don't know if the node is for and attribute or directory tree,
 244	 * so only fail if the count is outside both bounds
 245	 */
 246	if (ichdr.count > mp->m_dir_geo->node_ents &&
 247	    ichdr.count > mp->m_attr_geo->node_ents)
 248		return __this_address;
 249
 250	/* XXX: hash order check? */
 251
 252	return NULL;
 253}
 254
 255xfs_failaddr_t
 256xfs_da3_node_header_check(
 257	struct xfs_buf		*bp,
 258	xfs_ino_t		owner)
 259{
 260	struct xfs_mount	*mp = bp->b_mount;
 261
 262	if (xfs_has_crc(mp)) {
 263		struct xfs_da3_blkinfo *hdr3 = bp->b_addr;
 264
 265		if (hdr3->hdr.magic != cpu_to_be16(XFS_DA3_NODE_MAGIC))
 266			return __this_address;
 267
 268		if (be64_to_cpu(hdr3->owner) != owner)
 269			return __this_address;
 270	}
 271
 272	return NULL;
 273}
 274
 275xfs_failaddr_t
 276xfs_da3_header_check(
 277	struct xfs_buf		*bp,
 278	xfs_ino_t		owner)
 279{
 280	struct xfs_mount	*mp = bp->b_mount;
 281	struct xfs_da_blkinfo	*hdr = bp->b_addr;
 282
 283	if (!xfs_has_crc(mp))
 284		return NULL;
 285
 286	switch (hdr->magic) {
 287	case cpu_to_be16(XFS_ATTR3_LEAF_MAGIC):
 288		return xfs_attr3_leaf_header_check(bp, owner);
 289	case cpu_to_be16(XFS_DA3_NODE_MAGIC):
 290		return xfs_da3_node_header_check(bp, owner);
 291	case cpu_to_be16(XFS_DIR3_LEAF1_MAGIC):
 292	case cpu_to_be16(XFS_DIR3_LEAFN_MAGIC):
 293		return xfs_dir3_leaf_header_check(bp, owner);
 294	}
 295
 296	ASSERT(0);
 297	return NULL;
 298}
 299
 300static void
 301xfs_da3_node_write_verify(
 302	struct xfs_buf	*bp)
 303{
 304	struct xfs_mount	*mp = bp->b_mount;
 305	struct xfs_buf_log_item	*bip = bp->b_log_item;
 306	struct xfs_da3_node_hdr *hdr3 = bp->b_addr;
 307	xfs_failaddr_t		fa;
 308
 309	fa = xfs_da3_node_verify(bp);
 310	if (fa) {
 311		xfs_verifier_error(bp, -EFSCORRUPTED, fa);
 312		return;
 313	}
 314
 315	if (!xfs_has_crc(mp))
 316		return;
 317
 318	if (bip)
 319		hdr3->info.lsn = cpu_to_be64(bip->bli_item.li_lsn);
 320
 321	xfs_buf_update_cksum(bp, XFS_DA3_NODE_CRC_OFF);
 322}
 323
 324/*
 325 * leaf/node format detection on trees is sketchy, so a node read can be done on
 326 * leaf level blocks when detection identifies the tree as a node format tree
 327 * incorrectly. In this case, we need to swap the verifier to match the correct
 328 * format of the block being read.
 329 */
 330static void
 331xfs_da3_node_read_verify(
 332	struct xfs_buf		*bp)
 333{
 334	struct xfs_da_blkinfo	*info = bp->b_addr;
 335	xfs_failaddr_t		fa;
 336
 337	switch (be16_to_cpu(info->magic)) {
 338		case XFS_DA3_NODE_MAGIC:
 339			if (!xfs_buf_verify_cksum(bp, XFS_DA3_NODE_CRC_OFF)) {
 340				xfs_verifier_error(bp, -EFSBADCRC,
 341						__this_address);
 342				break;
 343			}
 344			fallthrough;
 345		case XFS_DA_NODE_MAGIC:
 346			fa = xfs_da3_node_verify(bp);
 347			if (fa)
 348				xfs_verifier_error(bp, -EFSCORRUPTED, fa);
 349			return;
 350		case XFS_ATTR_LEAF_MAGIC:
 351		case XFS_ATTR3_LEAF_MAGIC:
 352			bp->b_ops = &xfs_attr3_leaf_buf_ops;
 353			bp->b_ops->verify_read(bp);
 354			return;
 355		case XFS_DIR2_LEAFN_MAGIC:
 356		case XFS_DIR3_LEAFN_MAGIC:
 357			bp->b_ops = &xfs_dir3_leafn_buf_ops;
 358			bp->b_ops->verify_read(bp);
 359			return;
 360		default:
 361			xfs_verifier_error(bp, -EFSCORRUPTED, __this_address);
 362			break;
 363	}
 364}
 365
 366/* Verify the structure of a da3 block. */
 367static xfs_failaddr_t
 368xfs_da3_node_verify_struct(
 369	struct xfs_buf		*bp)
 370{
 371	struct xfs_da_blkinfo	*info = bp->b_addr;
 372
 373	switch (be16_to_cpu(info->magic)) {
 374	case XFS_DA3_NODE_MAGIC:
 375	case XFS_DA_NODE_MAGIC:
 376		return xfs_da3_node_verify(bp);
 377	case XFS_ATTR_LEAF_MAGIC:
 378	case XFS_ATTR3_LEAF_MAGIC:
 379		bp->b_ops = &xfs_attr3_leaf_buf_ops;
 380		return bp->b_ops->verify_struct(bp);
 381	case XFS_DIR2_LEAFN_MAGIC:
 382	case XFS_DIR3_LEAFN_MAGIC:
 383		bp->b_ops = &xfs_dir3_leafn_buf_ops;
 384		return bp->b_ops->verify_struct(bp);
 385	default:
 386		return __this_address;
 387	}
 388}
 389
 390const struct xfs_buf_ops xfs_da3_node_buf_ops = {
 391	.name = "xfs_da3_node",
 392	.magic16 = { cpu_to_be16(XFS_DA_NODE_MAGIC),
 393		     cpu_to_be16(XFS_DA3_NODE_MAGIC) },
 394	.verify_read = xfs_da3_node_read_verify,
 395	.verify_write = xfs_da3_node_write_verify,
 396	.verify_struct = xfs_da3_node_verify_struct,
 397};
 398
 399static int
 400xfs_da3_node_set_type(
 401	struct xfs_trans	*tp,
 402	struct xfs_inode	*dp,
 403	int			whichfork,
 404	struct xfs_buf		*bp)
 405{
 406	struct xfs_da_blkinfo	*info = bp->b_addr;
 407
 408	switch (be16_to_cpu(info->magic)) {
 409	case XFS_DA_NODE_MAGIC:
 410	case XFS_DA3_NODE_MAGIC:
 411		xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DA_NODE_BUF);
 412		return 0;
 413	case XFS_ATTR_LEAF_MAGIC:
 414	case XFS_ATTR3_LEAF_MAGIC:
 415		xfs_trans_buf_set_type(tp, bp, XFS_BLFT_ATTR_LEAF_BUF);
 416		return 0;
 417	case XFS_DIR2_LEAFN_MAGIC:
 418	case XFS_DIR3_LEAFN_MAGIC:
 419		xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DIR_LEAFN_BUF);
 420		return 0;
 421	default:
 422		XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, tp->t_mountp,
 423				info, sizeof(*info));
 424		xfs_trans_brelse(tp, bp);
 425		xfs_dirattr_mark_sick(dp, whichfork);
 426		return -EFSCORRUPTED;
 427	}
 428}
 429
 430int
 431xfs_da3_node_read(
 432	struct xfs_trans	*tp,
 433	struct xfs_inode	*dp,
 434	xfs_dablk_t		bno,
 435	struct xfs_buf		**bpp,
 436	int			whichfork)
 437{
 438	int			error;
 439
 440	error = xfs_da_read_buf(tp, dp, bno, 0, bpp, whichfork,
 441			&xfs_da3_node_buf_ops);
 442	if (error || !*bpp || !tp)
 443		return error;
 444	return xfs_da3_node_set_type(tp, dp, whichfork, *bpp);
 445}
 446
 447int
 448xfs_da3_node_read_mapped(
 449	struct xfs_trans	*tp,
 450	struct xfs_inode	*dp,
 451	xfs_daddr_t		mappedbno,
 452	struct xfs_buf		**bpp,
 453	int			whichfork)
 454{
 455	struct xfs_mount	*mp = dp->i_mount;
 456	int			error;
 457
 458	error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp, mappedbno,
 459			XFS_FSB_TO_BB(mp, xfs_dabuf_nfsb(mp, whichfork)), 0,
 460			bpp, &xfs_da3_node_buf_ops);
 461	if (xfs_metadata_is_sick(error))
 462		xfs_dirattr_mark_sick(dp, whichfork);
 463	if (error || !*bpp)
 464		return error;
 465
 466	if (whichfork == XFS_ATTR_FORK)
 467		xfs_buf_set_ref(*bpp, XFS_ATTR_BTREE_REF);
 468	else
 469		xfs_buf_set_ref(*bpp, XFS_DIR_BTREE_REF);
 470
 471	if (!tp)
 472		return 0;
 473	return xfs_da3_node_set_type(tp, dp, whichfork, *bpp);
 474}
 475
 476/*
 477 * Copy src directory/attr leaf/node buffer to the dst.
 478 * For v5 file systems make sure the right blkno is stamped in.
 479 */
 480void
 481xfs_da_buf_copy(
 482	struct xfs_buf *dst,
 483	struct xfs_buf *src,
 484	size_t size)
 485{
 486	struct xfs_da3_blkinfo *da3 = dst->b_addr;
 487
 488	memcpy(dst->b_addr, src->b_addr, size);
 489	dst->b_ops = src->b_ops;
 490	xfs_trans_buf_copy_type(dst, src);
 491	if (xfs_has_crc(dst->b_mount))
 492		da3->blkno = cpu_to_be64(xfs_buf_daddr(dst));
 493}
 494
 495/*========================================================================
 496 * Routines used for growing the Btree.
 497 *========================================================================*/
 498
 499/*
 500 * Create the initial contents of an intermediate node.
 501 */
 502int
 503xfs_da3_node_create(
 504	struct xfs_da_args	*args,
 505	xfs_dablk_t		blkno,
 506	int			level,
 507	struct xfs_buf		**bpp,
 508	int			whichfork)
 509{
 510	struct xfs_da_intnode	*node;
 511	struct xfs_trans	*tp = args->trans;
 512	struct xfs_mount	*mp = tp->t_mountp;
 513	struct xfs_da3_icnode_hdr ichdr = {0};
 514	struct xfs_buf		*bp;
 515	int			error;
 516	struct xfs_inode	*dp = args->dp;
 517
 518	trace_xfs_da_node_create(args);
 519	ASSERT(level <= XFS_DA_NODE_MAXDEPTH);
 520
 521	error = xfs_da_get_buf(tp, dp, blkno, &bp, whichfork);
 522	if (error)
 523		return error;
 524	bp->b_ops = &xfs_da3_node_buf_ops;
 525	xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DA_NODE_BUF);
 526	node = bp->b_addr;
 527
 528	if (xfs_has_crc(mp)) {
 529		struct xfs_da3_node_hdr *hdr3 = bp->b_addr;
 530
 531		memset(hdr3, 0, sizeof(struct xfs_da3_node_hdr));
 532		ichdr.magic = XFS_DA3_NODE_MAGIC;
 533		hdr3->info.blkno = cpu_to_be64(xfs_buf_daddr(bp));
 534		hdr3->info.owner = cpu_to_be64(args->owner);
 535		uuid_copy(&hdr3->info.uuid, &mp->m_sb.sb_meta_uuid);
 536	} else {
 537		ichdr.magic = XFS_DA_NODE_MAGIC;
 538	}
 539	ichdr.level = level;
 540
 541	xfs_da3_node_hdr_to_disk(dp->i_mount, node, &ichdr);
 542	xfs_trans_log_buf(tp, bp,
 543		XFS_DA_LOGRANGE(node, &node->hdr, args->geo->node_hdr_size));
 544
 545	*bpp = bp;
 546	return 0;
 547}
 548
 549/*
 550 * Split a leaf node, rebalance, then possibly split
 551 * intermediate nodes, rebalance, etc.
 552 */
 553int							/* error */
 554xfs_da3_split(
 555	struct xfs_da_state	*state)
 556{
 557	struct xfs_da_state_blk	*oldblk;
 558	struct xfs_da_state_blk	*newblk;
 559	struct xfs_da_state_blk	*addblk;
 560	struct xfs_da_intnode	*node;
 561	int			max;
 562	int			action = 0;
 563	int			error;
 564	int			i;
 565
 566	trace_xfs_da_split(state->args);
 567
 568	if (XFS_TEST_ERROR(state->mp, XFS_ERRTAG_DA_LEAF_SPLIT))
 569		return -EIO;
 570
 571	/*
 572	 * Walk back up the tree splitting/inserting/adjusting as necessary.
 573	 * If we need to insert and there isn't room, split the node, then
 574	 * decide which fragment to insert the new block from below into.
 575	 * Note that we may split the root this way, but we need more fixup.
 576	 */
 577	max = state->path.active - 1;
 578	ASSERT((max >= 0) && (max < XFS_DA_NODE_MAXDEPTH));
 579	ASSERT(state->path.blk[max].magic == XFS_ATTR_LEAF_MAGIC ||
 580	       state->path.blk[max].magic == XFS_DIR2_LEAFN_MAGIC);
 581
 582	addblk = &state->path.blk[max];		/* initial dummy value */
 583	for (i = max; (i >= 0) && addblk; state->path.active--, i--) {
 584		oldblk = &state->path.blk[i];
 585		newblk = &state->altpath.blk[i];
 586
 587		/*
 588		 * If a leaf node then
 589		 *     Allocate a new leaf node, then rebalance across them.
 590		 * else if an intermediate node then
 591		 *     We split on the last layer, must we split the node?
 592		 */
 593		switch (oldblk->magic) {
 594		case XFS_ATTR_LEAF_MAGIC:
 595			error = xfs_attr3_leaf_split(state, oldblk, newblk);
 596			if (error < 0)
 597				return error;	/* GROT: attr is inconsistent */
 598			if (!error) {
 599				addblk = newblk;
 600				break;
 601			}
 602			/*
 603			 * Entry wouldn't fit, split the leaf again. The new
 604			 * extrablk will be consumed by xfs_da3_node_split if
 605			 * the node is split.
 606			 */
 607			state->extravalid = 1;
 608			if (state->inleaf) {
 609				state->extraafter = 0;	/* before newblk */
 610				trace_xfs_attr_leaf_split_before(state->args);
 611				error = xfs_attr3_leaf_split(state, oldblk,
 612							    &state->extrablk);
 613			} else {
 614				state->extraafter = 1;	/* after newblk */
 615				trace_xfs_attr_leaf_split_after(state->args);
 616				error = xfs_attr3_leaf_split(state, newblk,
 617							    &state->extrablk);
 618			}
 619			if (error == 1)
 620				return -ENOSPC;
 621			if (error)
 622				return error;	/* GROT: attr inconsistent */
 623			addblk = newblk;
 624			break;
 625		case XFS_DIR2_LEAFN_MAGIC:
 626			error = xfs_dir2_leafn_split(state, oldblk, newblk);
 627			if (error)
 628				return error;
 629			addblk = newblk;
 630			break;
 631		case XFS_DA_NODE_MAGIC:
 632			error = xfs_da3_node_split(state, oldblk, newblk, addblk,
 633							 max - i, &action);
 634			addblk->bp = NULL;
 635			if (error)
 636				return error;	/* GROT: dir is inconsistent */
 637			/*
 638			 * Record the newly split block for the next time thru?
 639			 */
 640			if (action)
 641				addblk = newblk;
 642			else
 643				addblk = NULL;
 644			break;
 645		}
 646
 647		/*
 648		 * Update the btree to show the new hashval for this child.
 649		 */
 650		xfs_da3_fixhashpath(state, &state->path);
 651	}
 652	if (!addblk)
 653		return 0;
 654
 655	/*
 656	 * xfs_da3_node_split() should have consumed any extra blocks we added
 657	 * during a double leaf split in the attr fork. This is guaranteed as
 658	 * we can't be here if the attr fork only has a single leaf block.
 659	 */
 660	ASSERT(state->extravalid == 0 ||
 661	       state->path.blk[max].magic == XFS_DIR2_LEAFN_MAGIC);
 662
 663	/*
 664	 * Split the root node.
 665	 */
 666	ASSERT(state->path.active == 0);
 667	oldblk = &state->path.blk[0];
 668	error = xfs_da3_root_split(state, oldblk, addblk);
 669	if (error)
 670		goto out;
 671
 672	/*
 673	 * Update pointers to the node which used to be block 0 and just got
 674	 * bumped because of the addition of a new root node.  Note that the
 675	 * original block 0 could be at any position in the list of blocks in
 676	 * the tree.
 677	 *
 678	 * Note: the magic numbers and sibling pointers are in the same physical
 679	 * place for both v2 and v3 headers (by design). Hence it doesn't matter
 680	 * which version of the xfs_da_intnode structure we use here as the
 681	 * result will be the same using either structure.
 682	 */
 683	node = oldblk->bp->b_addr;
 684	if (node->hdr.info.forw) {
 685		if (be32_to_cpu(node->hdr.info.forw) != addblk->blkno) {
 686			xfs_buf_mark_corrupt(oldblk->bp);
 687			xfs_da_mark_sick(state->args);
 688			error = -EFSCORRUPTED;
 689			goto out;
 690		}
 691		node = addblk->bp->b_addr;
 692		node->hdr.info.back = cpu_to_be32(oldblk->blkno);
 693		xfs_trans_log_buf(state->args->trans, addblk->bp,
 694				  XFS_DA_LOGRANGE(node, &node->hdr.info,
 695				  sizeof(node->hdr.info)));
 696	}
 697	node = oldblk->bp->b_addr;
 698	if (node->hdr.info.back) {
 699		if (be32_to_cpu(node->hdr.info.back) != addblk->blkno) {
 700			xfs_buf_mark_corrupt(oldblk->bp);
 701			xfs_da_mark_sick(state->args);
 702			error = -EFSCORRUPTED;
 703			goto out;
 704		}
 705		node = addblk->bp->b_addr;
 706		node->hdr.info.forw = cpu_to_be32(oldblk->blkno);
 707		xfs_trans_log_buf(state->args->trans, addblk->bp,
 708				  XFS_DA_LOGRANGE(node, &node->hdr.info,
 709				  sizeof(node->hdr.info)));
 710	}
 711out:
 712	addblk->bp = NULL;
 713	return error;
 714}
 715
 716/*
 717 * Split the root.  We have to create a new root and point to the two
 718 * parts (the split old root) that we just created.  Copy block zero to
 719 * the EOF, extending the inode in process.
 720 */
 721STATIC int						/* error */
 722xfs_da3_root_split(
 723	struct xfs_da_state	*state,
 724	struct xfs_da_state_blk	*blk1,
 725	struct xfs_da_state_blk	*blk2)
 726{
 727	struct xfs_da_intnode	*node;
 728	struct xfs_da_intnode	*oldroot;
 729	struct xfs_da_node_entry *btree;
 730	struct xfs_da3_icnode_hdr nodehdr;
 731	struct xfs_da_args	*args;
 732	struct xfs_buf		*bp;
 733	struct xfs_inode	*dp;
 734	struct xfs_trans	*tp;
 735	struct xfs_dir2_leaf	*leaf;
 736	xfs_dablk_t		blkno;
 737	int			level;
 738	int			error;
 739	int			size;
 740
 741	trace_xfs_da_root_split(state->args);
 742
 743	/*
 744	 * Copy the existing (incorrect) block from the root node position
 745	 * to a free space somewhere.
 746	 */
 747	args = state->args;
 748	error = xfs_da_grow_inode(args, &blkno);
 749	if (error)
 750		return error;
 751
 752	dp = args->dp;
 753	tp = args->trans;
 754	error = xfs_da_get_buf(tp, dp, blkno, &bp, args->whichfork);
 755	if (error)
 756		return error;
 757	node = bp->b_addr;
 758	oldroot = blk1->bp->b_addr;
 759	if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
 760	    oldroot->hdr.info.magic == cpu_to_be16(XFS_DA3_NODE_MAGIC)) {
 761		struct xfs_da3_icnode_hdr icnodehdr;
 762
 763		xfs_da3_node_hdr_from_disk(dp->i_mount, &icnodehdr, oldroot);
 764		btree = icnodehdr.btree;
 765		size = (int)((char *)&btree[icnodehdr.count] - (char *)oldroot);
 766		level = icnodehdr.level;
 767	} else {
 768		struct xfs_dir3_icleaf_hdr leafhdr;
 769
 770		leaf = (xfs_dir2_leaf_t *)oldroot;
 771		xfs_dir2_leaf_hdr_from_disk(dp->i_mount, &leafhdr, leaf);
 772
 773		ASSERT(leafhdr.magic == XFS_DIR2_LEAFN_MAGIC ||
 774		       leafhdr.magic == XFS_DIR3_LEAFN_MAGIC);
 775		size = (int)((char *)&leafhdr.ents[leafhdr.count] -
 776			(char *)leaf);
 777		level = 0;
 778	}
 779
 780	/*
 781	 * Copy old root to new buffer and log it.
 782	 */
 783	xfs_da_buf_copy(bp, blk1->bp, size);
 784	xfs_trans_log_buf(tp, bp, 0, size - 1);
 785
 786	/*
 787	 * Update blk1 to point to new buffer.
 788	 */
 789	blk1->bp = bp;
 790	blk1->blkno = blkno;
 791
 792	/*
 793	 * Set up the new root node.
 794	 */
 795	error = xfs_da3_node_create(args,
 796		(args->whichfork == XFS_DATA_FORK) ? args->geo->leafblk : 0,
 797		level + 1, &bp, args->whichfork);
 798	if (error)
 799		return error;
 800
 801	node = bp->b_addr;
 802	xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
 803	btree = nodehdr.btree;
 804	btree[0].hashval = cpu_to_be32(blk1->hashval);
 805	btree[0].before = cpu_to_be32(blk1->blkno);
 806	btree[1].hashval = cpu_to_be32(blk2->hashval);
 807	btree[1].before = cpu_to_be32(blk2->blkno);
 808	nodehdr.count = 2;
 809	xfs_da3_node_hdr_to_disk(dp->i_mount, node, &nodehdr);
 810
 811#ifdef DEBUG
 812	if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
 813	    oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
 814		ASSERT(blk1->blkno >= args->geo->leafblk &&
 815		       blk1->blkno < args->geo->freeblk);
 816		ASSERT(blk2->blkno >= args->geo->leafblk &&
 817		       blk2->blkno < args->geo->freeblk);
 818	}
 819#endif
 820
 821	/* Header is already logged by xfs_da_node_create */
 822	xfs_trans_log_buf(tp, bp,
 823		XFS_DA_LOGRANGE(node, btree, sizeof(xfs_da_node_entry_t) * 2));
 824
 825	return 0;
 826}
 827
 828/*
 829 * Split the node, rebalance, then add the new entry.
 830 */
 831STATIC int						/* error */
 832xfs_da3_node_split(
 833	struct xfs_da_state	*state,
 834	struct xfs_da_state_blk	*oldblk,
 835	struct xfs_da_state_blk	*newblk,
 836	struct xfs_da_state_blk	*addblk,
 837	int			treelevel,
 838	int			*result)
 839{
 840	struct xfs_da_intnode	*node;
 841	struct xfs_da3_icnode_hdr nodehdr;
 842	xfs_dablk_t		blkno;
 843	int			newcount;
 844	int			error;
 845	int			useextra;
 846	struct xfs_inode	*dp = state->args->dp;
 847
 848	trace_xfs_da_node_split(state->args);
 849
 850	node = oldblk->bp->b_addr;
 851	xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
 852
 853	/*
 854	 * With V2 dirs the extra block is data or freespace.
 855	 */
 856	useextra = state->extravalid && state->args->whichfork == XFS_ATTR_FORK;
 857	newcount = 1 + useextra;
 858	/*
 859	 * Do we have to split the node?
 860	 */
 861	if (nodehdr.count + newcount > state->args->geo->node_ents) {
 862		/*
 863		 * Allocate a new node, add to the doubly linked chain of
 864		 * nodes, then move some of our excess entries into it.
 865		 */
 866		error = xfs_da_grow_inode(state->args, &blkno);
 867		if (error)
 868			return error;	/* GROT: dir is inconsistent */
 869
 870		error = xfs_da3_node_create(state->args, blkno, treelevel,
 871					   &newblk->bp, state->args->whichfork);
 872		if (error)
 873			return error;	/* GROT: dir is inconsistent */
 874		newblk->blkno = blkno;
 875		newblk->magic = XFS_DA_NODE_MAGIC;
 876		xfs_da3_node_rebalance(state, oldblk, newblk);
 877		error = xfs_da3_blk_link(state, oldblk, newblk);
 878		if (error)
 879			return error;
 880		*result = 1;
 881	} else {
 882		*result = 0;
 883	}
 884
 885	/*
 886	 * Insert the new entry(s) into the correct block
 887	 * (updating last hashval in the process).
 888	 *
 889	 * xfs_da3_node_add() inserts BEFORE the given index,
 890	 * and as a result of using node_lookup_int() we always
 891	 * point to a valid entry (not after one), but a split
 892	 * operation always results in a new block whose hashvals
 893	 * FOLLOW the current block.
 894	 *
 895	 * If we had double-split op below us, then add the extra block too.
 896	 */
 897	node = oldblk->bp->b_addr;
 898	xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
 899	if (oldblk->index <= nodehdr.count) {
 900		oldblk->index++;
 901		xfs_da3_node_add(state, oldblk, addblk);
 902		if (useextra) {
 903			if (state->extraafter)
 904				oldblk->index++;
 905			xfs_da3_node_add(state, oldblk, &state->extrablk);
 906			state->extravalid = 0;
 907		}
 908	} else {
 909		newblk->index++;
 910		xfs_da3_node_add(state, newblk, addblk);
 911		if (useextra) {
 912			if (state->extraafter)
 913				newblk->index++;
 914			xfs_da3_node_add(state, newblk, &state->extrablk);
 915			state->extravalid = 0;
 916		}
 917	}
 918
 919	return 0;
 920}
 921
 922/*
 923 * Balance the btree elements between two intermediate nodes,
 924 * usually one full and one empty.
 925 *
 926 * NOTE: if blk2 is empty, then it will get the upper half of blk1.
 927 */
 928STATIC void
 929xfs_da3_node_rebalance(
 930	struct xfs_da_state	*state,
 931	struct xfs_da_state_blk	*blk1,
 932	struct xfs_da_state_blk	*blk2)
 933{
 934	struct xfs_da_intnode	*node1;
 935	struct xfs_da_intnode	*node2;
 936	struct xfs_da_node_entry *btree1;
 937	struct xfs_da_node_entry *btree2;
 938	struct xfs_da_node_entry *btree_s;
 939	struct xfs_da_node_entry *btree_d;
 940	struct xfs_da3_icnode_hdr nodehdr1;
 941	struct xfs_da3_icnode_hdr nodehdr2;
 942	struct xfs_trans	*tp;
 943	int			count;
 944	int			tmp;
 945	int			swap = 0;
 946	struct xfs_inode	*dp = state->args->dp;
 947
 948	trace_xfs_da_node_rebalance(state->args);
 949
 950	node1 = blk1->bp->b_addr;
 951	node2 = blk2->bp->b_addr;
 952	xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr1, node1);
 953	xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr2, node2);
 954	btree1 = nodehdr1.btree;
 955	btree2 = nodehdr2.btree;
 956
 957	/*
 958	 * Figure out how many entries need to move, and in which direction.
 959	 * Swap the nodes around if that makes it simpler.
 960	 */
 961	if (nodehdr1.count > 0 && nodehdr2.count > 0 &&
 962	    ((be32_to_cpu(btree2[0].hashval) < be32_to_cpu(btree1[0].hashval)) ||
 963	     (be32_to_cpu(btree2[nodehdr2.count - 1].hashval) <
 964			be32_to_cpu(btree1[nodehdr1.count - 1].hashval)))) {
 965		swap(node1, node2);
 966		xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr1, node1);
 967		xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr2, node2);
 968		btree1 = nodehdr1.btree;
 969		btree2 = nodehdr2.btree;
 970		swap = 1;
 971	}
 972
 973	count = (nodehdr1.count - nodehdr2.count) / 2;
 974	if (count == 0)
 975		return;
 976	tp = state->args->trans;
 977	/*
 978	 * Two cases: high-to-low and low-to-high.
 979	 */
 980	if (count > 0) {
 981		/*
 982		 * Move elements in node2 up to make a hole.
 983		 */
 984		tmp = nodehdr2.count;
 985		if (tmp > 0) {
 986			tmp *= (uint)sizeof(xfs_da_node_entry_t);
 987			btree_s = &btree2[0];
 988			btree_d = &btree2[count];
 989			memmove(btree_d, btree_s, tmp);
 990		}
 991
 992		/*
 993		 * Move the req'd B-tree elements from high in node1 to
 994		 * low in node2.
 995		 */
 996		nodehdr2.count += count;
 997		tmp = count * (uint)sizeof(xfs_da_node_entry_t);
 998		btree_s = &btree1[nodehdr1.count - count];
 999		btree_d = &btree2[0];
1000		memcpy(btree_d, btree_s, tmp);
1001		nodehdr1.count -= count;
1002	} else {
1003		/*
1004		 * Move the req'd B-tree elements from low in node2 to
1005		 * high in node1.
1006		 */
1007		count = -count;
1008		tmp = count * (uint)sizeof(xfs_da_node_entry_t);
1009		btree_s = &btree2[0];
1010		btree_d = &btree1[nodehdr1.count];
1011		memcpy(btree_d, btree_s, tmp);
1012		nodehdr1.count += count;
1013
1014		xfs_trans_log_buf(tp, blk1->bp,
1015			XFS_DA_LOGRANGE(node1, btree_d, tmp));
1016
1017		/*
1018		 * Move elements in node2 down to fill the hole.
1019		 */
1020		tmp  = nodehdr2.count - count;
1021		tmp *= (uint)sizeof(xfs_da_node_entry_t);
1022		btree_s = &btree2[count];
1023		btree_d = &btree2[0];
1024		memmove(btree_d, btree_s, tmp);
1025		nodehdr2.count -= count;
1026	}
1027
1028	/*
1029	 * Log header of node 1 and all current bits of node 2.
1030	 */
1031	xfs_da3_node_hdr_to_disk(dp->i_mount, node1, &nodehdr1);
1032	xfs_trans_log_buf(tp, blk1->bp,
1033		XFS_DA_LOGRANGE(node1, &node1->hdr,
1034				state->args->geo->node_hdr_size));
1035
1036	xfs_da3_node_hdr_to_disk(dp->i_mount, node2, &nodehdr2);
1037	xfs_trans_log_buf(tp, blk2->bp,
1038		XFS_DA_LOGRANGE(node2, &node2->hdr,
1039				state->args->geo->node_hdr_size +
1040				(sizeof(btree2[0]) * nodehdr2.count)));
1041
1042	/*
1043	 * Record the last hashval from each block for upward propagation.
1044	 * (note: don't use the swapped node pointers)
1045	 */
1046	if (swap) {
1047		node1 = blk1->bp->b_addr;
1048		node2 = blk2->bp->b_addr;
1049		xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr1, node1);
1050		xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr2, node2);
1051		btree1 = nodehdr1.btree;
1052		btree2 = nodehdr2.btree;
1053	}
1054	blk1->hashval = be32_to_cpu(btree1[nodehdr1.count - 1].hashval);
1055	blk2->hashval = be32_to_cpu(btree2[nodehdr2.count - 1].hashval);
1056
1057	/*
1058	 * Adjust the expected index for insertion.
1059	 */
1060	if (blk1->index >= nodehdr1.count) {
1061		blk2->index = blk1->index - nodehdr1.count;
1062		blk1->index = nodehdr1.count + 1;	/* make it invalid */
1063	}
1064}
1065
1066/*
1067 * Add a new entry to an intermediate node.
1068 */
1069STATIC void
1070xfs_da3_node_add(
1071	struct xfs_da_state	*state,
1072	struct xfs_da_state_blk	*oldblk,
1073	struct xfs_da_state_blk	*newblk)
1074{
1075	struct xfs_da_intnode	*node;
1076	struct xfs_da3_icnode_hdr nodehdr;
1077	struct xfs_da_node_entry *btree;
1078	int			tmp;
1079	struct xfs_inode	*dp = state->args->dp;
1080
1081	trace_xfs_da_node_add(state->args);
1082
1083	node = oldblk->bp->b_addr;
1084	xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
1085	btree = nodehdr.btree;
1086
1087	ASSERT(oldblk->index >= 0 && oldblk->index <= nodehdr.count);
1088	ASSERT(newblk->blkno != 0);
1089	if (state->args->whichfork == XFS_DATA_FORK)
1090		ASSERT(newblk->blkno >= state->args->geo->leafblk &&
1091		       newblk->blkno < state->args->geo->freeblk);
1092
1093	/*
1094	 * We may need to make some room before we insert the new node.
1095	 */
1096	tmp = 0;
1097	if (oldblk->index < nodehdr.count) {
1098		tmp = (nodehdr.count - oldblk->index) * (uint)sizeof(*btree);
1099		memmove(&btree[oldblk->index + 1], &btree[oldblk->index], tmp);
1100	}
1101	btree[oldblk->index].hashval = cpu_to_be32(newblk->hashval);
1102	btree[oldblk->index].before = cpu_to_be32(newblk->blkno);
1103	xfs_trans_log_buf(state->args->trans, oldblk->bp,
1104		XFS_DA_LOGRANGE(node, &btree[oldblk->index],
1105				tmp + sizeof(*btree)));
1106
1107	nodehdr.count += 1;
1108	xfs_da3_node_hdr_to_disk(dp->i_mount, node, &nodehdr);
1109	xfs_trans_log_buf(state->args->trans, oldblk->bp,
1110		XFS_DA_LOGRANGE(node, &node->hdr,
1111				state->args->geo->node_hdr_size));
1112
1113	/*
1114	 * Copy the last hash value from the oldblk to propagate upwards.
1115	 */
1116	oldblk->hashval = be32_to_cpu(btree[nodehdr.count - 1].hashval);
1117}
1118
1119/*========================================================================
1120 * Routines used for shrinking the Btree.
1121 *========================================================================*/
1122
1123/*
1124 * Deallocate an empty leaf node, remove it from its parent,
1125 * possibly deallocating that block, etc...
1126 */
1127int
1128xfs_da3_join(
1129	struct xfs_da_state	*state)
1130{
1131	struct xfs_da_state_blk	*drop_blk;
1132	struct xfs_da_state_blk	*save_blk;
1133	int			action = 0;
1134	int			error;
1135
1136	trace_xfs_da_join(state->args);
1137
1138	drop_blk = &state->path.blk[ state->path.active-1 ];
1139	save_blk = &state->altpath.blk[ state->path.active-1 ];
1140	ASSERT(state->path.blk[0].magic == XFS_DA_NODE_MAGIC);
1141	ASSERT(drop_blk->magic == XFS_ATTR_LEAF_MAGIC ||
1142	       drop_blk->magic == XFS_DIR2_LEAFN_MAGIC);
1143
1144	/*
1145	 * Walk back up the tree joining/deallocating as necessary.
1146	 * When we stop dropping blocks, break out.
1147	 */
1148	for (  ; state->path.active >= 2; drop_blk--, save_blk--,
1149		 state->path.active--) {
1150		/*
1151		 * See if we can combine the block with a neighbor.
1152		 *   (action == 0) => no options, just leave
1153		 *   (action == 1) => coalesce, then unlink
1154		 *   (action == 2) => block empty, unlink it
1155		 */
1156		switch (drop_blk->magic) {
1157		case XFS_ATTR_LEAF_MAGIC:
1158			error = xfs_attr3_leaf_toosmall(state, &action);
1159			if (error)
1160				return error;
1161			if (action == 0)
1162				return 0;
1163			xfs_attr3_leaf_unbalance(state, drop_blk, save_blk);
1164			break;
1165		case XFS_DIR2_LEAFN_MAGIC:
1166			error = xfs_dir2_leafn_toosmall(state, &action);
1167			if (error)
1168				return error;
1169			if (action == 0)
1170				return 0;
1171			xfs_dir2_leafn_unbalance(state, drop_blk, save_blk);
1172			break;
1173		case XFS_DA_NODE_MAGIC:
1174			/*
1175			 * Remove the offending node, fixup hashvals,
1176			 * check for a toosmall neighbor.
1177			 */
1178			xfs_da3_node_remove(state, drop_blk);
1179			xfs_da3_fixhashpath(state, &state->path);
1180			error = xfs_da3_node_toosmall(state, &action);
1181			if (error)
1182				return error;
1183			if (action == 0)
1184				return 0;
1185			xfs_da3_node_unbalance(state, drop_blk, save_blk);
1186			break;
1187		}
1188		xfs_da3_fixhashpath(state, &state->altpath);
1189		error = xfs_da3_blk_unlink(state, drop_blk, save_blk);
1190		xfs_da_state_kill_altpath(state);
1191		if (error)
1192			return error;
1193		error = xfs_da_shrink_inode(state->args, drop_blk->blkno,
1194							 drop_blk->bp);
1195		drop_blk->bp = NULL;
1196		if (error)
1197			return error;
1198	}
1199	/*
1200	 * We joined all the way to the top.  If it turns out that
1201	 * we only have one entry in the root, make the child block
1202	 * the new root.
1203	 */
1204	xfs_da3_node_remove(state, drop_blk);
1205	xfs_da3_fixhashpath(state, &state->path);
1206	error = xfs_da3_root_join(state, &state->path.blk[0]);
1207	return error;
1208}
1209
1210#ifdef	DEBUG
1211static void
1212xfs_da_blkinfo_onlychild_validate(struct xfs_da_blkinfo *blkinfo, __u16 level)
1213{
1214	__be16	magic = blkinfo->magic;
1215
1216	if (level == 1) {
1217		ASSERT(magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
1218		       magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC) ||
1219		       magic == cpu_to_be16(XFS_ATTR_LEAF_MAGIC) ||
1220		       magic == cpu_to_be16(XFS_ATTR3_LEAF_MAGIC));
1221	} else {
1222		ASSERT(magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
1223		       magic == cpu_to_be16(XFS_DA3_NODE_MAGIC));
1224	}
1225	ASSERT(!blkinfo->forw);
1226	ASSERT(!blkinfo->back);
1227}
1228#else	/* !DEBUG */
1229#define	xfs_da_blkinfo_onlychild_validate(blkinfo, level)
1230#endif	/* !DEBUG */
1231
1232/*
1233 * We have only one entry in the root.  Copy the only remaining child of
1234 * the old root to block 0 as the new root node.
1235 */
1236STATIC int
1237xfs_da3_root_join(
1238	struct xfs_da_state	*state,
1239	struct xfs_da_state_blk	*root_blk)
1240{
1241	struct xfs_da_intnode	*oldroot;
1242	struct xfs_da_args	*args;
1243	xfs_dablk_t		child;
1244	struct xfs_buf		*bp;
1245	struct xfs_da3_icnode_hdr oldroothdr;
1246	int			error;
1247	struct xfs_inode	*dp = state->args->dp;
1248	xfs_failaddr_t		fa;
1249
1250	trace_xfs_da_root_join(state->args);
1251
1252	ASSERT(root_blk->magic == XFS_DA_NODE_MAGIC);
1253
1254	args = state->args;
1255	oldroot = root_blk->bp->b_addr;
1256	xfs_da3_node_hdr_from_disk(dp->i_mount, &oldroothdr, oldroot);
1257	ASSERT(oldroothdr.forw == 0);
1258	ASSERT(oldroothdr.back == 0);
1259
1260	/*
1261	 * If the root has more than one child, then don't do anything.
1262	 */
1263	if (oldroothdr.count > 1)
1264		return 0;
1265
1266	/*
1267	 * Read in the (only) child block, then copy those bytes into
1268	 * the root block's buffer and free the original child block.
1269	 */
1270	child = be32_to_cpu(oldroothdr.btree[0].before);
1271	ASSERT(child != 0);
1272	error = xfs_da3_node_read(args->trans, dp, child, &bp, args->whichfork);
1273	if (error)
1274		return error;
1275	fa = xfs_da3_header_check(bp, args->owner);
1276	if (fa) {
1277		__xfs_buf_mark_corrupt(bp, fa);
1278		xfs_trans_brelse(args->trans, bp);
1279		xfs_da_mark_sick(args);
1280		return -EFSCORRUPTED;
1281	}
1282	xfs_da_blkinfo_onlychild_validate(bp->b_addr, oldroothdr.level);
1283
1284	/*
1285	 * Copy child to root buffer and log it.
1286	 */
1287	xfs_da_buf_copy(root_blk->bp, bp, args->geo->blksize);
1288	xfs_trans_log_buf(args->trans, root_blk->bp, 0,
1289			  args->geo->blksize - 1);
1290	/*
1291	 * Now we can drop the child buffer.
1292	 */
1293	error = xfs_da_shrink_inode(args, child, bp);
1294	return error;
1295}
1296
1297/*
1298 * Check a node block and its neighbors to see if the block should be
1299 * collapsed into one or the other neighbor.  Always keep the block
1300 * with the smaller block number.
1301 * If the current block is over 50% full, don't try to join it, return 0.
1302 * If the block is empty, fill in the state structure and return 2.
1303 * If it can be collapsed, fill in the state structure and return 1.
1304 * If nothing can be done, return 0.
1305 */
1306STATIC int
1307xfs_da3_node_toosmall(
1308	struct xfs_da_state	*state,
1309	int			*action)
1310{
1311	struct xfs_da_intnode	*node;
1312	struct xfs_da_state_blk	*blk;
1313	struct xfs_da_blkinfo	*info;
1314	xfs_dablk_t		blkno;
1315	struct xfs_buf		*bp;
1316	xfs_failaddr_t		fa;
1317	struct xfs_da3_icnode_hdr nodehdr;
1318	int			count;
1319	int			forward;
1320	int			error;
1321	int			retval;
1322	int			i;
1323	struct xfs_inode	*dp = state->args->dp;
1324
1325	trace_xfs_da_node_toosmall(state->args);
1326
1327	/*
1328	 * Check for the degenerate case of the block being over 50% full.
1329	 * If so, it's not worth even looking to see if we might be able
1330	 * to coalesce with a sibling.
1331	 */
1332	blk = &state->path.blk[ state->path.active-1 ];
1333	info = blk->bp->b_addr;
1334	node = (xfs_da_intnode_t *)info;
1335	xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
1336	if (nodehdr.count > (state->args->geo->node_ents >> 1)) {
1337		*action = 0;	/* blk over 50%, don't try to join */
1338		return 0;	/* blk over 50%, don't try to join */
1339	}
1340
1341	/*
1342	 * Check for the degenerate case of the block being empty.
1343	 * If the block is empty, we'll simply delete it, no need to
1344	 * coalesce it with a sibling block.  We choose (arbitrarily)
1345	 * to merge with the forward block unless it is NULL.
1346	 */
1347	if (nodehdr.count == 0) {
1348		/*
1349		 * Make altpath point to the block we want to keep and
1350		 * path point to the block we want to drop (this one).
1351		 */
1352		forward = (info->forw != 0);
1353		memcpy(&state->altpath, &state->path, sizeof(state->path));
1354		error = xfs_da3_path_shift(state, &state->altpath, forward,
1355						 0, &retval);
1356		if (error)
1357			return error;
1358		if (retval) {
1359			*action = 0;
1360		} else {
1361			*action = 2;
1362		}
1363		return 0;
1364	}
1365
1366	/*
1367	 * Examine each sibling block to see if we can coalesce with
1368	 * at least 25% free space to spare.  We need to figure out
1369	 * whether to merge with the forward or the backward block.
1370	 * We prefer coalescing with the lower numbered sibling so as
1371	 * to shrink a directory over time.
1372	 */
1373	count  = state->args->geo->node_ents;
1374	count -= state->args->geo->node_ents >> 2;
1375	count -= nodehdr.count;
1376
1377	/* start with smaller blk num */
1378	forward = nodehdr.forw < nodehdr.back;
1379	for (i = 0; i < 2; forward = !forward, i++) {
1380		struct xfs_da3_icnode_hdr thdr;
1381		if (forward)
1382			blkno = nodehdr.forw;
1383		else
1384			blkno = nodehdr.back;
1385		if (blkno == 0)
1386			continue;
1387		error = xfs_da3_node_read(state->args->trans, dp, blkno, &bp,
1388				state->args->whichfork);
1389		if (error)
1390			return error;
1391		fa = xfs_da3_node_header_check(bp, state->args->owner);
1392		if (fa) {
1393			__xfs_buf_mark_corrupt(bp, fa);
1394			xfs_trans_brelse(state->args->trans, bp);
1395			xfs_da_mark_sick(state->args);
1396			return -EFSCORRUPTED;
1397		}
1398
1399		node = bp->b_addr;
1400		xfs_da3_node_hdr_from_disk(dp->i_mount, &thdr, node);
1401		xfs_trans_brelse(state->args->trans, bp);
1402
1403		if (count - thdr.count >= 0)
1404			break;	/* fits with at least 25% to spare */
1405	}
1406	if (i >= 2) {
1407		*action = 0;
1408		return 0;
1409	}
1410
1411	/*
1412	 * Make altpath point to the block we want to keep (the lower
1413	 * numbered block) and path point to the block we want to drop.
1414	 */
1415	memcpy(&state->altpath, &state->path, sizeof(state->path));
1416	if (blkno < blk->blkno) {
1417		error = xfs_da3_path_shift(state, &state->altpath, forward,
1418						 0, &retval);
1419	} else {
1420		error = xfs_da3_path_shift(state, &state->path, forward,
1421						 0, &retval);
1422	}
1423	if (error)
1424		return error;
1425	if (retval) {
1426		*action = 0;
1427		return 0;
1428	}
1429	*action = 1;
1430	return 0;
1431}
1432
1433/*
1434 * Pick up the last hashvalue from an intermediate node.
1435 */
1436STATIC uint
1437xfs_da3_node_lasthash(
1438	struct xfs_inode	*dp,
1439	struct xfs_buf		*bp,
1440	int			*count)
1441{
1442	struct xfs_da3_icnode_hdr nodehdr;
1443
1444	xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, bp->b_addr);
1445	if (count)
1446		*count = nodehdr.count;
1447	if (!nodehdr.count)
1448		return 0;
1449	return be32_to_cpu(nodehdr.btree[nodehdr.count - 1].hashval);
1450}
1451
1452/*
1453 * Walk back up the tree adjusting hash values as necessary,
1454 * when we stop making changes, return.
1455 */
1456void
1457xfs_da3_fixhashpath(
1458	struct xfs_da_state	*state,
1459	struct xfs_da_state_path *path)
1460{
1461	struct xfs_da_state_blk	*blk;
1462	struct xfs_da_intnode	*node;
1463	struct xfs_da_node_entry *btree;
1464	xfs_dahash_t		lasthash=0;
1465	int			level;
1466	int			count;
1467	struct xfs_inode	*dp = state->args->dp;
1468
1469	trace_xfs_da_fixhashpath(state->args);
1470
1471	level = path->active-1;
1472	blk = &path->blk[ level ];
1473	switch (blk->magic) {
1474	case XFS_ATTR_LEAF_MAGIC:
1475		lasthash = xfs_attr_leaf_lasthash(blk->bp, &count);
1476		if (count == 0)
1477			return;
1478		break;
1479	case XFS_DIR2_LEAFN_MAGIC:
1480		lasthash = xfs_dir2_leaf_lasthash(dp, blk->bp, &count);
1481		if (count == 0)
1482			return;
1483		break;
1484	case XFS_DA_NODE_MAGIC:
1485		lasthash = xfs_da3_node_lasthash(dp, blk->bp, &count);
1486		if (count == 0)
1487			return;
1488		break;
1489	}
1490	for (blk--, level--; level >= 0; blk--, level--) {
1491		struct xfs_da3_icnode_hdr nodehdr;
1492
1493		node = blk->bp->b_addr;
1494		xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
1495		btree = nodehdr.btree;
1496		if (be32_to_cpu(btree[blk->index].hashval) == lasthash)
1497			break;
1498		blk->hashval = lasthash;
1499		btree[blk->index].hashval = cpu_to_be32(lasthash);
1500		xfs_trans_log_buf(state->args->trans, blk->bp,
1501				  XFS_DA_LOGRANGE(node, &btree[blk->index],
1502						  sizeof(*btree)));
1503
1504		lasthash = be32_to_cpu(btree[nodehdr.count - 1].hashval);
1505	}
1506}
1507
1508/*
1509 * Internal implementation to remove an entry from an intermediate node.
1510 */
1511STATIC void
1512__xfs_da3_node_remove(
1513	struct xfs_trans	*tp,
1514	struct xfs_inode	*dp,
1515	struct xfs_da_geometry  *geo,
1516	struct xfs_da_state_blk *drop_blk)
1517{
1518	struct xfs_da_intnode	*node;
1519	struct xfs_da3_icnode_hdr nodehdr;
1520	struct xfs_da_node_entry *btree;
1521	int			index;
1522	int			tmp;
1523
1524	node = drop_blk->bp->b_addr;
1525	xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
1526	ASSERT(drop_blk->index < nodehdr.count);
1527	ASSERT(drop_blk->index >= 0);
1528
1529	/*
1530	 * Copy over the offending entry, or just zero it out.
1531	 */
1532	index = drop_blk->index;
1533	btree = nodehdr.btree;
1534	if (index < nodehdr.count - 1) {
1535		tmp  = nodehdr.count - index - 1;
1536		tmp *= (uint)sizeof(xfs_da_node_entry_t);
1537		memmove(&btree[index], &btree[index + 1], tmp);
1538		xfs_trans_log_buf(tp, drop_blk->bp,
1539		    XFS_DA_LOGRANGE(node, &btree[index], tmp));
1540		index = nodehdr.count - 1;
1541	}
1542	memset(&btree[index], 0, sizeof(xfs_da_node_entry_t));
1543	xfs_trans_log_buf(tp, drop_blk->bp,
1544	    XFS_DA_LOGRANGE(node, &btree[index], sizeof(btree[index])));
1545	nodehdr.count -= 1;
1546	xfs_da3_node_hdr_to_disk(dp->i_mount, node, &nodehdr);
1547	xfs_trans_log_buf(tp, drop_blk->bp,
1548	    XFS_DA_LOGRANGE(node, &node->hdr, geo->node_hdr_size));
1549
1550	/*
1551	 * Copy the last hash value from the block to propagate upwards.
1552	 */
1553	drop_blk->hashval = be32_to_cpu(btree[index - 1].hashval);
1554}
1555
1556/*
1557 * Remove an entry from an intermediate node.
1558 */
1559STATIC void
1560xfs_da3_node_remove(
1561	struct xfs_da_state	*state,
1562	struct xfs_da_state_blk	*drop_blk)
1563{
1564	trace_xfs_da_node_remove(state->args);
1565
1566	__xfs_da3_node_remove(state->args->trans, state->args->dp,
1567			state->args->geo, drop_blk);
1568}
1569
1570/*
1571 * Remove an entry from an intermediate attr node at the specified index.
1572 */
1573void
1574xfs_attr3_node_entry_remove(
1575	struct xfs_trans	*tp,
1576	struct xfs_inode	*dp,
1577	struct xfs_buf		*bp,
1578	int			index)
1579{
1580	struct xfs_da_state_blk blk = {
1581		.index		= index,
1582		.bp		= bp,
1583	};
1584
1585	__xfs_da3_node_remove(tp, dp, dp->i_mount->m_attr_geo, &blk);
1586}
1587
1588/*
1589 * Unbalance the elements between two intermediate nodes,
1590 * move all Btree elements from one node into another.
1591 */
1592STATIC void
1593xfs_da3_node_unbalance(
1594	struct xfs_da_state	*state,
1595	struct xfs_da_state_blk	*drop_blk,
1596	struct xfs_da_state_blk	*save_blk)
1597{
1598	struct xfs_da_intnode	*drop_node;
1599	struct xfs_da_intnode	*save_node;
1600	struct xfs_da_node_entry *drop_btree;
1601	struct xfs_da_node_entry *save_btree;
1602	struct xfs_da3_icnode_hdr drop_hdr;
1603	struct xfs_da3_icnode_hdr save_hdr;
1604	struct xfs_trans	*tp;
1605	int			sindex;
1606	int			tmp;
1607	struct xfs_inode	*dp = state->args->dp;
1608
1609	trace_xfs_da_node_unbalance(state->args);
1610
1611	drop_node = drop_blk->bp->b_addr;
1612	save_node = save_blk->bp->b_addr;
1613	xfs_da3_node_hdr_from_disk(dp->i_mount, &drop_hdr, drop_node);
1614	xfs_da3_node_hdr_from_disk(dp->i_mount, &save_hdr, save_node);
1615	drop_btree = drop_hdr.btree;
1616	save_btree = save_hdr.btree;
1617	tp = state->args->trans;
1618
1619	/*
1620	 * If the dying block has lower hashvals, then move all the
1621	 * elements in the remaining block up to make a hole.
1622	 */
1623	if ((be32_to_cpu(drop_btree[0].hashval) <
1624			be32_to_cpu(save_btree[0].hashval)) ||
1625	    (be32_to_cpu(drop_btree[drop_hdr.count - 1].hashval) <
1626			be32_to_cpu(save_btree[save_hdr.count - 1].hashval))) {
1627		/* XXX: check this - is memmove dst correct? */
1628		tmp = save_hdr.count * sizeof(xfs_da_node_entry_t);
1629		memmove(&save_btree[drop_hdr.count], &save_btree[0], tmp);
1630
1631		sindex = 0;
1632		xfs_trans_log_buf(tp, save_blk->bp,
1633			XFS_DA_LOGRANGE(save_node, &save_btree[0],
1634				(save_hdr.count + drop_hdr.count) *
1635						sizeof(xfs_da_node_entry_t)));
1636	} else {
1637		sindex = save_hdr.count;
1638		xfs_trans_log_buf(tp, save_blk->bp,
1639			XFS_DA_LOGRANGE(save_node, &save_btree[sindex],
1640				drop_hdr.count * sizeof(xfs_da_node_entry_t)));
1641	}
1642
1643	/*
1644	 * Move all the B-tree elements from drop_blk to save_blk.
1645	 */
1646	tmp = drop_hdr.count * (uint)sizeof(xfs_da_node_entry_t);
1647	memcpy(&save_btree[sindex], &drop_btree[0], tmp);
1648	save_hdr.count += drop_hdr.count;
1649
1650	xfs_da3_node_hdr_to_disk(dp->i_mount, save_node, &save_hdr);
1651	xfs_trans_log_buf(tp, save_blk->bp,
1652		XFS_DA_LOGRANGE(save_node, &save_node->hdr,
1653				state->args->geo->node_hdr_size));
1654
1655	/*
1656	 * Save the last hashval in the remaining block for upward propagation.
1657	 */
1658	save_blk->hashval = be32_to_cpu(save_btree[save_hdr.count - 1].hashval);
1659}
1660
1661/*========================================================================
1662 * Routines used for finding things in the Btree.
1663 *========================================================================*/
1664
1665/*
1666 * Walk down the Btree looking for a particular filename, filling
1667 * in the state structure as we go.
1668 *
1669 * We will set the state structure to point to each of the elements
1670 * in each of the nodes where either the hashval is or should be.
1671 *
1672 * We support duplicate hashval's so for each entry in the current
1673 * node that could contain the desired hashval, descend.  This is a
1674 * pruned depth-first tree search.
1675 */
1676int							/* error */
1677xfs_da3_node_lookup_int(
1678	struct xfs_da_state	*state,
1679	int			*result)
1680{
1681	struct xfs_da_state_blk	*blk;
1682	struct xfs_da_blkinfo	*curr;
1683	struct xfs_da_intnode	*node;
1684	struct xfs_da_node_entry *btree;
1685	struct xfs_da3_icnode_hdr nodehdr;
1686	struct xfs_da_args	*args;
1687	xfs_failaddr_t		fa;
1688	xfs_dablk_t		blkno;
1689	xfs_dahash_t		hashval;
1690	xfs_dahash_t		btreehashval;
1691	int			probe;
1692	int			span;
1693	int			max;
1694	int			error;
1695	int			retval;
1696	unsigned int		expected_level = 0;
1697	uint16_t		magic;
1698	struct xfs_inode	*dp = state->args->dp;
1699
1700	args = state->args;
1701
1702	/*
1703	 * Descend thru the B-tree searching each level for the right
1704	 * node to use, until the right hashval is found.
1705	 */
1706	blkno = args->geo->leafblk;
1707	for (blk = &state->path.blk[0], state->path.active = 1;
1708			 state->path.active <= XFS_DA_NODE_MAXDEPTH;
1709			 blk++, state->path.active++) {
1710		/*
1711		 * Read the next node down in the tree.
1712		 */
1713		blk->blkno = blkno;
1714		error = xfs_da3_node_read(args->trans, args->dp, blkno,
1715					&blk->bp, args->whichfork);
1716		if (error) {
1717			blk->blkno = 0;
1718			state->path.active--;
1719			return error;
1720		}
1721		curr = blk->bp->b_addr;
1722		magic = be16_to_cpu(curr->magic);
1723
1724		if (magic == XFS_ATTR_LEAF_MAGIC ||
1725		    magic == XFS_ATTR3_LEAF_MAGIC) {
1726			fa = xfs_attr3_leaf_header_check(blk->bp, args->owner);
1727			if (fa) {
1728				__xfs_buf_mark_corrupt(blk->bp, fa);
1729				xfs_da_mark_sick(args);
1730				return -EFSCORRUPTED;
1731			}
1732			blk->magic = XFS_ATTR_LEAF_MAGIC;
1733			blk->hashval = xfs_attr_leaf_lasthash(blk->bp, NULL);
1734			break;
1735		}
1736
1737		if (magic == XFS_DIR2_LEAFN_MAGIC ||
1738		    magic == XFS_DIR3_LEAFN_MAGIC) {
1739			fa = xfs_dir3_leaf_header_check(blk->bp, args->owner);
1740			if (fa) {
1741				__xfs_buf_mark_corrupt(blk->bp, fa);
1742				xfs_da_mark_sick(args);
1743				return -EFSCORRUPTED;
1744			}
1745			blk->magic = XFS_DIR2_LEAFN_MAGIC;
1746			blk->hashval = xfs_dir2_leaf_lasthash(args->dp,
1747							      blk->bp, NULL);
1748			break;
1749		}
1750
1751		if (magic != XFS_DA_NODE_MAGIC && magic != XFS_DA3_NODE_MAGIC) {
1752			xfs_buf_mark_corrupt(blk->bp);
1753			xfs_da_mark_sick(args);
1754			return -EFSCORRUPTED;
1755		}
1756
1757		fa = xfs_da3_node_header_check(blk->bp, args->owner);
1758		if (fa) {
1759			__xfs_buf_mark_corrupt(blk->bp, fa);
1760			xfs_da_mark_sick(args);
1761			return -EFSCORRUPTED;
1762		}
1763
1764		blk->magic = XFS_DA_NODE_MAGIC;
1765
1766		/*
1767		 * Search an intermediate node for a match.
1768		 */
1769		node = blk->bp->b_addr;
1770		xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr, node);
1771		btree = nodehdr.btree;
1772
1773		/* Tree taller than we can handle; bail out! */
1774		if (nodehdr.level >= XFS_DA_NODE_MAXDEPTH) {
1775			xfs_buf_mark_corrupt(blk->bp);
1776			xfs_da_mark_sick(args);
1777			return -EFSCORRUPTED;
1778		}
1779
1780		/* Check the level from the root. */
1781		if (blkno == args->geo->leafblk)
1782			expected_level = nodehdr.level - 1;
1783		else if (expected_level != nodehdr.level) {
1784			xfs_buf_mark_corrupt(blk->bp);
1785			xfs_da_mark_sick(args);
1786			return -EFSCORRUPTED;
1787		} else
1788			expected_level--;
1789
1790		max = nodehdr.count;
1791		blk->hashval = be32_to_cpu(btree[max - 1].hashval);
1792
1793		/*
1794		 * Binary search.  (note: small blocks will skip loop)
1795		 */
1796		probe = span = max / 2;
1797		hashval = args->hashval;
1798		while (span > 4) {
1799			span /= 2;
1800			btreehashval = be32_to_cpu(btree[probe].hashval);
1801			if (btreehashval < hashval)
1802				probe += span;
1803			else if (btreehashval > hashval)
1804				probe -= span;
1805			else
1806				break;
1807		}
1808		ASSERT((probe >= 0) && (probe < max));
1809		ASSERT((span <= 4) ||
1810			(be32_to_cpu(btree[probe].hashval) == hashval));
1811
1812		/*
1813		 * Since we may have duplicate hashval's, find the first
1814		 * matching hashval in the node.
1815		 */
1816		while (probe > 0 &&
1817		       be32_to_cpu(btree[probe].hashval) >= hashval) {
1818			probe--;
1819		}
1820		while (probe < max &&
1821		       be32_to_cpu(btree[probe].hashval) < hashval) {
1822			probe++;
1823		}
1824
1825		/*
1826		 * Pick the right block to descend on.
1827		 */
1828		if (probe == max) {
1829			blk->index = max - 1;
1830			blkno = be32_to_cpu(btree[max - 1].before);
1831		} else {
1832			blk->index = probe;
1833			blkno = be32_to_cpu(btree[probe].before);
1834		}
1835
1836		/* We can't point back to the root. */
1837		if (XFS_IS_CORRUPT(dp->i_mount, blkno == args->geo->leafblk)) {
1838			xfs_da_mark_sick(args);
1839			return -EFSCORRUPTED;
1840		}
1841	}
1842
1843	if (XFS_IS_CORRUPT(dp->i_mount, expected_level != 0)) {
1844		xfs_da_mark_sick(args);
1845		return -EFSCORRUPTED;
1846	}
1847
1848	/*
1849	 * A leaf block that ends in the hashval that we are interested in
1850	 * (final hashval == search hashval) means that the next block may
1851	 * contain more entries with the same hashval, shift upward to the
1852	 * next leaf and keep searching.
1853	 */
1854	for (;;) {
1855		if (blk->magic == XFS_DIR2_LEAFN_MAGIC) {
1856			retval = xfs_dir2_leafn_lookup_int(blk->bp, args,
1857							&blk->index, state);
1858		} else if (blk->magic == XFS_ATTR_LEAF_MAGIC) {
1859			retval = xfs_attr3_leaf_lookup_int(blk->bp, args);
1860			blk->index = args->index;
1861			args->blkno = blk->blkno;
1862		} else {
1863			ASSERT(0);
1864			xfs_da_mark_sick(args);
1865			return -EFSCORRUPTED;
1866		}
1867		if (((retval == -ENOENT) || (retval == -ENOATTR)) &&
1868		    (blk->hashval == args->hashval)) {
1869			error = xfs_da3_path_shift(state, &state->path, 1, 1,
1870							 &retval);
1871			if (error)
1872				return error;
1873			if (retval == 0) {
1874				continue;
1875			} else if (blk->magic == XFS_ATTR_LEAF_MAGIC) {
1876				/* path_shift() gives ENOENT */
1877				retval = -ENOATTR;
1878			}
1879		}
1880		break;
1881	}
1882	*result = retval;
1883	return 0;
1884}
1885
1886/*========================================================================
1887 * Utility routines.
1888 *========================================================================*/
1889
1890/*
1891 * Compare two intermediate nodes for "order".
1892 */
1893STATIC int
1894xfs_da3_node_order(
1895	struct xfs_inode *dp,
1896	struct xfs_buf	*node1_bp,
1897	struct xfs_buf	*node2_bp)
1898{
1899	struct xfs_da_intnode	*node1;
1900	struct xfs_da_intnode	*node2;
1901	struct xfs_da_node_entry *btree1;
1902	struct xfs_da_node_entry *btree2;
1903	struct xfs_da3_icnode_hdr node1hdr;
1904	struct xfs_da3_icnode_hdr node2hdr;
1905
1906	node1 = node1_bp->b_addr;
1907	node2 = node2_bp->b_addr;
1908	xfs_da3_node_hdr_from_disk(dp->i_mount, &node1hdr, node1);
1909	xfs_da3_node_hdr_from_disk(dp->i_mount, &node2hdr, node2);
1910	btree1 = node1hdr.btree;
1911	btree2 = node2hdr.btree;
1912
1913	if (node1hdr.count > 0 && node2hdr.count > 0 &&
1914	    ((be32_to_cpu(btree2[0].hashval) < be32_to_cpu(btree1[0].hashval)) ||
1915	     (be32_to_cpu(btree2[node2hdr.count - 1].hashval) <
1916	      be32_to_cpu(btree1[node1hdr.count - 1].hashval)))) {
1917		return 1;
1918	}
1919	return 0;
1920}
1921
1922/*
1923 * Link a new block into a doubly linked list of blocks (of whatever type).
1924 */
1925int							/* error */
1926xfs_da3_blk_link(
1927	struct xfs_da_state	*state,
1928	struct xfs_da_state_blk	*old_blk,
1929	struct xfs_da_state_blk	*new_blk)
1930{
1931	struct xfs_da_blkinfo	*old_info;
1932	struct xfs_da_blkinfo	*new_info;
1933	struct xfs_da_blkinfo	*tmp_info;
1934	struct xfs_da_args	*args;
1935	struct xfs_buf		*bp;
1936	xfs_failaddr_t		fa;
1937	int			before = 0;
1938	int			error;
1939	struct xfs_inode	*dp = state->args->dp;
1940
1941	/*
1942	 * Set up environment.
1943	 */
1944	args = state->args;
1945	ASSERT(args != NULL);
1946	old_info = old_blk->bp->b_addr;
1947	new_info = new_blk->bp->b_addr;
1948	ASSERT(old_blk->magic == XFS_DA_NODE_MAGIC ||
1949	       old_blk->magic == XFS_DIR2_LEAFN_MAGIC ||
1950	       old_blk->magic == XFS_ATTR_LEAF_MAGIC);
1951
1952	switch (old_blk->magic) {
1953	case XFS_ATTR_LEAF_MAGIC:
1954		before = xfs_attr_leaf_order(old_blk->bp, new_blk->bp);
1955		break;
1956	case XFS_DIR2_LEAFN_MAGIC:
1957		before = xfs_dir2_leafn_order(dp, old_blk->bp, new_blk->bp);
1958		break;
1959	case XFS_DA_NODE_MAGIC:
1960		before = xfs_da3_node_order(dp, old_blk->bp, new_blk->bp);
1961		break;
1962	}
1963
1964	/*
1965	 * Link blocks in appropriate order.
1966	 */
1967	if (before) {
1968		/*
1969		 * Link new block in before existing block.
1970		 */
1971		trace_xfs_da_link_before(args);
1972		new_info->forw = cpu_to_be32(old_blk->blkno);
1973		new_info->back = old_info->back;
1974		if (old_info->back) {
1975			error = xfs_da3_node_read(args->trans, dp,
1976						be32_to_cpu(old_info->back),
1977						&bp, args->whichfork);
1978			if (error)
1979				return error;
1980			fa = xfs_da3_header_check(bp, args->owner);
1981			if (fa) {
1982				__xfs_buf_mark_corrupt(bp, fa);
1983				xfs_trans_brelse(args->trans, bp);
1984				xfs_da_mark_sick(args);
1985				return -EFSCORRUPTED;
1986			}
1987			ASSERT(bp != NULL);
1988			tmp_info = bp->b_addr;
1989			ASSERT(tmp_info->magic == old_info->magic);
1990			ASSERT(be32_to_cpu(tmp_info->forw) == old_blk->blkno);
1991			tmp_info->forw = cpu_to_be32(new_blk->blkno);
1992			xfs_trans_log_buf(args->trans, bp, 0, sizeof(*tmp_info)-1);
1993		}
1994		old_info->back = cpu_to_be32(new_blk->blkno);
1995	} else {
1996		/*
1997		 * Link new block in after existing block.
1998		 */
1999		trace_xfs_da_link_after(args);
2000		new_info->forw = old_info->forw;
2001		new_info->back = cpu_to_be32(old_blk->blkno);
2002		if (old_info->forw) {
2003			error = xfs_da3_node_read(args->trans, dp,
2004						be32_to_cpu(old_info->forw),
2005						&bp, args->whichfork);
2006			if (error)
2007				return error;
2008			fa = xfs_da3_header_check(bp, args->owner);
2009			if (fa) {
2010				__xfs_buf_mark_corrupt(bp, fa);
2011				xfs_trans_brelse(args->trans, bp);
2012				xfs_da_mark_sick(args);
2013				return -EFSCORRUPTED;
2014			}
2015			ASSERT(bp != NULL);
2016			tmp_info = bp->b_addr;
2017			ASSERT(tmp_info->magic == old_info->magic);
2018			ASSERT(be32_to_cpu(tmp_info->back) == old_blk->blkno);
2019			tmp_info->back = cpu_to_be32(new_blk->blkno);
2020			xfs_trans_log_buf(args->trans, bp, 0, sizeof(*tmp_info)-1);
2021		}
2022		old_info->forw = cpu_to_be32(new_blk->blkno);
2023	}
2024
2025	xfs_trans_log_buf(args->trans, old_blk->bp, 0, sizeof(*tmp_info) - 1);
2026	xfs_trans_log_buf(args->trans, new_blk->bp, 0, sizeof(*tmp_info) - 1);
2027	return 0;
2028}
2029
2030/*
2031 * Unlink a block from a doubly linked list of blocks.
2032 */
2033STATIC int						/* error */
2034xfs_da3_blk_unlink(
2035	struct xfs_da_state	*state,
2036	struct xfs_da_state_blk	*drop_blk,
2037	struct xfs_da_state_blk	*save_blk)
2038{
2039	struct xfs_da_blkinfo	*drop_info;
2040	struct xfs_da_blkinfo	*save_info;
2041	struct xfs_da_blkinfo	*tmp_info;
2042	struct xfs_da_args	*args;
2043	struct xfs_buf		*bp;
2044	xfs_failaddr_t		fa;
2045	int			error;
2046
2047	/*
2048	 * Set up environment.
2049	 */
2050	args = state->args;
2051	ASSERT(args != NULL);
2052	save_info = save_blk->bp->b_addr;
2053	drop_info = drop_blk->bp->b_addr;
2054	ASSERT(save_blk->magic == XFS_DA_NODE_MAGIC ||
2055	       save_blk->magic == XFS_DIR2_LEAFN_MAGIC ||
2056	       save_blk->magic == XFS_ATTR_LEAF_MAGIC);
2057	ASSERT(save_blk->magic == drop_blk->magic);
2058	ASSERT((be32_to_cpu(save_info->forw) == drop_blk->blkno) ||
2059	       (be32_to_cpu(save_info->back) == drop_blk->blkno));
2060	ASSERT((be32_to_cpu(drop_info->forw) == save_blk->blkno) ||
2061	       (be32_to_cpu(drop_info->back) == save_blk->blkno));
2062
2063	/*
2064	 * Unlink the leaf block from the doubly linked chain of leaves.
2065	 */
2066	if (be32_to_cpu(save_info->back) == drop_blk->blkno) {
2067		trace_xfs_da_unlink_back(args);
2068		save_info->back = drop_info->back;
2069		if (drop_info->back) {
2070			error = xfs_da3_node_read(args->trans, args->dp,
2071						be32_to_cpu(drop_info->back),
2072						&bp, args->whichfork);
2073			if (error)
2074				return error;
2075			fa = xfs_da3_header_check(bp, args->owner);
2076			if (fa) {
2077				__xfs_buf_mark_corrupt(bp, fa);
2078				xfs_trans_brelse(args->trans, bp);
2079				xfs_da_mark_sick(args);
2080				return -EFSCORRUPTED;
2081			}
2082			ASSERT(bp != NULL);
2083			tmp_info = bp->b_addr;
2084			ASSERT(tmp_info->magic == save_info->magic);
2085			ASSERT(be32_to_cpu(tmp_info->forw) == drop_blk->blkno);
2086			tmp_info->forw = cpu_to_be32(save_blk->blkno);
2087			xfs_trans_log_buf(args->trans, bp, 0,
2088						    sizeof(*tmp_info) - 1);
2089		}
2090	} else {
2091		trace_xfs_da_unlink_forward(args);
2092		save_info->forw = drop_info->forw;
2093		if (drop_info->forw) {
2094			error = xfs_da3_node_read(args->trans, args->dp,
2095						be32_to_cpu(drop_info->forw),
2096						&bp, args->whichfork);
2097			if (error)
2098				return error;
2099			fa = xfs_da3_header_check(bp, args->owner);
2100			if (fa) {
2101				__xfs_buf_mark_corrupt(bp, fa);
2102				xfs_trans_brelse(args->trans, bp);
2103				xfs_da_mark_sick(args);
2104				return -EFSCORRUPTED;
2105			}
2106			ASSERT(bp != NULL);
2107			tmp_info = bp->b_addr;
2108			ASSERT(tmp_info->magic == save_info->magic);
2109			ASSERT(be32_to_cpu(tmp_info->back) == drop_blk->blkno);
2110			tmp_info->back = cpu_to_be32(save_blk->blkno);
2111			xfs_trans_log_buf(args->trans, bp, 0,
2112						    sizeof(*tmp_info) - 1);
2113		}
2114	}
2115
2116	xfs_trans_log_buf(args->trans, save_blk->bp, 0, sizeof(*save_info) - 1);
2117	return 0;
2118}
2119
2120/*
2121 * Move a path "forward" or "!forward" one block at the current level.
2122 *
2123 * This routine will adjust a "path" to point to the next block
2124 * "forward" (higher hashvalues) or "!forward" (lower hashvals) in the
2125 * Btree, including updating pointers to the intermediate nodes between
2126 * the new bottom and the root.
2127 */
2128int							/* error */
2129xfs_da3_path_shift(
2130	struct xfs_da_state	*state,
2131	struct xfs_da_state_path *path,
2132	int			forward,
2133	int			release,
2134	int			*result)
2135{
2136	struct xfs_da_state_blk	*blk;
2137	struct xfs_da_blkinfo	*info;
2138	struct xfs_da_args	*args;
2139	struct xfs_da_node_entry *btree;
2140	struct xfs_da3_icnode_hdr nodehdr;
2141	struct xfs_buf		*bp;
2142	xfs_failaddr_t		fa;
2143	xfs_dablk_t		blkno = 0;
2144	int			level;
2145	int			error;
2146	struct xfs_inode	*dp = state->args->dp;
2147
2148	trace_xfs_da_path_shift(state->args);
2149
2150	/*
2151	 * Roll up the Btree looking for the first block where our
2152	 * current index is not at the edge of the block.  Note that
2153	 * we skip the bottom layer because we want the sibling block.
2154	 */
2155	args = state->args;
2156	ASSERT(args != NULL);
2157	ASSERT(path != NULL);
2158	ASSERT((path->active > 0) && (path->active < XFS_DA_NODE_MAXDEPTH));
2159	level = (path->active-1) - 1;	/* skip bottom layer in path */
2160	for (; level >= 0; level--) {
2161		blk = &path->blk[level];
2162		xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr,
2163					   blk->bp->b_addr);
2164
2165		if (forward && (blk->index < nodehdr.count - 1)) {
2166			blk->index++;
2167			blkno = be32_to_cpu(nodehdr.btree[blk->index].before);
2168			break;
2169		} else if (!forward && (blk->index > 0)) {
2170			blk->index--;
2171			blkno = be32_to_cpu(nodehdr.btree[blk->index].before);
2172			break;
2173		}
2174	}
2175	if (level < 0) {
2176		*result = -ENOENT;	/* we're out of our tree */
2177		ASSERT(args->op_flags & XFS_DA_OP_OKNOENT);
2178		return 0;
2179	}
2180
2181	/*
2182	 * Roll down the edge of the subtree until we reach the
2183	 * same depth we were at originally.
2184	 */
2185	for (blk++, level++; level < path->active; blk++, level++) {
2186		/*
2187		 * Read the next child block into a local buffer.
2188		 */
2189		error = xfs_da3_node_read(args->trans, dp, blkno, &bp,
2190					  args->whichfork);
2191		if (error)
2192			return error;
2193
2194		/*
2195		 * Release the old block (if it's dirty, the trans doesn't
2196		 * actually let go) and swap the local buffer into the path
2197		 * structure. This ensures failure of the above read doesn't set
2198		 * a NULL buffer in an active slot in the path.
2199		 */
2200		if (release)
2201			xfs_trans_brelse(args->trans, blk->bp);
2202		blk->blkno = blkno;
2203		blk->bp = bp;
2204
2205		info = blk->bp->b_addr;
2206		ASSERT(info->magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
2207		       info->magic == cpu_to_be16(XFS_DA3_NODE_MAGIC) ||
2208		       info->magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
2209		       info->magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC) ||
2210		       info->magic == cpu_to_be16(XFS_ATTR_LEAF_MAGIC) ||
2211		       info->magic == cpu_to_be16(XFS_ATTR3_LEAF_MAGIC));
2212
2213
2214		/*
2215		 * Note: we flatten the magic number to a single type so we
2216		 * don't have to compare against crc/non-crc types elsewhere.
2217		 */
2218		switch (be16_to_cpu(info->magic)) {
2219		case XFS_DA_NODE_MAGIC:
2220		case XFS_DA3_NODE_MAGIC:
2221			fa = xfs_da3_node_header_check(blk->bp, args->owner);
2222			if (fa) {
2223				__xfs_buf_mark_corrupt(blk->bp, fa);
2224				xfs_da_mark_sick(args);
2225				return -EFSCORRUPTED;
2226			}
2227			blk->magic = XFS_DA_NODE_MAGIC;
2228			xfs_da3_node_hdr_from_disk(dp->i_mount, &nodehdr,
2229						   bp->b_addr);
2230			btree = nodehdr.btree;
2231			blk->hashval = be32_to_cpu(btree[nodehdr.count - 1].hashval);
2232			if (forward)
2233				blk->index = 0;
2234			else
2235				blk->index = nodehdr.count - 1;
2236			blkno = be32_to_cpu(btree[blk->index].before);
2237			break;
2238		case XFS_ATTR_LEAF_MAGIC:
2239		case XFS_ATTR3_LEAF_MAGIC:
2240			fa = xfs_attr3_leaf_header_check(blk->bp, args->owner);
2241			if (fa) {
2242				__xfs_buf_mark_corrupt(blk->bp, fa);
2243				xfs_da_mark_sick(args);
2244				return -EFSCORRUPTED;
2245			}
2246			blk->magic = XFS_ATTR_LEAF_MAGIC;
2247			ASSERT(level == path->active-1);
2248			blk->index = 0;
2249			blk->hashval = xfs_attr_leaf_lasthash(blk->bp, NULL);
2250			break;
2251		case XFS_DIR2_LEAFN_MAGIC:
2252		case XFS_DIR3_LEAFN_MAGIC:
2253			fa = xfs_dir3_leaf_header_check(blk->bp, args->owner);
2254			if (fa) {
2255				__xfs_buf_mark_corrupt(blk->bp, fa);
2256				xfs_da_mark_sick(args);
2257				return -EFSCORRUPTED;
2258			}
2259			blk->magic = XFS_DIR2_LEAFN_MAGIC;
2260			ASSERT(level == path->active-1);
2261			blk->index = 0;
2262			blk->hashval = xfs_dir2_leaf_lasthash(args->dp,
2263							      blk->bp, NULL);
2264			break;
2265		default:
2266			ASSERT(0);
2267			break;
2268		}
2269	}
2270	*result = 0;
2271	return 0;
2272}
2273
2274
2275/*========================================================================
2276 * Utility routines.
2277 *========================================================================*/
2278
2279/*
2280 * Implement a simple hash on a character string.
2281 * Rotate the hash value by 7 bits, then XOR each character in.
2282 * This is implemented with some source-level loop unrolling.
2283 */
2284xfs_dahash_t
2285xfs_da_hashname(const uint8_t *name, int namelen)
2286{
2287	xfs_dahash_t hash;
2288
2289	/*
2290	 * Do four characters at a time as long as we can.
2291	 */
2292	for (hash = 0; namelen >= 4; namelen -= 4, name += 4)
2293		hash = (name[0] << 21) ^ (name[1] << 14) ^ (name[2] << 7) ^
2294		       (name[3] << 0) ^ rol32(hash, 7 * 4);
2295
2296	/*
2297	 * Now do the rest of the characters.
2298	 */
2299	switch (namelen) {
2300	case 3:
2301		return (name[0] << 14) ^ (name[1] << 7) ^ (name[2] << 0) ^
2302		       rol32(hash, 7 * 3);
2303	case 2:
2304		return (name[0] << 7) ^ (name[1] << 0) ^ rol32(hash, 7 * 2);
2305	case 1:
2306		return (name[0] << 0) ^ rol32(hash, 7 * 1);
2307	default: /* case 0: */
2308		return hash;
2309	}
2310}
2311
2312enum xfs_dacmp
2313xfs_da_compname(
2314	struct xfs_da_args *args,
2315	const unsigned char *name,
2316	int		len)
2317{
2318	return (args->namelen == len && memcmp(args->name, name, len) == 0) ?
2319					XFS_CMP_EXACT : XFS_CMP_DIFFERENT;
2320}
2321
2322int
2323xfs_da_grow_inode_int(
2324	struct xfs_da_args	*args,
2325	xfs_fileoff_t		*bno,
2326	int			count)
2327{
2328	struct xfs_trans	*tp = args->trans;
2329	struct xfs_inode	*dp = args->dp;
2330	int			w = args->whichfork;
2331	xfs_rfsblock_t		nblks = dp->i_nblocks;
2332	struct xfs_bmbt_irec	map, *mapp = &map;
2333	int			nmap, error, got, i, mapi = 1;
2334
2335	/*
2336	 * Find a spot in the file space to put the new block.
2337	 */
2338	error = xfs_bmap_first_unused(tp, dp, count, bno, w);
2339	if (error)
2340		return error;
2341
2342	/*
2343	 * Try mapping it in one filesystem block.
2344	 */
2345	nmap = 1;
2346	error = xfs_bmapi_write(tp, dp, *bno, count,
2347			xfs_bmapi_aflag(w)|XFS_BMAPI_METADATA|XFS_BMAPI_CONTIG,
2348			args->total, &map, &nmap);
2349	if (error == -ENOSPC && count > 1) {
2350		xfs_fileoff_t		b;
2351		int			c;
2352
2353		/*
2354		 * If we didn't get it and the block might work if fragmented,
2355		 * try without the CONTIG flag.  Loop until we get it all.
2356		 */
2357		mapp = kmalloc(sizeof(*mapp) * count,
2358				GFP_KERNEL | __GFP_NOFAIL);
2359		for (b = *bno, mapi = 0; b < *bno + count; ) {
2360			c = (int)(*bno + count - b);
2361			nmap = min(XFS_BMAP_MAX_NMAP, c);
2362			error = xfs_bmapi_write(tp, dp, b, c,
2363					xfs_bmapi_aflag(w)|XFS_BMAPI_METADATA,
2364					args->total, &mapp[mapi], &nmap);
2365			if (error)
2366				goto out_free_map;
2367			mapi += nmap;
2368			b = mapp[mapi - 1].br_startoff +
2369			    mapp[mapi - 1].br_blockcount;
2370		}
2371	}
2372	if (error)
2373		goto out_free_map;
2374
2375	/*
2376	 * Count the blocks we got, make sure it matches the total.
2377	 */
2378	for (i = 0, got = 0; i < mapi; i++)
2379		got += mapp[i].br_blockcount;
2380	if (got != count || mapp[0].br_startoff != *bno ||
2381	    mapp[mapi - 1].br_startoff + mapp[mapi - 1].br_blockcount !=
2382	    *bno + count) {
2383		error = -ENOSPC;
2384		goto out_free_map;
2385	}
2386
2387	/* account for newly allocated blocks in reserved blocks total */
2388	args->total -= dp->i_nblocks - nblks;
2389
2390out_free_map:
2391	if (mapp != &map)
2392		kfree(mapp);
2393	return error;
2394}
2395
2396/*
2397 * Add a block to the btree ahead of the file.
2398 * Return the new block number to the caller.
2399 */
2400int
2401xfs_da_grow_inode(
2402	struct xfs_da_args	*args,
2403	xfs_dablk_t		*new_blkno)
2404{
2405	xfs_fileoff_t		bno;
2406	int			error;
2407
2408	trace_xfs_da_grow_inode(args);
2409
2410	bno = args->geo->leafblk;
2411	error = xfs_da_grow_inode_int(args, &bno, args->geo->fsbcount);
2412	if (!error)
2413		*new_blkno = (xfs_dablk_t)bno;
2414	return error;
2415}
2416
2417/*
2418 * Ick.  We need to always be able to remove a btree block, even
2419 * if there's no space reservation because the filesystem is full.
2420 * This is called if xfs_bunmapi on a btree block fails due to ENOSPC.
2421 * It swaps the target block with the last block in the file.  The
2422 * last block in the file can always be removed since it can't cause
2423 * a bmap btree split to do that.
2424 */
2425STATIC int
2426xfs_da3_swap_lastblock(
2427	struct xfs_da_args	*args,
2428	xfs_dablk_t		*dead_blknop,
2429	struct xfs_buf		**dead_bufp)
2430{
2431	struct xfs_da_blkinfo	*dead_info;
2432	struct xfs_da_blkinfo	*sib_info;
2433	struct xfs_da_intnode	*par_node;
2434	struct xfs_da_intnode	*dead_node;
2435	struct xfs_dir2_leaf	*dead_leaf2;
2436	struct xfs_da_node_entry *btree;
2437	struct xfs_da3_icnode_hdr par_hdr;
2438	struct xfs_inode	*dp;
2439	struct xfs_trans	*tp;
2440	struct xfs_mount	*mp;
2441	struct xfs_buf		*dead_buf;
2442	struct xfs_buf		*last_buf;
2443	struct xfs_buf		*sib_buf;
2444	struct xfs_buf		*par_buf;
2445	xfs_failaddr_t		fa;
2446	xfs_dahash_t		dead_hash;
2447	xfs_fileoff_t		lastoff;
2448	xfs_dablk_t		dead_blkno;
2449	xfs_dablk_t		last_blkno;
2450	xfs_dablk_t		sib_blkno;
2451	xfs_dablk_t		par_blkno;
2452	int			error;
2453	int			w;
2454	int			entno;
2455	int			level;
2456	int			dead_level;
2457
2458	trace_xfs_da_swap_lastblock(args);
2459
2460	dead_buf = *dead_bufp;
2461	dead_blkno = *dead_blknop;
2462	tp = args->trans;
2463	dp = args->dp;
2464	w = args->whichfork;
2465	ASSERT(w == XFS_DATA_FORK);
2466	mp = dp->i_mount;
2467	lastoff = args->geo->freeblk;
2468	error = xfs_bmap_last_before(tp, dp, &lastoff, w);
2469	if (error)
2470		return error;
2471	if (XFS_IS_CORRUPT(mp, lastoff == 0)) {
2472		xfs_da_mark_sick(args);
2473		return -EFSCORRUPTED;
2474	}
2475	/*
2476	 * Read the last block in the btree space.
2477	 */
2478	last_blkno = (xfs_dablk_t)lastoff - args->geo->fsbcount;
2479	error = xfs_da3_node_read(tp, dp, last_blkno, &last_buf, w);
2480	if (error)
2481		return error;
2482	fa = xfs_da3_header_check(last_buf, args->owner);
2483	if (fa) {
2484		__xfs_buf_mark_corrupt(last_buf, fa);
2485		xfs_trans_brelse(tp, last_buf);
2486		xfs_da_mark_sick(args);
2487		return -EFSCORRUPTED;
2488	}
2489
2490	/*
2491	 * Copy the last block into the dead buffer and log it.
2492	 */
2493	xfs_da_buf_copy(dead_buf, last_buf, args->geo->blksize);
2494	xfs_trans_log_buf(tp, dead_buf, 0, args->geo->blksize - 1);
2495	dead_info = dead_buf->b_addr;
2496
2497	/*
2498	 * Get values from the moved block.
2499	 */
2500	if (dead_info->magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
2501	    dead_info->magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
2502		struct xfs_dir3_icleaf_hdr leafhdr;
2503		struct xfs_dir2_leaf_entry *ents;
2504
2505		dead_leaf2 = (xfs_dir2_leaf_t *)dead_info;
2506		xfs_dir2_leaf_hdr_from_disk(dp->i_mount, &leafhdr,
2507					    dead_leaf2);
2508		ents = leafhdr.ents;
2509		dead_level = 0;
2510		dead_hash = be32_to_cpu(ents[leafhdr.count - 1].hashval);
2511	} else {
2512		struct xfs_da3_icnode_hdr deadhdr;
2513
2514		dead_node = (xfs_da_intnode_t *)dead_info;
2515		xfs_da3_node_hdr_from_disk(dp->i_mount, &deadhdr, dead_node);
2516		btree = deadhdr.btree;
2517		dead_level = deadhdr.level;
2518		dead_hash = be32_to_cpu(btree[deadhdr.count - 1].hashval);
2519	}
2520	sib_buf = par_buf = NULL;
2521	/*
2522	 * If the moved block has a left sibling, fix up the pointers.
2523	 */
2524	if ((sib_blkno = be32_to_cpu(dead_info->back))) {
2525		error = xfs_da3_node_read(tp, dp, sib_blkno, &sib_buf, w);
2526		if (error)
2527			goto done;
2528		fa = xfs_da3_header_check(sib_buf, args->owner);
2529		if (fa) {
2530			__xfs_buf_mark_corrupt(sib_buf, fa);
2531			xfs_da_mark_sick(args);
2532			error = -EFSCORRUPTED;
2533			goto done;
2534		}
2535		sib_info = sib_buf->b_addr;
2536		if (XFS_IS_CORRUPT(mp,
2537				   be32_to_cpu(sib_info->forw) != last_blkno ||
2538				   sib_info->magic != dead_info->magic)) {
2539			xfs_da_mark_sick(args);
2540			error = -EFSCORRUPTED;
2541			goto done;
2542		}
2543		sib_info->forw = cpu_to_be32(dead_blkno);
2544		xfs_trans_log_buf(tp, sib_buf,
2545			XFS_DA_LOGRANGE(sib_info, &sib_info->forw,
2546					sizeof(sib_info->forw)));
2547		sib_buf = NULL;
2548	}
2549	/*
2550	 * If the moved block has a right sibling, fix up the pointers.
2551	 */
2552	if ((sib_blkno = be32_to_cpu(dead_info->forw))) {
2553		error = xfs_da3_node_read(tp, dp, sib_blkno, &sib_buf, w);
2554		if (error)
2555			goto done;
2556		fa = xfs_da3_header_check(sib_buf, args->owner);
2557		if (fa) {
2558			__xfs_buf_mark_corrupt(sib_buf, fa);
2559			xfs_da_mark_sick(args);
2560			error = -EFSCORRUPTED;
2561			goto done;
2562		}
2563		sib_info = sib_buf->b_addr;
2564		if (XFS_IS_CORRUPT(mp,
2565				   be32_to_cpu(sib_info->back) != last_blkno ||
2566				   sib_info->magic != dead_info->magic)) {
2567			xfs_da_mark_sick(args);
2568			error = -EFSCORRUPTED;
2569			goto done;
2570		}
2571		sib_info->back = cpu_to_be32(dead_blkno);
2572		xfs_trans_log_buf(tp, sib_buf,
2573			XFS_DA_LOGRANGE(sib_info, &sib_info->back,
2574					sizeof(sib_info->back)));
2575		sib_buf = NULL;
2576	}
2577	par_blkno = args->geo->leafblk;
2578	level = -1;
2579	/*
2580	 * Walk down the tree looking for the parent of the moved block.
2581	 */
2582	for (;;) {
2583		error = xfs_da3_node_read(tp, dp, par_blkno, &par_buf, w);
2584		if (error)
2585			goto done;
2586		fa = xfs_da3_node_header_check(par_buf, args->owner);
2587		if (fa) {
2588			__xfs_buf_mark_corrupt(par_buf, fa);
2589			xfs_da_mark_sick(args);
2590			error = -EFSCORRUPTED;
2591			goto done;
2592		}
2593		par_node = par_buf->b_addr;
2594		xfs_da3_node_hdr_from_disk(dp->i_mount, &par_hdr, par_node);
2595		if (XFS_IS_CORRUPT(mp,
2596				   level >= 0 && level != par_hdr.level + 1)) {
2597			xfs_da_mark_sick(args);
2598			error = -EFSCORRUPTED;
2599			goto done;
2600		}
2601		level = par_hdr.level;
2602		btree = par_hdr.btree;
2603		for (entno = 0;
2604		     entno < par_hdr.count &&
2605		     be32_to_cpu(btree[entno].hashval) < dead_hash;
2606		     entno++)
2607			continue;
2608		if (XFS_IS_CORRUPT(mp, entno == par_hdr.count)) {
2609			xfs_da_mark_sick(args);
2610			error = -EFSCORRUPTED;
2611			goto done;
2612		}
2613		par_blkno = be32_to_cpu(btree[entno].before);
2614		if (level == dead_level + 1)
2615			break;
2616		xfs_trans_brelse(tp, par_buf);
2617		par_buf = NULL;
2618	}
2619	/*
2620	 * We're in the right parent block.
2621	 * Look for the right entry.
2622	 */
2623	for (;;) {
2624		for (;
2625		     entno < par_hdr.count &&
2626		     be32_to_cpu(btree[entno].before) != last_blkno;
2627		     entno++)
2628			continue;
2629		if (entno < par_hdr.count)
2630			break;
2631		par_blkno = par_hdr.forw;
2632		xfs_trans_brelse(tp, par_buf);
2633		par_buf = NULL;
2634		if (XFS_IS_CORRUPT(mp, par_blkno == 0)) {
2635			xfs_da_mark_sick(args);
2636			error = -EFSCORRUPTED;
2637			goto done;
2638		}
2639		error = xfs_da3_node_read(tp, dp, par_blkno, &par_buf, w);
2640		if (error)
2641			goto done;
2642		fa = xfs_da3_node_header_check(par_buf, args->owner);
2643		if (fa) {
2644			__xfs_buf_mark_corrupt(par_buf, fa);
2645			xfs_da_mark_sick(args);
2646			error = -EFSCORRUPTED;
2647			goto done;
2648		}
2649		par_node = par_buf->b_addr;
2650		xfs_da3_node_hdr_from_disk(dp->i_mount, &par_hdr, par_node);
2651		if (XFS_IS_CORRUPT(mp, par_hdr.level != level)) {
2652			xfs_da_mark_sick(args);
2653			error = -EFSCORRUPTED;
2654			goto done;
2655		}
2656		btree = par_hdr.btree;
2657		entno = 0;
2658	}
2659	/*
2660	 * Update the parent entry pointing to the moved block.
2661	 */
2662	btree[entno].before = cpu_to_be32(dead_blkno);
2663	xfs_trans_log_buf(tp, par_buf,
2664		XFS_DA_LOGRANGE(par_node, &btree[entno].before,
2665				sizeof(btree[entno].before)));
2666	*dead_blknop = last_blkno;
2667	*dead_bufp = last_buf;
2668	return 0;
2669done:
2670	if (par_buf)
2671		xfs_trans_brelse(tp, par_buf);
2672	if (sib_buf)
2673		xfs_trans_brelse(tp, sib_buf);
2674	xfs_trans_brelse(tp, last_buf);
2675	return error;
2676}
2677
2678/*
2679 * Remove a btree block from a directory or attribute.
2680 */
2681int
2682xfs_da_shrink_inode(
2683	struct xfs_da_args	*args,
2684	xfs_dablk_t		dead_blkno,
2685	struct xfs_buf		*dead_buf)
2686{
2687	struct xfs_inode	*dp;
2688	int			done, error, w, count;
2689	struct xfs_trans	*tp;
2690
2691	trace_xfs_da_shrink_inode(args);
2692
2693	dp = args->dp;
2694	w = args->whichfork;
2695	tp = args->trans;
2696	count = args->geo->fsbcount;
2697	for (;;) {
2698		/*
2699		 * Remove extents.  If we get ENOSPC for a dir we have to move
2700		 * the last block to the place we want to kill.
2701		 */
2702		error = xfs_bunmapi(tp, dp, dead_blkno, count,
2703				    xfs_bmapi_aflag(w), 0, &done);
2704		if (error == -ENOSPC) {
2705			if (w != XFS_DATA_FORK)
2706				break;
2707			error = xfs_da3_swap_lastblock(args, &dead_blkno,
2708						      &dead_buf);
2709			if (error)
2710				break;
2711		} else {
2712			break;
2713		}
2714	}
2715	xfs_trans_binval(tp, dead_buf);
2716	return error;
2717}
2718
2719static int
2720xfs_dabuf_map(
2721	struct xfs_inode	*dp,
2722	xfs_dablk_t		bno,
2723	unsigned int		flags,
2724	int			whichfork,
2725	struct xfs_buf_map	**mapp,
2726	int			*nmaps)
2727{
2728	struct xfs_mount	*mp = dp->i_mount;
2729	int			nfsb = xfs_dabuf_nfsb(mp, whichfork);
2730	struct xfs_bmbt_irec	irec, *irecs = &irec;
2731	struct xfs_buf_map	*map = *mapp;
2732	xfs_fileoff_t		off = bno;
2733	int			error = 0, nirecs, i;
2734
2735	if (nfsb > 1)
2736		irecs = kzalloc(sizeof(irec) * nfsb,
2737				GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOFAIL);
2738
2739	nirecs = nfsb;
2740	error = xfs_bmapi_read(dp, bno, nfsb, irecs, &nirecs,
2741			xfs_bmapi_aflag(whichfork));
2742	if (error)
2743		goto out_free_irecs;
2744
2745	/*
2746	 * Use the caller provided map for the single map case, else allocate a
2747	 * larger one that needs to be free by the caller.
2748	 */
2749	if (nirecs > 1) {
2750		map = kzalloc(nirecs * sizeof(struct xfs_buf_map),
2751				GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOFAIL);
2752		if (!map) {
2753			error = -ENOMEM;
2754			goto out_free_irecs;
2755		}
2756		*mapp = map;
2757	}
2758
2759	for (i = 0; i < nirecs; i++) {
2760		if (irecs[i].br_startblock == HOLESTARTBLOCK ||
2761		    irecs[i].br_startblock == DELAYSTARTBLOCK)
2762			goto invalid_mapping;
2763		if (off != irecs[i].br_startoff)
2764			goto invalid_mapping;
2765
2766		map[i].bm_bn = XFS_FSB_TO_DADDR(mp, irecs[i].br_startblock);
2767		map[i].bm_len = XFS_FSB_TO_BB(mp, irecs[i].br_blockcount);
2768		off += irecs[i].br_blockcount;
2769	}
2770
2771	if (off != bno + nfsb)
2772		goto invalid_mapping;
2773
2774	*nmaps = nirecs;
2775out_free_irecs:
2776	if (irecs != &irec)
2777		kfree(irecs);
2778	return error;
2779
2780invalid_mapping:
2781	/* Caller ok with no mapping. */
2782	if (XFS_IS_CORRUPT(mp, !(flags & XFS_DABUF_MAP_HOLE_OK))) {
2783		xfs_dirattr_mark_sick(dp, whichfork);
2784		error = -EFSCORRUPTED;
2785		if (xfs_error_level >= XFS_ERRLEVEL_LOW) {
2786			xfs_alert(mp, "%s: bno %u inode %llu",
2787					__func__, bno, dp->i_ino);
2788
2789			for (i = 0; i < nirecs; i++) {
2790				xfs_alert(mp,
2791"[%02d] br_startoff %lld br_startblock %lld br_blockcount %lld br_state %d",
2792					i, irecs[i].br_startoff,
2793					irecs[i].br_startblock,
2794					irecs[i].br_blockcount,
2795					irecs[i].br_state);
2796			}
2797		}
2798	} else {
2799		*nmaps = 0;
2800	}
2801	goto out_free_irecs;
2802}
2803
2804/*
2805 * Get a buffer for the dir/attr block.
2806 */
2807int
2808xfs_da_get_buf(
2809	struct xfs_trans	*tp,
2810	struct xfs_inode	*dp,
2811	xfs_dablk_t		bno,
2812	struct xfs_buf		**bpp,
2813	int			whichfork)
2814{
2815	struct xfs_mount	*mp = dp->i_mount;
2816	struct xfs_buf		*bp;
2817	struct xfs_buf_map	map, *mapp = &map;
2818	int			nmap = 1;
2819	int			error;
2820
2821	*bpp = NULL;
2822	error = xfs_dabuf_map(dp, bno, 0, whichfork, &mapp, &nmap);
2823	if (error || nmap == 0)
2824		goto out_free;
2825
2826	error = xfs_trans_get_buf_map(tp, mp->m_ddev_targp, mapp, nmap, 0, &bp);
2827	if (error)
2828		goto out_free;
2829
2830	*bpp = bp;
2831
2832out_free:
2833	if (mapp != &map)
2834		kfree(mapp);
2835
2836	return error;
2837}
2838
2839/*
2840 * Get a buffer for the dir/attr block, fill in the contents.
2841 */
2842int
2843xfs_da_read_buf(
2844	struct xfs_trans	*tp,
2845	struct xfs_inode	*dp,
2846	xfs_dablk_t		bno,
2847	unsigned int		flags,
2848	struct xfs_buf		**bpp,
2849	int			whichfork,
2850	const struct xfs_buf_ops *ops)
2851{
2852	struct xfs_mount	*mp = dp->i_mount;
2853	struct xfs_buf		*bp;
2854	struct xfs_buf_map	map, *mapp = &map;
2855	int			nmap = 1;
2856	int			error;
2857
2858	*bpp = NULL;
2859	error = xfs_dabuf_map(dp, bno, flags, whichfork, &mapp, &nmap);
2860	if (error || !nmap)
2861		goto out_free;
2862
2863	error = xfs_trans_read_buf_map(mp, tp, mp->m_ddev_targp, mapp, nmap, 0,
2864			&bp, ops);
2865	if (xfs_metadata_is_sick(error))
2866		xfs_dirattr_mark_sick(dp, whichfork);
2867	/*
2868	 * ENODATA from disk implies a disk medium failure; ENODATA for
2869	 * xattrs means attribute not found, so disambiguate that here.
2870	 */
2871	if (error == -ENODATA && whichfork == XFS_ATTR_FORK)
2872		error = -EIO;
2873	if (error)
2874		goto out_free;
2875
2876	if (whichfork == XFS_ATTR_FORK)
2877		xfs_buf_set_ref(bp, XFS_ATTR_BTREE_REF);
2878	else
2879		xfs_buf_set_ref(bp, XFS_DIR_BTREE_REF);
2880	*bpp = bp;
2881out_free:
2882	if (mapp != &map)
2883		kfree(mapp);
2884
2885	return error;
2886}
2887
2888/*
2889 * Readahead the dir/attr block.
2890 */
2891int
2892xfs_da_reada_buf(
2893	struct xfs_inode	*dp,
2894	xfs_dablk_t		bno,
2895	unsigned int		flags,
2896	int			whichfork,
2897	const struct xfs_buf_ops *ops)
2898{
2899	struct xfs_buf_map	map;
2900	struct xfs_buf_map	*mapp;
2901	int			nmap;
2902	int			error;
2903
2904	mapp = &map;
2905	nmap = 1;
2906	error = xfs_dabuf_map(dp, bno, flags, whichfork, &mapp, &nmap);
2907	if (error || !nmap)
2908		goto out_free;
2909
2910	xfs_buf_readahead_map(dp->i_mount->m_ddev_targp, mapp, nmap, ops);
2911
2912out_free:
2913	if (mapp != &map)
2914		kfree(mapp);
2915
2916	return error;
2917}