xref: /openbmc/linux/fs/xfs/libxfs/xfs_da_btree.c (revision 51b67a6e)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4  * Copyright (c) 2013 Red Hat, Inc.
5  * All Rights Reserved.
6  */
7 #include "xfs.h"
8 #include "xfs_fs.h"
9 #include "xfs_shared.h"
10 #include "xfs_format.h"
11 #include "xfs_log_format.h"
12 #include "xfs_trans_resv.h"
13 #include "xfs_bit.h"
14 #include "xfs_mount.h"
15 #include "xfs_dir2.h"
16 #include "xfs_dir2_priv.h"
17 #include "xfs_inode.h"
18 #include "xfs_trans.h"
19 #include "xfs_bmap.h"
20 #include "xfs_attr_leaf.h"
21 #include "xfs_error.h"
22 #include "xfs_trace.h"
23 #include "xfs_buf_item.h"
24 #include "xfs_log.h"
25 
26 /*
27  * xfs_da_btree.c
28  *
29  * Routines to implement directories as Btrees of hashed names.
30  */
31 
32 /*========================================================================
33  * Function prototypes for the kernel.
34  *========================================================================*/
35 
36 /*
37  * Routines used for growing the Btree.
38  */
39 STATIC int xfs_da3_root_split(xfs_da_state_t *state,
40 					    xfs_da_state_blk_t *existing_root,
41 					    xfs_da_state_blk_t *new_child);
42 STATIC int xfs_da3_node_split(xfs_da_state_t *state,
43 					    xfs_da_state_blk_t *existing_blk,
44 					    xfs_da_state_blk_t *split_blk,
45 					    xfs_da_state_blk_t *blk_to_add,
46 					    int treelevel,
47 					    int *result);
48 STATIC void xfs_da3_node_rebalance(xfs_da_state_t *state,
49 					 xfs_da_state_blk_t *node_blk_1,
50 					 xfs_da_state_blk_t *node_blk_2);
51 STATIC void xfs_da3_node_add(xfs_da_state_t *state,
52 				   xfs_da_state_blk_t *old_node_blk,
53 				   xfs_da_state_blk_t *new_node_blk);
54 
55 /*
56  * Routines used for shrinking the Btree.
57  */
58 STATIC int xfs_da3_root_join(xfs_da_state_t *state,
59 					   xfs_da_state_blk_t *root_blk);
60 STATIC int xfs_da3_node_toosmall(xfs_da_state_t *state, int *retval);
61 STATIC void xfs_da3_node_remove(xfs_da_state_t *state,
62 					      xfs_da_state_blk_t *drop_blk);
63 STATIC void xfs_da3_node_unbalance(xfs_da_state_t *state,
64 					 xfs_da_state_blk_t *src_node_blk,
65 					 xfs_da_state_blk_t *dst_node_blk);
66 
67 /*
68  * Utility routines.
69  */
70 STATIC int	xfs_da3_blk_unlink(xfs_da_state_t *state,
71 				  xfs_da_state_blk_t *drop_blk,
72 				  xfs_da_state_blk_t *save_blk);
73 
74 
75 kmem_zone_t *xfs_da_state_zone;	/* anchor for state struct zone */
76 
77 /*
78  * Allocate a dir-state structure.
79  * We don't put them on the stack since they're large.
80  */
81 xfs_da_state_t *
82 xfs_da_state_alloc(void)
83 {
84 	return kmem_zone_zalloc(xfs_da_state_zone, KM_NOFS);
85 }
86 
87 /*
88  * Kill the altpath contents of a da-state structure.
89  */
90 STATIC void
91 xfs_da_state_kill_altpath(xfs_da_state_t *state)
92 {
93 	int	i;
94 
95 	for (i = 0; i < state->altpath.active; i++)
96 		state->altpath.blk[i].bp = NULL;
97 	state->altpath.active = 0;
98 }
99 
100 /*
101  * Free a da-state structure.
102  */
103 void
104 xfs_da_state_free(xfs_da_state_t *state)
105 {
106 	xfs_da_state_kill_altpath(state);
107 #ifdef DEBUG
108 	memset((char *)state, 0, sizeof(*state));
109 #endif /* DEBUG */
110 	kmem_zone_free(xfs_da_state_zone, state);
111 }
112 
113 /*
114  * Verify an xfs_da3_blkinfo structure. Note that the da3 fields are only
115  * accessible on v5 filesystems. This header format is common across da node,
116  * attr leaf and dir leaf blocks.
117  */
118 xfs_failaddr_t
119 xfs_da3_blkinfo_verify(
120 	struct xfs_buf		*bp,
121 	struct xfs_da3_blkinfo	*hdr3)
122 {
123 	struct xfs_mount	*mp = bp->b_mount;
124 	struct xfs_da_blkinfo	*hdr = &hdr3->hdr;
125 
126 	if (!xfs_verify_magic16(bp, hdr->magic))
127 		return __this_address;
128 
129 	if (xfs_sb_version_hascrc(&mp->m_sb)) {
130 		if (!uuid_equal(&hdr3->uuid, &mp->m_sb.sb_meta_uuid))
131 			return __this_address;
132 		if (be64_to_cpu(hdr3->blkno) != bp->b_bn)
133 			return __this_address;
134 		if (!xfs_log_check_lsn(mp, be64_to_cpu(hdr3->lsn)))
135 			return __this_address;
136 	}
137 
138 	return NULL;
139 }
140 
141 static xfs_failaddr_t
142 xfs_da3_node_verify(
143 	struct xfs_buf		*bp)
144 {
145 	struct xfs_mount	*mp = bp->b_mount;
146 	struct xfs_da_intnode	*hdr = bp->b_addr;
147 	struct xfs_da3_icnode_hdr ichdr;
148 	const struct xfs_dir_ops *ops;
149 	xfs_failaddr_t		fa;
150 
151 	ops = xfs_dir_get_ops(mp, NULL);
152 
153 	ops->node_hdr_from_disk(&ichdr, hdr);
154 
155 	fa = xfs_da3_blkinfo_verify(bp, bp->b_addr);
156 	if (fa)
157 		return fa;
158 
159 	if (ichdr.level == 0)
160 		return __this_address;
161 	if (ichdr.level > XFS_DA_NODE_MAXDEPTH)
162 		return __this_address;
163 	if (ichdr.count == 0)
164 		return __this_address;
165 
166 	/*
167 	 * we don't know if the node is for and attribute or directory tree,
168 	 * so only fail if the count is outside both bounds
169 	 */
170 	if (ichdr.count > mp->m_dir_geo->node_ents &&
171 	    ichdr.count > mp->m_attr_geo->node_ents)
172 		return __this_address;
173 
174 	/* XXX: hash order check? */
175 
176 	return NULL;
177 }
178 
179 static void
180 xfs_da3_node_write_verify(
181 	struct xfs_buf	*bp)
182 {
183 	struct xfs_mount	*mp = bp->b_mount;
184 	struct xfs_buf_log_item	*bip = bp->b_log_item;
185 	struct xfs_da3_node_hdr *hdr3 = bp->b_addr;
186 	xfs_failaddr_t		fa;
187 
188 	fa = xfs_da3_node_verify(bp);
189 	if (fa) {
190 		xfs_verifier_error(bp, -EFSCORRUPTED, fa);
191 		return;
192 	}
193 
194 	if (!xfs_sb_version_hascrc(&mp->m_sb))
195 		return;
196 
197 	if (bip)
198 		hdr3->info.lsn = cpu_to_be64(bip->bli_item.li_lsn);
199 
200 	xfs_buf_update_cksum(bp, XFS_DA3_NODE_CRC_OFF);
201 }
202 
203 /*
204  * leaf/node format detection on trees is sketchy, so a node read can be done on
205  * leaf level blocks when detection identifies the tree as a node format tree
206  * incorrectly. In this case, we need to swap the verifier to match the correct
207  * format of the block being read.
208  */
209 static void
210 xfs_da3_node_read_verify(
211 	struct xfs_buf		*bp)
212 {
213 	struct xfs_da_blkinfo	*info = bp->b_addr;
214 	xfs_failaddr_t		fa;
215 
216 	switch (be16_to_cpu(info->magic)) {
217 		case XFS_DA3_NODE_MAGIC:
218 			if (!xfs_buf_verify_cksum(bp, XFS_DA3_NODE_CRC_OFF)) {
219 				xfs_verifier_error(bp, -EFSBADCRC,
220 						__this_address);
221 				break;
222 			}
223 			/* fall through */
224 		case XFS_DA_NODE_MAGIC:
225 			fa = xfs_da3_node_verify(bp);
226 			if (fa)
227 				xfs_verifier_error(bp, -EFSCORRUPTED, fa);
228 			return;
229 		case XFS_ATTR_LEAF_MAGIC:
230 		case XFS_ATTR3_LEAF_MAGIC:
231 			bp->b_ops = &xfs_attr3_leaf_buf_ops;
232 			bp->b_ops->verify_read(bp);
233 			return;
234 		case XFS_DIR2_LEAFN_MAGIC:
235 		case XFS_DIR3_LEAFN_MAGIC:
236 			bp->b_ops = &xfs_dir3_leafn_buf_ops;
237 			bp->b_ops->verify_read(bp);
238 			return;
239 		default:
240 			xfs_verifier_error(bp, -EFSCORRUPTED, __this_address);
241 			break;
242 	}
243 }
244 
245 /* Verify the structure of a da3 block. */
246 static xfs_failaddr_t
247 xfs_da3_node_verify_struct(
248 	struct xfs_buf		*bp)
249 {
250 	struct xfs_da_blkinfo	*info = bp->b_addr;
251 
252 	switch (be16_to_cpu(info->magic)) {
253 	case XFS_DA3_NODE_MAGIC:
254 	case XFS_DA_NODE_MAGIC:
255 		return xfs_da3_node_verify(bp);
256 	case XFS_ATTR_LEAF_MAGIC:
257 	case XFS_ATTR3_LEAF_MAGIC:
258 		bp->b_ops = &xfs_attr3_leaf_buf_ops;
259 		return bp->b_ops->verify_struct(bp);
260 	case XFS_DIR2_LEAFN_MAGIC:
261 	case XFS_DIR3_LEAFN_MAGIC:
262 		bp->b_ops = &xfs_dir3_leafn_buf_ops;
263 		return bp->b_ops->verify_struct(bp);
264 	default:
265 		return __this_address;
266 	}
267 }
268 
269 const struct xfs_buf_ops xfs_da3_node_buf_ops = {
270 	.name = "xfs_da3_node",
271 	.magic16 = { cpu_to_be16(XFS_DA_NODE_MAGIC),
272 		     cpu_to_be16(XFS_DA3_NODE_MAGIC) },
273 	.verify_read = xfs_da3_node_read_verify,
274 	.verify_write = xfs_da3_node_write_verify,
275 	.verify_struct = xfs_da3_node_verify_struct,
276 };
277 
278 int
279 xfs_da3_node_read(
280 	struct xfs_trans	*tp,
281 	struct xfs_inode	*dp,
282 	xfs_dablk_t		bno,
283 	xfs_daddr_t		mappedbno,
284 	struct xfs_buf		**bpp,
285 	int			which_fork)
286 {
287 	int			err;
288 
289 	err = xfs_da_read_buf(tp, dp, bno, mappedbno, bpp,
290 					which_fork, &xfs_da3_node_buf_ops);
291 	if (!err && tp && *bpp) {
292 		struct xfs_da_blkinfo	*info = (*bpp)->b_addr;
293 		int			type;
294 
295 		switch (be16_to_cpu(info->magic)) {
296 		case XFS_DA_NODE_MAGIC:
297 		case XFS_DA3_NODE_MAGIC:
298 			type = XFS_BLFT_DA_NODE_BUF;
299 			break;
300 		case XFS_ATTR_LEAF_MAGIC:
301 		case XFS_ATTR3_LEAF_MAGIC:
302 			type = XFS_BLFT_ATTR_LEAF_BUF;
303 			break;
304 		case XFS_DIR2_LEAFN_MAGIC:
305 		case XFS_DIR3_LEAFN_MAGIC:
306 			type = XFS_BLFT_DIR_LEAFN_BUF;
307 			break;
308 		default:
309 			XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW,
310 					tp->t_mountp, info, sizeof(*info));
311 			xfs_trans_brelse(tp, *bpp);
312 			*bpp = NULL;
313 			return -EFSCORRUPTED;
314 		}
315 		xfs_trans_buf_set_type(tp, *bpp, type);
316 	}
317 	return err;
318 }
319 
320 /*========================================================================
321  * Routines used for growing the Btree.
322  *========================================================================*/
323 
324 /*
325  * Create the initial contents of an intermediate node.
326  */
327 int
328 xfs_da3_node_create(
329 	struct xfs_da_args	*args,
330 	xfs_dablk_t		blkno,
331 	int			level,
332 	struct xfs_buf		**bpp,
333 	int			whichfork)
334 {
335 	struct xfs_da_intnode	*node;
336 	struct xfs_trans	*tp = args->trans;
337 	struct xfs_mount	*mp = tp->t_mountp;
338 	struct xfs_da3_icnode_hdr ichdr = {0};
339 	struct xfs_buf		*bp;
340 	int			error;
341 	struct xfs_inode	*dp = args->dp;
342 
343 	trace_xfs_da_node_create(args);
344 	ASSERT(level <= XFS_DA_NODE_MAXDEPTH);
345 
346 	error = xfs_da_get_buf(tp, dp, blkno, -1, &bp, whichfork);
347 	if (error)
348 		return error;
349 	bp->b_ops = &xfs_da3_node_buf_ops;
350 	xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DA_NODE_BUF);
351 	node = bp->b_addr;
352 
353 	if (xfs_sb_version_hascrc(&mp->m_sb)) {
354 		struct xfs_da3_node_hdr *hdr3 = bp->b_addr;
355 
356 		memset(hdr3, 0, sizeof(struct xfs_da3_node_hdr));
357 		ichdr.magic = XFS_DA3_NODE_MAGIC;
358 		hdr3->info.blkno = cpu_to_be64(bp->b_bn);
359 		hdr3->info.owner = cpu_to_be64(args->dp->i_ino);
360 		uuid_copy(&hdr3->info.uuid, &mp->m_sb.sb_meta_uuid);
361 	} else {
362 		ichdr.magic = XFS_DA_NODE_MAGIC;
363 	}
364 	ichdr.level = level;
365 
366 	dp->d_ops->node_hdr_to_disk(node, &ichdr);
367 	xfs_trans_log_buf(tp, bp,
368 		XFS_DA_LOGRANGE(node, &node->hdr, dp->d_ops->node_hdr_size));
369 
370 	*bpp = bp;
371 	return 0;
372 }
373 
374 /*
375  * Split a leaf node, rebalance, then possibly split
376  * intermediate nodes, rebalance, etc.
377  */
378 int							/* error */
379 xfs_da3_split(
380 	struct xfs_da_state	*state)
381 {
382 	struct xfs_da_state_blk	*oldblk;
383 	struct xfs_da_state_blk	*newblk;
384 	struct xfs_da_state_blk	*addblk;
385 	struct xfs_da_intnode	*node;
386 	int			max;
387 	int			action = 0;
388 	int			error;
389 	int			i;
390 
391 	trace_xfs_da_split(state->args);
392 
393 	/*
394 	 * Walk back up the tree splitting/inserting/adjusting as necessary.
395 	 * If we need to insert and there isn't room, split the node, then
396 	 * decide which fragment to insert the new block from below into.
397 	 * Note that we may split the root this way, but we need more fixup.
398 	 */
399 	max = state->path.active - 1;
400 	ASSERT((max >= 0) && (max < XFS_DA_NODE_MAXDEPTH));
401 	ASSERT(state->path.blk[max].magic == XFS_ATTR_LEAF_MAGIC ||
402 	       state->path.blk[max].magic == XFS_DIR2_LEAFN_MAGIC);
403 
404 	addblk = &state->path.blk[max];		/* initial dummy value */
405 	for (i = max; (i >= 0) && addblk; state->path.active--, i--) {
406 		oldblk = &state->path.blk[i];
407 		newblk = &state->altpath.blk[i];
408 
409 		/*
410 		 * If a leaf node then
411 		 *     Allocate a new leaf node, then rebalance across them.
412 		 * else if an intermediate node then
413 		 *     We split on the last layer, must we split the node?
414 		 */
415 		switch (oldblk->magic) {
416 		case XFS_ATTR_LEAF_MAGIC:
417 			error = xfs_attr3_leaf_split(state, oldblk, newblk);
418 			if ((error != 0) && (error != -ENOSPC)) {
419 				return error;	/* GROT: attr is inconsistent */
420 			}
421 			if (!error) {
422 				addblk = newblk;
423 				break;
424 			}
425 			/*
426 			 * Entry wouldn't fit, split the leaf again. The new
427 			 * extrablk will be consumed by xfs_da3_node_split if
428 			 * the node is split.
429 			 */
430 			state->extravalid = 1;
431 			if (state->inleaf) {
432 				state->extraafter = 0;	/* before newblk */
433 				trace_xfs_attr_leaf_split_before(state->args);
434 				error = xfs_attr3_leaf_split(state, oldblk,
435 							    &state->extrablk);
436 			} else {
437 				state->extraafter = 1;	/* after newblk */
438 				trace_xfs_attr_leaf_split_after(state->args);
439 				error = xfs_attr3_leaf_split(state, newblk,
440 							    &state->extrablk);
441 			}
442 			if (error)
443 				return error;	/* GROT: attr inconsistent */
444 			addblk = newblk;
445 			break;
446 		case XFS_DIR2_LEAFN_MAGIC:
447 			error = xfs_dir2_leafn_split(state, oldblk, newblk);
448 			if (error)
449 				return error;
450 			addblk = newblk;
451 			break;
452 		case XFS_DA_NODE_MAGIC:
453 			error = xfs_da3_node_split(state, oldblk, newblk, addblk,
454 							 max - i, &action);
455 			addblk->bp = NULL;
456 			if (error)
457 				return error;	/* GROT: dir is inconsistent */
458 			/*
459 			 * Record the newly split block for the next time thru?
460 			 */
461 			if (action)
462 				addblk = newblk;
463 			else
464 				addblk = NULL;
465 			break;
466 		}
467 
468 		/*
469 		 * Update the btree to show the new hashval for this child.
470 		 */
471 		xfs_da3_fixhashpath(state, &state->path);
472 	}
473 	if (!addblk)
474 		return 0;
475 
476 	/*
477 	 * xfs_da3_node_split() should have consumed any extra blocks we added
478 	 * during a double leaf split in the attr fork. This is guaranteed as
479 	 * we can't be here if the attr fork only has a single leaf block.
480 	 */
481 	ASSERT(state->extravalid == 0 ||
482 	       state->path.blk[max].magic == XFS_DIR2_LEAFN_MAGIC);
483 
484 	/*
485 	 * Split the root node.
486 	 */
487 	ASSERT(state->path.active == 0);
488 	oldblk = &state->path.blk[0];
489 	error = xfs_da3_root_split(state, oldblk, addblk);
490 	if (error)
491 		goto out;
492 
493 	/*
494 	 * Update pointers to the node which used to be block 0 and just got
495 	 * bumped because of the addition of a new root node.  Note that the
496 	 * original block 0 could be at any position in the list of blocks in
497 	 * the tree.
498 	 *
499 	 * Note: the magic numbers and sibling pointers are in the same physical
500 	 * place for both v2 and v3 headers (by design). Hence it doesn't matter
501 	 * which version of the xfs_da_intnode structure we use here as the
502 	 * result will be the same using either structure.
503 	 */
504 	node = oldblk->bp->b_addr;
505 	if (node->hdr.info.forw) {
506 		if (be32_to_cpu(node->hdr.info.forw) != addblk->blkno) {
507 			error = -EFSCORRUPTED;
508 			goto out;
509 		}
510 		node = addblk->bp->b_addr;
511 		node->hdr.info.back = cpu_to_be32(oldblk->blkno);
512 		xfs_trans_log_buf(state->args->trans, addblk->bp,
513 				  XFS_DA_LOGRANGE(node, &node->hdr.info,
514 				  sizeof(node->hdr.info)));
515 	}
516 	node = oldblk->bp->b_addr;
517 	if (node->hdr.info.back) {
518 		if (be32_to_cpu(node->hdr.info.back) != addblk->blkno) {
519 			error = -EFSCORRUPTED;
520 			goto out;
521 		}
522 		node = addblk->bp->b_addr;
523 		node->hdr.info.forw = cpu_to_be32(oldblk->blkno);
524 		xfs_trans_log_buf(state->args->trans, addblk->bp,
525 				  XFS_DA_LOGRANGE(node, &node->hdr.info,
526 				  sizeof(node->hdr.info)));
527 	}
528 out:
529 	addblk->bp = NULL;
530 	return error;
531 }
532 
533 /*
534  * Split the root.  We have to create a new root and point to the two
535  * parts (the split old root) that we just created.  Copy block zero to
536  * the EOF, extending the inode in process.
537  */
538 STATIC int						/* error */
539 xfs_da3_root_split(
540 	struct xfs_da_state	*state,
541 	struct xfs_da_state_blk	*blk1,
542 	struct xfs_da_state_blk	*blk2)
543 {
544 	struct xfs_da_intnode	*node;
545 	struct xfs_da_intnode	*oldroot;
546 	struct xfs_da_node_entry *btree;
547 	struct xfs_da3_icnode_hdr nodehdr;
548 	struct xfs_da_args	*args;
549 	struct xfs_buf		*bp;
550 	struct xfs_inode	*dp;
551 	struct xfs_trans	*tp;
552 	struct xfs_dir2_leaf	*leaf;
553 	xfs_dablk_t		blkno;
554 	int			level;
555 	int			error;
556 	int			size;
557 
558 	trace_xfs_da_root_split(state->args);
559 
560 	/*
561 	 * Copy the existing (incorrect) block from the root node position
562 	 * to a free space somewhere.
563 	 */
564 	args = state->args;
565 	error = xfs_da_grow_inode(args, &blkno);
566 	if (error)
567 		return error;
568 
569 	dp = args->dp;
570 	tp = args->trans;
571 	error = xfs_da_get_buf(tp, dp, blkno, -1, &bp, args->whichfork);
572 	if (error)
573 		return error;
574 	node = bp->b_addr;
575 	oldroot = blk1->bp->b_addr;
576 	if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
577 	    oldroot->hdr.info.magic == cpu_to_be16(XFS_DA3_NODE_MAGIC)) {
578 		struct xfs_da3_icnode_hdr icnodehdr;
579 
580 		dp->d_ops->node_hdr_from_disk(&icnodehdr, oldroot);
581 		btree = dp->d_ops->node_tree_p(oldroot);
582 		size = (int)((char *)&btree[icnodehdr.count] - (char *)oldroot);
583 		level = icnodehdr.level;
584 
585 		/*
586 		 * we are about to copy oldroot to bp, so set up the type
587 		 * of bp while we know exactly what it will be.
588 		 */
589 		xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DA_NODE_BUF);
590 	} else {
591 		struct xfs_dir3_icleaf_hdr leafhdr;
592 		struct xfs_dir2_leaf_entry *ents;
593 
594 		leaf = (xfs_dir2_leaf_t *)oldroot;
595 		dp->d_ops->leaf_hdr_from_disk(&leafhdr, leaf);
596 		ents = dp->d_ops->leaf_ents_p(leaf);
597 
598 		ASSERT(leafhdr.magic == XFS_DIR2_LEAFN_MAGIC ||
599 		       leafhdr.magic == XFS_DIR3_LEAFN_MAGIC);
600 		size = (int)((char *)&ents[leafhdr.count] - (char *)leaf);
601 		level = 0;
602 
603 		/*
604 		 * we are about to copy oldroot to bp, so set up the type
605 		 * of bp while we know exactly what it will be.
606 		 */
607 		xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DIR_LEAFN_BUF);
608 	}
609 
610 	/*
611 	 * we can copy most of the information in the node from one block to
612 	 * another, but for CRC enabled headers we have to make sure that the
613 	 * block specific identifiers are kept intact. We update the buffer
614 	 * directly for this.
615 	 */
616 	memcpy(node, oldroot, size);
617 	if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DA3_NODE_MAGIC) ||
618 	    oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
619 		struct xfs_da3_intnode *node3 = (struct xfs_da3_intnode *)node;
620 
621 		node3->hdr.info.blkno = cpu_to_be64(bp->b_bn);
622 	}
623 	xfs_trans_log_buf(tp, bp, 0, size - 1);
624 
625 	bp->b_ops = blk1->bp->b_ops;
626 	xfs_trans_buf_copy_type(bp, blk1->bp);
627 	blk1->bp = bp;
628 	blk1->blkno = blkno;
629 
630 	/*
631 	 * Set up the new root node.
632 	 */
633 	error = xfs_da3_node_create(args,
634 		(args->whichfork == XFS_DATA_FORK) ? args->geo->leafblk : 0,
635 		level + 1, &bp, args->whichfork);
636 	if (error)
637 		return error;
638 
639 	node = bp->b_addr;
640 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
641 	btree = dp->d_ops->node_tree_p(node);
642 	btree[0].hashval = cpu_to_be32(blk1->hashval);
643 	btree[0].before = cpu_to_be32(blk1->blkno);
644 	btree[1].hashval = cpu_to_be32(blk2->hashval);
645 	btree[1].before = cpu_to_be32(blk2->blkno);
646 	nodehdr.count = 2;
647 	dp->d_ops->node_hdr_to_disk(node, &nodehdr);
648 
649 #ifdef DEBUG
650 	if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
651 	    oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
652 		ASSERT(blk1->blkno >= args->geo->leafblk &&
653 		       blk1->blkno < args->geo->freeblk);
654 		ASSERT(blk2->blkno >= args->geo->leafblk &&
655 		       blk2->blkno < args->geo->freeblk);
656 	}
657 #endif
658 
659 	/* Header is already logged by xfs_da_node_create */
660 	xfs_trans_log_buf(tp, bp,
661 		XFS_DA_LOGRANGE(node, btree, sizeof(xfs_da_node_entry_t) * 2));
662 
663 	return 0;
664 }
665 
666 /*
667  * Split the node, rebalance, then add the new entry.
668  */
669 STATIC int						/* error */
670 xfs_da3_node_split(
671 	struct xfs_da_state	*state,
672 	struct xfs_da_state_blk	*oldblk,
673 	struct xfs_da_state_blk	*newblk,
674 	struct xfs_da_state_blk	*addblk,
675 	int			treelevel,
676 	int			*result)
677 {
678 	struct xfs_da_intnode	*node;
679 	struct xfs_da3_icnode_hdr nodehdr;
680 	xfs_dablk_t		blkno;
681 	int			newcount;
682 	int			error;
683 	int			useextra;
684 	struct xfs_inode	*dp = state->args->dp;
685 
686 	trace_xfs_da_node_split(state->args);
687 
688 	node = oldblk->bp->b_addr;
689 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
690 
691 	/*
692 	 * With V2 dirs the extra block is data or freespace.
693 	 */
694 	useextra = state->extravalid && state->args->whichfork == XFS_ATTR_FORK;
695 	newcount = 1 + useextra;
696 	/*
697 	 * Do we have to split the node?
698 	 */
699 	if (nodehdr.count + newcount > state->args->geo->node_ents) {
700 		/*
701 		 * Allocate a new node, add to the doubly linked chain of
702 		 * nodes, then move some of our excess entries into it.
703 		 */
704 		error = xfs_da_grow_inode(state->args, &blkno);
705 		if (error)
706 			return error;	/* GROT: dir is inconsistent */
707 
708 		error = xfs_da3_node_create(state->args, blkno, treelevel,
709 					   &newblk->bp, state->args->whichfork);
710 		if (error)
711 			return error;	/* GROT: dir is inconsistent */
712 		newblk->blkno = blkno;
713 		newblk->magic = XFS_DA_NODE_MAGIC;
714 		xfs_da3_node_rebalance(state, oldblk, newblk);
715 		error = xfs_da3_blk_link(state, oldblk, newblk);
716 		if (error)
717 			return error;
718 		*result = 1;
719 	} else {
720 		*result = 0;
721 	}
722 
723 	/*
724 	 * Insert the new entry(s) into the correct block
725 	 * (updating last hashval in the process).
726 	 *
727 	 * xfs_da3_node_add() inserts BEFORE the given index,
728 	 * and as a result of using node_lookup_int() we always
729 	 * point to a valid entry (not after one), but a split
730 	 * operation always results in a new block whose hashvals
731 	 * FOLLOW the current block.
732 	 *
733 	 * If we had double-split op below us, then add the extra block too.
734 	 */
735 	node = oldblk->bp->b_addr;
736 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
737 	if (oldblk->index <= nodehdr.count) {
738 		oldblk->index++;
739 		xfs_da3_node_add(state, oldblk, addblk);
740 		if (useextra) {
741 			if (state->extraafter)
742 				oldblk->index++;
743 			xfs_da3_node_add(state, oldblk, &state->extrablk);
744 			state->extravalid = 0;
745 		}
746 	} else {
747 		newblk->index++;
748 		xfs_da3_node_add(state, newblk, addblk);
749 		if (useextra) {
750 			if (state->extraafter)
751 				newblk->index++;
752 			xfs_da3_node_add(state, newblk, &state->extrablk);
753 			state->extravalid = 0;
754 		}
755 	}
756 
757 	return 0;
758 }
759 
760 /*
761  * Balance the btree elements between two intermediate nodes,
762  * usually one full and one empty.
763  *
764  * NOTE: if blk2 is empty, then it will get the upper half of blk1.
765  */
766 STATIC void
767 xfs_da3_node_rebalance(
768 	struct xfs_da_state	*state,
769 	struct xfs_da_state_blk	*blk1,
770 	struct xfs_da_state_blk	*blk2)
771 {
772 	struct xfs_da_intnode	*node1;
773 	struct xfs_da_intnode	*node2;
774 	struct xfs_da_intnode	*tmpnode;
775 	struct xfs_da_node_entry *btree1;
776 	struct xfs_da_node_entry *btree2;
777 	struct xfs_da_node_entry *btree_s;
778 	struct xfs_da_node_entry *btree_d;
779 	struct xfs_da3_icnode_hdr nodehdr1;
780 	struct xfs_da3_icnode_hdr nodehdr2;
781 	struct xfs_trans	*tp;
782 	int			count;
783 	int			tmp;
784 	int			swap = 0;
785 	struct xfs_inode	*dp = state->args->dp;
786 
787 	trace_xfs_da_node_rebalance(state->args);
788 
789 	node1 = blk1->bp->b_addr;
790 	node2 = blk2->bp->b_addr;
791 	dp->d_ops->node_hdr_from_disk(&nodehdr1, node1);
792 	dp->d_ops->node_hdr_from_disk(&nodehdr2, node2);
793 	btree1 = dp->d_ops->node_tree_p(node1);
794 	btree2 = dp->d_ops->node_tree_p(node2);
795 
796 	/*
797 	 * Figure out how many entries need to move, and in which direction.
798 	 * Swap the nodes around if that makes it simpler.
799 	 */
800 	if (nodehdr1.count > 0 && nodehdr2.count > 0 &&
801 	    ((be32_to_cpu(btree2[0].hashval) < be32_to_cpu(btree1[0].hashval)) ||
802 	     (be32_to_cpu(btree2[nodehdr2.count - 1].hashval) <
803 			be32_to_cpu(btree1[nodehdr1.count - 1].hashval)))) {
804 		tmpnode = node1;
805 		node1 = node2;
806 		node2 = tmpnode;
807 		dp->d_ops->node_hdr_from_disk(&nodehdr1, node1);
808 		dp->d_ops->node_hdr_from_disk(&nodehdr2, node2);
809 		btree1 = dp->d_ops->node_tree_p(node1);
810 		btree2 = dp->d_ops->node_tree_p(node2);
811 		swap = 1;
812 	}
813 
814 	count = (nodehdr1.count - nodehdr2.count) / 2;
815 	if (count == 0)
816 		return;
817 	tp = state->args->trans;
818 	/*
819 	 * Two cases: high-to-low and low-to-high.
820 	 */
821 	if (count > 0) {
822 		/*
823 		 * Move elements in node2 up to make a hole.
824 		 */
825 		tmp = nodehdr2.count;
826 		if (tmp > 0) {
827 			tmp *= (uint)sizeof(xfs_da_node_entry_t);
828 			btree_s = &btree2[0];
829 			btree_d = &btree2[count];
830 			memmove(btree_d, btree_s, tmp);
831 		}
832 
833 		/*
834 		 * Move the req'd B-tree elements from high in node1 to
835 		 * low in node2.
836 		 */
837 		nodehdr2.count += count;
838 		tmp = count * (uint)sizeof(xfs_da_node_entry_t);
839 		btree_s = &btree1[nodehdr1.count - count];
840 		btree_d = &btree2[0];
841 		memcpy(btree_d, btree_s, tmp);
842 		nodehdr1.count -= count;
843 	} else {
844 		/*
845 		 * Move the req'd B-tree elements from low in node2 to
846 		 * high in node1.
847 		 */
848 		count = -count;
849 		tmp = count * (uint)sizeof(xfs_da_node_entry_t);
850 		btree_s = &btree2[0];
851 		btree_d = &btree1[nodehdr1.count];
852 		memcpy(btree_d, btree_s, tmp);
853 		nodehdr1.count += count;
854 
855 		xfs_trans_log_buf(tp, blk1->bp,
856 			XFS_DA_LOGRANGE(node1, btree_d, tmp));
857 
858 		/*
859 		 * Move elements in node2 down to fill the hole.
860 		 */
861 		tmp  = nodehdr2.count - count;
862 		tmp *= (uint)sizeof(xfs_da_node_entry_t);
863 		btree_s = &btree2[count];
864 		btree_d = &btree2[0];
865 		memmove(btree_d, btree_s, tmp);
866 		nodehdr2.count -= count;
867 	}
868 
869 	/*
870 	 * Log header of node 1 and all current bits of node 2.
871 	 */
872 	dp->d_ops->node_hdr_to_disk(node1, &nodehdr1);
873 	xfs_trans_log_buf(tp, blk1->bp,
874 		XFS_DA_LOGRANGE(node1, &node1->hdr, dp->d_ops->node_hdr_size));
875 
876 	dp->d_ops->node_hdr_to_disk(node2, &nodehdr2);
877 	xfs_trans_log_buf(tp, blk2->bp,
878 		XFS_DA_LOGRANGE(node2, &node2->hdr,
879 				dp->d_ops->node_hdr_size +
880 				(sizeof(btree2[0]) * nodehdr2.count)));
881 
882 	/*
883 	 * Record the last hashval from each block for upward propagation.
884 	 * (note: don't use the swapped node pointers)
885 	 */
886 	if (swap) {
887 		node1 = blk1->bp->b_addr;
888 		node2 = blk2->bp->b_addr;
889 		dp->d_ops->node_hdr_from_disk(&nodehdr1, node1);
890 		dp->d_ops->node_hdr_from_disk(&nodehdr2, node2);
891 		btree1 = dp->d_ops->node_tree_p(node1);
892 		btree2 = dp->d_ops->node_tree_p(node2);
893 	}
894 	blk1->hashval = be32_to_cpu(btree1[nodehdr1.count - 1].hashval);
895 	blk2->hashval = be32_to_cpu(btree2[nodehdr2.count - 1].hashval);
896 
897 	/*
898 	 * Adjust the expected index for insertion.
899 	 */
900 	if (blk1->index >= nodehdr1.count) {
901 		blk2->index = blk1->index - nodehdr1.count;
902 		blk1->index = nodehdr1.count + 1;	/* make it invalid */
903 	}
904 }
905 
906 /*
907  * Add a new entry to an intermediate node.
908  */
909 STATIC void
910 xfs_da3_node_add(
911 	struct xfs_da_state	*state,
912 	struct xfs_da_state_blk	*oldblk,
913 	struct xfs_da_state_blk	*newblk)
914 {
915 	struct xfs_da_intnode	*node;
916 	struct xfs_da3_icnode_hdr nodehdr;
917 	struct xfs_da_node_entry *btree;
918 	int			tmp;
919 	struct xfs_inode	*dp = state->args->dp;
920 
921 	trace_xfs_da_node_add(state->args);
922 
923 	node = oldblk->bp->b_addr;
924 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
925 	btree = dp->d_ops->node_tree_p(node);
926 
927 	ASSERT(oldblk->index >= 0 && oldblk->index <= nodehdr.count);
928 	ASSERT(newblk->blkno != 0);
929 	if (state->args->whichfork == XFS_DATA_FORK)
930 		ASSERT(newblk->blkno >= state->args->geo->leafblk &&
931 		       newblk->blkno < state->args->geo->freeblk);
932 
933 	/*
934 	 * We may need to make some room before we insert the new node.
935 	 */
936 	tmp = 0;
937 	if (oldblk->index < nodehdr.count) {
938 		tmp = (nodehdr.count - oldblk->index) * (uint)sizeof(*btree);
939 		memmove(&btree[oldblk->index + 1], &btree[oldblk->index], tmp);
940 	}
941 	btree[oldblk->index].hashval = cpu_to_be32(newblk->hashval);
942 	btree[oldblk->index].before = cpu_to_be32(newblk->blkno);
943 	xfs_trans_log_buf(state->args->trans, oldblk->bp,
944 		XFS_DA_LOGRANGE(node, &btree[oldblk->index],
945 				tmp + sizeof(*btree)));
946 
947 	nodehdr.count += 1;
948 	dp->d_ops->node_hdr_to_disk(node, &nodehdr);
949 	xfs_trans_log_buf(state->args->trans, oldblk->bp,
950 		XFS_DA_LOGRANGE(node, &node->hdr, dp->d_ops->node_hdr_size));
951 
952 	/*
953 	 * Copy the last hash value from the oldblk to propagate upwards.
954 	 */
955 	oldblk->hashval = be32_to_cpu(btree[nodehdr.count - 1].hashval);
956 }
957 
958 /*========================================================================
959  * Routines used for shrinking the Btree.
960  *========================================================================*/
961 
962 /*
963  * Deallocate an empty leaf node, remove it from its parent,
964  * possibly deallocating that block, etc...
965  */
966 int
967 xfs_da3_join(
968 	struct xfs_da_state	*state)
969 {
970 	struct xfs_da_state_blk	*drop_blk;
971 	struct xfs_da_state_blk	*save_blk;
972 	int			action = 0;
973 	int			error;
974 
975 	trace_xfs_da_join(state->args);
976 
977 	drop_blk = &state->path.blk[ state->path.active-1 ];
978 	save_blk = &state->altpath.blk[ state->path.active-1 ];
979 	ASSERT(state->path.blk[0].magic == XFS_DA_NODE_MAGIC);
980 	ASSERT(drop_blk->magic == XFS_ATTR_LEAF_MAGIC ||
981 	       drop_blk->magic == XFS_DIR2_LEAFN_MAGIC);
982 
983 	/*
984 	 * Walk back up the tree joining/deallocating as necessary.
985 	 * When we stop dropping blocks, break out.
986 	 */
987 	for (  ; state->path.active >= 2; drop_blk--, save_blk--,
988 		 state->path.active--) {
989 		/*
990 		 * See if we can combine the block with a neighbor.
991 		 *   (action == 0) => no options, just leave
992 		 *   (action == 1) => coalesce, then unlink
993 		 *   (action == 2) => block empty, unlink it
994 		 */
995 		switch (drop_blk->magic) {
996 		case XFS_ATTR_LEAF_MAGIC:
997 			error = xfs_attr3_leaf_toosmall(state, &action);
998 			if (error)
999 				return error;
1000 			if (action == 0)
1001 				return 0;
1002 			xfs_attr3_leaf_unbalance(state, drop_blk, save_blk);
1003 			break;
1004 		case XFS_DIR2_LEAFN_MAGIC:
1005 			error = xfs_dir2_leafn_toosmall(state, &action);
1006 			if (error)
1007 				return error;
1008 			if (action == 0)
1009 				return 0;
1010 			xfs_dir2_leafn_unbalance(state, drop_blk, save_blk);
1011 			break;
1012 		case XFS_DA_NODE_MAGIC:
1013 			/*
1014 			 * Remove the offending node, fixup hashvals,
1015 			 * check for a toosmall neighbor.
1016 			 */
1017 			xfs_da3_node_remove(state, drop_blk);
1018 			xfs_da3_fixhashpath(state, &state->path);
1019 			error = xfs_da3_node_toosmall(state, &action);
1020 			if (error)
1021 				return error;
1022 			if (action == 0)
1023 				return 0;
1024 			xfs_da3_node_unbalance(state, drop_blk, save_blk);
1025 			break;
1026 		}
1027 		xfs_da3_fixhashpath(state, &state->altpath);
1028 		error = xfs_da3_blk_unlink(state, drop_blk, save_blk);
1029 		xfs_da_state_kill_altpath(state);
1030 		if (error)
1031 			return error;
1032 		error = xfs_da_shrink_inode(state->args, drop_blk->blkno,
1033 							 drop_blk->bp);
1034 		drop_blk->bp = NULL;
1035 		if (error)
1036 			return error;
1037 	}
1038 	/*
1039 	 * We joined all the way to the top.  If it turns out that
1040 	 * we only have one entry in the root, make the child block
1041 	 * the new root.
1042 	 */
1043 	xfs_da3_node_remove(state, drop_blk);
1044 	xfs_da3_fixhashpath(state, &state->path);
1045 	error = xfs_da3_root_join(state, &state->path.blk[0]);
1046 	return error;
1047 }
1048 
1049 #ifdef	DEBUG
1050 static void
1051 xfs_da_blkinfo_onlychild_validate(struct xfs_da_blkinfo *blkinfo, __u16 level)
1052 {
1053 	__be16	magic = blkinfo->magic;
1054 
1055 	if (level == 1) {
1056 		ASSERT(magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
1057 		       magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC) ||
1058 		       magic == cpu_to_be16(XFS_ATTR_LEAF_MAGIC) ||
1059 		       magic == cpu_to_be16(XFS_ATTR3_LEAF_MAGIC));
1060 	} else {
1061 		ASSERT(magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
1062 		       magic == cpu_to_be16(XFS_DA3_NODE_MAGIC));
1063 	}
1064 	ASSERT(!blkinfo->forw);
1065 	ASSERT(!blkinfo->back);
1066 }
1067 #else	/* !DEBUG */
1068 #define	xfs_da_blkinfo_onlychild_validate(blkinfo, level)
1069 #endif	/* !DEBUG */
1070 
1071 /*
1072  * We have only one entry in the root.  Copy the only remaining child of
1073  * the old root to block 0 as the new root node.
1074  */
1075 STATIC int
1076 xfs_da3_root_join(
1077 	struct xfs_da_state	*state,
1078 	struct xfs_da_state_blk	*root_blk)
1079 {
1080 	struct xfs_da_intnode	*oldroot;
1081 	struct xfs_da_args	*args;
1082 	xfs_dablk_t		child;
1083 	struct xfs_buf		*bp;
1084 	struct xfs_da3_icnode_hdr oldroothdr;
1085 	struct xfs_da_node_entry *btree;
1086 	int			error;
1087 	struct xfs_inode	*dp = state->args->dp;
1088 
1089 	trace_xfs_da_root_join(state->args);
1090 
1091 	ASSERT(root_blk->magic == XFS_DA_NODE_MAGIC);
1092 
1093 	args = state->args;
1094 	oldroot = root_blk->bp->b_addr;
1095 	dp->d_ops->node_hdr_from_disk(&oldroothdr, oldroot);
1096 	ASSERT(oldroothdr.forw == 0);
1097 	ASSERT(oldroothdr.back == 0);
1098 
1099 	/*
1100 	 * If the root has more than one child, then don't do anything.
1101 	 */
1102 	if (oldroothdr.count > 1)
1103 		return 0;
1104 
1105 	/*
1106 	 * Read in the (only) child block, then copy those bytes into
1107 	 * the root block's buffer and free the original child block.
1108 	 */
1109 	btree = dp->d_ops->node_tree_p(oldroot);
1110 	child = be32_to_cpu(btree[0].before);
1111 	ASSERT(child != 0);
1112 	error = xfs_da3_node_read(args->trans, dp, child, -1, &bp,
1113 					     args->whichfork);
1114 	if (error)
1115 		return error;
1116 	xfs_da_blkinfo_onlychild_validate(bp->b_addr, oldroothdr.level);
1117 
1118 	/*
1119 	 * This could be copying a leaf back into the root block in the case of
1120 	 * there only being a single leaf block left in the tree. Hence we have
1121 	 * to update the b_ops pointer as well to match the buffer type change
1122 	 * that could occur. For dir3 blocks we also need to update the block
1123 	 * number in the buffer header.
1124 	 */
1125 	memcpy(root_blk->bp->b_addr, bp->b_addr, args->geo->blksize);
1126 	root_blk->bp->b_ops = bp->b_ops;
1127 	xfs_trans_buf_copy_type(root_blk->bp, bp);
1128 	if (oldroothdr.magic == XFS_DA3_NODE_MAGIC) {
1129 		struct xfs_da3_blkinfo *da3 = root_blk->bp->b_addr;
1130 		da3->blkno = cpu_to_be64(root_blk->bp->b_bn);
1131 	}
1132 	xfs_trans_log_buf(args->trans, root_blk->bp, 0,
1133 			  args->geo->blksize - 1);
1134 	error = xfs_da_shrink_inode(args, child, bp);
1135 	return error;
1136 }
1137 
1138 /*
1139  * Check a node block and its neighbors to see if the block should be
1140  * collapsed into one or the other neighbor.  Always keep the block
1141  * with the smaller block number.
1142  * If the current block is over 50% full, don't try to join it, return 0.
1143  * If the block is empty, fill in the state structure and return 2.
1144  * If it can be collapsed, fill in the state structure and return 1.
1145  * If nothing can be done, return 0.
1146  */
1147 STATIC int
1148 xfs_da3_node_toosmall(
1149 	struct xfs_da_state	*state,
1150 	int			*action)
1151 {
1152 	struct xfs_da_intnode	*node;
1153 	struct xfs_da_state_blk	*blk;
1154 	struct xfs_da_blkinfo	*info;
1155 	xfs_dablk_t		blkno;
1156 	struct xfs_buf		*bp;
1157 	struct xfs_da3_icnode_hdr nodehdr;
1158 	int			count;
1159 	int			forward;
1160 	int			error;
1161 	int			retval;
1162 	int			i;
1163 	struct xfs_inode	*dp = state->args->dp;
1164 
1165 	trace_xfs_da_node_toosmall(state->args);
1166 
1167 	/*
1168 	 * Check for the degenerate case of the block being over 50% full.
1169 	 * If so, it's not worth even looking to see if we might be able
1170 	 * to coalesce with a sibling.
1171 	 */
1172 	blk = &state->path.blk[ state->path.active-1 ];
1173 	info = blk->bp->b_addr;
1174 	node = (xfs_da_intnode_t *)info;
1175 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1176 	if (nodehdr.count > (state->args->geo->node_ents >> 1)) {
1177 		*action = 0;	/* blk over 50%, don't try to join */
1178 		return 0;	/* blk over 50%, don't try to join */
1179 	}
1180 
1181 	/*
1182 	 * Check for the degenerate case of the block being empty.
1183 	 * If the block is empty, we'll simply delete it, no need to
1184 	 * coalesce it with a sibling block.  We choose (arbitrarily)
1185 	 * to merge with the forward block unless it is NULL.
1186 	 */
1187 	if (nodehdr.count == 0) {
1188 		/*
1189 		 * Make altpath point to the block we want to keep and
1190 		 * path point to the block we want to drop (this one).
1191 		 */
1192 		forward = (info->forw != 0);
1193 		memcpy(&state->altpath, &state->path, sizeof(state->path));
1194 		error = xfs_da3_path_shift(state, &state->altpath, forward,
1195 						 0, &retval);
1196 		if (error)
1197 			return error;
1198 		if (retval) {
1199 			*action = 0;
1200 		} else {
1201 			*action = 2;
1202 		}
1203 		return 0;
1204 	}
1205 
1206 	/*
1207 	 * Examine each sibling block to see if we can coalesce with
1208 	 * at least 25% free space to spare.  We need to figure out
1209 	 * whether to merge with the forward or the backward block.
1210 	 * We prefer coalescing with the lower numbered sibling so as
1211 	 * to shrink a directory over time.
1212 	 */
1213 	count  = state->args->geo->node_ents;
1214 	count -= state->args->geo->node_ents >> 2;
1215 	count -= nodehdr.count;
1216 
1217 	/* start with smaller blk num */
1218 	forward = nodehdr.forw < nodehdr.back;
1219 	for (i = 0; i < 2; forward = !forward, i++) {
1220 		struct xfs_da3_icnode_hdr thdr;
1221 		if (forward)
1222 			blkno = nodehdr.forw;
1223 		else
1224 			blkno = nodehdr.back;
1225 		if (blkno == 0)
1226 			continue;
1227 		error = xfs_da3_node_read(state->args->trans, dp,
1228 					blkno, -1, &bp, state->args->whichfork);
1229 		if (error)
1230 			return error;
1231 
1232 		node = bp->b_addr;
1233 		dp->d_ops->node_hdr_from_disk(&thdr, node);
1234 		xfs_trans_brelse(state->args->trans, bp);
1235 
1236 		if (count - thdr.count >= 0)
1237 			break;	/* fits with at least 25% to spare */
1238 	}
1239 	if (i >= 2) {
1240 		*action = 0;
1241 		return 0;
1242 	}
1243 
1244 	/*
1245 	 * Make altpath point to the block we want to keep (the lower
1246 	 * numbered block) and path point to the block we want to drop.
1247 	 */
1248 	memcpy(&state->altpath, &state->path, sizeof(state->path));
1249 	if (blkno < blk->blkno) {
1250 		error = xfs_da3_path_shift(state, &state->altpath, forward,
1251 						 0, &retval);
1252 	} else {
1253 		error = xfs_da3_path_shift(state, &state->path, forward,
1254 						 0, &retval);
1255 	}
1256 	if (error)
1257 		return error;
1258 	if (retval) {
1259 		*action = 0;
1260 		return 0;
1261 	}
1262 	*action = 1;
1263 	return 0;
1264 }
1265 
1266 /*
1267  * Pick up the last hashvalue from an intermediate node.
1268  */
1269 STATIC uint
1270 xfs_da3_node_lasthash(
1271 	struct xfs_inode	*dp,
1272 	struct xfs_buf		*bp,
1273 	int			*count)
1274 {
1275 	struct xfs_da_intnode	 *node;
1276 	struct xfs_da_node_entry *btree;
1277 	struct xfs_da3_icnode_hdr nodehdr;
1278 
1279 	node = bp->b_addr;
1280 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1281 	if (count)
1282 		*count = nodehdr.count;
1283 	if (!nodehdr.count)
1284 		return 0;
1285 	btree = dp->d_ops->node_tree_p(node);
1286 	return be32_to_cpu(btree[nodehdr.count - 1].hashval);
1287 }
1288 
1289 /*
1290  * Walk back up the tree adjusting hash values as necessary,
1291  * when we stop making changes, return.
1292  */
1293 void
1294 xfs_da3_fixhashpath(
1295 	struct xfs_da_state	*state,
1296 	struct xfs_da_state_path *path)
1297 {
1298 	struct xfs_da_state_blk	*blk;
1299 	struct xfs_da_intnode	*node;
1300 	struct xfs_da_node_entry *btree;
1301 	xfs_dahash_t		lasthash=0;
1302 	int			level;
1303 	int			count;
1304 	struct xfs_inode	*dp = state->args->dp;
1305 
1306 	trace_xfs_da_fixhashpath(state->args);
1307 
1308 	level = path->active-1;
1309 	blk = &path->blk[ level ];
1310 	switch (blk->magic) {
1311 	case XFS_ATTR_LEAF_MAGIC:
1312 		lasthash = xfs_attr_leaf_lasthash(blk->bp, &count);
1313 		if (count == 0)
1314 			return;
1315 		break;
1316 	case XFS_DIR2_LEAFN_MAGIC:
1317 		lasthash = xfs_dir2_leaf_lasthash(dp, blk->bp, &count);
1318 		if (count == 0)
1319 			return;
1320 		break;
1321 	case XFS_DA_NODE_MAGIC:
1322 		lasthash = xfs_da3_node_lasthash(dp, blk->bp, &count);
1323 		if (count == 0)
1324 			return;
1325 		break;
1326 	}
1327 	for (blk--, level--; level >= 0; blk--, level--) {
1328 		struct xfs_da3_icnode_hdr nodehdr;
1329 
1330 		node = blk->bp->b_addr;
1331 		dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1332 		btree = dp->d_ops->node_tree_p(node);
1333 		if (be32_to_cpu(btree[blk->index].hashval) == lasthash)
1334 			break;
1335 		blk->hashval = lasthash;
1336 		btree[blk->index].hashval = cpu_to_be32(lasthash);
1337 		xfs_trans_log_buf(state->args->trans, blk->bp,
1338 				  XFS_DA_LOGRANGE(node, &btree[blk->index],
1339 						  sizeof(*btree)));
1340 
1341 		lasthash = be32_to_cpu(btree[nodehdr.count - 1].hashval);
1342 	}
1343 }
1344 
1345 /*
1346  * Remove an entry from an intermediate node.
1347  */
1348 STATIC void
1349 xfs_da3_node_remove(
1350 	struct xfs_da_state	*state,
1351 	struct xfs_da_state_blk	*drop_blk)
1352 {
1353 	struct xfs_da_intnode	*node;
1354 	struct xfs_da3_icnode_hdr nodehdr;
1355 	struct xfs_da_node_entry *btree;
1356 	int			index;
1357 	int			tmp;
1358 	struct xfs_inode	*dp = state->args->dp;
1359 
1360 	trace_xfs_da_node_remove(state->args);
1361 
1362 	node = drop_blk->bp->b_addr;
1363 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1364 	ASSERT(drop_blk->index < nodehdr.count);
1365 	ASSERT(drop_blk->index >= 0);
1366 
1367 	/*
1368 	 * Copy over the offending entry, or just zero it out.
1369 	 */
1370 	index = drop_blk->index;
1371 	btree = dp->d_ops->node_tree_p(node);
1372 	if (index < nodehdr.count - 1) {
1373 		tmp  = nodehdr.count - index - 1;
1374 		tmp *= (uint)sizeof(xfs_da_node_entry_t);
1375 		memmove(&btree[index], &btree[index + 1], tmp);
1376 		xfs_trans_log_buf(state->args->trans, drop_blk->bp,
1377 		    XFS_DA_LOGRANGE(node, &btree[index], tmp));
1378 		index = nodehdr.count - 1;
1379 	}
1380 	memset(&btree[index], 0, sizeof(xfs_da_node_entry_t));
1381 	xfs_trans_log_buf(state->args->trans, drop_blk->bp,
1382 	    XFS_DA_LOGRANGE(node, &btree[index], sizeof(btree[index])));
1383 	nodehdr.count -= 1;
1384 	dp->d_ops->node_hdr_to_disk(node, &nodehdr);
1385 	xfs_trans_log_buf(state->args->trans, drop_blk->bp,
1386 	    XFS_DA_LOGRANGE(node, &node->hdr, dp->d_ops->node_hdr_size));
1387 
1388 	/*
1389 	 * Copy the last hash value from the block to propagate upwards.
1390 	 */
1391 	drop_blk->hashval = be32_to_cpu(btree[index - 1].hashval);
1392 }
1393 
1394 /*
1395  * Unbalance the elements between two intermediate nodes,
1396  * move all Btree elements from one node into another.
1397  */
1398 STATIC void
1399 xfs_da3_node_unbalance(
1400 	struct xfs_da_state	*state,
1401 	struct xfs_da_state_blk	*drop_blk,
1402 	struct xfs_da_state_blk	*save_blk)
1403 {
1404 	struct xfs_da_intnode	*drop_node;
1405 	struct xfs_da_intnode	*save_node;
1406 	struct xfs_da_node_entry *drop_btree;
1407 	struct xfs_da_node_entry *save_btree;
1408 	struct xfs_da3_icnode_hdr drop_hdr;
1409 	struct xfs_da3_icnode_hdr save_hdr;
1410 	struct xfs_trans	*tp;
1411 	int			sindex;
1412 	int			tmp;
1413 	struct xfs_inode	*dp = state->args->dp;
1414 
1415 	trace_xfs_da_node_unbalance(state->args);
1416 
1417 	drop_node = drop_blk->bp->b_addr;
1418 	save_node = save_blk->bp->b_addr;
1419 	dp->d_ops->node_hdr_from_disk(&drop_hdr, drop_node);
1420 	dp->d_ops->node_hdr_from_disk(&save_hdr, save_node);
1421 	drop_btree = dp->d_ops->node_tree_p(drop_node);
1422 	save_btree = dp->d_ops->node_tree_p(save_node);
1423 	tp = state->args->trans;
1424 
1425 	/*
1426 	 * If the dying block has lower hashvals, then move all the
1427 	 * elements in the remaining block up to make a hole.
1428 	 */
1429 	if ((be32_to_cpu(drop_btree[0].hashval) <
1430 			be32_to_cpu(save_btree[0].hashval)) ||
1431 	    (be32_to_cpu(drop_btree[drop_hdr.count - 1].hashval) <
1432 			be32_to_cpu(save_btree[save_hdr.count - 1].hashval))) {
1433 		/* XXX: check this - is memmove dst correct? */
1434 		tmp = save_hdr.count * sizeof(xfs_da_node_entry_t);
1435 		memmove(&save_btree[drop_hdr.count], &save_btree[0], tmp);
1436 
1437 		sindex = 0;
1438 		xfs_trans_log_buf(tp, save_blk->bp,
1439 			XFS_DA_LOGRANGE(save_node, &save_btree[0],
1440 				(save_hdr.count + drop_hdr.count) *
1441 						sizeof(xfs_da_node_entry_t)));
1442 	} else {
1443 		sindex = save_hdr.count;
1444 		xfs_trans_log_buf(tp, save_blk->bp,
1445 			XFS_DA_LOGRANGE(save_node, &save_btree[sindex],
1446 				drop_hdr.count * sizeof(xfs_da_node_entry_t)));
1447 	}
1448 
1449 	/*
1450 	 * Move all the B-tree elements from drop_blk to save_blk.
1451 	 */
1452 	tmp = drop_hdr.count * (uint)sizeof(xfs_da_node_entry_t);
1453 	memcpy(&save_btree[sindex], &drop_btree[0], tmp);
1454 	save_hdr.count += drop_hdr.count;
1455 
1456 	dp->d_ops->node_hdr_to_disk(save_node, &save_hdr);
1457 	xfs_trans_log_buf(tp, save_blk->bp,
1458 		XFS_DA_LOGRANGE(save_node, &save_node->hdr,
1459 				dp->d_ops->node_hdr_size));
1460 
1461 	/*
1462 	 * Save the last hashval in the remaining block for upward propagation.
1463 	 */
1464 	save_blk->hashval = be32_to_cpu(save_btree[save_hdr.count - 1].hashval);
1465 }
1466 
1467 /*========================================================================
1468  * Routines used for finding things in the Btree.
1469  *========================================================================*/
1470 
1471 /*
1472  * Walk down the Btree looking for a particular filename, filling
1473  * in the state structure as we go.
1474  *
1475  * We will set the state structure to point to each of the elements
1476  * in each of the nodes where either the hashval is or should be.
1477  *
1478  * We support duplicate hashval's so for each entry in the current
1479  * node that could contain the desired hashval, descend.  This is a
1480  * pruned depth-first tree search.
1481  */
1482 int							/* error */
1483 xfs_da3_node_lookup_int(
1484 	struct xfs_da_state	*state,
1485 	int			*result)
1486 {
1487 	struct xfs_da_state_blk	*blk;
1488 	struct xfs_da_blkinfo	*curr;
1489 	struct xfs_da_intnode	*node;
1490 	struct xfs_da_node_entry *btree;
1491 	struct xfs_da3_icnode_hdr nodehdr;
1492 	struct xfs_da_args	*args;
1493 	xfs_dablk_t		blkno;
1494 	xfs_dahash_t		hashval;
1495 	xfs_dahash_t		btreehashval;
1496 	int			probe;
1497 	int			span;
1498 	int			max;
1499 	int			error;
1500 	int			retval;
1501 	unsigned int		expected_level = 0;
1502 	uint16_t		magic;
1503 	struct xfs_inode	*dp = state->args->dp;
1504 
1505 	args = state->args;
1506 
1507 	/*
1508 	 * Descend thru the B-tree searching each level for the right
1509 	 * node to use, until the right hashval is found.
1510 	 */
1511 	blkno = args->geo->leafblk;
1512 	for (blk = &state->path.blk[0], state->path.active = 1;
1513 			 state->path.active <= XFS_DA_NODE_MAXDEPTH;
1514 			 blk++, state->path.active++) {
1515 		/*
1516 		 * Read the next node down in the tree.
1517 		 */
1518 		blk->blkno = blkno;
1519 		error = xfs_da3_node_read(args->trans, args->dp, blkno,
1520 					-1, &blk->bp, args->whichfork);
1521 		if (error) {
1522 			blk->blkno = 0;
1523 			state->path.active--;
1524 			return error;
1525 		}
1526 		curr = blk->bp->b_addr;
1527 		magic = be16_to_cpu(curr->magic);
1528 
1529 		if (magic == XFS_ATTR_LEAF_MAGIC ||
1530 		    magic == XFS_ATTR3_LEAF_MAGIC) {
1531 			blk->magic = XFS_ATTR_LEAF_MAGIC;
1532 			blk->hashval = xfs_attr_leaf_lasthash(blk->bp, NULL);
1533 			break;
1534 		}
1535 
1536 		if (magic == XFS_DIR2_LEAFN_MAGIC ||
1537 		    magic == XFS_DIR3_LEAFN_MAGIC) {
1538 			blk->magic = XFS_DIR2_LEAFN_MAGIC;
1539 			blk->hashval = xfs_dir2_leaf_lasthash(args->dp,
1540 							      blk->bp, NULL);
1541 			break;
1542 		}
1543 
1544 		if (magic != XFS_DA_NODE_MAGIC && magic != XFS_DA3_NODE_MAGIC)
1545 			return -EFSCORRUPTED;
1546 
1547 		blk->magic = XFS_DA_NODE_MAGIC;
1548 
1549 		/*
1550 		 * Search an intermediate node for a match.
1551 		 */
1552 		node = blk->bp->b_addr;
1553 		dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1554 		btree = dp->d_ops->node_tree_p(node);
1555 
1556 		/* Tree taller than we can handle; bail out! */
1557 		if (nodehdr.level >= XFS_DA_NODE_MAXDEPTH)
1558 			return -EFSCORRUPTED;
1559 
1560 		/* Check the level from the root. */
1561 		if (blkno == args->geo->leafblk)
1562 			expected_level = nodehdr.level - 1;
1563 		else if (expected_level != nodehdr.level)
1564 			return -EFSCORRUPTED;
1565 		else
1566 			expected_level--;
1567 
1568 		max = nodehdr.count;
1569 		blk->hashval = be32_to_cpu(btree[max - 1].hashval);
1570 
1571 		/*
1572 		 * Binary search.  (note: small blocks will skip loop)
1573 		 */
1574 		probe = span = max / 2;
1575 		hashval = args->hashval;
1576 		while (span > 4) {
1577 			span /= 2;
1578 			btreehashval = be32_to_cpu(btree[probe].hashval);
1579 			if (btreehashval < hashval)
1580 				probe += span;
1581 			else if (btreehashval > hashval)
1582 				probe -= span;
1583 			else
1584 				break;
1585 		}
1586 		ASSERT((probe >= 0) && (probe < max));
1587 		ASSERT((span <= 4) ||
1588 			(be32_to_cpu(btree[probe].hashval) == hashval));
1589 
1590 		/*
1591 		 * Since we may have duplicate hashval's, find the first
1592 		 * matching hashval in the node.
1593 		 */
1594 		while (probe > 0 &&
1595 		       be32_to_cpu(btree[probe].hashval) >= hashval) {
1596 			probe--;
1597 		}
1598 		while (probe < max &&
1599 		       be32_to_cpu(btree[probe].hashval) < hashval) {
1600 			probe++;
1601 		}
1602 
1603 		/*
1604 		 * Pick the right block to descend on.
1605 		 */
1606 		if (probe == max) {
1607 			blk->index = max - 1;
1608 			blkno = be32_to_cpu(btree[max - 1].before);
1609 		} else {
1610 			blk->index = probe;
1611 			blkno = be32_to_cpu(btree[probe].before);
1612 		}
1613 
1614 		/* We can't point back to the root. */
1615 		if (blkno == args->geo->leafblk)
1616 			return -EFSCORRUPTED;
1617 	}
1618 
1619 	if (expected_level != 0)
1620 		return -EFSCORRUPTED;
1621 
1622 	/*
1623 	 * A leaf block that ends in the hashval that we are interested in
1624 	 * (final hashval == search hashval) means that the next block may
1625 	 * contain more entries with the same hashval, shift upward to the
1626 	 * next leaf and keep searching.
1627 	 */
1628 	for (;;) {
1629 		if (blk->magic == XFS_DIR2_LEAFN_MAGIC) {
1630 			retval = xfs_dir2_leafn_lookup_int(blk->bp, args,
1631 							&blk->index, state);
1632 		} else if (blk->magic == XFS_ATTR_LEAF_MAGIC) {
1633 			retval = xfs_attr3_leaf_lookup_int(blk->bp, args);
1634 			blk->index = args->index;
1635 			args->blkno = blk->blkno;
1636 		} else {
1637 			ASSERT(0);
1638 			return -EFSCORRUPTED;
1639 		}
1640 		if (((retval == -ENOENT) || (retval == -ENOATTR)) &&
1641 		    (blk->hashval == args->hashval)) {
1642 			error = xfs_da3_path_shift(state, &state->path, 1, 1,
1643 							 &retval);
1644 			if (error)
1645 				return error;
1646 			if (retval == 0) {
1647 				continue;
1648 			} else if (blk->magic == XFS_ATTR_LEAF_MAGIC) {
1649 				/* path_shift() gives ENOENT */
1650 				retval = -ENOATTR;
1651 			}
1652 		}
1653 		break;
1654 	}
1655 	*result = retval;
1656 	return 0;
1657 }
1658 
1659 /*========================================================================
1660  * Utility routines.
1661  *========================================================================*/
1662 
1663 /*
1664  * Compare two intermediate nodes for "order".
1665  */
1666 STATIC int
1667 xfs_da3_node_order(
1668 	struct xfs_inode *dp,
1669 	struct xfs_buf	*node1_bp,
1670 	struct xfs_buf	*node2_bp)
1671 {
1672 	struct xfs_da_intnode	*node1;
1673 	struct xfs_da_intnode	*node2;
1674 	struct xfs_da_node_entry *btree1;
1675 	struct xfs_da_node_entry *btree2;
1676 	struct xfs_da3_icnode_hdr node1hdr;
1677 	struct xfs_da3_icnode_hdr node2hdr;
1678 
1679 	node1 = node1_bp->b_addr;
1680 	node2 = node2_bp->b_addr;
1681 	dp->d_ops->node_hdr_from_disk(&node1hdr, node1);
1682 	dp->d_ops->node_hdr_from_disk(&node2hdr, node2);
1683 	btree1 = dp->d_ops->node_tree_p(node1);
1684 	btree2 = dp->d_ops->node_tree_p(node2);
1685 
1686 	if (node1hdr.count > 0 && node2hdr.count > 0 &&
1687 	    ((be32_to_cpu(btree2[0].hashval) < be32_to_cpu(btree1[0].hashval)) ||
1688 	     (be32_to_cpu(btree2[node2hdr.count - 1].hashval) <
1689 	      be32_to_cpu(btree1[node1hdr.count - 1].hashval)))) {
1690 		return 1;
1691 	}
1692 	return 0;
1693 }
1694 
1695 /*
1696  * Link a new block into a doubly linked list of blocks (of whatever type).
1697  */
1698 int							/* error */
1699 xfs_da3_blk_link(
1700 	struct xfs_da_state	*state,
1701 	struct xfs_da_state_blk	*old_blk,
1702 	struct xfs_da_state_blk	*new_blk)
1703 {
1704 	struct xfs_da_blkinfo	*old_info;
1705 	struct xfs_da_blkinfo	*new_info;
1706 	struct xfs_da_blkinfo	*tmp_info;
1707 	struct xfs_da_args	*args;
1708 	struct xfs_buf		*bp;
1709 	int			before = 0;
1710 	int			error;
1711 	struct xfs_inode	*dp = state->args->dp;
1712 
1713 	/*
1714 	 * Set up environment.
1715 	 */
1716 	args = state->args;
1717 	ASSERT(args != NULL);
1718 	old_info = old_blk->bp->b_addr;
1719 	new_info = new_blk->bp->b_addr;
1720 	ASSERT(old_blk->magic == XFS_DA_NODE_MAGIC ||
1721 	       old_blk->magic == XFS_DIR2_LEAFN_MAGIC ||
1722 	       old_blk->magic == XFS_ATTR_LEAF_MAGIC);
1723 
1724 	switch (old_blk->magic) {
1725 	case XFS_ATTR_LEAF_MAGIC:
1726 		before = xfs_attr_leaf_order(old_blk->bp, new_blk->bp);
1727 		break;
1728 	case XFS_DIR2_LEAFN_MAGIC:
1729 		before = xfs_dir2_leafn_order(dp, old_blk->bp, new_blk->bp);
1730 		break;
1731 	case XFS_DA_NODE_MAGIC:
1732 		before = xfs_da3_node_order(dp, old_blk->bp, new_blk->bp);
1733 		break;
1734 	}
1735 
1736 	/*
1737 	 * Link blocks in appropriate order.
1738 	 */
1739 	if (before) {
1740 		/*
1741 		 * Link new block in before existing block.
1742 		 */
1743 		trace_xfs_da_link_before(args);
1744 		new_info->forw = cpu_to_be32(old_blk->blkno);
1745 		new_info->back = old_info->back;
1746 		if (old_info->back) {
1747 			error = xfs_da3_node_read(args->trans, dp,
1748 						be32_to_cpu(old_info->back),
1749 						-1, &bp, args->whichfork);
1750 			if (error)
1751 				return error;
1752 			ASSERT(bp != NULL);
1753 			tmp_info = bp->b_addr;
1754 			ASSERT(tmp_info->magic == old_info->magic);
1755 			ASSERT(be32_to_cpu(tmp_info->forw) == old_blk->blkno);
1756 			tmp_info->forw = cpu_to_be32(new_blk->blkno);
1757 			xfs_trans_log_buf(args->trans, bp, 0, sizeof(*tmp_info)-1);
1758 		}
1759 		old_info->back = cpu_to_be32(new_blk->blkno);
1760 	} else {
1761 		/*
1762 		 * Link new block in after existing block.
1763 		 */
1764 		trace_xfs_da_link_after(args);
1765 		new_info->forw = old_info->forw;
1766 		new_info->back = cpu_to_be32(old_blk->blkno);
1767 		if (old_info->forw) {
1768 			error = xfs_da3_node_read(args->trans, dp,
1769 						be32_to_cpu(old_info->forw),
1770 						-1, &bp, args->whichfork);
1771 			if (error)
1772 				return error;
1773 			ASSERT(bp != NULL);
1774 			tmp_info = bp->b_addr;
1775 			ASSERT(tmp_info->magic == old_info->magic);
1776 			ASSERT(be32_to_cpu(tmp_info->back) == old_blk->blkno);
1777 			tmp_info->back = cpu_to_be32(new_blk->blkno);
1778 			xfs_trans_log_buf(args->trans, bp, 0, sizeof(*tmp_info)-1);
1779 		}
1780 		old_info->forw = cpu_to_be32(new_blk->blkno);
1781 	}
1782 
1783 	xfs_trans_log_buf(args->trans, old_blk->bp, 0, sizeof(*tmp_info) - 1);
1784 	xfs_trans_log_buf(args->trans, new_blk->bp, 0, sizeof(*tmp_info) - 1);
1785 	return 0;
1786 }
1787 
1788 /*
1789  * Unlink a block from a doubly linked list of blocks.
1790  */
1791 STATIC int						/* error */
1792 xfs_da3_blk_unlink(
1793 	struct xfs_da_state	*state,
1794 	struct xfs_da_state_blk	*drop_blk,
1795 	struct xfs_da_state_blk	*save_blk)
1796 {
1797 	struct xfs_da_blkinfo	*drop_info;
1798 	struct xfs_da_blkinfo	*save_info;
1799 	struct xfs_da_blkinfo	*tmp_info;
1800 	struct xfs_da_args	*args;
1801 	struct xfs_buf		*bp;
1802 	int			error;
1803 
1804 	/*
1805 	 * Set up environment.
1806 	 */
1807 	args = state->args;
1808 	ASSERT(args != NULL);
1809 	save_info = save_blk->bp->b_addr;
1810 	drop_info = drop_blk->bp->b_addr;
1811 	ASSERT(save_blk->magic == XFS_DA_NODE_MAGIC ||
1812 	       save_blk->magic == XFS_DIR2_LEAFN_MAGIC ||
1813 	       save_blk->magic == XFS_ATTR_LEAF_MAGIC);
1814 	ASSERT(save_blk->magic == drop_blk->magic);
1815 	ASSERT((be32_to_cpu(save_info->forw) == drop_blk->blkno) ||
1816 	       (be32_to_cpu(save_info->back) == drop_blk->blkno));
1817 	ASSERT((be32_to_cpu(drop_info->forw) == save_blk->blkno) ||
1818 	       (be32_to_cpu(drop_info->back) == save_blk->blkno));
1819 
1820 	/*
1821 	 * Unlink the leaf block from the doubly linked chain of leaves.
1822 	 */
1823 	if (be32_to_cpu(save_info->back) == drop_blk->blkno) {
1824 		trace_xfs_da_unlink_back(args);
1825 		save_info->back = drop_info->back;
1826 		if (drop_info->back) {
1827 			error = xfs_da3_node_read(args->trans, args->dp,
1828 						be32_to_cpu(drop_info->back),
1829 						-1, &bp, args->whichfork);
1830 			if (error)
1831 				return error;
1832 			ASSERT(bp != NULL);
1833 			tmp_info = bp->b_addr;
1834 			ASSERT(tmp_info->magic == save_info->magic);
1835 			ASSERT(be32_to_cpu(tmp_info->forw) == drop_blk->blkno);
1836 			tmp_info->forw = cpu_to_be32(save_blk->blkno);
1837 			xfs_trans_log_buf(args->trans, bp, 0,
1838 						    sizeof(*tmp_info) - 1);
1839 		}
1840 	} else {
1841 		trace_xfs_da_unlink_forward(args);
1842 		save_info->forw = drop_info->forw;
1843 		if (drop_info->forw) {
1844 			error = xfs_da3_node_read(args->trans, args->dp,
1845 						be32_to_cpu(drop_info->forw),
1846 						-1, &bp, args->whichfork);
1847 			if (error)
1848 				return error;
1849 			ASSERT(bp != NULL);
1850 			tmp_info = bp->b_addr;
1851 			ASSERT(tmp_info->magic == save_info->magic);
1852 			ASSERT(be32_to_cpu(tmp_info->back) == drop_blk->blkno);
1853 			tmp_info->back = cpu_to_be32(save_blk->blkno);
1854 			xfs_trans_log_buf(args->trans, bp, 0,
1855 						    sizeof(*tmp_info) - 1);
1856 		}
1857 	}
1858 
1859 	xfs_trans_log_buf(args->trans, save_blk->bp, 0, sizeof(*save_info) - 1);
1860 	return 0;
1861 }
1862 
1863 /*
1864  * Move a path "forward" or "!forward" one block at the current level.
1865  *
1866  * This routine will adjust a "path" to point to the next block
1867  * "forward" (higher hashvalues) or "!forward" (lower hashvals) in the
1868  * Btree, including updating pointers to the intermediate nodes between
1869  * the new bottom and the root.
1870  */
1871 int							/* error */
1872 xfs_da3_path_shift(
1873 	struct xfs_da_state	*state,
1874 	struct xfs_da_state_path *path,
1875 	int			forward,
1876 	int			release,
1877 	int			*result)
1878 {
1879 	struct xfs_da_state_blk	*blk;
1880 	struct xfs_da_blkinfo	*info;
1881 	struct xfs_da_intnode	*node;
1882 	struct xfs_da_args	*args;
1883 	struct xfs_da_node_entry *btree;
1884 	struct xfs_da3_icnode_hdr nodehdr;
1885 	struct xfs_buf		*bp;
1886 	xfs_dablk_t		blkno = 0;
1887 	int			level;
1888 	int			error;
1889 	struct xfs_inode	*dp = state->args->dp;
1890 
1891 	trace_xfs_da_path_shift(state->args);
1892 
1893 	/*
1894 	 * Roll up the Btree looking for the first block where our
1895 	 * current index is not at the edge of the block.  Note that
1896 	 * we skip the bottom layer because we want the sibling block.
1897 	 */
1898 	args = state->args;
1899 	ASSERT(args != NULL);
1900 	ASSERT(path != NULL);
1901 	ASSERT((path->active > 0) && (path->active < XFS_DA_NODE_MAXDEPTH));
1902 	level = (path->active-1) - 1;	/* skip bottom layer in path */
1903 	for (blk = &path->blk[level]; level >= 0; blk--, level--) {
1904 		node = blk->bp->b_addr;
1905 		dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1906 		btree = dp->d_ops->node_tree_p(node);
1907 
1908 		if (forward && (blk->index < nodehdr.count - 1)) {
1909 			blk->index++;
1910 			blkno = be32_to_cpu(btree[blk->index].before);
1911 			break;
1912 		} else if (!forward && (blk->index > 0)) {
1913 			blk->index--;
1914 			blkno = be32_to_cpu(btree[blk->index].before);
1915 			break;
1916 		}
1917 	}
1918 	if (level < 0) {
1919 		*result = -ENOENT;	/* we're out of our tree */
1920 		ASSERT(args->op_flags & XFS_DA_OP_OKNOENT);
1921 		return 0;
1922 	}
1923 
1924 	/*
1925 	 * Roll down the edge of the subtree until we reach the
1926 	 * same depth we were at originally.
1927 	 */
1928 	for (blk++, level++; level < path->active; blk++, level++) {
1929 		/*
1930 		 * Read the next child block into a local buffer.
1931 		 */
1932 		error = xfs_da3_node_read(args->trans, dp, blkno, -1, &bp,
1933 					  args->whichfork);
1934 		if (error)
1935 			return error;
1936 
1937 		/*
1938 		 * Release the old block (if it's dirty, the trans doesn't
1939 		 * actually let go) and swap the local buffer into the path
1940 		 * structure. This ensures failure of the above read doesn't set
1941 		 * a NULL buffer in an active slot in the path.
1942 		 */
1943 		if (release)
1944 			xfs_trans_brelse(args->trans, blk->bp);
1945 		blk->blkno = blkno;
1946 		blk->bp = bp;
1947 
1948 		info = blk->bp->b_addr;
1949 		ASSERT(info->magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
1950 		       info->magic == cpu_to_be16(XFS_DA3_NODE_MAGIC) ||
1951 		       info->magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
1952 		       info->magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC) ||
1953 		       info->magic == cpu_to_be16(XFS_ATTR_LEAF_MAGIC) ||
1954 		       info->magic == cpu_to_be16(XFS_ATTR3_LEAF_MAGIC));
1955 
1956 
1957 		/*
1958 		 * Note: we flatten the magic number to a single type so we
1959 		 * don't have to compare against crc/non-crc types elsewhere.
1960 		 */
1961 		switch (be16_to_cpu(info->magic)) {
1962 		case XFS_DA_NODE_MAGIC:
1963 		case XFS_DA3_NODE_MAGIC:
1964 			blk->magic = XFS_DA_NODE_MAGIC;
1965 			node = (xfs_da_intnode_t *)info;
1966 			dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1967 			btree = dp->d_ops->node_tree_p(node);
1968 			blk->hashval = be32_to_cpu(btree[nodehdr.count - 1].hashval);
1969 			if (forward)
1970 				blk->index = 0;
1971 			else
1972 				blk->index = nodehdr.count - 1;
1973 			blkno = be32_to_cpu(btree[blk->index].before);
1974 			break;
1975 		case XFS_ATTR_LEAF_MAGIC:
1976 		case XFS_ATTR3_LEAF_MAGIC:
1977 			blk->magic = XFS_ATTR_LEAF_MAGIC;
1978 			ASSERT(level == path->active-1);
1979 			blk->index = 0;
1980 			blk->hashval = xfs_attr_leaf_lasthash(blk->bp, NULL);
1981 			break;
1982 		case XFS_DIR2_LEAFN_MAGIC:
1983 		case XFS_DIR3_LEAFN_MAGIC:
1984 			blk->magic = XFS_DIR2_LEAFN_MAGIC;
1985 			ASSERT(level == path->active-1);
1986 			blk->index = 0;
1987 			blk->hashval = xfs_dir2_leaf_lasthash(args->dp,
1988 							      blk->bp, NULL);
1989 			break;
1990 		default:
1991 			ASSERT(0);
1992 			break;
1993 		}
1994 	}
1995 	*result = 0;
1996 	return 0;
1997 }
1998 
1999 
2000 /*========================================================================
2001  * Utility routines.
2002  *========================================================================*/
2003 
2004 /*
2005  * Implement a simple hash on a character string.
2006  * Rotate the hash value by 7 bits, then XOR each character in.
2007  * This is implemented with some source-level loop unrolling.
2008  */
2009 xfs_dahash_t
2010 xfs_da_hashname(const uint8_t *name, int namelen)
2011 {
2012 	xfs_dahash_t hash;
2013 
2014 	/*
2015 	 * Do four characters at a time as long as we can.
2016 	 */
2017 	for (hash = 0; namelen >= 4; namelen -= 4, name += 4)
2018 		hash = (name[0] << 21) ^ (name[1] << 14) ^ (name[2] << 7) ^
2019 		       (name[3] << 0) ^ rol32(hash, 7 * 4);
2020 
2021 	/*
2022 	 * Now do the rest of the characters.
2023 	 */
2024 	switch (namelen) {
2025 	case 3:
2026 		return (name[0] << 14) ^ (name[1] << 7) ^ (name[2] << 0) ^
2027 		       rol32(hash, 7 * 3);
2028 	case 2:
2029 		return (name[0] << 7) ^ (name[1] << 0) ^ rol32(hash, 7 * 2);
2030 	case 1:
2031 		return (name[0] << 0) ^ rol32(hash, 7 * 1);
2032 	default: /* case 0: */
2033 		return hash;
2034 	}
2035 }
2036 
2037 enum xfs_dacmp
2038 xfs_da_compname(
2039 	struct xfs_da_args *args,
2040 	const unsigned char *name,
2041 	int		len)
2042 {
2043 	return (args->namelen == len && memcmp(args->name, name, len) == 0) ?
2044 					XFS_CMP_EXACT : XFS_CMP_DIFFERENT;
2045 }
2046 
2047 static xfs_dahash_t
2048 xfs_default_hashname(
2049 	struct xfs_name	*name)
2050 {
2051 	return xfs_da_hashname(name->name, name->len);
2052 }
2053 
2054 const struct xfs_nameops xfs_default_nameops = {
2055 	.hashname	= xfs_default_hashname,
2056 	.compname	= xfs_da_compname
2057 };
2058 
2059 int
2060 xfs_da_grow_inode_int(
2061 	struct xfs_da_args	*args,
2062 	xfs_fileoff_t		*bno,
2063 	int			count)
2064 {
2065 	struct xfs_trans	*tp = args->trans;
2066 	struct xfs_inode	*dp = args->dp;
2067 	int			w = args->whichfork;
2068 	xfs_rfsblock_t		nblks = dp->i_d.di_nblocks;
2069 	struct xfs_bmbt_irec	map, *mapp;
2070 	int			nmap, error, got, i, mapi;
2071 
2072 	/*
2073 	 * Find a spot in the file space to put the new block.
2074 	 */
2075 	error = xfs_bmap_first_unused(tp, dp, count, bno, w);
2076 	if (error)
2077 		return error;
2078 
2079 	/*
2080 	 * Try mapping it in one filesystem block.
2081 	 */
2082 	nmap = 1;
2083 	error = xfs_bmapi_write(tp, dp, *bno, count,
2084 			xfs_bmapi_aflag(w)|XFS_BMAPI_METADATA|XFS_BMAPI_CONTIG,
2085 			args->total, &map, &nmap);
2086 	if (error)
2087 		return error;
2088 
2089 	ASSERT(nmap <= 1);
2090 	if (nmap == 1) {
2091 		mapp = &map;
2092 		mapi = 1;
2093 	} else if (nmap == 0 && count > 1) {
2094 		xfs_fileoff_t		b;
2095 		int			c;
2096 
2097 		/*
2098 		 * If we didn't get it and the block might work if fragmented,
2099 		 * try without the CONTIG flag.  Loop until we get it all.
2100 		 */
2101 		mapp = kmem_alloc(sizeof(*mapp) * count, 0);
2102 		for (b = *bno, mapi = 0; b < *bno + count; ) {
2103 			nmap = min(XFS_BMAP_MAX_NMAP, count);
2104 			c = (int)(*bno + count - b);
2105 			error = xfs_bmapi_write(tp, dp, b, c,
2106 					xfs_bmapi_aflag(w)|XFS_BMAPI_METADATA,
2107 					args->total, &mapp[mapi], &nmap);
2108 			if (error)
2109 				goto out_free_map;
2110 			if (nmap < 1)
2111 				break;
2112 			mapi += nmap;
2113 			b = mapp[mapi - 1].br_startoff +
2114 			    mapp[mapi - 1].br_blockcount;
2115 		}
2116 	} else {
2117 		mapi = 0;
2118 		mapp = NULL;
2119 	}
2120 
2121 	/*
2122 	 * Count the blocks we got, make sure it matches the total.
2123 	 */
2124 	for (i = 0, got = 0; i < mapi; i++)
2125 		got += mapp[i].br_blockcount;
2126 	if (got != count || mapp[0].br_startoff != *bno ||
2127 	    mapp[mapi - 1].br_startoff + mapp[mapi - 1].br_blockcount !=
2128 	    *bno + count) {
2129 		error = -ENOSPC;
2130 		goto out_free_map;
2131 	}
2132 
2133 	/* account for newly allocated blocks in reserved blocks total */
2134 	args->total -= dp->i_d.di_nblocks - nblks;
2135 
2136 out_free_map:
2137 	if (mapp != &map)
2138 		kmem_free(mapp);
2139 	return error;
2140 }
2141 
2142 /*
2143  * Add a block to the btree ahead of the file.
2144  * Return the new block number to the caller.
2145  */
2146 int
2147 xfs_da_grow_inode(
2148 	struct xfs_da_args	*args,
2149 	xfs_dablk_t		*new_blkno)
2150 {
2151 	xfs_fileoff_t		bno;
2152 	int			error;
2153 
2154 	trace_xfs_da_grow_inode(args);
2155 
2156 	bno = args->geo->leafblk;
2157 	error = xfs_da_grow_inode_int(args, &bno, args->geo->fsbcount);
2158 	if (!error)
2159 		*new_blkno = (xfs_dablk_t)bno;
2160 	return error;
2161 }
2162 
2163 /*
2164  * Ick.  We need to always be able to remove a btree block, even
2165  * if there's no space reservation because the filesystem is full.
2166  * This is called if xfs_bunmapi on a btree block fails due to ENOSPC.
2167  * It swaps the target block with the last block in the file.  The
2168  * last block in the file can always be removed since it can't cause
2169  * a bmap btree split to do that.
2170  */
2171 STATIC int
2172 xfs_da3_swap_lastblock(
2173 	struct xfs_da_args	*args,
2174 	xfs_dablk_t		*dead_blknop,
2175 	struct xfs_buf		**dead_bufp)
2176 {
2177 	struct xfs_da_blkinfo	*dead_info;
2178 	struct xfs_da_blkinfo	*sib_info;
2179 	struct xfs_da_intnode	*par_node;
2180 	struct xfs_da_intnode	*dead_node;
2181 	struct xfs_dir2_leaf	*dead_leaf2;
2182 	struct xfs_da_node_entry *btree;
2183 	struct xfs_da3_icnode_hdr par_hdr;
2184 	struct xfs_inode	*dp;
2185 	struct xfs_trans	*tp;
2186 	struct xfs_mount	*mp;
2187 	struct xfs_buf		*dead_buf;
2188 	struct xfs_buf		*last_buf;
2189 	struct xfs_buf		*sib_buf;
2190 	struct xfs_buf		*par_buf;
2191 	xfs_dahash_t		dead_hash;
2192 	xfs_fileoff_t		lastoff;
2193 	xfs_dablk_t		dead_blkno;
2194 	xfs_dablk_t		last_blkno;
2195 	xfs_dablk_t		sib_blkno;
2196 	xfs_dablk_t		par_blkno;
2197 	int			error;
2198 	int			w;
2199 	int			entno;
2200 	int			level;
2201 	int			dead_level;
2202 
2203 	trace_xfs_da_swap_lastblock(args);
2204 
2205 	dead_buf = *dead_bufp;
2206 	dead_blkno = *dead_blknop;
2207 	tp = args->trans;
2208 	dp = args->dp;
2209 	w = args->whichfork;
2210 	ASSERT(w == XFS_DATA_FORK);
2211 	mp = dp->i_mount;
2212 	lastoff = args->geo->freeblk;
2213 	error = xfs_bmap_last_before(tp, dp, &lastoff, w);
2214 	if (error)
2215 		return error;
2216 	if (unlikely(lastoff == 0)) {
2217 		XFS_ERROR_REPORT("xfs_da_swap_lastblock(1)", XFS_ERRLEVEL_LOW,
2218 				 mp);
2219 		return -EFSCORRUPTED;
2220 	}
2221 	/*
2222 	 * Read the last block in the btree space.
2223 	 */
2224 	last_blkno = (xfs_dablk_t)lastoff - args->geo->fsbcount;
2225 	error = xfs_da3_node_read(tp, dp, last_blkno, -1, &last_buf, w);
2226 	if (error)
2227 		return error;
2228 	/*
2229 	 * Copy the last block into the dead buffer and log it.
2230 	 */
2231 	memcpy(dead_buf->b_addr, last_buf->b_addr, args->geo->blksize);
2232 	xfs_trans_log_buf(tp, dead_buf, 0, args->geo->blksize - 1);
2233 	dead_info = dead_buf->b_addr;
2234 	/*
2235 	 * Get values from the moved block.
2236 	 */
2237 	if (dead_info->magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
2238 	    dead_info->magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
2239 		struct xfs_dir3_icleaf_hdr leafhdr;
2240 		struct xfs_dir2_leaf_entry *ents;
2241 
2242 		dead_leaf2 = (xfs_dir2_leaf_t *)dead_info;
2243 		dp->d_ops->leaf_hdr_from_disk(&leafhdr, dead_leaf2);
2244 		ents = dp->d_ops->leaf_ents_p(dead_leaf2);
2245 		dead_level = 0;
2246 		dead_hash = be32_to_cpu(ents[leafhdr.count - 1].hashval);
2247 	} else {
2248 		struct xfs_da3_icnode_hdr deadhdr;
2249 
2250 		dead_node = (xfs_da_intnode_t *)dead_info;
2251 		dp->d_ops->node_hdr_from_disk(&deadhdr, dead_node);
2252 		btree = dp->d_ops->node_tree_p(dead_node);
2253 		dead_level = deadhdr.level;
2254 		dead_hash = be32_to_cpu(btree[deadhdr.count - 1].hashval);
2255 	}
2256 	sib_buf = par_buf = NULL;
2257 	/*
2258 	 * If the moved block has a left sibling, fix up the pointers.
2259 	 */
2260 	if ((sib_blkno = be32_to_cpu(dead_info->back))) {
2261 		error = xfs_da3_node_read(tp, dp, sib_blkno, -1, &sib_buf, w);
2262 		if (error)
2263 			goto done;
2264 		sib_info = sib_buf->b_addr;
2265 		if (unlikely(
2266 		    be32_to_cpu(sib_info->forw) != last_blkno ||
2267 		    sib_info->magic != dead_info->magic)) {
2268 			XFS_ERROR_REPORT("xfs_da_swap_lastblock(2)",
2269 					 XFS_ERRLEVEL_LOW, mp);
2270 			error = -EFSCORRUPTED;
2271 			goto done;
2272 		}
2273 		sib_info->forw = cpu_to_be32(dead_blkno);
2274 		xfs_trans_log_buf(tp, sib_buf,
2275 			XFS_DA_LOGRANGE(sib_info, &sib_info->forw,
2276 					sizeof(sib_info->forw)));
2277 		sib_buf = NULL;
2278 	}
2279 	/*
2280 	 * If the moved block has a right sibling, fix up the pointers.
2281 	 */
2282 	if ((sib_blkno = be32_to_cpu(dead_info->forw))) {
2283 		error = xfs_da3_node_read(tp, dp, sib_blkno, -1, &sib_buf, w);
2284 		if (error)
2285 			goto done;
2286 		sib_info = sib_buf->b_addr;
2287 		if (unlikely(
2288 		       be32_to_cpu(sib_info->back) != last_blkno ||
2289 		       sib_info->magic != dead_info->magic)) {
2290 			XFS_ERROR_REPORT("xfs_da_swap_lastblock(3)",
2291 					 XFS_ERRLEVEL_LOW, mp);
2292 			error = -EFSCORRUPTED;
2293 			goto done;
2294 		}
2295 		sib_info->back = cpu_to_be32(dead_blkno);
2296 		xfs_trans_log_buf(tp, sib_buf,
2297 			XFS_DA_LOGRANGE(sib_info, &sib_info->back,
2298 					sizeof(sib_info->back)));
2299 		sib_buf = NULL;
2300 	}
2301 	par_blkno = args->geo->leafblk;
2302 	level = -1;
2303 	/*
2304 	 * Walk down the tree looking for the parent of the moved block.
2305 	 */
2306 	for (;;) {
2307 		error = xfs_da3_node_read(tp, dp, par_blkno, -1, &par_buf, w);
2308 		if (error)
2309 			goto done;
2310 		par_node = par_buf->b_addr;
2311 		dp->d_ops->node_hdr_from_disk(&par_hdr, par_node);
2312 		if (level >= 0 && level != par_hdr.level + 1) {
2313 			XFS_ERROR_REPORT("xfs_da_swap_lastblock(4)",
2314 					 XFS_ERRLEVEL_LOW, mp);
2315 			error = -EFSCORRUPTED;
2316 			goto done;
2317 		}
2318 		level = par_hdr.level;
2319 		btree = dp->d_ops->node_tree_p(par_node);
2320 		for (entno = 0;
2321 		     entno < par_hdr.count &&
2322 		     be32_to_cpu(btree[entno].hashval) < dead_hash;
2323 		     entno++)
2324 			continue;
2325 		if (entno == par_hdr.count) {
2326 			XFS_ERROR_REPORT("xfs_da_swap_lastblock(5)",
2327 					 XFS_ERRLEVEL_LOW, mp);
2328 			error = -EFSCORRUPTED;
2329 			goto done;
2330 		}
2331 		par_blkno = be32_to_cpu(btree[entno].before);
2332 		if (level == dead_level + 1)
2333 			break;
2334 		xfs_trans_brelse(tp, par_buf);
2335 		par_buf = NULL;
2336 	}
2337 	/*
2338 	 * We're in the right parent block.
2339 	 * Look for the right entry.
2340 	 */
2341 	for (;;) {
2342 		for (;
2343 		     entno < par_hdr.count &&
2344 		     be32_to_cpu(btree[entno].before) != last_blkno;
2345 		     entno++)
2346 			continue;
2347 		if (entno < par_hdr.count)
2348 			break;
2349 		par_blkno = par_hdr.forw;
2350 		xfs_trans_brelse(tp, par_buf);
2351 		par_buf = NULL;
2352 		if (unlikely(par_blkno == 0)) {
2353 			XFS_ERROR_REPORT("xfs_da_swap_lastblock(6)",
2354 					 XFS_ERRLEVEL_LOW, mp);
2355 			error = -EFSCORRUPTED;
2356 			goto done;
2357 		}
2358 		error = xfs_da3_node_read(tp, dp, par_blkno, -1, &par_buf, w);
2359 		if (error)
2360 			goto done;
2361 		par_node = par_buf->b_addr;
2362 		dp->d_ops->node_hdr_from_disk(&par_hdr, par_node);
2363 		if (par_hdr.level != level) {
2364 			XFS_ERROR_REPORT("xfs_da_swap_lastblock(7)",
2365 					 XFS_ERRLEVEL_LOW, mp);
2366 			error = -EFSCORRUPTED;
2367 			goto done;
2368 		}
2369 		btree = dp->d_ops->node_tree_p(par_node);
2370 		entno = 0;
2371 	}
2372 	/*
2373 	 * Update the parent entry pointing to the moved block.
2374 	 */
2375 	btree[entno].before = cpu_to_be32(dead_blkno);
2376 	xfs_trans_log_buf(tp, par_buf,
2377 		XFS_DA_LOGRANGE(par_node, &btree[entno].before,
2378 				sizeof(btree[entno].before)));
2379 	*dead_blknop = last_blkno;
2380 	*dead_bufp = last_buf;
2381 	return 0;
2382 done:
2383 	if (par_buf)
2384 		xfs_trans_brelse(tp, par_buf);
2385 	if (sib_buf)
2386 		xfs_trans_brelse(tp, sib_buf);
2387 	xfs_trans_brelse(tp, last_buf);
2388 	return error;
2389 }
2390 
2391 /*
2392  * Remove a btree block from a directory or attribute.
2393  */
2394 int
2395 xfs_da_shrink_inode(
2396 	struct xfs_da_args	*args,
2397 	xfs_dablk_t		dead_blkno,
2398 	struct xfs_buf		*dead_buf)
2399 {
2400 	struct xfs_inode	*dp;
2401 	int			done, error, w, count;
2402 	struct xfs_trans	*tp;
2403 
2404 	trace_xfs_da_shrink_inode(args);
2405 
2406 	dp = args->dp;
2407 	w = args->whichfork;
2408 	tp = args->trans;
2409 	count = args->geo->fsbcount;
2410 	for (;;) {
2411 		/*
2412 		 * Remove extents.  If we get ENOSPC for a dir we have to move
2413 		 * the last block to the place we want to kill.
2414 		 */
2415 		error = xfs_bunmapi(tp, dp, dead_blkno, count,
2416 				    xfs_bmapi_aflag(w), 0, &done);
2417 		if (error == -ENOSPC) {
2418 			if (w != XFS_DATA_FORK)
2419 				break;
2420 			error = xfs_da3_swap_lastblock(args, &dead_blkno,
2421 						      &dead_buf);
2422 			if (error)
2423 				break;
2424 		} else {
2425 			break;
2426 		}
2427 	}
2428 	xfs_trans_binval(tp, dead_buf);
2429 	return error;
2430 }
2431 
2432 /*
2433  * See if the mapping(s) for this btree block are valid, i.e.
2434  * don't contain holes, are logically contiguous, and cover the whole range.
2435  */
2436 STATIC int
2437 xfs_da_map_covers_blocks(
2438 	int		nmap,
2439 	xfs_bmbt_irec_t	*mapp,
2440 	xfs_dablk_t	bno,
2441 	int		count)
2442 {
2443 	int		i;
2444 	xfs_fileoff_t	off;
2445 
2446 	for (i = 0, off = bno; i < nmap; i++) {
2447 		if (mapp[i].br_startblock == HOLESTARTBLOCK ||
2448 		    mapp[i].br_startblock == DELAYSTARTBLOCK) {
2449 			return 0;
2450 		}
2451 		if (off != mapp[i].br_startoff) {
2452 			return 0;
2453 		}
2454 		off += mapp[i].br_blockcount;
2455 	}
2456 	return off == bno + count;
2457 }
2458 
2459 /*
2460  * Convert a struct xfs_bmbt_irec to a struct xfs_buf_map.
2461  *
2462  * For the single map case, it is assumed that the caller has provided a pointer
2463  * to a valid xfs_buf_map.  For the multiple map case, this function will
2464  * allocate the xfs_buf_map to hold all the maps and replace the caller's single
2465  * map pointer with the allocated map.
2466  */
2467 static int
2468 xfs_buf_map_from_irec(
2469 	struct xfs_mount	*mp,
2470 	struct xfs_buf_map	**mapp,
2471 	int			*nmaps,
2472 	struct xfs_bmbt_irec	*irecs,
2473 	int			nirecs)
2474 {
2475 	struct xfs_buf_map	*map;
2476 	int			i;
2477 
2478 	ASSERT(*nmaps == 1);
2479 	ASSERT(nirecs >= 1);
2480 
2481 	if (nirecs > 1) {
2482 		map = kmem_zalloc(nirecs * sizeof(struct xfs_buf_map),
2483 				  KM_NOFS);
2484 		if (!map)
2485 			return -ENOMEM;
2486 		*mapp = map;
2487 	}
2488 
2489 	*nmaps = nirecs;
2490 	map = *mapp;
2491 	for (i = 0; i < *nmaps; i++) {
2492 		ASSERT(irecs[i].br_startblock != DELAYSTARTBLOCK &&
2493 		       irecs[i].br_startblock != HOLESTARTBLOCK);
2494 		map[i].bm_bn = XFS_FSB_TO_DADDR(mp, irecs[i].br_startblock);
2495 		map[i].bm_len = XFS_FSB_TO_BB(mp, irecs[i].br_blockcount);
2496 	}
2497 	return 0;
2498 }
2499 
2500 /*
2501  * Map the block we are given ready for reading. There are three possible return
2502  * values:
2503  *	-1 - will be returned if we land in a hole and mappedbno == -2 so the
2504  *	     caller knows not to execute a subsequent read.
2505  *	 0 - if we mapped the block successfully
2506  *	>0 - positive error number if there was an error.
2507  */
2508 static int
2509 xfs_dabuf_map(
2510 	struct xfs_inode	*dp,
2511 	xfs_dablk_t		bno,
2512 	xfs_daddr_t		mappedbno,
2513 	int			whichfork,
2514 	struct xfs_buf_map	**map,
2515 	int			*nmaps)
2516 {
2517 	struct xfs_mount	*mp = dp->i_mount;
2518 	int			nfsb;
2519 	int			error = 0;
2520 	struct xfs_bmbt_irec	irec;
2521 	struct xfs_bmbt_irec	*irecs = &irec;
2522 	int			nirecs;
2523 
2524 	ASSERT(map && *map);
2525 	ASSERT(*nmaps == 1);
2526 
2527 	if (whichfork == XFS_DATA_FORK)
2528 		nfsb = mp->m_dir_geo->fsbcount;
2529 	else
2530 		nfsb = mp->m_attr_geo->fsbcount;
2531 
2532 	/*
2533 	 * Caller doesn't have a mapping.  -2 means don't complain
2534 	 * if we land in a hole.
2535 	 */
2536 	if (mappedbno == -1 || mappedbno == -2) {
2537 		/*
2538 		 * Optimize the one-block case.
2539 		 */
2540 		if (nfsb != 1)
2541 			irecs = kmem_zalloc(sizeof(irec) * nfsb,
2542 					    KM_NOFS);
2543 
2544 		nirecs = nfsb;
2545 		error = xfs_bmapi_read(dp, (xfs_fileoff_t)bno, nfsb, irecs,
2546 				       &nirecs, xfs_bmapi_aflag(whichfork));
2547 		if (error)
2548 			goto out;
2549 	} else {
2550 		irecs->br_startblock = XFS_DADDR_TO_FSB(mp, mappedbno);
2551 		irecs->br_startoff = (xfs_fileoff_t)bno;
2552 		irecs->br_blockcount = nfsb;
2553 		irecs->br_state = 0;
2554 		nirecs = 1;
2555 	}
2556 
2557 	if (!xfs_da_map_covers_blocks(nirecs, irecs, bno, nfsb)) {
2558 		error = mappedbno == -2 ? -1 : -EFSCORRUPTED;
2559 		if (unlikely(error == -EFSCORRUPTED)) {
2560 			if (xfs_error_level >= XFS_ERRLEVEL_LOW) {
2561 				int i;
2562 				xfs_alert(mp, "%s: bno %lld dir: inode %lld",
2563 					__func__, (long long)bno,
2564 					(long long)dp->i_ino);
2565 				for (i = 0; i < *nmaps; i++) {
2566 					xfs_alert(mp,
2567 "[%02d] br_startoff %lld br_startblock %lld br_blockcount %lld br_state %d",
2568 						i,
2569 						(long long)irecs[i].br_startoff,
2570 						(long long)irecs[i].br_startblock,
2571 						(long long)irecs[i].br_blockcount,
2572 						irecs[i].br_state);
2573 				}
2574 			}
2575 			XFS_ERROR_REPORT("xfs_da_do_buf(1)",
2576 					 XFS_ERRLEVEL_LOW, mp);
2577 		}
2578 		goto out;
2579 	}
2580 	error = xfs_buf_map_from_irec(mp, map, nmaps, irecs, nirecs);
2581 out:
2582 	if (irecs != &irec)
2583 		kmem_free(irecs);
2584 	return error;
2585 }
2586 
2587 /*
2588  * Get a buffer for the dir/attr block.
2589  */
2590 int
2591 xfs_da_get_buf(
2592 	struct xfs_trans	*trans,
2593 	struct xfs_inode	*dp,
2594 	xfs_dablk_t		bno,
2595 	xfs_daddr_t		mappedbno,
2596 	struct xfs_buf		**bpp,
2597 	int			whichfork)
2598 {
2599 	struct xfs_buf		*bp;
2600 	struct xfs_buf_map	map;
2601 	struct xfs_buf_map	*mapp;
2602 	int			nmap;
2603 	int			error;
2604 
2605 	*bpp = NULL;
2606 	mapp = &map;
2607 	nmap = 1;
2608 	error = xfs_dabuf_map(dp, bno, mappedbno, whichfork,
2609 				&mapp, &nmap);
2610 	if (error) {
2611 		/* mapping a hole is not an error, but we don't continue */
2612 		if (error == -1)
2613 			error = 0;
2614 		goto out_free;
2615 	}
2616 
2617 	bp = xfs_trans_get_buf_map(trans, dp->i_mount->m_ddev_targp,
2618 				    mapp, nmap, 0);
2619 	error = bp ? bp->b_error : -EIO;
2620 	if (error) {
2621 		if (bp)
2622 			xfs_trans_brelse(trans, bp);
2623 		goto out_free;
2624 	}
2625 
2626 	*bpp = bp;
2627 
2628 out_free:
2629 	if (mapp != &map)
2630 		kmem_free(mapp);
2631 
2632 	return error;
2633 }
2634 
2635 /*
2636  * Get a buffer for the dir/attr block, fill in the contents.
2637  */
2638 int
2639 xfs_da_read_buf(
2640 	struct xfs_trans	*trans,
2641 	struct xfs_inode	*dp,
2642 	xfs_dablk_t		bno,
2643 	xfs_daddr_t		mappedbno,
2644 	struct xfs_buf		**bpp,
2645 	int			whichfork,
2646 	const struct xfs_buf_ops *ops)
2647 {
2648 	struct xfs_buf		*bp;
2649 	struct xfs_buf_map	map;
2650 	struct xfs_buf_map	*mapp;
2651 	int			nmap;
2652 	int			error;
2653 
2654 	*bpp = NULL;
2655 	mapp = &map;
2656 	nmap = 1;
2657 	error = xfs_dabuf_map(dp, bno, mappedbno, whichfork,
2658 				&mapp, &nmap);
2659 	if (error) {
2660 		/* mapping a hole is not an error, but we don't continue */
2661 		if (error == -1)
2662 			error = 0;
2663 		goto out_free;
2664 	}
2665 
2666 	error = xfs_trans_read_buf_map(dp->i_mount, trans,
2667 					dp->i_mount->m_ddev_targp,
2668 					mapp, nmap, 0, &bp, ops);
2669 	if (error)
2670 		goto out_free;
2671 
2672 	if (whichfork == XFS_ATTR_FORK)
2673 		xfs_buf_set_ref(bp, XFS_ATTR_BTREE_REF);
2674 	else
2675 		xfs_buf_set_ref(bp, XFS_DIR_BTREE_REF);
2676 	*bpp = bp;
2677 out_free:
2678 	if (mapp != &map)
2679 		kmem_free(mapp);
2680 
2681 	return error;
2682 }
2683 
2684 /*
2685  * Readahead the dir/attr block.
2686  */
2687 int
2688 xfs_da_reada_buf(
2689 	struct xfs_inode	*dp,
2690 	xfs_dablk_t		bno,
2691 	xfs_daddr_t		mappedbno,
2692 	int			whichfork,
2693 	const struct xfs_buf_ops *ops)
2694 {
2695 	struct xfs_buf_map	map;
2696 	struct xfs_buf_map	*mapp;
2697 	int			nmap;
2698 	int			error;
2699 
2700 	mapp = &map;
2701 	nmap = 1;
2702 	error = xfs_dabuf_map(dp, bno, mappedbno, whichfork,
2703 				&mapp, &nmap);
2704 	if (error) {
2705 		/* mapping a hole is not an error, but we don't continue */
2706 		if (error == -1)
2707 			error = 0;
2708 		goto out_free;
2709 	}
2710 
2711 	mappedbno = mapp[0].bm_bn;
2712 	xfs_buf_readahead_map(dp->i_mount->m_ddev_targp, mapp, nmap, ops);
2713 
2714 out_free:
2715 	if (mapp != &map)
2716 		kmem_free(mapp);
2717 
2718 	return error;
2719 }
2720