xref: /openbmc/linux/fs/xfs/libxfs/xfs_da_btree.c (revision e2c75e76)
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * Copyright (c) 2013 Red Hat, Inc.
4  * All Rights Reserved.
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License as
8  * published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it would be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License for more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program; if not, write the Free Software Foundation,
17  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
18  */
19 #include "xfs.h"
20 #include "xfs_fs.h"
21 #include "xfs_shared.h"
22 #include "xfs_format.h"
23 #include "xfs_log_format.h"
24 #include "xfs_trans_resv.h"
25 #include "xfs_bit.h"
26 #include "xfs_mount.h"
27 #include "xfs_da_format.h"
28 #include "xfs_da_btree.h"
29 #include "xfs_dir2.h"
30 #include "xfs_dir2_priv.h"
31 #include "xfs_inode.h"
32 #include "xfs_trans.h"
33 #include "xfs_inode_item.h"
34 #include "xfs_alloc.h"
35 #include "xfs_bmap.h"
36 #include "xfs_attr.h"
37 #include "xfs_attr_leaf.h"
38 #include "xfs_error.h"
39 #include "xfs_trace.h"
40 #include "xfs_cksum.h"
41 #include "xfs_buf_item.h"
42 #include "xfs_log.h"
43 
44 /*
45  * xfs_da_btree.c
46  *
47  * Routines to implement directories as Btrees of hashed names.
48  */
49 
50 /*========================================================================
51  * Function prototypes for the kernel.
52  *========================================================================*/
53 
54 /*
55  * Routines used for growing the Btree.
56  */
57 STATIC int xfs_da3_root_split(xfs_da_state_t *state,
58 					    xfs_da_state_blk_t *existing_root,
59 					    xfs_da_state_blk_t *new_child);
60 STATIC int xfs_da3_node_split(xfs_da_state_t *state,
61 					    xfs_da_state_blk_t *existing_blk,
62 					    xfs_da_state_blk_t *split_blk,
63 					    xfs_da_state_blk_t *blk_to_add,
64 					    int treelevel,
65 					    int *result);
66 STATIC void xfs_da3_node_rebalance(xfs_da_state_t *state,
67 					 xfs_da_state_blk_t *node_blk_1,
68 					 xfs_da_state_blk_t *node_blk_2);
69 STATIC void xfs_da3_node_add(xfs_da_state_t *state,
70 				   xfs_da_state_blk_t *old_node_blk,
71 				   xfs_da_state_blk_t *new_node_blk);
72 
73 /*
74  * Routines used for shrinking the Btree.
75  */
76 STATIC int xfs_da3_root_join(xfs_da_state_t *state,
77 					   xfs_da_state_blk_t *root_blk);
78 STATIC int xfs_da3_node_toosmall(xfs_da_state_t *state, int *retval);
79 STATIC void xfs_da3_node_remove(xfs_da_state_t *state,
80 					      xfs_da_state_blk_t *drop_blk);
81 STATIC void xfs_da3_node_unbalance(xfs_da_state_t *state,
82 					 xfs_da_state_blk_t *src_node_blk,
83 					 xfs_da_state_blk_t *dst_node_blk);
84 
85 /*
86  * Utility routines.
87  */
88 STATIC int	xfs_da3_blk_unlink(xfs_da_state_t *state,
89 				  xfs_da_state_blk_t *drop_blk,
90 				  xfs_da_state_blk_t *save_blk);
91 
92 
93 kmem_zone_t *xfs_da_state_zone;	/* anchor for state struct zone */
94 
95 /*
96  * Allocate a dir-state structure.
97  * We don't put them on the stack since they're large.
98  */
99 xfs_da_state_t *
100 xfs_da_state_alloc(void)
101 {
102 	return kmem_zone_zalloc(xfs_da_state_zone, KM_NOFS);
103 }
104 
105 /*
106  * Kill the altpath contents of a da-state structure.
107  */
108 STATIC void
109 xfs_da_state_kill_altpath(xfs_da_state_t *state)
110 {
111 	int	i;
112 
113 	for (i = 0; i < state->altpath.active; i++)
114 		state->altpath.blk[i].bp = NULL;
115 	state->altpath.active = 0;
116 }
117 
118 /*
119  * Free a da-state structure.
120  */
121 void
122 xfs_da_state_free(xfs_da_state_t *state)
123 {
124 	xfs_da_state_kill_altpath(state);
125 #ifdef DEBUG
126 	memset((char *)state, 0, sizeof(*state));
127 #endif /* DEBUG */
128 	kmem_zone_free(xfs_da_state_zone, state);
129 }
130 
131 static xfs_failaddr_t
132 xfs_da3_node_verify(
133 	struct xfs_buf		*bp)
134 {
135 	struct xfs_mount	*mp = bp->b_target->bt_mount;
136 	struct xfs_da_intnode	*hdr = bp->b_addr;
137 	struct xfs_da3_icnode_hdr ichdr;
138 	const struct xfs_dir_ops *ops;
139 
140 	ops = xfs_dir_get_ops(mp, NULL);
141 
142 	ops->node_hdr_from_disk(&ichdr, hdr);
143 
144 	if (xfs_sb_version_hascrc(&mp->m_sb)) {
145 		struct xfs_da3_node_hdr *hdr3 = bp->b_addr;
146 
147 		if (ichdr.magic != XFS_DA3_NODE_MAGIC)
148 			return __this_address;
149 
150 		if (!uuid_equal(&hdr3->info.uuid, &mp->m_sb.sb_meta_uuid))
151 			return __this_address;
152 		if (be64_to_cpu(hdr3->info.blkno) != bp->b_bn)
153 			return __this_address;
154 		if (!xfs_log_check_lsn(mp, be64_to_cpu(hdr3->info.lsn)))
155 			return __this_address;
156 	} else {
157 		if (ichdr.magic != XFS_DA_NODE_MAGIC)
158 			return __this_address;
159 	}
160 	if (ichdr.level == 0)
161 		return __this_address;
162 	if (ichdr.level > XFS_DA_NODE_MAXDEPTH)
163 		return __this_address;
164 	if (ichdr.count == 0)
165 		return __this_address;
166 
167 	/*
168 	 * we don't know if the node is for and attribute or directory tree,
169 	 * so only fail if the count is outside both bounds
170 	 */
171 	if (ichdr.count > mp->m_dir_geo->node_ents &&
172 	    ichdr.count > mp->m_attr_geo->node_ents)
173 		return __this_address;
174 
175 	/* XXX: hash order check? */
176 
177 	return NULL;
178 }
179 
180 static void
181 xfs_da3_node_write_verify(
182 	struct xfs_buf	*bp)
183 {
184 	struct xfs_mount	*mp = bp->b_target->bt_mount;
185 	struct xfs_buf_log_item	*bip = bp->b_log_item;
186 	struct xfs_da3_node_hdr *hdr3 = bp->b_addr;
187 	xfs_failaddr_t		fa;
188 
189 	fa = xfs_da3_node_verify(bp);
190 	if (fa) {
191 		xfs_verifier_error(bp, -EFSCORRUPTED, fa);
192 		return;
193 	}
194 
195 	if (!xfs_sb_version_hascrc(&mp->m_sb))
196 		return;
197 
198 	if (bip)
199 		hdr3->info.lsn = cpu_to_be64(bip->bli_item.li_lsn);
200 
201 	xfs_buf_update_cksum(bp, XFS_DA3_NODE_CRC_OFF);
202 }
203 
204 /*
205  * leaf/node format detection on trees is sketchy, so a node read can be done on
206  * leaf level blocks when detection identifies the tree as a node format tree
207  * incorrectly. In this case, we need to swap the verifier to match the correct
208  * format of the block being read.
209  */
210 static void
211 xfs_da3_node_read_verify(
212 	struct xfs_buf		*bp)
213 {
214 	struct xfs_da_blkinfo	*info = bp->b_addr;
215 	xfs_failaddr_t		fa;
216 
217 	switch (be16_to_cpu(info->magic)) {
218 		case XFS_DA3_NODE_MAGIC:
219 			if (!xfs_buf_verify_cksum(bp, XFS_DA3_NODE_CRC_OFF)) {
220 				xfs_verifier_error(bp, -EFSBADCRC,
221 						__this_address);
222 				break;
223 			}
224 			/* fall through */
225 		case XFS_DA_NODE_MAGIC:
226 			fa = xfs_da3_node_verify(bp);
227 			if (fa)
228 				xfs_verifier_error(bp, -EFSCORRUPTED, fa);
229 			return;
230 		case XFS_ATTR_LEAF_MAGIC:
231 		case XFS_ATTR3_LEAF_MAGIC:
232 			bp->b_ops = &xfs_attr3_leaf_buf_ops;
233 			bp->b_ops->verify_read(bp);
234 			return;
235 		case XFS_DIR2_LEAFN_MAGIC:
236 		case XFS_DIR3_LEAFN_MAGIC:
237 			bp->b_ops = &xfs_dir3_leafn_buf_ops;
238 			bp->b_ops->verify_read(bp);
239 			return;
240 		default:
241 			xfs_verifier_error(bp, -EFSCORRUPTED, __this_address);
242 			break;
243 	}
244 }
245 
246 /* Verify the structure of a da3 block. */
247 static xfs_failaddr_t
248 xfs_da3_node_verify_struct(
249 	struct xfs_buf		*bp)
250 {
251 	struct xfs_da_blkinfo	*info = bp->b_addr;
252 
253 	switch (be16_to_cpu(info->magic)) {
254 	case XFS_DA3_NODE_MAGIC:
255 	case XFS_DA_NODE_MAGIC:
256 		return xfs_da3_node_verify(bp);
257 	case XFS_ATTR_LEAF_MAGIC:
258 	case XFS_ATTR3_LEAF_MAGIC:
259 		bp->b_ops = &xfs_attr3_leaf_buf_ops;
260 		return bp->b_ops->verify_struct(bp);
261 	case XFS_DIR2_LEAFN_MAGIC:
262 	case XFS_DIR3_LEAFN_MAGIC:
263 		bp->b_ops = &xfs_dir3_leafn_buf_ops;
264 		return bp->b_ops->verify_struct(bp);
265 	default:
266 		return __this_address;
267 	}
268 }
269 
270 const struct xfs_buf_ops xfs_da3_node_buf_ops = {
271 	.name = "xfs_da3_node",
272 	.verify_read = xfs_da3_node_read_verify,
273 	.verify_write = xfs_da3_node_write_verify,
274 	.verify_struct = xfs_da3_node_verify_struct,
275 };
276 
277 int
278 xfs_da3_node_read(
279 	struct xfs_trans	*tp,
280 	struct xfs_inode	*dp,
281 	xfs_dablk_t		bno,
282 	xfs_daddr_t		mappedbno,
283 	struct xfs_buf		**bpp,
284 	int			which_fork)
285 {
286 	int			err;
287 
288 	err = xfs_da_read_buf(tp, dp, bno, mappedbno, bpp,
289 					which_fork, &xfs_da3_node_buf_ops);
290 	if (!err && tp && *bpp) {
291 		struct xfs_da_blkinfo	*info = (*bpp)->b_addr;
292 		int			type;
293 
294 		switch (be16_to_cpu(info->magic)) {
295 		case XFS_DA_NODE_MAGIC:
296 		case XFS_DA3_NODE_MAGIC:
297 			type = XFS_BLFT_DA_NODE_BUF;
298 			break;
299 		case XFS_ATTR_LEAF_MAGIC:
300 		case XFS_ATTR3_LEAF_MAGIC:
301 			type = XFS_BLFT_ATTR_LEAF_BUF;
302 			break;
303 		case XFS_DIR2_LEAFN_MAGIC:
304 		case XFS_DIR3_LEAFN_MAGIC:
305 			type = XFS_BLFT_DIR_LEAFN_BUF;
306 			break;
307 		default:
308 			type = 0;
309 			ASSERT(0);
310 			break;
311 		}
312 		xfs_trans_buf_set_type(tp, *bpp, type);
313 	}
314 	return err;
315 }
316 
317 /*========================================================================
318  * Routines used for growing the Btree.
319  *========================================================================*/
320 
321 /*
322  * Create the initial contents of an intermediate node.
323  */
324 int
325 xfs_da3_node_create(
326 	struct xfs_da_args	*args,
327 	xfs_dablk_t		blkno,
328 	int			level,
329 	struct xfs_buf		**bpp,
330 	int			whichfork)
331 {
332 	struct xfs_da_intnode	*node;
333 	struct xfs_trans	*tp = args->trans;
334 	struct xfs_mount	*mp = tp->t_mountp;
335 	struct xfs_da3_icnode_hdr ichdr = {0};
336 	struct xfs_buf		*bp;
337 	int			error;
338 	struct xfs_inode	*dp = args->dp;
339 
340 	trace_xfs_da_node_create(args);
341 	ASSERT(level <= XFS_DA_NODE_MAXDEPTH);
342 
343 	error = xfs_da_get_buf(tp, dp, blkno, -1, &bp, whichfork);
344 	if (error)
345 		return error;
346 	bp->b_ops = &xfs_da3_node_buf_ops;
347 	xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DA_NODE_BUF);
348 	node = bp->b_addr;
349 
350 	if (xfs_sb_version_hascrc(&mp->m_sb)) {
351 		struct xfs_da3_node_hdr *hdr3 = bp->b_addr;
352 
353 		memset(hdr3, 0, sizeof(struct xfs_da3_node_hdr));
354 		ichdr.magic = XFS_DA3_NODE_MAGIC;
355 		hdr3->info.blkno = cpu_to_be64(bp->b_bn);
356 		hdr3->info.owner = cpu_to_be64(args->dp->i_ino);
357 		uuid_copy(&hdr3->info.uuid, &mp->m_sb.sb_meta_uuid);
358 	} else {
359 		ichdr.magic = XFS_DA_NODE_MAGIC;
360 	}
361 	ichdr.level = level;
362 
363 	dp->d_ops->node_hdr_to_disk(node, &ichdr);
364 	xfs_trans_log_buf(tp, bp,
365 		XFS_DA_LOGRANGE(node, &node->hdr, dp->d_ops->node_hdr_size));
366 
367 	*bpp = bp;
368 	return 0;
369 }
370 
371 /*
372  * Split a leaf node, rebalance, then possibly split
373  * intermediate nodes, rebalance, etc.
374  */
375 int							/* error */
376 xfs_da3_split(
377 	struct xfs_da_state	*state)
378 {
379 	struct xfs_da_state_blk	*oldblk;
380 	struct xfs_da_state_blk	*newblk;
381 	struct xfs_da_state_blk	*addblk;
382 	struct xfs_da_intnode	*node;
383 	int			max;
384 	int			action = 0;
385 	int			error;
386 	int			i;
387 
388 	trace_xfs_da_split(state->args);
389 
390 	/*
391 	 * Walk back up the tree splitting/inserting/adjusting as necessary.
392 	 * If we need to insert and there isn't room, split the node, then
393 	 * decide which fragment to insert the new block from below into.
394 	 * Note that we may split the root this way, but we need more fixup.
395 	 */
396 	max = state->path.active - 1;
397 	ASSERT((max >= 0) && (max < XFS_DA_NODE_MAXDEPTH));
398 	ASSERT(state->path.blk[max].magic == XFS_ATTR_LEAF_MAGIC ||
399 	       state->path.blk[max].magic == XFS_DIR2_LEAFN_MAGIC);
400 
401 	addblk = &state->path.blk[max];		/* initial dummy value */
402 	for (i = max; (i >= 0) && addblk; state->path.active--, i--) {
403 		oldblk = &state->path.blk[i];
404 		newblk = &state->altpath.blk[i];
405 
406 		/*
407 		 * If a leaf node then
408 		 *     Allocate a new leaf node, then rebalance across them.
409 		 * else if an intermediate node then
410 		 *     We split on the last layer, must we split the node?
411 		 */
412 		switch (oldblk->magic) {
413 		case XFS_ATTR_LEAF_MAGIC:
414 			error = xfs_attr3_leaf_split(state, oldblk, newblk);
415 			if ((error != 0) && (error != -ENOSPC)) {
416 				return error;	/* GROT: attr is inconsistent */
417 			}
418 			if (!error) {
419 				addblk = newblk;
420 				break;
421 			}
422 			/*
423 			 * Entry wouldn't fit, split the leaf again. The new
424 			 * extrablk will be consumed by xfs_da3_node_split if
425 			 * the node is split.
426 			 */
427 			state->extravalid = 1;
428 			if (state->inleaf) {
429 				state->extraafter = 0;	/* before newblk */
430 				trace_xfs_attr_leaf_split_before(state->args);
431 				error = xfs_attr3_leaf_split(state, oldblk,
432 							    &state->extrablk);
433 			} else {
434 				state->extraafter = 1;	/* after newblk */
435 				trace_xfs_attr_leaf_split_after(state->args);
436 				error = xfs_attr3_leaf_split(state, newblk,
437 							    &state->extrablk);
438 			}
439 			if (error)
440 				return error;	/* GROT: attr inconsistent */
441 			addblk = newblk;
442 			break;
443 		case XFS_DIR2_LEAFN_MAGIC:
444 			error = xfs_dir2_leafn_split(state, oldblk, newblk);
445 			if (error)
446 				return error;
447 			addblk = newblk;
448 			break;
449 		case XFS_DA_NODE_MAGIC:
450 			error = xfs_da3_node_split(state, oldblk, newblk, addblk,
451 							 max - i, &action);
452 			addblk->bp = NULL;
453 			if (error)
454 				return error;	/* GROT: dir is inconsistent */
455 			/*
456 			 * Record the newly split block for the next time thru?
457 			 */
458 			if (action)
459 				addblk = newblk;
460 			else
461 				addblk = NULL;
462 			break;
463 		}
464 
465 		/*
466 		 * Update the btree to show the new hashval for this child.
467 		 */
468 		xfs_da3_fixhashpath(state, &state->path);
469 	}
470 	if (!addblk)
471 		return 0;
472 
473 	/*
474 	 * xfs_da3_node_split() should have consumed any extra blocks we added
475 	 * during a double leaf split in the attr fork. This is guaranteed as
476 	 * we can't be here if the attr fork only has a single leaf block.
477 	 */
478 	ASSERT(state->extravalid == 0 ||
479 	       state->path.blk[max].magic == XFS_DIR2_LEAFN_MAGIC);
480 
481 	/*
482 	 * Split the root node.
483 	 */
484 	ASSERT(state->path.active == 0);
485 	oldblk = &state->path.blk[0];
486 	error = xfs_da3_root_split(state, oldblk, addblk);
487 	if (error) {
488 		addblk->bp = NULL;
489 		return error;	/* GROT: dir is inconsistent */
490 	}
491 
492 	/*
493 	 * Update pointers to the node which used to be block 0 and just got
494 	 * bumped because of the addition of a new root node.  Note that the
495 	 * original block 0 could be at any position in the list of blocks in
496 	 * the tree.
497 	 *
498 	 * Note: the magic numbers and sibling pointers are in the same physical
499 	 * place for both v2 and v3 headers (by design). Hence it doesn't matter
500 	 * which version of the xfs_da_intnode structure we use here as the
501 	 * result will be the same using either structure.
502 	 */
503 	node = oldblk->bp->b_addr;
504 	if (node->hdr.info.forw) {
505 		ASSERT(be32_to_cpu(node->hdr.info.forw) == addblk->blkno);
506 		node = addblk->bp->b_addr;
507 		node->hdr.info.back = cpu_to_be32(oldblk->blkno);
508 		xfs_trans_log_buf(state->args->trans, addblk->bp,
509 				  XFS_DA_LOGRANGE(node, &node->hdr.info,
510 				  sizeof(node->hdr.info)));
511 	}
512 	node = oldblk->bp->b_addr;
513 	if (node->hdr.info.back) {
514 		ASSERT(be32_to_cpu(node->hdr.info.back) == addblk->blkno);
515 		node = addblk->bp->b_addr;
516 		node->hdr.info.forw = cpu_to_be32(oldblk->blkno);
517 		xfs_trans_log_buf(state->args->trans, addblk->bp,
518 				  XFS_DA_LOGRANGE(node, &node->hdr.info,
519 				  sizeof(node->hdr.info)));
520 	}
521 	addblk->bp = NULL;
522 	return 0;
523 }
524 
525 /*
526  * Split the root.  We have to create a new root and point to the two
527  * parts (the split old root) that we just created.  Copy block zero to
528  * the EOF, extending the inode in process.
529  */
530 STATIC int						/* error */
531 xfs_da3_root_split(
532 	struct xfs_da_state	*state,
533 	struct xfs_da_state_blk	*blk1,
534 	struct xfs_da_state_blk	*blk2)
535 {
536 	struct xfs_da_intnode	*node;
537 	struct xfs_da_intnode	*oldroot;
538 	struct xfs_da_node_entry *btree;
539 	struct xfs_da3_icnode_hdr nodehdr;
540 	struct xfs_da_args	*args;
541 	struct xfs_buf		*bp;
542 	struct xfs_inode	*dp;
543 	struct xfs_trans	*tp;
544 	struct xfs_dir2_leaf	*leaf;
545 	xfs_dablk_t		blkno;
546 	int			level;
547 	int			error;
548 	int			size;
549 
550 	trace_xfs_da_root_split(state->args);
551 
552 	/*
553 	 * Copy the existing (incorrect) block from the root node position
554 	 * to a free space somewhere.
555 	 */
556 	args = state->args;
557 	error = xfs_da_grow_inode(args, &blkno);
558 	if (error)
559 		return error;
560 
561 	dp = args->dp;
562 	tp = args->trans;
563 	error = xfs_da_get_buf(tp, dp, blkno, -1, &bp, args->whichfork);
564 	if (error)
565 		return error;
566 	node = bp->b_addr;
567 	oldroot = blk1->bp->b_addr;
568 	if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
569 	    oldroot->hdr.info.magic == cpu_to_be16(XFS_DA3_NODE_MAGIC)) {
570 		struct xfs_da3_icnode_hdr icnodehdr;
571 
572 		dp->d_ops->node_hdr_from_disk(&icnodehdr, oldroot);
573 		btree = dp->d_ops->node_tree_p(oldroot);
574 		size = (int)((char *)&btree[icnodehdr.count] - (char *)oldroot);
575 		level = icnodehdr.level;
576 
577 		/*
578 		 * we are about to copy oldroot to bp, so set up the type
579 		 * of bp while we know exactly what it will be.
580 		 */
581 		xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DA_NODE_BUF);
582 	} else {
583 		struct xfs_dir3_icleaf_hdr leafhdr;
584 		struct xfs_dir2_leaf_entry *ents;
585 
586 		leaf = (xfs_dir2_leaf_t *)oldroot;
587 		dp->d_ops->leaf_hdr_from_disk(&leafhdr, leaf);
588 		ents = dp->d_ops->leaf_ents_p(leaf);
589 
590 		ASSERT(leafhdr.magic == XFS_DIR2_LEAFN_MAGIC ||
591 		       leafhdr.magic == XFS_DIR3_LEAFN_MAGIC);
592 		size = (int)((char *)&ents[leafhdr.count] - (char *)leaf);
593 		level = 0;
594 
595 		/*
596 		 * we are about to copy oldroot to bp, so set up the type
597 		 * of bp while we know exactly what it will be.
598 		 */
599 		xfs_trans_buf_set_type(tp, bp, XFS_BLFT_DIR_LEAFN_BUF);
600 	}
601 
602 	/*
603 	 * we can copy most of the information in the node from one block to
604 	 * another, but for CRC enabled headers we have to make sure that the
605 	 * block specific identifiers are kept intact. We update the buffer
606 	 * directly for this.
607 	 */
608 	memcpy(node, oldroot, size);
609 	if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DA3_NODE_MAGIC) ||
610 	    oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
611 		struct xfs_da3_intnode *node3 = (struct xfs_da3_intnode *)node;
612 
613 		node3->hdr.info.blkno = cpu_to_be64(bp->b_bn);
614 	}
615 	xfs_trans_log_buf(tp, bp, 0, size - 1);
616 
617 	bp->b_ops = blk1->bp->b_ops;
618 	xfs_trans_buf_copy_type(bp, blk1->bp);
619 	blk1->bp = bp;
620 	blk1->blkno = blkno;
621 
622 	/*
623 	 * Set up the new root node.
624 	 */
625 	error = xfs_da3_node_create(args,
626 		(args->whichfork == XFS_DATA_FORK) ? args->geo->leafblk : 0,
627 		level + 1, &bp, args->whichfork);
628 	if (error)
629 		return error;
630 
631 	node = bp->b_addr;
632 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
633 	btree = dp->d_ops->node_tree_p(node);
634 	btree[0].hashval = cpu_to_be32(blk1->hashval);
635 	btree[0].before = cpu_to_be32(blk1->blkno);
636 	btree[1].hashval = cpu_to_be32(blk2->hashval);
637 	btree[1].before = cpu_to_be32(blk2->blkno);
638 	nodehdr.count = 2;
639 	dp->d_ops->node_hdr_to_disk(node, &nodehdr);
640 
641 #ifdef DEBUG
642 	if (oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
643 	    oldroot->hdr.info.magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
644 		ASSERT(blk1->blkno >= args->geo->leafblk &&
645 		       blk1->blkno < args->geo->freeblk);
646 		ASSERT(blk2->blkno >= args->geo->leafblk &&
647 		       blk2->blkno < args->geo->freeblk);
648 	}
649 #endif
650 
651 	/* Header is already logged by xfs_da_node_create */
652 	xfs_trans_log_buf(tp, bp,
653 		XFS_DA_LOGRANGE(node, btree, sizeof(xfs_da_node_entry_t) * 2));
654 
655 	return 0;
656 }
657 
658 /*
659  * Split the node, rebalance, then add the new entry.
660  */
661 STATIC int						/* error */
662 xfs_da3_node_split(
663 	struct xfs_da_state	*state,
664 	struct xfs_da_state_blk	*oldblk,
665 	struct xfs_da_state_blk	*newblk,
666 	struct xfs_da_state_blk	*addblk,
667 	int			treelevel,
668 	int			*result)
669 {
670 	struct xfs_da_intnode	*node;
671 	struct xfs_da3_icnode_hdr nodehdr;
672 	xfs_dablk_t		blkno;
673 	int			newcount;
674 	int			error;
675 	int			useextra;
676 	struct xfs_inode	*dp = state->args->dp;
677 
678 	trace_xfs_da_node_split(state->args);
679 
680 	node = oldblk->bp->b_addr;
681 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
682 
683 	/*
684 	 * With V2 dirs the extra block is data or freespace.
685 	 */
686 	useextra = state->extravalid && state->args->whichfork == XFS_ATTR_FORK;
687 	newcount = 1 + useextra;
688 	/*
689 	 * Do we have to split the node?
690 	 */
691 	if (nodehdr.count + newcount > state->args->geo->node_ents) {
692 		/*
693 		 * Allocate a new node, add to the doubly linked chain of
694 		 * nodes, then move some of our excess entries into it.
695 		 */
696 		error = xfs_da_grow_inode(state->args, &blkno);
697 		if (error)
698 			return error;	/* GROT: dir is inconsistent */
699 
700 		error = xfs_da3_node_create(state->args, blkno, treelevel,
701 					   &newblk->bp, state->args->whichfork);
702 		if (error)
703 			return error;	/* GROT: dir is inconsistent */
704 		newblk->blkno = blkno;
705 		newblk->magic = XFS_DA_NODE_MAGIC;
706 		xfs_da3_node_rebalance(state, oldblk, newblk);
707 		error = xfs_da3_blk_link(state, oldblk, newblk);
708 		if (error)
709 			return error;
710 		*result = 1;
711 	} else {
712 		*result = 0;
713 	}
714 
715 	/*
716 	 * Insert the new entry(s) into the correct block
717 	 * (updating last hashval in the process).
718 	 *
719 	 * xfs_da3_node_add() inserts BEFORE the given index,
720 	 * and as a result of using node_lookup_int() we always
721 	 * point to a valid entry (not after one), but a split
722 	 * operation always results in a new block whose hashvals
723 	 * FOLLOW the current block.
724 	 *
725 	 * If we had double-split op below us, then add the extra block too.
726 	 */
727 	node = oldblk->bp->b_addr;
728 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
729 	if (oldblk->index <= nodehdr.count) {
730 		oldblk->index++;
731 		xfs_da3_node_add(state, oldblk, addblk);
732 		if (useextra) {
733 			if (state->extraafter)
734 				oldblk->index++;
735 			xfs_da3_node_add(state, oldblk, &state->extrablk);
736 			state->extravalid = 0;
737 		}
738 	} else {
739 		newblk->index++;
740 		xfs_da3_node_add(state, newblk, addblk);
741 		if (useextra) {
742 			if (state->extraafter)
743 				newblk->index++;
744 			xfs_da3_node_add(state, newblk, &state->extrablk);
745 			state->extravalid = 0;
746 		}
747 	}
748 
749 	return 0;
750 }
751 
752 /*
753  * Balance the btree elements between two intermediate nodes,
754  * usually one full and one empty.
755  *
756  * NOTE: if blk2 is empty, then it will get the upper half of blk1.
757  */
758 STATIC void
759 xfs_da3_node_rebalance(
760 	struct xfs_da_state	*state,
761 	struct xfs_da_state_blk	*blk1,
762 	struct xfs_da_state_blk	*blk2)
763 {
764 	struct xfs_da_intnode	*node1;
765 	struct xfs_da_intnode	*node2;
766 	struct xfs_da_intnode	*tmpnode;
767 	struct xfs_da_node_entry *btree1;
768 	struct xfs_da_node_entry *btree2;
769 	struct xfs_da_node_entry *btree_s;
770 	struct xfs_da_node_entry *btree_d;
771 	struct xfs_da3_icnode_hdr nodehdr1;
772 	struct xfs_da3_icnode_hdr nodehdr2;
773 	struct xfs_trans	*tp;
774 	int			count;
775 	int			tmp;
776 	int			swap = 0;
777 	struct xfs_inode	*dp = state->args->dp;
778 
779 	trace_xfs_da_node_rebalance(state->args);
780 
781 	node1 = blk1->bp->b_addr;
782 	node2 = blk2->bp->b_addr;
783 	dp->d_ops->node_hdr_from_disk(&nodehdr1, node1);
784 	dp->d_ops->node_hdr_from_disk(&nodehdr2, node2);
785 	btree1 = dp->d_ops->node_tree_p(node1);
786 	btree2 = dp->d_ops->node_tree_p(node2);
787 
788 	/*
789 	 * Figure out how many entries need to move, and in which direction.
790 	 * Swap the nodes around if that makes it simpler.
791 	 */
792 	if (nodehdr1.count > 0 && nodehdr2.count > 0 &&
793 	    ((be32_to_cpu(btree2[0].hashval) < be32_to_cpu(btree1[0].hashval)) ||
794 	     (be32_to_cpu(btree2[nodehdr2.count - 1].hashval) <
795 			be32_to_cpu(btree1[nodehdr1.count - 1].hashval)))) {
796 		tmpnode = node1;
797 		node1 = node2;
798 		node2 = tmpnode;
799 		dp->d_ops->node_hdr_from_disk(&nodehdr1, node1);
800 		dp->d_ops->node_hdr_from_disk(&nodehdr2, node2);
801 		btree1 = dp->d_ops->node_tree_p(node1);
802 		btree2 = dp->d_ops->node_tree_p(node2);
803 		swap = 1;
804 	}
805 
806 	count = (nodehdr1.count - nodehdr2.count) / 2;
807 	if (count == 0)
808 		return;
809 	tp = state->args->trans;
810 	/*
811 	 * Two cases: high-to-low and low-to-high.
812 	 */
813 	if (count > 0) {
814 		/*
815 		 * Move elements in node2 up to make a hole.
816 		 */
817 		tmp = nodehdr2.count;
818 		if (tmp > 0) {
819 			tmp *= (uint)sizeof(xfs_da_node_entry_t);
820 			btree_s = &btree2[0];
821 			btree_d = &btree2[count];
822 			memmove(btree_d, btree_s, tmp);
823 		}
824 
825 		/*
826 		 * Move the req'd B-tree elements from high in node1 to
827 		 * low in node2.
828 		 */
829 		nodehdr2.count += count;
830 		tmp = count * (uint)sizeof(xfs_da_node_entry_t);
831 		btree_s = &btree1[nodehdr1.count - count];
832 		btree_d = &btree2[0];
833 		memcpy(btree_d, btree_s, tmp);
834 		nodehdr1.count -= count;
835 	} else {
836 		/*
837 		 * Move the req'd B-tree elements from low in node2 to
838 		 * high in node1.
839 		 */
840 		count = -count;
841 		tmp = count * (uint)sizeof(xfs_da_node_entry_t);
842 		btree_s = &btree2[0];
843 		btree_d = &btree1[nodehdr1.count];
844 		memcpy(btree_d, btree_s, tmp);
845 		nodehdr1.count += count;
846 
847 		xfs_trans_log_buf(tp, blk1->bp,
848 			XFS_DA_LOGRANGE(node1, btree_d, tmp));
849 
850 		/*
851 		 * Move elements in node2 down to fill the hole.
852 		 */
853 		tmp  = nodehdr2.count - count;
854 		tmp *= (uint)sizeof(xfs_da_node_entry_t);
855 		btree_s = &btree2[count];
856 		btree_d = &btree2[0];
857 		memmove(btree_d, btree_s, tmp);
858 		nodehdr2.count -= count;
859 	}
860 
861 	/*
862 	 * Log header of node 1 and all current bits of node 2.
863 	 */
864 	dp->d_ops->node_hdr_to_disk(node1, &nodehdr1);
865 	xfs_trans_log_buf(tp, blk1->bp,
866 		XFS_DA_LOGRANGE(node1, &node1->hdr, dp->d_ops->node_hdr_size));
867 
868 	dp->d_ops->node_hdr_to_disk(node2, &nodehdr2);
869 	xfs_trans_log_buf(tp, blk2->bp,
870 		XFS_DA_LOGRANGE(node2, &node2->hdr,
871 				dp->d_ops->node_hdr_size +
872 				(sizeof(btree2[0]) * nodehdr2.count)));
873 
874 	/*
875 	 * Record the last hashval from each block for upward propagation.
876 	 * (note: don't use the swapped node pointers)
877 	 */
878 	if (swap) {
879 		node1 = blk1->bp->b_addr;
880 		node2 = blk2->bp->b_addr;
881 		dp->d_ops->node_hdr_from_disk(&nodehdr1, node1);
882 		dp->d_ops->node_hdr_from_disk(&nodehdr2, node2);
883 		btree1 = dp->d_ops->node_tree_p(node1);
884 		btree2 = dp->d_ops->node_tree_p(node2);
885 	}
886 	blk1->hashval = be32_to_cpu(btree1[nodehdr1.count - 1].hashval);
887 	blk2->hashval = be32_to_cpu(btree2[nodehdr2.count - 1].hashval);
888 
889 	/*
890 	 * Adjust the expected index for insertion.
891 	 */
892 	if (blk1->index >= nodehdr1.count) {
893 		blk2->index = blk1->index - nodehdr1.count;
894 		blk1->index = nodehdr1.count + 1;	/* make it invalid */
895 	}
896 }
897 
898 /*
899  * Add a new entry to an intermediate node.
900  */
901 STATIC void
902 xfs_da3_node_add(
903 	struct xfs_da_state	*state,
904 	struct xfs_da_state_blk	*oldblk,
905 	struct xfs_da_state_blk	*newblk)
906 {
907 	struct xfs_da_intnode	*node;
908 	struct xfs_da3_icnode_hdr nodehdr;
909 	struct xfs_da_node_entry *btree;
910 	int			tmp;
911 	struct xfs_inode	*dp = state->args->dp;
912 
913 	trace_xfs_da_node_add(state->args);
914 
915 	node = oldblk->bp->b_addr;
916 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
917 	btree = dp->d_ops->node_tree_p(node);
918 
919 	ASSERT(oldblk->index >= 0 && oldblk->index <= nodehdr.count);
920 	ASSERT(newblk->blkno != 0);
921 	if (state->args->whichfork == XFS_DATA_FORK)
922 		ASSERT(newblk->blkno >= state->args->geo->leafblk &&
923 		       newblk->blkno < state->args->geo->freeblk);
924 
925 	/*
926 	 * We may need to make some room before we insert the new node.
927 	 */
928 	tmp = 0;
929 	if (oldblk->index < nodehdr.count) {
930 		tmp = (nodehdr.count - oldblk->index) * (uint)sizeof(*btree);
931 		memmove(&btree[oldblk->index + 1], &btree[oldblk->index], tmp);
932 	}
933 	btree[oldblk->index].hashval = cpu_to_be32(newblk->hashval);
934 	btree[oldblk->index].before = cpu_to_be32(newblk->blkno);
935 	xfs_trans_log_buf(state->args->trans, oldblk->bp,
936 		XFS_DA_LOGRANGE(node, &btree[oldblk->index],
937 				tmp + sizeof(*btree)));
938 
939 	nodehdr.count += 1;
940 	dp->d_ops->node_hdr_to_disk(node, &nodehdr);
941 	xfs_trans_log_buf(state->args->trans, oldblk->bp,
942 		XFS_DA_LOGRANGE(node, &node->hdr, dp->d_ops->node_hdr_size));
943 
944 	/*
945 	 * Copy the last hash value from the oldblk to propagate upwards.
946 	 */
947 	oldblk->hashval = be32_to_cpu(btree[nodehdr.count - 1].hashval);
948 }
949 
950 /*========================================================================
951  * Routines used for shrinking the Btree.
952  *========================================================================*/
953 
954 /*
955  * Deallocate an empty leaf node, remove it from its parent,
956  * possibly deallocating that block, etc...
957  */
958 int
959 xfs_da3_join(
960 	struct xfs_da_state	*state)
961 {
962 	struct xfs_da_state_blk	*drop_blk;
963 	struct xfs_da_state_blk	*save_blk;
964 	int			action = 0;
965 	int			error;
966 
967 	trace_xfs_da_join(state->args);
968 
969 	drop_blk = &state->path.blk[ state->path.active-1 ];
970 	save_blk = &state->altpath.blk[ state->path.active-1 ];
971 	ASSERT(state->path.blk[0].magic == XFS_DA_NODE_MAGIC);
972 	ASSERT(drop_blk->magic == XFS_ATTR_LEAF_MAGIC ||
973 	       drop_blk->magic == XFS_DIR2_LEAFN_MAGIC);
974 
975 	/*
976 	 * Walk back up the tree joining/deallocating as necessary.
977 	 * When we stop dropping blocks, break out.
978 	 */
979 	for (  ; state->path.active >= 2; drop_blk--, save_blk--,
980 		 state->path.active--) {
981 		/*
982 		 * See if we can combine the block with a neighbor.
983 		 *   (action == 0) => no options, just leave
984 		 *   (action == 1) => coalesce, then unlink
985 		 *   (action == 2) => block empty, unlink it
986 		 */
987 		switch (drop_blk->magic) {
988 		case XFS_ATTR_LEAF_MAGIC:
989 			error = xfs_attr3_leaf_toosmall(state, &action);
990 			if (error)
991 				return error;
992 			if (action == 0)
993 				return 0;
994 			xfs_attr3_leaf_unbalance(state, drop_blk, save_blk);
995 			break;
996 		case XFS_DIR2_LEAFN_MAGIC:
997 			error = xfs_dir2_leafn_toosmall(state, &action);
998 			if (error)
999 				return error;
1000 			if (action == 0)
1001 				return 0;
1002 			xfs_dir2_leafn_unbalance(state, drop_blk, save_blk);
1003 			break;
1004 		case XFS_DA_NODE_MAGIC:
1005 			/*
1006 			 * Remove the offending node, fixup hashvals,
1007 			 * check for a toosmall neighbor.
1008 			 */
1009 			xfs_da3_node_remove(state, drop_blk);
1010 			xfs_da3_fixhashpath(state, &state->path);
1011 			error = xfs_da3_node_toosmall(state, &action);
1012 			if (error)
1013 				return error;
1014 			if (action == 0)
1015 				return 0;
1016 			xfs_da3_node_unbalance(state, drop_blk, save_blk);
1017 			break;
1018 		}
1019 		xfs_da3_fixhashpath(state, &state->altpath);
1020 		error = xfs_da3_blk_unlink(state, drop_blk, save_blk);
1021 		xfs_da_state_kill_altpath(state);
1022 		if (error)
1023 			return error;
1024 		error = xfs_da_shrink_inode(state->args, drop_blk->blkno,
1025 							 drop_blk->bp);
1026 		drop_blk->bp = NULL;
1027 		if (error)
1028 			return error;
1029 	}
1030 	/*
1031 	 * We joined all the way to the top.  If it turns out that
1032 	 * we only have one entry in the root, make the child block
1033 	 * the new root.
1034 	 */
1035 	xfs_da3_node_remove(state, drop_blk);
1036 	xfs_da3_fixhashpath(state, &state->path);
1037 	error = xfs_da3_root_join(state, &state->path.blk[0]);
1038 	return error;
1039 }
1040 
1041 #ifdef	DEBUG
1042 static void
1043 xfs_da_blkinfo_onlychild_validate(struct xfs_da_blkinfo *blkinfo, __u16 level)
1044 {
1045 	__be16	magic = blkinfo->magic;
1046 
1047 	if (level == 1) {
1048 		ASSERT(magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
1049 		       magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC) ||
1050 		       magic == cpu_to_be16(XFS_ATTR_LEAF_MAGIC) ||
1051 		       magic == cpu_to_be16(XFS_ATTR3_LEAF_MAGIC));
1052 	} else {
1053 		ASSERT(magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
1054 		       magic == cpu_to_be16(XFS_DA3_NODE_MAGIC));
1055 	}
1056 	ASSERT(!blkinfo->forw);
1057 	ASSERT(!blkinfo->back);
1058 }
1059 #else	/* !DEBUG */
1060 #define	xfs_da_blkinfo_onlychild_validate(blkinfo, level)
1061 #endif	/* !DEBUG */
1062 
1063 /*
1064  * We have only one entry in the root.  Copy the only remaining child of
1065  * the old root to block 0 as the new root node.
1066  */
1067 STATIC int
1068 xfs_da3_root_join(
1069 	struct xfs_da_state	*state,
1070 	struct xfs_da_state_blk	*root_blk)
1071 {
1072 	struct xfs_da_intnode	*oldroot;
1073 	struct xfs_da_args	*args;
1074 	xfs_dablk_t		child;
1075 	struct xfs_buf		*bp;
1076 	struct xfs_da3_icnode_hdr oldroothdr;
1077 	struct xfs_da_node_entry *btree;
1078 	int			error;
1079 	struct xfs_inode	*dp = state->args->dp;
1080 
1081 	trace_xfs_da_root_join(state->args);
1082 
1083 	ASSERT(root_blk->magic == XFS_DA_NODE_MAGIC);
1084 
1085 	args = state->args;
1086 	oldroot = root_blk->bp->b_addr;
1087 	dp->d_ops->node_hdr_from_disk(&oldroothdr, oldroot);
1088 	ASSERT(oldroothdr.forw == 0);
1089 	ASSERT(oldroothdr.back == 0);
1090 
1091 	/*
1092 	 * If the root has more than one child, then don't do anything.
1093 	 */
1094 	if (oldroothdr.count > 1)
1095 		return 0;
1096 
1097 	/*
1098 	 * Read in the (only) child block, then copy those bytes into
1099 	 * the root block's buffer and free the original child block.
1100 	 */
1101 	btree = dp->d_ops->node_tree_p(oldroot);
1102 	child = be32_to_cpu(btree[0].before);
1103 	ASSERT(child != 0);
1104 	error = xfs_da3_node_read(args->trans, dp, child, -1, &bp,
1105 					     args->whichfork);
1106 	if (error)
1107 		return error;
1108 	xfs_da_blkinfo_onlychild_validate(bp->b_addr, oldroothdr.level);
1109 
1110 	/*
1111 	 * This could be copying a leaf back into the root block in the case of
1112 	 * there only being a single leaf block left in the tree. Hence we have
1113 	 * to update the b_ops pointer as well to match the buffer type change
1114 	 * that could occur. For dir3 blocks we also need to update the block
1115 	 * number in the buffer header.
1116 	 */
1117 	memcpy(root_blk->bp->b_addr, bp->b_addr, args->geo->blksize);
1118 	root_blk->bp->b_ops = bp->b_ops;
1119 	xfs_trans_buf_copy_type(root_blk->bp, bp);
1120 	if (oldroothdr.magic == XFS_DA3_NODE_MAGIC) {
1121 		struct xfs_da3_blkinfo *da3 = root_blk->bp->b_addr;
1122 		da3->blkno = cpu_to_be64(root_blk->bp->b_bn);
1123 	}
1124 	xfs_trans_log_buf(args->trans, root_blk->bp, 0,
1125 			  args->geo->blksize - 1);
1126 	error = xfs_da_shrink_inode(args, child, bp);
1127 	return error;
1128 }
1129 
1130 /*
1131  * Check a node block and its neighbors to see if the block should be
1132  * collapsed into one or the other neighbor.  Always keep the block
1133  * with the smaller block number.
1134  * If the current block is over 50% full, don't try to join it, return 0.
1135  * If the block is empty, fill in the state structure and return 2.
1136  * If it can be collapsed, fill in the state structure and return 1.
1137  * If nothing can be done, return 0.
1138  */
1139 STATIC int
1140 xfs_da3_node_toosmall(
1141 	struct xfs_da_state	*state,
1142 	int			*action)
1143 {
1144 	struct xfs_da_intnode	*node;
1145 	struct xfs_da_state_blk	*blk;
1146 	struct xfs_da_blkinfo	*info;
1147 	xfs_dablk_t		blkno;
1148 	struct xfs_buf		*bp;
1149 	struct xfs_da3_icnode_hdr nodehdr;
1150 	int			count;
1151 	int			forward;
1152 	int			error;
1153 	int			retval;
1154 	int			i;
1155 	struct xfs_inode	*dp = state->args->dp;
1156 
1157 	trace_xfs_da_node_toosmall(state->args);
1158 
1159 	/*
1160 	 * Check for the degenerate case of the block being over 50% full.
1161 	 * If so, it's not worth even looking to see if we might be able
1162 	 * to coalesce with a sibling.
1163 	 */
1164 	blk = &state->path.blk[ state->path.active-1 ];
1165 	info = blk->bp->b_addr;
1166 	node = (xfs_da_intnode_t *)info;
1167 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1168 	if (nodehdr.count > (state->args->geo->node_ents >> 1)) {
1169 		*action = 0;	/* blk over 50%, don't try to join */
1170 		return 0;	/* blk over 50%, don't try to join */
1171 	}
1172 
1173 	/*
1174 	 * Check for the degenerate case of the block being empty.
1175 	 * If the block is empty, we'll simply delete it, no need to
1176 	 * coalesce it with a sibling block.  We choose (arbitrarily)
1177 	 * to merge with the forward block unless it is NULL.
1178 	 */
1179 	if (nodehdr.count == 0) {
1180 		/*
1181 		 * Make altpath point to the block we want to keep and
1182 		 * path point to the block we want to drop (this one).
1183 		 */
1184 		forward = (info->forw != 0);
1185 		memcpy(&state->altpath, &state->path, sizeof(state->path));
1186 		error = xfs_da3_path_shift(state, &state->altpath, forward,
1187 						 0, &retval);
1188 		if (error)
1189 			return error;
1190 		if (retval) {
1191 			*action = 0;
1192 		} else {
1193 			*action = 2;
1194 		}
1195 		return 0;
1196 	}
1197 
1198 	/*
1199 	 * Examine each sibling block to see if we can coalesce with
1200 	 * at least 25% free space to spare.  We need to figure out
1201 	 * whether to merge with the forward or the backward block.
1202 	 * We prefer coalescing with the lower numbered sibling so as
1203 	 * to shrink a directory over time.
1204 	 */
1205 	count  = state->args->geo->node_ents;
1206 	count -= state->args->geo->node_ents >> 2;
1207 	count -= nodehdr.count;
1208 
1209 	/* start with smaller blk num */
1210 	forward = nodehdr.forw < nodehdr.back;
1211 	for (i = 0; i < 2; forward = !forward, i++) {
1212 		struct xfs_da3_icnode_hdr thdr;
1213 		if (forward)
1214 			blkno = nodehdr.forw;
1215 		else
1216 			blkno = nodehdr.back;
1217 		if (blkno == 0)
1218 			continue;
1219 		error = xfs_da3_node_read(state->args->trans, dp,
1220 					blkno, -1, &bp, state->args->whichfork);
1221 		if (error)
1222 			return error;
1223 
1224 		node = bp->b_addr;
1225 		dp->d_ops->node_hdr_from_disk(&thdr, node);
1226 		xfs_trans_brelse(state->args->trans, bp);
1227 
1228 		if (count - thdr.count >= 0)
1229 			break;	/* fits with at least 25% to spare */
1230 	}
1231 	if (i >= 2) {
1232 		*action = 0;
1233 		return 0;
1234 	}
1235 
1236 	/*
1237 	 * Make altpath point to the block we want to keep (the lower
1238 	 * numbered block) and path point to the block we want to drop.
1239 	 */
1240 	memcpy(&state->altpath, &state->path, sizeof(state->path));
1241 	if (blkno < blk->blkno) {
1242 		error = xfs_da3_path_shift(state, &state->altpath, forward,
1243 						 0, &retval);
1244 	} else {
1245 		error = xfs_da3_path_shift(state, &state->path, forward,
1246 						 0, &retval);
1247 	}
1248 	if (error)
1249 		return error;
1250 	if (retval) {
1251 		*action = 0;
1252 		return 0;
1253 	}
1254 	*action = 1;
1255 	return 0;
1256 }
1257 
1258 /*
1259  * Pick up the last hashvalue from an intermediate node.
1260  */
1261 STATIC uint
1262 xfs_da3_node_lasthash(
1263 	struct xfs_inode	*dp,
1264 	struct xfs_buf		*bp,
1265 	int			*count)
1266 {
1267 	struct xfs_da_intnode	 *node;
1268 	struct xfs_da_node_entry *btree;
1269 	struct xfs_da3_icnode_hdr nodehdr;
1270 
1271 	node = bp->b_addr;
1272 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1273 	if (count)
1274 		*count = nodehdr.count;
1275 	if (!nodehdr.count)
1276 		return 0;
1277 	btree = dp->d_ops->node_tree_p(node);
1278 	return be32_to_cpu(btree[nodehdr.count - 1].hashval);
1279 }
1280 
1281 /*
1282  * Walk back up the tree adjusting hash values as necessary,
1283  * when we stop making changes, return.
1284  */
1285 void
1286 xfs_da3_fixhashpath(
1287 	struct xfs_da_state	*state,
1288 	struct xfs_da_state_path *path)
1289 {
1290 	struct xfs_da_state_blk	*blk;
1291 	struct xfs_da_intnode	*node;
1292 	struct xfs_da_node_entry *btree;
1293 	xfs_dahash_t		lasthash=0;
1294 	int			level;
1295 	int			count;
1296 	struct xfs_inode	*dp = state->args->dp;
1297 
1298 	trace_xfs_da_fixhashpath(state->args);
1299 
1300 	level = path->active-1;
1301 	blk = &path->blk[ level ];
1302 	switch (blk->magic) {
1303 	case XFS_ATTR_LEAF_MAGIC:
1304 		lasthash = xfs_attr_leaf_lasthash(blk->bp, &count);
1305 		if (count == 0)
1306 			return;
1307 		break;
1308 	case XFS_DIR2_LEAFN_MAGIC:
1309 		lasthash = xfs_dir2_leaf_lasthash(dp, blk->bp, &count);
1310 		if (count == 0)
1311 			return;
1312 		break;
1313 	case XFS_DA_NODE_MAGIC:
1314 		lasthash = xfs_da3_node_lasthash(dp, blk->bp, &count);
1315 		if (count == 0)
1316 			return;
1317 		break;
1318 	}
1319 	for (blk--, level--; level >= 0; blk--, level--) {
1320 		struct xfs_da3_icnode_hdr nodehdr;
1321 
1322 		node = blk->bp->b_addr;
1323 		dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1324 		btree = dp->d_ops->node_tree_p(node);
1325 		if (be32_to_cpu(btree[blk->index].hashval) == lasthash)
1326 			break;
1327 		blk->hashval = lasthash;
1328 		btree[blk->index].hashval = cpu_to_be32(lasthash);
1329 		xfs_trans_log_buf(state->args->trans, blk->bp,
1330 				  XFS_DA_LOGRANGE(node, &btree[blk->index],
1331 						  sizeof(*btree)));
1332 
1333 		lasthash = be32_to_cpu(btree[nodehdr.count - 1].hashval);
1334 	}
1335 }
1336 
1337 /*
1338  * Remove an entry from an intermediate node.
1339  */
1340 STATIC void
1341 xfs_da3_node_remove(
1342 	struct xfs_da_state	*state,
1343 	struct xfs_da_state_blk	*drop_blk)
1344 {
1345 	struct xfs_da_intnode	*node;
1346 	struct xfs_da3_icnode_hdr nodehdr;
1347 	struct xfs_da_node_entry *btree;
1348 	int			index;
1349 	int			tmp;
1350 	struct xfs_inode	*dp = state->args->dp;
1351 
1352 	trace_xfs_da_node_remove(state->args);
1353 
1354 	node = drop_blk->bp->b_addr;
1355 	dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1356 	ASSERT(drop_blk->index < nodehdr.count);
1357 	ASSERT(drop_blk->index >= 0);
1358 
1359 	/*
1360 	 * Copy over the offending entry, or just zero it out.
1361 	 */
1362 	index = drop_blk->index;
1363 	btree = dp->d_ops->node_tree_p(node);
1364 	if (index < nodehdr.count - 1) {
1365 		tmp  = nodehdr.count - index - 1;
1366 		tmp *= (uint)sizeof(xfs_da_node_entry_t);
1367 		memmove(&btree[index], &btree[index + 1], tmp);
1368 		xfs_trans_log_buf(state->args->trans, drop_blk->bp,
1369 		    XFS_DA_LOGRANGE(node, &btree[index], tmp));
1370 		index = nodehdr.count - 1;
1371 	}
1372 	memset(&btree[index], 0, sizeof(xfs_da_node_entry_t));
1373 	xfs_trans_log_buf(state->args->trans, drop_blk->bp,
1374 	    XFS_DA_LOGRANGE(node, &btree[index], sizeof(btree[index])));
1375 	nodehdr.count -= 1;
1376 	dp->d_ops->node_hdr_to_disk(node, &nodehdr);
1377 	xfs_trans_log_buf(state->args->trans, drop_blk->bp,
1378 	    XFS_DA_LOGRANGE(node, &node->hdr, dp->d_ops->node_hdr_size));
1379 
1380 	/*
1381 	 * Copy the last hash value from the block to propagate upwards.
1382 	 */
1383 	drop_blk->hashval = be32_to_cpu(btree[index - 1].hashval);
1384 }
1385 
1386 /*
1387  * Unbalance the elements between two intermediate nodes,
1388  * move all Btree elements from one node into another.
1389  */
1390 STATIC void
1391 xfs_da3_node_unbalance(
1392 	struct xfs_da_state	*state,
1393 	struct xfs_da_state_blk	*drop_blk,
1394 	struct xfs_da_state_blk	*save_blk)
1395 {
1396 	struct xfs_da_intnode	*drop_node;
1397 	struct xfs_da_intnode	*save_node;
1398 	struct xfs_da_node_entry *drop_btree;
1399 	struct xfs_da_node_entry *save_btree;
1400 	struct xfs_da3_icnode_hdr drop_hdr;
1401 	struct xfs_da3_icnode_hdr save_hdr;
1402 	struct xfs_trans	*tp;
1403 	int			sindex;
1404 	int			tmp;
1405 	struct xfs_inode	*dp = state->args->dp;
1406 
1407 	trace_xfs_da_node_unbalance(state->args);
1408 
1409 	drop_node = drop_blk->bp->b_addr;
1410 	save_node = save_blk->bp->b_addr;
1411 	dp->d_ops->node_hdr_from_disk(&drop_hdr, drop_node);
1412 	dp->d_ops->node_hdr_from_disk(&save_hdr, save_node);
1413 	drop_btree = dp->d_ops->node_tree_p(drop_node);
1414 	save_btree = dp->d_ops->node_tree_p(save_node);
1415 	tp = state->args->trans;
1416 
1417 	/*
1418 	 * If the dying block has lower hashvals, then move all the
1419 	 * elements in the remaining block up to make a hole.
1420 	 */
1421 	if ((be32_to_cpu(drop_btree[0].hashval) <
1422 			be32_to_cpu(save_btree[0].hashval)) ||
1423 	    (be32_to_cpu(drop_btree[drop_hdr.count - 1].hashval) <
1424 			be32_to_cpu(save_btree[save_hdr.count - 1].hashval))) {
1425 		/* XXX: check this - is memmove dst correct? */
1426 		tmp = save_hdr.count * sizeof(xfs_da_node_entry_t);
1427 		memmove(&save_btree[drop_hdr.count], &save_btree[0], tmp);
1428 
1429 		sindex = 0;
1430 		xfs_trans_log_buf(tp, save_blk->bp,
1431 			XFS_DA_LOGRANGE(save_node, &save_btree[0],
1432 				(save_hdr.count + drop_hdr.count) *
1433 						sizeof(xfs_da_node_entry_t)));
1434 	} else {
1435 		sindex = save_hdr.count;
1436 		xfs_trans_log_buf(tp, save_blk->bp,
1437 			XFS_DA_LOGRANGE(save_node, &save_btree[sindex],
1438 				drop_hdr.count * sizeof(xfs_da_node_entry_t)));
1439 	}
1440 
1441 	/*
1442 	 * Move all the B-tree elements from drop_blk to save_blk.
1443 	 */
1444 	tmp = drop_hdr.count * (uint)sizeof(xfs_da_node_entry_t);
1445 	memcpy(&save_btree[sindex], &drop_btree[0], tmp);
1446 	save_hdr.count += drop_hdr.count;
1447 
1448 	dp->d_ops->node_hdr_to_disk(save_node, &save_hdr);
1449 	xfs_trans_log_buf(tp, save_blk->bp,
1450 		XFS_DA_LOGRANGE(save_node, &save_node->hdr,
1451 				dp->d_ops->node_hdr_size));
1452 
1453 	/*
1454 	 * Save the last hashval in the remaining block for upward propagation.
1455 	 */
1456 	save_blk->hashval = be32_to_cpu(save_btree[save_hdr.count - 1].hashval);
1457 }
1458 
1459 /*========================================================================
1460  * Routines used for finding things in the Btree.
1461  *========================================================================*/
1462 
1463 /*
1464  * Walk down the Btree looking for a particular filename, filling
1465  * in the state structure as we go.
1466  *
1467  * We will set the state structure to point to each of the elements
1468  * in each of the nodes where either the hashval is or should be.
1469  *
1470  * We support duplicate hashval's so for each entry in the current
1471  * node that could contain the desired hashval, descend.  This is a
1472  * pruned depth-first tree search.
1473  */
1474 int							/* error */
1475 xfs_da3_node_lookup_int(
1476 	struct xfs_da_state	*state,
1477 	int			*result)
1478 {
1479 	struct xfs_da_state_blk	*blk;
1480 	struct xfs_da_blkinfo	*curr;
1481 	struct xfs_da_intnode	*node;
1482 	struct xfs_da_node_entry *btree;
1483 	struct xfs_da3_icnode_hdr nodehdr;
1484 	struct xfs_da_args	*args;
1485 	xfs_dablk_t		blkno;
1486 	xfs_dahash_t		hashval;
1487 	xfs_dahash_t		btreehashval;
1488 	int			probe;
1489 	int			span;
1490 	int			max;
1491 	int			error;
1492 	int			retval;
1493 	unsigned int		expected_level = 0;
1494 	struct xfs_inode	*dp = state->args->dp;
1495 
1496 	args = state->args;
1497 
1498 	/*
1499 	 * Descend thru the B-tree searching each level for the right
1500 	 * node to use, until the right hashval is found.
1501 	 */
1502 	blkno = args->geo->leafblk;
1503 	for (blk = &state->path.blk[0], state->path.active = 1;
1504 			 state->path.active <= XFS_DA_NODE_MAXDEPTH;
1505 			 blk++, state->path.active++) {
1506 		/*
1507 		 * Read the next node down in the tree.
1508 		 */
1509 		blk->blkno = blkno;
1510 		error = xfs_da3_node_read(args->trans, args->dp, blkno,
1511 					-1, &blk->bp, args->whichfork);
1512 		if (error) {
1513 			blk->blkno = 0;
1514 			state->path.active--;
1515 			return error;
1516 		}
1517 		curr = blk->bp->b_addr;
1518 		blk->magic = be16_to_cpu(curr->magic);
1519 
1520 		if (blk->magic == XFS_ATTR_LEAF_MAGIC ||
1521 		    blk->magic == XFS_ATTR3_LEAF_MAGIC) {
1522 			blk->magic = XFS_ATTR_LEAF_MAGIC;
1523 			blk->hashval = xfs_attr_leaf_lasthash(blk->bp, NULL);
1524 			break;
1525 		}
1526 
1527 		if (blk->magic == XFS_DIR2_LEAFN_MAGIC ||
1528 		    blk->magic == XFS_DIR3_LEAFN_MAGIC) {
1529 			blk->magic = XFS_DIR2_LEAFN_MAGIC;
1530 			blk->hashval = xfs_dir2_leaf_lasthash(args->dp,
1531 							      blk->bp, NULL);
1532 			break;
1533 		}
1534 
1535 		blk->magic = XFS_DA_NODE_MAGIC;
1536 
1537 
1538 		/*
1539 		 * Search an intermediate node for a match.
1540 		 */
1541 		node = blk->bp->b_addr;
1542 		dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1543 		btree = dp->d_ops->node_tree_p(node);
1544 
1545 		/* Tree taller than we can handle; bail out! */
1546 		if (nodehdr.level >= XFS_DA_NODE_MAXDEPTH)
1547 			return -EFSCORRUPTED;
1548 
1549 		/* Check the level from the root. */
1550 		if (blkno == args->geo->leafblk)
1551 			expected_level = nodehdr.level - 1;
1552 		else if (expected_level != nodehdr.level)
1553 			return -EFSCORRUPTED;
1554 		else
1555 			expected_level--;
1556 
1557 		max = nodehdr.count;
1558 		blk->hashval = be32_to_cpu(btree[max - 1].hashval);
1559 
1560 		/*
1561 		 * Binary search.  (note: small blocks will skip loop)
1562 		 */
1563 		probe = span = max / 2;
1564 		hashval = args->hashval;
1565 		while (span > 4) {
1566 			span /= 2;
1567 			btreehashval = be32_to_cpu(btree[probe].hashval);
1568 			if (btreehashval < hashval)
1569 				probe += span;
1570 			else if (btreehashval > hashval)
1571 				probe -= span;
1572 			else
1573 				break;
1574 		}
1575 		ASSERT((probe >= 0) && (probe < max));
1576 		ASSERT((span <= 4) ||
1577 			(be32_to_cpu(btree[probe].hashval) == hashval));
1578 
1579 		/*
1580 		 * Since we may have duplicate hashval's, find the first
1581 		 * matching hashval in the node.
1582 		 */
1583 		while (probe > 0 &&
1584 		       be32_to_cpu(btree[probe].hashval) >= hashval) {
1585 			probe--;
1586 		}
1587 		while (probe < max &&
1588 		       be32_to_cpu(btree[probe].hashval) < hashval) {
1589 			probe++;
1590 		}
1591 
1592 		/*
1593 		 * Pick the right block to descend on.
1594 		 */
1595 		if (probe == max) {
1596 			blk->index = max - 1;
1597 			blkno = be32_to_cpu(btree[max - 1].before);
1598 		} else {
1599 			blk->index = probe;
1600 			blkno = be32_to_cpu(btree[probe].before);
1601 		}
1602 
1603 		/* We can't point back to the root. */
1604 		if (blkno == args->geo->leafblk)
1605 			return -EFSCORRUPTED;
1606 	}
1607 
1608 	if (expected_level != 0)
1609 		return -EFSCORRUPTED;
1610 
1611 	/*
1612 	 * A leaf block that ends in the hashval that we are interested in
1613 	 * (final hashval == search hashval) means that the next block may
1614 	 * contain more entries with the same hashval, shift upward to the
1615 	 * next leaf and keep searching.
1616 	 */
1617 	for (;;) {
1618 		if (blk->magic == XFS_DIR2_LEAFN_MAGIC) {
1619 			retval = xfs_dir2_leafn_lookup_int(blk->bp, args,
1620 							&blk->index, state);
1621 		} else if (blk->magic == XFS_ATTR_LEAF_MAGIC) {
1622 			retval = xfs_attr3_leaf_lookup_int(blk->bp, args);
1623 			blk->index = args->index;
1624 			args->blkno = blk->blkno;
1625 		} else {
1626 			ASSERT(0);
1627 			return -EFSCORRUPTED;
1628 		}
1629 		if (((retval == -ENOENT) || (retval == -ENOATTR)) &&
1630 		    (blk->hashval == args->hashval)) {
1631 			error = xfs_da3_path_shift(state, &state->path, 1, 1,
1632 							 &retval);
1633 			if (error)
1634 				return error;
1635 			if (retval == 0) {
1636 				continue;
1637 			} else if (blk->magic == XFS_ATTR_LEAF_MAGIC) {
1638 				/* path_shift() gives ENOENT */
1639 				retval = -ENOATTR;
1640 			}
1641 		}
1642 		break;
1643 	}
1644 	*result = retval;
1645 	return 0;
1646 }
1647 
1648 /*========================================================================
1649  * Utility routines.
1650  *========================================================================*/
1651 
1652 /*
1653  * Compare two intermediate nodes for "order".
1654  */
1655 STATIC int
1656 xfs_da3_node_order(
1657 	struct xfs_inode *dp,
1658 	struct xfs_buf	*node1_bp,
1659 	struct xfs_buf	*node2_bp)
1660 {
1661 	struct xfs_da_intnode	*node1;
1662 	struct xfs_da_intnode	*node2;
1663 	struct xfs_da_node_entry *btree1;
1664 	struct xfs_da_node_entry *btree2;
1665 	struct xfs_da3_icnode_hdr node1hdr;
1666 	struct xfs_da3_icnode_hdr node2hdr;
1667 
1668 	node1 = node1_bp->b_addr;
1669 	node2 = node2_bp->b_addr;
1670 	dp->d_ops->node_hdr_from_disk(&node1hdr, node1);
1671 	dp->d_ops->node_hdr_from_disk(&node2hdr, node2);
1672 	btree1 = dp->d_ops->node_tree_p(node1);
1673 	btree2 = dp->d_ops->node_tree_p(node2);
1674 
1675 	if (node1hdr.count > 0 && node2hdr.count > 0 &&
1676 	    ((be32_to_cpu(btree2[0].hashval) < be32_to_cpu(btree1[0].hashval)) ||
1677 	     (be32_to_cpu(btree2[node2hdr.count - 1].hashval) <
1678 	      be32_to_cpu(btree1[node1hdr.count - 1].hashval)))) {
1679 		return 1;
1680 	}
1681 	return 0;
1682 }
1683 
1684 /*
1685  * Link a new block into a doubly linked list of blocks (of whatever type).
1686  */
1687 int							/* error */
1688 xfs_da3_blk_link(
1689 	struct xfs_da_state	*state,
1690 	struct xfs_da_state_blk	*old_blk,
1691 	struct xfs_da_state_blk	*new_blk)
1692 {
1693 	struct xfs_da_blkinfo	*old_info;
1694 	struct xfs_da_blkinfo	*new_info;
1695 	struct xfs_da_blkinfo	*tmp_info;
1696 	struct xfs_da_args	*args;
1697 	struct xfs_buf		*bp;
1698 	int			before = 0;
1699 	int			error;
1700 	struct xfs_inode	*dp = state->args->dp;
1701 
1702 	/*
1703 	 * Set up environment.
1704 	 */
1705 	args = state->args;
1706 	ASSERT(args != NULL);
1707 	old_info = old_blk->bp->b_addr;
1708 	new_info = new_blk->bp->b_addr;
1709 	ASSERT(old_blk->magic == XFS_DA_NODE_MAGIC ||
1710 	       old_blk->magic == XFS_DIR2_LEAFN_MAGIC ||
1711 	       old_blk->magic == XFS_ATTR_LEAF_MAGIC);
1712 
1713 	switch (old_blk->magic) {
1714 	case XFS_ATTR_LEAF_MAGIC:
1715 		before = xfs_attr_leaf_order(old_blk->bp, new_blk->bp);
1716 		break;
1717 	case XFS_DIR2_LEAFN_MAGIC:
1718 		before = xfs_dir2_leafn_order(dp, old_blk->bp, new_blk->bp);
1719 		break;
1720 	case XFS_DA_NODE_MAGIC:
1721 		before = xfs_da3_node_order(dp, old_blk->bp, new_blk->bp);
1722 		break;
1723 	}
1724 
1725 	/*
1726 	 * Link blocks in appropriate order.
1727 	 */
1728 	if (before) {
1729 		/*
1730 		 * Link new block in before existing block.
1731 		 */
1732 		trace_xfs_da_link_before(args);
1733 		new_info->forw = cpu_to_be32(old_blk->blkno);
1734 		new_info->back = old_info->back;
1735 		if (old_info->back) {
1736 			error = xfs_da3_node_read(args->trans, dp,
1737 						be32_to_cpu(old_info->back),
1738 						-1, &bp, args->whichfork);
1739 			if (error)
1740 				return error;
1741 			ASSERT(bp != NULL);
1742 			tmp_info = bp->b_addr;
1743 			ASSERT(tmp_info->magic == old_info->magic);
1744 			ASSERT(be32_to_cpu(tmp_info->forw) == old_blk->blkno);
1745 			tmp_info->forw = cpu_to_be32(new_blk->blkno);
1746 			xfs_trans_log_buf(args->trans, bp, 0, sizeof(*tmp_info)-1);
1747 		}
1748 		old_info->back = cpu_to_be32(new_blk->blkno);
1749 	} else {
1750 		/*
1751 		 * Link new block in after existing block.
1752 		 */
1753 		trace_xfs_da_link_after(args);
1754 		new_info->forw = old_info->forw;
1755 		new_info->back = cpu_to_be32(old_blk->blkno);
1756 		if (old_info->forw) {
1757 			error = xfs_da3_node_read(args->trans, dp,
1758 						be32_to_cpu(old_info->forw),
1759 						-1, &bp, args->whichfork);
1760 			if (error)
1761 				return error;
1762 			ASSERT(bp != NULL);
1763 			tmp_info = bp->b_addr;
1764 			ASSERT(tmp_info->magic == old_info->magic);
1765 			ASSERT(be32_to_cpu(tmp_info->back) == old_blk->blkno);
1766 			tmp_info->back = cpu_to_be32(new_blk->blkno);
1767 			xfs_trans_log_buf(args->trans, bp, 0, sizeof(*tmp_info)-1);
1768 		}
1769 		old_info->forw = cpu_to_be32(new_blk->blkno);
1770 	}
1771 
1772 	xfs_trans_log_buf(args->trans, old_blk->bp, 0, sizeof(*tmp_info) - 1);
1773 	xfs_trans_log_buf(args->trans, new_blk->bp, 0, sizeof(*tmp_info) - 1);
1774 	return 0;
1775 }
1776 
1777 /*
1778  * Unlink a block from a doubly linked list of blocks.
1779  */
1780 STATIC int						/* error */
1781 xfs_da3_blk_unlink(
1782 	struct xfs_da_state	*state,
1783 	struct xfs_da_state_blk	*drop_blk,
1784 	struct xfs_da_state_blk	*save_blk)
1785 {
1786 	struct xfs_da_blkinfo	*drop_info;
1787 	struct xfs_da_blkinfo	*save_info;
1788 	struct xfs_da_blkinfo	*tmp_info;
1789 	struct xfs_da_args	*args;
1790 	struct xfs_buf		*bp;
1791 	int			error;
1792 
1793 	/*
1794 	 * Set up environment.
1795 	 */
1796 	args = state->args;
1797 	ASSERT(args != NULL);
1798 	save_info = save_blk->bp->b_addr;
1799 	drop_info = drop_blk->bp->b_addr;
1800 	ASSERT(save_blk->magic == XFS_DA_NODE_MAGIC ||
1801 	       save_blk->magic == XFS_DIR2_LEAFN_MAGIC ||
1802 	       save_blk->magic == XFS_ATTR_LEAF_MAGIC);
1803 	ASSERT(save_blk->magic == drop_blk->magic);
1804 	ASSERT((be32_to_cpu(save_info->forw) == drop_blk->blkno) ||
1805 	       (be32_to_cpu(save_info->back) == drop_blk->blkno));
1806 	ASSERT((be32_to_cpu(drop_info->forw) == save_blk->blkno) ||
1807 	       (be32_to_cpu(drop_info->back) == save_blk->blkno));
1808 
1809 	/*
1810 	 * Unlink the leaf block from the doubly linked chain of leaves.
1811 	 */
1812 	if (be32_to_cpu(save_info->back) == drop_blk->blkno) {
1813 		trace_xfs_da_unlink_back(args);
1814 		save_info->back = drop_info->back;
1815 		if (drop_info->back) {
1816 			error = xfs_da3_node_read(args->trans, args->dp,
1817 						be32_to_cpu(drop_info->back),
1818 						-1, &bp, args->whichfork);
1819 			if (error)
1820 				return error;
1821 			ASSERT(bp != NULL);
1822 			tmp_info = bp->b_addr;
1823 			ASSERT(tmp_info->magic == save_info->magic);
1824 			ASSERT(be32_to_cpu(tmp_info->forw) == drop_blk->blkno);
1825 			tmp_info->forw = cpu_to_be32(save_blk->blkno);
1826 			xfs_trans_log_buf(args->trans, bp, 0,
1827 						    sizeof(*tmp_info) - 1);
1828 		}
1829 	} else {
1830 		trace_xfs_da_unlink_forward(args);
1831 		save_info->forw = drop_info->forw;
1832 		if (drop_info->forw) {
1833 			error = xfs_da3_node_read(args->trans, args->dp,
1834 						be32_to_cpu(drop_info->forw),
1835 						-1, &bp, args->whichfork);
1836 			if (error)
1837 				return error;
1838 			ASSERT(bp != NULL);
1839 			tmp_info = bp->b_addr;
1840 			ASSERT(tmp_info->magic == save_info->magic);
1841 			ASSERT(be32_to_cpu(tmp_info->back) == drop_blk->blkno);
1842 			tmp_info->back = cpu_to_be32(save_blk->blkno);
1843 			xfs_trans_log_buf(args->trans, bp, 0,
1844 						    sizeof(*tmp_info) - 1);
1845 		}
1846 	}
1847 
1848 	xfs_trans_log_buf(args->trans, save_blk->bp, 0, sizeof(*save_info) - 1);
1849 	return 0;
1850 }
1851 
1852 /*
1853  * Move a path "forward" or "!forward" one block at the current level.
1854  *
1855  * This routine will adjust a "path" to point to the next block
1856  * "forward" (higher hashvalues) or "!forward" (lower hashvals) in the
1857  * Btree, including updating pointers to the intermediate nodes between
1858  * the new bottom and the root.
1859  */
1860 int							/* error */
1861 xfs_da3_path_shift(
1862 	struct xfs_da_state	*state,
1863 	struct xfs_da_state_path *path,
1864 	int			forward,
1865 	int			release,
1866 	int			*result)
1867 {
1868 	struct xfs_da_state_blk	*blk;
1869 	struct xfs_da_blkinfo	*info;
1870 	struct xfs_da_intnode	*node;
1871 	struct xfs_da_args	*args;
1872 	struct xfs_da_node_entry *btree;
1873 	struct xfs_da3_icnode_hdr nodehdr;
1874 	struct xfs_buf		*bp;
1875 	xfs_dablk_t		blkno = 0;
1876 	int			level;
1877 	int			error;
1878 	struct xfs_inode	*dp = state->args->dp;
1879 
1880 	trace_xfs_da_path_shift(state->args);
1881 
1882 	/*
1883 	 * Roll up the Btree looking for the first block where our
1884 	 * current index is not at the edge of the block.  Note that
1885 	 * we skip the bottom layer because we want the sibling block.
1886 	 */
1887 	args = state->args;
1888 	ASSERT(args != NULL);
1889 	ASSERT(path != NULL);
1890 	ASSERT((path->active > 0) && (path->active < XFS_DA_NODE_MAXDEPTH));
1891 	level = (path->active-1) - 1;	/* skip bottom layer in path */
1892 	for (blk = &path->blk[level]; level >= 0; blk--, level--) {
1893 		node = blk->bp->b_addr;
1894 		dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1895 		btree = dp->d_ops->node_tree_p(node);
1896 
1897 		if (forward && (blk->index < nodehdr.count - 1)) {
1898 			blk->index++;
1899 			blkno = be32_to_cpu(btree[blk->index].before);
1900 			break;
1901 		} else if (!forward && (blk->index > 0)) {
1902 			blk->index--;
1903 			blkno = be32_to_cpu(btree[blk->index].before);
1904 			break;
1905 		}
1906 	}
1907 	if (level < 0) {
1908 		*result = -ENOENT;	/* we're out of our tree */
1909 		ASSERT(args->op_flags & XFS_DA_OP_OKNOENT);
1910 		return 0;
1911 	}
1912 
1913 	/*
1914 	 * Roll down the edge of the subtree until we reach the
1915 	 * same depth we were at originally.
1916 	 */
1917 	for (blk++, level++; level < path->active; blk++, level++) {
1918 		/*
1919 		 * Read the next child block into a local buffer.
1920 		 */
1921 		error = xfs_da3_node_read(args->trans, dp, blkno, -1, &bp,
1922 					  args->whichfork);
1923 		if (error)
1924 			return error;
1925 
1926 		/*
1927 		 * Release the old block (if it's dirty, the trans doesn't
1928 		 * actually let go) and swap the local buffer into the path
1929 		 * structure. This ensures failure of the above read doesn't set
1930 		 * a NULL buffer in an active slot in the path.
1931 		 */
1932 		if (release)
1933 			xfs_trans_brelse(args->trans, blk->bp);
1934 		blk->blkno = blkno;
1935 		blk->bp = bp;
1936 
1937 		info = blk->bp->b_addr;
1938 		ASSERT(info->magic == cpu_to_be16(XFS_DA_NODE_MAGIC) ||
1939 		       info->magic == cpu_to_be16(XFS_DA3_NODE_MAGIC) ||
1940 		       info->magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
1941 		       info->magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC) ||
1942 		       info->magic == cpu_to_be16(XFS_ATTR_LEAF_MAGIC) ||
1943 		       info->magic == cpu_to_be16(XFS_ATTR3_LEAF_MAGIC));
1944 
1945 
1946 		/*
1947 		 * Note: we flatten the magic number to a single type so we
1948 		 * don't have to compare against crc/non-crc types elsewhere.
1949 		 */
1950 		switch (be16_to_cpu(info->magic)) {
1951 		case XFS_DA_NODE_MAGIC:
1952 		case XFS_DA3_NODE_MAGIC:
1953 			blk->magic = XFS_DA_NODE_MAGIC;
1954 			node = (xfs_da_intnode_t *)info;
1955 			dp->d_ops->node_hdr_from_disk(&nodehdr, node);
1956 			btree = dp->d_ops->node_tree_p(node);
1957 			blk->hashval = be32_to_cpu(btree[nodehdr.count - 1].hashval);
1958 			if (forward)
1959 				blk->index = 0;
1960 			else
1961 				blk->index = nodehdr.count - 1;
1962 			blkno = be32_to_cpu(btree[blk->index].before);
1963 			break;
1964 		case XFS_ATTR_LEAF_MAGIC:
1965 		case XFS_ATTR3_LEAF_MAGIC:
1966 			blk->magic = XFS_ATTR_LEAF_MAGIC;
1967 			ASSERT(level == path->active-1);
1968 			blk->index = 0;
1969 			blk->hashval = xfs_attr_leaf_lasthash(blk->bp, NULL);
1970 			break;
1971 		case XFS_DIR2_LEAFN_MAGIC:
1972 		case XFS_DIR3_LEAFN_MAGIC:
1973 			blk->magic = XFS_DIR2_LEAFN_MAGIC;
1974 			ASSERT(level == path->active-1);
1975 			blk->index = 0;
1976 			blk->hashval = xfs_dir2_leaf_lasthash(args->dp,
1977 							      blk->bp, NULL);
1978 			break;
1979 		default:
1980 			ASSERT(0);
1981 			break;
1982 		}
1983 	}
1984 	*result = 0;
1985 	return 0;
1986 }
1987 
1988 
1989 /*========================================================================
1990  * Utility routines.
1991  *========================================================================*/
1992 
1993 /*
1994  * Implement a simple hash on a character string.
1995  * Rotate the hash value by 7 bits, then XOR each character in.
1996  * This is implemented with some source-level loop unrolling.
1997  */
1998 xfs_dahash_t
1999 xfs_da_hashname(const uint8_t *name, int namelen)
2000 {
2001 	xfs_dahash_t hash;
2002 
2003 	/*
2004 	 * Do four characters at a time as long as we can.
2005 	 */
2006 	for (hash = 0; namelen >= 4; namelen -= 4, name += 4)
2007 		hash = (name[0] << 21) ^ (name[1] << 14) ^ (name[2] << 7) ^
2008 		       (name[3] << 0) ^ rol32(hash, 7 * 4);
2009 
2010 	/*
2011 	 * Now do the rest of the characters.
2012 	 */
2013 	switch (namelen) {
2014 	case 3:
2015 		return (name[0] << 14) ^ (name[1] << 7) ^ (name[2] << 0) ^
2016 		       rol32(hash, 7 * 3);
2017 	case 2:
2018 		return (name[0] << 7) ^ (name[1] << 0) ^ rol32(hash, 7 * 2);
2019 	case 1:
2020 		return (name[0] << 0) ^ rol32(hash, 7 * 1);
2021 	default: /* case 0: */
2022 		return hash;
2023 	}
2024 }
2025 
2026 enum xfs_dacmp
2027 xfs_da_compname(
2028 	struct xfs_da_args *args,
2029 	const unsigned char *name,
2030 	int		len)
2031 {
2032 	return (args->namelen == len && memcmp(args->name, name, len) == 0) ?
2033 					XFS_CMP_EXACT : XFS_CMP_DIFFERENT;
2034 }
2035 
2036 static xfs_dahash_t
2037 xfs_default_hashname(
2038 	struct xfs_name	*name)
2039 {
2040 	return xfs_da_hashname(name->name, name->len);
2041 }
2042 
2043 const struct xfs_nameops xfs_default_nameops = {
2044 	.hashname	= xfs_default_hashname,
2045 	.compname	= xfs_da_compname
2046 };
2047 
2048 int
2049 xfs_da_grow_inode_int(
2050 	struct xfs_da_args	*args,
2051 	xfs_fileoff_t		*bno,
2052 	int			count)
2053 {
2054 	struct xfs_trans	*tp = args->trans;
2055 	struct xfs_inode	*dp = args->dp;
2056 	int			w = args->whichfork;
2057 	xfs_rfsblock_t		nblks = dp->i_d.di_nblocks;
2058 	struct xfs_bmbt_irec	map, *mapp;
2059 	int			nmap, error, got, i, mapi;
2060 
2061 	/*
2062 	 * Find a spot in the file space to put the new block.
2063 	 */
2064 	error = xfs_bmap_first_unused(tp, dp, count, bno, w);
2065 	if (error)
2066 		return error;
2067 
2068 	/*
2069 	 * Try mapping it in one filesystem block.
2070 	 */
2071 	nmap = 1;
2072 	ASSERT(args->firstblock != NULL);
2073 	error = xfs_bmapi_write(tp, dp, *bno, count,
2074 			xfs_bmapi_aflag(w)|XFS_BMAPI_METADATA|XFS_BMAPI_CONTIG,
2075 			args->firstblock, args->total, &map, &nmap,
2076 			args->dfops);
2077 	if (error)
2078 		return error;
2079 
2080 	ASSERT(nmap <= 1);
2081 	if (nmap == 1) {
2082 		mapp = &map;
2083 		mapi = 1;
2084 	} else if (nmap == 0 && count > 1) {
2085 		xfs_fileoff_t		b;
2086 		int			c;
2087 
2088 		/*
2089 		 * If we didn't get it and the block might work if fragmented,
2090 		 * try without the CONTIG flag.  Loop until we get it all.
2091 		 */
2092 		mapp = kmem_alloc(sizeof(*mapp) * count, KM_SLEEP);
2093 		for (b = *bno, mapi = 0; b < *bno + count; ) {
2094 			nmap = MIN(XFS_BMAP_MAX_NMAP, count);
2095 			c = (int)(*bno + count - b);
2096 			error = xfs_bmapi_write(tp, dp, b, c,
2097 					xfs_bmapi_aflag(w)|XFS_BMAPI_METADATA,
2098 					args->firstblock, args->total,
2099 					&mapp[mapi], &nmap, args->dfops);
2100 			if (error)
2101 				goto out_free_map;
2102 			if (nmap < 1)
2103 				break;
2104 			mapi += nmap;
2105 			b = mapp[mapi - 1].br_startoff +
2106 			    mapp[mapi - 1].br_blockcount;
2107 		}
2108 	} else {
2109 		mapi = 0;
2110 		mapp = NULL;
2111 	}
2112 
2113 	/*
2114 	 * Count the blocks we got, make sure it matches the total.
2115 	 */
2116 	for (i = 0, got = 0; i < mapi; i++)
2117 		got += mapp[i].br_blockcount;
2118 	if (got != count || mapp[0].br_startoff != *bno ||
2119 	    mapp[mapi - 1].br_startoff + mapp[mapi - 1].br_blockcount !=
2120 	    *bno + count) {
2121 		error = -ENOSPC;
2122 		goto out_free_map;
2123 	}
2124 
2125 	/* account for newly allocated blocks in reserved blocks total */
2126 	args->total -= dp->i_d.di_nblocks - nblks;
2127 
2128 out_free_map:
2129 	if (mapp != &map)
2130 		kmem_free(mapp);
2131 	return error;
2132 }
2133 
2134 /*
2135  * Add a block to the btree ahead of the file.
2136  * Return the new block number to the caller.
2137  */
2138 int
2139 xfs_da_grow_inode(
2140 	struct xfs_da_args	*args,
2141 	xfs_dablk_t		*new_blkno)
2142 {
2143 	xfs_fileoff_t		bno;
2144 	int			error;
2145 
2146 	trace_xfs_da_grow_inode(args);
2147 
2148 	bno = args->geo->leafblk;
2149 	error = xfs_da_grow_inode_int(args, &bno, args->geo->fsbcount);
2150 	if (!error)
2151 		*new_blkno = (xfs_dablk_t)bno;
2152 	return error;
2153 }
2154 
2155 /*
2156  * Ick.  We need to always be able to remove a btree block, even
2157  * if there's no space reservation because the filesystem is full.
2158  * This is called if xfs_bunmapi on a btree block fails due to ENOSPC.
2159  * It swaps the target block with the last block in the file.  The
2160  * last block in the file can always be removed since it can't cause
2161  * a bmap btree split to do that.
2162  */
2163 STATIC int
2164 xfs_da3_swap_lastblock(
2165 	struct xfs_da_args	*args,
2166 	xfs_dablk_t		*dead_blknop,
2167 	struct xfs_buf		**dead_bufp)
2168 {
2169 	struct xfs_da_blkinfo	*dead_info;
2170 	struct xfs_da_blkinfo	*sib_info;
2171 	struct xfs_da_intnode	*par_node;
2172 	struct xfs_da_intnode	*dead_node;
2173 	struct xfs_dir2_leaf	*dead_leaf2;
2174 	struct xfs_da_node_entry *btree;
2175 	struct xfs_da3_icnode_hdr par_hdr;
2176 	struct xfs_inode	*dp;
2177 	struct xfs_trans	*tp;
2178 	struct xfs_mount	*mp;
2179 	struct xfs_buf		*dead_buf;
2180 	struct xfs_buf		*last_buf;
2181 	struct xfs_buf		*sib_buf;
2182 	struct xfs_buf		*par_buf;
2183 	xfs_dahash_t		dead_hash;
2184 	xfs_fileoff_t		lastoff;
2185 	xfs_dablk_t		dead_blkno;
2186 	xfs_dablk_t		last_blkno;
2187 	xfs_dablk_t		sib_blkno;
2188 	xfs_dablk_t		par_blkno;
2189 	int			error;
2190 	int			w;
2191 	int			entno;
2192 	int			level;
2193 	int			dead_level;
2194 
2195 	trace_xfs_da_swap_lastblock(args);
2196 
2197 	dead_buf = *dead_bufp;
2198 	dead_blkno = *dead_blknop;
2199 	tp = args->trans;
2200 	dp = args->dp;
2201 	w = args->whichfork;
2202 	ASSERT(w == XFS_DATA_FORK);
2203 	mp = dp->i_mount;
2204 	lastoff = args->geo->freeblk;
2205 	error = xfs_bmap_last_before(tp, dp, &lastoff, w);
2206 	if (error)
2207 		return error;
2208 	if (unlikely(lastoff == 0)) {
2209 		XFS_ERROR_REPORT("xfs_da_swap_lastblock(1)", XFS_ERRLEVEL_LOW,
2210 				 mp);
2211 		return -EFSCORRUPTED;
2212 	}
2213 	/*
2214 	 * Read the last block in the btree space.
2215 	 */
2216 	last_blkno = (xfs_dablk_t)lastoff - args->geo->fsbcount;
2217 	error = xfs_da3_node_read(tp, dp, last_blkno, -1, &last_buf, w);
2218 	if (error)
2219 		return error;
2220 	/*
2221 	 * Copy the last block into the dead buffer and log it.
2222 	 */
2223 	memcpy(dead_buf->b_addr, last_buf->b_addr, args->geo->blksize);
2224 	xfs_trans_log_buf(tp, dead_buf, 0, args->geo->blksize - 1);
2225 	dead_info = dead_buf->b_addr;
2226 	/*
2227 	 * Get values from the moved block.
2228 	 */
2229 	if (dead_info->magic == cpu_to_be16(XFS_DIR2_LEAFN_MAGIC) ||
2230 	    dead_info->magic == cpu_to_be16(XFS_DIR3_LEAFN_MAGIC)) {
2231 		struct xfs_dir3_icleaf_hdr leafhdr;
2232 		struct xfs_dir2_leaf_entry *ents;
2233 
2234 		dead_leaf2 = (xfs_dir2_leaf_t *)dead_info;
2235 		dp->d_ops->leaf_hdr_from_disk(&leafhdr, dead_leaf2);
2236 		ents = dp->d_ops->leaf_ents_p(dead_leaf2);
2237 		dead_level = 0;
2238 		dead_hash = be32_to_cpu(ents[leafhdr.count - 1].hashval);
2239 	} else {
2240 		struct xfs_da3_icnode_hdr deadhdr;
2241 
2242 		dead_node = (xfs_da_intnode_t *)dead_info;
2243 		dp->d_ops->node_hdr_from_disk(&deadhdr, dead_node);
2244 		btree = dp->d_ops->node_tree_p(dead_node);
2245 		dead_level = deadhdr.level;
2246 		dead_hash = be32_to_cpu(btree[deadhdr.count - 1].hashval);
2247 	}
2248 	sib_buf = par_buf = NULL;
2249 	/*
2250 	 * If the moved block has a left sibling, fix up the pointers.
2251 	 */
2252 	if ((sib_blkno = be32_to_cpu(dead_info->back))) {
2253 		error = xfs_da3_node_read(tp, dp, sib_blkno, -1, &sib_buf, w);
2254 		if (error)
2255 			goto done;
2256 		sib_info = sib_buf->b_addr;
2257 		if (unlikely(
2258 		    be32_to_cpu(sib_info->forw) != last_blkno ||
2259 		    sib_info->magic != dead_info->magic)) {
2260 			XFS_ERROR_REPORT("xfs_da_swap_lastblock(2)",
2261 					 XFS_ERRLEVEL_LOW, mp);
2262 			error = -EFSCORRUPTED;
2263 			goto done;
2264 		}
2265 		sib_info->forw = cpu_to_be32(dead_blkno);
2266 		xfs_trans_log_buf(tp, sib_buf,
2267 			XFS_DA_LOGRANGE(sib_info, &sib_info->forw,
2268 					sizeof(sib_info->forw)));
2269 		sib_buf = NULL;
2270 	}
2271 	/*
2272 	 * If the moved block has a right sibling, fix up the pointers.
2273 	 */
2274 	if ((sib_blkno = be32_to_cpu(dead_info->forw))) {
2275 		error = xfs_da3_node_read(tp, dp, sib_blkno, -1, &sib_buf, w);
2276 		if (error)
2277 			goto done;
2278 		sib_info = sib_buf->b_addr;
2279 		if (unlikely(
2280 		       be32_to_cpu(sib_info->back) != last_blkno ||
2281 		       sib_info->magic != dead_info->magic)) {
2282 			XFS_ERROR_REPORT("xfs_da_swap_lastblock(3)",
2283 					 XFS_ERRLEVEL_LOW, mp);
2284 			error = -EFSCORRUPTED;
2285 			goto done;
2286 		}
2287 		sib_info->back = cpu_to_be32(dead_blkno);
2288 		xfs_trans_log_buf(tp, sib_buf,
2289 			XFS_DA_LOGRANGE(sib_info, &sib_info->back,
2290 					sizeof(sib_info->back)));
2291 		sib_buf = NULL;
2292 	}
2293 	par_blkno = args->geo->leafblk;
2294 	level = -1;
2295 	/*
2296 	 * Walk down the tree looking for the parent of the moved block.
2297 	 */
2298 	for (;;) {
2299 		error = xfs_da3_node_read(tp, dp, par_blkno, -1, &par_buf, w);
2300 		if (error)
2301 			goto done;
2302 		par_node = par_buf->b_addr;
2303 		dp->d_ops->node_hdr_from_disk(&par_hdr, par_node);
2304 		if (level >= 0 && level != par_hdr.level + 1) {
2305 			XFS_ERROR_REPORT("xfs_da_swap_lastblock(4)",
2306 					 XFS_ERRLEVEL_LOW, mp);
2307 			error = -EFSCORRUPTED;
2308 			goto done;
2309 		}
2310 		level = par_hdr.level;
2311 		btree = dp->d_ops->node_tree_p(par_node);
2312 		for (entno = 0;
2313 		     entno < par_hdr.count &&
2314 		     be32_to_cpu(btree[entno].hashval) < dead_hash;
2315 		     entno++)
2316 			continue;
2317 		if (entno == par_hdr.count) {
2318 			XFS_ERROR_REPORT("xfs_da_swap_lastblock(5)",
2319 					 XFS_ERRLEVEL_LOW, mp);
2320 			error = -EFSCORRUPTED;
2321 			goto done;
2322 		}
2323 		par_blkno = be32_to_cpu(btree[entno].before);
2324 		if (level == dead_level + 1)
2325 			break;
2326 		xfs_trans_brelse(tp, par_buf);
2327 		par_buf = NULL;
2328 	}
2329 	/*
2330 	 * We're in the right parent block.
2331 	 * Look for the right entry.
2332 	 */
2333 	for (;;) {
2334 		for (;
2335 		     entno < par_hdr.count &&
2336 		     be32_to_cpu(btree[entno].before) != last_blkno;
2337 		     entno++)
2338 			continue;
2339 		if (entno < par_hdr.count)
2340 			break;
2341 		par_blkno = par_hdr.forw;
2342 		xfs_trans_brelse(tp, par_buf);
2343 		par_buf = NULL;
2344 		if (unlikely(par_blkno == 0)) {
2345 			XFS_ERROR_REPORT("xfs_da_swap_lastblock(6)",
2346 					 XFS_ERRLEVEL_LOW, mp);
2347 			error = -EFSCORRUPTED;
2348 			goto done;
2349 		}
2350 		error = xfs_da3_node_read(tp, dp, par_blkno, -1, &par_buf, w);
2351 		if (error)
2352 			goto done;
2353 		par_node = par_buf->b_addr;
2354 		dp->d_ops->node_hdr_from_disk(&par_hdr, par_node);
2355 		if (par_hdr.level != level) {
2356 			XFS_ERROR_REPORT("xfs_da_swap_lastblock(7)",
2357 					 XFS_ERRLEVEL_LOW, mp);
2358 			error = -EFSCORRUPTED;
2359 			goto done;
2360 		}
2361 		btree = dp->d_ops->node_tree_p(par_node);
2362 		entno = 0;
2363 	}
2364 	/*
2365 	 * Update the parent entry pointing to the moved block.
2366 	 */
2367 	btree[entno].before = cpu_to_be32(dead_blkno);
2368 	xfs_trans_log_buf(tp, par_buf,
2369 		XFS_DA_LOGRANGE(par_node, &btree[entno].before,
2370 				sizeof(btree[entno].before)));
2371 	*dead_blknop = last_blkno;
2372 	*dead_bufp = last_buf;
2373 	return 0;
2374 done:
2375 	if (par_buf)
2376 		xfs_trans_brelse(tp, par_buf);
2377 	if (sib_buf)
2378 		xfs_trans_brelse(tp, sib_buf);
2379 	xfs_trans_brelse(tp, last_buf);
2380 	return error;
2381 }
2382 
2383 /*
2384  * Remove a btree block from a directory or attribute.
2385  */
2386 int
2387 xfs_da_shrink_inode(
2388 	xfs_da_args_t	*args,
2389 	xfs_dablk_t	dead_blkno,
2390 	struct xfs_buf	*dead_buf)
2391 {
2392 	xfs_inode_t *dp;
2393 	int done, error, w, count;
2394 	xfs_trans_t *tp;
2395 
2396 	trace_xfs_da_shrink_inode(args);
2397 
2398 	dp = args->dp;
2399 	w = args->whichfork;
2400 	tp = args->trans;
2401 	count = args->geo->fsbcount;
2402 	for (;;) {
2403 		/*
2404 		 * Remove extents.  If we get ENOSPC for a dir we have to move
2405 		 * the last block to the place we want to kill.
2406 		 */
2407 		error = xfs_bunmapi(tp, dp, dead_blkno, count,
2408 				    xfs_bmapi_aflag(w), 0, args->firstblock,
2409 				    args->dfops, &done);
2410 		if (error == -ENOSPC) {
2411 			if (w != XFS_DATA_FORK)
2412 				break;
2413 			error = xfs_da3_swap_lastblock(args, &dead_blkno,
2414 						      &dead_buf);
2415 			if (error)
2416 				break;
2417 		} else {
2418 			break;
2419 		}
2420 	}
2421 	xfs_trans_binval(tp, dead_buf);
2422 	return error;
2423 }
2424 
2425 /*
2426  * See if the mapping(s) for this btree block are valid, i.e.
2427  * don't contain holes, are logically contiguous, and cover the whole range.
2428  */
2429 STATIC int
2430 xfs_da_map_covers_blocks(
2431 	int		nmap,
2432 	xfs_bmbt_irec_t	*mapp,
2433 	xfs_dablk_t	bno,
2434 	int		count)
2435 {
2436 	int		i;
2437 	xfs_fileoff_t	off;
2438 
2439 	for (i = 0, off = bno; i < nmap; i++) {
2440 		if (mapp[i].br_startblock == HOLESTARTBLOCK ||
2441 		    mapp[i].br_startblock == DELAYSTARTBLOCK) {
2442 			return 0;
2443 		}
2444 		if (off != mapp[i].br_startoff) {
2445 			return 0;
2446 		}
2447 		off += mapp[i].br_blockcount;
2448 	}
2449 	return off == bno + count;
2450 }
2451 
2452 /*
2453  * Convert a struct xfs_bmbt_irec to a struct xfs_buf_map.
2454  *
2455  * For the single map case, it is assumed that the caller has provided a pointer
2456  * to a valid xfs_buf_map.  For the multiple map case, this function will
2457  * allocate the xfs_buf_map to hold all the maps and replace the caller's single
2458  * map pointer with the allocated map.
2459  */
2460 static int
2461 xfs_buf_map_from_irec(
2462 	struct xfs_mount	*mp,
2463 	struct xfs_buf_map	**mapp,
2464 	int			*nmaps,
2465 	struct xfs_bmbt_irec	*irecs,
2466 	int			nirecs)
2467 {
2468 	struct xfs_buf_map	*map;
2469 	int			i;
2470 
2471 	ASSERT(*nmaps == 1);
2472 	ASSERT(nirecs >= 1);
2473 
2474 	if (nirecs > 1) {
2475 		map = kmem_zalloc(nirecs * sizeof(struct xfs_buf_map),
2476 				  KM_SLEEP | KM_NOFS);
2477 		if (!map)
2478 			return -ENOMEM;
2479 		*mapp = map;
2480 	}
2481 
2482 	*nmaps = nirecs;
2483 	map = *mapp;
2484 	for (i = 0; i < *nmaps; i++) {
2485 		ASSERT(irecs[i].br_startblock != DELAYSTARTBLOCK &&
2486 		       irecs[i].br_startblock != HOLESTARTBLOCK);
2487 		map[i].bm_bn = XFS_FSB_TO_DADDR(mp, irecs[i].br_startblock);
2488 		map[i].bm_len = XFS_FSB_TO_BB(mp, irecs[i].br_blockcount);
2489 	}
2490 	return 0;
2491 }
2492 
2493 /*
2494  * Map the block we are given ready for reading. There are three possible return
2495  * values:
2496  *	-1 - will be returned if we land in a hole and mappedbno == -2 so the
2497  *	     caller knows not to execute a subsequent read.
2498  *	 0 - if we mapped the block successfully
2499  *	>0 - positive error number if there was an error.
2500  */
2501 static int
2502 xfs_dabuf_map(
2503 	struct xfs_inode	*dp,
2504 	xfs_dablk_t		bno,
2505 	xfs_daddr_t		mappedbno,
2506 	int			whichfork,
2507 	struct xfs_buf_map	**map,
2508 	int			*nmaps)
2509 {
2510 	struct xfs_mount	*mp = dp->i_mount;
2511 	int			nfsb;
2512 	int			error = 0;
2513 	struct xfs_bmbt_irec	irec;
2514 	struct xfs_bmbt_irec	*irecs = &irec;
2515 	int			nirecs;
2516 
2517 	ASSERT(map && *map);
2518 	ASSERT(*nmaps == 1);
2519 
2520 	if (whichfork == XFS_DATA_FORK)
2521 		nfsb = mp->m_dir_geo->fsbcount;
2522 	else
2523 		nfsb = mp->m_attr_geo->fsbcount;
2524 
2525 	/*
2526 	 * Caller doesn't have a mapping.  -2 means don't complain
2527 	 * if we land in a hole.
2528 	 */
2529 	if (mappedbno == -1 || mappedbno == -2) {
2530 		/*
2531 		 * Optimize the one-block case.
2532 		 */
2533 		if (nfsb != 1)
2534 			irecs = kmem_zalloc(sizeof(irec) * nfsb,
2535 					    KM_SLEEP | KM_NOFS);
2536 
2537 		nirecs = nfsb;
2538 		error = xfs_bmapi_read(dp, (xfs_fileoff_t)bno, nfsb, irecs,
2539 				       &nirecs, xfs_bmapi_aflag(whichfork));
2540 		if (error)
2541 			goto out;
2542 	} else {
2543 		irecs->br_startblock = XFS_DADDR_TO_FSB(mp, mappedbno);
2544 		irecs->br_startoff = (xfs_fileoff_t)bno;
2545 		irecs->br_blockcount = nfsb;
2546 		irecs->br_state = 0;
2547 		nirecs = 1;
2548 	}
2549 
2550 	if (!xfs_da_map_covers_blocks(nirecs, irecs, bno, nfsb)) {
2551 		error = mappedbno == -2 ? -1 : -EFSCORRUPTED;
2552 		if (unlikely(error == -EFSCORRUPTED)) {
2553 			if (xfs_error_level >= XFS_ERRLEVEL_LOW) {
2554 				int i;
2555 				xfs_alert(mp, "%s: bno %lld dir: inode %lld",
2556 					__func__, (long long)bno,
2557 					(long long)dp->i_ino);
2558 				for (i = 0; i < *nmaps; i++) {
2559 					xfs_alert(mp,
2560 "[%02d] br_startoff %lld br_startblock %lld br_blockcount %lld br_state %d",
2561 						i,
2562 						(long long)irecs[i].br_startoff,
2563 						(long long)irecs[i].br_startblock,
2564 						(long long)irecs[i].br_blockcount,
2565 						irecs[i].br_state);
2566 				}
2567 			}
2568 			XFS_ERROR_REPORT("xfs_da_do_buf(1)",
2569 					 XFS_ERRLEVEL_LOW, mp);
2570 		}
2571 		goto out;
2572 	}
2573 	error = xfs_buf_map_from_irec(mp, map, nmaps, irecs, nirecs);
2574 out:
2575 	if (irecs != &irec)
2576 		kmem_free(irecs);
2577 	return error;
2578 }
2579 
2580 /*
2581  * Get a buffer for the dir/attr block.
2582  */
2583 int
2584 xfs_da_get_buf(
2585 	struct xfs_trans	*trans,
2586 	struct xfs_inode	*dp,
2587 	xfs_dablk_t		bno,
2588 	xfs_daddr_t		mappedbno,
2589 	struct xfs_buf		**bpp,
2590 	int			whichfork)
2591 {
2592 	struct xfs_buf		*bp;
2593 	struct xfs_buf_map	map;
2594 	struct xfs_buf_map	*mapp;
2595 	int			nmap;
2596 	int			error;
2597 
2598 	*bpp = NULL;
2599 	mapp = &map;
2600 	nmap = 1;
2601 	error = xfs_dabuf_map(dp, bno, mappedbno, whichfork,
2602 				&mapp, &nmap);
2603 	if (error) {
2604 		/* mapping a hole is not an error, but we don't continue */
2605 		if (error == -1)
2606 			error = 0;
2607 		goto out_free;
2608 	}
2609 
2610 	bp = xfs_trans_get_buf_map(trans, dp->i_mount->m_ddev_targp,
2611 				    mapp, nmap, 0);
2612 	error = bp ? bp->b_error : -EIO;
2613 	if (error) {
2614 		if (bp)
2615 			xfs_trans_brelse(trans, bp);
2616 		goto out_free;
2617 	}
2618 
2619 	*bpp = bp;
2620 
2621 out_free:
2622 	if (mapp != &map)
2623 		kmem_free(mapp);
2624 
2625 	return error;
2626 }
2627 
2628 /*
2629  * Get a buffer for the dir/attr block, fill in the contents.
2630  */
2631 int
2632 xfs_da_read_buf(
2633 	struct xfs_trans	*trans,
2634 	struct xfs_inode	*dp,
2635 	xfs_dablk_t		bno,
2636 	xfs_daddr_t		mappedbno,
2637 	struct xfs_buf		**bpp,
2638 	int			whichfork,
2639 	const struct xfs_buf_ops *ops)
2640 {
2641 	struct xfs_buf		*bp;
2642 	struct xfs_buf_map	map;
2643 	struct xfs_buf_map	*mapp;
2644 	int			nmap;
2645 	int			error;
2646 
2647 	*bpp = NULL;
2648 	mapp = &map;
2649 	nmap = 1;
2650 	error = xfs_dabuf_map(dp, bno, mappedbno, whichfork,
2651 				&mapp, &nmap);
2652 	if (error) {
2653 		/* mapping a hole is not an error, but we don't continue */
2654 		if (error == -1)
2655 			error = 0;
2656 		goto out_free;
2657 	}
2658 
2659 	error = xfs_trans_read_buf_map(dp->i_mount, trans,
2660 					dp->i_mount->m_ddev_targp,
2661 					mapp, nmap, 0, &bp, ops);
2662 	if (error)
2663 		goto out_free;
2664 
2665 	if (whichfork == XFS_ATTR_FORK)
2666 		xfs_buf_set_ref(bp, XFS_ATTR_BTREE_REF);
2667 	else
2668 		xfs_buf_set_ref(bp, XFS_DIR_BTREE_REF);
2669 	*bpp = bp;
2670 out_free:
2671 	if (mapp != &map)
2672 		kmem_free(mapp);
2673 
2674 	return error;
2675 }
2676 
2677 /*
2678  * Readahead the dir/attr block.
2679  */
2680 int
2681 xfs_da_reada_buf(
2682 	struct xfs_inode	*dp,
2683 	xfs_dablk_t		bno,
2684 	xfs_daddr_t		mappedbno,
2685 	int			whichfork,
2686 	const struct xfs_buf_ops *ops)
2687 {
2688 	struct xfs_buf_map	map;
2689 	struct xfs_buf_map	*mapp;
2690 	int			nmap;
2691 	int			error;
2692 
2693 	mapp = &map;
2694 	nmap = 1;
2695 	error = xfs_dabuf_map(dp, bno, mappedbno, whichfork,
2696 				&mapp, &nmap);
2697 	if (error) {
2698 		/* mapping a hole is not an error, but we don't continue */
2699 		if (error == -1)
2700 			error = 0;
2701 		goto out_free;
2702 	}
2703 
2704 	mappedbno = mapp[0].bm_bn;
2705 	xfs_buf_readahead_map(dp->i_mount->m_ddev_targp, mapp, nmap, ops);
2706 
2707 out_free:
2708 	if (mapp != &map)
2709 		kmem_free(mapp);
2710 
2711 	return error;
2712 }
2713