Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/ext4/indirect.c
4 *
5 * from
6 *
7 * linux/fs/ext4/inode.c
8 *
9 * Copyright (C) 1992, 1993, 1994, 1995
10 * Remy Card (card@masi.ibp.fr)
11 * Laboratoire MASI - Institut Blaise Pascal
12 * Universite Pierre et Marie Curie (Paris VI)
13 *
14 * from
15 *
16 * linux/fs/minix/inode.c
17 *
18 * Copyright (C) 1991, 1992 Linus Torvalds
19 *
20 * Goal-directed block allocation by Stephen Tweedie
21 * (sct@redhat.com), 1993, 1998
22 */
23
24#include "ext4_jbd2.h"
25#include "truncate.h"
26#include <linux/dax.h>
27#include <linux/uio.h>
28
29#include <trace/events/ext4.h>
30
31typedef struct {
32 __le32 *p;
33 __le32 key;
34 struct buffer_head *bh;
35} Indirect;
36
37static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
38{
39 p->key = *(p->p = v);
40 p->bh = bh;
41}
42
43/**
44 * ext4_block_to_path - parse the block number into array of offsets
45 * @inode: inode in question (we are only interested in its superblock)
46 * @i_block: block number to be parsed
47 * @offsets: array to store the offsets in
48 * @boundary: set this non-zero if the referred-to block is likely to be
49 * followed (on disk) by an indirect block.
50 *
51 * To store the locations of file's data ext4 uses a data structure common
52 * for UNIX filesystems - tree of pointers anchored in the inode, with
53 * data blocks at leaves and indirect blocks in intermediate nodes.
54 * This function translates the block number into path in that tree -
55 * return value is the path length and @offsets[n] is the offset of
56 * pointer to (n+1)th node in the nth one. If @block is out of range
57 * (negative or too large) warning is printed and zero returned.
58 *
59 * Note: function doesn't find node addresses, so no IO is needed. All
60 * we need to know is the capacity of indirect blocks (taken from the
61 * inode->i_sb).
62 */
63
64/*
65 * Portability note: the last comparison (check that we fit into triple
66 * indirect block) is spelled differently, because otherwise on an
67 * architecture with 32-bit longs and 8Kb pages we might get into trouble
68 * if our filesystem had 8Kb blocks. We might use long long, but that would
69 * kill us on x86. Oh, well, at least the sign propagation does not matter -
70 * i_block would have to be negative in the very beginning, so we would not
71 * get there at all.
72 */
73
74static int ext4_block_to_path(struct inode *inode,
75 ext4_lblk_t i_block,
76 ext4_lblk_t offsets[4], int *boundary)
77{
78 int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
79 int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
80 const long direct_blocks = EXT4_NDIR_BLOCKS,
81 indirect_blocks = ptrs,
82 double_blocks = (1 << (ptrs_bits * 2));
83 int n = 0;
84 int final = 0;
85
86 if (i_block < direct_blocks) {
87 offsets[n++] = i_block;
88 final = direct_blocks;
89 } else if ((i_block -= direct_blocks) < indirect_blocks) {
90 offsets[n++] = EXT4_IND_BLOCK;
91 offsets[n++] = i_block;
92 final = ptrs;
93 } else if ((i_block -= indirect_blocks) < double_blocks) {
94 offsets[n++] = EXT4_DIND_BLOCK;
95 offsets[n++] = i_block >> ptrs_bits;
96 offsets[n++] = i_block & (ptrs - 1);
97 final = ptrs;
98 } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
99 offsets[n++] = EXT4_TIND_BLOCK;
100 offsets[n++] = i_block >> (ptrs_bits * 2);
101 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
102 offsets[n++] = i_block & (ptrs - 1);
103 final = ptrs;
104 } else {
105 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
106 i_block + direct_blocks +
107 indirect_blocks + double_blocks, inode->i_ino);
108 }
109 if (boundary)
110 *boundary = final - 1 - (i_block & (ptrs - 1));
111 return n;
112}
113
114/**
115 * ext4_get_branch - read the chain of indirect blocks leading to data
116 * @inode: inode in question
117 * @depth: depth of the chain (1 - direct pointer, etc.)
118 * @offsets: offsets of pointers in inode/indirect blocks
119 * @chain: place to store the result
120 * @err: here we store the error value
121 *
122 * Function fills the array of triples <key, p, bh> and returns %NULL
123 * if everything went OK or the pointer to the last filled triple
124 * (incomplete one) otherwise. Upon the return chain[i].key contains
125 * the number of (i+1)-th block in the chain (as it is stored in memory,
126 * i.e. little-endian 32-bit), chain[i].p contains the address of that
127 * number (it points into struct inode for i==0 and into the bh->b_data
128 * for i>0) and chain[i].bh points to the buffer_head of i-th indirect
129 * block for i>0 and NULL for i==0. In other words, it holds the block
130 * numbers of the chain, addresses they were taken from (and where we can
131 * verify that chain did not change) and buffer_heads hosting these
132 * numbers.
133 *
134 * Function stops when it stumbles upon zero pointer (absent block)
135 * (pointer to last triple returned, *@err == 0)
136 * or when it gets an IO error reading an indirect block
137 * (ditto, *@err == -EIO)
138 * or when it reads all @depth-1 indirect blocks successfully and finds
139 * the whole chain, all way to the data (returns %NULL, *err == 0).
140 *
141 * Need to be called with
142 * down_read(&EXT4_I(inode)->i_data_sem)
143 */
144static Indirect *ext4_get_branch(struct inode *inode, int depth,
145 ext4_lblk_t *offsets,
146 Indirect chain[4], int *err)
147{
148 struct super_block *sb = inode->i_sb;
149 Indirect *p = chain;
150 struct buffer_head *bh;
151 int ret = -EIO;
152
153 *err = 0;
154 /* i_data is not going away, no lock needed */
155 add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
156 if (!p->key)
157 goto no_block;
158 while (--depth) {
159 bh = sb_getblk(sb, le32_to_cpu(p->key));
160 if (unlikely(!bh)) {
161 ret = -ENOMEM;
162 goto failure;
163 }
164
165 if (!bh_uptodate_or_lock(bh)) {
166 if (bh_submit_read(bh) < 0) {
167 put_bh(bh);
168 goto failure;
169 }
170 /* validate block references */
171 if (ext4_check_indirect_blockref(inode, bh)) {
172 put_bh(bh);
173 goto failure;
174 }
175 }
176
177 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
178 /* Reader: end */
179 if (!p->key)
180 goto no_block;
181 }
182 return NULL;
183
184failure:
185 *err = ret;
186no_block:
187 return p;
188}
189
190/**
191 * ext4_find_near - find a place for allocation with sufficient locality
192 * @inode: owner
193 * @ind: descriptor of indirect block.
194 *
195 * This function returns the preferred place for block allocation.
196 * It is used when heuristic for sequential allocation fails.
197 * Rules are:
198 * + if there is a block to the left of our position - allocate near it.
199 * + if pointer will live in indirect block - allocate near that block.
200 * + if pointer will live in inode - allocate in the same
201 * cylinder group.
202 *
203 * In the latter case we colour the starting block by the callers PID to
204 * prevent it from clashing with concurrent allocations for a different inode
205 * in the same block group. The PID is used here so that functionally related
206 * files will be close-by on-disk.
207 *
208 * Caller must make sure that @ind is valid and will stay that way.
209 */
210static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
211{
212 struct ext4_inode_info *ei = EXT4_I(inode);
213 __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
214 __le32 *p;
215
216 /* Try to find previous block */
217 for (p = ind->p - 1; p >= start; p--) {
218 if (*p)
219 return le32_to_cpu(*p);
220 }
221
222 /* No such thing, so let's try location of indirect block */
223 if (ind->bh)
224 return ind->bh->b_blocknr;
225
226 /*
227 * It is going to be referred to from the inode itself? OK, just put it
228 * into the same cylinder group then.
229 */
230 return ext4_inode_to_goal_block(inode);
231}
232
233/**
234 * ext4_find_goal - find a preferred place for allocation.
235 * @inode: owner
236 * @block: block we want
237 * @partial: pointer to the last triple within a chain
238 *
239 * Normally this function find the preferred place for block allocation,
240 * returns it.
241 * Because this is only used for non-extent files, we limit the block nr
242 * to 32 bits.
243 */
244static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
245 Indirect *partial)
246{
247 ext4_fsblk_t goal;
248
249 /*
250 * XXX need to get goal block from mballoc's data structures
251 */
252
253 goal = ext4_find_near(inode, partial);
254 goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
255 return goal;
256}
257
258/**
259 * ext4_blks_to_allocate - Look up the block map and count the number
260 * of direct blocks need to be allocated for the given branch.
261 *
262 * @branch: chain of indirect blocks
263 * @k: number of blocks need for indirect blocks
264 * @blks: number of data blocks to be mapped.
265 * @blocks_to_boundary: the offset in the indirect block
266 *
267 * return the total number of blocks to be allocate, including the
268 * direct and indirect blocks.
269 */
270static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
271 int blocks_to_boundary)
272{
273 unsigned int count = 0;
274
275 /*
276 * Simple case, [t,d]Indirect block(s) has not allocated yet
277 * then it's clear blocks on that path have not allocated
278 */
279 if (k > 0) {
280 /* right now we don't handle cross boundary allocation */
281 if (blks < blocks_to_boundary + 1)
282 count += blks;
283 else
284 count += blocks_to_boundary + 1;
285 return count;
286 }
287
288 count++;
289 while (count < blks && count <= blocks_to_boundary &&
290 le32_to_cpu(*(branch[0].p + count)) == 0) {
291 count++;
292 }
293 return count;
294}
295
296/**
297 * ext4_alloc_branch - allocate and set up a chain of blocks.
298 * @handle: handle for this transaction
299 * @inode: owner
300 * @indirect_blks: number of allocated indirect blocks
301 * @blks: number of allocated direct blocks
302 * @goal: preferred place for allocation
303 * @offsets: offsets (in the blocks) to store the pointers to next.
304 * @branch: place to store the chain in.
305 *
306 * This function allocates blocks, zeroes out all but the last one,
307 * links them into chain and (if we are synchronous) writes them to disk.
308 * In other words, it prepares a branch that can be spliced onto the
309 * inode. It stores the information about that chain in the branch[], in
310 * the same format as ext4_get_branch() would do. We are calling it after
311 * we had read the existing part of chain and partial points to the last
312 * triple of that (one with zero ->key). Upon the exit we have the same
313 * picture as after the successful ext4_get_block(), except that in one
314 * place chain is disconnected - *branch->p is still zero (we did not
315 * set the last link), but branch->key contains the number that should
316 * be placed into *branch->p to fill that gap.
317 *
318 * If allocation fails we free all blocks we've allocated (and forget
319 * their buffer_heads) and return the error value the from failed
320 * ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
321 * as described above and return 0.
322 */
323static int ext4_alloc_branch(handle_t *handle,
324 struct ext4_allocation_request *ar,
325 int indirect_blks, ext4_lblk_t *offsets,
326 Indirect *branch)
327{
328 struct buffer_head * bh;
329 ext4_fsblk_t b, new_blocks[4];
330 __le32 *p;
331 int i, j, err, len = 1;
332
333 for (i = 0; i <= indirect_blks; i++) {
334 if (i == indirect_blks) {
335 new_blocks[i] = ext4_mb_new_blocks(handle, ar, &err);
336 } else
337 ar->goal = new_blocks[i] = ext4_new_meta_blocks(handle,
338 ar->inode, ar->goal,
339 ar->flags & EXT4_MB_DELALLOC_RESERVED,
340 NULL, &err);
341 if (err) {
342 i--;
343 goto failed;
344 }
345 branch[i].key = cpu_to_le32(new_blocks[i]);
346 if (i == 0)
347 continue;
348
349 bh = branch[i].bh = sb_getblk(ar->inode->i_sb, new_blocks[i-1]);
350 if (unlikely(!bh)) {
351 err = -ENOMEM;
352 goto failed;
353 }
354 lock_buffer(bh);
355 BUFFER_TRACE(bh, "call get_create_access");
356 err = ext4_journal_get_create_access(handle, bh);
357 if (err) {
358 unlock_buffer(bh);
359 goto failed;
360 }
361
362 memset(bh->b_data, 0, bh->b_size);
363 p = branch[i].p = (__le32 *) bh->b_data + offsets[i];
364 b = new_blocks[i];
365
366 if (i == indirect_blks)
367 len = ar->len;
368 for (j = 0; j < len; j++)
369 *p++ = cpu_to_le32(b++);
370
371 BUFFER_TRACE(bh, "marking uptodate");
372 set_buffer_uptodate(bh);
373 unlock_buffer(bh);
374
375 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
376 err = ext4_handle_dirty_metadata(handle, ar->inode, bh);
377 if (err)
378 goto failed;
379 }
380 return 0;
381failed:
382 for (; i >= 0; i--) {
383 /*
384 * We want to ext4_forget() only freshly allocated indirect
385 * blocks. Buffer for new_blocks[i-1] is at branch[i].bh and
386 * buffer at branch[0].bh is indirect block / inode already
387 * existing before ext4_alloc_branch() was called.
388 */
389 if (i > 0 && i != indirect_blks && branch[i].bh)
390 ext4_forget(handle, 1, ar->inode, branch[i].bh,
391 branch[i].bh->b_blocknr);
392 ext4_free_blocks(handle, ar->inode, NULL, new_blocks[i],
393 (i == indirect_blks) ? ar->len : 1, 0);
394 }
395 return err;
396}
397
398/**
399 * ext4_splice_branch - splice the allocated branch onto inode.
400 * @handle: handle for this transaction
401 * @inode: owner
402 * @block: (logical) number of block we are adding
403 * @chain: chain of indirect blocks (with a missing link - see
404 * ext4_alloc_branch)
405 * @where: location of missing link
406 * @num: number of indirect blocks we are adding
407 * @blks: number of direct blocks we are adding
408 *
409 * This function fills the missing link and does all housekeeping needed in
410 * inode (->i_blocks, etc.). In case of success we end up with the full
411 * chain to new block and return 0.
412 */
413static int ext4_splice_branch(handle_t *handle,
414 struct ext4_allocation_request *ar,
415 Indirect *where, int num)
416{
417 int i;
418 int err = 0;
419 ext4_fsblk_t current_block;
420
421 /*
422 * If we're splicing into a [td]indirect block (as opposed to the
423 * inode) then we need to get write access to the [td]indirect block
424 * before the splice.
425 */
426 if (where->bh) {
427 BUFFER_TRACE(where->bh, "get_write_access");
428 err = ext4_journal_get_write_access(handle, where->bh);
429 if (err)
430 goto err_out;
431 }
432 /* That's it */
433
434 *where->p = where->key;
435
436 /*
437 * Update the host buffer_head or inode to point to more just allocated
438 * direct blocks blocks
439 */
440 if (num == 0 && ar->len > 1) {
441 current_block = le32_to_cpu(where->key) + 1;
442 for (i = 1; i < ar->len; i++)
443 *(where->p + i) = cpu_to_le32(current_block++);
444 }
445
446 /* We are done with atomic stuff, now do the rest of housekeeping */
447 /* had we spliced it onto indirect block? */
448 if (where->bh) {
449 /*
450 * If we spliced it onto an indirect block, we haven't
451 * altered the inode. Note however that if it is being spliced
452 * onto an indirect block at the very end of the file (the
453 * file is growing) then we *will* alter the inode to reflect
454 * the new i_size. But that is not done here - it is done in
455 * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
456 */
457 jbd_debug(5, "splicing indirect only\n");
458 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
459 err = ext4_handle_dirty_metadata(handle, ar->inode, where->bh);
460 if (err)
461 goto err_out;
462 } else {
463 /*
464 * OK, we spliced it into the inode itself on a direct block.
465 */
466 ext4_mark_inode_dirty(handle, ar->inode);
467 jbd_debug(5, "splicing direct\n");
468 }
469 return err;
470
471err_out:
472 for (i = 1; i <= num; i++) {
473 /*
474 * branch[i].bh is newly allocated, so there is no
475 * need to revoke the block, which is why we don't
476 * need to set EXT4_FREE_BLOCKS_METADATA.
477 */
478 ext4_free_blocks(handle, ar->inode, where[i].bh, 0, 1,
479 EXT4_FREE_BLOCKS_FORGET);
480 }
481 ext4_free_blocks(handle, ar->inode, NULL, le32_to_cpu(where[num].key),
482 ar->len, 0);
483
484 return err;
485}
486
487/*
488 * The ext4_ind_map_blocks() function handles non-extents inodes
489 * (i.e., using the traditional indirect/double-indirect i_blocks
490 * scheme) for ext4_map_blocks().
491 *
492 * Allocation strategy is simple: if we have to allocate something, we will
493 * have to go the whole way to leaf. So let's do it before attaching anything
494 * to tree, set linkage between the newborn blocks, write them if sync is
495 * required, recheck the path, free and repeat if check fails, otherwise
496 * set the last missing link (that will protect us from any truncate-generated
497 * removals - all blocks on the path are immune now) and possibly force the
498 * write on the parent block.
499 * That has a nice additional property: no special recovery from the failed
500 * allocations is needed - we simply release blocks and do not touch anything
501 * reachable from inode.
502 *
503 * `handle' can be NULL if create == 0.
504 *
505 * return > 0, # of blocks mapped or allocated.
506 * return = 0, if plain lookup failed.
507 * return < 0, error case.
508 *
509 * The ext4_ind_get_blocks() function should be called with
510 * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
511 * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
512 * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
513 * blocks.
514 */
515int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
516 struct ext4_map_blocks *map,
517 int flags)
518{
519 struct ext4_allocation_request ar;
520 int err = -EIO;
521 ext4_lblk_t offsets[4];
522 Indirect chain[4];
523 Indirect *partial;
524 int indirect_blks;
525 int blocks_to_boundary = 0;
526 int depth;
527 int count = 0;
528 ext4_fsblk_t first_block = 0;
529
530 trace_ext4_ind_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
531 J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
532 J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
533 depth = ext4_block_to_path(inode, map->m_lblk, offsets,
534 &blocks_to_boundary);
535
536 if (depth == 0)
537 goto out;
538
539 partial = ext4_get_branch(inode, depth, offsets, chain, &err);
540
541 /* Simplest case - block found, no allocation needed */
542 if (!partial) {
543 first_block = le32_to_cpu(chain[depth - 1].key);
544 count++;
545 /*map more blocks*/
546 while (count < map->m_len && count <= blocks_to_boundary) {
547 ext4_fsblk_t blk;
548
549 blk = le32_to_cpu(*(chain[depth-1].p + count));
550
551 if (blk == first_block + count)
552 count++;
553 else
554 break;
555 }
556 goto got_it;
557 }
558
559 /* Next simple case - plain lookup failed */
560 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
561 unsigned epb = inode->i_sb->s_blocksize / sizeof(u32);
562 int i;
563
564 /* Count number blocks in a subtree under 'partial' */
565 count = 1;
566 for (i = 0; partial + i != chain + depth - 1; i++)
567 count *= epb;
568 /* Fill in size of a hole we found */
569 map->m_pblk = 0;
570 map->m_len = min_t(unsigned int, map->m_len, count);
571 goto cleanup;
572 }
573
574 /* Failed read of indirect block */
575 if (err == -EIO)
576 goto cleanup;
577
578 /*
579 * Okay, we need to do block allocation.
580 */
581 if (ext4_has_feature_bigalloc(inode->i_sb)) {
582 EXT4_ERROR_INODE(inode, "Can't allocate blocks for "
583 "non-extent mapped inodes with bigalloc");
584 return -EFSCORRUPTED;
585 }
586
587 /* Set up for the direct block allocation */
588 memset(&ar, 0, sizeof(ar));
589 ar.inode = inode;
590 ar.logical = map->m_lblk;
591 if (S_ISREG(inode->i_mode))
592 ar.flags = EXT4_MB_HINT_DATA;
593 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
594 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
595 if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
596 ar.flags |= EXT4_MB_USE_RESERVED;
597
598 ar.goal = ext4_find_goal(inode, map->m_lblk, partial);
599
600 /* the number of blocks need to allocate for [d,t]indirect blocks */
601 indirect_blks = (chain + depth) - partial - 1;
602
603 /*
604 * Next look up the indirect map to count the totoal number of
605 * direct blocks to allocate for this branch.
606 */
607 ar.len = ext4_blks_to_allocate(partial, indirect_blks,
608 map->m_len, blocks_to_boundary);
609
610 /*
611 * Block out ext4_truncate while we alter the tree
612 */
613 err = ext4_alloc_branch(handle, &ar, indirect_blks,
614 offsets + (partial - chain), partial);
615
616 /*
617 * The ext4_splice_branch call will free and forget any buffers
618 * on the new chain if there is a failure, but that risks using
619 * up transaction credits, especially for bitmaps where the
620 * credits cannot be returned. Can we handle this somehow? We
621 * may need to return -EAGAIN upwards in the worst case. --sct
622 */
623 if (!err)
624 err = ext4_splice_branch(handle, &ar, partial, indirect_blks);
625 if (err)
626 goto cleanup;
627
628 map->m_flags |= EXT4_MAP_NEW;
629
630 ext4_update_inode_fsync_trans(handle, inode, 1);
631 count = ar.len;
632got_it:
633 map->m_flags |= EXT4_MAP_MAPPED;
634 map->m_pblk = le32_to_cpu(chain[depth-1].key);
635 map->m_len = count;
636 if (count > blocks_to_boundary)
637 map->m_flags |= EXT4_MAP_BOUNDARY;
638 err = count;
639 /* Clean up and exit */
640 partial = chain + depth - 1; /* the whole chain */
641cleanup:
642 while (partial > chain) {
643 BUFFER_TRACE(partial->bh, "call brelse");
644 brelse(partial->bh);
645 partial--;
646 }
647out:
648 trace_ext4_ind_map_blocks_exit(inode, flags, map, err);
649 return err;
650}
651
652/*
653 * Calculate the number of metadata blocks need to reserve
654 * to allocate a new block at @lblocks for non extent file based file
655 */
656int ext4_ind_calc_metadata_amount(struct inode *inode, sector_t lblock)
657{
658 struct ext4_inode_info *ei = EXT4_I(inode);
659 sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
660 int blk_bits;
661
662 if (lblock < EXT4_NDIR_BLOCKS)
663 return 0;
664
665 lblock -= EXT4_NDIR_BLOCKS;
666
667 if (ei->i_da_metadata_calc_len &&
668 (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
669 ei->i_da_metadata_calc_len++;
670 return 0;
671 }
672 ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
673 ei->i_da_metadata_calc_len = 1;
674 blk_bits = order_base_2(lblock);
675 return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
676}
677
678/*
679 * Calculate number of indirect blocks touched by mapping @nrblocks logically
680 * contiguous blocks
681 */
682int ext4_ind_trans_blocks(struct inode *inode, int nrblocks)
683{
684 /*
685 * With N contiguous data blocks, we need at most
686 * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
687 * 2 dindirect blocks, and 1 tindirect block
688 */
689 return DIV_ROUND_UP(nrblocks, EXT4_ADDR_PER_BLOCK(inode->i_sb)) + 4;
690}
691
692/*
693 * Truncate transactions can be complex and absolutely huge. So we need to
694 * be able to restart the transaction at a conventient checkpoint to make
695 * sure we don't overflow the journal.
696 *
697 * Try to extend this transaction for the purposes of truncation. If
698 * extend fails, we need to propagate the failure up and restart the
699 * transaction in the top-level truncate loop. --sct
700 *
701 * Returns 0 if we managed to create more room. If we can't create more
702 * room, and the transaction must be restarted we return 1.
703 */
704static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
705{
706 if (!ext4_handle_valid(handle))
707 return 0;
708 if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
709 return 0;
710 if (!ext4_journal_extend(handle, ext4_blocks_for_truncate(inode)))
711 return 0;
712 return 1;
713}
714
715/*
716 * Probably it should be a library function... search for first non-zero word
717 * or memcmp with zero_page, whatever is better for particular architecture.
718 * Linus?
719 */
720static inline int all_zeroes(__le32 *p, __le32 *q)
721{
722 while (p < q)
723 if (*p++)
724 return 0;
725 return 1;
726}
727
728/**
729 * ext4_find_shared - find the indirect blocks for partial truncation.
730 * @inode: inode in question
731 * @depth: depth of the affected branch
732 * @offsets: offsets of pointers in that branch (see ext4_block_to_path)
733 * @chain: place to store the pointers to partial indirect blocks
734 * @top: place to the (detached) top of branch
735 *
736 * This is a helper function used by ext4_truncate().
737 *
738 * When we do truncate() we may have to clean the ends of several
739 * indirect blocks but leave the blocks themselves alive. Block is
740 * partially truncated if some data below the new i_size is referred
741 * from it (and it is on the path to the first completely truncated
742 * data block, indeed). We have to free the top of that path along
743 * with everything to the right of the path. Since no allocation
744 * past the truncation point is possible until ext4_truncate()
745 * finishes, we may safely do the latter, but top of branch may
746 * require special attention - pageout below the truncation point
747 * might try to populate it.
748 *
749 * We atomically detach the top of branch from the tree, store the
750 * block number of its root in *@top, pointers to buffer_heads of
751 * partially truncated blocks - in @chain[].bh and pointers to
752 * their last elements that should not be removed - in
753 * @chain[].p. Return value is the pointer to last filled element
754 * of @chain.
755 *
756 * The work left to caller to do the actual freeing of subtrees:
757 * a) free the subtree starting from *@top
758 * b) free the subtrees whose roots are stored in
759 * (@chain[i].p+1 .. end of @chain[i].bh->b_data)
760 * c) free the subtrees growing from the inode past the @chain[0].
761 * (no partially truncated stuff there). */
762
763static Indirect *ext4_find_shared(struct inode *inode, int depth,
764 ext4_lblk_t offsets[4], Indirect chain[4],
765 __le32 *top)
766{
767 Indirect *partial, *p;
768 int k, err;
769
770 *top = 0;
771 /* Make k index the deepest non-null offset + 1 */
772 for (k = depth; k > 1 && !offsets[k-1]; k--)
773 ;
774 partial = ext4_get_branch(inode, k, offsets, chain, &err);
775 /* Writer: pointers */
776 if (!partial)
777 partial = chain + k-1;
778 /*
779 * If the branch acquired continuation since we've looked at it -
780 * fine, it should all survive and (new) top doesn't belong to us.
781 */
782 if (!partial->key && *partial->p)
783 /* Writer: end */
784 goto no_top;
785 for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
786 ;
787 /*
788 * OK, we've found the last block that must survive. The rest of our
789 * branch should be detached before unlocking. However, if that rest
790 * of branch is all ours and does not grow immediately from the inode
791 * it's easier to cheat and just decrement partial->p.
792 */
793 if (p == chain + k - 1 && p > chain) {
794 p->p--;
795 } else {
796 *top = *p->p;
797 /* Nope, don't do this in ext4. Must leave the tree intact */
798#if 0
799 *p->p = 0;
800#endif
801 }
802 /* Writer: end */
803
804 while (partial > p) {
805 brelse(partial->bh);
806 partial--;
807 }
808no_top:
809 return partial;
810}
811
812/*
813 * Zero a number of block pointers in either an inode or an indirect block.
814 * If we restart the transaction we must again get write access to the
815 * indirect block for further modification.
816 *
817 * We release `count' blocks on disk, but (last - first) may be greater
818 * than `count' because there can be holes in there.
819 *
820 * Return 0 on success, 1 on invalid block range
821 * and < 0 on fatal error.
822 */
823static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
824 struct buffer_head *bh,
825 ext4_fsblk_t block_to_free,
826 unsigned long count, __le32 *first,
827 __le32 *last)
828{
829 __le32 *p;
830 int flags = EXT4_FREE_BLOCKS_VALIDATED;
831 int err;
832
833 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
834 ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
835 flags |= EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_METADATA;
836 else if (ext4_should_journal_data(inode))
837 flags |= EXT4_FREE_BLOCKS_FORGET;
838
839 if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
840 count)) {
841 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
842 "blocks %llu len %lu",
843 (unsigned long long) block_to_free, count);
844 return 1;
845 }
846
847 if (try_to_extend_transaction(handle, inode)) {
848 if (bh) {
849 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
850 err = ext4_handle_dirty_metadata(handle, inode, bh);
851 if (unlikely(err))
852 goto out_err;
853 }
854 err = ext4_mark_inode_dirty(handle, inode);
855 if (unlikely(err))
856 goto out_err;
857 err = ext4_truncate_restart_trans(handle, inode,
858 ext4_blocks_for_truncate(inode));
859 if (unlikely(err))
860 goto out_err;
861 if (bh) {
862 BUFFER_TRACE(bh, "retaking write access");
863 err = ext4_journal_get_write_access(handle, bh);
864 if (unlikely(err))
865 goto out_err;
866 }
867 }
868
869 for (p = first; p < last; p++)
870 *p = 0;
871
872 ext4_free_blocks(handle, inode, NULL, block_to_free, count, flags);
873 return 0;
874out_err:
875 ext4_std_error(inode->i_sb, err);
876 return err;
877}
878
879/**
880 * ext4_free_data - free a list of data blocks
881 * @handle: handle for this transaction
882 * @inode: inode we are dealing with
883 * @this_bh: indirect buffer_head which contains *@first and *@last
884 * @first: array of block numbers
885 * @last: points immediately past the end of array
886 *
887 * We are freeing all blocks referred from that array (numbers are stored as
888 * little-endian 32-bit) and updating @inode->i_blocks appropriately.
889 *
890 * We accumulate contiguous runs of blocks to free. Conveniently, if these
891 * blocks are contiguous then releasing them at one time will only affect one
892 * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
893 * actually use a lot of journal space.
894 *
895 * @this_bh will be %NULL if @first and @last point into the inode's direct
896 * block pointers.
897 */
898static void ext4_free_data(handle_t *handle, struct inode *inode,
899 struct buffer_head *this_bh,
900 __le32 *first, __le32 *last)
901{
902 ext4_fsblk_t block_to_free = 0; /* Starting block # of a run */
903 unsigned long count = 0; /* Number of blocks in the run */
904 __le32 *block_to_free_p = NULL; /* Pointer into inode/ind
905 corresponding to
906 block_to_free */
907 ext4_fsblk_t nr; /* Current block # */
908 __le32 *p; /* Pointer into inode/ind
909 for current block */
910 int err = 0;
911
912 if (this_bh) { /* For indirect block */
913 BUFFER_TRACE(this_bh, "get_write_access");
914 err = ext4_journal_get_write_access(handle, this_bh);
915 /* Important: if we can't update the indirect pointers
916 * to the blocks, we can't free them. */
917 if (err)
918 return;
919 }
920
921 for (p = first; p < last; p++) {
922 nr = le32_to_cpu(*p);
923 if (nr) {
924 /* accumulate blocks to free if they're contiguous */
925 if (count == 0) {
926 block_to_free = nr;
927 block_to_free_p = p;
928 count = 1;
929 } else if (nr == block_to_free + count) {
930 count++;
931 } else {
932 err = ext4_clear_blocks(handle, inode, this_bh,
933 block_to_free, count,
934 block_to_free_p, p);
935 if (err)
936 break;
937 block_to_free = nr;
938 block_to_free_p = p;
939 count = 1;
940 }
941 }
942 }
943
944 if (!err && count > 0)
945 err = ext4_clear_blocks(handle, inode, this_bh, block_to_free,
946 count, block_to_free_p, p);
947 if (err < 0)
948 /* fatal error */
949 return;
950
951 if (this_bh) {
952 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
953
954 /*
955 * The buffer head should have an attached journal head at this
956 * point. However, if the data is corrupted and an indirect
957 * block pointed to itself, it would have been detached when
958 * the block was cleared. Check for this instead of OOPSing.
959 */
960 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
961 ext4_handle_dirty_metadata(handle, inode, this_bh);
962 else
963 EXT4_ERROR_INODE(inode,
964 "circular indirect block detected at "
965 "block %llu",
966 (unsigned long long) this_bh->b_blocknr);
967 }
968}
969
970/**
971 * ext4_free_branches - free an array of branches
972 * @handle: JBD handle for this transaction
973 * @inode: inode we are dealing with
974 * @parent_bh: the buffer_head which contains *@first and *@last
975 * @first: array of block numbers
976 * @last: pointer immediately past the end of array
977 * @depth: depth of the branches to free
978 *
979 * We are freeing all blocks referred from these branches (numbers are
980 * stored as little-endian 32-bit) and updating @inode->i_blocks
981 * appropriately.
982 */
983static void ext4_free_branches(handle_t *handle, struct inode *inode,
984 struct buffer_head *parent_bh,
985 __le32 *first, __le32 *last, int depth)
986{
987 ext4_fsblk_t nr;
988 __le32 *p;
989
990 if (ext4_handle_is_aborted(handle))
991 return;
992
993 if (depth--) {
994 struct buffer_head *bh;
995 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
996 p = last;
997 while (--p >= first) {
998 nr = le32_to_cpu(*p);
999 if (!nr)
1000 continue; /* A hole */
1001
1002 if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
1003 nr, 1)) {
1004 EXT4_ERROR_INODE(inode,
1005 "invalid indirect mapped "
1006 "block %lu (level %d)",
1007 (unsigned long) nr, depth);
1008 break;
1009 }
1010
1011 /* Go read the buffer for the next level down */
1012 bh = sb_bread(inode->i_sb, nr);
1013
1014 /*
1015 * A read failure? Report error and clear slot
1016 * (should be rare).
1017 */
1018 if (!bh) {
1019 EXT4_ERROR_INODE_BLOCK(inode, nr,
1020 "Read failure");
1021 continue;
1022 }
1023
1024 /* This zaps the entire block. Bottom up. */
1025 BUFFER_TRACE(bh, "free child branches");
1026 ext4_free_branches(handle, inode, bh,
1027 (__le32 *) bh->b_data,
1028 (__le32 *) bh->b_data + addr_per_block,
1029 depth);
1030 brelse(bh);
1031
1032 /*
1033 * Everything below this this pointer has been
1034 * released. Now let this top-of-subtree go.
1035 *
1036 * We want the freeing of this indirect block to be
1037 * atomic in the journal with the updating of the
1038 * bitmap block which owns it. So make some room in
1039 * the journal.
1040 *
1041 * We zero the parent pointer *after* freeing its
1042 * pointee in the bitmaps, so if extend_transaction()
1043 * for some reason fails to put the bitmap changes and
1044 * the release into the same transaction, recovery
1045 * will merely complain about releasing a free block,
1046 * rather than leaking blocks.
1047 */
1048 if (ext4_handle_is_aborted(handle))
1049 return;
1050 if (try_to_extend_transaction(handle, inode)) {
1051 ext4_mark_inode_dirty(handle, inode);
1052 ext4_truncate_restart_trans(handle, inode,
1053 ext4_blocks_for_truncate(inode));
1054 }
1055
1056 /*
1057 * The forget flag here is critical because if
1058 * we are journaling (and not doing data
1059 * journaling), we have to make sure a revoke
1060 * record is written to prevent the journal
1061 * replay from overwriting the (former)
1062 * indirect block if it gets reallocated as a
1063 * data block. This must happen in the same
1064 * transaction where the data blocks are
1065 * actually freed.
1066 */
1067 ext4_free_blocks(handle, inode, NULL, nr, 1,
1068 EXT4_FREE_BLOCKS_METADATA|
1069 EXT4_FREE_BLOCKS_FORGET);
1070
1071 if (parent_bh) {
1072 /*
1073 * The block which we have just freed is
1074 * pointed to by an indirect block: journal it
1075 */
1076 BUFFER_TRACE(parent_bh, "get_write_access");
1077 if (!ext4_journal_get_write_access(handle,
1078 parent_bh)){
1079 *p = 0;
1080 BUFFER_TRACE(parent_bh,
1081 "call ext4_handle_dirty_metadata");
1082 ext4_handle_dirty_metadata(handle,
1083 inode,
1084 parent_bh);
1085 }
1086 }
1087 }
1088 } else {
1089 /* We have reached the bottom of the tree. */
1090 BUFFER_TRACE(parent_bh, "free data blocks");
1091 ext4_free_data(handle, inode, parent_bh, first, last);
1092 }
1093}
1094
1095void ext4_ind_truncate(handle_t *handle, struct inode *inode)
1096{
1097 struct ext4_inode_info *ei = EXT4_I(inode);
1098 __le32 *i_data = ei->i_data;
1099 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1100 ext4_lblk_t offsets[4];
1101 Indirect chain[4];
1102 Indirect *partial;
1103 __le32 nr = 0;
1104 int n = 0;
1105 ext4_lblk_t last_block, max_block;
1106 unsigned blocksize = inode->i_sb->s_blocksize;
1107
1108 last_block = (inode->i_size + blocksize-1)
1109 >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1110 max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1111 >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1112
1113 if (last_block != max_block) {
1114 n = ext4_block_to_path(inode, last_block, offsets, NULL);
1115 if (n == 0)
1116 return;
1117 }
1118
1119 ext4_es_remove_extent(inode, last_block, EXT_MAX_BLOCKS - last_block);
1120
1121 /*
1122 * The orphan list entry will now protect us from any crash which
1123 * occurs before the truncate completes, so it is now safe to propagate
1124 * the new, shorter inode size (held for now in i_size) into the
1125 * on-disk inode. We do this via i_disksize, which is the value which
1126 * ext4 *really* writes onto the disk inode.
1127 */
1128 ei->i_disksize = inode->i_size;
1129
1130 if (last_block == max_block) {
1131 /*
1132 * It is unnecessary to free any data blocks if last_block is
1133 * equal to the indirect block limit.
1134 */
1135 return;
1136 } else if (n == 1) { /* direct blocks */
1137 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
1138 i_data + EXT4_NDIR_BLOCKS);
1139 goto do_indirects;
1140 }
1141
1142 partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1143 /* Kill the top of shared branch (not detached) */
1144 if (nr) {
1145 if (partial == chain) {
1146 /* Shared branch grows from the inode */
1147 ext4_free_branches(handle, inode, NULL,
1148 &nr, &nr+1, (chain+n-1) - partial);
1149 *partial->p = 0;
1150 /*
1151 * We mark the inode dirty prior to restart,
1152 * and prior to stop. No need for it here.
1153 */
1154 } else {
1155 /* Shared branch grows from an indirect block */
1156 BUFFER_TRACE(partial->bh, "get_write_access");
1157 ext4_free_branches(handle, inode, partial->bh,
1158 partial->p,
1159 partial->p+1, (chain+n-1) - partial);
1160 }
1161 }
1162 /* Clear the ends of indirect blocks on the shared branch */
1163 while (partial > chain) {
1164 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
1165 (__le32*)partial->bh->b_data+addr_per_block,
1166 (chain+n-1) - partial);
1167 BUFFER_TRACE(partial->bh, "call brelse");
1168 brelse(partial->bh);
1169 partial--;
1170 }
1171do_indirects:
1172 /* Kill the remaining (whole) subtrees */
1173 switch (offsets[0]) {
1174 default:
1175 nr = i_data[EXT4_IND_BLOCK];
1176 if (nr) {
1177 ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1178 i_data[EXT4_IND_BLOCK] = 0;
1179 }
1180 case EXT4_IND_BLOCK:
1181 nr = i_data[EXT4_DIND_BLOCK];
1182 if (nr) {
1183 ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1184 i_data[EXT4_DIND_BLOCK] = 0;
1185 }
1186 case EXT4_DIND_BLOCK:
1187 nr = i_data[EXT4_TIND_BLOCK];
1188 if (nr) {
1189 ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1190 i_data[EXT4_TIND_BLOCK] = 0;
1191 }
1192 case EXT4_TIND_BLOCK:
1193 ;
1194 }
1195}
1196
1197/**
1198 * ext4_ind_remove_space - remove space from the range
1199 * @handle: JBD handle for this transaction
1200 * @inode: inode we are dealing with
1201 * @start: First block to remove
1202 * @end: One block after the last block to remove (exclusive)
1203 *
1204 * Free the blocks in the defined range (end is exclusive endpoint of
1205 * range). This is used by ext4_punch_hole().
1206 */
1207int ext4_ind_remove_space(handle_t *handle, struct inode *inode,
1208 ext4_lblk_t start, ext4_lblk_t end)
1209{
1210 struct ext4_inode_info *ei = EXT4_I(inode);
1211 __le32 *i_data = ei->i_data;
1212 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1213 ext4_lblk_t offsets[4], offsets2[4];
1214 Indirect chain[4], chain2[4];
1215 Indirect *partial, *partial2;
1216 ext4_lblk_t max_block;
1217 __le32 nr = 0, nr2 = 0;
1218 int n = 0, n2 = 0;
1219 unsigned blocksize = inode->i_sb->s_blocksize;
1220
1221 max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1222 >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1223 if (end >= max_block)
1224 end = max_block;
1225 if ((start >= end) || (start > max_block))
1226 return 0;
1227
1228 n = ext4_block_to_path(inode, start, offsets, NULL);
1229 n2 = ext4_block_to_path(inode, end, offsets2, NULL);
1230
1231 BUG_ON(n > n2);
1232
1233 if ((n == 1) && (n == n2)) {
1234 /* We're punching only within direct block range */
1235 ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1236 i_data + offsets2[0]);
1237 return 0;
1238 } else if (n2 > n) {
1239 /*
1240 * Start and end are on a different levels so we're going to
1241 * free partial block at start, and partial block at end of
1242 * the range. If there are some levels in between then
1243 * do_indirects label will take care of that.
1244 */
1245
1246 if (n == 1) {
1247 /*
1248 * Start is at the direct block level, free
1249 * everything to the end of the level.
1250 */
1251 ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1252 i_data + EXT4_NDIR_BLOCKS);
1253 goto end_range;
1254 }
1255
1256
1257 partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1258 if (nr) {
1259 if (partial == chain) {
1260 /* Shared branch grows from the inode */
1261 ext4_free_branches(handle, inode, NULL,
1262 &nr, &nr+1, (chain+n-1) - partial);
1263 *partial->p = 0;
1264 } else {
1265 /* Shared branch grows from an indirect block */
1266 BUFFER_TRACE(partial->bh, "get_write_access");
1267 ext4_free_branches(handle, inode, partial->bh,
1268 partial->p,
1269 partial->p+1, (chain+n-1) - partial);
1270 }
1271 }
1272
1273 /*
1274 * Clear the ends of indirect blocks on the shared branch
1275 * at the start of the range
1276 */
1277 while (partial > chain) {
1278 ext4_free_branches(handle, inode, partial->bh,
1279 partial->p + 1,
1280 (__le32 *)partial->bh->b_data+addr_per_block,
1281 (chain+n-1) - partial);
1282 BUFFER_TRACE(partial->bh, "call brelse");
1283 brelse(partial->bh);
1284 partial--;
1285 }
1286
1287end_range:
1288 partial2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1289 if (nr2) {
1290 if (partial2 == chain2) {
1291 /*
1292 * Remember, end is exclusive so here we're at
1293 * the start of the next level we're not going
1294 * to free. Everything was covered by the start
1295 * of the range.
1296 */
1297 goto do_indirects;
1298 }
1299 } else {
1300 /*
1301 * ext4_find_shared returns Indirect structure which
1302 * points to the last element which should not be
1303 * removed by truncate. But this is end of the range
1304 * in punch_hole so we need to point to the next element
1305 */
1306 partial2->p++;
1307 }
1308
1309 /*
1310 * Clear the ends of indirect blocks on the shared branch
1311 * at the end of the range
1312 */
1313 while (partial2 > chain2) {
1314 ext4_free_branches(handle, inode, partial2->bh,
1315 (__le32 *)partial2->bh->b_data,
1316 partial2->p,
1317 (chain2+n2-1) - partial2);
1318 BUFFER_TRACE(partial2->bh, "call brelse");
1319 brelse(partial2->bh);
1320 partial2--;
1321 }
1322 goto do_indirects;
1323 }
1324
1325 /* Punch happened within the same level (n == n2) */
1326 partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1327 partial2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1328
1329 /* Free top, but only if partial2 isn't its subtree. */
1330 if (nr) {
1331 int level = min(partial - chain, partial2 - chain2);
1332 int i;
1333 int subtree = 1;
1334
1335 for (i = 0; i <= level; i++) {
1336 if (offsets[i] != offsets2[i]) {
1337 subtree = 0;
1338 break;
1339 }
1340 }
1341
1342 if (!subtree) {
1343 if (partial == chain) {
1344 /* Shared branch grows from the inode */
1345 ext4_free_branches(handle, inode, NULL,
1346 &nr, &nr+1,
1347 (chain+n-1) - partial);
1348 *partial->p = 0;
1349 } else {
1350 /* Shared branch grows from an indirect block */
1351 BUFFER_TRACE(partial->bh, "get_write_access");
1352 ext4_free_branches(handle, inode, partial->bh,
1353 partial->p,
1354 partial->p+1,
1355 (chain+n-1) - partial);
1356 }
1357 }
1358 }
1359
1360 if (!nr2) {
1361 /*
1362 * ext4_find_shared returns Indirect structure which
1363 * points to the last element which should not be
1364 * removed by truncate. But this is end of the range
1365 * in punch_hole so we need to point to the next element
1366 */
1367 partial2->p++;
1368 }
1369
1370 while (partial > chain || partial2 > chain2) {
1371 int depth = (chain+n-1) - partial;
1372 int depth2 = (chain2+n2-1) - partial2;
1373
1374 if (partial > chain && partial2 > chain2 &&
1375 partial->bh->b_blocknr == partial2->bh->b_blocknr) {
1376 /*
1377 * We've converged on the same block. Clear the range,
1378 * then we're done.
1379 */
1380 ext4_free_branches(handle, inode, partial->bh,
1381 partial->p + 1,
1382 partial2->p,
1383 (chain+n-1) - partial);
1384 BUFFER_TRACE(partial->bh, "call brelse");
1385 brelse(partial->bh);
1386 BUFFER_TRACE(partial2->bh, "call brelse");
1387 brelse(partial2->bh);
1388 return 0;
1389 }
1390
1391 /*
1392 * The start and end partial branches may not be at the same
1393 * level even though the punch happened within one level. So, we
1394 * give them a chance to arrive at the same level, then walk
1395 * them in step with each other until we converge on the same
1396 * block.
1397 */
1398 if (partial > chain && depth <= depth2) {
1399 ext4_free_branches(handle, inode, partial->bh,
1400 partial->p + 1,
1401 (__le32 *)partial->bh->b_data+addr_per_block,
1402 (chain+n-1) - partial);
1403 BUFFER_TRACE(partial->bh, "call brelse");
1404 brelse(partial->bh);
1405 partial--;
1406 }
1407 if (partial2 > chain2 && depth2 <= depth) {
1408 ext4_free_branches(handle, inode, partial2->bh,
1409 (__le32 *)partial2->bh->b_data,
1410 partial2->p,
1411 (chain2+n2-1) - partial2);
1412 BUFFER_TRACE(partial2->bh, "call brelse");
1413 brelse(partial2->bh);
1414 partial2--;
1415 }
1416 }
1417 return 0;
1418
1419do_indirects:
1420 /* Kill the remaining (whole) subtrees */
1421 switch (offsets[0]) {
1422 default:
1423 if (++n >= n2)
1424 return 0;
1425 nr = i_data[EXT4_IND_BLOCK];
1426 if (nr) {
1427 ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1428 i_data[EXT4_IND_BLOCK] = 0;
1429 }
1430 case EXT4_IND_BLOCK:
1431 if (++n >= n2)
1432 return 0;
1433 nr = i_data[EXT4_DIND_BLOCK];
1434 if (nr) {
1435 ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1436 i_data[EXT4_DIND_BLOCK] = 0;
1437 }
1438 case EXT4_DIND_BLOCK:
1439 if (++n >= n2)
1440 return 0;
1441 nr = i_data[EXT4_TIND_BLOCK];
1442 if (nr) {
1443 ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1444 i_data[EXT4_TIND_BLOCK] = 0;
1445 }
1446 case EXT4_TIND_BLOCK:
1447 ;
1448 }
1449 return 0;
1450}
1/*
2 * linux/fs/ext4/indirect.c
3 *
4 * from
5 *
6 * linux/fs/ext4/inode.c
7 *
8 * Copyright (C) 1992, 1993, 1994, 1995
9 * Remy Card (card@masi.ibp.fr)
10 * Laboratoire MASI - Institut Blaise Pascal
11 * Universite Pierre et Marie Curie (Paris VI)
12 *
13 * from
14 *
15 * linux/fs/minix/inode.c
16 *
17 * Copyright (C) 1991, 1992 Linus Torvalds
18 *
19 * Goal-directed block allocation by Stephen Tweedie
20 * (sct@redhat.com), 1993, 1998
21 */
22
23#include <linux/aio.h>
24#include "ext4_jbd2.h"
25#include "truncate.h"
26
27#include <trace/events/ext4.h>
28
29typedef struct {
30 __le32 *p;
31 __le32 key;
32 struct buffer_head *bh;
33} Indirect;
34
35static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
36{
37 p->key = *(p->p = v);
38 p->bh = bh;
39}
40
41/**
42 * ext4_block_to_path - parse the block number into array of offsets
43 * @inode: inode in question (we are only interested in its superblock)
44 * @i_block: block number to be parsed
45 * @offsets: array to store the offsets in
46 * @boundary: set this non-zero if the referred-to block is likely to be
47 * followed (on disk) by an indirect block.
48 *
49 * To store the locations of file's data ext4 uses a data structure common
50 * for UNIX filesystems - tree of pointers anchored in the inode, with
51 * data blocks at leaves and indirect blocks in intermediate nodes.
52 * This function translates the block number into path in that tree -
53 * return value is the path length and @offsets[n] is the offset of
54 * pointer to (n+1)th node in the nth one. If @block is out of range
55 * (negative or too large) warning is printed and zero returned.
56 *
57 * Note: function doesn't find node addresses, so no IO is needed. All
58 * we need to know is the capacity of indirect blocks (taken from the
59 * inode->i_sb).
60 */
61
62/*
63 * Portability note: the last comparison (check that we fit into triple
64 * indirect block) is spelled differently, because otherwise on an
65 * architecture with 32-bit longs and 8Kb pages we might get into trouble
66 * if our filesystem had 8Kb blocks. We might use long long, but that would
67 * kill us on x86. Oh, well, at least the sign propagation does not matter -
68 * i_block would have to be negative in the very beginning, so we would not
69 * get there at all.
70 */
71
72static int ext4_block_to_path(struct inode *inode,
73 ext4_lblk_t i_block,
74 ext4_lblk_t offsets[4], int *boundary)
75{
76 int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
77 int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
78 const long direct_blocks = EXT4_NDIR_BLOCKS,
79 indirect_blocks = ptrs,
80 double_blocks = (1 << (ptrs_bits * 2));
81 int n = 0;
82 int final = 0;
83
84 if (i_block < direct_blocks) {
85 offsets[n++] = i_block;
86 final = direct_blocks;
87 } else if ((i_block -= direct_blocks) < indirect_blocks) {
88 offsets[n++] = EXT4_IND_BLOCK;
89 offsets[n++] = i_block;
90 final = ptrs;
91 } else if ((i_block -= indirect_blocks) < double_blocks) {
92 offsets[n++] = EXT4_DIND_BLOCK;
93 offsets[n++] = i_block >> ptrs_bits;
94 offsets[n++] = i_block & (ptrs - 1);
95 final = ptrs;
96 } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
97 offsets[n++] = EXT4_TIND_BLOCK;
98 offsets[n++] = i_block >> (ptrs_bits * 2);
99 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
100 offsets[n++] = i_block & (ptrs - 1);
101 final = ptrs;
102 } else {
103 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
104 i_block + direct_blocks +
105 indirect_blocks + double_blocks, inode->i_ino);
106 }
107 if (boundary)
108 *boundary = final - 1 - (i_block & (ptrs - 1));
109 return n;
110}
111
112/**
113 * ext4_get_branch - read the chain of indirect blocks leading to data
114 * @inode: inode in question
115 * @depth: depth of the chain (1 - direct pointer, etc.)
116 * @offsets: offsets of pointers in inode/indirect blocks
117 * @chain: place to store the result
118 * @err: here we store the error value
119 *
120 * Function fills the array of triples <key, p, bh> and returns %NULL
121 * if everything went OK or the pointer to the last filled triple
122 * (incomplete one) otherwise. Upon the return chain[i].key contains
123 * the number of (i+1)-th block in the chain (as it is stored in memory,
124 * i.e. little-endian 32-bit), chain[i].p contains the address of that
125 * number (it points into struct inode for i==0 and into the bh->b_data
126 * for i>0) and chain[i].bh points to the buffer_head of i-th indirect
127 * block for i>0 and NULL for i==0. In other words, it holds the block
128 * numbers of the chain, addresses they were taken from (and where we can
129 * verify that chain did not change) and buffer_heads hosting these
130 * numbers.
131 *
132 * Function stops when it stumbles upon zero pointer (absent block)
133 * (pointer to last triple returned, *@err == 0)
134 * or when it gets an IO error reading an indirect block
135 * (ditto, *@err == -EIO)
136 * or when it reads all @depth-1 indirect blocks successfully and finds
137 * the whole chain, all way to the data (returns %NULL, *err == 0).
138 *
139 * Need to be called with
140 * down_read(&EXT4_I(inode)->i_data_sem)
141 */
142static Indirect *ext4_get_branch(struct inode *inode, int depth,
143 ext4_lblk_t *offsets,
144 Indirect chain[4], int *err)
145{
146 struct super_block *sb = inode->i_sb;
147 Indirect *p = chain;
148 struct buffer_head *bh;
149 int ret = -EIO;
150
151 *err = 0;
152 /* i_data is not going away, no lock needed */
153 add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
154 if (!p->key)
155 goto no_block;
156 while (--depth) {
157 bh = sb_getblk(sb, le32_to_cpu(p->key));
158 if (unlikely(!bh)) {
159 ret = -ENOMEM;
160 goto failure;
161 }
162
163 if (!bh_uptodate_or_lock(bh)) {
164 if (bh_submit_read(bh) < 0) {
165 put_bh(bh);
166 goto failure;
167 }
168 /* validate block references */
169 if (ext4_check_indirect_blockref(inode, bh)) {
170 put_bh(bh);
171 goto failure;
172 }
173 }
174
175 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
176 /* Reader: end */
177 if (!p->key)
178 goto no_block;
179 }
180 return NULL;
181
182failure:
183 *err = ret;
184no_block:
185 return p;
186}
187
188/**
189 * ext4_find_near - find a place for allocation with sufficient locality
190 * @inode: owner
191 * @ind: descriptor of indirect block.
192 *
193 * This function returns the preferred place for block allocation.
194 * It is used when heuristic for sequential allocation fails.
195 * Rules are:
196 * + if there is a block to the left of our position - allocate near it.
197 * + if pointer will live in indirect block - allocate near that block.
198 * + if pointer will live in inode - allocate in the same
199 * cylinder group.
200 *
201 * In the latter case we colour the starting block by the callers PID to
202 * prevent it from clashing with concurrent allocations for a different inode
203 * in the same block group. The PID is used here so that functionally related
204 * files will be close-by on-disk.
205 *
206 * Caller must make sure that @ind is valid and will stay that way.
207 */
208static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
209{
210 struct ext4_inode_info *ei = EXT4_I(inode);
211 __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
212 __le32 *p;
213
214 /* Try to find previous block */
215 for (p = ind->p - 1; p >= start; p--) {
216 if (*p)
217 return le32_to_cpu(*p);
218 }
219
220 /* No such thing, so let's try location of indirect block */
221 if (ind->bh)
222 return ind->bh->b_blocknr;
223
224 /*
225 * It is going to be referred to from the inode itself? OK, just put it
226 * into the same cylinder group then.
227 */
228 return ext4_inode_to_goal_block(inode);
229}
230
231/**
232 * ext4_find_goal - find a preferred place for allocation.
233 * @inode: owner
234 * @block: block we want
235 * @partial: pointer to the last triple within a chain
236 *
237 * Normally this function find the preferred place for block allocation,
238 * returns it.
239 * Because this is only used for non-extent files, we limit the block nr
240 * to 32 bits.
241 */
242static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
243 Indirect *partial)
244{
245 ext4_fsblk_t goal;
246
247 /*
248 * XXX need to get goal block from mballoc's data structures
249 */
250
251 goal = ext4_find_near(inode, partial);
252 goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
253 return goal;
254}
255
256/**
257 * ext4_blks_to_allocate - Look up the block map and count the number
258 * of direct blocks need to be allocated for the given branch.
259 *
260 * @branch: chain of indirect blocks
261 * @k: number of blocks need for indirect blocks
262 * @blks: number of data blocks to be mapped.
263 * @blocks_to_boundary: the offset in the indirect block
264 *
265 * return the total number of blocks to be allocate, including the
266 * direct and indirect blocks.
267 */
268static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
269 int blocks_to_boundary)
270{
271 unsigned int count = 0;
272
273 /*
274 * Simple case, [t,d]Indirect block(s) has not allocated yet
275 * then it's clear blocks on that path have not allocated
276 */
277 if (k > 0) {
278 /* right now we don't handle cross boundary allocation */
279 if (blks < blocks_to_boundary + 1)
280 count += blks;
281 else
282 count += blocks_to_boundary + 1;
283 return count;
284 }
285
286 count++;
287 while (count < blks && count <= blocks_to_boundary &&
288 le32_to_cpu(*(branch[0].p + count)) == 0) {
289 count++;
290 }
291 return count;
292}
293
294/**
295 * ext4_alloc_branch - allocate and set up a chain of blocks.
296 * @handle: handle for this transaction
297 * @inode: owner
298 * @indirect_blks: number of allocated indirect blocks
299 * @blks: number of allocated direct blocks
300 * @goal: preferred place for allocation
301 * @offsets: offsets (in the blocks) to store the pointers to next.
302 * @branch: place to store the chain in.
303 *
304 * This function allocates blocks, zeroes out all but the last one,
305 * links them into chain and (if we are synchronous) writes them to disk.
306 * In other words, it prepares a branch that can be spliced onto the
307 * inode. It stores the information about that chain in the branch[], in
308 * the same format as ext4_get_branch() would do. We are calling it after
309 * we had read the existing part of chain and partial points to the last
310 * triple of that (one with zero ->key). Upon the exit we have the same
311 * picture as after the successful ext4_get_block(), except that in one
312 * place chain is disconnected - *branch->p is still zero (we did not
313 * set the last link), but branch->key contains the number that should
314 * be placed into *branch->p to fill that gap.
315 *
316 * If allocation fails we free all blocks we've allocated (and forget
317 * their buffer_heads) and return the error value the from failed
318 * ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
319 * as described above and return 0.
320 */
321static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
322 ext4_lblk_t iblock, int indirect_blks,
323 int *blks, ext4_fsblk_t goal,
324 ext4_lblk_t *offsets, Indirect *branch)
325{
326 struct ext4_allocation_request ar;
327 struct buffer_head * bh;
328 ext4_fsblk_t b, new_blocks[4];
329 __le32 *p;
330 int i, j, err, len = 1;
331
332 /*
333 * Set up for the direct block allocation
334 */
335 memset(&ar, 0, sizeof(ar));
336 ar.inode = inode;
337 ar.len = *blks;
338 ar.logical = iblock;
339 if (S_ISREG(inode->i_mode))
340 ar.flags = EXT4_MB_HINT_DATA;
341
342 for (i = 0; i <= indirect_blks; i++) {
343 if (i == indirect_blks) {
344 ar.goal = goal;
345 new_blocks[i] = ext4_mb_new_blocks(handle, &ar, &err);
346 } else
347 goal = new_blocks[i] = ext4_new_meta_blocks(handle, inode,
348 goal, 0, NULL, &err);
349 if (err) {
350 i--;
351 goto failed;
352 }
353 branch[i].key = cpu_to_le32(new_blocks[i]);
354 if (i == 0)
355 continue;
356
357 bh = branch[i].bh = sb_getblk(inode->i_sb, new_blocks[i-1]);
358 if (unlikely(!bh)) {
359 err = -ENOMEM;
360 goto failed;
361 }
362 lock_buffer(bh);
363 BUFFER_TRACE(bh, "call get_create_access");
364 err = ext4_journal_get_create_access(handle, bh);
365 if (err) {
366 unlock_buffer(bh);
367 goto failed;
368 }
369
370 memset(bh->b_data, 0, bh->b_size);
371 p = branch[i].p = (__le32 *) bh->b_data + offsets[i];
372 b = new_blocks[i];
373
374 if (i == indirect_blks)
375 len = ar.len;
376 for (j = 0; j < len; j++)
377 *p++ = cpu_to_le32(b++);
378
379 BUFFER_TRACE(bh, "marking uptodate");
380 set_buffer_uptodate(bh);
381 unlock_buffer(bh);
382
383 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
384 err = ext4_handle_dirty_metadata(handle, inode, bh);
385 if (err)
386 goto failed;
387 }
388 *blks = ar.len;
389 return 0;
390failed:
391 for (; i >= 0; i--) {
392 if (i != indirect_blks && branch[i].bh)
393 ext4_forget(handle, 1, inode, branch[i].bh,
394 branch[i].bh->b_blocknr);
395 ext4_free_blocks(handle, inode, NULL, new_blocks[i],
396 (i == indirect_blks) ? ar.len : 1, 0);
397 }
398 return err;
399}
400
401/**
402 * ext4_splice_branch - splice the allocated branch onto inode.
403 * @handle: handle for this transaction
404 * @inode: owner
405 * @block: (logical) number of block we are adding
406 * @chain: chain of indirect blocks (with a missing link - see
407 * ext4_alloc_branch)
408 * @where: location of missing link
409 * @num: number of indirect blocks we are adding
410 * @blks: number of direct blocks we are adding
411 *
412 * This function fills the missing link and does all housekeeping needed in
413 * inode (->i_blocks, etc.). In case of success we end up with the full
414 * chain to new block and return 0.
415 */
416static int ext4_splice_branch(handle_t *handle, struct inode *inode,
417 ext4_lblk_t block, Indirect *where, int num,
418 int blks)
419{
420 int i;
421 int err = 0;
422 ext4_fsblk_t current_block;
423
424 /*
425 * If we're splicing into a [td]indirect block (as opposed to the
426 * inode) then we need to get write access to the [td]indirect block
427 * before the splice.
428 */
429 if (where->bh) {
430 BUFFER_TRACE(where->bh, "get_write_access");
431 err = ext4_journal_get_write_access(handle, where->bh);
432 if (err)
433 goto err_out;
434 }
435 /* That's it */
436
437 *where->p = where->key;
438
439 /*
440 * Update the host buffer_head or inode to point to more just allocated
441 * direct blocks blocks
442 */
443 if (num == 0 && blks > 1) {
444 current_block = le32_to_cpu(where->key) + 1;
445 for (i = 1; i < blks; i++)
446 *(where->p + i) = cpu_to_le32(current_block++);
447 }
448
449 /* We are done with atomic stuff, now do the rest of housekeeping */
450 /* had we spliced it onto indirect block? */
451 if (where->bh) {
452 /*
453 * If we spliced it onto an indirect block, we haven't
454 * altered the inode. Note however that if it is being spliced
455 * onto an indirect block at the very end of the file (the
456 * file is growing) then we *will* alter the inode to reflect
457 * the new i_size. But that is not done here - it is done in
458 * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
459 */
460 jbd_debug(5, "splicing indirect only\n");
461 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
462 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
463 if (err)
464 goto err_out;
465 } else {
466 /*
467 * OK, we spliced it into the inode itself on a direct block.
468 */
469 ext4_mark_inode_dirty(handle, inode);
470 jbd_debug(5, "splicing direct\n");
471 }
472 return err;
473
474err_out:
475 for (i = 1; i <= num; i++) {
476 /*
477 * branch[i].bh is newly allocated, so there is no
478 * need to revoke the block, which is why we don't
479 * need to set EXT4_FREE_BLOCKS_METADATA.
480 */
481 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
482 EXT4_FREE_BLOCKS_FORGET);
483 }
484 ext4_free_blocks(handle, inode, NULL, le32_to_cpu(where[num].key),
485 blks, 0);
486
487 return err;
488}
489
490/*
491 * The ext4_ind_map_blocks() function handles non-extents inodes
492 * (i.e., using the traditional indirect/double-indirect i_blocks
493 * scheme) for ext4_map_blocks().
494 *
495 * Allocation strategy is simple: if we have to allocate something, we will
496 * have to go the whole way to leaf. So let's do it before attaching anything
497 * to tree, set linkage between the newborn blocks, write them if sync is
498 * required, recheck the path, free and repeat if check fails, otherwise
499 * set the last missing link (that will protect us from any truncate-generated
500 * removals - all blocks on the path are immune now) and possibly force the
501 * write on the parent block.
502 * That has a nice additional property: no special recovery from the failed
503 * allocations is needed - we simply release blocks and do not touch anything
504 * reachable from inode.
505 *
506 * `handle' can be NULL if create == 0.
507 *
508 * return > 0, # of blocks mapped or allocated.
509 * return = 0, if plain lookup failed.
510 * return < 0, error case.
511 *
512 * The ext4_ind_get_blocks() function should be called with
513 * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
514 * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
515 * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
516 * blocks.
517 */
518int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
519 struct ext4_map_blocks *map,
520 int flags)
521{
522 int err = -EIO;
523 ext4_lblk_t offsets[4];
524 Indirect chain[4];
525 Indirect *partial;
526 ext4_fsblk_t goal;
527 int indirect_blks;
528 int blocks_to_boundary = 0;
529 int depth;
530 int count = 0;
531 ext4_fsblk_t first_block = 0;
532
533 trace_ext4_ind_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
534 J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
535 J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
536 depth = ext4_block_to_path(inode, map->m_lblk, offsets,
537 &blocks_to_boundary);
538
539 if (depth == 0)
540 goto out;
541
542 partial = ext4_get_branch(inode, depth, offsets, chain, &err);
543
544 /* Simplest case - block found, no allocation needed */
545 if (!partial) {
546 first_block = le32_to_cpu(chain[depth - 1].key);
547 count++;
548 /*map more blocks*/
549 while (count < map->m_len && count <= blocks_to_boundary) {
550 ext4_fsblk_t blk;
551
552 blk = le32_to_cpu(*(chain[depth-1].p + count));
553
554 if (blk == first_block + count)
555 count++;
556 else
557 break;
558 }
559 goto got_it;
560 }
561
562 /* Next simple case - plain lookup or failed read of indirect block */
563 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
564 goto cleanup;
565
566 /*
567 * Okay, we need to do block allocation.
568 */
569 if (EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
570 EXT4_FEATURE_RO_COMPAT_BIGALLOC)) {
571 EXT4_ERROR_INODE(inode, "Can't allocate blocks for "
572 "non-extent mapped inodes with bigalloc");
573 return -ENOSPC;
574 }
575
576 goal = ext4_find_goal(inode, map->m_lblk, partial);
577
578 /* the number of blocks need to allocate for [d,t]indirect blocks */
579 indirect_blks = (chain + depth) - partial - 1;
580
581 /*
582 * Next look up the indirect map to count the totoal number of
583 * direct blocks to allocate for this branch.
584 */
585 count = ext4_blks_to_allocate(partial, indirect_blks,
586 map->m_len, blocks_to_boundary);
587 /*
588 * Block out ext4_truncate while we alter the tree
589 */
590 err = ext4_alloc_branch(handle, inode, map->m_lblk, indirect_blks,
591 &count, goal,
592 offsets + (partial - chain), partial);
593
594 /*
595 * The ext4_splice_branch call will free and forget any buffers
596 * on the new chain if there is a failure, but that risks using
597 * up transaction credits, especially for bitmaps where the
598 * credits cannot be returned. Can we handle this somehow? We
599 * may need to return -EAGAIN upwards in the worst case. --sct
600 */
601 if (!err)
602 err = ext4_splice_branch(handle, inode, map->m_lblk,
603 partial, indirect_blks, count);
604 if (err)
605 goto cleanup;
606
607 map->m_flags |= EXT4_MAP_NEW;
608
609 ext4_update_inode_fsync_trans(handle, inode, 1);
610got_it:
611 map->m_flags |= EXT4_MAP_MAPPED;
612 map->m_pblk = le32_to_cpu(chain[depth-1].key);
613 map->m_len = count;
614 if (count > blocks_to_boundary)
615 map->m_flags |= EXT4_MAP_BOUNDARY;
616 err = count;
617 /* Clean up and exit */
618 partial = chain + depth - 1; /* the whole chain */
619cleanup:
620 while (partial > chain) {
621 BUFFER_TRACE(partial->bh, "call brelse");
622 brelse(partial->bh);
623 partial--;
624 }
625out:
626 trace_ext4_ind_map_blocks_exit(inode, flags, map, err);
627 return err;
628}
629
630/*
631 * O_DIRECT for ext3 (or indirect map) based files
632 *
633 * If the O_DIRECT write will extend the file then add this inode to the
634 * orphan list. So recovery will truncate it back to the original size
635 * if the machine crashes during the write.
636 *
637 * If the O_DIRECT write is intantiating holes inside i_size and the machine
638 * crashes then stale disk data _may_ be exposed inside the file. But current
639 * VFS code falls back into buffered path in that case so we are safe.
640 */
641ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
642 const struct iovec *iov, loff_t offset,
643 unsigned long nr_segs)
644{
645 struct file *file = iocb->ki_filp;
646 struct inode *inode = file->f_mapping->host;
647 struct ext4_inode_info *ei = EXT4_I(inode);
648 handle_t *handle;
649 ssize_t ret;
650 int orphan = 0;
651 size_t count = iov_length(iov, nr_segs);
652 int retries = 0;
653
654 if (rw == WRITE) {
655 loff_t final_size = offset + count;
656
657 if (final_size > inode->i_size) {
658 /* Credits for sb + inode write */
659 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
660 if (IS_ERR(handle)) {
661 ret = PTR_ERR(handle);
662 goto out;
663 }
664 ret = ext4_orphan_add(handle, inode);
665 if (ret) {
666 ext4_journal_stop(handle);
667 goto out;
668 }
669 orphan = 1;
670 ei->i_disksize = inode->i_size;
671 ext4_journal_stop(handle);
672 }
673 }
674
675retry:
676 if (rw == READ && ext4_should_dioread_nolock(inode)) {
677 /*
678 * Nolock dioread optimization may be dynamically disabled
679 * via ext4_inode_block_unlocked_dio(). Check inode's state
680 * while holding extra i_dio_count ref.
681 */
682 atomic_inc(&inode->i_dio_count);
683 smp_mb();
684 if (unlikely(ext4_test_inode_state(inode,
685 EXT4_STATE_DIOREAD_LOCK))) {
686 inode_dio_done(inode);
687 goto locked;
688 }
689 ret = __blockdev_direct_IO(rw, iocb, inode,
690 inode->i_sb->s_bdev, iov,
691 offset, nr_segs,
692 ext4_get_block, NULL, NULL, 0);
693 inode_dio_done(inode);
694 } else {
695locked:
696 ret = blockdev_direct_IO(rw, iocb, inode, iov,
697 offset, nr_segs, ext4_get_block);
698
699 if (unlikely((rw & WRITE) && ret < 0)) {
700 loff_t isize = i_size_read(inode);
701 loff_t end = offset + iov_length(iov, nr_segs);
702
703 if (end > isize)
704 ext4_truncate_failed_write(inode);
705 }
706 }
707 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
708 goto retry;
709
710 if (orphan) {
711 int err;
712
713 /* Credits for sb + inode write */
714 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
715 if (IS_ERR(handle)) {
716 /* This is really bad luck. We've written the data
717 * but cannot extend i_size. Bail out and pretend
718 * the write failed... */
719 ret = PTR_ERR(handle);
720 if (inode->i_nlink)
721 ext4_orphan_del(NULL, inode);
722
723 goto out;
724 }
725 if (inode->i_nlink)
726 ext4_orphan_del(handle, inode);
727 if (ret > 0) {
728 loff_t end = offset + ret;
729 if (end > inode->i_size) {
730 ei->i_disksize = end;
731 i_size_write(inode, end);
732 /*
733 * We're going to return a positive `ret'
734 * here due to non-zero-length I/O, so there's
735 * no way of reporting error returns from
736 * ext4_mark_inode_dirty() to userspace. So
737 * ignore it.
738 */
739 ext4_mark_inode_dirty(handle, inode);
740 }
741 }
742 err = ext4_journal_stop(handle);
743 if (ret == 0)
744 ret = err;
745 }
746out:
747 return ret;
748}
749
750/*
751 * Calculate the number of metadata blocks need to reserve
752 * to allocate a new block at @lblocks for non extent file based file
753 */
754int ext4_ind_calc_metadata_amount(struct inode *inode, sector_t lblock)
755{
756 struct ext4_inode_info *ei = EXT4_I(inode);
757 sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
758 int blk_bits;
759
760 if (lblock < EXT4_NDIR_BLOCKS)
761 return 0;
762
763 lblock -= EXT4_NDIR_BLOCKS;
764
765 if (ei->i_da_metadata_calc_len &&
766 (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
767 ei->i_da_metadata_calc_len++;
768 return 0;
769 }
770 ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
771 ei->i_da_metadata_calc_len = 1;
772 blk_bits = order_base_2(lblock);
773 return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
774}
775
776/*
777 * Calculate number of indirect blocks touched by mapping @nrblocks logically
778 * contiguous blocks
779 */
780int ext4_ind_trans_blocks(struct inode *inode, int nrblocks)
781{
782 /*
783 * With N contiguous data blocks, we need at most
784 * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
785 * 2 dindirect blocks, and 1 tindirect block
786 */
787 return DIV_ROUND_UP(nrblocks, EXT4_ADDR_PER_BLOCK(inode->i_sb)) + 4;
788}
789
790/*
791 * Truncate transactions can be complex and absolutely huge. So we need to
792 * be able to restart the transaction at a conventient checkpoint to make
793 * sure we don't overflow the journal.
794 *
795 * Try to extend this transaction for the purposes of truncation. If
796 * extend fails, we need to propagate the failure up and restart the
797 * transaction in the top-level truncate loop. --sct
798 *
799 * Returns 0 if we managed to create more room. If we can't create more
800 * room, and the transaction must be restarted we return 1.
801 */
802static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
803{
804 if (!ext4_handle_valid(handle))
805 return 0;
806 if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
807 return 0;
808 if (!ext4_journal_extend(handle, ext4_blocks_for_truncate(inode)))
809 return 0;
810 return 1;
811}
812
813/*
814 * Probably it should be a library function... search for first non-zero word
815 * or memcmp with zero_page, whatever is better for particular architecture.
816 * Linus?
817 */
818static inline int all_zeroes(__le32 *p, __le32 *q)
819{
820 while (p < q)
821 if (*p++)
822 return 0;
823 return 1;
824}
825
826/**
827 * ext4_find_shared - find the indirect blocks for partial truncation.
828 * @inode: inode in question
829 * @depth: depth of the affected branch
830 * @offsets: offsets of pointers in that branch (see ext4_block_to_path)
831 * @chain: place to store the pointers to partial indirect blocks
832 * @top: place to the (detached) top of branch
833 *
834 * This is a helper function used by ext4_truncate().
835 *
836 * When we do truncate() we may have to clean the ends of several
837 * indirect blocks but leave the blocks themselves alive. Block is
838 * partially truncated if some data below the new i_size is referred
839 * from it (and it is on the path to the first completely truncated
840 * data block, indeed). We have to free the top of that path along
841 * with everything to the right of the path. Since no allocation
842 * past the truncation point is possible until ext4_truncate()
843 * finishes, we may safely do the latter, but top of branch may
844 * require special attention - pageout below the truncation point
845 * might try to populate it.
846 *
847 * We atomically detach the top of branch from the tree, store the
848 * block number of its root in *@top, pointers to buffer_heads of
849 * partially truncated blocks - in @chain[].bh and pointers to
850 * their last elements that should not be removed - in
851 * @chain[].p. Return value is the pointer to last filled element
852 * of @chain.
853 *
854 * The work left to caller to do the actual freeing of subtrees:
855 * a) free the subtree starting from *@top
856 * b) free the subtrees whose roots are stored in
857 * (@chain[i].p+1 .. end of @chain[i].bh->b_data)
858 * c) free the subtrees growing from the inode past the @chain[0].
859 * (no partially truncated stuff there). */
860
861static Indirect *ext4_find_shared(struct inode *inode, int depth,
862 ext4_lblk_t offsets[4], Indirect chain[4],
863 __le32 *top)
864{
865 Indirect *partial, *p;
866 int k, err;
867
868 *top = 0;
869 /* Make k index the deepest non-null offset + 1 */
870 for (k = depth; k > 1 && !offsets[k-1]; k--)
871 ;
872 partial = ext4_get_branch(inode, k, offsets, chain, &err);
873 /* Writer: pointers */
874 if (!partial)
875 partial = chain + k-1;
876 /*
877 * If the branch acquired continuation since we've looked at it -
878 * fine, it should all survive and (new) top doesn't belong to us.
879 */
880 if (!partial->key && *partial->p)
881 /* Writer: end */
882 goto no_top;
883 for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
884 ;
885 /*
886 * OK, we've found the last block that must survive. The rest of our
887 * branch should be detached before unlocking. However, if that rest
888 * of branch is all ours and does not grow immediately from the inode
889 * it's easier to cheat and just decrement partial->p.
890 */
891 if (p == chain + k - 1 && p > chain) {
892 p->p--;
893 } else {
894 *top = *p->p;
895 /* Nope, don't do this in ext4. Must leave the tree intact */
896#if 0
897 *p->p = 0;
898#endif
899 }
900 /* Writer: end */
901
902 while (partial > p) {
903 brelse(partial->bh);
904 partial--;
905 }
906no_top:
907 return partial;
908}
909
910/*
911 * Zero a number of block pointers in either an inode or an indirect block.
912 * If we restart the transaction we must again get write access to the
913 * indirect block for further modification.
914 *
915 * We release `count' blocks on disk, but (last - first) may be greater
916 * than `count' because there can be holes in there.
917 *
918 * Return 0 on success, 1 on invalid block range
919 * and < 0 on fatal error.
920 */
921static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
922 struct buffer_head *bh,
923 ext4_fsblk_t block_to_free,
924 unsigned long count, __le32 *first,
925 __le32 *last)
926{
927 __le32 *p;
928 int flags = EXT4_FREE_BLOCKS_VALIDATED;
929 int err;
930
931 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
932 flags |= EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_METADATA;
933 else if (ext4_should_journal_data(inode))
934 flags |= EXT4_FREE_BLOCKS_FORGET;
935
936 if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
937 count)) {
938 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
939 "blocks %llu len %lu",
940 (unsigned long long) block_to_free, count);
941 return 1;
942 }
943
944 if (try_to_extend_transaction(handle, inode)) {
945 if (bh) {
946 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
947 err = ext4_handle_dirty_metadata(handle, inode, bh);
948 if (unlikely(err))
949 goto out_err;
950 }
951 err = ext4_mark_inode_dirty(handle, inode);
952 if (unlikely(err))
953 goto out_err;
954 err = ext4_truncate_restart_trans(handle, inode,
955 ext4_blocks_for_truncate(inode));
956 if (unlikely(err))
957 goto out_err;
958 if (bh) {
959 BUFFER_TRACE(bh, "retaking write access");
960 err = ext4_journal_get_write_access(handle, bh);
961 if (unlikely(err))
962 goto out_err;
963 }
964 }
965
966 for (p = first; p < last; p++)
967 *p = 0;
968
969 ext4_free_blocks(handle, inode, NULL, block_to_free, count, flags);
970 return 0;
971out_err:
972 ext4_std_error(inode->i_sb, err);
973 return err;
974}
975
976/**
977 * ext4_free_data - free a list of data blocks
978 * @handle: handle for this transaction
979 * @inode: inode we are dealing with
980 * @this_bh: indirect buffer_head which contains *@first and *@last
981 * @first: array of block numbers
982 * @last: points immediately past the end of array
983 *
984 * We are freeing all blocks referred from that array (numbers are stored as
985 * little-endian 32-bit) and updating @inode->i_blocks appropriately.
986 *
987 * We accumulate contiguous runs of blocks to free. Conveniently, if these
988 * blocks are contiguous then releasing them at one time will only affect one
989 * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
990 * actually use a lot of journal space.
991 *
992 * @this_bh will be %NULL if @first and @last point into the inode's direct
993 * block pointers.
994 */
995static void ext4_free_data(handle_t *handle, struct inode *inode,
996 struct buffer_head *this_bh,
997 __le32 *first, __le32 *last)
998{
999 ext4_fsblk_t block_to_free = 0; /* Starting block # of a run */
1000 unsigned long count = 0; /* Number of blocks in the run */
1001 __le32 *block_to_free_p = NULL; /* Pointer into inode/ind
1002 corresponding to
1003 block_to_free */
1004 ext4_fsblk_t nr; /* Current block # */
1005 __le32 *p; /* Pointer into inode/ind
1006 for current block */
1007 int err = 0;
1008
1009 if (this_bh) { /* For indirect block */
1010 BUFFER_TRACE(this_bh, "get_write_access");
1011 err = ext4_journal_get_write_access(handle, this_bh);
1012 /* Important: if we can't update the indirect pointers
1013 * to the blocks, we can't free them. */
1014 if (err)
1015 return;
1016 }
1017
1018 for (p = first; p < last; p++) {
1019 nr = le32_to_cpu(*p);
1020 if (nr) {
1021 /* accumulate blocks to free if they're contiguous */
1022 if (count == 0) {
1023 block_to_free = nr;
1024 block_to_free_p = p;
1025 count = 1;
1026 } else if (nr == block_to_free + count) {
1027 count++;
1028 } else {
1029 err = ext4_clear_blocks(handle, inode, this_bh,
1030 block_to_free, count,
1031 block_to_free_p, p);
1032 if (err)
1033 break;
1034 block_to_free = nr;
1035 block_to_free_p = p;
1036 count = 1;
1037 }
1038 }
1039 }
1040
1041 if (!err && count > 0)
1042 err = ext4_clear_blocks(handle, inode, this_bh, block_to_free,
1043 count, block_to_free_p, p);
1044 if (err < 0)
1045 /* fatal error */
1046 return;
1047
1048 if (this_bh) {
1049 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
1050
1051 /*
1052 * The buffer head should have an attached journal head at this
1053 * point. However, if the data is corrupted and an indirect
1054 * block pointed to itself, it would have been detached when
1055 * the block was cleared. Check for this instead of OOPSing.
1056 */
1057 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
1058 ext4_handle_dirty_metadata(handle, inode, this_bh);
1059 else
1060 EXT4_ERROR_INODE(inode,
1061 "circular indirect block detected at "
1062 "block %llu",
1063 (unsigned long long) this_bh->b_blocknr);
1064 }
1065}
1066
1067/**
1068 * ext4_free_branches - free an array of branches
1069 * @handle: JBD handle for this transaction
1070 * @inode: inode we are dealing with
1071 * @parent_bh: the buffer_head which contains *@first and *@last
1072 * @first: array of block numbers
1073 * @last: pointer immediately past the end of array
1074 * @depth: depth of the branches to free
1075 *
1076 * We are freeing all blocks referred from these branches (numbers are
1077 * stored as little-endian 32-bit) and updating @inode->i_blocks
1078 * appropriately.
1079 */
1080static void ext4_free_branches(handle_t *handle, struct inode *inode,
1081 struct buffer_head *parent_bh,
1082 __le32 *first, __le32 *last, int depth)
1083{
1084 ext4_fsblk_t nr;
1085 __le32 *p;
1086
1087 if (ext4_handle_is_aborted(handle))
1088 return;
1089
1090 if (depth--) {
1091 struct buffer_head *bh;
1092 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1093 p = last;
1094 while (--p >= first) {
1095 nr = le32_to_cpu(*p);
1096 if (!nr)
1097 continue; /* A hole */
1098
1099 if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
1100 nr, 1)) {
1101 EXT4_ERROR_INODE(inode,
1102 "invalid indirect mapped "
1103 "block %lu (level %d)",
1104 (unsigned long) nr, depth);
1105 break;
1106 }
1107
1108 /* Go read the buffer for the next level down */
1109 bh = sb_bread(inode->i_sb, nr);
1110
1111 /*
1112 * A read failure? Report error and clear slot
1113 * (should be rare).
1114 */
1115 if (!bh) {
1116 EXT4_ERROR_INODE_BLOCK(inode, nr,
1117 "Read failure");
1118 continue;
1119 }
1120
1121 /* This zaps the entire block. Bottom up. */
1122 BUFFER_TRACE(bh, "free child branches");
1123 ext4_free_branches(handle, inode, bh,
1124 (__le32 *) bh->b_data,
1125 (__le32 *) bh->b_data + addr_per_block,
1126 depth);
1127 brelse(bh);
1128
1129 /*
1130 * Everything below this this pointer has been
1131 * released. Now let this top-of-subtree go.
1132 *
1133 * We want the freeing of this indirect block to be
1134 * atomic in the journal with the updating of the
1135 * bitmap block which owns it. So make some room in
1136 * the journal.
1137 *
1138 * We zero the parent pointer *after* freeing its
1139 * pointee in the bitmaps, so if extend_transaction()
1140 * for some reason fails to put the bitmap changes and
1141 * the release into the same transaction, recovery
1142 * will merely complain about releasing a free block,
1143 * rather than leaking blocks.
1144 */
1145 if (ext4_handle_is_aborted(handle))
1146 return;
1147 if (try_to_extend_transaction(handle, inode)) {
1148 ext4_mark_inode_dirty(handle, inode);
1149 ext4_truncate_restart_trans(handle, inode,
1150 ext4_blocks_for_truncate(inode));
1151 }
1152
1153 /*
1154 * The forget flag here is critical because if
1155 * we are journaling (and not doing data
1156 * journaling), we have to make sure a revoke
1157 * record is written to prevent the journal
1158 * replay from overwriting the (former)
1159 * indirect block if it gets reallocated as a
1160 * data block. This must happen in the same
1161 * transaction where the data blocks are
1162 * actually freed.
1163 */
1164 ext4_free_blocks(handle, inode, NULL, nr, 1,
1165 EXT4_FREE_BLOCKS_METADATA|
1166 EXT4_FREE_BLOCKS_FORGET);
1167
1168 if (parent_bh) {
1169 /*
1170 * The block which we have just freed is
1171 * pointed to by an indirect block: journal it
1172 */
1173 BUFFER_TRACE(parent_bh, "get_write_access");
1174 if (!ext4_journal_get_write_access(handle,
1175 parent_bh)){
1176 *p = 0;
1177 BUFFER_TRACE(parent_bh,
1178 "call ext4_handle_dirty_metadata");
1179 ext4_handle_dirty_metadata(handle,
1180 inode,
1181 parent_bh);
1182 }
1183 }
1184 }
1185 } else {
1186 /* We have reached the bottom of the tree. */
1187 BUFFER_TRACE(parent_bh, "free data blocks");
1188 ext4_free_data(handle, inode, parent_bh, first, last);
1189 }
1190}
1191
1192void ext4_ind_truncate(handle_t *handle, struct inode *inode)
1193{
1194 struct ext4_inode_info *ei = EXT4_I(inode);
1195 __le32 *i_data = ei->i_data;
1196 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1197 ext4_lblk_t offsets[4];
1198 Indirect chain[4];
1199 Indirect *partial;
1200 __le32 nr = 0;
1201 int n = 0;
1202 ext4_lblk_t last_block, max_block;
1203 unsigned blocksize = inode->i_sb->s_blocksize;
1204
1205 last_block = (inode->i_size + blocksize-1)
1206 >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1207 max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1208 >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1209
1210 if (last_block != max_block) {
1211 n = ext4_block_to_path(inode, last_block, offsets, NULL);
1212 if (n == 0)
1213 return;
1214 }
1215
1216 ext4_es_remove_extent(inode, last_block, EXT_MAX_BLOCKS - last_block);
1217
1218 /*
1219 * The orphan list entry will now protect us from any crash which
1220 * occurs before the truncate completes, so it is now safe to propagate
1221 * the new, shorter inode size (held for now in i_size) into the
1222 * on-disk inode. We do this via i_disksize, which is the value which
1223 * ext4 *really* writes onto the disk inode.
1224 */
1225 ei->i_disksize = inode->i_size;
1226
1227 if (last_block == max_block) {
1228 /*
1229 * It is unnecessary to free any data blocks if last_block is
1230 * equal to the indirect block limit.
1231 */
1232 return;
1233 } else if (n == 1) { /* direct blocks */
1234 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
1235 i_data + EXT4_NDIR_BLOCKS);
1236 goto do_indirects;
1237 }
1238
1239 partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1240 /* Kill the top of shared branch (not detached) */
1241 if (nr) {
1242 if (partial == chain) {
1243 /* Shared branch grows from the inode */
1244 ext4_free_branches(handle, inode, NULL,
1245 &nr, &nr+1, (chain+n-1) - partial);
1246 *partial->p = 0;
1247 /*
1248 * We mark the inode dirty prior to restart,
1249 * and prior to stop. No need for it here.
1250 */
1251 } else {
1252 /* Shared branch grows from an indirect block */
1253 BUFFER_TRACE(partial->bh, "get_write_access");
1254 ext4_free_branches(handle, inode, partial->bh,
1255 partial->p,
1256 partial->p+1, (chain+n-1) - partial);
1257 }
1258 }
1259 /* Clear the ends of indirect blocks on the shared branch */
1260 while (partial > chain) {
1261 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
1262 (__le32*)partial->bh->b_data+addr_per_block,
1263 (chain+n-1) - partial);
1264 BUFFER_TRACE(partial->bh, "call brelse");
1265 brelse(partial->bh);
1266 partial--;
1267 }
1268do_indirects:
1269 /* Kill the remaining (whole) subtrees */
1270 switch (offsets[0]) {
1271 default:
1272 nr = i_data[EXT4_IND_BLOCK];
1273 if (nr) {
1274 ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1275 i_data[EXT4_IND_BLOCK] = 0;
1276 }
1277 case EXT4_IND_BLOCK:
1278 nr = i_data[EXT4_DIND_BLOCK];
1279 if (nr) {
1280 ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1281 i_data[EXT4_DIND_BLOCK] = 0;
1282 }
1283 case EXT4_DIND_BLOCK:
1284 nr = i_data[EXT4_TIND_BLOCK];
1285 if (nr) {
1286 ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1287 i_data[EXT4_TIND_BLOCK] = 0;
1288 }
1289 case EXT4_TIND_BLOCK:
1290 ;
1291 }
1292}
1293
1294static int free_hole_blocks(handle_t *handle, struct inode *inode,
1295 struct buffer_head *parent_bh, __le32 *i_data,
1296 int level, ext4_lblk_t first,
1297 ext4_lblk_t count, int max)
1298{
1299 struct buffer_head *bh = NULL;
1300 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1301 int ret = 0;
1302 int i, inc;
1303 ext4_lblk_t offset;
1304 __le32 blk;
1305
1306 inc = 1 << ((EXT4_BLOCK_SIZE_BITS(inode->i_sb) - 2) * level);
1307 for (i = 0, offset = 0; i < max; i++, i_data++, offset += inc) {
1308 if (offset >= count + first)
1309 break;
1310 if (*i_data == 0 || (offset + inc) <= first)
1311 continue;
1312 blk = *i_data;
1313 if (level > 0) {
1314 ext4_lblk_t first2;
1315 bh = sb_bread(inode->i_sb, le32_to_cpu(blk));
1316 if (!bh) {
1317 EXT4_ERROR_INODE_BLOCK(inode, le32_to_cpu(blk),
1318 "Read failure");
1319 return -EIO;
1320 }
1321 first2 = (first > offset) ? first - offset : 0;
1322 ret = free_hole_blocks(handle, inode, bh,
1323 (__le32 *)bh->b_data, level - 1,
1324 first2, count - offset,
1325 inode->i_sb->s_blocksize >> 2);
1326 if (ret) {
1327 brelse(bh);
1328 goto err;
1329 }
1330 }
1331 if (level == 0 ||
1332 (bh && all_zeroes((__le32 *)bh->b_data,
1333 (__le32 *)bh->b_data + addr_per_block))) {
1334 ext4_free_data(handle, inode, parent_bh, &blk, &blk+1);
1335 *i_data = 0;
1336 }
1337 brelse(bh);
1338 bh = NULL;
1339 }
1340
1341err:
1342 return ret;
1343}
1344
1345int ext4_free_hole_blocks(handle_t *handle, struct inode *inode,
1346 ext4_lblk_t first, ext4_lblk_t stop)
1347{
1348 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1349 int level, ret = 0;
1350 int num = EXT4_NDIR_BLOCKS;
1351 ext4_lblk_t count, max = EXT4_NDIR_BLOCKS;
1352 __le32 *i_data = EXT4_I(inode)->i_data;
1353
1354 count = stop - first;
1355 for (level = 0; level < 4; level++, max *= addr_per_block) {
1356 if (first < max) {
1357 ret = free_hole_blocks(handle, inode, NULL, i_data,
1358 level, first, count, num);
1359 if (ret)
1360 goto err;
1361 if (count > max - first)
1362 count -= max - first;
1363 else
1364 break;
1365 first = 0;
1366 } else {
1367 first -= max;
1368 }
1369 i_data += num;
1370 if (level == 0) {
1371 num = 1;
1372 max = 1;
1373 }
1374 }
1375
1376err:
1377 return ret;
1378}
1379