Loading...
1/*
2 * This file is part of UBIFS.
3 *
4 * Copyright (C) 2006-2008 Nokia Corporation.
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published by
8 * the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
13 * more details.
14 *
15 * You should have received a copy of the GNU General Public License along with
16 * this program; if not, write to the Free Software Foundation, Inc., 51
17 * Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
18 *
19 * Authors: Adrian Hunter
20 * Artem Bityutskiy (Битюцкий Артём)
21 */
22
23/* This file implements TNC functions for committing */
24
25#include <linux/random.h>
26#include "ubifs.h"
27
28/**
29 * make_idx_node - make an index node for fill-the-gaps method of TNC commit.
30 * @c: UBIFS file-system description object
31 * @idx: buffer in which to place new index node
32 * @znode: znode from which to make new index node
33 * @lnum: LEB number where new index node will be written
34 * @offs: offset where new index node will be written
35 * @len: length of new index node
36 */
37static int make_idx_node(struct ubifs_info *c, struct ubifs_idx_node *idx,
38 struct ubifs_znode *znode, int lnum, int offs, int len)
39{
40 struct ubifs_znode *zp;
41 int i, err;
42
43 /* Make index node */
44 idx->ch.node_type = UBIFS_IDX_NODE;
45 idx->child_cnt = cpu_to_le16(znode->child_cnt);
46 idx->level = cpu_to_le16(znode->level);
47 for (i = 0; i < znode->child_cnt; i++) {
48 struct ubifs_branch *br = ubifs_idx_branch(c, idx, i);
49 struct ubifs_zbranch *zbr = &znode->zbranch[i];
50
51 key_write_idx(c, &zbr->key, &br->key);
52 br->lnum = cpu_to_le32(zbr->lnum);
53 br->offs = cpu_to_le32(zbr->offs);
54 br->len = cpu_to_le32(zbr->len);
55 if (!zbr->lnum || !zbr->len) {
56 ubifs_err(c, "bad ref in znode");
57 ubifs_dump_znode(c, znode);
58 if (zbr->znode)
59 ubifs_dump_znode(c, zbr->znode);
60 }
61 }
62 ubifs_prepare_node(c, idx, len, 0);
63
64 znode->lnum = lnum;
65 znode->offs = offs;
66 znode->len = len;
67
68 err = insert_old_idx_znode(c, znode);
69
70 /* Update the parent */
71 zp = znode->parent;
72 if (zp) {
73 struct ubifs_zbranch *zbr;
74
75 zbr = &zp->zbranch[znode->iip];
76 zbr->lnum = lnum;
77 zbr->offs = offs;
78 zbr->len = len;
79 } else {
80 c->zroot.lnum = lnum;
81 c->zroot.offs = offs;
82 c->zroot.len = len;
83 }
84 c->calc_idx_sz += ALIGN(len, 8);
85
86 atomic_long_dec(&c->dirty_zn_cnt);
87
88 ubifs_assert(ubifs_zn_dirty(znode));
89 ubifs_assert(ubifs_zn_cow(znode));
90
91 /*
92 * Note, unlike 'write_index()' we do not add memory barriers here
93 * because this function is called with @c->tnc_mutex locked.
94 */
95 __clear_bit(DIRTY_ZNODE, &znode->flags);
96 __clear_bit(COW_ZNODE, &znode->flags);
97
98 return err;
99}
100
101/**
102 * fill_gap - make index nodes in gaps in dirty index LEBs.
103 * @c: UBIFS file-system description object
104 * @lnum: LEB number that gap appears in
105 * @gap_start: offset of start of gap
106 * @gap_end: offset of end of gap
107 * @dirt: adds dirty space to this
108 *
109 * This function returns the number of index nodes written into the gap.
110 */
111static int fill_gap(struct ubifs_info *c, int lnum, int gap_start, int gap_end,
112 int *dirt)
113{
114 int len, gap_remains, gap_pos, written, pad_len;
115
116 ubifs_assert((gap_start & 7) == 0);
117 ubifs_assert((gap_end & 7) == 0);
118 ubifs_assert(gap_end >= gap_start);
119
120 gap_remains = gap_end - gap_start;
121 if (!gap_remains)
122 return 0;
123 gap_pos = gap_start;
124 written = 0;
125 while (c->enext) {
126 len = ubifs_idx_node_sz(c, c->enext->child_cnt);
127 if (len < gap_remains) {
128 struct ubifs_znode *znode = c->enext;
129 const int alen = ALIGN(len, 8);
130 int err;
131
132 ubifs_assert(alen <= gap_remains);
133 err = make_idx_node(c, c->ileb_buf + gap_pos, znode,
134 lnum, gap_pos, len);
135 if (err)
136 return err;
137 gap_remains -= alen;
138 gap_pos += alen;
139 c->enext = znode->cnext;
140 if (c->enext == c->cnext)
141 c->enext = NULL;
142 written += 1;
143 } else
144 break;
145 }
146 if (gap_end == c->leb_size) {
147 c->ileb_len = ALIGN(gap_pos, c->min_io_size);
148 /* Pad to end of min_io_size */
149 pad_len = c->ileb_len - gap_pos;
150 } else
151 /* Pad to end of gap */
152 pad_len = gap_remains;
153 dbg_gc("LEB %d:%d to %d len %d nodes written %d wasted bytes %d",
154 lnum, gap_start, gap_end, gap_end - gap_start, written, pad_len);
155 ubifs_pad(c, c->ileb_buf + gap_pos, pad_len);
156 *dirt += pad_len;
157 return written;
158}
159
160/**
161 * find_old_idx - find an index node obsoleted since the last commit start.
162 * @c: UBIFS file-system description object
163 * @lnum: LEB number of obsoleted index node
164 * @offs: offset of obsoleted index node
165 *
166 * Returns %1 if found and %0 otherwise.
167 */
168static int find_old_idx(struct ubifs_info *c, int lnum, int offs)
169{
170 struct ubifs_old_idx *o;
171 struct rb_node *p;
172
173 p = c->old_idx.rb_node;
174 while (p) {
175 o = rb_entry(p, struct ubifs_old_idx, rb);
176 if (lnum < o->lnum)
177 p = p->rb_left;
178 else if (lnum > o->lnum)
179 p = p->rb_right;
180 else if (offs < o->offs)
181 p = p->rb_left;
182 else if (offs > o->offs)
183 p = p->rb_right;
184 else
185 return 1;
186 }
187 return 0;
188}
189
190/**
191 * is_idx_node_in_use - determine if an index node can be overwritten.
192 * @c: UBIFS file-system description object
193 * @key: key of index node
194 * @level: index node level
195 * @lnum: LEB number of index node
196 * @offs: offset of index node
197 *
198 * If @key / @lnum / @offs identify an index node that was not part of the old
199 * index, then this function returns %0 (obsolete). Else if the index node was
200 * part of the old index but is now dirty %1 is returned, else if it is clean %2
201 * is returned. A negative error code is returned on failure.
202 */
203static int is_idx_node_in_use(struct ubifs_info *c, union ubifs_key *key,
204 int level, int lnum, int offs)
205{
206 int ret;
207
208 ret = is_idx_node_in_tnc(c, key, level, lnum, offs);
209 if (ret < 0)
210 return ret; /* Error code */
211 if (ret == 0)
212 if (find_old_idx(c, lnum, offs))
213 return 1;
214 return ret;
215}
216
217/**
218 * layout_leb_in_gaps - layout index nodes using in-the-gaps method.
219 * @c: UBIFS file-system description object
220 * @p: return LEB number here
221 *
222 * This function lays out new index nodes for dirty znodes using in-the-gaps
223 * method of TNC commit.
224 * This function merely puts the next znode into the next gap, making no attempt
225 * to try to maximise the number of znodes that fit.
226 * This function returns the number of index nodes written into the gaps, or a
227 * negative error code on failure.
228 */
229static int layout_leb_in_gaps(struct ubifs_info *c, int *p)
230{
231 struct ubifs_scan_leb *sleb;
232 struct ubifs_scan_node *snod;
233 int lnum, dirt = 0, gap_start, gap_end, err, written, tot_written;
234
235 tot_written = 0;
236 /* Get an index LEB with lots of obsolete index nodes */
237 lnum = ubifs_find_dirty_idx_leb(c);
238 if (lnum < 0)
239 /*
240 * There also may be dirt in the index head that could be
241 * filled, however we do not check there at present.
242 */
243 return lnum; /* Error code */
244 *p = lnum;
245 dbg_gc("LEB %d", lnum);
246 /*
247 * Scan the index LEB. We use the generic scan for this even though
248 * it is more comprehensive and less efficient than is needed for this
249 * purpose.
250 */
251 sleb = ubifs_scan(c, lnum, 0, c->ileb_buf, 0);
252 c->ileb_len = 0;
253 if (IS_ERR(sleb))
254 return PTR_ERR(sleb);
255 gap_start = 0;
256 list_for_each_entry(snod, &sleb->nodes, list) {
257 struct ubifs_idx_node *idx;
258 int in_use, level;
259
260 ubifs_assert(snod->type == UBIFS_IDX_NODE);
261 idx = snod->node;
262 key_read(c, ubifs_idx_key(c, idx), &snod->key);
263 level = le16_to_cpu(idx->level);
264 /* Determine if the index node is in use (not obsolete) */
265 in_use = is_idx_node_in_use(c, &snod->key, level, lnum,
266 snod->offs);
267 if (in_use < 0) {
268 ubifs_scan_destroy(sleb);
269 return in_use; /* Error code */
270 }
271 if (in_use) {
272 if (in_use == 1)
273 dirt += ALIGN(snod->len, 8);
274 /*
275 * The obsolete index nodes form gaps that can be
276 * overwritten. This gap has ended because we have
277 * found an index node that is still in use
278 * i.e. not obsolete
279 */
280 gap_end = snod->offs;
281 /* Try to fill gap */
282 written = fill_gap(c, lnum, gap_start, gap_end, &dirt);
283 if (written < 0) {
284 ubifs_scan_destroy(sleb);
285 return written; /* Error code */
286 }
287 tot_written += written;
288 gap_start = ALIGN(snod->offs + snod->len, 8);
289 }
290 }
291 ubifs_scan_destroy(sleb);
292 c->ileb_len = c->leb_size;
293 gap_end = c->leb_size;
294 /* Try to fill gap */
295 written = fill_gap(c, lnum, gap_start, gap_end, &dirt);
296 if (written < 0)
297 return written; /* Error code */
298 tot_written += written;
299 if (tot_written == 0) {
300 struct ubifs_lprops lp;
301
302 dbg_gc("LEB %d wrote %d index nodes", lnum, tot_written);
303 err = ubifs_read_one_lp(c, lnum, &lp);
304 if (err)
305 return err;
306 if (lp.free == c->leb_size) {
307 /*
308 * We must have snatched this LEB from the idx_gc list
309 * so we need to correct the free and dirty space.
310 */
311 err = ubifs_change_one_lp(c, lnum,
312 c->leb_size - c->ileb_len,
313 dirt, 0, 0, 0);
314 if (err)
315 return err;
316 }
317 return 0;
318 }
319 err = ubifs_change_one_lp(c, lnum, c->leb_size - c->ileb_len, dirt,
320 0, 0, 0);
321 if (err)
322 return err;
323 err = ubifs_leb_change(c, lnum, c->ileb_buf, c->ileb_len);
324 if (err)
325 return err;
326 dbg_gc("LEB %d wrote %d index nodes", lnum, tot_written);
327 return tot_written;
328}
329
330/**
331 * get_leb_cnt - calculate the number of empty LEBs needed to commit.
332 * @c: UBIFS file-system description object
333 * @cnt: number of znodes to commit
334 *
335 * This function returns the number of empty LEBs needed to commit @cnt znodes
336 * to the current index head. The number is not exact and may be more than
337 * needed.
338 */
339static int get_leb_cnt(struct ubifs_info *c, int cnt)
340{
341 int d;
342
343 /* Assume maximum index node size (i.e. overestimate space needed) */
344 cnt -= (c->leb_size - c->ihead_offs) / c->max_idx_node_sz;
345 if (cnt < 0)
346 cnt = 0;
347 d = c->leb_size / c->max_idx_node_sz;
348 return DIV_ROUND_UP(cnt, d);
349}
350
351/**
352 * layout_in_gaps - in-the-gaps method of committing TNC.
353 * @c: UBIFS file-system description object
354 * @cnt: number of dirty znodes to commit.
355 *
356 * This function lays out new index nodes for dirty znodes using in-the-gaps
357 * method of TNC commit.
358 *
359 * This function returns %0 on success and a negative error code on failure.
360 */
361static int layout_in_gaps(struct ubifs_info *c, int cnt)
362{
363 int err, leb_needed_cnt, written, *p;
364
365 dbg_gc("%d znodes to write", cnt);
366
367 c->gap_lebs = kmalloc(sizeof(int) * (c->lst.idx_lebs + 1), GFP_NOFS);
368 if (!c->gap_lebs)
369 return -ENOMEM;
370
371 p = c->gap_lebs;
372 do {
373 ubifs_assert(p < c->gap_lebs + sizeof(int) * c->lst.idx_lebs);
374 written = layout_leb_in_gaps(c, p);
375 if (written < 0) {
376 err = written;
377 if (err != -ENOSPC) {
378 kfree(c->gap_lebs);
379 c->gap_lebs = NULL;
380 return err;
381 }
382 if (!dbg_is_chk_index(c)) {
383 /*
384 * Do not print scary warnings if the debugging
385 * option which forces in-the-gaps is enabled.
386 */
387 ubifs_warn(c, "out of space");
388 ubifs_dump_budg(c, &c->bi);
389 ubifs_dump_lprops(c);
390 }
391 /* Try to commit anyway */
392 break;
393 }
394 p++;
395 cnt -= written;
396 leb_needed_cnt = get_leb_cnt(c, cnt);
397 dbg_gc("%d znodes remaining, need %d LEBs, have %d", cnt,
398 leb_needed_cnt, c->ileb_cnt);
399 } while (leb_needed_cnt > c->ileb_cnt);
400
401 *p = -1;
402 return 0;
403}
404
405/**
406 * layout_in_empty_space - layout index nodes in empty space.
407 * @c: UBIFS file-system description object
408 *
409 * This function lays out new index nodes for dirty znodes using empty LEBs.
410 *
411 * This function returns %0 on success and a negative error code on failure.
412 */
413static int layout_in_empty_space(struct ubifs_info *c)
414{
415 struct ubifs_znode *znode, *cnext, *zp;
416 int lnum, offs, len, next_len, buf_len, buf_offs, used, avail;
417 int wlen, blen, err;
418
419 cnext = c->enext;
420 if (!cnext)
421 return 0;
422
423 lnum = c->ihead_lnum;
424 buf_offs = c->ihead_offs;
425
426 buf_len = ubifs_idx_node_sz(c, c->fanout);
427 buf_len = ALIGN(buf_len, c->min_io_size);
428 used = 0;
429 avail = buf_len;
430
431 /* Ensure there is enough room for first write */
432 next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
433 if (buf_offs + next_len > c->leb_size)
434 lnum = -1;
435
436 while (1) {
437 znode = cnext;
438
439 len = ubifs_idx_node_sz(c, znode->child_cnt);
440
441 /* Determine the index node position */
442 if (lnum == -1) {
443 if (c->ileb_nxt >= c->ileb_cnt) {
444 ubifs_err(c, "out of space");
445 return -ENOSPC;
446 }
447 lnum = c->ilebs[c->ileb_nxt++];
448 buf_offs = 0;
449 used = 0;
450 avail = buf_len;
451 }
452
453 offs = buf_offs + used;
454
455 znode->lnum = lnum;
456 znode->offs = offs;
457 znode->len = len;
458
459 /* Update the parent */
460 zp = znode->parent;
461 if (zp) {
462 struct ubifs_zbranch *zbr;
463 int i;
464
465 i = znode->iip;
466 zbr = &zp->zbranch[i];
467 zbr->lnum = lnum;
468 zbr->offs = offs;
469 zbr->len = len;
470 } else {
471 c->zroot.lnum = lnum;
472 c->zroot.offs = offs;
473 c->zroot.len = len;
474 }
475 c->calc_idx_sz += ALIGN(len, 8);
476
477 /*
478 * Once lprops is updated, we can decrease the dirty znode count
479 * but it is easier to just do it here.
480 */
481 atomic_long_dec(&c->dirty_zn_cnt);
482
483 /*
484 * Calculate the next index node length to see if there is
485 * enough room for it
486 */
487 cnext = znode->cnext;
488 if (cnext == c->cnext)
489 next_len = 0;
490 else
491 next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
492
493 /* Update buffer positions */
494 wlen = used + len;
495 used += ALIGN(len, 8);
496 avail -= ALIGN(len, 8);
497
498 if (next_len != 0 &&
499 buf_offs + used + next_len <= c->leb_size &&
500 avail > 0)
501 continue;
502
503 if (avail <= 0 && next_len &&
504 buf_offs + used + next_len <= c->leb_size)
505 blen = buf_len;
506 else
507 blen = ALIGN(wlen, c->min_io_size);
508
509 /* The buffer is full or there are no more znodes to do */
510 buf_offs += blen;
511 if (next_len) {
512 if (buf_offs + next_len > c->leb_size) {
513 err = ubifs_update_one_lp(c, lnum,
514 c->leb_size - buf_offs, blen - used,
515 0, 0);
516 if (err)
517 return err;
518 lnum = -1;
519 }
520 used -= blen;
521 if (used < 0)
522 used = 0;
523 avail = buf_len - used;
524 continue;
525 }
526 err = ubifs_update_one_lp(c, lnum, c->leb_size - buf_offs,
527 blen - used, 0, 0);
528 if (err)
529 return err;
530 break;
531 }
532
533 c->dbg->new_ihead_lnum = lnum;
534 c->dbg->new_ihead_offs = buf_offs;
535
536 return 0;
537}
538
539/**
540 * layout_commit - determine positions of index nodes to commit.
541 * @c: UBIFS file-system description object
542 * @no_space: indicates that insufficient empty LEBs were allocated
543 * @cnt: number of znodes to commit
544 *
545 * Calculate and update the positions of index nodes to commit. If there were
546 * an insufficient number of empty LEBs allocated, then index nodes are placed
547 * into the gaps created by obsolete index nodes in non-empty index LEBs. For
548 * this purpose, an obsolete index node is one that was not in the index as at
549 * the end of the last commit. To write "in-the-gaps" requires that those index
550 * LEBs are updated atomically in-place.
551 */
552static int layout_commit(struct ubifs_info *c, int no_space, int cnt)
553{
554 int err;
555
556 if (no_space) {
557 err = layout_in_gaps(c, cnt);
558 if (err)
559 return err;
560 }
561 err = layout_in_empty_space(c);
562 return err;
563}
564
565/**
566 * find_first_dirty - find first dirty znode.
567 * @znode: znode to begin searching from
568 */
569static struct ubifs_znode *find_first_dirty(struct ubifs_znode *znode)
570{
571 int i, cont;
572
573 if (!znode)
574 return NULL;
575
576 while (1) {
577 if (znode->level == 0) {
578 if (ubifs_zn_dirty(znode))
579 return znode;
580 return NULL;
581 }
582 cont = 0;
583 for (i = 0; i < znode->child_cnt; i++) {
584 struct ubifs_zbranch *zbr = &znode->zbranch[i];
585
586 if (zbr->znode && ubifs_zn_dirty(zbr->znode)) {
587 znode = zbr->znode;
588 cont = 1;
589 break;
590 }
591 }
592 if (!cont) {
593 if (ubifs_zn_dirty(znode))
594 return znode;
595 return NULL;
596 }
597 }
598}
599
600/**
601 * find_next_dirty - find next dirty znode.
602 * @znode: znode to begin searching from
603 */
604static struct ubifs_znode *find_next_dirty(struct ubifs_znode *znode)
605{
606 int n = znode->iip + 1;
607
608 znode = znode->parent;
609 if (!znode)
610 return NULL;
611 for (; n < znode->child_cnt; n++) {
612 struct ubifs_zbranch *zbr = &znode->zbranch[n];
613
614 if (zbr->znode && ubifs_zn_dirty(zbr->znode))
615 return find_first_dirty(zbr->znode);
616 }
617 return znode;
618}
619
620/**
621 * get_znodes_to_commit - create list of dirty znodes to commit.
622 * @c: UBIFS file-system description object
623 *
624 * This function returns the number of znodes to commit.
625 */
626static int get_znodes_to_commit(struct ubifs_info *c)
627{
628 struct ubifs_znode *znode, *cnext;
629 int cnt = 0;
630
631 c->cnext = find_first_dirty(c->zroot.znode);
632 znode = c->enext = c->cnext;
633 if (!znode) {
634 dbg_cmt("no znodes to commit");
635 return 0;
636 }
637 cnt += 1;
638 while (1) {
639 ubifs_assert(!ubifs_zn_cow(znode));
640 __set_bit(COW_ZNODE, &znode->flags);
641 znode->alt = 0;
642 cnext = find_next_dirty(znode);
643 if (!cnext) {
644 znode->cnext = c->cnext;
645 break;
646 }
647 znode->cnext = cnext;
648 znode = cnext;
649 cnt += 1;
650 }
651 dbg_cmt("committing %d znodes", cnt);
652 ubifs_assert(cnt == atomic_long_read(&c->dirty_zn_cnt));
653 return cnt;
654}
655
656/**
657 * alloc_idx_lebs - allocate empty LEBs to be used to commit.
658 * @c: UBIFS file-system description object
659 * @cnt: number of znodes to commit
660 *
661 * This function returns %-ENOSPC if it cannot allocate a sufficient number of
662 * empty LEBs. %0 is returned on success, otherwise a negative error code
663 * is returned.
664 */
665static int alloc_idx_lebs(struct ubifs_info *c, int cnt)
666{
667 int i, leb_cnt, lnum;
668
669 c->ileb_cnt = 0;
670 c->ileb_nxt = 0;
671 leb_cnt = get_leb_cnt(c, cnt);
672 dbg_cmt("need about %d empty LEBS for TNC commit", leb_cnt);
673 if (!leb_cnt)
674 return 0;
675 c->ilebs = kmalloc(leb_cnt * sizeof(int), GFP_NOFS);
676 if (!c->ilebs)
677 return -ENOMEM;
678 for (i = 0; i < leb_cnt; i++) {
679 lnum = ubifs_find_free_leb_for_idx(c);
680 if (lnum < 0)
681 return lnum;
682 c->ilebs[c->ileb_cnt++] = lnum;
683 dbg_cmt("LEB %d", lnum);
684 }
685 if (dbg_is_chk_index(c) && !(prandom_u32() & 7))
686 return -ENOSPC;
687 return 0;
688}
689
690/**
691 * free_unused_idx_lebs - free unused LEBs that were allocated for the commit.
692 * @c: UBIFS file-system description object
693 *
694 * It is possible that we allocate more empty LEBs for the commit than we need.
695 * This functions frees the surplus.
696 *
697 * This function returns %0 on success and a negative error code on failure.
698 */
699static int free_unused_idx_lebs(struct ubifs_info *c)
700{
701 int i, err = 0, lnum, er;
702
703 for (i = c->ileb_nxt; i < c->ileb_cnt; i++) {
704 lnum = c->ilebs[i];
705 dbg_cmt("LEB %d", lnum);
706 er = ubifs_change_one_lp(c, lnum, LPROPS_NC, LPROPS_NC, 0,
707 LPROPS_INDEX | LPROPS_TAKEN, 0);
708 if (!err)
709 err = er;
710 }
711 return err;
712}
713
714/**
715 * free_idx_lebs - free unused LEBs after commit end.
716 * @c: UBIFS file-system description object
717 *
718 * This function returns %0 on success and a negative error code on failure.
719 */
720static int free_idx_lebs(struct ubifs_info *c)
721{
722 int err;
723
724 err = free_unused_idx_lebs(c);
725 kfree(c->ilebs);
726 c->ilebs = NULL;
727 return err;
728}
729
730/**
731 * ubifs_tnc_start_commit - start TNC commit.
732 * @c: UBIFS file-system description object
733 * @zroot: new index root position is returned here
734 *
735 * This function prepares the list of indexing nodes to commit and lays out
736 * their positions on flash. If there is not enough free space it uses the
737 * in-gap commit method. Returns zero in case of success and a negative error
738 * code in case of failure.
739 */
740int ubifs_tnc_start_commit(struct ubifs_info *c, struct ubifs_zbranch *zroot)
741{
742 int err = 0, cnt;
743
744 mutex_lock(&c->tnc_mutex);
745 err = dbg_check_tnc(c, 1);
746 if (err)
747 goto out;
748 cnt = get_znodes_to_commit(c);
749 if (cnt != 0) {
750 int no_space = 0;
751
752 err = alloc_idx_lebs(c, cnt);
753 if (err == -ENOSPC)
754 no_space = 1;
755 else if (err)
756 goto out_free;
757 err = layout_commit(c, no_space, cnt);
758 if (err)
759 goto out_free;
760 ubifs_assert(atomic_long_read(&c->dirty_zn_cnt) == 0);
761 err = free_unused_idx_lebs(c);
762 if (err)
763 goto out;
764 }
765 destroy_old_idx(c);
766 memcpy(zroot, &c->zroot, sizeof(struct ubifs_zbranch));
767
768 err = ubifs_save_dirty_idx_lnums(c);
769 if (err)
770 goto out;
771
772 spin_lock(&c->space_lock);
773 /*
774 * Although we have not finished committing yet, update size of the
775 * committed index ('c->bi.old_idx_sz') and zero out the index growth
776 * budget. It is OK to do this now, because we've reserved all the
777 * space which is needed to commit the index, and it is save for the
778 * budgeting subsystem to assume the index is already committed,
779 * even though it is not.
780 */
781 ubifs_assert(c->bi.min_idx_lebs == ubifs_calc_min_idx_lebs(c));
782 c->bi.old_idx_sz = c->calc_idx_sz;
783 c->bi.uncommitted_idx = 0;
784 c->bi.min_idx_lebs = ubifs_calc_min_idx_lebs(c);
785 spin_unlock(&c->space_lock);
786 mutex_unlock(&c->tnc_mutex);
787
788 dbg_cmt("number of index LEBs %d", c->lst.idx_lebs);
789 dbg_cmt("size of index %llu", c->calc_idx_sz);
790 return err;
791
792out_free:
793 free_idx_lebs(c);
794out:
795 mutex_unlock(&c->tnc_mutex);
796 return err;
797}
798
799/**
800 * write_index - write index nodes.
801 * @c: UBIFS file-system description object
802 *
803 * This function writes the index nodes whose positions were laid out in the
804 * layout_in_empty_space function.
805 */
806static int write_index(struct ubifs_info *c)
807{
808 struct ubifs_idx_node *idx;
809 struct ubifs_znode *znode, *cnext;
810 int i, lnum, offs, len, next_len, buf_len, buf_offs, used;
811 int avail, wlen, err, lnum_pos = 0, blen, nxt_offs;
812
813 cnext = c->enext;
814 if (!cnext)
815 return 0;
816
817 /*
818 * Always write index nodes to the index head so that index nodes and
819 * other types of nodes are never mixed in the same erase block.
820 */
821 lnum = c->ihead_lnum;
822 buf_offs = c->ihead_offs;
823
824 /* Allocate commit buffer */
825 buf_len = ALIGN(c->max_idx_node_sz, c->min_io_size);
826 used = 0;
827 avail = buf_len;
828
829 /* Ensure there is enough room for first write */
830 next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
831 if (buf_offs + next_len > c->leb_size) {
832 err = ubifs_update_one_lp(c, lnum, LPROPS_NC, 0, 0,
833 LPROPS_TAKEN);
834 if (err)
835 return err;
836 lnum = -1;
837 }
838
839 while (1) {
840 cond_resched();
841
842 znode = cnext;
843 idx = c->cbuf + used;
844
845 /* Make index node */
846 idx->ch.node_type = UBIFS_IDX_NODE;
847 idx->child_cnt = cpu_to_le16(znode->child_cnt);
848 idx->level = cpu_to_le16(znode->level);
849 for (i = 0; i < znode->child_cnt; i++) {
850 struct ubifs_branch *br = ubifs_idx_branch(c, idx, i);
851 struct ubifs_zbranch *zbr = &znode->zbranch[i];
852
853 key_write_idx(c, &zbr->key, &br->key);
854 br->lnum = cpu_to_le32(zbr->lnum);
855 br->offs = cpu_to_le32(zbr->offs);
856 br->len = cpu_to_le32(zbr->len);
857 if (!zbr->lnum || !zbr->len) {
858 ubifs_err(c, "bad ref in znode");
859 ubifs_dump_znode(c, znode);
860 if (zbr->znode)
861 ubifs_dump_znode(c, zbr->znode);
862 }
863 }
864 len = ubifs_idx_node_sz(c, znode->child_cnt);
865 ubifs_prepare_node(c, idx, len, 0);
866
867 /* Determine the index node position */
868 if (lnum == -1) {
869 lnum = c->ilebs[lnum_pos++];
870 buf_offs = 0;
871 used = 0;
872 avail = buf_len;
873 }
874 offs = buf_offs + used;
875
876 if (lnum != znode->lnum || offs != znode->offs ||
877 len != znode->len) {
878 ubifs_err(c, "inconsistent znode posn");
879 return -EINVAL;
880 }
881
882 /* Grab some stuff from znode while we still can */
883 cnext = znode->cnext;
884
885 ubifs_assert(ubifs_zn_dirty(znode));
886 ubifs_assert(ubifs_zn_cow(znode));
887
888 /*
889 * It is important that other threads should see %DIRTY_ZNODE
890 * flag cleared before %COW_ZNODE. Specifically, it matters in
891 * the 'dirty_cow_znode()' function. This is the reason for the
892 * first barrier. Also, we want the bit changes to be seen to
893 * other threads ASAP, to avoid unnecesarry copying, which is
894 * the reason for the second barrier.
895 */
896 clear_bit(DIRTY_ZNODE, &znode->flags);
897 smp_mb__before_atomic();
898 clear_bit(COW_ZNODE, &znode->flags);
899 smp_mb__after_atomic();
900
901 /*
902 * We have marked the znode as clean but have not updated the
903 * @c->clean_zn_cnt counter. If this znode becomes dirty again
904 * before 'free_obsolete_znodes()' is called, then
905 * @c->clean_zn_cnt will be decremented before it gets
906 * incremented (resulting in 2 decrements for the same znode).
907 * This means that @c->clean_zn_cnt may become negative for a
908 * while.
909 *
910 * Q: why we cannot increment @c->clean_zn_cnt?
911 * A: because we do not have the @c->tnc_mutex locked, and the
912 * following code would be racy and buggy:
913 *
914 * if (!ubifs_zn_obsolete(znode)) {
915 * atomic_long_inc(&c->clean_zn_cnt);
916 * atomic_long_inc(&ubifs_clean_zn_cnt);
917 * }
918 *
919 * Thus, we just delay the @c->clean_zn_cnt update until we
920 * have the mutex locked.
921 */
922
923 /* Do not access znode from this point on */
924
925 /* Update buffer positions */
926 wlen = used + len;
927 used += ALIGN(len, 8);
928 avail -= ALIGN(len, 8);
929
930 /*
931 * Calculate the next index node length to see if there is
932 * enough room for it
933 */
934 if (cnext == c->cnext)
935 next_len = 0;
936 else
937 next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
938
939 nxt_offs = buf_offs + used + next_len;
940 if (next_len && nxt_offs <= c->leb_size) {
941 if (avail > 0)
942 continue;
943 else
944 blen = buf_len;
945 } else {
946 wlen = ALIGN(wlen, 8);
947 blen = ALIGN(wlen, c->min_io_size);
948 ubifs_pad(c, c->cbuf + wlen, blen - wlen);
949 }
950
951 /* The buffer is full or there are no more znodes to do */
952 err = ubifs_leb_write(c, lnum, c->cbuf, buf_offs, blen);
953 if (err)
954 return err;
955 buf_offs += blen;
956 if (next_len) {
957 if (nxt_offs > c->leb_size) {
958 err = ubifs_update_one_lp(c, lnum, LPROPS_NC, 0,
959 0, LPROPS_TAKEN);
960 if (err)
961 return err;
962 lnum = -1;
963 }
964 used -= blen;
965 if (used < 0)
966 used = 0;
967 avail = buf_len - used;
968 memmove(c->cbuf, c->cbuf + blen, used);
969 continue;
970 }
971 break;
972 }
973
974 if (lnum != c->dbg->new_ihead_lnum ||
975 buf_offs != c->dbg->new_ihead_offs) {
976 ubifs_err(c, "inconsistent ihead");
977 return -EINVAL;
978 }
979
980 c->ihead_lnum = lnum;
981 c->ihead_offs = buf_offs;
982
983 return 0;
984}
985
986/**
987 * free_obsolete_znodes - free obsolete znodes.
988 * @c: UBIFS file-system description object
989 *
990 * At the end of commit end, obsolete znodes are freed.
991 */
992static void free_obsolete_znodes(struct ubifs_info *c)
993{
994 struct ubifs_znode *znode, *cnext;
995
996 cnext = c->cnext;
997 do {
998 znode = cnext;
999 cnext = znode->cnext;
1000 if (ubifs_zn_obsolete(znode))
1001 kfree(znode);
1002 else {
1003 znode->cnext = NULL;
1004 atomic_long_inc(&c->clean_zn_cnt);
1005 atomic_long_inc(&ubifs_clean_zn_cnt);
1006 }
1007 } while (cnext != c->cnext);
1008}
1009
1010/**
1011 * return_gap_lebs - return LEBs used by the in-gap commit method.
1012 * @c: UBIFS file-system description object
1013 *
1014 * This function clears the "taken" flag for the LEBs which were used by the
1015 * "commit in-the-gaps" method.
1016 */
1017static int return_gap_lebs(struct ubifs_info *c)
1018{
1019 int *p, err;
1020
1021 if (!c->gap_lebs)
1022 return 0;
1023
1024 dbg_cmt("");
1025 for (p = c->gap_lebs; *p != -1; p++) {
1026 err = ubifs_change_one_lp(c, *p, LPROPS_NC, LPROPS_NC, 0,
1027 LPROPS_TAKEN, 0);
1028 if (err)
1029 return err;
1030 }
1031
1032 kfree(c->gap_lebs);
1033 c->gap_lebs = NULL;
1034 return 0;
1035}
1036
1037/**
1038 * ubifs_tnc_end_commit - update the TNC for commit end.
1039 * @c: UBIFS file-system description object
1040 *
1041 * Write the dirty znodes.
1042 */
1043int ubifs_tnc_end_commit(struct ubifs_info *c)
1044{
1045 int err;
1046
1047 if (!c->cnext)
1048 return 0;
1049
1050 err = return_gap_lebs(c);
1051 if (err)
1052 return err;
1053
1054 err = write_index(c);
1055 if (err)
1056 return err;
1057
1058 mutex_lock(&c->tnc_mutex);
1059
1060 dbg_cmt("TNC height is %d", c->zroot.znode->level + 1);
1061
1062 free_obsolete_znodes(c);
1063
1064 c->cnext = NULL;
1065 kfree(c->ilebs);
1066 c->ilebs = NULL;
1067
1068 mutex_unlock(&c->tnc_mutex);
1069
1070 return 0;
1071}
1/*
2 * This file is part of UBIFS.
3 *
4 * Copyright (C) 2006-2008 Nokia Corporation.
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published by
8 * the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
13 * more details.
14 *
15 * You should have received a copy of the GNU General Public License along with
16 * this program; if not, write to the Free Software Foundation, Inc., 51
17 * Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
18 *
19 * Authors: Adrian Hunter
20 * Artem Bityutskiy (Битюцкий Артём)
21 */
22
23/* This file implements TNC functions for committing */
24
25#include <linux/random.h>
26#include "ubifs.h"
27
28/**
29 * make_idx_node - make an index node for fill-the-gaps method of TNC commit.
30 * @c: UBIFS file-system description object
31 * @idx: buffer in which to place new index node
32 * @znode: znode from which to make new index node
33 * @lnum: LEB number where new index node will be written
34 * @offs: offset where new index node will be written
35 * @len: length of new index node
36 */
37static int make_idx_node(struct ubifs_info *c, struct ubifs_idx_node *idx,
38 struct ubifs_znode *znode, int lnum, int offs, int len)
39{
40 struct ubifs_znode *zp;
41 int i, err;
42
43 /* Make index node */
44 idx->ch.node_type = UBIFS_IDX_NODE;
45 idx->child_cnt = cpu_to_le16(znode->child_cnt);
46 idx->level = cpu_to_le16(znode->level);
47 for (i = 0; i < znode->child_cnt; i++) {
48 struct ubifs_branch *br = ubifs_idx_branch(c, idx, i);
49 struct ubifs_zbranch *zbr = &znode->zbranch[i];
50
51 key_write_idx(c, &zbr->key, &br->key);
52 br->lnum = cpu_to_le32(zbr->lnum);
53 br->offs = cpu_to_le32(zbr->offs);
54 br->len = cpu_to_le32(zbr->len);
55 if (!zbr->lnum || !zbr->len) {
56 ubifs_err("bad ref in znode");
57 dbg_dump_znode(c, znode);
58 if (zbr->znode)
59 dbg_dump_znode(c, zbr->znode);
60 }
61 }
62 ubifs_prepare_node(c, idx, len, 0);
63
64#ifdef CONFIG_UBIFS_FS_DEBUG
65 znode->lnum = lnum;
66 znode->offs = offs;
67 znode->len = len;
68#endif
69
70 err = insert_old_idx_znode(c, znode);
71
72 /* Update the parent */
73 zp = znode->parent;
74 if (zp) {
75 struct ubifs_zbranch *zbr;
76
77 zbr = &zp->zbranch[znode->iip];
78 zbr->lnum = lnum;
79 zbr->offs = offs;
80 zbr->len = len;
81 } else {
82 c->zroot.lnum = lnum;
83 c->zroot.offs = offs;
84 c->zroot.len = len;
85 }
86 c->calc_idx_sz += ALIGN(len, 8);
87
88 atomic_long_dec(&c->dirty_zn_cnt);
89
90 ubifs_assert(ubifs_zn_dirty(znode));
91 ubifs_assert(ubifs_zn_cow(znode));
92
93 /*
94 * Note, unlike 'write_index()' we do not add memory barriers here
95 * because this function is called with @c->tnc_mutex locked.
96 */
97 __clear_bit(DIRTY_ZNODE, &znode->flags);
98 __clear_bit(COW_ZNODE, &znode->flags);
99
100 return err;
101}
102
103/**
104 * fill_gap - make index nodes in gaps in dirty index LEBs.
105 * @c: UBIFS file-system description object
106 * @lnum: LEB number that gap appears in
107 * @gap_start: offset of start of gap
108 * @gap_end: offset of end of gap
109 * @dirt: adds dirty space to this
110 *
111 * This function returns the number of index nodes written into the gap.
112 */
113static int fill_gap(struct ubifs_info *c, int lnum, int gap_start, int gap_end,
114 int *dirt)
115{
116 int len, gap_remains, gap_pos, written, pad_len;
117
118 ubifs_assert((gap_start & 7) == 0);
119 ubifs_assert((gap_end & 7) == 0);
120 ubifs_assert(gap_end >= gap_start);
121
122 gap_remains = gap_end - gap_start;
123 if (!gap_remains)
124 return 0;
125 gap_pos = gap_start;
126 written = 0;
127 while (c->enext) {
128 len = ubifs_idx_node_sz(c, c->enext->child_cnt);
129 if (len < gap_remains) {
130 struct ubifs_znode *znode = c->enext;
131 const int alen = ALIGN(len, 8);
132 int err;
133
134 ubifs_assert(alen <= gap_remains);
135 err = make_idx_node(c, c->ileb_buf + gap_pos, znode,
136 lnum, gap_pos, len);
137 if (err)
138 return err;
139 gap_remains -= alen;
140 gap_pos += alen;
141 c->enext = znode->cnext;
142 if (c->enext == c->cnext)
143 c->enext = NULL;
144 written += 1;
145 } else
146 break;
147 }
148 if (gap_end == c->leb_size) {
149 c->ileb_len = ALIGN(gap_pos, c->min_io_size);
150 /* Pad to end of min_io_size */
151 pad_len = c->ileb_len - gap_pos;
152 } else
153 /* Pad to end of gap */
154 pad_len = gap_remains;
155 dbg_gc("LEB %d:%d to %d len %d nodes written %d wasted bytes %d",
156 lnum, gap_start, gap_end, gap_end - gap_start, written, pad_len);
157 ubifs_pad(c, c->ileb_buf + gap_pos, pad_len);
158 *dirt += pad_len;
159 return written;
160}
161
162/**
163 * find_old_idx - find an index node obsoleted since the last commit start.
164 * @c: UBIFS file-system description object
165 * @lnum: LEB number of obsoleted index node
166 * @offs: offset of obsoleted index node
167 *
168 * Returns %1 if found and %0 otherwise.
169 */
170static int find_old_idx(struct ubifs_info *c, int lnum, int offs)
171{
172 struct ubifs_old_idx *o;
173 struct rb_node *p;
174
175 p = c->old_idx.rb_node;
176 while (p) {
177 o = rb_entry(p, struct ubifs_old_idx, rb);
178 if (lnum < o->lnum)
179 p = p->rb_left;
180 else if (lnum > o->lnum)
181 p = p->rb_right;
182 else if (offs < o->offs)
183 p = p->rb_left;
184 else if (offs > o->offs)
185 p = p->rb_right;
186 else
187 return 1;
188 }
189 return 0;
190}
191
192/**
193 * is_idx_node_in_use - determine if an index node can be overwritten.
194 * @c: UBIFS file-system description object
195 * @key: key of index node
196 * @level: index node level
197 * @lnum: LEB number of index node
198 * @offs: offset of index node
199 *
200 * If @key / @lnum / @offs identify an index node that was not part of the old
201 * index, then this function returns %0 (obsolete). Else if the index node was
202 * part of the old index but is now dirty %1 is returned, else if it is clean %2
203 * is returned. A negative error code is returned on failure.
204 */
205static int is_idx_node_in_use(struct ubifs_info *c, union ubifs_key *key,
206 int level, int lnum, int offs)
207{
208 int ret;
209
210 ret = is_idx_node_in_tnc(c, key, level, lnum, offs);
211 if (ret < 0)
212 return ret; /* Error code */
213 if (ret == 0)
214 if (find_old_idx(c, lnum, offs))
215 return 1;
216 return ret;
217}
218
219/**
220 * layout_leb_in_gaps - layout index nodes using in-the-gaps method.
221 * @c: UBIFS file-system description object
222 * @p: return LEB number here
223 *
224 * This function lays out new index nodes for dirty znodes using in-the-gaps
225 * method of TNC commit.
226 * This function merely puts the next znode into the next gap, making no attempt
227 * to try to maximise the number of znodes that fit.
228 * This function returns the number of index nodes written into the gaps, or a
229 * negative error code on failure.
230 */
231static int layout_leb_in_gaps(struct ubifs_info *c, int *p)
232{
233 struct ubifs_scan_leb *sleb;
234 struct ubifs_scan_node *snod;
235 int lnum, dirt = 0, gap_start, gap_end, err, written, tot_written;
236
237 tot_written = 0;
238 /* Get an index LEB with lots of obsolete index nodes */
239 lnum = ubifs_find_dirty_idx_leb(c);
240 if (lnum < 0)
241 /*
242 * There also may be dirt in the index head that could be
243 * filled, however we do not check there at present.
244 */
245 return lnum; /* Error code */
246 *p = lnum;
247 dbg_gc("LEB %d", lnum);
248 /*
249 * Scan the index LEB. We use the generic scan for this even though
250 * it is more comprehensive and less efficient than is needed for this
251 * purpose.
252 */
253 sleb = ubifs_scan(c, lnum, 0, c->ileb_buf, 0);
254 c->ileb_len = 0;
255 if (IS_ERR(sleb))
256 return PTR_ERR(sleb);
257 gap_start = 0;
258 list_for_each_entry(snod, &sleb->nodes, list) {
259 struct ubifs_idx_node *idx;
260 int in_use, level;
261
262 ubifs_assert(snod->type == UBIFS_IDX_NODE);
263 idx = snod->node;
264 key_read(c, ubifs_idx_key(c, idx), &snod->key);
265 level = le16_to_cpu(idx->level);
266 /* Determine if the index node is in use (not obsolete) */
267 in_use = is_idx_node_in_use(c, &snod->key, level, lnum,
268 snod->offs);
269 if (in_use < 0) {
270 ubifs_scan_destroy(sleb);
271 return in_use; /* Error code */
272 }
273 if (in_use) {
274 if (in_use == 1)
275 dirt += ALIGN(snod->len, 8);
276 /*
277 * The obsolete index nodes form gaps that can be
278 * overwritten. This gap has ended because we have
279 * found an index node that is still in use
280 * i.e. not obsolete
281 */
282 gap_end = snod->offs;
283 /* Try to fill gap */
284 written = fill_gap(c, lnum, gap_start, gap_end, &dirt);
285 if (written < 0) {
286 ubifs_scan_destroy(sleb);
287 return written; /* Error code */
288 }
289 tot_written += written;
290 gap_start = ALIGN(snod->offs + snod->len, 8);
291 }
292 }
293 ubifs_scan_destroy(sleb);
294 c->ileb_len = c->leb_size;
295 gap_end = c->leb_size;
296 /* Try to fill gap */
297 written = fill_gap(c, lnum, gap_start, gap_end, &dirt);
298 if (written < 0)
299 return written; /* Error code */
300 tot_written += written;
301 if (tot_written == 0) {
302 struct ubifs_lprops lp;
303
304 dbg_gc("LEB %d wrote %d index nodes", lnum, tot_written);
305 err = ubifs_read_one_lp(c, lnum, &lp);
306 if (err)
307 return err;
308 if (lp.free == c->leb_size) {
309 /*
310 * We must have snatched this LEB from the idx_gc list
311 * so we need to correct the free and dirty space.
312 */
313 err = ubifs_change_one_lp(c, lnum,
314 c->leb_size - c->ileb_len,
315 dirt, 0, 0, 0);
316 if (err)
317 return err;
318 }
319 return 0;
320 }
321 err = ubifs_change_one_lp(c, lnum, c->leb_size - c->ileb_len, dirt,
322 0, 0, 0);
323 if (err)
324 return err;
325 err = ubifs_leb_change(c, lnum, c->ileb_buf, c->ileb_len,
326 UBI_SHORTTERM);
327 if (err)
328 return err;
329 dbg_gc("LEB %d wrote %d index nodes", lnum, tot_written);
330 return tot_written;
331}
332
333/**
334 * get_leb_cnt - calculate the number of empty LEBs needed to commit.
335 * @c: UBIFS file-system description object
336 * @cnt: number of znodes to commit
337 *
338 * This function returns the number of empty LEBs needed to commit @cnt znodes
339 * to the current index head. The number is not exact and may be more than
340 * needed.
341 */
342static int get_leb_cnt(struct ubifs_info *c, int cnt)
343{
344 int d;
345
346 /* Assume maximum index node size (i.e. overestimate space needed) */
347 cnt -= (c->leb_size - c->ihead_offs) / c->max_idx_node_sz;
348 if (cnt < 0)
349 cnt = 0;
350 d = c->leb_size / c->max_idx_node_sz;
351 return DIV_ROUND_UP(cnt, d);
352}
353
354/**
355 * layout_in_gaps - in-the-gaps method of committing TNC.
356 * @c: UBIFS file-system description object
357 * @cnt: number of dirty znodes to commit.
358 *
359 * This function lays out new index nodes for dirty znodes using in-the-gaps
360 * method of TNC commit.
361 *
362 * This function returns %0 on success and a negative error code on failure.
363 */
364static int layout_in_gaps(struct ubifs_info *c, int cnt)
365{
366 int err, leb_needed_cnt, written, *p;
367
368 dbg_gc("%d znodes to write", cnt);
369
370 c->gap_lebs = kmalloc(sizeof(int) * (c->lst.idx_lebs + 1), GFP_NOFS);
371 if (!c->gap_lebs)
372 return -ENOMEM;
373
374 p = c->gap_lebs;
375 do {
376 ubifs_assert(p < c->gap_lebs + sizeof(int) * c->lst.idx_lebs);
377 written = layout_leb_in_gaps(c, p);
378 if (written < 0) {
379 err = written;
380 if (err != -ENOSPC) {
381 kfree(c->gap_lebs);
382 c->gap_lebs = NULL;
383 return err;
384 }
385 if (!dbg_is_chk_index(c)) {
386 /*
387 * Do not print scary warnings if the debugging
388 * option which forces in-the-gaps is enabled.
389 */
390 ubifs_warn("out of space");
391 dbg_dump_budg(c, &c->bi);
392 dbg_dump_lprops(c);
393 }
394 /* Try to commit anyway */
395 err = 0;
396 break;
397 }
398 p++;
399 cnt -= written;
400 leb_needed_cnt = get_leb_cnt(c, cnt);
401 dbg_gc("%d znodes remaining, need %d LEBs, have %d", cnt,
402 leb_needed_cnt, c->ileb_cnt);
403 } while (leb_needed_cnt > c->ileb_cnt);
404
405 *p = -1;
406 return 0;
407}
408
409/**
410 * layout_in_empty_space - layout index nodes in empty space.
411 * @c: UBIFS file-system description object
412 *
413 * This function lays out new index nodes for dirty znodes using empty LEBs.
414 *
415 * This function returns %0 on success and a negative error code on failure.
416 */
417static int layout_in_empty_space(struct ubifs_info *c)
418{
419 struct ubifs_znode *znode, *cnext, *zp;
420 int lnum, offs, len, next_len, buf_len, buf_offs, used, avail;
421 int wlen, blen, err;
422
423 cnext = c->enext;
424 if (!cnext)
425 return 0;
426
427 lnum = c->ihead_lnum;
428 buf_offs = c->ihead_offs;
429
430 buf_len = ubifs_idx_node_sz(c, c->fanout);
431 buf_len = ALIGN(buf_len, c->min_io_size);
432 used = 0;
433 avail = buf_len;
434
435 /* Ensure there is enough room for first write */
436 next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
437 if (buf_offs + next_len > c->leb_size)
438 lnum = -1;
439
440 while (1) {
441 znode = cnext;
442
443 len = ubifs_idx_node_sz(c, znode->child_cnt);
444
445 /* Determine the index node position */
446 if (lnum == -1) {
447 if (c->ileb_nxt >= c->ileb_cnt) {
448 ubifs_err("out of space");
449 return -ENOSPC;
450 }
451 lnum = c->ilebs[c->ileb_nxt++];
452 buf_offs = 0;
453 used = 0;
454 avail = buf_len;
455 }
456
457 offs = buf_offs + used;
458
459#ifdef CONFIG_UBIFS_FS_DEBUG
460 znode->lnum = lnum;
461 znode->offs = offs;
462 znode->len = len;
463#endif
464
465 /* Update the parent */
466 zp = znode->parent;
467 if (zp) {
468 struct ubifs_zbranch *zbr;
469 int i;
470
471 i = znode->iip;
472 zbr = &zp->zbranch[i];
473 zbr->lnum = lnum;
474 zbr->offs = offs;
475 zbr->len = len;
476 } else {
477 c->zroot.lnum = lnum;
478 c->zroot.offs = offs;
479 c->zroot.len = len;
480 }
481 c->calc_idx_sz += ALIGN(len, 8);
482
483 /*
484 * Once lprops is updated, we can decrease the dirty znode count
485 * but it is easier to just do it here.
486 */
487 atomic_long_dec(&c->dirty_zn_cnt);
488
489 /*
490 * Calculate the next index node length to see if there is
491 * enough room for it
492 */
493 cnext = znode->cnext;
494 if (cnext == c->cnext)
495 next_len = 0;
496 else
497 next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
498
499 /* Update buffer positions */
500 wlen = used + len;
501 used += ALIGN(len, 8);
502 avail -= ALIGN(len, 8);
503
504 if (next_len != 0 &&
505 buf_offs + used + next_len <= c->leb_size &&
506 avail > 0)
507 continue;
508
509 if (avail <= 0 && next_len &&
510 buf_offs + used + next_len <= c->leb_size)
511 blen = buf_len;
512 else
513 blen = ALIGN(wlen, c->min_io_size);
514
515 /* The buffer is full or there are no more znodes to do */
516 buf_offs += blen;
517 if (next_len) {
518 if (buf_offs + next_len > c->leb_size) {
519 err = ubifs_update_one_lp(c, lnum,
520 c->leb_size - buf_offs, blen - used,
521 0, 0);
522 if (err)
523 return err;
524 lnum = -1;
525 }
526 used -= blen;
527 if (used < 0)
528 used = 0;
529 avail = buf_len - used;
530 continue;
531 }
532 err = ubifs_update_one_lp(c, lnum, c->leb_size - buf_offs,
533 blen - used, 0, 0);
534 if (err)
535 return err;
536 break;
537 }
538
539#ifdef CONFIG_UBIFS_FS_DEBUG
540 c->dbg->new_ihead_lnum = lnum;
541 c->dbg->new_ihead_offs = buf_offs;
542#endif
543
544 return 0;
545}
546
547/**
548 * layout_commit - determine positions of index nodes to commit.
549 * @c: UBIFS file-system description object
550 * @no_space: indicates that insufficient empty LEBs were allocated
551 * @cnt: number of znodes to commit
552 *
553 * Calculate and update the positions of index nodes to commit. If there were
554 * an insufficient number of empty LEBs allocated, then index nodes are placed
555 * into the gaps created by obsolete index nodes in non-empty index LEBs. For
556 * this purpose, an obsolete index node is one that was not in the index as at
557 * the end of the last commit. To write "in-the-gaps" requires that those index
558 * LEBs are updated atomically in-place.
559 */
560static int layout_commit(struct ubifs_info *c, int no_space, int cnt)
561{
562 int err;
563
564 if (no_space) {
565 err = layout_in_gaps(c, cnt);
566 if (err)
567 return err;
568 }
569 err = layout_in_empty_space(c);
570 return err;
571}
572
573/**
574 * find_first_dirty - find first dirty znode.
575 * @znode: znode to begin searching from
576 */
577static struct ubifs_znode *find_first_dirty(struct ubifs_znode *znode)
578{
579 int i, cont;
580
581 if (!znode)
582 return NULL;
583
584 while (1) {
585 if (znode->level == 0) {
586 if (ubifs_zn_dirty(znode))
587 return znode;
588 return NULL;
589 }
590 cont = 0;
591 for (i = 0; i < znode->child_cnt; i++) {
592 struct ubifs_zbranch *zbr = &znode->zbranch[i];
593
594 if (zbr->znode && ubifs_zn_dirty(zbr->znode)) {
595 znode = zbr->znode;
596 cont = 1;
597 break;
598 }
599 }
600 if (!cont) {
601 if (ubifs_zn_dirty(znode))
602 return znode;
603 return NULL;
604 }
605 }
606}
607
608/**
609 * find_next_dirty - find next dirty znode.
610 * @znode: znode to begin searching from
611 */
612static struct ubifs_znode *find_next_dirty(struct ubifs_znode *znode)
613{
614 int n = znode->iip + 1;
615
616 znode = znode->parent;
617 if (!znode)
618 return NULL;
619 for (; n < znode->child_cnt; n++) {
620 struct ubifs_zbranch *zbr = &znode->zbranch[n];
621
622 if (zbr->znode && ubifs_zn_dirty(zbr->znode))
623 return find_first_dirty(zbr->znode);
624 }
625 return znode;
626}
627
628/**
629 * get_znodes_to_commit - create list of dirty znodes to commit.
630 * @c: UBIFS file-system description object
631 *
632 * This function returns the number of znodes to commit.
633 */
634static int get_znodes_to_commit(struct ubifs_info *c)
635{
636 struct ubifs_znode *znode, *cnext;
637 int cnt = 0;
638
639 c->cnext = find_first_dirty(c->zroot.znode);
640 znode = c->enext = c->cnext;
641 if (!znode) {
642 dbg_cmt("no znodes to commit");
643 return 0;
644 }
645 cnt += 1;
646 while (1) {
647 ubifs_assert(!ubifs_zn_cow(znode));
648 __set_bit(COW_ZNODE, &znode->flags);
649 znode->alt = 0;
650 cnext = find_next_dirty(znode);
651 if (!cnext) {
652 znode->cnext = c->cnext;
653 break;
654 }
655 znode->cnext = cnext;
656 znode = cnext;
657 cnt += 1;
658 }
659 dbg_cmt("committing %d znodes", cnt);
660 ubifs_assert(cnt == atomic_long_read(&c->dirty_zn_cnt));
661 return cnt;
662}
663
664/**
665 * alloc_idx_lebs - allocate empty LEBs to be used to commit.
666 * @c: UBIFS file-system description object
667 * @cnt: number of znodes to commit
668 *
669 * This function returns %-ENOSPC if it cannot allocate a sufficient number of
670 * empty LEBs. %0 is returned on success, otherwise a negative error code
671 * is returned.
672 */
673static int alloc_idx_lebs(struct ubifs_info *c, int cnt)
674{
675 int i, leb_cnt, lnum;
676
677 c->ileb_cnt = 0;
678 c->ileb_nxt = 0;
679 leb_cnt = get_leb_cnt(c, cnt);
680 dbg_cmt("need about %d empty LEBS for TNC commit", leb_cnt);
681 if (!leb_cnt)
682 return 0;
683 c->ilebs = kmalloc(leb_cnt * sizeof(int), GFP_NOFS);
684 if (!c->ilebs)
685 return -ENOMEM;
686 for (i = 0; i < leb_cnt; i++) {
687 lnum = ubifs_find_free_leb_for_idx(c);
688 if (lnum < 0)
689 return lnum;
690 c->ilebs[c->ileb_cnt++] = lnum;
691 dbg_cmt("LEB %d", lnum);
692 }
693 if (dbg_is_chk_index(c) && !(random32() & 7))
694 return -ENOSPC;
695 return 0;
696}
697
698/**
699 * free_unused_idx_lebs - free unused LEBs that were allocated for the commit.
700 * @c: UBIFS file-system description object
701 *
702 * It is possible that we allocate more empty LEBs for the commit than we need.
703 * This functions frees the surplus.
704 *
705 * This function returns %0 on success and a negative error code on failure.
706 */
707static int free_unused_idx_lebs(struct ubifs_info *c)
708{
709 int i, err = 0, lnum, er;
710
711 for (i = c->ileb_nxt; i < c->ileb_cnt; i++) {
712 lnum = c->ilebs[i];
713 dbg_cmt("LEB %d", lnum);
714 er = ubifs_change_one_lp(c, lnum, LPROPS_NC, LPROPS_NC, 0,
715 LPROPS_INDEX | LPROPS_TAKEN, 0);
716 if (!err)
717 err = er;
718 }
719 return err;
720}
721
722/**
723 * free_idx_lebs - free unused LEBs after commit end.
724 * @c: UBIFS file-system description object
725 *
726 * This function returns %0 on success and a negative error code on failure.
727 */
728static int free_idx_lebs(struct ubifs_info *c)
729{
730 int err;
731
732 err = free_unused_idx_lebs(c);
733 kfree(c->ilebs);
734 c->ilebs = NULL;
735 return err;
736}
737
738/**
739 * ubifs_tnc_start_commit - start TNC commit.
740 * @c: UBIFS file-system description object
741 * @zroot: new index root position is returned here
742 *
743 * This function prepares the list of indexing nodes to commit and lays out
744 * their positions on flash. If there is not enough free space it uses the
745 * in-gap commit method. Returns zero in case of success and a negative error
746 * code in case of failure.
747 */
748int ubifs_tnc_start_commit(struct ubifs_info *c, struct ubifs_zbranch *zroot)
749{
750 int err = 0, cnt;
751
752 mutex_lock(&c->tnc_mutex);
753 err = dbg_check_tnc(c, 1);
754 if (err)
755 goto out;
756 cnt = get_znodes_to_commit(c);
757 if (cnt != 0) {
758 int no_space = 0;
759
760 err = alloc_idx_lebs(c, cnt);
761 if (err == -ENOSPC)
762 no_space = 1;
763 else if (err)
764 goto out_free;
765 err = layout_commit(c, no_space, cnt);
766 if (err)
767 goto out_free;
768 ubifs_assert(atomic_long_read(&c->dirty_zn_cnt) == 0);
769 err = free_unused_idx_lebs(c);
770 if (err)
771 goto out;
772 }
773 destroy_old_idx(c);
774 memcpy(zroot, &c->zroot, sizeof(struct ubifs_zbranch));
775
776 err = ubifs_save_dirty_idx_lnums(c);
777 if (err)
778 goto out;
779
780 spin_lock(&c->space_lock);
781 /*
782 * Although we have not finished committing yet, update size of the
783 * committed index ('c->bi.old_idx_sz') and zero out the index growth
784 * budget. It is OK to do this now, because we've reserved all the
785 * space which is needed to commit the index, and it is save for the
786 * budgeting subsystem to assume the index is already committed,
787 * even though it is not.
788 */
789 ubifs_assert(c->bi.min_idx_lebs == ubifs_calc_min_idx_lebs(c));
790 c->bi.old_idx_sz = c->calc_idx_sz;
791 c->bi.uncommitted_idx = 0;
792 c->bi.min_idx_lebs = ubifs_calc_min_idx_lebs(c);
793 spin_unlock(&c->space_lock);
794 mutex_unlock(&c->tnc_mutex);
795
796 dbg_cmt("number of index LEBs %d", c->lst.idx_lebs);
797 dbg_cmt("size of index %llu", c->calc_idx_sz);
798 return err;
799
800out_free:
801 free_idx_lebs(c);
802out:
803 mutex_unlock(&c->tnc_mutex);
804 return err;
805}
806
807/**
808 * write_index - write index nodes.
809 * @c: UBIFS file-system description object
810 *
811 * This function writes the index nodes whose positions were laid out in the
812 * layout_in_empty_space function.
813 */
814static int write_index(struct ubifs_info *c)
815{
816 struct ubifs_idx_node *idx;
817 struct ubifs_znode *znode, *cnext;
818 int i, lnum, offs, len, next_len, buf_len, buf_offs, used;
819 int avail, wlen, err, lnum_pos = 0, blen, nxt_offs;
820
821 cnext = c->enext;
822 if (!cnext)
823 return 0;
824
825 /*
826 * Always write index nodes to the index head so that index nodes and
827 * other types of nodes are never mixed in the same erase block.
828 */
829 lnum = c->ihead_lnum;
830 buf_offs = c->ihead_offs;
831
832 /* Allocate commit buffer */
833 buf_len = ALIGN(c->max_idx_node_sz, c->min_io_size);
834 used = 0;
835 avail = buf_len;
836
837 /* Ensure there is enough room for first write */
838 next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
839 if (buf_offs + next_len > c->leb_size) {
840 err = ubifs_update_one_lp(c, lnum, LPROPS_NC, 0, 0,
841 LPROPS_TAKEN);
842 if (err)
843 return err;
844 lnum = -1;
845 }
846
847 while (1) {
848 cond_resched();
849
850 znode = cnext;
851 idx = c->cbuf + used;
852
853 /* Make index node */
854 idx->ch.node_type = UBIFS_IDX_NODE;
855 idx->child_cnt = cpu_to_le16(znode->child_cnt);
856 idx->level = cpu_to_le16(znode->level);
857 for (i = 0; i < znode->child_cnt; i++) {
858 struct ubifs_branch *br = ubifs_idx_branch(c, idx, i);
859 struct ubifs_zbranch *zbr = &znode->zbranch[i];
860
861 key_write_idx(c, &zbr->key, &br->key);
862 br->lnum = cpu_to_le32(zbr->lnum);
863 br->offs = cpu_to_le32(zbr->offs);
864 br->len = cpu_to_le32(zbr->len);
865 if (!zbr->lnum || !zbr->len) {
866 ubifs_err("bad ref in znode");
867 dbg_dump_znode(c, znode);
868 if (zbr->znode)
869 dbg_dump_znode(c, zbr->znode);
870 }
871 }
872 len = ubifs_idx_node_sz(c, znode->child_cnt);
873 ubifs_prepare_node(c, idx, len, 0);
874
875 /* Determine the index node position */
876 if (lnum == -1) {
877 lnum = c->ilebs[lnum_pos++];
878 buf_offs = 0;
879 used = 0;
880 avail = buf_len;
881 }
882 offs = buf_offs + used;
883
884#ifdef CONFIG_UBIFS_FS_DEBUG
885 if (lnum != znode->lnum || offs != znode->offs ||
886 len != znode->len) {
887 ubifs_err("inconsistent znode posn");
888 return -EINVAL;
889 }
890#endif
891
892 /* Grab some stuff from znode while we still can */
893 cnext = znode->cnext;
894
895 ubifs_assert(ubifs_zn_dirty(znode));
896 ubifs_assert(ubifs_zn_cow(znode));
897
898 /*
899 * It is important that other threads should see %DIRTY_ZNODE
900 * flag cleared before %COW_ZNODE. Specifically, it matters in
901 * the 'dirty_cow_znode()' function. This is the reason for the
902 * first barrier. Also, we want the bit changes to be seen to
903 * other threads ASAP, to avoid unnecesarry copying, which is
904 * the reason for the second barrier.
905 */
906 clear_bit(DIRTY_ZNODE, &znode->flags);
907 smp_mb__before_clear_bit();
908 clear_bit(COW_ZNODE, &znode->flags);
909 smp_mb__after_clear_bit();
910
911 /*
912 * We have marked the znode as clean but have not updated the
913 * @c->clean_zn_cnt counter. If this znode becomes dirty again
914 * before 'free_obsolete_znodes()' is called, then
915 * @c->clean_zn_cnt will be decremented before it gets
916 * incremented (resulting in 2 decrements for the same znode).
917 * This means that @c->clean_zn_cnt may become negative for a
918 * while.
919 *
920 * Q: why we cannot increment @c->clean_zn_cnt?
921 * A: because we do not have the @c->tnc_mutex locked, and the
922 * following code would be racy and buggy:
923 *
924 * if (!ubifs_zn_obsolete(znode)) {
925 * atomic_long_inc(&c->clean_zn_cnt);
926 * atomic_long_inc(&ubifs_clean_zn_cnt);
927 * }
928 *
929 * Thus, we just delay the @c->clean_zn_cnt update until we
930 * have the mutex locked.
931 */
932
933 /* Do not access znode from this point on */
934
935 /* Update buffer positions */
936 wlen = used + len;
937 used += ALIGN(len, 8);
938 avail -= ALIGN(len, 8);
939
940 /*
941 * Calculate the next index node length to see if there is
942 * enough room for it
943 */
944 if (cnext == c->cnext)
945 next_len = 0;
946 else
947 next_len = ubifs_idx_node_sz(c, cnext->child_cnt);
948
949 nxt_offs = buf_offs + used + next_len;
950 if (next_len && nxt_offs <= c->leb_size) {
951 if (avail > 0)
952 continue;
953 else
954 blen = buf_len;
955 } else {
956 wlen = ALIGN(wlen, 8);
957 blen = ALIGN(wlen, c->min_io_size);
958 ubifs_pad(c, c->cbuf + wlen, blen - wlen);
959 }
960
961 /* The buffer is full or there are no more znodes to do */
962 err = ubifs_leb_write(c, lnum, c->cbuf, buf_offs, blen,
963 UBI_SHORTTERM);
964 if (err)
965 return err;
966 buf_offs += blen;
967 if (next_len) {
968 if (nxt_offs > c->leb_size) {
969 err = ubifs_update_one_lp(c, lnum, LPROPS_NC, 0,
970 0, LPROPS_TAKEN);
971 if (err)
972 return err;
973 lnum = -1;
974 }
975 used -= blen;
976 if (used < 0)
977 used = 0;
978 avail = buf_len - used;
979 memmove(c->cbuf, c->cbuf + blen, used);
980 continue;
981 }
982 break;
983 }
984
985#ifdef CONFIG_UBIFS_FS_DEBUG
986 if (lnum != c->dbg->new_ihead_lnum ||
987 buf_offs != c->dbg->new_ihead_offs) {
988 ubifs_err("inconsistent ihead");
989 return -EINVAL;
990 }
991#endif
992
993 c->ihead_lnum = lnum;
994 c->ihead_offs = buf_offs;
995
996 return 0;
997}
998
999/**
1000 * free_obsolete_znodes - free obsolete znodes.
1001 * @c: UBIFS file-system description object
1002 *
1003 * At the end of commit end, obsolete znodes are freed.
1004 */
1005static void free_obsolete_znodes(struct ubifs_info *c)
1006{
1007 struct ubifs_znode *znode, *cnext;
1008
1009 cnext = c->cnext;
1010 do {
1011 znode = cnext;
1012 cnext = znode->cnext;
1013 if (ubifs_zn_obsolete(znode))
1014 kfree(znode);
1015 else {
1016 znode->cnext = NULL;
1017 atomic_long_inc(&c->clean_zn_cnt);
1018 atomic_long_inc(&ubifs_clean_zn_cnt);
1019 }
1020 } while (cnext != c->cnext);
1021}
1022
1023/**
1024 * return_gap_lebs - return LEBs used by the in-gap commit method.
1025 * @c: UBIFS file-system description object
1026 *
1027 * This function clears the "taken" flag for the LEBs which were used by the
1028 * "commit in-the-gaps" method.
1029 */
1030static int return_gap_lebs(struct ubifs_info *c)
1031{
1032 int *p, err;
1033
1034 if (!c->gap_lebs)
1035 return 0;
1036
1037 dbg_cmt("");
1038 for (p = c->gap_lebs; *p != -1; p++) {
1039 err = ubifs_change_one_lp(c, *p, LPROPS_NC, LPROPS_NC, 0,
1040 LPROPS_TAKEN, 0);
1041 if (err)
1042 return err;
1043 }
1044
1045 kfree(c->gap_lebs);
1046 c->gap_lebs = NULL;
1047 return 0;
1048}
1049
1050/**
1051 * ubifs_tnc_end_commit - update the TNC for commit end.
1052 * @c: UBIFS file-system description object
1053 *
1054 * Write the dirty znodes.
1055 */
1056int ubifs_tnc_end_commit(struct ubifs_info *c)
1057{
1058 int err;
1059
1060 if (!c->cnext)
1061 return 0;
1062
1063 err = return_gap_lebs(c);
1064 if (err)
1065 return err;
1066
1067 err = write_index(c);
1068 if (err)
1069 return err;
1070
1071 mutex_lock(&c->tnc_mutex);
1072
1073 dbg_cmt("TNC height is %d", c->zroot.znode->level + 1);
1074
1075 free_obsolete_znodes(c);
1076
1077 c->cnext = NULL;
1078 kfree(c->ilebs);
1079 c->ilebs = NULL;
1080
1081 mutex_unlock(&c->tnc_mutex);
1082
1083 return 0;
1084}