blob: 73db464cd08b7a213365c0331824ce478a69e4d1 [file] [log] [blame]
Artem Bityutskiy1e517642008-07-14 19:08:37 +03001/*
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/*
24 * This file implements the budgeting sub-system which is responsible for UBIFS
25 * space management.
26 *
27 * Factors such as compression, wasted space at the ends of LEBs, space in other
28 * journal heads, the effect of updates on the index, and so on, make it
29 * impossible to accurately predict the amount of space needed. Consequently
30 * approximations are used.
31 */
32
33#include "ubifs.h"
34#include <linux/writeback.h>
35#include <asm/div64.h>
36
37/*
38 * When pessimistic budget calculations say that there is no enough space,
39 * UBIFS starts writing back dirty inodes and pages, doing garbage collection,
40 * or committing. The below constants define maximum number of times UBIFS
41 * repeats the operations.
42 */
43#define MAX_SHRINK_RETRIES 8
44#define MAX_GC_RETRIES 4
45#define MAX_CMT_RETRIES 2
46#define MAX_NOSPC_RETRIES 1
47
48/*
49 * The below constant defines amount of dirty pages which should be written
50 * back at when trying to shrink the liability.
51 */
52#define NR_TO_WRITE 16
53
54/**
55 * struct retries_info - information about re-tries while making free space.
56 * @prev_liability: previous liability
57 * @shrink_cnt: how many times the liability was shrinked
58 * @shrink_retries: count of liability shrink re-tries (increased when
59 * liability does not shrink)
60 * @try_gc: GC should be tried first
61 * @gc_retries: how many times GC was run
62 * @cmt_retries: how many times commit has been done
63 * @nospc_retries: how many times GC returned %-ENOSPC
64 *
65 * Since we consider budgeting to be the fast-path, and this structure has to
66 * be allocated on stack and zeroed out, we make it smaller using bit-fields.
67 */
68struct retries_info {
69 long long prev_liability;
70 unsigned int shrink_cnt;
71 unsigned int shrink_retries:5;
72 unsigned int try_gc:1;
73 unsigned int gc_retries:4;
74 unsigned int cmt_retries:3;
75 unsigned int nospc_retries:1;
76};
77
78/**
79 * shrink_liability - write-back some dirty pages/inodes.
80 * @c: UBIFS file-system description object
81 * @nr_to_write: how many dirty pages to write-back
82 *
83 * This function shrinks UBIFS liability by means of writing back some amount
84 * of dirty inodes and their pages. Returns the amount of pages which were
85 * written back. The returned value does not include dirty inodes which were
86 * synchronized.
87 *
88 * Note, this function synchronizes even VFS inodes which are locked
89 * (@i_mutex) by the caller of the budgeting function, because write-back does
90 * not touch @i_mutex.
91 */
92static int shrink_liability(struct ubifs_info *c, int nr_to_write)
93{
94 int nr_written;
95 struct writeback_control wbc = {
96 .sync_mode = WB_SYNC_NONE,
97 .range_end = LLONG_MAX,
98 .nr_to_write = nr_to_write,
99 };
100
101 generic_sync_sb_inodes(c->vfs_sb, &wbc);
102 nr_written = nr_to_write - wbc.nr_to_write;
103
104 if (!nr_written) {
105 /*
106 * Re-try again but wait on pages/inodes which are being
107 * written-back concurrently (e.g., by pdflush).
108 */
109 memset(&wbc, 0, sizeof(struct writeback_control));
110 wbc.sync_mode = WB_SYNC_ALL;
111 wbc.range_end = LLONG_MAX;
112 wbc.nr_to_write = nr_to_write;
113 generic_sync_sb_inodes(c->vfs_sb, &wbc);
114 nr_written = nr_to_write - wbc.nr_to_write;
115 }
116
117 dbg_budg("%d pages were written back", nr_written);
118 return nr_written;
119}
120
121
122/**
123 * run_gc - run garbage collector.
124 * @c: UBIFS file-system description object
125 *
126 * This function runs garbage collector to make some more free space. Returns
127 * zero if a free LEB has been produced, %-EAGAIN if commit is required, and a
128 * negative error code in case of failure.
129 */
130static int run_gc(struct ubifs_info *c)
131{
132 int err, lnum;
133
134 /* Make some free space by garbage-collecting dirty space */
135 down_read(&c->commit_sem);
136 lnum = ubifs_garbage_collect(c, 1);
137 up_read(&c->commit_sem);
138 if (lnum < 0)
139 return lnum;
140
141 /* GC freed one LEB, return it to lprops */
142 dbg_budg("GC freed LEB %d", lnum);
143 err = ubifs_return_leb(c, lnum);
144 if (err)
145 return err;
146 return 0;
147}
148
149/**
150 * make_free_space - make more free space on the file-system.
151 * @c: UBIFS file-system description object
152 * @ri: information about previous invocations of this function
153 *
154 * This function is called when an operation cannot be budgeted because there
155 * is supposedly no free space. But in most cases there is some free space:
156 * o budgeting is pessimistic, so it always budgets more then it is actually
157 * needed, so shrinking the liability is one way to make free space - the
158 * cached data will take less space then it was budgeted for;
159 * o GC may turn some dark space into free space (budgeting treats dark space
160 * as not available);
161 * o commit may free some LEB, i.e., turn freeable LEBs into free LEBs.
162 *
163 * So this function tries to do the above. Returns %-EAGAIN if some free space
164 * was presumably made and the caller has to re-try budgeting the operation.
165 * Returns %-ENOSPC if it couldn't do more free space, and other negative error
166 * codes on failures.
167 */
168static int make_free_space(struct ubifs_info *c, struct retries_info *ri)
169{
170 int err;
171
172 /*
173 * If we have some dirty pages and inodes (liability), try to write
174 * them back unless this was tried too many times without effect
175 * already.
176 */
177 if (ri->shrink_retries < MAX_SHRINK_RETRIES && !ri->try_gc) {
178 long long liability;
179
180 spin_lock(&c->space_lock);
181 liability = c->budg_idx_growth + c->budg_data_growth +
182 c->budg_dd_growth;
183 spin_unlock(&c->space_lock);
184
185 if (ri->prev_liability >= liability) {
186 /* Liability does not shrink, next time try GC then */
187 ri->shrink_retries += 1;
188 if (ri->gc_retries < MAX_GC_RETRIES)
189 ri->try_gc = 1;
190 dbg_budg("liability did not shrink: retries %d of %d",
191 ri->shrink_retries, MAX_SHRINK_RETRIES);
192 }
193
194 dbg_budg("force write-back (count %d)", ri->shrink_cnt);
195 shrink_liability(c, NR_TO_WRITE + ri->shrink_cnt);
196
197 ri->prev_liability = liability;
198 ri->shrink_cnt += 1;
199 return -EAGAIN;
200 }
201
202 /*
203 * Try to run garbage collector unless it was already tried too many
204 * times.
205 */
206 if (ri->gc_retries < MAX_GC_RETRIES) {
207 ri->gc_retries += 1;
208 dbg_budg("run GC, retries %d of %d",
209 ri->gc_retries, MAX_GC_RETRIES);
210
211 ri->try_gc = 0;
212 err = run_gc(c);
213 if (!err)
214 return -EAGAIN;
215
216 if (err == -EAGAIN) {
217 dbg_budg("GC asked to commit");
218 err = ubifs_run_commit(c);
219 if (err)
220 return err;
221 return -EAGAIN;
222 }
223
224 if (err != -ENOSPC)
225 return err;
226
227 /*
228 * GC could not make any progress. If this is the first time,
229 * then it makes sense to try to commit, because it might make
230 * some dirty space.
231 */
232 dbg_budg("GC returned -ENOSPC, retries %d",
233 ri->nospc_retries);
234 if (ri->nospc_retries >= MAX_NOSPC_RETRIES)
235 return err;
236 ri->nospc_retries += 1;
237 }
238
239 /* Neither GC nor write-back helped, try to commit */
240 if (ri->cmt_retries < MAX_CMT_RETRIES) {
241 ri->cmt_retries += 1;
242 dbg_budg("run commit, retries %d of %d",
243 ri->cmt_retries, MAX_CMT_RETRIES);
244 err = ubifs_run_commit(c);
245 if (err)
246 return err;
247 return -EAGAIN;
248 }
249 return -ENOSPC;
250}
251
252/**
253 * ubifs_calc_min_idx_lebs - calculate amount of eraseblocks for the index.
254 * @c: UBIFS file-system description object
255 *
256 * This function calculates and returns the number of eraseblocks which should
257 * be kept for index usage.
258 */
259int ubifs_calc_min_idx_lebs(struct ubifs_info *c)
260{
261 int ret;
262 uint64_t idx_size;
263
264 idx_size = c->old_idx_sz + c->budg_idx_growth + c->budg_uncommitted_idx;
265
Adrian Hunter3a132522008-07-30 12:18:02 +0300266 /* And make sure we have thrice the index size of space reserved */
Artem Bityutskiyb364b412008-07-25 14:38:51 +0300267 idx_size = idx_size + (idx_size << 1);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300268
269 /*
270 * We do not maintain 'old_idx_size' as 'old_idx_lebs'/'old_idx_bytes'
271 * pair, nor similarly the two variables for the new index size, so we
272 * have to do this costly 64-bit division on fast-path.
273 */
274 if (do_div(idx_size, c->leb_size - c->max_idx_node_sz))
275 ret = idx_size + 1;
276 else
277 ret = idx_size;
278 /*
279 * The index head is not available for the in-the-gaps method, so add an
280 * extra LEB to compensate.
281 */
282 ret += 1;
283 /*
284 * At present the index needs at least 2 LEBs: one for the index head
285 * and one for in-the-gaps method (which currently does not cater for
286 * the index head and so excludes it from consideration).
287 */
288 if (ret < 2)
289 ret = 2;
290 return ret;
291}
292
293/**
294 * ubifs_calc_available - calculate available FS space.
295 * @c: UBIFS file-system description object
296 * @min_idx_lebs: minimum number of LEBs reserved for the index
297 *
298 * This function calculates and returns amount of FS space available for use.
299 */
300long long ubifs_calc_available(const struct ubifs_info *c, int min_idx_lebs)
301{
302 int subtract_lebs;
303 long long available;
304
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300305 available = c->main_bytes - c->lst.total_used;
306
307 /*
308 * Now 'available' contains theoretically available flash space
309 * assuming there is no index, so we have to subtract the space which
310 * is reserved for the index.
311 */
312 subtract_lebs = min_idx_lebs;
313
314 /* Take into account that GC reserves one LEB for its own needs */
315 subtract_lebs += 1;
316
317 /*
318 * The GC journal head LEB is not really accessible. And since
319 * different write types go to different heads, we may count only on
320 * one head's space.
321 */
322 subtract_lebs += c->jhead_cnt - 1;
323
324 /* We also reserve one LEB for deletions, which bypass budgeting */
325 subtract_lebs += 1;
326
327 available -= (long long)subtract_lebs * c->leb_size;
328
329 /* Subtract the dead space which is not available for use */
330 available -= c->lst.total_dead;
331
332 /*
333 * Subtract dark space, which might or might not be usable - it depends
334 * on the data which we have on the media and which will be written. If
335 * this is a lot of uncompressed or not-compressible data, the dark
336 * space cannot be used.
337 */
338 available -= c->lst.total_dark;
339
340 /*
341 * However, there is more dark space. The index may be bigger than
342 * @min_idx_lebs. Those extra LEBs are assumed to be available, but
343 * their dark space is not included in total_dark, so it is subtracted
344 * here.
345 */
346 if (c->lst.idx_lebs > min_idx_lebs) {
347 subtract_lebs = c->lst.idx_lebs - min_idx_lebs;
348 available -= subtract_lebs * c->dark_wm;
349 }
350
351 /* The calculations are rough and may end up with a negative number */
352 return available > 0 ? available : 0;
353}
354
355/**
356 * can_use_rp - check whether the user is allowed to use reserved pool.
357 * @c: UBIFS file-system description object
358 *
359 * UBIFS has so-called "reserved pool" which is flash space reserved
360 * for the superuser and for uses whose UID/GID is recorded in UBIFS superblock.
361 * This function checks whether current user is allowed to use reserved pool.
362 * Returns %1 current user is allowed to use reserved pool and %0 otherwise.
363 */
364static int can_use_rp(struct ubifs_info *c)
365{
366 if (current->fsuid == c->rp_uid || capable(CAP_SYS_RESOURCE) ||
367 (c->rp_gid != 0 && in_group_p(c->rp_gid)))
368 return 1;
369 return 0;
370}
371
372/**
373 * do_budget_space - reserve flash space for index and data growth.
374 * @c: UBIFS file-system description object
375 *
376 * This function makes sure UBIFS has enough free eraseblocks for index growth
377 * and data.
378 *
Adrian Hunter3a132522008-07-30 12:18:02 +0300379 * When budgeting index space, UBIFS reserves thrice as many LEBs as the index
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300380 * would take if it was consolidated and written to the flash. This guarantees
381 * that the "in-the-gaps" commit method always succeeds and UBIFS will always
382 * be able to commit dirty index. So this function basically adds amount of
Artem Bityutskiyb364b412008-07-25 14:38:51 +0300383 * budgeted index space to the size of the current index, multiplies this by 3,
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300384 * and makes sure this does not exceed the amount of free eraseblocks.
385 *
386 * Notes about @c->min_idx_lebs and @c->lst.idx_lebs variables:
387 * o @c->lst.idx_lebs is the number of LEBs the index currently uses. It might
388 * be large, because UBIFS does not do any index consolidation as long as
389 * there is free space. IOW, the index may take a lot of LEBs, but the LEBs
390 * will contain a lot of dirt.
391 * o @c->min_idx_lebs is the the index presumably takes. IOW, the index may be
392 * consolidated to take up to @c->min_idx_lebs LEBs.
393 *
394 * This function returns zero in case of success, and %-ENOSPC in case of
395 * failure.
396 */
397static int do_budget_space(struct ubifs_info *c)
398{
399 long long outstanding, available;
400 int lebs, rsvd_idx_lebs, min_idx_lebs;
401
402 /* First budget index space */
403 min_idx_lebs = ubifs_calc_min_idx_lebs(c);
404
405 /* Now 'min_idx_lebs' contains number of LEBs to reserve */
406 if (min_idx_lebs > c->lst.idx_lebs)
407 rsvd_idx_lebs = min_idx_lebs - c->lst.idx_lebs;
408 else
409 rsvd_idx_lebs = 0;
410
411 /*
412 * The number of LEBs that are available to be used by the index is:
413 *
414 * @c->lst.empty_lebs + @c->freeable_cnt + @c->idx_gc_cnt -
415 * @c->lst.taken_empty_lebs
416 *
417 * @empty_lebs are available because they are empty. @freeable_cnt are
418 * available because they contain only free and dirty space and the
419 * index allocation always occurs after wbufs are synch'ed.
420 * @idx_gc_cnt are available because they are index LEBs that have been
421 * garbage collected (including trivial GC) and are awaiting the commit
422 * before they can be unmapped - note that the in-the-gaps method will
423 * grab these if it needs them. @taken_empty_lebs are empty_lebs that
424 * have already been allocated for some purpose (also includes those
425 * LEBs on the @idx_gc list).
426 *
427 * Note, @taken_empty_lebs may temporarily be higher by one because of
428 * the way we serialize LEB allocations and budgeting. See a comment in
429 * 'ubifs_find_free_space()'.
430 */
431 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
432 c->lst.taken_empty_lebs;
433 if (unlikely(rsvd_idx_lebs > lebs)) {
434 dbg_budg("out of indexing space: min_idx_lebs %d (old %d), "
435 "rsvd_idx_lebs %d", min_idx_lebs, c->min_idx_lebs,
436 rsvd_idx_lebs);
437 return -ENOSPC;
438 }
439
440 available = ubifs_calc_available(c, min_idx_lebs);
441 outstanding = c->budg_data_growth + c->budg_dd_growth;
442
443 if (unlikely(available < outstanding)) {
444 dbg_budg("out of data space: available %lld, outstanding %lld",
445 available, outstanding);
446 return -ENOSPC;
447 }
448
449 if (available - outstanding <= c->rp_size && !can_use_rp(c))
450 return -ENOSPC;
451
452 c->min_idx_lebs = min_idx_lebs;
453 return 0;
454}
455
456/**
457 * calc_idx_growth - calculate approximate index growth from budgeting request.
458 * @c: UBIFS file-system description object
459 * @req: budgeting request
460 *
461 * For now we assume each new node adds one znode. But this is rather poor
462 * approximation, though.
463 */
464static int calc_idx_growth(const struct ubifs_info *c,
465 const struct ubifs_budget_req *req)
466{
467 int znodes;
468
469 znodes = req->new_ino + (req->new_page << UBIFS_BLOCKS_PER_PAGE_SHIFT) +
470 req->new_dent;
471 return znodes * c->max_idx_node_sz;
472}
473
474/**
475 * calc_data_growth - calculate approximate amount of new data from budgeting
476 * request.
477 * @c: UBIFS file-system description object
478 * @req: budgeting request
479 */
480static int calc_data_growth(const struct ubifs_info *c,
481 const struct ubifs_budget_req *req)
482{
483 int data_growth;
484
485 data_growth = req->new_ino ? c->inode_budget : 0;
486 if (req->new_page)
487 data_growth += c->page_budget;
488 if (req->new_dent)
489 data_growth += c->dent_budget;
490 data_growth += req->new_ino_d;
491 return data_growth;
492}
493
494/**
495 * calc_dd_growth - calculate approximate amount of data which makes other data
496 * dirty from budgeting request.
497 * @c: UBIFS file-system description object
498 * @req: budgeting request
499 */
500static int calc_dd_growth(const struct ubifs_info *c,
501 const struct ubifs_budget_req *req)
502{
503 int dd_growth;
504
505 dd_growth = req->dirtied_page ? c->page_budget : 0;
506
507 if (req->dirtied_ino)
508 dd_growth += c->inode_budget << (req->dirtied_ino - 1);
509 if (req->mod_dent)
510 dd_growth += c->dent_budget;
511 dd_growth += req->dirtied_ino_d;
512 return dd_growth;
513}
514
515/**
516 * ubifs_budget_space - ensure there is enough space to complete an operation.
517 * @c: UBIFS file-system description object
518 * @req: budget request
519 *
520 * This function allocates budget for an operation. It uses pessimistic
521 * approximation of how much flash space the operation needs. The goal of this
522 * function is to make sure UBIFS always has flash space to flush all dirty
523 * pages, dirty inodes, and dirty znodes (liability). This function may force
524 * commit, garbage-collection or write-back. Returns zero in case of success,
525 * %-ENOSPC if there is no free space and other negative error codes in case of
526 * failures.
527 */
528int ubifs_budget_space(struct ubifs_info *c, struct ubifs_budget_req *req)
529{
530 int uninitialized_var(cmt_retries), uninitialized_var(wb_retries);
531 int err, idx_growth, data_growth, dd_growth;
532 struct retries_info ri;
533
Artem Bityutskiy547000d2008-07-24 14:42:05 +0300534 ubifs_assert(req->new_page <= 1);
535 ubifs_assert(req->dirtied_page <= 1);
536 ubifs_assert(req->new_dent <= 1);
537 ubifs_assert(req->mod_dent <= 1);
538 ubifs_assert(req->new_ino <= 1);
539 ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300540 ubifs_assert(req->dirtied_ino <= 4);
541 ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300542 ubifs_assert(!(req->new_ino_d & 7));
543 ubifs_assert(!(req->dirtied_ino_d & 7));
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300544
545 data_growth = calc_data_growth(c, req);
546 dd_growth = calc_dd_growth(c, req);
547 if (!data_growth && !dd_growth)
548 return 0;
549 idx_growth = calc_idx_growth(c, req);
550 memset(&ri, 0, sizeof(struct retries_info));
551
552again:
553 spin_lock(&c->space_lock);
554 ubifs_assert(c->budg_idx_growth >= 0);
555 ubifs_assert(c->budg_data_growth >= 0);
556 ubifs_assert(c->budg_dd_growth >= 0);
557
558 if (unlikely(c->nospace) && (c->nospace_rp || !can_use_rp(c))) {
559 dbg_budg("no space");
560 spin_unlock(&c->space_lock);
561 return -ENOSPC;
562 }
563
564 c->budg_idx_growth += idx_growth;
565 c->budg_data_growth += data_growth;
566 c->budg_dd_growth += dd_growth;
567
568 err = do_budget_space(c);
569 if (likely(!err)) {
570 req->idx_growth = idx_growth;
571 req->data_growth = data_growth;
572 req->dd_growth = dd_growth;
573 spin_unlock(&c->space_lock);
574 return 0;
575 }
576
577 /* Restore the old values */
578 c->budg_idx_growth -= idx_growth;
579 c->budg_data_growth -= data_growth;
580 c->budg_dd_growth -= dd_growth;
581 spin_unlock(&c->space_lock);
582
583 if (req->fast) {
584 dbg_budg("no space for fast budgeting");
585 return err;
586 }
587
588 err = make_free_space(c, &ri);
589 if (err == -EAGAIN) {
590 dbg_budg("try again");
591 cond_resched();
592 goto again;
593 } else if (err == -ENOSPC) {
594 dbg_budg("FS is full, -ENOSPC");
595 c->nospace = 1;
596 if (can_use_rp(c) || c->rp_size == 0)
597 c->nospace_rp = 1;
598 smp_wmb();
599 } else
600 ubifs_err("cannot budget space, error %d", err);
601 return err;
602}
603
604/**
605 * ubifs_release_budget - release budgeted free space.
606 * @c: UBIFS file-system description object
607 * @req: budget request
608 *
609 * This function releases the space budgeted by 'ubifs_budget_space()'. Note,
610 * since the index changes (which were budgeted for in @req->idx_growth) will
611 * only be written to the media on commit, this function moves the index budget
612 * from @c->budg_idx_growth to @c->budg_uncommitted_idx. The latter will be
613 * zeroed by the commit operation.
614 */
615void ubifs_release_budget(struct ubifs_info *c, struct ubifs_budget_req *req)
616{
Artem Bityutskiy547000d2008-07-24 14:42:05 +0300617 ubifs_assert(req->new_page <= 1);
618 ubifs_assert(req->dirtied_page <= 1);
619 ubifs_assert(req->new_dent <= 1);
620 ubifs_assert(req->mod_dent <= 1);
621 ubifs_assert(req->new_ino <= 1);
622 ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300623 ubifs_assert(req->dirtied_ino <= 4);
624 ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300625 ubifs_assert(!(req->new_ino_d & 7));
626 ubifs_assert(!(req->dirtied_ino_d & 7));
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300627 if (!req->recalculate) {
628 ubifs_assert(req->idx_growth >= 0);
629 ubifs_assert(req->data_growth >= 0);
630 ubifs_assert(req->dd_growth >= 0);
631 }
632
633 if (req->recalculate) {
634 req->data_growth = calc_data_growth(c, req);
635 req->dd_growth = calc_dd_growth(c, req);
636 req->idx_growth = calc_idx_growth(c, req);
637 }
638
639 if (!req->data_growth && !req->dd_growth)
640 return;
641
642 c->nospace = c->nospace_rp = 0;
643 smp_wmb();
644
645 spin_lock(&c->space_lock);
646 c->budg_idx_growth -= req->idx_growth;
647 c->budg_uncommitted_idx += req->idx_growth;
648 c->budg_data_growth -= req->data_growth;
649 c->budg_dd_growth -= req->dd_growth;
650 c->min_idx_lebs = ubifs_calc_min_idx_lebs(c);
651
652 ubifs_assert(c->budg_idx_growth >= 0);
653 ubifs_assert(c->budg_data_growth >= 0);
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300654 ubifs_assert(c->budg_dd_growth >= 0);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300655 ubifs_assert(c->min_idx_lebs < c->main_lebs);
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300656 ubifs_assert(!(c->budg_idx_growth & 7));
657 ubifs_assert(!(c->budg_data_growth & 7));
658 ubifs_assert(!(c->budg_dd_growth & 7));
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300659 spin_unlock(&c->space_lock);
660}
661
662/**
663 * ubifs_convert_page_budget - convert budget of a new page.
664 * @c: UBIFS file-system description object
665 *
666 * This function converts budget which was allocated for a new page of data to
667 * the budget of changing an existing page of data. The latter is smaller then
668 * the former, so this function only does simple re-calculation and does not
669 * involve any write-back.
670 */
671void ubifs_convert_page_budget(struct ubifs_info *c)
672{
673 spin_lock(&c->space_lock);
674 /* Release the index growth reservation */
675 c->budg_idx_growth -= c->max_idx_node_sz << UBIFS_BLOCKS_PER_PAGE_SHIFT;
676 /* Release the data growth reservation */
677 c->budg_data_growth -= c->page_budget;
678 /* Increase the dirty data growth reservation instead */
679 c->budg_dd_growth += c->page_budget;
680 /* And re-calculate the indexing space reservation */
681 c->min_idx_lebs = ubifs_calc_min_idx_lebs(c);
682 spin_unlock(&c->space_lock);
683}
684
685/**
686 * ubifs_release_dirty_inode_budget - release dirty inode budget.
687 * @c: UBIFS file-system description object
688 * @ui: UBIFS inode to release the budget for
689 *
690 * This function releases budget corresponding to a dirty inode. It is usually
691 * called when after the inode has been written to the media and marked as
692 * clean.
693 */
694void ubifs_release_dirty_inode_budget(struct ubifs_info *c,
695 struct ubifs_inode *ui)
696{
Artem Bityutskiy182854b2008-07-18 18:54:29 +0300697 struct ubifs_budget_req req;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300698
Artem Bityutskiy182854b2008-07-18 18:54:29 +0300699 memset(&req, 0, sizeof(struct ubifs_budget_req));
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300700 req.dd_growth = c->inode_budget + ALIGN(ui->data_len, 8);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300701 ubifs_release_budget(c, &req);
702}
703
704/**
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300705 * ubifs_reported_space - calculate reported free space.
706 * @c: the UBIFS file-system description object
707 * @free: amount of free space
708 *
709 * This function calculates amount of free space which will be reported to
710 * user-space. User-space application tend to expect that if the file-system
711 * (e.g., via the 'statfs()' call) reports that it has N bytes available, they
712 * are able to write a file of size N. UBIFS attaches node headers to each data
713 * node and it has to write indexind nodes as well. This introduces additional
714 * overhead, and UBIFS it has to report sligtly less free space to meet the
715 * above expectetion.
716 *
717 * This function assumes free space is made up of uncompressed data nodes and
718 * full index nodes (one per data node, tripled because we always allow enough
719 * space to write the index thrice).
720 *
721 * Note, the calculation is pessimistic, which means that most of the time
722 * UBIFS reports less space than it actually has.
723 */
724long long ubifs_reported_space(const struct ubifs_info *c, uint64_t free)
725{
Artem Bityutskiyf171d4d2008-09-03 16:17:14 +0300726 int divisor, factor, f;
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300727
728 /*
729 * Reported space size is @free * X, where X is UBIFS block size
730 * divided by UBIFS block size + all overhead one data block
731 * introduces. The overhead is the node header + indexing overhead.
732 *
Artem Bityutskiyf171d4d2008-09-03 16:17:14 +0300733 * Indexing overhead calculations are based on the following formula:
734 * I = N/(f - 1) + 1, where I - number of indexing nodes, N - number
735 * of data nodes, f - fanout. Because effective UBIFS fanout is twice
736 * as less than maximum fanout, we assume that each data node
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300737 * introduces 3 * @c->max_idx_node_sz / (@c->fanout/2 - 1) bytes.
738 * Note, the multiplier 3 is because UBIFS reseves thrice as more space
739 * for the index.
740 */
Artem Bityutskiyf171d4d2008-09-03 16:17:14 +0300741 f = c->fanout > 3 ? c->fanout >> 1 : 2;
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300742 factor = UBIFS_BLOCK_SIZE;
743 divisor = UBIFS_MAX_DATA_NODE_SZ;
Artem Bityutskiyf171d4d2008-09-03 16:17:14 +0300744 divisor += (c->max_idx_node_sz * 3) / (f - 1);
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300745 free *= factor;
746 do_div(free, divisor);
747 return free;
748}
749
750/**
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300751 * ubifs_get_free_space - return amount of free space.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300752 * @c: UBIFS file-system description object
753 *
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300754 * This function calculates amount of free space to report to user-space.
755 *
756 * Because UBIFS may introduce substantial overhead (the index, node headers,
757 * alighment, wastage at the end of eraseblocks, etc), it cannot report real
758 * amount of free flash space it has (well, because not all dirty space is
759 * reclamable, UBIFS does not actually know the real amount). If UBIFS did so,
760 * it would bread user expectetion about what free space is. Users seem to
761 * accustomed to assume that if the file-system reports N bytes of free space,
762 * they would be able to fit a file of N bytes to the FS. This almost works for
763 * traditional file-systems, because they have way less overhead than UBIFS.
764 * So, to keep users happy, UBIFS tries to take the overhead into account.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300765 */
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300766long long ubifs_get_free_space(struct ubifs_info *c)
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300767{
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300768 int min_idx_lebs, rsvd_idx_lebs, lebs;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300769 long long available, outstanding, free;
770
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300771 spin_lock(&c->space_lock);
772 min_idx_lebs = ubifs_calc_min_idx_lebs(c);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300773 outstanding = c->budg_data_growth + c->budg_dd_growth;
Artem Bityutskiy9e5de352008-08-25 17:29:43 +0300774
775 /*
776 * Force the amount available to the total size reported if the used
777 * space is zero.
778 */
779 if (c->lst.total_used <= UBIFS_INO_NODE_SZ && !outstanding) {
780 spin_unlock(&c->space_lock);
781 return (long long)c->block_cnt << UBIFS_BLOCK_SHIFT;
782 }
783
784 available = ubifs_calc_available(c, min_idx_lebs);
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300785
786 /*
787 * When reporting free space to user-space, UBIFS guarantees that it is
788 * possible to write a file of free space size. This means that for
789 * empty LEBs we may use more precise calculations than
790 * 'ubifs_calc_available()' is using. Namely, we know that in empty
791 * LEBs we would waste only @c->leb_overhead bytes, not @c->dark_wm.
792 * Thus, amend the available space.
793 *
794 * Note, the calculations below are similar to what we have in
795 * 'do_budget_space()', so refer there for comments.
796 */
797 if (min_idx_lebs > c->lst.idx_lebs)
798 rsvd_idx_lebs = min_idx_lebs - c->lst.idx_lebs;
799 else
800 rsvd_idx_lebs = 0;
801 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
802 c->lst.taken_empty_lebs;
803 lebs -= rsvd_idx_lebs;
804 available += lebs * (c->dark_wm - c->leb_overhead);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300805 spin_unlock(&c->space_lock);
806
807 if (available > outstanding)
808 free = ubifs_reported_space(c, available - outstanding);
809 else
810 free = 0;
811 return free;
812}