Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * This file is part of UBIFS.
4 *
5 * Copyright (C) 2006-2008 Nokia Corporation.
6 *
7 * Authors: Adrian Hunter
8 * Artem Bityutskiy (Битюцкий Артём)
9 */
10
11/*
12 * This file implements the budgeting sub-system which is responsible for UBIFS
13 * space management.
14 *
15 * Factors such as compression, wasted space at the ends of LEBs, space in other
16 * journal heads, the effect of updates on the index, and so on, make it
17 * impossible to accurately predict the amount of space needed. Consequently
18 * approximations are used.
19 */
20
21#include "ubifs.h"
22#include <linux/writeback.h>
23#include <linux/math64.h>
24
25/*
26 * When pessimistic budget calculations say that there is no enough space,
27 * UBIFS starts writing back dirty inodes and pages, doing garbage collection,
28 * or committing. The below constant defines maximum number of times UBIFS
29 * repeats the operations.
30 */
31#define MAX_MKSPC_RETRIES 3
32
33/*
34 * The below constant defines amount of dirty pages which should be written
35 * back at when trying to shrink the liability.
36 */
37#define NR_TO_WRITE 16
38
39/**
40 * shrink_liability - write-back some dirty pages/inodes.
41 * @c: UBIFS file-system description object
42 * @nr_to_write: how many dirty pages to write-back
43 *
44 * This function shrinks UBIFS liability by means of writing back some amount
45 * of dirty inodes and their pages.
46 *
47 * Note, this function synchronizes even VFS inodes which are locked
48 * (@i_mutex) by the caller of the budgeting function, because write-back does
49 * not touch @i_mutex.
50 */
51static void shrink_liability(struct ubifs_info *c, int nr_to_write)
52{
53 down_read(&c->vfs_sb->s_umount);
54 writeback_inodes_sb_nr(c->vfs_sb, nr_to_write, WB_REASON_FS_FREE_SPACE);
55 up_read(&c->vfs_sb->s_umount);
56}
57
58/**
59 * run_gc - run garbage collector.
60 * @c: UBIFS file-system description object
61 *
62 * This function runs garbage collector to make some more free space. Returns
63 * zero if a free LEB has been produced, %-EAGAIN if commit is required, and a
64 * negative error code in case of failure.
65 */
66static int run_gc(struct ubifs_info *c)
67{
68 int err, lnum;
69
70 /* Make some free space by garbage-collecting dirty space */
71 down_read(&c->commit_sem);
72 lnum = ubifs_garbage_collect(c, 1);
73 up_read(&c->commit_sem);
74 if (lnum < 0)
75 return lnum;
76
77 /* GC freed one LEB, return it to lprops */
78 dbg_budg("GC freed LEB %d", lnum);
79 err = ubifs_return_leb(c, lnum);
80 if (err)
81 return err;
82 return 0;
83}
84
85/**
86 * get_liability - calculate current liability.
87 * @c: UBIFS file-system description object
88 *
89 * This function calculates and returns current UBIFS liability, i.e. the
90 * amount of bytes UBIFS has "promised" to write to the media.
91 */
92static long long get_liability(struct ubifs_info *c)
93{
94 long long liab;
95
96 spin_lock(&c->space_lock);
97 liab = c->bi.idx_growth + c->bi.data_growth + c->bi.dd_growth;
98 spin_unlock(&c->space_lock);
99 return liab;
100}
101
102/**
103 * make_free_space - make more free space on the file-system.
104 * @c: UBIFS file-system description object
105 *
106 * This function is called when an operation cannot be budgeted because there
107 * is supposedly no free space. But in most cases there is some free space:
108 * o budgeting is pessimistic, so it always budgets more than it is actually
109 * needed, so shrinking the liability is one way to make free space - the
110 * cached data will take less space then it was budgeted for;
111 * o GC may turn some dark space into free space (budgeting treats dark space
112 * as not available);
113 * o commit may free some LEB, i.e., turn freeable LEBs into free LEBs.
114 *
115 * So this function tries to do the above. Returns %-EAGAIN if some free space
116 * was presumably made and the caller has to re-try budgeting the operation.
117 * Returns %-ENOSPC if it couldn't do more free space, and other negative error
118 * codes on failures.
119 */
120static int make_free_space(struct ubifs_info *c)
121{
122 int err, retries = 0;
123 long long liab1, liab2;
124
125 do {
126 liab1 = get_liability(c);
127 /*
128 * We probably have some dirty pages or inodes (liability), try
129 * to write them back.
130 */
131 dbg_budg("liability %lld, run write-back", liab1);
132 shrink_liability(c, NR_TO_WRITE);
133
134 liab2 = get_liability(c);
135 if (liab2 < liab1)
136 return -EAGAIN;
137
138 dbg_budg("new liability %lld (not shrunk)", liab2);
139
140 /* Liability did not shrink again, try GC */
141 dbg_budg("Run GC");
142 err = run_gc(c);
143 if (!err)
144 return -EAGAIN;
145
146 if (err != -EAGAIN && err != -ENOSPC)
147 /* Some real error happened */
148 return err;
149
150 dbg_budg("Run commit (retries %d)", retries);
151 err = ubifs_run_commit(c);
152 if (err)
153 return err;
154 } while (retries++ < MAX_MKSPC_RETRIES);
155
156 return -ENOSPC;
157}
158
159/**
160 * ubifs_calc_min_idx_lebs - calculate amount of LEBs for the index.
161 * @c: UBIFS file-system description object
162 *
163 * This function calculates and returns the number of LEBs which should be kept
164 * for index usage.
165 */
166int ubifs_calc_min_idx_lebs(struct ubifs_info *c)
167{
168 int idx_lebs;
169 long long idx_size;
170
171 idx_size = c->bi.old_idx_sz + c->bi.idx_growth + c->bi.uncommitted_idx;
172 /* And make sure we have thrice the index size of space reserved */
173 idx_size += idx_size << 1;
174 /*
175 * We do not maintain 'old_idx_size' as 'old_idx_lebs'/'old_idx_bytes'
176 * pair, nor similarly the two variables for the new index size, so we
177 * have to do this costly 64-bit division on fast-path.
178 */
179 idx_lebs = div_u64(idx_size + c->idx_leb_size - 1, c->idx_leb_size);
180 /*
181 * The index head is not available for the in-the-gaps method, so add an
182 * extra LEB to compensate.
183 */
184 idx_lebs += 1;
185 if (idx_lebs < MIN_INDEX_LEBS)
186 idx_lebs = MIN_INDEX_LEBS;
187 return idx_lebs;
188}
189
190/**
191 * ubifs_calc_available - calculate available FS space.
192 * @c: UBIFS file-system description object
193 * @min_idx_lebs: minimum number of LEBs reserved for the index
194 *
195 * This function calculates and returns amount of FS space available for use.
196 */
197long long ubifs_calc_available(const struct ubifs_info *c, int min_idx_lebs)
198{
199 int subtract_lebs;
200 long long available;
201
202 available = c->main_bytes - c->lst.total_used;
203
204 /*
205 * Now 'available' contains theoretically available flash space
206 * assuming there is no index, so we have to subtract the space which
207 * is reserved for the index.
208 */
209 subtract_lebs = min_idx_lebs;
210
211 /* Take into account that GC reserves one LEB for its own needs */
212 subtract_lebs += 1;
213
214 /*
215 * The GC journal head LEB is not really accessible. And since
216 * different write types go to different heads, we may count only on
217 * one head's space.
218 */
219 subtract_lebs += c->jhead_cnt - 1;
220
221 /* We also reserve one LEB for deletions, which bypass budgeting */
222 subtract_lebs += 1;
223
224 available -= (long long)subtract_lebs * c->leb_size;
225
226 /* Subtract the dead space which is not available for use */
227 available -= c->lst.total_dead;
228
229 /*
230 * Subtract dark space, which might or might not be usable - it depends
231 * on the data which we have on the media and which will be written. If
232 * this is a lot of uncompressed or not-compressible data, the dark
233 * space cannot be used.
234 */
235 available -= c->lst.total_dark;
236
237 /*
238 * However, there is more dark space. The index may be bigger than
239 * @min_idx_lebs. Those extra LEBs are assumed to be available, but
240 * their dark space is not included in total_dark, so it is subtracted
241 * here.
242 */
243 if (c->lst.idx_lebs > min_idx_lebs) {
244 subtract_lebs = c->lst.idx_lebs - min_idx_lebs;
245 available -= subtract_lebs * c->dark_wm;
246 }
247
248 /* The calculations are rough and may end up with a negative number */
249 return available > 0 ? available : 0;
250}
251
252/**
253 * can_use_rp - check whether the user is allowed to use reserved pool.
254 * @c: UBIFS file-system description object
255 *
256 * UBIFS has so-called "reserved pool" which is flash space reserved
257 * for the superuser and for uses whose UID/GID is recorded in UBIFS superblock.
258 * This function checks whether current user is allowed to use reserved pool.
259 * Returns %1 current user is allowed to use reserved pool and %0 otherwise.
260 */
261static int can_use_rp(struct ubifs_info *c)
262{
263 if (uid_eq(current_fsuid(), c->rp_uid) || capable(CAP_SYS_RESOURCE) ||
264 (!gid_eq(c->rp_gid, GLOBAL_ROOT_GID) && in_group_p(c->rp_gid)))
265 return 1;
266 return 0;
267}
268
269/**
270 * do_budget_space - reserve flash space for index and data growth.
271 * @c: UBIFS file-system description object
272 *
273 * This function makes sure UBIFS has enough free LEBs for index growth and
274 * data.
275 *
276 * When budgeting index space, UBIFS reserves thrice as many LEBs as the index
277 * would take if it was consolidated and written to the flash. This guarantees
278 * that the "in-the-gaps" commit method always succeeds and UBIFS will always
279 * be able to commit dirty index. So this function basically adds amount of
280 * budgeted index space to the size of the current index, multiplies this by 3,
281 * and makes sure this does not exceed the amount of free LEBs.
282 *
283 * Notes about @c->bi.min_idx_lebs and @c->lst.idx_lebs variables:
284 * o @c->lst.idx_lebs is the number of LEBs the index currently uses. It might
285 * be large, because UBIFS does not do any index consolidation as long as
286 * there is free space. IOW, the index may take a lot of LEBs, but the LEBs
287 * will contain a lot of dirt.
288 * o @c->bi.min_idx_lebs is the number of LEBS the index presumably takes. IOW,
289 * the index may be consolidated to take up to @c->bi.min_idx_lebs LEBs.
290 *
291 * This function returns zero in case of success, and %-ENOSPC in case of
292 * failure.
293 */
294static int do_budget_space(struct ubifs_info *c)
295{
296 long long outstanding, available;
297 int lebs, rsvd_idx_lebs, min_idx_lebs;
298
299 /* First budget index space */
300 min_idx_lebs = ubifs_calc_min_idx_lebs(c);
301
302 /* Now 'min_idx_lebs' contains number of LEBs to reserve */
303 if (min_idx_lebs > c->lst.idx_lebs)
304 rsvd_idx_lebs = min_idx_lebs - c->lst.idx_lebs;
305 else
306 rsvd_idx_lebs = 0;
307
308 /*
309 * The number of LEBs that are available to be used by the index is:
310 *
311 * @c->lst.empty_lebs + @c->freeable_cnt + @c->idx_gc_cnt -
312 * @c->lst.taken_empty_lebs
313 *
314 * @c->lst.empty_lebs are available because they are empty.
315 * @c->freeable_cnt are available because they contain only free and
316 * dirty space, @c->idx_gc_cnt are available because they are index
317 * LEBs that have been garbage collected and are awaiting the commit
318 * before they can be used. And the in-the-gaps method will grab these
319 * if it needs them. @c->lst.taken_empty_lebs are empty LEBs that have
320 * already been allocated for some purpose.
321 *
322 * Note, @c->idx_gc_cnt is included to both @c->lst.empty_lebs (because
323 * these LEBs are empty) and to @c->lst.taken_empty_lebs (because they
324 * are taken until after the commit).
325 *
326 * Note, @c->lst.taken_empty_lebs may temporarily be higher by one
327 * because of the way we serialize LEB allocations and budgeting. See a
328 * comment in 'ubifs_find_free_space()'.
329 */
330 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
331 c->lst.taken_empty_lebs;
332 if (unlikely(rsvd_idx_lebs > lebs)) {
333 dbg_budg("out of indexing space: min_idx_lebs %d (old %d), rsvd_idx_lebs %d",
334 min_idx_lebs, c->bi.min_idx_lebs, rsvd_idx_lebs);
335 return -ENOSPC;
336 }
337
338 available = ubifs_calc_available(c, min_idx_lebs);
339 outstanding = c->bi.data_growth + c->bi.dd_growth;
340
341 if (unlikely(available < outstanding)) {
342 dbg_budg("out of data space: available %lld, outstanding %lld",
343 available, outstanding);
344 return -ENOSPC;
345 }
346
347 if (available - outstanding <= c->rp_size && !can_use_rp(c))
348 return -ENOSPC;
349
350 c->bi.min_idx_lebs = min_idx_lebs;
351 return 0;
352}
353
354/**
355 * calc_idx_growth - calculate approximate index growth from budgeting request.
356 * @c: UBIFS file-system description object
357 * @req: budgeting request
358 *
359 * For now we assume each new node adds one znode. But this is rather poor
360 * approximation, though.
361 */
362static int calc_idx_growth(const struct ubifs_info *c,
363 const struct ubifs_budget_req *req)
364{
365 int znodes;
366
367 znodes = req->new_ino + (req->new_page << UBIFS_BLOCKS_PER_PAGE_SHIFT) +
368 req->new_dent;
369 return znodes * c->max_idx_node_sz;
370}
371
372/**
373 * calc_data_growth - calculate approximate amount of new data from budgeting
374 * request.
375 * @c: UBIFS file-system description object
376 * @req: budgeting request
377 */
378static int calc_data_growth(const struct ubifs_info *c,
379 const struct ubifs_budget_req *req)
380{
381 int data_growth;
382
383 data_growth = req->new_ino ? c->bi.inode_budget : 0;
384 if (req->new_page)
385 data_growth += c->bi.page_budget;
386 if (req->new_dent)
387 data_growth += c->bi.dent_budget;
388 data_growth += req->new_ino_d;
389 return data_growth;
390}
391
392/**
393 * calc_dd_growth - calculate approximate amount of data which makes other data
394 * dirty from budgeting request.
395 * @c: UBIFS file-system description object
396 * @req: budgeting request
397 */
398static int calc_dd_growth(const struct ubifs_info *c,
399 const struct ubifs_budget_req *req)
400{
401 int dd_growth;
402
403 dd_growth = req->dirtied_page ? c->bi.page_budget : 0;
404
405 if (req->dirtied_ino)
406 dd_growth += c->bi.inode_budget << (req->dirtied_ino - 1);
407 if (req->mod_dent)
408 dd_growth += c->bi.dent_budget;
409 dd_growth += req->dirtied_ino_d;
410 return dd_growth;
411}
412
413/**
414 * ubifs_budget_space - ensure there is enough space to complete an operation.
415 * @c: UBIFS file-system description object
416 * @req: budget request
417 *
418 * This function allocates budget for an operation. It uses pessimistic
419 * approximation of how much flash space the operation needs. The goal of this
420 * function is to make sure UBIFS always has flash space to flush all dirty
421 * pages, dirty inodes, and dirty znodes (liability). This function may force
422 * commit, garbage-collection or write-back. Returns zero in case of success,
423 * %-ENOSPC if there is no free space and other negative error codes in case of
424 * failures.
425 */
426int ubifs_budget_space(struct ubifs_info *c, struct ubifs_budget_req *req)
427{
428 int err, idx_growth, data_growth, dd_growth, retried = 0;
429
430 ubifs_assert(c, req->new_page <= 1);
431 ubifs_assert(c, req->dirtied_page <= 1);
432 ubifs_assert(c, req->new_dent <= 1);
433 ubifs_assert(c, req->mod_dent <= 1);
434 ubifs_assert(c, req->new_ino <= 1);
435 ubifs_assert(c, req->new_ino_d <= UBIFS_MAX_INO_DATA);
436 ubifs_assert(c, req->dirtied_ino <= 4);
437 ubifs_assert(c, req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
438 ubifs_assert(c, !(req->new_ino_d & 7));
439 ubifs_assert(c, !(req->dirtied_ino_d & 7));
440
441 data_growth = calc_data_growth(c, req);
442 dd_growth = calc_dd_growth(c, req);
443 if (!data_growth && !dd_growth)
444 return 0;
445 idx_growth = calc_idx_growth(c, req);
446
447again:
448 spin_lock(&c->space_lock);
449 ubifs_assert(c, c->bi.idx_growth >= 0);
450 ubifs_assert(c, c->bi.data_growth >= 0);
451 ubifs_assert(c, c->bi.dd_growth >= 0);
452
453 if (unlikely(c->bi.nospace) && (c->bi.nospace_rp || !can_use_rp(c))) {
454 dbg_budg("no space");
455 spin_unlock(&c->space_lock);
456 return -ENOSPC;
457 }
458
459 c->bi.idx_growth += idx_growth;
460 c->bi.data_growth += data_growth;
461 c->bi.dd_growth += dd_growth;
462
463 err = do_budget_space(c);
464 if (likely(!err)) {
465 req->idx_growth = idx_growth;
466 req->data_growth = data_growth;
467 req->dd_growth = dd_growth;
468 spin_unlock(&c->space_lock);
469 return 0;
470 }
471
472 /* Restore the old values */
473 c->bi.idx_growth -= idx_growth;
474 c->bi.data_growth -= data_growth;
475 c->bi.dd_growth -= dd_growth;
476 spin_unlock(&c->space_lock);
477
478 if (req->fast) {
479 dbg_budg("no space for fast budgeting");
480 return err;
481 }
482
483 err = make_free_space(c);
484 cond_resched();
485 if (err == -EAGAIN) {
486 dbg_budg("try again");
487 goto again;
488 } else if (err == -ENOSPC) {
489 if (!retried) {
490 retried = 1;
491 dbg_budg("-ENOSPC, but anyway try once again");
492 goto again;
493 }
494 dbg_budg("FS is full, -ENOSPC");
495 c->bi.nospace = 1;
496 if (can_use_rp(c) || c->rp_size == 0)
497 c->bi.nospace_rp = 1;
498 smp_wmb();
499 } else
500 ubifs_err(c, "cannot budget space, error %d", err);
501 return err;
502}
503
504/**
505 * ubifs_release_budget - release budgeted free space.
506 * @c: UBIFS file-system description object
507 * @req: budget request
508 *
509 * This function releases the space budgeted by 'ubifs_budget_space()'. Note,
510 * since the index changes (which were budgeted for in @req->idx_growth) will
511 * only be written to the media on commit, this function moves the index budget
512 * from @c->bi.idx_growth to @c->bi.uncommitted_idx. The latter will be zeroed
513 * by the commit operation.
514 */
515void ubifs_release_budget(struct ubifs_info *c, struct ubifs_budget_req *req)
516{
517 ubifs_assert(c, req->new_page <= 1);
518 ubifs_assert(c, req->dirtied_page <= 1);
519 ubifs_assert(c, req->new_dent <= 1);
520 ubifs_assert(c, req->mod_dent <= 1);
521 ubifs_assert(c, req->new_ino <= 1);
522 ubifs_assert(c, req->new_ino_d <= UBIFS_MAX_INO_DATA);
523 ubifs_assert(c, req->dirtied_ino <= 4);
524 ubifs_assert(c, req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
525 ubifs_assert(c, !(req->new_ino_d & 7));
526 ubifs_assert(c, !(req->dirtied_ino_d & 7));
527 if (!req->recalculate) {
528 ubifs_assert(c, req->idx_growth >= 0);
529 ubifs_assert(c, req->data_growth >= 0);
530 ubifs_assert(c, req->dd_growth >= 0);
531 }
532
533 if (req->recalculate) {
534 req->data_growth = calc_data_growth(c, req);
535 req->dd_growth = calc_dd_growth(c, req);
536 req->idx_growth = calc_idx_growth(c, req);
537 }
538
539 if (!req->data_growth && !req->dd_growth)
540 return;
541
542 c->bi.nospace = c->bi.nospace_rp = 0;
543 smp_wmb();
544
545 spin_lock(&c->space_lock);
546 c->bi.idx_growth -= req->idx_growth;
547 c->bi.uncommitted_idx += req->idx_growth;
548 c->bi.data_growth -= req->data_growth;
549 c->bi.dd_growth -= req->dd_growth;
550 c->bi.min_idx_lebs = ubifs_calc_min_idx_lebs(c);
551
552 ubifs_assert(c, c->bi.idx_growth >= 0);
553 ubifs_assert(c, c->bi.data_growth >= 0);
554 ubifs_assert(c, c->bi.dd_growth >= 0);
555 ubifs_assert(c, c->bi.min_idx_lebs < c->main_lebs);
556 ubifs_assert(c, !(c->bi.idx_growth & 7));
557 ubifs_assert(c, !(c->bi.data_growth & 7));
558 ubifs_assert(c, !(c->bi.dd_growth & 7));
559 spin_unlock(&c->space_lock);
560}
561
562/**
563 * ubifs_convert_page_budget - convert budget of a new page.
564 * @c: UBIFS file-system description object
565 *
566 * This function converts budget which was allocated for a new page of data to
567 * the budget of changing an existing page of data. The latter is smaller than
568 * the former, so this function only does simple re-calculation and does not
569 * involve any write-back.
570 */
571void ubifs_convert_page_budget(struct ubifs_info *c)
572{
573 spin_lock(&c->space_lock);
574 /* Release the index growth reservation */
575 c->bi.idx_growth -= c->max_idx_node_sz << UBIFS_BLOCKS_PER_PAGE_SHIFT;
576 /* Release the data growth reservation */
577 c->bi.data_growth -= c->bi.page_budget;
578 /* Increase the dirty data growth reservation instead */
579 c->bi.dd_growth += c->bi.page_budget;
580 /* And re-calculate the indexing space reservation */
581 c->bi.min_idx_lebs = ubifs_calc_min_idx_lebs(c);
582 spin_unlock(&c->space_lock);
583}
584
585/**
586 * ubifs_release_dirty_inode_budget - release dirty inode budget.
587 * @c: UBIFS file-system description object
588 * @ui: UBIFS inode to release the budget for
589 *
590 * This function releases budget corresponding to a dirty inode. It is usually
591 * called when after the inode has been written to the media and marked as
592 * clean. It also causes the "no space" flags to be cleared.
593 */
594void ubifs_release_dirty_inode_budget(struct ubifs_info *c,
595 struct ubifs_inode *ui)
596{
597 struct ubifs_budget_req req;
598
599 memset(&req, 0, sizeof(struct ubifs_budget_req));
600 /* The "no space" flags will be cleared because dd_growth is > 0 */
601 req.dd_growth = c->bi.inode_budget + ALIGN(ui->data_len, 8);
602 ubifs_release_budget(c, &req);
603}
604
605/**
606 * ubifs_reported_space - calculate reported free space.
607 * @c: the UBIFS file-system description object
608 * @free: amount of free space
609 *
610 * This function calculates amount of free space which will be reported to
611 * user-space. User-space application tend to expect that if the file-system
612 * (e.g., via the 'statfs()' call) reports that it has N bytes available, they
613 * are able to write a file of size N. UBIFS attaches node headers to each data
614 * node and it has to write indexing nodes as well. This introduces additional
615 * overhead, and UBIFS has to report slightly less free space to meet the above
616 * expectations.
617 *
618 * This function assumes free space is made up of uncompressed data nodes and
619 * full index nodes (one per data node, tripled because we always allow enough
620 * space to write the index thrice).
621 *
622 * Note, the calculation is pessimistic, which means that most of the time
623 * UBIFS reports less space than it actually has.
624 */
625long long ubifs_reported_space(const struct ubifs_info *c, long long free)
626{
627 int divisor, factor, f;
628
629 /*
630 * Reported space size is @free * X, where X is UBIFS block size
631 * divided by UBIFS block size + all overhead one data block
632 * introduces. The overhead is the node header + indexing overhead.
633 *
634 * Indexing overhead calculations are based on the following formula:
635 * I = N/(f - 1) + 1, where I - number of indexing nodes, N - number
636 * of data nodes, f - fanout. Because effective UBIFS fanout is twice
637 * as less than maximum fanout, we assume that each data node
638 * introduces 3 * @c->max_idx_node_sz / (@c->fanout/2 - 1) bytes.
639 * Note, the multiplier 3 is because UBIFS reserves thrice as more space
640 * for the index.
641 */
642 f = c->fanout > 3 ? c->fanout >> 1 : 2;
643 factor = UBIFS_BLOCK_SIZE;
644 divisor = UBIFS_MAX_DATA_NODE_SZ;
645 divisor += (c->max_idx_node_sz * 3) / (f - 1);
646 free *= factor;
647 return div_u64(free, divisor);
648}
649
650/**
651 * ubifs_get_free_space_nolock - return amount of free space.
652 * @c: UBIFS file-system description object
653 *
654 * This function calculates amount of free space to report to user-space.
655 *
656 * Because UBIFS may introduce substantial overhead (the index, node headers,
657 * alignment, wastage at the end of LEBs, etc), it cannot report real amount of
658 * free flash space it has (well, because not all dirty space is reclaimable,
659 * UBIFS does not actually know the real amount). If UBIFS did so, it would
660 * bread user expectations about what free space is. Users seem to accustomed
661 * to assume that if the file-system reports N bytes of free space, they would
662 * be able to fit a file of N bytes to the FS. This almost works for
663 * traditional file-systems, because they have way less overhead than UBIFS.
664 * So, to keep users happy, UBIFS tries to take the overhead into account.
665 */
666long long ubifs_get_free_space_nolock(struct ubifs_info *c)
667{
668 int rsvd_idx_lebs, lebs;
669 long long available, outstanding, free;
670
671 ubifs_assert(c, c->bi.min_idx_lebs == ubifs_calc_min_idx_lebs(c));
672 outstanding = c->bi.data_growth + c->bi.dd_growth;
673 available = ubifs_calc_available(c, c->bi.min_idx_lebs);
674
675 /*
676 * When reporting free space to user-space, UBIFS guarantees that it is
677 * possible to write a file of free space size. This means that for
678 * empty LEBs we may use more precise calculations than
679 * 'ubifs_calc_available()' is using. Namely, we know that in empty
680 * LEBs we would waste only @c->leb_overhead bytes, not @c->dark_wm.
681 * Thus, amend the available space.
682 *
683 * Note, the calculations below are similar to what we have in
684 * 'do_budget_space()', so refer there for comments.
685 */
686 if (c->bi.min_idx_lebs > c->lst.idx_lebs)
687 rsvd_idx_lebs = c->bi.min_idx_lebs - c->lst.idx_lebs;
688 else
689 rsvd_idx_lebs = 0;
690 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
691 c->lst.taken_empty_lebs;
692 lebs -= rsvd_idx_lebs;
693 available += lebs * (c->dark_wm - c->leb_overhead);
694
695 if (available > outstanding)
696 free = ubifs_reported_space(c, available - outstanding);
697 else
698 free = 0;
699 return free;
700}
701
702/**
703 * ubifs_get_free_space - return amount of free space.
704 * @c: UBIFS file-system description object
705 *
706 * This function calculates and returns amount of free space to report to
707 * user-space.
708 */
709long long ubifs_get_free_space(struct ubifs_info *c)
710{
711 long long free;
712
713 spin_lock(&c->space_lock);
714 free = ubifs_get_free_space_nolock(c);
715 spin_unlock(&c->space_lock);
716
717 return free;
718}
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/*
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 <linux/math64.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 constant defines maximum number of times UBIFS
41 * repeats the operations.
42 */
43#define MAX_MKSPC_RETRIES 3
44
45/*
46 * The below constant defines amount of dirty pages which should be written
47 * back at when trying to shrink the liability.
48 */
49#define NR_TO_WRITE 16
50
51/**
52 * shrink_liability - write-back some dirty pages/inodes.
53 * @c: UBIFS file-system description object
54 * @nr_to_write: how many dirty pages to write-back
55 *
56 * This function shrinks UBIFS liability by means of writing back some amount
57 * of dirty inodes and their pages.
58 *
59 * Note, this function synchronizes even VFS inodes which are locked
60 * (@i_mutex) by the caller of the budgeting function, because write-back does
61 * not touch @i_mutex.
62 */
63static void shrink_liability(struct ubifs_info *c, int nr_to_write)
64{
65 down_read(&c->vfs_sb->s_umount);
66 writeback_inodes_sb(c->vfs_sb, WB_REASON_FS_FREE_SPACE);
67 up_read(&c->vfs_sb->s_umount);
68}
69
70/**
71 * run_gc - run garbage collector.
72 * @c: UBIFS file-system description object
73 *
74 * This function runs garbage collector to make some more free space. Returns
75 * zero if a free LEB has been produced, %-EAGAIN if commit is required, and a
76 * negative error code in case of failure.
77 */
78static int run_gc(struct ubifs_info *c)
79{
80 int err, lnum;
81
82 /* Make some free space by garbage-collecting dirty space */
83 down_read(&c->commit_sem);
84 lnum = ubifs_garbage_collect(c, 1);
85 up_read(&c->commit_sem);
86 if (lnum < 0)
87 return lnum;
88
89 /* GC freed one LEB, return it to lprops */
90 dbg_budg("GC freed LEB %d", lnum);
91 err = ubifs_return_leb(c, lnum);
92 if (err)
93 return err;
94 return 0;
95}
96
97/**
98 * get_liability - calculate current liability.
99 * @c: UBIFS file-system description object
100 *
101 * This function calculates and returns current UBIFS liability, i.e. the
102 * amount of bytes UBIFS has "promised" to write to the media.
103 */
104static long long get_liability(struct ubifs_info *c)
105{
106 long long liab;
107
108 spin_lock(&c->space_lock);
109 liab = c->bi.idx_growth + c->bi.data_growth + c->bi.dd_growth;
110 spin_unlock(&c->space_lock);
111 return liab;
112}
113
114/**
115 * make_free_space - make more free space on the file-system.
116 * @c: UBIFS file-system description object
117 *
118 * This function is called when an operation cannot be budgeted because there
119 * is supposedly no free space. But in most cases there is some free space:
120 * o budgeting is pessimistic, so it always budgets more than it is actually
121 * needed, so shrinking the liability is one way to make free space - the
122 * cached data will take less space then it was budgeted for;
123 * o GC may turn some dark space into free space (budgeting treats dark space
124 * as not available);
125 * o commit may free some LEB, i.e., turn freeable LEBs into free LEBs.
126 *
127 * So this function tries to do the above. Returns %-EAGAIN if some free space
128 * was presumably made and the caller has to re-try budgeting the operation.
129 * Returns %-ENOSPC if it couldn't do more free space, and other negative error
130 * codes on failures.
131 */
132static int make_free_space(struct ubifs_info *c)
133{
134 int err, retries = 0;
135 long long liab1, liab2;
136
137 do {
138 liab1 = get_liability(c);
139 /*
140 * We probably have some dirty pages or inodes (liability), try
141 * to write them back.
142 */
143 dbg_budg("liability %lld, run write-back", liab1);
144 shrink_liability(c, NR_TO_WRITE);
145
146 liab2 = get_liability(c);
147 if (liab2 < liab1)
148 return -EAGAIN;
149
150 dbg_budg("new liability %lld (not shrunk)", liab2);
151
152 /* Liability did not shrink again, try GC */
153 dbg_budg("Run GC");
154 err = run_gc(c);
155 if (!err)
156 return -EAGAIN;
157
158 if (err != -EAGAIN && err != -ENOSPC)
159 /* Some real error happened */
160 return err;
161
162 dbg_budg("Run commit (retries %d)", retries);
163 err = ubifs_run_commit(c);
164 if (err)
165 return err;
166 } while (retries++ < MAX_MKSPC_RETRIES);
167
168 return -ENOSPC;
169}
170
171/**
172 * ubifs_calc_min_idx_lebs - calculate amount of LEBs for the index.
173 * @c: UBIFS file-system description object
174 *
175 * This function calculates and returns the number of LEBs which should be kept
176 * for index usage.
177 */
178int ubifs_calc_min_idx_lebs(struct ubifs_info *c)
179{
180 int idx_lebs;
181 long long idx_size;
182
183 idx_size = c->bi.old_idx_sz + c->bi.idx_growth + c->bi.uncommitted_idx;
184 /* And make sure we have thrice the index size of space reserved */
185 idx_size += idx_size << 1;
186 /*
187 * We do not maintain 'old_idx_size' as 'old_idx_lebs'/'old_idx_bytes'
188 * pair, nor similarly the two variables for the new index size, so we
189 * have to do this costly 64-bit division on fast-path.
190 */
191 idx_lebs = div_u64(idx_size + c->idx_leb_size - 1, c->idx_leb_size);
192 /*
193 * The index head is not available for the in-the-gaps method, so add an
194 * extra LEB to compensate.
195 */
196 idx_lebs += 1;
197 if (idx_lebs < MIN_INDEX_LEBS)
198 idx_lebs = MIN_INDEX_LEBS;
199 return idx_lebs;
200}
201
202/**
203 * ubifs_calc_available - calculate available FS space.
204 * @c: UBIFS file-system description object
205 * @min_idx_lebs: minimum number of LEBs reserved for the index
206 *
207 * This function calculates and returns amount of FS space available for use.
208 */
209long long ubifs_calc_available(const struct ubifs_info *c, int min_idx_lebs)
210{
211 int subtract_lebs;
212 long long available;
213
214 available = c->main_bytes - c->lst.total_used;
215
216 /*
217 * Now 'available' contains theoretically available flash space
218 * assuming there is no index, so we have to subtract the space which
219 * is reserved for the index.
220 */
221 subtract_lebs = min_idx_lebs;
222
223 /* Take into account that GC reserves one LEB for its own needs */
224 subtract_lebs += 1;
225
226 /*
227 * The GC journal head LEB is not really accessible. And since
228 * different write types go to different heads, we may count only on
229 * one head's space.
230 */
231 subtract_lebs += c->jhead_cnt - 1;
232
233 /* We also reserve one LEB for deletions, which bypass budgeting */
234 subtract_lebs += 1;
235
236 available -= (long long)subtract_lebs * c->leb_size;
237
238 /* Subtract the dead space which is not available for use */
239 available -= c->lst.total_dead;
240
241 /*
242 * Subtract dark space, which might or might not be usable - it depends
243 * on the data which we have on the media and which will be written. If
244 * this is a lot of uncompressed or not-compressible data, the dark
245 * space cannot be used.
246 */
247 available -= c->lst.total_dark;
248
249 /*
250 * However, there is more dark space. The index may be bigger than
251 * @min_idx_lebs. Those extra LEBs are assumed to be available, but
252 * their dark space is not included in total_dark, so it is subtracted
253 * here.
254 */
255 if (c->lst.idx_lebs > min_idx_lebs) {
256 subtract_lebs = c->lst.idx_lebs - min_idx_lebs;
257 available -= subtract_lebs * c->dark_wm;
258 }
259
260 /* The calculations are rough and may end up with a negative number */
261 return available > 0 ? available : 0;
262}
263
264/**
265 * can_use_rp - check whether the user is allowed to use reserved pool.
266 * @c: UBIFS file-system description object
267 *
268 * UBIFS has so-called "reserved pool" which is flash space reserved
269 * for the superuser and for uses whose UID/GID is recorded in UBIFS superblock.
270 * This function checks whether current user is allowed to use reserved pool.
271 * Returns %1 current user is allowed to use reserved pool and %0 otherwise.
272 */
273static int can_use_rp(struct ubifs_info *c)
274{
275 if (uid_eq(current_fsuid(), c->rp_uid) || capable(CAP_SYS_RESOURCE) ||
276 (!gid_eq(c->rp_gid, GLOBAL_ROOT_GID) && in_group_p(c->rp_gid)))
277 return 1;
278 return 0;
279}
280
281/**
282 * do_budget_space - reserve flash space for index and data growth.
283 * @c: UBIFS file-system description object
284 *
285 * This function makes sure UBIFS has enough free LEBs for index growth and
286 * data.
287 *
288 * When budgeting index space, UBIFS reserves thrice as many LEBs as the index
289 * would take if it was consolidated and written to the flash. This guarantees
290 * that the "in-the-gaps" commit method always succeeds and UBIFS will always
291 * be able to commit dirty index. So this function basically adds amount of
292 * budgeted index space to the size of the current index, multiplies this by 3,
293 * and makes sure this does not exceed the amount of free LEBs.
294 *
295 * Notes about @c->bi.min_idx_lebs and @c->lst.idx_lebs variables:
296 * o @c->lst.idx_lebs is the number of LEBs the index currently uses. It might
297 * be large, because UBIFS does not do any index consolidation as long as
298 * there is free space. IOW, the index may take a lot of LEBs, but the LEBs
299 * will contain a lot of dirt.
300 * o @c->bi.min_idx_lebs is the number of LEBS the index presumably takes. IOW,
301 * the index may be consolidated to take up to @c->bi.min_idx_lebs LEBs.
302 *
303 * This function returns zero in case of success, and %-ENOSPC in case of
304 * failure.
305 */
306static int do_budget_space(struct ubifs_info *c)
307{
308 long long outstanding, available;
309 int lebs, rsvd_idx_lebs, min_idx_lebs;
310
311 /* First budget index space */
312 min_idx_lebs = ubifs_calc_min_idx_lebs(c);
313
314 /* Now 'min_idx_lebs' contains number of LEBs to reserve */
315 if (min_idx_lebs > c->lst.idx_lebs)
316 rsvd_idx_lebs = min_idx_lebs - c->lst.idx_lebs;
317 else
318 rsvd_idx_lebs = 0;
319
320 /*
321 * The number of LEBs that are available to be used by the index is:
322 *
323 * @c->lst.empty_lebs + @c->freeable_cnt + @c->idx_gc_cnt -
324 * @c->lst.taken_empty_lebs
325 *
326 * @c->lst.empty_lebs are available because they are empty.
327 * @c->freeable_cnt are available because they contain only free and
328 * dirty space, @c->idx_gc_cnt are available because they are index
329 * LEBs that have been garbage collected and are awaiting the commit
330 * before they can be used. And the in-the-gaps method will grab these
331 * if it needs them. @c->lst.taken_empty_lebs are empty LEBs that have
332 * already been allocated for some purpose.
333 *
334 * Note, @c->idx_gc_cnt is included to both @c->lst.empty_lebs (because
335 * these LEBs are empty) and to @c->lst.taken_empty_lebs (because they
336 * are taken until after the commit).
337 *
338 * Note, @c->lst.taken_empty_lebs may temporarily be higher by one
339 * because of the way we serialize LEB allocations and budgeting. See a
340 * comment in 'ubifs_find_free_space()'.
341 */
342 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
343 c->lst.taken_empty_lebs;
344 if (unlikely(rsvd_idx_lebs > lebs)) {
345 dbg_budg("out of indexing space: min_idx_lebs %d (old %d), rsvd_idx_lebs %d",
346 min_idx_lebs, c->bi.min_idx_lebs, rsvd_idx_lebs);
347 return -ENOSPC;
348 }
349
350 available = ubifs_calc_available(c, min_idx_lebs);
351 outstanding = c->bi.data_growth + c->bi.dd_growth;
352
353 if (unlikely(available < outstanding)) {
354 dbg_budg("out of data space: available %lld, outstanding %lld",
355 available, outstanding);
356 return -ENOSPC;
357 }
358
359 if (available - outstanding <= c->rp_size && !can_use_rp(c))
360 return -ENOSPC;
361
362 c->bi.min_idx_lebs = min_idx_lebs;
363 return 0;
364}
365
366/**
367 * calc_idx_growth - calculate approximate index growth from budgeting request.
368 * @c: UBIFS file-system description object
369 * @req: budgeting request
370 *
371 * For now we assume each new node adds one znode. But this is rather poor
372 * approximation, though.
373 */
374static int calc_idx_growth(const struct ubifs_info *c,
375 const struct ubifs_budget_req *req)
376{
377 int znodes;
378
379 znodes = req->new_ino + (req->new_page << UBIFS_BLOCKS_PER_PAGE_SHIFT) +
380 req->new_dent;
381 return znodes * c->max_idx_node_sz;
382}
383
384/**
385 * calc_data_growth - calculate approximate amount of new data from budgeting
386 * request.
387 * @c: UBIFS file-system description object
388 * @req: budgeting request
389 */
390static int calc_data_growth(const struct ubifs_info *c,
391 const struct ubifs_budget_req *req)
392{
393 int data_growth;
394
395 data_growth = req->new_ino ? c->bi.inode_budget : 0;
396 if (req->new_page)
397 data_growth += c->bi.page_budget;
398 if (req->new_dent)
399 data_growth += c->bi.dent_budget;
400 data_growth += req->new_ino_d;
401 return data_growth;
402}
403
404/**
405 * calc_dd_growth - calculate approximate amount of data which makes other data
406 * dirty from budgeting request.
407 * @c: UBIFS file-system description object
408 * @req: budgeting request
409 */
410static int calc_dd_growth(const struct ubifs_info *c,
411 const struct ubifs_budget_req *req)
412{
413 int dd_growth;
414
415 dd_growth = req->dirtied_page ? c->bi.page_budget : 0;
416
417 if (req->dirtied_ino)
418 dd_growth += c->bi.inode_budget << (req->dirtied_ino - 1);
419 if (req->mod_dent)
420 dd_growth += c->bi.dent_budget;
421 dd_growth += req->dirtied_ino_d;
422 return dd_growth;
423}
424
425/**
426 * ubifs_budget_space - ensure there is enough space to complete an operation.
427 * @c: UBIFS file-system description object
428 * @req: budget request
429 *
430 * This function allocates budget for an operation. It uses pessimistic
431 * approximation of how much flash space the operation needs. The goal of this
432 * function is to make sure UBIFS always has flash space to flush all dirty
433 * pages, dirty inodes, and dirty znodes (liability). This function may force
434 * commit, garbage-collection or write-back. Returns zero in case of success,
435 * %-ENOSPC if there is no free space and other negative error codes in case of
436 * failures.
437 */
438int ubifs_budget_space(struct ubifs_info *c, struct ubifs_budget_req *req)
439{
440 int uninitialized_var(cmt_retries), uninitialized_var(wb_retries);
441 int err, idx_growth, data_growth, dd_growth, retried = 0;
442
443 ubifs_assert(req->new_page <= 1);
444 ubifs_assert(req->dirtied_page <= 1);
445 ubifs_assert(req->new_dent <= 1);
446 ubifs_assert(req->mod_dent <= 1);
447 ubifs_assert(req->new_ino <= 1);
448 ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA);
449 ubifs_assert(req->dirtied_ino <= 4);
450 ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
451 ubifs_assert(!(req->new_ino_d & 7));
452 ubifs_assert(!(req->dirtied_ino_d & 7));
453
454 data_growth = calc_data_growth(c, req);
455 dd_growth = calc_dd_growth(c, req);
456 if (!data_growth && !dd_growth)
457 return 0;
458 idx_growth = calc_idx_growth(c, req);
459
460again:
461 spin_lock(&c->space_lock);
462 ubifs_assert(c->bi.idx_growth >= 0);
463 ubifs_assert(c->bi.data_growth >= 0);
464 ubifs_assert(c->bi.dd_growth >= 0);
465
466 if (unlikely(c->bi.nospace) && (c->bi.nospace_rp || !can_use_rp(c))) {
467 dbg_budg("no space");
468 spin_unlock(&c->space_lock);
469 return -ENOSPC;
470 }
471
472 c->bi.idx_growth += idx_growth;
473 c->bi.data_growth += data_growth;
474 c->bi.dd_growth += dd_growth;
475
476 err = do_budget_space(c);
477 if (likely(!err)) {
478 req->idx_growth = idx_growth;
479 req->data_growth = data_growth;
480 req->dd_growth = dd_growth;
481 spin_unlock(&c->space_lock);
482 return 0;
483 }
484
485 /* Restore the old values */
486 c->bi.idx_growth -= idx_growth;
487 c->bi.data_growth -= data_growth;
488 c->bi.dd_growth -= dd_growth;
489 spin_unlock(&c->space_lock);
490
491 if (req->fast) {
492 dbg_budg("no space for fast budgeting");
493 return err;
494 }
495
496 err = make_free_space(c);
497 cond_resched();
498 if (err == -EAGAIN) {
499 dbg_budg("try again");
500 goto again;
501 } else if (err == -ENOSPC) {
502 if (!retried) {
503 retried = 1;
504 dbg_budg("-ENOSPC, but anyway try once again");
505 goto again;
506 }
507 dbg_budg("FS is full, -ENOSPC");
508 c->bi.nospace = 1;
509 if (can_use_rp(c) || c->rp_size == 0)
510 c->bi.nospace_rp = 1;
511 smp_wmb();
512 } else
513 ubifs_err("cannot budget space, error %d", err);
514 return err;
515}
516
517/**
518 * ubifs_release_budget - release budgeted free space.
519 * @c: UBIFS file-system description object
520 * @req: budget request
521 *
522 * This function releases the space budgeted by 'ubifs_budget_space()'. Note,
523 * since the index changes (which were budgeted for in @req->idx_growth) will
524 * only be written to the media on commit, this function moves the index budget
525 * from @c->bi.idx_growth to @c->bi.uncommitted_idx. The latter will be zeroed
526 * by the commit operation.
527 */
528void ubifs_release_budget(struct ubifs_info *c, struct ubifs_budget_req *req)
529{
530 ubifs_assert(req->new_page <= 1);
531 ubifs_assert(req->dirtied_page <= 1);
532 ubifs_assert(req->new_dent <= 1);
533 ubifs_assert(req->mod_dent <= 1);
534 ubifs_assert(req->new_ino <= 1);
535 ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA);
536 ubifs_assert(req->dirtied_ino <= 4);
537 ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
538 ubifs_assert(!(req->new_ino_d & 7));
539 ubifs_assert(!(req->dirtied_ino_d & 7));
540 if (!req->recalculate) {
541 ubifs_assert(req->idx_growth >= 0);
542 ubifs_assert(req->data_growth >= 0);
543 ubifs_assert(req->dd_growth >= 0);
544 }
545
546 if (req->recalculate) {
547 req->data_growth = calc_data_growth(c, req);
548 req->dd_growth = calc_dd_growth(c, req);
549 req->idx_growth = calc_idx_growth(c, req);
550 }
551
552 if (!req->data_growth && !req->dd_growth)
553 return;
554
555 c->bi.nospace = c->bi.nospace_rp = 0;
556 smp_wmb();
557
558 spin_lock(&c->space_lock);
559 c->bi.idx_growth -= req->idx_growth;
560 c->bi.uncommitted_idx += req->idx_growth;
561 c->bi.data_growth -= req->data_growth;
562 c->bi.dd_growth -= req->dd_growth;
563 c->bi.min_idx_lebs = ubifs_calc_min_idx_lebs(c);
564
565 ubifs_assert(c->bi.idx_growth >= 0);
566 ubifs_assert(c->bi.data_growth >= 0);
567 ubifs_assert(c->bi.dd_growth >= 0);
568 ubifs_assert(c->bi.min_idx_lebs < c->main_lebs);
569 ubifs_assert(!(c->bi.idx_growth & 7));
570 ubifs_assert(!(c->bi.data_growth & 7));
571 ubifs_assert(!(c->bi.dd_growth & 7));
572 spin_unlock(&c->space_lock);
573}
574
575/**
576 * ubifs_convert_page_budget - convert budget of a new page.
577 * @c: UBIFS file-system description object
578 *
579 * This function converts budget which was allocated for a new page of data to
580 * the budget of changing an existing page of data. The latter is smaller than
581 * the former, so this function only does simple re-calculation and does not
582 * involve any write-back.
583 */
584void ubifs_convert_page_budget(struct ubifs_info *c)
585{
586 spin_lock(&c->space_lock);
587 /* Release the index growth reservation */
588 c->bi.idx_growth -= c->max_idx_node_sz << UBIFS_BLOCKS_PER_PAGE_SHIFT;
589 /* Release the data growth reservation */
590 c->bi.data_growth -= c->bi.page_budget;
591 /* Increase the dirty data growth reservation instead */
592 c->bi.dd_growth += c->bi.page_budget;
593 /* And re-calculate the indexing space reservation */
594 c->bi.min_idx_lebs = ubifs_calc_min_idx_lebs(c);
595 spin_unlock(&c->space_lock);
596}
597
598/**
599 * ubifs_release_dirty_inode_budget - release dirty inode budget.
600 * @c: UBIFS file-system description object
601 * @ui: UBIFS inode to release the budget for
602 *
603 * This function releases budget corresponding to a dirty inode. It is usually
604 * called when after the inode has been written to the media and marked as
605 * clean. It also causes the "no space" flags to be cleared.
606 */
607void ubifs_release_dirty_inode_budget(struct ubifs_info *c,
608 struct ubifs_inode *ui)
609{
610 struct ubifs_budget_req req;
611
612 memset(&req, 0, sizeof(struct ubifs_budget_req));
613 /* The "no space" flags will be cleared because dd_growth is > 0 */
614 req.dd_growth = c->bi.inode_budget + ALIGN(ui->data_len, 8);
615 ubifs_release_budget(c, &req);
616}
617
618/**
619 * ubifs_reported_space - calculate reported free space.
620 * @c: the UBIFS file-system description object
621 * @free: amount of free space
622 *
623 * This function calculates amount of free space which will be reported to
624 * user-space. User-space application tend to expect that if the file-system
625 * (e.g., via the 'statfs()' call) reports that it has N bytes available, they
626 * are able to write a file of size N. UBIFS attaches node headers to each data
627 * node and it has to write indexing nodes as well. This introduces additional
628 * overhead, and UBIFS has to report slightly less free space to meet the above
629 * expectations.
630 *
631 * This function assumes free space is made up of uncompressed data nodes and
632 * full index nodes (one per data node, tripled because we always allow enough
633 * space to write the index thrice).
634 *
635 * Note, the calculation is pessimistic, which means that most of the time
636 * UBIFS reports less space than it actually has.
637 */
638long long ubifs_reported_space(const struct ubifs_info *c, long long free)
639{
640 int divisor, factor, f;
641
642 /*
643 * Reported space size is @free * X, where X is UBIFS block size
644 * divided by UBIFS block size + all overhead one data block
645 * introduces. The overhead is the node header + indexing overhead.
646 *
647 * Indexing overhead calculations are based on the following formula:
648 * I = N/(f - 1) + 1, where I - number of indexing nodes, N - number
649 * of data nodes, f - fanout. Because effective UBIFS fanout is twice
650 * as less than maximum fanout, we assume that each data node
651 * introduces 3 * @c->max_idx_node_sz / (@c->fanout/2 - 1) bytes.
652 * Note, the multiplier 3 is because UBIFS reserves thrice as more space
653 * for the index.
654 */
655 f = c->fanout > 3 ? c->fanout >> 1 : 2;
656 factor = UBIFS_BLOCK_SIZE;
657 divisor = UBIFS_MAX_DATA_NODE_SZ;
658 divisor += (c->max_idx_node_sz * 3) / (f - 1);
659 free *= factor;
660 return div_u64(free, divisor);
661}
662
663/**
664 * ubifs_get_free_space_nolock - return amount of free space.
665 * @c: UBIFS file-system description object
666 *
667 * This function calculates amount of free space to report to user-space.
668 *
669 * Because UBIFS may introduce substantial overhead (the index, node headers,
670 * alignment, wastage at the end of LEBs, etc), it cannot report real amount of
671 * free flash space it has (well, because not all dirty space is reclaimable,
672 * UBIFS does not actually know the real amount). If UBIFS did so, it would
673 * bread user expectations about what free space is. Users seem to accustomed
674 * to assume that if the file-system reports N bytes of free space, they would
675 * be able to fit a file of N bytes to the FS. This almost works for
676 * traditional file-systems, because they have way less overhead than UBIFS.
677 * So, to keep users happy, UBIFS tries to take the overhead into account.
678 */
679long long ubifs_get_free_space_nolock(struct ubifs_info *c)
680{
681 int rsvd_idx_lebs, lebs;
682 long long available, outstanding, free;
683
684 ubifs_assert(c->bi.min_idx_lebs == ubifs_calc_min_idx_lebs(c));
685 outstanding = c->bi.data_growth + c->bi.dd_growth;
686 available = ubifs_calc_available(c, c->bi.min_idx_lebs);
687
688 /*
689 * When reporting free space to user-space, UBIFS guarantees that it is
690 * possible to write a file of free space size. This means that for
691 * empty LEBs we may use more precise calculations than
692 * 'ubifs_calc_available()' is using. Namely, we know that in empty
693 * LEBs we would waste only @c->leb_overhead bytes, not @c->dark_wm.
694 * Thus, amend the available space.
695 *
696 * Note, the calculations below are similar to what we have in
697 * 'do_budget_space()', so refer there for comments.
698 */
699 if (c->bi.min_idx_lebs > c->lst.idx_lebs)
700 rsvd_idx_lebs = c->bi.min_idx_lebs - c->lst.idx_lebs;
701 else
702 rsvd_idx_lebs = 0;
703 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
704 c->lst.taken_empty_lebs;
705 lebs -= rsvd_idx_lebs;
706 available += lebs * (c->dark_wm - c->leb_overhead);
707
708 if (available > outstanding)
709 free = ubifs_reported_space(c, available - outstanding);
710 else
711 free = 0;
712 return free;
713}
714
715/**
716 * ubifs_get_free_space - return amount of free space.
717 * @c: UBIFS file-system description object
718 *
719 * This function calculates and returns amount of free space to report to
720 * user-space.
721 */
722long long ubifs_get_free_space(struct ubifs_info *c)
723{
724 long long free;
725
726 spin_lock(&c->space_lock);
727 free = ubifs_get_free_space_nolock(c);
728 spin_unlock(&c->space_lock);
729
730 return free;
731}