2 * This file is part of UBIFS.
4 * Copyright (C) 2006-2008 Nokia Corporation.
6 * SPDX-License-Identifier: GPL-2.0+
8 * Authors: Adrian Hunter
9 * Artem Bityutskiy (Битюцкий Артём)
13 * This file implements the budgeting sub-system which is responsible for UBIFS
16 * Factors such as compression, wasted space at the ends of LEBs, space in other
17 * journal heads, the effect of updates on the index, and so on, make it
18 * impossible to accurately predict the amount of space needed. Consequently
19 * approximations are used.
24 #include <linux/writeback.h>
26 #include <linux/err.h>
28 #include <linux/math64.h>
31 * When pessimistic budget calculations say that there is no enough space,
32 * UBIFS starts writing back dirty inodes and pages, doing garbage collection,
33 * or committing. The below constant defines maximum number of times UBIFS
34 * repeats the operations.
36 #define MAX_MKSPC_RETRIES 3
39 * The below constant defines amount of dirty pages which should be written
40 * back at when trying to shrink the liability.
42 #define NR_TO_WRITE 16
46 * shrink_liability - write-back some dirty pages/inodes.
47 * @c: UBIFS file-system description object
48 * @nr_to_write: how many dirty pages to write-back
50 * This function shrinks UBIFS liability by means of writing back some amount
51 * of dirty inodes and their pages.
53 * Note, this function synchronizes even VFS inodes which are locked
54 * (@i_mutex) by the caller of the budgeting function, because write-back does
57 static void shrink_liability(struct ubifs_info *c, int nr_to_write)
59 down_read(&c->vfs_sb->s_umount);
60 writeback_inodes_sb(c->vfs_sb, WB_REASON_FS_FREE_SPACE);
61 up_read(&c->vfs_sb->s_umount);
65 * run_gc - run garbage collector.
66 * @c: UBIFS file-system description object
68 * This function runs garbage collector to make some more free space. Returns
69 * zero if a free LEB has been produced, %-EAGAIN if commit is required, and a
70 * negative error code in case of failure.
72 static int run_gc(struct ubifs_info *c)
76 /* Make some free space by garbage-collecting dirty space */
77 down_read(&c->commit_sem);
78 lnum = ubifs_garbage_collect(c, 1);
79 up_read(&c->commit_sem);
83 /* GC freed one LEB, return it to lprops */
84 dbg_budg("GC freed LEB %d", lnum);
85 err = ubifs_return_leb(c, lnum);
92 * get_liability - calculate current liability.
93 * @c: UBIFS file-system description object
95 * This function calculates and returns current UBIFS liability, i.e. the
96 * amount of bytes UBIFS has "promised" to write to the media.
98 static long long get_liability(struct ubifs_info *c)
102 spin_lock(&c->space_lock);
103 liab = c->bi.idx_growth + c->bi.data_growth + c->bi.dd_growth;
104 spin_unlock(&c->space_lock);
109 * make_free_space - make more free space on the file-system.
110 * @c: UBIFS file-system description object
112 * This function is called when an operation cannot be budgeted because there
113 * is supposedly no free space. But in most cases there is some free space:
114 * o budgeting is pessimistic, so it always budgets more than it is actually
115 * needed, so shrinking the liability is one way to make free space - the
116 * cached data will take less space then it was budgeted for;
117 * o GC may turn some dark space into free space (budgeting treats dark space
119 * o commit may free some LEB, i.e., turn freeable LEBs into free LEBs.
121 * So this function tries to do the above. Returns %-EAGAIN if some free space
122 * was presumably made and the caller has to re-try budgeting the operation.
123 * Returns %-ENOSPC if it couldn't do more free space, and other negative error
126 static int make_free_space(struct ubifs_info *c)
128 int err, retries = 0;
129 long long liab1, liab2;
132 liab1 = get_liability(c);
134 * We probably have some dirty pages or inodes (liability), try
135 * to write them back.
137 dbg_budg("liability %lld, run write-back", liab1);
138 shrink_liability(c, NR_TO_WRITE);
140 liab2 = get_liability(c);
144 dbg_budg("new liability %lld (not shrunk)", liab2);
146 /* Liability did not shrink again, try GC */
152 if (err != -EAGAIN && err != -ENOSPC)
153 /* Some real error happened */
156 dbg_budg("Run commit (retries %d)", retries);
157 err = ubifs_run_commit(c);
160 } while (retries++ < MAX_MKSPC_RETRIES);
167 * ubifs_calc_min_idx_lebs - calculate amount of LEBs for the index.
168 * @c: UBIFS file-system description object
170 * This function calculates and returns the number of LEBs which should be kept
173 int ubifs_calc_min_idx_lebs(struct ubifs_info *c)
178 idx_size = c->bi.old_idx_sz + c->bi.idx_growth + c->bi.uncommitted_idx;
179 /* And make sure we have thrice the index size of space reserved */
180 idx_size += idx_size << 1;
182 * We do not maintain 'old_idx_size' as 'old_idx_lebs'/'old_idx_bytes'
183 * pair, nor similarly the two variables for the new index size, so we
184 * have to do this costly 64-bit division on fast-path.
186 idx_lebs = div_u64(idx_size + c->idx_leb_size - 1, c->idx_leb_size);
188 * The index head is not available for the in-the-gaps method, so add an
189 * extra LEB to compensate.
192 if (idx_lebs < MIN_INDEX_LEBS)
193 idx_lebs = MIN_INDEX_LEBS;
199 * ubifs_calc_available - calculate available FS space.
200 * @c: UBIFS file-system description object
201 * @min_idx_lebs: minimum number of LEBs reserved for the index
203 * This function calculates and returns amount of FS space available for use.
205 long long ubifs_calc_available(const struct ubifs_info *c, int min_idx_lebs)
210 available = c->main_bytes - c->lst.total_used;
213 * Now 'available' contains theoretically available flash space
214 * assuming there is no index, so we have to subtract the space which
215 * is reserved for the index.
217 subtract_lebs = min_idx_lebs;
219 /* Take into account that GC reserves one LEB for its own needs */
223 * The GC journal head LEB is not really accessible. And since
224 * different write types go to different heads, we may count only on
227 subtract_lebs += c->jhead_cnt - 1;
229 /* We also reserve one LEB for deletions, which bypass budgeting */
232 available -= (long long)subtract_lebs * c->leb_size;
234 /* Subtract the dead space which is not available for use */
235 available -= c->lst.total_dead;
238 * Subtract dark space, which might or might not be usable - it depends
239 * on the data which we have on the media and which will be written. If
240 * this is a lot of uncompressed or not-compressible data, the dark
241 * space cannot be used.
243 available -= c->lst.total_dark;
246 * However, there is more dark space. The index may be bigger than
247 * @min_idx_lebs. Those extra LEBs are assumed to be available, but
248 * their dark space is not included in total_dark, so it is subtracted
251 if (c->lst.idx_lebs > min_idx_lebs) {
252 subtract_lebs = c->lst.idx_lebs - min_idx_lebs;
253 available -= subtract_lebs * c->dark_wm;
256 /* The calculations are rough and may end up with a negative number */
257 return available > 0 ? available : 0;
261 * can_use_rp - check whether the user is allowed to use reserved pool.
262 * @c: UBIFS file-system description object
264 * UBIFS has so-called "reserved pool" which is flash space reserved
265 * for the superuser and for uses whose UID/GID is recorded in UBIFS superblock.
266 * This function checks whether current user is allowed to use reserved pool.
267 * Returns %1 current user is allowed to use reserved pool and %0 otherwise.
269 static int can_use_rp(struct ubifs_info *c)
271 if (uid_eq(current_fsuid(), c->rp_uid) || capable(CAP_SYS_RESOURCE) ||
272 (!gid_eq(c->rp_gid, GLOBAL_ROOT_GID) && in_group_p(c->rp_gid)))
278 * do_budget_space - reserve flash space for index and data growth.
279 * @c: UBIFS file-system description object
281 * This function makes sure UBIFS has enough free LEBs for index growth and
284 * When budgeting index space, UBIFS reserves thrice as many LEBs as the index
285 * would take if it was consolidated and written to the flash. This guarantees
286 * that the "in-the-gaps" commit method always succeeds and UBIFS will always
287 * be able to commit dirty index. So this function basically adds amount of
288 * budgeted index space to the size of the current index, multiplies this by 3,
289 * and makes sure this does not exceed the amount of free LEBs.
291 * Notes about @c->bi.min_idx_lebs and @c->lst.idx_lebs variables:
292 * o @c->lst.idx_lebs is the number of LEBs the index currently uses. It might
293 * be large, because UBIFS does not do any index consolidation as long as
294 * there is free space. IOW, the index may take a lot of LEBs, but the LEBs
295 * will contain a lot of dirt.
296 * o @c->bi.min_idx_lebs is the number of LEBS the index presumably takes. IOW,
297 * the index may be consolidated to take up to @c->bi.min_idx_lebs LEBs.
299 * This function returns zero in case of success, and %-ENOSPC in case of
302 static int do_budget_space(struct ubifs_info *c)
304 long long outstanding, available;
305 int lebs, rsvd_idx_lebs, min_idx_lebs;
307 /* First budget index space */
308 min_idx_lebs = ubifs_calc_min_idx_lebs(c);
310 /* Now 'min_idx_lebs' contains number of LEBs to reserve */
311 if (min_idx_lebs > c->lst.idx_lebs)
312 rsvd_idx_lebs = min_idx_lebs - c->lst.idx_lebs;
317 * The number of LEBs that are available to be used by the index is:
319 * @c->lst.empty_lebs + @c->freeable_cnt + @c->idx_gc_cnt -
320 * @c->lst.taken_empty_lebs
322 * @c->lst.empty_lebs are available because they are empty.
323 * @c->freeable_cnt are available because they contain only free and
324 * dirty space, @c->idx_gc_cnt are available because they are index
325 * LEBs that have been garbage collected and are awaiting the commit
326 * before they can be used. And the in-the-gaps method will grab these
327 * if it needs them. @c->lst.taken_empty_lebs are empty LEBs that have
328 * already been allocated for some purpose.
330 * Note, @c->idx_gc_cnt is included to both @c->lst.empty_lebs (because
331 * these LEBs are empty) and to @c->lst.taken_empty_lebs (because they
332 * are taken until after the commit).
334 * Note, @c->lst.taken_empty_lebs may temporarily be higher by one
335 * because of the way we serialize LEB allocations and budgeting. See a
336 * comment in 'ubifs_find_free_space()'.
338 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
339 c->lst.taken_empty_lebs;
340 if (unlikely(rsvd_idx_lebs > lebs)) {
341 dbg_budg("out of indexing space: min_idx_lebs %d (old %d), rsvd_idx_lebs %d",
342 min_idx_lebs, c->bi.min_idx_lebs, rsvd_idx_lebs);
346 available = ubifs_calc_available(c, min_idx_lebs);
347 outstanding = c->bi.data_growth + c->bi.dd_growth;
349 if (unlikely(available < outstanding)) {
350 dbg_budg("out of data space: available %lld, outstanding %lld",
351 available, outstanding);
355 if (available - outstanding <= c->rp_size && !can_use_rp(c))
358 c->bi.min_idx_lebs = min_idx_lebs;
363 * calc_idx_growth - calculate approximate index growth from budgeting request.
364 * @c: UBIFS file-system description object
365 * @req: budgeting request
367 * For now we assume each new node adds one znode. But this is rather poor
368 * approximation, though.
370 static int calc_idx_growth(const struct ubifs_info *c,
371 const struct ubifs_budget_req *req)
375 znodes = req->new_ino + (req->new_page << UBIFS_BLOCKS_PER_PAGE_SHIFT) +
377 return znodes * c->max_idx_node_sz;
381 * calc_data_growth - calculate approximate amount of new data from budgeting
383 * @c: UBIFS file-system description object
384 * @req: budgeting request
386 static int calc_data_growth(const struct ubifs_info *c,
387 const struct ubifs_budget_req *req)
391 data_growth = req->new_ino ? c->bi.inode_budget : 0;
393 data_growth += c->bi.page_budget;
395 data_growth += c->bi.dent_budget;
396 data_growth += req->new_ino_d;
401 * calc_dd_growth - calculate approximate amount of data which makes other data
402 * dirty from budgeting request.
403 * @c: UBIFS file-system description object
404 * @req: budgeting request
406 static int calc_dd_growth(const struct ubifs_info *c,
407 const struct ubifs_budget_req *req)
411 dd_growth = req->dirtied_page ? c->bi.page_budget : 0;
413 if (req->dirtied_ino)
414 dd_growth += c->bi.inode_budget << (req->dirtied_ino - 1);
416 dd_growth += c->bi.dent_budget;
417 dd_growth += req->dirtied_ino_d;
422 * ubifs_budget_space - ensure there is enough space to complete an operation.
423 * @c: UBIFS file-system description object
424 * @req: budget request
426 * This function allocates budget for an operation. It uses pessimistic
427 * approximation of how much flash space the operation needs. The goal of this
428 * function is to make sure UBIFS always has flash space to flush all dirty
429 * pages, dirty inodes, and dirty znodes (liability). This function may force
430 * commit, garbage-collection or write-back. Returns zero in case of success,
431 * %-ENOSPC if there is no free space and other negative error codes in case of
434 int ubifs_budget_space(struct ubifs_info *c, struct ubifs_budget_req *req)
436 int err, idx_growth, data_growth, dd_growth, retried = 0;
438 ubifs_assert(req->new_page <= 1);
439 ubifs_assert(req->dirtied_page <= 1);
440 ubifs_assert(req->new_dent <= 1);
441 ubifs_assert(req->mod_dent <= 1);
442 ubifs_assert(req->new_ino <= 1);
443 ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA);
444 ubifs_assert(req->dirtied_ino <= 4);
445 ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
446 ubifs_assert(!(req->new_ino_d & 7));
447 ubifs_assert(!(req->dirtied_ino_d & 7));
449 data_growth = calc_data_growth(c, req);
450 dd_growth = calc_dd_growth(c, req);
451 if (!data_growth && !dd_growth)
453 idx_growth = calc_idx_growth(c, req);
456 spin_lock(&c->space_lock);
457 ubifs_assert(c->bi.idx_growth >= 0);
458 ubifs_assert(c->bi.data_growth >= 0);
459 ubifs_assert(c->bi.dd_growth >= 0);
461 if (unlikely(c->bi.nospace) && (c->bi.nospace_rp || !can_use_rp(c))) {
462 dbg_budg("no space");
463 spin_unlock(&c->space_lock);
467 c->bi.idx_growth += idx_growth;
468 c->bi.data_growth += data_growth;
469 c->bi.dd_growth += dd_growth;
471 err = do_budget_space(c);
473 req->idx_growth = idx_growth;
474 req->data_growth = data_growth;
475 req->dd_growth = dd_growth;
476 spin_unlock(&c->space_lock);
480 /* Restore the old values */
481 c->bi.idx_growth -= idx_growth;
482 c->bi.data_growth -= data_growth;
483 c->bi.dd_growth -= dd_growth;
484 spin_unlock(&c->space_lock);
487 dbg_budg("no space for fast budgeting");
491 err = make_free_space(c);
493 if (err == -EAGAIN) {
494 dbg_budg("try again");
496 } else if (err == -ENOSPC) {
499 dbg_budg("-ENOSPC, but anyway try once again");
502 dbg_budg("FS is full, -ENOSPC");
504 if (can_use_rp(c) || c->rp_size == 0)
505 c->bi.nospace_rp = 1;
508 ubifs_err(c, "cannot budget space, error %d", err);
513 * ubifs_release_budget - release budgeted free space.
514 * @c: UBIFS file-system description object
515 * @req: budget request
517 * This function releases the space budgeted by 'ubifs_budget_space()'. Note,
518 * since the index changes (which were budgeted for in @req->idx_growth) will
519 * only be written to the media on commit, this function moves the index budget
520 * from @c->bi.idx_growth to @c->bi.uncommitted_idx. The latter will be zeroed
521 * by the commit operation.
523 void ubifs_release_budget(struct ubifs_info *c, struct ubifs_budget_req *req)
525 ubifs_assert(req->new_page <= 1);
526 ubifs_assert(req->dirtied_page <= 1);
527 ubifs_assert(req->new_dent <= 1);
528 ubifs_assert(req->mod_dent <= 1);
529 ubifs_assert(req->new_ino <= 1);
530 ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA);
531 ubifs_assert(req->dirtied_ino <= 4);
532 ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
533 ubifs_assert(!(req->new_ino_d & 7));
534 ubifs_assert(!(req->dirtied_ino_d & 7));
535 if (!req->recalculate) {
536 ubifs_assert(req->idx_growth >= 0);
537 ubifs_assert(req->data_growth >= 0);
538 ubifs_assert(req->dd_growth >= 0);
541 if (req->recalculate) {
542 req->data_growth = calc_data_growth(c, req);
543 req->dd_growth = calc_dd_growth(c, req);
544 req->idx_growth = calc_idx_growth(c, req);
547 if (!req->data_growth && !req->dd_growth)
550 c->bi.nospace = c->bi.nospace_rp = 0;
553 spin_lock(&c->space_lock);
554 c->bi.idx_growth -= req->idx_growth;
555 c->bi.uncommitted_idx += req->idx_growth;
556 c->bi.data_growth -= req->data_growth;
557 c->bi.dd_growth -= req->dd_growth;
558 c->bi.min_idx_lebs = ubifs_calc_min_idx_lebs(c);
560 ubifs_assert(c->bi.idx_growth >= 0);
561 ubifs_assert(c->bi.data_growth >= 0);
562 ubifs_assert(c->bi.dd_growth >= 0);
563 ubifs_assert(c->bi.min_idx_lebs < c->main_lebs);
564 ubifs_assert(!(c->bi.idx_growth & 7));
565 ubifs_assert(!(c->bi.data_growth & 7));
566 ubifs_assert(!(c->bi.dd_growth & 7));
567 spin_unlock(&c->space_lock);
571 * ubifs_convert_page_budget - convert budget of a new page.
572 * @c: UBIFS file-system description object
574 * This function converts budget which was allocated for a new page of data to
575 * the budget of changing an existing page of data. The latter is smaller than
576 * the former, so this function only does simple re-calculation and does not
577 * involve any write-back.
579 void ubifs_convert_page_budget(struct ubifs_info *c)
581 spin_lock(&c->space_lock);
582 /* Release the index growth reservation */
583 c->bi.idx_growth -= c->max_idx_node_sz << UBIFS_BLOCKS_PER_PAGE_SHIFT;
584 /* Release the data growth reservation */
585 c->bi.data_growth -= c->bi.page_budget;
586 /* Increase the dirty data growth reservation instead */
587 c->bi.dd_growth += c->bi.page_budget;
588 /* And re-calculate the indexing space reservation */
589 c->bi.min_idx_lebs = ubifs_calc_min_idx_lebs(c);
590 spin_unlock(&c->space_lock);
594 * ubifs_release_dirty_inode_budget - release dirty inode budget.
595 * @c: UBIFS file-system description object
596 * @ui: UBIFS inode to release the budget for
598 * This function releases budget corresponding to a dirty inode. It is usually
599 * called when after the inode has been written to the media and marked as
600 * clean. It also causes the "no space" flags to be cleared.
602 void ubifs_release_dirty_inode_budget(struct ubifs_info *c,
603 struct ubifs_inode *ui)
605 struct ubifs_budget_req req;
607 memset(&req, 0, sizeof(struct ubifs_budget_req));
608 /* The "no space" flags will be cleared because dd_growth is > 0 */
609 req.dd_growth = c->bi.inode_budget + ALIGN(ui->data_len, 8);
610 ubifs_release_budget(c, &req);
615 * ubifs_reported_space - calculate reported free space.
616 * @c: the UBIFS file-system description object
617 * @free: amount of free space
619 * This function calculates amount of free space which will be reported to
620 * user-space. User-space application tend to expect that if the file-system
621 * (e.g., via the 'statfs()' call) reports that it has N bytes available, they
622 * are able to write a file of size N. UBIFS attaches node headers to each data
623 * node and it has to write indexing nodes as well. This introduces additional
624 * overhead, and UBIFS has to report slightly less free space to meet the above
627 * This function assumes free space is made up of uncompressed data nodes and
628 * full index nodes (one per data node, tripled because we always allow enough
629 * space to write the index thrice).
631 * Note, the calculation is pessimistic, which means that most of the time
632 * UBIFS reports less space than it actually has.
634 long long ubifs_reported_space(const struct ubifs_info *c, long long free)
636 int divisor, factor, f;
639 * Reported space size is @free * X, where X is UBIFS block size
640 * divided by UBIFS block size + all overhead one data block
641 * introduces. The overhead is the node header + indexing overhead.
643 * Indexing overhead calculations are based on the following formula:
644 * I = N/(f - 1) + 1, where I - number of indexing nodes, N - number
645 * of data nodes, f - fanout. Because effective UBIFS fanout is twice
646 * as less than maximum fanout, we assume that each data node
647 * introduces 3 * @c->max_idx_node_sz / (@c->fanout/2 - 1) bytes.
648 * Note, the multiplier 3 is because UBIFS reserves thrice as more space
651 f = c->fanout > 3 ? c->fanout >> 1 : 2;
652 factor = UBIFS_BLOCK_SIZE;
653 divisor = UBIFS_MAX_DATA_NODE_SZ;
654 divisor += (c->max_idx_node_sz * 3) / (f - 1);
656 return div_u64(free, divisor);
661 * ubifs_get_free_space_nolock - return amount of free space.
662 * @c: UBIFS file-system description object
664 * This function calculates amount of free space to report to user-space.
666 * Because UBIFS may introduce substantial overhead (the index, node headers,
667 * alignment, wastage at the end of LEBs, etc), it cannot report real amount of
668 * free flash space it has (well, because not all dirty space is reclaimable,
669 * UBIFS does not actually know the real amount). If UBIFS did so, it would
670 * bread user expectations about what free space is. Users seem to accustomed
671 * to assume that if the file-system reports N bytes of free space, they would
672 * be able to fit a file of N bytes to the FS. This almost works for
673 * traditional file-systems, because they have way less overhead than UBIFS.
674 * So, to keep users happy, UBIFS tries to take the overhead into account.
676 long long ubifs_get_free_space_nolock(struct ubifs_info *c)
678 int rsvd_idx_lebs, lebs;
679 long long available, outstanding, free;
681 ubifs_assert(c->bi.min_idx_lebs == ubifs_calc_min_idx_lebs(c));
682 outstanding = c->bi.data_growth + c->bi.dd_growth;
683 available = ubifs_calc_available(c, c->bi.min_idx_lebs);
686 * When reporting free space to user-space, UBIFS guarantees that it is
687 * possible to write a file of free space size. This means that for
688 * empty LEBs we may use more precise calculations than
689 * 'ubifs_calc_available()' is using. Namely, we know that in empty
690 * LEBs we would waste only @c->leb_overhead bytes, not @c->dark_wm.
691 * Thus, amend the available space.
693 * Note, the calculations below are similar to what we have in
694 * 'do_budget_space()', so refer there for comments.
696 if (c->bi.min_idx_lebs > c->lst.idx_lebs)
697 rsvd_idx_lebs = c->bi.min_idx_lebs - c->lst.idx_lebs;
700 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
701 c->lst.taken_empty_lebs;
702 lebs -= rsvd_idx_lebs;
703 available += lebs * (c->dark_wm - c->leb_overhead);
705 if (available > outstanding)
706 free = ubifs_reported_space(c, available - outstanding);
713 * ubifs_get_free_space - return amount of free space.
714 * @c: UBIFS file-system description object
716 * This function calculates and returns amount of free space to report to
719 long long ubifs_get_free_space(struct ubifs_info *c)
723 spin_lock(&c->space_lock);
724 free = ubifs_get_free_space_nolock(c);
725 spin_unlock(&c->space_lock);