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.
25 #include <linux/writeback.h>
27 #include <linux/err.h>
29 #include <linux/math64.h>
32 * When pessimistic budget calculations say that there is no enough space,
33 * UBIFS starts writing back dirty inodes and pages, doing garbage collection,
34 * or committing. The below constant defines maximum number of times UBIFS
35 * repeats the operations.
37 #define MAX_MKSPC_RETRIES 3
40 * The below constant defines amount of dirty pages which should be written
41 * back at when trying to shrink the liability.
43 #define NR_TO_WRITE 16
47 * shrink_liability - write-back some dirty pages/inodes.
48 * @c: UBIFS file-system description object
49 * @nr_to_write: how many dirty pages to write-back
51 * This function shrinks UBIFS liability by means of writing back some amount
52 * of dirty inodes and their pages.
54 * Note, this function synchronizes even VFS inodes which are locked
55 * (@i_mutex) by the caller of the budgeting function, because write-back does
58 static void shrink_liability(struct ubifs_info *c, int nr_to_write)
60 down_read(&c->vfs_sb->s_umount);
61 writeback_inodes_sb(c->vfs_sb, WB_REASON_FS_FREE_SPACE);
62 up_read(&c->vfs_sb->s_umount);
66 * run_gc - run garbage collector.
67 * @c: UBIFS file-system description object
69 * This function runs garbage collector to make some more free space. Returns
70 * zero if a free LEB has been produced, %-EAGAIN if commit is required, and a
71 * negative error code in case of failure.
73 static int run_gc(struct ubifs_info *c)
77 /* Make some free space by garbage-collecting dirty space */
78 down_read(&c->commit_sem);
79 lnum = ubifs_garbage_collect(c, 1);
80 up_read(&c->commit_sem);
84 /* GC freed one LEB, return it to lprops */
85 dbg_budg("GC freed LEB %d", lnum);
86 err = ubifs_return_leb(c, lnum);
93 * get_liability - calculate current liability.
94 * @c: UBIFS file-system description object
96 * This function calculates and returns current UBIFS liability, i.e. the
97 * amount of bytes UBIFS has "promised" to write to the media.
99 static long long get_liability(struct ubifs_info *c)
103 spin_lock(&c->space_lock);
104 liab = c->bi.idx_growth + c->bi.data_growth + c->bi.dd_growth;
105 spin_unlock(&c->space_lock);
110 * make_free_space - make more free space on the file-system.
111 * @c: UBIFS file-system description object
113 * This function is called when an operation cannot be budgeted because there
114 * is supposedly no free space. But in most cases there is some free space:
115 * o budgeting is pessimistic, so it always budgets more than it is actually
116 * needed, so shrinking the liability is one way to make free space - the
117 * cached data will take less space then it was budgeted for;
118 * o GC may turn some dark space into free space (budgeting treats dark space
120 * o commit may free some LEB, i.e., turn freeable LEBs into free LEBs.
122 * So this function tries to do the above. Returns %-EAGAIN if some free space
123 * was presumably made and the caller has to re-try budgeting the operation.
124 * Returns %-ENOSPC if it couldn't do more free space, and other negative error
127 static int make_free_space(struct ubifs_info *c)
129 int err, retries = 0;
130 long long liab1, liab2;
133 liab1 = get_liability(c);
135 * We probably have some dirty pages or inodes (liability), try
136 * to write them back.
138 dbg_budg("liability %lld, run write-back", liab1);
139 shrink_liability(c, NR_TO_WRITE);
141 liab2 = get_liability(c);
145 dbg_budg("new liability %lld (not shrunk)", liab2);
147 /* Liability did not shrink again, try GC */
153 if (err != -EAGAIN && err != -ENOSPC)
154 /* Some real error happened */
157 dbg_budg("Run commit (retries %d)", retries);
158 err = ubifs_run_commit(c);
161 } while (retries++ < MAX_MKSPC_RETRIES);
168 * ubifs_calc_min_idx_lebs - calculate amount of LEBs for the index.
169 * @c: UBIFS file-system description object
171 * This function calculates and returns the number of LEBs which should be kept
174 int ubifs_calc_min_idx_lebs(struct ubifs_info *c)
179 idx_size = c->bi.old_idx_sz + c->bi.idx_growth + c->bi.uncommitted_idx;
180 /* And make sure we have thrice the index size of space reserved */
181 idx_size += idx_size << 1;
183 * We do not maintain 'old_idx_size' as 'old_idx_lebs'/'old_idx_bytes'
184 * pair, nor similarly the two variables for the new index size, so we
185 * have to do this costly 64-bit division on fast-path.
187 idx_lebs = div_u64(idx_size + c->idx_leb_size - 1, c->idx_leb_size);
189 * The index head is not available for the in-the-gaps method, so add an
190 * extra LEB to compensate.
193 if (idx_lebs < MIN_INDEX_LEBS)
194 idx_lebs = MIN_INDEX_LEBS;
200 * ubifs_calc_available - calculate available FS space.
201 * @c: UBIFS file-system description object
202 * @min_idx_lebs: minimum number of LEBs reserved for the index
204 * This function calculates and returns amount of FS space available for use.
206 long long ubifs_calc_available(const struct ubifs_info *c, int min_idx_lebs)
211 available = c->main_bytes - c->lst.total_used;
214 * Now 'available' contains theoretically available flash space
215 * assuming there is no index, so we have to subtract the space which
216 * is reserved for the index.
218 subtract_lebs = min_idx_lebs;
220 /* Take into account that GC reserves one LEB for its own needs */
224 * The GC journal head LEB is not really accessible. And since
225 * different write types go to different heads, we may count only on
228 subtract_lebs += c->jhead_cnt - 1;
230 /* We also reserve one LEB for deletions, which bypass budgeting */
233 available -= (long long)subtract_lebs * c->leb_size;
235 /* Subtract the dead space which is not available for use */
236 available -= c->lst.total_dead;
239 * Subtract dark space, which might or might not be usable - it depends
240 * on the data which we have on the media and which will be written. If
241 * this is a lot of uncompressed or not-compressible data, the dark
242 * space cannot be used.
244 available -= c->lst.total_dark;
247 * However, there is more dark space. The index may be bigger than
248 * @min_idx_lebs. Those extra LEBs are assumed to be available, but
249 * their dark space is not included in total_dark, so it is subtracted
252 if (c->lst.idx_lebs > min_idx_lebs) {
253 subtract_lebs = c->lst.idx_lebs - min_idx_lebs;
254 available -= subtract_lebs * c->dark_wm;
257 /* The calculations are rough and may end up with a negative number */
258 return available > 0 ? available : 0;
262 * can_use_rp - check whether the user is allowed to use reserved pool.
263 * @c: UBIFS file-system description object
265 * UBIFS has so-called "reserved pool" which is flash space reserved
266 * for the superuser and for uses whose UID/GID is recorded in UBIFS superblock.
267 * This function checks whether current user is allowed to use reserved pool.
268 * Returns %1 current user is allowed to use reserved pool and %0 otherwise.
270 static int can_use_rp(struct ubifs_info *c)
272 if (uid_eq(current_fsuid(), c->rp_uid) || capable(CAP_SYS_RESOURCE) ||
273 (!gid_eq(c->rp_gid, GLOBAL_ROOT_GID) && in_group_p(c->rp_gid)))
279 * do_budget_space - reserve flash space for index and data growth.
280 * @c: UBIFS file-system description object
282 * This function makes sure UBIFS has enough free LEBs for index growth and
285 * When budgeting index space, UBIFS reserves thrice as many LEBs as the index
286 * would take if it was consolidated and written to the flash. This guarantees
287 * that the "in-the-gaps" commit method always succeeds and UBIFS will always
288 * be able to commit dirty index. So this function basically adds amount of
289 * budgeted index space to the size of the current index, multiplies this by 3,
290 * and makes sure this does not exceed the amount of free LEBs.
292 * Notes about @c->bi.min_idx_lebs and @c->lst.idx_lebs variables:
293 * o @c->lst.idx_lebs is the number of LEBs the index currently uses. It might
294 * be large, because UBIFS does not do any index consolidation as long as
295 * there is free space. IOW, the index may take a lot of LEBs, but the LEBs
296 * will contain a lot of dirt.
297 * o @c->bi.min_idx_lebs is the number of LEBS the index presumably takes. IOW,
298 * the index may be consolidated to take up to @c->bi.min_idx_lebs LEBs.
300 * This function returns zero in case of success, and %-ENOSPC in case of
303 static int do_budget_space(struct ubifs_info *c)
305 long long outstanding, available;
306 int lebs, rsvd_idx_lebs, min_idx_lebs;
308 /* First budget index space */
309 min_idx_lebs = ubifs_calc_min_idx_lebs(c);
311 /* Now 'min_idx_lebs' contains number of LEBs to reserve */
312 if (min_idx_lebs > c->lst.idx_lebs)
313 rsvd_idx_lebs = min_idx_lebs - c->lst.idx_lebs;
318 * The number of LEBs that are available to be used by the index is:
320 * @c->lst.empty_lebs + @c->freeable_cnt + @c->idx_gc_cnt -
321 * @c->lst.taken_empty_lebs
323 * @c->lst.empty_lebs are available because they are empty.
324 * @c->freeable_cnt are available because they contain only free and
325 * dirty space, @c->idx_gc_cnt are available because they are index
326 * LEBs that have been garbage collected and are awaiting the commit
327 * before they can be used. And the in-the-gaps method will grab these
328 * if it needs them. @c->lst.taken_empty_lebs are empty LEBs that have
329 * already been allocated for some purpose.
331 * Note, @c->idx_gc_cnt is included to both @c->lst.empty_lebs (because
332 * these LEBs are empty) and to @c->lst.taken_empty_lebs (because they
333 * are taken until after the commit).
335 * Note, @c->lst.taken_empty_lebs may temporarily be higher by one
336 * because of the way we serialize LEB allocations and budgeting. See a
337 * comment in 'ubifs_find_free_space()'.
339 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
340 c->lst.taken_empty_lebs;
341 if (unlikely(rsvd_idx_lebs > lebs)) {
342 dbg_budg("out of indexing space: min_idx_lebs %d (old %d), rsvd_idx_lebs %d",
343 min_idx_lebs, c->bi.min_idx_lebs, rsvd_idx_lebs);
347 available = ubifs_calc_available(c, min_idx_lebs);
348 outstanding = c->bi.data_growth + c->bi.dd_growth;
350 if (unlikely(available < outstanding)) {
351 dbg_budg("out of data space: available %lld, outstanding %lld",
352 available, outstanding);
356 if (available - outstanding <= c->rp_size && !can_use_rp(c))
359 c->bi.min_idx_lebs = min_idx_lebs;
364 * calc_idx_growth - calculate approximate index growth from budgeting request.
365 * @c: UBIFS file-system description object
366 * @req: budgeting request
368 * For now we assume each new node adds one znode. But this is rather poor
369 * approximation, though.
371 static int calc_idx_growth(const struct ubifs_info *c,
372 const struct ubifs_budget_req *req)
376 znodes = req->new_ino + (req->new_page << UBIFS_BLOCKS_PER_PAGE_SHIFT) +
378 return znodes * c->max_idx_node_sz;
382 * calc_data_growth - calculate approximate amount of new data from budgeting
384 * @c: UBIFS file-system description object
385 * @req: budgeting request
387 static int calc_data_growth(const struct ubifs_info *c,
388 const struct ubifs_budget_req *req)
392 data_growth = req->new_ino ? c->bi.inode_budget : 0;
394 data_growth += c->bi.page_budget;
396 data_growth += c->bi.dent_budget;
397 data_growth += req->new_ino_d;
402 * calc_dd_growth - calculate approximate amount of data which makes other data
403 * dirty from budgeting request.
404 * @c: UBIFS file-system description object
405 * @req: budgeting request
407 static int calc_dd_growth(const struct ubifs_info *c,
408 const struct ubifs_budget_req *req)
412 dd_growth = req->dirtied_page ? c->bi.page_budget : 0;
414 if (req->dirtied_ino)
415 dd_growth += c->bi.inode_budget << (req->dirtied_ino - 1);
417 dd_growth += c->bi.dent_budget;
418 dd_growth += req->dirtied_ino_d;
423 * ubifs_budget_space - ensure there is enough space to complete an operation.
424 * @c: UBIFS file-system description object
425 * @req: budget request
427 * This function allocates budget for an operation. It uses pessimistic
428 * approximation of how much flash space the operation needs. The goal of this
429 * function is to make sure UBIFS always has flash space to flush all dirty
430 * pages, dirty inodes, and dirty znodes (liability). This function may force
431 * commit, garbage-collection or write-back. Returns zero in case of success,
432 * %-ENOSPC if there is no free space and other negative error codes in case of
435 int ubifs_budget_space(struct ubifs_info *c, struct ubifs_budget_req *req)
437 int uninitialized_var(cmt_retries), uninitialized_var(wb_retries);
438 int err, idx_growth, data_growth, dd_growth, retried = 0;
440 ubifs_assert(req->new_page <= 1);
441 ubifs_assert(req->dirtied_page <= 1);
442 ubifs_assert(req->new_dent <= 1);
443 ubifs_assert(req->mod_dent <= 1);
444 ubifs_assert(req->new_ino <= 1);
445 ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA);
446 ubifs_assert(req->dirtied_ino <= 4);
447 ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
448 ubifs_assert(!(req->new_ino_d & 7));
449 ubifs_assert(!(req->dirtied_ino_d & 7));
451 data_growth = calc_data_growth(c, req);
452 dd_growth = calc_dd_growth(c, req);
453 if (!data_growth && !dd_growth)
455 idx_growth = calc_idx_growth(c, req);
458 spin_lock(&c->space_lock);
459 ubifs_assert(c->bi.idx_growth >= 0);
460 ubifs_assert(c->bi.data_growth >= 0);
461 ubifs_assert(c->bi.dd_growth >= 0);
463 if (unlikely(c->bi.nospace) && (c->bi.nospace_rp || !can_use_rp(c))) {
464 dbg_budg("no space");
465 spin_unlock(&c->space_lock);
469 c->bi.idx_growth += idx_growth;
470 c->bi.data_growth += data_growth;
471 c->bi.dd_growth += dd_growth;
473 err = do_budget_space(c);
475 req->idx_growth = idx_growth;
476 req->data_growth = data_growth;
477 req->dd_growth = dd_growth;
478 spin_unlock(&c->space_lock);
482 /* Restore the old values */
483 c->bi.idx_growth -= idx_growth;
484 c->bi.data_growth -= data_growth;
485 c->bi.dd_growth -= dd_growth;
486 spin_unlock(&c->space_lock);
489 dbg_budg("no space for fast budgeting");
493 err = make_free_space(c);
495 if (err == -EAGAIN) {
496 dbg_budg("try again");
498 } else if (err == -ENOSPC) {
501 dbg_budg("-ENOSPC, but anyway try once again");
504 dbg_budg("FS is full, -ENOSPC");
506 if (can_use_rp(c) || c->rp_size == 0)
507 c->bi.nospace_rp = 1;
510 ubifs_err("cannot budget space, error %d", err);
515 * ubifs_release_budget - release budgeted free space.
516 * @c: UBIFS file-system description object
517 * @req: budget request
519 * This function releases the space budgeted by 'ubifs_budget_space()'. Note,
520 * since the index changes (which were budgeted for in @req->idx_growth) will
521 * only be written to the media on commit, this function moves the index budget
522 * from @c->bi.idx_growth to @c->bi.uncommitted_idx. The latter will be zeroed
523 * by the commit operation.
525 void ubifs_release_budget(struct ubifs_info *c, struct ubifs_budget_req *req)
527 ubifs_assert(req->new_page <= 1);
528 ubifs_assert(req->dirtied_page <= 1);
529 ubifs_assert(req->new_dent <= 1);
530 ubifs_assert(req->mod_dent <= 1);
531 ubifs_assert(req->new_ino <= 1);
532 ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA);
533 ubifs_assert(req->dirtied_ino <= 4);
534 ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
535 ubifs_assert(!(req->new_ino_d & 7));
536 ubifs_assert(!(req->dirtied_ino_d & 7));
537 if (!req->recalculate) {
538 ubifs_assert(req->idx_growth >= 0);
539 ubifs_assert(req->data_growth >= 0);
540 ubifs_assert(req->dd_growth >= 0);
543 if (req->recalculate) {
544 req->data_growth = calc_data_growth(c, req);
545 req->dd_growth = calc_dd_growth(c, req);
546 req->idx_growth = calc_idx_growth(c, req);
549 if (!req->data_growth && !req->dd_growth)
552 c->bi.nospace = c->bi.nospace_rp = 0;
555 spin_lock(&c->space_lock);
556 c->bi.idx_growth -= req->idx_growth;
557 c->bi.uncommitted_idx += req->idx_growth;
558 c->bi.data_growth -= req->data_growth;
559 c->bi.dd_growth -= req->dd_growth;
560 c->bi.min_idx_lebs = ubifs_calc_min_idx_lebs(c);
562 ubifs_assert(c->bi.idx_growth >= 0);
563 ubifs_assert(c->bi.data_growth >= 0);
564 ubifs_assert(c->bi.dd_growth >= 0);
565 ubifs_assert(c->bi.min_idx_lebs < c->main_lebs);
566 ubifs_assert(!(c->bi.idx_growth & 7));
567 ubifs_assert(!(c->bi.data_growth & 7));
568 ubifs_assert(!(c->bi.dd_growth & 7));
569 spin_unlock(&c->space_lock);
573 * ubifs_convert_page_budget - convert budget of a new page.
574 * @c: UBIFS file-system description object
576 * This function converts budget which was allocated for a new page of data to
577 * the budget of changing an existing page of data. The latter is smaller than
578 * the former, so this function only does simple re-calculation and does not
579 * involve any write-back.
581 void ubifs_convert_page_budget(struct ubifs_info *c)
583 spin_lock(&c->space_lock);
584 /* Release the index growth reservation */
585 c->bi.idx_growth -= c->max_idx_node_sz << UBIFS_BLOCKS_PER_PAGE_SHIFT;
586 /* Release the data growth reservation */
587 c->bi.data_growth -= c->bi.page_budget;
588 /* Increase the dirty data growth reservation instead */
589 c->bi.dd_growth += c->bi.page_budget;
590 /* And re-calculate the indexing space reservation */
591 c->bi.min_idx_lebs = ubifs_calc_min_idx_lebs(c);
592 spin_unlock(&c->space_lock);
596 * ubifs_release_dirty_inode_budget - release dirty inode budget.
597 * @c: UBIFS file-system description object
598 * @ui: UBIFS inode to release the budget for
600 * This function releases budget corresponding to a dirty inode. It is usually
601 * called when after the inode has been written to the media and marked as
602 * clean. It also causes the "no space" flags to be cleared.
604 void ubifs_release_dirty_inode_budget(struct ubifs_info *c,
605 struct ubifs_inode *ui)
607 struct ubifs_budget_req req;
609 memset(&req, 0, sizeof(struct ubifs_budget_req));
610 /* The "no space" flags will be cleared because dd_growth is > 0 */
611 req.dd_growth = c->bi.inode_budget + ALIGN(ui->data_len, 8);
612 ubifs_release_budget(c, &req);
617 * ubifs_reported_space - calculate reported free space.
618 * @c: the UBIFS file-system description object
619 * @free: amount of free space
621 * This function calculates amount of free space which will be reported to
622 * user-space. User-space application tend to expect that if the file-system
623 * (e.g., via the 'statfs()' call) reports that it has N bytes available, they
624 * are able to write a file of size N. UBIFS attaches node headers to each data
625 * node and it has to write indexing nodes as well. This introduces additional
626 * overhead, and UBIFS has to report slightly less free space to meet the above
629 * This function assumes free space is made up of uncompressed data nodes and
630 * full index nodes (one per data node, tripled because we always allow enough
631 * space to write the index thrice).
633 * Note, the calculation is pessimistic, which means that most of the time
634 * UBIFS reports less space than it actually has.
636 long long ubifs_reported_space(const struct ubifs_info *c, long long free)
638 int divisor, factor, f;
641 * Reported space size is @free * X, where X is UBIFS block size
642 * divided by UBIFS block size + all overhead one data block
643 * introduces. The overhead is the node header + indexing overhead.
645 * Indexing overhead calculations are based on the following formula:
646 * I = N/(f - 1) + 1, where I - number of indexing nodes, N - number
647 * of data nodes, f - fanout. Because effective UBIFS fanout is twice
648 * as less than maximum fanout, we assume that each data node
649 * introduces 3 * @c->max_idx_node_sz / (@c->fanout/2 - 1) bytes.
650 * Note, the multiplier 3 is because UBIFS reserves thrice as more space
653 f = c->fanout > 3 ? c->fanout >> 1 : 2;
654 factor = UBIFS_BLOCK_SIZE;
655 divisor = UBIFS_MAX_DATA_NODE_SZ;
656 divisor += (c->max_idx_node_sz * 3) / (f - 1);
658 return div_u64(free, divisor);
663 * ubifs_get_free_space_nolock - return amount of free space.
664 * @c: UBIFS file-system description object
666 * This function calculates amount of free space to report to user-space.
668 * Because UBIFS may introduce substantial overhead (the index, node headers,
669 * alignment, wastage at the end of LEBs, etc), it cannot report real amount of
670 * free flash space it has (well, because not all dirty space is reclaimable,
671 * UBIFS does not actually know the real amount). If UBIFS did so, it would
672 * bread user expectations about what free space is. Users seem to accustomed
673 * to assume that if the file-system reports N bytes of free space, they would
674 * be able to fit a file of N bytes to the FS. This almost works for
675 * traditional file-systems, because they have way less overhead than UBIFS.
676 * So, to keep users happy, UBIFS tries to take the overhead into account.
678 long long ubifs_get_free_space_nolock(struct ubifs_info *c)
680 int rsvd_idx_lebs, lebs;
681 long long available, outstanding, free;
683 ubifs_assert(c->bi.min_idx_lebs == ubifs_calc_min_idx_lebs(c));
684 outstanding = c->bi.data_growth + c->bi.dd_growth;
685 available = ubifs_calc_available(c, c->bi.min_idx_lebs);
688 * When reporting free space to user-space, UBIFS guarantees that it is
689 * possible to write a file of free space size. This means that for
690 * empty LEBs we may use more precise calculations than
691 * 'ubifs_calc_available()' is using. Namely, we know that in empty
692 * LEBs we would waste only @c->leb_overhead bytes, not @c->dark_wm.
693 * Thus, amend the available space.
695 * Note, the calculations below are similar to what we have in
696 * 'do_budget_space()', so refer there for comments.
698 if (c->bi.min_idx_lebs > c->lst.idx_lebs)
699 rsvd_idx_lebs = c->bi.min_idx_lebs - c->lst.idx_lebs;
702 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
703 c->lst.taken_empty_lebs;
704 lebs -= rsvd_idx_lebs;
705 available += lebs * (c->dark_wm - c->leb_overhead);
707 if (available > outstanding)
708 free = ubifs_reported_space(c, available - outstanding);
715 * ubifs_get_free_space - return amount of free space.
716 * @c: UBIFS file-system description object
718 * This function calculates and returns amount of free space to report to
721 long long ubifs_get_free_space(struct ubifs_info *c)
725 spin_lock(&c->space_lock);
726 free = ubifs_get_free_space_nolock(c);
727 spin_unlock(&c->space_lock);