2 * This file is part of UBIFS.
4 * Copyright (C) 2006-2008 Nokia Corporation.
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.
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
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
19 * Authors: Adrian Hunter
20 * Artem Bityutskiy (Битюцкий Артём)
24 * This file implements the budgeting sub-system which is responsible for UBIFS
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.
34 #include <linux/writeback.h>
35 #include <linux/math64.h>
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.
43 #define MAX_MKSPC_RETRIES 3
46 * The below constant defines amount of dirty pages which should be written
47 * back at when trying to shrink the liability.
49 #define NR_TO_WRITE 16
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
56 * This function shrinks UBIFS liability by means of writing back some amount
57 * of dirty inodes and their pages.
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
63 static void shrink_liability(struct ubifs_info *c, int nr_to_write)
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);
71 * run_gc - run garbage collector.
72 * @c: UBIFS file-system description object
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.
78 static int run_gc(struct ubifs_info *c)
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);
89 /* GC freed one LEB, return it to lprops */
90 dbg_budg("GC freed LEB %d", lnum);
91 err = ubifs_return_leb(c, lnum);
98 * get_liability - calculate current liability.
99 * @c: UBIFS file-system description object
101 * This function calculates and returns current UBIFS liability, i.e. the
102 * amount of bytes UBIFS has "promised" to write to the media.
104 static long long get_liability(struct ubifs_info *c)
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);
115 * make_free_space - make more free space on the file-system.
116 * @c: UBIFS file-system description object
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
125 * o commit may free some LEB, i.e., turn freeable LEBs into free LEBs.
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
132 static int make_free_space(struct ubifs_info *c)
134 int err, retries = 0;
135 long long liab1, liab2;
138 liab1 = get_liability(c);
140 * We probably have some dirty pages or inodes (liability), try
141 * to write them back.
143 dbg_budg("liability %lld, run write-back", liab1);
144 shrink_liability(c, NR_TO_WRITE);
146 liab2 = get_liability(c);
150 dbg_budg("new liability %lld (not shrunk)", liab2);
152 /* Liability did not shrink again, try GC */
158 if (err != -EAGAIN && err != -ENOSPC)
159 /* Some real error happened */
162 dbg_budg("Run commit (retries %d)", retries);
163 err = ubifs_run_commit(c);
166 } while (retries++ < MAX_MKSPC_RETRIES);
172 * ubifs_calc_min_idx_lebs - calculate amount of LEBs for the index.
173 * @c: UBIFS file-system description object
175 * This function calculates and returns the number of LEBs which should be kept
178 int ubifs_calc_min_idx_lebs(struct ubifs_info *c)
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;
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.
191 idx_lebs = div_u64(idx_size + c->idx_leb_size - 1, c->idx_leb_size);
193 * The index head is not available for the in-the-gaps method, so add an
194 * extra LEB to compensate.
197 if (idx_lebs < MIN_INDEX_LEBS)
198 idx_lebs = MIN_INDEX_LEBS;
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
207 * This function calculates and returns amount of FS space available for use.
209 long long ubifs_calc_available(const struct ubifs_info *c, int min_idx_lebs)
214 available = c->main_bytes - c->lst.total_used;
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.
221 subtract_lebs = min_idx_lebs;
223 /* Take into account that GC reserves one LEB for its own needs */
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
231 subtract_lebs += c->jhead_cnt - 1;
233 /* We also reserve one LEB for deletions, which bypass budgeting */
236 available -= (long long)subtract_lebs * c->leb_size;
238 /* Subtract the dead space which is not available for use */
239 available -= c->lst.total_dead;
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.
247 available -= c->lst.total_dark;
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
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;
260 /* The calculations are rough and may end up with a negative number */
261 return available > 0 ? available : 0;
265 * can_use_rp - check whether the user is allowed to use reserved pool.
266 * @c: UBIFS file-system description object
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.
273 static int can_use_rp(struct ubifs_info *c)
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)))
282 * do_budget_space - reserve flash space for index and data growth.
283 * @c: UBIFS file-system description object
285 * This function makes sure UBIFS has enough free LEBs for index growth and
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.
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.
303 * This function returns zero in case of success, and %-ENOSPC in case of
306 static int do_budget_space(struct ubifs_info *c)
308 long long outstanding, available;
309 int lebs, rsvd_idx_lebs, min_idx_lebs;
311 /* First budget index space */
312 min_idx_lebs = ubifs_calc_min_idx_lebs(c);
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;
321 * The number of LEBs that are available to be used by the index is:
323 * @c->lst.empty_lebs + @c->freeable_cnt + @c->idx_gc_cnt -
324 * @c->lst.taken_empty_lebs
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.
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).
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()'.
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);
350 available = ubifs_calc_available(c, min_idx_lebs);
351 outstanding = c->bi.data_growth + c->bi.dd_growth;
353 if (unlikely(available < outstanding)) {
354 dbg_budg("out of data space: available %lld, outstanding %lld",
355 available, outstanding);
359 if (available - outstanding <= c->rp_size && !can_use_rp(c))
362 c->bi.min_idx_lebs = min_idx_lebs;
367 * calc_idx_growth - calculate approximate index growth from budgeting request.
368 * @c: UBIFS file-system description object
369 * @req: budgeting request
371 * For now we assume each new node adds one znode. But this is rather poor
372 * approximation, though.
374 static int calc_idx_growth(const struct ubifs_info *c,
375 const struct ubifs_budget_req *req)
379 znodes = req->new_ino + (req->new_page << UBIFS_BLOCKS_PER_PAGE_SHIFT) +
381 return znodes * c->max_idx_node_sz;
385 * calc_data_growth - calculate approximate amount of new data from budgeting
387 * @c: UBIFS file-system description object
388 * @req: budgeting request
390 static int calc_data_growth(const struct ubifs_info *c,
391 const struct ubifs_budget_req *req)
395 data_growth = req->new_ino ? c->bi.inode_budget : 0;
397 data_growth += c->bi.page_budget;
399 data_growth += c->bi.dent_budget;
400 data_growth += req->new_ino_d;
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
410 static int calc_dd_growth(const struct ubifs_info *c,
411 const struct ubifs_budget_req *req)
415 dd_growth = req->dirtied_page ? c->bi.page_budget : 0;
417 if (req->dirtied_ino)
418 dd_growth += c->bi.inode_budget << (req->dirtied_ino - 1);
420 dd_growth += c->bi.dent_budget;
421 dd_growth += req->dirtied_ino_d;
426 * ubifs_budget_space - ensure there is enough space to complete an operation.
427 * @c: UBIFS file-system description object
428 * @req: budget request
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
438 int ubifs_budget_space(struct ubifs_info *c, struct ubifs_budget_req *req)
440 int uninitialized_var(cmt_retries), uninitialized_var(wb_retries);
441 int err, idx_growth, data_growth, dd_growth, retried = 0;
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));
454 data_growth = calc_data_growth(c, req);
455 dd_growth = calc_dd_growth(c, req);
456 if (!data_growth && !dd_growth)
458 idx_growth = calc_idx_growth(c, req);
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);
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);
472 c->bi.idx_growth += idx_growth;
473 c->bi.data_growth += data_growth;
474 c->bi.dd_growth += dd_growth;
476 err = do_budget_space(c);
478 req->idx_growth = idx_growth;
479 req->data_growth = data_growth;
480 req->dd_growth = dd_growth;
481 spin_unlock(&c->space_lock);
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);
492 dbg_budg("no space for fast budgeting");
496 err = make_free_space(c);
498 if (err == -EAGAIN) {
499 dbg_budg("try again");
501 } else if (err == -ENOSPC) {
504 dbg_budg("-ENOSPC, but anyway try once again");
507 dbg_budg("FS is full, -ENOSPC");
509 if (can_use_rp(c) || c->rp_size == 0)
510 c->bi.nospace_rp = 1;
513 ubifs_err("cannot budget space, error %d", err);
518 * ubifs_release_budget - release budgeted free space.
519 * @c: UBIFS file-system description object
520 * @req: budget request
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.
528 void ubifs_release_budget(struct ubifs_info *c, struct ubifs_budget_req *req)
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);
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);
552 if (!req->data_growth && !req->dd_growth)
555 c->bi.nospace = c->bi.nospace_rp = 0;
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);
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);
576 * ubifs_convert_page_budget - convert budget of a new page.
577 * @c: UBIFS file-system description object
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.
584 void ubifs_convert_page_budget(struct ubifs_info *c)
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);
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
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.
607 void ubifs_release_dirty_inode_budget(struct ubifs_info *c,
608 struct ubifs_inode *ui)
610 struct ubifs_budget_req req;
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);
619 * ubifs_reported_space - calculate reported free space.
620 * @c: the UBIFS file-system description object
621 * @free: amount of free space
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
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).
635 * Note, the calculation is pessimistic, which means that most of the time
636 * UBIFS reports less space than it actually has.
638 long long ubifs_reported_space(const struct ubifs_info *c, long long free)
640 int divisor, factor, f;
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.
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
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);
660 return div_u64(free, divisor);
664 * ubifs_get_free_space_nolock - return amount of free space.
665 * @c: UBIFS file-system description object
667 * This function calculates amount of free space to report to user-space.
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.
679 long long ubifs_get_free_space_nolock(struct ubifs_info *c)
681 int rsvd_idx_lebs, lebs;
682 long long available, outstanding, free;
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);
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.
696 * Note, the calculations below are similar to what we have in
697 * 'do_budget_space()', so refer there for comments.
699 if (c->bi.min_idx_lebs > c->lst.idx_lebs)
700 rsvd_idx_lebs = c->bi.min_idx_lebs - c->lst.idx_lebs;
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);
708 if (available > outstanding)
709 free = ubifs_reported_space(c, available - outstanding);
716 * ubifs_get_free_space - return amount of free space.
717 * @c: UBIFS file-system description object
719 * This function calculates and returns amount of free space to report to
722 long long ubifs_get_free_space(struct ubifs_info *c)
726 spin_lock(&c->space_lock);
727 free = ubifs_get_free_space_nolock(c);
728 spin_unlock(&c->space_lock);