blob: 1a4973e106643dd2badce2eaeb805f82be0744a4 [file] [log] [blame]
Artem Bityutskiy1e517642008-07-14 19:08:37 +03001/*
2 * This file is part of UBIFS.
3 *
4 * Copyright (C) 2006-2008 Nokia Corporation.
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published by
8 * the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
13 * more details.
14 *
15 * You should have received a copy of the GNU General Public License along with
16 * this program; if not, write to the Free Software Foundation, Inc., 51
17 * Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
18 *
19 * Authors: Adrian Hunter
20 * Artem Bityutskiy (Битюцкий Артём)
21 */
22
23/*
24 * This file implements the budgeting sub-system which is responsible for UBIFS
25 * space management.
26 *
27 * Factors such as compression, wasted space at the ends of LEBs, space in other
28 * journal heads, the effect of updates on the index, and so on, make it
29 * impossible to accurately predict the amount of space needed. Consequently
30 * approximations are used.
31 */
32
33#include "ubifs.h"
34#include <linux/writeback.h>
35#include <asm/div64.h>
36
37/*
38 * When pessimistic budget calculations say that there is no enough space,
39 * UBIFS starts writing back dirty inodes and pages, doing garbage collection,
40 * or committing. The below constants define maximum number of times UBIFS
41 * repeats the operations.
42 */
43#define MAX_SHRINK_RETRIES 8
44#define MAX_GC_RETRIES 4
45#define MAX_CMT_RETRIES 2
46#define MAX_NOSPC_RETRIES 1
47
48/*
49 * The below constant defines amount of dirty pages which should be written
50 * back at when trying to shrink the liability.
51 */
52#define NR_TO_WRITE 16
53
54/**
55 * struct retries_info - information about re-tries while making free space.
56 * @prev_liability: previous liability
57 * @shrink_cnt: how many times the liability was shrinked
58 * @shrink_retries: count of liability shrink re-tries (increased when
59 * liability does not shrink)
60 * @try_gc: GC should be tried first
61 * @gc_retries: how many times GC was run
62 * @cmt_retries: how many times commit has been done
63 * @nospc_retries: how many times GC returned %-ENOSPC
64 *
65 * Since we consider budgeting to be the fast-path, and this structure has to
66 * be allocated on stack and zeroed out, we make it smaller using bit-fields.
67 */
68struct retries_info {
69 long long prev_liability;
70 unsigned int shrink_cnt;
71 unsigned int shrink_retries:5;
72 unsigned int try_gc:1;
73 unsigned int gc_retries:4;
74 unsigned int cmt_retries:3;
75 unsigned int nospc_retries:1;
76};
77
78/**
79 * shrink_liability - write-back some dirty pages/inodes.
80 * @c: UBIFS file-system description object
81 * @nr_to_write: how many dirty pages to write-back
82 *
83 * This function shrinks UBIFS liability by means of writing back some amount
84 * of dirty inodes and their pages. Returns the amount of pages which were
85 * written back. The returned value does not include dirty inodes which were
86 * synchronized.
87 *
88 * Note, this function synchronizes even VFS inodes which are locked
89 * (@i_mutex) by the caller of the budgeting function, because write-back does
90 * not touch @i_mutex.
91 */
92static int shrink_liability(struct ubifs_info *c, int nr_to_write)
93{
94 int nr_written;
95 struct writeback_control wbc = {
96 .sync_mode = WB_SYNC_NONE,
97 .range_end = LLONG_MAX,
98 .nr_to_write = nr_to_write,
99 };
100
101 generic_sync_sb_inodes(c->vfs_sb, &wbc);
102 nr_written = nr_to_write - wbc.nr_to_write;
103
104 if (!nr_written) {
105 /*
106 * Re-try again but wait on pages/inodes which are being
107 * written-back concurrently (e.g., by pdflush).
108 */
109 memset(&wbc, 0, sizeof(struct writeback_control));
110 wbc.sync_mode = WB_SYNC_ALL;
111 wbc.range_end = LLONG_MAX;
112 wbc.nr_to_write = nr_to_write;
113 generic_sync_sb_inodes(c->vfs_sb, &wbc);
114 nr_written = nr_to_write - wbc.nr_to_write;
115 }
116
117 dbg_budg("%d pages were written back", nr_written);
118 return nr_written;
119}
120
121
122/**
123 * run_gc - run garbage collector.
124 * @c: UBIFS file-system description object
125 *
126 * This function runs garbage collector to make some more free space. Returns
127 * zero if a free LEB has been produced, %-EAGAIN if commit is required, and a
128 * negative error code in case of failure.
129 */
130static int run_gc(struct ubifs_info *c)
131{
132 int err, lnum;
133
134 /* Make some free space by garbage-collecting dirty space */
135 down_read(&c->commit_sem);
136 lnum = ubifs_garbage_collect(c, 1);
137 up_read(&c->commit_sem);
138 if (lnum < 0)
139 return lnum;
140
141 /* GC freed one LEB, return it to lprops */
142 dbg_budg("GC freed LEB %d", lnum);
143 err = ubifs_return_leb(c, lnum);
144 if (err)
145 return err;
146 return 0;
147}
148
149/**
150 * make_free_space - make more free space on the file-system.
151 * @c: UBIFS file-system description object
152 * @ri: information about previous invocations of this function
153 *
154 * This function is called when an operation cannot be budgeted because there
155 * is supposedly no free space. But in most cases there is some free space:
156 * o budgeting is pessimistic, so it always budgets more then it is actually
157 * needed, so shrinking the liability is one way to make free space - the
158 * cached data will take less space then it was budgeted for;
159 * o GC may turn some dark space into free space (budgeting treats dark space
160 * as not available);
161 * o commit may free some LEB, i.e., turn freeable LEBs into free LEBs.
162 *
163 * So this function tries to do the above. Returns %-EAGAIN if some free space
164 * was presumably made and the caller has to re-try budgeting the operation.
165 * Returns %-ENOSPC if it couldn't do more free space, and other negative error
166 * codes on failures.
167 */
168static int make_free_space(struct ubifs_info *c, struct retries_info *ri)
169{
170 int err;
171
172 /*
173 * If we have some dirty pages and inodes (liability), try to write
174 * them back unless this was tried too many times without effect
175 * already.
176 */
177 if (ri->shrink_retries < MAX_SHRINK_RETRIES && !ri->try_gc) {
178 long long liability;
179
180 spin_lock(&c->space_lock);
181 liability = c->budg_idx_growth + c->budg_data_growth +
182 c->budg_dd_growth;
183 spin_unlock(&c->space_lock);
184
185 if (ri->prev_liability >= liability) {
186 /* Liability does not shrink, next time try GC then */
187 ri->shrink_retries += 1;
188 if (ri->gc_retries < MAX_GC_RETRIES)
189 ri->try_gc = 1;
190 dbg_budg("liability did not shrink: retries %d of %d",
191 ri->shrink_retries, MAX_SHRINK_RETRIES);
192 }
193
194 dbg_budg("force write-back (count %d)", ri->shrink_cnt);
195 shrink_liability(c, NR_TO_WRITE + ri->shrink_cnt);
196
197 ri->prev_liability = liability;
198 ri->shrink_cnt += 1;
199 return -EAGAIN;
200 }
201
202 /*
203 * Try to run garbage collector unless it was already tried too many
204 * times.
205 */
206 if (ri->gc_retries < MAX_GC_RETRIES) {
207 ri->gc_retries += 1;
208 dbg_budg("run GC, retries %d of %d",
209 ri->gc_retries, MAX_GC_RETRIES);
210
211 ri->try_gc = 0;
212 err = run_gc(c);
213 if (!err)
214 return -EAGAIN;
215
216 if (err == -EAGAIN) {
217 dbg_budg("GC asked to commit");
218 err = ubifs_run_commit(c);
219 if (err)
220 return err;
221 return -EAGAIN;
222 }
223
224 if (err != -ENOSPC)
225 return err;
226
227 /*
228 * GC could not make any progress. If this is the first time,
229 * then it makes sense to try to commit, because it might make
230 * some dirty space.
231 */
232 dbg_budg("GC returned -ENOSPC, retries %d",
233 ri->nospc_retries);
234 if (ri->nospc_retries >= MAX_NOSPC_RETRIES)
235 return err;
236 ri->nospc_retries += 1;
237 }
238
239 /* Neither GC nor write-back helped, try to commit */
240 if (ri->cmt_retries < MAX_CMT_RETRIES) {
241 ri->cmt_retries += 1;
242 dbg_budg("run commit, retries %d of %d",
243 ri->cmt_retries, MAX_CMT_RETRIES);
244 err = ubifs_run_commit(c);
245 if (err)
246 return err;
247 return -EAGAIN;
248 }
249 return -ENOSPC;
250}
251
252/**
253 * ubifs_calc_min_idx_lebs - calculate amount of eraseblocks for the index.
254 * @c: UBIFS file-system description object
255 *
256 * This function calculates and returns the number of eraseblocks which should
257 * be kept for index usage.
258 */
259int ubifs_calc_min_idx_lebs(struct ubifs_info *c)
260{
261 int ret;
262 uint64_t idx_size;
263
264 idx_size = c->old_idx_sz + c->budg_idx_growth + c->budg_uncommitted_idx;
265
Adrian Hunter3a132522008-07-30 12:18:02 +0300266 /* And make sure we have thrice the index size of space reserved */
Artem Bityutskiyb364b412008-07-25 14:38:51 +0300267 idx_size = idx_size + (idx_size << 1);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300268
269 /*
270 * We do not maintain 'old_idx_size' as 'old_idx_lebs'/'old_idx_bytes'
271 * pair, nor similarly the two variables for the new index size, so we
272 * have to do this costly 64-bit division on fast-path.
273 */
274 if (do_div(idx_size, c->leb_size - c->max_idx_node_sz))
275 ret = idx_size + 1;
276 else
277 ret = idx_size;
278 /*
279 * The index head is not available for the in-the-gaps method, so add an
280 * extra LEB to compensate.
281 */
282 ret += 1;
283 /*
284 * At present the index needs at least 2 LEBs: one for the index head
285 * and one for in-the-gaps method (which currently does not cater for
286 * the index head and so excludes it from consideration).
287 */
288 if (ret < 2)
289 ret = 2;
290 return ret;
291}
292
293/**
294 * ubifs_calc_available - calculate available FS space.
295 * @c: UBIFS file-system description object
296 * @min_idx_lebs: minimum number of LEBs reserved for the index
297 *
298 * This function calculates and returns amount of FS space available for use.
299 */
300long long ubifs_calc_available(const struct ubifs_info *c, int min_idx_lebs)
301{
302 int subtract_lebs;
303 long long available;
304
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300305 available = c->main_bytes - c->lst.total_used;
306
307 /*
308 * Now 'available' contains theoretically available flash space
309 * assuming there is no index, so we have to subtract the space which
310 * is reserved for the index.
311 */
312 subtract_lebs = min_idx_lebs;
313
314 /* Take into account that GC reserves one LEB for its own needs */
315 subtract_lebs += 1;
316
317 /*
318 * The GC journal head LEB is not really accessible. And since
319 * different write types go to different heads, we may count only on
320 * one head's space.
321 */
322 subtract_lebs += c->jhead_cnt - 1;
323
324 /* We also reserve one LEB for deletions, which bypass budgeting */
325 subtract_lebs += 1;
326
327 available -= (long long)subtract_lebs * c->leb_size;
328
329 /* Subtract the dead space which is not available for use */
330 available -= c->lst.total_dead;
331
332 /*
333 * Subtract dark space, which might or might not be usable - it depends
334 * on the data which we have on the media and which will be written. If
335 * this is a lot of uncompressed or not-compressible data, the dark
336 * space cannot be used.
337 */
338 available -= c->lst.total_dark;
339
340 /*
341 * However, there is more dark space. The index may be bigger than
342 * @min_idx_lebs. Those extra LEBs are assumed to be available, but
343 * their dark space is not included in total_dark, so it is subtracted
344 * here.
345 */
346 if (c->lst.idx_lebs > min_idx_lebs) {
347 subtract_lebs = c->lst.idx_lebs - min_idx_lebs;
348 available -= subtract_lebs * c->dark_wm;
349 }
350
351 /* The calculations are rough and may end up with a negative number */
352 return available > 0 ? available : 0;
353}
354
355/**
356 * can_use_rp - check whether the user is allowed to use reserved pool.
357 * @c: UBIFS file-system description object
358 *
359 * UBIFS has so-called "reserved pool" which is flash space reserved
360 * for the superuser and for uses whose UID/GID is recorded in UBIFS superblock.
361 * This function checks whether current user is allowed to use reserved pool.
362 * Returns %1 current user is allowed to use reserved pool and %0 otherwise.
363 */
364static int can_use_rp(struct ubifs_info *c)
365{
366 if (current->fsuid == c->rp_uid || capable(CAP_SYS_RESOURCE) ||
367 (c->rp_gid != 0 && in_group_p(c->rp_gid)))
368 return 1;
369 return 0;
370}
371
372/**
373 * do_budget_space - reserve flash space for index and data growth.
374 * @c: UBIFS file-system description object
375 *
376 * This function makes sure UBIFS has enough free eraseblocks for index growth
377 * and data.
378 *
Adrian Hunter3a132522008-07-30 12:18:02 +0300379 * When budgeting index space, UBIFS reserves thrice as many LEBs as the index
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300380 * would take if it was consolidated and written to the flash. This guarantees
381 * that the "in-the-gaps" commit method always succeeds and UBIFS will always
382 * be able to commit dirty index. So this function basically adds amount of
Artem Bityutskiyb364b412008-07-25 14:38:51 +0300383 * budgeted index space to the size of the current index, multiplies this by 3,
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300384 * and makes sure this does not exceed the amount of free eraseblocks.
385 *
386 * Notes about @c->min_idx_lebs and @c->lst.idx_lebs variables:
387 * o @c->lst.idx_lebs is the number of LEBs the index currently uses. It might
388 * be large, because UBIFS does not do any index consolidation as long as
389 * there is free space. IOW, the index may take a lot of LEBs, but the LEBs
390 * will contain a lot of dirt.
391 * o @c->min_idx_lebs is the the index presumably takes. IOW, the index may be
392 * consolidated to take up to @c->min_idx_lebs LEBs.
393 *
394 * This function returns zero in case of success, and %-ENOSPC in case of
395 * failure.
396 */
397static int do_budget_space(struct ubifs_info *c)
398{
399 long long outstanding, available;
400 int lebs, rsvd_idx_lebs, min_idx_lebs;
401
402 /* First budget index space */
403 min_idx_lebs = ubifs_calc_min_idx_lebs(c);
404
405 /* Now 'min_idx_lebs' contains number of LEBs to reserve */
406 if (min_idx_lebs > c->lst.idx_lebs)
407 rsvd_idx_lebs = min_idx_lebs - c->lst.idx_lebs;
408 else
409 rsvd_idx_lebs = 0;
410
411 /*
412 * The number of LEBs that are available to be used by the index is:
413 *
414 * @c->lst.empty_lebs + @c->freeable_cnt + @c->idx_gc_cnt -
415 * @c->lst.taken_empty_lebs
416 *
Artem Bityutskiy948cfb22008-08-20 11:56:33 +0300417 * @c->lst.empty_lebs are available because they are empty.
418 * @c->freeable_cnt are available because they contain only free and
419 * dirty space, @c->idx_gc_cnt are available because they are index
420 * LEBs that have been garbage collected and are awaiting the commit
421 * before they can be used. And the in-the-gaps method will grab these
422 * if it needs them. @c->lst.taken_empty_lebs are empty LEBs that have
423 * already been allocated for some purpose.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300424 *
Artem Bityutskiy948cfb22008-08-20 11:56:33 +0300425 * Note, @c->idx_gc_cnt is included to both @c->lst.empty_lebs (because
426 * these LEBs are empty) and to @c->lst.taken_empty_lebs (because they
427 * are taken until after the commit).
428 *
429 * Note, @c->lst.taken_empty_lebs may temporarily be higher by one
430 * because of the way we serialize LEB allocations and budgeting. See a
431 * comment in 'ubifs_find_free_space()'.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300432 */
433 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
434 c->lst.taken_empty_lebs;
435 if (unlikely(rsvd_idx_lebs > lebs)) {
436 dbg_budg("out of indexing space: min_idx_lebs %d (old %d), "
437 "rsvd_idx_lebs %d", min_idx_lebs, c->min_idx_lebs,
438 rsvd_idx_lebs);
439 return -ENOSPC;
440 }
441
442 available = ubifs_calc_available(c, min_idx_lebs);
443 outstanding = c->budg_data_growth + c->budg_dd_growth;
444
445 if (unlikely(available < outstanding)) {
446 dbg_budg("out of data space: available %lld, outstanding %lld",
447 available, outstanding);
448 return -ENOSPC;
449 }
450
451 if (available - outstanding <= c->rp_size && !can_use_rp(c))
452 return -ENOSPC;
453
454 c->min_idx_lebs = min_idx_lebs;
455 return 0;
456}
457
458/**
459 * calc_idx_growth - calculate approximate index growth from budgeting request.
460 * @c: UBIFS file-system description object
461 * @req: budgeting request
462 *
463 * For now we assume each new node adds one znode. But this is rather poor
464 * approximation, though.
465 */
466static int calc_idx_growth(const struct ubifs_info *c,
467 const struct ubifs_budget_req *req)
468{
469 int znodes;
470
471 znodes = req->new_ino + (req->new_page << UBIFS_BLOCKS_PER_PAGE_SHIFT) +
472 req->new_dent;
473 return znodes * c->max_idx_node_sz;
474}
475
476/**
477 * calc_data_growth - calculate approximate amount of new data from budgeting
478 * request.
479 * @c: UBIFS file-system description object
480 * @req: budgeting request
481 */
482static int calc_data_growth(const struct ubifs_info *c,
483 const struct ubifs_budget_req *req)
484{
485 int data_growth;
486
487 data_growth = req->new_ino ? c->inode_budget : 0;
488 if (req->new_page)
489 data_growth += c->page_budget;
490 if (req->new_dent)
491 data_growth += c->dent_budget;
492 data_growth += req->new_ino_d;
493 return data_growth;
494}
495
496/**
497 * calc_dd_growth - calculate approximate amount of data which makes other data
498 * dirty from budgeting request.
499 * @c: UBIFS file-system description object
500 * @req: budgeting request
501 */
502static int calc_dd_growth(const struct ubifs_info *c,
503 const struct ubifs_budget_req *req)
504{
505 int dd_growth;
506
507 dd_growth = req->dirtied_page ? c->page_budget : 0;
508
509 if (req->dirtied_ino)
510 dd_growth += c->inode_budget << (req->dirtied_ino - 1);
511 if (req->mod_dent)
512 dd_growth += c->dent_budget;
513 dd_growth += req->dirtied_ino_d;
514 return dd_growth;
515}
516
517/**
518 * ubifs_budget_space - ensure there is enough space to complete an operation.
519 * @c: UBIFS file-system description object
520 * @req: budget request
521 *
522 * This function allocates budget for an operation. It uses pessimistic
523 * approximation of how much flash space the operation needs. The goal of this
524 * function is to make sure UBIFS always has flash space to flush all dirty
525 * pages, dirty inodes, and dirty znodes (liability). This function may force
526 * commit, garbage-collection or write-back. Returns zero in case of success,
527 * %-ENOSPC if there is no free space and other negative error codes in case of
528 * failures.
529 */
530int ubifs_budget_space(struct ubifs_info *c, struct ubifs_budget_req *req)
531{
532 int uninitialized_var(cmt_retries), uninitialized_var(wb_retries);
533 int err, idx_growth, data_growth, dd_growth;
534 struct retries_info ri;
535
Artem Bityutskiy547000d2008-07-24 14:42:05 +0300536 ubifs_assert(req->new_page <= 1);
537 ubifs_assert(req->dirtied_page <= 1);
538 ubifs_assert(req->new_dent <= 1);
539 ubifs_assert(req->mod_dent <= 1);
540 ubifs_assert(req->new_ino <= 1);
541 ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300542 ubifs_assert(req->dirtied_ino <= 4);
543 ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300544 ubifs_assert(!(req->new_ino_d & 7));
545 ubifs_assert(!(req->dirtied_ino_d & 7));
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300546
547 data_growth = calc_data_growth(c, req);
548 dd_growth = calc_dd_growth(c, req);
549 if (!data_growth && !dd_growth)
550 return 0;
551 idx_growth = calc_idx_growth(c, req);
552 memset(&ri, 0, sizeof(struct retries_info));
553
554again:
555 spin_lock(&c->space_lock);
556 ubifs_assert(c->budg_idx_growth >= 0);
557 ubifs_assert(c->budg_data_growth >= 0);
558 ubifs_assert(c->budg_dd_growth >= 0);
559
560 if (unlikely(c->nospace) && (c->nospace_rp || !can_use_rp(c))) {
561 dbg_budg("no space");
562 spin_unlock(&c->space_lock);
563 return -ENOSPC;
564 }
565
566 c->budg_idx_growth += idx_growth;
567 c->budg_data_growth += data_growth;
568 c->budg_dd_growth += dd_growth;
569
570 err = do_budget_space(c);
571 if (likely(!err)) {
572 req->idx_growth = idx_growth;
573 req->data_growth = data_growth;
574 req->dd_growth = dd_growth;
575 spin_unlock(&c->space_lock);
576 return 0;
577 }
578
579 /* Restore the old values */
580 c->budg_idx_growth -= idx_growth;
581 c->budg_data_growth -= data_growth;
582 c->budg_dd_growth -= dd_growth;
583 spin_unlock(&c->space_lock);
584
585 if (req->fast) {
586 dbg_budg("no space for fast budgeting");
587 return err;
588 }
589
590 err = make_free_space(c, &ri);
591 if (err == -EAGAIN) {
592 dbg_budg("try again");
593 cond_resched();
594 goto again;
595 } else if (err == -ENOSPC) {
596 dbg_budg("FS is full, -ENOSPC");
597 c->nospace = 1;
598 if (can_use_rp(c) || c->rp_size == 0)
599 c->nospace_rp = 1;
600 smp_wmb();
601 } else
602 ubifs_err("cannot budget space, error %d", err);
603 return err;
604}
605
606/**
607 * ubifs_release_budget - release budgeted free space.
608 * @c: UBIFS file-system description object
609 * @req: budget request
610 *
611 * This function releases the space budgeted by 'ubifs_budget_space()'. Note,
612 * since the index changes (which were budgeted for in @req->idx_growth) will
613 * only be written to the media on commit, this function moves the index budget
614 * from @c->budg_idx_growth to @c->budg_uncommitted_idx. The latter will be
615 * zeroed by the commit operation.
616 */
617void ubifs_release_budget(struct ubifs_info *c, struct ubifs_budget_req *req)
618{
Artem Bityutskiy547000d2008-07-24 14:42:05 +0300619 ubifs_assert(req->new_page <= 1);
620 ubifs_assert(req->dirtied_page <= 1);
621 ubifs_assert(req->new_dent <= 1);
622 ubifs_assert(req->mod_dent <= 1);
623 ubifs_assert(req->new_ino <= 1);
624 ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300625 ubifs_assert(req->dirtied_ino <= 4);
626 ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300627 ubifs_assert(!(req->new_ino_d & 7));
628 ubifs_assert(!(req->dirtied_ino_d & 7));
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300629 if (!req->recalculate) {
630 ubifs_assert(req->idx_growth >= 0);
631 ubifs_assert(req->data_growth >= 0);
632 ubifs_assert(req->dd_growth >= 0);
633 }
634
635 if (req->recalculate) {
636 req->data_growth = calc_data_growth(c, req);
637 req->dd_growth = calc_dd_growth(c, req);
638 req->idx_growth = calc_idx_growth(c, req);
639 }
640
641 if (!req->data_growth && !req->dd_growth)
642 return;
643
644 c->nospace = c->nospace_rp = 0;
645 smp_wmb();
646
647 spin_lock(&c->space_lock);
648 c->budg_idx_growth -= req->idx_growth;
649 c->budg_uncommitted_idx += req->idx_growth;
650 c->budg_data_growth -= req->data_growth;
651 c->budg_dd_growth -= req->dd_growth;
652 c->min_idx_lebs = ubifs_calc_min_idx_lebs(c);
653
654 ubifs_assert(c->budg_idx_growth >= 0);
655 ubifs_assert(c->budg_data_growth >= 0);
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300656 ubifs_assert(c->budg_dd_growth >= 0);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300657 ubifs_assert(c->min_idx_lebs < c->main_lebs);
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300658 ubifs_assert(!(c->budg_idx_growth & 7));
659 ubifs_assert(!(c->budg_data_growth & 7));
660 ubifs_assert(!(c->budg_dd_growth & 7));
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300661 spin_unlock(&c->space_lock);
662}
663
664/**
665 * ubifs_convert_page_budget - convert budget of a new page.
666 * @c: UBIFS file-system description object
667 *
668 * This function converts budget which was allocated for a new page of data to
669 * the budget of changing an existing page of data. The latter is smaller then
670 * the former, so this function only does simple re-calculation and does not
671 * involve any write-back.
672 */
673void ubifs_convert_page_budget(struct ubifs_info *c)
674{
675 spin_lock(&c->space_lock);
676 /* Release the index growth reservation */
677 c->budg_idx_growth -= c->max_idx_node_sz << UBIFS_BLOCKS_PER_PAGE_SHIFT;
678 /* Release the data growth reservation */
679 c->budg_data_growth -= c->page_budget;
680 /* Increase the dirty data growth reservation instead */
681 c->budg_dd_growth += c->page_budget;
682 /* And re-calculate the indexing space reservation */
683 c->min_idx_lebs = ubifs_calc_min_idx_lebs(c);
684 spin_unlock(&c->space_lock);
685}
686
687/**
688 * ubifs_release_dirty_inode_budget - release dirty inode budget.
689 * @c: UBIFS file-system description object
690 * @ui: UBIFS inode to release the budget for
691 *
692 * This function releases budget corresponding to a dirty inode. It is usually
693 * called when after the inode has been written to the media and marked as
694 * clean.
695 */
696void ubifs_release_dirty_inode_budget(struct ubifs_info *c,
697 struct ubifs_inode *ui)
698{
Artem Bityutskiy182854b2008-07-18 18:54:29 +0300699 struct ubifs_budget_req req;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300700
Artem Bityutskiy182854b2008-07-18 18:54:29 +0300701 memset(&req, 0, sizeof(struct ubifs_budget_req));
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300702 req.dd_growth = c->inode_budget + ALIGN(ui->data_len, 8);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300703 ubifs_release_budget(c, &req);
704}
705
706/**
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300707 * ubifs_reported_space - calculate reported free space.
708 * @c: the UBIFS file-system description object
709 * @free: amount of free space
710 *
711 * This function calculates amount of free space which will be reported to
712 * user-space. User-space application tend to expect that if the file-system
713 * (e.g., via the 'statfs()' call) reports that it has N bytes available, they
714 * are able to write a file of size N. UBIFS attaches node headers to each data
715 * node and it has to write indexind nodes as well. This introduces additional
716 * overhead, and UBIFS it has to report sligtly less free space to meet the
717 * above expectetion.
718 *
719 * This function assumes free space is made up of uncompressed data nodes and
720 * full index nodes (one per data node, tripled because we always allow enough
721 * space to write the index thrice).
722 *
723 * Note, the calculation is pessimistic, which means that most of the time
724 * UBIFS reports less space than it actually has.
725 */
726long long ubifs_reported_space(const struct ubifs_info *c, uint64_t free)
727{
Artem Bityutskiyf171d4d2008-09-03 16:17:14 +0300728 int divisor, factor, f;
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300729
730 /*
731 * Reported space size is @free * X, where X is UBIFS block size
732 * divided by UBIFS block size + all overhead one data block
733 * introduces. The overhead is the node header + indexing overhead.
734 *
Artem Bityutskiyf171d4d2008-09-03 16:17:14 +0300735 * Indexing overhead calculations are based on the following formula:
736 * I = N/(f - 1) + 1, where I - number of indexing nodes, N - number
737 * of data nodes, f - fanout. Because effective UBIFS fanout is twice
738 * as less than maximum fanout, we assume that each data node
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300739 * introduces 3 * @c->max_idx_node_sz / (@c->fanout/2 - 1) bytes.
740 * Note, the multiplier 3 is because UBIFS reseves thrice as more space
741 * for the index.
742 */
Artem Bityutskiyf171d4d2008-09-03 16:17:14 +0300743 f = c->fanout > 3 ? c->fanout >> 1 : 2;
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300744 factor = UBIFS_BLOCK_SIZE;
745 divisor = UBIFS_MAX_DATA_NODE_SZ;
Artem Bityutskiyf171d4d2008-09-03 16:17:14 +0300746 divisor += (c->max_idx_node_sz * 3) / (f - 1);
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300747 free *= factor;
748 do_div(free, divisor);
749 return free;
750}
751
752/**
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300753 * ubifs_get_free_space - return amount of free space.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300754 * @c: UBIFS file-system description object
755 *
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300756 * This function calculates amount of free space to report to user-space.
757 *
758 * Because UBIFS may introduce substantial overhead (the index, node headers,
759 * alighment, wastage at the end of eraseblocks, etc), it cannot report real
760 * amount of free flash space it has (well, because not all dirty space is
761 * reclamable, UBIFS does not actually know the real amount). If UBIFS did so,
762 * it would bread user expectetion about what free space is. Users seem to
763 * accustomed to assume that if the file-system reports N bytes of free space,
764 * they would be able to fit a file of N bytes to the FS. This almost works for
765 * traditional file-systems, because they have way less overhead than UBIFS.
766 * So, to keep users happy, UBIFS tries to take the overhead into account.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300767 */
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300768long long ubifs_get_free_space(struct ubifs_info *c)
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300769{
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300770 int min_idx_lebs, rsvd_idx_lebs, lebs;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300771 long long available, outstanding, free;
772
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300773 spin_lock(&c->space_lock);
774 min_idx_lebs = ubifs_calc_min_idx_lebs(c);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300775 outstanding = c->budg_data_growth + c->budg_dd_growth;
Artem Bityutskiy9e5de352008-08-25 17:29:43 +0300776
777 /*
778 * Force the amount available to the total size reported if the used
779 * space is zero.
780 */
781 if (c->lst.total_used <= UBIFS_INO_NODE_SZ && !outstanding) {
782 spin_unlock(&c->space_lock);
783 return (long long)c->block_cnt << UBIFS_BLOCK_SHIFT;
784 }
785
786 available = ubifs_calc_available(c, min_idx_lebs);
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300787
788 /*
789 * When reporting free space to user-space, UBIFS guarantees that it is
790 * possible to write a file of free space size. This means that for
791 * empty LEBs we may use more precise calculations than
792 * 'ubifs_calc_available()' is using. Namely, we know that in empty
793 * LEBs we would waste only @c->leb_overhead bytes, not @c->dark_wm.
794 * Thus, amend the available space.
795 *
796 * Note, the calculations below are similar to what we have in
797 * 'do_budget_space()', so refer there for comments.
798 */
799 if (min_idx_lebs > c->lst.idx_lebs)
800 rsvd_idx_lebs = min_idx_lebs - c->lst.idx_lebs;
801 else
802 rsvd_idx_lebs = 0;
803 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
804 c->lst.taken_empty_lebs;
805 lebs -= rsvd_idx_lebs;
806 available += lebs * (c->dark_wm - c->leb_overhead);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300807 spin_unlock(&c->space_lock);
808
809 if (available > outstanding)
810 free = ubifs_reported_space(c, available - outstanding);
811 else
812 free = 0;
813 return free;
814}