blob: 0bcb8031ca18f319c967c7688f61173fd1471641 [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>
Artem Bityutskiy4d61db42008-12-18 14:06:51 +020035#include <linux/math64.h>
Artem Bityutskiy1e517642008-07-14 19:08:37 +030036
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{
Artem Bityutskiy4d61db42008-12-18 14:06:51 +0200261 int idx_lebs, eff_leb_size = c->leb_size - c->max_idx_node_sz;
262 long long idx_size;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300263
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 */
Artem Bityutskiy4d61db42008-12-18 14:06:51 +0200274 idx_size += eff_leb_size - 1;
275 idx_lebs = div_u64(idx_size, eff_leb_size);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300276 /*
277 * The index head is not available for the in-the-gaps method, so add an
278 * extra LEB to compensate.
279 */
Artem Bityutskiy4d61db42008-12-18 14:06:51 +0200280 idx_lebs += 1;
281 if (idx_lebs < MIN_INDEX_LEBS)
282 idx_lebs = MIN_INDEX_LEBS;
283 return idx_lebs;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300284}
285
286/**
287 * ubifs_calc_available - calculate available FS space.
288 * @c: UBIFS file-system description object
289 * @min_idx_lebs: minimum number of LEBs reserved for the index
290 *
291 * This function calculates and returns amount of FS space available for use.
292 */
293long long ubifs_calc_available(const struct ubifs_info *c, int min_idx_lebs)
294{
295 int subtract_lebs;
296 long long available;
297
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300298 available = c->main_bytes - c->lst.total_used;
299
300 /*
301 * Now 'available' contains theoretically available flash space
302 * assuming there is no index, so we have to subtract the space which
303 * is reserved for the index.
304 */
305 subtract_lebs = min_idx_lebs;
306
307 /* Take into account that GC reserves one LEB for its own needs */
308 subtract_lebs += 1;
309
310 /*
311 * The GC journal head LEB is not really accessible. And since
312 * different write types go to different heads, we may count only on
313 * one head's space.
314 */
315 subtract_lebs += c->jhead_cnt - 1;
316
317 /* We also reserve one LEB for deletions, which bypass budgeting */
318 subtract_lebs += 1;
319
320 available -= (long long)subtract_lebs * c->leb_size;
321
322 /* Subtract the dead space which is not available for use */
323 available -= c->lst.total_dead;
324
325 /*
326 * Subtract dark space, which might or might not be usable - it depends
327 * on the data which we have on the media and which will be written. If
328 * this is a lot of uncompressed or not-compressible data, the dark
329 * space cannot be used.
330 */
331 available -= c->lst.total_dark;
332
333 /*
334 * However, there is more dark space. The index may be bigger than
335 * @min_idx_lebs. Those extra LEBs are assumed to be available, but
336 * their dark space is not included in total_dark, so it is subtracted
337 * here.
338 */
339 if (c->lst.idx_lebs > min_idx_lebs) {
340 subtract_lebs = c->lst.idx_lebs - min_idx_lebs;
341 available -= subtract_lebs * c->dark_wm;
342 }
343
344 /* The calculations are rough and may end up with a negative number */
345 return available > 0 ? available : 0;
346}
347
348/**
349 * can_use_rp - check whether the user is allowed to use reserved pool.
350 * @c: UBIFS file-system description object
351 *
352 * UBIFS has so-called "reserved pool" which is flash space reserved
353 * for the superuser and for uses whose UID/GID is recorded in UBIFS superblock.
354 * This function checks whether current user is allowed to use reserved pool.
355 * Returns %1 current user is allowed to use reserved pool and %0 otherwise.
356 */
357static int can_use_rp(struct ubifs_info *c)
358{
359 if (current->fsuid == c->rp_uid || capable(CAP_SYS_RESOURCE) ||
360 (c->rp_gid != 0 && in_group_p(c->rp_gid)))
361 return 1;
362 return 0;
363}
364
365/**
366 * do_budget_space - reserve flash space for index and data growth.
367 * @c: UBIFS file-system description object
368 *
369 * This function makes sure UBIFS has enough free eraseblocks for index growth
370 * and data.
371 *
Adrian Hunter3a132522008-07-30 12:18:02 +0300372 * When budgeting index space, UBIFS reserves thrice as many LEBs as the index
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300373 * would take if it was consolidated and written to the flash. This guarantees
374 * that the "in-the-gaps" commit method always succeeds and UBIFS will always
375 * be able to commit dirty index. So this function basically adds amount of
Artem Bityutskiyb364b412008-07-25 14:38:51 +0300376 * budgeted index space to the size of the current index, multiplies this by 3,
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300377 * and makes sure this does not exceed the amount of free eraseblocks.
378 *
379 * Notes about @c->min_idx_lebs and @c->lst.idx_lebs variables:
380 * o @c->lst.idx_lebs is the number of LEBs the index currently uses. It might
381 * be large, because UBIFS does not do any index consolidation as long as
382 * there is free space. IOW, the index may take a lot of LEBs, but the LEBs
383 * will contain a lot of dirt.
384 * o @c->min_idx_lebs is the the index presumably takes. IOW, the index may be
385 * consolidated to take up to @c->min_idx_lebs LEBs.
386 *
387 * This function returns zero in case of success, and %-ENOSPC in case of
388 * failure.
389 */
390static int do_budget_space(struct ubifs_info *c)
391{
392 long long outstanding, available;
393 int lebs, rsvd_idx_lebs, min_idx_lebs;
394
395 /* First budget index space */
396 min_idx_lebs = ubifs_calc_min_idx_lebs(c);
397
398 /* Now 'min_idx_lebs' contains number of LEBs to reserve */
399 if (min_idx_lebs > c->lst.idx_lebs)
400 rsvd_idx_lebs = min_idx_lebs - c->lst.idx_lebs;
401 else
402 rsvd_idx_lebs = 0;
403
404 /*
405 * The number of LEBs that are available to be used by the index is:
406 *
407 * @c->lst.empty_lebs + @c->freeable_cnt + @c->idx_gc_cnt -
408 * @c->lst.taken_empty_lebs
409 *
Artem Bityutskiy948cfb22008-08-20 11:56:33 +0300410 * @c->lst.empty_lebs are available because they are empty.
411 * @c->freeable_cnt are available because they contain only free and
412 * dirty space, @c->idx_gc_cnt are available because they are index
413 * LEBs that have been garbage collected and are awaiting the commit
414 * before they can be used. And the in-the-gaps method will grab these
415 * if it needs them. @c->lst.taken_empty_lebs are empty LEBs that have
416 * already been allocated for some purpose.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300417 *
Artem Bityutskiy948cfb22008-08-20 11:56:33 +0300418 * Note, @c->idx_gc_cnt is included to both @c->lst.empty_lebs (because
419 * these LEBs are empty) and to @c->lst.taken_empty_lebs (because they
420 * are taken until after the commit).
421 *
422 * Note, @c->lst.taken_empty_lebs may temporarily be higher by one
423 * because of the way we serialize LEB allocations and budgeting. See a
424 * comment in 'ubifs_find_free_space()'.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300425 */
426 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
427 c->lst.taken_empty_lebs;
428 if (unlikely(rsvd_idx_lebs > lebs)) {
429 dbg_budg("out of indexing space: min_idx_lebs %d (old %d), "
430 "rsvd_idx_lebs %d", min_idx_lebs, c->min_idx_lebs,
431 rsvd_idx_lebs);
432 return -ENOSPC;
433 }
434
435 available = ubifs_calc_available(c, min_idx_lebs);
436 outstanding = c->budg_data_growth + c->budg_dd_growth;
437
438 if (unlikely(available < outstanding)) {
439 dbg_budg("out of data space: available %lld, outstanding %lld",
440 available, outstanding);
441 return -ENOSPC;
442 }
443
444 if (available - outstanding <= c->rp_size && !can_use_rp(c))
445 return -ENOSPC;
446
447 c->min_idx_lebs = min_idx_lebs;
448 return 0;
449}
450
451/**
452 * calc_idx_growth - calculate approximate index growth from budgeting request.
453 * @c: UBIFS file-system description object
454 * @req: budgeting request
455 *
456 * For now we assume each new node adds one znode. But this is rather poor
457 * approximation, though.
458 */
459static int calc_idx_growth(const struct ubifs_info *c,
460 const struct ubifs_budget_req *req)
461{
462 int znodes;
463
464 znodes = req->new_ino + (req->new_page << UBIFS_BLOCKS_PER_PAGE_SHIFT) +
465 req->new_dent;
466 return znodes * c->max_idx_node_sz;
467}
468
469/**
470 * calc_data_growth - calculate approximate amount of new data from budgeting
471 * request.
472 * @c: UBIFS file-system description object
473 * @req: budgeting request
474 */
475static int calc_data_growth(const struct ubifs_info *c,
476 const struct ubifs_budget_req *req)
477{
478 int data_growth;
479
480 data_growth = req->new_ino ? c->inode_budget : 0;
481 if (req->new_page)
482 data_growth += c->page_budget;
483 if (req->new_dent)
484 data_growth += c->dent_budget;
485 data_growth += req->new_ino_d;
486 return data_growth;
487}
488
489/**
490 * calc_dd_growth - calculate approximate amount of data which makes other data
491 * dirty from budgeting request.
492 * @c: UBIFS file-system description object
493 * @req: budgeting request
494 */
495static int calc_dd_growth(const struct ubifs_info *c,
496 const struct ubifs_budget_req *req)
497{
498 int dd_growth;
499
500 dd_growth = req->dirtied_page ? c->page_budget : 0;
501
502 if (req->dirtied_ino)
503 dd_growth += c->inode_budget << (req->dirtied_ino - 1);
504 if (req->mod_dent)
505 dd_growth += c->dent_budget;
506 dd_growth += req->dirtied_ino_d;
507 return dd_growth;
508}
509
510/**
511 * ubifs_budget_space - ensure there is enough space to complete an operation.
512 * @c: UBIFS file-system description object
513 * @req: budget request
514 *
515 * This function allocates budget for an operation. It uses pessimistic
516 * approximation of how much flash space the operation needs. The goal of this
517 * function is to make sure UBIFS always has flash space to flush all dirty
518 * pages, dirty inodes, and dirty znodes (liability). This function may force
519 * commit, garbage-collection or write-back. Returns zero in case of success,
520 * %-ENOSPC if there is no free space and other negative error codes in case of
521 * failures.
522 */
523int ubifs_budget_space(struct ubifs_info *c, struct ubifs_budget_req *req)
524{
525 int uninitialized_var(cmt_retries), uninitialized_var(wb_retries);
526 int err, idx_growth, data_growth, dd_growth;
527 struct retries_info ri;
528
Artem Bityutskiy547000d2008-07-24 14:42:05 +0300529 ubifs_assert(req->new_page <= 1);
530 ubifs_assert(req->dirtied_page <= 1);
531 ubifs_assert(req->new_dent <= 1);
532 ubifs_assert(req->mod_dent <= 1);
533 ubifs_assert(req->new_ino <= 1);
534 ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300535 ubifs_assert(req->dirtied_ino <= 4);
536 ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300537 ubifs_assert(!(req->new_ino_d & 7));
538 ubifs_assert(!(req->dirtied_ino_d & 7));
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300539
540 data_growth = calc_data_growth(c, req);
541 dd_growth = calc_dd_growth(c, req);
542 if (!data_growth && !dd_growth)
543 return 0;
544 idx_growth = calc_idx_growth(c, req);
545 memset(&ri, 0, sizeof(struct retries_info));
546
547again:
548 spin_lock(&c->space_lock);
549 ubifs_assert(c->budg_idx_growth >= 0);
550 ubifs_assert(c->budg_data_growth >= 0);
551 ubifs_assert(c->budg_dd_growth >= 0);
552
553 if (unlikely(c->nospace) && (c->nospace_rp || !can_use_rp(c))) {
554 dbg_budg("no space");
555 spin_unlock(&c->space_lock);
556 return -ENOSPC;
557 }
558
559 c->budg_idx_growth += idx_growth;
560 c->budg_data_growth += data_growth;
561 c->budg_dd_growth += dd_growth;
562
563 err = do_budget_space(c);
564 if (likely(!err)) {
565 req->idx_growth = idx_growth;
566 req->data_growth = data_growth;
567 req->dd_growth = dd_growth;
568 spin_unlock(&c->space_lock);
569 return 0;
570 }
571
572 /* Restore the old values */
573 c->budg_idx_growth -= idx_growth;
574 c->budg_data_growth -= data_growth;
575 c->budg_dd_growth -= dd_growth;
576 spin_unlock(&c->space_lock);
577
578 if (req->fast) {
579 dbg_budg("no space for fast budgeting");
580 return err;
581 }
582
583 err = make_free_space(c, &ri);
584 if (err == -EAGAIN) {
585 dbg_budg("try again");
586 cond_resched();
587 goto again;
588 } else if (err == -ENOSPC) {
589 dbg_budg("FS is full, -ENOSPC");
590 c->nospace = 1;
591 if (can_use_rp(c) || c->rp_size == 0)
592 c->nospace_rp = 1;
593 smp_wmb();
594 } else
595 ubifs_err("cannot budget space, error %d", err);
596 return err;
597}
598
599/**
600 * ubifs_release_budget - release budgeted free space.
601 * @c: UBIFS file-system description object
602 * @req: budget request
603 *
604 * This function releases the space budgeted by 'ubifs_budget_space()'. Note,
605 * since the index changes (which were budgeted for in @req->idx_growth) will
606 * only be written to the media on commit, this function moves the index budget
607 * from @c->budg_idx_growth to @c->budg_uncommitted_idx. The latter will be
608 * zeroed by the commit operation.
609 */
610void ubifs_release_budget(struct ubifs_info *c, struct ubifs_budget_req *req)
611{
Artem Bityutskiy547000d2008-07-24 14:42:05 +0300612 ubifs_assert(req->new_page <= 1);
613 ubifs_assert(req->dirtied_page <= 1);
614 ubifs_assert(req->new_dent <= 1);
615 ubifs_assert(req->mod_dent <= 1);
616 ubifs_assert(req->new_ino <= 1);
617 ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300618 ubifs_assert(req->dirtied_ino <= 4);
619 ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4);
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300620 ubifs_assert(!(req->new_ino_d & 7));
621 ubifs_assert(!(req->dirtied_ino_d & 7));
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300622 if (!req->recalculate) {
623 ubifs_assert(req->idx_growth >= 0);
624 ubifs_assert(req->data_growth >= 0);
625 ubifs_assert(req->dd_growth >= 0);
626 }
627
628 if (req->recalculate) {
629 req->data_growth = calc_data_growth(c, req);
630 req->dd_growth = calc_dd_growth(c, req);
631 req->idx_growth = calc_idx_growth(c, req);
632 }
633
634 if (!req->data_growth && !req->dd_growth)
635 return;
636
637 c->nospace = c->nospace_rp = 0;
638 smp_wmb();
639
640 spin_lock(&c->space_lock);
641 c->budg_idx_growth -= req->idx_growth;
642 c->budg_uncommitted_idx += req->idx_growth;
643 c->budg_data_growth -= req->data_growth;
644 c->budg_dd_growth -= req->dd_growth;
645 c->min_idx_lebs = ubifs_calc_min_idx_lebs(c);
646
647 ubifs_assert(c->budg_idx_growth >= 0);
648 ubifs_assert(c->budg_data_growth >= 0);
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300649 ubifs_assert(c->budg_dd_growth >= 0);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300650 ubifs_assert(c->min_idx_lebs < c->main_lebs);
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300651 ubifs_assert(!(c->budg_idx_growth & 7));
652 ubifs_assert(!(c->budg_data_growth & 7));
653 ubifs_assert(!(c->budg_dd_growth & 7));
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300654 spin_unlock(&c->space_lock);
655}
656
657/**
658 * ubifs_convert_page_budget - convert budget of a new page.
659 * @c: UBIFS file-system description object
660 *
661 * This function converts budget which was allocated for a new page of data to
662 * the budget of changing an existing page of data. The latter is smaller then
663 * the former, so this function only does simple re-calculation and does not
664 * involve any write-back.
665 */
666void ubifs_convert_page_budget(struct ubifs_info *c)
667{
668 spin_lock(&c->space_lock);
669 /* Release the index growth reservation */
670 c->budg_idx_growth -= c->max_idx_node_sz << UBIFS_BLOCKS_PER_PAGE_SHIFT;
671 /* Release the data growth reservation */
672 c->budg_data_growth -= c->page_budget;
673 /* Increase the dirty data growth reservation instead */
674 c->budg_dd_growth += c->page_budget;
675 /* And re-calculate the indexing space reservation */
676 c->min_idx_lebs = ubifs_calc_min_idx_lebs(c);
677 spin_unlock(&c->space_lock);
678}
679
680/**
681 * ubifs_release_dirty_inode_budget - release dirty inode budget.
682 * @c: UBIFS file-system description object
683 * @ui: UBIFS inode to release the budget for
684 *
685 * This function releases budget corresponding to a dirty inode. It is usually
686 * called when after the inode has been written to the media and marked as
687 * clean.
688 */
689void ubifs_release_dirty_inode_budget(struct ubifs_info *c,
690 struct ubifs_inode *ui)
691{
Artem Bityutskiy182854b2008-07-18 18:54:29 +0300692 struct ubifs_budget_req req;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300693
Artem Bityutskiy182854b2008-07-18 18:54:29 +0300694 memset(&req, 0, sizeof(struct ubifs_budget_req));
Artem Bityutskiydab4b4d2008-07-24 14:52:45 +0300695 req.dd_growth = c->inode_budget + ALIGN(ui->data_len, 8);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300696 ubifs_release_budget(c, &req);
697}
698
699/**
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300700 * ubifs_reported_space - calculate reported free space.
701 * @c: the UBIFS file-system description object
702 * @free: amount of free space
703 *
704 * This function calculates amount of free space which will be reported to
705 * user-space. User-space application tend to expect that if the file-system
706 * (e.g., via the 'statfs()' call) reports that it has N bytes available, they
707 * are able to write a file of size N. UBIFS attaches node headers to each data
708 * node and it has to write indexind nodes as well. This introduces additional
Artem Bityutskiy21a60252008-12-12 11:13:17 -0500709 * overhead, and UBIFS has to report sligtly less free space to meet the above
710 * expectetions.
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300711 *
712 * This function assumes free space is made up of uncompressed data nodes and
713 * full index nodes (one per data node, tripled because we always allow enough
714 * space to write the index thrice).
715 *
716 * Note, the calculation is pessimistic, which means that most of the time
717 * UBIFS reports less space than it actually has.
718 */
Artem Bityutskiy4d61db42008-12-18 14:06:51 +0200719long long ubifs_reported_space(const struct ubifs_info *c, long long free)
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300720{
Artem Bityutskiyf171d4d2008-09-03 16:17:14 +0300721 int divisor, factor, f;
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300722
723 /*
724 * Reported space size is @free * X, where X is UBIFS block size
725 * divided by UBIFS block size + all overhead one data block
726 * introduces. The overhead is the node header + indexing overhead.
727 *
Artem Bityutskiyf171d4d2008-09-03 16:17:14 +0300728 * Indexing overhead calculations are based on the following formula:
729 * I = N/(f - 1) + 1, where I - number of indexing nodes, N - number
730 * of data nodes, f - fanout. Because effective UBIFS fanout is twice
731 * as less than maximum fanout, we assume that each data node
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300732 * introduces 3 * @c->max_idx_node_sz / (@c->fanout/2 - 1) bytes.
733 * Note, the multiplier 3 is because UBIFS reseves thrice as more space
734 * for the index.
735 */
Artem Bityutskiyf171d4d2008-09-03 16:17:14 +0300736 f = c->fanout > 3 ? c->fanout >> 1 : 2;
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300737 factor = UBIFS_BLOCK_SIZE;
738 divisor = UBIFS_MAX_DATA_NODE_SZ;
Artem Bityutskiyf171d4d2008-09-03 16:17:14 +0300739 divisor += (c->max_idx_node_sz * 3) / (f - 1);
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300740 free *= factor;
Artem Bityutskiy4d61db42008-12-18 14:06:51 +0200741 return div_u64(free, divisor);
Artem Bityutskiy4b5f2762008-08-25 16:15:56 +0300742}
743
744/**
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300745 * ubifs_get_free_space - return amount of free space.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300746 * @c: UBIFS file-system description object
747 *
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300748 * This function calculates amount of free space to report to user-space.
749 *
750 * Because UBIFS may introduce substantial overhead (the index, node headers,
751 * alighment, wastage at the end of eraseblocks, etc), it cannot report real
752 * amount of free flash space it has (well, because not all dirty space is
753 * reclamable, UBIFS does not actually know the real amount). If UBIFS did so,
754 * it would bread user expectetion about what free space is. Users seem to
755 * accustomed to assume that if the file-system reports N bytes of free space,
756 * they would be able to fit a file of N bytes to the FS. This almost works for
757 * traditional file-systems, because they have way less overhead than UBIFS.
758 * So, to keep users happy, UBIFS tries to take the overhead into account.
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300759 */
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300760long long ubifs_get_free_space(struct ubifs_info *c)
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300761{
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300762 int min_idx_lebs, rsvd_idx_lebs, lebs;
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300763 long long available, outstanding, free;
764
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300765 spin_lock(&c->space_lock);
766 min_idx_lebs = ubifs_calc_min_idx_lebs(c);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300767 outstanding = c->budg_data_growth + c->budg_dd_growth;
Artem Bityutskiy9e5de352008-08-25 17:29:43 +0300768
769 /*
770 * Force the amount available to the total size reported if the used
771 * space is zero.
772 */
773 if (c->lst.total_used <= UBIFS_INO_NODE_SZ && !outstanding) {
774 spin_unlock(&c->space_lock);
775 return (long long)c->block_cnt << UBIFS_BLOCK_SHIFT;
776 }
777
778 available = ubifs_calc_available(c, min_idx_lebs);
Artem Bityutskiy7dad1812008-08-25 18:58:19 +0300779
780 /*
781 * When reporting free space to user-space, UBIFS guarantees that it is
782 * possible to write a file of free space size. This means that for
783 * empty LEBs we may use more precise calculations than
784 * 'ubifs_calc_available()' is using. Namely, we know that in empty
785 * LEBs we would waste only @c->leb_overhead bytes, not @c->dark_wm.
786 * Thus, amend the available space.
787 *
788 * Note, the calculations below are similar to what we have in
789 * 'do_budget_space()', so refer there for comments.
790 */
791 if (min_idx_lebs > c->lst.idx_lebs)
792 rsvd_idx_lebs = min_idx_lebs - c->lst.idx_lebs;
793 else
794 rsvd_idx_lebs = 0;
795 lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt -
796 c->lst.taken_empty_lebs;
797 lebs -= rsvd_idx_lebs;
798 available += lebs * (c->dark_wm - c->leb_overhead);
Artem Bityutskiy1e517642008-07-14 19:08:37 +0300799 spin_unlock(&c->space_lock);
800
801 if (available > outstanding)
802 free = ubifs_reported_space(c, available - outstanding);
803 else
804 free = 0;
805 return free;
806}