Commit | Line | Data |
---|---|---|
1e51764a AB |
1 | /* |
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 | */ | |
68 | struct 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 | */ | |
92 | static 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 | */ | |
130 | static 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 | */ | |
168 | static 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 | */ | |
259 | int 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 | ||
3a13252c | 266 | /* And make sure we have thrice the index size of space reserved */ |
b364b41a | 267 | idx_size = idx_size + (idx_size << 1); |
1e51764a AB |
268 | |
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; | |
af14a1ad AB |
283 | if (ret < MIN_INDEX_LEBS) |
284 | ret = MIN_INDEX_LEBS; | |
1e51764a AB |
285 | return ret; |
286 | } | |
287 | ||
288 | /** | |
289 | * ubifs_calc_available - calculate available FS space. | |
290 | * @c: UBIFS file-system description object | |
291 | * @min_idx_lebs: minimum number of LEBs reserved for the index | |
292 | * | |
293 | * This function calculates and returns amount of FS space available for use. | |
294 | */ | |
295 | long long ubifs_calc_available(const struct ubifs_info *c, int min_idx_lebs) | |
296 | { | |
297 | int subtract_lebs; | |
298 | long long available; | |
299 | ||
1e51764a AB |
300 | available = c->main_bytes - c->lst.total_used; |
301 | ||
302 | /* | |
303 | * Now 'available' contains theoretically available flash space | |
304 | * assuming there is no index, so we have to subtract the space which | |
305 | * is reserved for the index. | |
306 | */ | |
307 | subtract_lebs = min_idx_lebs; | |
308 | ||
309 | /* Take into account that GC reserves one LEB for its own needs */ | |
310 | subtract_lebs += 1; | |
311 | ||
312 | /* | |
313 | * The GC journal head LEB is not really accessible. And since | |
314 | * different write types go to different heads, we may count only on | |
315 | * one head's space. | |
316 | */ | |
317 | subtract_lebs += c->jhead_cnt - 1; | |
318 | ||
319 | /* We also reserve one LEB for deletions, which bypass budgeting */ | |
320 | subtract_lebs += 1; | |
321 | ||
322 | available -= (long long)subtract_lebs * c->leb_size; | |
323 | ||
324 | /* Subtract the dead space which is not available for use */ | |
325 | available -= c->lst.total_dead; | |
326 | ||
327 | /* | |
328 | * Subtract dark space, which might or might not be usable - it depends | |
329 | * on the data which we have on the media and which will be written. If | |
330 | * this is a lot of uncompressed or not-compressible data, the dark | |
331 | * space cannot be used. | |
332 | */ | |
333 | available -= c->lst.total_dark; | |
334 | ||
335 | /* | |
336 | * However, there is more dark space. The index may be bigger than | |
337 | * @min_idx_lebs. Those extra LEBs are assumed to be available, but | |
338 | * their dark space is not included in total_dark, so it is subtracted | |
339 | * here. | |
340 | */ | |
341 | if (c->lst.idx_lebs > min_idx_lebs) { | |
342 | subtract_lebs = c->lst.idx_lebs - min_idx_lebs; | |
343 | available -= subtract_lebs * c->dark_wm; | |
344 | } | |
345 | ||
346 | /* The calculations are rough and may end up with a negative number */ | |
347 | return available > 0 ? available : 0; | |
348 | } | |
349 | ||
350 | /** | |
351 | * can_use_rp - check whether the user is allowed to use reserved pool. | |
352 | * @c: UBIFS file-system description object | |
353 | * | |
354 | * UBIFS has so-called "reserved pool" which is flash space reserved | |
355 | * for the superuser and for uses whose UID/GID is recorded in UBIFS superblock. | |
356 | * This function checks whether current user is allowed to use reserved pool. | |
357 | * Returns %1 current user is allowed to use reserved pool and %0 otherwise. | |
358 | */ | |
359 | static int can_use_rp(struct ubifs_info *c) | |
360 | { | |
361 | if (current->fsuid == c->rp_uid || capable(CAP_SYS_RESOURCE) || | |
362 | (c->rp_gid != 0 && in_group_p(c->rp_gid))) | |
363 | return 1; | |
364 | return 0; | |
365 | } | |
366 | ||
367 | /** | |
368 | * do_budget_space - reserve flash space for index and data growth. | |
369 | * @c: UBIFS file-system description object | |
370 | * | |
371 | * This function makes sure UBIFS has enough free eraseblocks for index growth | |
372 | * and data. | |
373 | * | |
3a13252c | 374 | * When budgeting index space, UBIFS reserves thrice as many LEBs as the index |
1e51764a AB |
375 | * would take if it was consolidated and written to the flash. This guarantees |
376 | * that the "in-the-gaps" commit method always succeeds and UBIFS will always | |
377 | * be able to commit dirty index. So this function basically adds amount of | |
b364b41a | 378 | * budgeted index space to the size of the current index, multiplies this by 3, |
1e51764a AB |
379 | * and makes sure this does not exceed the amount of free eraseblocks. |
380 | * | |
381 | * Notes about @c->min_idx_lebs and @c->lst.idx_lebs variables: | |
382 | * o @c->lst.idx_lebs is the number of LEBs the index currently uses. It might | |
383 | * be large, because UBIFS does not do any index consolidation as long as | |
384 | * there is free space. IOW, the index may take a lot of LEBs, but the LEBs | |
385 | * will contain a lot of dirt. | |
386 | * o @c->min_idx_lebs is the the index presumably takes. IOW, the index may be | |
387 | * consolidated to take up to @c->min_idx_lebs LEBs. | |
388 | * | |
389 | * This function returns zero in case of success, and %-ENOSPC in case of | |
390 | * failure. | |
391 | */ | |
392 | static int do_budget_space(struct ubifs_info *c) | |
393 | { | |
394 | long long outstanding, available; | |
395 | int lebs, rsvd_idx_lebs, min_idx_lebs; | |
396 | ||
397 | /* First budget index space */ | |
398 | min_idx_lebs = ubifs_calc_min_idx_lebs(c); | |
399 | ||
400 | /* Now 'min_idx_lebs' contains number of LEBs to reserve */ | |
401 | if (min_idx_lebs > c->lst.idx_lebs) | |
402 | rsvd_idx_lebs = min_idx_lebs - c->lst.idx_lebs; | |
403 | else | |
404 | rsvd_idx_lebs = 0; | |
405 | ||
406 | /* | |
407 | * The number of LEBs that are available to be used by the index is: | |
408 | * | |
409 | * @c->lst.empty_lebs + @c->freeable_cnt + @c->idx_gc_cnt - | |
410 | * @c->lst.taken_empty_lebs | |
411 | * | |
948cfb21 AB |
412 | * @c->lst.empty_lebs are available because they are empty. |
413 | * @c->freeable_cnt are available because they contain only free and | |
414 | * dirty space, @c->idx_gc_cnt are available because they are index | |
415 | * LEBs that have been garbage collected and are awaiting the commit | |
416 | * before they can be used. And the in-the-gaps method will grab these | |
417 | * if it needs them. @c->lst.taken_empty_lebs are empty LEBs that have | |
418 | * already been allocated for some purpose. | |
1e51764a | 419 | * |
948cfb21 AB |
420 | * Note, @c->idx_gc_cnt is included to both @c->lst.empty_lebs (because |
421 | * these LEBs are empty) and to @c->lst.taken_empty_lebs (because they | |
422 | * are taken until after the commit). | |
423 | * | |
424 | * Note, @c->lst.taken_empty_lebs may temporarily be higher by one | |
425 | * because of the way we serialize LEB allocations and budgeting. See a | |
426 | * comment in 'ubifs_find_free_space()'. | |
1e51764a AB |
427 | */ |
428 | lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt - | |
429 | c->lst.taken_empty_lebs; | |
430 | if (unlikely(rsvd_idx_lebs > lebs)) { | |
431 | dbg_budg("out of indexing space: min_idx_lebs %d (old %d), " | |
432 | "rsvd_idx_lebs %d", min_idx_lebs, c->min_idx_lebs, | |
433 | rsvd_idx_lebs); | |
434 | return -ENOSPC; | |
435 | } | |
436 | ||
437 | available = ubifs_calc_available(c, min_idx_lebs); | |
438 | outstanding = c->budg_data_growth + c->budg_dd_growth; | |
439 | ||
440 | if (unlikely(available < outstanding)) { | |
441 | dbg_budg("out of data space: available %lld, outstanding %lld", | |
442 | available, outstanding); | |
443 | return -ENOSPC; | |
444 | } | |
445 | ||
446 | if (available - outstanding <= c->rp_size && !can_use_rp(c)) | |
447 | return -ENOSPC; | |
448 | ||
449 | c->min_idx_lebs = min_idx_lebs; | |
450 | return 0; | |
451 | } | |
452 | ||
453 | /** | |
454 | * calc_idx_growth - calculate approximate index growth from budgeting request. | |
455 | * @c: UBIFS file-system description object | |
456 | * @req: budgeting request | |
457 | * | |
458 | * For now we assume each new node adds one znode. But this is rather poor | |
459 | * approximation, though. | |
460 | */ | |
461 | static int calc_idx_growth(const struct ubifs_info *c, | |
462 | const struct ubifs_budget_req *req) | |
463 | { | |
464 | int znodes; | |
465 | ||
466 | znodes = req->new_ino + (req->new_page << UBIFS_BLOCKS_PER_PAGE_SHIFT) + | |
467 | req->new_dent; | |
468 | return znodes * c->max_idx_node_sz; | |
469 | } | |
470 | ||
471 | /** | |
472 | * calc_data_growth - calculate approximate amount of new data from budgeting | |
473 | * request. | |
474 | * @c: UBIFS file-system description object | |
475 | * @req: budgeting request | |
476 | */ | |
477 | static int calc_data_growth(const struct ubifs_info *c, | |
478 | const struct ubifs_budget_req *req) | |
479 | { | |
480 | int data_growth; | |
481 | ||
482 | data_growth = req->new_ino ? c->inode_budget : 0; | |
483 | if (req->new_page) | |
484 | data_growth += c->page_budget; | |
485 | if (req->new_dent) | |
486 | data_growth += c->dent_budget; | |
487 | data_growth += req->new_ino_d; | |
488 | return data_growth; | |
489 | } | |
490 | ||
491 | /** | |
492 | * calc_dd_growth - calculate approximate amount of data which makes other data | |
493 | * dirty from budgeting request. | |
494 | * @c: UBIFS file-system description object | |
495 | * @req: budgeting request | |
496 | */ | |
497 | static int calc_dd_growth(const struct ubifs_info *c, | |
498 | const struct ubifs_budget_req *req) | |
499 | { | |
500 | int dd_growth; | |
501 | ||
502 | dd_growth = req->dirtied_page ? c->page_budget : 0; | |
503 | ||
504 | if (req->dirtied_ino) | |
505 | dd_growth += c->inode_budget << (req->dirtied_ino - 1); | |
506 | if (req->mod_dent) | |
507 | dd_growth += c->dent_budget; | |
508 | dd_growth += req->dirtied_ino_d; | |
509 | return dd_growth; | |
510 | } | |
511 | ||
512 | /** | |
513 | * ubifs_budget_space - ensure there is enough space to complete an operation. | |
514 | * @c: UBIFS file-system description object | |
515 | * @req: budget request | |
516 | * | |
517 | * This function allocates budget for an operation. It uses pessimistic | |
518 | * approximation of how much flash space the operation needs. The goal of this | |
519 | * function is to make sure UBIFS always has flash space to flush all dirty | |
520 | * pages, dirty inodes, and dirty znodes (liability). This function may force | |
521 | * commit, garbage-collection or write-back. Returns zero in case of success, | |
522 | * %-ENOSPC if there is no free space and other negative error codes in case of | |
523 | * failures. | |
524 | */ | |
525 | int ubifs_budget_space(struct ubifs_info *c, struct ubifs_budget_req *req) | |
526 | { | |
527 | int uninitialized_var(cmt_retries), uninitialized_var(wb_retries); | |
528 | int err, idx_growth, data_growth, dd_growth; | |
529 | struct retries_info ri; | |
530 | ||
547000da AB |
531 | ubifs_assert(req->new_page <= 1); |
532 | ubifs_assert(req->dirtied_page <= 1); | |
533 | ubifs_assert(req->new_dent <= 1); | |
534 | ubifs_assert(req->mod_dent <= 1); | |
535 | ubifs_assert(req->new_ino <= 1); | |
536 | ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA); | |
1e51764a AB |
537 | ubifs_assert(req->dirtied_ino <= 4); |
538 | ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4); | |
dab4b4d2 AB |
539 | ubifs_assert(!(req->new_ino_d & 7)); |
540 | ubifs_assert(!(req->dirtied_ino_d & 7)); | |
1e51764a AB |
541 | |
542 | data_growth = calc_data_growth(c, req); | |
543 | dd_growth = calc_dd_growth(c, req); | |
544 | if (!data_growth && !dd_growth) | |
545 | return 0; | |
546 | idx_growth = calc_idx_growth(c, req); | |
547 | memset(&ri, 0, sizeof(struct retries_info)); | |
548 | ||
549 | again: | |
550 | spin_lock(&c->space_lock); | |
551 | ubifs_assert(c->budg_idx_growth >= 0); | |
552 | ubifs_assert(c->budg_data_growth >= 0); | |
553 | ubifs_assert(c->budg_dd_growth >= 0); | |
554 | ||
555 | if (unlikely(c->nospace) && (c->nospace_rp || !can_use_rp(c))) { | |
556 | dbg_budg("no space"); | |
557 | spin_unlock(&c->space_lock); | |
558 | return -ENOSPC; | |
559 | } | |
560 | ||
561 | c->budg_idx_growth += idx_growth; | |
562 | c->budg_data_growth += data_growth; | |
563 | c->budg_dd_growth += dd_growth; | |
564 | ||
565 | err = do_budget_space(c); | |
566 | if (likely(!err)) { | |
567 | req->idx_growth = idx_growth; | |
568 | req->data_growth = data_growth; | |
569 | req->dd_growth = dd_growth; | |
570 | spin_unlock(&c->space_lock); | |
571 | return 0; | |
572 | } | |
573 | ||
574 | /* Restore the old values */ | |
575 | c->budg_idx_growth -= idx_growth; | |
576 | c->budg_data_growth -= data_growth; | |
577 | c->budg_dd_growth -= dd_growth; | |
578 | spin_unlock(&c->space_lock); | |
579 | ||
580 | if (req->fast) { | |
581 | dbg_budg("no space for fast budgeting"); | |
582 | return err; | |
583 | } | |
584 | ||
585 | err = make_free_space(c, &ri); | |
586 | if (err == -EAGAIN) { | |
587 | dbg_budg("try again"); | |
588 | cond_resched(); | |
589 | goto again; | |
590 | } else if (err == -ENOSPC) { | |
591 | dbg_budg("FS is full, -ENOSPC"); | |
592 | c->nospace = 1; | |
593 | if (can_use_rp(c) || c->rp_size == 0) | |
594 | c->nospace_rp = 1; | |
595 | smp_wmb(); | |
596 | } else | |
597 | ubifs_err("cannot budget space, error %d", err); | |
598 | return err; | |
599 | } | |
600 | ||
601 | /** | |
602 | * ubifs_release_budget - release budgeted free space. | |
603 | * @c: UBIFS file-system description object | |
604 | * @req: budget request | |
605 | * | |
606 | * This function releases the space budgeted by 'ubifs_budget_space()'. Note, | |
607 | * since the index changes (which were budgeted for in @req->idx_growth) will | |
608 | * only be written to the media on commit, this function moves the index budget | |
609 | * from @c->budg_idx_growth to @c->budg_uncommitted_idx. The latter will be | |
610 | * zeroed by the commit operation. | |
611 | */ | |
612 | void ubifs_release_budget(struct ubifs_info *c, struct ubifs_budget_req *req) | |
613 | { | |
547000da AB |
614 | ubifs_assert(req->new_page <= 1); |
615 | ubifs_assert(req->dirtied_page <= 1); | |
616 | ubifs_assert(req->new_dent <= 1); | |
617 | ubifs_assert(req->mod_dent <= 1); | |
618 | ubifs_assert(req->new_ino <= 1); | |
619 | ubifs_assert(req->new_ino_d <= UBIFS_MAX_INO_DATA); | |
1e51764a AB |
620 | ubifs_assert(req->dirtied_ino <= 4); |
621 | ubifs_assert(req->dirtied_ino_d <= UBIFS_MAX_INO_DATA * 4); | |
dab4b4d2 AB |
622 | ubifs_assert(!(req->new_ino_d & 7)); |
623 | ubifs_assert(!(req->dirtied_ino_d & 7)); | |
1e51764a AB |
624 | if (!req->recalculate) { |
625 | ubifs_assert(req->idx_growth >= 0); | |
626 | ubifs_assert(req->data_growth >= 0); | |
627 | ubifs_assert(req->dd_growth >= 0); | |
628 | } | |
629 | ||
630 | if (req->recalculate) { | |
631 | req->data_growth = calc_data_growth(c, req); | |
632 | req->dd_growth = calc_dd_growth(c, req); | |
633 | req->idx_growth = calc_idx_growth(c, req); | |
634 | } | |
635 | ||
636 | if (!req->data_growth && !req->dd_growth) | |
637 | return; | |
638 | ||
639 | c->nospace = c->nospace_rp = 0; | |
640 | smp_wmb(); | |
641 | ||
642 | spin_lock(&c->space_lock); | |
643 | c->budg_idx_growth -= req->idx_growth; | |
644 | c->budg_uncommitted_idx += req->idx_growth; | |
645 | c->budg_data_growth -= req->data_growth; | |
646 | c->budg_dd_growth -= req->dd_growth; | |
647 | c->min_idx_lebs = ubifs_calc_min_idx_lebs(c); | |
648 | ||
649 | ubifs_assert(c->budg_idx_growth >= 0); | |
650 | ubifs_assert(c->budg_data_growth >= 0); | |
dab4b4d2 | 651 | ubifs_assert(c->budg_dd_growth >= 0); |
1e51764a | 652 | ubifs_assert(c->min_idx_lebs < c->main_lebs); |
dab4b4d2 AB |
653 | ubifs_assert(!(c->budg_idx_growth & 7)); |
654 | ubifs_assert(!(c->budg_data_growth & 7)); | |
655 | ubifs_assert(!(c->budg_dd_growth & 7)); | |
1e51764a AB |
656 | spin_unlock(&c->space_lock); |
657 | } | |
658 | ||
659 | /** | |
660 | * ubifs_convert_page_budget - convert budget of a new page. | |
661 | * @c: UBIFS file-system description object | |
662 | * | |
663 | * This function converts budget which was allocated for a new page of data to | |
664 | * the budget of changing an existing page of data. The latter is smaller then | |
665 | * the former, so this function only does simple re-calculation and does not | |
666 | * involve any write-back. | |
667 | */ | |
668 | void ubifs_convert_page_budget(struct ubifs_info *c) | |
669 | { | |
670 | spin_lock(&c->space_lock); | |
671 | /* Release the index growth reservation */ | |
672 | c->budg_idx_growth -= c->max_idx_node_sz << UBIFS_BLOCKS_PER_PAGE_SHIFT; | |
673 | /* Release the data growth reservation */ | |
674 | c->budg_data_growth -= c->page_budget; | |
675 | /* Increase the dirty data growth reservation instead */ | |
676 | c->budg_dd_growth += c->page_budget; | |
677 | /* And re-calculate the indexing space reservation */ | |
678 | c->min_idx_lebs = ubifs_calc_min_idx_lebs(c); | |
679 | spin_unlock(&c->space_lock); | |
680 | } | |
681 | ||
682 | /** | |
683 | * ubifs_release_dirty_inode_budget - release dirty inode budget. | |
684 | * @c: UBIFS file-system description object | |
685 | * @ui: UBIFS inode to release the budget for | |
686 | * | |
687 | * This function releases budget corresponding to a dirty inode. It is usually | |
688 | * called when after the inode has been written to the media and marked as | |
689 | * clean. | |
690 | */ | |
691 | void ubifs_release_dirty_inode_budget(struct ubifs_info *c, | |
692 | struct ubifs_inode *ui) | |
693 | { | |
182854b4 | 694 | struct ubifs_budget_req req; |
1e51764a | 695 | |
182854b4 | 696 | memset(&req, 0, sizeof(struct ubifs_budget_req)); |
dab4b4d2 | 697 | req.dd_growth = c->inode_budget + ALIGN(ui->data_len, 8); |
1e51764a AB |
698 | ubifs_release_budget(c, &req); |
699 | } | |
700 | ||
4b5f2762 AB |
701 | /** |
702 | * ubifs_reported_space - calculate reported free space. | |
703 | * @c: the UBIFS file-system description object | |
704 | * @free: amount of free space | |
705 | * | |
706 | * This function calculates amount of free space which will be reported to | |
707 | * user-space. User-space application tend to expect that if the file-system | |
708 | * (e.g., via the 'statfs()' call) reports that it has N bytes available, they | |
709 | * are able to write a file of size N. UBIFS attaches node headers to each data | |
710 | * node and it has to write indexind nodes as well. This introduces additional | |
21a60258 AB |
711 | * overhead, and UBIFS has to report sligtly less free space to meet the above |
712 | * expectetions. | |
4b5f2762 AB |
713 | * |
714 | * This function assumes free space is made up of uncompressed data nodes and | |
715 | * full index nodes (one per data node, tripled because we always allow enough | |
716 | * space to write the index thrice). | |
717 | * | |
718 | * Note, the calculation is pessimistic, which means that most of the time | |
719 | * UBIFS reports less space than it actually has. | |
720 | */ | |
721 | long long ubifs_reported_space(const struct ubifs_info *c, uint64_t free) | |
722 | { | |
f171d4d7 | 723 | int divisor, factor, f; |
4b5f2762 AB |
724 | |
725 | /* | |
726 | * Reported space size is @free * X, where X is UBIFS block size | |
727 | * divided by UBIFS block size + all overhead one data block | |
728 | * introduces. The overhead is the node header + indexing overhead. | |
729 | * | |
f171d4d7 AB |
730 | * Indexing overhead calculations are based on the following formula: |
731 | * I = N/(f - 1) + 1, where I - number of indexing nodes, N - number | |
732 | * of data nodes, f - fanout. Because effective UBIFS fanout is twice | |
733 | * as less than maximum fanout, we assume that each data node | |
4b5f2762 AB |
734 | * introduces 3 * @c->max_idx_node_sz / (@c->fanout/2 - 1) bytes. |
735 | * Note, the multiplier 3 is because UBIFS reseves thrice as more space | |
736 | * for the index. | |
737 | */ | |
f171d4d7 | 738 | f = c->fanout > 3 ? c->fanout >> 1 : 2; |
4b5f2762 AB |
739 | factor = UBIFS_BLOCK_SIZE; |
740 | divisor = UBIFS_MAX_DATA_NODE_SZ; | |
f171d4d7 | 741 | divisor += (c->max_idx_node_sz * 3) / (f - 1); |
4b5f2762 AB |
742 | free *= factor; |
743 | do_div(free, divisor); | |
744 | return free; | |
745 | } | |
746 | ||
1e51764a | 747 | /** |
7dad181b | 748 | * ubifs_get_free_space - return amount of free space. |
1e51764a AB |
749 | * @c: UBIFS file-system description object |
750 | * | |
7dad181b AB |
751 | * This function calculates amount of free space to report to user-space. |
752 | * | |
753 | * Because UBIFS may introduce substantial overhead (the index, node headers, | |
754 | * alighment, wastage at the end of eraseblocks, etc), it cannot report real | |
755 | * amount of free flash space it has (well, because not all dirty space is | |
756 | * reclamable, UBIFS does not actually know the real amount). If UBIFS did so, | |
757 | * it would bread user expectetion about what free space is. Users seem to | |
758 | * accustomed to assume that if the file-system reports N bytes of free space, | |
759 | * they would be able to fit a file of N bytes to the FS. This almost works for | |
760 | * traditional file-systems, because they have way less overhead than UBIFS. | |
761 | * So, to keep users happy, UBIFS tries to take the overhead into account. | |
1e51764a | 762 | */ |
7dad181b | 763 | long long ubifs_get_free_space(struct ubifs_info *c) |
1e51764a | 764 | { |
7dad181b | 765 | int min_idx_lebs, rsvd_idx_lebs, lebs; |
1e51764a AB |
766 | long long available, outstanding, free; |
767 | ||
1e51764a AB |
768 | spin_lock(&c->space_lock); |
769 | min_idx_lebs = ubifs_calc_min_idx_lebs(c); | |
1e51764a | 770 | outstanding = c->budg_data_growth + c->budg_dd_growth; |
9e5de354 AB |
771 | |
772 | /* | |
773 | * Force the amount available to the total size reported if the used | |
774 | * space is zero. | |
775 | */ | |
776 | if (c->lst.total_used <= UBIFS_INO_NODE_SZ && !outstanding) { | |
777 | spin_unlock(&c->space_lock); | |
778 | return (long long)c->block_cnt << UBIFS_BLOCK_SHIFT; | |
779 | } | |
780 | ||
781 | available = ubifs_calc_available(c, min_idx_lebs); | |
7dad181b AB |
782 | |
783 | /* | |
784 | * When reporting free space to user-space, UBIFS guarantees that it is | |
785 | * possible to write a file of free space size. This means that for | |
786 | * empty LEBs we may use more precise calculations than | |
787 | * 'ubifs_calc_available()' is using. Namely, we know that in empty | |
788 | * LEBs we would waste only @c->leb_overhead bytes, not @c->dark_wm. | |
789 | * Thus, amend the available space. | |
790 | * | |
791 | * Note, the calculations below are similar to what we have in | |
792 | * 'do_budget_space()', so refer there for comments. | |
793 | */ | |
794 | if (min_idx_lebs > c->lst.idx_lebs) | |
795 | rsvd_idx_lebs = min_idx_lebs - c->lst.idx_lebs; | |
796 | else | |
797 | rsvd_idx_lebs = 0; | |
798 | lebs = c->lst.empty_lebs + c->freeable_cnt + c->idx_gc_cnt - | |
799 | c->lst.taken_empty_lebs; | |
800 | lebs -= rsvd_idx_lebs; | |
801 | available += lebs * (c->dark_wm - c->leb_overhead); | |
1e51764a AB |
802 | spin_unlock(&c->space_lock); |
803 | ||
804 | if (available > outstanding) | |
805 | free = ubifs_reported_space(c, available - outstanding); | |
806 | else | |
807 | free = 0; | |
808 | return free; | |
809 | } |