Artem Bityutskiy | 1e51764 | 2008-07-14 19:08:37 +0300 | [diff] [blame] | 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 garbage collection. The procedure for garbage collection |
| 25 | * is different depending on whether a LEB as an index LEB (contains index |
| 26 | * nodes) or not. For non-index LEBs, garbage collection finds a LEB which |
| 27 | * contains a lot of dirty space (obsolete nodes), and copies the non-obsolete |
| 28 | * nodes to the journal, at which point the garbage-collected LEB is free to be |
| 29 | * reused. For index LEBs, garbage collection marks the non-obsolete index nodes |
| 30 | * dirty in the TNC, and after the next commit, the garbage-collected LEB is |
| 31 | * to be reused. Garbage collection will cause the number of dirty index nodes |
| 32 | * to grow, however sufficient space is reserved for the index to ensure the |
| 33 | * commit will never run out of space. |
| 34 | */ |
| 35 | |
| 36 | #include <linux/pagemap.h> |
| 37 | #include "ubifs.h" |
| 38 | |
| 39 | /* |
| 40 | * GC tries to optimize the way it fit nodes to available space, and it sorts |
| 41 | * nodes a little. The below constants are watermarks which define "large", |
| 42 | * "medium", and "small" nodes. |
| 43 | */ |
| 44 | #define MEDIUM_NODE_WM (UBIFS_BLOCK_SIZE / 4) |
| 45 | #define SMALL_NODE_WM UBIFS_MAX_DENT_NODE_SZ |
| 46 | |
| 47 | /* |
| 48 | * GC may need to move more then one LEB to make progress. The below constants |
| 49 | * define "soft" and "hard" limits on the number of LEBs the garbage collector |
| 50 | * may move. |
| 51 | */ |
| 52 | #define SOFT_LEBS_LIMIT 4 |
| 53 | #define HARD_LEBS_LIMIT 32 |
| 54 | |
| 55 | /** |
| 56 | * switch_gc_head - switch the garbage collection journal head. |
| 57 | * @c: UBIFS file-system description object |
| 58 | * @buf: buffer to write |
| 59 | * @len: length of the buffer to write |
| 60 | * @lnum: LEB number written is returned here |
| 61 | * @offs: offset written is returned here |
| 62 | * |
| 63 | * This function switch the GC head to the next LEB which is reserved in |
| 64 | * @c->gc_lnum. Returns %0 in case of success, %-EAGAIN if commit is required, |
| 65 | * and other negative error code in case of failures. |
| 66 | */ |
| 67 | static int switch_gc_head(struct ubifs_info *c) |
| 68 | { |
| 69 | int err, gc_lnum = c->gc_lnum; |
| 70 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; |
| 71 | |
| 72 | ubifs_assert(gc_lnum != -1); |
| 73 | dbg_gc("switch GC head from LEB %d:%d to LEB %d (waste %d bytes)", |
| 74 | wbuf->lnum, wbuf->offs + wbuf->used, gc_lnum, |
| 75 | c->leb_size - wbuf->offs - wbuf->used); |
| 76 | |
| 77 | err = ubifs_wbuf_sync_nolock(wbuf); |
| 78 | if (err) |
| 79 | return err; |
| 80 | |
| 81 | /* |
| 82 | * The GC write-buffer was synchronized, we may safely unmap |
| 83 | * 'c->gc_lnum'. |
| 84 | */ |
| 85 | err = ubifs_leb_unmap(c, gc_lnum); |
| 86 | if (err) |
| 87 | return err; |
| 88 | |
| 89 | err = ubifs_add_bud_to_log(c, GCHD, gc_lnum, 0); |
| 90 | if (err) |
| 91 | return err; |
| 92 | |
| 93 | c->gc_lnum = -1; |
| 94 | err = ubifs_wbuf_seek_nolock(wbuf, gc_lnum, 0, UBI_LONGTERM); |
| 95 | return err; |
| 96 | } |
| 97 | |
| 98 | /** |
| 99 | * move_nodes - move nodes. |
| 100 | * @c: UBIFS file-system description object |
| 101 | * @sleb: describes nodes to move |
| 102 | * |
| 103 | * This function moves valid nodes from data LEB described by @sleb to the GC |
| 104 | * journal head. The obsolete nodes are dropped. |
| 105 | * |
| 106 | * When moving nodes we have to deal with classical bin-packing problem: the |
| 107 | * space in the current GC journal head LEB and in @c->gc_lnum are the "bins", |
| 108 | * where the nodes in the @sleb->nodes list are the elements which should be |
| 109 | * fit optimally to the bins. This function uses the "first fit decreasing" |
| 110 | * strategy, although it does not really sort the nodes but just split them on |
| 111 | * 3 classes - large, medium, and small, so they are roughly sorted. |
| 112 | * |
| 113 | * This function returns zero in case of success, %-EAGAIN if commit is |
| 114 | * required, and other negative error codes in case of other failures. |
| 115 | */ |
| 116 | static int move_nodes(struct ubifs_info *c, struct ubifs_scan_leb *sleb) |
| 117 | { |
| 118 | struct ubifs_scan_node *snod, *tmp; |
| 119 | struct list_head large, medium, small; |
| 120 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; |
| 121 | int avail, err, min = INT_MAX; |
| 122 | |
| 123 | INIT_LIST_HEAD(&large); |
| 124 | INIT_LIST_HEAD(&medium); |
| 125 | INIT_LIST_HEAD(&small); |
| 126 | |
| 127 | list_for_each_entry_safe(snod, tmp, &sleb->nodes, list) { |
| 128 | struct list_head *lst; |
| 129 | |
| 130 | ubifs_assert(snod->type != UBIFS_IDX_NODE); |
| 131 | ubifs_assert(snod->type != UBIFS_REF_NODE); |
| 132 | ubifs_assert(snod->type != UBIFS_CS_NODE); |
| 133 | |
| 134 | err = ubifs_tnc_has_node(c, &snod->key, 0, sleb->lnum, |
| 135 | snod->offs, 0); |
| 136 | if (err < 0) |
| 137 | goto out; |
| 138 | |
| 139 | lst = &snod->list; |
| 140 | list_del(lst); |
| 141 | if (!err) { |
| 142 | /* The node is obsolete, remove it from the list */ |
| 143 | kfree(snod); |
| 144 | continue; |
| 145 | } |
| 146 | |
| 147 | /* |
| 148 | * Sort the list of nodes so that large nodes go first, and |
| 149 | * small nodes go last. |
| 150 | */ |
| 151 | if (snod->len > MEDIUM_NODE_WM) |
| 152 | list_add(lst, &large); |
| 153 | else if (snod->len > SMALL_NODE_WM) |
| 154 | list_add(lst, &medium); |
| 155 | else |
| 156 | list_add(lst, &small); |
| 157 | |
| 158 | /* And find the smallest node */ |
| 159 | if (snod->len < min) |
| 160 | min = snod->len; |
| 161 | } |
| 162 | |
| 163 | /* |
| 164 | * Join the tree lists so that we'd have one roughly sorted list |
| 165 | * ('large' will be the head of the joined list). |
| 166 | */ |
| 167 | list_splice(&medium, large.prev); |
| 168 | list_splice(&small, large.prev); |
| 169 | |
| 170 | if (wbuf->lnum == -1) { |
| 171 | /* |
| 172 | * The GC journal head is not set, because it is the first GC |
| 173 | * invocation since mount. |
| 174 | */ |
| 175 | err = switch_gc_head(c); |
| 176 | if (err) |
| 177 | goto out; |
| 178 | } |
| 179 | |
| 180 | /* Write nodes to their new location. Use the first-fit strategy */ |
| 181 | while (1) { |
| 182 | avail = c->leb_size - wbuf->offs - wbuf->used; |
| 183 | list_for_each_entry_safe(snod, tmp, &large, list) { |
| 184 | int new_lnum, new_offs; |
| 185 | |
| 186 | if (avail < min) |
| 187 | break; |
| 188 | |
| 189 | if (snod->len > avail) |
| 190 | /* This node does not fit */ |
| 191 | continue; |
| 192 | |
| 193 | cond_resched(); |
| 194 | |
| 195 | new_lnum = wbuf->lnum; |
| 196 | new_offs = wbuf->offs + wbuf->used; |
| 197 | err = ubifs_wbuf_write_nolock(wbuf, snod->node, |
| 198 | snod->len); |
| 199 | if (err) |
| 200 | goto out; |
| 201 | err = ubifs_tnc_replace(c, &snod->key, sleb->lnum, |
| 202 | snod->offs, new_lnum, new_offs, |
| 203 | snod->len); |
| 204 | if (err) |
| 205 | goto out; |
| 206 | |
| 207 | avail = c->leb_size - wbuf->offs - wbuf->used; |
| 208 | list_del(&snod->list); |
| 209 | kfree(snod); |
| 210 | } |
| 211 | |
| 212 | if (list_empty(&large)) |
| 213 | break; |
| 214 | |
| 215 | /* |
| 216 | * Waste the rest of the space in the LEB and switch to the |
| 217 | * next LEB. |
| 218 | */ |
| 219 | err = switch_gc_head(c); |
| 220 | if (err) |
| 221 | goto out; |
| 222 | } |
| 223 | |
| 224 | return 0; |
| 225 | |
| 226 | out: |
| 227 | list_for_each_entry_safe(snod, tmp, &large, list) { |
| 228 | list_del(&snod->list); |
| 229 | kfree(snod); |
| 230 | } |
| 231 | return err; |
| 232 | } |
| 233 | |
| 234 | /** |
| 235 | * gc_sync_wbufs - sync write-buffers for GC. |
| 236 | * @c: UBIFS file-system description object |
| 237 | * |
| 238 | * We must guarantee that obsoleting nodes are on flash. Unfortunately they may |
| 239 | * be in a write-buffer instead. That is, a node could be written to a |
| 240 | * write-buffer, obsoleting another node in a LEB that is GC'd. If that LEB is |
| 241 | * erased before the write-buffer is sync'd and then there is an unclean |
| 242 | * unmount, then an existing node is lost. To avoid this, we sync all |
| 243 | * write-buffers. |
| 244 | * |
| 245 | * This function returns %0 on success or a negative error code on failure. |
| 246 | */ |
| 247 | static int gc_sync_wbufs(struct ubifs_info *c) |
| 248 | { |
| 249 | int err, i; |
| 250 | |
| 251 | for (i = 0; i < c->jhead_cnt; i++) { |
| 252 | if (i == GCHD) |
| 253 | continue; |
| 254 | err = ubifs_wbuf_sync(&c->jheads[i].wbuf); |
| 255 | if (err) |
| 256 | return err; |
| 257 | } |
| 258 | return 0; |
| 259 | } |
| 260 | |
| 261 | /** |
| 262 | * ubifs_garbage_collect_leb - garbage-collect a logical eraseblock. |
| 263 | * @c: UBIFS file-system description object |
| 264 | * @lp: describes the LEB to garbage collect |
| 265 | * |
| 266 | * This function garbage-collects an LEB and returns one of the @LEB_FREED, |
| 267 | * @LEB_RETAINED, etc positive codes in case of success, %-EAGAIN if commit is |
| 268 | * required, and other negative error codes in case of failures. |
| 269 | */ |
| 270 | int ubifs_garbage_collect_leb(struct ubifs_info *c, struct ubifs_lprops *lp) |
| 271 | { |
| 272 | struct ubifs_scan_leb *sleb; |
| 273 | struct ubifs_scan_node *snod; |
| 274 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; |
| 275 | int err = 0, lnum = lp->lnum; |
| 276 | |
| 277 | ubifs_assert(c->gc_lnum != -1 || wbuf->offs + wbuf->used == 0 || |
| 278 | c->need_recovery); |
| 279 | ubifs_assert(c->gc_lnum != lnum); |
| 280 | ubifs_assert(wbuf->lnum != lnum); |
| 281 | |
| 282 | /* |
| 283 | * We scan the entire LEB even though we only really need to scan up to |
| 284 | * (c->leb_size - lp->free). |
| 285 | */ |
| 286 | sleb = ubifs_scan(c, lnum, 0, c->sbuf); |
| 287 | if (IS_ERR(sleb)) |
| 288 | return PTR_ERR(sleb); |
| 289 | |
| 290 | ubifs_assert(!list_empty(&sleb->nodes)); |
| 291 | snod = list_entry(sleb->nodes.next, struct ubifs_scan_node, list); |
| 292 | |
| 293 | if (snod->type == UBIFS_IDX_NODE) { |
| 294 | struct ubifs_gced_idx_leb *idx_gc; |
| 295 | |
| 296 | dbg_gc("indexing LEB %d (free %d, dirty %d)", |
| 297 | lnum, lp->free, lp->dirty); |
| 298 | list_for_each_entry(snod, &sleb->nodes, list) { |
| 299 | struct ubifs_idx_node *idx = snod->node; |
| 300 | int level = le16_to_cpu(idx->level); |
| 301 | |
| 302 | ubifs_assert(snod->type == UBIFS_IDX_NODE); |
| 303 | key_read(c, ubifs_idx_key(c, idx), &snod->key); |
| 304 | err = ubifs_dirty_idx_node(c, &snod->key, level, lnum, |
| 305 | snod->offs); |
| 306 | if (err) |
| 307 | goto out; |
| 308 | } |
| 309 | |
| 310 | idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS); |
| 311 | if (!idx_gc) { |
| 312 | err = -ENOMEM; |
| 313 | goto out; |
| 314 | } |
| 315 | |
| 316 | idx_gc->lnum = lnum; |
| 317 | idx_gc->unmap = 0; |
| 318 | list_add(&idx_gc->list, &c->idx_gc); |
| 319 | |
| 320 | /* |
| 321 | * Don't release the LEB until after the next commit, because |
| 322 | * it may contain date which is needed for recovery. So |
| 323 | * although we freed this LEB, it will become usable only after |
| 324 | * the commit. |
| 325 | */ |
| 326 | err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, |
| 327 | LPROPS_INDEX, 1); |
| 328 | if (err) |
| 329 | goto out; |
| 330 | err = LEB_FREED_IDX; |
| 331 | } else { |
| 332 | dbg_gc("data LEB %d (free %d, dirty %d)", |
| 333 | lnum, lp->free, lp->dirty); |
| 334 | |
| 335 | err = move_nodes(c, sleb); |
| 336 | if (err) |
Adrian Hunter | 6dcfac4 | 2008-09-12 12:27:47 +0300 | [diff] [blame^] | 337 | goto out_inc_seq; |
Artem Bityutskiy | 1e51764 | 2008-07-14 19:08:37 +0300 | [diff] [blame] | 338 | |
| 339 | err = gc_sync_wbufs(c); |
| 340 | if (err) |
Adrian Hunter | 6dcfac4 | 2008-09-12 12:27:47 +0300 | [diff] [blame^] | 341 | goto out_inc_seq; |
Artem Bityutskiy | 1e51764 | 2008-07-14 19:08:37 +0300 | [diff] [blame] | 342 | |
| 343 | err = ubifs_change_one_lp(c, lnum, c->leb_size, 0, 0, 0, 0); |
| 344 | if (err) |
Adrian Hunter | 6dcfac4 | 2008-09-12 12:27:47 +0300 | [diff] [blame^] | 345 | goto out_inc_seq; |
Artem Bityutskiy | 1e51764 | 2008-07-14 19:08:37 +0300 | [diff] [blame] | 346 | |
Adrian Hunter | 601c0bc | 2008-08-22 14:23:35 +0300 | [diff] [blame] | 347 | /* Allow for races with TNC */ |
| 348 | c->gced_lnum = lnum; |
| 349 | smp_wmb(); |
| 350 | c->gc_seq += 1; |
| 351 | smp_wmb(); |
| 352 | |
Artem Bityutskiy | 1e51764 | 2008-07-14 19:08:37 +0300 | [diff] [blame] | 353 | if (c->gc_lnum == -1) { |
| 354 | c->gc_lnum = lnum; |
| 355 | err = LEB_RETAINED; |
| 356 | } else { |
| 357 | err = ubifs_wbuf_sync_nolock(wbuf); |
| 358 | if (err) |
| 359 | goto out; |
| 360 | |
| 361 | err = ubifs_leb_unmap(c, lnum); |
| 362 | if (err) |
| 363 | goto out; |
| 364 | |
| 365 | err = LEB_FREED; |
| 366 | } |
| 367 | } |
| 368 | |
| 369 | out: |
| 370 | ubifs_scan_destroy(sleb); |
| 371 | return err; |
Adrian Hunter | 6dcfac4 | 2008-09-12 12:27:47 +0300 | [diff] [blame^] | 372 | |
| 373 | out_inc_seq: |
| 374 | /* We may have moved at least some nodes so allow for races with TNC */ |
| 375 | c->gced_lnum = lnum; |
| 376 | smp_wmb(); |
| 377 | c->gc_seq += 1; |
| 378 | smp_wmb(); |
| 379 | goto out; |
Artem Bityutskiy | 1e51764 | 2008-07-14 19:08:37 +0300 | [diff] [blame] | 380 | } |
| 381 | |
| 382 | /** |
| 383 | * ubifs_garbage_collect - UBIFS garbage collector. |
| 384 | * @c: UBIFS file-system description object |
| 385 | * @anyway: do GC even if there are free LEBs |
| 386 | * |
| 387 | * This function does out-of-place garbage collection. The return codes are: |
| 388 | * o positive LEB number if the LEB has been freed and may be used; |
| 389 | * o %-EAGAIN if the caller has to run commit; |
| 390 | * o %-ENOSPC if GC failed to make any progress; |
| 391 | * o other negative error codes in case of other errors. |
| 392 | * |
| 393 | * Garbage collector writes data to the journal when GC'ing data LEBs, and just |
| 394 | * marking indexing nodes dirty when GC'ing indexing LEBs. Thus, at some point |
| 395 | * commit may be required. But commit cannot be run from inside GC, because the |
| 396 | * caller might be holding the commit lock, so %-EAGAIN is returned instead; |
| 397 | * And this error code means that the caller has to run commit, and re-run GC |
| 398 | * if there is still no free space. |
| 399 | * |
| 400 | * There are many reasons why this function may return %-EAGAIN: |
| 401 | * o the log is full and there is no space to write an LEB reference for |
| 402 | * @c->gc_lnum; |
| 403 | * o the journal is too large and exceeds size limitations; |
| 404 | * o GC moved indexing LEBs, but they can be used only after the commit; |
| 405 | * o the shrinker fails to find clean znodes to free and requests the commit; |
| 406 | * o etc. |
| 407 | * |
| 408 | * Note, if the file-system is close to be full, this function may return |
| 409 | * %-EAGAIN infinitely, so the caller has to limit amount of re-invocations of |
| 410 | * the function. E.g., this happens if the limits on the journal size are too |
| 411 | * tough and GC writes too much to the journal before an LEB is freed. This |
| 412 | * might also mean that the journal is too large, and the TNC becomes to big, |
| 413 | * so that the shrinker is constantly called, finds not clean znodes to free, |
| 414 | * and requests commit. Well, this may also happen if the journal is all right, |
| 415 | * but another kernel process consumes too much memory. Anyway, infinite |
| 416 | * %-EAGAIN may happen, but in some extreme/misconfiguration cases. |
| 417 | */ |
| 418 | int ubifs_garbage_collect(struct ubifs_info *c, int anyway) |
| 419 | { |
| 420 | int i, err, ret, min_space = c->dead_wm; |
| 421 | struct ubifs_lprops lp; |
| 422 | struct ubifs_wbuf *wbuf = &c->jheads[GCHD].wbuf; |
| 423 | |
| 424 | ubifs_assert_cmt_locked(c); |
| 425 | |
| 426 | if (ubifs_gc_should_commit(c)) |
| 427 | return -EAGAIN; |
| 428 | |
| 429 | mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead); |
| 430 | |
| 431 | if (c->ro_media) { |
| 432 | ret = -EROFS; |
| 433 | goto out_unlock; |
| 434 | } |
| 435 | |
| 436 | /* We expect the write-buffer to be empty on entry */ |
| 437 | ubifs_assert(!wbuf->used); |
| 438 | |
| 439 | for (i = 0; ; i++) { |
| 440 | int space_before = c->leb_size - wbuf->offs - wbuf->used; |
| 441 | int space_after; |
| 442 | |
| 443 | cond_resched(); |
| 444 | |
| 445 | /* Give the commit an opportunity to run */ |
| 446 | if (ubifs_gc_should_commit(c)) { |
| 447 | ret = -EAGAIN; |
| 448 | break; |
| 449 | } |
| 450 | |
| 451 | if (i > SOFT_LEBS_LIMIT && !list_empty(&c->idx_gc)) { |
| 452 | /* |
| 453 | * We've done enough iterations. Indexing LEBs were |
| 454 | * moved and will be available after the commit. |
| 455 | */ |
| 456 | dbg_gc("soft limit, some index LEBs GC'ed, -EAGAIN"); |
| 457 | ubifs_commit_required(c); |
| 458 | ret = -EAGAIN; |
| 459 | break; |
| 460 | } |
| 461 | |
| 462 | if (i > HARD_LEBS_LIMIT) { |
| 463 | /* |
| 464 | * We've moved too many LEBs and have not made |
| 465 | * progress, give up. |
| 466 | */ |
| 467 | dbg_gc("hard limit, -ENOSPC"); |
| 468 | ret = -ENOSPC; |
| 469 | break; |
| 470 | } |
| 471 | |
| 472 | /* |
| 473 | * Empty and freeable LEBs can turn up while we waited for |
| 474 | * the wbuf lock, or while we have been running GC. In that |
| 475 | * case, we should just return one of those instead of |
| 476 | * continuing to GC dirty LEBs. Hence we request |
| 477 | * 'ubifs_find_dirty_leb()' to return an empty LEB if it can. |
| 478 | */ |
| 479 | ret = ubifs_find_dirty_leb(c, &lp, min_space, anyway ? 0 : 1); |
| 480 | if (ret) { |
| 481 | if (ret == -ENOSPC) |
| 482 | dbg_gc("no more dirty LEBs"); |
| 483 | break; |
| 484 | } |
| 485 | |
| 486 | dbg_gc("found LEB %d: free %d, dirty %d, sum %d " |
| 487 | "(min. space %d)", lp.lnum, lp.free, lp.dirty, |
| 488 | lp.free + lp.dirty, min_space); |
| 489 | |
| 490 | if (lp.free + lp.dirty == c->leb_size) { |
| 491 | /* An empty LEB was returned */ |
| 492 | dbg_gc("LEB %d is free, return it", lp.lnum); |
| 493 | /* |
| 494 | * ubifs_find_dirty_leb() doesn't return freeable index |
| 495 | * LEBs. |
| 496 | */ |
| 497 | ubifs_assert(!(lp.flags & LPROPS_INDEX)); |
| 498 | if (lp.free != c->leb_size) { |
| 499 | /* |
| 500 | * Write buffers must be sync'd before |
| 501 | * unmapping freeable LEBs, because one of them |
| 502 | * may contain data which obsoletes something |
| 503 | * in 'lp.pnum'. |
| 504 | */ |
| 505 | ret = gc_sync_wbufs(c); |
| 506 | if (ret) |
| 507 | goto out; |
| 508 | ret = ubifs_change_one_lp(c, lp.lnum, |
| 509 | c->leb_size, 0, 0, 0, |
| 510 | 0); |
| 511 | if (ret) |
| 512 | goto out; |
| 513 | } |
| 514 | ret = ubifs_leb_unmap(c, lp.lnum); |
| 515 | if (ret) |
| 516 | goto out; |
| 517 | ret = lp.lnum; |
| 518 | break; |
| 519 | } |
| 520 | |
| 521 | space_before = c->leb_size - wbuf->offs - wbuf->used; |
| 522 | if (wbuf->lnum == -1) |
| 523 | space_before = 0; |
| 524 | |
| 525 | ret = ubifs_garbage_collect_leb(c, &lp); |
| 526 | if (ret < 0) { |
| 527 | if (ret == -EAGAIN || ret == -ENOSPC) { |
| 528 | /* |
| 529 | * These codes are not errors, so we have to |
| 530 | * return the LEB to lprops. But if the |
| 531 | * 'ubifs_return_leb()' function fails, its |
| 532 | * failure code is propagated to the caller |
| 533 | * instead of the original '-EAGAIN' or |
| 534 | * '-ENOSPC'. |
| 535 | */ |
| 536 | err = ubifs_return_leb(c, lp.lnum); |
| 537 | if (err) |
| 538 | ret = err; |
| 539 | break; |
| 540 | } |
| 541 | goto out; |
| 542 | } |
| 543 | |
| 544 | if (ret == LEB_FREED) { |
| 545 | /* An LEB has been freed and is ready for use */ |
| 546 | dbg_gc("LEB %d freed, return", lp.lnum); |
| 547 | ret = lp.lnum; |
| 548 | break; |
| 549 | } |
| 550 | |
| 551 | if (ret == LEB_FREED_IDX) { |
| 552 | /* |
| 553 | * This was an indexing LEB and it cannot be |
| 554 | * immediately used. And instead of requesting the |
| 555 | * commit straight away, we try to garbage collect some |
| 556 | * more. |
| 557 | */ |
| 558 | dbg_gc("indexing LEB %d freed, continue", lp.lnum); |
| 559 | continue; |
| 560 | } |
| 561 | |
| 562 | ubifs_assert(ret == LEB_RETAINED); |
| 563 | space_after = c->leb_size - wbuf->offs - wbuf->used; |
| 564 | dbg_gc("LEB %d retained, freed %d bytes", lp.lnum, |
| 565 | space_after - space_before); |
| 566 | |
| 567 | if (space_after > space_before) { |
| 568 | /* GC makes progress, keep working */ |
| 569 | min_space >>= 1; |
| 570 | if (min_space < c->dead_wm) |
| 571 | min_space = c->dead_wm; |
| 572 | continue; |
| 573 | } |
| 574 | |
| 575 | dbg_gc("did not make progress"); |
| 576 | |
| 577 | /* |
| 578 | * GC moved an LEB bud have not done any progress. This means |
| 579 | * that the previous GC head LEB contained too few free space |
| 580 | * and the LEB which was GC'ed contained only large nodes which |
| 581 | * did not fit that space. |
| 582 | * |
| 583 | * We can do 2 things: |
| 584 | * 1. pick another LEB in a hope it'll contain a small node |
| 585 | * which will fit the space we have at the end of current GC |
| 586 | * head LEB, but there is no guarantee, so we try this out |
| 587 | * unless we have already been working for too long; |
| 588 | * 2. request an LEB with more dirty space, which will force |
| 589 | * 'ubifs_find_dirty_leb()' to start scanning the lprops |
| 590 | * table, instead of just picking one from the heap |
| 591 | * (previously it already picked the dirtiest LEB). |
| 592 | */ |
| 593 | if (i < SOFT_LEBS_LIMIT) { |
| 594 | dbg_gc("try again"); |
| 595 | continue; |
| 596 | } |
| 597 | |
| 598 | min_space <<= 1; |
| 599 | if (min_space > c->dark_wm) |
| 600 | min_space = c->dark_wm; |
| 601 | dbg_gc("set min. space to %d", min_space); |
| 602 | } |
| 603 | |
| 604 | if (ret == -ENOSPC && !list_empty(&c->idx_gc)) { |
| 605 | dbg_gc("no space, some index LEBs GC'ed, -EAGAIN"); |
| 606 | ubifs_commit_required(c); |
| 607 | ret = -EAGAIN; |
| 608 | } |
| 609 | |
| 610 | err = ubifs_wbuf_sync_nolock(wbuf); |
| 611 | if (!err) |
| 612 | err = ubifs_leb_unmap(c, c->gc_lnum); |
| 613 | if (err) { |
| 614 | ret = err; |
| 615 | goto out; |
| 616 | } |
| 617 | out_unlock: |
| 618 | mutex_unlock(&wbuf->io_mutex); |
| 619 | return ret; |
| 620 | |
| 621 | out: |
| 622 | ubifs_assert(ret < 0); |
| 623 | ubifs_assert(ret != -ENOSPC && ret != -EAGAIN); |
| 624 | ubifs_ro_mode(c, ret); |
| 625 | ubifs_wbuf_sync_nolock(wbuf); |
| 626 | mutex_unlock(&wbuf->io_mutex); |
| 627 | ubifs_return_leb(c, lp.lnum); |
| 628 | return ret; |
| 629 | } |
| 630 | |
| 631 | /** |
| 632 | * ubifs_gc_start_commit - garbage collection at start of commit. |
| 633 | * @c: UBIFS file-system description object |
| 634 | * |
| 635 | * If a LEB has only dirty and free space, then we may safely unmap it and make |
| 636 | * it free. Note, we cannot do this with indexing LEBs because dirty space may |
| 637 | * correspond index nodes that are required for recovery. In that case, the |
| 638 | * LEB cannot be unmapped until after the next commit. |
| 639 | * |
| 640 | * This function returns %0 upon success and a negative error code upon failure. |
| 641 | */ |
| 642 | int ubifs_gc_start_commit(struct ubifs_info *c) |
| 643 | { |
| 644 | struct ubifs_gced_idx_leb *idx_gc; |
| 645 | const struct ubifs_lprops *lp; |
| 646 | int err = 0, flags; |
| 647 | |
| 648 | ubifs_get_lprops(c); |
| 649 | |
| 650 | /* |
| 651 | * Unmap (non-index) freeable LEBs. Note that recovery requires that all |
| 652 | * wbufs are sync'd before this, which is done in 'do_commit()'. |
| 653 | */ |
| 654 | while (1) { |
| 655 | lp = ubifs_fast_find_freeable(c); |
| 656 | if (unlikely(IS_ERR(lp))) { |
| 657 | err = PTR_ERR(lp); |
| 658 | goto out; |
| 659 | } |
| 660 | if (!lp) |
| 661 | break; |
| 662 | ubifs_assert(!(lp->flags & LPROPS_TAKEN)); |
| 663 | ubifs_assert(!(lp->flags & LPROPS_INDEX)); |
| 664 | err = ubifs_leb_unmap(c, lp->lnum); |
| 665 | if (err) |
| 666 | goto out; |
| 667 | lp = ubifs_change_lp(c, lp, c->leb_size, 0, lp->flags, 0); |
| 668 | if (unlikely(IS_ERR(lp))) { |
| 669 | err = PTR_ERR(lp); |
| 670 | goto out; |
| 671 | } |
| 672 | ubifs_assert(!(lp->flags & LPROPS_TAKEN)); |
| 673 | ubifs_assert(!(lp->flags & LPROPS_INDEX)); |
| 674 | } |
| 675 | |
| 676 | /* Mark GC'd index LEBs OK to unmap after this commit finishes */ |
| 677 | list_for_each_entry(idx_gc, &c->idx_gc, list) |
| 678 | idx_gc->unmap = 1; |
| 679 | |
| 680 | /* Record index freeable LEBs for unmapping after commit */ |
| 681 | while (1) { |
| 682 | lp = ubifs_fast_find_frdi_idx(c); |
| 683 | if (unlikely(IS_ERR(lp))) { |
| 684 | err = PTR_ERR(lp); |
| 685 | goto out; |
| 686 | } |
| 687 | if (!lp) |
| 688 | break; |
| 689 | idx_gc = kmalloc(sizeof(struct ubifs_gced_idx_leb), GFP_NOFS); |
| 690 | if (!idx_gc) { |
| 691 | err = -ENOMEM; |
| 692 | goto out; |
| 693 | } |
| 694 | ubifs_assert(!(lp->flags & LPROPS_TAKEN)); |
| 695 | ubifs_assert(lp->flags & LPROPS_INDEX); |
| 696 | /* Don't release the LEB until after the next commit */ |
| 697 | flags = (lp->flags | LPROPS_TAKEN) ^ LPROPS_INDEX; |
| 698 | lp = ubifs_change_lp(c, lp, c->leb_size, 0, flags, 1); |
| 699 | if (unlikely(IS_ERR(lp))) { |
| 700 | err = PTR_ERR(lp); |
| 701 | kfree(idx_gc); |
| 702 | goto out; |
| 703 | } |
| 704 | ubifs_assert(lp->flags & LPROPS_TAKEN); |
| 705 | ubifs_assert(!(lp->flags & LPROPS_INDEX)); |
| 706 | idx_gc->lnum = lp->lnum; |
| 707 | idx_gc->unmap = 1; |
| 708 | list_add(&idx_gc->list, &c->idx_gc); |
| 709 | } |
| 710 | out: |
| 711 | ubifs_release_lprops(c); |
| 712 | return err; |
| 713 | } |
| 714 | |
| 715 | /** |
| 716 | * ubifs_gc_end_commit - garbage collection at end of commit. |
| 717 | * @c: UBIFS file-system description object |
| 718 | * |
| 719 | * This function completes out-of-place garbage collection of index LEBs. |
| 720 | */ |
| 721 | int ubifs_gc_end_commit(struct ubifs_info *c) |
| 722 | { |
| 723 | struct ubifs_gced_idx_leb *idx_gc, *tmp; |
| 724 | struct ubifs_wbuf *wbuf; |
| 725 | int err = 0; |
| 726 | |
| 727 | wbuf = &c->jheads[GCHD].wbuf; |
| 728 | mutex_lock_nested(&wbuf->io_mutex, wbuf->jhead); |
| 729 | list_for_each_entry_safe(idx_gc, tmp, &c->idx_gc, list) |
| 730 | if (idx_gc->unmap) { |
| 731 | dbg_gc("LEB %d", idx_gc->lnum); |
| 732 | err = ubifs_leb_unmap(c, idx_gc->lnum); |
| 733 | if (err) |
| 734 | goto out; |
| 735 | err = ubifs_change_one_lp(c, idx_gc->lnum, LPROPS_NC, |
| 736 | LPROPS_NC, 0, LPROPS_TAKEN, -1); |
| 737 | if (err) |
| 738 | goto out; |
| 739 | list_del(&idx_gc->list); |
| 740 | kfree(idx_gc); |
| 741 | } |
| 742 | out: |
| 743 | mutex_unlock(&wbuf->io_mutex); |
| 744 | return err; |
| 745 | } |
| 746 | |
| 747 | /** |
| 748 | * ubifs_destroy_idx_gc - destroy idx_gc list. |
| 749 | * @c: UBIFS file-system description object |
| 750 | * |
| 751 | * This function destroys the idx_gc list. It is called when unmounting or |
| 752 | * remounting read-only so locks are not needed. |
| 753 | */ |
| 754 | void ubifs_destroy_idx_gc(struct ubifs_info *c) |
| 755 | { |
| 756 | while (!list_empty(&c->idx_gc)) { |
| 757 | struct ubifs_gced_idx_leb *idx_gc; |
| 758 | |
| 759 | idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, |
| 760 | list); |
| 761 | c->idx_gc_cnt -= 1; |
| 762 | list_del(&idx_gc->list); |
| 763 | kfree(idx_gc); |
| 764 | } |
| 765 | |
| 766 | } |
| 767 | |
| 768 | /** |
| 769 | * ubifs_get_idx_gc_leb - get a LEB from GC'd index LEB list. |
| 770 | * @c: UBIFS file-system description object |
| 771 | * |
| 772 | * Called during start commit so locks are not needed. |
| 773 | */ |
| 774 | int ubifs_get_idx_gc_leb(struct ubifs_info *c) |
| 775 | { |
| 776 | struct ubifs_gced_idx_leb *idx_gc; |
| 777 | int lnum; |
| 778 | |
| 779 | if (list_empty(&c->idx_gc)) |
| 780 | return -ENOSPC; |
| 781 | idx_gc = list_entry(c->idx_gc.next, struct ubifs_gced_idx_leb, list); |
| 782 | lnum = idx_gc->lnum; |
| 783 | /* c->idx_gc_cnt is updated by the caller when lprops are updated */ |
| 784 | list_del(&idx_gc->list); |
| 785 | kfree(idx_gc); |
| 786 | return lnum; |
| 787 | } |