Zheng Liu | 654598b | 2012-11-08 21:57:20 -0500 | [diff] [blame] | 1 | /* |
| 2 | * fs/ext4/extents_status.c |
| 3 | * |
| 4 | * Written by Yongqiang Yang <xiaoqiangnk@gmail.com> |
| 5 | * Modified by |
| 6 | * Allison Henderson <achender@linux.vnet.ibm.com> |
| 7 | * Hugh Dickins <hughd@google.com> |
| 8 | * Zheng Liu <wenqing.lz@taobao.com> |
| 9 | * |
| 10 | * Ext4 extents status tree core functions. |
| 11 | */ |
| 12 | #include <linux/rbtree.h> |
| 13 | #include "ext4.h" |
| 14 | #include "extents_status.h" |
| 15 | #include "ext4_extents.h" |
| 16 | |
Zheng Liu | 992e9fd | 2012-11-08 21:57:33 -0500 | [diff] [blame^] | 17 | #include <trace/events/ext4.h> |
| 18 | |
Zheng Liu | 654598b | 2012-11-08 21:57:20 -0500 | [diff] [blame] | 19 | /* |
| 20 | * According to previous discussion in Ext4 Developer Workshop, we |
| 21 | * will introduce a new structure called io tree to track all extent |
| 22 | * status in order to solve some problems that we have met |
| 23 | * (e.g. Reservation space warning), and provide extent-level locking. |
| 24 | * Delay extent tree is the first step to achieve this goal. It is |
| 25 | * original built by Yongqiang Yang. At that time it is called delay |
| 26 | * extent tree, whose goal is only track delay extent in memory to |
| 27 | * simplify the implementation of fiemap and bigalloc, and introduce |
| 28 | * lseek SEEK_DATA/SEEK_HOLE support. That is why it is still called |
| 29 | * delay extent tree at the following comment. But for better |
| 30 | * understand what it does, it has been rename to extent status tree. |
| 31 | * |
| 32 | * Currently the first step has been done. All delay extents are |
| 33 | * tracked in the tree. It maintains the delay extent when a delay |
| 34 | * allocation is issued, and the delay extent is written out or |
| 35 | * invalidated. Therefore the implementation of fiemap and bigalloc |
| 36 | * are simplified, and SEEK_DATA/SEEK_HOLE are introduced. |
| 37 | * |
| 38 | * The following comment describes the implemenmtation of extent |
| 39 | * status tree and future works. |
| 40 | */ |
| 41 | |
| 42 | /* |
| 43 | * extents status tree implementation for ext4. |
| 44 | * |
| 45 | * |
| 46 | * ========================================================================== |
| 47 | * Extents status encompass delayed extents and extent locks |
| 48 | * |
| 49 | * 1. Why delayed extent implementation ? |
| 50 | * |
| 51 | * Without delayed extent, ext4 identifies a delayed extent by looking |
| 52 | * up page cache, this has several deficiencies - complicated, buggy, |
| 53 | * and inefficient code. |
| 54 | * |
| 55 | * FIEMAP, SEEK_HOLE/DATA, bigalloc, punch hole and writeout all need |
| 56 | * to know if a block or a range of blocks are belonged to a delayed |
| 57 | * extent. |
| 58 | * |
| 59 | * Let us have a look at how they do without delayed extents implementation. |
| 60 | * -- FIEMAP |
| 61 | * FIEMAP looks up page cache to identify delayed allocations from holes. |
| 62 | * |
| 63 | * -- SEEK_HOLE/DATA |
| 64 | * SEEK_HOLE/DATA has the same problem as FIEMAP. |
| 65 | * |
| 66 | * -- bigalloc |
| 67 | * bigalloc looks up page cache to figure out if a block is |
| 68 | * already under delayed allocation or not to determine whether |
| 69 | * quota reserving is needed for the cluster. |
| 70 | * |
| 71 | * -- punch hole |
| 72 | * punch hole looks up page cache to identify a delayed extent. |
| 73 | * |
| 74 | * -- writeout |
| 75 | * Writeout looks up whole page cache to see if a buffer is |
| 76 | * mapped, If there are not very many delayed buffers, then it is |
| 77 | * time comsuming. |
| 78 | * |
| 79 | * With delayed extents implementation, FIEMAP, SEEK_HOLE/DATA, |
| 80 | * bigalloc and writeout can figure out if a block or a range of |
| 81 | * blocks is under delayed allocation(belonged to a delayed extent) or |
| 82 | * not by searching the delayed extent tree. |
| 83 | * |
| 84 | * |
| 85 | * ========================================================================== |
| 86 | * 2. ext4 delayed extents impelmentation |
| 87 | * |
| 88 | * -- delayed extent |
| 89 | * A delayed extent is a range of blocks which are contiguous |
| 90 | * logically and under delayed allocation. Unlike extent in |
| 91 | * ext4, delayed extent in ext4 is a in-memory struct, there is |
| 92 | * no corresponding on-disk data. There is no limit on length of |
| 93 | * delayed extent, so a delayed extent can contain as many blocks |
| 94 | * as they are contiguous logically. |
| 95 | * |
| 96 | * -- delayed extent tree |
| 97 | * Every inode has a delayed extent tree and all under delayed |
| 98 | * allocation blocks are added to the tree as delayed extents. |
| 99 | * Delayed extents in the tree are ordered by logical block no. |
| 100 | * |
| 101 | * -- operations on a delayed extent tree |
| 102 | * There are three operations on a delayed extent tree: find next |
| 103 | * delayed extent, adding a space(a range of blocks) and removing |
| 104 | * a space. |
| 105 | * |
| 106 | * -- race on a delayed extent tree |
| 107 | * Delayed extent tree is protected inode->i_es_lock. |
| 108 | * |
| 109 | * |
| 110 | * ========================================================================== |
| 111 | * 3. performance analysis |
| 112 | * -- overhead |
| 113 | * 1. There is a cache extent for write access, so if writes are |
| 114 | * not very random, adding space operaions are in O(1) time. |
| 115 | * |
| 116 | * -- gain |
| 117 | * 2. Code is much simpler, more readable, more maintainable and |
| 118 | * more efficient. |
| 119 | * |
| 120 | * |
| 121 | * ========================================================================== |
| 122 | * 4. TODO list |
| 123 | * -- Track all extent status |
| 124 | * |
| 125 | * -- Improve get block process |
| 126 | * |
| 127 | * -- Extent-level locking |
| 128 | */ |
| 129 | |
| 130 | static struct kmem_cache *ext4_es_cachep; |
| 131 | |
| 132 | int __init ext4_init_es(void) |
| 133 | { |
| 134 | ext4_es_cachep = KMEM_CACHE(extent_status, SLAB_RECLAIM_ACCOUNT); |
| 135 | if (ext4_es_cachep == NULL) |
| 136 | return -ENOMEM; |
| 137 | return 0; |
| 138 | } |
| 139 | |
| 140 | void ext4_exit_es(void) |
| 141 | { |
| 142 | if (ext4_es_cachep) |
| 143 | kmem_cache_destroy(ext4_es_cachep); |
| 144 | } |
| 145 | |
| 146 | void ext4_es_init_tree(struct ext4_es_tree *tree) |
| 147 | { |
| 148 | tree->root = RB_ROOT; |
| 149 | tree->cache_es = NULL; |
| 150 | } |
| 151 | |
| 152 | #ifdef ES_DEBUG__ |
| 153 | static void ext4_es_print_tree(struct inode *inode) |
| 154 | { |
| 155 | struct ext4_es_tree *tree; |
| 156 | struct rb_node *node; |
| 157 | |
| 158 | printk(KERN_DEBUG "status extents for inode %lu:", inode->i_ino); |
| 159 | tree = &EXT4_I(inode)->i_es_tree; |
| 160 | node = rb_first(&tree->root); |
| 161 | while (node) { |
| 162 | struct extent_status *es; |
| 163 | es = rb_entry(node, struct extent_status, rb_node); |
| 164 | printk(KERN_DEBUG " [%u/%u)", es->start, es->len); |
| 165 | node = rb_next(node); |
| 166 | } |
| 167 | printk(KERN_DEBUG "\n"); |
| 168 | } |
| 169 | #else |
| 170 | #define ext4_es_print_tree(inode) |
| 171 | #endif |
| 172 | |
| 173 | static inline ext4_lblk_t extent_status_end(struct extent_status *es) |
| 174 | { |
| 175 | BUG_ON(es->start + es->len < es->start); |
| 176 | return es->start + es->len - 1; |
| 177 | } |
| 178 | |
| 179 | /* |
| 180 | * search through the tree for an delayed extent with a given offset. If |
| 181 | * it can't be found, try to find next extent. |
| 182 | */ |
| 183 | static struct extent_status *__es_tree_search(struct rb_root *root, |
| 184 | ext4_lblk_t offset) |
| 185 | { |
| 186 | struct rb_node *node = root->rb_node; |
| 187 | struct extent_status *es = NULL; |
| 188 | |
| 189 | while (node) { |
| 190 | es = rb_entry(node, struct extent_status, rb_node); |
| 191 | if (offset < es->start) |
| 192 | node = node->rb_left; |
| 193 | else if (offset > extent_status_end(es)) |
| 194 | node = node->rb_right; |
| 195 | else |
| 196 | return es; |
| 197 | } |
| 198 | |
| 199 | if (es && offset < es->start) |
| 200 | return es; |
| 201 | |
| 202 | if (es && offset > extent_status_end(es)) { |
| 203 | node = rb_next(&es->rb_node); |
| 204 | return node ? rb_entry(node, struct extent_status, rb_node) : |
| 205 | NULL; |
| 206 | } |
| 207 | |
| 208 | return NULL; |
| 209 | } |
| 210 | |
| 211 | /* |
| 212 | * ext4_es_find_extent: find the 1st delayed extent covering @es->start |
| 213 | * if it exists, otherwise, the next extent after @es->start. |
| 214 | * |
| 215 | * @inode: the inode which owns delayed extents |
| 216 | * @es: delayed extent that we found |
| 217 | * |
| 218 | * Returns the first block of the next extent after es, otherwise |
| 219 | * EXT_MAX_BLOCKS if no delay extent is found. |
| 220 | * Delayed extent is returned via @es. |
| 221 | */ |
| 222 | ext4_lblk_t ext4_es_find_extent(struct inode *inode, struct extent_status *es) |
| 223 | { |
| 224 | struct ext4_es_tree *tree = NULL; |
| 225 | struct extent_status *es1 = NULL; |
| 226 | struct rb_node *node; |
| 227 | ext4_lblk_t ret = EXT_MAX_BLOCKS; |
| 228 | |
Zheng Liu | 992e9fd | 2012-11-08 21:57:33 -0500 | [diff] [blame^] | 229 | trace_ext4_es_find_extent_enter(inode, es->start); |
| 230 | |
Zheng Liu | 654598b | 2012-11-08 21:57:20 -0500 | [diff] [blame] | 231 | read_lock(&EXT4_I(inode)->i_es_lock); |
| 232 | tree = &EXT4_I(inode)->i_es_tree; |
| 233 | |
| 234 | /* find delay extent in cache firstly */ |
| 235 | if (tree->cache_es) { |
| 236 | es1 = tree->cache_es; |
| 237 | if (in_range(es->start, es1->start, es1->len)) { |
| 238 | es_debug("%u cached by [%u/%u)\n", |
| 239 | es->start, es1->start, es1->len); |
| 240 | goto out; |
| 241 | } |
| 242 | } |
| 243 | |
| 244 | es->len = 0; |
| 245 | es1 = __es_tree_search(&tree->root, es->start); |
| 246 | |
| 247 | out: |
| 248 | if (es1) { |
| 249 | tree->cache_es = es1; |
| 250 | es->start = es1->start; |
| 251 | es->len = es1->len; |
| 252 | node = rb_next(&es1->rb_node); |
| 253 | if (node) { |
| 254 | es1 = rb_entry(node, struct extent_status, rb_node); |
| 255 | ret = es1->start; |
| 256 | } |
| 257 | } |
| 258 | |
| 259 | read_unlock(&EXT4_I(inode)->i_es_lock); |
Zheng Liu | 992e9fd | 2012-11-08 21:57:33 -0500 | [diff] [blame^] | 260 | |
| 261 | trace_ext4_es_find_extent_exit(inode, es, ret); |
Zheng Liu | 654598b | 2012-11-08 21:57:20 -0500 | [diff] [blame] | 262 | return ret; |
| 263 | } |
| 264 | |
| 265 | static struct extent_status * |
| 266 | ext4_es_alloc_extent(ext4_lblk_t start, ext4_lblk_t len) |
| 267 | { |
| 268 | struct extent_status *es; |
| 269 | es = kmem_cache_alloc(ext4_es_cachep, GFP_ATOMIC); |
| 270 | if (es == NULL) |
| 271 | return NULL; |
| 272 | es->start = start; |
| 273 | es->len = len; |
| 274 | return es; |
| 275 | } |
| 276 | |
| 277 | static void ext4_es_free_extent(struct extent_status *es) |
| 278 | { |
| 279 | kmem_cache_free(ext4_es_cachep, es); |
| 280 | } |
| 281 | |
| 282 | static struct extent_status * |
| 283 | ext4_es_try_to_merge_left(struct ext4_es_tree *tree, struct extent_status *es) |
| 284 | { |
| 285 | struct extent_status *es1; |
| 286 | struct rb_node *node; |
| 287 | |
| 288 | node = rb_prev(&es->rb_node); |
| 289 | if (!node) |
| 290 | return es; |
| 291 | |
| 292 | es1 = rb_entry(node, struct extent_status, rb_node); |
| 293 | if (es->start == extent_status_end(es1) + 1) { |
| 294 | es1->len += es->len; |
| 295 | rb_erase(&es->rb_node, &tree->root); |
| 296 | ext4_es_free_extent(es); |
| 297 | es = es1; |
| 298 | } |
| 299 | |
| 300 | return es; |
| 301 | } |
| 302 | |
| 303 | static struct extent_status * |
| 304 | ext4_es_try_to_merge_right(struct ext4_es_tree *tree, struct extent_status *es) |
| 305 | { |
| 306 | struct extent_status *es1; |
| 307 | struct rb_node *node; |
| 308 | |
| 309 | node = rb_next(&es->rb_node); |
| 310 | if (!node) |
| 311 | return es; |
| 312 | |
| 313 | es1 = rb_entry(node, struct extent_status, rb_node); |
| 314 | if (es1->start == extent_status_end(es) + 1) { |
| 315 | es->len += es1->len; |
| 316 | rb_erase(node, &tree->root); |
| 317 | ext4_es_free_extent(es1); |
| 318 | } |
| 319 | |
| 320 | return es; |
| 321 | } |
| 322 | |
| 323 | static int __es_insert_extent(struct ext4_es_tree *tree, ext4_lblk_t offset, |
| 324 | ext4_lblk_t len) |
| 325 | { |
| 326 | struct rb_node **p = &tree->root.rb_node; |
| 327 | struct rb_node *parent = NULL; |
| 328 | struct extent_status *es; |
| 329 | ext4_lblk_t end = offset + len - 1; |
| 330 | |
| 331 | BUG_ON(end < offset); |
| 332 | es = tree->cache_es; |
| 333 | if (es && offset == (extent_status_end(es) + 1)) { |
| 334 | es_debug("cached by [%u/%u)\n", es->start, es->len); |
| 335 | es->len += len; |
| 336 | es = ext4_es_try_to_merge_right(tree, es); |
| 337 | goto out; |
| 338 | } else if (es && es->start == end + 1) { |
| 339 | es_debug("cached by [%u/%u)\n", es->start, es->len); |
| 340 | es->start = offset; |
| 341 | es->len += len; |
| 342 | es = ext4_es_try_to_merge_left(tree, es); |
| 343 | goto out; |
| 344 | } else if (es && es->start <= offset && |
| 345 | end <= extent_status_end(es)) { |
| 346 | es_debug("cached by [%u/%u)\n", es->start, es->len); |
| 347 | goto out; |
| 348 | } |
| 349 | |
| 350 | while (*p) { |
| 351 | parent = *p; |
| 352 | es = rb_entry(parent, struct extent_status, rb_node); |
| 353 | |
| 354 | if (offset < es->start) { |
| 355 | if (es->start == end + 1) { |
| 356 | es->start = offset; |
| 357 | es->len += len; |
| 358 | es = ext4_es_try_to_merge_left(tree, es); |
| 359 | goto out; |
| 360 | } |
| 361 | p = &(*p)->rb_left; |
| 362 | } else if (offset > extent_status_end(es)) { |
| 363 | if (offset == extent_status_end(es) + 1) { |
| 364 | es->len += len; |
| 365 | es = ext4_es_try_to_merge_right(tree, es); |
| 366 | goto out; |
| 367 | } |
| 368 | p = &(*p)->rb_right; |
| 369 | } else { |
| 370 | if (extent_status_end(es) <= end) |
| 371 | es->len = offset - es->start + len; |
| 372 | goto out; |
| 373 | } |
| 374 | } |
| 375 | |
| 376 | es = ext4_es_alloc_extent(offset, len); |
| 377 | if (!es) |
| 378 | return -ENOMEM; |
| 379 | rb_link_node(&es->rb_node, parent, p); |
| 380 | rb_insert_color(&es->rb_node, &tree->root); |
| 381 | |
| 382 | out: |
| 383 | tree->cache_es = es; |
| 384 | return 0; |
| 385 | } |
| 386 | |
| 387 | /* |
| 388 | * ext4_es_insert_extent() adds a space to a delayed extent tree. |
| 389 | * Caller holds inode->i_es_lock. |
| 390 | * |
| 391 | * ext4_es_insert_extent is called by ext4_da_write_begin and |
| 392 | * ext4_es_remove_extent. |
| 393 | * |
| 394 | * Return 0 on success, error code on failure. |
| 395 | */ |
| 396 | int ext4_es_insert_extent(struct inode *inode, ext4_lblk_t offset, |
| 397 | ext4_lblk_t len) |
| 398 | { |
| 399 | struct ext4_es_tree *tree; |
| 400 | int err = 0; |
| 401 | |
Zheng Liu | 992e9fd | 2012-11-08 21:57:33 -0500 | [diff] [blame^] | 402 | trace_ext4_es_insert_extent(inode, offset, len); |
Zheng Liu | 654598b | 2012-11-08 21:57:20 -0500 | [diff] [blame] | 403 | es_debug("add [%u/%u) to extent status tree of inode %lu\n", |
| 404 | offset, len, inode->i_ino); |
| 405 | |
| 406 | write_lock(&EXT4_I(inode)->i_es_lock); |
| 407 | tree = &EXT4_I(inode)->i_es_tree; |
| 408 | err = __es_insert_extent(tree, offset, len); |
| 409 | write_unlock(&EXT4_I(inode)->i_es_lock); |
| 410 | |
| 411 | ext4_es_print_tree(inode); |
| 412 | |
| 413 | return err; |
| 414 | } |
| 415 | |
| 416 | /* |
| 417 | * ext4_es_remove_extent() removes a space from a delayed extent tree. |
| 418 | * Caller holds inode->i_es_lock. |
| 419 | * |
| 420 | * Return 0 on success, error code on failure. |
| 421 | */ |
| 422 | int ext4_es_remove_extent(struct inode *inode, ext4_lblk_t offset, |
| 423 | ext4_lblk_t len) |
| 424 | { |
| 425 | struct rb_node *node; |
| 426 | struct ext4_es_tree *tree; |
| 427 | struct extent_status *es; |
| 428 | struct extent_status orig_es; |
| 429 | ext4_lblk_t len1, len2, end; |
| 430 | int err = 0; |
| 431 | |
Zheng Liu | 992e9fd | 2012-11-08 21:57:33 -0500 | [diff] [blame^] | 432 | trace_ext4_es_remove_extent(inode, offset, len); |
Zheng Liu | 654598b | 2012-11-08 21:57:20 -0500 | [diff] [blame] | 433 | es_debug("remove [%u/%u) from extent status tree of inode %lu\n", |
| 434 | offset, len, inode->i_ino); |
| 435 | |
| 436 | end = offset + len - 1; |
| 437 | BUG_ON(end < offset); |
| 438 | write_lock(&EXT4_I(inode)->i_es_lock); |
| 439 | tree = &EXT4_I(inode)->i_es_tree; |
| 440 | es = __es_tree_search(&tree->root, offset); |
| 441 | if (!es) |
| 442 | goto out; |
| 443 | if (es->start > end) |
| 444 | goto out; |
| 445 | |
| 446 | /* Simply invalidate cache_es. */ |
| 447 | tree->cache_es = NULL; |
| 448 | |
| 449 | orig_es.start = es->start; |
| 450 | orig_es.len = es->len; |
| 451 | len1 = offset > es->start ? offset - es->start : 0; |
| 452 | len2 = extent_status_end(es) > end ? |
| 453 | extent_status_end(es) - end : 0; |
| 454 | if (len1 > 0) |
| 455 | es->len = len1; |
| 456 | if (len2 > 0) { |
| 457 | if (len1 > 0) { |
| 458 | err = __es_insert_extent(tree, end + 1, len2); |
| 459 | if (err) { |
| 460 | es->start = orig_es.start; |
| 461 | es->len = orig_es.len; |
| 462 | goto out; |
| 463 | } |
| 464 | } else { |
| 465 | es->start = end + 1; |
| 466 | es->len = len2; |
| 467 | } |
| 468 | goto out; |
| 469 | } |
| 470 | |
| 471 | if (len1 > 0) { |
| 472 | node = rb_next(&es->rb_node); |
| 473 | if (node) |
| 474 | es = rb_entry(node, struct extent_status, rb_node); |
| 475 | else |
| 476 | es = NULL; |
| 477 | } |
| 478 | |
| 479 | while (es && extent_status_end(es) <= end) { |
| 480 | node = rb_next(&es->rb_node); |
| 481 | rb_erase(&es->rb_node, &tree->root); |
| 482 | ext4_es_free_extent(es); |
| 483 | if (!node) { |
| 484 | es = NULL; |
| 485 | break; |
| 486 | } |
| 487 | es = rb_entry(node, struct extent_status, rb_node); |
| 488 | } |
| 489 | |
| 490 | if (es && es->start < end + 1) { |
| 491 | len1 = extent_status_end(es) - end; |
| 492 | es->start = end + 1; |
| 493 | es->len = len1; |
| 494 | } |
| 495 | |
| 496 | out: |
| 497 | write_unlock(&EXT4_I(inode)->i_es_lock); |
| 498 | ext4_es_print_tree(inode); |
| 499 | return err; |
| 500 | } |