| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * Copyright (C) 2011 Red Hat, Inc. |
| 4 | * |
| 5 | * This file is released under the GPL. |
| 6 | */ |
| 7 | #include "dm-transaction-manager.h" |
| 8 | #include "dm-space-map.h" |
| 9 | #include "dm-space-map-disk.h" |
| 10 | #include "dm-space-map-metadata.h" |
| 11 | #include "dm-persistent-data-internal.h" |
| 12 | |
| 13 | #include <linux/export.h> |
| 14 | #include <linux/mutex.h> |
| 15 | #include <linux/hash.h> |
| 16 | #include <linux/rbtree.h> |
| 17 | #include <linux/slab.h> |
| 18 | #include <linux/device-mapper.h> |
| 19 | |
| 20 | #define DM_MSG_PREFIX "transaction manager" |
| 21 | |
| 22 | /*----------------------------------------------------------------*/ |
| 23 | |
| 24 | #define PREFETCH_SIZE 128 |
| 25 | #define PREFETCH_BITS 7 |
| 26 | #define PREFETCH_SENTINEL ((dm_block_t) -1ULL) |
| 27 | |
| 28 | struct prefetch_set { |
| 29 | struct mutex lock; |
| 30 | dm_block_t blocks[PREFETCH_SIZE]; |
| 31 | }; |
| 32 | |
| 33 | static unsigned int prefetch_hash(dm_block_t b) |
| 34 | { |
| 35 | return hash_64(val: b, PREFETCH_BITS); |
| 36 | } |
| 37 | |
| 38 | static void prefetch_wipe(struct prefetch_set *p) |
| 39 | { |
| 40 | unsigned int i; |
| 41 | |
| 42 | for (i = 0; i < PREFETCH_SIZE; i++) |
| 43 | p->blocks[i] = PREFETCH_SENTINEL; |
| 44 | } |
| 45 | |
| 46 | static void prefetch_init(struct prefetch_set *p) |
| 47 | { |
| 48 | mutex_init(&p->lock); |
| 49 | prefetch_wipe(p); |
| 50 | } |
| 51 | |
| 52 | static void prefetch_add(struct prefetch_set *p, dm_block_t b) |
| 53 | { |
| 54 | unsigned int h = prefetch_hash(b); |
| 55 | |
| 56 | mutex_lock(&p->lock); |
| 57 | if (p->blocks[h] == PREFETCH_SENTINEL) |
| 58 | p->blocks[h] = b; |
| 59 | |
| 60 | mutex_unlock(lock: &p->lock); |
| 61 | } |
| 62 | |
| 63 | static void prefetch_issue(struct prefetch_set *p, struct dm_block_manager *bm) |
| 64 | { |
| 65 | unsigned int i; |
| 66 | |
| 67 | mutex_lock(&p->lock); |
| 68 | |
| 69 | for (i = 0; i < PREFETCH_SIZE; i++) |
| 70 | if (p->blocks[i] != PREFETCH_SENTINEL) { |
| 71 | dm_bm_prefetch(bm, b: p->blocks[i]); |
| 72 | p->blocks[i] = PREFETCH_SENTINEL; |
| 73 | } |
| 74 | |
| 75 | mutex_unlock(lock: &p->lock); |
| 76 | } |
| 77 | |
| 78 | /*----------------------------------------------------------------*/ |
| 79 | |
| 80 | struct shadow_info { |
| 81 | struct rb_node node; |
| 82 | dm_block_t where; |
| 83 | }; |
| 84 | |
| 85 | /* |
| 86 | * It would be nice if we scaled with the size of transaction. |
| 87 | */ |
| 88 | #define DM_HASH_SIZE 256 |
| 89 | #define DM_HASH_MASK (DM_HASH_SIZE - 1) |
| 90 | |
| 91 | struct dm_transaction_manager { |
| 92 | int is_clone; |
| 93 | struct dm_transaction_manager *real; |
| 94 | |
| 95 | struct dm_block_manager *bm; |
| 96 | struct dm_space_map *sm; |
| 97 | |
| 98 | spinlock_t lock; |
| 99 | struct rb_root buckets[DM_HASH_SIZE]; |
| 100 | |
| 101 | struct prefetch_set prefetches; |
| 102 | }; |
| 103 | |
| 104 | /*----------------------------------------------------------------*/ |
| 105 | |
| 106 | static int is_shadow(struct dm_transaction_manager *tm, dm_block_t b) |
| 107 | { |
| 108 | int r = 0; |
| 109 | unsigned int bucket = dm_hash_block(b, DM_HASH_MASK); |
| 110 | struct rb_node **node; |
| 111 | |
| 112 | spin_lock(lock: &tm->lock); |
| 113 | node = &tm->buckets[bucket].rb_node; |
| 114 | while (*node) { |
| 115 | struct shadow_info *si = |
| 116 | rb_entry(*node, struct shadow_info, node); |
| 117 | if (b == si->where) { |
| 118 | r = 1; |
| 119 | break; |
| 120 | } |
| 121 | if (b < si->where) |
| 122 | node = &si->node.rb_left; |
| 123 | else |
| 124 | node = &si->node.rb_right; |
| 125 | } |
| 126 | spin_unlock(lock: &tm->lock); |
| 127 | |
| 128 | return r; |
| 129 | } |
| 130 | |
| 131 | /* |
| 132 | * This can silently fail if there's no memory. We're ok with this since |
| 133 | * creating redundant shadows causes no harm. |
| 134 | */ |
| 135 | static void insert_shadow(struct dm_transaction_manager *tm, dm_block_t b) |
| 136 | { |
| 137 | unsigned int bucket; |
| 138 | struct shadow_info *si; |
| 139 | |
| 140 | si = kmalloc(sizeof(*si), GFP_NOIO); |
| 141 | if (si) { |
| 142 | struct rb_node **node, *parent; |
| 143 | si->where = b; |
| 144 | bucket = dm_hash_block(b, DM_HASH_MASK); |
| 145 | |
| 146 | spin_lock(lock: &tm->lock); |
| 147 | node = &tm->buckets[bucket].rb_node; |
| 148 | parent = NULL; |
| 149 | while (*node) { |
| 150 | struct shadow_info *si = |
| 151 | rb_entry(*node, struct shadow_info, node); |
| 152 | parent = *node; |
| 153 | if (b < si->where) |
| 154 | node = &si->node.rb_left; |
| 155 | else |
| 156 | node = &si->node.rb_right; |
| 157 | } |
| 158 | rb_link_node(node: &si->node, parent, rb_link: node); |
| 159 | rb_insert_color(&si->node, &tm->buckets[bucket]); |
| 160 | spin_unlock(lock: &tm->lock); |
| 161 | } |
| 162 | } |
| 163 | |
| 164 | static void wipe_shadow_table(struct dm_transaction_manager *tm) |
| 165 | { |
| 166 | unsigned int i; |
| 167 | |
| 168 | spin_lock(lock: &tm->lock); |
| 169 | for (i = 0; i < DM_HASH_SIZE; i++) { |
| 170 | while (!RB_EMPTY_ROOT(&tm->buckets[i])) { |
| 171 | struct shadow_info *si = |
| 172 | rb_entry(tm->buckets[i].rb_node, struct shadow_info, node); |
| 173 | rb_erase(&si->node, &tm->buckets[i]); |
| 174 | kfree(objp: si); |
| 175 | } |
| 176 | } |
| 177 | spin_unlock(lock: &tm->lock); |
| 178 | } |
| 179 | |
| 180 | /*----------------------------------------------------------------*/ |
| 181 | |
| 182 | static struct dm_transaction_manager *dm_tm_create(struct dm_block_manager *bm, |
| 183 | struct dm_space_map *sm) |
| 184 | { |
| 185 | unsigned int i; |
| 186 | struct dm_transaction_manager *tm; |
| 187 | |
| 188 | tm = kmalloc(sizeof(*tm), GFP_KERNEL); |
| 189 | if (!tm) |
| 190 | return ERR_PTR(error: -ENOMEM); |
| 191 | |
| 192 | tm->is_clone = 0; |
| 193 | tm->real = NULL; |
| 194 | tm->bm = bm; |
| 195 | tm->sm = sm; |
| 196 | |
| 197 | spin_lock_init(&tm->lock); |
| 198 | for (i = 0; i < DM_HASH_SIZE; i++) |
| 199 | tm->buckets[i] = RB_ROOT; |
| 200 | |
| 201 | prefetch_init(p: &tm->prefetches); |
| 202 | |
| 203 | return tm; |
| 204 | } |
| 205 | |
| 206 | struct dm_transaction_manager *dm_tm_create_non_blocking_clone(struct dm_transaction_manager *real) |
| 207 | { |
| 208 | struct dm_transaction_manager *tm; |
| 209 | |
| 210 | tm = kmalloc(sizeof(*tm), GFP_KERNEL); |
| 211 | if (tm) { |
| 212 | tm->is_clone = 1; |
| 213 | tm->real = real; |
| 214 | } |
| 215 | |
| 216 | return tm; |
| 217 | } |
| 218 | EXPORT_SYMBOL_GPL(dm_tm_create_non_blocking_clone); |
| 219 | |
| 220 | void dm_tm_destroy(struct dm_transaction_manager *tm) |
| 221 | { |
| 222 | if (!tm) |
| 223 | return; |
| 224 | |
| 225 | if (!tm->is_clone) |
| 226 | wipe_shadow_table(tm); |
| 227 | |
| 228 | kfree(objp: tm); |
| 229 | } |
| 230 | EXPORT_SYMBOL_GPL(dm_tm_destroy); |
| 231 | |
| 232 | int dm_tm_pre_commit(struct dm_transaction_manager *tm) |
| 233 | { |
| 234 | int r; |
| 235 | |
| 236 | if (tm->is_clone) |
| 237 | return -EWOULDBLOCK; |
| 238 | |
| 239 | r = dm_sm_commit(sm: tm->sm); |
| 240 | if (r < 0) |
| 241 | return r; |
| 242 | |
| 243 | return dm_bm_flush(bm: tm->bm); |
| 244 | } |
| 245 | EXPORT_SYMBOL_GPL(dm_tm_pre_commit); |
| 246 | |
| 247 | int dm_tm_commit(struct dm_transaction_manager *tm, struct dm_block *root) |
| 248 | { |
| 249 | if (tm->is_clone) |
| 250 | return -EWOULDBLOCK; |
| 251 | |
| 252 | wipe_shadow_table(tm); |
| 253 | dm_bm_unlock(b: root); |
| 254 | |
| 255 | return dm_bm_flush(bm: tm->bm); |
| 256 | } |
| 257 | EXPORT_SYMBOL_GPL(dm_tm_commit); |
| 258 | |
| 259 | int dm_tm_new_block(struct dm_transaction_manager *tm, |
| 260 | const struct dm_block_validator *v, |
| 261 | struct dm_block **result) |
| 262 | { |
| 263 | int r; |
| 264 | dm_block_t new_block; |
| 265 | |
| 266 | if (tm->is_clone) |
| 267 | return -EWOULDBLOCK; |
| 268 | |
| 269 | r = dm_sm_new_block(sm: tm->sm, b: &new_block); |
| 270 | if (r < 0) |
| 271 | return r; |
| 272 | |
| 273 | r = dm_bm_write_lock_zero(bm: tm->bm, b: new_block, v, result); |
| 274 | if (r < 0) { |
| 275 | dm_sm_dec_block(sm: tm->sm, b: new_block); |
| 276 | return r; |
| 277 | } |
| 278 | |
| 279 | /* |
| 280 | * New blocks count as shadows in that they don't need to be |
| 281 | * shadowed again. |
| 282 | */ |
| 283 | insert_shadow(tm, b: new_block); |
| 284 | |
| 285 | return 0; |
| 286 | } |
| 287 | |
| 288 | static int __shadow_block(struct dm_transaction_manager *tm, dm_block_t orig, |
| 289 | const struct dm_block_validator *v, |
| 290 | struct dm_block **result) |
| 291 | { |
| 292 | int r; |
| 293 | dm_block_t new; |
| 294 | struct dm_block *orig_block; |
| 295 | |
| 296 | r = dm_sm_new_block(sm: tm->sm, b: &new); |
| 297 | if (r < 0) |
| 298 | return r; |
| 299 | |
| 300 | r = dm_sm_dec_block(sm: tm->sm, b: orig); |
| 301 | if (r < 0) |
| 302 | return r; |
| 303 | |
| 304 | r = dm_bm_read_lock(bm: tm->bm, b: orig, v, result: &orig_block); |
| 305 | if (r < 0) |
| 306 | return r; |
| 307 | |
| 308 | /* |
| 309 | * It would be tempting to use dm_bm_unlock_move here, but some |
| 310 | * code, such as the space maps, keeps using the old data structures |
| 311 | * secure in the knowledge they won't be changed until the next |
| 312 | * transaction. Using unlock_move would force a synchronous read |
| 313 | * since the old block would no longer be in the cache. |
| 314 | */ |
| 315 | r = dm_bm_write_lock_zero(bm: tm->bm, b: new, v, result); |
| 316 | if (r) { |
| 317 | dm_bm_unlock(b: orig_block); |
| 318 | return r; |
| 319 | } |
| 320 | |
| 321 | memcpy(dm_block_data(*result), dm_block_data(orig_block), |
| 322 | dm_bm_block_size(tm->bm)); |
| 323 | |
| 324 | dm_bm_unlock(b: orig_block); |
| 325 | return r; |
| 326 | } |
| 327 | |
| 328 | int dm_tm_shadow_block(struct dm_transaction_manager *tm, dm_block_t orig, |
| 329 | const struct dm_block_validator *v, struct dm_block **result, |
| 330 | int *inc_children) |
| 331 | { |
| 332 | int r; |
| 333 | |
| 334 | if (tm->is_clone) |
| 335 | return -EWOULDBLOCK; |
| 336 | |
| 337 | r = dm_sm_count_is_more_than_one(sm: tm->sm, b: orig, result: inc_children); |
| 338 | if (r < 0) |
| 339 | return r; |
| 340 | |
| 341 | if (is_shadow(tm, b: orig) && !*inc_children) |
| 342 | return dm_bm_write_lock(bm: tm->bm, b: orig, v, result); |
| 343 | |
| 344 | r = __shadow_block(tm, orig, v, result); |
| 345 | if (r < 0) |
| 346 | return r; |
| 347 | insert_shadow(tm, b: dm_block_location(b: *result)); |
| 348 | |
| 349 | return r; |
| 350 | } |
| 351 | EXPORT_SYMBOL_GPL(dm_tm_shadow_block); |
| 352 | |
| 353 | int dm_tm_read_lock(struct dm_transaction_manager *tm, dm_block_t b, |
| 354 | const struct dm_block_validator *v, |
| 355 | struct dm_block **blk) |
| 356 | { |
| 357 | if (tm->is_clone) { |
| 358 | int r = dm_bm_read_try_lock(bm: tm->real->bm, b, v, result: blk); |
| 359 | |
| 360 | if (r == -EWOULDBLOCK) |
| 361 | prefetch_add(p: &tm->real->prefetches, b); |
| 362 | |
| 363 | return r; |
| 364 | } |
| 365 | |
| 366 | return dm_bm_read_lock(bm: tm->bm, b, v, result: blk); |
| 367 | } |
| 368 | EXPORT_SYMBOL_GPL(dm_tm_read_lock); |
| 369 | |
| 370 | void dm_tm_unlock(struct dm_transaction_manager *tm, struct dm_block *b) |
| 371 | { |
| 372 | dm_bm_unlock(b); |
| 373 | } |
| 374 | EXPORT_SYMBOL_GPL(dm_tm_unlock); |
| 375 | |
| 376 | void dm_tm_inc(struct dm_transaction_manager *tm, dm_block_t b) |
| 377 | { |
| 378 | /* |
| 379 | * The non-blocking clone doesn't support this. |
| 380 | */ |
| 381 | BUG_ON(tm->is_clone); |
| 382 | |
| 383 | dm_sm_inc_block(sm: tm->sm, b); |
| 384 | } |
| 385 | EXPORT_SYMBOL_GPL(dm_tm_inc); |
| 386 | |
| 387 | void dm_tm_inc_range(struct dm_transaction_manager *tm, dm_block_t b, dm_block_t e) |
| 388 | { |
| 389 | /* |
| 390 | * The non-blocking clone doesn't support this. |
| 391 | */ |
| 392 | BUG_ON(tm->is_clone); |
| 393 | |
| 394 | dm_sm_inc_blocks(sm: tm->sm, b, e); |
| 395 | } |
| 396 | EXPORT_SYMBOL_GPL(dm_tm_inc_range); |
| 397 | |
| 398 | void dm_tm_dec(struct dm_transaction_manager *tm, dm_block_t b) |
| 399 | { |
| 400 | /* |
| 401 | * The non-blocking clone doesn't support this. |
| 402 | */ |
| 403 | BUG_ON(tm->is_clone); |
| 404 | |
| 405 | dm_sm_dec_block(sm: tm->sm, b); |
| 406 | } |
| 407 | EXPORT_SYMBOL_GPL(dm_tm_dec); |
| 408 | |
| 409 | void dm_tm_dec_range(struct dm_transaction_manager *tm, dm_block_t b, dm_block_t e) |
| 410 | { |
| 411 | /* |
| 412 | * The non-blocking clone doesn't support this. |
| 413 | */ |
| 414 | BUG_ON(tm->is_clone); |
| 415 | |
| 416 | dm_sm_dec_blocks(sm: tm->sm, b, e); |
| 417 | } |
| 418 | EXPORT_SYMBOL_GPL(dm_tm_dec_range); |
| 419 | |
| 420 | void dm_tm_with_runs(struct dm_transaction_manager *tm, |
| 421 | const __le64 *value_le, unsigned int count, dm_tm_run_fn fn) |
| 422 | { |
| 423 | uint64_t b, begin, end; |
| 424 | bool in_run = false; |
| 425 | unsigned int i; |
| 426 | |
| 427 | for (i = 0; i < count; i++, value_le++) { |
| 428 | b = le64_to_cpu(*value_le); |
| 429 | |
| 430 | if (in_run) { |
| 431 | if (b == end) |
| 432 | end++; |
| 433 | else { |
| 434 | fn(tm, begin, end); |
| 435 | begin = b; |
| 436 | end = b + 1; |
| 437 | } |
| 438 | } else { |
| 439 | in_run = true; |
| 440 | begin = b; |
| 441 | end = b + 1; |
| 442 | } |
| 443 | } |
| 444 | |
| 445 | if (in_run) |
| 446 | fn(tm, begin, end); |
| 447 | } |
| 448 | EXPORT_SYMBOL_GPL(dm_tm_with_runs); |
| 449 | |
| 450 | int dm_tm_ref(struct dm_transaction_manager *tm, dm_block_t b, |
| 451 | uint32_t *result) |
| 452 | { |
| 453 | if (tm->is_clone) |
| 454 | return -EWOULDBLOCK; |
| 455 | |
| 456 | return dm_sm_get_count(sm: tm->sm, b, result); |
| 457 | } |
| 458 | |
| 459 | int dm_tm_block_is_shared(struct dm_transaction_manager *tm, dm_block_t b, |
| 460 | int *result) |
| 461 | { |
| 462 | if (tm->is_clone) |
| 463 | return -EWOULDBLOCK; |
| 464 | |
| 465 | return dm_sm_count_is_more_than_one(sm: tm->sm, b, result); |
| 466 | } |
| 467 | |
| 468 | struct dm_block_manager *dm_tm_get_bm(struct dm_transaction_manager *tm) |
| 469 | { |
| 470 | return tm->bm; |
| 471 | } |
| 472 | |
| 473 | void dm_tm_issue_prefetches(struct dm_transaction_manager *tm) |
| 474 | { |
| 475 | prefetch_issue(p: &tm->prefetches, bm: tm->bm); |
| 476 | } |
| 477 | EXPORT_SYMBOL_GPL(dm_tm_issue_prefetches); |
| 478 | |
| 479 | /*----------------------------------------------------------------*/ |
| 480 | |
| 481 | static int dm_tm_create_internal(struct dm_block_manager *bm, |
| 482 | dm_block_t sb_location, |
| 483 | struct dm_transaction_manager **tm, |
| 484 | struct dm_space_map **sm, |
| 485 | int create, |
| 486 | void *sm_root, size_t sm_len) |
| 487 | { |
| 488 | int r; |
| 489 | |
| 490 | *sm = dm_sm_metadata_init(); |
| 491 | if (IS_ERR(ptr: *sm)) |
| 492 | return PTR_ERR(ptr: *sm); |
| 493 | |
| 494 | *tm = dm_tm_create(bm, sm: *sm); |
| 495 | if (IS_ERR(ptr: *tm)) { |
| 496 | dm_sm_destroy(sm: *sm); |
| 497 | return PTR_ERR(ptr: *tm); |
| 498 | } |
| 499 | |
| 500 | if (create) { |
| 501 | r = dm_sm_metadata_create(sm: *sm, tm: *tm, nr_blocks: dm_bm_nr_blocks(bm), |
| 502 | superblock: sb_location); |
| 503 | if (r) { |
| 504 | DMERR("couldn't create metadata space map" ); |
| 505 | goto bad; |
| 506 | } |
| 507 | |
| 508 | } else { |
| 509 | r = dm_sm_metadata_open(sm: *sm, tm: *tm, root_le: sm_root, len: sm_len); |
| 510 | if (r) { |
| 511 | DMERR("couldn't open metadata space map" ); |
| 512 | goto bad; |
| 513 | } |
| 514 | } |
| 515 | |
| 516 | return 0; |
| 517 | |
| 518 | bad: |
| 519 | dm_tm_destroy(*tm); |
| 520 | dm_sm_destroy(sm: *sm); |
| 521 | return r; |
| 522 | } |
| 523 | |
| 524 | int dm_tm_create_with_sm(struct dm_block_manager *bm, dm_block_t sb_location, |
| 525 | struct dm_transaction_manager **tm, |
| 526 | struct dm_space_map **sm) |
| 527 | { |
| 528 | return dm_tm_create_internal(bm, sb_location, tm, sm, create: 1, NULL, sm_len: 0); |
| 529 | } |
| 530 | EXPORT_SYMBOL_GPL(dm_tm_create_with_sm); |
| 531 | |
| 532 | int dm_tm_open_with_sm(struct dm_block_manager *bm, dm_block_t sb_location, |
| 533 | void *sm_root, size_t root_len, |
| 534 | struct dm_transaction_manager **tm, |
| 535 | struct dm_space_map **sm) |
| 536 | { |
| 537 | return dm_tm_create_internal(bm, sb_location, tm, sm, create: 0, sm_root, sm_len: root_len); |
| 538 | } |
| 539 | EXPORT_SYMBOL_GPL(dm_tm_open_with_sm); |
| 540 | |
| 541 | /*----------------------------------------------------------------*/ |
| 542 | |