Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1 | /* |
| 2 | * Primary bucket allocation code |
| 3 | * |
| 4 | * Copyright 2012 Google, Inc. |
| 5 | * |
| 6 | * Allocation in bcache is done in terms of buckets: |
| 7 | * |
| 8 | * Each bucket has associated an 8 bit gen; this gen corresponds to the gen in |
| 9 | * btree pointers - they must match for the pointer to be considered valid. |
| 10 | * |
| 11 | * Thus (assuming a bucket has no dirty data or metadata in it) we can reuse a |
| 12 | * bucket simply by incrementing its gen. |
| 13 | * |
| 14 | * The gens (along with the priorities; it's really the gens are important but |
| 15 | * the code is named as if it's the priorities) are written in an arbitrary list |
| 16 | * of buckets on disk, with a pointer to them in the journal header. |
| 17 | * |
| 18 | * When we invalidate a bucket, we have to write its new gen to disk and wait |
| 19 | * for that write to complete before we use it - otherwise after a crash we |
| 20 | * could have pointers that appeared to be good but pointed to data that had |
| 21 | * been overwritten. |
| 22 | * |
| 23 | * Since the gens and priorities are all stored contiguously on disk, we can |
| 24 | * batch this up: We fill up the free_inc list with freshly invalidated buckets, |
| 25 | * call prio_write(), and when prio_write() finishes we pull buckets off the |
| 26 | * free_inc list and optionally discard them. |
| 27 | * |
| 28 | * free_inc isn't the only freelist - if it was, we'd often to sleep while |
| 29 | * priorities and gens were being written before we could allocate. c->free is a |
| 30 | * smaller freelist, and buckets on that list are always ready to be used. |
| 31 | * |
| 32 | * If we've got discards enabled, that happens when a bucket moves from the |
| 33 | * free_inc list to the free list. |
| 34 | * |
| 35 | * There is another freelist, because sometimes we have buckets that we know |
| 36 | * have nothing pointing into them - these we can reuse without waiting for |
| 37 | * priorities to be rewritten. These come from freed btree nodes and buckets |
| 38 | * that garbage collection discovered no longer had valid keys pointing into |
| 39 | * them (because they were overwritten). That's the unused list - buckets on the |
| 40 | * unused list move to the free list, optionally being discarded in the process. |
| 41 | * |
| 42 | * It's also important to ensure that gens don't wrap around - with respect to |
| 43 | * either the oldest gen in the btree or the gen on disk. This is quite |
| 44 | * difficult to do in practice, but we explicitly guard against it anyways - if |
| 45 | * a bucket is in danger of wrapping around we simply skip invalidating it that |
| 46 | * time around, and we garbage collect or rewrite the priorities sooner than we |
| 47 | * would have otherwise. |
| 48 | * |
| 49 | * bch_bucket_alloc() allocates a single bucket from a specific cache. |
| 50 | * |
| 51 | * bch_bucket_alloc_set() allocates one or more buckets from different caches |
| 52 | * out of a cache set. |
| 53 | * |
| 54 | * free_some_buckets() drives all the processes described above. It's called |
| 55 | * from bch_bucket_alloc() and a few other places that need to make sure free |
| 56 | * buckets are ready. |
| 57 | * |
| 58 | * invalidate_buckets_(lru|fifo)() find buckets that are available to be |
| 59 | * invalidated, and then invalidate them and stick them on the free_inc list - |
| 60 | * in either lru or fifo order. |
| 61 | */ |
| 62 | |
| 63 | #include "bcache.h" |
| 64 | #include "btree.h" |
| 65 | |
Kent Overstreet | 49b1212 | 2013-07-24 17:16:09 -0700 | [diff] [blame] | 66 | #include <linux/blkdev.h> |
Kent Overstreet | 79826c3 | 2013-07-10 18:31:58 -0700 | [diff] [blame] | 67 | #include <linux/freezer.h> |
Kent Overstreet | 119ba0f | 2013-04-24 19:01:12 -0700 | [diff] [blame] | 68 | #include <linux/kthread.h> |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 69 | #include <linux/random.h> |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 70 | #include <trace/events/bcache.h> |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 71 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 72 | /* Bucket heap / gen */ |
| 73 | |
| 74 | uint8_t bch_inc_gen(struct cache *ca, struct bucket *b) |
| 75 | { |
| 76 | uint8_t ret = ++b->gen; |
| 77 | |
| 78 | ca->set->need_gc = max(ca->set->need_gc, bucket_gc_gen(b)); |
| 79 | WARN_ON_ONCE(ca->set->need_gc > BUCKET_GC_GEN_MAX); |
| 80 | |
| 81 | if (CACHE_SYNC(&ca->set->sb)) { |
| 82 | ca->need_save_prio = max(ca->need_save_prio, |
| 83 | bucket_disk_gen(b)); |
| 84 | WARN_ON_ONCE(ca->need_save_prio > BUCKET_DISK_GEN_MAX); |
| 85 | } |
| 86 | |
| 87 | return ret; |
| 88 | } |
| 89 | |
| 90 | void bch_rescale_priorities(struct cache_set *c, int sectors) |
| 91 | { |
| 92 | struct cache *ca; |
| 93 | struct bucket *b; |
| 94 | unsigned next = c->nbuckets * c->sb.bucket_size / 1024; |
| 95 | unsigned i; |
| 96 | int r; |
| 97 | |
| 98 | atomic_sub(sectors, &c->rescale); |
| 99 | |
| 100 | do { |
| 101 | r = atomic_read(&c->rescale); |
| 102 | |
| 103 | if (r >= 0) |
| 104 | return; |
| 105 | } while (atomic_cmpxchg(&c->rescale, r, r + next) != r); |
| 106 | |
| 107 | mutex_lock(&c->bucket_lock); |
| 108 | |
| 109 | c->min_prio = USHRT_MAX; |
| 110 | |
| 111 | for_each_cache(ca, c, i) |
| 112 | for_each_bucket(b, ca) |
| 113 | if (b->prio && |
| 114 | b->prio != BTREE_PRIO && |
| 115 | !atomic_read(&b->pin)) { |
| 116 | b->prio--; |
| 117 | c->min_prio = min(c->min_prio, b->prio); |
| 118 | } |
| 119 | |
| 120 | mutex_unlock(&c->bucket_lock); |
| 121 | } |
| 122 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 123 | /* Allocation */ |
| 124 | |
| 125 | static inline bool can_inc_bucket_gen(struct bucket *b) |
| 126 | { |
| 127 | return bucket_gc_gen(b) < BUCKET_GC_GEN_MAX && |
| 128 | bucket_disk_gen(b) < BUCKET_DISK_GEN_MAX; |
| 129 | } |
| 130 | |
| 131 | bool bch_bucket_add_unused(struct cache *ca, struct bucket *b) |
| 132 | { |
| 133 | BUG_ON(GC_MARK(b) || GC_SECTORS_USED(b)); |
| 134 | |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 135 | if (CACHE_REPLACEMENT(&ca->sb) == CACHE_REPLACEMENT_FIFO) { |
| 136 | unsigned i; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 137 | |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 138 | for (i = 0; i < RESERVE_NONE; i++) |
| 139 | if (!fifo_full(&ca->free[i])) |
| 140 | goto add; |
| 141 | |
| 142 | return false; |
| 143 | } |
| 144 | add: |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 145 | b->prio = 0; |
| 146 | |
| 147 | if (can_inc_bucket_gen(b) && |
| 148 | fifo_push(&ca->unused, b - ca->buckets)) { |
| 149 | atomic_inc(&b->pin); |
| 150 | return true; |
| 151 | } |
| 152 | |
| 153 | return false; |
| 154 | } |
| 155 | |
| 156 | static bool can_invalidate_bucket(struct cache *ca, struct bucket *b) |
| 157 | { |
| 158 | return GC_MARK(b) == GC_MARK_RECLAIMABLE && |
| 159 | !atomic_read(&b->pin) && |
| 160 | can_inc_bucket_gen(b); |
| 161 | } |
| 162 | |
| 163 | static void invalidate_one_bucket(struct cache *ca, struct bucket *b) |
| 164 | { |
| 165 | bch_inc_gen(ca, b); |
| 166 | b->prio = INITIAL_PRIO; |
| 167 | atomic_inc(&b->pin); |
| 168 | fifo_push(&ca->free_inc, b - ca->buckets); |
| 169 | } |
| 170 | |
Kent Overstreet | e0a985a | 2013-11-12 13:49:10 -0800 | [diff] [blame^] | 171 | /* |
| 172 | * Determines what order we're going to reuse buckets, smallest bucket_prio() |
| 173 | * first: we also take into account the number of sectors of live data in that |
| 174 | * bucket, and in order for that multiply to make sense we have to scale bucket |
| 175 | * |
| 176 | * Thus, we scale the bucket priorities so that the bucket with the smallest |
| 177 | * prio is worth 1/8th of what INITIAL_PRIO is worth. |
| 178 | */ |
| 179 | |
| 180 | #define bucket_prio(b) \ |
| 181 | ({ \ |
| 182 | unsigned min_prio = (INITIAL_PRIO - ca->set->min_prio) / 8; \ |
| 183 | \ |
| 184 | (b->prio - ca->set->min_prio + min_prio) * GC_SECTORS_USED(b); \ |
| 185 | }) |
Kent Overstreet | b1a67b0 | 2013-03-25 11:46:44 -0700 | [diff] [blame] | 186 | |
| 187 | #define bucket_max_cmp(l, r) (bucket_prio(l) < bucket_prio(r)) |
| 188 | #define bucket_min_cmp(l, r) (bucket_prio(l) > bucket_prio(r)) |
| 189 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 190 | static void invalidate_buckets_lru(struct cache *ca) |
| 191 | { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 192 | struct bucket *b; |
| 193 | ssize_t i; |
| 194 | |
| 195 | ca->heap.used = 0; |
| 196 | |
| 197 | for_each_bucket(b, ca) { |
Kent Overstreet | 86b26b8 | 2013-04-30 19:14:40 -0700 | [diff] [blame] | 198 | /* |
| 199 | * If we fill up the unused list, if we then return before |
| 200 | * adding anything to the free_inc list we'll skip writing |
| 201 | * prios/gens and just go back to allocating from the unused |
| 202 | * list: |
| 203 | */ |
| 204 | if (fifo_full(&ca->unused)) |
| 205 | return; |
| 206 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 207 | if (!can_invalidate_bucket(ca, b)) |
| 208 | continue; |
| 209 | |
Kent Overstreet | 86b26b8 | 2013-04-30 19:14:40 -0700 | [diff] [blame] | 210 | if (!GC_SECTORS_USED(b) && |
| 211 | bch_bucket_add_unused(ca, b)) |
| 212 | continue; |
| 213 | |
| 214 | if (!heap_full(&ca->heap)) |
| 215 | heap_add(&ca->heap, b, bucket_max_cmp); |
| 216 | else if (bucket_max_cmp(b, heap_peek(&ca->heap))) { |
| 217 | ca->heap.data[0] = b; |
| 218 | heap_sift(&ca->heap, 0, bucket_max_cmp); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 219 | } |
| 220 | } |
| 221 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 222 | for (i = ca->heap.used / 2 - 1; i >= 0; --i) |
| 223 | heap_sift(&ca->heap, i, bucket_min_cmp); |
| 224 | |
| 225 | while (!fifo_full(&ca->free_inc)) { |
| 226 | if (!heap_pop(&ca->heap, b, bucket_min_cmp)) { |
Kent Overstreet | 86b26b8 | 2013-04-30 19:14:40 -0700 | [diff] [blame] | 227 | /* |
| 228 | * We don't want to be calling invalidate_buckets() |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 229 | * multiple times when it can't do anything |
| 230 | */ |
| 231 | ca->invalidate_needs_gc = 1; |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 232 | wake_up_gc(ca->set); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 233 | return; |
| 234 | } |
| 235 | |
| 236 | invalidate_one_bucket(ca, b); |
| 237 | } |
| 238 | } |
| 239 | |
| 240 | static void invalidate_buckets_fifo(struct cache *ca) |
| 241 | { |
| 242 | struct bucket *b; |
| 243 | size_t checked = 0; |
| 244 | |
| 245 | while (!fifo_full(&ca->free_inc)) { |
| 246 | if (ca->fifo_last_bucket < ca->sb.first_bucket || |
| 247 | ca->fifo_last_bucket >= ca->sb.nbuckets) |
| 248 | ca->fifo_last_bucket = ca->sb.first_bucket; |
| 249 | |
| 250 | b = ca->buckets + ca->fifo_last_bucket++; |
| 251 | |
| 252 | if (can_invalidate_bucket(ca, b)) |
| 253 | invalidate_one_bucket(ca, b); |
| 254 | |
| 255 | if (++checked >= ca->sb.nbuckets) { |
| 256 | ca->invalidate_needs_gc = 1; |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 257 | wake_up_gc(ca->set); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 258 | return; |
| 259 | } |
| 260 | } |
| 261 | } |
| 262 | |
| 263 | static void invalidate_buckets_random(struct cache *ca) |
| 264 | { |
| 265 | struct bucket *b; |
| 266 | size_t checked = 0; |
| 267 | |
| 268 | while (!fifo_full(&ca->free_inc)) { |
| 269 | size_t n; |
| 270 | get_random_bytes(&n, sizeof(n)); |
| 271 | |
| 272 | n %= (size_t) (ca->sb.nbuckets - ca->sb.first_bucket); |
| 273 | n += ca->sb.first_bucket; |
| 274 | |
| 275 | b = ca->buckets + n; |
| 276 | |
| 277 | if (can_invalidate_bucket(ca, b)) |
| 278 | invalidate_one_bucket(ca, b); |
| 279 | |
| 280 | if (++checked >= ca->sb.nbuckets / 2) { |
| 281 | ca->invalidate_needs_gc = 1; |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 282 | wake_up_gc(ca->set); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 283 | return; |
| 284 | } |
| 285 | } |
| 286 | } |
| 287 | |
| 288 | static void invalidate_buckets(struct cache *ca) |
| 289 | { |
| 290 | if (ca->invalidate_needs_gc) |
| 291 | return; |
| 292 | |
| 293 | switch (CACHE_REPLACEMENT(&ca->sb)) { |
| 294 | case CACHE_REPLACEMENT_LRU: |
| 295 | invalidate_buckets_lru(ca); |
| 296 | break; |
| 297 | case CACHE_REPLACEMENT_FIFO: |
| 298 | invalidate_buckets_fifo(ca); |
| 299 | break; |
| 300 | case CACHE_REPLACEMENT_RANDOM: |
| 301 | invalidate_buckets_random(ca); |
| 302 | break; |
| 303 | } |
Kent Overstreet | 86b26b8 | 2013-04-30 19:14:40 -0700 | [diff] [blame] | 304 | |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 305 | trace_bcache_alloc_invalidate(ca); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 306 | } |
| 307 | |
| 308 | #define allocator_wait(ca, cond) \ |
| 309 | do { \ |
Kent Overstreet | 86b26b8 | 2013-04-30 19:14:40 -0700 | [diff] [blame] | 310 | while (1) { \ |
Kent Overstreet | 119ba0f | 2013-04-24 19:01:12 -0700 | [diff] [blame] | 311 | set_current_state(TASK_INTERRUPTIBLE); \ |
Kent Overstreet | 86b26b8 | 2013-04-30 19:14:40 -0700 | [diff] [blame] | 312 | if (cond) \ |
| 313 | break; \ |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 314 | \ |
| 315 | mutex_unlock(&(ca)->set->bucket_lock); \ |
Kent Overstreet | 79826c3 | 2013-07-10 18:31:58 -0700 | [diff] [blame] | 316 | if (kthread_should_stop()) \ |
Kent Overstreet | 119ba0f | 2013-04-24 19:01:12 -0700 | [diff] [blame] | 317 | return 0; \ |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 318 | \ |
Kent Overstreet | 79826c3 | 2013-07-10 18:31:58 -0700 | [diff] [blame] | 319 | try_to_freeze(); \ |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 320 | schedule(); \ |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 321 | mutex_lock(&(ca)->set->bucket_lock); \ |
| 322 | } \ |
Kent Overstreet | 119ba0f | 2013-04-24 19:01:12 -0700 | [diff] [blame] | 323 | __set_current_state(TASK_RUNNING); \ |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 324 | } while (0) |
| 325 | |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 326 | static int bch_allocator_push(struct cache *ca, long bucket) |
| 327 | { |
| 328 | unsigned i; |
| 329 | |
| 330 | /* Prios/gens are actually the most important reserve */ |
| 331 | if (fifo_push(&ca->free[RESERVE_PRIO], bucket)) |
| 332 | return true; |
| 333 | |
| 334 | for (i = 0; i < RESERVE_NR; i++) |
| 335 | if (fifo_push(&ca->free[i], bucket)) |
| 336 | return true; |
| 337 | |
| 338 | return false; |
| 339 | } |
| 340 | |
Kent Overstreet | 119ba0f | 2013-04-24 19:01:12 -0700 | [diff] [blame] | 341 | static int bch_allocator_thread(void *arg) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 342 | { |
Kent Overstreet | 119ba0f | 2013-04-24 19:01:12 -0700 | [diff] [blame] | 343 | struct cache *ca = arg; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 344 | |
| 345 | mutex_lock(&ca->set->bucket_lock); |
| 346 | |
| 347 | while (1) { |
Kent Overstreet | 86b26b8 | 2013-04-30 19:14:40 -0700 | [diff] [blame] | 348 | /* |
| 349 | * First, we pull buckets off of the unused and free_inc lists, |
| 350 | * possibly issue discards to them, then we add the bucket to |
| 351 | * the free list: |
| 352 | */ |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 353 | while (1) { |
| 354 | long bucket; |
| 355 | |
| 356 | if ((!atomic_read(&ca->set->prio_blocked) || |
| 357 | !CACHE_SYNC(&ca->set->sb)) && |
| 358 | !fifo_empty(&ca->unused)) |
| 359 | fifo_pop(&ca->unused, bucket); |
| 360 | else if (!fifo_empty(&ca->free_inc)) |
| 361 | fifo_pop(&ca->free_inc, bucket); |
| 362 | else |
| 363 | break; |
| 364 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 365 | if (ca->discard) { |
Kent Overstreet | 49b1212 | 2013-07-24 17:16:09 -0700 | [diff] [blame] | 366 | mutex_unlock(&ca->set->bucket_lock); |
| 367 | blkdev_issue_discard(ca->bdev, |
| 368 | bucket_to_sector(ca->set, bucket), |
| 369 | ca->sb.block_size, GFP_KERNEL, 0); |
| 370 | mutex_lock(&ca->set->bucket_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 371 | } |
Kent Overstreet | 49b1212 | 2013-07-24 17:16:09 -0700 | [diff] [blame] | 372 | |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 373 | allocator_wait(ca, bch_allocator_push(ca, bucket)); |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 374 | wake_up(&ca->set->bucket_wait); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 375 | } |
| 376 | |
Kent Overstreet | 86b26b8 | 2013-04-30 19:14:40 -0700 | [diff] [blame] | 377 | /* |
| 378 | * We've run out of free buckets, we need to find some buckets |
| 379 | * we can invalidate. First, invalidate them in memory and add |
| 380 | * them to the free_inc list: |
| 381 | */ |
| 382 | |
| 383 | allocator_wait(ca, ca->set->gc_mark_valid && |
| 384 | (ca->need_save_prio > 64 || |
| 385 | !ca->invalidate_needs_gc)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 386 | invalidate_buckets(ca); |
| 387 | |
Kent Overstreet | 86b26b8 | 2013-04-30 19:14:40 -0700 | [diff] [blame] | 388 | /* |
| 389 | * Now, we write their new gens to disk so we can start writing |
| 390 | * new stuff to them: |
| 391 | */ |
| 392 | allocator_wait(ca, !atomic_read(&ca->set->prio_blocked)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 393 | if (CACHE_SYNC(&ca->set->sb) && |
| 394 | (!fifo_empty(&ca->free_inc) || |
Kent Overstreet | 86b26b8 | 2013-04-30 19:14:40 -0700 | [diff] [blame] | 395 | ca->need_save_prio > 64)) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 396 | bch_prio_write(ca); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 397 | } |
| 398 | } |
| 399 | |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 400 | long bch_bucket_alloc(struct cache *ca, unsigned reserve, bool wait) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 401 | { |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 402 | DEFINE_WAIT(w); |
| 403 | struct bucket *b; |
| 404 | long r; |
| 405 | |
| 406 | /* fastpath */ |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 407 | if (fifo_pop(&ca->free[RESERVE_NONE], r) || |
| 408 | fifo_pop(&ca->free[reserve], r)) |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 409 | goto out; |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 410 | |
| 411 | if (!wait) |
| 412 | return -1; |
| 413 | |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 414 | do { |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 415 | prepare_to_wait(&ca->set->bucket_wait, &w, |
| 416 | TASK_UNINTERRUPTIBLE); |
| 417 | |
| 418 | mutex_unlock(&ca->set->bucket_lock); |
| 419 | schedule(); |
| 420 | mutex_lock(&ca->set->bucket_lock); |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 421 | } while (!fifo_pop(&ca->free[RESERVE_NONE], r) && |
| 422 | !fifo_pop(&ca->free[reserve], r)); |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 423 | |
| 424 | finish_wait(&ca->set->bucket_wait, &w); |
| 425 | out: |
Kent Overstreet | 119ba0f | 2013-04-24 19:01:12 -0700 | [diff] [blame] | 426 | wake_up_process(ca->alloc_thread); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 427 | |
Kent Overstreet | 280481d | 2013-10-24 16:36:03 -0700 | [diff] [blame] | 428 | if (expensive_debug_checks(ca->set)) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 429 | size_t iter; |
| 430 | long i; |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 431 | unsigned j; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 432 | |
| 433 | for (iter = 0; iter < prio_buckets(ca) * 2; iter++) |
| 434 | BUG_ON(ca->prio_buckets[iter] == (uint64_t) r); |
| 435 | |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 436 | for (j = 0; j < RESERVE_NR; j++) |
| 437 | fifo_for_each(i, &ca->free[j], iter) |
| 438 | BUG_ON(i == r); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 439 | fifo_for_each(i, &ca->free_inc, iter) |
| 440 | BUG_ON(i == r); |
| 441 | fifo_for_each(i, &ca->unused, iter) |
| 442 | BUG_ON(i == r); |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 443 | } |
Kent Overstreet | 280481d | 2013-10-24 16:36:03 -0700 | [diff] [blame] | 444 | |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 445 | b = ca->buckets + r; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 446 | |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 447 | BUG_ON(atomic_read(&b->pin) != 1); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 448 | |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 449 | SET_GC_SECTORS_USED(b, ca->sb.bucket_size); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 450 | |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 451 | if (reserve <= RESERVE_PRIO) { |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 452 | SET_GC_MARK(b, GC_MARK_METADATA); |
Nicholas Swenson | 981aa8c | 2013-11-07 17:53:19 -0800 | [diff] [blame] | 453 | SET_GC_MOVE(b, 0); |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 454 | b->prio = BTREE_PRIO; |
| 455 | } else { |
| 456 | SET_GC_MARK(b, GC_MARK_RECLAIMABLE); |
Nicholas Swenson | 981aa8c | 2013-11-07 17:53:19 -0800 | [diff] [blame] | 457 | SET_GC_MOVE(b, 0); |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 458 | b->prio = INITIAL_PRIO; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 459 | } |
| 460 | |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 461 | return r; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 462 | } |
| 463 | |
| 464 | void bch_bucket_free(struct cache_set *c, struct bkey *k) |
| 465 | { |
| 466 | unsigned i; |
| 467 | |
| 468 | for (i = 0; i < KEY_PTRS(k); i++) { |
| 469 | struct bucket *b = PTR_BUCKET(c, k, i); |
| 470 | |
Kent Overstreet | 86b26b8 | 2013-04-30 19:14:40 -0700 | [diff] [blame] | 471 | SET_GC_MARK(b, GC_MARK_RECLAIMABLE); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 472 | SET_GC_SECTORS_USED(b, 0); |
| 473 | bch_bucket_add_unused(PTR_CACHE(c, k, i), b); |
| 474 | } |
| 475 | } |
| 476 | |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 477 | int __bch_bucket_alloc_set(struct cache_set *c, unsigned reserve, |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 478 | struct bkey *k, int n, bool wait) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 479 | { |
| 480 | int i; |
| 481 | |
| 482 | lockdep_assert_held(&c->bucket_lock); |
| 483 | BUG_ON(!n || n > c->caches_loaded || n > 8); |
| 484 | |
| 485 | bkey_init(k); |
| 486 | |
| 487 | /* sort by free space/prio of oldest data in caches */ |
| 488 | |
| 489 | for (i = 0; i < n; i++) { |
| 490 | struct cache *ca = c->cache_by_alloc[i]; |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 491 | long b = bch_bucket_alloc(ca, reserve, wait); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 492 | |
| 493 | if (b == -1) |
| 494 | goto err; |
| 495 | |
| 496 | k->ptr[i] = PTR(ca->buckets[b].gen, |
| 497 | bucket_to_sector(c, b), |
| 498 | ca->sb.nr_this_dev); |
| 499 | |
| 500 | SET_KEY_PTRS(k, i + 1); |
| 501 | } |
| 502 | |
| 503 | return 0; |
| 504 | err: |
| 505 | bch_bucket_free(c, k); |
Kent Overstreet | 3a3b6a4 | 2013-07-24 16:46:42 -0700 | [diff] [blame] | 506 | bkey_put(c, k); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 507 | return -1; |
| 508 | } |
| 509 | |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 510 | int bch_bucket_alloc_set(struct cache_set *c, unsigned reserve, |
Kent Overstreet | 35fcd84 | 2013-07-24 17:29:09 -0700 | [diff] [blame] | 511 | struct bkey *k, int n, bool wait) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 512 | { |
| 513 | int ret; |
| 514 | mutex_lock(&c->bucket_lock); |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 515 | ret = __bch_bucket_alloc_set(c, reserve, k, n, wait); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 516 | mutex_unlock(&c->bucket_lock); |
| 517 | return ret; |
| 518 | } |
| 519 | |
Kent Overstreet | 2599b53 | 2013-07-24 18:11:11 -0700 | [diff] [blame] | 520 | /* Sector allocator */ |
| 521 | |
| 522 | struct open_bucket { |
| 523 | struct list_head list; |
| 524 | unsigned last_write_point; |
| 525 | unsigned sectors_free; |
| 526 | BKEY_PADDED(key); |
| 527 | }; |
| 528 | |
| 529 | /* |
| 530 | * We keep multiple buckets open for writes, and try to segregate different |
| 531 | * write streams for better cache utilization: first we look for a bucket where |
| 532 | * the last write to it was sequential with the current write, and failing that |
| 533 | * we look for a bucket that was last used by the same task. |
| 534 | * |
| 535 | * The ideas is if you've got multiple tasks pulling data into the cache at the |
| 536 | * same time, you'll get better cache utilization if you try to segregate their |
| 537 | * data and preserve locality. |
| 538 | * |
| 539 | * For example, say you've starting Firefox at the same time you're copying a |
| 540 | * bunch of files. Firefox will likely end up being fairly hot and stay in the |
| 541 | * cache awhile, but the data you copied might not be; if you wrote all that |
| 542 | * data to the same buckets it'd get invalidated at the same time. |
| 543 | * |
| 544 | * Both of those tasks will be doing fairly random IO so we can't rely on |
| 545 | * detecting sequential IO to segregate their data, but going off of the task |
| 546 | * should be a sane heuristic. |
| 547 | */ |
| 548 | static struct open_bucket *pick_data_bucket(struct cache_set *c, |
| 549 | const struct bkey *search, |
| 550 | unsigned write_point, |
| 551 | struct bkey *alloc) |
| 552 | { |
| 553 | struct open_bucket *ret, *ret_task = NULL; |
| 554 | |
| 555 | list_for_each_entry_reverse(ret, &c->data_buckets, list) |
| 556 | if (!bkey_cmp(&ret->key, search)) |
| 557 | goto found; |
| 558 | else if (ret->last_write_point == write_point) |
| 559 | ret_task = ret; |
| 560 | |
| 561 | ret = ret_task ?: list_first_entry(&c->data_buckets, |
| 562 | struct open_bucket, list); |
| 563 | found: |
| 564 | if (!ret->sectors_free && KEY_PTRS(alloc)) { |
| 565 | ret->sectors_free = c->sb.bucket_size; |
| 566 | bkey_copy(&ret->key, alloc); |
| 567 | bkey_init(alloc); |
| 568 | } |
| 569 | |
| 570 | if (!ret->sectors_free) |
| 571 | ret = NULL; |
| 572 | |
| 573 | return ret; |
| 574 | } |
| 575 | |
| 576 | /* |
| 577 | * Allocates some space in the cache to write to, and k to point to the newly |
| 578 | * allocated space, and updates KEY_SIZE(k) and KEY_OFFSET(k) (to point to the |
| 579 | * end of the newly allocated space). |
| 580 | * |
| 581 | * May allocate fewer sectors than @sectors, KEY_SIZE(k) indicates how many |
| 582 | * sectors were actually allocated. |
| 583 | * |
| 584 | * If s->writeback is true, will not fail. |
| 585 | */ |
| 586 | bool bch_alloc_sectors(struct cache_set *c, struct bkey *k, unsigned sectors, |
| 587 | unsigned write_point, unsigned write_prio, bool wait) |
| 588 | { |
| 589 | struct open_bucket *b; |
| 590 | BKEY_PADDED(key) alloc; |
| 591 | unsigned i; |
| 592 | |
| 593 | /* |
| 594 | * We might have to allocate a new bucket, which we can't do with a |
| 595 | * spinlock held. So if we have to allocate, we drop the lock, allocate |
| 596 | * and then retry. KEY_PTRS() indicates whether alloc points to |
| 597 | * allocated bucket(s). |
| 598 | */ |
| 599 | |
| 600 | bkey_init(&alloc.key); |
| 601 | spin_lock(&c->data_bucket_lock); |
| 602 | |
| 603 | while (!(b = pick_data_bucket(c, k, write_point, &alloc.key))) { |
| 604 | unsigned watermark = write_prio |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 605 | ? RESERVE_MOVINGGC |
| 606 | : RESERVE_NONE; |
Kent Overstreet | 2599b53 | 2013-07-24 18:11:11 -0700 | [diff] [blame] | 607 | |
| 608 | spin_unlock(&c->data_bucket_lock); |
| 609 | |
| 610 | if (bch_bucket_alloc_set(c, watermark, &alloc.key, 1, wait)) |
| 611 | return false; |
| 612 | |
| 613 | spin_lock(&c->data_bucket_lock); |
| 614 | } |
| 615 | |
| 616 | /* |
| 617 | * If we had to allocate, we might race and not need to allocate the |
| 618 | * second time we call find_data_bucket(). If we allocated a bucket but |
| 619 | * didn't use it, drop the refcount bch_bucket_alloc_set() took: |
| 620 | */ |
| 621 | if (KEY_PTRS(&alloc.key)) |
Kent Overstreet | 3a3b6a4 | 2013-07-24 16:46:42 -0700 | [diff] [blame] | 622 | bkey_put(c, &alloc.key); |
Kent Overstreet | 2599b53 | 2013-07-24 18:11:11 -0700 | [diff] [blame] | 623 | |
| 624 | for (i = 0; i < KEY_PTRS(&b->key); i++) |
| 625 | EBUG_ON(ptr_stale(c, &b->key, i)); |
| 626 | |
| 627 | /* Set up the pointer to the space we're allocating: */ |
| 628 | |
| 629 | for (i = 0; i < KEY_PTRS(&b->key); i++) |
| 630 | k->ptr[i] = b->key.ptr[i]; |
| 631 | |
| 632 | sectors = min(sectors, b->sectors_free); |
| 633 | |
| 634 | SET_KEY_OFFSET(k, KEY_OFFSET(k) + sectors); |
| 635 | SET_KEY_SIZE(k, sectors); |
| 636 | SET_KEY_PTRS(k, KEY_PTRS(&b->key)); |
| 637 | |
| 638 | /* |
| 639 | * Move b to the end of the lru, and keep track of what this bucket was |
| 640 | * last used for: |
| 641 | */ |
| 642 | list_move_tail(&b->list, &c->data_buckets); |
| 643 | bkey_copy_key(&b->key, k); |
| 644 | b->last_write_point = write_point; |
| 645 | |
| 646 | b->sectors_free -= sectors; |
| 647 | |
| 648 | for (i = 0; i < KEY_PTRS(&b->key); i++) { |
| 649 | SET_PTR_OFFSET(&b->key, i, PTR_OFFSET(&b->key, i) + sectors); |
| 650 | |
| 651 | atomic_long_add(sectors, |
| 652 | &PTR_CACHE(c, &b->key, i)->sectors_written); |
| 653 | } |
| 654 | |
| 655 | if (b->sectors_free < c->sb.block_size) |
| 656 | b->sectors_free = 0; |
| 657 | |
| 658 | /* |
| 659 | * k takes refcounts on the buckets it points to until it's inserted |
| 660 | * into the btree, but if we're done with this bucket we just transfer |
| 661 | * get_data_bucket()'s refcount. |
| 662 | */ |
| 663 | if (b->sectors_free) |
| 664 | for (i = 0; i < KEY_PTRS(&b->key); i++) |
| 665 | atomic_inc(&PTR_BUCKET(c, &b->key, i)->pin); |
| 666 | |
| 667 | spin_unlock(&c->data_bucket_lock); |
| 668 | return true; |
| 669 | } |
| 670 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 671 | /* Init */ |
| 672 | |
Kent Overstreet | 2599b53 | 2013-07-24 18:11:11 -0700 | [diff] [blame] | 673 | void bch_open_buckets_free(struct cache_set *c) |
| 674 | { |
| 675 | struct open_bucket *b; |
| 676 | |
| 677 | while (!list_empty(&c->data_buckets)) { |
| 678 | b = list_first_entry(&c->data_buckets, |
| 679 | struct open_bucket, list); |
| 680 | list_del(&b->list); |
| 681 | kfree(b); |
| 682 | } |
| 683 | } |
| 684 | |
| 685 | int bch_open_buckets_alloc(struct cache_set *c) |
| 686 | { |
| 687 | int i; |
| 688 | |
| 689 | spin_lock_init(&c->data_bucket_lock); |
| 690 | |
| 691 | for (i = 0; i < 6; i++) { |
| 692 | struct open_bucket *b = kzalloc(sizeof(*b), GFP_KERNEL); |
| 693 | if (!b) |
| 694 | return -ENOMEM; |
| 695 | |
| 696 | list_add(&b->list, &c->data_buckets); |
| 697 | } |
| 698 | |
| 699 | return 0; |
| 700 | } |
| 701 | |
Kent Overstreet | 119ba0f | 2013-04-24 19:01:12 -0700 | [diff] [blame] | 702 | int bch_cache_allocator_start(struct cache *ca) |
| 703 | { |
Kent Overstreet | 79826c3 | 2013-07-10 18:31:58 -0700 | [diff] [blame] | 704 | struct task_struct *k = kthread_run(bch_allocator_thread, |
| 705 | ca, "bcache_allocator"); |
| 706 | if (IS_ERR(k)) |
| 707 | return PTR_ERR(k); |
Kent Overstreet | 119ba0f | 2013-04-24 19:01:12 -0700 | [diff] [blame] | 708 | |
Kent Overstreet | 79826c3 | 2013-07-10 18:31:58 -0700 | [diff] [blame] | 709 | ca->alloc_thread = k; |
Kent Overstreet | 119ba0f | 2013-04-24 19:01:12 -0700 | [diff] [blame] | 710 | return 0; |
| 711 | } |
| 712 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 713 | int bch_cache_allocator_init(struct cache *ca) |
| 714 | { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 715 | /* |
| 716 | * Reserve: |
| 717 | * Prio/gen writes first |
| 718 | * Then 8 for btree allocations |
| 719 | * Then half for the moving garbage collector |
| 720 | */ |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 721 | #if 0 |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 722 | ca->watermark[WATERMARK_PRIO] = 0; |
| 723 | |
| 724 | ca->watermark[WATERMARK_METADATA] = prio_buckets(ca); |
| 725 | |
| 726 | ca->watermark[WATERMARK_MOVINGGC] = 8 + |
| 727 | ca->watermark[WATERMARK_METADATA]; |
| 728 | |
| 729 | ca->watermark[WATERMARK_NONE] = ca->free.size / 2 + |
| 730 | ca->watermark[WATERMARK_MOVINGGC]; |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 731 | #endif |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 732 | return 0; |
| 733 | } |