1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Moving/copying garbage collector 4 * 5 * Copyright 2012 Google, Inc. 6 */ 7 8 #include "bcachefs.h" 9 #include "alloc_background.h" 10 #include "alloc_foreground.h" 11 #include "btree_iter.h" 12 #include "btree_update.h" 13 #include "btree_write_buffer.h" 14 #include "buckets.h" 15 #include "clock.h" 16 #include "errcode.h" 17 #include "error.h" 18 #include "lru.h" 19 #include "move.h" 20 #include "movinggc.h" 21 #include "trace.h" 22 23 #include <linux/freezer.h> 24 #include <linux/kthread.h> 25 #include <linux/math64.h> 26 #include <linux/sched/task.h> 27 #include <linux/wait.h> 28 29 struct buckets_in_flight { 30 struct rhashtable table; 31 struct move_bucket_in_flight *first; 32 struct move_bucket_in_flight *last; 33 size_t nr; 34 size_t sectors; 35 }; 36 37 static const struct rhashtable_params bch_move_bucket_params = { 38 .head_offset = offsetof(struct move_bucket_in_flight, hash), 39 .key_offset = offsetof(struct move_bucket_in_flight, bucket.k), 40 .key_len = sizeof(struct move_bucket_key), 41 .automatic_shrinking = true, 42 }; 43 44 static struct move_bucket_in_flight * 45 move_bucket_in_flight_add(struct buckets_in_flight *list, struct move_bucket b) 46 { 47 struct move_bucket_in_flight *new = kzalloc(sizeof(*new), GFP_KERNEL); 48 int ret; 49 50 if (!new) 51 return ERR_PTR(-ENOMEM); 52 53 new->bucket = b; 54 55 ret = rhashtable_lookup_insert_fast(&list->table, &new->hash, 56 bch_move_bucket_params); 57 if (ret) { 58 kfree(new); 59 return ERR_PTR(ret); 60 } 61 62 if (!list->first) 63 list->first = new; 64 else 65 list->last->next = new; 66 67 list->last = new; 68 list->nr++; 69 list->sectors += b.sectors; 70 return new; 71 } 72 73 static int bch2_bucket_is_movable(struct btree_trans *trans, 74 struct move_bucket *b, u64 time) 75 { 76 struct bch_fs *c = trans->c; 77 struct btree_iter iter; 78 struct bkey_s_c k; 79 struct bch_alloc_v4 _a; 80 const struct bch_alloc_v4 *a; 81 int ret; 82 83 if (bch2_bucket_is_open(trans->c, 84 b->k.bucket.inode, 85 b->k.bucket.offset)) 86 return 0; 87 88 k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc, 89 b->k.bucket, BTREE_ITER_cached); 90 ret = bkey_err(k); 91 if (ret) 92 return ret; 93 94 struct bch_dev *ca = bch2_dev_tryget(c, k.k->p.inode); 95 if (!ca) 96 goto out; 97 98 a = bch2_alloc_to_v4(k, &_a); 99 b->k.gen = a->gen; 100 b->sectors = bch2_bucket_sectors_dirty(*a); 101 u64 lru_idx = alloc_lru_idx_fragmentation(*a, ca); 102 103 ret = lru_idx && lru_idx <= time; 104 105 bch2_dev_put(ca); 106 out: 107 bch2_trans_iter_exit(trans, &iter); 108 return ret; 109 } 110 111 static void move_buckets_wait(struct moving_context *ctxt, 112 struct buckets_in_flight *list, 113 bool flush) 114 { 115 struct move_bucket_in_flight *i; 116 int ret; 117 118 while ((i = list->first)) { 119 if (flush) 120 move_ctxt_wait_event(ctxt, !atomic_read(&i->count)); 121 122 if (atomic_read(&i->count)) 123 break; 124 125 list->first = i->next; 126 if (!list->first) 127 list->last = NULL; 128 129 list->nr--; 130 list->sectors -= i->bucket.sectors; 131 132 ret = rhashtable_remove_fast(&list->table, &i->hash, 133 bch_move_bucket_params); 134 BUG_ON(ret); 135 kfree(i); 136 } 137 138 bch2_trans_unlock_long(ctxt->trans); 139 } 140 141 static bool bucket_in_flight(struct buckets_in_flight *list, 142 struct move_bucket_key k) 143 { 144 return rhashtable_lookup_fast(&list->table, &k, bch_move_bucket_params); 145 } 146 147 typedef DARRAY(struct move_bucket) move_buckets; 148 149 static int bch2_copygc_get_buckets(struct moving_context *ctxt, 150 struct buckets_in_flight *buckets_in_flight, 151 move_buckets *buckets) 152 { 153 struct btree_trans *trans = ctxt->trans; 154 struct bch_fs *c = trans->c; 155 size_t nr_to_get = max_t(size_t, 16U, buckets_in_flight->nr / 4); 156 size_t saw = 0, in_flight = 0, not_movable = 0, sectors = 0; 157 int ret; 158 159 move_buckets_wait(ctxt, buckets_in_flight, false); 160 161 ret = bch2_btree_write_buffer_tryflush(trans); 162 if (bch2_err_matches(ret, EROFS)) 163 return ret; 164 165 if (bch2_fs_fatal_err_on(ret, c, "%s: from bch2_btree_write_buffer_tryflush()", bch2_err_str(ret))) 166 return ret; 167 168 bch2_trans_begin(trans); 169 170 ret = for_each_btree_key_max(trans, iter, BTREE_ID_lru, 171 lru_pos(BCH_LRU_FRAGMENTATION_START, 0, 0), 172 lru_pos(BCH_LRU_FRAGMENTATION_START, U64_MAX, LRU_TIME_MAX), 173 0, k, ({ 174 struct move_bucket b = { .k.bucket = u64_to_bucket(k.k->p.offset) }; 175 int ret2 = 0; 176 177 saw++; 178 179 ret2 = bch2_bucket_is_movable(trans, &b, lru_pos_time(k.k->p)); 180 if (ret2 < 0) 181 goto err; 182 183 if (!ret2) 184 not_movable++; 185 else if (bucket_in_flight(buckets_in_flight, b.k)) 186 in_flight++; 187 else { 188 ret2 = darray_push(buckets, b); 189 if (ret2) 190 goto err; 191 sectors += b.sectors; 192 } 193 194 ret2 = buckets->nr >= nr_to_get; 195 err: 196 ret2; 197 })); 198 199 pr_debug("have: %zu (%zu) saw %zu in flight %zu not movable %zu got %zu (%zu)/%zu buckets ret %i", 200 buckets_in_flight->nr, buckets_in_flight->sectors, 201 saw, in_flight, not_movable, buckets->nr, sectors, nr_to_get, ret); 202 203 return ret < 0 ? ret : 0; 204 } 205 206 noinline 207 static int bch2_copygc(struct moving_context *ctxt, 208 struct buckets_in_flight *buckets_in_flight, 209 bool *did_work) 210 { 211 struct btree_trans *trans = ctxt->trans; 212 struct bch_fs *c = trans->c; 213 struct data_update_opts data_opts = { 214 .btree_insert_flags = BCH_WATERMARK_copygc, 215 }; 216 move_buckets buckets = { 0 }; 217 struct move_bucket_in_flight *f; 218 u64 sectors_seen = atomic64_read(&ctxt->stats->sectors_seen); 219 u64 sectors_moved = atomic64_read(&ctxt->stats->sectors_moved); 220 int ret = 0; 221 222 ret = bch2_copygc_get_buckets(ctxt, buckets_in_flight, &buckets); 223 if (ret) 224 goto err; 225 226 darray_for_each(buckets, i) { 227 if (kthread_should_stop() || freezing(current)) 228 break; 229 230 f = move_bucket_in_flight_add(buckets_in_flight, *i); 231 ret = PTR_ERR_OR_ZERO(f); 232 if (ret == -EEXIST) { /* rare race: copygc_get_buckets returned same bucket more than once */ 233 ret = 0; 234 continue; 235 } 236 if (ret == -ENOMEM) { /* flush IO, continue later */ 237 ret = 0; 238 break; 239 } 240 241 ret = bch2_evacuate_bucket(ctxt, f, f->bucket.k.bucket, 242 f->bucket.k.gen, data_opts); 243 if (ret) 244 goto err; 245 246 *did_work = true; 247 } 248 err: 249 250 /* no entries in LRU btree found, or got to end: */ 251 if (bch2_err_matches(ret, ENOENT)) 252 ret = 0; 253 254 if (ret < 0 && !bch2_err_matches(ret, EROFS)) 255 bch_err_msg(c, ret, "from bch2_move_data()"); 256 257 sectors_seen = atomic64_read(&ctxt->stats->sectors_seen) - sectors_seen; 258 sectors_moved = atomic64_read(&ctxt->stats->sectors_moved) - sectors_moved; 259 trace_and_count(c, copygc, c, buckets.nr, sectors_seen, sectors_moved); 260 261 darray_exit(&buckets); 262 return ret; 263 } 264 265 /* 266 * Copygc runs when the amount of fragmented data is above some arbitrary 267 * threshold: 268 * 269 * The threshold at the limit - when the device is full - is the amount of space 270 * we reserved in bch2_recalc_capacity; we can't have more than that amount of 271 * disk space stranded due to fragmentation and store everything we have 272 * promised to store. 273 * 274 * But we don't want to be running copygc unnecessarily when the device still 275 * has plenty of free space - rather, we want copygc to smoothly run every so 276 * often and continually reduce the amount of fragmented space as the device 277 * fills up. So, we increase the threshold by half the current free space. 278 */ 279 unsigned long bch2_copygc_wait_amount(struct bch_fs *c) 280 { 281 s64 wait = S64_MAX, fragmented_allowed, fragmented; 282 283 for_each_rw_member(c, ca) { 284 struct bch_dev_usage usage = bch2_dev_usage_read(ca); 285 286 fragmented_allowed = ((__dev_buckets_available(ca, usage, BCH_WATERMARK_stripe) * 287 ca->mi.bucket_size) >> 1); 288 fragmented = 0; 289 290 for (unsigned i = 0; i < BCH_DATA_NR; i++) 291 if (data_type_movable(i)) 292 fragmented += usage.d[i].fragmented; 293 294 wait = min(wait, max(0LL, fragmented_allowed - fragmented)); 295 } 296 297 return wait; 298 } 299 300 void bch2_copygc_wait_to_text(struct printbuf *out, struct bch_fs *c) 301 { 302 printbuf_tabstop_push(out, 32); 303 prt_printf(out, "running:\t%u\n", c->copygc_running); 304 prt_printf(out, "copygc_wait:\t%llu\n", c->copygc_wait); 305 prt_printf(out, "copygc_wait_at:\t%llu\n", c->copygc_wait_at); 306 307 prt_printf(out, "Currently waiting for:\t"); 308 prt_human_readable_u64(out, max(0LL, c->copygc_wait - 309 atomic64_read(&c->io_clock[WRITE].now)) << 9); 310 prt_newline(out); 311 312 prt_printf(out, "Currently waiting since:\t"); 313 prt_human_readable_u64(out, max(0LL, 314 atomic64_read(&c->io_clock[WRITE].now) - 315 c->copygc_wait_at) << 9); 316 prt_newline(out); 317 318 prt_printf(out, "Currently calculated wait:\t"); 319 prt_human_readable_u64(out, bch2_copygc_wait_amount(c)); 320 prt_newline(out); 321 } 322 323 static int bch2_copygc_thread(void *arg) 324 { 325 struct bch_fs *c = arg; 326 struct moving_context ctxt; 327 struct bch_move_stats move_stats; 328 struct io_clock *clock = &c->io_clock[WRITE]; 329 struct buckets_in_flight *buckets; 330 u64 last, wait; 331 int ret = 0; 332 333 buckets = kzalloc(sizeof(struct buckets_in_flight), GFP_KERNEL); 334 if (!buckets) 335 return -ENOMEM; 336 ret = rhashtable_init(&buckets->table, &bch_move_bucket_params); 337 bch_err_msg(c, ret, "allocating copygc buckets in flight"); 338 if (ret) { 339 kfree(buckets); 340 return ret; 341 } 342 343 set_freezable(); 344 345 bch2_move_stats_init(&move_stats, "copygc"); 346 bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats, 347 writepoint_ptr(&c->copygc_write_point), 348 false); 349 350 while (!ret && !kthread_should_stop()) { 351 bool did_work = false; 352 353 bch2_trans_unlock_long(ctxt.trans); 354 cond_resched(); 355 356 if (!c->opts.copygc_enabled) { 357 move_buckets_wait(&ctxt, buckets, true); 358 kthread_wait_freezable(c->opts.copygc_enabled || 359 kthread_should_stop()); 360 } 361 362 if (unlikely(freezing(current))) { 363 move_buckets_wait(&ctxt, buckets, true); 364 __refrigerator(false); 365 continue; 366 } 367 368 last = atomic64_read(&clock->now); 369 wait = bch2_copygc_wait_amount(c); 370 371 if (wait > clock->max_slop) { 372 c->copygc_wait_at = last; 373 c->copygc_wait = last + wait; 374 move_buckets_wait(&ctxt, buckets, true); 375 trace_and_count(c, copygc_wait, c, wait, last + wait); 376 bch2_kthread_io_clock_wait(clock, last + wait, 377 MAX_SCHEDULE_TIMEOUT); 378 continue; 379 } 380 381 c->copygc_wait = 0; 382 383 c->copygc_running = true; 384 ret = bch2_copygc(&ctxt, buckets, &did_work); 385 c->copygc_running = false; 386 387 wake_up(&c->copygc_running_wq); 388 389 if (!wait && !did_work) { 390 u64 min_member_capacity = bch2_min_rw_member_capacity(c); 391 392 if (min_member_capacity == U64_MAX) 393 min_member_capacity = 128 * 2048; 394 395 move_buckets_wait(&ctxt, buckets, true); 396 bch2_kthread_io_clock_wait(clock, last + (min_member_capacity >> 6), 397 MAX_SCHEDULE_TIMEOUT); 398 } 399 } 400 401 move_buckets_wait(&ctxt, buckets, true); 402 403 rhashtable_destroy(&buckets->table); 404 kfree(buckets); 405 bch2_moving_ctxt_exit(&ctxt); 406 bch2_move_stats_exit(&move_stats, c); 407 408 return 0; 409 } 410 411 void bch2_copygc_stop(struct bch_fs *c) 412 { 413 if (c->copygc_thread) { 414 kthread_stop(c->copygc_thread); 415 put_task_struct(c->copygc_thread); 416 } 417 c->copygc_thread = NULL; 418 } 419 420 int bch2_copygc_start(struct bch_fs *c) 421 { 422 struct task_struct *t; 423 int ret; 424 425 if (c->copygc_thread) 426 return 0; 427 428 if (c->opts.nochanges) 429 return 0; 430 431 if (bch2_fs_init_fault("copygc_start")) 432 return -ENOMEM; 433 434 t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name); 435 ret = PTR_ERR_OR_ZERO(t); 436 bch_err_msg(c, ret, "creating copygc thread"); 437 if (ret) 438 return ret; 439 440 get_task_struct(t); 441 442 c->copygc_thread = t; 443 wake_up_process(c->copygc_thread); 444 445 return 0; 446 } 447 448 void bch2_fs_copygc_init(struct bch_fs *c) 449 { 450 init_waitqueue_head(&c->copygc_running_wq); 451 c->copygc_running = false; 452 } 453