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_upto(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 moved = atomic64_read(&ctxt->stats->sectors_moved); 219 int ret = 0; 220 221 ret = bch2_copygc_get_buckets(ctxt, buckets_in_flight, &buckets); 222 if (ret) 223 goto err; 224 225 darray_for_each(buckets, i) { 226 if (kthread_should_stop() || freezing(current)) 227 break; 228 229 f = move_bucket_in_flight_add(buckets_in_flight, *i); 230 ret = PTR_ERR_OR_ZERO(f); 231 if (ret == -EEXIST) { /* rare race: copygc_get_buckets returned same bucket more than once */ 232 ret = 0; 233 continue; 234 } 235 if (ret == -ENOMEM) { /* flush IO, continue later */ 236 ret = 0; 237 break; 238 } 239 240 ret = bch2_evacuate_bucket(ctxt, f, f->bucket.k.bucket, 241 f->bucket.k.gen, data_opts); 242 if (ret) 243 goto err; 244 245 *did_work = true; 246 } 247 err: 248 darray_exit(&buckets); 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 moved = atomic64_read(&ctxt->stats->sectors_moved) - moved; 258 trace_and_count(c, copygc, c, moved, 0, 0, 0); 259 return ret; 260 } 261 262 /* 263 * Copygc runs when the amount of fragmented data is above some arbitrary 264 * threshold: 265 * 266 * The threshold at the limit - when the device is full - is the amount of space 267 * we reserved in bch2_recalc_capacity; we can't have more than that amount of 268 * disk space stranded due to fragmentation and store everything we have 269 * promised to store. 270 * 271 * But we don't want to be running copygc unnecessarily when the device still 272 * has plenty of free space - rather, we want copygc to smoothly run every so 273 * often and continually reduce the amount of fragmented space as the device 274 * fills up. So, we increase the threshold by half the current free space. 275 */ 276 unsigned long bch2_copygc_wait_amount(struct bch_fs *c) 277 { 278 s64 wait = S64_MAX, fragmented_allowed, fragmented; 279 280 for_each_rw_member(c, ca) { 281 struct bch_dev_usage usage = bch2_dev_usage_read(ca); 282 283 fragmented_allowed = ((__dev_buckets_available(ca, usage, BCH_WATERMARK_stripe) * 284 ca->mi.bucket_size) >> 1); 285 fragmented = 0; 286 287 for (unsigned i = 0; i < BCH_DATA_NR; i++) 288 if (data_type_movable(i)) 289 fragmented += usage.d[i].fragmented; 290 291 wait = min(wait, max(0LL, fragmented_allowed - fragmented)); 292 } 293 294 return wait; 295 } 296 297 void bch2_copygc_wait_to_text(struct printbuf *out, struct bch_fs *c) 298 { 299 printbuf_tabstop_push(out, 32); 300 prt_printf(out, "running:\t%u\n", c->copygc_running); 301 prt_printf(out, "copygc_wait:\t%llu\n", c->copygc_wait); 302 prt_printf(out, "copygc_wait_at:\t%llu\n", c->copygc_wait_at); 303 304 prt_printf(out, "Currently waiting for:\t"); 305 prt_human_readable_u64(out, max(0LL, c->copygc_wait - 306 atomic64_read(&c->io_clock[WRITE].now)) << 9); 307 prt_newline(out); 308 309 prt_printf(out, "Currently waiting since:\t"); 310 prt_human_readable_u64(out, max(0LL, 311 atomic64_read(&c->io_clock[WRITE].now) - 312 c->copygc_wait_at) << 9); 313 prt_newline(out); 314 315 prt_printf(out, "Currently calculated wait:\t"); 316 prt_human_readable_u64(out, bch2_copygc_wait_amount(c)); 317 prt_newline(out); 318 } 319 320 static int bch2_copygc_thread(void *arg) 321 { 322 struct bch_fs *c = arg; 323 struct moving_context ctxt; 324 struct bch_move_stats move_stats; 325 struct io_clock *clock = &c->io_clock[WRITE]; 326 struct buckets_in_flight *buckets; 327 u64 last, wait; 328 int ret = 0; 329 330 buckets = kzalloc(sizeof(struct buckets_in_flight), GFP_KERNEL); 331 if (!buckets) 332 return -ENOMEM; 333 ret = rhashtable_init(&buckets->table, &bch_move_bucket_params); 334 bch_err_msg(c, ret, "allocating copygc buckets in flight"); 335 if (ret) { 336 kfree(buckets); 337 return ret; 338 } 339 340 set_freezable(); 341 342 bch2_move_stats_init(&move_stats, "copygc"); 343 bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats, 344 writepoint_ptr(&c->copygc_write_point), 345 false); 346 347 while (!ret && !kthread_should_stop()) { 348 bool did_work = false; 349 350 bch2_trans_unlock_long(ctxt.trans); 351 cond_resched(); 352 353 if (!c->copy_gc_enabled) { 354 move_buckets_wait(&ctxt, buckets, true); 355 kthread_wait_freezable(c->copy_gc_enabled || 356 kthread_should_stop()); 357 } 358 359 if (unlikely(freezing(current))) { 360 move_buckets_wait(&ctxt, buckets, true); 361 __refrigerator(false); 362 continue; 363 } 364 365 last = atomic64_read(&clock->now); 366 wait = bch2_copygc_wait_amount(c); 367 368 if (wait > clock->max_slop) { 369 c->copygc_wait_at = last; 370 c->copygc_wait = last + wait; 371 move_buckets_wait(&ctxt, buckets, true); 372 trace_and_count(c, copygc_wait, c, wait, last + wait); 373 bch2_kthread_io_clock_wait(clock, last + wait, 374 MAX_SCHEDULE_TIMEOUT); 375 continue; 376 } 377 378 c->copygc_wait = 0; 379 380 c->copygc_running = true; 381 ret = bch2_copygc(&ctxt, buckets, &did_work); 382 c->copygc_running = false; 383 384 wake_up(&c->copygc_running_wq); 385 386 if (!wait && !did_work) { 387 u64 min_member_capacity = bch2_min_rw_member_capacity(c); 388 389 if (min_member_capacity == U64_MAX) 390 min_member_capacity = 128 * 2048; 391 392 move_buckets_wait(&ctxt, buckets, true); 393 bch2_kthread_io_clock_wait(clock, last + (min_member_capacity >> 6), 394 MAX_SCHEDULE_TIMEOUT); 395 } 396 } 397 398 move_buckets_wait(&ctxt, buckets, true); 399 400 rhashtable_destroy(&buckets->table); 401 kfree(buckets); 402 bch2_moving_ctxt_exit(&ctxt); 403 bch2_move_stats_exit(&move_stats, c); 404 405 return 0; 406 } 407 408 void bch2_copygc_stop(struct bch_fs *c) 409 { 410 if (c->copygc_thread) { 411 kthread_stop(c->copygc_thread); 412 put_task_struct(c->copygc_thread); 413 } 414 c->copygc_thread = NULL; 415 } 416 417 int bch2_copygc_start(struct bch_fs *c) 418 { 419 struct task_struct *t; 420 int ret; 421 422 if (c->copygc_thread) 423 return 0; 424 425 if (c->opts.nochanges) 426 return 0; 427 428 if (bch2_fs_init_fault("copygc_start")) 429 return -ENOMEM; 430 431 t = kthread_create(bch2_copygc_thread, c, "bch-copygc/%s", c->name); 432 ret = PTR_ERR_OR_ZERO(t); 433 bch_err_msg(c, ret, "creating copygc thread"); 434 if (ret) 435 return ret; 436 437 get_task_struct(t); 438 439 c->copygc_thread = t; 440 wake_up_process(c->copygc_thread); 441 442 return 0; 443 } 444 445 void bch2_fs_copygc_init(struct bch_fs *c) 446 { 447 init_waitqueue_head(&c->copygc_running_wq); 448 c->copygc_running = false; 449 } 450