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