1 // SPDX-License-Identifier: GPL-2.0 2 3 #include "bcachefs.h" 4 #include "alloc_background.h" 5 #include "alloc_foreground.h" 6 #include "btree_iter.h" 7 #include "btree_update.h" 8 #include "btree_write_buffer.h" 9 #include "buckets.h" 10 #include "clock.h" 11 #include "compress.h" 12 #include "disk_groups.h" 13 #include "errcode.h" 14 #include "error.h" 15 #include "inode.h" 16 #include "move.h" 17 #include "rebalance.h" 18 #include "subvolume.h" 19 #include "super-io.h" 20 #include "trace.h" 21 22 #include <linux/freezer.h> 23 #include <linux/kthread.h> 24 #include <linux/sched/cputime.h> 25 26 #define REBALANCE_WORK_SCAN_OFFSET (U64_MAX - 1) 27 28 static const char * const bch2_rebalance_state_strs[] = { 29 #define x(t) #t, 30 BCH_REBALANCE_STATES() 31 NULL 32 #undef x 33 }; 34 35 static int __bch2_set_rebalance_needs_scan(struct btree_trans *trans, u64 inum) 36 { 37 struct btree_iter iter; 38 struct bkey_s_c k; 39 struct bkey_i_cookie *cookie; 40 u64 v; 41 int ret; 42 43 bch2_trans_iter_init(trans, &iter, BTREE_ID_rebalance_work, 44 SPOS(inum, REBALANCE_WORK_SCAN_OFFSET, U32_MAX), 45 BTREE_ITER_INTENT); 46 k = bch2_btree_iter_peek_slot(&iter); 47 ret = bkey_err(k); 48 if (ret) 49 goto err; 50 51 v = k.k->type == KEY_TYPE_cookie 52 ? le64_to_cpu(bkey_s_c_to_cookie(k).v->cookie) 53 : 0; 54 55 cookie = bch2_trans_kmalloc(trans, sizeof(*cookie)); 56 ret = PTR_ERR_OR_ZERO(cookie); 57 if (ret) 58 goto err; 59 60 bkey_cookie_init(&cookie->k_i); 61 cookie->k.p = iter.pos; 62 cookie->v.cookie = cpu_to_le64(v + 1); 63 64 ret = bch2_trans_update(trans, &iter, &cookie->k_i, 0); 65 err: 66 bch2_trans_iter_exit(trans, &iter); 67 return ret; 68 } 69 70 int bch2_set_rebalance_needs_scan(struct bch_fs *c, u64 inum) 71 { 72 int ret = bch2_trans_do(c, NULL, NULL, BCH_TRANS_COMMIT_no_enospc|BCH_TRANS_COMMIT_lazy_rw, 73 __bch2_set_rebalance_needs_scan(trans, inum)); 74 rebalance_wakeup(c); 75 return ret; 76 } 77 78 int bch2_set_fs_needs_rebalance(struct bch_fs *c) 79 { 80 return bch2_set_rebalance_needs_scan(c, 0); 81 } 82 83 static int bch2_clear_rebalance_needs_scan(struct btree_trans *trans, u64 inum, u64 cookie) 84 { 85 struct btree_iter iter; 86 struct bkey_s_c k; 87 u64 v; 88 int ret; 89 90 bch2_trans_iter_init(trans, &iter, BTREE_ID_rebalance_work, 91 SPOS(inum, REBALANCE_WORK_SCAN_OFFSET, U32_MAX), 92 BTREE_ITER_INTENT); 93 k = bch2_btree_iter_peek_slot(&iter); 94 ret = bkey_err(k); 95 if (ret) 96 goto err; 97 98 v = k.k->type == KEY_TYPE_cookie 99 ? le64_to_cpu(bkey_s_c_to_cookie(k).v->cookie) 100 : 0; 101 102 if (v == cookie) 103 ret = bch2_btree_delete_at(trans, &iter, 0); 104 err: 105 bch2_trans_iter_exit(trans, &iter); 106 return ret; 107 } 108 109 static struct bkey_s_c next_rebalance_entry(struct btree_trans *trans, 110 struct btree_iter *work_iter) 111 { 112 return !kthread_should_stop() 113 ? bch2_btree_iter_peek(work_iter) 114 : bkey_s_c_null; 115 } 116 117 static int bch2_bkey_clear_needs_rebalance(struct btree_trans *trans, 118 struct btree_iter *iter, 119 struct bkey_s_c k) 120 { 121 struct bkey_i *n = bch2_bkey_make_mut(trans, iter, &k, 0); 122 int ret = PTR_ERR_OR_ZERO(n); 123 if (ret) 124 return ret; 125 126 extent_entry_drop(bkey_i_to_s(n), 127 (void *) bch2_bkey_rebalance_opts(bkey_i_to_s_c(n))); 128 return bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc); 129 } 130 131 static struct bkey_s_c next_rebalance_extent(struct btree_trans *trans, 132 struct bpos work_pos, 133 struct btree_iter *extent_iter, 134 struct data_update_opts *data_opts) 135 { 136 struct bch_fs *c = trans->c; 137 struct bkey_s_c k; 138 139 bch2_trans_iter_exit(trans, extent_iter); 140 bch2_trans_iter_init(trans, extent_iter, 141 work_pos.inode ? BTREE_ID_extents : BTREE_ID_reflink, 142 work_pos, 143 BTREE_ITER_ALL_SNAPSHOTS); 144 k = bch2_btree_iter_peek_slot(extent_iter); 145 if (bkey_err(k)) 146 return k; 147 148 const struct bch_extent_rebalance *r = k.k ? bch2_bkey_rebalance_opts(k) : NULL; 149 if (!r) { 150 /* raced due to btree write buffer, nothing to do */ 151 return bkey_s_c_null; 152 } 153 154 memset(data_opts, 0, sizeof(*data_opts)); 155 156 data_opts->rewrite_ptrs = 157 bch2_bkey_ptrs_need_rebalance(c, k, r->target, r->compression); 158 data_opts->target = r->target; 159 160 if (!data_opts->rewrite_ptrs) { 161 /* 162 * device we would want to write to offline? devices in target 163 * changed? 164 * 165 * We'll now need a full scan before this extent is picked up 166 * again: 167 */ 168 int ret = bch2_bkey_clear_needs_rebalance(trans, extent_iter, k); 169 if (ret) 170 return bkey_s_c_err(ret); 171 return bkey_s_c_null; 172 } 173 174 if (trace_rebalance_extent_enabled()) { 175 struct printbuf buf = PRINTBUF; 176 177 prt_str(&buf, "target="); 178 bch2_target_to_text(&buf, c, r->target); 179 prt_str(&buf, " compression="); 180 struct bch_compression_opt opt = __bch2_compression_decode(r->compression); 181 prt_str(&buf, bch2_compression_opts[opt.type]); 182 prt_str(&buf, " "); 183 bch2_bkey_val_to_text(&buf, c, k); 184 185 trace_rebalance_extent(c, buf.buf); 186 printbuf_exit(&buf); 187 } 188 189 return k; 190 } 191 192 noinline_for_stack 193 static int do_rebalance_extent(struct moving_context *ctxt, 194 struct bpos work_pos, 195 struct btree_iter *extent_iter) 196 { 197 struct btree_trans *trans = ctxt->trans; 198 struct bch_fs *c = trans->c; 199 struct bch_fs_rebalance *r = &trans->c->rebalance; 200 struct data_update_opts data_opts; 201 struct bch_io_opts io_opts; 202 struct bkey_s_c k; 203 struct bkey_buf sk; 204 int ret; 205 206 ctxt->stats = &r->work_stats; 207 r->state = BCH_REBALANCE_working; 208 209 bch2_bkey_buf_init(&sk); 210 211 ret = bkey_err(k = next_rebalance_extent(trans, work_pos, 212 extent_iter, &data_opts)); 213 if (ret || !k.k) 214 goto out; 215 216 ret = bch2_move_get_io_opts_one(trans, &io_opts, k); 217 if (ret) 218 goto out; 219 220 atomic64_add(k.k->size, &ctxt->stats->sectors_seen); 221 222 /* 223 * The iterator gets unlocked by __bch2_read_extent - need to 224 * save a copy of @k elsewhere: 225 */ 226 bch2_bkey_buf_reassemble(&sk, c, k); 227 k = bkey_i_to_s_c(sk.k); 228 229 ret = bch2_move_extent(ctxt, NULL, extent_iter, k, io_opts, data_opts); 230 if (ret) { 231 if (bch2_err_matches(ret, ENOMEM)) { 232 /* memory allocation failure, wait for some IO to finish */ 233 bch2_move_ctxt_wait_for_io(ctxt); 234 ret = -BCH_ERR_transaction_restart_nested; 235 } 236 237 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) 238 goto out; 239 240 /* skip it and continue, XXX signal failure */ 241 ret = 0; 242 } 243 out: 244 bch2_bkey_buf_exit(&sk, c); 245 return ret; 246 } 247 248 static bool rebalance_pred(struct bch_fs *c, void *arg, 249 struct bkey_s_c k, 250 struct bch_io_opts *io_opts, 251 struct data_update_opts *data_opts) 252 { 253 unsigned target, compression; 254 255 if (k.k->p.inode) { 256 target = io_opts->background_target; 257 compression = io_opts->background_compression ?: io_opts->compression; 258 } else { 259 const struct bch_extent_rebalance *r = bch2_bkey_rebalance_opts(k); 260 261 target = r ? r->target : io_opts->background_target; 262 compression = r ? r->compression : 263 (io_opts->background_compression ?: io_opts->compression); 264 } 265 266 data_opts->rewrite_ptrs = bch2_bkey_ptrs_need_rebalance(c, k, target, compression); 267 data_opts->target = target; 268 return data_opts->rewrite_ptrs != 0; 269 } 270 271 static int do_rebalance_scan(struct moving_context *ctxt, u64 inum, u64 cookie) 272 { 273 struct btree_trans *trans = ctxt->trans; 274 struct bch_fs_rebalance *r = &trans->c->rebalance; 275 int ret; 276 277 bch2_move_stats_init(&r->scan_stats, "rebalance_scan"); 278 ctxt->stats = &r->scan_stats; 279 280 if (!inum) { 281 r->scan_start = BBPOS_MIN; 282 r->scan_end = BBPOS_MAX; 283 } else { 284 r->scan_start = BBPOS(BTREE_ID_extents, POS(inum, 0)); 285 r->scan_end = BBPOS(BTREE_ID_extents, POS(inum, U64_MAX)); 286 } 287 288 r->state = BCH_REBALANCE_scanning; 289 290 ret = __bch2_move_data(ctxt, r->scan_start, r->scan_end, rebalance_pred, NULL) ?: 291 commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc, 292 bch2_clear_rebalance_needs_scan(trans, inum, cookie)); 293 294 bch2_move_stats_exit(&r->scan_stats, trans->c); 295 return ret; 296 } 297 298 static void rebalance_wait(struct bch_fs *c) 299 { 300 struct bch_fs_rebalance *r = &c->rebalance; 301 struct io_clock *clock = &c->io_clock[WRITE]; 302 u64 now = atomic64_read(&clock->now); 303 u64 min_member_capacity = bch2_min_rw_member_capacity(c); 304 305 if (min_member_capacity == U64_MAX) 306 min_member_capacity = 128 * 2048; 307 308 r->wait_iotime_end = now + (min_member_capacity >> 6); 309 310 if (r->state != BCH_REBALANCE_waiting) { 311 r->wait_iotime_start = now; 312 r->wait_wallclock_start = ktime_get_real_ns(); 313 r->state = BCH_REBALANCE_waiting; 314 } 315 316 bch2_kthread_io_clock_wait(clock, r->wait_iotime_end, MAX_SCHEDULE_TIMEOUT); 317 } 318 319 static int do_rebalance(struct moving_context *ctxt) 320 { 321 struct btree_trans *trans = ctxt->trans; 322 struct bch_fs *c = trans->c; 323 struct bch_fs_rebalance *r = &c->rebalance; 324 struct btree_iter rebalance_work_iter, extent_iter = { NULL }; 325 struct bkey_s_c k; 326 int ret = 0; 327 328 bch2_move_stats_init(&r->work_stats, "rebalance_work"); 329 bch2_move_stats_init(&r->scan_stats, "rebalance_scan"); 330 331 bch2_trans_iter_init(trans, &rebalance_work_iter, 332 BTREE_ID_rebalance_work, POS_MIN, 333 BTREE_ITER_ALL_SNAPSHOTS); 334 335 while (!bch2_move_ratelimit(ctxt)) { 336 if (!r->enabled) { 337 bch2_moving_ctxt_flush_all(ctxt); 338 kthread_wait_freezable(r->enabled || 339 kthread_should_stop()); 340 } 341 342 if (kthread_should_stop()) 343 break; 344 345 bch2_trans_begin(trans); 346 347 ret = bkey_err(k = next_rebalance_entry(trans, &rebalance_work_iter)); 348 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) 349 continue; 350 if (ret || !k.k) 351 break; 352 353 ret = k.k->type == KEY_TYPE_cookie 354 ? do_rebalance_scan(ctxt, k.k->p.inode, 355 le64_to_cpu(bkey_s_c_to_cookie(k).v->cookie)) 356 : do_rebalance_extent(ctxt, k.k->p, &extent_iter); 357 358 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) 359 continue; 360 if (ret) 361 break; 362 363 bch2_btree_iter_advance(&rebalance_work_iter); 364 } 365 366 bch2_trans_iter_exit(trans, &extent_iter); 367 bch2_trans_iter_exit(trans, &rebalance_work_iter); 368 bch2_move_stats_exit(&r->scan_stats, c); 369 370 if (!ret && 371 !kthread_should_stop() && 372 !atomic64_read(&r->work_stats.sectors_seen) && 373 !atomic64_read(&r->scan_stats.sectors_seen)) { 374 bch2_trans_unlock_long(trans); 375 rebalance_wait(c); 376 } 377 378 if (!bch2_err_matches(ret, EROFS)) 379 bch_err_fn(c, ret); 380 return ret; 381 } 382 383 static int bch2_rebalance_thread(void *arg) 384 { 385 struct bch_fs *c = arg; 386 struct bch_fs_rebalance *r = &c->rebalance; 387 struct moving_context ctxt; 388 int ret; 389 390 set_freezable(); 391 392 bch2_moving_ctxt_init(&ctxt, c, NULL, &r->work_stats, 393 writepoint_ptr(&c->rebalance_write_point), 394 true); 395 396 while (!kthread_should_stop() && 397 !(ret = do_rebalance(&ctxt))) 398 ; 399 400 bch2_moving_ctxt_exit(&ctxt); 401 402 return 0; 403 } 404 405 void bch2_rebalance_status_to_text(struct printbuf *out, struct bch_fs *c) 406 { 407 struct bch_fs_rebalance *r = &c->rebalance; 408 409 prt_str(out, bch2_rebalance_state_strs[r->state]); 410 prt_newline(out); 411 printbuf_indent_add(out, 2); 412 413 switch (r->state) { 414 case BCH_REBALANCE_waiting: { 415 u64 now = atomic64_read(&c->io_clock[WRITE].now); 416 417 prt_str(out, "io wait duration: "); 418 bch2_prt_human_readable_s64(out, r->wait_iotime_end - r->wait_iotime_start); 419 prt_newline(out); 420 421 prt_str(out, "io wait remaining: "); 422 bch2_prt_human_readable_s64(out, r->wait_iotime_end - now); 423 prt_newline(out); 424 425 prt_str(out, "duration waited: "); 426 bch2_pr_time_units(out, ktime_get_real_ns() - r->wait_wallclock_start); 427 prt_newline(out); 428 break; 429 } 430 case BCH_REBALANCE_working: 431 bch2_move_stats_to_text(out, &r->work_stats); 432 break; 433 case BCH_REBALANCE_scanning: 434 bch2_move_stats_to_text(out, &r->scan_stats); 435 break; 436 } 437 prt_newline(out); 438 printbuf_indent_sub(out, 2); 439 } 440 441 void bch2_rebalance_stop(struct bch_fs *c) 442 { 443 struct task_struct *p; 444 445 c->rebalance.pd.rate.rate = UINT_MAX; 446 bch2_ratelimit_reset(&c->rebalance.pd.rate); 447 448 p = rcu_dereference_protected(c->rebalance.thread, 1); 449 c->rebalance.thread = NULL; 450 451 if (p) { 452 /* for sychronizing with rebalance_wakeup() */ 453 synchronize_rcu(); 454 455 kthread_stop(p); 456 put_task_struct(p); 457 } 458 } 459 460 int bch2_rebalance_start(struct bch_fs *c) 461 { 462 struct task_struct *p; 463 int ret; 464 465 if (c->rebalance.thread) 466 return 0; 467 468 if (c->opts.nochanges) 469 return 0; 470 471 p = kthread_create(bch2_rebalance_thread, c, "bch-rebalance/%s", c->name); 472 ret = PTR_ERR_OR_ZERO(p); 473 bch_err_msg(c, ret, "creating rebalance thread"); 474 if (ret) 475 return ret; 476 477 get_task_struct(p); 478 rcu_assign_pointer(c->rebalance.thread, p); 479 wake_up_process(p); 480 return 0; 481 } 482 483 void bch2_fs_rebalance_init(struct bch_fs *c) 484 { 485 bch2_pd_controller_init(&c->rebalance.pd); 486 } 487