1 // SPDX-License-Identifier: GPL-2.0 2 3 #include "bcachefs.h" 4 #include "alloc_foreground.h" 5 #include "bkey_buf.h" 6 #include "btree_update.h" 7 #include "buckets.h" 8 #include "data_update.h" 9 #include "ec.h" 10 #include "error.h" 11 #include "extents.h" 12 #include "io_write.h" 13 #include "keylist.h" 14 #include "move.h" 15 #include "nocow_locking.h" 16 #include "subvolume.h" 17 #include "trace.h" 18 19 static void trace_move_extent_finish2(struct bch_fs *c, struct bkey_s_c k) 20 { 21 if (trace_move_extent_finish_enabled()) { 22 struct printbuf buf = PRINTBUF; 23 24 bch2_bkey_val_to_text(&buf, c, k); 25 trace_move_extent_finish(c, buf.buf); 26 printbuf_exit(&buf); 27 } 28 } 29 30 static void trace_move_extent_fail2(struct data_update *m, 31 struct bkey_s_c new, 32 struct bkey_s_c wrote, 33 struct bkey_i *insert, 34 const char *msg) 35 { 36 struct bch_fs *c = m->op.c; 37 struct bkey_s_c old = bkey_i_to_s_c(m->k.k); 38 const union bch_extent_entry *entry; 39 struct bch_extent_ptr *ptr; 40 struct extent_ptr_decoded p; 41 struct printbuf buf = PRINTBUF; 42 unsigned i, rewrites_found = 0; 43 44 if (!trace_move_extent_fail_enabled()) 45 return; 46 47 prt_str(&buf, msg); 48 49 if (insert) { 50 i = 0; 51 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) { 52 if (((1U << i) & m->data_opts.rewrite_ptrs) && 53 (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) && 54 !ptr->cached) 55 rewrites_found |= 1U << i; 56 i++; 57 } 58 } 59 60 prt_printf(&buf, "\nrewrite ptrs: %u%u%u%u", 61 (m->data_opts.rewrite_ptrs & (1 << 0)) != 0, 62 (m->data_opts.rewrite_ptrs & (1 << 1)) != 0, 63 (m->data_opts.rewrite_ptrs & (1 << 2)) != 0, 64 (m->data_opts.rewrite_ptrs & (1 << 3)) != 0); 65 66 prt_printf(&buf, "\nrewrites found: %u%u%u%u", 67 (rewrites_found & (1 << 0)) != 0, 68 (rewrites_found & (1 << 1)) != 0, 69 (rewrites_found & (1 << 2)) != 0, 70 (rewrites_found & (1 << 3)) != 0); 71 72 prt_str(&buf, "\nold: "); 73 bch2_bkey_val_to_text(&buf, c, old); 74 75 prt_str(&buf, "\nnew: "); 76 bch2_bkey_val_to_text(&buf, c, new); 77 78 prt_str(&buf, "\nwrote: "); 79 bch2_bkey_val_to_text(&buf, c, wrote); 80 81 if (insert) { 82 prt_str(&buf, "\ninsert: "); 83 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert)); 84 } 85 86 trace_move_extent_fail(c, buf.buf); 87 printbuf_exit(&buf); 88 } 89 90 static int __bch2_data_update_index_update(struct btree_trans *trans, 91 struct bch_write_op *op) 92 { 93 struct bch_fs *c = op->c; 94 struct btree_iter iter; 95 struct data_update *m = 96 container_of(op, struct data_update, op); 97 struct keylist *keys = &op->insert_keys; 98 struct bkey_buf _new, _insert; 99 int ret = 0; 100 101 bch2_bkey_buf_init(&_new); 102 bch2_bkey_buf_init(&_insert); 103 bch2_bkey_buf_realloc(&_insert, c, U8_MAX); 104 105 bch2_trans_iter_init(trans, &iter, m->btree_id, 106 bkey_start_pos(&bch2_keylist_front(keys)->k), 107 BTREE_ITER_SLOTS|BTREE_ITER_INTENT); 108 109 while (1) { 110 struct bkey_s_c k; 111 struct bkey_s_c old = bkey_i_to_s_c(m->k.k); 112 struct bkey_i *insert = NULL; 113 struct bkey_i_extent *new; 114 const union bch_extent_entry *entry_c; 115 union bch_extent_entry *entry; 116 struct extent_ptr_decoded p; 117 struct bch_extent_ptr *ptr; 118 const struct bch_extent_ptr *ptr_c; 119 struct bpos next_pos; 120 bool should_check_enospc; 121 s64 i_sectors_delta = 0, disk_sectors_delta = 0; 122 unsigned rewrites_found = 0, durability, i; 123 124 bch2_trans_begin(trans); 125 126 k = bch2_btree_iter_peek_slot(&iter); 127 ret = bkey_err(k); 128 if (ret) 129 goto err; 130 131 new = bkey_i_to_extent(bch2_keylist_front(keys)); 132 133 if (!bch2_extents_match(k, old)) { 134 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), 135 NULL, "no match:"); 136 goto nowork; 137 } 138 139 bkey_reassemble(_insert.k, k); 140 insert = _insert.k; 141 142 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys)); 143 new = bkey_i_to_extent(_new.k); 144 bch2_cut_front(iter.pos, &new->k_i); 145 146 bch2_cut_front(iter.pos, insert); 147 bch2_cut_back(new->k.p, insert); 148 bch2_cut_back(insert->k.p, &new->k_i); 149 150 /* 151 * @old: extent that we read from 152 * @insert: key that we're going to update, initialized from 153 * extent currently in btree - same as @old unless we raced with 154 * other updates 155 * @new: extent with new pointers that we'll be adding to @insert 156 * 157 * Fist, drop rewrite_ptrs from @new: 158 */ 159 i = 0; 160 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry_c) { 161 if (((1U << i) & m->data_opts.rewrite_ptrs) && 162 (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) && 163 !ptr->cached) { 164 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr); 165 /* 166 * See comment below: 167 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), ptr); 168 */ 169 rewrites_found |= 1U << i; 170 } 171 i++; 172 } 173 174 if (m->data_opts.rewrite_ptrs && 175 !rewrites_found && 176 bch2_bkey_durability(c, k) >= m->op.opts.data_replicas) { 177 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "no rewrites found:"); 178 goto nowork; 179 } 180 181 /* 182 * A replica that we just wrote might conflict with a replica 183 * that we want to keep, due to racing with another move: 184 */ 185 restart_drop_conflicting_replicas: 186 extent_for_each_ptr(extent_i_to_s(new), ptr) 187 if ((ptr_c = bch2_bkey_has_device_c(bkey_i_to_s_c(insert), ptr->dev)) && 188 !ptr_c->cached) { 189 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(&new->k_i), ptr); 190 goto restart_drop_conflicting_replicas; 191 } 192 193 if (!bkey_val_u64s(&new->k)) { 194 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "new replicas conflicted:"); 195 goto nowork; 196 } 197 198 /* Now, drop pointers that conflict with what we just wrote: */ 199 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) 200 if ((ptr = bch2_bkey_has_device(bkey_i_to_s(insert), p.ptr.dev))) 201 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr); 202 203 durability = bch2_bkey_durability(c, bkey_i_to_s_c(insert)) + 204 bch2_bkey_durability(c, bkey_i_to_s_c(&new->k_i)); 205 206 /* Now, drop excess replicas: */ 207 restart_drop_extra_replicas: 208 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs(bkey_i_to_s(insert)), p, entry) { 209 unsigned ptr_durability = bch2_extent_ptr_durability(c, &p); 210 211 if (!p.ptr.cached && 212 durability - ptr_durability >= m->op.opts.data_replicas) { 213 durability -= ptr_durability; 214 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), &entry->ptr); 215 /* 216 * Currently, we're dropping unneeded replicas 217 * instead of marking them as cached, since 218 * cached data in stripe buckets prevents them 219 * from being reused: 220 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), &entry->ptr); 221 */ 222 goto restart_drop_extra_replicas; 223 } 224 } 225 226 /* Finally, add the pointers we just wrote: */ 227 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) 228 bch2_extent_ptr_decoded_append(insert, &p); 229 230 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 }); 231 bch2_extent_normalize(c, bkey_i_to_s(insert)); 232 233 ret = bch2_sum_sector_overwrites(trans, &iter, insert, 234 &should_check_enospc, 235 &i_sectors_delta, 236 &disk_sectors_delta); 237 if (ret) 238 goto err; 239 240 if (disk_sectors_delta > (s64) op->res.sectors) { 241 ret = bch2_disk_reservation_add(c, &op->res, 242 disk_sectors_delta - op->res.sectors, 243 !should_check_enospc 244 ? BCH_DISK_RESERVATION_NOFAIL : 0); 245 if (ret) 246 goto out; 247 } 248 249 next_pos = insert->k.p; 250 251 ret = bch2_insert_snapshot_whiteouts(trans, m->btree_id, 252 k.k->p, bkey_start_pos(&insert->k)) ?: 253 bch2_insert_snapshot_whiteouts(trans, m->btree_id, 254 k.k->p, insert->k.p); 255 if (ret) 256 goto err; 257 258 ret = bch2_trans_update(trans, &iter, insert, 259 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?: 260 bch2_trans_commit(trans, &op->res, 261 NULL, 262 BTREE_INSERT_NOCHECK_RW| 263 BTREE_INSERT_NOFAIL| 264 m->data_opts.btree_insert_flags); 265 if (!ret) { 266 bch2_btree_iter_set_pos(&iter, next_pos); 267 268 this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size); 269 trace_move_extent_finish2(c, bkey_i_to_s_c(&new->k_i)); 270 } 271 err: 272 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) 273 ret = 0; 274 if (ret) 275 break; 276 next: 277 while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) { 278 bch2_keylist_pop_front(keys); 279 if (bch2_keylist_empty(keys)) 280 goto out; 281 } 282 continue; 283 nowork: 284 if (m->ctxt && m->ctxt->stats) { 285 BUG_ON(k.k->p.offset <= iter.pos.offset); 286 atomic64_inc(&m->ctxt->stats->keys_raced); 287 atomic64_add(k.k->p.offset - iter.pos.offset, 288 &m->ctxt->stats->sectors_raced); 289 } 290 291 this_cpu_inc(c->counters[BCH_COUNTER_move_extent_fail]); 292 293 bch2_btree_iter_advance(&iter); 294 goto next; 295 } 296 out: 297 bch2_trans_iter_exit(trans, &iter); 298 bch2_bkey_buf_exit(&_insert, c); 299 bch2_bkey_buf_exit(&_new, c); 300 BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart)); 301 return ret; 302 } 303 304 int bch2_data_update_index_update(struct bch_write_op *op) 305 { 306 return bch2_trans_run(op->c, __bch2_data_update_index_update(trans, op)); 307 } 308 309 void bch2_data_update_read_done(struct data_update *m, 310 struct bch_extent_crc_unpacked crc) 311 { 312 /* write bio must own pages: */ 313 BUG_ON(!m->op.wbio.bio.bi_vcnt); 314 315 m->op.crc = crc; 316 m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9; 317 318 closure_call(&m->op.cl, bch2_write, NULL, NULL); 319 } 320 321 void bch2_data_update_exit(struct data_update *update) 322 { 323 struct bch_fs *c = update->op.c; 324 struct bkey_ptrs_c ptrs = 325 bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k)); 326 const struct bch_extent_ptr *ptr; 327 328 bkey_for_each_ptr(ptrs, ptr) { 329 if (c->opts.nocow_enabled) 330 bch2_bucket_nocow_unlock(&c->nocow_locks, 331 PTR_BUCKET_POS(c, ptr), 0); 332 percpu_ref_put(&bch_dev_bkey_exists(c, ptr->dev)->ref); 333 } 334 335 bch2_bkey_buf_exit(&update->k, c); 336 bch2_disk_reservation_put(c, &update->op.res); 337 bch2_bio_free_pages_pool(c, &update->op.wbio.bio); 338 } 339 340 void bch2_update_unwritten_extent(struct btree_trans *trans, 341 struct data_update *update) 342 { 343 struct bch_fs *c = update->op.c; 344 struct bio *bio = &update->op.wbio.bio; 345 struct bkey_i_extent *e; 346 struct write_point *wp; 347 struct bch_extent_ptr *ptr; 348 struct closure cl; 349 struct btree_iter iter; 350 struct bkey_s_c k; 351 int ret; 352 353 closure_init_stack(&cl); 354 bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys); 355 356 while (bio_sectors(bio)) { 357 unsigned sectors = bio_sectors(bio); 358 359 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos, 360 BTREE_ITER_SLOTS); 361 ret = lockrestart_do(trans, ({ 362 k = bch2_btree_iter_peek_slot(&iter); 363 bkey_err(k); 364 })); 365 bch2_trans_iter_exit(trans, &iter); 366 367 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k))) 368 break; 369 370 e = bkey_extent_init(update->op.insert_keys.top); 371 e->k.p = update->op.pos; 372 373 ret = bch2_alloc_sectors_start_trans(trans, 374 update->op.target, 375 false, 376 update->op.write_point, 377 &update->op.devs_have, 378 update->op.nr_replicas, 379 update->op.nr_replicas, 380 update->op.watermark, 381 0, &cl, &wp); 382 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) { 383 bch2_trans_unlock(trans); 384 closure_sync(&cl); 385 continue; 386 } 387 388 if (ret) 389 return; 390 391 sectors = min(sectors, wp->sectors_free); 392 393 bch2_key_resize(&e->k, sectors); 394 395 bch2_open_bucket_get(c, wp, &update->op.open_buckets); 396 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false); 397 bch2_alloc_sectors_done(c, wp); 398 399 bio_advance(bio, sectors << 9); 400 update->op.pos.offset += sectors; 401 402 extent_for_each_ptr(extent_i_to_s(e), ptr) 403 ptr->unwritten = true; 404 bch2_keylist_push(&update->op.insert_keys); 405 406 ret = __bch2_data_update_index_update(trans, &update->op); 407 408 bch2_open_buckets_put(c, &update->op.open_buckets); 409 410 if (ret) 411 break; 412 } 413 414 if (closure_nr_remaining(&cl) != 1) { 415 bch2_trans_unlock(trans); 416 closure_sync(&cl); 417 } 418 } 419 420 int bch2_data_update_init(struct btree_trans *trans, 421 struct moving_context *ctxt, 422 struct data_update *m, 423 struct write_point_specifier wp, 424 struct bch_io_opts io_opts, 425 struct data_update_opts data_opts, 426 enum btree_id btree_id, 427 struct bkey_s_c k) 428 { 429 struct bch_fs *c = trans->c; 430 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k); 431 const union bch_extent_entry *entry; 432 struct extent_ptr_decoded p; 433 const struct bch_extent_ptr *ptr; 434 unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas; 435 unsigned ptrs_locked = 0; 436 int ret; 437 438 bch2_bkey_buf_init(&m->k); 439 bch2_bkey_buf_reassemble(&m->k, c, k); 440 m->btree_id = btree_id; 441 m->data_opts = data_opts; 442 443 bch2_write_op_init(&m->op, c, io_opts); 444 m->op.pos = bkey_start_pos(k.k); 445 m->op.version = k.k->version; 446 m->op.target = data_opts.target; 447 m->op.write_point = wp; 448 m->op.nr_replicas = 0; 449 m->op.flags |= BCH_WRITE_PAGES_STABLE| 450 BCH_WRITE_PAGES_OWNED| 451 BCH_WRITE_DATA_ENCODED| 452 BCH_WRITE_MOVE| 453 m->data_opts.write_flags; 454 m->op.compression_opt = io_opts.background_compression ?: io_opts.compression; 455 m->op.watermark = m->data_opts.btree_insert_flags & BCH_WATERMARK_MASK; 456 457 bkey_for_each_ptr(ptrs, ptr) 458 percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref); 459 460 i = 0; 461 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) { 462 bool locked; 463 464 if (((1U << i) & m->data_opts.rewrite_ptrs)) { 465 BUG_ON(p.ptr.cached); 466 467 if (crc_is_compressed(p.crc)) 468 reserve_sectors += k.k->size; 469 470 m->op.nr_replicas += bch2_extent_ptr_desired_durability(c, &p); 471 } else if (!p.ptr.cached) { 472 bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev); 473 } 474 475 /* 476 * op->csum_type is normally initialized from the fs/file's 477 * current options - but if an extent is encrypted, we require 478 * that it stays encrypted: 479 */ 480 if (bch2_csum_type_is_encryption(p.crc.csum_type)) { 481 m->op.nonce = p.crc.nonce + p.crc.offset; 482 m->op.csum_type = p.crc.csum_type; 483 } 484 485 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible) 486 m->op.incompressible = true; 487 488 if (c->opts.nocow_enabled) { 489 if (ctxt) { 490 move_ctxt_wait_event(ctxt, trans, 491 (locked = bch2_bucket_nocow_trylock(&c->nocow_locks, 492 PTR_BUCKET_POS(c, &p.ptr), 0)) || 493 !atomic_read(&ctxt->read_sectors)); 494 495 if (!locked) 496 bch2_bucket_nocow_lock(&c->nocow_locks, 497 PTR_BUCKET_POS(c, &p.ptr), 0); 498 } else { 499 if (!bch2_bucket_nocow_trylock(&c->nocow_locks, 500 PTR_BUCKET_POS(c, &p.ptr), 0)) { 501 ret = -BCH_ERR_nocow_lock_blocked; 502 goto err; 503 } 504 } 505 ptrs_locked |= (1U << i); 506 } 507 508 i++; 509 } 510 511 if (reserve_sectors) { 512 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors, 513 m->data_opts.extra_replicas 514 ? 0 515 : BCH_DISK_RESERVATION_NOFAIL); 516 if (ret) 517 goto err; 518 } 519 520 m->op.nr_replicas += m->data_opts.extra_replicas; 521 m->op.nr_replicas_required = m->op.nr_replicas; 522 523 BUG_ON(!m->op.nr_replicas); 524 525 /* Special handling required: */ 526 if (bkey_extent_is_unwritten(k)) 527 return -BCH_ERR_unwritten_extent_update; 528 return 0; 529 err: 530 i = 0; 531 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) { 532 if ((1U << i) & ptrs_locked) 533 bch2_bucket_nocow_unlock(&c->nocow_locks, 534 PTR_BUCKET_POS(c, &p.ptr), 0); 535 percpu_ref_put(&bch_dev_bkey_exists(c, p.ptr.dev)->ref); 536 i++; 537 } 538 539 bch2_bkey_buf_exit(&m->k, c); 540 bch2_bio_free_pages_pool(c, &m->op.wbio.bio); 541 return ret; 542 } 543 544 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts) 545 { 546 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k); 547 const struct bch_extent_ptr *ptr; 548 unsigned i = 0; 549 550 bkey_for_each_ptr(ptrs, ptr) { 551 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) { 552 opts->kill_ptrs |= 1U << i; 553 opts->rewrite_ptrs ^= 1U << i; 554 } 555 556 i++; 557 } 558 } 559