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 "rebalance.h" 17 #include "subvolume.h" 18 #include "trace.h" 19 20 static void trace_move_extent_finish2(struct bch_fs *c, struct bkey_s_c k) 21 { 22 if (trace_move_extent_finish_enabled()) { 23 struct printbuf buf = PRINTBUF; 24 25 bch2_bkey_val_to_text(&buf, c, k); 26 trace_move_extent_finish(c, buf.buf); 27 printbuf_exit(&buf); 28 } 29 } 30 31 static void trace_move_extent_fail2(struct data_update *m, 32 struct bkey_s_c new, 33 struct bkey_s_c wrote, 34 struct bkey_i *insert, 35 const char *msg) 36 { 37 struct bch_fs *c = m->op.c; 38 struct bkey_s_c old = bkey_i_to_s_c(m->k.k); 39 const union bch_extent_entry *entry; 40 struct bch_extent_ptr *ptr; 41 struct extent_ptr_decoded p; 42 struct printbuf buf = PRINTBUF; 43 unsigned i, rewrites_found = 0; 44 45 if (!trace_move_extent_fail_enabled()) 46 return; 47 48 prt_str(&buf, msg); 49 50 if (insert) { 51 i = 0; 52 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) { 53 if (((1U << i) & m->data_opts.rewrite_ptrs) && 54 (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) && 55 !ptr->cached) 56 rewrites_found |= 1U << i; 57 i++; 58 } 59 } 60 61 prt_printf(&buf, "\nrewrite ptrs: %u%u%u%u", 62 (m->data_opts.rewrite_ptrs & (1 << 0)) != 0, 63 (m->data_opts.rewrite_ptrs & (1 << 1)) != 0, 64 (m->data_opts.rewrite_ptrs & (1 << 2)) != 0, 65 (m->data_opts.rewrite_ptrs & (1 << 3)) != 0); 66 67 prt_printf(&buf, "\nrewrites found: %u%u%u%u", 68 (rewrites_found & (1 << 0)) != 0, 69 (rewrites_found & (1 << 1)) != 0, 70 (rewrites_found & (1 << 2)) != 0, 71 (rewrites_found & (1 << 3)) != 0); 72 73 prt_str(&buf, "\nold: "); 74 bch2_bkey_val_to_text(&buf, c, old); 75 76 prt_str(&buf, "\nnew: "); 77 bch2_bkey_val_to_text(&buf, c, new); 78 79 prt_str(&buf, "\nwrote: "); 80 bch2_bkey_val_to_text(&buf, c, wrote); 81 82 if (insert) { 83 prt_str(&buf, "\ninsert: "); 84 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert)); 85 } 86 87 trace_move_extent_fail(c, buf.buf); 88 printbuf_exit(&buf); 89 } 90 91 static int __bch2_data_update_index_update(struct btree_trans *trans, 92 struct bch_write_op *op) 93 { 94 struct bch_fs *c = op->c; 95 struct btree_iter iter; 96 struct data_update *m = 97 container_of(op, struct data_update, op); 98 struct keylist *keys = &op->insert_keys; 99 struct bkey_buf _new, _insert; 100 int ret = 0; 101 102 bch2_bkey_buf_init(&_new); 103 bch2_bkey_buf_init(&_insert); 104 bch2_bkey_buf_realloc(&_insert, c, U8_MAX); 105 106 bch2_trans_iter_init(trans, &iter, m->btree_id, 107 bkey_start_pos(&bch2_keylist_front(keys)->k), 108 BTREE_ITER_SLOTS|BTREE_ITER_INTENT); 109 110 while (1) { 111 struct bkey_s_c k; 112 struct bkey_s_c old = bkey_i_to_s_c(m->k.k); 113 struct bkey_i *insert = NULL; 114 struct bkey_i_extent *new; 115 const union bch_extent_entry *entry_c; 116 union bch_extent_entry *entry; 117 struct extent_ptr_decoded p; 118 struct bch_extent_ptr *ptr; 119 const struct bch_extent_ptr *ptr_c; 120 struct bpos next_pos; 121 bool should_check_enospc; 122 s64 i_sectors_delta = 0, disk_sectors_delta = 0; 123 unsigned rewrites_found = 0, durability, i; 124 125 bch2_trans_begin(trans); 126 127 k = bch2_btree_iter_peek_slot(&iter); 128 ret = bkey_err(k); 129 if (ret) 130 goto err; 131 132 new = bkey_i_to_extent(bch2_keylist_front(keys)); 133 134 if (!bch2_extents_match(k, old)) { 135 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), 136 NULL, "no match:"); 137 goto nowork; 138 } 139 140 bkey_reassemble(_insert.k, k); 141 insert = _insert.k; 142 143 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys)); 144 new = bkey_i_to_extent(_new.k); 145 bch2_cut_front(iter.pos, &new->k_i); 146 147 bch2_cut_front(iter.pos, insert); 148 bch2_cut_back(new->k.p, insert); 149 bch2_cut_back(insert->k.p, &new->k_i); 150 151 /* 152 * @old: extent that we read from 153 * @insert: key that we're going to update, initialized from 154 * extent currently in btree - same as @old unless we raced with 155 * other updates 156 * @new: extent with new pointers that we'll be adding to @insert 157 * 158 * Fist, drop rewrite_ptrs from @new: 159 */ 160 i = 0; 161 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry_c) { 162 if (((1U << i) & m->data_opts.rewrite_ptrs) && 163 (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) && 164 !ptr->cached) { 165 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), ptr); 166 rewrites_found |= 1U << i; 167 } 168 i++; 169 } 170 171 if (m->data_opts.rewrite_ptrs && 172 !rewrites_found && 173 bch2_bkey_durability(c, k) >= m->op.opts.data_replicas) { 174 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "no rewrites found:"); 175 goto nowork; 176 } 177 178 /* 179 * A replica that we just wrote might conflict with a replica 180 * that we want to keep, due to racing with another move: 181 */ 182 restart_drop_conflicting_replicas: 183 extent_for_each_ptr(extent_i_to_s(new), ptr) 184 if ((ptr_c = bch2_bkey_has_device_c(bkey_i_to_s_c(insert), ptr->dev)) && 185 !ptr_c->cached) { 186 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(&new->k_i), ptr); 187 goto restart_drop_conflicting_replicas; 188 } 189 190 if (!bkey_val_u64s(&new->k)) { 191 trace_move_extent_fail2(m, k, bkey_i_to_s_c(&new->k_i), insert, "new replicas conflicted:"); 192 goto nowork; 193 } 194 195 /* Now, drop pointers that conflict with what we just wrote: */ 196 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) 197 if ((ptr = bch2_bkey_has_device(bkey_i_to_s(insert), p.ptr.dev))) 198 bch2_bkey_drop_ptr_noerror(bkey_i_to_s(insert), ptr); 199 200 durability = bch2_bkey_durability(c, bkey_i_to_s_c(insert)) + 201 bch2_bkey_durability(c, bkey_i_to_s_c(&new->k_i)); 202 203 /* Now, drop excess replicas: */ 204 restart_drop_extra_replicas: 205 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs(bkey_i_to_s(insert)), p, entry) { 206 unsigned ptr_durability = bch2_extent_ptr_durability(c, &p); 207 208 if (!p.ptr.cached && 209 durability - ptr_durability >= m->op.opts.data_replicas) { 210 durability -= ptr_durability; 211 212 bch2_extent_ptr_set_cached(bkey_i_to_s(insert), &entry->ptr); 213 goto restart_drop_extra_replicas; 214 } 215 } 216 217 /* Finally, add the pointers we just wrote: */ 218 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) 219 bch2_extent_ptr_decoded_append(insert, &p); 220 221 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 }); 222 bch2_extent_normalize(c, bkey_i_to_s(insert)); 223 224 ret = bch2_sum_sector_overwrites(trans, &iter, insert, 225 &should_check_enospc, 226 &i_sectors_delta, 227 &disk_sectors_delta); 228 if (ret) 229 goto err; 230 231 if (disk_sectors_delta > (s64) op->res.sectors) { 232 ret = bch2_disk_reservation_add(c, &op->res, 233 disk_sectors_delta - op->res.sectors, 234 !should_check_enospc 235 ? BCH_DISK_RESERVATION_NOFAIL : 0); 236 if (ret) 237 goto out; 238 } 239 240 next_pos = insert->k.p; 241 242 ret = bch2_insert_snapshot_whiteouts(trans, m->btree_id, 243 k.k->p, bkey_start_pos(&insert->k)) ?: 244 bch2_insert_snapshot_whiteouts(trans, m->btree_id, 245 k.k->p, insert->k.p) ?: 246 bch2_bkey_set_needs_rebalance(c, insert, 247 op->opts.background_target, 248 op->opts.background_compression) ?: 249 bch2_trans_update(trans, &iter, insert, 250 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?: 251 bch2_trans_commit(trans, &op->res, 252 NULL, 253 BTREE_INSERT_NOCHECK_RW| 254 BTREE_INSERT_NOFAIL| 255 m->data_opts.btree_insert_flags); 256 if (!ret) { 257 bch2_btree_iter_set_pos(&iter, next_pos); 258 259 this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size); 260 trace_move_extent_finish2(c, bkey_i_to_s_c(&new->k_i)); 261 } 262 err: 263 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) 264 ret = 0; 265 if (ret) 266 break; 267 next: 268 while (bkey_ge(iter.pos, bch2_keylist_front(keys)->k.p)) { 269 bch2_keylist_pop_front(keys); 270 if (bch2_keylist_empty(keys)) 271 goto out; 272 } 273 continue; 274 nowork: 275 if (m->stats && m->stats) { 276 BUG_ON(k.k->p.offset <= iter.pos.offset); 277 atomic64_inc(&m->stats->keys_raced); 278 atomic64_add(k.k->p.offset - iter.pos.offset, 279 &m->stats->sectors_raced); 280 } 281 282 this_cpu_inc(c->counters[BCH_COUNTER_move_extent_fail]); 283 284 bch2_btree_iter_advance(&iter); 285 goto next; 286 } 287 out: 288 bch2_trans_iter_exit(trans, &iter); 289 bch2_bkey_buf_exit(&_insert, c); 290 bch2_bkey_buf_exit(&_new, c); 291 BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart)); 292 return ret; 293 } 294 295 int bch2_data_update_index_update(struct bch_write_op *op) 296 { 297 return bch2_trans_run(op->c, __bch2_data_update_index_update(trans, op)); 298 } 299 300 void bch2_data_update_read_done(struct data_update *m, 301 struct bch_extent_crc_unpacked crc) 302 { 303 /* write bio must own pages: */ 304 BUG_ON(!m->op.wbio.bio.bi_vcnt); 305 306 m->op.crc = crc; 307 m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9; 308 309 closure_call(&m->op.cl, bch2_write, NULL, NULL); 310 } 311 312 void bch2_data_update_exit(struct data_update *update) 313 { 314 struct bch_fs *c = update->op.c; 315 struct bkey_ptrs_c ptrs = 316 bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k)); 317 const struct bch_extent_ptr *ptr; 318 319 bkey_for_each_ptr(ptrs, ptr) { 320 if (c->opts.nocow_enabled) 321 bch2_bucket_nocow_unlock(&c->nocow_locks, 322 PTR_BUCKET_POS(c, ptr), 0); 323 percpu_ref_put(&bch_dev_bkey_exists(c, ptr->dev)->ref); 324 } 325 326 bch2_bkey_buf_exit(&update->k, c); 327 bch2_disk_reservation_put(c, &update->op.res); 328 bch2_bio_free_pages_pool(c, &update->op.wbio.bio); 329 } 330 331 void bch2_update_unwritten_extent(struct btree_trans *trans, 332 struct data_update *update) 333 { 334 struct bch_fs *c = update->op.c; 335 struct bio *bio = &update->op.wbio.bio; 336 struct bkey_i_extent *e; 337 struct write_point *wp; 338 struct bch_extent_ptr *ptr; 339 struct closure cl; 340 struct btree_iter iter; 341 struct bkey_s_c k; 342 int ret; 343 344 closure_init_stack(&cl); 345 bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys); 346 347 while (bio_sectors(bio)) { 348 unsigned sectors = bio_sectors(bio); 349 350 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos, 351 BTREE_ITER_SLOTS); 352 ret = lockrestart_do(trans, ({ 353 k = bch2_btree_iter_peek_slot(&iter); 354 bkey_err(k); 355 })); 356 bch2_trans_iter_exit(trans, &iter); 357 358 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k))) 359 break; 360 361 e = bkey_extent_init(update->op.insert_keys.top); 362 e->k.p = update->op.pos; 363 364 ret = bch2_alloc_sectors_start_trans(trans, 365 update->op.target, 366 false, 367 update->op.write_point, 368 &update->op.devs_have, 369 update->op.nr_replicas, 370 update->op.nr_replicas, 371 update->op.watermark, 372 0, &cl, &wp); 373 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) { 374 bch2_trans_unlock(trans); 375 closure_sync(&cl); 376 continue; 377 } 378 379 if (ret) 380 return; 381 382 sectors = min(sectors, wp->sectors_free); 383 384 bch2_key_resize(&e->k, sectors); 385 386 bch2_open_bucket_get(c, wp, &update->op.open_buckets); 387 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false); 388 bch2_alloc_sectors_done(c, wp); 389 390 bio_advance(bio, sectors << 9); 391 update->op.pos.offset += sectors; 392 393 extent_for_each_ptr(extent_i_to_s(e), ptr) 394 ptr->unwritten = true; 395 bch2_keylist_push(&update->op.insert_keys); 396 397 ret = __bch2_data_update_index_update(trans, &update->op); 398 399 bch2_open_buckets_put(c, &update->op.open_buckets); 400 401 if (ret) 402 break; 403 } 404 405 if (closure_nr_remaining(&cl) != 1) { 406 bch2_trans_unlock(trans); 407 closure_sync(&cl); 408 } 409 } 410 411 int bch2_data_update_init(struct btree_trans *trans, 412 struct moving_context *ctxt, 413 struct data_update *m, 414 struct write_point_specifier wp, 415 struct bch_io_opts io_opts, 416 struct data_update_opts data_opts, 417 enum btree_id btree_id, 418 struct bkey_s_c k) 419 { 420 struct bch_fs *c = trans->c; 421 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k); 422 const union bch_extent_entry *entry; 423 struct extent_ptr_decoded p; 424 const struct bch_extent_ptr *ptr; 425 unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas; 426 unsigned ptrs_locked = 0; 427 int ret; 428 429 bch2_bkey_buf_init(&m->k); 430 bch2_bkey_buf_reassemble(&m->k, c, k); 431 m->btree_id = btree_id; 432 m->data_opts = data_opts; 433 m->ctxt = ctxt; 434 m->stats = ctxt ? ctxt->stats : NULL; 435 436 bch2_write_op_init(&m->op, c, io_opts); 437 m->op.pos = bkey_start_pos(k.k); 438 m->op.version = k.k->version; 439 m->op.target = data_opts.target; 440 m->op.write_point = wp; 441 m->op.nr_replicas = 0; 442 m->op.flags |= BCH_WRITE_PAGES_STABLE| 443 BCH_WRITE_PAGES_OWNED| 444 BCH_WRITE_DATA_ENCODED| 445 BCH_WRITE_MOVE| 446 m->data_opts.write_flags; 447 m->op.compression_opt = io_opts.background_compression ?: io_opts.compression; 448 m->op.watermark = m->data_opts.btree_insert_flags & BCH_WATERMARK_MASK; 449 450 bkey_for_each_ptr(ptrs, ptr) 451 percpu_ref_get(&bch_dev_bkey_exists(c, ptr->dev)->ref); 452 453 i = 0; 454 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) { 455 bool locked; 456 457 if (((1U << i) & m->data_opts.rewrite_ptrs)) { 458 BUG_ON(p.ptr.cached); 459 460 if (crc_is_compressed(p.crc)) 461 reserve_sectors += k.k->size; 462 463 m->op.nr_replicas += bch2_extent_ptr_desired_durability(c, &p); 464 } else if (!p.ptr.cached) { 465 bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev); 466 } 467 468 /* 469 * op->csum_type is normally initialized from the fs/file's 470 * current options - but if an extent is encrypted, we require 471 * that it stays encrypted: 472 */ 473 if (bch2_csum_type_is_encryption(p.crc.csum_type)) { 474 m->op.nonce = p.crc.nonce + p.crc.offset; 475 m->op.csum_type = p.crc.csum_type; 476 } 477 478 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible) 479 m->op.incompressible = true; 480 481 if (c->opts.nocow_enabled) { 482 if (ctxt) { 483 move_ctxt_wait_event(ctxt, 484 (locked = bch2_bucket_nocow_trylock(&c->nocow_locks, 485 PTR_BUCKET_POS(c, &p.ptr), 0)) || 486 !atomic_read(&ctxt->read_sectors)); 487 488 if (!locked) 489 bch2_bucket_nocow_lock(&c->nocow_locks, 490 PTR_BUCKET_POS(c, &p.ptr), 0); 491 } else { 492 if (!bch2_bucket_nocow_trylock(&c->nocow_locks, 493 PTR_BUCKET_POS(c, &p.ptr), 0)) { 494 ret = -BCH_ERR_nocow_lock_blocked; 495 goto err; 496 } 497 } 498 ptrs_locked |= (1U << i); 499 } 500 501 i++; 502 } 503 504 if (reserve_sectors) { 505 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors, 506 m->data_opts.extra_replicas 507 ? 0 508 : BCH_DISK_RESERVATION_NOFAIL); 509 if (ret) 510 goto err; 511 } 512 513 m->op.nr_replicas += m->data_opts.extra_replicas; 514 m->op.nr_replicas_required = m->op.nr_replicas; 515 516 BUG_ON(!m->op.nr_replicas); 517 518 /* Special handling required: */ 519 if (bkey_extent_is_unwritten(k)) 520 return -BCH_ERR_unwritten_extent_update; 521 return 0; 522 err: 523 i = 0; 524 bkey_for_each_ptr_decode(k.k, ptrs, p, entry) { 525 if ((1U << i) & ptrs_locked) 526 bch2_bucket_nocow_unlock(&c->nocow_locks, 527 PTR_BUCKET_POS(c, &p.ptr), 0); 528 percpu_ref_put(&bch_dev_bkey_exists(c, p.ptr.dev)->ref); 529 i++; 530 } 531 532 bch2_bkey_buf_exit(&m->k, c); 533 bch2_bio_free_pages_pool(c, &m->op.wbio.bio); 534 return ret; 535 } 536 537 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts) 538 { 539 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k); 540 const struct bch_extent_ptr *ptr; 541 unsigned i = 0; 542 543 bkey_for_each_ptr(ptrs, ptr) { 544 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) { 545 opts->kill_ptrs |= 1U << i; 546 opts->rewrite_ptrs ^= 1U << i; 547 } 548 549 i++; 550 } 551 } 552