1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * io_misc.c - fallocate, fpunch, truncate: 4 */ 5 6 #include "bcachefs.h" 7 #include "alloc_foreground.h" 8 #include "bkey_buf.h" 9 #include "btree_update.h" 10 #include "buckets.h" 11 #include "clock.h" 12 #include "error.h" 13 #include "extents.h" 14 #include "extent_update.h" 15 #include "inode.h" 16 #include "io_misc.h" 17 #include "io_write.h" 18 #include "logged_ops.h" 19 #include "rebalance.h" 20 #include "subvolume.h" 21 22 /* Overwrites whatever was present with zeroes: */ 23 int bch2_extent_fallocate(struct btree_trans *trans, 24 subvol_inum inum, 25 struct btree_iter *iter, 26 u64 sectors, 27 struct bch_io_opts opts, 28 s64 *i_sectors_delta, 29 struct write_point_specifier write_point) 30 { 31 struct bch_fs *c = trans->c; 32 struct disk_reservation disk_res = { 0 }; 33 struct closure cl; 34 struct open_buckets open_buckets = { 0 }; 35 struct bkey_s_c k; 36 struct bkey_buf old, new; 37 unsigned sectors_allocated = 0, new_replicas; 38 bool unwritten = opts.nocow && 39 c->sb.version >= bcachefs_metadata_version_unwritten_extents; 40 int ret; 41 42 bch2_bkey_buf_init(&old); 43 bch2_bkey_buf_init(&new); 44 closure_init_stack(&cl); 45 46 k = bch2_btree_iter_peek_slot(iter); 47 ret = bkey_err(k); 48 if (ret) 49 return ret; 50 51 sectors = min_t(u64, sectors, k.k->p.offset - iter->pos.offset); 52 new_replicas = max(0, (int) opts.data_replicas - 53 (int) bch2_bkey_nr_ptrs_fully_allocated(k)); 54 55 /* 56 * Get a disk reservation before (in the nocow case) calling 57 * into the allocator: 58 */ 59 ret = bch2_disk_reservation_get(c, &disk_res, sectors, new_replicas, 0); 60 if (unlikely(ret)) 61 goto err_noprint; 62 63 bch2_bkey_buf_reassemble(&old, c, k); 64 65 if (!unwritten) { 66 struct bkey_i_reservation *reservation; 67 68 bch2_bkey_buf_realloc(&new, c, sizeof(*reservation) / sizeof(u64)); 69 reservation = bkey_reservation_init(new.k); 70 reservation->k.p = iter->pos; 71 bch2_key_resize(&reservation->k, sectors); 72 reservation->v.nr_replicas = opts.data_replicas; 73 } else { 74 struct bkey_i_extent *e; 75 struct bch_devs_list devs_have; 76 struct write_point *wp; 77 78 devs_have.nr = 0; 79 80 bch2_bkey_buf_realloc(&new, c, BKEY_EXTENT_U64s_MAX); 81 82 e = bkey_extent_init(new.k); 83 e->k.p = iter->pos; 84 85 ret = bch2_alloc_sectors_start_trans(trans, 86 opts.foreground_target, 87 false, 88 write_point, 89 &devs_have, 90 opts.data_replicas, 91 opts.data_replicas, 92 BCH_WATERMARK_normal, 0, &cl, &wp); 93 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) 94 ret = -BCH_ERR_transaction_restart_nested; 95 if (ret) 96 goto err; 97 98 sectors = min_t(u64, sectors, wp->sectors_free); 99 sectors_allocated = sectors; 100 101 bch2_key_resize(&e->k, sectors); 102 103 bch2_open_bucket_get(c, wp, &open_buckets); 104 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false); 105 bch2_alloc_sectors_done(c, wp); 106 107 extent_for_each_ptr(extent_i_to_s(e), ptr) 108 ptr->unwritten = true; 109 } 110 111 ret = bch2_extent_update(trans, inum, iter, new.k, &disk_res, 112 0, i_sectors_delta, true); 113 err: 114 if (!ret && sectors_allocated) 115 bch2_increment_clock(c, sectors_allocated, WRITE); 116 if (should_print_err(ret)) 117 bch_err_inum_offset_ratelimited(c, 118 inum.inum, 119 iter->pos.offset << 9, 120 "%s(): error: %s", __func__, bch2_err_str(ret)); 121 err_noprint: 122 bch2_open_buckets_put(c, &open_buckets); 123 bch2_disk_reservation_put(c, &disk_res); 124 bch2_bkey_buf_exit(&new, c); 125 bch2_bkey_buf_exit(&old, c); 126 127 if (closure_nr_remaining(&cl) != 1) { 128 bch2_trans_unlock_long(trans); 129 bch2_wait_on_allocator(c, &cl); 130 } 131 132 return ret; 133 } 134 135 /* 136 * Returns -BCH_ERR_transacton_restart if we had to drop locks: 137 */ 138 int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter, 139 subvol_inum inum, u64 end, 140 s64 *i_sectors_delta) 141 { 142 struct bch_fs *c = trans->c; 143 unsigned max_sectors = KEY_SIZE_MAX & (~0 << c->block_bits); 144 struct bpos end_pos = POS(inum.inum, end); 145 struct bkey_s_c k; 146 int ret = 0, ret2 = 0; 147 u32 snapshot; 148 149 while (!ret || 150 bch2_err_matches(ret, BCH_ERR_transaction_restart)) { 151 struct disk_reservation disk_res = 152 bch2_disk_reservation_init(c, 0); 153 struct bkey_i delete; 154 155 if (ret) 156 ret2 = ret; 157 158 bch2_trans_begin(trans); 159 160 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot); 161 if (ret) 162 continue; 163 164 bch2_btree_iter_set_snapshot(iter, snapshot); 165 166 /* 167 * peek_upto() doesn't have ideal semantics for extents: 168 */ 169 k = bch2_btree_iter_peek_upto(iter, end_pos); 170 if (!k.k) 171 break; 172 173 ret = bkey_err(k); 174 if (ret) 175 continue; 176 177 bkey_init(&delete.k); 178 delete.k.p = iter->pos; 179 180 /* create the biggest key we can */ 181 bch2_key_resize(&delete.k, max_sectors); 182 bch2_cut_back(end_pos, &delete); 183 184 ret = bch2_extent_update(trans, inum, iter, &delete, 185 &disk_res, 0, i_sectors_delta, false); 186 bch2_disk_reservation_put(c, &disk_res); 187 } 188 189 return ret ?: ret2; 190 } 191 192 int bch2_fpunch(struct bch_fs *c, subvol_inum inum, u64 start, u64 end, 193 s64 *i_sectors_delta) 194 { 195 struct btree_trans *trans = bch2_trans_get(c); 196 struct btree_iter iter; 197 int ret; 198 199 bch2_trans_iter_init(trans, &iter, BTREE_ID_extents, 200 POS(inum.inum, start), 201 BTREE_ITER_intent); 202 203 ret = bch2_fpunch_at(trans, &iter, inum, end, i_sectors_delta); 204 205 bch2_trans_iter_exit(trans, &iter); 206 bch2_trans_put(trans); 207 208 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) 209 ret = 0; 210 211 return ret; 212 } 213 214 /* truncate: */ 215 216 void bch2_logged_op_truncate_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k) 217 { 218 struct bkey_s_c_logged_op_truncate op = bkey_s_c_to_logged_op_truncate(k); 219 220 prt_printf(out, "subvol=%u", le32_to_cpu(op.v->subvol)); 221 prt_printf(out, " inum=%llu", le64_to_cpu(op.v->inum)); 222 prt_printf(out, " new_i_size=%llu", le64_to_cpu(op.v->new_i_size)); 223 } 224 225 static int truncate_set_isize(struct btree_trans *trans, 226 subvol_inum inum, 227 u64 new_i_size, 228 bool warn) 229 { 230 struct btree_iter iter = { NULL }; 231 struct bch_inode_unpacked inode_u; 232 int ret; 233 234 ret = __bch2_inode_peek(trans, &iter, &inode_u, inum, BTREE_ITER_intent, warn) ?: 235 (inode_u.bi_size = new_i_size, 0) ?: 236 bch2_inode_write(trans, &iter, &inode_u); 237 238 bch2_trans_iter_exit(trans, &iter); 239 return ret; 240 } 241 242 static int __bch2_resume_logged_op_truncate(struct btree_trans *trans, 243 struct bkey_i *op_k, 244 u64 *i_sectors_delta) 245 { 246 struct bch_fs *c = trans->c; 247 struct btree_iter fpunch_iter; 248 struct bkey_i_logged_op_truncate *op = bkey_i_to_logged_op_truncate(op_k); 249 subvol_inum inum = { le32_to_cpu(op->v.subvol), le64_to_cpu(op->v.inum) }; 250 u64 new_i_size = le64_to_cpu(op->v.new_i_size); 251 bool warn_errors = i_sectors_delta != NULL; 252 int ret; 253 254 ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc, 255 truncate_set_isize(trans, inum, new_i_size, i_sectors_delta != NULL)); 256 if (ret) 257 goto err; 258 259 bch2_trans_iter_init(trans, &fpunch_iter, BTREE_ID_extents, 260 POS(inum.inum, round_up(new_i_size, block_bytes(c)) >> 9), 261 BTREE_ITER_intent); 262 ret = bch2_fpunch_at(trans, &fpunch_iter, inum, U64_MAX, i_sectors_delta); 263 bch2_trans_iter_exit(trans, &fpunch_iter); 264 265 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) 266 ret = 0; 267 err: 268 if (warn_errors) 269 bch_err_fn(c, ret); 270 return ret; 271 } 272 273 int bch2_resume_logged_op_truncate(struct btree_trans *trans, struct bkey_i *op_k) 274 { 275 return __bch2_resume_logged_op_truncate(trans, op_k, NULL); 276 } 277 278 int bch2_truncate(struct bch_fs *c, subvol_inum inum, u64 new_i_size, u64 *i_sectors_delta) 279 { 280 struct bkey_i_logged_op_truncate op; 281 282 bkey_logged_op_truncate_init(&op.k_i); 283 op.v.subvol = cpu_to_le32(inum.subvol); 284 op.v.inum = cpu_to_le64(inum.inum); 285 op.v.new_i_size = cpu_to_le64(new_i_size); 286 287 /* 288 * Logged ops aren't atomic w.r.t. snapshot creation: creating a 289 * snapshot while they're in progress, then crashing, will result in the 290 * resume only proceeding in one of the snapshots 291 */ 292 down_read(&c->snapshot_create_lock); 293 struct btree_trans *trans = bch2_trans_get(c); 294 int ret = bch2_logged_op_start(trans, &op.k_i); 295 if (ret) 296 goto out; 297 ret = __bch2_resume_logged_op_truncate(trans, &op.k_i, i_sectors_delta); 298 ret = bch2_logged_op_finish(trans, &op.k_i) ?: ret; 299 out: 300 bch2_trans_put(trans); 301 up_read(&c->snapshot_create_lock); 302 303 return ret; 304 } 305 306 /* finsert/fcollapse: */ 307 308 void bch2_logged_op_finsert_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k) 309 { 310 struct bkey_s_c_logged_op_finsert op = bkey_s_c_to_logged_op_finsert(k); 311 312 prt_printf(out, "subvol=%u", le32_to_cpu(op.v->subvol)); 313 prt_printf(out, " inum=%llu", le64_to_cpu(op.v->inum)); 314 prt_printf(out, " dst_offset=%lli", le64_to_cpu(op.v->dst_offset)); 315 prt_printf(out, " src_offset=%llu", le64_to_cpu(op.v->src_offset)); 316 } 317 318 static int adjust_i_size(struct btree_trans *trans, subvol_inum inum, 319 u64 offset, s64 len, bool warn) 320 { 321 struct btree_iter iter; 322 struct bch_inode_unpacked inode_u; 323 int ret; 324 325 offset <<= 9; 326 len <<= 9; 327 328 ret = __bch2_inode_peek(trans, &iter, &inode_u, inum, BTREE_ITER_intent, warn); 329 if (ret) 330 return ret; 331 332 if (len > 0) { 333 if (MAX_LFS_FILESIZE - inode_u.bi_size < len) { 334 ret = -EFBIG; 335 goto err; 336 } 337 338 if (offset >= inode_u.bi_size) { 339 ret = -EINVAL; 340 goto err; 341 } 342 } 343 344 inode_u.bi_size += len; 345 inode_u.bi_mtime = inode_u.bi_ctime = bch2_current_time(trans->c); 346 347 ret = bch2_inode_write(trans, &iter, &inode_u); 348 err: 349 bch2_trans_iter_exit(trans, &iter); 350 return ret; 351 } 352 353 static int __bch2_resume_logged_op_finsert(struct btree_trans *trans, 354 struct bkey_i *op_k, 355 u64 *i_sectors_delta) 356 { 357 struct bch_fs *c = trans->c; 358 struct btree_iter iter; 359 struct bkey_i_logged_op_finsert *op = bkey_i_to_logged_op_finsert(op_k); 360 subvol_inum inum = { le32_to_cpu(op->v.subvol), le64_to_cpu(op->v.inum) }; 361 struct bch_io_opts opts; 362 u64 dst_offset = le64_to_cpu(op->v.dst_offset); 363 u64 src_offset = le64_to_cpu(op->v.src_offset); 364 s64 shift = dst_offset - src_offset; 365 u64 len = abs(shift); 366 u64 pos = le64_to_cpu(op->v.pos); 367 bool insert = shift > 0; 368 u32 snapshot; 369 bool warn_errors = i_sectors_delta != NULL; 370 int ret = 0; 371 372 ret = bch2_inum_opts_get(trans, inum, &opts); 373 if (ret) 374 return ret; 375 376 /* 377 * check for missing subvolume before fpunch, as in resume we don't want 378 * it to be a fatal error 379 */ 380 ret = lockrestart_do(trans, __bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot, warn_errors)); 381 if (ret) 382 return ret; 383 384 bch2_trans_iter_init(trans, &iter, BTREE_ID_extents, 385 POS(inum.inum, 0), 386 BTREE_ITER_intent); 387 388 switch (op->v.state) { 389 case LOGGED_OP_FINSERT_start: 390 op->v.state = LOGGED_OP_FINSERT_shift_extents; 391 392 if (insert) { 393 ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc, 394 adjust_i_size(trans, inum, src_offset, len, warn_errors) ?: 395 bch2_logged_op_update(trans, &op->k_i)); 396 if (ret) 397 goto err; 398 } else { 399 bch2_btree_iter_set_pos(&iter, POS(inum.inum, src_offset)); 400 401 ret = bch2_fpunch_at(trans, &iter, inum, src_offset + len, i_sectors_delta); 402 if (ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart)) 403 goto err; 404 405 ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc, 406 bch2_logged_op_update(trans, &op->k_i)); 407 } 408 409 fallthrough; 410 case LOGGED_OP_FINSERT_shift_extents: 411 while (1) { 412 struct disk_reservation disk_res = 413 bch2_disk_reservation_init(c, 0); 414 struct bkey_i delete, *copy; 415 struct bkey_s_c k; 416 struct bpos src_pos = POS(inum.inum, src_offset); 417 418 bch2_trans_begin(trans); 419 420 ret = __bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot, 421 warn_errors); 422 if (ret) 423 goto btree_err; 424 425 bch2_btree_iter_set_snapshot(&iter, snapshot); 426 bch2_btree_iter_set_pos(&iter, SPOS(inum.inum, pos, snapshot)); 427 428 k = insert 429 ? bch2_btree_iter_peek_prev(&iter) 430 : bch2_btree_iter_peek_upto(&iter, POS(inum.inum, U64_MAX)); 431 if ((ret = bkey_err(k))) 432 goto btree_err; 433 434 if (!k.k || 435 k.k->p.inode != inum.inum || 436 bkey_le(k.k->p, POS(inum.inum, src_offset))) 437 break; 438 439 copy = bch2_bkey_make_mut_noupdate(trans, k); 440 if ((ret = PTR_ERR_OR_ZERO(copy))) 441 goto btree_err; 442 443 if (insert && 444 bkey_lt(bkey_start_pos(k.k), src_pos)) { 445 bch2_cut_front(src_pos, copy); 446 447 /* Splitting compressed extent? */ 448 bch2_disk_reservation_add(c, &disk_res, 449 copy->k.size * 450 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy)), 451 BCH_DISK_RESERVATION_NOFAIL); 452 } 453 454 bkey_init(&delete.k); 455 delete.k.p = copy->k.p; 456 delete.k.p.snapshot = snapshot; 457 delete.k.size = copy->k.size; 458 459 copy->k.p.offset += shift; 460 copy->k.p.snapshot = snapshot; 461 462 op->v.pos = cpu_to_le64(insert ? bkey_start_offset(&delete.k) : delete.k.p.offset); 463 464 ret = bch2_bkey_set_needs_rebalance(c, copy, &opts) ?: 465 bch2_btree_insert_trans(trans, BTREE_ID_extents, &delete, 0) ?: 466 bch2_btree_insert_trans(trans, BTREE_ID_extents, copy, 0) ?: 467 bch2_logged_op_update(trans, &op->k_i) ?: 468 bch2_trans_commit(trans, &disk_res, NULL, BCH_TRANS_COMMIT_no_enospc); 469 btree_err: 470 bch2_disk_reservation_put(c, &disk_res); 471 472 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) 473 continue; 474 if (ret) 475 goto err; 476 477 pos = le64_to_cpu(op->v.pos); 478 } 479 480 op->v.state = LOGGED_OP_FINSERT_finish; 481 482 if (!insert) { 483 ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc, 484 adjust_i_size(trans, inum, src_offset, shift, warn_errors) ?: 485 bch2_logged_op_update(trans, &op->k_i)); 486 } else { 487 /* We need an inode update to update bi_journal_seq for fsync: */ 488 ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc, 489 adjust_i_size(trans, inum, 0, 0, warn_errors) ?: 490 bch2_logged_op_update(trans, &op->k_i)); 491 } 492 493 break; 494 case LOGGED_OP_FINSERT_finish: 495 break; 496 } 497 err: 498 bch2_trans_iter_exit(trans, &iter); 499 if (warn_errors) 500 bch_err_fn(c, ret); 501 return ret; 502 } 503 504 int bch2_resume_logged_op_finsert(struct btree_trans *trans, struct bkey_i *op_k) 505 { 506 return __bch2_resume_logged_op_finsert(trans, op_k, NULL); 507 } 508 509 int bch2_fcollapse_finsert(struct bch_fs *c, subvol_inum inum, 510 u64 offset, u64 len, bool insert, 511 s64 *i_sectors_delta) 512 { 513 struct bkey_i_logged_op_finsert op; 514 s64 shift = insert ? len : -len; 515 516 bkey_logged_op_finsert_init(&op.k_i); 517 op.v.subvol = cpu_to_le32(inum.subvol); 518 op.v.inum = cpu_to_le64(inum.inum); 519 op.v.dst_offset = cpu_to_le64(offset + shift); 520 op.v.src_offset = cpu_to_le64(offset); 521 op.v.pos = cpu_to_le64(insert ? U64_MAX : offset); 522 523 /* 524 * Logged ops aren't atomic w.r.t. snapshot creation: creating a 525 * snapshot while they're in progress, then crashing, will result in the 526 * resume only proceeding in one of the snapshots 527 */ 528 down_read(&c->snapshot_create_lock); 529 struct btree_trans *trans = bch2_trans_get(c); 530 int ret = bch2_logged_op_start(trans, &op.k_i); 531 if (ret) 532 goto out; 533 ret = __bch2_resume_logged_op_finsert(trans, &op.k_i, i_sectors_delta); 534 ret = bch2_logged_op_finish(trans, &op.k_i) ?: ret; 535 out: 536 bch2_trans_put(trans); 537 up_read(&c->snapshot_create_lock); 538 539 return ret; 540 } 541