1 // SPDX-License-Identifier: GPL-2.0 2 #ifndef NO_BCACHEFS_FS 3 4 #include "bcachefs.h" 5 #include "alloc_foreground.h" 6 #include "fs.h" 7 #include "fs-io.h" 8 #include "fs-io-direct.h" 9 #include "fs-io-pagecache.h" 10 #include "io_read.h" 11 #include "io_write.h" 12 13 #include <linux/kthread.h> 14 #include <linux/pagemap.h> 15 #include <linux/prefetch.h> 16 #include <linux/task_io_accounting_ops.h> 17 18 /* O_DIRECT reads */ 19 20 struct dio_read { 21 struct closure cl; 22 struct kiocb *req; 23 long ret; 24 bool should_dirty; 25 struct bch_read_bio rbio; 26 }; 27 28 static void bio_check_or_release(struct bio *bio, bool check_dirty) 29 { 30 if (check_dirty) { 31 bio_check_pages_dirty(bio); 32 } else { 33 bio_release_pages(bio, false); 34 bio_put(bio); 35 } 36 } 37 38 static void bch2_dio_read_complete(struct closure *cl) 39 { 40 struct dio_read *dio = container_of(cl, struct dio_read, cl); 41 42 dio->req->ki_complete(dio->req, dio->ret); 43 bio_check_or_release(&dio->rbio.bio, dio->should_dirty); 44 } 45 46 static void bch2_direct_IO_read_endio(struct bio *bio) 47 { 48 struct dio_read *dio = bio->bi_private; 49 50 if (bio->bi_status) 51 dio->ret = blk_status_to_errno(bio->bi_status); 52 53 closure_put(&dio->cl); 54 } 55 56 static void bch2_direct_IO_read_split_endio(struct bio *bio) 57 { 58 struct dio_read *dio = bio->bi_private; 59 bool should_dirty = dio->should_dirty; 60 61 bch2_direct_IO_read_endio(bio); 62 bio_check_or_release(bio, should_dirty); 63 } 64 65 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter) 66 { 67 struct file *file = req->ki_filp; 68 struct bch_inode_info *inode = file_bch_inode(file); 69 struct bch_fs *c = inode->v.i_sb->s_fs_info; 70 struct bch_io_opts opts; 71 struct dio_read *dio; 72 struct bio *bio; 73 loff_t offset = req->ki_pos; 74 bool sync = is_sync_kiocb(req); 75 size_t shorten; 76 ssize_t ret; 77 78 bch2_inode_opts_get(&opts, c, &inode->ei_inode); 79 80 if ((offset|iter->count) & (block_bytes(c) - 1)) 81 return -EINVAL; 82 83 ret = min_t(loff_t, iter->count, 84 max_t(loff_t, 0, i_size_read(&inode->v) - offset)); 85 86 if (!ret) 87 return ret; 88 89 shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c)); 90 iter->count -= shorten; 91 92 bio = bio_alloc_bioset(NULL, 93 bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS), 94 REQ_OP_READ, 95 GFP_KERNEL, 96 &c->dio_read_bioset); 97 98 bio->bi_end_io = bch2_direct_IO_read_endio; 99 100 dio = container_of(bio, struct dio_read, rbio.bio); 101 closure_init(&dio->cl, NULL); 102 103 /* 104 * this is a _really_ horrible hack just to avoid an atomic sub at the 105 * end: 106 */ 107 if (!sync) { 108 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL); 109 atomic_set(&dio->cl.remaining, 110 CLOSURE_REMAINING_INITIALIZER - 111 CLOSURE_RUNNING + 112 CLOSURE_DESTRUCTOR); 113 } else { 114 atomic_set(&dio->cl.remaining, 115 CLOSURE_REMAINING_INITIALIZER + 1); 116 } 117 118 dio->req = req; 119 dio->ret = ret; 120 /* 121 * This is one of the sketchier things I've encountered: we have to skip 122 * the dirtying of requests that are internal from the kernel (i.e. from 123 * loopback), because we'll deadlock on page_lock. 124 */ 125 dio->should_dirty = iter_is_iovec(iter); 126 127 goto start; 128 while (iter->count) { 129 bio = bio_alloc_bioset(NULL, 130 bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS), 131 REQ_OP_READ, 132 GFP_KERNEL, 133 &c->bio_read); 134 bio->bi_end_io = bch2_direct_IO_read_split_endio; 135 start: 136 bio->bi_opf = REQ_OP_READ|REQ_SYNC; 137 bio->bi_iter.bi_sector = offset >> 9; 138 bio->bi_private = dio; 139 140 ret = bio_iov_iter_get_pages(bio, iter); 141 if (ret < 0) { 142 /* XXX: fault inject this path */ 143 bio->bi_status = BLK_STS_RESOURCE; 144 bio_endio(bio); 145 break; 146 } 147 148 offset += bio->bi_iter.bi_size; 149 150 if (dio->should_dirty) 151 bio_set_pages_dirty(bio); 152 153 if (iter->count) 154 closure_get(&dio->cl); 155 156 bch2_read(c, rbio_init(bio, opts), inode_inum(inode)); 157 } 158 159 iter->count += shorten; 160 161 if (sync) { 162 closure_sync(&dio->cl); 163 closure_debug_destroy(&dio->cl); 164 ret = dio->ret; 165 bio_check_or_release(&dio->rbio.bio, dio->should_dirty); 166 return ret; 167 } else { 168 return -EIOCBQUEUED; 169 } 170 } 171 172 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter) 173 { 174 struct file *file = iocb->ki_filp; 175 struct bch_inode_info *inode = file_bch_inode(file); 176 struct address_space *mapping = file->f_mapping; 177 size_t count = iov_iter_count(iter); 178 ssize_t ret; 179 180 if (!count) 181 return 0; /* skip atime */ 182 183 if (iocb->ki_flags & IOCB_DIRECT) { 184 struct blk_plug plug; 185 186 if (unlikely(mapping->nrpages)) { 187 ret = filemap_write_and_wait_range(mapping, 188 iocb->ki_pos, 189 iocb->ki_pos + count - 1); 190 if (ret < 0) 191 goto out; 192 } 193 194 file_accessed(file); 195 196 blk_start_plug(&plug); 197 ret = bch2_direct_IO_read(iocb, iter); 198 blk_finish_plug(&plug); 199 200 if (ret >= 0) 201 iocb->ki_pos += ret; 202 } else { 203 bch2_pagecache_add_get(inode); 204 ret = generic_file_read_iter(iocb, iter); 205 bch2_pagecache_add_put(inode); 206 } 207 out: 208 return bch2_err_class(ret); 209 } 210 211 /* O_DIRECT writes */ 212 213 struct dio_write { 214 struct kiocb *req; 215 struct address_space *mapping; 216 struct bch_inode_info *inode; 217 struct mm_struct *mm; 218 unsigned loop:1, 219 extending:1, 220 sync:1, 221 flush:1, 222 free_iov:1; 223 struct quota_res quota_res; 224 u64 written; 225 226 struct iov_iter iter; 227 struct iovec inline_vecs[2]; 228 229 /* must be last: */ 230 struct bch_write_op op; 231 }; 232 233 static bool bch2_check_range_allocated(struct bch_fs *c, subvol_inum inum, 234 u64 offset, u64 size, 235 unsigned nr_replicas, bool compressed) 236 { 237 struct btree_trans *trans = bch2_trans_get(c); 238 struct btree_iter iter; 239 struct bkey_s_c k; 240 u64 end = offset + size; 241 u32 snapshot; 242 bool ret = true; 243 int err; 244 retry: 245 bch2_trans_begin(trans); 246 247 err = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot); 248 if (err) 249 goto err; 250 251 for_each_btree_key_norestart(trans, iter, BTREE_ID_extents, 252 SPOS(inum.inum, offset, snapshot), 253 BTREE_ITER_SLOTS, k, err) { 254 if (bkey_ge(bkey_start_pos(k.k), POS(inum.inum, end))) 255 break; 256 257 if (k.k->p.snapshot != snapshot || 258 nr_replicas > bch2_bkey_replicas(c, k) || 259 (!compressed && bch2_bkey_sectors_compressed(k))) { 260 ret = false; 261 break; 262 } 263 } 264 265 offset = iter.pos.offset; 266 bch2_trans_iter_exit(trans, &iter); 267 err: 268 if (bch2_err_matches(err, BCH_ERR_transaction_restart)) 269 goto retry; 270 bch2_trans_put(trans); 271 272 return err ? false : ret; 273 } 274 275 static noinline bool bch2_dio_write_check_allocated(struct dio_write *dio) 276 { 277 struct bch_fs *c = dio->op.c; 278 struct bch_inode_info *inode = dio->inode; 279 struct bio *bio = &dio->op.wbio.bio; 280 281 return bch2_check_range_allocated(c, inode_inum(inode), 282 dio->op.pos.offset, bio_sectors(bio), 283 dio->op.opts.data_replicas, 284 dio->op.opts.compression != 0); 285 } 286 287 static void bch2_dio_write_loop_async(struct bch_write_op *); 288 static __always_inline long bch2_dio_write_done(struct dio_write *dio); 289 290 /* 291 * We're going to return -EIOCBQUEUED, but we haven't finished consuming the 292 * iov_iter yet, so we need to stash a copy of the iovec: it might be on the 293 * caller's stack, we're not guaranteed that it will live for the duration of 294 * the IO: 295 */ 296 static noinline int bch2_dio_write_copy_iov(struct dio_write *dio) 297 { 298 struct iovec *iov = dio->inline_vecs; 299 300 /* 301 * iov_iter has a single embedded iovec - nothing to do: 302 */ 303 if (iter_is_ubuf(&dio->iter)) 304 return 0; 305 306 /* 307 * We don't currently handle non-iovec iov_iters here - return an error, 308 * and we'll fall back to doing the IO synchronously: 309 */ 310 if (!iter_is_iovec(&dio->iter)) 311 return -1; 312 313 if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) { 314 iov = kmalloc_array(dio->iter.nr_segs, sizeof(*iov), 315 GFP_KERNEL); 316 if (unlikely(!iov)) 317 return -ENOMEM; 318 319 dio->free_iov = true; 320 } 321 322 memcpy(iov, dio->iter.__iov, dio->iter.nr_segs * sizeof(*iov)); 323 dio->iter.__iov = iov; 324 return 0; 325 } 326 327 static void bch2_dio_write_flush_done(struct closure *cl) 328 { 329 struct dio_write *dio = container_of(cl, struct dio_write, op.cl); 330 struct bch_fs *c = dio->op.c; 331 332 closure_debug_destroy(cl); 333 334 dio->op.error = bch2_journal_error(&c->journal); 335 336 bch2_dio_write_done(dio); 337 } 338 339 static noinline void bch2_dio_write_flush(struct dio_write *dio) 340 { 341 struct bch_fs *c = dio->op.c; 342 struct bch_inode_unpacked inode; 343 int ret; 344 345 dio->flush = 0; 346 347 closure_init(&dio->op.cl, NULL); 348 349 if (!dio->op.error) { 350 ret = bch2_inode_find_by_inum(c, inode_inum(dio->inode), &inode); 351 if (ret) { 352 dio->op.error = ret; 353 } else { 354 bch2_journal_flush_seq_async(&c->journal, inode.bi_journal_seq, 355 &dio->op.cl); 356 bch2_inode_flush_nocow_writes_async(c, dio->inode, &dio->op.cl); 357 } 358 } 359 360 if (dio->sync) { 361 closure_sync(&dio->op.cl); 362 closure_debug_destroy(&dio->op.cl); 363 } else { 364 continue_at(&dio->op.cl, bch2_dio_write_flush_done, NULL); 365 } 366 } 367 368 static __always_inline long bch2_dio_write_done(struct dio_write *dio) 369 { 370 struct kiocb *req = dio->req; 371 struct bch_inode_info *inode = dio->inode; 372 bool sync = dio->sync; 373 long ret; 374 375 if (unlikely(dio->flush)) { 376 bch2_dio_write_flush(dio); 377 if (!sync) 378 return -EIOCBQUEUED; 379 } 380 381 bch2_pagecache_block_put(inode); 382 383 if (dio->free_iov) 384 kfree(dio->iter.__iov); 385 386 ret = dio->op.error ?: ((long) dio->written << 9); 387 bio_put(&dio->op.wbio.bio); 388 389 /* inode->i_dio_count is our ref on inode and thus bch_fs */ 390 inode_dio_end(&inode->v); 391 392 if (ret < 0) 393 ret = bch2_err_class(ret); 394 395 if (!sync) { 396 req->ki_complete(req, ret); 397 ret = -EIOCBQUEUED; 398 } 399 return ret; 400 } 401 402 static __always_inline void bch2_dio_write_end(struct dio_write *dio) 403 { 404 struct bch_fs *c = dio->op.c; 405 struct kiocb *req = dio->req; 406 struct bch_inode_info *inode = dio->inode; 407 struct bio *bio = &dio->op.wbio.bio; 408 409 req->ki_pos += (u64) dio->op.written << 9; 410 dio->written += dio->op.written; 411 412 if (dio->extending) { 413 spin_lock(&inode->v.i_lock); 414 if (req->ki_pos > inode->v.i_size) 415 i_size_write(&inode->v, req->ki_pos); 416 spin_unlock(&inode->v.i_lock); 417 } 418 419 if (dio->op.i_sectors_delta || dio->quota_res.sectors) { 420 mutex_lock(&inode->ei_quota_lock); 421 __bch2_i_sectors_acct(c, inode, &dio->quota_res, dio->op.i_sectors_delta); 422 __bch2_quota_reservation_put(c, inode, &dio->quota_res); 423 mutex_unlock(&inode->ei_quota_lock); 424 } 425 426 bio_release_pages(bio, false); 427 428 if (unlikely(dio->op.error)) 429 set_bit(EI_INODE_ERROR, &inode->ei_flags); 430 } 431 432 static __always_inline long bch2_dio_write_loop(struct dio_write *dio) 433 { 434 struct bch_fs *c = dio->op.c; 435 struct kiocb *req = dio->req; 436 struct address_space *mapping = dio->mapping; 437 struct bch_inode_info *inode = dio->inode; 438 struct bch_io_opts opts; 439 struct bio *bio = &dio->op.wbio.bio; 440 unsigned unaligned, iter_count; 441 bool sync = dio->sync, dropped_locks; 442 long ret; 443 444 bch2_inode_opts_get(&opts, c, &inode->ei_inode); 445 446 while (1) { 447 iter_count = dio->iter.count; 448 449 EBUG_ON(current->faults_disabled_mapping); 450 current->faults_disabled_mapping = mapping; 451 452 ret = bio_iov_iter_get_pages(bio, &dio->iter); 453 454 dropped_locks = fdm_dropped_locks(); 455 456 current->faults_disabled_mapping = NULL; 457 458 /* 459 * If the fault handler returned an error but also signalled 460 * that it dropped & retook ei_pagecache_lock, we just need to 461 * re-shoot down the page cache and retry: 462 */ 463 if (dropped_locks && ret) 464 ret = 0; 465 466 if (unlikely(ret < 0)) 467 goto err; 468 469 if (unlikely(dropped_locks)) { 470 ret = bch2_write_invalidate_inode_pages_range(mapping, 471 req->ki_pos, 472 req->ki_pos + iter_count - 1); 473 if (unlikely(ret)) 474 goto err; 475 476 if (!bio->bi_iter.bi_size) 477 continue; 478 } 479 480 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1); 481 bio->bi_iter.bi_size -= unaligned; 482 iov_iter_revert(&dio->iter, unaligned); 483 484 if (!bio->bi_iter.bi_size) { 485 /* 486 * bio_iov_iter_get_pages was only able to get < 487 * blocksize worth of pages: 488 */ 489 ret = -EFAULT; 490 goto err; 491 } 492 493 bch2_write_op_init(&dio->op, c, opts); 494 dio->op.end_io = sync 495 ? NULL 496 : bch2_dio_write_loop_async; 497 dio->op.target = dio->op.opts.foreground_target; 498 dio->op.write_point = writepoint_hashed((unsigned long) current); 499 dio->op.nr_replicas = dio->op.opts.data_replicas; 500 dio->op.subvol = inode->ei_subvol; 501 dio->op.pos = POS(inode->v.i_ino, (u64) req->ki_pos >> 9); 502 dio->op.devs_need_flush = &inode->ei_devs_need_flush; 503 504 if (sync) 505 dio->op.flags |= BCH_WRITE_SYNC; 506 dio->op.flags |= BCH_WRITE_CHECK_ENOSPC; 507 508 ret = bch2_quota_reservation_add(c, inode, &dio->quota_res, 509 bio_sectors(bio), true); 510 if (unlikely(ret)) 511 goto err; 512 513 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio), 514 dio->op.opts.data_replicas, 0); 515 if (unlikely(ret) && 516 !bch2_dio_write_check_allocated(dio)) 517 goto err; 518 519 task_io_account_write(bio->bi_iter.bi_size); 520 521 if (unlikely(dio->iter.count) && 522 !dio->sync && 523 !dio->loop && 524 bch2_dio_write_copy_iov(dio)) 525 dio->sync = sync = true; 526 527 dio->loop = true; 528 closure_call(&dio->op.cl, bch2_write, NULL, NULL); 529 530 if (!sync) 531 return -EIOCBQUEUED; 532 533 bch2_dio_write_end(dio); 534 535 if (likely(!dio->iter.count) || dio->op.error) 536 break; 537 538 bio_reset(bio, NULL, REQ_OP_WRITE); 539 } 540 out: 541 return bch2_dio_write_done(dio); 542 err: 543 dio->op.error = ret; 544 545 bio_release_pages(bio, false); 546 547 bch2_quota_reservation_put(c, inode, &dio->quota_res); 548 goto out; 549 } 550 551 static noinline __cold void bch2_dio_write_continue(struct dio_write *dio) 552 { 553 struct mm_struct *mm = dio->mm; 554 555 bio_reset(&dio->op.wbio.bio, NULL, REQ_OP_WRITE); 556 557 if (mm) 558 kthread_use_mm(mm); 559 bch2_dio_write_loop(dio); 560 if (mm) 561 kthread_unuse_mm(mm); 562 } 563 564 static void bch2_dio_write_loop_async(struct bch_write_op *op) 565 { 566 struct dio_write *dio = container_of(op, struct dio_write, op); 567 568 bch2_dio_write_end(dio); 569 570 if (likely(!dio->iter.count) || dio->op.error) 571 bch2_dio_write_done(dio); 572 else 573 bch2_dio_write_continue(dio); 574 } 575 576 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter) 577 { 578 struct file *file = req->ki_filp; 579 struct address_space *mapping = file->f_mapping; 580 struct bch_inode_info *inode = file_bch_inode(file); 581 struct bch_fs *c = inode->v.i_sb->s_fs_info; 582 struct dio_write *dio; 583 struct bio *bio; 584 bool locked = true, extending; 585 ssize_t ret; 586 587 prefetch(&c->opts); 588 prefetch((void *) &c->opts + 64); 589 prefetch(&inode->ei_inode); 590 prefetch((void *) &inode->ei_inode + 64); 591 592 inode_lock(&inode->v); 593 594 ret = generic_write_checks(req, iter); 595 if (unlikely(ret <= 0)) 596 goto err; 597 598 ret = file_remove_privs(file); 599 if (unlikely(ret)) 600 goto err; 601 602 ret = file_update_time(file); 603 if (unlikely(ret)) 604 goto err; 605 606 if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1))) 607 goto err; 608 609 inode_dio_begin(&inode->v); 610 bch2_pagecache_block_get(inode); 611 612 extending = req->ki_pos + iter->count > inode->v.i_size; 613 if (!extending) { 614 inode_unlock(&inode->v); 615 locked = false; 616 } 617 618 bio = bio_alloc_bioset(NULL, 619 bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS), 620 REQ_OP_WRITE, 621 GFP_KERNEL, 622 &c->dio_write_bioset); 623 dio = container_of(bio, struct dio_write, op.wbio.bio); 624 dio->req = req; 625 dio->mapping = mapping; 626 dio->inode = inode; 627 dio->mm = current->mm; 628 dio->loop = false; 629 dio->extending = extending; 630 dio->sync = is_sync_kiocb(req) || extending; 631 dio->flush = iocb_is_dsync(req) && !c->opts.journal_flush_disabled; 632 dio->free_iov = false; 633 dio->quota_res.sectors = 0; 634 dio->written = 0; 635 dio->iter = *iter; 636 dio->op.c = c; 637 638 if (unlikely(mapping->nrpages)) { 639 ret = bch2_write_invalidate_inode_pages_range(mapping, 640 req->ki_pos, 641 req->ki_pos + iter->count - 1); 642 if (unlikely(ret)) 643 goto err_put_bio; 644 } 645 646 ret = bch2_dio_write_loop(dio); 647 err: 648 if (locked) 649 inode_unlock(&inode->v); 650 return ret; 651 err_put_bio: 652 bch2_pagecache_block_put(inode); 653 bio_put(bio); 654 inode_dio_end(&inode->v); 655 goto err; 656 } 657 658 void bch2_fs_fs_io_direct_exit(struct bch_fs *c) 659 { 660 bioset_exit(&c->dio_write_bioset); 661 bioset_exit(&c->dio_read_bioset); 662 } 663 664 int bch2_fs_fs_io_direct_init(struct bch_fs *c) 665 { 666 if (bioset_init(&c->dio_read_bioset, 667 4, offsetof(struct dio_read, rbio.bio), 668 BIOSET_NEED_BVECS)) 669 return -BCH_ERR_ENOMEM_dio_read_bioset_init; 670 671 if (bioset_init(&c->dio_write_bioset, 672 4, offsetof(struct dio_write, op.wbio.bio), 673 BIOSET_NEED_BVECS)) 674 return -BCH_ERR_ENOMEM_dio_write_bioset_init; 675 676 return 0; 677 } 678 679 #endif /* NO_BCACHEFS_FS */ 680