1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 1991, 1992 Linus Torvalds 4 * Copyright (C) 2001 Andrea Arcangeli <andrea@suse.de> SuSE 5 * Copyright (C) 2016 - 2020 Christoph Hellwig 6 */ 7 #include <linux/init.h> 8 #include <linux/mm.h> 9 #include <linux/blkdev.h> 10 #include <linux/buffer_head.h> 11 #include <linux/mpage.h> 12 #include <linux/uio.h> 13 #include <linux/namei.h> 14 #include <linux/task_io_accounting_ops.h> 15 #include <linux/falloc.h> 16 #include <linux/suspend.h> 17 #include <linux/fs.h> 18 #include <linux/module.h> 19 #include "blk.h" 20 21 static inline struct inode *bdev_file_inode(struct file *file) 22 { 23 return file->f_mapping->host; 24 } 25 26 static int blkdev_get_block(struct inode *inode, sector_t iblock, 27 struct buffer_head *bh, int create) 28 { 29 bh->b_bdev = I_BDEV(inode); 30 bh->b_blocknr = iblock; 31 set_buffer_mapped(bh); 32 return 0; 33 } 34 35 static blk_opf_t dio_bio_write_op(struct kiocb *iocb) 36 { 37 blk_opf_t opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE; 38 39 /* avoid the need for a I/O completion work item */ 40 if (iocb_is_dsync(iocb)) 41 opf |= REQ_FUA; 42 return opf; 43 } 44 45 static bool blkdev_dio_unaligned(struct block_device *bdev, loff_t pos, 46 struct iov_iter *iter) 47 { 48 return pos & (bdev_logical_block_size(bdev) - 1) || 49 !bdev_iter_is_aligned(bdev, iter); 50 } 51 52 #define DIO_INLINE_BIO_VECS 4 53 54 static ssize_t __blkdev_direct_IO_simple(struct kiocb *iocb, 55 struct iov_iter *iter, unsigned int nr_pages) 56 { 57 struct block_device *bdev = iocb->ki_filp->private_data; 58 struct bio_vec inline_vecs[DIO_INLINE_BIO_VECS], *vecs; 59 loff_t pos = iocb->ki_pos; 60 bool should_dirty = false; 61 struct bio bio; 62 ssize_t ret; 63 64 if (blkdev_dio_unaligned(bdev, pos, iter)) 65 return -EINVAL; 66 67 if (nr_pages <= DIO_INLINE_BIO_VECS) 68 vecs = inline_vecs; 69 else { 70 vecs = kmalloc_array(nr_pages, sizeof(struct bio_vec), 71 GFP_KERNEL); 72 if (!vecs) 73 return -ENOMEM; 74 } 75 76 if (iov_iter_rw(iter) == READ) { 77 bio_init(&bio, bdev, vecs, nr_pages, REQ_OP_READ); 78 if (user_backed_iter(iter)) 79 should_dirty = true; 80 } else { 81 bio_init(&bio, bdev, vecs, nr_pages, dio_bio_write_op(iocb)); 82 } 83 bio.bi_iter.bi_sector = pos >> SECTOR_SHIFT; 84 bio.bi_ioprio = iocb->ki_ioprio; 85 86 ret = bio_iov_iter_get_pages(&bio, iter); 87 if (unlikely(ret)) 88 goto out; 89 ret = bio.bi_iter.bi_size; 90 91 if (iov_iter_rw(iter) == WRITE) 92 task_io_account_write(ret); 93 94 if (iocb->ki_flags & IOCB_NOWAIT) 95 bio.bi_opf |= REQ_NOWAIT; 96 97 submit_bio_wait(&bio); 98 99 bio_release_pages(&bio, should_dirty); 100 if (unlikely(bio.bi_status)) 101 ret = blk_status_to_errno(bio.bi_status); 102 103 out: 104 if (vecs != inline_vecs) 105 kfree(vecs); 106 107 bio_uninit(&bio); 108 109 return ret; 110 } 111 112 enum { 113 DIO_SHOULD_DIRTY = 1, 114 DIO_IS_SYNC = 2, 115 }; 116 117 struct blkdev_dio { 118 union { 119 struct kiocb *iocb; 120 struct task_struct *waiter; 121 }; 122 size_t size; 123 atomic_t ref; 124 unsigned int flags; 125 struct bio bio ____cacheline_aligned_in_smp; 126 }; 127 128 static struct bio_set blkdev_dio_pool; 129 130 static void blkdev_bio_end_io(struct bio *bio) 131 { 132 struct blkdev_dio *dio = bio->bi_private; 133 bool should_dirty = dio->flags & DIO_SHOULD_DIRTY; 134 135 if (bio->bi_status && !dio->bio.bi_status) 136 dio->bio.bi_status = bio->bi_status; 137 138 if (atomic_dec_and_test(&dio->ref)) { 139 if (!(dio->flags & DIO_IS_SYNC)) { 140 struct kiocb *iocb = dio->iocb; 141 ssize_t ret; 142 143 WRITE_ONCE(iocb->private, NULL); 144 145 if (likely(!dio->bio.bi_status)) { 146 ret = dio->size; 147 iocb->ki_pos += ret; 148 } else { 149 ret = blk_status_to_errno(dio->bio.bi_status); 150 } 151 152 dio->iocb->ki_complete(iocb, ret); 153 bio_put(&dio->bio); 154 } else { 155 struct task_struct *waiter = dio->waiter; 156 157 WRITE_ONCE(dio->waiter, NULL); 158 blk_wake_io_task(waiter); 159 } 160 } 161 162 if (should_dirty) { 163 bio_check_pages_dirty(bio); 164 } else { 165 bio_release_pages(bio, false); 166 bio_put(bio); 167 } 168 } 169 170 static ssize_t __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, 171 unsigned int nr_pages) 172 { 173 struct block_device *bdev = iocb->ki_filp->private_data; 174 struct blk_plug plug; 175 struct blkdev_dio *dio; 176 struct bio *bio; 177 bool is_read = (iov_iter_rw(iter) == READ), is_sync; 178 blk_opf_t opf = is_read ? REQ_OP_READ : dio_bio_write_op(iocb); 179 loff_t pos = iocb->ki_pos; 180 int ret = 0; 181 182 if (blkdev_dio_unaligned(bdev, pos, iter)) 183 return -EINVAL; 184 185 if (iocb->ki_flags & IOCB_ALLOC_CACHE) 186 opf |= REQ_ALLOC_CACHE; 187 bio = bio_alloc_bioset(bdev, nr_pages, opf, GFP_KERNEL, 188 &blkdev_dio_pool); 189 dio = container_of(bio, struct blkdev_dio, bio); 190 atomic_set(&dio->ref, 1); 191 /* 192 * Grab an extra reference to ensure the dio structure which is embedded 193 * into the first bio stays around. 194 */ 195 bio_get(bio); 196 197 is_sync = is_sync_kiocb(iocb); 198 if (is_sync) { 199 dio->flags = DIO_IS_SYNC; 200 dio->waiter = current; 201 } else { 202 dio->flags = 0; 203 dio->iocb = iocb; 204 } 205 206 dio->size = 0; 207 if (is_read && user_backed_iter(iter)) 208 dio->flags |= DIO_SHOULD_DIRTY; 209 210 blk_start_plug(&plug); 211 212 for (;;) { 213 bio->bi_iter.bi_sector = pos >> SECTOR_SHIFT; 214 bio->bi_private = dio; 215 bio->bi_end_io = blkdev_bio_end_io; 216 bio->bi_ioprio = iocb->ki_ioprio; 217 218 ret = bio_iov_iter_get_pages(bio, iter); 219 if (unlikely(ret)) { 220 bio->bi_status = BLK_STS_IOERR; 221 bio_endio(bio); 222 break; 223 } 224 225 if (is_read) { 226 if (dio->flags & DIO_SHOULD_DIRTY) 227 bio_set_pages_dirty(bio); 228 } else { 229 task_io_account_write(bio->bi_iter.bi_size); 230 } 231 if (iocb->ki_flags & IOCB_NOWAIT) 232 bio->bi_opf |= REQ_NOWAIT; 233 234 dio->size += bio->bi_iter.bi_size; 235 pos += bio->bi_iter.bi_size; 236 237 nr_pages = bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS); 238 if (!nr_pages) { 239 submit_bio(bio); 240 break; 241 } 242 atomic_inc(&dio->ref); 243 submit_bio(bio); 244 bio = bio_alloc(bdev, nr_pages, opf, GFP_KERNEL); 245 } 246 247 blk_finish_plug(&plug); 248 249 if (!is_sync) 250 return -EIOCBQUEUED; 251 252 for (;;) { 253 set_current_state(TASK_UNINTERRUPTIBLE); 254 if (!READ_ONCE(dio->waiter)) 255 break; 256 blk_io_schedule(); 257 } 258 __set_current_state(TASK_RUNNING); 259 260 if (!ret) 261 ret = blk_status_to_errno(dio->bio.bi_status); 262 if (likely(!ret)) 263 ret = dio->size; 264 265 bio_put(&dio->bio); 266 return ret; 267 } 268 269 static void blkdev_bio_end_io_async(struct bio *bio) 270 { 271 struct blkdev_dio *dio = container_of(bio, struct blkdev_dio, bio); 272 struct kiocb *iocb = dio->iocb; 273 ssize_t ret; 274 275 WRITE_ONCE(iocb->private, NULL); 276 277 if (likely(!bio->bi_status)) { 278 ret = dio->size; 279 iocb->ki_pos += ret; 280 } else { 281 ret = blk_status_to_errno(bio->bi_status); 282 } 283 284 iocb->ki_complete(iocb, ret); 285 286 if (dio->flags & DIO_SHOULD_DIRTY) { 287 bio_check_pages_dirty(bio); 288 } else { 289 bio_release_pages(bio, false); 290 bio_put(bio); 291 } 292 } 293 294 static ssize_t __blkdev_direct_IO_async(struct kiocb *iocb, 295 struct iov_iter *iter, 296 unsigned int nr_pages) 297 { 298 struct block_device *bdev = iocb->ki_filp->private_data; 299 bool is_read = iov_iter_rw(iter) == READ; 300 blk_opf_t opf = is_read ? REQ_OP_READ : dio_bio_write_op(iocb); 301 struct blkdev_dio *dio; 302 struct bio *bio; 303 loff_t pos = iocb->ki_pos; 304 int ret = 0; 305 306 if (blkdev_dio_unaligned(bdev, pos, iter)) 307 return -EINVAL; 308 309 if (iocb->ki_flags & IOCB_ALLOC_CACHE) 310 opf |= REQ_ALLOC_CACHE; 311 bio = bio_alloc_bioset(bdev, nr_pages, opf, GFP_KERNEL, 312 &blkdev_dio_pool); 313 dio = container_of(bio, struct blkdev_dio, bio); 314 dio->flags = 0; 315 dio->iocb = iocb; 316 bio->bi_iter.bi_sector = pos >> SECTOR_SHIFT; 317 bio->bi_end_io = blkdev_bio_end_io_async; 318 bio->bi_ioprio = iocb->ki_ioprio; 319 320 if (iov_iter_is_bvec(iter)) { 321 /* 322 * Users don't rely on the iterator being in any particular 323 * state for async I/O returning -EIOCBQUEUED, hence we can 324 * avoid expensive iov_iter_advance(). Bypass 325 * bio_iov_iter_get_pages() and set the bvec directly. 326 */ 327 bio_iov_bvec_set(bio, iter); 328 } else { 329 ret = bio_iov_iter_get_pages(bio, iter); 330 if (unlikely(ret)) { 331 bio_put(bio); 332 return ret; 333 } 334 } 335 dio->size = bio->bi_iter.bi_size; 336 337 if (is_read) { 338 if (user_backed_iter(iter)) { 339 dio->flags |= DIO_SHOULD_DIRTY; 340 bio_set_pages_dirty(bio); 341 } 342 } else { 343 task_io_account_write(bio->bi_iter.bi_size); 344 } 345 346 if (iocb->ki_flags & IOCB_HIPRI) { 347 bio->bi_opf |= REQ_POLLED | REQ_NOWAIT; 348 submit_bio(bio); 349 WRITE_ONCE(iocb->private, bio); 350 } else { 351 if (iocb->ki_flags & IOCB_NOWAIT) 352 bio->bi_opf |= REQ_NOWAIT; 353 submit_bio(bio); 354 } 355 return -EIOCBQUEUED; 356 } 357 358 static ssize_t blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter) 359 { 360 unsigned int nr_pages; 361 362 if (!iov_iter_count(iter)) 363 return 0; 364 365 nr_pages = bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS + 1); 366 if (likely(nr_pages <= BIO_MAX_VECS)) { 367 if (is_sync_kiocb(iocb)) 368 return __blkdev_direct_IO_simple(iocb, iter, nr_pages); 369 return __blkdev_direct_IO_async(iocb, iter, nr_pages); 370 } 371 return __blkdev_direct_IO(iocb, iter, bio_max_segs(nr_pages)); 372 } 373 374 static int blkdev_writepage(struct page *page, struct writeback_control *wbc) 375 { 376 return block_write_full_page(page, blkdev_get_block, wbc); 377 } 378 379 static int blkdev_read_folio(struct file *file, struct folio *folio) 380 { 381 return block_read_full_folio(folio, blkdev_get_block); 382 } 383 384 static void blkdev_readahead(struct readahead_control *rac) 385 { 386 mpage_readahead(rac, blkdev_get_block); 387 } 388 389 static int blkdev_write_begin(struct file *file, struct address_space *mapping, 390 loff_t pos, unsigned len, struct page **pagep, void **fsdata) 391 { 392 return block_write_begin(mapping, pos, len, pagep, blkdev_get_block); 393 } 394 395 static int blkdev_write_end(struct file *file, struct address_space *mapping, 396 loff_t pos, unsigned len, unsigned copied, struct page *page, 397 void *fsdata) 398 { 399 int ret; 400 ret = block_write_end(file, mapping, pos, len, copied, page, fsdata); 401 402 unlock_page(page); 403 put_page(page); 404 405 return ret; 406 } 407 408 const struct address_space_operations def_blk_aops = { 409 .dirty_folio = block_dirty_folio, 410 .invalidate_folio = block_invalidate_folio, 411 .read_folio = blkdev_read_folio, 412 .readahead = blkdev_readahead, 413 .writepage = blkdev_writepage, 414 .write_begin = blkdev_write_begin, 415 .write_end = blkdev_write_end, 416 .direct_IO = blkdev_direct_IO, 417 .migrate_folio = buffer_migrate_folio_norefs, 418 .is_dirty_writeback = buffer_check_dirty_writeback, 419 }; 420 421 /* 422 * for a block special file file_inode(file)->i_size is zero 423 * so we compute the size by hand (just as in block_read/write above) 424 */ 425 static loff_t blkdev_llseek(struct file *file, loff_t offset, int whence) 426 { 427 struct inode *bd_inode = bdev_file_inode(file); 428 loff_t retval; 429 430 inode_lock(bd_inode); 431 retval = fixed_size_llseek(file, offset, whence, i_size_read(bd_inode)); 432 inode_unlock(bd_inode); 433 return retval; 434 } 435 436 static int blkdev_fsync(struct file *filp, loff_t start, loff_t end, 437 int datasync) 438 { 439 struct block_device *bdev = filp->private_data; 440 int error; 441 442 error = file_write_and_wait_range(filp, start, end); 443 if (error) 444 return error; 445 446 /* 447 * There is no need to serialise calls to blkdev_issue_flush with 448 * i_mutex and doing so causes performance issues with concurrent 449 * O_SYNC writers to a block device. 450 */ 451 error = blkdev_issue_flush(bdev); 452 if (error == -EOPNOTSUPP) 453 error = 0; 454 455 return error; 456 } 457 458 static int blkdev_open(struct inode *inode, struct file *filp) 459 { 460 struct block_device *bdev; 461 462 /* 463 * Preserve backwards compatibility and allow large file access 464 * even if userspace doesn't ask for it explicitly. Some mkfs 465 * binary needs it. We might want to drop this workaround 466 * during an unstable branch. 467 */ 468 filp->f_flags |= O_LARGEFILE; 469 filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC; 470 471 if (filp->f_flags & O_NDELAY) 472 filp->f_mode |= FMODE_NDELAY; 473 if (filp->f_flags & O_EXCL) 474 filp->f_mode |= FMODE_EXCL; 475 if ((filp->f_flags & O_ACCMODE) == 3) 476 filp->f_mode |= FMODE_WRITE_IOCTL; 477 478 bdev = blkdev_get_by_dev(inode->i_rdev, filp->f_mode, filp); 479 if (IS_ERR(bdev)) 480 return PTR_ERR(bdev); 481 482 filp->private_data = bdev; 483 filp->f_mapping = bdev->bd_inode->i_mapping; 484 filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping); 485 return 0; 486 } 487 488 static int blkdev_close(struct inode *inode, struct file *filp) 489 { 490 struct block_device *bdev = filp->private_data; 491 492 blkdev_put(bdev, filp->f_mode); 493 return 0; 494 } 495 496 /* 497 * Write data to the block device. Only intended for the block device itself 498 * and the raw driver which basically is a fake block device. 499 * 500 * Does not take i_mutex for the write and thus is not for general purpose 501 * use. 502 */ 503 static ssize_t blkdev_write_iter(struct kiocb *iocb, struct iov_iter *from) 504 { 505 struct block_device *bdev = iocb->ki_filp->private_data; 506 struct inode *bd_inode = bdev->bd_inode; 507 loff_t size = bdev_nr_bytes(bdev); 508 struct blk_plug plug; 509 size_t shorted = 0; 510 ssize_t ret; 511 512 if (bdev_read_only(bdev)) 513 return -EPERM; 514 515 if (IS_SWAPFILE(bd_inode) && !is_hibernate_resume_dev(bd_inode->i_rdev)) 516 return -ETXTBSY; 517 518 if (!iov_iter_count(from)) 519 return 0; 520 521 if (iocb->ki_pos >= size) 522 return -ENOSPC; 523 524 if ((iocb->ki_flags & (IOCB_NOWAIT | IOCB_DIRECT)) == IOCB_NOWAIT) 525 return -EOPNOTSUPP; 526 527 size -= iocb->ki_pos; 528 if (iov_iter_count(from) > size) { 529 shorted = iov_iter_count(from) - size; 530 iov_iter_truncate(from, size); 531 } 532 533 blk_start_plug(&plug); 534 ret = __generic_file_write_iter(iocb, from); 535 if (ret > 0) 536 ret = generic_write_sync(iocb, ret); 537 iov_iter_reexpand(from, iov_iter_count(from) + shorted); 538 blk_finish_plug(&plug); 539 return ret; 540 } 541 542 static ssize_t blkdev_read_iter(struct kiocb *iocb, struct iov_iter *to) 543 { 544 struct block_device *bdev = iocb->ki_filp->private_data; 545 loff_t size = bdev_nr_bytes(bdev); 546 loff_t pos = iocb->ki_pos; 547 size_t shorted = 0; 548 ssize_t ret = 0; 549 size_t count; 550 551 if (unlikely(pos + iov_iter_count(to) > size)) { 552 if (pos >= size) 553 return 0; 554 size -= pos; 555 shorted = iov_iter_count(to) - size; 556 iov_iter_truncate(to, size); 557 } 558 559 count = iov_iter_count(to); 560 if (!count) 561 goto reexpand; /* skip atime */ 562 563 if (iocb->ki_flags & IOCB_DIRECT) { 564 struct address_space *mapping = iocb->ki_filp->f_mapping; 565 566 if (iocb->ki_flags & IOCB_NOWAIT) { 567 if (filemap_range_needs_writeback(mapping, pos, 568 pos + count - 1)) { 569 ret = -EAGAIN; 570 goto reexpand; 571 } 572 } else { 573 ret = filemap_write_and_wait_range(mapping, pos, 574 pos + count - 1); 575 if (ret < 0) 576 goto reexpand; 577 } 578 579 file_accessed(iocb->ki_filp); 580 581 ret = blkdev_direct_IO(iocb, to); 582 if (ret >= 0) { 583 iocb->ki_pos += ret; 584 count -= ret; 585 } 586 iov_iter_revert(to, count - iov_iter_count(to)); 587 if (ret < 0 || !count) 588 goto reexpand; 589 } 590 591 ret = filemap_read(iocb, to, ret); 592 593 reexpand: 594 if (unlikely(shorted)) 595 iov_iter_reexpand(to, iov_iter_count(to) + shorted); 596 return ret; 597 } 598 599 #define BLKDEV_FALLOC_FL_SUPPORTED \ 600 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \ 601 FALLOC_FL_ZERO_RANGE | FALLOC_FL_NO_HIDE_STALE) 602 603 static long blkdev_fallocate(struct file *file, int mode, loff_t start, 604 loff_t len) 605 { 606 struct inode *inode = bdev_file_inode(file); 607 struct block_device *bdev = I_BDEV(inode); 608 loff_t end = start + len - 1; 609 loff_t isize; 610 int error; 611 612 /* Fail if we don't recognize the flags. */ 613 if (mode & ~BLKDEV_FALLOC_FL_SUPPORTED) 614 return -EOPNOTSUPP; 615 616 /* Don't go off the end of the device. */ 617 isize = bdev_nr_bytes(bdev); 618 if (start >= isize) 619 return -EINVAL; 620 if (end >= isize) { 621 if (mode & FALLOC_FL_KEEP_SIZE) { 622 len = isize - start; 623 end = start + len - 1; 624 } else 625 return -EINVAL; 626 } 627 628 /* 629 * Don't allow IO that isn't aligned to logical block size. 630 */ 631 if ((start | len) & (bdev_logical_block_size(bdev) - 1)) 632 return -EINVAL; 633 634 filemap_invalidate_lock(inode->i_mapping); 635 636 /* Invalidate the page cache, including dirty pages. */ 637 error = truncate_bdev_range(bdev, file->f_mode, start, end); 638 if (error) 639 goto fail; 640 641 switch (mode) { 642 case FALLOC_FL_ZERO_RANGE: 643 case FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE: 644 error = blkdev_issue_zeroout(bdev, start >> SECTOR_SHIFT, 645 len >> SECTOR_SHIFT, GFP_KERNEL, 646 BLKDEV_ZERO_NOUNMAP); 647 break; 648 case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE: 649 error = blkdev_issue_zeroout(bdev, start >> SECTOR_SHIFT, 650 len >> SECTOR_SHIFT, GFP_KERNEL, 651 BLKDEV_ZERO_NOFALLBACK); 652 break; 653 case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE | FALLOC_FL_NO_HIDE_STALE: 654 error = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT, 655 len >> SECTOR_SHIFT, GFP_KERNEL); 656 break; 657 default: 658 error = -EOPNOTSUPP; 659 } 660 661 fail: 662 filemap_invalidate_unlock(inode->i_mapping); 663 return error; 664 } 665 666 const struct file_operations def_blk_fops = { 667 .open = blkdev_open, 668 .release = blkdev_close, 669 .llseek = blkdev_llseek, 670 .read_iter = blkdev_read_iter, 671 .write_iter = blkdev_write_iter, 672 .iopoll = iocb_bio_iopoll, 673 .mmap = generic_file_mmap, 674 .fsync = blkdev_fsync, 675 .unlocked_ioctl = blkdev_ioctl, 676 #ifdef CONFIG_COMPAT 677 .compat_ioctl = compat_blkdev_ioctl, 678 #endif 679 .splice_read = generic_file_splice_read, 680 .splice_write = iter_file_splice_write, 681 .fallocate = blkdev_fallocate, 682 }; 683 684 static __init int blkdev_init(void) 685 { 686 return bioset_init(&blkdev_dio_pool, 4, 687 offsetof(struct blkdev_dio, bio), 688 BIOSET_NEED_BVECS|BIOSET_PERCPU_CACHE); 689 } 690 module_init(blkdev_init); 691