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/blk-integrity.h> 11 #include <linux/buffer_head.h> 12 #include <linux/mpage.h> 13 #include <linux/uio.h> 14 #include <linux/namei.h> 15 #include <linux/task_io_accounting_ops.h> 16 #include <linux/falloc.h> 17 #include <linux/suspend.h> 18 #include <linux/fs.h> 19 #include <linux/iomap.h> 20 #include <linux/module.h> 21 #include <linux/io_uring/cmd.h> 22 #include "blk.h" 23 24 static inline struct inode *bdev_file_inode(struct file *file) 25 { 26 return file->f_mapping->host; 27 } 28 29 static blk_opf_t dio_bio_write_op(struct kiocb *iocb) 30 { 31 blk_opf_t opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE; 32 33 /* avoid the need for a I/O completion work item */ 34 if (iocb_is_dsync(iocb)) 35 opf |= REQ_FUA; 36 return opf; 37 } 38 39 static bool blkdev_dio_invalid(struct block_device *bdev, struct kiocb *iocb, 40 struct iov_iter *iter) 41 { 42 return (iocb->ki_pos | iov_iter_count(iter)) & 43 (bdev_logical_block_size(bdev) - 1); 44 } 45 46 static inline int blkdev_iov_iter_get_pages(struct bio *bio, 47 struct iov_iter *iter, struct block_device *bdev) 48 { 49 return bio_iov_iter_get_pages(bio, iter, 50 bdev_logical_block_size(bdev) - 1); 51 } 52 53 #define DIO_INLINE_BIO_VECS 4 54 55 static ssize_t __blkdev_direct_IO_simple(struct kiocb *iocb, 56 struct iov_iter *iter, struct block_device *bdev, 57 unsigned int nr_pages) 58 { 59 struct bio_vec inline_vecs[DIO_INLINE_BIO_VECS], *vecs; 60 loff_t pos = iocb->ki_pos; 61 bool should_dirty = false; 62 struct bio bio; 63 ssize_t ret; 64 65 if (nr_pages <= DIO_INLINE_BIO_VECS) 66 vecs = inline_vecs; 67 else { 68 vecs = kmalloc_array(nr_pages, sizeof(struct bio_vec), 69 GFP_KERNEL); 70 if (!vecs) 71 return -ENOMEM; 72 } 73 74 if (iov_iter_rw(iter) == READ) { 75 bio_init(&bio, bdev, vecs, nr_pages, REQ_OP_READ); 76 if (user_backed_iter(iter)) 77 should_dirty = true; 78 } else { 79 bio_init(&bio, bdev, vecs, nr_pages, dio_bio_write_op(iocb)); 80 } 81 bio.bi_iter.bi_sector = pos >> SECTOR_SHIFT; 82 bio.bi_write_hint = file_inode(iocb->ki_filp)->i_write_hint; 83 bio.bi_write_stream = iocb->ki_write_stream; 84 bio.bi_ioprio = iocb->ki_ioprio; 85 if (iocb->ki_flags & IOCB_ATOMIC) 86 bio.bi_opf |= REQ_ATOMIC; 87 88 ret = blkdev_iov_iter_get_pages(&bio, iter, bdev); 89 if (unlikely(ret)) 90 goto out; 91 ret = bio.bi_iter.bi_size; 92 93 if (iov_iter_rw(iter) == WRITE) 94 task_io_account_write(ret); 95 96 if (iocb->ki_flags & IOCB_NOWAIT) 97 bio.bi_opf |= REQ_NOWAIT; 98 99 submit_bio_wait(&bio); 100 101 bio_release_pages(&bio, should_dirty); 102 if (unlikely(bio.bi_status)) 103 ret = blk_status_to_errno(bio.bi_status); 104 105 out: 106 if (vecs != inline_vecs) 107 kfree(vecs); 108 109 bio_uninit(&bio); 110 111 return ret; 112 } 113 114 enum { 115 DIO_SHOULD_DIRTY = 1, 116 DIO_IS_SYNC = 2, 117 }; 118 119 struct blkdev_dio { 120 union { 121 struct kiocb *iocb; 122 struct task_struct *waiter; 123 }; 124 size_t size; 125 atomic_t ref; 126 unsigned int flags; 127 struct bio bio ____cacheline_aligned_in_smp; 128 }; 129 130 static struct bio_set blkdev_dio_pool; 131 132 static void blkdev_bio_end_io(struct bio *bio) 133 { 134 struct blkdev_dio *dio = bio->bi_private; 135 bool should_dirty = dio->flags & DIO_SHOULD_DIRTY; 136 bool is_sync = dio->flags & DIO_IS_SYNC; 137 138 if (bio->bi_status && !dio->bio.bi_status) 139 dio->bio.bi_status = bio->bi_status; 140 141 if (bio_integrity(bio)) 142 bio_integrity_unmap_user(bio); 143 144 if (atomic_dec_and_test(&dio->ref)) { 145 if (!is_sync) { 146 struct kiocb *iocb = dio->iocb; 147 ssize_t ret; 148 149 WRITE_ONCE(iocb->private, NULL); 150 151 if (likely(!dio->bio.bi_status)) { 152 ret = dio->size; 153 iocb->ki_pos += ret; 154 } else { 155 ret = blk_status_to_errno(dio->bio.bi_status); 156 } 157 158 dio->iocb->ki_complete(iocb, ret); 159 bio_put(&dio->bio); 160 } else { 161 struct task_struct *waiter = dio->waiter; 162 163 WRITE_ONCE(dio->waiter, NULL); 164 blk_wake_io_task(waiter); 165 } 166 } 167 168 if (should_dirty) { 169 bio_check_pages_dirty(bio); 170 } else { 171 bio_release_pages(bio, false); 172 bio_put(bio); 173 } 174 } 175 176 static ssize_t __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, 177 struct block_device *bdev, unsigned int nr_pages) 178 { 179 struct blk_plug plug; 180 struct blkdev_dio *dio; 181 struct bio *bio; 182 bool is_read = (iov_iter_rw(iter) == READ), is_sync; 183 blk_opf_t opf = is_read ? REQ_OP_READ : dio_bio_write_op(iocb); 184 loff_t pos = iocb->ki_pos; 185 int ret = 0; 186 187 if (iocb->ki_flags & IOCB_ALLOC_CACHE) 188 opf |= REQ_ALLOC_CACHE; 189 bio = bio_alloc_bioset(bdev, nr_pages, opf, GFP_KERNEL, 190 &blkdev_dio_pool); 191 dio = container_of(bio, struct blkdev_dio, bio); 192 atomic_set(&dio->ref, 1); 193 /* 194 * Grab an extra reference to ensure the dio structure which is embedded 195 * into the first bio stays around. 196 */ 197 bio_get(bio); 198 199 is_sync = is_sync_kiocb(iocb); 200 if (is_sync) { 201 dio->flags = DIO_IS_SYNC; 202 dio->waiter = current; 203 } else { 204 dio->flags = 0; 205 dio->iocb = iocb; 206 } 207 208 dio->size = 0; 209 if (is_read && user_backed_iter(iter)) 210 dio->flags |= DIO_SHOULD_DIRTY; 211 212 blk_start_plug(&plug); 213 214 for (;;) { 215 bio->bi_iter.bi_sector = pos >> SECTOR_SHIFT; 216 bio->bi_write_hint = file_inode(iocb->ki_filp)->i_write_hint; 217 bio->bi_write_stream = iocb->ki_write_stream; 218 bio->bi_private = dio; 219 bio->bi_end_io = blkdev_bio_end_io; 220 bio->bi_ioprio = iocb->ki_ioprio; 221 222 ret = blkdev_iov_iter_get_pages(bio, iter, bdev); 223 if (unlikely(ret)) { 224 bio->bi_status = BLK_STS_IOERR; 225 bio_endio(bio); 226 break; 227 } 228 if (iocb->ki_flags & IOCB_NOWAIT) { 229 /* 230 * This is nonblocking IO, and we need to allocate 231 * another bio if we have data left to map. As we 232 * cannot guarantee that one of the sub bios will not 233 * fail getting issued FOR NOWAIT and as error results 234 * are coalesced across all of them, be safe and ask for 235 * a retry of this from blocking context. 236 */ 237 if (unlikely(iov_iter_count(iter))) { 238 ret = -EAGAIN; 239 goto fail; 240 } 241 bio->bi_opf |= REQ_NOWAIT; 242 } 243 if (iocb->ki_flags & IOCB_HAS_METADATA) { 244 ret = bio_integrity_map_iter(bio, iocb->private); 245 if (unlikely(ret)) 246 goto fail; 247 } 248 249 if (is_read) { 250 if (dio->flags & DIO_SHOULD_DIRTY) 251 bio_set_pages_dirty(bio); 252 } else { 253 task_io_account_write(bio->bi_iter.bi_size); 254 } 255 dio->size += bio->bi_iter.bi_size; 256 pos += bio->bi_iter.bi_size; 257 258 nr_pages = bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS); 259 if (!nr_pages) { 260 submit_bio(bio); 261 break; 262 } 263 atomic_inc(&dio->ref); 264 submit_bio(bio); 265 bio = bio_alloc(bdev, nr_pages, opf, GFP_KERNEL); 266 } 267 268 blk_finish_plug(&plug); 269 270 if (!is_sync) 271 return -EIOCBQUEUED; 272 273 for (;;) { 274 set_current_state(TASK_UNINTERRUPTIBLE); 275 if (!READ_ONCE(dio->waiter)) 276 break; 277 blk_io_schedule(); 278 } 279 __set_current_state(TASK_RUNNING); 280 281 if (!ret) 282 ret = blk_status_to_errno(dio->bio.bi_status); 283 if (likely(!ret)) 284 ret = dio->size; 285 286 bio_put(&dio->bio); 287 return ret; 288 fail: 289 bio_release_pages(bio, false); 290 bio_clear_flag(bio, BIO_REFFED); 291 bio_put(bio); 292 blk_finish_plug(&plug); 293 return ret; 294 } 295 296 static void blkdev_bio_end_io_async(struct bio *bio) 297 { 298 struct blkdev_dio *dio = container_of(bio, struct blkdev_dio, bio); 299 struct kiocb *iocb = dio->iocb; 300 ssize_t ret; 301 302 WRITE_ONCE(iocb->private, NULL); 303 304 if (likely(!bio->bi_status)) { 305 ret = dio->size; 306 iocb->ki_pos += ret; 307 } else { 308 ret = blk_status_to_errno(bio->bi_status); 309 } 310 311 if (bio_integrity(bio)) 312 bio_integrity_unmap_user(bio); 313 314 iocb->ki_complete(iocb, ret); 315 316 if (dio->flags & DIO_SHOULD_DIRTY) { 317 bio_check_pages_dirty(bio); 318 } else { 319 bio_release_pages(bio, false); 320 bio_put(bio); 321 } 322 } 323 324 static ssize_t __blkdev_direct_IO_async(struct kiocb *iocb, 325 struct iov_iter *iter, 326 struct block_device *bdev, 327 unsigned int nr_pages) 328 { 329 bool is_read = iov_iter_rw(iter) == READ; 330 blk_opf_t opf = is_read ? REQ_OP_READ : dio_bio_write_op(iocb); 331 struct blkdev_dio *dio; 332 struct bio *bio; 333 loff_t pos = iocb->ki_pos; 334 int ret = 0; 335 336 if (iocb->ki_flags & IOCB_ALLOC_CACHE) 337 opf |= REQ_ALLOC_CACHE; 338 bio = bio_alloc_bioset(bdev, nr_pages, opf, GFP_KERNEL, 339 &blkdev_dio_pool); 340 dio = container_of(bio, struct blkdev_dio, bio); 341 dio->flags = 0; 342 dio->iocb = iocb; 343 bio->bi_iter.bi_sector = pos >> SECTOR_SHIFT; 344 bio->bi_write_hint = file_inode(iocb->ki_filp)->i_write_hint; 345 bio->bi_write_stream = iocb->ki_write_stream; 346 bio->bi_end_io = blkdev_bio_end_io_async; 347 bio->bi_ioprio = iocb->ki_ioprio; 348 349 if (iov_iter_is_bvec(iter)) { 350 /* 351 * Users don't rely on the iterator being in any particular 352 * state for async I/O returning -EIOCBQUEUED, hence we can 353 * avoid expensive iov_iter_advance(). Bypass 354 * bio_iov_iter_get_pages() and set the bvec directly. 355 */ 356 bio_iov_bvec_set(bio, iter); 357 } else { 358 ret = blkdev_iov_iter_get_pages(bio, iter, bdev); 359 if (unlikely(ret)) 360 goto out_bio_put; 361 } 362 dio->size = bio->bi_iter.bi_size; 363 364 if (is_read) { 365 if (user_backed_iter(iter)) { 366 dio->flags |= DIO_SHOULD_DIRTY; 367 bio_set_pages_dirty(bio); 368 } 369 } else { 370 task_io_account_write(bio->bi_iter.bi_size); 371 } 372 373 if (iocb->ki_flags & IOCB_HAS_METADATA) { 374 ret = bio_integrity_map_iter(bio, iocb->private); 375 WRITE_ONCE(iocb->private, NULL); 376 if (unlikely(ret)) 377 goto out_bio_put; 378 } 379 380 if (iocb->ki_flags & IOCB_ATOMIC) 381 bio->bi_opf |= REQ_ATOMIC; 382 383 if (iocb->ki_flags & IOCB_NOWAIT) 384 bio->bi_opf |= REQ_NOWAIT; 385 386 if (iocb->ki_flags & IOCB_HIPRI) { 387 bio->bi_opf |= REQ_POLLED; 388 submit_bio(bio); 389 WRITE_ONCE(iocb->private, bio); 390 } else { 391 submit_bio(bio); 392 } 393 return -EIOCBQUEUED; 394 395 out_bio_put: 396 bio_put(bio); 397 return ret; 398 } 399 400 static ssize_t blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter) 401 { 402 struct block_device *bdev = I_BDEV(iocb->ki_filp->f_mapping->host); 403 unsigned int nr_pages; 404 405 if (!iov_iter_count(iter)) 406 return 0; 407 408 if (blkdev_dio_invalid(bdev, iocb, iter)) 409 return -EINVAL; 410 411 if (iov_iter_rw(iter) == WRITE) { 412 u16 max_write_streams = bdev_max_write_streams(bdev); 413 414 if (iocb->ki_write_stream) { 415 if (iocb->ki_write_stream > max_write_streams) 416 return -EINVAL; 417 } else if (max_write_streams) { 418 enum rw_hint write_hint = 419 file_inode(iocb->ki_filp)->i_write_hint; 420 421 /* 422 * Just use the write hint as write stream for block 423 * device writes. This assumes no file system is 424 * mounted that would use the streams differently. 425 */ 426 if (write_hint <= max_write_streams) 427 iocb->ki_write_stream = write_hint; 428 } 429 } 430 431 nr_pages = bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS + 1); 432 if (likely(nr_pages <= BIO_MAX_VECS && 433 !(iocb->ki_flags & IOCB_HAS_METADATA))) { 434 if (is_sync_kiocb(iocb)) 435 return __blkdev_direct_IO_simple(iocb, iter, bdev, 436 nr_pages); 437 return __blkdev_direct_IO_async(iocb, iter, bdev, nr_pages); 438 } else if (iocb->ki_flags & IOCB_ATOMIC) { 439 return -EINVAL; 440 } 441 return __blkdev_direct_IO(iocb, iter, bdev, bio_max_segs(nr_pages)); 442 } 443 444 static int blkdev_iomap_begin(struct inode *inode, loff_t offset, loff_t length, 445 unsigned int flags, struct iomap *iomap, struct iomap *srcmap) 446 { 447 struct block_device *bdev = I_BDEV(inode); 448 loff_t isize = i_size_read(inode); 449 450 if (offset >= isize) 451 return -EIO; 452 453 iomap->bdev = bdev; 454 iomap->offset = ALIGN_DOWN(offset, bdev_logical_block_size(bdev)); 455 iomap->type = IOMAP_MAPPED; 456 iomap->addr = iomap->offset; 457 iomap->length = isize - iomap->offset; 458 iomap->flags |= IOMAP_F_BUFFER_HEAD; /* noop for !CONFIG_BUFFER_HEAD */ 459 return 0; 460 } 461 462 static const struct iomap_ops blkdev_iomap_ops = { 463 .iomap_begin = blkdev_iomap_begin, 464 }; 465 466 #ifdef CONFIG_BUFFER_HEAD 467 static int blkdev_get_block(struct inode *inode, sector_t iblock, 468 struct buffer_head *bh, int create) 469 { 470 bh->b_bdev = I_BDEV(inode); 471 bh->b_blocknr = iblock; 472 set_buffer_mapped(bh); 473 return 0; 474 } 475 476 /* 477 * We cannot call mpage_writepages() as it does not take the buffer lock. 478 * We must use block_write_full_folio() directly which holds the buffer 479 * lock. The buffer lock provides the synchronisation with writeback 480 * that filesystems rely on when they use the blockdev's mapping. 481 */ 482 static int blkdev_writepages(struct address_space *mapping, 483 struct writeback_control *wbc) 484 { 485 struct folio *folio = NULL; 486 struct blk_plug plug; 487 int err; 488 489 blk_start_plug(&plug); 490 while ((folio = writeback_iter(mapping, wbc, folio, &err))) 491 err = block_write_full_folio(folio, wbc, blkdev_get_block); 492 blk_finish_plug(&plug); 493 494 return err; 495 } 496 497 static int blkdev_read_folio(struct file *file, struct folio *folio) 498 { 499 return block_read_full_folio(folio, blkdev_get_block); 500 } 501 502 static void blkdev_readahead(struct readahead_control *rac) 503 { 504 mpage_readahead(rac, blkdev_get_block); 505 } 506 507 static int blkdev_write_begin(const struct kiocb *iocb, 508 struct address_space *mapping, loff_t pos, 509 unsigned len, struct folio **foliop, 510 void **fsdata) 511 { 512 return block_write_begin(mapping, pos, len, foliop, blkdev_get_block); 513 } 514 515 static int blkdev_write_end(const struct kiocb *iocb, 516 struct address_space *mapping, 517 loff_t pos, unsigned len, unsigned copied, 518 struct folio *folio, void *fsdata) 519 { 520 int ret; 521 ret = block_write_end(pos, len, copied, folio); 522 523 folio_unlock(folio); 524 folio_put(folio); 525 526 return ret; 527 } 528 529 const struct address_space_operations def_blk_aops = { 530 .dirty_folio = block_dirty_folio, 531 .invalidate_folio = block_invalidate_folio, 532 .read_folio = blkdev_read_folio, 533 .readahead = blkdev_readahead, 534 .writepages = blkdev_writepages, 535 .write_begin = blkdev_write_begin, 536 .write_end = blkdev_write_end, 537 .migrate_folio = buffer_migrate_folio_norefs, 538 .is_dirty_writeback = buffer_check_dirty_writeback, 539 }; 540 #else /* CONFIG_BUFFER_HEAD */ 541 static int blkdev_read_folio(struct file *file, struct folio *folio) 542 { 543 return iomap_read_folio(folio, &blkdev_iomap_ops); 544 } 545 546 static void blkdev_readahead(struct readahead_control *rac) 547 { 548 iomap_readahead(rac, &blkdev_iomap_ops); 549 } 550 551 static ssize_t blkdev_writeback_range(struct iomap_writepage_ctx *wpc, 552 struct folio *folio, u64 offset, unsigned int len, u64 end_pos) 553 { 554 loff_t isize = i_size_read(wpc->inode); 555 556 if (WARN_ON_ONCE(offset >= isize)) 557 return -EIO; 558 559 if (offset < wpc->iomap.offset || 560 offset >= wpc->iomap.offset + wpc->iomap.length) { 561 int error; 562 563 error = blkdev_iomap_begin(wpc->inode, offset, isize - offset, 564 IOMAP_WRITE, &wpc->iomap, NULL); 565 if (error) 566 return error; 567 } 568 569 return iomap_add_to_ioend(wpc, folio, offset, end_pos, len); 570 } 571 572 static const struct iomap_writeback_ops blkdev_writeback_ops = { 573 .writeback_range = blkdev_writeback_range, 574 .writeback_submit = iomap_ioend_writeback_submit, 575 }; 576 577 static int blkdev_writepages(struct address_space *mapping, 578 struct writeback_control *wbc) 579 { 580 struct iomap_writepage_ctx wpc = { 581 .inode = mapping->host, 582 .wbc = wbc, 583 .ops = &blkdev_writeback_ops 584 }; 585 586 return iomap_writepages(&wpc); 587 } 588 589 const struct address_space_operations def_blk_aops = { 590 .dirty_folio = filemap_dirty_folio, 591 .release_folio = iomap_release_folio, 592 .invalidate_folio = iomap_invalidate_folio, 593 .read_folio = blkdev_read_folio, 594 .readahead = blkdev_readahead, 595 .writepages = blkdev_writepages, 596 .is_partially_uptodate = iomap_is_partially_uptodate, 597 .error_remove_folio = generic_error_remove_folio, 598 .migrate_folio = filemap_migrate_folio, 599 }; 600 #endif /* CONFIG_BUFFER_HEAD */ 601 602 /* 603 * for a block special file file_inode(file)->i_size is zero 604 * so we compute the size by hand (just as in block_read/write above) 605 */ 606 static loff_t blkdev_llseek(struct file *file, loff_t offset, int whence) 607 { 608 struct inode *bd_inode = bdev_file_inode(file); 609 loff_t retval; 610 611 inode_lock(bd_inode); 612 retval = fixed_size_llseek(file, offset, whence, i_size_read(bd_inode)); 613 inode_unlock(bd_inode); 614 return retval; 615 } 616 617 static int blkdev_fsync(struct file *filp, loff_t start, loff_t end, 618 int datasync) 619 { 620 struct block_device *bdev = I_BDEV(filp->f_mapping->host); 621 int error; 622 623 error = file_write_and_wait_range(filp, start, end); 624 if (error) 625 return error; 626 627 /* 628 * There is no need to serialise calls to blkdev_issue_flush with 629 * i_mutex and doing so causes performance issues with concurrent 630 * O_SYNC writers to a block device. 631 */ 632 error = blkdev_issue_flush(bdev); 633 if (error == -EOPNOTSUPP) 634 error = 0; 635 636 return error; 637 } 638 639 /** 640 * file_to_blk_mode - get block open flags from file flags 641 * @file: file whose open flags should be converted 642 * 643 * Look at file open flags and generate corresponding block open flags from 644 * them. The function works both for file just being open (e.g. during ->open 645 * callback) and for file that is already open. This is actually non-trivial 646 * (see comment in the function). 647 */ 648 blk_mode_t file_to_blk_mode(struct file *file) 649 { 650 blk_mode_t mode = 0; 651 652 if (file->f_mode & FMODE_READ) 653 mode |= BLK_OPEN_READ; 654 if (file->f_mode & FMODE_WRITE) 655 mode |= BLK_OPEN_WRITE; 656 /* 657 * do_dentry_open() clears O_EXCL from f_flags, use file->private_data 658 * to determine whether the open was exclusive for already open files. 659 */ 660 if (file->private_data) 661 mode |= BLK_OPEN_EXCL; 662 else if (file->f_flags & O_EXCL) 663 mode |= BLK_OPEN_EXCL; 664 if (file->f_flags & O_NDELAY) 665 mode |= BLK_OPEN_NDELAY; 666 667 /* 668 * If all bits in O_ACCMODE set (aka O_RDWR | O_WRONLY), the floppy 669 * driver has historically allowed ioctls as if the file was opened for 670 * writing, but does not allow and actual reads or writes. 671 */ 672 if ((file->f_flags & O_ACCMODE) == (O_RDWR | O_WRONLY)) 673 mode |= BLK_OPEN_WRITE_IOCTL; 674 675 return mode; 676 } 677 678 static int blkdev_open(struct inode *inode, struct file *filp) 679 { 680 struct block_device *bdev; 681 blk_mode_t mode; 682 int ret; 683 684 mode = file_to_blk_mode(filp); 685 /* Use the file as the holder. */ 686 if (mode & BLK_OPEN_EXCL) 687 filp->private_data = filp; 688 ret = bdev_permission(inode->i_rdev, mode, filp->private_data); 689 if (ret) 690 return ret; 691 692 bdev = blkdev_get_no_open(inode->i_rdev, true); 693 if (!bdev) 694 return -ENXIO; 695 696 if (bdev_can_atomic_write(bdev)) 697 filp->f_mode |= FMODE_CAN_ATOMIC_WRITE; 698 if (blk_get_integrity(bdev->bd_disk)) 699 filp->f_mode |= FMODE_HAS_METADATA; 700 701 ret = bdev_open(bdev, mode, filp->private_data, NULL, filp); 702 if (ret) 703 blkdev_put_no_open(bdev); 704 return ret; 705 } 706 707 static int blkdev_release(struct inode *inode, struct file *filp) 708 { 709 bdev_release(filp); 710 return 0; 711 } 712 713 static ssize_t 714 blkdev_direct_write(struct kiocb *iocb, struct iov_iter *from) 715 { 716 size_t count = iov_iter_count(from); 717 ssize_t written; 718 719 written = kiocb_invalidate_pages(iocb, count); 720 if (written) { 721 if (written == -EBUSY) 722 return 0; 723 return written; 724 } 725 726 written = blkdev_direct_IO(iocb, from); 727 if (written > 0) { 728 kiocb_invalidate_post_direct_write(iocb, count); 729 iocb->ki_pos += written; 730 count -= written; 731 } 732 if (written != -EIOCBQUEUED) 733 iov_iter_revert(from, count - iov_iter_count(from)); 734 return written; 735 } 736 737 static ssize_t blkdev_buffered_write(struct kiocb *iocb, struct iov_iter *from) 738 { 739 return iomap_file_buffered_write(iocb, from, &blkdev_iomap_ops, NULL, 740 NULL); 741 } 742 743 /* 744 * Write data to the block device. Only intended for the block device itself 745 * and the raw driver which basically is a fake block device. 746 * 747 * Does not take i_mutex for the write and thus is not for general purpose 748 * use. 749 */ 750 static ssize_t blkdev_write_iter(struct kiocb *iocb, struct iov_iter *from) 751 { 752 struct file *file = iocb->ki_filp; 753 struct inode *bd_inode = bdev_file_inode(file); 754 struct block_device *bdev = I_BDEV(bd_inode); 755 bool atomic = iocb->ki_flags & IOCB_ATOMIC; 756 loff_t size = bdev_nr_bytes(bdev); 757 size_t shorted = 0; 758 ssize_t ret; 759 760 if (bdev_read_only(bdev)) 761 return -EPERM; 762 763 if (IS_SWAPFILE(bd_inode) && !is_hibernate_resume_dev(bd_inode->i_rdev)) 764 return -ETXTBSY; 765 766 if (!iov_iter_count(from)) 767 return 0; 768 769 if (iocb->ki_pos >= size) 770 return -ENOSPC; 771 772 if ((iocb->ki_flags & (IOCB_NOWAIT | IOCB_DIRECT)) == IOCB_NOWAIT) 773 return -EOPNOTSUPP; 774 775 if (atomic) { 776 ret = generic_atomic_write_valid(iocb, from); 777 if (ret) 778 return ret; 779 } 780 781 size -= iocb->ki_pos; 782 if (iov_iter_count(from) > size) { 783 if (atomic) 784 return -EINVAL; 785 shorted = iov_iter_count(from) - size; 786 iov_iter_truncate(from, size); 787 } 788 789 ret = file_update_time(file); 790 if (ret) 791 return ret; 792 793 if (iocb->ki_flags & IOCB_DIRECT) { 794 ret = blkdev_direct_write(iocb, from); 795 if (ret >= 0 && iov_iter_count(from)) 796 ret = direct_write_fallback(iocb, from, ret, 797 blkdev_buffered_write(iocb, from)); 798 } else { 799 /* 800 * Take i_rwsem and invalidate_lock to avoid racing with 801 * set_blocksize changing i_blkbits/folio order and punching 802 * out the pagecache. 803 */ 804 inode_lock_shared(bd_inode); 805 ret = blkdev_buffered_write(iocb, from); 806 inode_unlock_shared(bd_inode); 807 } 808 809 if (ret > 0) 810 ret = generic_write_sync(iocb, ret); 811 iov_iter_reexpand(from, iov_iter_count(from) + shorted); 812 return ret; 813 } 814 815 static ssize_t blkdev_read_iter(struct kiocb *iocb, struct iov_iter *to) 816 { 817 struct inode *bd_inode = bdev_file_inode(iocb->ki_filp); 818 struct block_device *bdev = I_BDEV(iocb->ki_filp->f_mapping->host); 819 loff_t size = bdev_nr_bytes(bdev); 820 loff_t pos = iocb->ki_pos; 821 size_t shorted = 0; 822 ssize_t ret = 0; 823 size_t count; 824 825 if (unlikely(pos + iov_iter_count(to) > size)) { 826 if (pos >= size) 827 return 0; 828 size -= pos; 829 shorted = iov_iter_count(to) - size; 830 iov_iter_truncate(to, size); 831 } 832 833 count = iov_iter_count(to); 834 if (!count) 835 goto reexpand; /* skip atime */ 836 837 if (iocb->ki_flags & IOCB_DIRECT) { 838 ret = kiocb_write_and_wait(iocb, count); 839 if (ret < 0) 840 goto reexpand; 841 file_accessed(iocb->ki_filp); 842 843 ret = blkdev_direct_IO(iocb, to); 844 if (ret > 0) { 845 iocb->ki_pos += ret; 846 count -= ret; 847 } 848 if (ret != -EIOCBQUEUED) 849 iov_iter_revert(to, count - iov_iter_count(to)); 850 if (ret < 0 || !count) 851 goto reexpand; 852 } 853 854 /* 855 * Take i_rwsem and invalidate_lock to avoid racing with set_blocksize 856 * changing i_blkbits/folio order and punching out the pagecache. 857 */ 858 inode_lock_shared(bd_inode); 859 ret = filemap_read(iocb, to, ret); 860 inode_unlock_shared(bd_inode); 861 862 reexpand: 863 if (unlikely(shorted)) 864 iov_iter_reexpand(to, iov_iter_count(to) + shorted); 865 return ret; 866 } 867 868 #define BLKDEV_FALLOC_FL_SUPPORTED \ 869 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \ 870 FALLOC_FL_ZERO_RANGE | FALLOC_FL_WRITE_ZEROES) 871 872 static long blkdev_fallocate(struct file *file, int mode, loff_t start, 873 loff_t len) 874 { 875 struct inode *inode = bdev_file_inode(file); 876 struct block_device *bdev = I_BDEV(inode); 877 loff_t end = start + len - 1; 878 loff_t isize; 879 unsigned int flags; 880 int error; 881 882 /* Fail if we don't recognize the flags. */ 883 if (mode & ~BLKDEV_FALLOC_FL_SUPPORTED) 884 return -EOPNOTSUPP; 885 /* 886 * Don't allow writing zeroes if the device does not enable the 887 * unmap write zeroes operation. 888 */ 889 if ((mode & FALLOC_FL_WRITE_ZEROES) && 890 !bdev_write_zeroes_unmap_sectors(bdev)) 891 return -EOPNOTSUPP; 892 893 /* Don't go off the end of the device. */ 894 isize = bdev_nr_bytes(bdev); 895 if (start >= isize) 896 return -EINVAL; 897 if (end >= isize) { 898 if (mode & FALLOC_FL_KEEP_SIZE) { 899 len = isize - start; 900 end = start + len - 1; 901 } else 902 return -EINVAL; 903 } 904 905 /* 906 * Don't allow IO that isn't aligned to logical block size. 907 */ 908 if ((start | len) & (bdev_logical_block_size(bdev) - 1)) 909 return -EINVAL; 910 911 inode_lock(inode); 912 filemap_invalidate_lock(inode->i_mapping); 913 914 switch (mode) { 915 case FALLOC_FL_ZERO_RANGE: 916 case FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE: 917 flags = BLKDEV_ZERO_NOUNMAP; 918 break; 919 case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE: 920 flags = BLKDEV_ZERO_NOFALLBACK; 921 break; 922 case FALLOC_FL_WRITE_ZEROES: 923 flags = 0; 924 break; 925 default: 926 error = -EOPNOTSUPP; 927 goto fail; 928 } 929 930 /* 931 * Invalidate the page cache, including dirty pages, for valid 932 * de-allocate mode calls to fallocate(). 933 */ 934 error = truncate_bdev_range(bdev, file_to_blk_mode(file), start, end); 935 if (error) 936 goto fail; 937 938 error = blkdev_issue_zeroout(bdev, start >> SECTOR_SHIFT, 939 len >> SECTOR_SHIFT, GFP_KERNEL, flags); 940 fail: 941 filemap_invalidate_unlock(inode->i_mapping); 942 inode_unlock(inode); 943 return error; 944 } 945 946 static int blkdev_mmap_prepare(struct vm_area_desc *desc) 947 { 948 struct file *file = desc->file; 949 950 if (bdev_read_only(I_BDEV(bdev_file_inode(file)))) 951 return generic_file_readonly_mmap_prepare(desc); 952 953 return generic_file_mmap_prepare(desc); 954 } 955 956 const struct file_operations def_blk_fops = { 957 .open = blkdev_open, 958 .release = blkdev_release, 959 .llseek = blkdev_llseek, 960 .read_iter = blkdev_read_iter, 961 .write_iter = blkdev_write_iter, 962 .iopoll = iocb_bio_iopoll, 963 .mmap_prepare = blkdev_mmap_prepare, 964 .fsync = blkdev_fsync, 965 .unlocked_ioctl = blkdev_ioctl, 966 #ifdef CONFIG_COMPAT 967 .compat_ioctl = compat_blkdev_ioctl, 968 #endif 969 .splice_read = filemap_splice_read, 970 .splice_write = iter_file_splice_write, 971 .fallocate = blkdev_fallocate, 972 .uring_cmd = blkdev_uring_cmd, 973 .fop_flags = FOP_BUFFER_RASYNC, 974 }; 975 976 static __init int blkdev_init(void) 977 { 978 return bioset_init(&blkdev_dio_pool, 4, 979 offsetof(struct blkdev_dio, bio), 980 BIOSET_NEED_BVECS|BIOSET_PERCPU_CACHE); 981 } 982 module_init(blkdev_init); 983