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, BIO_MAX_SIZE, 50 bdev_dma_alignment(bdev), 51 bdev_logical_block_size(bdev) - 1); 52 } 53 54 #define DIO_INLINE_BIO_VECS 4 55 56 static ssize_t __blkdev_direct_IO_simple(struct kiocb *iocb, 57 struct iov_iter *iter, struct block_device *bdev, 58 unsigned int nr_pages) 59 { 60 struct bio_vec inline_vecs[DIO_INLINE_BIO_VECS], *vecs; 61 loff_t pos = iocb->ki_pos; 62 bool should_dirty = false; 63 struct bio bio; 64 ssize_t ret; 65 66 if (nr_pages <= DIO_INLINE_BIO_VECS) 67 vecs = inline_vecs; 68 else { 69 vecs = kmalloc_objs(struct bio_vec, nr_pages); 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 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_write_hint = file_inode(iocb->ki_filp)->i_write_hint; 215 bio->bi_write_stream = iocb->ki_write_stream; 216 bio->bi_private = dio; 217 bio->bi_end_io = blkdev_bio_end_io; 218 bio->bi_ioprio = iocb->ki_ioprio; 219 220 ret = blkdev_iov_iter_get_pages(bio, iter, bdev); 221 if (unlikely(ret)) { 222 bio_endio_status(bio, errno_to_blk_status(ret)); 223 break; 224 } 225 if (iocb->ki_flags & IOCB_NOWAIT) { 226 /* 227 * This is nonblocking IO, and we need to allocate 228 * another bio if we have data left to map. As we 229 * cannot guarantee that one of the sub bios will not 230 * fail getting issued FOR NOWAIT and as error results 231 * are coalesced across all of them, be safe and ask for 232 * a retry of this from blocking context. 233 */ 234 if (unlikely(iov_iter_count(iter))) { 235 ret = -EAGAIN; 236 goto fail; 237 } 238 bio->bi_opf |= REQ_NOWAIT; 239 } 240 if (iocb->ki_flags & IOCB_HAS_METADATA) { 241 ret = bio_integrity_map_iter(bio, iocb->private); 242 if (unlikely(ret)) { 243 bio_endio_status(bio, errno_to_blk_status(ret)); 244 break; 245 } 246 } 247 248 if (is_read) { 249 if (dio->flags & DIO_SHOULD_DIRTY) 250 bio_set_pages_dirty(bio); 251 } else { 252 task_io_account_write(bio->bi_iter.bi_size); 253 } 254 dio->size += bio->bi_iter.bi_size; 255 pos += bio->bi_iter.bi_size; 256 257 nr_pages = bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS); 258 if (!nr_pages) { 259 submit_bio(bio); 260 break; 261 } 262 atomic_inc(&dio->ref); 263 submit_bio(bio); 264 bio = bio_alloc(bdev, nr_pages, opf, GFP_KERNEL); 265 } 266 267 blk_finish_plug(&plug); 268 269 if (!is_sync) 270 return -EIOCBQUEUED; 271 272 for (;;) { 273 set_current_state(TASK_UNINTERRUPTIBLE); 274 if (!READ_ONCE(dio->waiter)) 275 break; 276 blk_io_schedule(); 277 } 278 __set_current_state(TASK_RUNNING); 279 280 if (!ret) 281 ret = blk_status_to_errno(dio->bio.bi_status); 282 if (likely(!ret)) 283 ret = dio->size; 284 285 bio_put(&dio->bio); 286 return ret; 287 fail: 288 bio_release_pages(bio, false); 289 bio_clear_flag(bio, BIO_REFFED); 290 bio_put(bio); 291 blk_finish_plug(&plug); 292 return ret; 293 } 294 295 static void blkdev_bio_end_io_async(struct bio *bio) 296 { 297 struct blkdev_dio *dio = container_of(bio, struct blkdev_dio, bio); 298 struct kiocb *iocb = dio->iocb; 299 ssize_t ret; 300 301 WRITE_ONCE(iocb->private, NULL); 302 303 if (likely(!bio->bi_status)) { 304 ret = dio->size; 305 iocb->ki_pos += ret; 306 } else { 307 ret = blk_status_to_errno(bio->bi_status); 308 } 309 310 if (bio_integrity(bio)) 311 bio_integrity_unmap_user(bio); 312 313 iocb->ki_complete(iocb, ret); 314 315 if (dio->flags & DIO_SHOULD_DIRTY) { 316 bio_check_pages_dirty(bio); 317 } else { 318 bio_release_pages(bio, false); 319 bio_put(bio); 320 } 321 } 322 323 static ssize_t __blkdev_direct_IO_async(struct kiocb *iocb, 324 struct iov_iter *iter, 325 struct block_device *bdev, 326 unsigned int nr_pages) 327 { 328 bool is_read = iov_iter_rw(iter) == READ; 329 blk_opf_t opf = is_read ? REQ_OP_READ : dio_bio_write_op(iocb); 330 struct blkdev_dio *dio; 331 struct bio *bio; 332 loff_t pos = iocb->ki_pos; 333 int ret = 0; 334 335 bio = bio_alloc_bioset(bdev, nr_pages, opf, GFP_KERNEL, 336 &blkdev_dio_pool); 337 dio = container_of(bio, struct blkdev_dio, bio); 338 dio->flags = 0; 339 dio->iocb = iocb; 340 bio->bi_iter.bi_sector = pos >> SECTOR_SHIFT; 341 bio->bi_write_hint = file_inode(iocb->ki_filp)->i_write_hint; 342 bio->bi_write_stream = iocb->ki_write_stream; 343 bio->bi_end_io = blkdev_bio_end_io_async; 344 bio->bi_ioprio = iocb->ki_ioprio; 345 346 /* 347 * Users don't rely on the iterator being in any particular 348 * state for async I/O returning -EIOCBQUEUED, hence we can 349 * avoid expensive iov_iter_advance(). Bypass 350 * bio_iov_iter_get_pages() and set the bvec directly. 351 */ 352 if (!bio_iov_iter_set(bio, iter)) { 353 ret = blkdev_iov_iter_get_pages(bio, iter, bdev); 354 if (unlikely(ret)) 355 goto out_bio_put; 356 } 357 dio->size = bio->bi_iter.bi_size; 358 359 if (is_read) { 360 if (user_backed_iter(iter)) { 361 dio->flags |= DIO_SHOULD_DIRTY; 362 bio_set_pages_dirty(bio); 363 } 364 } else { 365 task_io_account_write(bio->bi_iter.bi_size); 366 } 367 368 if (iocb->ki_flags & IOCB_HAS_METADATA) { 369 ret = bio_integrity_map_iter(bio, iocb->private); 370 WRITE_ONCE(iocb->private, NULL); 371 if (unlikely(ret)) 372 goto out_bio_put; 373 } 374 375 if (iocb->ki_flags & IOCB_ATOMIC) 376 bio->bi_opf |= REQ_ATOMIC; 377 378 if (iocb->ki_flags & IOCB_NOWAIT) 379 bio->bi_opf |= REQ_NOWAIT; 380 381 if (iocb->ki_flags & IOCB_HIPRI) { 382 bio->bi_opf |= REQ_POLLED; 383 submit_bio(bio); 384 WRITE_ONCE(iocb->private, bio); 385 } else { 386 submit_bio(bio); 387 } 388 return -EIOCBQUEUED; 389 390 out_bio_put: 391 bio_put(bio); 392 return ret; 393 } 394 395 static ssize_t blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter) 396 { 397 struct block_device *bdev = I_BDEV(iocb->ki_filp->f_mapping->host); 398 unsigned int nr_pages; 399 400 if (!iov_iter_count(iter)) 401 return 0; 402 403 if (blkdev_dio_invalid(bdev, iocb, iter)) 404 return -EINVAL; 405 406 /* HIPRI needs private as bio; HAS_METADATA keeps it as uio_meta */ 407 if ((iocb->ki_flags & IOCB_HIPRI) && 408 (iocb->ki_flags & IOCB_HAS_METADATA)) 409 return -EOPNOTSUPP; 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 DEFINE_IOMAP_ITER_NEXT(blkdev_iomap_next, blkdev_iomap_begin); 463 464 static const struct iomap_ops blkdev_iomap_ops = { 465 .iomap_next = blkdev_iomap_next, 466 }; 467 468 #ifdef CONFIG_BUFFER_HEAD 469 static int blkdev_get_block(struct inode *inode, sector_t iblock, 470 struct buffer_head *bh, int create) 471 { 472 bh->b_bdev = I_BDEV(inode); 473 bh->b_blocknr = iblock; 474 set_buffer_mapped(bh); 475 return 0; 476 } 477 478 /* 479 * We cannot call mpage_writepages() as it does not take the buffer lock. 480 * We must use block_write_full_folio() directly which holds the buffer 481 * lock. The buffer lock provides the synchronisation with writeback 482 * that filesystems rely on when they use the blockdev's mapping. 483 */ 484 static int blkdev_writepages(struct address_space *mapping, 485 struct writeback_control *wbc) 486 { 487 struct folio *folio = NULL; 488 struct blk_plug plug; 489 int err; 490 491 blk_start_plug(&plug); 492 while ((folio = writeback_iter(mapping, wbc, folio, &err))) 493 err = block_write_full_folio(folio, wbc, blkdev_get_block); 494 blk_finish_plug(&plug); 495 496 return err; 497 } 498 499 static int blkdev_read_folio(struct file *file, struct folio *folio) 500 { 501 return block_read_full_folio(folio, blkdev_get_block); 502 } 503 504 static void blkdev_readahead(struct readahead_control *rac) 505 { 506 mpage_readahead(rac, blkdev_get_block); 507 } 508 509 const struct address_space_operations def_blk_aops = { 510 .dirty_folio = block_dirty_folio, 511 .invalidate_folio = block_invalidate_folio, 512 .read_folio = blkdev_read_folio, 513 .readahead = blkdev_readahead, 514 .writepages = blkdev_writepages, 515 .migrate_folio = buffer_migrate_folio_norefs, 516 .is_dirty_writeback = buffer_check_dirty_writeback, 517 }; 518 #else /* CONFIG_BUFFER_HEAD */ 519 static int blkdev_read_folio(struct file *file, struct folio *folio) 520 { 521 iomap_bio_read_folio(folio, &blkdev_iomap_ops); 522 return 0; 523 } 524 525 static void blkdev_readahead(struct readahead_control *rac) 526 { 527 iomap_bio_readahead(rac, &blkdev_iomap_ops); 528 } 529 530 static ssize_t blkdev_writeback_range(struct iomap_writepage_ctx *wpc, 531 struct folio *folio, u64 offset, unsigned int len, u64 end_pos) 532 { 533 loff_t isize = i_size_read(wpc->inode); 534 535 if (WARN_ON_ONCE(offset >= isize)) 536 return -EIO; 537 538 if (offset < wpc->iomap.offset || 539 offset >= wpc->iomap.offset + wpc->iomap.length) { 540 int error; 541 542 error = blkdev_iomap_begin(wpc->inode, offset, isize - offset, 543 IOMAP_WRITE, &wpc->iomap, NULL); 544 if (error) 545 return error; 546 } 547 548 return iomap_add_to_ioend(wpc, folio, offset, end_pos, len); 549 } 550 551 static const struct iomap_writeback_ops blkdev_writeback_ops = { 552 .writeback_range = blkdev_writeback_range, 553 .writeback_submit = iomap_ioend_writeback_submit, 554 }; 555 556 static int blkdev_writepages(struct address_space *mapping, 557 struct writeback_control *wbc) 558 { 559 struct iomap_writepage_ctx wpc = { 560 .inode = mapping->host, 561 .wbc = wbc, 562 .ops = &blkdev_writeback_ops 563 }; 564 565 return iomap_writepages(&wpc); 566 } 567 568 const struct address_space_operations def_blk_aops = { 569 .dirty_folio = filemap_dirty_folio, 570 .release_folio = iomap_release_folio, 571 .invalidate_folio = iomap_invalidate_folio, 572 .read_folio = blkdev_read_folio, 573 .readahead = blkdev_readahead, 574 .writepages = blkdev_writepages, 575 .is_partially_uptodate = iomap_is_partially_uptodate, 576 .error_remove_folio = generic_error_remove_folio, 577 .migrate_folio = filemap_migrate_folio, 578 }; 579 #endif /* CONFIG_BUFFER_HEAD */ 580 581 /* 582 * for a block special file file_inode(file)->i_size is zero 583 * so we compute the size by hand (just as in block_read/write above) 584 */ 585 static loff_t blkdev_llseek(struct file *file, loff_t offset, int whence) 586 { 587 struct inode *bd_inode = bdev_file_inode(file); 588 loff_t retval; 589 590 inode_lock(bd_inode); 591 retval = fixed_size_llseek(file, offset, whence, i_size_read(bd_inode)); 592 inode_unlock(bd_inode); 593 return retval; 594 } 595 596 static int blkdev_fsync(struct file *filp, loff_t start, loff_t end, 597 int datasync) 598 { 599 struct block_device *bdev = I_BDEV(filp->f_mapping->host); 600 int error; 601 602 error = file_write_and_wait_range(filp, start, end); 603 if (error) 604 return error; 605 606 /* 607 * There is no need to serialise calls to blkdev_issue_flush with 608 * i_mutex and doing so causes performance issues with concurrent 609 * O_SYNC writers to a block device. 610 */ 611 error = blkdev_issue_flush(bdev); 612 if (error == -EOPNOTSUPP) 613 error = 0; 614 615 return error; 616 } 617 618 /** 619 * file_to_blk_mode - get block open flags from file flags 620 * @file: file whose open flags should be converted 621 * 622 * Look at file open flags and generate corresponding block open flags from 623 * them. The function works both for file just being open (e.g. during ->open 624 * callback) and for file that is already open. This is actually non-trivial 625 * (see comment in the function). 626 */ 627 blk_mode_t file_to_blk_mode(struct file *file) 628 { 629 blk_mode_t mode = 0; 630 631 if (file->f_mode & FMODE_READ) 632 mode |= BLK_OPEN_READ; 633 if (file->f_mode & FMODE_WRITE) 634 mode |= BLK_OPEN_WRITE; 635 /* 636 * do_dentry_open() clears O_EXCL from f_flags, use file->private_data 637 * to determine whether the open was exclusive for already open files. 638 */ 639 if (file->private_data) 640 mode |= BLK_OPEN_EXCL; 641 else if (file->f_flags & O_EXCL) 642 mode |= BLK_OPEN_EXCL; 643 if (file->f_flags & O_NDELAY) 644 mode |= BLK_OPEN_NDELAY; 645 646 /* 647 * If all bits in O_ACCMODE set (aka O_RDWR | O_WRONLY), the floppy 648 * driver has historically allowed ioctls as if the file was opened for 649 * writing, but does not allow and actual reads or writes. 650 */ 651 if ((file->f_flags & O_ACCMODE) == (O_RDWR | O_WRONLY)) 652 mode |= BLK_OPEN_WRITE_IOCTL; 653 654 return mode; 655 } 656 657 static int blkdev_open(struct inode *inode, struct file *filp) 658 { 659 struct block_device *bdev; 660 blk_mode_t mode; 661 int ret; 662 663 mode = file_to_blk_mode(filp); 664 /* Use the file as the holder. */ 665 if (mode & BLK_OPEN_EXCL) 666 filp->private_data = filp; 667 ret = bdev_permission(inode->i_rdev, mode, filp->private_data); 668 if (ret) 669 return ret; 670 671 bdev = blkdev_get_no_open(inode->i_rdev, true); 672 if (!bdev) 673 return -ENXIO; 674 675 if (bdev_can_atomic_write(bdev)) 676 filp->f_mode |= FMODE_CAN_ATOMIC_WRITE; 677 if (blk_get_integrity(bdev->bd_disk)) 678 filp->f_mode |= FMODE_HAS_METADATA; 679 680 ret = bdev_open(bdev, mode, filp->private_data, NULL, filp); 681 if (ret) 682 blkdev_put_no_open(bdev); 683 return ret; 684 } 685 686 static int blkdev_release(struct inode *inode, struct file *filp) 687 { 688 bdev_release(filp); 689 return 0; 690 } 691 692 static ssize_t 693 blkdev_direct_write(struct kiocb *iocb, struct iov_iter *from) 694 { 695 size_t count = iov_iter_count(from); 696 ssize_t written; 697 698 written = kiocb_invalidate_pages(iocb, count); 699 if (written) { 700 if (written == -EBUSY) 701 return 0; 702 return written; 703 } 704 705 written = blkdev_direct_IO(iocb, from); 706 if (written > 0) { 707 kiocb_invalidate_post_direct_write(iocb, count); 708 iocb->ki_pos += written; 709 count -= written; 710 } 711 if (written != -EIOCBQUEUED) 712 iov_iter_revert(from, count - iov_iter_count(from)); 713 return written; 714 } 715 716 static ssize_t blkdev_buffered_write(struct kiocb *iocb, struct iov_iter *from) 717 { 718 return iomap_file_buffered_write(iocb, from, &blkdev_iomap_ops, NULL, 719 NULL); 720 } 721 722 /* 723 * Write data to the block device. Only intended for the block device itself 724 * and the raw driver which basically is a fake block device. 725 * 726 * Does not take i_mutex for the write and thus is not for general purpose 727 * use. 728 */ 729 static ssize_t blkdev_write_iter(struct kiocb *iocb, struct iov_iter *from) 730 { 731 struct file *file = iocb->ki_filp; 732 struct inode *bd_inode = bdev_file_inode(file); 733 struct block_device *bdev = I_BDEV(bd_inode); 734 bool atomic = iocb->ki_flags & IOCB_ATOMIC; 735 loff_t size = bdev_nr_bytes(bdev); 736 size_t shorted = 0; 737 ssize_t ret; 738 739 if (bdev_read_only(bdev)) 740 return -EPERM; 741 742 if (IS_SWAPFILE(bd_inode) && !is_hibernate_resume_dev(bd_inode->i_rdev)) 743 return -ETXTBSY; 744 745 if (!iov_iter_count(from)) 746 return 0; 747 748 if (iocb->ki_pos >= size) 749 return -ENOSPC; 750 751 if ((iocb->ki_flags & (IOCB_NOWAIT | IOCB_DIRECT)) == IOCB_NOWAIT) 752 return -EOPNOTSUPP; 753 754 if (atomic) { 755 ret = generic_atomic_write_valid(iocb, from); 756 if (ret) 757 return ret; 758 } 759 760 size -= iocb->ki_pos; 761 if (iov_iter_count(from) > size) { 762 if (atomic) 763 return -EINVAL; 764 shorted = iov_iter_count(from) - size; 765 iov_iter_truncate(from, size); 766 } 767 768 ret = file_update_time(file); 769 if (ret) 770 return ret; 771 772 if (iocb->ki_flags & IOCB_DIRECT) { 773 ret = blkdev_direct_write(iocb, from); 774 if (ret >= 0 && iov_iter_count(from)) 775 ret = direct_write_fallback(iocb, from, ret, 776 blkdev_buffered_write(iocb, from)); 777 } else { 778 /* 779 * Take i_rwsem and invalidate_lock to avoid racing with 780 * set_blocksize changing i_blkbits/folio order and punching 781 * out the pagecache. 782 */ 783 inode_lock_shared(bd_inode); 784 ret = blkdev_buffered_write(iocb, from); 785 inode_unlock_shared(bd_inode); 786 } 787 788 if (ret > 0) 789 ret = generic_write_sync(iocb, ret); 790 iov_iter_reexpand(from, iov_iter_count(from) + shorted); 791 return ret; 792 } 793 794 static ssize_t blkdev_read_iter(struct kiocb *iocb, struct iov_iter *to) 795 { 796 struct inode *bd_inode = bdev_file_inode(iocb->ki_filp); 797 struct block_device *bdev = I_BDEV(iocb->ki_filp->f_mapping->host); 798 loff_t size = bdev_nr_bytes(bdev); 799 loff_t pos = iocb->ki_pos; 800 size_t shorted = 0; 801 ssize_t ret = 0; 802 size_t count; 803 804 if (unlikely(pos + iov_iter_count(to) > size)) { 805 if (pos >= size) 806 return 0; 807 size -= pos; 808 shorted = iov_iter_count(to) - size; 809 iov_iter_truncate(to, size); 810 } 811 812 count = iov_iter_count(to); 813 if (!count) 814 goto reexpand; /* skip atime */ 815 816 if (iocb->ki_flags & IOCB_DIRECT) { 817 ret = kiocb_write_and_wait(iocb, count); 818 if (ret < 0) 819 goto reexpand; 820 file_accessed(iocb->ki_filp); 821 822 ret = blkdev_direct_IO(iocb, to); 823 if (ret > 0) { 824 iocb->ki_pos += ret; 825 count -= ret; 826 } 827 if (ret != -EIOCBQUEUED) 828 iov_iter_revert(to, count - iov_iter_count(to)); 829 if (ret < 0 || !count) 830 goto reexpand; 831 } 832 833 /* 834 * Take i_rwsem and invalidate_lock to avoid racing with set_blocksize 835 * changing i_blkbits/folio order and punching out the pagecache. 836 */ 837 inode_lock_shared(bd_inode); 838 ret = filemap_read(iocb, to, ret); 839 inode_unlock_shared(bd_inode); 840 841 reexpand: 842 if (unlikely(shorted)) 843 iov_iter_reexpand(to, iov_iter_count(to) + shorted); 844 return ret; 845 } 846 847 #define BLKDEV_FALLOC_FL_SUPPORTED \ 848 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \ 849 FALLOC_FL_ZERO_RANGE | FALLOC_FL_WRITE_ZEROES) 850 851 static long blkdev_fallocate(struct file *file, int mode, loff_t start, 852 loff_t len) 853 { 854 struct inode *inode = bdev_file_inode(file); 855 struct block_device *bdev = I_BDEV(inode); 856 loff_t end = start + len - 1; 857 loff_t isize; 858 unsigned int flags; 859 int error; 860 861 /* Fail if we don't recognize the flags. */ 862 if (mode & ~BLKDEV_FALLOC_FL_SUPPORTED) 863 return -EOPNOTSUPP; 864 /* 865 * Don't allow writing zeroes if the device does not enable the 866 * unmap write zeroes operation. 867 */ 868 if ((mode & FALLOC_FL_WRITE_ZEROES) && 869 !bdev_write_zeroes_unmap_sectors(bdev)) 870 return -EOPNOTSUPP; 871 872 /* Don't go off the end of the device. */ 873 isize = bdev_nr_bytes(bdev); 874 if (start >= isize) 875 return -EINVAL; 876 if (end >= isize) { 877 if (mode & FALLOC_FL_KEEP_SIZE) { 878 len = isize - start; 879 end = start + len - 1; 880 } else 881 return -EINVAL; 882 } 883 884 /* 885 * Don't allow IO that isn't aligned to logical block size. 886 */ 887 if ((start | len) & (bdev_logical_block_size(bdev) - 1)) 888 return -EINVAL; 889 890 inode_lock(inode); 891 filemap_invalidate_lock(inode->i_mapping); 892 893 switch (mode) { 894 case FALLOC_FL_ZERO_RANGE: 895 case FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE: 896 flags = BLKDEV_ZERO_NOUNMAP; 897 break; 898 case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE: 899 flags = BLKDEV_ZERO_NOFALLBACK; 900 break; 901 case FALLOC_FL_WRITE_ZEROES: 902 flags = 0; 903 break; 904 default: 905 error = -EOPNOTSUPP; 906 goto fail; 907 } 908 909 /* 910 * Invalidate the page cache, including dirty pages, for valid 911 * de-allocate mode calls to fallocate(). 912 */ 913 error = truncate_bdev_range(bdev, file_to_blk_mode(file), start, end); 914 if (error) 915 goto fail; 916 917 error = blkdev_issue_zeroout(bdev, start >> SECTOR_SHIFT, 918 len >> SECTOR_SHIFT, GFP_KERNEL, flags); 919 fail: 920 filemap_invalidate_unlock(inode->i_mapping); 921 inode_unlock(inode); 922 return error; 923 } 924 925 static int blkdev_mmap_prepare(struct vm_area_desc *desc) 926 { 927 struct file *file = desc->file; 928 929 if (bdev_read_only(I_BDEV(bdev_file_inode(file)))) 930 return generic_file_readonly_mmap_prepare(desc); 931 932 return generic_file_mmap_prepare(desc); 933 } 934 935 const struct file_operations def_blk_fops = { 936 .open = blkdev_open, 937 .release = blkdev_release, 938 .llseek = blkdev_llseek, 939 .read_iter = blkdev_read_iter, 940 .write_iter = blkdev_write_iter, 941 .iopoll = iocb_bio_iopoll, 942 .mmap_prepare = blkdev_mmap_prepare, 943 .fsync = blkdev_fsync, 944 .unlocked_ioctl = blkdev_ioctl, 945 #ifdef CONFIG_COMPAT 946 .compat_ioctl = compat_blkdev_ioctl, 947 #endif 948 .splice_read = filemap_splice_read, 949 .splice_write = iter_file_splice_write, 950 .fallocate = blkdev_fallocate, 951 .uring_cmd = blkdev_uring_cmd, 952 .fop_flags = FOP_BUFFER_RASYNC | FOP_DONTCACHE, 953 }; 954 955 static __init int blkdev_init(void) 956 { 957 return bioset_init(&blkdev_dio_pool, 4, 958 offsetof(struct blkdev_dio, bio), 959 BIOSET_NEED_BVECS|BIOSET_PERCPU_CACHE); 960 } 961 module_init(blkdev_init); 962