block/fops.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 1991, 1992  Linus Torvalds
   4  * Copyright (C) 2001  Andrea Arcangeli <[email protected]> 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/iomap.h>
  19 #include <linux/module.h>
  20 #include "blk.h"
  21
  22 static inline struct inode *bdev_file_inode(struct file *file)
  23 {
  24         return file->f_mapping->host;
  25 }
  26
  27 static blk_opf_t dio_bio_write_op(struct kiocb *iocb)
  28 {
  29         blk_opf_t opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
  30
  31         /* avoid the need for a I/O completion work item */
  32         if (iocb_is_dsync(iocb))
  33                 opf |= REQ_FUA;
  34         return opf;
  35 }
  36
  37 static bool blkdev_dio_unaligned(struct block_device *bdev, loff_t pos,
  38                               struct iov_iter *iter)
  39 {
  40         return pos & (bdev_logical_block_size(bdev) - 1) ||
  41                 !bdev_iter_is_aligned(bdev, iter);
  42 }
  43
  44 #define DIO_INLINE_BIO_VECS 4
  45
  46 static ssize_t __blkdev_direct_IO_simple(struct kiocb *iocb,
  47                 struct iov_iter *iter, struct block_device *bdev,
  48                 unsigned int nr_pages)
  49 {
  50         struct bio_vec inline_vecs[DIO_INLINE_BIO_VECS], *vecs;
  51         loff_t pos = iocb->ki_pos;
  52         bool should_dirty = false;
  53         struct bio bio;
  54         ssize_t ret;
  55
  56         if (nr_pages <= DIO_INLINE_BIO_VECS)
  57                 vecs = inline_vecs;
  58         else {
  59                 vecs = kmalloc_array(nr_pages, sizeof(struct bio_vec),
  60                                      GFP_KERNEL);
  61                 if (!vecs)
  62                         return -ENOMEM;
  63         }
  64
  65         if (iov_iter_rw(iter) == READ) {
  66                 bio_init(&bio, bdev, vecs, nr_pages, REQ_OP_READ);
  67                 if (user_backed_iter(iter))
  68                         should_dirty = true;
  69         } else {
  70                 bio_init(&bio, bdev, vecs, nr_pages, dio_bio_write_op(iocb));
  71         }
  72         bio.bi_iter.bi_sector = pos >> SECTOR_SHIFT;
  73         bio.bi_write_hint = file_inode(iocb->ki_filp)->i_write_hint;
  74         bio.bi_ioprio = iocb->ki_ioprio;
  75
  76         ret = bio_iov_iter_get_pages(&bio, iter);
  77         if (unlikely(ret))
  78                 goto out;
  79         ret = bio.bi_iter.bi_size;
  80
  81         if (iov_iter_rw(iter) == WRITE)
  82                 task_io_account_write(ret);
  83
  84         if (iocb->ki_flags & IOCB_NOWAIT)
  85                 bio.bi_opf |= REQ_NOWAIT;
  86
  87         submit_bio_wait(&bio);
  88
  89         bio_release_pages(&bio, should_dirty);
  90         if (unlikely(bio.bi_status))
  91                 ret = blk_status_to_errno(bio.bi_status);
  92
  93 out:
  94         if (vecs != inline_vecs)
  95                 kfree(vecs);
  96
  97         bio_uninit(&bio);
  98
  99         return ret;
 100 }
 101
 102 enum {
 103         DIO_SHOULD_DIRTY        = 1,
 104         DIO_IS_SYNC             = 2,
 105 };
 106
 107 struct blkdev_dio {
 108         union {
 109                 struct kiocb            *iocb;
 110                 struct task_struct      *waiter;
 111         };
 112         size_t                  size;
 113         atomic_t                ref;
 114         unsigned int            flags;
 115         struct bio              bio ____cacheline_aligned_in_smp;
 116 };
 117
 118 static struct bio_set blkdev_dio_pool;
 119
 120 static void blkdev_bio_end_io(struct bio *bio)
 121 {
 122         struct blkdev_dio *dio = bio->bi_private;
 123         bool should_dirty = dio->flags & DIO_SHOULD_DIRTY;
 124
 125         if (bio->bi_status && !dio->bio.bi_status)
 126                 dio->bio.bi_status = bio->bi_status;
 127
 128         if (atomic_dec_and_test(&dio->ref)) {
 129                 if (!(dio->flags & DIO_IS_SYNC)) {
 130                         struct kiocb *iocb = dio->iocb;
 131                         ssize_t ret;
 132
 133                         WRITE_ONCE(iocb->private, NULL);
 134
 135                         if (likely(!dio->bio.bi_status)) {
 136                                 ret = dio->size;
 137                                 iocb->ki_pos += ret;
 138                         } else {
 139                                 ret = blk_status_to_errno(dio->bio.bi_status);
 140                         }
 141
 142                         dio->iocb->ki_complete(iocb, ret);
 143                         bio_put(&dio->bio);
 144                 } else {
 145                         struct task_struct *waiter = dio->waiter;
 146
 147                         WRITE_ONCE(dio->waiter, NULL);
 148                         blk_wake_io_task(waiter);
 149                 }
 150         }
 151
 152         if (should_dirty) {
 153                 bio_check_pages_dirty(bio);
 154         } else {
 155                 bio_release_pages(bio, false);
 156                 bio_put(bio);
 157         }
 158 }
 159
 160 static ssize_t __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
 161                 struct block_device *bdev, unsigned int nr_pages)
 162 {
 163         struct blk_plug plug;
 164         struct blkdev_dio *dio;
 165         struct bio *bio;
 166         bool is_read = (iov_iter_rw(iter) == READ), is_sync;
 167         blk_opf_t opf = is_read ? REQ_OP_READ : dio_bio_write_op(iocb);
 168         loff_t pos = iocb->ki_pos;
 169         int ret = 0;
 170
 171         if (iocb->ki_flags & IOCB_ALLOC_CACHE)
 172                 opf |= REQ_ALLOC_CACHE;
 173         bio = bio_alloc_bioset(bdev, nr_pages, opf, GFP_KERNEL,
 174                                &blkdev_dio_pool);
 175         dio = container_of(bio, struct blkdev_dio, bio);
 176         atomic_set(&dio->ref, 1);
 177         /*
 178          * Grab an extra reference to ensure the dio structure which is embedded
 179          * into the first bio stays around.
 180          */
 181         bio_get(bio);
 182
 183         is_sync = is_sync_kiocb(iocb);
 184         if (is_sync) {
 185                 dio->flags = DIO_IS_SYNC;
 186                 dio->waiter = current;
 187         } else {
 188                 dio->flags = 0;
 189                 dio->iocb = iocb;
 190         }
 191
 192         dio->size = 0;
 193         if (is_read && user_backed_iter(iter))
 194                 dio->flags |= DIO_SHOULD_DIRTY;
 195
 196         blk_start_plug(&plug);
 197
 198         for (;;) {
 199                 bio->bi_iter.bi_sector = pos >> SECTOR_SHIFT;
 200                 bio->bi_write_hint = file_inode(iocb->ki_filp)->i_write_hint;
 201                 bio->bi_private = dio;
 202                 bio->bi_end_io = blkdev_bio_end_io;
 203                 bio->bi_ioprio = iocb->ki_ioprio;
 204
 205                 ret = bio_iov_iter_get_pages(bio, iter);
 206                 if (unlikely(ret)) {
 207                         bio->bi_status = BLK_STS_IOERR;
 208                         bio_endio(bio);
 209                         break;
 210                 }
 211                 if (iocb->ki_flags & IOCB_NOWAIT) {
 212                         /*
 213                          * This is nonblocking IO, and we need to allocate
 214                          * another bio if we have data left to map. As we
 215                          * cannot guarantee that one of the sub bios will not
 216                          * fail getting issued FOR NOWAIT and as error results
 217                          * are coalesced across all of them, be safe and ask for
 218                          * a retry of this from blocking context.
 219                          */
 220                         if (unlikely(iov_iter_count(iter))) {
 221                                 bio_release_pages(bio, false);
 222                                 bio_clear_flag(bio, BIO_REFFED);
 223                                 bio_put(bio);
 224                                 blk_finish_plug(&plug);
 225                                 return -EAGAIN;
 226                         }
 227                         bio->bi_opf |= REQ_NOWAIT;
 228                 }
 229
 230                 if (is_read) {
 231                         if (dio->flags & DIO_SHOULD_DIRTY)
 232                                 bio_set_pages_dirty(bio);
 233                 } else {
 234                         task_io_account_write(bio->bi_iter.bi_size);
 235                 }
 236                 dio->size += bio->bi_iter.bi_size;
 237                 pos += bio->bi_iter.bi_size;
 238
 239                 nr_pages = bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS);
 240                 if (!nr_pages) {
 241                         submit_bio(bio);
 242                         break;
 243                 }
 244                 atomic_inc(&dio->ref);
 245                 submit_bio(bio);
 246                 bio = bio_alloc(bdev, nr_pages, opf, GFP_KERNEL);
 247         }
 248
 249         blk_finish_plug(&plug);
 250
 251         if (!is_sync)
 252                 return -EIOCBQUEUED;
 253
 254         for (;;) {
 255                 set_current_state(TASK_UNINTERRUPTIBLE);
 256                 if (!READ_ONCE(dio->waiter))
 257                         break;
 258                 blk_io_schedule();
 259         }
 260         __set_current_state(TASK_RUNNING);
 261
 262         if (!ret)
 263                 ret = blk_status_to_errno(dio->bio.bi_status);
 264         if (likely(!ret))
 265                 ret = dio->size;
 266
 267         bio_put(&dio->bio);
 268         return ret;
 269 }
 270
 271 static void blkdev_bio_end_io_async(struct bio *bio)
 272 {
 273         struct blkdev_dio *dio = container_of(bio, struct blkdev_dio, bio);
 274         struct kiocb *iocb = dio->iocb;
 275         ssize_t ret;
 276
 277         WRITE_ONCE(iocb->private, NULL);
 278
 279         if (likely(!bio->bi_status)) {
 280                 ret = dio->size;
 281                 iocb->ki_pos += ret;
 282         } else {
 283                 ret = blk_status_to_errno(bio->bi_status);
 284         }
 285
 286         iocb->ki_complete(iocb, ret);
 287
 288         if (dio->flags & DIO_SHOULD_DIRTY) {
 289                 bio_check_pages_dirty(bio);
 290         } else {
 291                 bio_release_pages(bio, false);
 292                 bio_put(bio);
 293         }
 294 }
 295
 296 static ssize_t __blkdev_direct_IO_async(struct kiocb *iocb,
 297                                         struct iov_iter *iter,
 298                                         struct block_device *bdev,
 299                                         unsigned int nr_pages)
 300 {
 301         bool is_read = iov_iter_rw(iter) == READ;
 302         blk_opf_t opf = is_read ? REQ_OP_READ : dio_bio_write_op(iocb);
 303         struct blkdev_dio *dio;
 304         struct bio *bio;
 305         loff_t pos = iocb->ki_pos;
 306         int ret = 0;
 307
 308         if (iocb->ki_flags & IOCB_ALLOC_CACHE)
 309                 opf |= REQ_ALLOC_CACHE;
 310         bio = bio_alloc_bioset(bdev, nr_pages, opf, GFP_KERNEL,
 311                                &blkdev_dio_pool);
 312         dio = container_of(bio, struct blkdev_dio, bio);
 313         dio->flags = 0;
 314         dio->iocb = iocb;
 315         bio->bi_iter.bi_sector = pos >> SECTOR_SHIFT;
 316         bio->bi_write_hint = file_inode(iocb->ki_filp)->i_write_hint;
 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_NOWAIT)
 347                 bio->bi_opf |= REQ_NOWAIT;
 348
 349         if (iocb->ki_flags & IOCB_HIPRI) {
 350                 bio->bi_opf |= REQ_POLLED;
 351                 submit_bio(bio);
 352                 WRITE_ONCE(iocb->private, bio);
 353         } else {
 354                 submit_bio(bio);
 355         }
 356         return -EIOCBQUEUED;
 357 }
 358
 359 static ssize_t blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 360 {
 361         struct block_device *bdev = I_BDEV(iocb->ki_filp->f_mapping->host);
 362         unsigned int nr_pages;
 363
 364         if (!iov_iter_count(iter))
 365                 return 0;
 366
 367         if (blkdev_dio_unaligned(bdev, iocb->ki_pos, iter))
 368                 return -EINVAL;
 369
 370         nr_pages = bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS + 1);
 371         if (likely(nr_pages <= BIO_MAX_VECS)) {
 372                 if (is_sync_kiocb(iocb))
 373                         return __blkdev_direct_IO_simple(iocb, iter, bdev,
 374                                                         nr_pages);
 375                 return __blkdev_direct_IO_async(iocb, iter, bdev, nr_pages);
 376         }
 377         return __blkdev_direct_IO(iocb, iter, bdev, bio_max_segs(nr_pages));
 378 }
 379
 380 static int blkdev_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
 381                 unsigned int flags, struct iomap *iomap, struct iomap *srcmap)
 382 {
 383         struct block_device *bdev = I_BDEV(inode);
 384         loff_t isize = i_size_read(inode);
 385
 386         iomap->bdev = bdev;
 387         iomap->offset = ALIGN_DOWN(offset, bdev_logical_block_size(bdev));
 388         if (offset >= isize)
 389                 return -EIO;
 390         iomap->type = IOMAP_MAPPED;
 391         iomap->addr = iomap->offset;
 392         iomap->length = isize - iomap->offset;
 393         iomap->flags |= IOMAP_F_BUFFER_HEAD; /* noop for !CONFIG_BUFFER_HEAD */
 394         return 0;
 395 }
 396
 397 static const struct iomap_ops blkdev_iomap_ops = {
 398         .iomap_begin            = blkdev_iomap_begin,
 399 };
 400
 401 #ifdef CONFIG_BUFFER_HEAD
 402 static int blkdev_get_block(struct inode *inode, sector_t iblock,
 403                 struct buffer_head *bh, int create)
 404 {
 405         bh->b_bdev = I_BDEV(inode);
 406         bh->b_blocknr = iblock;
 407         set_buffer_mapped(bh);
 408         return 0;
 409 }
 410
 411 /*
 412  * We cannot call mpage_writepages() as it does not take the buffer lock.
 413  * We must use block_write_full_folio() directly which holds the buffer
 414  * lock.  The buffer lock provides the synchronisation with writeback
 415  * that filesystems rely on when they use the blockdev's mapping.
 416  */
 417 static int blkdev_writepages(struct address_space *mapping,
 418                 struct writeback_control *wbc)
 419 {
 420         struct blk_plug plug;
 421         int err;
 422
 423         blk_start_plug(&plug);
 424         err = write_cache_pages(mapping, wbc, block_write_full_folio,
 425                         blkdev_get_block);
 426         blk_finish_plug(&plug);
 427
 428         return err;
 429 }
 430
 431 static int blkdev_read_folio(struct file *file, struct folio *folio)
 432 {
 433         return block_read_full_folio(folio, blkdev_get_block);
 434 }
 435
 436 static void blkdev_readahead(struct readahead_control *rac)
 437 {
 438         mpage_readahead(rac, blkdev_get_block);
 439 }
 440
 441 static int blkdev_write_begin(struct file *file, struct address_space *mapping,
 442                 loff_t pos, unsigned len, struct page **pagep, void **fsdata)
 443 {
 444         return block_write_begin(mapping, pos, len, pagep, blkdev_get_block);
 445 }
 446
 447 static int blkdev_write_end(struct file *file, struct address_space *mapping,
 448                 loff_t pos, unsigned len, unsigned copied, struct page *page,
 449                 void *fsdata)
 450 {
 451         int ret;
 452         ret = block_write_end(file, mapping, pos, len, copied, page, fsdata);
 453
 454         unlock_page(page);
 455         put_page(page);
 456
 457         return ret;
 458 }
 459
 460 const struct address_space_operations def_blk_aops = {
 461         .dirty_folio    = block_dirty_folio,
 462         .invalidate_folio = block_invalidate_folio,
 463         .read_folio     = blkdev_read_folio,
 464         .readahead      = blkdev_readahead,
 465         .writepages     = blkdev_writepages,
 466         .write_begin    = blkdev_write_begin,
 467         .write_end      = blkdev_write_end,
 468         .migrate_folio  = buffer_migrate_folio_norefs,
 469         .is_dirty_writeback = buffer_check_dirty_writeback,
 470 };
 471 #else /* CONFIG_BUFFER_HEAD */
 472 static int blkdev_read_folio(struct file *file, struct folio *folio)
 473 {
 474         return iomap_read_folio(folio, &blkdev_iomap_ops);
 475 }
 476
 477 static void blkdev_readahead(struct readahead_control *rac)
 478 {
 479         iomap_readahead(rac, &blkdev_iomap_ops);
 480 }
 481
 482 static int blkdev_map_blocks(struct iomap_writepage_ctx *wpc,
 483                 struct inode *inode, loff_t offset, unsigned int len)
 484 {
 485         loff_t isize = i_size_read(inode);
 486
 487         if (WARN_ON_ONCE(offset >= isize))
 488                 return -EIO;
 489         if (offset >= wpc->iomap.offset &&
 490             offset < wpc->iomap.offset + wpc->iomap.length)
 491                 return 0;
 492         return blkdev_iomap_begin(inode, offset, isize - offset,
 493                                   IOMAP_WRITE, &wpc->iomap, NULL);
 494 }
 495
 496 static const struct iomap_writeback_ops blkdev_writeback_ops = {
 497         .map_blocks             = blkdev_map_blocks,
 498 };
 499
 500 static int blkdev_writepages(struct address_space *mapping,
 501                 struct writeback_control *wbc)
 502 {
 503         struct iomap_writepage_ctx wpc = { };
 504
 505         return iomap_writepages(mapping, wbc, &wpc, &blkdev_writeback_ops);
 506 }
 507
 508 const struct address_space_operations def_blk_aops = {
 509         .dirty_folio    = filemap_dirty_folio,
 510         .release_folio          = iomap_release_folio,
 511         .invalidate_folio       = iomap_invalidate_folio,
 512         .read_folio             = blkdev_read_folio,
 513         .readahead              = blkdev_readahead,
 514         .writepages             = blkdev_writepages,
 515         .is_partially_uptodate  = iomap_is_partially_uptodate,
 516         .error_remove_folio     = generic_error_remove_folio,
 517         .migrate_folio          = filemap_migrate_folio,
 518 };
 519 #endif /* CONFIG_BUFFER_HEAD */
 520
 521 /*
 522  * for a block special file file_inode(file)->i_size is zero
 523  * so we compute the size by hand (just as in block_read/write above)
 524  */
 525 static loff_t blkdev_llseek(struct file *file, loff_t offset, int whence)
 526 {
 527         struct inode *bd_inode = bdev_file_inode(file);
 528         loff_t retval;
 529
 530         inode_lock(bd_inode);
 531         retval = fixed_size_llseek(file, offset, whence, i_size_read(bd_inode));
 532         inode_unlock(bd_inode);
 533         return retval;
 534 }
 535
 536 static int blkdev_fsync(struct file *filp, loff_t start, loff_t end,
 537                 int datasync)
 538 {
 539         struct block_device *bdev = I_BDEV(filp->f_mapping->host);
 540         int error;
 541
 542         error = file_write_and_wait_range(filp, start, end);
 543         if (error)
 544                 return error;
 545
 546         /*
 547          * There is no need to serialise calls to blkdev_issue_flush with
 548          * i_mutex and doing so causes performance issues with concurrent
 549          * O_SYNC writers to a block device.
 550          */
 551         error = blkdev_issue_flush(bdev);
 552         if (error == -EOPNOTSUPP)
 553                 error = 0;
 554
 555         return error;
 556 }
 557
 558 /**
 559  * file_to_blk_mode - get block open flags from file flags
 560  * @file: file whose open flags should be converted
 561  *
 562  * Look at file open flags and generate corresponding block open flags from
 563  * them. The function works both for file just being open (e.g. during ->open
 564  * callback) and for file that is already open. This is actually non-trivial
 565  * (see comment in the function).
 566  */
 567 blk_mode_t file_to_blk_mode(struct file *file)
 568 {
 569         blk_mode_t mode = 0;
 570
 571         if (file->f_mode & FMODE_READ)
 572                 mode |= BLK_OPEN_READ;
 573         if (file->f_mode & FMODE_WRITE)
 574                 mode |= BLK_OPEN_WRITE;
 575         /*
 576          * do_dentry_open() clears O_EXCL from f_flags, use file->private_data
 577          * to determine whether the open was exclusive for already open files.
 578          */
 579         if (file->private_data)
 580                 mode |= BLK_OPEN_EXCL;
 581         else if (file->f_flags & O_EXCL)
 582                 mode |= BLK_OPEN_EXCL;
 583         if (file->f_flags & O_NDELAY)
 584                 mode |= BLK_OPEN_NDELAY;
 585
 586         /*
 587          * If all bits in O_ACCMODE set (aka O_RDWR | O_WRONLY), the floppy
 588          * driver has historically allowed ioctls as if the file was opened for
 589          * writing, but does not allow and actual reads or writes.
 590          */
 591         if ((file->f_flags & O_ACCMODE) == (O_RDWR | O_WRONLY))
 592                 mode |= BLK_OPEN_WRITE_IOCTL;
 593
 594         return mode;
 595 }
 596
 597 static int blkdev_open(struct inode *inode, struct file *filp)
 598 {
 599         struct block_device *bdev;
 600         blk_mode_t mode;
 601         int ret;
 602
 603         mode = file_to_blk_mode(filp);
 604         /* Use the file as the holder. */
 605         if (mode & BLK_OPEN_EXCL)
 606                 filp->private_data = filp;
 607         ret = bdev_permission(inode->i_rdev, mode, filp->private_data);
 608         if (ret)
 609                 return ret;
 610
 611         bdev = blkdev_get_no_open(inode->i_rdev);
 612         if (!bdev)
 613                 return -ENXIO;
 614
 615         ret = bdev_open(bdev, mode, filp->private_data, NULL, filp);
 616         if (ret)
 617                 blkdev_put_no_open(bdev);
 618         return ret;
 619 }
 620
 621 static int blkdev_release(struct inode *inode, struct file *filp)
 622 {
 623         bdev_release(filp);
 624         return 0;
 625 }
 626
 627 static ssize_t
 628 blkdev_direct_write(struct kiocb *iocb, struct iov_iter *from)
 629 {
 630         size_t count = iov_iter_count(from);
 631         ssize_t written;
 632
 633         written = kiocb_invalidate_pages(iocb, count);
 634         if (written) {
 635                 if (written == -EBUSY)
 636                         return 0;
 637                 return written;
 638         }
 639
 640         written = blkdev_direct_IO(iocb, from);
 641         if (written > 0) {
 642                 kiocb_invalidate_post_direct_write(iocb, count);
 643                 iocb->ki_pos += written;
 644                 count -= written;
 645         }
 646         if (written != -EIOCBQUEUED)
 647                 iov_iter_revert(from, count - iov_iter_count(from));
 648         return written;
 649 }
 650
 651 static ssize_t blkdev_buffered_write(struct kiocb *iocb, struct iov_iter *from)
 652 {
 653         return iomap_file_buffered_write(iocb, from, &blkdev_iomap_ops);
 654 }
 655
 656 /*
 657  * Write data to the block device.  Only intended for the block device itself
 658  * and the raw driver which basically is a fake block device.
 659  *
 660  * Does not take i_mutex for the write and thus is not for general purpose
 661  * use.
 662  */
 663 static ssize_t blkdev_write_iter(struct kiocb *iocb, struct iov_iter *from)
 664 {
 665         struct file *file = iocb->ki_filp;
 666         struct inode *bd_inode = bdev_file_inode(file);
 667         struct block_device *bdev = I_BDEV(bd_inode);
 668         loff_t size = bdev_nr_bytes(bdev);
 669         size_t shorted = 0;
 670         ssize_t ret;
 671
 672         if (bdev_read_only(bdev))
 673                 return -EPERM;
 674
 675         if (IS_SWAPFILE(bd_inode) && !is_hibernate_resume_dev(bd_inode->i_rdev))
 676                 return -ETXTBSY;
 677
 678         if (!iov_iter_count(from))
 679                 return 0;
 680
 681         if (iocb->ki_pos >= size)
 682                 return -ENOSPC;
 683
 684         if ((iocb->ki_flags & (IOCB_NOWAIT | IOCB_DIRECT)) == IOCB_NOWAIT)
 685                 return -EOPNOTSUPP;
 686
 687         size -= iocb->ki_pos;
 688         if (iov_iter_count(from) > size) {
 689                 shorted = iov_iter_count(from) - size;
 690                 iov_iter_truncate(from, size);
 691         }
 692
 693         ret = file_update_time(file);
 694         if (ret)
 695                 return ret;
 696
 697         if (iocb->ki_flags & IOCB_DIRECT) {
 698                 ret = blkdev_direct_write(iocb, from);
 699                 if (ret >= 0 && iov_iter_count(from))
 700                         ret = direct_write_fallback(iocb, from, ret,
 701                                         blkdev_buffered_write(iocb, from));
 702         } else {
 703                 ret = blkdev_buffered_write(iocb, from);
 704         }
 705
 706         if (ret > 0)
 707                 ret = generic_write_sync(iocb, ret);
 708         iov_iter_reexpand(from, iov_iter_count(from) + shorted);
 709         return ret;
 710 }
 711
 712 static ssize_t blkdev_read_iter(struct kiocb *iocb, struct iov_iter *to)
 713 {
 714         struct block_device *bdev = I_BDEV(iocb->ki_filp->f_mapping->host);
 715         loff_t size = bdev_nr_bytes(bdev);
 716         loff_t pos = iocb->ki_pos;
 717         size_t shorted = 0;
 718         ssize_t ret = 0;
 719         size_t count;
 720
 721         if (unlikely(pos + iov_iter_count(to) > size)) {
 722                 if (pos >= size)
 723                         return 0;
 724                 size -= pos;
 725                 shorted = iov_iter_count(to) - size;
 726                 iov_iter_truncate(to, size);
 727         }
 728
 729         count = iov_iter_count(to);
 730         if (!count)
 731                 goto reexpand; /* skip atime */
 732
 733         if (iocb->ki_flags & IOCB_DIRECT) {
 734                 ret = kiocb_write_and_wait(iocb, count);
 735                 if (ret < 0)
 736                         goto reexpand;
 737                 file_accessed(iocb->ki_filp);
 738
 739                 ret = blkdev_direct_IO(iocb, to);
 740                 if (ret >= 0) {
 741                         iocb->ki_pos += ret;
 742                         count -= ret;
 743                 }
 744                 iov_iter_revert(to, count - iov_iter_count(to));
 745                 if (ret < 0 || !count)
 746                         goto reexpand;
 747         }
 748
 749         ret = filemap_read(iocb, to, ret);
 750
 751 reexpand:
 752         if (unlikely(shorted))
 753                 iov_iter_reexpand(to, iov_iter_count(to) + shorted);
 754         return ret;
 755 }
 756
 757 #define BLKDEV_FALLOC_FL_SUPPORTED                                      \
 758                 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |           \
 759                  FALLOC_FL_ZERO_RANGE | FALLOC_FL_NO_HIDE_STALE)
 760
 761 static long blkdev_fallocate(struct file *file, int mode, loff_t start,
 762                              loff_t len)
 763 {
 764         struct inode *inode = bdev_file_inode(file);
 765         struct block_device *bdev = I_BDEV(inode);
 766         loff_t end = start + len - 1;
 767         loff_t isize;
 768         int error;
 769
 770         /* Fail if we don't recognize the flags. */
 771         if (mode & ~BLKDEV_FALLOC_FL_SUPPORTED)
 772                 return -EOPNOTSUPP;
 773
 774         /* Don't go off the end of the device. */
 775         isize = bdev_nr_bytes(bdev);
 776         if (start >= isize)
 777                 return -EINVAL;
 778         if (end >= isize) {
 779                 if (mode & FALLOC_FL_KEEP_SIZE) {
 780                         len = isize - start;
 781                         end = start + len - 1;
 782                 } else
 783                         return -EINVAL;
 784         }
 785
 786         /*
 787          * Don't allow IO that isn't aligned to logical block size.
 788          */
 789         if ((start | len) & (bdev_logical_block_size(bdev) - 1))
 790                 return -EINVAL;
 791
 792         filemap_invalidate_lock(inode->i_mapping);
 793
 794         /*
 795          * Invalidate the page cache, including dirty pages, for valid
 796          * de-allocate mode calls to fallocate().
 797          */
 798         switch (mode) {
 799         case FALLOC_FL_ZERO_RANGE:
 800         case FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE:
 801                 error = truncate_bdev_range(bdev, file_to_blk_mode(file), start, end);
 802                 if (error)
 803                         goto fail;
 804
 805                 error = blkdev_issue_zeroout(bdev, start >> SECTOR_SHIFT,
 806                                              len >> SECTOR_SHIFT, GFP_KERNEL,
 807                                              BLKDEV_ZERO_NOUNMAP);
 808                 break;
 809         case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE:
 810                 error = truncate_bdev_range(bdev, file_to_blk_mode(file), start, end);
 811                 if (error)
 812                         goto fail;
 813
 814                 error = blkdev_issue_zeroout(bdev, start >> SECTOR_SHIFT,
 815                                              len >> SECTOR_SHIFT, GFP_KERNEL,
 816                                              BLKDEV_ZERO_NOFALLBACK);
 817                 break;
 818         case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE | FALLOC_FL_NO_HIDE_STALE:
 819                 error = truncate_bdev_range(bdev, file_to_blk_mode(file), start, end);
 820                 if (error)
 821                         goto fail;
 822
 823                 error = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
 824                                              len >> SECTOR_SHIFT, GFP_KERNEL);
 825                 break;
 826         default:
 827                 error = -EOPNOTSUPP;
 828         }
 829
 830  fail:
 831         filemap_invalidate_unlock(inode->i_mapping);
 832         return error;
 833 }
 834
 835 static int blkdev_mmap(struct file *file, struct vm_area_struct *vma)
 836 {
 837         struct inode *bd_inode = bdev_file_inode(file);
 838
 839         if (bdev_read_only(I_BDEV(bd_inode)))
 840                 return generic_file_readonly_mmap(file, vma);
 841
 842         return generic_file_mmap(file, vma);
 843 }
 844
 845 const struct file_operations def_blk_fops = {
 846         .open           = blkdev_open,
 847         .release        = blkdev_release,
 848         .llseek         = blkdev_llseek,
 849         .read_iter      = blkdev_read_iter,
 850         .write_iter     = blkdev_write_iter,
 851         .iopoll         = iocb_bio_iopoll,
 852         .mmap           = blkdev_mmap,
 853         .fsync          = blkdev_fsync,
 854         .unlocked_ioctl = blkdev_ioctl,
 855 #ifdef CONFIG_COMPAT
 856         .compat_ioctl   = compat_blkdev_ioctl,
 857 #endif
 858         .splice_read    = filemap_splice_read,
 859         .splice_write   = iter_file_splice_write,
 860         .fallocate      = blkdev_fallocate,
 861         .fop_flags      = FOP_BUFFER_RASYNC,
 862 };
 863
 864 static __init int blkdev_init(void)
 865 {
 866         return bioset_init(&blkdev_dio_pool, 4,
 867                                 offsetof(struct blkdev_dio, bio),
 868                                 BIOSET_NEED_BVECS|BIOSET_PERCPU_CACHE);
 869 }
 870 module_init(blkdev_init);