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 <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 (iter_is_iovec(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 && iter_is_iovec(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 (iter_is_iovec(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 static int blkdev_writepages(struct address_space *mapping,
 409                              struct writeback_control *wbc)
 410 {
 411         return generic_writepages(mapping, wbc);
 412 }
 413
 414 const struct address_space_operations def_blk_aops = {
 415         .dirty_folio    = block_dirty_folio,
 416         .invalidate_folio = block_invalidate_folio,
 417         .read_folio     = blkdev_read_folio,
 418         .readahead      = blkdev_readahead,
 419         .writepage      = blkdev_writepage,
 420         .write_begin    = blkdev_write_begin,
 421         .write_end      = blkdev_write_end,
 422         .writepages     = blkdev_writepages,
 423         .direct_IO      = blkdev_direct_IO,
 424         .migrate_folio  = buffer_migrate_folio_norefs,
 425         .is_dirty_writeback = buffer_check_dirty_writeback,
 426 };
 427
 428 /*
 429  * for a block special file file_inode(file)->i_size is zero
 430  * so we compute the size by hand (just as in block_read/write above)
 431  */
 432 static loff_t blkdev_llseek(struct file *file, loff_t offset, int whence)
 433 {
 434         struct inode *bd_inode = bdev_file_inode(file);
 435         loff_t retval;
 436
 437         inode_lock(bd_inode);
 438         retval = fixed_size_llseek(file, offset, whence, i_size_read(bd_inode));
 439         inode_unlock(bd_inode);
 440         return retval;
 441 }
 442
 443 static int blkdev_fsync(struct file *filp, loff_t start, loff_t end,
 444                 int datasync)
 445 {
 446         struct block_device *bdev = filp->private_data;
 447         int error;
 448
 449         error = file_write_and_wait_range(filp, start, end);
 450         if (error)
 451                 return error;
 452
 453         /*
 454          * There is no need to serialise calls to blkdev_issue_flush with
 455          * i_mutex and doing so causes performance issues with concurrent
 456          * O_SYNC writers to a block device.
 457          */
 458         error = blkdev_issue_flush(bdev);
 459         if (error == -EOPNOTSUPP)
 460                 error = 0;
 461
 462         return error;
 463 }
 464
 465 static int blkdev_open(struct inode *inode, struct file *filp)
 466 {
 467         struct block_device *bdev;
 468
 469         /*
 470          * Preserve backwards compatibility and allow large file access
 471          * even if userspace doesn't ask for it explicitly. Some mkfs
 472          * binary needs it. We might want to drop this workaround
 473          * during an unstable branch.
 474          */
 475         filp->f_flags |= O_LARGEFILE;
 476         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
 477
 478         if (filp->f_flags & O_NDELAY)
 479                 filp->f_mode |= FMODE_NDELAY;
 480         if (filp->f_flags & O_EXCL)
 481                 filp->f_mode |= FMODE_EXCL;
 482         if ((filp->f_flags & O_ACCMODE) == 3)
 483                 filp->f_mode |= FMODE_WRITE_IOCTL;
 484
 485         bdev = blkdev_get_by_dev(inode->i_rdev, filp->f_mode, filp);
 486         if (IS_ERR(bdev))
 487                 return PTR_ERR(bdev);
 488
 489         filp->private_data = bdev;
 490         filp->f_mapping = bdev->bd_inode->i_mapping;
 491         filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping);
 492         return 0;
 493 }
 494
 495 static int blkdev_close(struct inode *inode, struct file *filp)
 496 {
 497         struct block_device *bdev = filp->private_data;
 498
 499         blkdev_put(bdev, filp->f_mode);
 500         return 0;
 501 }
 502
 503 /*
 504  * Write data to the block device.  Only intended for the block device itself
 505  * and the raw driver which basically is a fake block device.
 506  *
 507  * Does not take i_mutex for the write and thus is not for general purpose
 508  * use.
 509  */
 510 static ssize_t blkdev_write_iter(struct kiocb *iocb, struct iov_iter *from)
 511 {
 512         struct block_device *bdev = iocb->ki_filp->private_data;
 513         struct inode *bd_inode = bdev->bd_inode;
 514         loff_t size = bdev_nr_bytes(bdev);
 515         struct blk_plug plug;
 516         size_t shorted = 0;
 517         ssize_t ret;
 518
 519         if (bdev_read_only(bdev))
 520                 return -EPERM;
 521
 522         if (IS_SWAPFILE(bd_inode) && !is_hibernate_resume_dev(bd_inode->i_rdev))
 523                 return -ETXTBSY;
 524
 525         if (!iov_iter_count(from))
 526                 return 0;
 527
 528         if (iocb->ki_pos >= size)
 529                 return -ENOSPC;
 530
 531         if ((iocb->ki_flags & (IOCB_NOWAIT | IOCB_DIRECT)) == IOCB_NOWAIT)
 532                 return -EOPNOTSUPP;
 533
 534         size -= iocb->ki_pos;
 535         if (iov_iter_count(from) > size) {
 536                 shorted = iov_iter_count(from) - size;
 537                 iov_iter_truncate(from, size);
 538         }
 539
 540         blk_start_plug(&plug);
 541         ret = __generic_file_write_iter(iocb, from);
 542         if (ret > 0)
 543                 ret = generic_write_sync(iocb, ret);
 544         iov_iter_reexpand(from, iov_iter_count(from) + shorted);
 545         blk_finish_plug(&plug);
 546         return ret;
 547 }
 548
 549 static ssize_t blkdev_read_iter(struct kiocb *iocb, struct iov_iter *to)
 550 {
 551         struct block_device *bdev = iocb->ki_filp->private_data;
 552         loff_t size = bdev_nr_bytes(bdev);
 553         loff_t pos = iocb->ki_pos;
 554         size_t shorted = 0;
 555         ssize_t ret = 0;
 556         size_t count;
 557
 558         if (unlikely(pos + iov_iter_count(to) > size)) {
 559                 if (pos >= size)
 560                         return 0;
 561                 size -= pos;
 562                 shorted = iov_iter_count(to) - size;
 563                 iov_iter_truncate(to, size);
 564         }
 565
 566         count = iov_iter_count(to);
 567         if (!count)
 568                 goto reexpand; /* skip atime */
 569
 570         if (iocb->ki_flags & IOCB_DIRECT) {
 571                 struct address_space *mapping = iocb->ki_filp->f_mapping;
 572
 573                 if (iocb->ki_flags & IOCB_NOWAIT) {
 574                         if (filemap_range_needs_writeback(mapping, pos,
 575                                                           pos + count - 1)) {
 576                                 ret = -EAGAIN;
 577                                 goto reexpand;
 578                         }
 579                 } else {
 580                         ret = filemap_write_and_wait_range(mapping, pos,
 581                                                            pos + count - 1);
 582                         if (ret < 0)
 583                                 goto reexpand;
 584                 }
 585
 586                 file_accessed(iocb->ki_filp);
 587
 588                 ret = blkdev_direct_IO(iocb, to);
 589                 if (ret >= 0) {
 590                         iocb->ki_pos += ret;
 591                         count -= ret;
 592                 }
 593                 iov_iter_revert(to, count - iov_iter_count(to));
 594                 if (ret < 0 || !count)
 595                         goto reexpand;
 596         }
 597
 598         ret = filemap_read(iocb, to, ret);
 599
 600 reexpand:
 601         if (unlikely(shorted))
 602                 iov_iter_reexpand(to, iov_iter_count(to) + shorted);
 603         return ret;
 604 }
 605
 606 #define BLKDEV_FALLOC_FL_SUPPORTED                                      \
 607                 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |           \
 608                  FALLOC_FL_ZERO_RANGE | FALLOC_FL_NO_HIDE_STALE)
 609
 610 static long blkdev_fallocate(struct file *file, int mode, loff_t start,
 611                              loff_t len)
 612 {
 613         struct inode *inode = bdev_file_inode(file);
 614         struct block_device *bdev = I_BDEV(inode);
 615         loff_t end = start + len - 1;
 616         loff_t isize;
 617         int error;
 618
 619         /* Fail if we don't recognize the flags. */
 620         if (mode & ~BLKDEV_FALLOC_FL_SUPPORTED)
 621                 return -EOPNOTSUPP;
 622
 623         /* Don't go off the end of the device. */
 624         isize = bdev_nr_bytes(bdev);
 625         if (start >= isize)
 626                 return -EINVAL;
 627         if (end >= isize) {
 628                 if (mode & FALLOC_FL_KEEP_SIZE) {
 629                         len = isize - start;
 630                         end = start + len - 1;
 631                 } else
 632                         return -EINVAL;
 633         }
 634
 635         /*
 636          * Don't allow IO that isn't aligned to logical block size.
 637          */
 638         if ((start | len) & (bdev_logical_block_size(bdev) - 1))
 639                 return -EINVAL;
 640
 641         filemap_invalidate_lock(inode->i_mapping);
 642
 643         /* Invalidate the page cache, including dirty pages. */
 644         error = truncate_bdev_range(bdev, file->f_mode, start, end);
 645         if (error)
 646                 goto fail;
 647
 648         switch (mode) {
 649         case FALLOC_FL_ZERO_RANGE:
 650         case FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE:
 651                 error = blkdev_issue_zeroout(bdev, start >> SECTOR_SHIFT,
 652                                              len >> SECTOR_SHIFT, GFP_KERNEL,
 653                                              BLKDEV_ZERO_NOUNMAP);
 654                 break;
 655         case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE:
 656                 error = blkdev_issue_zeroout(bdev, start >> SECTOR_SHIFT,
 657                                              len >> SECTOR_SHIFT, GFP_KERNEL,
 658                                              BLKDEV_ZERO_NOFALLBACK);
 659                 break;
 660         case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE | FALLOC_FL_NO_HIDE_STALE:
 661                 error = blkdev_issue_discard(bdev, start >> SECTOR_SHIFT,
 662                                              len >> SECTOR_SHIFT, GFP_KERNEL);
 663                 break;
 664         default:
 665                 error = -EOPNOTSUPP;
 666         }
 667
 668  fail:
 669         filemap_invalidate_unlock(inode->i_mapping);
 670         return error;
 671 }
 672
 673 const struct file_operations def_blk_fops = {
 674         .open           = blkdev_open,
 675         .release        = blkdev_close,
 676         .llseek         = blkdev_llseek,
 677         .read_iter      = blkdev_read_iter,
 678         .write_iter     = blkdev_write_iter,
 679         .iopoll         = iocb_bio_iopoll,
 680         .mmap           = generic_file_mmap,
 681         .fsync          = blkdev_fsync,
 682         .unlocked_ioctl = blkdev_ioctl,
 683 #ifdef CONFIG_COMPAT
 684         .compat_ioctl   = compat_blkdev_ioctl,
 685 #endif
 686         .splice_read    = generic_file_splice_read,
 687         .splice_write   = iter_file_splice_write,
 688         .fallocate      = blkdev_fallocate,
 689 };
 690
 691 static __init int blkdev_init(void)
 692 {
 693         return bioset_init(&blkdev_dio_pool, 4,
 694                                 offsetof(struct blkdev_dio, bio),
 695                                 BIOSET_NEED_BVECS|BIOSET_PERCPU_CACHE);
 696 }
 697 module_init(blkdev_init);