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