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