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 "blk.h"
  19
  20 static inline struct inode *bdev_file_inode(struct file *file)
  21 {
  22         return file->f_mapping->host;
  23 }
  24
  25 static int blkdev_get_block(struct inode *inode, sector_t iblock,
  26                 struct buffer_head *bh, int create)
  27 {
  28         bh->b_bdev = I_BDEV(inode);
  29         bh->b_blocknr = iblock;
  30         set_buffer_mapped(bh);
  31         return 0;
  32 }
  33
  34 static unsigned int dio_bio_write_op(struct kiocb *iocb)
  35 {
  36         unsigned int op = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
  37
  38         /* avoid the need for a I/O completion work item */
  39         if (iocb->ki_flags & IOCB_DSYNC)
  40                 op |= REQ_FUA;
  41         return op;
  42 }
  43
  44 #define DIO_INLINE_BIO_VECS 4
  45
  46 static void blkdev_bio_end_io_simple(struct bio *bio)
  47 {
  48         struct task_struct *waiter = bio->bi_private;
  49
  50         WRITE_ONCE(bio->bi_private, NULL);
  51         blk_wake_io_task(waiter);
  52 }
  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 ((pos | iov_iter_alignment(iter)) &
  65             (bdev_logical_block_size(bdev) - 1))
  66                 return -EINVAL;
  67
  68         if (nr_pages <= DIO_INLINE_BIO_VECS)
  69                 vecs = inline_vecs;
  70         else {
  71                 vecs = kmalloc_array(nr_pages, sizeof(struct bio_vec),
  72                                      GFP_KERNEL);
  73                 if (!vecs)
  74                         return -ENOMEM;
  75         }
  76
  77         bio_init(&bio, vecs, nr_pages);
  78         bio_set_dev(&bio, bdev);
  79         bio.bi_iter.bi_sector = pos >> 9;
  80         bio.bi_write_hint = iocb->ki_hint;
  81         bio.bi_private = current;
  82         bio.bi_end_io = blkdev_bio_end_io_simple;
  83         bio.bi_ioprio = iocb->ki_ioprio;
  84
  85         ret = bio_iov_iter_get_pages(&bio, iter);
  86         if (unlikely(ret))
  87                 goto out;
  88         ret = bio.bi_iter.bi_size;
  89
  90         if (iov_iter_rw(iter) == READ) {
  91                 bio.bi_opf = REQ_OP_READ;
  92                 if (iter_is_iovec(iter))
  93                         should_dirty = true;
  94         } else {
  95                 bio.bi_opf = dio_bio_write_op(iocb);
  96                 task_io_account_write(ret);
  97         }
  98         if (iocb->ki_flags & IOCB_NOWAIT)
  99                 bio.bi_opf |= REQ_NOWAIT;
 100         if (iocb->ki_flags & IOCB_HIPRI)
 101                 bio_set_polled(&bio, iocb);
 102
 103         submit_bio(&bio);
 104         for (;;) {
 105                 set_current_state(TASK_UNINTERRUPTIBLE);
 106                 if (!READ_ONCE(bio.bi_private))
 107                         break;
 108                 if (!(iocb->ki_flags & IOCB_HIPRI) || !bio_poll(&bio, NULL, 0))
 109                         blk_io_schedule();
 110         }
 111         __set_current_state(TASK_RUNNING);
 112
 113         bio_release_pages(&bio, should_dirty);
 114         if (unlikely(bio.bi_status))
 115                 ret = blk_status_to_errno(bio.bi_status);
 116
 117 out:
 118         if (vecs != inline_vecs)
 119                 kfree(vecs);
 120
 121         bio_uninit(&bio);
 122
 123         return ret;
 124 }
 125
 126 enum {
 127         DIO_MULTI_BIO           = 1,
 128         DIO_SHOULD_DIRTY        = 2,
 129         DIO_IS_SYNC             = 4,
 130 };
 131
 132 struct blkdev_dio {
 133         union {
 134                 struct kiocb            *iocb;
 135                 struct task_struct      *waiter;
 136         };
 137         size_t                  size;
 138         atomic_t                ref;
 139         unsigned int            flags;
 140         struct bio              bio ____cacheline_aligned_in_smp;
 141 };
 142
 143 static struct bio_set blkdev_dio_pool;
 144
 145 static void blkdev_bio_end_io(struct bio *bio)
 146 {
 147         struct blkdev_dio *dio = bio->bi_private;
 148         bool should_dirty = dio->flags & DIO_SHOULD_DIRTY;
 149
 150         if (bio->bi_status && !dio->bio.bi_status)
 151                 dio->bio.bi_status = bio->bi_status;
 152
 153         if (!(dio->flags & DIO_MULTI_BIO) || atomic_dec_and_test(&dio->ref)) {
 154                 if (!(dio->flags & DIO_IS_SYNC)) {
 155                         struct kiocb *iocb = dio->iocb;
 156                         ssize_t ret;
 157
 158                         WRITE_ONCE(iocb->private, NULL);
 159
 160                         if (likely(!dio->bio.bi_status)) {
 161                                 ret = dio->size;
 162                                 iocb->ki_pos += ret;
 163                         } else {
 164                                 ret = blk_status_to_errno(dio->bio.bi_status);
 165                         }
 166
 167                         dio->iocb->ki_complete(iocb, ret, 0);
 168                         if (dio->flags & DIO_MULTI_BIO)
 169                                 bio_put(&dio->bio);
 170                 } else {
 171                         struct task_struct *waiter = dio->waiter;
 172
 173                         WRITE_ONCE(dio->waiter, NULL);
 174                         blk_wake_io_task(waiter);
 175                 }
 176         }
 177
 178         if (should_dirty) {
 179                 bio_check_pages_dirty(bio);
 180         } else {
 181                 bio_release_pages(bio, false);
 182                 bio_put(bio);
 183         }
 184 }
 185
 186 static ssize_t __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
 187                 unsigned int nr_pages)
 188 {
 189         struct block_device *bdev = iocb->ki_filp->private_data;
 190         struct blk_plug plug;
 191         struct blkdev_dio *dio;
 192         struct bio *bio;
 193         bool do_poll = (iocb->ki_flags & IOCB_HIPRI);
 194         bool is_read = (iov_iter_rw(iter) == READ), is_sync;
 195         loff_t pos = iocb->ki_pos;
 196         int ret = 0;
 197
 198         if ((pos | iov_iter_alignment(iter)) &
 199             (bdev_logical_block_size(bdev) - 1))
 200                 return -EINVAL;
 201
 202         bio = bio_alloc_kiocb(iocb, nr_pages, &blkdev_dio_pool);
 203
 204         dio = container_of(bio, struct blkdev_dio, bio);
 205         is_sync = is_sync_kiocb(iocb);
 206         if (is_sync) {
 207                 dio->flags = DIO_IS_SYNC;
 208                 dio->waiter = current;
 209                 bio_get(bio);
 210         } else {
 211                 dio->flags = 0;
 212                 dio->iocb = iocb;
 213         }
 214
 215         dio->size = 0;
 216         if (is_read && iter_is_iovec(iter))
 217                 dio->flags |= DIO_SHOULD_DIRTY;
 218
 219         /*
 220          * Don't plug for HIPRI/polled IO, as those should go straight
 221          * to issue
 222          */
 223         if (!(iocb->ki_flags & IOCB_HIPRI))
 224                 blk_start_plug(&plug);
 225
 226         for (;;) {
 227                 bio_set_dev(bio, bdev);
 228                 bio->bi_iter.bi_sector = pos >> 9;
 229                 bio->bi_write_hint = iocb->ki_hint;
 230                 bio->bi_private = dio;
 231                 bio->bi_end_io = blkdev_bio_end_io;
 232                 bio->bi_ioprio = iocb->ki_ioprio;
 233
 234                 ret = bio_iov_iter_get_pages(bio, iter);
 235                 if (unlikely(ret)) {
 236                         bio->bi_status = BLK_STS_IOERR;
 237                         bio_endio(bio);
 238                         break;
 239                 }
 240
 241                 if (is_read) {
 242                         bio->bi_opf = REQ_OP_READ;
 243                         if (dio->flags & DIO_SHOULD_DIRTY)
 244                                 bio_set_pages_dirty(bio);
 245                 } else {
 246                         bio->bi_opf = dio_bio_write_op(iocb);
 247                         task_io_account_write(bio->bi_iter.bi_size);
 248                 }
 249                 if (iocb->ki_flags & IOCB_NOWAIT)
 250                         bio->bi_opf |= REQ_NOWAIT;
 251
 252                 dio->size += bio->bi_iter.bi_size;
 253                 pos += bio->bi_iter.bi_size;
 254
 255                 nr_pages = bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS);
 256                 if (!nr_pages) {
 257                         if (do_poll)
 258                                 bio_set_polled(bio, iocb);
 259                         submit_bio(bio);
 260                         if (do_poll)
 261                                 WRITE_ONCE(iocb->private, bio);
 262                         break;
 263                 }
 264                 if (!(dio->flags & DIO_MULTI_BIO)) {
 265                         /*
 266                          * AIO needs an extra reference to ensure the dio
 267                          * structure which is embedded into the first bio
 268                          * stays around.
 269                          */
 270                         if (!is_sync)
 271                                 bio_get(bio);
 272                         dio->flags |= DIO_MULTI_BIO;
 273                         atomic_set(&dio->ref, 2);
 274                         do_poll = false;
 275                 } else {
 276                         atomic_inc(&dio->ref);
 277                 }
 278
 279                 submit_bio(bio);
 280                 bio = bio_alloc(GFP_KERNEL, nr_pages);
 281         }
 282
 283         if (!(iocb->ki_flags & IOCB_HIPRI))
 284                 blk_finish_plug(&plug);
 285
 286         if (!is_sync)
 287                 return -EIOCBQUEUED;
 288
 289         for (;;) {
 290                 set_current_state(TASK_UNINTERRUPTIBLE);
 291                 if (!READ_ONCE(dio->waiter))
 292                         break;
 293
 294                 if (!do_poll || !bio_poll(bio, NULL, 0))
 295                         blk_io_schedule();
 296         }
 297         __set_current_state(TASK_RUNNING);
 298
 299         if (!ret)
 300                 ret = blk_status_to_errno(dio->bio.bi_status);
 301         if (likely(!ret))
 302                 ret = dio->size;
 303
 304         bio_put(&dio->bio);
 305         return ret;
 306 }
 307
 308 static ssize_t blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 309 {
 310         unsigned int nr_pages;
 311
 312         if (!iov_iter_count(iter))
 313                 return 0;
 314
 315         nr_pages = bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS + 1);
 316         if (is_sync_kiocb(iocb) && nr_pages <= BIO_MAX_VECS)
 317                 return __blkdev_direct_IO_simple(iocb, iter, nr_pages);
 318
 319         return __blkdev_direct_IO(iocb, iter, bio_max_segs(nr_pages));
 320 }
 321
 322 static int blkdev_writepage(struct page *page, struct writeback_control *wbc)
 323 {
 324         return block_write_full_page(page, blkdev_get_block, wbc);
 325 }
 326
 327 static int blkdev_readpage(struct file * file, struct page * page)
 328 {
 329         return block_read_full_page(page, blkdev_get_block);
 330 }
 331
 332 static void blkdev_readahead(struct readahead_control *rac)
 333 {
 334         mpage_readahead(rac, blkdev_get_block);
 335 }
 336
 337 static int blkdev_write_begin(struct file *file, struct address_space *mapping,
 338                 loff_t pos, unsigned len, unsigned flags, struct page **pagep,
 339                 void **fsdata)
 340 {
 341         return block_write_begin(mapping, pos, len, flags, pagep,
 342                                  blkdev_get_block);
 343 }
 344
 345 static int blkdev_write_end(struct file *file, struct address_space *mapping,
 346                 loff_t pos, unsigned len, unsigned copied, struct page *page,
 347                 void *fsdata)
 348 {
 349         int ret;
 350         ret = block_write_end(file, mapping, pos, len, copied, page, fsdata);
 351
 352         unlock_page(page);
 353         put_page(page);
 354
 355         return ret;
 356 }
 357
 358 static int blkdev_writepages(struct address_space *mapping,
 359                              struct writeback_control *wbc)
 360 {
 361         return generic_writepages(mapping, wbc);
 362 }
 363
 364 const struct address_space_operations def_blk_aops = {
 365         .set_page_dirty = __set_page_dirty_buffers,
 366         .readpage       = blkdev_readpage,
 367         .readahead      = blkdev_readahead,
 368         .writepage      = blkdev_writepage,
 369         .write_begin    = blkdev_write_begin,
 370         .write_end      = blkdev_write_end,
 371         .writepages     = blkdev_writepages,
 372         .direct_IO      = blkdev_direct_IO,
 373         .migratepage    = buffer_migrate_page_norefs,
 374         .is_dirty_writeback = buffer_check_dirty_writeback,
 375 };
 376
 377 /*
 378  * for a block special file file_inode(file)->i_size is zero
 379  * so we compute the size by hand (just as in block_read/write above)
 380  */
 381 static loff_t blkdev_llseek(struct file *file, loff_t offset, int whence)
 382 {
 383         struct inode *bd_inode = bdev_file_inode(file);
 384         loff_t retval;
 385
 386         inode_lock(bd_inode);
 387         retval = fixed_size_llseek(file, offset, whence, i_size_read(bd_inode));
 388         inode_unlock(bd_inode);
 389         return retval;
 390 }
 391
 392 static int blkdev_fsync(struct file *filp, loff_t start, loff_t end,
 393                 int datasync)
 394 {
 395         struct block_device *bdev = filp->private_data;
 396         int error;
 397
 398         error = file_write_and_wait_range(filp, start, end);
 399         if (error)
 400                 return error;
 401
 402         /*
 403          * There is no need to serialise calls to blkdev_issue_flush with
 404          * i_mutex and doing so causes performance issues with concurrent
 405          * O_SYNC writers to a block device.
 406          */
 407         error = blkdev_issue_flush(bdev);
 408         if (error == -EOPNOTSUPP)
 409                 error = 0;
 410
 411         return error;
 412 }
 413
 414 static int blkdev_open(struct inode *inode, struct file *filp)
 415 {
 416         struct block_device *bdev;
 417
 418         /*
 419          * Preserve backwards compatibility and allow large file access
 420          * even if userspace doesn't ask for it explicitly. Some mkfs
 421          * binary needs it. We might want to drop this workaround
 422          * during an unstable branch.
 423          */
 424         filp->f_flags |= O_LARGEFILE;
 425         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
 426
 427         if (filp->f_flags & O_NDELAY)
 428                 filp->f_mode |= FMODE_NDELAY;
 429         if (filp->f_flags & O_EXCL)
 430                 filp->f_mode |= FMODE_EXCL;
 431         if ((filp->f_flags & O_ACCMODE) == 3)
 432                 filp->f_mode |= FMODE_WRITE_IOCTL;
 433
 434         bdev = blkdev_get_by_dev(inode->i_rdev, filp->f_mode, filp);
 435         if (IS_ERR(bdev))
 436                 return PTR_ERR(bdev);
 437
 438         filp->private_data = bdev;
 439         filp->f_mapping = bdev->bd_inode->i_mapping;
 440         filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping);
 441         return 0;
 442 }
 443
 444 static int blkdev_close(struct inode *inode, struct file *filp)
 445 {
 446         struct block_device *bdev = filp->private_data;
 447
 448         blkdev_put(bdev, filp->f_mode);
 449         return 0;
 450 }
 451
 452 /*
 453  * Write data to the block device.  Only intended for the block device itself
 454  * and the raw driver which basically is a fake block device.
 455  *
 456  * Does not take i_mutex for the write and thus is not for general purpose
 457  * use.
 458  */
 459 static ssize_t blkdev_write_iter(struct kiocb *iocb, struct iov_iter *from)
 460 {
 461         struct block_device *bdev = iocb->ki_filp->private_data;
 462         struct inode *bd_inode = bdev->bd_inode;
 463         loff_t size = i_size_read(bd_inode);
 464         struct blk_plug plug;
 465         size_t shorted = 0;
 466         ssize_t ret;
 467
 468         if (bdev_read_only(bdev))
 469                 return -EPERM;
 470
 471         if (IS_SWAPFILE(bd_inode) && !is_hibernate_resume_dev(bd_inode->i_rdev))
 472                 return -ETXTBSY;
 473
 474         if (!iov_iter_count(from))
 475                 return 0;
 476
 477         if (iocb->ki_pos >= size)
 478                 return -ENOSPC;
 479
 480         if ((iocb->ki_flags & (IOCB_NOWAIT | IOCB_DIRECT)) == IOCB_NOWAIT)
 481                 return -EOPNOTSUPP;
 482
 483         size -= iocb->ki_pos;
 484         if (iov_iter_count(from) > size) {
 485                 shorted = iov_iter_count(from) - size;
 486                 iov_iter_truncate(from, size);
 487         }
 488
 489         blk_start_plug(&plug);
 490         ret = __generic_file_write_iter(iocb, from);
 491         if (ret > 0)
 492                 ret = generic_write_sync(iocb, ret);
 493         iov_iter_reexpand(from, iov_iter_count(from) + shorted);
 494         blk_finish_plug(&plug);
 495         return ret;
 496 }
 497
 498 static ssize_t blkdev_read_iter(struct kiocb *iocb, struct iov_iter *to)
 499 {
 500         struct block_device *bdev = iocb->ki_filp->private_data;
 501         loff_t size = i_size_read(bdev->bd_inode);
 502         loff_t pos = iocb->ki_pos;
 503         size_t shorted = 0;
 504         ssize_t ret;
 505
 506         if (pos >= size)
 507                 return 0;
 508
 509         size -= pos;
 510         if (iov_iter_count(to) > size) {
 511                 shorted = iov_iter_count(to) - size;
 512                 iov_iter_truncate(to, size);
 513         }
 514
 515         ret = generic_file_read_iter(iocb, to);
 516         iov_iter_reexpand(to, iov_iter_count(to) + shorted);
 517         return ret;
 518 }
 519
 520 #define BLKDEV_FALLOC_FL_SUPPORTED                                      \
 521                 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |           \
 522                  FALLOC_FL_ZERO_RANGE | FALLOC_FL_NO_HIDE_STALE)
 523
 524 static long blkdev_fallocate(struct file *file, int mode, loff_t start,
 525                              loff_t len)
 526 {
 527         struct inode *inode = bdev_file_inode(file);
 528         struct block_device *bdev = I_BDEV(inode);
 529         loff_t end = start + len - 1;
 530         loff_t isize;
 531         int error;
 532
 533         /* Fail if we don't recognize the flags. */
 534         if (mode & ~BLKDEV_FALLOC_FL_SUPPORTED)
 535                 return -EOPNOTSUPP;
 536
 537         /* Don't go off the end of the device. */
 538         isize = i_size_read(bdev->bd_inode);
 539         if (start >= isize)
 540                 return -EINVAL;
 541         if (end >= isize) {
 542                 if (mode & FALLOC_FL_KEEP_SIZE) {
 543                         len = isize - start;
 544                         end = start + len - 1;
 545                 } else
 546                         return -EINVAL;
 547         }
 548
 549         /*
 550          * Don't allow IO that isn't aligned to logical block size.
 551          */
 552         if ((start | len) & (bdev_logical_block_size(bdev) - 1))
 553                 return -EINVAL;
 554
 555         filemap_invalidate_lock(inode->i_mapping);
 556
 557         /* Invalidate the page cache, including dirty pages. */
 558         error = truncate_bdev_range(bdev, file->f_mode, start, end);
 559         if (error)
 560                 goto fail;
 561
 562         switch (mode) {
 563         case FALLOC_FL_ZERO_RANGE:
 564         case FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE:
 565                 error = blkdev_issue_zeroout(bdev, start >> 9, len >> 9,
 566                                             GFP_KERNEL, BLKDEV_ZERO_NOUNMAP);
 567                 break;
 568         case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE:
 569                 error = blkdev_issue_zeroout(bdev, start >> 9, len >> 9,
 570                                              GFP_KERNEL, BLKDEV_ZERO_NOFALLBACK);
 571                 break;
 572         case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE | FALLOC_FL_NO_HIDE_STALE:
 573                 error = blkdev_issue_discard(bdev, start >> 9, len >> 9,
 574                                              GFP_KERNEL, 0);
 575                 break;
 576         default:
 577                 error = -EOPNOTSUPP;
 578         }
 579
 580  fail:
 581         filemap_invalidate_unlock(inode->i_mapping);
 582         return error;
 583 }
 584
 585 const struct file_operations def_blk_fops = {
 586         .open           = blkdev_open,
 587         .release        = blkdev_close,
 588         .llseek         = blkdev_llseek,
 589         .read_iter      = blkdev_read_iter,
 590         .write_iter     = blkdev_write_iter,
 591         .iopoll         = iocb_bio_iopoll,
 592         .mmap           = generic_file_mmap,
 593         .fsync          = blkdev_fsync,
 594         .unlocked_ioctl = blkdev_ioctl,
 595 #ifdef CONFIG_COMPAT
 596         .compat_ioctl   = compat_blkdev_ioctl,
 597 #endif
 598         .splice_read    = generic_file_splice_read,
 599         .splice_write   = iter_file_splice_write,
 600         .fallocate      = blkdev_fallocate,
 601 };
 602
 603 static __init int blkdev_init(void)
 604 {
 605         return bioset_init(&blkdev_dio_pool, 4,
 606                                 offsetof(struct blkdev_dio, bio),
 607                                 BIOSET_NEED_BVECS|BIOSET_PERCPU_CACHE);
 608 }
 609 module_init(blkdev_init);