fs/iomap/direct-io.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Red Hat, Inc.
   4  * Copyright (c) 2016-2021 Christoph Hellwig.
   5  */
   6 #include <linux/module.h>
   7 #include <linux/compiler.h>
   8 #include <linux/fs.h>
   9 #include <linux/fscrypt.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/iomap.h>
  12 #include <linux/backing-dev.h>
  13 #include <linux/uio.h>
  14 #include <linux/task_io_accounting_ops.h>
  15 #include "trace.h"
  16
  17 #include "../internal.h"
  18
  19 /*
  20  * Private flags for iomap_dio, must not overlap with the public ones in
  21  * iomap.h:
  22  */
  23 #define IOMAP_DIO_WRITE_FUA     (1 << 28)
  24 #define IOMAP_DIO_NEED_SYNC     (1 << 29)
  25 #define IOMAP_DIO_WRITE         (1 << 30)
  26 #define IOMAP_DIO_DIRTY         (1 << 31)
  27
  28 struct iomap_dio {
  29         struct kiocb            *iocb;
  30         const struct iomap_dio_ops *dops;
  31         loff_t                  i_size;
  32         loff_t                  size;
  33         atomic_t                ref;
  34         unsigned                flags;
  35         int                     error;
  36         size_t                  done_before;
  37         bool                    wait_for_completion;
  38
  39         union {
  40                 /* used during submission and for synchronous completion: */
  41                 struct {
  42                         struct iov_iter         *iter;
  43                         struct task_struct      *waiter;
  44                         struct bio              *poll_bio;
  45                 } submit;
  46
  47                 /* used for aio completion: */
  48                 struct {
  49                         struct work_struct      work;
  50                 } aio;
  51         };
  52 };
  53
  54 static struct bio *iomap_dio_alloc_bio(const struct iomap_iter *iter,
  55                 struct iomap_dio *dio, unsigned short nr_vecs, blk_opf_t opf)
  56 {
  57         if (dio->dops && dio->dops->bio_set)
  58                 return bio_alloc_bioset(iter->iomap.bdev, nr_vecs, opf,
  59                                         GFP_KERNEL, dio->dops->bio_set);
  60         return bio_alloc(iter->iomap.bdev, nr_vecs, opf, GFP_KERNEL);
  61 }
  62
  63 static void iomap_dio_submit_bio(const struct iomap_iter *iter,
  64                 struct iomap_dio *dio, struct bio *bio, loff_t pos)
  65 {
  66         atomic_inc(&dio->ref);
  67
  68         /* Sync dio can't be polled reliably */
  69         if ((dio->iocb->ki_flags & IOCB_HIPRI) && !is_sync_kiocb(dio->iocb)) {
  70                 bio_set_polled(bio, dio->iocb);
  71                 dio->submit.poll_bio = bio;
  72         }
  73
  74         if (dio->dops && dio->dops->submit_io)
  75                 dio->dops->submit_io(iter, bio, pos);
  76         else
  77                 submit_bio(bio);
  78 }
  79
  80 ssize_t iomap_dio_complete(struct iomap_dio *dio)
  81 {
  82         const struct iomap_dio_ops *dops = dio->dops;
  83         struct kiocb *iocb = dio->iocb;
  84         struct inode *inode = file_inode(iocb->ki_filp);
  85         loff_t offset = iocb->ki_pos;
  86         ssize_t ret = dio->error;
  87
  88         if (dops && dops->end_io)
  89                 ret = dops->end_io(iocb, dio->size, ret, dio->flags);
  90
  91         if (likely(!ret)) {
  92                 ret = dio->size;
  93                 /* check for short read */
  94                 if (offset + ret > dio->i_size &&
  95                     !(dio->flags & IOMAP_DIO_WRITE))
  96                         ret = dio->i_size - offset;
  97                 iocb->ki_pos += ret;
  98         }
  99
 100         /*
 101          * Try again to invalidate clean pages which might have been cached by
 102          * non-direct readahead, or faulted in by get_user_pages() if the source
 103          * of the write was an mmap'ed region of the file we're writing.  Either
 104          * one is a pretty crazy thing to do, so we don't support it 100%.  If
 105          * this invalidation fails, tough, the write still worked...
 106          *
 107          * And this page cache invalidation has to be after ->end_io(), as some
 108          * filesystems convert unwritten extents to real allocations in
 109          * ->end_io() when necessary, otherwise a racing buffer read would cache
 110          * zeros from unwritten extents.
 111          */
 112         if (!dio->error && dio->size &&
 113             (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
 114                 int err;
 115                 err = invalidate_inode_pages2_range(inode->i_mapping,
 116                                 offset >> PAGE_SHIFT,
 117                                 (offset + dio->size - 1) >> PAGE_SHIFT);
 118                 if (err)
 119                         dio_warn_stale_pagecache(iocb->ki_filp);
 120         }
 121
 122         inode_dio_end(file_inode(iocb->ki_filp));
 123         /*
 124          * If this is a DSYNC write, make sure we push it to stable storage now
 125          * that we've written data.
 126          */
 127         if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
 128                 ret = generic_write_sync(iocb, ret);
 129
 130         if (ret > 0)
 131                 ret += dio->done_before;
 132
 133         trace_iomap_dio_complete(iocb, dio->error, ret);
 134         kfree(dio);
 135
 136         return ret;
 137 }
 138 EXPORT_SYMBOL_GPL(iomap_dio_complete);
 139
 140 static void iomap_dio_complete_work(struct work_struct *work)
 141 {
 142         struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
 143         struct kiocb *iocb = dio->iocb;
 144
 145         iocb->ki_complete(iocb, iomap_dio_complete(dio));
 146 }
 147
 148 /*
 149  * Set an error in the dio if none is set yet.  We have to use cmpxchg
 150  * as the submission context and the completion context(s) can race to
 151  * update the error.
 152  */
 153 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
 154 {
 155         cmpxchg(&dio->error, 0, ret);
 156 }
 157
 158 void iomap_dio_bio_end_io(struct bio *bio)
 159 {
 160         struct iomap_dio *dio = bio->bi_private;
 161         bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
 162
 163         if (bio->bi_status)
 164                 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
 165
 166         if (atomic_dec_and_test(&dio->ref)) {
 167                 if (dio->wait_for_completion) {
 168                         struct task_struct *waiter = dio->submit.waiter;
 169                         WRITE_ONCE(dio->submit.waiter, NULL);
 170                         blk_wake_io_task(waiter);
 171                 } else if (dio->flags & IOMAP_DIO_WRITE) {
 172                         struct inode *inode = file_inode(dio->iocb->ki_filp);
 173
 174                         WRITE_ONCE(dio->iocb->private, NULL);
 175                         INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
 176                         queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
 177                 } else {
 178                         WRITE_ONCE(dio->iocb->private, NULL);
 179                         iomap_dio_complete_work(&dio->aio.work);
 180                 }
 181         }
 182
 183         if (should_dirty) {
 184                 bio_check_pages_dirty(bio);
 185         } else {
 186                 bio_release_pages(bio, false);
 187                 bio_put(bio);
 188         }
 189 }
 190 EXPORT_SYMBOL_GPL(iomap_dio_bio_end_io);
 191
 192 static void iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio,
 193                 loff_t pos, unsigned len)
 194 {
 195         struct inode *inode = file_inode(dio->iocb->ki_filp);
 196         struct page *page = ZERO_PAGE(0);
 197         struct bio *bio;
 198
 199         bio = iomap_dio_alloc_bio(iter, dio, 1, REQ_OP_WRITE | REQ_SYNC | REQ_IDLE);
 200         fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
 201                                   GFP_KERNEL);
 202         bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos);
 203         bio->bi_private = dio;
 204         bio->bi_end_io = iomap_dio_bio_end_io;
 205
 206         get_page(page);
 207         __bio_add_page(bio, page, len, 0);
 208         iomap_dio_submit_bio(iter, dio, bio, pos);
 209 }
 210
 211 /*
 212  * Figure out the bio's operation flags from the dio request, the
 213  * mapping, and whether or not we want FUA.  Note that we can end up
 214  * clearing the WRITE_FUA flag in the dio request.
 215  */
 216 static inline blk_opf_t iomap_dio_bio_opflags(struct iomap_dio *dio,
 217                 const struct iomap *iomap, bool use_fua)
 218 {
 219         blk_opf_t opflags = REQ_SYNC | REQ_IDLE;
 220
 221         if (!(dio->flags & IOMAP_DIO_WRITE))
 222                 return REQ_OP_READ;
 223
 224         opflags |= REQ_OP_WRITE;
 225         if (use_fua)
 226                 opflags |= REQ_FUA;
 227         else
 228                 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
 229
 230         return opflags;
 231 }
 232
 233 static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
 234                 struct iomap_dio *dio)
 235 {
 236         const struct iomap *iomap = &iter->iomap;
 237         struct inode *inode = iter->inode;
 238         unsigned int fs_block_size = i_blocksize(inode), pad;
 239         loff_t length = iomap_length(iter);
 240         loff_t pos = iter->pos;
 241         blk_opf_t bio_opf;
 242         struct bio *bio;
 243         bool need_zeroout = false;
 244         bool use_fua = false;
 245         int nr_pages, ret = 0;
 246         size_t copied = 0;
 247         size_t orig_count;
 248
 249         if ((pos | length) & (bdev_logical_block_size(iomap->bdev) - 1) ||
 250             !bdev_iter_is_aligned(iomap->bdev, dio->submit.iter))
 251                 return -EINVAL;
 252
 253         if (iomap->type == IOMAP_UNWRITTEN) {
 254                 dio->flags |= IOMAP_DIO_UNWRITTEN;
 255                 need_zeroout = true;
 256         }
 257
 258         if (iomap->flags & IOMAP_F_SHARED)
 259                 dio->flags |= IOMAP_DIO_COW;
 260
 261         if (iomap->flags & IOMAP_F_NEW) {
 262                 need_zeroout = true;
 263         } else if (iomap->type == IOMAP_MAPPED) {
 264                 /*
 265                  * Use a FUA write if we need datasync semantics, this is a pure
 266                  * data IO that doesn't require any metadata updates (including
 267                  * after IO completion such as unwritten extent conversion) and
 268                  * the underlying device supports FUA. This allows us to avoid
 269                  * cache flushes on IO completion.
 270                  */
 271                 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
 272                     (dio->flags & IOMAP_DIO_WRITE_FUA) && bdev_fua(iomap->bdev))
 273                         use_fua = true;
 274         }
 275
 276         /*
 277          * Save the original count and trim the iter to just the extent we
 278          * are operating on right now.  The iter will be re-expanded once
 279          * we are done.
 280          */
 281         orig_count = iov_iter_count(dio->submit.iter);
 282         iov_iter_truncate(dio->submit.iter, length);
 283
 284         if (!iov_iter_count(dio->submit.iter))
 285                 goto out;
 286
 287         /*
 288          * We can only poll for single bio I/Os.
 289          */
 290         if (need_zeroout ||
 291             ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode)))
 292                 dio->iocb->ki_flags &= ~IOCB_HIPRI;
 293
 294         if (need_zeroout) {
 295                 /* zero out from the start of the block to the write offset */
 296                 pad = pos & (fs_block_size - 1);
 297                 if (pad)
 298                         iomap_dio_zero(iter, dio, pos - pad, pad);
 299         }
 300
 301         /*
 302          * Set the operation flags early so that bio_iov_iter_get_pages
 303          * can set up the page vector appropriately for a ZONE_APPEND
 304          * operation.
 305          */
 306         bio_opf = iomap_dio_bio_opflags(dio, iomap, use_fua);
 307
 308         nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS);
 309         do {
 310                 size_t n;
 311                 if (dio->error) {
 312                         iov_iter_revert(dio->submit.iter, copied);
 313                         copied = ret = 0;
 314                         goto out;
 315                 }
 316
 317                 bio = iomap_dio_alloc_bio(iter, dio, nr_pages, bio_opf);
 318                 fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
 319                                           GFP_KERNEL);
 320                 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
 321                 bio->bi_ioprio = dio->iocb->ki_ioprio;
 322                 bio->bi_private = dio;
 323                 bio->bi_end_io = iomap_dio_bio_end_io;
 324
 325                 ret = bio_iov_iter_get_pages(bio, dio->submit.iter);
 326                 if (unlikely(ret)) {
 327                         /*
 328                          * We have to stop part way through an IO. We must fall
 329                          * through to the sub-block tail zeroing here, otherwise
 330                          * this short IO may expose stale data in the tail of
 331                          * the block we haven't written data to.
 332                          */
 333                         bio_put(bio);
 334                         goto zero_tail;
 335                 }
 336
 337                 n = bio->bi_iter.bi_size;
 338                 if (dio->flags & IOMAP_DIO_WRITE) {
 339                         task_io_account_write(n);
 340                 } else {
 341                         if (dio->flags & IOMAP_DIO_DIRTY)
 342                                 bio_set_pages_dirty(bio);
 343                 }
 344
 345                 dio->size += n;
 346                 copied += n;
 347
 348                 nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter,
 349                                                  BIO_MAX_VECS);
 350                 /*
 351                  * We can only poll for single bio I/Os.
 352                  */
 353                 if (nr_pages)
 354                         dio->iocb->ki_flags &= ~IOCB_HIPRI;
 355                 iomap_dio_submit_bio(iter, dio, bio, pos);
 356                 pos += n;
 357         } while (nr_pages);
 358
 359         /*
 360          * We need to zeroout the tail of a sub-block write if the extent type
 361          * requires zeroing or the write extends beyond EOF. If we don't zero
 362          * the block tail in the latter case, we can expose stale data via mmap
 363          * reads of the EOF block.
 364          */
 365 zero_tail:
 366         if (need_zeroout ||
 367             ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
 368                 /* zero out from the end of the write to the end of the block */
 369                 pad = pos & (fs_block_size - 1);
 370                 if (pad)
 371                         iomap_dio_zero(iter, dio, pos, fs_block_size - pad);
 372         }
 373 out:
 374         /* Undo iter limitation to current extent */
 375         iov_iter_reexpand(dio->submit.iter, orig_count - copied);
 376         if (copied)
 377                 return copied;
 378         return ret;
 379 }
 380
 381 static loff_t iomap_dio_hole_iter(const struct iomap_iter *iter,
 382                 struct iomap_dio *dio)
 383 {
 384         loff_t length = iov_iter_zero(iomap_length(iter), dio->submit.iter);
 385
 386         dio->size += length;
 387         if (!length)
 388                 return -EFAULT;
 389         return length;
 390 }
 391
 392 static loff_t iomap_dio_inline_iter(const struct iomap_iter *iomi,
 393                 struct iomap_dio *dio)
 394 {
 395         const struct iomap *iomap = &iomi->iomap;
 396         struct iov_iter *iter = dio->submit.iter;
 397         void *inline_data = iomap_inline_data(iomap, iomi->pos);
 398         loff_t length = iomap_length(iomi);
 399         loff_t pos = iomi->pos;
 400         size_t copied;
 401
 402         if (WARN_ON_ONCE(!iomap_inline_data_valid(iomap)))
 403                 return -EIO;
 404
 405         if (dio->flags & IOMAP_DIO_WRITE) {
 406                 loff_t size = iomi->inode->i_size;
 407
 408                 if (pos > size)
 409                         memset(iomap_inline_data(iomap, size), 0, pos - size);
 410                 copied = copy_from_iter(inline_data, length, iter);
 411                 if (copied) {
 412                         if (pos + copied > size)
 413                                 i_size_write(iomi->inode, pos + copied);
 414                         mark_inode_dirty(iomi->inode);
 415                 }
 416         } else {
 417                 copied = copy_to_iter(inline_data, length, iter);
 418         }
 419         dio->size += copied;
 420         if (!copied)
 421                 return -EFAULT;
 422         return copied;
 423 }
 424
 425 static loff_t iomap_dio_iter(const struct iomap_iter *iter,
 426                 struct iomap_dio *dio)
 427 {
 428         switch (iter->iomap.type) {
 429         case IOMAP_HOLE:
 430                 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
 431                         return -EIO;
 432                 return iomap_dio_hole_iter(iter, dio);
 433         case IOMAP_UNWRITTEN:
 434                 if (!(dio->flags & IOMAP_DIO_WRITE))
 435                         return iomap_dio_hole_iter(iter, dio);
 436                 return iomap_dio_bio_iter(iter, dio);
 437         case IOMAP_MAPPED:
 438                 return iomap_dio_bio_iter(iter, dio);
 439         case IOMAP_INLINE:
 440                 return iomap_dio_inline_iter(iter, dio);
 441         case IOMAP_DELALLOC:
 442                 /*
 443                  * DIO is not serialised against mmap() access at all, and so
 444                  * if the page_mkwrite occurs between the writeback and the
 445                  * iomap_iter() call in the DIO path, then it will see the
 446                  * DELALLOC block that the page-mkwrite allocated.
 447                  */
 448                 pr_warn_ratelimited("Direct I/O collision with buffered writes! File: %pD4 Comm: %.20s\n",
 449                                     dio->iocb->ki_filp, current->comm);
 450                 return -EIO;
 451         default:
 452                 WARN_ON_ONCE(1);
 453                 return -EIO;
 454         }
 455 }
 456
 457 /*
 458  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
 459  * is being issued as AIO or not.  This allows us to optimise pure data writes
 460  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
 461  * REQ_FLUSH post write. This is slightly tricky because a single request here
 462  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
 463  * may be pure data writes. In that case, we still need to do a full data sync
 464  * completion.
 465  *
 466  * When page faults are disabled and @dio_flags includes IOMAP_DIO_PARTIAL,
 467  * __iomap_dio_rw can return a partial result if it encounters a non-resident
 468  * page in @iter after preparing a transfer.  In that case, the non-resident
 469  * pages can be faulted in and the request resumed with @done_before set to the
 470  * number of bytes previously transferred.  The request will then complete with
 471  * the correct total number of bytes transferred; this is essential for
 472  * completing partial requests asynchronously.
 473  *
 474  * Returns -ENOTBLK In case of a page invalidation invalidation failure for
 475  * writes.  The callers needs to fall back to buffered I/O in this case.
 476  */
 477 struct iomap_dio *
 478 __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 479                 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
 480                 unsigned int dio_flags, void *private, size_t done_before)
 481 {
 482         struct address_space *mapping = iocb->ki_filp->f_mapping;
 483         struct inode *inode = file_inode(iocb->ki_filp);
 484         struct iomap_iter iomi = {
 485                 .inode          = inode,
 486                 .pos            = iocb->ki_pos,
 487                 .len            = iov_iter_count(iter),
 488                 .flags          = IOMAP_DIRECT,
 489                 .private        = private,
 490         };
 491         loff_t end = iomi.pos + iomi.len - 1, ret = 0;
 492         bool wait_for_completion =
 493                 is_sync_kiocb(iocb) || (dio_flags & IOMAP_DIO_FORCE_WAIT);
 494         struct blk_plug plug;
 495         struct iomap_dio *dio;
 496
 497         trace_iomap_dio_rw_begin(iocb, iter, dio_flags, done_before);
 498
 499         if (!iomi.len)
 500                 return NULL;
 501
 502         dio = kmalloc(sizeof(*dio), GFP_KERNEL);
 503         if (!dio)
 504                 return ERR_PTR(-ENOMEM);
 505
 506         dio->iocb = iocb;
 507         atomic_set(&dio->ref, 1);
 508         dio->size = 0;
 509         dio->i_size = i_size_read(inode);
 510         dio->dops = dops;
 511         dio->error = 0;
 512         dio->flags = 0;
 513         dio->done_before = done_before;
 514
 515         dio->submit.iter = iter;
 516         dio->submit.waiter = current;
 517         dio->submit.poll_bio = NULL;
 518
 519         if (iov_iter_rw(iter) == READ) {
 520                 if (iomi.pos >= dio->i_size)
 521                         goto out_free_dio;
 522
 523                 if (iocb->ki_flags & IOCB_NOWAIT) {
 524                         if (filemap_range_needs_writeback(mapping, iomi.pos,
 525                                         end)) {
 526                                 ret = -EAGAIN;
 527                                 goto out_free_dio;
 528                         }
 529                         iomi.flags |= IOMAP_NOWAIT;
 530                 }
 531
 532                 if (user_backed_iter(iter))
 533                         dio->flags |= IOMAP_DIO_DIRTY;
 534         } else {
 535                 iomi.flags |= IOMAP_WRITE;
 536                 dio->flags |= IOMAP_DIO_WRITE;
 537
 538                 if (iocb->ki_flags & IOCB_NOWAIT) {
 539                         if (filemap_range_has_page(mapping, iomi.pos, end)) {
 540                                 ret = -EAGAIN;
 541                                 goto out_free_dio;
 542                         }
 543                         iomi.flags |= IOMAP_NOWAIT;
 544                 }
 545
 546                 /* for data sync or sync, we need sync completion processing */
 547                 if (iocb_is_dsync(iocb)) {
 548                         dio->flags |= IOMAP_DIO_NEED_SYNC;
 549
 550                        /*
 551                         * For datasync only writes, we optimistically try
 552                         * using FUA for this IO.  Any non-FUA write that
 553                         * occurs will clear this flag, hence we know before
 554                         * completion whether a cache flush is necessary.
 555                         */
 556                         if (!(iocb->ki_flags & IOCB_SYNC))
 557                                 dio->flags |= IOMAP_DIO_WRITE_FUA;
 558                 }
 559         }
 560
 561         if (dio_flags & IOMAP_DIO_OVERWRITE_ONLY) {
 562                 ret = -EAGAIN;
 563                 if (iomi.pos >= dio->i_size ||
 564                     iomi.pos + iomi.len > dio->i_size)
 565                         goto out_free_dio;
 566                 iomi.flags |= IOMAP_OVERWRITE_ONLY;
 567         }
 568
 569         ret = filemap_write_and_wait_range(mapping, iomi.pos, end);
 570         if (ret)
 571                 goto out_free_dio;
 572
 573         if (iov_iter_rw(iter) == WRITE) {
 574                 /*
 575                  * Try to invalidate cache pages for the range we are writing.
 576                  * If this invalidation fails, let the caller fall back to
 577                  * buffered I/O.
 578                  */
 579                 if (invalidate_inode_pages2_range(mapping,
 580                                 iomi.pos >> PAGE_SHIFT, end >> PAGE_SHIFT)) {
 581                         trace_iomap_dio_invalidate_fail(inode, iomi.pos,
 582                                                         iomi.len);
 583                         ret = -ENOTBLK;
 584                         goto out_free_dio;
 585                 }
 586
 587                 if (!wait_for_completion && !inode->i_sb->s_dio_done_wq) {
 588                         ret = sb_init_dio_done_wq(inode->i_sb);
 589                         if (ret < 0)
 590                                 goto out_free_dio;
 591                 }
 592         }
 593
 594         inode_dio_begin(inode);
 595
 596         blk_start_plug(&plug);
 597         while ((ret = iomap_iter(&iomi, ops)) > 0) {
 598                 iomi.processed = iomap_dio_iter(&iomi, dio);
 599
 600                 /*
 601                  * We can only poll for single bio I/Os.
 602                  */
 603                 iocb->ki_flags &= ~IOCB_HIPRI;
 604         }
 605
 606         blk_finish_plug(&plug);
 607
 608         /*
 609          * We only report that we've read data up to i_size.
 610          * Revert iter to a state corresponding to that as some callers (such
 611          * as the splice code) rely on it.
 612          */
 613         if (iov_iter_rw(iter) == READ && iomi.pos >= dio->i_size)
 614                 iov_iter_revert(iter, iomi.pos - dio->i_size);
 615
 616         if (ret == -EFAULT && dio->size && (dio_flags & IOMAP_DIO_PARTIAL)) {
 617                 if (!(iocb->ki_flags & IOCB_NOWAIT))
 618                         wait_for_completion = true;
 619                 ret = 0;
 620         }
 621
 622         /* magic error code to fall back to buffered I/O */
 623         if (ret == -ENOTBLK) {
 624                 wait_for_completion = true;
 625                 ret = 0;
 626         }
 627         if (ret < 0)
 628                 iomap_dio_set_error(dio, ret);
 629
 630         /*
 631          * If all the writes we issued were FUA, we don't need to flush the
 632          * cache on IO completion. Clear the sync flag for this case.
 633          */
 634         if (dio->flags & IOMAP_DIO_WRITE_FUA)
 635                 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
 636
 637         WRITE_ONCE(iocb->private, dio->submit.poll_bio);
 638
 639         /*
 640          * We are about to drop our additional submission reference, which
 641          * might be the last reference to the dio.  There are three different
 642          * ways we can progress here:
 643          *
 644          *  (a) If this is the last reference we will always complete and free
 645          *      the dio ourselves.
 646          *  (b) If this is not the last reference, and we serve an asynchronous
 647          *      iocb, we must never touch the dio after the decrement, the
 648          *      I/O completion handler will complete and free it.
 649          *  (c) If this is not the last reference, but we serve a synchronous
 650          *      iocb, the I/O completion handler will wake us up on the drop
 651          *      of the final reference, and we will complete and free it here
 652          *      after we got woken by the I/O completion handler.
 653          */
 654         dio->wait_for_completion = wait_for_completion;
 655         if (!atomic_dec_and_test(&dio->ref)) {
 656                 if (!wait_for_completion) {
 657                         trace_iomap_dio_rw_queued(inode, iomi.pos, iomi.len);
 658                         return ERR_PTR(-EIOCBQUEUED);
 659                 }
 660
 661                 for (;;) {
 662                         set_current_state(TASK_UNINTERRUPTIBLE);
 663                         if (!READ_ONCE(dio->submit.waiter))
 664                                 break;
 665
 666                         blk_io_schedule();
 667                 }
 668                 __set_current_state(TASK_RUNNING);
 669         }
 670
 671         return dio;
 672
 673 out_free_dio:
 674         kfree(dio);
 675         if (ret)
 676                 return ERR_PTR(ret);
 677         return NULL;
 678 }
 679 EXPORT_SYMBOL_GPL(__iomap_dio_rw);
 680
 681 ssize_t
 682 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 683                 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
 684                 unsigned int dio_flags, void *private, size_t done_before)
 685 {
 686         struct iomap_dio *dio;
 687
 688         dio = __iomap_dio_rw(iocb, iter, ops, dops, dio_flags, private,
 689                              done_before);
 690         if (IS_ERR_OR_NULL(dio))
 691                 return PTR_ERR_OR_ZERO(dio);
 692         return iomap_dio_complete(dio);
 693 }
 694 EXPORT_SYMBOL_GPL(iomap_dio_rw);