fs/iomap/direct-io.c

0001 // SPDX-License-Identifier: GPL-2.0
0002 /*
0003  * Copyright (C) 2010 Red Hat, Inc.
0004  * Copyright (c) 2016-2021 Christoph Hellwig.
0005  */
0006 #include <linux/module.h>
0007 #include <linux/compiler.h>
0008 #include <linux/fs.h>
0009 #include <linux/fscrypt.h>
0010 #include <linux/pagemap.h>
0011 #include <linux/iomap.h>
0012 #include <linux/backing-dev.h>
0013 #include <linux/uio.h>
0014 #include <linux/task_io_accounting_ops.h>
0015 #include "trace.h"
0016
0017 #include "../internal.h"
0018
0019 /*
0020  * Private flags for iomap_dio, must not overlap with the public ones in
0021  * iomap.h:
0022  */
0023 #define IOMAP_DIO_WRITE_FUA (1 << 28)
0024 #define IOMAP_DIO_NEED_SYNC (1 << 29)
0025 #define IOMAP_DIO_WRITE     (1 << 30)
0026 #define IOMAP_DIO_DIRTY     (1 << 31)
0027
0028 struct iomap_dio {
0029     struct kiocb        *iocb;
0030     const struct iomap_dio_ops *dops;
0031     loff_t          i_size;
0032     loff_t          size;
0033     atomic_t        ref;
0034     unsigned        flags;
0035     int         error;
0036     size_t          done_before;
0037     bool            wait_for_completion;
0038
0039     union {
0040         /* used during submission and for synchronous completion: */
0041         struct {
0042             struct iov_iter     *iter;
0043             struct task_struct  *waiter;
0044             struct bio      *poll_bio;
0045         } submit;
0046
0047         /* used for aio completion: */
0048         struct {
0049             struct work_struct  work;
0050         } aio;
0051     };
0052 };
0053
0054 static struct bio *iomap_dio_alloc_bio(const struct iomap_iter *iter,
0055         struct iomap_dio *dio, unsigned short nr_vecs, blk_opf_t opf)
0056 {
0057     if (dio->dops && dio->dops->bio_set)
0058         return bio_alloc_bioset(iter->iomap.bdev, nr_vecs, opf,
0059                     GFP_KERNEL, dio->dops->bio_set);
0060     return bio_alloc(iter->iomap.bdev, nr_vecs, opf, GFP_KERNEL);
0061 }
0062
0063 static void iomap_dio_submit_bio(const struct iomap_iter *iter,
0064         struct iomap_dio *dio, struct bio *bio, loff_t pos)
0065 {
0066     atomic_inc(&dio->ref);
0067
0068     /* Sync dio can't be polled reliably */
0069     if ((dio->iocb->ki_flags & IOCB_HIPRI) && !is_sync_kiocb(dio->iocb)) {
0070         bio_set_polled(bio, dio->iocb);
0071         dio->submit.poll_bio = bio;
0072     }
0073
0074     if (dio->dops && dio->dops->submit_io)
0075         dio->dops->submit_io(iter, bio, pos);
0076     else
0077         submit_bio(bio);
0078 }
0079
0080 ssize_t iomap_dio_complete(struct iomap_dio *dio)
0081 {
0082     const struct iomap_dio_ops *dops = dio->dops;
0083     struct kiocb *iocb = dio->iocb;
0084     struct inode *inode = file_inode(iocb->ki_filp);
0085     loff_t offset = iocb->ki_pos;
0086     ssize_t ret = dio->error;
0087
0088     if (dops && dops->end_io)
0089         ret = dops->end_io(iocb, dio->size, ret, dio->flags);
0090
0091     if (likely(!ret)) {
0092         ret = dio->size;
0093         /* check for short read */
0094         if (offset + ret > dio->i_size &&
0095             !(dio->flags & IOMAP_DIO_WRITE))
0096             ret = dio->i_size - offset;
0097         iocb->ki_pos += ret;
0098     }
0099
0100     /*
0101      * Try again to invalidate clean pages which might have been cached by
0102      * non-direct readahead, or faulted in by get_user_pages() if the source
0103      * of the write was an mmap'ed region of the file we're writing.  Either
0104      * one is a pretty crazy thing to do, so we don't support it 100%.  If
0105      * this invalidation fails, tough, the write still worked...
0106      *
0107      * And this page cache invalidation has to be after ->end_io(), as some
0108      * filesystems convert unwritten extents to real allocations in
0109      * ->end_io() when necessary, otherwise a racing buffer read would cache
0110      * zeros from unwritten extents.
0111      */
0112     if (!dio->error && dio->size &&
0113         (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
0114         int err;
0115         err = invalidate_inode_pages2_range(inode->i_mapping,
0116                 offset >> PAGE_SHIFT,
0117                 (offset + dio->size - 1) >> PAGE_SHIFT);
0118         if (err)
0119             dio_warn_stale_pagecache(iocb->ki_filp);
0120     }
0121
0122     inode_dio_end(file_inode(iocb->ki_filp));
0123     /*
0124      * If this is a DSYNC write, make sure we push it to stable storage now
0125      * that we've written data.
0126      */
0127     if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
0128         ret = generic_write_sync(iocb, ret);
0129
0130     if (ret > 0)
0131         ret += dio->done_before;
0132
0133     kfree(dio);
0134
0135     return ret;
0136 }
0137 EXPORT_SYMBOL_GPL(iomap_dio_complete);
0138
0139 static void iomap_dio_complete_work(struct work_struct *work)
0140 {
0141     struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
0142     struct kiocb *iocb = dio->iocb;
0143
0144     iocb->ki_complete(iocb, iomap_dio_complete(dio));
0145 }
0146
0147 /*
0148  * Set an error in the dio if none is set yet.  We have to use cmpxchg
0149  * as the submission context and the completion context(s) can race to
0150  * update the error.
0151  */
0152 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
0153 {
0154     cmpxchg(&dio->error, 0, ret);
0155 }
0156
0157 void iomap_dio_bio_end_io(struct bio *bio)
0158 {
0159     struct iomap_dio *dio = bio->bi_private;
0160     bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
0161
0162     if (bio->bi_status)
0163         iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
0164
0165     if (atomic_dec_and_test(&dio->ref)) {
0166         if (dio->wait_for_completion) {
0167             struct task_struct *waiter = dio->submit.waiter;
0168             WRITE_ONCE(dio->submit.waiter, NULL);
0169             blk_wake_io_task(waiter);
0170         } else if (dio->flags & IOMAP_DIO_WRITE) {
0171             struct inode *inode = file_inode(dio->iocb->ki_filp);
0172
0173             WRITE_ONCE(dio->iocb->private, NULL);
0174             INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
0175             queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
0176         } else {
0177             WRITE_ONCE(dio->iocb->private, NULL);
0178             iomap_dio_complete_work(&dio->aio.work);
0179         }
0180     }
0181
0182     if (should_dirty) {
0183         bio_check_pages_dirty(bio);
0184     } else {
0185         bio_release_pages(bio, false);
0186         bio_put(bio);
0187     }
0188 }
0189 EXPORT_SYMBOL_GPL(iomap_dio_bio_end_io);
0190
0191 static void iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio,
0192         loff_t pos, unsigned len)
0193 {
0194     struct inode *inode = file_inode(dio->iocb->ki_filp);
0195     struct page *page = ZERO_PAGE(0);
0196     struct bio *bio;
0197
0198     bio = iomap_dio_alloc_bio(iter, dio, 1, REQ_OP_WRITE | REQ_SYNC | REQ_IDLE);
0199     fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
0200                   GFP_KERNEL);
0201     bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos);
0202     bio->bi_private = dio;
0203     bio->bi_end_io = iomap_dio_bio_end_io;
0204
0205     get_page(page);
0206     __bio_add_page(bio, page, len, 0);
0207     iomap_dio_submit_bio(iter, dio, bio, pos);
0208 }
0209
0210 /*
0211  * Figure out the bio's operation flags from the dio request, the
0212  * mapping, and whether or not we want FUA.  Note that we can end up
0213  * clearing the WRITE_FUA flag in the dio request.
0214  */
0215 static inline blk_opf_t iomap_dio_bio_opflags(struct iomap_dio *dio,
0216         const struct iomap *iomap, bool use_fua)
0217 {
0218     blk_opf_t opflags = REQ_SYNC | REQ_IDLE;
0219
0220     if (!(dio->flags & IOMAP_DIO_WRITE)) {
0221         WARN_ON_ONCE(iomap->flags & IOMAP_F_ZONE_APPEND);
0222         return REQ_OP_READ;
0223     }
0224
0225     if (iomap->flags & IOMAP_F_ZONE_APPEND)
0226         opflags |= REQ_OP_ZONE_APPEND;
0227     else
0228         opflags |= REQ_OP_WRITE;
0229
0230     if (use_fua)
0231         opflags |= REQ_FUA;
0232     else
0233         dio->flags &= ~IOMAP_DIO_WRITE_FUA;
0234
0235     return opflags;
0236 }
0237
0238 static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
0239         struct iomap_dio *dio)
0240 {
0241     const struct iomap *iomap = &iter->iomap;
0242     struct inode *inode = iter->inode;
0243     unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
0244     unsigned int fs_block_size = i_blocksize(inode), pad;
0245     loff_t length = iomap_length(iter);
0246     loff_t pos = iter->pos;
0247     blk_opf_t bio_opf;
0248     struct bio *bio;
0249     bool need_zeroout = false;
0250     bool use_fua = false;
0251     int nr_pages, ret = 0;
0252     size_t copied = 0;
0253     size_t orig_count;
0254
0255     if ((pos | length) & ((1 << blkbits) - 1) ||
0256         !bdev_iter_is_aligned(iomap->bdev, dio->submit.iter))
0257         return -EINVAL;
0258
0259     if (iomap->type == IOMAP_UNWRITTEN) {
0260         dio->flags |= IOMAP_DIO_UNWRITTEN;
0261         need_zeroout = true;
0262     }
0263
0264     if (iomap->flags & IOMAP_F_SHARED)
0265         dio->flags |= IOMAP_DIO_COW;
0266
0267     if (iomap->flags & IOMAP_F_NEW) {
0268         need_zeroout = true;
0269     } else if (iomap->type == IOMAP_MAPPED) {
0270         /*
0271          * Use a FUA write if we need datasync semantics, this is a pure
0272          * data IO that doesn't require any metadata updates (including
0273          * after IO completion such as unwritten extent conversion) and
0274          * the underlying device supports FUA. This allows us to avoid
0275          * cache flushes on IO completion.
0276          */
0277         if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
0278             (dio->flags & IOMAP_DIO_WRITE_FUA) && bdev_fua(iomap->bdev))
0279             use_fua = true;
0280     }
0281
0282     /*
0283      * Save the original count and trim the iter to just the extent we
0284      * are operating on right now.  The iter will be re-expanded once
0285      * we are done.
0286      */
0287     orig_count = iov_iter_count(dio->submit.iter);
0288     iov_iter_truncate(dio->submit.iter, length);
0289
0290     if (!iov_iter_count(dio->submit.iter))
0291         goto out;
0292
0293     /*
0294      * We can only poll for single bio I/Os.
0295      */
0296     if (need_zeroout ||
0297         ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode)))
0298         dio->iocb->ki_flags &= ~IOCB_HIPRI;
0299
0300     if (need_zeroout) {
0301         /* zero out from the start of the block to the write offset */
0302         pad = pos & (fs_block_size - 1);
0303         if (pad)
0304             iomap_dio_zero(iter, dio, pos - pad, pad);
0305     }
0306
0307     /*
0308      * Set the operation flags early so that bio_iov_iter_get_pages
0309      * can set up the page vector appropriately for a ZONE_APPEND
0310      * operation.
0311      */
0312     bio_opf = iomap_dio_bio_opflags(dio, iomap, use_fua);
0313
0314     nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS);
0315     do {
0316         size_t n;
0317         if (dio->error) {
0318             iov_iter_revert(dio->submit.iter, copied);
0319             copied = ret = 0;
0320             goto out;
0321         }
0322
0323         bio = iomap_dio_alloc_bio(iter, dio, nr_pages, bio_opf);
0324         fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
0325                       GFP_KERNEL);
0326         bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
0327         bio->bi_ioprio = dio->iocb->ki_ioprio;
0328         bio->bi_private = dio;
0329         bio->bi_end_io = iomap_dio_bio_end_io;
0330
0331         ret = bio_iov_iter_get_pages(bio, dio->submit.iter);
0332         if (unlikely(ret)) {
0333             /*
0334              * We have to stop part way through an IO. We must fall
0335              * through to the sub-block tail zeroing here, otherwise
0336              * this short IO may expose stale data in the tail of
0337              * the block we haven't written data to.
0338              */
0339             bio_put(bio);
0340             goto zero_tail;
0341         }
0342
0343         n = bio->bi_iter.bi_size;
0344         if (dio->flags & IOMAP_DIO_WRITE) {
0345             task_io_account_write(n);
0346         } else {
0347             if (dio->flags & IOMAP_DIO_DIRTY)
0348                 bio_set_pages_dirty(bio);
0349         }
0350
0351         dio->size += n;
0352         copied += n;
0353
0354         nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter,
0355                          BIO_MAX_VECS);
0356         /*
0357          * We can only poll for single bio I/Os.
0358          */
0359         if (nr_pages)
0360             dio->iocb->ki_flags &= ~IOCB_HIPRI;
0361         iomap_dio_submit_bio(iter, dio, bio, pos);
0362         pos += n;
0363     } while (nr_pages);
0364
0365     /*
0366      * We need to zeroout the tail of a sub-block write if the extent type
0367      * requires zeroing or the write extends beyond EOF. If we don't zero
0368      * the block tail in the latter case, we can expose stale data via mmap
0369      * reads of the EOF block.
0370      */
0371 zero_tail:
0372     if (need_zeroout ||
0373         ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
0374         /* zero out from the end of the write to the end of the block */
0375         pad = pos & (fs_block_size - 1);
0376         if (pad)
0377             iomap_dio_zero(iter, dio, pos, fs_block_size - pad);
0378     }
0379 out:
0380     /* Undo iter limitation to current extent */
0381     iov_iter_reexpand(dio->submit.iter, orig_count - copied);
0382     if (copied)
0383         return copied;
0384     return ret;
0385 }
0386
0387 static loff_t iomap_dio_hole_iter(const struct iomap_iter *iter,
0388         struct iomap_dio *dio)
0389 {
0390     loff_t length = iov_iter_zero(iomap_length(iter), dio->submit.iter);
0391
0392     dio->size += length;
0393     if (!length)
0394         return -EFAULT;
0395     return length;
0396 }
0397
0398 static loff_t iomap_dio_inline_iter(const struct iomap_iter *iomi,
0399         struct iomap_dio *dio)
0400 {
0401     const struct iomap *iomap = &iomi->iomap;
0402     struct iov_iter *iter = dio->submit.iter;
0403     void *inline_data = iomap_inline_data(iomap, iomi->pos);
0404     loff_t length = iomap_length(iomi);
0405     loff_t pos = iomi->pos;
0406     size_t copied;
0407
0408     if (WARN_ON_ONCE(!iomap_inline_data_valid(iomap)))
0409         return -EIO;
0410
0411     if (dio->flags & IOMAP_DIO_WRITE) {
0412         loff_t size = iomi->inode->i_size;
0413
0414         if (pos > size)
0415             memset(iomap_inline_data(iomap, size), 0, pos - size);
0416         copied = copy_from_iter(inline_data, length, iter);
0417         if (copied) {
0418             if (pos + copied > size)
0419                 i_size_write(iomi->inode, pos + copied);
0420             mark_inode_dirty(iomi->inode);
0421         }
0422     } else {
0423         copied = copy_to_iter(inline_data, length, iter);
0424     }
0425     dio->size += copied;
0426     if (!copied)
0427         return -EFAULT;
0428     return copied;
0429 }
0430
0431 static loff_t iomap_dio_iter(const struct iomap_iter *iter,
0432         struct iomap_dio *dio)
0433 {
0434     switch (iter->iomap.type) {
0435     case IOMAP_HOLE:
0436         if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
0437             return -EIO;
0438         return iomap_dio_hole_iter(iter, dio);
0439     case IOMAP_UNWRITTEN:
0440         if (!(dio->flags & IOMAP_DIO_WRITE))
0441             return iomap_dio_hole_iter(iter, dio);
0442         return iomap_dio_bio_iter(iter, dio);
0443     case IOMAP_MAPPED:
0444         return iomap_dio_bio_iter(iter, dio);
0445     case IOMAP_INLINE:
0446         return iomap_dio_inline_iter(iter, dio);
0447     case IOMAP_DELALLOC:
0448         /*
0449          * DIO is not serialised against mmap() access at all, and so
0450          * if the page_mkwrite occurs between the writeback and the
0451          * iomap_iter() call in the DIO path, then it will see the
0452          * DELALLOC block that the page-mkwrite allocated.
0453          */
0454         pr_warn_ratelimited("Direct I/O collision with buffered writes! File: %pD4 Comm: %.20s\n",
0455                     dio->iocb->ki_filp, current->comm);
0456         return -EIO;
0457     default:
0458         WARN_ON_ONCE(1);
0459         return -EIO;
0460     }
0461 }
0462
0463 /*
0464  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
0465  * is being issued as AIO or not.  This allows us to optimise pure data writes
0466  * to use REQ_FUA rather than requiring generic_write_sync() to issue a
0467  * REQ_FLUSH post write. This is slightly tricky because a single request here
0468  * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
0469  * may be pure data writes. In that case, we still need to do a full data sync
0470  * completion.
0471  *
0472  * When page faults are disabled and @dio_flags includes IOMAP_DIO_PARTIAL,
0473  * __iomap_dio_rw can return a partial result if it encounters a non-resident
0474  * page in @iter after preparing a transfer.  In that case, the non-resident
0475  * pages can be faulted in and the request resumed with @done_before set to the
0476  * number of bytes previously transferred.  The request will then complete with
0477  * the correct total number of bytes transferred; this is essential for
0478  * completing partial requests asynchronously.
0479  *
0480  * Returns -ENOTBLK In case of a page invalidation invalidation failure for
0481  * writes.  The callers needs to fall back to buffered I/O in this case.
0482  */
0483 struct iomap_dio *
0484 __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
0485         const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
0486         unsigned int dio_flags, void *private, size_t done_before)
0487 {
0488     struct address_space *mapping = iocb->ki_filp->f_mapping;
0489     struct inode *inode = file_inode(iocb->ki_filp);
0490     struct iomap_iter iomi = {
0491         .inode      = inode,
0492         .pos        = iocb->ki_pos,
0493         .len        = iov_iter_count(iter),
0494         .flags      = IOMAP_DIRECT,
0495         .private    = private,
0496     };
0497     loff_t end = iomi.pos + iomi.len - 1, ret = 0;
0498     bool wait_for_completion =
0499         is_sync_kiocb(iocb) || (dio_flags & IOMAP_DIO_FORCE_WAIT);
0500     struct blk_plug plug;
0501     struct iomap_dio *dio;
0502
0503     if (!iomi.len)
0504         return NULL;
0505
0506     dio = kmalloc(sizeof(*dio), GFP_KERNEL);
0507     if (!dio)
0508         return ERR_PTR(-ENOMEM);
0509
0510     dio->iocb = iocb;
0511     atomic_set(&dio->ref, 1);
0512     dio->size = 0;
0513     dio->i_size = i_size_read(inode);
0514     dio->dops = dops;
0515     dio->error = 0;
0516     dio->flags = 0;
0517     dio->done_before = done_before;
0518
0519     dio->submit.iter = iter;
0520     dio->submit.waiter = current;
0521     dio->submit.poll_bio = NULL;
0522
0523     if (iov_iter_rw(iter) == READ) {
0524         if (iomi.pos >= dio->i_size)
0525             goto out_free_dio;
0526
0527         if (iocb->ki_flags & IOCB_NOWAIT) {
0528             if (filemap_range_needs_writeback(mapping, iomi.pos,
0529                     end)) {
0530                 ret = -EAGAIN;
0531                 goto out_free_dio;
0532             }
0533             iomi.flags |= IOMAP_NOWAIT;
0534         }
0535
0536         if (user_backed_iter(iter))
0537             dio->flags |= IOMAP_DIO_DIRTY;
0538     } else {
0539         iomi.flags |= IOMAP_WRITE;
0540         dio->flags |= IOMAP_DIO_WRITE;
0541
0542         if (iocb->ki_flags & IOCB_NOWAIT) {
0543             if (filemap_range_has_page(mapping, iomi.pos, end)) {
0544                 ret = -EAGAIN;
0545                 goto out_free_dio;
0546             }
0547             iomi.flags |= IOMAP_NOWAIT;
0548         }
0549
0550         /* for data sync or sync, we need sync completion processing */
0551         if (iocb_is_dsync(iocb) && !(dio_flags & IOMAP_DIO_NOSYNC)) {
0552             dio->flags |= IOMAP_DIO_NEED_SYNC;
0553
0554                /*
0555             * For datasync only writes, we optimistically try
0556             * using FUA for this IO.  Any non-FUA write that
0557             * occurs will clear this flag, hence we know before
0558             * completion whether a cache flush is necessary.
0559             */
0560             if (!(iocb->ki_flags & IOCB_SYNC))
0561                 dio->flags |= IOMAP_DIO_WRITE_FUA;
0562         }
0563     }
0564
0565     if (dio_flags & IOMAP_DIO_OVERWRITE_ONLY) {
0566         ret = -EAGAIN;
0567         if (iomi.pos >= dio->i_size ||
0568             iomi.pos + iomi.len > dio->i_size)
0569             goto out_free_dio;
0570         iomi.flags |= IOMAP_OVERWRITE_ONLY;
0571     }
0572
0573     ret = filemap_write_and_wait_range(mapping, iomi.pos, end);
0574     if (ret)
0575         goto out_free_dio;
0576
0577     if (iov_iter_rw(iter) == WRITE) {
0578         /*
0579          * Try to invalidate cache pages for the range we are writing.
0580          * If this invalidation fails, let the caller fall back to
0581          * buffered I/O.
0582          */
0583         if (invalidate_inode_pages2_range(mapping,
0584                 iomi.pos >> PAGE_SHIFT, end >> PAGE_SHIFT)) {
0585             trace_iomap_dio_invalidate_fail(inode, iomi.pos,
0586                             iomi.len);
0587             ret = -ENOTBLK;
0588             goto out_free_dio;
0589         }
0590
0591         if (!wait_for_completion && !inode->i_sb->s_dio_done_wq) {
0592             ret = sb_init_dio_done_wq(inode->i_sb);
0593             if (ret < 0)
0594                 goto out_free_dio;
0595         }
0596     }
0597
0598     inode_dio_begin(inode);
0599
0600     blk_start_plug(&plug);
0601     while ((ret = iomap_iter(&iomi, ops)) > 0) {
0602         iomi.processed = iomap_dio_iter(&iomi, dio);
0603
0604         /*
0605          * We can only poll for single bio I/Os.
0606          */
0607         iocb->ki_flags &= ~IOCB_HIPRI;
0608     }
0609
0610     blk_finish_plug(&plug);
0611
0612     /*
0613      * We only report that we've read data up to i_size.
0614      * Revert iter to a state corresponding to that as some callers (such
0615      * as the splice code) rely on it.
0616      */
0617     if (iov_iter_rw(iter) == READ && iomi.pos >= dio->i_size)
0618         iov_iter_revert(iter, iomi.pos - dio->i_size);
0619
0620     if (ret == -EFAULT && dio->size && (dio_flags & IOMAP_DIO_PARTIAL)) {
0621         if (!(iocb->ki_flags & IOCB_NOWAIT))
0622             wait_for_completion = true;
0623         ret = 0;
0624     }
0625
0626     /* magic error code to fall back to buffered I/O */
0627     if (ret == -ENOTBLK) {
0628         wait_for_completion = true;
0629         ret = 0;
0630     }
0631     if (ret < 0)
0632         iomap_dio_set_error(dio, ret);
0633
0634     /*
0635      * If all the writes we issued were FUA, we don't need to flush the
0636      * cache on IO completion. Clear the sync flag for this case.
0637      */
0638     if (dio->flags & IOMAP_DIO_WRITE_FUA)
0639         dio->flags &= ~IOMAP_DIO_NEED_SYNC;
0640
0641     WRITE_ONCE(iocb->private, dio->submit.poll_bio);
0642
0643     /*
0644      * We are about to drop our additional submission reference, which
0645      * might be the last reference to the dio.  There are three different
0646      * ways we can progress here:
0647      *
0648      *  (a) If this is the last reference we will always complete and free
0649      *  the dio ourselves.
0650      *  (b) If this is not the last reference, and we serve an asynchronous
0651      *  iocb, we must never touch the dio after the decrement, the
0652      *  I/O completion handler will complete and free it.
0653      *  (c) If this is not the last reference, but we serve a synchronous
0654      *  iocb, the I/O completion handler will wake us up on the drop
0655      *  of the final reference, and we will complete and free it here
0656      *  after we got woken by the I/O completion handler.
0657      */
0658     dio->wait_for_completion = wait_for_completion;
0659     if (!atomic_dec_and_test(&dio->ref)) {
0660         if (!wait_for_completion)
0661             return ERR_PTR(-EIOCBQUEUED);
0662
0663         for (;;) {
0664             set_current_state(TASK_UNINTERRUPTIBLE);
0665             if (!READ_ONCE(dio->submit.waiter))
0666                 break;
0667
0668             blk_io_schedule();
0669         }
0670         __set_current_state(TASK_RUNNING);
0671     }
0672
0673     return dio;
0674
0675 out_free_dio:
0676     kfree(dio);
0677     if (ret)
0678         return ERR_PTR(ret);
0679     return NULL;
0680 }
0681 EXPORT_SYMBOL_GPL(__iomap_dio_rw);
0682
0683 ssize_t
0684 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
0685         const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
0686         unsigned int dio_flags, void *private, size_t done_before)
0687 {
0688     struct iomap_dio *dio;
0689
0690     dio = __iomap_dio_rw(iocb, iter, ops, dops, dio_flags, private,
0691                  done_before);
0692     if (IS_ERR_OR_NULL(dio))
0693         return PTR_ERR_OR_ZERO(dio);
0694     return iomap_dio_complete(dio);
0695 }
0696 EXPORT_SYMBOL_GPL(iomap_dio_rw);