1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2010 Red Hat, Inc.
4 * Copyright (c) 2016-2021 Christoph Hellwig.
6 #include <linux/module.h>
7 #include <linux/compiler.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>
17 #include "../internal.h"
20 * Private flags for iomap_dio, must not overlap with the public ones in
23 #define IOMAP_DIO_CALLER_COMP (1U << 26)
24 #define IOMAP_DIO_INLINE_COMP (1U << 27)
25 #define IOMAP_DIO_WRITE_THROUGH (1U << 28)
26 #define IOMAP_DIO_NEED_SYNC (1U << 29)
27 #define IOMAP_DIO_WRITE (1U << 30)
28 #define IOMAP_DIO_DIRTY (1U << 31)
32 const struct iomap_dio_ops *dops;
39 bool wait_for_completion;
42 /* used during submission and for synchronous completion: */
44 struct iov_iter *iter;
45 struct task_struct *waiter;
48 /* used for aio completion: */
50 struct work_struct work;
55 static struct bio *iomap_dio_alloc_bio(const struct iomap_iter *iter,
56 struct iomap_dio *dio, unsigned short nr_vecs, blk_opf_t opf)
58 if (dio->dops && dio->dops->bio_set)
59 return bio_alloc_bioset(iter->iomap.bdev, nr_vecs, opf,
60 GFP_KERNEL, dio->dops->bio_set);
61 return bio_alloc(iter->iomap.bdev, nr_vecs, opf, GFP_KERNEL);
64 static void iomap_dio_submit_bio(const struct iomap_iter *iter,
65 struct iomap_dio *dio, struct bio *bio, loff_t pos)
67 struct kiocb *iocb = dio->iocb;
69 atomic_inc(&dio->ref);
71 /* Sync dio can't be polled reliably */
72 if ((iocb->ki_flags & IOCB_HIPRI) && !is_sync_kiocb(iocb)) {
73 bio_set_polled(bio, iocb);
74 WRITE_ONCE(iocb->private, bio);
77 if (dio->dops && dio->dops->submit_io)
78 dio->dops->submit_io(iter, bio, pos);
83 ssize_t iomap_dio_complete(struct iomap_dio *dio)
85 const struct iomap_dio_ops *dops = dio->dops;
86 struct kiocb *iocb = dio->iocb;
87 loff_t offset = iocb->ki_pos;
88 ssize_t ret = dio->error;
90 if (dops && dops->end_io)
91 ret = dops->end_io(iocb, dio->size, ret, dio->flags);
95 /* check for short read */
96 if (offset + ret > dio->i_size &&
97 !(dio->flags & IOMAP_DIO_WRITE))
98 ret = dio->i_size - offset;
102 * Try again to invalidate clean pages which might have been cached by
103 * non-direct readahead, or faulted in by get_user_pages() if the source
104 * of the write was an mmap'ed region of the file we're writing. Either
105 * one is a pretty crazy thing to do, so we don't support it 100%. If
106 * this invalidation fails, tough, the write still worked...
108 * And this page cache invalidation has to be after ->end_io(), as some
109 * filesystems convert unwritten extents to real allocations in
110 * ->end_io() when necessary, otherwise a racing buffer read would cache
111 * zeros from unwritten extents.
113 if (!dio->error && dio->size && (dio->flags & IOMAP_DIO_WRITE))
114 kiocb_invalidate_post_direct_write(iocb, dio->size);
116 inode_dio_end(file_inode(iocb->ki_filp));
122 * If this is a DSYNC write, make sure we push it to stable
123 * storage now that we've written data.
125 if (dio->flags & IOMAP_DIO_NEED_SYNC)
126 ret = generic_write_sync(iocb, ret);
128 ret += dio->done_before;
130 trace_iomap_dio_complete(iocb, dio->error, ret);
134 EXPORT_SYMBOL_GPL(iomap_dio_complete);
136 static ssize_t iomap_dio_deferred_complete(void *data)
138 return iomap_dio_complete(data);
141 static void iomap_dio_complete_work(struct work_struct *work)
143 struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
144 struct kiocb *iocb = dio->iocb;
146 iocb->ki_complete(iocb, iomap_dio_complete(dio));
150 * Set an error in the dio if none is set yet. We have to use cmpxchg
151 * as the submission context and the completion context(s) can race to
154 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
156 cmpxchg(&dio->error, 0, ret);
159 void iomap_dio_bio_end_io(struct bio *bio)
161 struct iomap_dio *dio = bio->bi_private;
162 bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
163 struct kiocb *iocb = dio->iocb;
166 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
167 if (!atomic_dec_and_test(&dio->ref))
171 * Synchronous dio, task itself will handle any completion work
172 * that needs after IO. All we need to do is wake the task.
174 if (dio->wait_for_completion) {
175 struct task_struct *waiter = dio->submit.waiter;
177 WRITE_ONCE(dio->submit.waiter, NULL);
178 blk_wake_io_task(waiter);
183 * Flagged with IOMAP_DIO_INLINE_COMP, we can complete it inline
185 if (dio->flags & IOMAP_DIO_INLINE_COMP) {
186 WRITE_ONCE(iocb->private, NULL);
187 iomap_dio_complete_work(&dio->aio.work);
192 * If this dio is flagged with IOMAP_DIO_CALLER_COMP, then schedule
193 * our completion that way to avoid an async punt to a workqueue.
195 if (dio->flags & IOMAP_DIO_CALLER_COMP) {
196 /* only polled IO cares about private cleared */
198 iocb->dio_complete = iomap_dio_deferred_complete;
201 * Invoke ->ki_complete() directly. We've assigned our
202 * dio_complete callback handler, and since the issuer set
203 * IOCB_DIO_CALLER_COMP, we know their ki_complete handler will
204 * notice ->dio_complete being set and will defer calling that
205 * handler until it can be done from a safe task context.
207 * Note that the 'res' being passed in here is not important
208 * for this case. The actual completion value of the request
209 * will be gotten from dio_complete when that is run by the
212 iocb->ki_complete(iocb, 0);
217 * Async DIO completion that requires filesystem level completion work
218 * gets punted to a work queue to complete as the operation may require
219 * more IO to be issued to finalise filesystem metadata changes or
220 * guarantee data integrity.
222 INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
223 queue_work(file_inode(iocb->ki_filp)->i_sb->s_dio_done_wq,
227 bio_check_pages_dirty(bio);
229 bio_release_pages(bio, false);
233 EXPORT_SYMBOL_GPL(iomap_dio_bio_end_io);
235 static void iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio,
236 loff_t pos, unsigned len)
238 struct inode *inode = file_inode(dio->iocb->ki_filp);
239 struct page *page = ZERO_PAGE(0);
242 bio = iomap_dio_alloc_bio(iter, dio, 1, REQ_OP_WRITE | REQ_SYNC | REQ_IDLE);
243 fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
245 bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos);
246 bio->bi_private = dio;
247 bio->bi_end_io = iomap_dio_bio_end_io;
249 __bio_add_page(bio, page, len, 0);
250 iomap_dio_submit_bio(iter, dio, bio, pos);
254 * Figure out the bio's operation flags from the dio request, the
255 * mapping, and whether or not we want FUA. Note that we can end up
256 * clearing the WRITE_THROUGH flag in the dio request.
258 static inline blk_opf_t iomap_dio_bio_opflags(struct iomap_dio *dio,
259 const struct iomap *iomap, bool use_fua)
261 blk_opf_t opflags = REQ_SYNC | REQ_IDLE;
263 if (!(dio->flags & IOMAP_DIO_WRITE))
266 opflags |= REQ_OP_WRITE;
270 dio->flags &= ~IOMAP_DIO_WRITE_THROUGH;
275 static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
276 struct iomap_dio *dio)
278 const struct iomap *iomap = &iter->iomap;
279 struct inode *inode = iter->inode;
280 unsigned int fs_block_size = i_blocksize(inode), pad;
281 loff_t length = iomap_length(iter);
282 loff_t pos = iter->pos;
285 bool need_zeroout = false;
286 bool use_fua = false;
287 int nr_pages, ret = 0;
291 if ((pos | length) & (bdev_logical_block_size(iomap->bdev) - 1) ||
292 !bdev_iter_is_aligned(iomap->bdev, dio->submit.iter))
295 if (iomap->type == IOMAP_UNWRITTEN) {
296 dio->flags |= IOMAP_DIO_UNWRITTEN;
300 if (iomap->flags & IOMAP_F_SHARED)
301 dio->flags |= IOMAP_DIO_COW;
303 if (iomap->flags & IOMAP_F_NEW) {
305 } else if (iomap->type == IOMAP_MAPPED) {
307 * Use a FUA write if we need datasync semantics, this is a pure
308 * data IO that doesn't require any metadata updates (including
309 * after IO completion such as unwritten extent conversion) and
310 * the underlying device either supports FUA or doesn't have
311 * a volatile write cache. This allows us to avoid cache flushes
312 * on IO completion. If we can't use writethrough and need to
313 * sync, disable in-task completions as dio completion will
314 * need to call generic_write_sync() which will do a blocking
315 * fsync / cache flush call.
317 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
318 (dio->flags & IOMAP_DIO_WRITE_THROUGH) &&
319 (bdev_fua(iomap->bdev) || !bdev_write_cache(iomap->bdev)))
321 else if (dio->flags & IOMAP_DIO_NEED_SYNC)
322 dio->flags &= ~IOMAP_DIO_CALLER_COMP;
326 * Save the original count and trim the iter to just the extent we
327 * are operating on right now. The iter will be re-expanded once
330 orig_count = iov_iter_count(dio->submit.iter);
331 iov_iter_truncate(dio->submit.iter, length);
333 if (!iov_iter_count(dio->submit.iter))
337 * We can only do deferred completion for pure overwrites that
338 * don't require additional IO at completion. This rules out
339 * writes that need zeroing or extent conversion, extend
340 * the file size, or issue journal IO or cache flushes
341 * during completion processing.
344 ((dio->flags & IOMAP_DIO_NEED_SYNC) && !use_fua) ||
345 ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode)))
346 dio->flags &= ~IOMAP_DIO_CALLER_COMP;
349 * The rules for polled IO completions follow the guidelines as the
350 * ones we set for inline and deferred completions. If none of those
351 * are available for this IO, clear the polled flag.
353 if (!(dio->flags & (IOMAP_DIO_INLINE_COMP|IOMAP_DIO_CALLER_COMP)))
354 dio->iocb->ki_flags &= ~IOCB_HIPRI;
357 /* zero out from the start of the block to the write offset */
358 pad = pos & (fs_block_size - 1);
360 iomap_dio_zero(iter, dio, pos - pad, pad);
364 * Set the operation flags early so that bio_iov_iter_get_pages
365 * can set up the page vector appropriately for a ZONE_APPEND
368 bio_opf = iomap_dio_bio_opflags(dio, iomap, use_fua);
370 nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS);
374 iov_iter_revert(dio->submit.iter, copied);
379 bio = iomap_dio_alloc_bio(iter, dio, nr_pages, bio_opf);
380 fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
382 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
383 bio->bi_ioprio = dio->iocb->ki_ioprio;
384 bio->bi_private = dio;
385 bio->bi_end_io = iomap_dio_bio_end_io;
387 ret = bio_iov_iter_get_pages(bio, dio->submit.iter);
390 * We have to stop part way through an IO. We must fall
391 * through to the sub-block tail zeroing here, otherwise
392 * this short IO may expose stale data in the tail of
393 * the block we haven't written data to.
399 n = bio->bi_iter.bi_size;
400 if (dio->flags & IOMAP_DIO_WRITE) {
401 task_io_account_write(n);
403 if (dio->flags & IOMAP_DIO_DIRTY)
404 bio_set_pages_dirty(bio);
410 nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter,
413 * We can only poll for single bio I/Os.
416 dio->iocb->ki_flags &= ~IOCB_HIPRI;
417 iomap_dio_submit_bio(iter, dio, bio, pos);
422 * We need to zeroout the tail of a sub-block write if the extent type
423 * requires zeroing or the write extends beyond EOF. If we don't zero
424 * the block tail in the latter case, we can expose stale data via mmap
425 * reads of the EOF block.
429 ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
430 /* zero out from the end of the write to the end of the block */
431 pad = pos & (fs_block_size - 1);
433 iomap_dio_zero(iter, dio, pos, fs_block_size - pad);
436 /* Undo iter limitation to current extent */
437 iov_iter_reexpand(dio->submit.iter, orig_count - copied);
443 static loff_t iomap_dio_hole_iter(const struct iomap_iter *iter,
444 struct iomap_dio *dio)
446 loff_t length = iov_iter_zero(iomap_length(iter), dio->submit.iter);
454 static loff_t iomap_dio_inline_iter(const struct iomap_iter *iomi,
455 struct iomap_dio *dio)
457 const struct iomap *iomap = &iomi->iomap;
458 struct iov_iter *iter = dio->submit.iter;
459 void *inline_data = iomap_inline_data(iomap, iomi->pos);
460 loff_t length = iomap_length(iomi);
461 loff_t pos = iomi->pos;
464 if (WARN_ON_ONCE(!iomap_inline_data_valid(iomap)))
467 if (dio->flags & IOMAP_DIO_WRITE) {
468 loff_t size = iomi->inode->i_size;
471 memset(iomap_inline_data(iomap, size), 0, pos - size);
472 copied = copy_from_iter(inline_data, length, iter);
474 if (pos + copied > size)
475 i_size_write(iomi->inode, pos + copied);
476 mark_inode_dirty(iomi->inode);
479 copied = copy_to_iter(inline_data, length, iter);
487 static loff_t iomap_dio_iter(const struct iomap_iter *iter,
488 struct iomap_dio *dio)
490 switch (iter->iomap.type) {
492 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
494 return iomap_dio_hole_iter(iter, dio);
495 case IOMAP_UNWRITTEN:
496 if (!(dio->flags & IOMAP_DIO_WRITE))
497 return iomap_dio_hole_iter(iter, dio);
498 return iomap_dio_bio_iter(iter, dio);
500 return iomap_dio_bio_iter(iter, dio);
502 return iomap_dio_inline_iter(iter, dio);
505 * DIO is not serialised against mmap() access at all, and so
506 * if the page_mkwrite occurs between the writeback and the
507 * iomap_iter() call in the DIO path, then it will see the
508 * DELALLOC block that the page-mkwrite allocated.
510 pr_warn_ratelimited("Direct I/O collision with buffered writes! File: %pD4 Comm: %.20s\n",
511 dio->iocb->ki_filp, current->comm);
520 * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
521 * is being issued as AIO or not. This allows us to optimise pure data writes
522 * to use REQ_FUA rather than requiring generic_write_sync() to issue a
523 * REQ_FLUSH post write. This is slightly tricky because a single request here
524 * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
525 * may be pure data writes. In that case, we still need to do a full data sync
528 * When page faults are disabled and @dio_flags includes IOMAP_DIO_PARTIAL,
529 * __iomap_dio_rw can return a partial result if it encounters a non-resident
530 * page in @iter after preparing a transfer. In that case, the non-resident
531 * pages can be faulted in and the request resumed with @done_before set to the
532 * number of bytes previously transferred. The request will then complete with
533 * the correct total number of bytes transferred; this is essential for
534 * completing partial requests asynchronously.
536 * Returns -ENOTBLK In case of a page invalidation invalidation failure for
537 * writes. The callers needs to fall back to buffered I/O in this case.
540 __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
541 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
542 unsigned int dio_flags, void *private, size_t done_before)
544 struct inode *inode = file_inode(iocb->ki_filp);
545 struct iomap_iter iomi = {
548 .len = iov_iter_count(iter),
549 .flags = IOMAP_DIRECT,
552 bool wait_for_completion =
553 is_sync_kiocb(iocb) || (dio_flags & IOMAP_DIO_FORCE_WAIT);
554 struct blk_plug plug;
555 struct iomap_dio *dio;
558 trace_iomap_dio_rw_begin(iocb, iter, dio_flags, done_before);
563 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
565 return ERR_PTR(-ENOMEM);
568 atomic_set(&dio->ref, 1);
570 dio->i_size = i_size_read(inode);
574 dio->done_before = done_before;
576 dio->submit.iter = iter;
577 dio->submit.waiter = current;
579 if (iocb->ki_flags & IOCB_NOWAIT)
580 iomi.flags |= IOMAP_NOWAIT;
582 if (iov_iter_rw(iter) == READ) {
583 /* reads can always complete inline */
584 dio->flags |= IOMAP_DIO_INLINE_COMP;
586 if (iomi.pos >= dio->i_size)
589 if (user_backed_iter(iter))
590 dio->flags |= IOMAP_DIO_DIRTY;
592 ret = kiocb_write_and_wait(iocb, iomi.len);
596 iomi.flags |= IOMAP_WRITE;
597 dio->flags |= IOMAP_DIO_WRITE;
600 * Flag as supporting deferred completions, if the issuer
601 * groks it. This can avoid a workqueue punt for writes.
602 * We may later clear this flag if we need to do other IO
603 * as part of this IO completion.
605 if (iocb->ki_flags & IOCB_DIO_CALLER_COMP)
606 dio->flags |= IOMAP_DIO_CALLER_COMP;
608 if (dio_flags & IOMAP_DIO_OVERWRITE_ONLY) {
610 if (iomi.pos >= dio->i_size ||
611 iomi.pos + iomi.len > dio->i_size)
613 iomi.flags |= IOMAP_OVERWRITE_ONLY;
616 /* for data sync or sync, we need sync completion processing */
617 if (iocb_is_dsync(iocb)) {
618 dio->flags |= IOMAP_DIO_NEED_SYNC;
621 * For datasync only writes, we optimistically try using
622 * WRITE_THROUGH for this IO. This flag requires either
623 * FUA writes through the device's write cache, or a
624 * normal write to a device without a volatile write
625 * cache. For the former, Any non-FUA write that occurs
626 * will clear this flag, hence we know before completion
627 * whether a cache flush is necessary.
629 if (!(iocb->ki_flags & IOCB_SYNC))
630 dio->flags |= IOMAP_DIO_WRITE_THROUGH;
634 * Try to invalidate cache pages for the range we are writing.
635 * If this invalidation fails, let the caller fall back to
638 ret = kiocb_invalidate_pages(iocb, iomi.len);
640 if (ret != -EAGAIN) {
641 trace_iomap_dio_invalidate_fail(inode, iomi.pos,
648 if (!wait_for_completion && !inode->i_sb->s_dio_done_wq) {
649 ret = sb_init_dio_done_wq(inode->i_sb);
655 inode_dio_begin(inode);
657 blk_start_plug(&plug);
658 while ((ret = iomap_iter(&iomi, ops)) > 0) {
659 iomi.processed = iomap_dio_iter(&iomi, dio);
662 * We can only poll for single bio I/Os.
664 iocb->ki_flags &= ~IOCB_HIPRI;
667 blk_finish_plug(&plug);
670 * We only report that we've read data up to i_size.
671 * Revert iter to a state corresponding to that as some callers (such
672 * as the splice code) rely on it.
674 if (iov_iter_rw(iter) == READ && iomi.pos >= dio->i_size)
675 iov_iter_revert(iter, iomi.pos - dio->i_size);
677 if (ret == -EFAULT && dio->size && (dio_flags & IOMAP_DIO_PARTIAL)) {
678 if (!(iocb->ki_flags & IOCB_NOWAIT))
679 wait_for_completion = true;
683 /* magic error code to fall back to buffered I/O */
684 if (ret == -ENOTBLK) {
685 wait_for_completion = true;
689 iomap_dio_set_error(dio, ret);
692 * If all the writes we issued were already written through to the
693 * media, we don't need to flush the cache on IO completion. Clear the
694 * sync flag for this case.
696 if (dio->flags & IOMAP_DIO_WRITE_THROUGH)
697 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
700 * We are about to drop our additional submission reference, which
701 * might be the last reference to the dio. There are three different
702 * ways we can progress here:
704 * (a) If this is the last reference we will always complete and free
706 * (b) If this is not the last reference, and we serve an asynchronous
707 * iocb, we must never touch the dio after the decrement, the
708 * I/O completion handler will complete and free it.
709 * (c) If this is not the last reference, but we serve a synchronous
710 * iocb, the I/O completion handler will wake us up on the drop
711 * of the final reference, and we will complete and free it here
712 * after we got woken by the I/O completion handler.
714 dio->wait_for_completion = wait_for_completion;
715 if (!atomic_dec_and_test(&dio->ref)) {
716 if (!wait_for_completion) {
717 trace_iomap_dio_rw_queued(inode, iomi.pos, iomi.len);
718 return ERR_PTR(-EIOCBQUEUED);
722 set_current_state(TASK_UNINTERRUPTIBLE);
723 if (!READ_ONCE(dio->submit.waiter))
728 __set_current_state(TASK_RUNNING);
739 EXPORT_SYMBOL_GPL(__iomap_dio_rw);
742 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
743 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
744 unsigned int dio_flags, void *private, size_t done_before)
746 struct iomap_dio *dio;
748 dio = __iomap_dio_rw(iocb, iter, ops, dops, dio_flags, private,
750 if (IS_ERR_OR_NULL(dio))
751 return PTR_ERR_OR_ZERO(dio);
752 return iomap_dio_complete(dio);
754 EXPORT_SYMBOL_GPL(iomap_dio_rw);