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_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)
30 const struct iomap_dio_ops *dops;
37 bool wait_for_completion;
40 /* used during submission and for synchronous completion: */
42 struct iov_iter *iter;
43 struct task_struct *waiter;
47 /* used for aio completion: */
49 struct work_struct work;
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)
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);
63 static void iomap_dio_submit_bio(const struct iomap_iter *iter,
64 struct iomap_dio *dio, struct bio *bio, loff_t pos)
66 atomic_inc(&dio->ref);
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;
74 if (dio->dops && dio->dops->submit_io)
75 dio->dops->submit_io(iter, bio, pos);
80 ssize_t iomap_dio_complete(struct iomap_dio *dio)
82 const struct iomap_dio_ops *dops = dio->dops;
83 struct kiocb *iocb = dio->iocb;
84 loff_t offset = iocb->ki_pos;
85 ssize_t ret = dio->error;
87 if (dops && dops->end_io)
88 ret = dops->end_io(iocb, dio->size, ret, dio->flags);
92 /* check for short read */
93 if (offset + ret > dio->i_size &&
94 !(dio->flags & IOMAP_DIO_WRITE))
95 ret = dio->i_size - offset;
99 * Try again to invalidate clean pages which might have been cached by
100 * non-direct readahead, or faulted in by get_user_pages() if the source
101 * of the write was an mmap'ed region of the file we're writing. Either
102 * one is a pretty crazy thing to do, so we don't support it 100%. If
103 * this invalidation fails, tough, the write still worked...
105 * And this page cache invalidation has to be after ->end_io(), as some
106 * filesystems convert unwritten extents to real allocations in
107 * ->end_io() when necessary, otherwise a racing buffer read would cache
108 * zeros from unwritten extents.
110 if (!dio->error && dio->size && (dio->flags & IOMAP_DIO_WRITE))
111 kiocb_invalidate_post_direct_write(iocb, dio->size);
113 inode_dio_end(file_inode(iocb->ki_filp));
119 * If this is a DSYNC write, make sure we push it to stable
120 * storage now that we've written data.
122 if (dio->flags & IOMAP_DIO_NEED_SYNC)
123 ret = generic_write_sync(iocb, ret);
125 ret += dio->done_before;
127 trace_iomap_dio_complete(iocb, dio->error, ret);
131 EXPORT_SYMBOL_GPL(iomap_dio_complete);
133 static void iomap_dio_complete_work(struct work_struct *work)
135 struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
136 struct kiocb *iocb = dio->iocb;
138 iocb->ki_complete(iocb, iomap_dio_complete(dio));
142 * Set an error in the dio if none is set yet. We have to use cmpxchg
143 * as the submission context and the completion context(s) can race to
146 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
148 cmpxchg(&dio->error, 0, ret);
151 void iomap_dio_bio_end_io(struct bio *bio)
153 struct iomap_dio *dio = bio->bi_private;
154 bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
157 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
159 if (atomic_dec_and_test(&dio->ref)) {
160 if (dio->wait_for_completion) {
161 struct task_struct *waiter = dio->submit.waiter;
162 WRITE_ONCE(dio->submit.waiter, NULL);
163 blk_wake_io_task(waiter);
164 } else if (dio->flags & IOMAP_DIO_WRITE) {
165 struct inode *inode = file_inode(dio->iocb->ki_filp);
167 WRITE_ONCE(dio->iocb->private, NULL);
168 INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
169 queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
171 WRITE_ONCE(dio->iocb->private, NULL);
172 iomap_dio_complete_work(&dio->aio.work);
177 bio_check_pages_dirty(bio);
179 bio_release_pages(bio, false);
183 EXPORT_SYMBOL_GPL(iomap_dio_bio_end_io);
185 static void iomap_dio_zero(const struct iomap_iter *iter, struct iomap_dio *dio,
186 loff_t pos, unsigned len)
188 struct inode *inode = file_inode(dio->iocb->ki_filp);
189 struct page *page = ZERO_PAGE(0);
192 bio = iomap_dio_alloc_bio(iter, dio, 1, REQ_OP_WRITE | REQ_SYNC | REQ_IDLE);
193 fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
195 bio->bi_iter.bi_sector = iomap_sector(&iter->iomap, pos);
196 bio->bi_private = dio;
197 bio->bi_end_io = iomap_dio_bio_end_io;
199 __bio_add_page(bio, page, len, 0);
200 iomap_dio_submit_bio(iter, dio, bio, pos);
204 * Figure out the bio's operation flags from the dio request, the
205 * mapping, and whether or not we want FUA. Note that we can end up
206 * clearing the WRITE_FUA flag in the dio request.
208 static inline blk_opf_t iomap_dio_bio_opflags(struct iomap_dio *dio,
209 const struct iomap *iomap, bool use_fua)
211 blk_opf_t opflags = REQ_SYNC | REQ_IDLE;
213 if (!(dio->flags & IOMAP_DIO_WRITE))
216 opflags |= REQ_OP_WRITE;
220 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
225 static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
226 struct iomap_dio *dio)
228 const struct iomap *iomap = &iter->iomap;
229 struct inode *inode = iter->inode;
230 unsigned int fs_block_size = i_blocksize(inode), pad;
231 loff_t length = iomap_length(iter);
232 loff_t pos = iter->pos;
235 bool need_zeroout = false;
236 bool use_fua = false;
237 int nr_pages, ret = 0;
241 if ((pos | length) & (bdev_logical_block_size(iomap->bdev) - 1) ||
242 !bdev_iter_is_aligned(iomap->bdev, dio->submit.iter))
245 if (iomap->type == IOMAP_UNWRITTEN) {
246 dio->flags |= IOMAP_DIO_UNWRITTEN;
250 if (iomap->flags & IOMAP_F_SHARED)
251 dio->flags |= IOMAP_DIO_COW;
253 if (iomap->flags & IOMAP_F_NEW) {
255 } else if (iomap->type == IOMAP_MAPPED) {
257 * Use a FUA write if we need datasync semantics, this is a pure
258 * data IO that doesn't require any metadata updates (including
259 * after IO completion such as unwritten extent conversion) and
260 * the underlying device supports FUA. This allows us to avoid
261 * cache flushes on IO completion.
263 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
264 (dio->flags & IOMAP_DIO_WRITE_FUA) && bdev_fua(iomap->bdev))
269 * Save the original count and trim the iter to just the extent we
270 * are operating on right now. The iter will be re-expanded once
273 orig_count = iov_iter_count(dio->submit.iter);
274 iov_iter_truncate(dio->submit.iter, length);
276 if (!iov_iter_count(dio->submit.iter))
280 * We can only poll for single bio I/Os.
283 ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode)))
284 dio->iocb->ki_flags &= ~IOCB_HIPRI;
287 /* zero out from the start of the block to the write offset */
288 pad = pos & (fs_block_size - 1);
290 iomap_dio_zero(iter, dio, pos - pad, pad);
294 * Set the operation flags early so that bio_iov_iter_get_pages
295 * can set up the page vector appropriately for a ZONE_APPEND
298 bio_opf = iomap_dio_bio_opflags(dio, iomap, use_fua);
300 nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter, BIO_MAX_VECS);
304 iov_iter_revert(dio->submit.iter, copied);
309 bio = iomap_dio_alloc_bio(iter, dio, nr_pages, bio_opf);
310 fscrypt_set_bio_crypt_ctx(bio, inode, pos >> inode->i_blkbits,
312 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
313 bio->bi_ioprio = dio->iocb->ki_ioprio;
314 bio->bi_private = dio;
315 bio->bi_end_io = iomap_dio_bio_end_io;
317 ret = bio_iov_iter_get_pages(bio, dio->submit.iter);
320 * We have to stop part way through an IO. We must fall
321 * through to the sub-block tail zeroing here, otherwise
322 * this short IO may expose stale data in the tail of
323 * the block we haven't written data to.
329 n = bio->bi_iter.bi_size;
330 if (dio->flags & IOMAP_DIO_WRITE) {
331 task_io_account_write(n);
333 if (dio->flags & IOMAP_DIO_DIRTY)
334 bio_set_pages_dirty(bio);
340 nr_pages = bio_iov_vecs_to_alloc(dio->submit.iter,
343 * We can only poll for single bio I/Os.
346 dio->iocb->ki_flags &= ~IOCB_HIPRI;
347 iomap_dio_submit_bio(iter, dio, bio, pos);
352 * We need to zeroout the tail of a sub-block write if the extent type
353 * requires zeroing or the write extends beyond EOF. If we don't zero
354 * the block tail in the latter case, we can expose stale data via mmap
355 * reads of the EOF block.
359 ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
360 /* zero out from the end of the write to the end of the block */
361 pad = pos & (fs_block_size - 1);
363 iomap_dio_zero(iter, dio, pos, fs_block_size - pad);
366 /* Undo iter limitation to current extent */
367 iov_iter_reexpand(dio->submit.iter, orig_count - copied);
373 static loff_t iomap_dio_hole_iter(const struct iomap_iter *iter,
374 struct iomap_dio *dio)
376 loff_t length = iov_iter_zero(iomap_length(iter), dio->submit.iter);
384 static loff_t iomap_dio_inline_iter(const struct iomap_iter *iomi,
385 struct iomap_dio *dio)
387 const struct iomap *iomap = &iomi->iomap;
388 struct iov_iter *iter = dio->submit.iter;
389 void *inline_data = iomap_inline_data(iomap, iomi->pos);
390 loff_t length = iomap_length(iomi);
391 loff_t pos = iomi->pos;
394 if (WARN_ON_ONCE(!iomap_inline_data_valid(iomap)))
397 if (dio->flags & IOMAP_DIO_WRITE) {
398 loff_t size = iomi->inode->i_size;
401 memset(iomap_inline_data(iomap, size), 0, pos - size);
402 copied = copy_from_iter(inline_data, length, iter);
404 if (pos + copied > size)
405 i_size_write(iomi->inode, pos + copied);
406 mark_inode_dirty(iomi->inode);
409 copied = copy_to_iter(inline_data, length, iter);
417 static loff_t iomap_dio_iter(const struct iomap_iter *iter,
418 struct iomap_dio *dio)
420 switch (iter->iomap.type) {
422 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
424 return iomap_dio_hole_iter(iter, dio);
425 case IOMAP_UNWRITTEN:
426 if (!(dio->flags & IOMAP_DIO_WRITE))
427 return iomap_dio_hole_iter(iter, dio);
428 return iomap_dio_bio_iter(iter, dio);
430 return iomap_dio_bio_iter(iter, dio);
432 return iomap_dio_inline_iter(iter, dio);
435 * DIO is not serialised against mmap() access at all, and so
436 * if the page_mkwrite occurs between the writeback and the
437 * iomap_iter() call in the DIO path, then it will see the
438 * DELALLOC block that the page-mkwrite allocated.
440 pr_warn_ratelimited("Direct I/O collision with buffered writes! File: %pD4 Comm: %.20s\n",
441 dio->iocb->ki_filp, current->comm);
450 * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
451 * is being issued as AIO or not. This allows us to optimise pure data writes
452 * to use REQ_FUA rather than requiring generic_write_sync() to issue a
453 * REQ_FLUSH post write. This is slightly tricky because a single request here
454 * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
455 * may be pure data writes. In that case, we still need to do a full data sync
458 * When page faults are disabled and @dio_flags includes IOMAP_DIO_PARTIAL,
459 * __iomap_dio_rw can return a partial result if it encounters a non-resident
460 * page in @iter after preparing a transfer. In that case, the non-resident
461 * pages can be faulted in and the request resumed with @done_before set to the
462 * number of bytes previously transferred. The request will then complete with
463 * the correct total number of bytes transferred; this is essential for
464 * completing partial requests asynchronously.
466 * Returns -ENOTBLK In case of a page invalidation invalidation failure for
467 * writes. The callers needs to fall back to buffered I/O in this case.
470 __iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
471 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
472 unsigned int dio_flags, void *private, size_t done_before)
474 struct inode *inode = file_inode(iocb->ki_filp);
475 struct iomap_iter iomi = {
478 .len = iov_iter_count(iter),
479 .flags = IOMAP_DIRECT,
482 bool wait_for_completion =
483 is_sync_kiocb(iocb) || (dio_flags & IOMAP_DIO_FORCE_WAIT);
484 struct blk_plug plug;
485 struct iomap_dio *dio;
488 trace_iomap_dio_rw_begin(iocb, iter, dio_flags, done_before);
493 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
495 return ERR_PTR(-ENOMEM);
498 atomic_set(&dio->ref, 1);
500 dio->i_size = i_size_read(inode);
504 dio->done_before = done_before;
506 dio->submit.iter = iter;
507 dio->submit.waiter = current;
508 dio->submit.poll_bio = NULL;
510 if (iocb->ki_flags & IOCB_NOWAIT)
511 iomi.flags |= IOMAP_NOWAIT;
513 if (iov_iter_rw(iter) == READ) {
514 if (iomi.pos >= dio->i_size)
517 if (user_backed_iter(iter))
518 dio->flags |= IOMAP_DIO_DIRTY;
520 ret = kiocb_write_and_wait(iocb, iomi.len);
524 iomi.flags |= IOMAP_WRITE;
525 dio->flags |= IOMAP_DIO_WRITE;
527 if (dio_flags & IOMAP_DIO_OVERWRITE_ONLY) {
529 if (iomi.pos >= dio->i_size ||
530 iomi.pos + iomi.len > dio->i_size)
532 iomi.flags |= IOMAP_OVERWRITE_ONLY;
535 /* for data sync or sync, we need sync completion processing */
536 if (iocb_is_dsync(iocb)) {
537 dio->flags |= IOMAP_DIO_NEED_SYNC;
540 * For datasync only writes, we optimistically try
541 * using FUA for this IO. Any non-FUA write that
542 * occurs will clear this flag, hence we know before
543 * completion whether a cache flush is necessary.
545 if (!(iocb->ki_flags & IOCB_SYNC))
546 dio->flags |= IOMAP_DIO_WRITE_FUA;
550 * Try to invalidate cache pages for the range we are writing.
551 * If this invalidation fails, let the caller fall back to
554 ret = kiocb_invalidate_pages(iocb, iomi.len);
556 if (ret != -EAGAIN) {
557 trace_iomap_dio_invalidate_fail(inode, iomi.pos,
564 if (!wait_for_completion && !inode->i_sb->s_dio_done_wq) {
565 ret = sb_init_dio_done_wq(inode->i_sb);
571 inode_dio_begin(inode);
573 blk_start_plug(&plug);
574 while ((ret = iomap_iter(&iomi, ops)) > 0) {
575 iomi.processed = iomap_dio_iter(&iomi, dio);
578 * We can only poll for single bio I/Os.
580 iocb->ki_flags &= ~IOCB_HIPRI;
583 blk_finish_plug(&plug);
586 * We only report that we've read data up to i_size.
587 * Revert iter to a state corresponding to that as some callers (such
588 * as the splice code) rely on it.
590 if (iov_iter_rw(iter) == READ && iomi.pos >= dio->i_size)
591 iov_iter_revert(iter, iomi.pos - dio->i_size);
593 if (ret == -EFAULT && dio->size && (dio_flags & IOMAP_DIO_PARTIAL)) {
594 if (!(iocb->ki_flags & IOCB_NOWAIT))
595 wait_for_completion = true;
599 /* magic error code to fall back to buffered I/O */
600 if (ret == -ENOTBLK) {
601 wait_for_completion = true;
605 iomap_dio_set_error(dio, ret);
608 * If all the writes we issued were FUA, we don't need to flush the
609 * cache on IO completion. Clear the sync flag for this case.
611 if (dio->flags & IOMAP_DIO_WRITE_FUA)
612 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
614 WRITE_ONCE(iocb->private, dio->submit.poll_bio);
617 * We are about to drop our additional submission reference, which
618 * might be the last reference to the dio. There are three different
619 * ways we can progress here:
621 * (a) If this is the last reference we will always complete and free
623 * (b) If this is not the last reference, and we serve an asynchronous
624 * iocb, we must never touch the dio after the decrement, the
625 * I/O completion handler will complete and free it.
626 * (c) If this is not the last reference, but we serve a synchronous
627 * iocb, the I/O completion handler will wake us up on the drop
628 * of the final reference, and we will complete and free it here
629 * after we got woken by the I/O completion handler.
631 dio->wait_for_completion = wait_for_completion;
632 if (!atomic_dec_and_test(&dio->ref)) {
633 if (!wait_for_completion) {
634 trace_iomap_dio_rw_queued(inode, iomi.pos, iomi.len);
635 return ERR_PTR(-EIOCBQUEUED);
639 set_current_state(TASK_UNINTERRUPTIBLE);
640 if (!READ_ONCE(dio->submit.waiter))
645 __set_current_state(TASK_RUNNING);
656 EXPORT_SYMBOL_GPL(__iomap_dio_rw);
659 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
660 const struct iomap_ops *ops, const struct iomap_dio_ops *dops,
661 unsigned int dio_flags, void *private, size_t done_before)
663 struct iomap_dio *dio;
665 dio = __iomap_dio_rw(iocb, iter, ops, dops, dio_flags, private,
667 if (IS_ERR_OR_NULL(dio))
668 return PTR_ERR_OR_ZERO(dio);
669 return iomap_dio_complete(dio);
671 EXPORT_SYMBOL_GPL(iomap_dio_rw);