1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2010 Red Hat, Inc.
4 * Copyright (c) 2016-2018 Christoph Hellwig.
6 #include <linux/module.h>
7 #include <linux/compiler.h>
9 #include <linux/iomap.h>
10 #include <linux/backing-dev.h>
11 #include <linux/uio.h>
12 #include <linux/task_io_accounting_ops.h>
14 #include "../internal.h"
17 * Private flags for iomap_dio, must not overlap with the public ones in
20 #define IOMAP_DIO_WRITE_FUA (1 << 28)
21 #define IOMAP_DIO_NEED_SYNC (1 << 29)
22 #define IOMAP_DIO_WRITE (1 << 30)
23 #define IOMAP_DIO_DIRTY (1 << 31)
27 const struct iomap_dio_ops *dops;
33 bool wait_for_completion;
36 /* used during submission and for synchronous completion: */
38 struct iov_iter *iter;
39 struct task_struct *waiter;
40 struct request_queue *last_queue;
44 /* used for aio completion: */
46 struct work_struct work;
51 int iomap_dio_iopoll(struct kiocb *kiocb, bool spin)
53 struct request_queue *q = READ_ONCE(kiocb->private);
57 return blk_poll(q, READ_ONCE(kiocb->ki_cookie), spin);
59 EXPORT_SYMBOL_GPL(iomap_dio_iopoll);
61 static void iomap_dio_submit_bio(struct iomap_dio *dio, struct iomap *iomap,
64 atomic_inc(&dio->ref);
66 if (dio->iocb->ki_flags & IOCB_HIPRI)
67 bio_set_polled(bio, dio->iocb);
69 dio->submit.last_queue = bdev_get_queue(iomap->bdev);
70 dio->submit.cookie = submit_bio(bio);
73 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
75 const struct iomap_dio_ops *dops = dio->dops;
76 struct kiocb *iocb = dio->iocb;
77 struct inode *inode = file_inode(iocb->ki_filp);
78 loff_t offset = iocb->ki_pos;
79 ssize_t ret = dio->error;
81 if (dops && dops->end_io)
82 ret = dops->end_io(iocb, dio->size, ret, dio->flags);
86 /* check for short read */
87 if (offset + ret > dio->i_size &&
88 !(dio->flags & IOMAP_DIO_WRITE))
89 ret = dio->i_size - offset;
94 * Try again to invalidate clean pages which might have been cached by
95 * non-direct readahead, or faulted in by get_user_pages() if the source
96 * of the write was an mmap'ed region of the file we're writing. Either
97 * one is a pretty crazy thing to do, so we don't support it 100%. If
98 * this invalidation fails, tough, the write still worked...
100 * And this page cache invalidation has to be after ->end_io(), as some
101 * filesystems convert unwritten extents to real allocations in
102 * ->end_io() when necessary, otherwise a racing buffer read would cache
103 * zeros from unwritten extents.
106 (dio->flags & IOMAP_DIO_WRITE) && inode->i_mapping->nrpages) {
108 err = invalidate_inode_pages2_range(inode->i_mapping,
109 offset >> PAGE_SHIFT,
110 (offset + dio->size - 1) >> PAGE_SHIFT);
112 dio_warn_stale_pagecache(iocb->ki_filp);
116 * If this is a DSYNC write, make sure we push it to stable storage now
117 * that we've written data.
119 if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
120 ret = generic_write_sync(iocb, ret);
122 inode_dio_end(file_inode(iocb->ki_filp));
128 static void iomap_dio_complete_work(struct work_struct *work)
130 struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
131 struct kiocb *iocb = dio->iocb;
133 iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
137 * Set an error in the dio if none is set yet. We have to use cmpxchg
138 * as the submission context and the completion context(s) can race to
141 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
143 cmpxchg(&dio->error, 0, ret);
146 static void iomap_dio_bio_end_io(struct bio *bio)
148 struct iomap_dio *dio = bio->bi_private;
149 bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
152 iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
154 if (atomic_dec_and_test(&dio->ref)) {
155 if (dio->wait_for_completion) {
156 struct task_struct *waiter = dio->submit.waiter;
157 WRITE_ONCE(dio->submit.waiter, NULL);
158 blk_wake_io_task(waiter);
159 } else if (dio->flags & IOMAP_DIO_WRITE) {
160 struct inode *inode = file_inode(dio->iocb->ki_filp);
162 INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
163 queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
165 iomap_dio_complete_work(&dio->aio.work);
170 bio_check_pages_dirty(bio);
172 bio_release_pages(bio, false);
178 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
181 struct page *page = ZERO_PAGE(0);
182 int flags = REQ_SYNC | REQ_IDLE;
185 bio = bio_alloc(GFP_KERNEL, 1);
186 bio_set_dev(bio, iomap->bdev);
187 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
188 bio->bi_private = dio;
189 bio->bi_end_io = iomap_dio_bio_end_io;
192 __bio_add_page(bio, page, len, 0);
193 bio_set_op_attrs(bio, REQ_OP_WRITE, flags);
194 iomap_dio_submit_bio(dio, iomap, bio);
198 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
199 struct iomap_dio *dio, struct iomap *iomap)
201 unsigned int blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
202 unsigned int fs_block_size = i_blocksize(inode), pad;
203 unsigned int align = iov_iter_alignment(dio->submit.iter);
204 struct iov_iter iter;
206 bool need_zeroout = false;
207 bool use_fua = false;
208 int nr_pages, ret = 0;
211 if ((pos | length | align) & ((1 << blkbits) - 1))
214 if (iomap->type == IOMAP_UNWRITTEN) {
215 dio->flags |= IOMAP_DIO_UNWRITTEN;
219 if (iomap->flags & IOMAP_F_SHARED)
220 dio->flags |= IOMAP_DIO_COW;
222 if (iomap->flags & IOMAP_F_NEW) {
224 } else if (iomap->type == IOMAP_MAPPED) {
226 * Use a FUA write if we need datasync semantics, this is a pure
227 * data IO that doesn't require any metadata updates (including
228 * after IO completion such as unwritten extent conversion) and
229 * the underlying device supports FUA. This allows us to avoid
230 * cache flushes on IO completion.
232 if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
233 (dio->flags & IOMAP_DIO_WRITE_FUA) &&
234 blk_queue_fua(bdev_get_queue(iomap->bdev)))
239 * Operate on a partial iter trimmed to the extent we were called for.
240 * We'll update the iter in the dio once we're done with this extent.
242 iter = *dio->submit.iter;
243 iov_iter_truncate(&iter, length);
245 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
250 /* zero out from the start of the block to the write offset */
251 pad = pos & (fs_block_size - 1);
253 iomap_dio_zero(dio, iomap, pos - pad, pad);
259 iov_iter_revert(dio->submit.iter, copied);
263 bio = bio_alloc(GFP_KERNEL, nr_pages);
264 bio_set_dev(bio, iomap->bdev);
265 bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
266 bio->bi_write_hint = dio->iocb->ki_hint;
267 bio->bi_ioprio = dio->iocb->ki_ioprio;
268 bio->bi_private = dio;
269 bio->bi_end_io = iomap_dio_bio_end_io;
271 ret = bio_iov_iter_get_pages(bio, &iter);
274 * We have to stop part way through an IO. We must fall
275 * through to the sub-block tail zeroing here, otherwise
276 * this short IO may expose stale data in the tail of
277 * the block we haven't written data to.
283 n = bio->bi_iter.bi_size;
284 if (dio->flags & IOMAP_DIO_WRITE) {
285 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
287 bio->bi_opf |= REQ_FUA;
289 dio->flags &= ~IOMAP_DIO_WRITE_FUA;
290 task_io_account_write(n);
292 bio->bi_opf = REQ_OP_READ;
293 if (dio->flags & IOMAP_DIO_DIRTY)
294 bio_set_pages_dirty(bio);
297 iov_iter_advance(dio->submit.iter, n);
303 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
304 iomap_dio_submit_bio(dio, iomap, bio);
308 * We need to zeroout the tail of a sub-block write if the extent type
309 * requires zeroing or the write extends beyond EOF. If we don't zero
310 * the block tail in the latter case, we can expose stale data via mmap
311 * reads of the EOF block.
315 ((dio->flags & IOMAP_DIO_WRITE) && pos >= i_size_read(inode))) {
316 /* zero out from the end of the write to the end of the block */
317 pad = pos & (fs_block_size - 1);
319 iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
327 iomap_dio_hole_actor(loff_t length, struct iomap_dio *dio)
329 length = iov_iter_zero(length, dio->submit.iter);
335 iomap_dio_inline_actor(struct inode *inode, loff_t pos, loff_t length,
336 struct iomap_dio *dio, struct iomap *iomap)
338 struct iov_iter *iter = dio->submit.iter;
341 BUG_ON(pos + length > PAGE_SIZE - offset_in_page(iomap->inline_data));
343 if (dio->flags & IOMAP_DIO_WRITE) {
344 loff_t size = inode->i_size;
347 memset(iomap->inline_data + size, 0, pos - size);
348 copied = copy_from_iter(iomap->inline_data + pos, length, iter);
350 if (pos + copied > size)
351 i_size_write(inode, pos + copied);
352 mark_inode_dirty(inode);
355 copied = copy_to_iter(iomap->inline_data + pos, length, iter);
362 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
363 void *data, struct iomap *iomap)
365 struct iomap_dio *dio = data;
367 switch (iomap->type) {
369 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
371 return iomap_dio_hole_actor(length, dio);
372 case IOMAP_UNWRITTEN:
373 if (!(dio->flags & IOMAP_DIO_WRITE))
374 return iomap_dio_hole_actor(length, dio);
375 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
377 return iomap_dio_bio_actor(inode, pos, length, dio, iomap);
379 return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
382 * DIO is not serialised against mmap() access at all, and so
383 * if the page_mkwrite occurs between the writeback and the
384 * iomap_apply() call in the DIO path, then it will see the
385 * DELALLOC block that the page-mkwrite allocated.
387 pr_warn_ratelimited("Direct I/O collision with buffered writes! File: %pD4 Comm: %.20s\n",
388 dio->iocb->ki_filp, current->comm);
397 * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
398 * is being issued as AIO or not. This allows us to optimise pure data writes
399 * to use REQ_FUA rather than requiring generic_write_sync() to issue a
400 * REQ_FLUSH post write. This is slightly tricky because a single request here
401 * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
402 * may be pure data writes. In that case, we still need to do a full data sync
406 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
407 const struct iomap_ops *ops, const struct iomap_dio_ops *dops)
409 struct address_space *mapping = iocb->ki_filp->f_mapping;
410 struct inode *inode = file_inode(iocb->ki_filp);
411 size_t count = iov_iter_count(iter);
412 loff_t pos = iocb->ki_pos, start = pos;
413 loff_t end = iocb->ki_pos + count - 1, ret = 0;
414 unsigned int flags = IOMAP_DIRECT;
415 bool wait_for_completion = is_sync_kiocb(iocb);
416 struct blk_plug plug;
417 struct iomap_dio *dio;
419 lockdep_assert_held(&inode->i_rwsem);
424 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
429 atomic_set(&dio->ref, 1);
431 dio->i_size = i_size_read(inode);
436 dio->submit.iter = iter;
437 dio->submit.waiter = current;
438 dio->submit.cookie = BLK_QC_T_NONE;
439 dio->submit.last_queue = NULL;
441 if (iov_iter_rw(iter) == READ) {
442 if (pos >= dio->i_size)
445 if (iter_is_iovec(iter) && iov_iter_rw(iter) == READ)
446 dio->flags |= IOMAP_DIO_DIRTY;
448 flags |= IOMAP_WRITE;
449 dio->flags |= IOMAP_DIO_WRITE;
451 /* for data sync or sync, we need sync completion processing */
452 if (iocb->ki_flags & IOCB_DSYNC)
453 dio->flags |= IOMAP_DIO_NEED_SYNC;
456 * For datasync only writes, we optimistically try using FUA for
457 * this IO. Any non-FUA write that occurs will clear this flag,
458 * hence we know before completion whether a cache flush is
461 if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
462 dio->flags |= IOMAP_DIO_WRITE_FUA;
465 if (iocb->ki_flags & IOCB_NOWAIT) {
466 if (filemap_range_has_page(mapping, start, end)) {
470 flags |= IOMAP_NOWAIT;
473 ret = filemap_write_and_wait_range(mapping, start, end);
478 * Try to invalidate cache pages for the range we're direct
479 * writing. If this invalidation fails, tough, the write will
480 * still work, but racing two incompatible write paths is a
481 * pretty crazy thing to do, so we don't support it 100%.
483 ret = invalidate_inode_pages2_range(mapping,
484 start >> PAGE_SHIFT, end >> PAGE_SHIFT);
486 dio_warn_stale_pagecache(iocb->ki_filp);
489 if (iov_iter_rw(iter) == WRITE && !wait_for_completion &&
490 !inode->i_sb->s_dio_done_wq) {
491 ret = sb_init_dio_done_wq(inode->i_sb);
496 inode_dio_begin(inode);
498 blk_start_plug(&plug);
500 ret = iomap_apply(inode, pos, count, flags, ops, dio,
503 /* magic error code to fall back to buffered I/O */
504 if (ret == -ENOTBLK) {
505 wait_for_completion = true;
512 if (iov_iter_rw(iter) == READ && pos >= dio->i_size) {
514 * We only report that we've read data up to i_size.
515 * Revert iter to a state corresponding to that as
516 * some callers (such as splice code) rely on it.
518 iov_iter_revert(iter, pos - dio->i_size);
521 } while ((count = iov_iter_count(iter)) > 0);
522 blk_finish_plug(&plug);
525 iomap_dio_set_error(dio, ret);
528 * If all the writes we issued were FUA, we don't need to flush the
529 * cache on IO completion. Clear the sync flag for this case.
531 if (dio->flags & IOMAP_DIO_WRITE_FUA)
532 dio->flags &= ~IOMAP_DIO_NEED_SYNC;
534 WRITE_ONCE(iocb->ki_cookie, dio->submit.cookie);
535 WRITE_ONCE(iocb->private, dio->submit.last_queue);
538 * We are about to drop our additional submission reference, which
539 * might be the last reference to the dio. There are three three
540 * different ways we can progress here:
542 * (a) If this is the last reference we will always complete and free
544 * (b) If this is not the last reference, and we serve an asynchronous
545 * iocb, we must never touch the dio after the decrement, the
546 * I/O completion handler will complete and free it.
547 * (c) If this is not the last reference, but we serve a synchronous
548 * iocb, the I/O completion handler will wake us up on the drop
549 * of the final reference, and we will complete and free it here
550 * after we got woken by the I/O completion handler.
552 dio->wait_for_completion = wait_for_completion;
553 if (!atomic_dec_and_test(&dio->ref)) {
554 if (!wait_for_completion)
558 set_current_state(TASK_UNINTERRUPTIBLE);
559 if (!READ_ONCE(dio->submit.waiter))
562 if (!(iocb->ki_flags & IOCB_HIPRI) ||
563 !dio->submit.last_queue ||
564 !blk_poll(dio->submit.last_queue,
565 dio->submit.cookie, true))
568 __set_current_state(TASK_RUNNING);
571 return iomap_dio_complete(dio);
577 EXPORT_SYMBOL_GPL(iomap_dio_rw);