2 * Copyright (C) 2010 Red Hat, Inc.
3 * Copyright (c) 2016 Christoph Hellwig.
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 #include <linux/module.h>
15 #include <linux/compiler.h>
17 #include <linux/iomap.h>
18 #include <linux/uaccess.h>
19 #include <linux/gfp.h>
21 #include <linux/swap.h>
22 #include <linux/pagemap.h>
23 #include <linux/file.h>
24 #include <linux/uio.h>
25 #include <linux/backing-dev.h>
26 #include <linux/buffer_head.h>
27 #include <linux/task_io_accounting_ops.h>
28 #include <linux/dax.h>
32 * Execute a iomap write on a segment of the mapping that spans a
33 * contiguous range of pages that have identical block mapping state.
35 * This avoids the need to map pages individually, do individual allocations
36 * for each page and most importantly avoid the need for filesystem specific
37 * locking per page. Instead, all the operations are amortised over the entire
38 * range of pages. It is assumed that the filesystems will lock whatever
39 * resources they require in the iomap_begin call, and release them in the
43 iomap_apply(struct inode *inode, loff_t pos, loff_t length, unsigned flags,
44 struct iomap_ops *ops, void *data, iomap_actor_t actor)
46 struct iomap iomap = { 0 };
47 loff_t written = 0, ret;
50 * Need to map a range from start position for length bytes. This can
51 * span multiple pages - it is only guaranteed to return a range of a
52 * single type of pages (e.g. all into a hole, all mapped or all
53 * unwritten). Failure at this point has nothing to undo.
55 * If allocation is required for this range, reserve the space now so
56 * that the allocation is guaranteed to succeed later on. Once we copy
57 * the data into the page cache pages, then we cannot fail otherwise we
58 * expose transient stale data. If the reserve fails, we can safely
59 * back out at this point as there is nothing to undo.
61 ret = ops->iomap_begin(inode, pos, length, flags, &iomap);
64 if (WARN_ON(iomap.offset > pos))
68 * Cut down the length to the one actually provided by the filesystem,
69 * as it might not be able to give us the whole size that we requested.
71 if (iomap.offset + iomap.length < pos + length)
72 length = iomap.offset + iomap.length - pos;
75 * Now that we have guaranteed that the space allocation will succeed.
76 * we can do the copy-in page by page without having to worry about
77 * failures exposing transient data.
79 written = actor(inode, pos, length, data, &iomap);
82 * Now the data has been copied, commit the range we've copied. This
83 * should not fail unless the filesystem has had a fatal error.
86 ret = ops->iomap_end(inode, pos, length,
87 written > 0 ? written : 0,
91 return written ? written : ret;
95 iomap_write_failed(struct inode *inode, loff_t pos, unsigned len)
97 loff_t i_size = i_size_read(inode);
100 * Only truncate newly allocated pages beyoned EOF, even if the
101 * write started inside the existing inode size.
103 if (pos + len > i_size)
104 truncate_pagecache_range(inode, max(pos, i_size), pos + len);
108 iomap_write_begin(struct inode *inode, loff_t pos, unsigned len, unsigned flags,
109 struct page **pagep, struct iomap *iomap)
111 pgoff_t index = pos >> PAGE_SHIFT;
115 BUG_ON(pos + len > iomap->offset + iomap->length);
117 page = grab_cache_page_write_begin(inode->i_mapping, index, flags);
121 status = __block_write_begin_int(page, pos, len, NULL, iomap);
122 if (unlikely(status)) {
127 iomap_write_failed(inode, pos, len);
135 iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
136 unsigned copied, struct page *page)
140 ret = generic_write_end(NULL, inode->i_mapping, pos, len,
143 iomap_write_failed(inode, pos, len);
148 iomap_write_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
151 struct iov_iter *i = data;
154 unsigned int flags = AOP_FLAG_NOFS;
157 * Copies from kernel address space cannot fail (NFSD is a big user).
159 if (!iter_is_iovec(i))
160 flags |= AOP_FLAG_UNINTERRUPTIBLE;
164 unsigned long offset; /* Offset into pagecache page */
165 unsigned long bytes; /* Bytes to write to page */
166 size_t copied; /* Bytes copied from user */
168 offset = (pos & (PAGE_SIZE - 1));
169 bytes = min_t(unsigned long, PAGE_SIZE - offset,
176 * Bring in the user page that we will copy from _first_.
177 * Otherwise there's a nasty deadlock on copying from the
178 * same page as we're writing to, without it being marked
181 * Not only is this an optimisation, but it is also required
182 * to check that the address is actually valid, when atomic
183 * usercopies are used, below.
185 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
190 status = iomap_write_begin(inode, pos, bytes, flags, &page,
192 if (unlikely(status))
195 if (mapping_writably_mapped(inode->i_mapping))
196 flush_dcache_page(page);
198 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
200 flush_dcache_page(page);
202 status = iomap_write_end(inode, pos, bytes, copied, page);
203 if (unlikely(status < 0))
209 iov_iter_advance(i, copied);
210 if (unlikely(copied == 0)) {
212 * If we were unable to copy any data at all, we must
213 * fall back to a single segment length write.
215 * If we didn't fallback here, we could livelock
216 * because not all segments in the iov can be copied at
217 * once without a pagefault.
219 bytes = min_t(unsigned long, PAGE_SIZE - offset,
220 iov_iter_single_seg_count(i));
227 balance_dirty_pages_ratelimited(inode->i_mapping);
228 } while (iov_iter_count(i) && length);
230 return written ? written : status;
234 iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *iter,
235 struct iomap_ops *ops)
237 struct inode *inode = iocb->ki_filp->f_mapping->host;
238 loff_t pos = iocb->ki_pos, ret = 0, written = 0;
240 while (iov_iter_count(iter)) {
241 ret = iomap_apply(inode, pos, iov_iter_count(iter),
242 IOMAP_WRITE, ops, iter, iomap_write_actor);
249 return written ? written : ret;
251 EXPORT_SYMBOL_GPL(iomap_file_buffered_write);
254 __iomap_read_page(struct inode *inode, loff_t offset)
256 struct address_space *mapping = inode->i_mapping;
259 page = read_mapping_page(mapping, offset >> PAGE_SHIFT, NULL);
262 if (!PageUptodate(page)) {
264 return ERR_PTR(-EIO);
270 iomap_dirty_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
277 struct page *page, *rpage;
278 unsigned long offset; /* Offset into pagecache page */
279 unsigned long bytes; /* Bytes to write to page */
281 offset = (pos & (PAGE_SIZE - 1));
282 bytes = min_t(unsigned long, PAGE_SIZE - offset, length);
284 rpage = __iomap_read_page(inode, pos);
286 return PTR_ERR(rpage);
288 status = iomap_write_begin(inode, pos, bytes,
289 AOP_FLAG_NOFS | AOP_FLAG_UNINTERRUPTIBLE,
292 if (unlikely(status))
295 WARN_ON_ONCE(!PageUptodate(page));
297 status = iomap_write_end(inode, pos, bytes, bytes, page);
298 if (unlikely(status <= 0)) {
299 if (WARN_ON_ONCE(status == 0))
310 balance_dirty_pages_ratelimited(inode->i_mapping);
317 iomap_file_dirty(struct inode *inode, loff_t pos, loff_t len,
318 struct iomap_ops *ops)
323 ret = iomap_apply(inode, pos, len, IOMAP_WRITE, ops, NULL,
333 EXPORT_SYMBOL_GPL(iomap_file_dirty);
335 static int iomap_zero(struct inode *inode, loff_t pos, unsigned offset,
336 unsigned bytes, struct iomap *iomap)
341 status = iomap_write_begin(inode, pos, bytes,
342 AOP_FLAG_UNINTERRUPTIBLE | AOP_FLAG_NOFS, &page, iomap);
346 zero_user(page, offset, bytes);
347 mark_page_accessed(page);
349 return iomap_write_end(inode, pos, bytes, bytes, page);
352 static int iomap_dax_zero(loff_t pos, unsigned offset, unsigned bytes,
355 sector_t sector = iomap->blkno +
356 (((pos & ~(PAGE_SIZE - 1)) - iomap->offset) >> 9);
358 return __dax_zero_page_range(iomap->bdev, sector, offset, bytes);
362 iomap_zero_range_actor(struct inode *inode, loff_t pos, loff_t count,
363 void *data, struct iomap *iomap)
365 bool *did_zero = data;
369 /* already zeroed? we're done. */
370 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
374 unsigned offset, bytes;
376 offset = pos & (PAGE_SIZE - 1); /* Within page */
377 bytes = min_t(unsigned, PAGE_SIZE - offset, count);
380 status = iomap_dax_zero(pos, offset, bytes, iomap);
382 status = iomap_zero(inode, pos, offset, bytes, iomap);
397 iomap_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
398 struct iomap_ops *ops)
403 ret = iomap_apply(inode, pos, len, IOMAP_ZERO,
404 ops, did_zero, iomap_zero_range_actor);
414 EXPORT_SYMBOL_GPL(iomap_zero_range);
417 iomap_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
418 struct iomap_ops *ops)
420 unsigned blocksize = (1 << inode->i_blkbits);
421 unsigned off = pos & (blocksize - 1);
423 /* Block boundary? Nothing to do */
426 return iomap_zero_range(inode, pos, blocksize - off, did_zero, ops);
428 EXPORT_SYMBOL_GPL(iomap_truncate_page);
431 iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
432 void *data, struct iomap *iomap)
434 struct page *page = data;
437 ret = __block_write_begin_int(page, pos, length, NULL, iomap);
441 block_commit_write(page, 0, length);
445 int iomap_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
446 struct iomap_ops *ops)
448 struct page *page = vmf->page;
449 struct inode *inode = file_inode(vma->vm_file);
450 unsigned long length;
455 size = i_size_read(inode);
456 if ((page->mapping != inode->i_mapping) ||
457 (page_offset(page) > size)) {
458 /* We overload EFAULT to mean page got truncated */
463 /* page is wholly or partially inside EOF */
464 if (((page->index + 1) << PAGE_SHIFT) > size)
465 length = size & ~PAGE_MASK;
469 offset = page_offset(page);
471 ret = iomap_apply(inode, offset, length,
472 IOMAP_WRITE | IOMAP_FAULT, ops, page,
473 iomap_page_mkwrite_actor);
474 if (unlikely(ret <= 0))
480 set_page_dirty(page);
481 wait_for_stable_page(page);
487 EXPORT_SYMBOL_GPL(iomap_page_mkwrite);
490 struct fiemap_extent_info *fi;
494 static int iomap_to_fiemap(struct fiemap_extent_info *fi,
495 struct iomap *iomap, u32 flags)
497 switch (iomap->type) {
502 flags |= FIEMAP_EXTENT_DELALLOC | FIEMAP_EXTENT_UNKNOWN;
504 case IOMAP_UNWRITTEN:
505 flags |= FIEMAP_EXTENT_UNWRITTEN;
511 if (iomap->flags & IOMAP_F_MERGED)
512 flags |= FIEMAP_EXTENT_MERGED;
513 if (iomap->flags & IOMAP_F_SHARED)
514 flags |= FIEMAP_EXTENT_SHARED;
516 return fiemap_fill_next_extent(fi, iomap->offset,
517 iomap->blkno != IOMAP_NULL_BLOCK ? iomap->blkno << 9: 0,
518 iomap->length, flags);
523 iomap_fiemap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
526 struct fiemap_ctx *ctx = data;
529 if (iomap->type == IOMAP_HOLE)
532 ret = iomap_to_fiemap(ctx->fi, &ctx->prev, 0);
535 case 0: /* success */
537 case 1: /* extent array full */
544 int iomap_fiemap(struct inode *inode, struct fiemap_extent_info *fi,
545 loff_t start, loff_t len, struct iomap_ops *ops)
547 struct fiemap_ctx ctx;
550 memset(&ctx, 0, sizeof(ctx));
552 ctx.prev.type = IOMAP_HOLE;
554 ret = fiemap_check_flags(fi, FIEMAP_FLAG_SYNC);
558 if (fi->fi_flags & FIEMAP_FLAG_SYNC) {
559 ret = filemap_write_and_wait(inode->i_mapping);
565 ret = iomap_apply(inode, start, len, IOMAP_REPORT, ops, &ctx,
567 /* inode with no (attribute) mapping will give ENOENT */
579 if (ctx.prev.type != IOMAP_HOLE) {
580 ret = iomap_to_fiemap(fi, &ctx.prev, FIEMAP_EXTENT_LAST);
587 EXPORT_SYMBOL_GPL(iomap_fiemap);
590 * Private flags for iomap_dio, must not overlap with the public ones in
593 #define IOMAP_DIO_WRITE (1 << 30)
594 #define IOMAP_DIO_DIRTY (1 << 31)
598 iomap_dio_end_io_t *end_io;
606 /* used during submission and for synchronous completion: */
608 struct iov_iter *iter;
609 struct task_struct *waiter;
610 struct request_queue *last_queue;
614 /* used for aio completion: */
616 struct work_struct work;
621 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
623 struct kiocb *iocb = dio->iocb;
627 ret = dio->end_io(iocb,
628 dio->error ? dio->error : dio->size,
636 /* check for short read */
637 if (iocb->ki_pos + ret > dio->i_size &&
638 !(dio->flags & IOMAP_DIO_WRITE))
639 ret = dio->i_size - iocb->ki_pos;
643 inode_dio_end(file_inode(iocb->ki_filp));
649 static void iomap_dio_complete_work(struct work_struct *work)
651 struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
652 struct kiocb *iocb = dio->iocb;
653 bool is_write = (dio->flags & IOMAP_DIO_WRITE);
656 ret = iomap_dio_complete(dio);
657 if (is_write && ret > 0)
658 ret = generic_write_sync(iocb, ret);
659 iocb->ki_complete(iocb, ret, 0);
663 * Set an error in the dio if none is set yet. We have to use cmpxchg
664 * as the submission context and the completion context(s) can race to
667 static inline void iomap_dio_set_error(struct iomap_dio *dio, int ret)
669 cmpxchg(&dio->error, 0, ret);
672 static void iomap_dio_bio_end_io(struct bio *bio)
674 struct iomap_dio *dio = bio->bi_private;
675 bool should_dirty = (dio->flags & IOMAP_DIO_DIRTY);
678 iomap_dio_set_error(dio, bio->bi_error);
680 if (atomic_dec_and_test(&dio->ref)) {
681 if (is_sync_kiocb(dio->iocb)) {
682 struct task_struct *waiter = dio->submit.waiter;
684 WRITE_ONCE(dio->submit.waiter, NULL);
685 wake_up_process(waiter);
686 } else if (dio->flags & IOMAP_DIO_WRITE) {
687 struct inode *inode = file_inode(dio->iocb->ki_filp);
689 INIT_WORK(&dio->aio.work, iomap_dio_complete_work);
690 queue_work(inode->i_sb->s_dio_done_wq, &dio->aio.work);
692 iomap_dio_complete_work(&dio->aio.work);
697 bio_check_pages_dirty(bio);
699 struct bio_vec *bvec;
702 bio_for_each_segment_all(bvec, bio, i)
703 put_page(bvec->bv_page);
709 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
712 struct page *page = ZERO_PAGE(0);
715 bio = bio_alloc(GFP_KERNEL, 1);
716 bio->bi_bdev = iomap->bdev;
717 bio->bi_iter.bi_sector =
718 iomap->blkno + ((pos - iomap->offset) >> 9);
719 bio->bi_private = dio;
720 bio->bi_end_io = iomap_dio_bio_end_io;
723 if (bio_add_page(bio, page, len, 0) != len)
725 bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
727 atomic_inc(&dio->ref);
728 return submit_bio(bio);
732 iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
733 void *data, struct iomap *iomap)
735 struct iomap_dio *dio = data;
736 unsigned blkbits = blksize_bits(bdev_logical_block_size(iomap->bdev));
737 unsigned fs_block_size = (1 << inode->i_blkbits), pad;
738 unsigned align = iov_iter_alignment(dio->submit.iter);
739 struct iov_iter iter;
741 bool need_zeroout = false;
744 if ((pos | length | align) & ((1 << blkbits) - 1))
747 switch (iomap->type) {
749 if (WARN_ON_ONCE(dio->flags & IOMAP_DIO_WRITE))
752 case IOMAP_UNWRITTEN:
753 if (!(dio->flags & IOMAP_DIO_WRITE)) {
754 iov_iter_zero(length, dio->submit.iter);
758 dio->flags |= IOMAP_DIO_UNWRITTEN;
762 if (iomap->flags & IOMAP_F_SHARED)
763 dio->flags |= IOMAP_DIO_COW;
764 if (iomap->flags & IOMAP_F_NEW)
773 * Operate on a partial iter trimmed to the extent we were called for.
774 * We'll update the iter in the dio once we're done with this extent.
776 iter = *dio->submit.iter;
777 iov_iter_truncate(&iter, length);
779 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
784 /* zero out from the start of the block to the write offset */
785 pad = pos & (fs_block_size - 1);
787 iomap_dio_zero(dio, iomap, pos - pad, pad);
794 bio = bio_alloc(GFP_KERNEL, nr_pages);
795 bio->bi_bdev = iomap->bdev;
796 bio->bi_iter.bi_sector =
797 iomap->blkno + ((pos - iomap->offset) >> 9);
798 bio->bi_private = dio;
799 bio->bi_end_io = iomap_dio_bio_end_io;
801 ret = bio_iov_iter_get_pages(bio, &iter);
807 if (dio->flags & IOMAP_DIO_WRITE) {
808 bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
809 task_io_account_write(bio->bi_iter.bi_size);
811 bio_set_op_attrs(bio, REQ_OP_READ, 0);
812 if (dio->flags & IOMAP_DIO_DIRTY)
813 bio_set_pages_dirty(bio);
816 dio->size += bio->bi_iter.bi_size;
817 pos += bio->bi_iter.bi_size;
819 nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
821 atomic_inc(&dio->ref);
823 dio->submit.last_queue = bdev_get_queue(iomap->bdev);
824 dio->submit.cookie = submit_bio(bio);
828 /* zero out from the end of the write to the end of the block */
829 pad = pos & (fs_block_size - 1);
831 iomap_dio_zero(dio, iomap, pos, fs_block_size - pad);
834 iov_iter_advance(dio->submit.iter, length);
839 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter, struct iomap_ops *ops,
840 iomap_dio_end_io_t end_io)
842 struct address_space *mapping = iocb->ki_filp->f_mapping;
843 struct inode *inode = file_inode(iocb->ki_filp);
844 size_t count = iov_iter_count(iter);
845 loff_t pos = iocb->ki_pos, end = iocb->ki_pos + count - 1, ret = 0;
846 unsigned int flags = IOMAP_DIRECT;
847 struct blk_plug plug;
848 struct iomap_dio *dio;
850 lockdep_assert_held(&inode->i_rwsem);
855 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
860 atomic_set(&dio->ref, 1);
862 dio->i_size = i_size_read(inode);
863 dio->end_io = end_io;
867 dio->submit.iter = iter;
868 if (is_sync_kiocb(iocb)) {
869 dio->submit.waiter = current;
870 dio->submit.cookie = BLK_QC_T_NONE;
871 dio->submit.last_queue = NULL;
874 if (iov_iter_rw(iter) == READ) {
875 if (pos >= dio->i_size)
878 if (iter->type == ITER_IOVEC)
879 dio->flags |= IOMAP_DIO_DIRTY;
881 dio->flags |= IOMAP_DIO_WRITE;
882 flags |= IOMAP_WRITE;
885 if (mapping->nrpages) {
886 ret = filemap_write_and_wait_range(mapping, iocb->ki_pos, end);
890 ret = invalidate_inode_pages2_range(mapping,
891 iocb->ki_pos >> PAGE_SHIFT, end >> PAGE_SHIFT);
896 inode_dio_begin(inode);
898 blk_start_plug(&plug);
900 ret = iomap_apply(inode, pos, count, flags, ops, dio,
903 /* magic error code to fall back to buffered I/O */
909 } while ((count = iov_iter_count(iter)) > 0);
910 blk_finish_plug(&plug);
913 iomap_dio_set_error(dio, ret);
915 if (ret >= 0 && iov_iter_rw(iter) == WRITE && !is_sync_kiocb(iocb) &&
916 !inode->i_sb->s_dio_done_wq) {
917 ret = sb_init_dio_done_wq(inode->i_sb);
919 iomap_dio_set_error(dio, ret);
922 if (!atomic_dec_and_test(&dio->ref)) {
923 if (!is_sync_kiocb(iocb))
927 set_current_state(TASK_UNINTERRUPTIBLE);
928 if (!READ_ONCE(dio->submit.waiter))
931 if (!(iocb->ki_flags & IOCB_HIPRI) ||
932 !dio->submit.last_queue ||
933 !blk_mq_poll(dio->submit.last_queue,
937 __set_current_state(TASK_RUNNING);
941 * Try again to invalidate clean pages which might have been cached by
942 * non-direct readahead, or faulted in by get_user_pages() if the source
943 * of the write was an mmap'ed region of the file we're writing. Either
944 * one is a pretty crazy thing to do, so we don't support it 100%. If
945 * this invalidation fails, tough, the write still worked...
947 if (iov_iter_rw(iter) == WRITE && mapping->nrpages) {
948 ret = invalidate_inode_pages2_range(mapping,
949 iocb->ki_pos >> PAGE_SHIFT, end >> PAGE_SHIFT);
953 return iomap_dio_complete(dio);
959 EXPORT_SYMBOL_GPL(iomap_dio_rw);