Loading...
1/*
2 * fs/dax.c - Direct Access filesystem code
3 * Copyright (c) 2013-2014 Intel Corporation
4 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
5 * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
6 *
7 * This program is free software; you can redistribute it and/or modify it
8 * under the terms and conditions of the GNU General Public License,
9 * version 2, as published by the Free Software Foundation.
10 *
11 * This program is distributed in the hope it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 * more details.
15 */
16
17#include <linux/atomic.h>
18#include <linux/blkdev.h>
19#include <linux/buffer_head.h>
20#include <linux/dax.h>
21#include <linux/fs.h>
22#include <linux/genhd.h>
23#include <linux/highmem.h>
24#include <linux/memcontrol.h>
25#include <linux/mm.h>
26#include <linux/mutex.h>
27#include <linux/pagevec.h>
28#include <linux/pmem.h>
29#include <linux/sched.h>
30#include <linux/uio.h>
31#include <linux/vmstat.h>
32#include <linux/pfn_t.h>
33#include <linux/sizes.h>
34
35static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
36{
37 struct request_queue *q = bdev->bd_queue;
38 long rc = -EIO;
39
40 dax->addr = (void __pmem *) ERR_PTR(-EIO);
41 if (blk_queue_enter(q, true) != 0)
42 return rc;
43
44 rc = bdev_direct_access(bdev, dax);
45 if (rc < 0) {
46 dax->addr = (void __pmem *) ERR_PTR(rc);
47 blk_queue_exit(q);
48 return rc;
49 }
50 return rc;
51}
52
53static void dax_unmap_atomic(struct block_device *bdev,
54 const struct blk_dax_ctl *dax)
55{
56 if (IS_ERR(dax->addr))
57 return;
58 blk_queue_exit(bdev->bd_queue);
59}
60
61struct page *read_dax_sector(struct block_device *bdev, sector_t n)
62{
63 struct page *page = alloc_pages(GFP_KERNEL, 0);
64 struct blk_dax_ctl dax = {
65 .size = PAGE_SIZE,
66 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
67 };
68 long rc;
69
70 if (!page)
71 return ERR_PTR(-ENOMEM);
72
73 rc = dax_map_atomic(bdev, &dax);
74 if (rc < 0)
75 return ERR_PTR(rc);
76 memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
77 dax_unmap_atomic(bdev, &dax);
78 return page;
79}
80
81/*
82 * dax_clear_sectors() is called from within transaction context from XFS,
83 * and hence this means the stack from this point must follow GFP_NOFS
84 * semantics for all operations.
85 */
86int dax_clear_sectors(struct block_device *bdev, sector_t _sector, long _size)
87{
88 struct blk_dax_ctl dax = {
89 .sector = _sector,
90 .size = _size,
91 };
92
93 might_sleep();
94 do {
95 long count, sz;
96
97 count = dax_map_atomic(bdev, &dax);
98 if (count < 0)
99 return count;
100 sz = min_t(long, count, SZ_128K);
101 clear_pmem(dax.addr, sz);
102 dax.size -= sz;
103 dax.sector += sz / 512;
104 dax_unmap_atomic(bdev, &dax);
105 cond_resched();
106 } while (dax.size);
107
108 wmb_pmem();
109 return 0;
110}
111EXPORT_SYMBOL_GPL(dax_clear_sectors);
112
113/* the clear_pmem() calls are ordered by a wmb_pmem() in the caller */
114static void dax_new_buf(void __pmem *addr, unsigned size, unsigned first,
115 loff_t pos, loff_t end)
116{
117 loff_t final = end - pos + first; /* The final byte of the buffer */
118
119 if (first > 0)
120 clear_pmem(addr, first);
121 if (final < size)
122 clear_pmem(addr + final, size - final);
123}
124
125static bool buffer_written(struct buffer_head *bh)
126{
127 return buffer_mapped(bh) && !buffer_unwritten(bh);
128}
129
130/*
131 * When ext4 encounters a hole, it returns without modifying the buffer_head
132 * which means that we can't trust b_size. To cope with this, we set b_state
133 * to 0 before calling get_block and, if any bit is set, we know we can trust
134 * b_size. Unfortunate, really, since ext4 knows precisely how long a hole is
135 * and would save us time calling get_block repeatedly.
136 */
137static bool buffer_size_valid(struct buffer_head *bh)
138{
139 return bh->b_state != 0;
140}
141
142
143static sector_t to_sector(const struct buffer_head *bh,
144 const struct inode *inode)
145{
146 sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
147
148 return sector;
149}
150
151static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
152 loff_t start, loff_t end, get_block_t get_block,
153 struct buffer_head *bh)
154{
155 loff_t pos = start, max = start, bh_max = start;
156 bool hole = false, need_wmb = false;
157 struct block_device *bdev = NULL;
158 int rw = iov_iter_rw(iter), rc;
159 long map_len = 0;
160 struct blk_dax_ctl dax = {
161 .addr = (void __pmem *) ERR_PTR(-EIO),
162 };
163
164 if (rw == READ)
165 end = min(end, i_size_read(inode));
166
167 while (pos < end) {
168 size_t len;
169 if (pos == max) {
170 unsigned blkbits = inode->i_blkbits;
171 long page = pos >> PAGE_SHIFT;
172 sector_t block = page << (PAGE_SHIFT - blkbits);
173 unsigned first = pos - (block << blkbits);
174 long size;
175
176 if (pos == bh_max) {
177 bh->b_size = PAGE_ALIGN(end - pos);
178 bh->b_state = 0;
179 rc = get_block(inode, block, bh, rw == WRITE);
180 if (rc)
181 break;
182 if (!buffer_size_valid(bh))
183 bh->b_size = 1 << blkbits;
184 bh_max = pos - first + bh->b_size;
185 bdev = bh->b_bdev;
186 } else {
187 unsigned done = bh->b_size -
188 (bh_max - (pos - first));
189 bh->b_blocknr += done >> blkbits;
190 bh->b_size -= done;
191 }
192
193 hole = rw == READ && !buffer_written(bh);
194 if (hole) {
195 size = bh->b_size - first;
196 } else {
197 dax_unmap_atomic(bdev, &dax);
198 dax.sector = to_sector(bh, inode);
199 dax.size = bh->b_size;
200 map_len = dax_map_atomic(bdev, &dax);
201 if (map_len < 0) {
202 rc = map_len;
203 break;
204 }
205 if (buffer_unwritten(bh) || buffer_new(bh)) {
206 dax_new_buf(dax.addr, map_len, first,
207 pos, end);
208 need_wmb = true;
209 }
210 dax.addr += first;
211 size = map_len - first;
212 }
213 max = min(pos + size, end);
214 }
215
216 if (iov_iter_rw(iter) == WRITE) {
217 len = copy_from_iter_pmem(dax.addr, max - pos, iter);
218 need_wmb = true;
219 } else if (!hole)
220 len = copy_to_iter((void __force *) dax.addr, max - pos,
221 iter);
222 else
223 len = iov_iter_zero(max - pos, iter);
224
225 if (!len) {
226 rc = -EFAULT;
227 break;
228 }
229
230 pos += len;
231 if (!IS_ERR(dax.addr))
232 dax.addr += len;
233 }
234
235 if (need_wmb)
236 wmb_pmem();
237 dax_unmap_atomic(bdev, &dax);
238
239 return (pos == start) ? rc : pos - start;
240}
241
242/**
243 * dax_do_io - Perform I/O to a DAX file
244 * @iocb: The control block for this I/O
245 * @inode: The file which the I/O is directed at
246 * @iter: The addresses to do I/O from or to
247 * @pos: The file offset where the I/O starts
248 * @get_block: The filesystem method used to translate file offsets to blocks
249 * @end_io: A filesystem callback for I/O completion
250 * @flags: See below
251 *
252 * This function uses the same locking scheme as do_blockdev_direct_IO:
253 * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
254 * caller for writes. For reads, we take and release the i_mutex ourselves.
255 * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
256 * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
257 * is in progress.
258 */
259ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
260 struct iov_iter *iter, loff_t pos, get_block_t get_block,
261 dio_iodone_t end_io, int flags)
262{
263 struct buffer_head bh;
264 ssize_t retval = -EINVAL;
265 loff_t end = pos + iov_iter_count(iter);
266
267 memset(&bh, 0, sizeof(bh));
268 bh.b_bdev = inode->i_sb->s_bdev;
269
270 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) {
271 struct address_space *mapping = inode->i_mapping;
272 inode_lock(inode);
273 retval = filemap_write_and_wait_range(mapping, pos, end - 1);
274 if (retval) {
275 inode_unlock(inode);
276 goto out;
277 }
278 }
279
280 /* Protects against truncate */
281 if (!(flags & DIO_SKIP_DIO_COUNT))
282 inode_dio_begin(inode);
283
284 retval = dax_io(inode, iter, pos, end, get_block, &bh);
285
286 if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
287 inode_unlock(inode);
288
289 if (end_io) {
290 int err;
291
292 err = end_io(iocb, pos, retval, bh.b_private);
293 if (err)
294 retval = err;
295 }
296
297 if (!(flags & DIO_SKIP_DIO_COUNT))
298 inode_dio_end(inode);
299 out:
300 return retval;
301}
302EXPORT_SYMBOL_GPL(dax_do_io);
303
304/*
305 * The user has performed a load from a hole in the file. Allocating
306 * a new page in the file would cause excessive storage usage for
307 * workloads with sparse files. We allocate a page cache page instead.
308 * We'll kick it out of the page cache if it's ever written to,
309 * otherwise it will simply fall out of the page cache under memory
310 * pressure without ever having been dirtied.
311 */
312static int dax_load_hole(struct address_space *mapping, struct page *page,
313 struct vm_fault *vmf)
314{
315 unsigned long size;
316 struct inode *inode = mapping->host;
317 if (!page)
318 page = find_or_create_page(mapping, vmf->pgoff,
319 GFP_KERNEL | __GFP_ZERO);
320 if (!page)
321 return VM_FAULT_OOM;
322 /* Recheck i_size under page lock to avoid truncate race */
323 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
324 if (vmf->pgoff >= size) {
325 unlock_page(page);
326 put_page(page);
327 return VM_FAULT_SIGBUS;
328 }
329
330 vmf->page = page;
331 return VM_FAULT_LOCKED;
332}
333
334static int copy_user_bh(struct page *to, struct inode *inode,
335 struct buffer_head *bh, unsigned long vaddr)
336{
337 struct blk_dax_ctl dax = {
338 .sector = to_sector(bh, inode),
339 .size = bh->b_size,
340 };
341 struct block_device *bdev = bh->b_bdev;
342 void *vto;
343
344 if (dax_map_atomic(bdev, &dax) < 0)
345 return PTR_ERR(dax.addr);
346 vto = kmap_atomic(to);
347 copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
348 kunmap_atomic(vto);
349 dax_unmap_atomic(bdev, &dax);
350 return 0;
351}
352
353#define NO_SECTOR -1
354#define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT))
355
356static int dax_radix_entry(struct address_space *mapping, pgoff_t index,
357 sector_t sector, bool pmd_entry, bool dirty)
358{
359 struct radix_tree_root *page_tree = &mapping->page_tree;
360 pgoff_t pmd_index = DAX_PMD_INDEX(index);
361 int type, error = 0;
362 void *entry;
363
364 WARN_ON_ONCE(pmd_entry && !dirty);
365 if (dirty)
366 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
367
368 spin_lock_irq(&mapping->tree_lock);
369
370 entry = radix_tree_lookup(page_tree, pmd_index);
371 if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) {
372 index = pmd_index;
373 goto dirty;
374 }
375
376 entry = radix_tree_lookup(page_tree, index);
377 if (entry) {
378 type = RADIX_DAX_TYPE(entry);
379 if (WARN_ON_ONCE(type != RADIX_DAX_PTE &&
380 type != RADIX_DAX_PMD)) {
381 error = -EIO;
382 goto unlock;
383 }
384
385 if (!pmd_entry || type == RADIX_DAX_PMD)
386 goto dirty;
387
388 /*
389 * We only insert dirty PMD entries into the radix tree. This
390 * means we don't need to worry about removing a dirty PTE
391 * entry and inserting a clean PMD entry, thus reducing the
392 * range we would flush with a follow-up fsync/msync call.
393 */
394 radix_tree_delete(&mapping->page_tree, index);
395 mapping->nrexceptional--;
396 }
397
398 if (sector == NO_SECTOR) {
399 /*
400 * This can happen during correct operation if our pfn_mkwrite
401 * fault raced against a hole punch operation. If this
402 * happens the pte that was hole punched will have been
403 * unmapped and the radix tree entry will have been removed by
404 * the time we are called, but the call will still happen. We
405 * will return all the way up to wp_pfn_shared(), where the
406 * pte_same() check will fail, eventually causing page fault
407 * to be retried by the CPU.
408 */
409 goto unlock;
410 }
411
412 error = radix_tree_insert(page_tree, index,
413 RADIX_DAX_ENTRY(sector, pmd_entry));
414 if (error)
415 goto unlock;
416
417 mapping->nrexceptional++;
418 dirty:
419 if (dirty)
420 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
421 unlock:
422 spin_unlock_irq(&mapping->tree_lock);
423 return error;
424}
425
426static int dax_writeback_one(struct block_device *bdev,
427 struct address_space *mapping, pgoff_t index, void *entry)
428{
429 struct radix_tree_root *page_tree = &mapping->page_tree;
430 int type = RADIX_DAX_TYPE(entry);
431 struct radix_tree_node *node;
432 struct blk_dax_ctl dax;
433 void **slot;
434 int ret = 0;
435
436 spin_lock_irq(&mapping->tree_lock);
437 /*
438 * Regular page slots are stabilized by the page lock even
439 * without the tree itself locked. These unlocked entries
440 * need verification under the tree lock.
441 */
442 if (!__radix_tree_lookup(page_tree, index, &node, &slot))
443 goto unlock;
444 if (*slot != entry)
445 goto unlock;
446
447 /* another fsync thread may have already written back this entry */
448 if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
449 goto unlock;
450
451 if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
452 ret = -EIO;
453 goto unlock;
454 }
455
456 dax.sector = RADIX_DAX_SECTOR(entry);
457 dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
458 spin_unlock_irq(&mapping->tree_lock);
459
460 /*
461 * We cannot hold tree_lock while calling dax_map_atomic() because it
462 * eventually calls cond_resched().
463 */
464 ret = dax_map_atomic(bdev, &dax);
465 if (ret < 0)
466 return ret;
467
468 if (WARN_ON_ONCE(ret < dax.size)) {
469 ret = -EIO;
470 goto unmap;
471 }
472
473 wb_cache_pmem(dax.addr, dax.size);
474
475 spin_lock_irq(&mapping->tree_lock);
476 radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
477 spin_unlock_irq(&mapping->tree_lock);
478 unmap:
479 dax_unmap_atomic(bdev, &dax);
480 return ret;
481
482 unlock:
483 spin_unlock_irq(&mapping->tree_lock);
484 return ret;
485}
486
487/*
488 * Flush the mapping to the persistent domain within the byte range of [start,
489 * end]. This is required by data integrity operations to ensure file data is
490 * on persistent storage prior to completion of the operation.
491 */
492int dax_writeback_mapping_range(struct address_space *mapping,
493 struct block_device *bdev, struct writeback_control *wbc)
494{
495 struct inode *inode = mapping->host;
496 pgoff_t start_index, end_index, pmd_index;
497 pgoff_t indices[PAGEVEC_SIZE];
498 struct pagevec pvec;
499 bool done = false;
500 int i, ret = 0;
501 void *entry;
502
503 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
504 return -EIO;
505
506 if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
507 return 0;
508
509 start_index = wbc->range_start >> PAGE_SHIFT;
510 end_index = wbc->range_end >> PAGE_SHIFT;
511 pmd_index = DAX_PMD_INDEX(start_index);
512
513 rcu_read_lock();
514 entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
515 rcu_read_unlock();
516
517 /* see if the start of our range is covered by a PMD entry */
518 if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
519 start_index = pmd_index;
520
521 tag_pages_for_writeback(mapping, start_index, end_index);
522
523 pagevec_init(&pvec, 0);
524 while (!done) {
525 pvec.nr = find_get_entries_tag(mapping, start_index,
526 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
527 pvec.pages, indices);
528
529 if (pvec.nr == 0)
530 break;
531
532 for (i = 0; i < pvec.nr; i++) {
533 if (indices[i] > end_index) {
534 done = true;
535 break;
536 }
537
538 ret = dax_writeback_one(bdev, mapping, indices[i],
539 pvec.pages[i]);
540 if (ret < 0)
541 return ret;
542 }
543 }
544 wmb_pmem();
545 return 0;
546}
547EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
548
549static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
550 struct vm_area_struct *vma, struct vm_fault *vmf)
551{
552 unsigned long vaddr = (unsigned long)vmf->virtual_address;
553 struct address_space *mapping = inode->i_mapping;
554 struct block_device *bdev = bh->b_bdev;
555 struct blk_dax_ctl dax = {
556 .sector = to_sector(bh, inode),
557 .size = bh->b_size,
558 };
559 pgoff_t size;
560 int error;
561
562 i_mmap_lock_read(mapping);
563
564 /*
565 * Check truncate didn't happen while we were allocating a block.
566 * If it did, this block may or may not be still allocated to the
567 * file. We can't tell the filesystem to free it because we can't
568 * take i_mutex here. In the worst case, the file still has blocks
569 * allocated past the end of the file.
570 */
571 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
572 if (unlikely(vmf->pgoff >= size)) {
573 error = -EIO;
574 goto out;
575 }
576
577 if (dax_map_atomic(bdev, &dax) < 0) {
578 error = PTR_ERR(dax.addr);
579 goto out;
580 }
581
582 if (buffer_unwritten(bh) || buffer_new(bh)) {
583 clear_pmem(dax.addr, PAGE_SIZE);
584 wmb_pmem();
585 }
586 dax_unmap_atomic(bdev, &dax);
587
588 error = dax_radix_entry(mapping, vmf->pgoff, dax.sector, false,
589 vmf->flags & FAULT_FLAG_WRITE);
590 if (error)
591 goto out;
592
593 error = vm_insert_mixed(vma, vaddr, dax.pfn);
594
595 out:
596 i_mmap_unlock_read(mapping);
597
598 return error;
599}
600
601/**
602 * __dax_fault - handle a page fault on a DAX file
603 * @vma: The virtual memory area where the fault occurred
604 * @vmf: The description of the fault
605 * @get_block: The filesystem method used to translate file offsets to blocks
606 * @complete_unwritten: The filesystem method used to convert unwritten blocks
607 * to written so the data written to them is exposed. This is required for
608 * required by write faults for filesystems that will return unwritten
609 * extent mappings from @get_block, but it is optional for reads as
610 * dax_insert_mapping() will always zero unwritten blocks. If the fs does
611 * not support unwritten extents, the it should pass NULL.
612 *
613 * When a page fault occurs, filesystems may call this helper in their
614 * fault handler for DAX files. __dax_fault() assumes the caller has done all
615 * the necessary locking for the page fault to proceed successfully.
616 */
617int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
618 get_block_t get_block, dax_iodone_t complete_unwritten)
619{
620 struct file *file = vma->vm_file;
621 struct address_space *mapping = file->f_mapping;
622 struct inode *inode = mapping->host;
623 struct page *page;
624 struct buffer_head bh;
625 unsigned long vaddr = (unsigned long)vmf->virtual_address;
626 unsigned blkbits = inode->i_blkbits;
627 sector_t block;
628 pgoff_t size;
629 int error;
630 int major = 0;
631
632 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
633 if (vmf->pgoff >= size)
634 return VM_FAULT_SIGBUS;
635
636 memset(&bh, 0, sizeof(bh));
637 block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
638 bh.b_bdev = inode->i_sb->s_bdev;
639 bh.b_size = PAGE_SIZE;
640
641 repeat:
642 page = find_get_page(mapping, vmf->pgoff);
643 if (page) {
644 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
645 put_page(page);
646 return VM_FAULT_RETRY;
647 }
648 if (unlikely(page->mapping != mapping)) {
649 unlock_page(page);
650 put_page(page);
651 goto repeat;
652 }
653 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
654 if (unlikely(vmf->pgoff >= size)) {
655 /*
656 * We have a struct page covering a hole in the file
657 * from a read fault and we've raced with a truncate
658 */
659 error = -EIO;
660 goto unlock_page;
661 }
662 }
663
664 error = get_block(inode, block, &bh, 0);
665 if (!error && (bh.b_size < PAGE_SIZE))
666 error = -EIO; /* fs corruption? */
667 if (error)
668 goto unlock_page;
669
670 if (!buffer_mapped(&bh) && !buffer_unwritten(&bh) && !vmf->cow_page) {
671 if (vmf->flags & FAULT_FLAG_WRITE) {
672 error = get_block(inode, block, &bh, 1);
673 count_vm_event(PGMAJFAULT);
674 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
675 major = VM_FAULT_MAJOR;
676 if (!error && (bh.b_size < PAGE_SIZE))
677 error = -EIO;
678 if (error)
679 goto unlock_page;
680 } else {
681 return dax_load_hole(mapping, page, vmf);
682 }
683 }
684
685 if (vmf->cow_page) {
686 struct page *new_page = vmf->cow_page;
687 if (buffer_written(&bh))
688 error = copy_user_bh(new_page, inode, &bh, vaddr);
689 else
690 clear_user_highpage(new_page, vaddr);
691 if (error)
692 goto unlock_page;
693 vmf->page = page;
694 if (!page) {
695 i_mmap_lock_read(mapping);
696 /* Check we didn't race with truncate */
697 size = (i_size_read(inode) + PAGE_SIZE - 1) >>
698 PAGE_SHIFT;
699 if (vmf->pgoff >= size) {
700 i_mmap_unlock_read(mapping);
701 error = -EIO;
702 goto out;
703 }
704 }
705 return VM_FAULT_LOCKED;
706 }
707
708 /* Check we didn't race with a read fault installing a new page */
709 if (!page && major)
710 page = find_lock_page(mapping, vmf->pgoff);
711
712 if (page) {
713 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
714 PAGE_SIZE, 0);
715 delete_from_page_cache(page);
716 unlock_page(page);
717 put_page(page);
718 page = NULL;
719 }
720
721 /*
722 * If we successfully insert the new mapping over an unwritten extent,
723 * we need to ensure we convert the unwritten extent. If there is an
724 * error inserting the mapping, the filesystem needs to leave it as
725 * unwritten to prevent exposure of the stale underlying data to
726 * userspace, but we still need to call the completion function so
727 * the private resources on the mapping buffer can be released. We
728 * indicate what the callback should do via the uptodate variable, same
729 * as for normal BH based IO completions.
730 */
731 error = dax_insert_mapping(inode, &bh, vma, vmf);
732 if (buffer_unwritten(&bh)) {
733 if (complete_unwritten)
734 complete_unwritten(&bh, !error);
735 else
736 WARN_ON_ONCE(!(vmf->flags & FAULT_FLAG_WRITE));
737 }
738
739 out:
740 if (error == -ENOMEM)
741 return VM_FAULT_OOM | major;
742 /* -EBUSY is fine, somebody else faulted on the same PTE */
743 if ((error < 0) && (error != -EBUSY))
744 return VM_FAULT_SIGBUS | major;
745 return VM_FAULT_NOPAGE | major;
746
747 unlock_page:
748 if (page) {
749 unlock_page(page);
750 put_page(page);
751 }
752 goto out;
753}
754EXPORT_SYMBOL(__dax_fault);
755
756/**
757 * dax_fault - handle a page fault on a DAX file
758 * @vma: The virtual memory area where the fault occurred
759 * @vmf: The description of the fault
760 * @get_block: The filesystem method used to translate file offsets to blocks
761 *
762 * When a page fault occurs, filesystems may call this helper in their
763 * fault handler for DAX files.
764 */
765int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
766 get_block_t get_block, dax_iodone_t complete_unwritten)
767{
768 int result;
769 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
770
771 if (vmf->flags & FAULT_FLAG_WRITE) {
772 sb_start_pagefault(sb);
773 file_update_time(vma->vm_file);
774 }
775 result = __dax_fault(vma, vmf, get_block, complete_unwritten);
776 if (vmf->flags & FAULT_FLAG_WRITE)
777 sb_end_pagefault(sb);
778
779 return result;
780}
781EXPORT_SYMBOL_GPL(dax_fault);
782
783#ifdef CONFIG_TRANSPARENT_HUGEPAGE
784/*
785 * The 'colour' (ie low bits) within a PMD of a page offset. This comes up
786 * more often than one might expect in the below function.
787 */
788#define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
789
790static void __dax_dbg(struct buffer_head *bh, unsigned long address,
791 const char *reason, const char *fn)
792{
793 if (bh) {
794 char bname[BDEVNAME_SIZE];
795 bdevname(bh->b_bdev, bname);
796 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
797 "length %zd fallback: %s\n", fn, current->comm,
798 address, bname, bh->b_state, (u64)bh->b_blocknr,
799 bh->b_size, reason);
800 } else {
801 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
802 current->comm, address, reason);
803 }
804}
805
806#define dax_pmd_dbg(bh, address, reason) __dax_dbg(bh, address, reason, "dax_pmd")
807
808int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
809 pmd_t *pmd, unsigned int flags, get_block_t get_block,
810 dax_iodone_t complete_unwritten)
811{
812 struct file *file = vma->vm_file;
813 struct address_space *mapping = file->f_mapping;
814 struct inode *inode = mapping->host;
815 struct buffer_head bh;
816 unsigned blkbits = inode->i_blkbits;
817 unsigned long pmd_addr = address & PMD_MASK;
818 bool write = flags & FAULT_FLAG_WRITE;
819 struct block_device *bdev;
820 pgoff_t size, pgoff;
821 sector_t block;
822 int error, result = 0;
823 bool alloc = false;
824
825 /* dax pmd mappings require pfn_t_devmap() */
826 if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
827 return VM_FAULT_FALLBACK;
828
829 /* Fall back to PTEs if we're going to COW */
830 if (write && !(vma->vm_flags & VM_SHARED)) {
831 split_huge_pmd(vma, pmd, address);
832 dax_pmd_dbg(NULL, address, "cow write");
833 return VM_FAULT_FALLBACK;
834 }
835 /* If the PMD would extend outside the VMA */
836 if (pmd_addr < vma->vm_start) {
837 dax_pmd_dbg(NULL, address, "vma start unaligned");
838 return VM_FAULT_FALLBACK;
839 }
840 if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
841 dax_pmd_dbg(NULL, address, "vma end unaligned");
842 return VM_FAULT_FALLBACK;
843 }
844
845 pgoff = linear_page_index(vma, pmd_addr);
846 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
847 if (pgoff >= size)
848 return VM_FAULT_SIGBUS;
849 /* If the PMD would cover blocks out of the file */
850 if ((pgoff | PG_PMD_COLOUR) >= size) {
851 dax_pmd_dbg(NULL, address,
852 "offset + huge page size > file size");
853 return VM_FAULT_FALLBACK;
854 }
855
856 memset(&bh, 0, sizeof(bh));
857 bh.b_bdev = inode->i_sb->s_bdev;
858 block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
859
860 bh.b_size = PMD_SIZE;
861
862 if (get_block(inode, block, &bh, 0) != 0)
863 return VM_FAULT_SIGBUS;
864
865 if (!buffer_mapped(&bh) && write) {
866 if (get_block(inode, block, &bh, 1) != 0)
867 return VM_FAULT_SIGBUS;
868 alloc = true;
869 }
870
871 bdev = bh.b_bdev;
872
873 /*
874 * If the filesystem isn't willing to tell us the length of a hole,
875 * just fall back to PTEs. Calling get_block 512 times in a loop
876 * would be silly.
877 */
878 if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
879 dax_pmd_dbg(&bh, address, "allocated block too small");
880 return VM_FAULT_FALLBACK;
881 }
882
883 /*
884 * If we allocated new storage, make sure no process has any
885 * zero pages covering this hole
886 */
887 if (alloc) {
888 loff_t lstart = pgoff << PAGE_SHIFT;
889 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
890
891 truncate_pagecache_range(inode, lstart, lend);
892 }
893
894 i_mmap_lock_read(mapping);
895
896 /*
897 * If a truncate happened while we were allocating blocks, we may
898 * leave blocks allocated to the file that are beyond EOF. We can't
899 * take i_mutex here, so just leave them hanging; they'll be freed
900 * when the file is deleted.
901 */
902 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
903 if (pgoff >= size) {
904 result = VM_FAULT_SIGBUS;
905 goto out;
906 }
907 if ((pgoff | PG_PMD_COLOUR) >= size) {
908 dax_pmd_dbg(&bh, address,
909 "offset + huge page size > file size");
910 goto fallback;
911 }
912
913 if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
914 spinlock_t *ptl;
915 pmd_t entry;
916 struct page *zero_page = get_huge_zero_page();
917
918 if (unlikely(!zero_page)) {
919 dax_pmd_dbg(&bh, address, "no zero page");
920 goto fallback;
921 }
922
923 ptl = pmd_lock(vma->vm_mm, pmd);
924 if (!pmd_none(*pmd)) {
925 spin_unlock(ptl);
926 dax_pmd_dbg(&bh, address, "pmd already present");
927 goto fallback;
928 }
929
930 dev_dbg(part_to_dev(bdev->bd_part),
931 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
932 __func__, current->comm, address,
933 (unsigned long long) to_sector(&bh, inode));
934
935 entry = mk_pmd(zero_page, vma->vm_page_prot);
936 entry = pmd_mkhuge(entry);
937 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
938 result = VM_FAULT_NOPAGE;
939 spin_unlock(ptl);
940 } else {
941 struct blk_dax_ctl dax = {
942 .sector = to_sector(&bh, inode),
943 .size = PMD_SIZE,
944 };
945 long length = dax_map_atomic(bdev, &dax);
946
947 if (length < 0) {
948 result = VM_FAULT_SIGBUS;
949 goto out;
950 }
951 if (length < PMD_SIZE) {
952 dax_pmd_dbg(&bh, address, "dax-length too small");
953 dax_unmap_atomic(bdev, &dax);
954 goto fallback;
955 }
956 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
957 dax_pmd_dbg(&bh, address, "pfn unaligned");
958 dax_unmap_atomic(bdev, &dax);
959 goto fallback;
960 }
961
962 if (!pfn_t_devmap(dax.pfn)) {
963 dax_unmap_atomic(bdev, &dax);
964 dax_pmd_dbg(&bh, address, "pfn not in memmap");
965 goto fallback;
966 }
967
968 if (buffer_unwritten(&bh) || buffer_new(&bh)) {
969 clear_pmem(dax.addr, PMD_SIZE);
970 wmb_pmem();
971 count_vm_event(PGMAJFAULT);
972 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
973 result |= VM_FAULT_MAJOR;
974 }
975 dax_unmap_atomic(bdev, &dax);
976
977 /*
978 * For PTE faults we insert a radix tree entry for reads, and
979 * leave it clean. Then on the first write we dirty the radix
980 * tree entry via the dax_pfn_mkwrite() path. This sequence
981 * allows the dax_pfn_mkwrite() call to be simpler and avoid a
982 * call into get_block() to translate the pgoff to a sector in
983 * order to be able to create a new radix tree entry.
984 *
985 * The PMD path doesn't have an equivalent to
986 * dax_pfn_mkwrite(), though, so for a read followed by a
987 * write we traverse all the way through __dax_pmd_fault()
988 * twice. This means we can just skip inserting a radix tree
989 * entry completely on the initial read and just wait until
990 * the write to insert a dirty entry.
991 */
992 if (write) {
993 error = dax_radix_entry(mapping, pgoff, dax.sector,
994 true, true);
995 if (error) {
996 dax_pmd_dbg(&bh, address,
997 "PMD radix insertion failed");
998 goto fallback;
999 }
1000 }
1001
1002 dev_dbg(part_to_dev(bdev->bd_part),
1003 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
1004 __func__, current->comm, address,
1005 pfn_t_to_pfn(dax.pfn),
1006 (unsigned long long) dax.sector);
1007 result |= vmf_insert_pfn_pmd(vma, address, pmd,
1008 dax.pfn, write);
1009 }
1010
1011 out:
1012 i_mmap_unlock_read(mapping);
1013
1014 if (buffer_unwritten(&bh))
1015 complete_unwritten(&bh, !(result & VM_FAULT_ERROR));
1016
1017 return result;
1018
1019 fallback:
1020 count_vm_event(THP_FAULT_FALLBACK);
1021 result = VM_FAULT_FALLBACK;
1022 goto out;
1023}
1024EXPORT_SYMBOL_GPL(__dax_pmd_fault);
1025
1026/**
1027 * dax_pmd_fault - handle a PMD fault on a DAX file
1028 * @vma: The virtual memory area where the fault occurred
1029 * @vmf: The description of the fault
1030 * @get_block: The filesystem method used to translate file offsets to blocks
1031 *
1032 * When a page fault occurs, filesystems may call this helper in their
1033 * pmd_fault handler for DAX files.
1034 */
1035int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1036 pmd_t *pmd, unsigned int flags, get_block_t get_block,
1037 dax_iodone_t complete_unwritten)
1038{
1039 int result;
1040 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
1041
1042 if (flags & FAULT_FLAG_WRITE) {
1043 sb_start_pagefault(sb);
1044 file_update_time(vma->vm_file);
1045 }
1046 result = __dax_pmd_fault(vma, address, pmd, flags, get_block,
1047 complete_unwritten);
1048 if (flags & FAULT_FLAG_WRITE)
1049 sb_end_pagefault(sb);
1050
1051 return result;
1052}
1053EXPORT_SYMBOL_GPL(dax_pmd_fault);
1054#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1055
1056/**
1057 * dax_pfn_mkwrite - handle first write to DAX page
1058 * @vma: The virtual memory area where the fault occurred
1059 * @vmf: The description of the fault
1060 */
1061int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1062{
1063 struct file *file = vma->vm_file;
1064 int error;
1065
1066 /*
1067 * We pass NO_SECTOR to dax_radix_entry() because we expect that a
1068 * RADIX_DAX_PTE entry already exists in the radix tree from a
1069 * previous call to __dax_fault(). We just want to look up that PTE
1070 * entry using vmf->pgoff and make sure the dirty tag is set. This
1071 * saves us from having to make a call to get_block() here to look
1072 * up the sector.
1073 */
1074 error = dax_radix_entry(file->f_mapping, vmf->pgoff, NO_SECTOR, false,
1075 true);
1076
1077 if (error == -ENOMEM)
1078 return VM_FAULT_OOM;
1079 if (error)
1080 return VM_FAULT_SIGBUS;
1081 return VM_FAULT_NOPAGE;
1082}
1083EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1084
1085/**
1086 * dax_zero_page_range - zero a range within a page of a DAX file
1087 * @inode: The file being truncated
1088 * @from: The file offset that is being truncated to
1089 * @length: The number of bytes to zero
1090 * @get_block: The filesystem method used to translate file offsets to blocks
1091 *
1092 * This function can be called by a filesystem when it is zeroing part of a
1093 * page in a DAX file. This is intended for hole-punch operations. If
1094 * you are truncating a file, the helper function dax_truncate_page() may be
1095 * more convenient.
1096 *
1097 * We work in terms of PAGE_SIZE here for commonality with
1098 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1099 * took care of disposing of the unnecessary blocks. Even if the filesystem
1100 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1101 * since the file might be mmapped.
1102 */
1103int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1104 get_block_t get_block)
1105{
1106 struct buffer_head bh;
1107 pgoff_t index = from >> PAGE_SHIFT;
1108 unsigned offset = from & (PAGE_SIZE-1);
1109 int err;
1110
1111 /* Block boundary? Nothing to do */
1112 if (!length)
1113 return 0;
1114 BUG_ON((offset + length) > PAGE_SIZE);
1115
1116 memset(&bh, 0, sizeof(bh));
1117 bh.b_bdev = inode->i_sb->s_bdev;
1118 bh.b_size = PAGE_SIZE;
1119 err = get_block(inode, index, &bh, 0);
1120 if (err < 0)
1121 return err;
1122 if (buffer_written(&bh)) {
1123 struct block_device *bdev = bh.b_bdev;
1124 struct blk_dax_ctl dax = {
1125 .sector = to_sector(&bh, inode),
1126 .size = PAGE_SIZE,
1127 };
1128
1129 if (dax_map_atomic(bdev, &dax) < 0)
1130 return PTR_ERR(dax.addr);
1131 clear_pmem(dax.addr + offset, length);
1132 wmb_pmem();
1133 dax_unmap_atomic(bdev, &dax);
1134 }
1135
1136 return 0;
1137}
1138EXPORT_SYMBOL_GPL(dax_zero_page_range);
1139
1140/**
1141 * dax_truncate_page - handle a partial page being truncated in a DAX file
1142 * @inode: The file being truncated
1143 * @from: The file offset that is being truncated to
1144 * @get_block: The filesystem method used to translate file offsets to blocks
1145 *
1146 * Similar to block_truncate_page(), this function can be called by a
1147 * filesystem when it is truncating a DAX file to handle the partial page.
1148 *
1149 * We work in terms of PAGE_SIZE here for commonality with
1150 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1151 * took care of disposing of the unnecessary blocks. Even if the filesystem
1152 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1153 * since the file might be mmapped.
1154 */
1155int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1156{
1157 unsigned length = PAGE_ALIGN(from) - from;
1158 return dax_zero_page_range(inode, from, length, get_block);
1159}
1160EXPORT_SYMBOL_GPL(dax_truncate_page);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * fs/dax.c - Direct Access filesystem code
4 * Copyright (c) 2013-2014 Intel Corporation
5 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
6 * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
7 */
8
9#include <linux/atomic.h>
10#include <linux/blkdev.h>
11#include <linux/buffer_head.h>
12#include <linux/dax.h>
13#include <linux/fs.h>
14#include <linux/highmem.h>
15#include <linux/memcontrol.h>
16#include <linux/mm.h>
17#include <linux/mutex.h>
18#include <linux/pagevec.h>
19#include <linux/sched.h>
20#include <linux/sched/signal.h>
21#include <linux/uio.h>
22#include <linux/vmstat.h>
23#include <linux/pfn_t.h>
24#include <linux/sizes.h>
25#include <linux/mmu_notifier.h>
26#include <linux/iomap.h>
27#include <linux/rmap.h>
28#include <asm/pgalloc.h>
29
30#define CREATE_TRACE_POINTS
31#include <trace/events/fs_dax.h>
32
33/* We choose 4096 entries - same as per-zone page wait tables */
34#define DAX_WAIT_TABLE_BITS 12
35#define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
36
37/* The 'colour' (ie low bits) within a PMD of a page offset. */
38#define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
39#define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT)
40
41static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
42
43static int __init init_dax_wait_table(void)
44{
45 int i;
46
47 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
48 init_waitqueue_head(wait_table + i);
49 return 0;
50}
51fs_initcall(init_dax_wait_table);
52
53/*
54 * DAX pagecache entries use XArray value entries so they can't be mistaken
55 * for pages. We use one bit for locking, one bit for the entry size (PMD)
56 * and two more to tell us if the entry is a zero page or an empty entry that
57 * is just used for locking. In total four special bits.
58 *
59 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
60 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
61 * block allocation.
62 */
63#define DAX_SHIFT (4)
64#define DAX_LOCKED (1UL << 0)
65#define DAX_PMD (1UL << 1)
66#define DAX_ZERO_PAGE (1UL << 2)
67#define DAX_EMPTY (1UL << 3)
68
69static unsigned long dax_to_pfn(void *entry)
70{
71 return xa_to_value(entry) >> DAX_SHIFT;
72}
73
74static void *dax_make_entry(pfn_t pfn, unsigned long flags)
75{
76 return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
77}
78
79static bool dax_is_locked(void *entry)
80{
81 return xa_to_value(entry) & DAX_LOCKED;
82}
83
84static unsigned int dax_entry_order(void *entry)
85{
86 if (xa_to_value(entry) & DAX_PMD)
87 return PMD_ORDER;
88 return 0;
89}
90
91static unsigned long dax_is_pmd_entry(void *entry)
92{
93 return xa_to_value(entry) & DAX_PMD;
94}
95
96static bool dax_is_pte_entry(void *entry)
97{
98 return !(xa_to_value(entry) & DAX_PMD);
99}
100
101static int dax_is_zero_entry(void *entry)
102{
103 return xa_to_value(entry) & DAX_ZERO_PAGE;
104}
105
106static int dax_is_empty_entry(void *entry)
107{
108 return xa_to_value(entry) & DAX_EMPTY;
109}
110
111/*
112 * true if the entry that was found is of a smaller order than the entry
113 * we were looking for
114 */
115static bool dax_is_conflict(void *entry)
116{
117 return entry == XA_RETRY_ENTRY;
118}
119
120/*
121 * DAX page cache entry locking
122 */
123struct exceptional_entry_key {
124 struct xarray *xa;
125 pgoff_t entry_start;
126};
127
128struct wait_exceptional_entry_queue {
129 wait_queue_entry_t wait;
130 struct exceptional_entry_key key;
131};
132
133/**
134 * enum dax_wake_mode: waitqueue wakeup behaviour
135 * @WAKE_ALL: wake all waiters in the waitqueue
136 * @WAKE_NEXT: wake only the first waiter in the waitqueue
137 */
138enum dax_wake_mode {
139 WAKE_ALL,
140 WAKE_NEXT,
141};
142
143static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas,
144 void *entry, struct exceptional_entry_key *key)
145{
146 unsigned long hash;
147 unsigned long index = xas->xa_index;
148
149 /*
150 * If 'entry' is a PMD, align the 'index' that we use for the wait
151 * queue to the start of that PMD. This ensures that all offsets in
152 * the range covered by the PMD map to the same bit lock.
153 */
154 if (dax_is_pmd_entry(entry))
155 index &= ~PG_PMD_COLOUR;
156 key->xa = xas->xa;
157 key->entry_start = index;
158
159 hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS);
160 return wait_table + hash;
161}
162
163static int wake_exceptional_entry_func(wait_queue_entry_t *wait,
164 unsigned int mode, int sync, void *keyp)
165{
166 struct exceptional_entry_key *key = keyp;
167 struct wait_exceptional_entry_queue *ewait =
168 container_of(wait, struct wait_exceptional_entry_queue, wait);
169
170 if (key->xa != ewait->key.xa ||
171 key->entry_start != ewait->key.entry_start)
172 return 0;
173 return autoremove_wake_function(wait, mode, sync, NULL);
174}
175
176/*
177 * @entry may no longer be the entry at the index in the mapping.
178 * The important information it's conveying is whether the entry at
179 * this index used to be a PMD entry.
180 */
181static void dax_wake_entry(struct xa_state *xas, void *entry,
182 enum dax_wake_mode mode)
183{
184 struct exceptional_entry_key key;
185 wait_queue_head_t *wq;
186
187 wq = dax_entry_waitqueue(xas, entry, &key);
188
189 /*
190 * Checking for locked entry and prepare_to_wait_exclusive() happens
191 * under the i_pages lock, ditto for entry handling in our callers.
192 * So at this point all tasks that could have seen our entry locked
193 * must be in the waitqueue and the following check will see them.
194 */
195 if (waitqueue_active(wq))
196 __wake_up(wq, TASK_NORMAL, mode == WAKE_ALL ? 0 : 1, &key);
197}
198
199/*
200 * Look up entry in page cache, wait for it to become unlocked if it
201 * is a DAX entry and return it. The caller must subsequently call
202 * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
203 * if it did. The entry returned may have a larger order than @order.
204 * If @order is larger than the order of the entry found in i_pages, this
205 * function returns a dax_is_conflict entry.
206 *
207 * Must be called with the i_pages lock held.
208 */
209static void *get_unlocked_entry(struct xa_state *xas, unsigned int order)
210{
211 void *entry;
212 struct wait_exceptional_entry_queue ewait;
213 wait_queue_head_t *wq;
214
215 init_wait(&ewait.wait);
216 ewait.wait.func = wake_exceptional_entry_func;
217
218 for (;;) {
219 entry = xas_find_conflict(xas);
220 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
221 return entry;
222 if (dax_entry_order(entry) < order)
223 return XA_RETRY_ENTRY;
224 if (!dax_is_locked(entry))
225 return entry;
226
227 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
228 prepare_to_wait_exclusive(wq, &ewait.wait,
229 TASK_UNINTERRUPTIBLE);
230 xas_unlock_irq(xas);
231 xas_reset(xas);
232 schedule();
233 finish_wait(wq, &ewait.wait);
234 xas_lock_irq(xas);
235 }
236}
237
238/*
239 * The only thing keeping the address space around is the i_pages lock
240 * (it's cycled in clear_inode() after removing the entries from i_pages)
241 * After we call xas_unlock_irq(), we cannot touch xas->xa.
242 */
243static void wait_entry_unlocked(struct xa_state *xas, void *entry)
244{
245 struct wait_exceptional_entry_queue ewait;
246 wait_queue_head_t *wq;
247
248 init_wait(&ewait.wait);
249 ewait.wait.func = wake_exceptional_entry_func;
250
251 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
252 /*
253 * Unlike get_unlocked_entry() there is no guarantee that this
254 * path ever successfully retrieves an unlocked entry before an
255 * inode dies. Perform a non-exclusive wait in case this path
256 * never successfully performs its own wake up.
257 */
258 prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
259 xas_unlock_irq(xas);
260 schedule();
261 finish_wait(wq, &ewait.wait);
262}
263
264static void put_unlocked_entry(struct xa_state *xas, void *entry,
265 enum dax_wake_mode mode)
266{
267 if (entry && !dax_is_conflict(entry))
268 dax_wake_entry(xas, entry, mode);
269}
270
271/*
272 * We used the xa_state to get the entry, but then we locked the entry and
273 * dropped the xa_lock, so we know the xa_state is stale and must be reset
274 * before use.
275 */
276static void dax_unlock_entry(struct xa_state *xas, void *entry)
277{
278 void *old;
279
280 BUG_ON(dax_is_locked(entry));
281 xas_reset(xas);
282 xas_lock_irq(xas);
283 old = xas_store(xas, entry);
284 xas_unlock_irq(xas);
285 BUG_ON(!dax_is_locked(old));
286 dax_wake_entry(xas, entry, WAKE_NEXT);
287}
288
289/*
290 * Return: The entry stored at this location before it was locked.
291 */
292static void *dax_lock_entry(struct xa_state *xas, void *entry)
293{
294 unsigned long v = xa_to_value(entry);
295 return xas_store(xas, xa_mk_value(v | DAX_LOCKED));
296}
297
298static unsigned long dax_entry_size(void *entry)
299{
300 if (dax_is_zero_entry(entry))
301 return 0;
302 else if (dax_is_empty_entry(entry))
303 return 0;
304 else if (dax_is_pmd_entry(entry))
305 return PMD_SIZE;
306 else
307 return PAGE_SIZE;
308}
309
310static unsigned long dax_end_pfn(void *entry)
311{
312 return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
313}
314
315/*
316 * Iterate through all mapped pfns represented by an entry, i.e. skip
317 * 'empty' and 'zero' entries.
318 */
319#define for_each_mapped_pfn(entry, pfn) \
320 for (pfn = dax_to_pfn(entry); \
321 pfn < dax_end_pfn(entry); pfn++)
322
323static inline bool dax_page_is_shared(struct page *page)
324{
325 return page->mapping == PAGE_MAPPING_DAX_SHARED;
326}
327
328/*
329 * Set the page->mapping with PAGE_MAPPING_DAX_SHARED flag, increase the
330 * refcount.
331 */
332static inline void dax_page_share_get(struct page *page)
333{
334 if (page->mapping != PAGE_MAPPING_DAX_SHARED) {
335 /*
336 * Reset the index if the page was already mapped
337 * regularly before.
338 */
339 if (page->mapping)
340 page->share = 1;
341 page->mapping = PAGE_MAPPING_DAX_SHARED;
342 }
343 page->share++;
344}
345
346static inline unsigned long dax_page_share_put(struct page *page)
347{
348 return --page->share;
349}
350
351/*
352 * When it is called in dax_insert_entry(), the shared flag will indicate that
353 * whether this entry is shared by multiple files. If so, set the page->mapping
354 * PAGE_MAPPING_DAX_SHARED, and use page->share as refcount.
355 */
356static void dax_associate_entry(void *entry, struct address_space *mapping,
357 struct vm_area_struct *vma, unsigned long address, bool shared)
358{
359 unsigned long size = dax_entry_size(entry), pfn, index;
360 int i = 0;
361
362 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
363 return;
364
365 index = linear_page_index(vma, address & ~(size - 1));
366 for_each_mapped_pfn(entry, pfn) {
367 struct page *page = pfn_to_page(pfn);
368
369 if (shared) {
370 dax_page_share_get(page);
371 } else {
372 WARN_ON_ONCE(page->mapping);
373 page->mapping = mapping;
374 page->index = index + i++;
375 }
376 }
377}
378
379static void dax_disassociate_entry(void *entry, struct address_space *mapping,
380 bool trunc)
381{
382 unsigned long pfn;
383
384 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
385 return;
386
387 for_each_mapped_pfn(entry, pfn) {
388 struct page *page = pfn_to_page(pfn);
389
390 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
391 if (dax_page_is_shared(page)) {
392 /* keep the shared flag if this page is still shared */
393 if (dax_page_share_put(page) > 0)
394 continue;
395 } else
396 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
397 page->mapping = NULL;
398 page->index = 0;
399 }
400}
401
402static struct page *dax_busy_page(void *entry)
403{
404 unsigned long pfn;
405
406 for_each_mapped_pfn(entry, pfn) {
407 struct page *page = pfn_to_page(pfn);
408
409 if (page_ref_count(page) > 1)
410 return page;
411 }
412 return NULL;
413}
414
415/**
416 * dax_lock_folio - Lock the DAX entry corresponding to a folio
417 * @folio: The folio whose entry we want to lock
418 *
419 * Context: Process context.
420 * Return: A cookie to pass to dax_unlock_folio() or 0 if the entry could
421 * not be locked.
422 */
423dax_entry_t dax_lock_folio(struct folio *folio)
424{
425 XA_STATE(xas, NULL, 0);
426 void *entry;
427
428 /* Ensure folio->mapping isn't freed while we look at it */
429 rcu_read_lock();
430 for (;;) {
431 struct address_space *mapping = READ_ONCE(folio->mapping);
432
433 entry = NULL;
434 if (!mapping || !dax_mapping(mapping))
435 break;
436
437 /*
438 * In the device-dax case there's no need to lock, a
439 * struct dev_pagemap pin is sufficient to keep the
440 * inode alive, and we assume we have dev_pagemap pin
441 * otherwise we would not have a valid pfn_to_page()
442 * translation.
443 */
444 entry = (void *)~0UL;
445 if (S_ISCHR(mapping->host->i_mode))
446 break;
447
448 xas.xa = &mapping->i_pages;
449 xas_lock_irq(&xas);
450 if (mapping != folio->mapping) {
451 xas_unlock_irq(&xas);
452 continue;
453 }
454 xas_set(&xas, folio->index);
455 entry = xas_load(&xas);
456 if (dax_is_locked(entry)) {
457 rcu_read_unlock();
458 wait_entry_unlocked(&xas, entry);
459 rcu_read_lock();
460 continue;
461 }
462 dax_lock_entry(&xas, entry);
463 xas_unlock_irq(&xas);
464 break;
465 }
466 rcu_read_unlock();
467 return (dax_entry_t)entry;
468}
469
470void dax_unlock_folio(struct folio *folio, dax_entry_t cookie)
471{
472 struct address_space *mapping = folio->mapping;
473 XA_STATE(xas, &mapping->i_pages, folio->index);
474
475 if (S_ISCHR(mapping->host->i_mode))
476 return;
477
478 dax_unlock_entry(&xas, (void *)cookie);
479}
480
481/*
482 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a mapping
483 * @mapping: the file's mapping whose entry we want to lock
484 * @index: the offset within this file
485 * @page: output the dax page corresponding to this dax entry
486 *
487 * Return: A cookie to pass to dax_unlock_mapping_entry() or 0 if the entry
488 * could not be locked.
489 */
490dax_entry_t dax_lock_mapping_entry(struct address_space *mapping, pgoff_t index,
491 struct page **page)
492{
493 XA_STATE(xas, NULL, 0);
494 void *entry;
495
496 rcu_read_lock();
497 for (;;) {
498 entry = NULL;
499 if (!dax_mapping(mapping))
500 break;
501
502 xas.xa = &mapping->i_pages;
503 xas_lock_irq(&xas);
504 xas_set(&xas, index);
505 entry = xas_load(&xas);
506 if (dax_is_locked(entry)) {
507 rcu_read_unlock();
508 wait_entry_unlocked(&xas, entry);
509 rcu_read_lock();
510 continue;
511 }
512 if (!entry ||
513 dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
514 /*
515 * Because we are looking for entry from file's mapping
516 * and index, so the entry may not be inserted for now,
517 * or even a zero/empty entry. We don't think this is
518 * an error case. So, return a special value and do
519 * not output @page.
520 */
521 entry = (void *)~0UL;
522 } else {
523 *page = pfn_to_page(dax_to_pfn(entry));
524 dax_lock_entry(&xas, entry);
525 }
526 xas_unlock_irq(&xas);
527 break;
528 }
529 rcu_read_unlock();
530 return (dax_entry_t)entry;
531}
532
533void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index,
534 dax_entry_t cookie)
535{
536 XA_STATE(xas, &mapping->i_pages, index);
537
538 if (cookie == ~0UL)
539 return;
540
541 dax_unlock_entry(&xas, (void *)cookie);
542}
543
544/*
545 * Find page cache entry at given index. If it is a DAX entry, return it
546 * with the entry locked. If the page cache doesn't contain an entry at
547 * that index, add a locked empty entry.
548 *
549 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
550 * either return that locked entry or will return VM_FAULT_FALLBACK.
551 * This will happen if there are any PTE entries within the PMD range
552 * that we are requesting.
553 *
554 * We always favor PTE entries over PMD entries. There isn't a flow where we
555 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD
556 * insertion will fail if it finds any PTE entries already in the tree, and a
557 * PTE insertion will cause an existing PMD entry to be unmapped and
558 * downgraded to PTE entries. This happens for both PMD zero pages as
559 * well as PMD empty entries.
560 *
561 * The exception to this downgrade path is for PMD entries that have
562 * real storage backing them. We will leave these real PMD entries in
563 * the tree, and PTE writes will simply dirty the entire PMD entry.
564 *
565 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
566 * persistent memory the benefit is doubtful. We can add that later if we can
567 * show it helps.
568 *
569 * On error, this function does not return an ERR_PTR. Instead it returns
570 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values
571 * overlap with xarray value entries.
572 */
573static void *grab_mapping_entry(struct xa_state *xas,
574 struct address_space *mapping, unsigned int order)
575{
576 unsigned long index = xas->xa_index;
577 bool pmd_downgrade; /* splitting PMD entry into PTE entries? */
578 void *entry;
579
580retry:
581 pmd_downgrade = false;
582 xas_lock_irq(xas);
583 entry = get_unlocked_entry(xas, order);
584
585 if (entry) {
586 if (dax_is_conflict(entry))
587 goto fallback;
588 if (!xa_is_value(entry)) {
589 xas_set_err(xas, -EIO);
590 goto out_unlock;
591 }
592
593 if (order == 0) {
594 if (dax_is_pmd_entry(entry) &&
595 (dax_is_zero_entry(entry) ||
596 dax_is_empty_entry(entry))) {
597 pmd_downgrade = true;
598 }
599 }
600 }
601
602 if (pmd_downgrade) {
603 /*
604 * Make sure 'entry' remains valid while we drop
605 * the i_pages lock.
606 */
607 dax_lock_entry(xas, entry);
608
609 /*
610 * Besides huge zero pages the only other thing that gets
611 * downgraded are empty entries which don't need to be
612 * unmapped.
613 */
614 if (dax_is_zero_entry(entry)) {
615 xas_unlock_irq(xas);
616 unmap_mapping_pages(mapping,
617 xas->xa_index & ~PG_PMD_COLOUR,
618 PG_PMD_NR, false);
619 xas_reset(xas);
620 xas_lock_irq(xas);
621 }
622
623 dax_disassociate_entry(entry, mapping, false);
624 xas_store(xas, NULL); /* undo the PMD join */
625 dax_wake_entry(xas, entry, WAKE_ALL);
626 mapping->nrpages -= PG_PMD_NR;
627 entry = NULL;
628 xas_set(xas, index);
629 }
630
631 if (entry) {
632 dax_lock_entry(xas, entry);
633 } else {
634 unsigned long flags = DAX_EMPTY;
635
636 if (order > 0)
637 flags |= DAX_PMD;
638 entry = dax_make_entry(pfn_to_pfn_t(0), flags);
639 dax_lock_entry(xas, entry);
640 if (xas_error(xas))
641 goto out_unlock;
642 mapping->nrpages += 1UL << order;
643 }
644
645out_unlock:
646 xas_unlock_irq(xas);
647 if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM))
648 goto retry;
649 if (xas->xa_node == XA_ERROR(-ENOMEM))
650 return xa_mk_internal(VM_FAULT_OOM);
651 if (xas_error(xas))
652 return xa_mk_internal(VM_FAULT_SIGBUS);
653 return entry;
654fallback:
655 xas_unlock_irq(xas);
656 return xa_mk_internal(VM_FAULT_FALLBACK);
657}
658
659/**
660 * dax_layout_busy_page_range - find first pinned page in @mapping
661 * @mapping: address space to scan for a page with ref count > 1
662 * @start: Starting offset. Page containing 'start' is included.
663 * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX,
664 * pages from 'start' till the end of file are included.
665 *
666 * DAX requires ZONE_DEVICE mapped pages. These pages are never
667 * 'onlined' to the page allocator so they are considered idle when
668 * page->count == 1. A filesystem uses this interface to determine if
669 * any page in the mapping is busy, i.e. for DMA, or other
670 * get_user_pages() usages.
671 *
672 * It is expected that the filesystem is holding locks to block the
673 * establishment of new mappings in this address_space. I.e. it expects
674 * to be able to run unmap_mapping_range() and subsequently not race
675 * mapping_mapped() becoming true.
676 */
677struct page *dax_layout_busy_page_range(struct address_space *mapping,
678 loff_t start, loff_t end)
679{
680 void *entry;
681 unsigned int scanned = 0;
682 struct page *page = NULL;
683 pgoff_t start_idx = start >> PAGE_SHIFT;
684 pgoff_t end_idx;
685 XA_STATE(xas, &mapping->i_pages, start_idx);
686
687 /*
688 * In the 'limited' case get_user_pages() for dax is disabled.
689 */
690 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
691 return NULL;
692
693 if (!dax_mapping(mapping) || !mapping_mapped(mapping))
694 return NULL;
695
696 /* If end == LLONG_MAX, all pages from start to till end of file */
697 if (end == LLONG_MAX)
698 end_idx = ULONG_MAX;
699 else
700 end_idx = end >> PAGE_SHIFT;
701 /*
702 * If we race get_user_pages_fast() here either we'll see the
703 * elevated page count in the iteration and wait, or
704 * get_user_pages_fast() will see that the page it took a reference
705 * against is no longer mapped in the page tables and bail to the
706 * get_user_pages() slow path. The slow path is protected by
707 * pte_lock() and pmd_lock(). New references are not taken without
708 * holding those locks, and unmap_mapping_pages() will not zero the
709 * pte or pmd without holding the respective lock, so we are
710 * guaranteed to either see new references or prevent new
711 * references from being established.
712 */
713 unmap_mapping_pages(mapping, start_idx, end_idx - start_idx + 1, 0);
714
715 xas_lock_irq(&xas);
716 xas_for_each(&xas, entry, end_idx) {
717 if (WARN_ON_ONCE(!xa_is_value(entry)))
718 continue;
719 if (unlikely(dax_is_locked(entry)))
720 entry = get_unlocked_entry(&xas, 0);
721 if (entry)
722 page = dax_busy_page(entry);
723 put_unlocked_entry(&xas, entry, WAKE_NEXT);
724 if (page)
725 break;
726 if (++scanned % XA_CHECK_SCHED)
727 continue;
728
729 xas_pause(&xas);
730 xas_unlock_irq(&xas);
731 cond_resched();
732 xas_lock_irq(&xas);
733 }
734 xas_unlock_irq(&xas);
735 return page;
736}
737EXPORT_SYMBOL_GPL(dax_layout_busy_page_range);
738
739struct page *dax_layout_busy_page(struct address_space *mapping)
740{
741 return dax_layout_busy_page_range(mapping, 0, LLONG_MAX);
742}
743EXPORT_SYMBOL_GPL(dax_layout_busy_page);
744
745static int __dax_invalidate_entry(struct address_space *mapping,
746 pgoff_t index, bool trunc)
747{
748 XA_STATE(xas, &mapping->i_pages, index);
749 int ret = 0;
750 void *entry;
751
752 xas_lock_irq(&xas);
753 entry = get_unlocked_entry(&xas, 0);
754 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
755 goto out;
756 if (!trunc &&
757 (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) ||
758 xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE)))
759 goto out;
760 dax_disassociate_entry(entry, mapping, trunc);
761 xas_store(&xas, NULL);
762 mapping->nrpages -= 1UL << dax_entry_order(entry);
763 ret = 1;
764out:
765 put_unlocked_entry(&xas, entry, WAKE_ALL);
766 xas_unlock_irq(&xas);
767 return ret;
768}
769
770static int __dax_clear_dirty_range(struct address_space *mapping,
771 pgoff_t start, pgoff_t end)
772{
773 XA_STATE(xas, &mapping->i_pages, start);
774 unsigned int scanned = 0;
775 void *entry;
776
777 xas_lock_irq(&xas);
778 xas_for_each(&xas, entry, end) {
779 entry = get_unlocked_entry(&xas, 0);
780 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
781 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
782 put_unlocked_entry(&xas, entry, WAKE_NEXT);
783
784 if (++scanned % XA_CHECK_SCHED)
785 continue;
786
787 xas_pause(&xas);
788 xas_unlock_irq(&xas);
789 cond_resched();
790 xas_lock_irq(&xas);
791 }
792 xas_unlock_irq(&xas);
793
794 return 0;
795}
796
797/*
798 * Delete DAX entry at @index from @mapping. Wait for it
799 * to be unlocked before deleting it.
800 */
801int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
802{
803 int ret = __dax_invalidate_entry(mapping, index, true);
804
805 /*
806 * This gets called from truncate / punch_hole path. As such, the caller
807 * must hold locks protecting against concurrent modifications of the
808 * page cache (usually fs-private i_mmap_sem for writing). Since the
809 * caller has seen a DAX entry for this index, we better find it
810 * at that index as well...
811 */
812 WARN_ON_ONCE(!ret);
813 return ret;
814}
815
816/*
817 * Invalidate DAX entry if it is clean.
818 */
819int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
820 pgoff_t index)
821{
822 return __dax_invalidate_entry(mapping, index, false);
823}
824
825static pgoff_t dax_iomap_pgoff(const struct iomap *iomap, loff_t pos)
826{
827 return PHYS_PFN(iomap->addr + (pos & PAGE_MASK) - iomap->offset);
828}
829
830static int copy_cow_page_dax(struct vm_fault *vmf, const struct iomap_iter *iter)
831{
832 pgoff_t pgoff = dax_iomap_pgoff(&iter->iomap, iter->pos);
833 void *vto, *kaddr;
834 long rc;
835 int id;
836
837 id = dax_read_lock();
838 rc = dax_direct_access(iter->iomap.dax_dev, pgoff, 1, DAX_ACCESS,
839 &kaddr, NULL);
840 if (rc < 0) {
841 dax_read_unlock(id);
842 return rc;
843 }
844 vto = kmap_atomic(vmf->cow_page);
845 copy_user_page(vto, kaddr, vmf->address, vmf->cow_page);
846 kunmap_atomic(vto);
847 dax_read_unlock(id);
848 return 0;
849}
850
851/*
852 * MAP_SYNC on a dax mapping guarantees dirty metadata is
853 * flushed on write-faults (non-cow), but not read-faults.
854 */
855static bool dax_fault_is_synchronous(const struct iomap_iter *iter,
856 struct vm_area_struct *vma)
857{
858 return (iter->flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) &&
859 (iter->iomap.flags & IOMAP_F_DIRTY);
860}
861
862/*
863 * By this point grab_mapping_entry() has ensured that we have a locked entry
864 * of the appropriate size so we don't have to worry about downgrading PMDs to
865 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
866 * already in the tree, we will skip the insertion and just dirty the PMD as
867 * appropriate.
868 */
869static void *dax_insert_entry(struct xa_state *xas, struct vm_fault *vmf,
870 const struct iomap_iter *iter, void *entry, pfn_t pfn,
871 unsigned long flags)
872{
873 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
874 void *new_entry = dax_make_entry(pfn, flags);
875 bool write = iter->flags & IOMAP_WRITE;
876 bool dirty = write && !dax_fault_is_synchronous(iter, vmf->vma);
877 bool shared = iter->iomap.flags & IOMAP_F_SHARED;
878
879 if (dirty)
880 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
881
882 if (shared || (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE))) {
883 unsigned long index = xas->xa_index;
884 /* we are replacing a zero page with block mapping */
885 if (dax_is_pmd_entry(entry))
886 unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
887 PG_PMD_NR, false);
888 else /* pte entry */
889 unmap_mapping_pages(mapping, index, 1, false);
890 }
891
892 xas_reset(xas);
893 xas_lock_irq(xas);
894 if (shared || dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
895 void *old;
896
897 dax_disassociate_entry(entry, mapping, false);
898 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address,
899 shared);
900 /*
901 * Only swap our new entry into the page cache if the current
902 * entry is a zero page or an empty entry. If a normal PTE or
903 * PMD entry is already in the cache, we leave it alone. This
904 * means that if we are trying to insert a PTE and the
905 * existing entry is a PMD, we will just leave the PMD in the
906 * tree and dirty it if necessary.
907 */
908 old = dax_lock_entry(xas, new_entry);
909 WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) |
910 DAX_LOCKED));
911 entry = new_entry;
912 } else {
913 xas_load(xas); /* Walk the xa_state */
914 }
915
916 if (dirty)
917 xas_set_mark(xas, PAGECACHE_TAG_DIRTY);
918
919 if (write && shared)
920 xas_set_mark(xas, PAGECACHE_TAG_TOWRITE);
921
922 xas_unlock_irq(xas);
923 return entry;
924}
925
926static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev,
927 struct address_space *mapping, void *entry)
928{
929 unsigned long pfn, index, count, end;
930 long ret = 0;
931 struct vm_area_struct *vma;
932
933 /*
934 * A page got tagged dirty in DAX mapping? Something is seriously
935 * wrong.
936 */
937 if (WARN_ON(!xa_is_value(entry)))
938 return -EIO;
939
940 if (unlikely(dax_is_locked(entry))) {
941 void *old_entry = entry;
942
943 entry = get_unlocked_entry(xas, 0);
944
945 /* Entry got punched out / reallocated? */
946 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
947 goto put_unlocked;
948 /*
949 * Entry got reallocated elsewhere? No need to writeback.
950 * We have to compare pfns as we must not bail out due to
951 * difference in lockbit or entry type.
952 */
953 if (dax_to_pfn(old_entry) != dax_to_pfn(entry))
954 goto put_unlocked;
955 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
956 dax_is_zero_entry(entry))) {
957 ret = -EIO;
958 goto put_unlocked;
959 }
960
961 /* Another fsync thread may have already done this entry */
962 if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE))
963 goto put_unlocked;
964 }
965
966 /* Lock the entry to serialize with page faults */
967 dax_lock_entry(xas, entry);
968
969 /*
970 * We can clear the tag now but we have to be careful so that concurrent
971 * dax_writeback_one() calls for the same index cannot finish before we
972 * actually flush the caches. This is achieved as the calls will look
973 * at the entry only under the i_pages lock and once they do that
974 * they will see the entry locked and wait for it to unlock.
975 */
976 xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE);
977 xas_unlock_irq(xas);
978
979 /*
980 * If dax_writeback_mapping_range() was given a wbc->range_start
981 * in the middle of a PMD, the 'index' we use needs to be
982 * aligned to the start of the PMD.
983 * This allows us to flush for PMD_SIZE and not have to worry about
984 * partial PMD writebacks.
985 */
986 pfn = dax_to_pfn(entry);
987 count = 1UL << dax_entry_order(entry);
988 index = xas->xa_index & ~(count - 1);
989 end = index + count - 1;
990
991 /* Walk all mappings of a given index of a file and writeprotect them */
992 i_mmap_lock_read(mapping);
993 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, end) {
994 pfn_mkclean_range(pfn, count, index, vma);
995 cond_resched();
996 }
997 i_mmap_unlock_read(mapping);
998
999 dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE);
1000 /*
1001 * After we have flushed the cache, we can clear the dirty tag. There
1002 * cannot be new dirty data in the pfn after the flush has completed as
1003 * the pfn mappings are writeprotected and fault waits for mapping
1004 * entry lock.
1005 */
1006 xas_reset(xas);
1007 xas_lock_irq(xas);
1008 xas_store(xas, entry);
1009 xas_clear_mark(xas, PAGECACHE_TAG_DIRTY);
1010 dax_wake_entry(xas, entry, WAKE_NEXT);
1011
1012 trace_dax_writeback_one(mapping->host, index, count);
1013 return ret;
1014
1015 put_unlocked:
1016 put_unlocked_entry(xas, entry, WAKE_NEXT);
1017 return ret;
1018}
1019
1020/*
1021 * Flush the mapping to the persistent domain within the byte range of [start,
1022 * end]. This is required by data integrity operations to ensure file data is
1023 * on persistent storage prior to completion of the operation.
1024 */
1025int dax_writeback_mapping_range(struct address_space *mapping,
1026 struct dax_device *dax_dev, struct writeback_control *wbc)
1027{
1028 XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT);
1029 struct inode *inode = mapping->host;
1030 pgoff_t end_index = wbc->range_end >> PAGE_SHIFT;
1031 void *entry;
1032 int ret = 0;
1033 unsigned int scanned = 0;
1034
1035 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
1036 return -EIO;
1037
1038 if (mapping_empty(mapping) || wbc->sync_mode != WB_SYNC_ALL)
1039 return 0;
1040
1041 trace_dax_writeback_range(inode, xas.xa_index, end_index);
1042
1043 tag_pages_for_writeback(mapping, xas.xa_index, end_index);
1044
1045 xas_lock_irq(&xas);
1046 xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) {
1047 ret = dax_writeback_one(&xas, dax_dev, mapping, entry);
1048 if (ret < 0) {
1049 mapping_set_error(mapping, ret);
1050 break;
1051 }
1052 if (++scanned % XA_CHECK_SCHED)
1053 continue;
1054
1055 xas_pause(&xas);
1056 xas_unlock_irq(&xas);
1057 cond_resched();
1058 xas_lock_irq(&xas);
1059 }
1060 xas_unlock_irq(&xas);
1061 trace_dax_writeback_range_done(inode, xas.xa_index, end_index);
1062 return ret;
1063}
1064EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
1065
1066static int dax_iomap_direct_access(const struct iomap *iomap, loff_t pos,
1067 size_t size, void **kaddr, pfn_t *pfnp)
1068{
1069 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1070 int id, rc = 0;
1071 long length;
1072
1073 id = dax_read_lock();
1074 length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
1075 DAX_ACCESS, kaddr, pfnp);
1076 if (length < 0) {
1077 rc = length;
1078 goto out;
1079 }
1080 if (!pfnp)
1081 goto out_check_addr;
1082 rc = -EINVAL;
1083 if (PFN_PHYS(length) < size)
1084 goto out;
1085 if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1086 goto out;
1087 /* For larger pages we need devmap */
1088 if (length > 1 && !pfn_t_devmap(*pfnp))
1089 goto out;
1090 rc = 0;
1091
1092out_check_addr:
1093 if (!kaddr)
1094 goto out;
1095 if (!*kaddr)
1096 rc = -EFAULT;
1097out:
1098 dax_read_unlock(id);
1099 return rc;
1100}
1101
1102/**
1103 * dax_iomap_copy_around - Prepare for an unaligned write to a shared/cow page
1104 * by copying the data before and after the range to be written.
1105 * @pos: address to do copy from.
1106 * @length: size of copy operation.
1107 * @align_size: aligned w.r.t align_size (either PMD_SIZE or PAGE_SIZE)
1108 * @srcmap: iomap srcmap
1109 * @daddr: destination address to copy to.
1110 *
1111 * This can be called from two places. Either during DAX write fault (page
1112 * aligned), to copy the length size data to daddr. Or, while doing normal DAX
1113 * write operation, dax_iomap_iter() might call this to do the copy of either
1114 * start or end unaligned address. In the latter case the rest of the copy of
1115 * aligned ranges is taken care by dax_iomap_iter() itself.
1116 * If the srcmap contains invalid data, such as HOLE and UNWRITTEN, zero the
1117 * area to make sure no old data remains.
1118 */
1119static int dax_iomap_copy_around(loff_t pos, uint64_t length, size_t align_size,
1120 const struct iomap *srcmap, void *daddr)
1121{
1122 loff_t head_off = pos & (align_size - 1);
1123 size_t size = ALIGN(head_off + length, align_size);
1124 loff_t end = pos + length;
1125 loff_t pg_end = round_up(end, align_size);
1126 /* copy_all is usually in page fault case */
1127 bool copy_all = head_off == 0 && end == pg_end;
1128 /* zero the edges if srcmap is a HOLE or IOMAP_UNWRITTEN */
1129 bool zero_edge = srcmap->flags & IOMAP_F_SHARED ||
1130 srcmap->type == IOMAP_UNWRITTEN;
1131 void *saddr = NULL;
1132 int ret = 0;
1133
1134 if (!zero_edge) {
1135 ret = dax_iomap_direct_access(srcmap, pos, size, &saddr, NULL);
1136 if (ret)
1137 return dax_mem2blk_err(ret);
1138 }
1139
1140 if (copy_all) {
1141 if (zero_edge)
1142 memset(daddr, 0, size);
1143 else
1144 ret = copy_mc_to_kernel(daddr, saddr, length);
1145 goto out;
1146 }
1147
1148 /* Copy the head part of the range */
1149 if (head_off) {
1150 if (zero_edge)
1151 memset(daddr, 0, head_off);
1152 else {
1153 ret = copy_mc_to_kernel(daddr, saddr, head_off);
1154 if (ret)
1155 return -EIO;
1156 }
1157 }
1158
1159 /* Copy the tail part of the range */
1160 if (end < pg_end) {
1161 loff_t tail_off = head_off + length;
1162 loff_t tail_len = pg_end - end;
1163
1164 if (zero_edge)
1165 memset(daddr + tail_off, 0, tail_len);
1166 else {
1167 ret = copy_mc_to_kernel(daddr + tail_off,
1168 saddr + tail_off, tail_len);
1169 if (ret)
1170 return -EIO;
1171 }
1172 }
1173out:
1174 if (zero_edge)
1175 dax_flush(srcmap->dax_dev, daddr, size);
1176 return ret ? -EIO : 0;
1177}
1178
1179/*
1180 * The user has performed a load from a hole in the file. Allocating a new
1181 * page in the file would cause excessive storage usage for workloads with
1182 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1183 * If this page is ever written to we will re-fault and change the mapping to
1184 * point to real DAX storage instead.
1185 */
1186static vm_fault_t dax_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1187 const struct iomap_iter *iter, void **entry)
1188{
1189 struct inode *inode = iter->inode;
1190 unsigned long vaddr = vmf->address;
1191 pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1192 vm_fault_t ret;
1193
1194 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, DAX_ZERO_PAGE);
1195
1196 ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1197 trace_dax_load_hole(inode, vmf, ret);
1198 return ret;
1199}
1200
1201#ifdef CONFIG_FS_DAX_PMD
1202static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1203 const struct iomap_iter *iter, void **entry)
1204{
1205 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1206 unsigned long pmd_addr = vmf->address & PMD_MASK;
1207 struct vm_area_struct *vma = vmf->vma;
1208 struct inode *inode = mapping->host;
1209 pgtable_t pgtable = NULL;
1210 struct page *zero_page;
1211 spinlock_t *ptl;
1212 pmd_t pmd_entry;
1213 pfn_t pfn;
1214
1215 zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1216
1217 if (unlikely(!zero_page))
1218 goto fallback;
1219
1220 pfn = page_to_pfn_t(zero_page);
1221 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn,
1222 DAX_PMD | DAX_ZERO_PAGE);
1223
1224 if (arch_needs_pgtable_deposit()) {
1225 pgtable = pte_alloc_one(vma->vm_mm);
1226 if (!pgtable)
1227 return VM_FAULT_OOM;
1228 }
1229
1230 ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1231 if (!pmd_none(*(vmf->pmd))) {
1232 spin_unlock(ptl);
1233 goto fallback;
1234 }
1235
1236 if (pgtable) {
1237 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1238 mm_inc_nr_ptes(vma->vm_mm);
1239 }
1240 pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1241 pmd_entry = pmd_mkhuge(pmd_entry);
1242 set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1243 spin_unlock(ptl);
1244 trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry);
1245 return VM_FAULT_NOPAGE;
1246
1247fallback:
1248 if (pgtable)
1249 pte_free(vma->vm_mm, pgtable);
1250 trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry);
1251 return VM_FAULT_FALLBACK;
1252}
1253#else
1254static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1255 const struct iomap_iter *iter, void **entry)
1256{
1257 return VM_FAULT_FALLBACK;
1258}
1259#endif /* CONFIG_FS_DAX_PMD */
1260
1261static s64 dax_unshare_iter(struct iomap_iter *iter)
1262{
1263 struct iomap *iomap = &iter->iomap;
1264 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1265 loff_t pos = iter->pos;
1266 loff_t length = iomap_length(iter);
1267 int id = 0;
1268 s64 ret = 0;
1269 void *daddr = NULL, *saddr = NULL;
1270
1271 /* don't bother with blocks that are not shared to start with */
1272 if (!(iomap->flags & IOMAP_F_SHARED))
1273 return length;
1274
1275 id = dax_read_lock();
1276 ret = dax_iomap_direct_access(iomap, pos, length, &daddr, NULL);
1277 if (ret < 0)
1278 goto out_unlock;
1279
1280 /* zero the distance if srcmap is HOLE or UNWRITTEN */
1281 if (srcmap->flags & IOMAP_F_SHARED || srcmap->type == IOMAP_UNWRITTEN) {
1282 memset(daddr, 0, length);
1283 dax_flush(iomap->dax_dev, daddr, length);
1284 ret = length;
1285 goto out_unlock;
1286 }
1287
1288 ret = dax_iomap_direct_access(srcmap, pos, length, &saddr, NULL);
1289 if (ret < 0)
1290 goto out_unlock;
1291
1292 if (copy_mc_to_kernel(daddr, saddr, length) == 0)
1293 ret = length;
1294 else
1295 ret = -EIO;
1296
1297out_unlock:
1298 dax_read_unlock(id);
1299 return dax_mem2blk_err(ret);
1300}
1301
1302int dax_file_unshare(struct inode *inode, loff_t pos, loff_t len,
1303 const struct iomap_ops *ops)
1304{
1305 struct iomap_iter iter = {
1306 .inode = inode,
1307 .pos = pos,
1308 .len = len,
1309 .flags = IOMAP_WRITE | IOMAP_UNSHARE | IOMAP_DAX,
1310 };
1311 int ret;
1312
1313 while ((ret = iomap_iter(&iter, ops)) > 0)
1314 iter.processed = dax_unshare_iter(&iter);
1315 return ret;
1316}
1317EXPORT_SYMBOL_GPL(dax_file_unshare);
1318
1319static int dax_memzero(struct iomap_iter *iter, loff_t pos, size_t size)
1320{
1321 const struct iomap *iomap = &iter->iomap;
1322 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1323 unsigned offset = offset_in_page(pos);
1324 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1325 void *kaddr;
1326 long ret;
1327
1328 ret = dax_direct_access(iomap->dax_dev, pgoff, 1, DAX_ACCESS, &kaddr,
1329 NULL);
1330 if (ret < 0)
1331 return dax_mem2blk_err(ret);
1332
1333 memset(kaddr + offset, 0, size);
1334 if (iomap->flags & IOMAP_F_SHARED)
1335 ret = dax_iomap_copy_around(pos, size, PAGE_SIZE, srcmap,
1336 kaddr);
1337 else
1338 dax_flush(iomap->dax_dev, kaddr + offset, size);
1339 return ret;
1340}
1341
1342static s64 dax_zero_iter(struct iomap_iter *iter, bool *did_zero)
1343{
1344 const struct iomap *iomap = &iter->iomap;
1345 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1346 loff_t pos = iter->pos;
1347 u64 length = iomap_length(iter);
1348 s64 written = 0;
1349
1350 /* already zeroed? we're done. */
1351 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
1352 return length;
1353
1354 /*
1355 * invalidate the pages whose sharing state is to be changed
1356 * because of CoW.
1357 */
1358 if (iomap->flags & IOMAP_F_SHARED)
1359 invalidate_inode_pages2_range(iter->inode->i_mapping,
1360 pos >> PAGE_SHIFT,
1361 (pos + length - 1) >> PAGE_SHIFT);
1362
1363 do {
1364 unsigned offset = offset_in_page(pos);
1365 unsigned size = min_t(u64, PAGE_SIZE - offset, length);
1366 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1367 long rc;
1368 int id;
1369
1370 id = dax_read_lock();
1371 if (IS_ALIGNED(pos, PAGE_SIZE) && size == PAGE_SIZE)
1372 rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1);
1373 else
1374 rc = dax_memzero(iter, pos, size);
1375 dax_read_unlock(id);
1376
1377 if (rc < 0)
1378 return rc;
1379 pos += size;
1380 length -= size;
1381 written += size;
1382 } while (length > 0);
1383
1384 if (did_zero)
1385 *did_zero = true;
1386 return written;
1387}
1388
1389int dax_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1390 const struct iomap_ops *ops)
1391{
1392 struct iomap_iter iter = {
1393 .inode = inode,
1394 .pos = pos,
1395 .len = len,
1396 .flags = IOMAP_DAX | IOMAP_ZERO,
1397 };
1398 int ret;
1399
1400 while ((ret = iomap_iter(&iter, ops)) > 0)
1401 iter.processed = dax_zero_iter(&iter, did_zero);
1402 return ret;
1403}
1404EXPORT_SYMBOL_GPL(dax_zero_range);
1405
1406int dax_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1407 const struct iomap_ops *ops)
1408{
1409 unsigned int blocksize = i_blocksize(inode);
1410 unsigned int off = pos & (blocksize - 1);
1411
1412 /* Block boundary? Nothing to do */
1413 if (!off)
1414 return 0;
1415 return dax_zero_range(inode, pos, blocksize - off, did_zero, ops);
1416}
1417EXPORT_SYMBOL_GPL(dax_truncate_page);
1418
1419static loff_t dax_iomap_iter(const struct iomap_iter *iomi,
1420 struct iov_iter *iter)
1421{
1422 const struct iomap *iomap = &iomi->iomap;
1423 const struct iomap *srcmap = iomap_iter_srcmap(iomi);
1424 loff_t length = iomap_length(iomi);
1425 loff_t pos = iomi->pos;
1426 struct dax_device *dax_dev = iomap->dax_dev;
1427 loff_t end = pos + length, done = 0;
1428 bool write = iov_iter_rw(iter) == WRITE;
1429 bool cow = write && iomap->flags & IOMAP_F_SHARED;
1430 ssize_t ret = 0;
1431 size_t xfer;
1432 int id;
1433
1434 if (!write) {
1435 end = min(end, i_size_read(iomi->inode));
1436 if (pos >= end)
1437 return 0;
1438
1439 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1440 return iov_iter_zero(min(length, end - pos), iter);
1441 }
1442
1443 /*
1444 * In DAX mode, enforce either pure overwrites of written extents, or
1445 * writes to unwritten extents as part of a copy-on-write operation.
1446 */
1447 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED &&
1448 !(iomap->flags & IOMAP_F_SHARED)))
1449 return -EIO;
1450
1451 /*
1452 * Write can allocate block for an area which has a hole page mapped
1453 * into page tables. We have to tear down these mappings so that data
1454 * written by write(2) is visible in mmap.
1455 */
1456 if (iomap->flags & IOMAP_F_NEW || cow) {
1457 /*
1458 * Filesystem allows CoW on non-shared extents. The src extents
1459 * may have been mmapped with dirty mark before. To be able to
1460 * invalidate its dax entries, we need to clear the dirty mark
1461 * in advance.
1462 */
1463 if (cow)
1464 __dax_clear_dirty_range(iomi->inode->i_mapping,
1465 pos >> PAGE_SHIFT,
1466 (end - 1) >> PAGE_SHIFT);
1467 invalidate_inode_pages2_range(iomi->inode->i_mapping,
1468 pos >> PAGE_SHIFT,
1469 (end - 1) >> PAGE_SHIFT);
1470 }
1471
1472 id = dax_read_lock();
1473 while (pos < end) {
1474 unsigned offset = pos & (PAGE_SIZE - 1);
1475 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1476 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1477 ssize_t map_len;
1478 bool recovery = false;
1479 void *kaddr;
1480
1481 if (fatal_signal_pending(current)) {
1482 ret = -EINTR;
1483 break;
1484 }
1485
1486 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1487 DAX_ACCESS, &kaddr, NULL);
1488 if (map_len == -EHWPOISON && iov_iter_rw(iter) == WRITE) {
1489 map_len = dax_direct_access(dax_dev, pgoff,
1490 PHYS_PFN(size), DAX_RECOVERY_WRITE,
1491 &kaddr, NULL);
1492 if (map_len > 0)
1493 recovery = true;
1494 }
1495 if (map_len < 0) {
1496 ret = dax_mem2blk_err(map_len);
1497 break;
1498 }
1499
1500 if (cow) {
1501 ret = dax_iomap_copy_around(pos, length, PAGE_SIZE,
1502 srcmap, kaddr);
1503 if (ret)
1504 break;
1505 }
1506
1507 map_len = PFN_PHYS(map_len);
1508 kaddr += offset;
1509 map_len -= offset;
1510 if (map_len > end - pos)
1511 map_len = end - pos;
1512
1513 if (recovery)
1514 xfer = dax_recovery_write(dax_dev, pgoff, kaddr,
1515 map_len, iter);
1516 else if (write)
1517 xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1518 map_len, iter);
1519 else
1520 xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1521 map_len, iter);
1522
1523 pos += xfer;
1524 length -= xfer;
1525 done += xfer;
1526
1527 if (xfer == 0)
1528 ret = -EFAULT;
1529 if (xfer < map_len)
1530 break;
1531 }
1532 dax_read_unlock(id);
1533
1534 return done ? done : ret;
1535}
1536
1537/**
1538 * dax_iomap_rw - Perform I/O to a DAX file
1539 * @iocb: The control block for this I/O
1540 * @iter: The addresses to do I/O from or to
1541 * @ops: iomap ops passed from the file system
1542 *
1543 * This function performs read and write operations to directly mapped
1544 * persistent memory. The callers needs to take care of read/write exclusion
1545 * and evicting any page cache pages in the region under I/O.
1546 */
1547ssize_t
1548dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1549 const struct iomap_ops *ops)
1550{
1551 struct iomap_iter iomi = {
1552 .inode = iocb->ki_filp->f_mapping->host,
1553 .pos = iocb->ki_pos,
1554 .len = iov_iter_count(iter),
1555 .flags = IOMAP_DAX,
1556 };
1557 loff_t done = 0;
1558 int ret;
1559
1560 if (!iomi.len)
1561 return 0;
1562
1563 if (iov_iter_rw(iter) == WRITE) {
1564 lockdep_assert_held_write(&iomi.inode->i_rwsem);
1565 iomi.flags |= IOMAP_WRITE;
1566 } else {
1567 lockdep_assert_held(&iomi.inode->i_rwsem);
1568 }
1569
1570 if (iocb->ki_flags & IOCB_NOWAIT)
1571 iomi.flags |= IOMAP_NOWAIT;
1572
1573 while ((ret = iomap_iter(&iomi, ops)) > 0)
1574 iomi.processed = dax_iomap_iter(&iomi, iter);
1575
1576 done = iomi.pos - iocb->ki_pos;
1577 iocb->ki_pos = iomi.pos;
1578 return done ? done : ret;
1579}
1580EXPORT_SYMBOL_GPL(dax_iomap_rw);
1581
1582static vm_fault_t dax_fault_return(int error)
1583{
1584 if (error == 0)
1585 return VM_FAULT_NOPAGE;
1586 return vmf_error(error);
1587}
1588
1589/*
1590 * When handling a synchronous page fault and the inode need a fsync, we can
1591 * insert the PTE/PMD into page tables only after that fsync happened. Skip
1592 * insertion for now and return the pfn so that caller can insert it after the
1593 * fsync is done.
1594 */
1595static vm_fault_t dax_fault_synchronous_pfnp(pfn_t *pfnp, pfn_t pfn)
1596{
1597 if (WARN_ON_ONCE(!pfnp))
1598 return VM_FAULT_SIGBUS;
1599 *pfnp = pfn;
1600 return VM_FAULT_NEEDDSYNC;
1601}
1602
1603static vm_fault_t dax_fault_cow_page(struct vm_fault *vmf,
1604 const struct iomap_iter *iter)
1605{
1606 vm_fault_t ret;
1607 int error = 0;
1608
1609 switch (iter->iomap.type) {
1610 case IOMAP_HOLE:
1611 case IOMAP_UNWRITTEN:
1612 clear_user_highpage(vmf->cow_page, vmf->address);
1613 break;
1614 case IOMAP_MAPPED:
1615 error = copy_cow_page_dax(vmf, iter);
1616 break;
1617 default:
1618 WARN_ON_ONCE(1);
1619 error = -EIO;
1620 break;
1621 }
1622
1623 if (error)
1624 return dax_fault_return(error);
1625
1626 __SetPageUptodate(vmf->cow_page);
1627 ret = finish_fault(vmf);
1628 if (!ret)
1629 return VM_FAULT_DONE_COW;
1630 return ret;
1631}
1632
1633/**
1634 * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault.
1635 * @vmf: vm fault instance
1636 * @iter: iomap iter
1637 * @pfnp: pfn to be returned
1638 * @xas: the dax mapping tree of a file
1639 * @entry: an unlocked dax entry to be inserted
1640 * @pmd: distinguish whether it is a pmd fault
1641 */
1642static vm_fault_t dax_fault_iter(struct vm_fault *vmf,
1643 const struct iomap_iter *iter, pfn_t *pfnp,
1644 struct xa_state *xas, void **entry, bool pmd)
1645{
1646 const struct iomap *iomap = &iter->iomap;
1647 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1648 size_t size = pmd ? PMD_SIZE : PAGE_SIZE;
1649 loff_t pos = (loff_t)xas->xa_index << PAGE_SHIFT;
1650 bool write = iter->flags & IOMAP_WRITE;
1651 unsigned long entry_flags = pmd ? DAX_PMD : 0;
1652 int err = 0;
1653 pfn_t pfn;
1654 void *kaddr;
1655
1656 if (!pmd && vmf->cow_page)
1657 return dax_fault_cow_page(vmf, iter);
1658
1659 /* if we are reading UNWRITTEN and HOLE, return a hole. */
1660 if (!write &&
1661 (iomap->type == IOMAP_UNWRITTEN || iomap->type == IOMAP_HOLE)) {
1662 if (!pmd)
1663 return dax_load_hole(xas, vmf, iter, entry);
1664 return dax_pmd_load_hole(xas, vmf, iter, entry);
1665 }
1666
1667 if (iomap->type != IOMAP_MAPPED && !(iomap->flags & IOMAP_F_SHARED)) {
1668 WARN_ON_ONCE(1);
1669 return pmd ? VM_FAULT_FALLBACK : VM_FAULT_SIGBUS;
1670 }
1671
1672 err = dax_iomap_direct_access(iomap, pos, size, &kaddr, &pfn);
1673 if (err)
1674 return pmd ? VM_FAULT_FALLBACK : dax_fault_return(err);
1675
1676 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, entry_flags);
1677
1678 if (write && iomap->flags & IOMAP_F_SHARED) {
1679 err = dax_iomap_copy_around(pos, size, size, srcmap, kaddr);
1680 if (err)
1681 return dax_fault_return(err);
1682 }
1683
1684 if (dax_fault_is_synchronous(iter, vmf->vma))
1685 return dax_fault_synchronous_pfnp(pfnp, pfn);
1686
1687 /* insert PMD pfn */
1688 if (pmd)
1689 return vmf_insert_pfn_pmd(vmf, pfn, write);
1690
1691 /* insert PTE pfn */
1692 if (write)
1693 return vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1694 return vmf_insert_mixed(vmf->vma, vmf->address, pfn);
1695}
1696
1697static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1698 int *iomap_errp, const struct iomap_ops *ops)
1699{
1700 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1701 XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
1702 struct iomap_iter iter = {
1703 .inode = mapping->host,
1704 .pos = (loff_t)vmf->pgoff << PAGE_SHIFT,
1705 .len = PAGE_SIZE,
1706 .flags = IOMAP_DAX | IOMAP_FAULT,
1707 };
1708 vm_fault_t ret = 0;
1709 void *entry;
1710 int error;
1711
1712 trace_dax_pte_fault(iter.inode, vmf, ret);
1713 /*
1714 * Check whether offset isn't beyond end of file now. Caller is supposed
1715 * to hold locks serializing us with truncate / punch hole so this is
1716 * a reliable test.
1717 */
1718 if (iter.pos >= i_size_read(iter.inode)) {
1719 ret = VM_FAULT_SIGBUS;
1720 goto out;
1721 }
1722
1723 if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1724 iter.flags |= IOMAP_WRITE;
1725
1726 entry = grab_mapping_entry(&xas, mapping, 0);
1727 if (xa_is_internal(entry)) {
1728 ret = xa_to_internal(entry);
1729 goto out;
1730 }
1731
1732 /*
1733 * It is possible, particularly with mixed reads & writes to private
1734 * mappings, that we have raced with a PMD fault that overlaps with
1735 * the PTE we need to set up. If so just return and the fault will be
1736 * retried.
1737 */
1738 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1739 ret = VM_FAULT_NOPAGE;
1740 goto unlock_entry;
1741 }
1742
1743 while ((error = iomap_iter(&iter, ops)) > 0) {
1744 if (WARN_ON_ONCE(iomap_length(&iter) < PAGE_SIZE)) {
1745 iter.processed = -EIO; /* fs corruption? */
1746 continue;
1747 }
1748
1749 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, false);
1750 if (ret != VM_FAULT_SIGBUS &&
1751 (iter.iomap.flags & IOMAP_F_NEW)) {
1752 count_vm_event(PGMAJFAULT);
1753 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
1754 ret |= VM_FAULT_MAJOR;
1755 }
1756
1757 if (!(ret & VM_FAULT_ERROR))
1758 iter.processed = PAGE_SIZE;
1759 }
1760
1761 if (iomap_errp)
1762 *iomap_errp = error;
1763 if (!ret && error)
1764 ret = dax_fault_return(error);
1765
1766unlock_entry:
1767 dax_unlock_entry(&xas, entry);
1768out:
1769 trace_dax_pte_fault_done(iter.inode, vmf, ret);
1770 return ret;
1771}
1772
1773#ifdef CONFIG_FS_DAX_PMD
1774static bool dax_fault_check_fallback(struct vm_fault *vmf, struct xa_state *xas,
1775 pgoff_t max_pgoff)
1776{
1777 unsigned long pmd_addr = vmf->address & PMD_MASK;
1778 bool write = vmf->flags & FAULT_FLAG_WRITE;
1779
1780 /*
1781 * Make sure that the faulting address's PMD offset (color) matches
1782 * the PMD offset from the start of the file. This is necessary so
1783 * that a PMD range in the page table overlaps exactly with a PMD
1784 * range in the page cache.
1785 */
1786 if ((vmf->pgoff & PG_PMD_COLOUR) !=
1787 ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1788 return true;
1789
1790 /* Fall back to PTEs if we're going to COW */
1791 if (write && !(vmf->vma->vm_flags & VM_SHARED))
1792 return true;
1793
1794 /* If the PMD would extend outside the VMA */
1795 if (pmd_addr < vmf->vma->vm_start)
1796 return true;
1797 if ((pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
1798 return true;
1799
1800 /* If the PMD would extend beyond the file size */
1801 if ((xas->xa_index | PG_PMD_COLOUR) >= max_pgoff)
1802 return true;
1803
1804 return false;
1805}
1806
1807static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1808 const struct iomap_ops *ops)
1809{
1810 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1811 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
1812 struct iomap_iter iter = {
1813 .inode = mapping->host,
1814 .len = PMD_SIZE,
1815 .flags = IOMAP_DAX | IOMAP_FAULT,
1816 };
1817 vm_fault_t ret = VM_FAULT_FALLBACK;
1818 pgoff_t max_pgoff;
1819 void *entry;
1820
1821 if (vmf->flags & FAULT_FLAG_WRITE)
1822 iter.flags |= IOMAP_WRITE;
1823
1824 /*
1825 * Check whether offset isn't beyond end of file now. Caller is
1826 * supposed to hold locks serializing us with truncate / punch hole so
1827 * this is a reliable test.
1828 */
1829 max_pgoff = DIV_ROUND_UP(i_size_read(iter.inode), PAGE_SIZE);
1830
1831 trace_dax_pmd_fault(iter.inode, vmf, max_pgoff, 0);
1832
1833 if (xas.xa_index >= max_pgoff) {
1834 ret = VM_FAULT_SIGBUS;
1835 goto out;
1836 }
1837
1838 if (dax_fault_check_fallback(vmf, &xas, max_pgoff))
1839 goto fallback;
1840
1841 /*
1842 * grab_mapping_entry() will make sure we get an empty PMD entry,
1843 * a zero PMD entry or a DAX PMD. If it can't (because a PTE
1844 * entry is already in the array, for instance), it will return
1845 * VM_FAULT_FALLBACK.
1846 */
1847 entry = grab_mapping_entry(&xas, mapping, PMD_ORDER);
1848 if (xa_is_internal(entry)) {
1849 ret = xa_to_internal(entry);
1850 goto fallback;
1851 }
1852
1853 /*
1854 * It is possible, particularly with mixed reads & writes to private
1855 * mappings, that we have raced with a PTE fault that overlaps with
1856 * the PMD we need to set up. If so just return and the fault will be
1857 * retried.
1858 */
1859 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1860 !pmd_devmap(*vmf->pmd)) {
1861 ret = 0;
1862 goto unlock_entry;
1863 }
1864
1865 iter.pos = (loff_t)xas.xa_index << PAGE_SHIFT;
1866 while (iomap_iter(&iter, ops) > 0) {
1867 if (iomap_length(&iter) < PMD_SIZE)
1868 continue; /* actually breaks out of the loop */
1869
1870 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, true);
1871 if (ret != VM_FAULT_FALLBACK)
1872 iter.processed = PMD_SIZE;
1873 }
1874
1875unlock_entry:
1876 dax_unlock_entry(&xas, entry);
1877fallback:
1878 if (ret == VM_FAULT_FALLBACK) {
1879 split_huge_pmd(vmf->vma, vmf->pmd, vmf->address);
1880 count_vm_event(THP_FAULT_FALLBACK);
1881 }
1882out:
1883 trace_dax_pmd_fault_done(iter.inode, vmf, max_pgoff, ret);
1884 return ret;
1885}
1886#else
1887static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1888 const struct iomap_ops *ops)
1889{
1890 return VM_FAULT_FALLBACK;
1891}
1892#endif /* CONFIG_FS_DAX_PMD */
1893
1894/**
1895 * dax_iomap_fault - handle a page fault on a DAX file
1896 * @vmf: The description of the fault
1897 * @order: Order of the page to fault in
1898 * @pfnp: PFN to insert for synchronous faults if fsync is required
1899 * @iomap_errp: Storage for detailed error code in case of error
1900 * @ops: Iomap ops passed from the file system
1901 *
1902 * When a page fault occurs, filesystems may call this helper in
1903 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1904 * has done all the necessary locking for page fault to proceed
1905 * successfully.
1906 */
1907vm_fault_t dax_iomap_fault(struct vm_fault *vmf, unsigned int order,
1908 pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1909{
1910 if (order == 0)
1911 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1912 else if (order == PMD_ORDER)
1913 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1914 else
1915 return VM_FAULT_FALLBACK;
1916}
1917EXPORT_SYMBOL_GPL(dax_iomap_fault);
1918
1919/*
1920 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1921 * @vmf: The description of the fault
1922 * @pfn: PFN to insert
1923 * @order: Order of entry to insert.
1924 *
1925 * This function inserts a writeable PTE or PMD entry into the page tables
1926 * for an mmaped DAX file. It also marks the page cache entry as dirty.
1927 */
1928static vm_fault_t
1929dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
1930{
1931 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1932 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1933 void *entry;
1934 vm_fault_t ret;
1935
1936 xas_lock_irq(&xas);
1937 entry = get_unlocked_entry(&xas, order);
1938 /* Did we race with someone splitting entry or so? */
1939 if (!entry || dax_is_conflict(entry) ||
1940 (order == 0 && !dax_is_pte_entry(entry))) {
1941 put_unlocked_entry(&xas, entry, WAKE_NEXT);
1942 xas_unlock_irq(&xas);
1943 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1944 VM_FAULT_NOPAGE);
1945 return VM_FAULT_NOPAGE;
1946 }
1947 xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1948 dax_lock_entry(&xas, entry);
1949 xas_unlock_irq(&xas);
1950 if (order == 0)
1951 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1952#ifdef CONFIG_FS_DAX_PMD
1953 else if (order == PMD_ORDER)
1954 ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
1955#endif
1956 else
1957 ret = VM_FAULT_FALLBACK;
1958 dax_unlock_entry(&xas, entry);
1959 trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1960 return ret;
1961}
1962
1963/**
1964 * dax_finish_sync_fault - finish synchronous page fault
1965 * @vmf: The description of the fault
1966 * @order: Order of entry to be inserted
1967 * @pfn: PFN to insert
1968 *
1969 * This function ensures that the file range touched by the page fault is
1970 * stored persistently on the media and handles inserting of appropriate page
1971 * table entry.
1972 */
1973vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf, unsigned int order,
1974 pfn_t pfn)
1975{
1976 int err;
1977 loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1978 size_t len = PAGE_SIZE << order;
1979
1980 err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1981 if (err)
1982 return VM_FAULT_SIGBUS;
1983 return dax_insert_pfn_mkwrite(vmf, pfn, order);
1984}
1985EXPORT_SYMBOL_GPL(dax_finish_sync_fault);
1986
1987static loff_t dax_range_compare_iter(struct iomap_iter *it_src,
1988 struct iomap_iter *it_dest, u64 len, bool *same)
1989{
1990 const struct iomap *smap = &it_src->iomap;
1991 const struct iomap *dmap = &it_dest->iomap;
1992 loff_t pos1 = it_src->pos, pos2 = it_dest->pos;
1993 void *saddr, *daddr;
1994 int id, ret;
1995
1996 len = min(len, min(smap->length, dmap->length));
1997
1998 if (smap->type == IOMAP_HOLE && dmap->type == IOMAP_HOLE) {
1999 *same = true;
2000 return len;
2001 }
2002
2003 if (smap->type == IOMAP_HOLE || dmap->type == IOMAP_HOLE) {
2004 *same = false;
2005 return 0;
2006 }
2007
2008 id = dax_read_lock();
2009 ret = dax_iomap_direct_access(smap, pos1, ALIGN(pos1 + len, PAGE_SIZE),
2010 &saddr, NULL);
2011 if (ret < 0)
2012 goto out_unlock;
2013
2014 ret = dax_iomap_direct_access(dmap, pos2, ALIGN(pos2 + len, PAGE_SIZE),
2015 &daddr, NULL);
2016 if (ret < 0)
2017 goto out_unlock;
2018
2019 *same = !memcmp(saddr, daddr, len);
2020 if (!*same)
2021 len = 0;
2022 dax_read_unlock(id);
2023 return len;
2024
2025out_unlock:
2026 dax_read_unlock(id);
2027 return -EIO;
2028}
2029
2030int dax_dedupe_file_range_compare(struct inode *src, loff_t srcoff,
2031 struct inode *dst, loff_t dstoff, loff_t len, bool *same,
2032 const struct iomap_ops *ops)
2033{
2034 struct iomap_iter src_iter = {
2035 .inode = src,
2036 .pos = srcoff,
2037 .len = len,
2038 .flags = IOMAP_DAX,
2039 };
2040 struct iomap_iter dst_iter = {
2041 .inode = dst,
2042 .pos = dstoff,
2043 .len = len,
2044 .flags = IOMAP_DAX,
2045 };
2046 int ret, compared = 0;
2047
2048 while ((ret = iomap_iter(&src_iter, ops)) > 0 &&
2049 (ret = iomap_iter(&dst_iter, ops)) > 0) {
2050 compared = dax_range_compare_iter(&src_iter, &dst_iter,
2051 min(src_iter.len, dst_iter.len), same);
2052 if (compared < 0)
2053 return ret;
2054 src_iter.processed = dst_iter.processed = compared;
2055 }
2056 return ret;
2057}
2058
2059int dax_remap_file_range_prep(struct file *file_in, loff_t pos_in,
2060 struct file *file_out, loff_t pos_out,
2061 loff_t *len, unsigned int remap_flags,
2062 const struct iomap_ops *ops)
2063{
2064 return __generic_remap_file_range_prep(file_in, pos_in, file_out,
2065 pos_out, len, remap_flags, ops);
2066}
2067EXPORT_SYMBOL_GPL(dax_remap_file_range_prep);