a345c168acaae4956dad1cba8707061b5d22f534
[sfrench/cifs-2.6.git] / fs / dax.c
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
35 #define RADIX_DAX_MASK  0xf
36 #define RADIX_DAX_SHIFT 4
37 #define RADIX_DAX_PTE  (0x4 | RADIX_TREE_EXCEPTIONAL_ENTRY)
38 #define RADIX_DAX_PMD  (0x8 | RADIX_TREE_EXCEPTIONAL_ENTRY)
39 #define RADIX_DAX_TYPE(entry) ((unsigned long)entry & RADIX_DAX_MASK)
40 #define RADIX_DAX_SECTOR(entry) (((unsigned long)entry >> RADIX_DAX_SHIFT))
41 #define RADIX_DAX_ENTRY(sector, pmd) ((void *)((unsigned long)sector << \
42                 RADIX_DAX_SHIFT | (pmd ? RADIX_DAX_PMD : RADIX_DAX_PTE)))
43
44 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
45 {
46         struct request_queue *q = bdev->bd_queue;
47         long rc = -EIO;
48
49         dax->addr = (void __pmem *) ERR_PTR(-EIO);
50         if (blk_queue_enter(q, true) != 0)
51                 return rc;
52
53         rc = bdev_direct_access(bdev, dax);
54         if (rc < 0) {
55                 dax->addr = (void __pmem *) ERR_PTR(rc);
56                 blk_queue_exit(q);
57                 return rc;
58         }
59         return rc;
60 }
61
62 static void dax_unmap_atomic(struct block_device *bdev,
63                 const struct blk_dax_ctl *dax)
64 {
65         if (IS_ERR(dax->addr))
66                 return;
67         blk_queue_exit(bdev->bd_queue);
68 }
69
70 struct page *read_dax_sector(struct block_device *bdev, sector_t n)
71 {
72         struct page *page = alloc_pages(GFP_KERNEL, 0);
73         struct blk_dax_ctl dax = {
74                 .size = PAGE_SIZE,
75                 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
76         };
77         long rc;
78
79         if (!page)
80                 return ERR_PTR(-ENOMEM);
81
82         rc = dax_map_atomic(bdev, &dax);
83         if (rc < 0)
84                 return ERR_PTR(rc);
85         memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
86         dax_unmap_atomic(bdev, &dax);
87         return page;
88 }
89
90 /*
91  * dax_clear_sectors() is called from within transaction context from XFS,
92  * and hence this means the stack from this point must follow GFP_NOFS
93  * semantics for all operations.
94  */
95 int dax_clear_sectors(struct block_device *bdev, sector_t _sector, long _size)
96 {
97         struct blk_dax_ctl dax = {
98                 .sector = _sector,
99                 .size = _size,
100         };
101
102         might_sleep();
103         do {
104                 long count, sz;
105
106                 count = dax_map_atomic(bdev, &dax);
107                 if (count < 0)
108                         return count;
109                 sz = min_t(long, count, SZ_128K);
110                 clear_pmem(dax.addr, sz);
111                 dax.size -= sz;
112                 dax.sector += sz / 512;
113                 dax_unmap_atomic(bdev, &dax);
114                 cond_resched();
115         } while (dax.size);
116
117         wmb_pmem();
118         return 0;
119 }
120 EXPORT_SYMBOL_GPL(dax_clear_sectors);
121
122 /* the clear_pmem() calls are ordered by a wmb_pmem() in the caller */
123 static void dax_new_buf(void __pmem *addr, unsigned size, unsigned first,
124                 loff_t pos, loff_t end)
125 {
126         loff_t final = end - pos + first; /* The final byte of the buffer */
127
128         if (first > 0)
129                 clear_pmem(addr, first);
130         if (final < size)
131                 clear_pmem(addr + final, size - final);
132 }
133
134 static bool buffer_written(struct buffer_head *bh)
135 {
136         return buffer_mapped(bh) && !buffer_unwritten(bh);
137 }
138
139 /*
140  * When ext4 encounters a hole, it returns without modifying the buffer_head
141  * which means that we can't trust b_size.  To cope with this, we set b_state
142  * to 0 before calling get_block and, if any bit is set, we know we can trust
143  * b_size.  Unfortunate, really, since ext4 knows precisely how long a hole is
144  * and would save us time calling get_block repeatedly.
145  */
146 static bool buffer_size_valid(struct buffer_head *bh)
147 {
148         return bh->b_state != 0;
149 }
150
151
152 static sector_t to_sector(const struct buffer_head *bh,
153                 const struct inode *inode)
154 {
155         sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
156
157         return sector;
158 }
159
160 static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
161                       loff_t start, loff_t end, get_block_t get_block,
162                       struct buffer_head *bh)
163 {
164         loff_t pos = start, max = start, bh_max = start;
165         bool hole = false, need_wmb = false;
166         struct block_device *bdev = NULL;
167         int rw = iov_iter_rw(iter), rc;
168         long map_len = 0;
169         struct blk_dax_ctl dax = {
170                 .addr = (void __pmem *) ERR_PTR(-EIO),
171         };
172
173         if (rw == READ)
174                 end = min(end, i_size_read(inode));
175
176         while (pos < end) {
177                 size_t len;
178                 if (pos == max) {
179                         unsigned blkbits = inode->i_blkbits;
180                         long page = pos >> PAGE_SHIFT;
181                         sector_t block = page << (PAGE_SHIFT - blkbits);
182                         unsigned first = pos - (block << blkbits);
183                         long size;
184
185                         if (pos == bh_max) {
186                                 bh->b_size = PAGE_ALIGN(end - pos);
187                                 bh->b_state = 0;
188                                 rc = get_block(inode, block, bh, rw == WRITE);
189                                 if (rc)
190                                         break;
191                                 if (!buffer_size_valid(bh))
192                                         bh->b_size = 1 << blkbits;
193                                 bh_max = pos - first + bh->b_size;
194                                 bdev = bh->b_bdev;
195                         } else {
196                                 unsigned done = bh->b_size -
197                                                 (bh_max - (pos - first));
198                                 bh->b_blocknr += done >> blkbits;
199                                 bh->b_size -= done;
200                         }
201
202                         hole = rw == READ && !buffer_written(bh);
203                         if (hole) {
204                                 size = bh->b_size - first;
205                         } else {
206                                 dax_unmap_atomic(bdev, &dax);
207                                 dax.sector = to_sector(bh, inode);
208                                 dax.size = bh->b_size;
209                                 map_len = dax_map_atomic(bdev, &dax);
210                                 if (map_len < 0) {
211                                         rc = map_len;
212                                         break;
213                                 }
214                                 if (buffer_unwritten(bh) || buffer_new(bh)) {
215                                         dax_new_buf(dax.addr, map_len, first,
216                                                         pos, end);
217                                         need_wmb = true;
218                                 }
219                                 dax.addr += first;
220                                 size = map_len - first;
221                         }
222                         max = min(pos + size, end);
223                 }
224
225                 if (iov_iter_rw(iter) == WRITE) {
226                         len = copy_from_iter_pmem(dax.addr, max - pos, iter);
227                         need_wmb = true;
228                 } else if (!hole)
229                         len = copy_to_iter((void __force *) dax.addr, max - pos,
230                                         iter);
231                 else
232                         len = iov_iter_zero(max - pos, iter);
233
234                 if (!len) {
235                         rc = -EFAULT;
236                         break;
237                 }
238
239                 pos += len;
240                 if (!IS_ERR(dax.addr))
241                         dax.addr += len;
242         }
243
244         if (need_wmb)
245                 wmb_pmem();
246         dax_unmap_atomic(bdev, &dax);
247
248         return (pos == start) ? rc : pos - start;
249 }
250
251 /**
252  * dax_do_io - Perform I/O to a DAX file
253  * @iocb: The control block for this I/O
254  * @inode: The file which the I/O is directed at
255  * @iter: The addresses to do I/O from or to
256  * @get_block: The filesystem method used to translate file offsets to blocks
257  * @end_io: A filesystem callback for I/O completion
258  * @flags: See below
259  *
260  * This function uses the same locking scheme as do_blockdev_direct_IO:
261  * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
262  * caller for writes.  For reads, we take and release the i_mutex ourselves.
263  * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
264  * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
265  * is in progress.
266  */
267 ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
268                   struct iov_iter *iter, get_block_t get_block,
269                   dio_iodone_t end_io, int flags)
270 {
271         struct buffer_head bh;
272         ssize_t retval = -EINVAL;
273         loff_t pos = iocb->ki_pos;
274         loff_t end = pos + iov_iter_count(iter);
275
276         memset(&bh, 0, sizeof(bh));
277         bh.b_bdev = inode->i_sb->s_bdev;
278
279         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) {
280                 struct address_space *mapping = inode->i_mapping;
281                 inode_lock(inode);
282                 retval = filemap_write_and_wait_range(mapping, pos, end - 1);
283                 if (retval) {
284                         inode_unlock(inode);
285                         goto out;
286                 }
287         }
288
289         /* Protects against truncate */
290         if (!(flags & DIO_SKIP_DIO_COUNT))
291                 inode_dio_begin(inode);
292
293         retval = dax_io(inode, iter, pos, end, get_block, &bh);
294
295         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
296                 inode_unlock(inode);
297
298         if (end_io) {
299                 int err;
300
301                 err = end_io(iocb, pos, retval, bh.b_private);
302                 if (err)
303                         retval = err;
304         }
305
306         if (!(flags & DIO_SKIP_DIO_COUNT))
307                 inode_dio_end(inode);
308  out:
309         return retval;
310 }
311 EXPORT_SYMBOL_GPL(dax_do_io);
312
313 /*
314  * The user has performed a load from a hole in the file.  Allocating
315  * a new page in the file would cause excessive storage usage for
316  * workloads with sparse files.  We allocate a page cache page instead.
317  * We'll kick it out of the page cache if it's ever written to,
318  * otherwise it will simply fall out of the page cache under memory
319  * pressure without ever having been dirtied.
320  */
321 static int dax_load_hole(struct address_space *mapping, struct page *page,
322                                                         struct vm_fault *vmf)
323 {
324         unsigned long size;
325         struct inode *inode = mapping->host;
326         if (!page)
327                 page = find_or_create_page(mapping, vmf->pgoff,
328                                                 GFP_KERNEL | __GFP_ZERO);
329         if (!page)
330                 return VM_FAULT_OOM;
331         /* Recheck i_size under page lock to avoid truncate race */
332         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
333         if (vmf->pgoff >= size) {
334                 unlock_page(page);
335                 put_page(page);
336                 return VM_FAULT_SIGBUS;
337         }
338
339         vmf->page = page;
340         return VM_FAULT_LOCKED;
341 }
342
343 static int copy_user_bh(struct page *to, struct inode *inode,
344                 struct buffer_head *bh, unsigned long vaddr)
345 {
346         struct blk_dax_ctl dax = {
347                 .sector = to_sector(bh, inode),
348                 .size = bh->b_size,
349         };
350         struct block_device *bdev = bh->b_bdev;
351         void *vto;
352
353         if (dax_map_atomic(bdev, &dax) < 0)
354                 return PTR_ERR(dax.addr);
355         vto = kmap_atomic(to);
356         copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
357         kunmap_atomic(vto);
358         dax_unmap_atomic(bdev, &dax);
359         return 0;
360 }
361
362 #define NO_SECTOR -1
363 #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT))
364
365 static int dax_radix_entry(struct address_space *mapping, pgoff_t index,
366                 sector_t sector, bool pmd_entry, bool dirty)
367 {
368         struct radix_tree_root *page_tree = &mapping->page_tree;
369         pgoff_t pmd_index = DAX_PMD_INDEX(index);
370         int type, error = 0;
371         void *entry;
372
373         WARN_ON_ONCE(pmd_entry && !dirty);
374         if (dirty)
375                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
376
377         spin_lock_irq(&mapping->tree_lock);
378
379         entry = radix_tree_lookup(page_tree, pmd_index);
380         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) {
381                 index = pmd_index;
382                 goto dirty;
383         }
384
385         entry = radix_tree_lookup(page_tree, index);
386         if (entry) {
387                 type = RADIX_DAX_TYPE(entry);
388                 if (WARN_ON_ONCE(type != RADIX_DAX_PTE &&
389                                         type != RADIX_DAX_PMD)) {
390                         error = -EIO;
391                         goto unlock;
392                 }
393
394                 if (!pmd_entry || type == RADIX_DAX_PMD)
395                         goto dirty;
396
397                 /*
398                  * We only insert dirty PMD entries into the radix tree.  This
399                  * means we don't need to worry about removing a dirty PTE
400                  * entry and inserting a clean PMD entry, thus reducing the
401                  * range we would flush with a follow-up fsync/msync call.
402                  */
403                 radix_tree_delete(&mapping->page_tree, index);
404                 mapping->nrexceptional--;
405         }
406
407         if (sector == NO_SECTOR) {
408                 /*
409                  * This can happen during correct operation if our pfn_mkwrite
410                  * fault raced against a hole punch operation.  If this
411                  * happens the pte that was hole punched will have been
412                  * unmapped and the radix tree entry will have been removed by
413                  * the time we are called, but the call will still happen.  We
414                  * will return all the way up to wp_pfn_shared(), where the
415                  * pte_same() check will fail, eventually causing page fault
416                  * to be retried by the CPU.
417                  */
418                 goto unlock;
419         }
420
421         error = radix_tree_insert(page_tree, index,
422                         RADIX_DAX_ENTRY(sector, pmd_entry));
423         if (error)
424                 goto unlock;
425
426         mapping->nrexceptional++;
427  dirty:
428         if (dirty)
429                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
430  unlock:
431         spin_unlock_irq(&mapping->tree_lock);
432         return error;
433 }
434
435 static int dax_writeback_one(struct block_device *bdev,
436                 struct address_space *mapping, pgoff_t index, void *entry)
437 {
438         struct radix_tree_root *page_tree = &mapping->page_tree;
439         int type = RADIX_DAX_TYPE(entry);
440         struct radix_tree_node *node;
441         struct blk_dax_ctl dax;
442         void **slot;
443         int ret = 0;
444
445         spin_lock_irq(&mapping->tree_lock);
446         /*
447          * Regular page slots are stabilized by the page lock even
448          * without the tree itself locked.  These unlocked entries
449          * need verification under the tree lock.
450          */
451         if (!__radix_tree_lookup(page_tree, index, &node, &slot))
452                 goto unlock;
453         if (*slot != entry)
454                 goto unlock;
455
456         /* another fsync thread may have already written back this entry */
457         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
458                 goto unlock;
459
460         if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
461                 ret = -EIO;
462                 goto unlock;
463         }
464
465         dax.sector = RADIX_DAX_SECTOR(entry);
466         dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
467         spin_unlock_irq(&mapping->tree_lock);
468
469         /*
470          * We cannot hold tree_lock while calling dax_map_atomic() because it
471          * eventually calls cond_resched().
472          */
473         ret = dax_map_atomic(bdev, &dax);
474         if (ret < 0)
475                 return ret;
476
477         if (WARN_ON_ONCE(ret < dax.size)) {
478                 ret = -EIO;
479                 goto unmap;
480         }
481
482         wb_cache_pmem(dax.addr, dax.size);
483
484         spin_lock_irq(&mapping->tree_lock);
485         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
486         spin_unlock_irq(&mapping->tree_lock);
487  unmap:
488         dax_unmap_atomic(bdev, &dax);
489         return ret;
490
491  unlock:
492         spin_unlock_irq(&mapping->tree_lock);
493         return ret;
494 }
495
496 /*
497  * Flush the mapping to the persistent domain within the byte range of [start,
498  * end]. This is required by data integrity operations to ensure file data is
499  * on persistent storage prior to completion of the operation.
500  */
501 int dax_writeback_mapping_range(struct address_space *mapping,
502                 struct block_device *bdev, struct writeback_control *wbc)
503 {
504         struct inode *inode = mapping->host;
505         pgoff_t start_index, end_index, pmd_index;
506         pgoff_t indices[PAGEVEC_SIZE];
507         struct pagevec pvec;
508         bool done = false;
509         int i, ret = 0;
510         void *entry;
511
512         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
513                 return -EIO;
514
515         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
516                 return 0;
517
518         start_index = wbc->range_start >> PAGE_SHIFT;
519         end_index = wbc->range_end >> PAGE_SHIFT;
520         pmd_index = DAX_PMD_INDEX(start_index);
521
522         rcu_read_lock();
523         entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
524         rcu_read_unlock();
525
526         /* see if the start of our range is covered by a PMD entry */
527         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
528                 start_index = pmd_index;
529
530         tag_pages_for_writeback(mapping, start_index, end_index);
531
532         pagevec_init(&pvec, 0);
533         while (!done) {
534                 pvec.nr = find_get_entries_tag(mapping, start_index,
535                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
536                                 pvec.pages, indices);
537
538                 if (pvec.nr == 0)
539                         break;
540
541                 for (i = 0; i < pvec.nr; i++) {
542                         if (indices[i] > end_index) {
543                                 done = true;
544                                 break;
545                         }
546
547                         ret = dax_writeback_one(bdev, mapping, indices[i],
548                                         pvec.pages[i]);
549                         if (ret < 0)
550                                 return ret;
551                 }
552         }
553         wmb_pmem();
554         return 0;
555 }
556 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
557
558 static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
559                         struct vm_area_struct *vma, struct vm_fault *vmf)
560 {
561         unsigned long vaddr = (unsigned long)vmf->virtual_address;
562         struct address_space *mapping = inode->i_mapping;
563         struct block_device *bdev = bh->b_bdev;
564         struct blk_dax_ctl dax = {
565                 .sector = to_sector(bh, inode),
566                 .size = bh->b_size,
567         };
568         pgoff_t size;
569         int error;
570
571         i_mmap_lock_read(mapping);
572
573         /*
574          * Check truncate didn't happen while we were allocating a block.
575          * If it did, this block may or may not be still allocated to the
576          * file.  We can't tell the filesystem to free it because we can't
577          * take i_mutex here.  In the worst case, the file still has blocks
578          * allocated past the end of the file.
579          */
580         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
581         if (unlikely(vmf->pgoff >= size)) {
582                 error = -EIO;
583                 goto out;
584         }
585
586         if (dax_map_atomic(bdev, &dax) < 0) {
587                 error = PTR_ERR(dax.addr);
588                 goto out;
589         }
590
591         if (buffer_unwritten(bh) || buffer_new(bh)) {
592                 clear_pmem(dax.addr, PAGE_SIZE);
593                 wmb_pmem();
594         }
595         dax_unmap_atomic(bdev, &dax);
596
597         error = dax_radix_entry(mapping, vmf->pgoff, dax.sector, false,
598                         vmf->flags & FAULT_FLAG_WRITE);
599         if (error)
600                 goto out;
601
602         error = vm_insert_mixed(vma, vaddr, dax.pfn);
603
604  out:
605         i_mmap_unlock_read(mapping);
606
607         return error;
608 }
609
610 /**
611  * __dax_fault - handle a page fault on a DAX file
612  * @vma: The virtual memory area where the fault occurred
613  * @vmf: The description of the fault
614  * @get_block: The filesystem method used to translate file offsets to blocks
615  * @complete_unwritten: The filesystem method used to convert unwritten blocks
616  *      to written so the data written to them is exposed. This is required for
617  *      required by write faults for filesystems that will return unwritten
618  *      extent mappings from @get_block, but it is optional for reads as
619  *      dax_insert_mapping() will always zero unwritten blocks. If the fs does
620  *      not support unwritten extents, the it should pass NULL.
621  *
622  * When a page fault occurs, filesystems may call this helper in their
623  * fault handler for DAX files. __dax_fault() assumes the caller has done all
624  * the necessary locking for the page fault to proceed successfully.
625  */
626 int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
627                         get_block_t get_block, dax_iodone_t complete_unwritten)
628 {
629         struct file *file = vma->vm_file;
630         struct address_space *mapping = file->f_mapping;
631         struct inode *inode = mapping->host;
632         struct page *page;
633         struct buffer_head bh;
634         unsigned long vaddr = (unsigned long)vmf->virtual_address;
635         unsigned blkbits = inode->i_blkbits;
636         sector_t block;
637         pgoff_t size;
638         int error;
639         int major = 0;
640
641         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
642         if (vmf->pgoff >= size)
643                 return VM_FAULT_SIGBUS;
644
645         memset(&bh, 0, sizeof(bh));
646         block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
647         bh.b_bdev = inode->i_sb->s_bdev;
648         bh.b_size = PAGE_SIZE;
649
650  repeat:
651         page = find_get_page(mapping, vmf->pgoff);
652         if (page) {
653                 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
654                         put_page(page);
655                         return VM_FAULT_RETRY;
656                 }
657                 if (unlikely(page->mapping != mapping)) {
658                         unlock_page(page);
659                         put_page(page);
660                         goto repeat;
661                 }
662                 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
663                 if (unlikely(vmf->pgoff >= size)) {
664                         /*
665                          * We have a struct page covering a hole in the file
666                          * from a read fault and we've raced with a truncate
667                          */
668                         error = -EIO;
669                         goto unlock_page;
670                 }
671         }
672
673         error = get_block(inode, block, &bh, 0);
674         if (!error && (bh.b_size < PAGE_SIZE))
675                 error = -EIO;           /* fs corruption? */
676         if (error)
677                 goto unlock_page;
678
679         if (!buffer_mapped(&bh) && !buffer_unwritten(&bh) && !vmf->cow_page) {
680                 if (vmf->flags & FAULT_FLAG_WRITE) {
681                         error = get_block(inode, block, &bh, 1);
682                         count_vm_event(PGMAJFAULT);
683                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
684                         major = VM_FAULT_MAJOR;
685                         if (!error && (bh.b_size < PAGE_SIZE))
686                                 error = -EIO;
687                         if (error)
688                                 goto unlock_page;
689                 } else {
690                         return dax_load_hole(mapping, page, vmf);
691                 }
692         }
693
694         if (vmf->cow_page) {
695                 struct page *new_page = vmf->cow_page;
696                 if (buffer_written(&bh))
697                         error = copy_user_bh(new_page, inode, &bh, vaddr);
698                 else
699                         clear_user_highpage(new_page, vaddr);
700                 if (error)
701                         goto unlock_page;
702                 vmf->page = page;
703                 if (!page) {
704                         i_mmap_lock_read(mapping);
705                         /* Check we didn't race with truncate */
706                         size = (i_size_read(inode) + PAGE_SIZE - 1) >>
707                                                                 PAGE_SHIFT;
708                         if (vmf->pgoff >= size) {
709                                 i_mmap_unlock_read(mapping);
710                                 error = -EIO;
711                                 goto out;
712                         }
713                 }
714                 return VM_FAULT_LOCKED;
715         }
716
717         /* Check we didn't race with a read fault installing a new page */
718         if (!page && major)
719                 page = find_lock_page(mapping, vmf->pgoff);
720
721         if (page) {
722                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
723                                                         PAGE_SIZE, 0);
724                 delete_from_page_cache(page);
725                 unlock_page(page);
726                 put_page(page);
727                 page = NULL;
728         }
729
730         /*
731          * If we successfully insert the new mapping over an unwritten extent,
732          * we need to ensure we convert the unwritten extent. If there is an
733          * error inserting the mapping, the filesystem needs to leave it as
734          * unwritten to prevent exposure of the stale underlying data to
735          * userspace, but we still need to call the completion function so
736          * the private resources on the mapping buffer can be released. We
737          * indicate what the callback should do via the uptodate variable, same
738          * as for normal BH based IO completions.
739          */
740         error = dax_insert_mapping(inode, &bh, vma, vmf);
741         if (buffer_unwritten(&bh)) {
742                 if (complete_unwritten)
743                         complete_unwritten(&bh, !error);
744                 else
745                         WARN_ON_ONCE(!(vmf->flags & FAULT_FLAG_WRITE));
746         }
747
748  out:
749         if (error == -ENOMEM)
750                 return VM_FAULT_OOM | major;
751         /* -EBUSY is fine, somebody else faulted on the same PTE */
752         if ((error < 0) && (error != -EBUSY))
753                 return VM_FAULT_SIGBUS | major;
754         return VM_FAULT_NOPAGE | major;
755
756  unlock_page:
757         if (page) {
758                 unlock_page(page);
759                 put_page(page);
760         }
761         goto out;
762 }
763 EXPORT_SYMBOL(__dax_fault);
764
765 /**
766  * dax_fault - handle a page fault on a DAX file
767  * @vma: The virtual memory area where the fault occurred
768  * @vmf: The description of the fault
769  * @get_block: The filesystem method used to translate file offsets to blocks
770  *
771  * When a page fault occurs, filesystems may call this helper in their
772  * fault handler for DAX files.
773  */
774 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
775               get_block_t get_block, dax_iodone_t complete_unwritten)
776 {
777         int result;
778         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
779
780         if (vmf->flags & FAULT_FLAG_WRITE) {
781                 sb_start_pagefault(sb);
782                 file_update_time(vma->vm_file);
783         }
784         result = __dax_fault(vma, vmf, get_block, complete_unwritten);
785         if (vmf->flags & FAULT_FLAG_WRITE)
786                 sb_end_pagefault(sb);
787
788         return result;
789 }
790 EXPORT_SYMBOL_GPL(dax_fault);
791
792 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
793 /*
794  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
795  * more often than one might expect in the below function.
796  */
797 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
798
799 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
800                 const char *reason, const char *fn)
801 {
802         if (bh) {
803                 char bname[BDEVNAME_SIZE];
804                 bdevname(bh->b_bdev, bname);
805                 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
806                         "length %zd fallback: %s\n", fn, current->comm,
807                         address, bname, bh->b_state, (u64)bh->b_blocknr,
808                         bh->b_size, reason);
809         } else {
810                 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
811                         current->comm, address, reason);
812         }
813 }
814
815 #define dax_pmd_dbg(bh, address, reason)        __dax_dbg(bh, address, reason, "dax_pmd")
816
817 int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
818                 pmd_t *pmd, unsigned int flags, get_block_t get_block,
819                 dax_iodone_t complete_unwritten)
820 {
821         struct file *file = vma->vm_file;
822         struct address_space *mapping = file->f_mapping;
823         struct inode *inode = mapping->host;
824         struct buffer_head bh;
825         unsigned blkbits = inode->i_blkbits;
826         unsigned long pmd_addr = address & PMD_MASK;
827         bool write = flags & FAULT_FLAG_WRITE;
828         struct block_device *bdev;
829         pgoff_t size, pgoff;
830         sector_t block;
831         int error, result = 0;
832         bool alloc = false;
833
834         /* dax pmd mappings require pfn_t_devmap() */
835         if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
836                 return VM_FAULT_FALLBACK;
837
838         /* Fall back to PTEs if we're going to COW */
839         if (write && !(vma->vm_flags & VM_SHARED)) {
840                 split_huge_pmd(vma, pmd, address);
841                 dax_pmd_dbg(NULL, address, "cow write");
842                 return VM_FAULT_FALLBACK;
843         }
844         /* If the PMD would extend outside the VMA */
845         if (pmd_addr < vma->vm_start) {
846                 dax_pmd_dbg(NULL, address, "vma start unaligned");
847                 return VM_FAULT_FALLBACK;
848         }
849         if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
850                 dax_pmd_dbg(NULL, address, "vma end unaligned");
851                 return VM_FAULT_FALLBACK;
852         }
853
854         pgoff = linear_page_index(vma, pmd_addr);
855         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
856         if (pgoff >= size)
857                 return VM_FAULT_SIGBUS;
858         /* If the PMD would cover blocks out of the file */
859         if ((pgoff | PG_PMD_COLOUR) >= size) {
860                 dax_pmd_dbg(NULL, address,
861                                 "offset + huge page size > file size");
862                 return VM_FAULT_FALLBACK;
863         }
864
865         memset(&bh, 0, sizeof(bh));
866         bh.b_bdev = inode->i_sb->s_bdev;
867         block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
868
869         bh.b_size = PMD_SIZE;
870
871         if (get_block(inode, block, &bh, 0) != 0)
872                 return VM_FAULT_SIGBUS;
873
874         if (!buffer_mapped(&bh) && write) {
875                 if (get_block(inode, block, &bh, 1) != 0)
876                         return VM_FAULT_SIGBUS;
877                 alloc = true;
878         }
879
880         bdev = bh.b_bdev;
881
882         /*
883          * If the filesystem isn't willing to tell us the length of a hole,
884          * just fall back to PTEs.  Calling get_block 512 times in a loop
885          * would be silly.
886          */
887         if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
888                 dax_pmd_dbg(&bh, address, "allocated block too small");
889                 return VM_FAULT_FALLBACK;
890         }
891
892         /*
893          * If we allocated new storage, make sure no process has any
894          * zero pages covering this hole
895          */
896         if (alloc) {
897                 loff_t lstart = pgoff << PAGE_SHIFT;
898                 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
899
900                 truncate_pagecache_range(inode, lstart, lend);
901         }
902
903         i_mmap_lock_read(mapping);
904
905         /*
906          * If a truncate happened while we were allocating blocks, we may
907          * leave blocks allocated to the file that are beyond EOF.  We can't
908          * take i_mutex here, so just leave them hanging; they'll be freed
909          * when the file is deleted.
910          */
911         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
912         if (pgoff >= size) {
913                 result = VM_FAULT_SIGBUS;
914                 goto out;
915         }
916         if ((pgoff | PG_PMD_COLOUR) >= size) {
917                 dax_pmd_dbg(&bh, address,
918                                 "offset + huge page size > file size");
919                 goto fallback;
920         }
921
922         if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
923                 spinlock_t *ptl;
924                 pmd_t entry;
925                 struct page *zero_page = get_huge_zero_page();
926
927                 if (unlikely(!zero_page)) {
928                         dax_pmd_dbg(&bh, address, "no zero page");
929                         goto fallback;
930                 }
931
932                 ptl = pmd_lock(vma->vm_mm, pmd);
933                 if (!pmd_none(*pmd)) {
934                         spin_unlock(ptl);
935                         dax_pmd_dbg(&bh, address, "pmd already present");
936                         goto fallback;
937                 }
938
939                 dev_dbg(part_to_dev(bdev->bd_part),
940                                 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
941                                 __func__, current->comm, address,
942                                 (unsigned long long) to_sector(&bh, inode));
943
944                 entry = mk_pmd(zero_page, vma->vm_page_prot);
945                 entry = pmd_mkhuge(entry);
946                 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
947                 result = VM_FAULT_NOPAGE;
948                 spin_unlock(ptl);
949         } else {
950                 struct blk_dax_ctl dax = {
951                         .sector = to_sector(&bh, inode),
952                         .size = PMD_SIZE,
953                 };
954                 long length = dax_map_atomic(bdev, &dax);
955
956                 if (length < 0) {
957                         result = VM_FAULT_SIGBUS;
958                         goto out;
959                 }
960                 if (length < PMD_SIZE) {
961                         dax_pmd_dbg(&bh, address, "dax-length too small");
962                         dax_unmap_atomic(bdev, &dax);
963                         goto fallback;
964                 }
965                 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
966                         dax_pmd_dbg(&bh, address, "pfn unaligned");
967                         dax_unmap_atomic(bdev, &dax);
968                         goto fallback;
969                 }
970
971                 if (!pfn_t_devmap(dax.pfn)) {
972                         dax_unmap_atomic(bdev, &dax);
973                         dax_pmd_dbg(&bh, address, "pfn not in memmap");
974                         goto fallback;
975                 }
976
977                 if (buffer_unwritten(&bh) || buffer_new(&bh)) {
978                         clear_pmem(dax.addr, PMD_SIZE);
979                         wmb_pmem();
980                         count_vm_event(PGMAJFAULT);
981                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
982                         result |= VM_FAULT_MAJOR;
983                 }
984                 dax_unmap_atomic(bdev, &dax);
985
986                 /*
987                  * For PTE faults we insert a radix tree entry for reads, and
988                  * leave it clean.  Then on the first write we dirty the radix
989                  * tree entry via the dax_pfn_mkwrite() path.  This sequence
990                  * allows the dax_pfn_mkwrite() call to be simpler and avoid a
991                  * call into get_block() to translate the pgoff to a sector in
992                  * order to be able to create a new radix tree entry.
993                  *
994                  * The PMD path doesn't have an equivalent to
995                  * dax_pfn_mkwrite(), though, so for a read followed by a
996                  * write we traverse all the way through __dax_pmd_fault()
997                  * twice.  This means we can just skip inserting a radix tree
998                  * entry completely on the initial read and just wait until
999                  * the write to insert a dirty entry.
1000                  */
1001                 if (write) {
1002                         error = dax_radix_entry(mapping, pgoff, dax.sector,
1003                                         true, true);
1004                         if (error) {
1005                                 dax_pmd_dbg(&bh, address,
1006                                                 "PMD radix insertion failed");
1007                                 goto fallback;
1008                         }
1009                 }
1010
1011                 dev_dbg(part_to_dev(bdev->bd_part),
1012                                 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
1013                                 __func__, current->comm, address,
1014                                 pfn_t_to_pfn(dax.pfn),
1015                                 (unsigned long long) dax.sector);
1016                 result |= vmf_insert_pfn_pmd(vma, address, pmd,
1017                                 dax.pfn, write);
1018         }
1019
1020  out:
1021         i_mmap_unlock_read(mapping);
1022
1023         if (buffer_unwritten(&bh))
1024                 complete_unwritten(&bh, !(result & VM_FAULT_ERROR));
1025
1026         return result;
1027
1028  fallback:
1029         count_vm_event(THP_FAULT_FALLBACK);
1030         result = VM_FAULT_FALLBACK;
1031         goto out;
1032 }
1033 EXPORT_SYMBOL_GPL(__dax_pmd_fault);
1034
1035 /**
1036  * dax_pmd_fault - handle a PMD fault on a DAX file
1037  * @vma: The virtual memory area where the fault occurred
1038  * @vmf: The description of the fault
1039  * @get_block: The filesystem method used to translate file offsets to blocks
1040  *
1041  * When a page fault occurs, filesystems may call this helper in their
1042  * pmd_fault handler for DAX files.
1043  */
1044 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1045                         pmd_t *pmd, unsigned int flags, get_block_t get_block,
1046                         dax_iodone_t complete_unwritten)
1047 {
1048         int result;
1049         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
1050
1051         if (flags & FAULT_FLAG_WRITE) {
1052                 sb_start_pagefault(sb);
1053                 file_update_time(vma->vm_file);
1054         }
1055         result = __dax_pmd_fault(vma, address, pmd, flags, get_block,
1056                                 complete_unwritten);
1057         if (flags & FAULT_FLAG_WRITE)
1058                 sb_end_pagefault(sb);
1059
1060         return result;
1061 }
1062 EXPORT_SYMBOL_GPL(dax_pmd_fault);
1063 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1064
1065 /**
1066  * dax_pfn_mkwrite - handle first write to DAX page
1067  * @vma: The virtual memory area where the fault occurred
1068  * @vmf: The description of the fault
1069  */
1070 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1071 {
1072         struct file *file = vma->vm_file;
1073         int error;
1074
1075         /*
1076          * We pass NO_SECTOR to dax_radix_entry() because we expect that a
1077          * RADIX_DAX_PTE entry already exists in the radix tree from a
1078          * previous call to __dax_fault().  We just want to look up that PTE
1079          * entry using vmf->pgoff and make sure the dirty tag is set.  This
1080          * saves us from having to make a call to get_block() here to look
1081          * up the sector.
1082          */
1083         error = dax_radix_entry(file->f_mapping, vmf->pgoff, NO_SECTOR, false,
1084                         true);
1085
1086         if (error == -ENOMEM)
1087                 return VM_FAULT_OOM;
1088         if (error)
1089                 return VM_FAULT_SIGBUS;
1090         return VM_FAULT_NOPAGE;
1091 }
1092 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1093
1094 /**
1095  * dax_zero_page_range - zero a range within a page of a DAX file
1096  * @inode: The file being truncated
1097  * @from: The file offset that is being truncated to
1098  * @length: The number of bytes to zero
1099  * @get_block: The filesystem method used to translate file offsets to blocks
1100  *
1101  * This function can be called by a filesystem when it is zeroing part of a
1102  * page in a DAX file.  This is intended for hole-punch operations.  If
1103  * you are truncating a file, the helper function dax_truncate_page() may be
1104  * more convenient.
1105  *
1106  * We work in terms of PAGE_SIZE here for commonality with
1107  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1108  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1109  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1110  * since the file might be mmapped.
1111  */
1112 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1113                                                         get_block_t get_block)
1114 {
1115         struct buffer_head bh;
1116         pgoff_t index = from >> PAGE_SHIFT;
1117         unsigned offset = from & (PAGE_SIZE-1);
1118         int err;
1119
1120         /* Block boundary? Nothing to do */
1121         if (!length)
1122                 return 0;
1123         BUG_ON((offset + length) > PAGE_SIZE);
1124
1125         memset(&bh, 0, sizeof(bh));
1126         bh.b_bdev = inode->i_sb->s_bdev;
1127         bh.b_size = PAGE_SIZE;
1128         err = get_block(inode, index, &bh, 0);
1129         if (err < 0)
1130                 return err;
1131         if (buffer_written(&bh)) {
1132                 struct block_device *bdev = bh.b_bdev;
1133                 struct blk_dax_ctl dax = {
1134                         .sector = to_sector(&bh, inode),
1135                         .size = PAGE_SIZE,
1136                 };
1137
1138                 if (dax_map_atomic(bdev, &dax) < 0)
1139                         return PTR_ERR(dax.addr);
1140                 clear_pmem(dax.addr + offset, length);
1141                 wmb_pmem();
1142                 dax_unmap_atomic(bdev, &dax);
1143         }
1144
1145         return 0;
1146 }
1147 EXPORT_SYMBOL_GPL(dax_zero_page_range);
1148
1149 /**
1150  * dax_truncate_page - handle a partial page being truncated in a DAX file
1151  * @inode: The file being truncated
1152  * @from: The file offset that is being truncated to
1153  * @get_block: The filesystem method used to translate file offsets to blocks
1154  *
1155  * Similar to block_truncate_page(), this function can be called by a
1156  * filesystem when it is truncating a DAX file to handle the partial page.
1157  *
1158  * We work in terms of PAGE_SIZE here for commonality with
1159  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1160  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1161  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1162  * since the file might be mmapped.
1163  */
1164 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1165 {
1166         unsigned length = PAGE_ALIGN(from) - from;
1167         return dax_zero_page_range(inode, from, length, get_block);
1168 }
1169 EXPORT_SYMBOL_GPL(dax_truncate_page);