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/genhd.h>
15 #include <linux/highmem.h>
16 #include <linux/memcontrol.h>
17 #include <linux/mm.h>
18 #include <linux/mutex.h>
19 #include <linux/pagevec.h>
20 #include <linux/sched.h>
21 #include <linux/sched/signal.h>
22 #include <linux/uio.h>
23 #include <linux/vmstat.h>
24 #include <linux/pfn_t.h>
25 #include <linux/sizes.h>
26 #include <linux/mmu_notifier.h>
27 #include <linux/iomap.h>
28 #include <asm/pgalloc.h>
29 
30 #define CREATE_TRACE_POINTS
31 #include <trace/events/fs_dax.h>
32 
pe_order(enum page_entry_size pe_size)33 static inline unsigned int pe_order(enum page_entry_size pe_size)
34 {
35 	if (pe_size == PE_SIZE_PTE)
36 		return PAGE_SHIFT - PAGE_SHIFT;
37 	if (pe_size == PE_SIZE_PMD)
38 		return PMD_SHIFT - PAGE_SHIFT;
39 	if (pe_size == PE_SIZE_PUD)
40 		return PUD_SHIFT - PAGE_SHIFT;
41 	return ~0;
42 }
43 
44 /* We choose 4096 entries - same as per-zone page wait tables */
45 #define DAX_WAIT_TABLE_BITS 12
46 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
47 
48 /* The 'colour' (ie low bits) within a PMD of a page offset.  */
49 #define PG_PMD_COLOUR	((PMD_SIZE >> PAGE_SHIFT) - 1)
50 #define PG_PMD_NR	(PMD_SIZE >> PAGE_SHIFT)
51 
52 /* The order of a PMD entry */
53 #define PMD_ORDER	(PMD_SHIFT - PAGE_SHIFT)
54 
55 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
56 
init_dax_wait_table(void)57 static int __init init_dax_wait_table(void)
58 {
59 	int i;
60 
61 	for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
62 		init_waitqueue_head(wait_table + i);
63 	return 0;
64 }
65 fs_initcall(init_dax_wait_table);
66 
67 /*
68  * DAX pagecache entries use XArray value entries so they can't be mistaken
69  * for pages.  We use one bit for locking, one bit for the entry size (PMD)
70  * and two more to tell us if the entry is a zero page or an empty entry that
71  * is just used for locking.  In total four special bits.
72  *
73  * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
74  * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
75  * block allocation.
76  */
77 #define DAX_SHIFT	(4)
78 #define DAX_LOCKED	(1UL << 0)
79 #define DAX_PMD		(1UL << 1)
80 #define DAX_ZERO_PAGE	(1UL << 2)
81 #define DAX_EMPTY	(1UL << 3)
82 
dax_to_pfn(void * entry)83 static unsigned long dax_to_pfn(void *entry)
84 {
85 	return xa_to_value(entry) >> DAX_SHIFT;
86 }
87 
dax_make_entry(pfn_t pfn,unsigned long flags)88 static void *dax_make_entry(pfn_t pfn, unsigned long flags)
89 {
90 	return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
91 }
92 
dax_is_locked(void * entry)93 static bool dax_is_locked(void *entry)
94 {
95 	return xa_to_value(entry) & DAX_LOCKED;
96 }
97 
dax_entry_order(void * entry)98 static unsigned int dax_entry_order(void *entry)
99 {
100 	if (xa_to_value(entry) & DAX_PMD)
101 		return PMD_ORDER;
102 	return 0;
103 }
104 
dax_is_pmd_entry(void * entry)105 static unsigned long dax_is_pmd_entry(void *entry)
106 {
107 	return xa_to_value(entry) & DAX_PMD;
108 }
109 
dax_is_pte_entry(void * entry)110 static bool dax_is_pte_entry(void *entry)
111 {
112 	return !(xa_to_value(entry) & DAX_PMD);
113 }
114 
dax_is_zero_entry(void * entry)115 static int dax_is_zero_entry(void *entry)
116 {
117 	return xa_to_value(entry) & DAX_ZERO_PAGE;
118 }
119 
dax_is_empty_entry(void * entry)120 static int dax_is_empty_entry(void *entry)
121 {
122 	return xa_to_value(entry) & DAX_EMPTY;
123 }
124 
125 /*
126  * true if the entry that was found is of a smaller order than the entry
127  * we were looking for
128  */
dax_is_conflict(void * entry)129 static bool dax_is_conflict(void *entry)
130 {
131 	return entry == XA_RETRY_ENTRY;
132 }
133 
134 /*
135  * DAX page cache entry locking
136  */
137 struct exceptional_entry_key {
138 	struct xarray *xa;
139 	pgoff_t entry_start;
140 };
141 
142 struct wait_exceptional_entry_queue {
143 	wait_queue_entry_t wait;
144 	struct exceptional_entry_key key;
145 };
146 
147 /**
148  * enum dax_wake_mode: waitqueue wakeup behaviour
149  * @WAKE_ALL: wake all waiters in the waitqueue
150  * @WAKE_NEXT: wake only the first waiter in the waitqueue
151  */
152 enum dax_wake_mode {
153 	WAKE_ALL,
154 	WAKE_NEXT,
155 };
156 
dax_entry_waitqueue(struct xa_state * xas,void * entry,struct exceptional_entry_key * key)157 static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas,
158 		void *entry, struct exceptional_entry_key *key)
159 {
160 	unsigned long hash;
161 	unsigned long index = xas->xa_index;
162 
163 	/*
164 	 * If 'entry' is a PMD, align the 'index' that we use for the wait
165 	 * queue to the start of that PMD.  This ensures that all offsets in
166 	 * the range covered by the PMD map to the same bit lock.
167 	 */
168 	if (dax_is_pmd_entry(entry))
169 		index &= ~PG_PMD_COLOUR;
170 	key->xa = xas->xa;
171 	key->entry_start = index;
172 
173 	hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS);
174 	return wait_table + hash;
175 }
176 
wake_exceptional_entry_func(wait_queue_entry_t * wait,unsigned int mode,int sync,void * keyp)177 static int wake_exceptional_entry_func(wait_queue_entry_t *wait,
178 		unsigned int mode, int sync, void *keyp)
179 {
180 	struct exceptional_entry_key *key = keyp;
181 	struct wait_exceptional_entry_queue *ewait =
182 		container_of(wait, struct wait_exceptional_entry_queue, wait);
183 
184 	if (key->xa != ewait->key.xa ||
185 	    key->entry_start != ewait->key.entry_start)
186 		return 0;
187 	return autoremove_wake_function(wait, mode, sync, NULL);
188 }
189 
190 /*
191  * @entry may no longer be the entry at the index in the mapping.
192  * The important information it's conveying is whether the entry at
193  * this index used to be a PMD entry.
194  */
dax_wake_entry(struct xa_state * xas,void * entry,enum dax_wake_mode mode)195 static void dax_wake_entry(struct xa_state *xas, void *entry,
196 			   enum dax_wake_mode mode)
197 {
198 	struct exceptional_entry_key key;
199 	wait_queue_head_t *wq;
200 
201 	wq = dax_entry_waitqueue(xas, entry, &key);
202 
203 	/*
204 	 * Checking for locked entry and prepare_to_wait_exclusive() happens
205 	 * under the i_pages lock, ditto for entry handling in our callers.
206 	 * So at this point all tasks that could have seen our entry locked
207 	 * must be in the waitqueue and the following check will see them.
208 	 */
209 	if (waitqueue_active(wq))
210 		__wake_up(wq, TASK_NORMAL, mode == WAKE_ALL ? 0 : 1, &key);
211 }
212 
213 /*
214  * Look up entry in page cache, wait for it to become unlocked if it
215  * is a DAX entry and return it.  The caller must subsequently call
216  * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
217  * if it did.  The entry returned may have a larger order than @order.
218  * If @order is larger than the order of the entry found in i_pages, this
219  * function returns a dax_is_conflict entry.
220  *
221  * Must be called with the i_pages lock held.
222  */
get_unlocked_entry(struct xa_state * xas,unsigned int order)223 static void *get_unlocked_entry(struct xa_state *xas, unsigned int order)
224 {
225 	void *entry;
226 	struct wait_exceptional_entry_queue ewait;
227 	wait_queue_head_t *wq;
228 
229 	init_wait(&ewait.wait);
230 	ewait.wait.func = wake_exceptional_entry_func;
231 
232 	for (;;) {
233 		entry = xas_find_conflict(xas);
234 		if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
235 			return entry;
236 		if (dax_entry_order(entry) < order)
237 			return XA_RETRY_ENTRY;
238 		if (!dax_is_locked(entry))
239 			return entry;
240 
241 		wq = dax_entry_waitqueue(xas, entry, &ewait.key);
242 		prepare_to_wait_exclusive(wq, &ewait.wait,
243 					  TASK_UNINTERRUPTIBLE);
244 		xas_unlock_irq(xas);
245 		xas_reset(xas);
246 		schedule();
247 		finish_wait(wq, &ewait.wait);
248 		xas_lock_irq(xas);
249 	}
250 }
251 
252 /*
253  * The only thing keeping the address space around is the i_pages lock
254  * (it's cycled in clear_inode() after removing the entries from i_pages)
255  * After we call xas_unlock_irq(), we cannot touch xas->xa.
256  */
wait_entry_unlocked(struct xa_state * xas,void * entry)257 static void wait_entry_unlocked(struct xa_state *xas, void *entry)
258 {
259 	struct wait_exceptional_entry_queue ewait;
260 	wait_queue_head_t *wq;
261 
262 	init_wait(&ewait.wait);
263 	ewait.wait.func = wake_exceptional_entry_func;
264 
265 	wq = dax_entry_waitqueue(xas, entry, &ewait.key);
266 	/*
267 	 * Unlike get_unlocked_entry() there is no guarantee that this
268 	 * path ever successfully retrieves an unlocked entry before an
269 	 * inode dies. Perform a non-exclusive wait in case this path
270 	 * never successfully performs its own wake up.
271 	 */
272 	prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
273 	xas_unlock_irq(xas);
274 	schedule();
275 	finish_wait(wq, &ewait.wait);
276 }
277 
put_unlocked_entry(struct xa_state * xas,void * entry,enum dax_wake_mode mode)278 static void put_unlocked_entry(struct xa_state *xas, void *entry,
279 			       enum dax_wake_mode mode)
280 {
281 	if (entry && !dax_is_conflict(entry))
282 		dax_wake_entry(xas, entry, mode);
283 }
284 
285 /*
286  * We used the xa_state to get the entry, but then we locked the entry and
287  * dropped the xa_lock, so we know the xa_state is stale and must be reset
288  * before use.
289  */
dax_unlock_entry(struct xa_state * xas,void * entry)290 static void dax_unlock_entry(struct xa_state *xas, void *entry)
291 {
292 	void *old;
293 
294 	BUG_ON(dax_is_locked(entry));
295 	xas_reset(xas);
296 	xas_lock_irq(xas);
297 	old = xas_store(xas, entry);
298 	xas_unlock_irq(xas);
299 	BUG_ON(!dax_is_locked(old));
300 	dax_wake_entry(xas, entry, WAKE_NEXT);
301 }
302 
303 /*
304  * Return: The entry stored at this location before it was locked.
305  */
dax_lock_entry(struct xa_state * xas,void * entry)306 static void *dax_lock_entry(struct xa_state *xas, void *entry)
307 {
308 	unsigned long v = xa_to_value(entry);
309 	return xas_store(xas, xa_mk_value(v | DAX_LOCKED));
310 }
311 
dax_entry_size(void * entry)312 static unsigned long dax_entry_size(void *entry)
313 {
314 	if (dax_is_zero_entry(entry))
315 		return 0;
316 	else if (dax_is_empty_entry(entry))
317 		return 0;
318 	else if (dax_is_pmd_entry(entry))
319 		return PMD_SIZE;
320 	else
321 		return PAGE_SIZE;
322 }
323 
dax_end_pfn(void * entry)324 static unsigned long dax_end_pfn(void *entry)
325 {
326 	return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
327 }
328 
329 /*
330  * Iterate through all mapped pfns represented by an entry, i.e. skip
331  * 'empty' and 'zero' entries.
332  */
333 #define for_each_mapped_pfn(entry, pfn) \
334 	for (pfn = dax_to_pfn(entry); \
335 			pfn < dax_end_pfn(entry); pfn++)
336 
337 /*
338  * TODO: for reflink+dax we need a way to associate a single page with
339  * multiple address_space instances at different linear_page_index()
340  * offsets.
341  */
dax_associate_entry(void * entry,struct address_space * mapping,struct vm_area_struct * vma,unsigned long address)342 static void dax_associate_entry(void *entry, struct address_space *mapping,
343 		struct vm_area_struct *vma, unsigned long address)
344 {
345 	unsigned long size = dax_entry_size(entry), pfn, index;
346 	int i = 0;
347 
348 	if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
349 		return;
350 
351 	index = linear_page_index(vma, address & ~(size - 1));
352 	for_each_mapped_pfn(entry, pfn) {
353 		struct page *page = pfn_to_page(pfn);
354 
355 		WARN_ON_ONCE(page->mapping);
356 		page->mapping = mapping;
357 		page->index = index + i++;
358 	}
359 }
360 
dax_disassociate_entry(void * entry,struct address_space * mapping,bool trunc)361 static void dax_disassociate_entry(void *entry, struct address_space *mapping,
362 		bool trunc)
363 {
364 	unsigned long pfn;
365 
366 	if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
367 		return;
368 
369 	for_each_mapped_pfn(entry, pfn) {
370 		struct page *page = pfn_to_page(pfn);
371 
372 		WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
373 		WARN_ON_ONCE(page->mapping && page->mapping != mapping);
374 		page->mapping = NULL;
375 		page->index = 0;
376 	}
377 }
378 
dax_busy_page(void * entry)379 static struct page *dax_busy_page(void *entry)
380 {
381 	unsigned long pfn;
382 
383 	for_each_mapped_pfn(entry, pfn) {
384 		struct page *page = pfn_to_page(pfn);
385 
386 		if (page_ref_count(page) > 1)
387 			return page;
388 	}
389 	return NULL;
390 }
391 
392 /*
393  * dax_lock_mapping_entry - Lock the DAX entry corresponding to a page
394  * @page: The page whose entry we want to lock
395  *
396  * Context: Process context.
397  * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could
398  * not be locked.
399  */
dax_lock_page(struct page * page)400 dax_entry_t dax_lock_page(struct page *page)
401 {
402 	XA_STATE(xas, NULL, 0);
403 	void *entry;
404 
405 	/* Ensure page->mapping isn't freed while we look at it */
406 	rcu_read_lock();
407 	for (;;) {
408 		struct address_space *mapping = READ_ONCE(page->mapping);
409 
410 		entry = NULL;
411 		if (!mapping || !dax_mapping(mapping))
412 			break;
413 
414 		/*
415 		 * In the device-dax case there's no need to lock, a
416 		 * struct dev_pagemap pin is sufficient to keep the
417 		 * inode alive, and we assume we have dev_pagemap pin
418 		 * otherwise we would not have a valid pfn_to_page()
419 		 * translation.
420 		 */
421 		entry = (void *)~0UL;
422 		if (S_ISCHR(mapping->host->i_mode))
423 			break;
424 
425 		xas.xa = &mapping->i_pages;
426 		xas_lock_irq(&xas);
427 		if (mapping != page->mapping) {
428 			xas_unlock_irq(&xas);
429 			continue;
430 		}
431 		xas_set(&xas, page->index);
432 		entry = xas_load(&xas);
433 		if (dax_is_locked(entry)) {
434 			rcu_read_unlock();
435 			wait_entry_unlocked(&xas, entry);
436 			rcu_read_lock();
437 			continue;
438 		}
439 		dax_lock_entry(&xas, entry);
440 		xas_unlock_irq(&xas);
441 		break;
442 	}
443 	rcu_read_unlock();
444 	return (dax_entry_t)entry;
445 }
446 
dax_unlock_page(struct page * page,dax_entry_t cookie)447 void dax_unlock_page(struct page *page, dax_entry_t cookie)
448 {
449 	struct address_space *mapping = page->mapping;
450 	XA_STATE(xas, &mapping->i_pages, page->index);
451 
452 	if (S_ISCHR(mapping->host->i_mode))
453 		return;
454 
455 	dax_unlock_entry(&xas, (void *)cookie);
456 }
457 
458 /*
459  * Find page cache entry at given index. If it is a DAX entry, return it
460  * with the entry locked. If the page cache doesn't contain an entry at
461  * that index, add a locked empty entry.
462  *
463  * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
464  * either return that locked entry or will return VM_FAULT_FALLBACK.
465  * This will happen if there are any PTE entries within the PMD range
466  * that we are requesting.
467  *
468  * We always favor PTE entries over PMD entries. There isn't a flow where we
469  * evict PTE entries in order to 'upgrade' them to a PMD entry.  A PMD
470  * insertion will fail if it finds any PTE entries already in the tree, and a
471  * PTE insertion will cause an existing PMD entry to be unmapped and
472  * downgraded to PTE entries.  This happens for both PMD zero pages as
473  * well as PMD empty entries.
474  *
475  * The exception to this downgrade path is for PMD entries that have
476  * real storage backing them.  We will leave these real PMD entries in
477  * the tree, and PTE writes will simply dirty the entire PMD entry.
478  *
479  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
480  * persistent memory the benefit is doubtful. We can add that later if we can
481  * show it helps.
482  *
483  * On error, this function does not return an ERR_PTR.  Instead it returns
484  * a VM_FAULT code, encoded as an xarray internal entry.  The ERR_PTR values
485  * overlap with xarray value entries.
486  */
grab_mapping_entry(struct xa_state * xas,struct address_space * mapping,unsigned int order)487 static void *grab_mapping_entry(struct xa_state *xas,
488 		struct address_space *mapping, unsigned int order)
489 {
490 	unsigned long index = xas->xa_index;
491 	bool pmd_downgrade = false; /* splitting PMD entry into PTE entries? */
492 	void *entry;
493 
494 retry:
495 	xas_lock_irq(xas);
496 	entry = get_unlocked_entry(xas, order);
497 
498 	if (entry) {
499 		if (dax_is_conflict(entry))
500 			goto fallback;
501 		if (!xa_is_value(entry)) {
502 			xas_set_err(xas, -EIO);
503 			goto out_unlock;
504 		}
505 
506 		if (order == 0) {
507 			if (dax_is_pmd_entry(entry) &&
508 			    (dax_is_zero_entry(entry) ||
509 			     dax_is_empty_entry(entry))) {
510 				pmd_downgrade = true;
511 			}
512 		}
513 	}
514 
515 	if (pmd_downgrade) {
516 		/*
517 		 * Make sure 'entry' remains valid while we drop
518 		 * the i_pages lock.
519 		 */
520 		dax_lock_entry(xas, entry);
521 
522 		/*
523 		 * Besides huge zero pages the only other thing that gets
524 		 * downgraded are empty entries which don't need to be
525 		 * unmapped.
526 		 */
527 		if (dax_is_zero_entry(entry)) {
528 			xas_unlock_irq(xas);
529 			unmap_mapping_pages(mapping,
530 					xas->xa_index & ~PG_PMD_COLOUR,
531 					PG_PMD_NR, false);
532 			xas_reset(xas);
533 			xas_lock_irq(xas);
534 		}
535 
536 		dax_disassociate_entry(entry, mapping, false);
537 		xas_store(xas, NULL);	/* undo the PMD join */
538 		dax_wake_entry(xas, entry, WAKE_ALL);
539 		mapping->nrpages -= PG_PMD_NR;
540 		entry = NULL;
541 		xas_set(xas, index);
542 	}
543 
544 	if (entry) {
545 		dax_lock_entry(xas, entry);
546 	} else {
547 		unsigned long flags = DAX_EMPTY;
548 
549 		if (order > 0)
550 			flags |= DAX_PMD;
551 		entry = dax_make_entry(pfn_to_pfn_t(0), flags);
552 		dax_lock_entry(xas, entry);
553 		if (xas_error(xas))
554 			goto out_unlock;
555 		mapping->nrpages += 1UL << order;
556 	}
557 
558 out_unlock:
559 	xas_unlock_irq(xas);
560 	if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM))
561 		goto retry;
562 	if (xas->xa_node == XA_ERROR(-ENOMEM))
563 		return xa_mk_internal(VM_FAULT_OOM);
564 	if (xas_error(xas))
565 		return xa_mk_internal(VM_FAULT_SIGBUS);
566 	return entry;
567 fallback:
568 	xas_unlock_irq(xas);
569 	return xa_mk_internal(VM_FAULT_FALLBACK);
570 }
571 
572 /**
573  * dax_layout_busy_page_range - find first pinned page in @mapping
574  * @mapping: address space to scan for a page with ref count > 1
575  * @start: Starting offset. Page containing 'start' is included.
576  * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX,
577  *       pages from 'start' till the end of file are included.
578  *
579  * DAX requires ZONE_DEVICE mapped pages. These pages are never
580  * 'onlined' to the page allocator so they are considered idle when
581  * page->count == 1. A filesystem uses this interface to determine if
582  * any page in the mapping is busy, i.e. for DMA, or other
583  * get_user_pages() usages.
584  *
585  * It is expected that the filesystem is holding locks to block the
586  * establishment of new mappings in this address_space. I.e. it expects
587  * to be able to run unmap_mapping_range() and subsequently not race
588  * mapping_mapped() becoming true.
589  */
dax_layout_busy_page_range(struct address_space * mapping,loff_t start,loff_t end)590 struct page *dax_layout_busy_page_range(struct address_space *mapping,
591 					loff_t start, loff_t end)
592 {
593 	void *entry;
594 	unsigned int scanned = 0;
595 	struct page *page = NULL;
596 	pgoff_t start_idx = start >> PAGE_SHIFT;
597 	pgoff_t end_idx;
598 	XA_STATE(xas, &mapping->i_pages, start_idx);
599 
600 	/*
601 	 * In the 'limited' case get_user_pages() for dax is disabled.
602 	 */
603 	if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
604 		return NULL;
605 
606 	if (!dax_mapping(mapping) || !mapping_mapped(mapping))
607 		return NULL;
608 
609 	/* If end == LLONG_MAX, all pages from start to till end of file */
610 	if (end == LLONG_MAX)
611 		end_idx = ULONG_MAX;
612 	else
613 		end_idx = end >> PAGE_SHIFT;
614 	/*
615 	 * If we race get_user_pages_fast() here either we'll see the
616 	 * elevated page count in the iteration and wait, or
617 	 * get_user_pages_fast() will see that the page it took a reference
618 	 * against is no longer mapped in the page tables and bail to the
619 	 * get_user_pages() slow path.  The slow path is protected by
620 	 * pte_lock() and pmd_lock(). New references are not taken without
621 	 * holding those locks, and unmap_mapping_pages() will not zero the
622 	 * pte or pmd without holding the respective lock, so we are
623 	 * guaranteed to either see new references or prevent new
624 	 * references from being established.
625 	 */
626 	unmap_mapping_pages(mapping, start_idx, end_idx - start_idx + 1, 0);
627 
628 	xas_lock_irq(&xas);
629 	xas_for_each(&xas, entry, end_idx) {
630 		if (WARN_ON_ONCE(!xa_is_value(entry)))
631 			continue;
632 		if (unlikely(dax_is_locked(entry)))
633 			entry = get_unlocked_entry(&xas, 0);
634 		if (entry)
635 			page = dax_busy_page(entry);
636 		put_unlocked_entry(&xas, entry, WAKE_NEXT);
637 		if (page)
638 			break;
639 		if (++scanned % XA_CHECK_SCHED)
640 			continue;
641 
642 		xas_pause(&xas);
643 		xas_unlock_irq(&xas);
644 		cond_resched();
645 		xas_lock_irq(&xas);
646 	}
647 	xas_unlock_irq(&xas);
648 	return page;
649 }
650 EXPORT_SYMBOL_GPL(dax_layout_busy_page_range);
651 
dax_layout_busy_page(struct address_space * mapping)652 struct page *dax_layout_busy_page(struct address_space *mapping)
653 {
654 	return dax_layout_busy_page_range(mapping, 0, LLONG_MAX);
655 }
656 EXPORT_SYMBOL_GPL(dax_layout_busy_page);
657 
__dax_invalidate_entry(struct address_space * mapping,pgoff_t index,bool trunc)658 static int __dax_invalidate_entry(struct address_space *mapping,
659 					  pgoff_t index, bool trunc)
660 {
661 	XA_STATE(xas, &mapping->i_pages, index);
662 	int ret = 0;
663 	void *entry;
664 
665 	xas_lock_irq(&xas);
666 	entry = get_unlocked_entry(&xas, 0);
667 	if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
668 		goto out;
669 	if (!trunc &&
670 	    (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) ||
671 	     xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE)))
672 		goto out;
673 	dax_disassociate_entry(entry, mapping, trunc);
674 	xas_store(&xas, NULL);
675 	mapping->nrpages -= 1UL << dax_entry_order(entry);
676 	ret = 1;
677 out:
678 	put_unlocked_entry(&xas, entry, WAKE_ALL);
679 	xas_unlock_irq(&xas);
680 	return ret;
681 }
682 
683 /*
684  * Delete DAX entry at @index from @mapping.  Wait for it
685  * to be unlocked before deleting it.
686  */
dax_delete_mapping_entry(struct address_space * mapping,pgoff_t index)687 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
688 {
689 	int ret = __dax_invalidate_entry(mapping, index, true);
690 
691 	/*
692 	 * This gets called from truncate / punch_hole path. As such, the caller
693 	 * must hold locks protecting against concurrent modifications of the
694 	 * page cache (usually fs-private i_mmap_sem for writing). Since the
695 	 * caller has seen a DAX entry for this index, we better find it
696 	 * at that index as well...
697 	 */
698 	WARN_ON_ONCE(!ret);
699 	return ret;
700 }
701 
702 /*
703  * Invalidate DAX entry if it is clean.
704  */
dax_invalidate_mapping_entry_sync(struct address_space * mapping,pgoff_t index)705 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
706 				      pgoff_t index)
707 {
708 	return __dax_invalidate_entry(mapping, index, false);
709 }
710 
copy_cow_page_dax(struct block_device * bdev,struct dax_device * dax_dev,sector_t sector,struct page * to,unsigned long vaddr)711 static int copy_cow_page_dax(struct block_device *bdev, struct dax_device *dax_dev,
712 			     sector_t sector, struct page *to, unsigned long vaddr)
713 {
714 	void *vto, *kaddr;
715 	pgoff_t pgoff;
716 	long rc;
717 	int id;
718 
719 	rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
720 	if (rc)
721 		return rc;
722 
723 	id = dax_read_lock();
724 	rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(PAGE_SIZE), &kaddr, NULL);
725 	if (rc < 0) {
726 		dax_read_unlock(id);
727 		return rc;
728 	}
729 	vto = kmap_atomic(to);
730 	copy_user_page(vto, (void __force *)kaddr, vaddr, to);
731 	kunmap_atomic(vto);
732 	dax_read_unlock(id);
733 	return 0;
734 }
735 
736 /*
737  * By this point grab_mapping_entry() has ensured that we have a locked entry
738  * of the appropriate size so we don't have to worry about downgrading PMDs to
739  * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
740  * already in the tree, we will skip the insertion and just dirty the PMD as
741  * appropriate.
742  */
dax_insert_entry(struct xa_state * xas,struct address_space * mapping,struct vm_fault * vmf,void * entry,pfn_t pfn,unsigned long flags,bool dirty)743 static void *dax_insert_entry(struct xa_state *xas,
744 		struct address_space *mapping, struct vm_fault *vmf,
745 		void *entry, pfn_t pfn, unsigned long flags, bool dirty)
746 {
747 	void *new_entry = dax_make_entry(pfn, flags);
748 
749 	if (dirty)
750 		__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
751 
752 	if (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE)) {
753 		unsigned long index = xas->xa_index;
754 		/* we are replacing a zero page with block mapping */
755 		if (dax_is_pmd_entry(entry))
756 			unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
757 					PG_PMD_NR, false);
758 		else /* pte entry */
759 			unmap_mapping_pages(mapping, index, 1, false);
760 	}
761 
762 	xas_reset(xas);
763 	xas_lock_irq(xas);
764 	if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
765 		void *old;
766 
767 		dax_disassociate_entry(entry, mapping, false);
768 		dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address);
769 		/*
770 		 * Only swap our new entry into the page cache if the current
771 		 * entry is a zero page or an empty entry.  If a normal PTE or
772 		 * PMD entry is already in the cache, we leave it alone.  This
773 		 * means that if we are trying to insert a PTE and the
774 		 * existing entry is a PMD, we will just leave the PMD in the
775 		 * tree and dirty it if necessary.
776 		 */
777 		old = dax_lock_entry(xas, new_entry);
778 		WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) |
779 					DAX_LOCKED));
780 		entry = new_entry;
781 	} else {
782 		xas_load(xas);	/* Walk the xa_state */
783 	}
784 
785 	if (dirty)
786 		xas_set_mark(xas, PAGECACHE_TAG_DIRTY);
787 
788 	xas_unlock_irq(xas);
789 	return entry;
790 }
791 
792 static inline
pgoff_address(pgoff_t pgoff,struct vm_area_struct * vma)793 unsigned long pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
794 {
795 	unsigned long address;
796 
797 	address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
798 	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
799 	return address;
800 }
801 
802 /* Walk all mappings of a given index of a file and writeprotect them */
dax_entry_mkclean(struct address_space * mapping,pgoff_t index,unsigned long pfn)803 static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index,
804 		unsigned long pfn)
805 {
806 	struct vm_area_struct *vma;
807 	pte_t pte, *ptep = NULL;
808 	pmd_t *pmdp = NULL;
809 	spinlock_t *ptl;
810 
811 	i_mmap_lock_read(mapping);
812 	vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
813 		struct mmu_notifier_range range;
814 		unsigned long address;
815 
816 		cond_resched();
817 
818 		if (!(vma->vm_flags & VM_SHARED))
819 			continue;
820 
821 		address = pgoff_address(index, vma);
822 
823 		/*
824 		 * follow_invalidate_pte() will use the range to call
825 		 * mmu_notifier_invalidate_range_start() on our behalf before
826 		 * taking any lock.
827 		 */
828 		if (follow_invalidate_pte(vma->vm_mm, address, &range, &ptep,
829 					  &pmdp, &ptl))
830 			continue;
831 
832 		/*
833 		 * No need to call mmu_notifier_invalidate_range() as we are
834 		 * downgrading page table protection not changing it to point
835 		 * to a new page.
836 		 *
837 		 * See Documentation/vm/mmu_notifier.rst
838 		 */
839 		if (pmdp) {
840 #ifdef CONFIG_FS_DAX_PMD
841 			pmd_t pmd;
842 
843 			if (pfn != pmd_pfn(*pmdp))
844 				goto unlock_pmd;
845 			if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
846 				goto unlock_pmd;
847 
848 			flush_cache_page(vma, address, pfn);
849 			pmd = pmdp_invalidate(vma, address, pmdp);
850 			pmd = pmd_wrprotect(pmd);
851 			pmd = pmd_mkclean(pmd);
852 			set_pmd_at(vma->vm_mm, address, pmdp, pmd);
853 unlock_pmd:
854 #endif
855 			spin_unlock(ptl);
856 		} else {
857 			if (pfn != pte_pfn(*ptep))
858 				goto unlock_pte;
859 			if (!pte_dirty(*ptep) && !pte_write(*ptep))
860 				goto unlock_pte;
861 
862 			flush_cache_page(vma, address, pfn);
863 			pte = ptep_clear_flush(vma, address, ptep);
864 			pte = pte_wrprotect(pte);
865 			pte = pte_mkclean(pte);
866 			set_pte_at(vma->vm_mm, address, ptep, pte);
867 unlock_pte:
868 			pte_unmap_unlock(ptep, ptl);
869 		}
870 
871 		mmu_notifier_invalidate_range_end(&range);
872 	}
873 	i_mmap_unlock_read(mapping);
874 }
875 
dax_writeback_one(struct xa_state * xas,struct dax_device * dax_dev,struct address_space * mapping,void * entry)876 static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev,
877 		struct address_space *mapping, void *entry)
878 {
879 	unsigned long pfn, index, count;
880 	long ret = 0;
881 
882 	/*
883 	 * A page got tagged dirty in DAX mapping? Something is seriously
884 	 * wrong.
885 	 */
886 	if (WARN_ON(!xa_is_value(entry)))
887 		return -EIO;
888 
889 	if (unlikely(dax_is_locked(entry))) {
890 		void *old_entry = entry;
891 
892 		entry = get_unlocked_entry(xas, 0);
893 
894 		/* Entry got punched out / reallocated? */
895 		if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
896 			goto put_unlocked;
897 		/*
898 		 * Entry got reallocated elsewhere? No need to writeback.
899 		 * We have to compare pfns as we must not bail out due to
900 		 * difference in lockbit or entry type.
901 		 */
902 		if (dax_to_pfn(old_entry) != dax_to_pfn(entry))
903 			goto put_unlocked;
904 		if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
905 					dax_is_zero_entry(entry))) {
906 			ret = -EIO;
907 			goto put_unlocked;
908 		}
909 
910 		/* Another fsync thread may have already done this entry */
911 		if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE))
912 			goto put_unlocked;
913 	}
914 
915 	/* Lock the entry to serialize with page faults */
916 	dax_lock_entry(xas, entry);
917 
918 	/*
919 	 * We can clear the tag now but we have to be careful so that concurrent
920 	 * dax_writeback_one() calls for the same index cannot finish before we
921 	 * actually flush the caches. This is achieved as the calls will look
922 	 * at the entry only under the i_pages lock and once they do that
923 	 * they will see the entry locked and wait for it to unlock.
924 	 */
925 	xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE);
926 	xas_unlock_irq(xas);
927 
928 	/*
929 	 * If dax_writeback_mapping_range() was given a wbc->range_start
930 	 * in the middle of a PMD, the 'index' we use needs to be
931 	 * aligned to the start of the PMD.
932 	 * This allows us to flush for PMD_SIZE and not have to worry about
933 	 * partial PMD writebacks.
934 	 */
935 	pfn = dax_to_pfn(entry);
936 	count = 1UL << dax_entry_order(entry);
937 	index = xas->xa_index & ~(count - 1);
938 
939 	dax_entry_mkclean(mapping, index, pfn);
940 	dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE);
941 	/*
942 	 * After we have flushed the cache, we can clear the dirty tag. There
943 	 * cannot be new dirty data in the pfn after the flush has completed as
944 	 * the pfn mappings are writeprotected and fault waits for mapping
945 	 * entry lock.
946 	 */
947 	xas_reset(xas);
948 	xas_lock_irq(xas);
949 	xas_store(xas, entry);
950 	xas_clear_mark(xas, PAGECACHE_TAG_DIRTY);
951 	dax_wake_entry(xas, entry, WAKE_NEXT);
952 
953 	trace_dax_writeback_one(mapping->host, index, count);
954 	return ret;
955 
956  put_unlocked:
957 	put_unlocked_entry(xas, entry, WAKE_NEXT);
958 	return ret;
959 }
960 
961 /*
962  * Flush the mapping to the persistent domain within the byte range of [start,
963  * end]. This is required by data integrity operations to ensure file data is
964  * on persistent storage prior to completion of the operation.
965  */
dax_writeback_mapping_range(struct address_space * mapping,struct dax_device * dax_dev,struct writeback_control * wbc)966 int dax_writeback_mapping_range(struct address_space *mapping,
967 		struct dax_device *dax_dev, struct writeback_control *wbc)
968 {
969 	XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT);
970 	struct inode *inode = mapping->host;
971 	pgoff_t end_index = wbc->range_end >> PAGE_SHIFT;
972 	void *entry;
973 	int ret = 0;
974 	unsigned int scanned = 0;
975 
976 	if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
977 		return -EIO;
978 
979 	if (mapping_empty(mapping) || wbc->sync_mode != WB_SYNC_ALL)
980 		return 0;
981 
982 	trace_dax_writeback_range(inode, xas.xa_index, end_index);
983 
984 	tag_pages_for_writeback(mapping, xas.xa_index, end_index);
985 
986 	xas_lock_irq(&xas);
987 	xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) {
988 		ret = dax_writeback_one(&xas, dax_dev, mapping, entry);
989 		if (ret < 0) {
990 			mapping_set_error(mapping, ret);
991 			break;
992 		}
993 		if (++scanned % XA_CHECK_SCHED)
994 			continue;
995 
996 		xas_pause(&xas);
997 		xas_unlock_irq(&xas);
998 		cond_resched();
999 		xas_lock_irq(&xas);
1000 	}
1001 	xas_unlock_irq(&xas);
1002 	trace_dax_writeback_range_done(inode, xas.xa_index, end_index);
1003 	return ret;
1004 }
1005 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
1006 
dax_iomap_sector(struct iomap * iomap,loff_t pos)1007 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
1008 {
1009 	return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9;
1010 }
1011 
dax_iomap_pfn(struct iomap * iomap,loff_t pos,size_t size,pfn_t * pfnp)1012 static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size,
1013 			 pfn_t *pfnp)
1014 {
1015 	const sector_t sector = dax_iomap_sector(iomap, pos);
1016 	pgoff_t pgoff;
1017 	int id, rc;
1018 	long length;
1019 
1020 	rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff);
1021 	if (rc)
1022 		return rc;
1023 	id = dax_read_lock();
1024 	length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
1025 				   NULL, pfnp);
1026 	if (length < 0) {
1027 		rc = length;
1028 		goto out;
1029 	}
1030 	rc = -EINVAL;
1031 	if (PFN_PHYS(length) < size)
1032 		goto out;
1033 	if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1034 		goto out;
1035 	/* For larger pages we need devmap */
1036 	if (length > 1 && !pfn_t_devmap(*pfnp))
1037 		goto out;
1038 	rc = 0;
1039 out:
1040 	dax_read_unlock(id);
1041 	return rc;
1042 }
1043 
1044 /*
1045  * The user has performed a load from a hole in the file.  Allocating a new
1046  * page in the file would cause excessive storage usage for workloads with
1047  * sparse files.  Instead we insert a read-only mapping of the 4k zero page.
1048  * If this page is ever written to we will re-fault and change the mapping to
1049  * point to real DAX storage instead.
1050  */
dax_load_hole(struct xa_state * xas,struct address_space * mapping,void ** entry,struct vm_fault * vmf)1051 static vm_fault_t dax_load_hole(struct xa_state *xas,
1052 		struct address_space *mapping, void **entry,
1053 		struct vm_fault *vmf)
1054 {
1055 	struct inode *inode = mapping->host;
1056 	unsigned long vaddr = vmf->address;
1057 	pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1058 	vm_fault_t ret;
1059 
1060 	*entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1061 			DAX_ZERO_PAGE, false);
1062 
1063 	ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1064 	trace_dax_load_hole(inode, vmf, ret);
1065 	return ret;
1066 }
1067 
dax_iomap_zero(loff_t pos,u64 length,struct iomap * iomap)1068 s64 dax_iomap_zero(loff_t pos, u64 length, struct iomap *iomap)
1069 {
1070 	sector_t sector = iomap_sector(iomap, pos & PAGE_MASK);
1071 	pgoff_t pgoff;
1072 	long rc, id;
1073 	void *kaddr;
1074 	bool page_aligned = false;
1075 	unsigned offset = offset_in_page(pos);
1076 	unsigned size = min_t(u64, PAGE_SIZE - offset, length);
1077 
1078 	if (IS_ALIGNED(sector << SECTOR_SHIFT, PAGE_SIZE) &&
1079 	    (size == PAGE_SIZE))
1080 		page_aligned = true;
1081 
1082 	rc = bdev_dax_pgoff(iomap->bdev, sector, PAGE_SIZE, &pgoff);
1083 	if (rc)
1084 		return rc;
1085 
1086 	id = dax_read_lock();
1087 
1088 	if (page_aligned)
1089 		rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1);
1090 	else
1091 		rc = dax_direct_access(iomap->dax_dev, pgoff, 1, &kaddr, NULL);
1092 	if (rc < 0) {
1093 		dax_read_unlock(id);
1094 		return rc;
1095 	}
1096 
1097 	if (!page_aligned) {
1098 		memset(kaddr + offset, 0, size);
1099 		dax_flush(iomap->dax_dev, kaddr + offset, size);
1100 	}
1101 	dax_read_unlock(id);
1102 	return size;
1103 }
1104 
1105 static loff_t
dax_iomap_actor(struct inode * inode,loff_t pos,loff_t length,void * data,struct iomap * iomap,struct iomap * srcmap)1106 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1107 		struct iomap *iomap, struct iomap *srcmap)
1108 {
1109 	struct block_device *bdev = iomap->bdev;
1110 	struct dax_device *dax_dev = iomap->dax_dev;
1111 	struct iov_iter *iter = data;
1112 	loff_t end = pos + length, done = 0;
1113 	ssize_t ret = 0;
1114 	size_t xfer;
1115 	int id;
1116 
1117 	if (iov_iter_rw(iter) == READ) {
1118 		end = min(end, i_size_read(inode));
1119 		if (pos >= end)
1120 			return 0;
1121 
1122 		if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1123 			return iov_iter_zero(min(length, end - pos), iter);
1124 	}
1125 
1126 	if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1127 		return -EIO;
1128 
1129 	/*
1130 	 * Write can allocate block for an area which has a hole page mapped
1131 	 * into page tables. We have to tear down these mappings so that data
1132 	 * written by write(2) is visible in mmap.
1133 	 */
1134 	if (iomap->flags & IOMAP_F_NEW) {
1135 		invalidate_inode_pages2_range(inode->i_mapping,
1136 					      pos >> PAGE_SHIFT,
1137 					      (end - 1) >> PAGE_SHIFT);
1138 	}
1139 
1140 	id = dax_read_lock();
1141 	while (pos < end) {
1142 		unsigned offset = pos & (PAGE_SIZE - 1);
1143 		const size_t size = ALIGN(length + offset, PAGE_SIZE);
1144 		const sector_t sector = dax_iomap_sector(iomap, pos);
1145 		ssize_t map_len;
1146 		pgoff_t pgoff;
1147 		void *kaddr;
1148 
1149 		if (fatal_signal_pending(current)) {
1150 			ret = -EINTR;
1151 			break;
1152 		}
1153 
1154 		ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1155 		if (ret)
1156 			break;
1157 
1158 		map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1159 				&kaddr, NULL);
1160 		if (map_len < 0) {
1161 			ret = map_len;
1162 			break;
1163 		}
1164 
1165 		map_len = PFN_PHYS(map_len);
1166 		kaddr += offset;
1167 		map_len -= offset;
1168 		if (map_len > end - pos)
1169 			map_len = end - pos;
1170 
1171 		/*
1172 		 * The userspace address for the memory copy has already been
1173 		 * validated via access_ok() in either vfs_read() or
1174 		 * vfs_write(), depending on which operation we are doing.
1175 		 */
1176 		if (iov_iter_rw(iter) == WRITE)
1177 			xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1178 					map_len, iter);
1179 		else
1180 			xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1181 					map_len, iter);
1182 
1183 		pos += xfer;
1184 		length -= xfer;
1185 		done += xfer;
1186 
1187 		if (xfer == 0)
1188 			ret = -EFAULT;
1189 		if (xfer < map_len)
1190 			break;
1191 	}
1192 	dax_read_unlock(id);
1193 
1194 	return done ? done : ret;
1195 }
1196 
1197 /**
1198  * dax_iomap_rw - Perform I/O to a DAX file
1199  * @iocb:	The control block for this I/O
1200  * @iter:	The addresses to do I/O from or to
1201  * @ops:	iomap ops passed from the file system
1202  *
1203  * This function performs read and write operations to directly mapped
1204  * persistent memory.  The callers needs to take care of read/write exclusion
1205  * and evicting any page cache pages in the region under I/O.
1206  */
1207 ssize_t
dax_iomap_rw(struct kiocb * iocb,struct iov_iter * iter,const struct iomap_ops * ops)1208 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1209 		const struct iomap_ops *ops)
1210 {
1211 	struct address_space *mapping = iocb->ki_filp->f_mapping;
1212 	struct inode *inode = mapping->host;
1213 	loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1214 	unsigned flags = 0;
1215 
1216 	if (iov_iter_rw(iter) == WRITE) {
1217 		lockdep_assert_held_write(&inode->i_rwsem);
1218 		flags |= IOMAP_WRITE;
1219 	} else {
1220 		lockdep_assert_held(&inode->i_rwsem);
1221 	}
1222 
1223 	if (iocb->ki_flags & IOCB_NOWAIT)
1224 		flags |= IOMAP_NOWAIT;
1225 
1226 	while (iov_iter_count(iter)) {
1227 		ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1228 				iter, dax_iomap_actor);
1229 		if (ret <= 0)
1230 			break;
1231 		pos += ret;
1232 		done += ret;
1233 	}
1234 
1235 	iocb->ki_pos += done;
1236 	return done ? done : ret;
1237 }
1238 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1239 
dax_fault_return(int error)1240 static vm_fault_t dax_fault_return(int error)
1241 {
1242 	if (error == 0)
1243 		return VM_FAULT_NOPAGE;
1244 	return vmf_error(error);
1245 }
1246 
1247 /*
1248  * MAP_SYNC on a dax mapping guarantees dirty metadata is
1249  * flushed on write-faults (non-cow), but not read-faults.
1250  */
dax_fault_is_synchronous(unsigned long flags,struct vm_area_struct * vma,struct iomap * iomap)1251 static bool dax_fault_is_synchronous(unsigned long flags,
1252 		struct vm_area_struct *vma, struct iomap *iomap)
1253 {
1254 	return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC)
1255 		&& (iomap->flags & IOMAP_F_DIRTY);
1256 }
1257 
dax_iomap_pte_fault(struct vm_fault * vmf,pfn_t * pfnp,int * iomap_errp,const struct iomap_ops * ops)1258 static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1259 			       int *iomap_errp, const struct iomap_ops *ops)
1260 {
1261 	struct vm_area_struct *vma = vmf->vma;
1262 	struct address_space *mapping = vma->vm_file->f_mapping;
1263 	XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
1264 	struct inode *inode = mapping->host;
1265 	unsigned long vaddr = vmf->address;
1266 	loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1267 	struct iomap iomap = { .type = IOMAP_HOLE };
1268 	struct iomap srcmap = { .type = IOMAP_HOLE };
1269 	unsigned flags = IOMAP_FAULT;
1270 	int error, major = 0;
1271 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1272 	bool sync;
1273 	vm_fault_t ret = 0;
1274 	void *entry;
1275 	pfn_t pfn;
1276 
1277 	trace_dax_pte_fault(inode, vmf, ret);
1278 	/*
1279 	 * Check whether offset isn't beyond end of file now. Caller is supposed
1280 	 * to hold locks serializing us with truncate / punch hole so this is
1281 	 * a reliable test.
1282 	 */
1283 	if (pos >= i_size_read(inode)) {
1284 		ret = VM_FAULT_SIGBUS;
1285 		goto out;
1286 	}
1287 
1288 	if (write && !vmf->cow_page)
1289 		flags |= IOMAP_WRITE;
1290 
1291 	entry = grab_mapping_entry(&xas, mapping, 0);
1292 	if (xa_is_internal(entry)) {
1293 		ret = xa_to_internal(entry);
1294 		goto out;
1295 	}
1296 
1297 	/*
1298 	 * It is possible, particularly with mixed reads & writes to private
1299 	 * mappings, that we have raced with a PMD fault that overlaps with
1300 	 * the PTE we need to set up.  If so just return and the fault will be
1301 	 * retried.
1302 	 */
1303 	if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1304 		ret = VM_FAULT_NOPAGE;
1305 		goto unlock_entry;
1306 	}
1307 
1308 	/*
1309 	 * Note that we don't bother to use iomap_apply here: DAX required
1310 	 * the file system block size to be equal the page size, which means
1311 	 * that we never have to deal with more than a single extent here.
1312 	 */
1313 	error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap, &srcmap);
1314 	if (iomap_errp)
1315 		*iomap_errp = error;
1316 	if (error) {
1317 		ret = dax_fault_return(error);
1318 		goto unlock_entry;
1319 	}
1320 	if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1321 		error = -EIO;	/* fs corruption? */
1322 		goto error_finish_iomap;
1323 	}
1324 
1325 	if (vmf->cow_page) {
1326 		sector_t sector = dax_iomap_sector(&iomap, pos);
1327 
1328 		switch (iomap.type) {
1329 		case IOMAP_HOLE:
1330 		case IOMAP_UNWRITTEN:
1331 			clear_user_highpage(vmf->cow_page, vaddr);
1332 			break;
1333 		case IOMAP_MAPPED:
1334 			error = copy_cow_page_dax(iomap.bdev, iomap.dax_dev,
1335 						  sector, vmf->cow_page, vaddr);
1336 			break;
1337 		default:
1338 			WARN_ON_ONCE(1);
1339 			error = -EIO;
1340 			break;
1341 		}
1342 
1343 		if (error)
1344 			goto error_finish_iomap;
1345 
1346 		__SetPageUptodate(vmf->cow_page);
1347 		ret = finish_fault(vmf);
1348 		if (!ret)
1349 			ret = VM_FAULT_DONE_COW;
1350 		goto finish_iomap;
1351 	}
1352 
1353 	sync = dax_fault_is_synchronous(flags, vma, &iomap);
1354 
1355 	switch (iomap.type) {
1356 	case IOMAP_MAPPED:
1357 		if (iomap.flags & IOMAP_F_NEW) {
1358 			count_vm_event(PGMAJFAULT);
1359 			count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
1360 			major = VM_FAULT_MAJOR;
1361 		}
1362 		error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn);
1363 		if (error < 0)
1364 			goto error_finish_iomap;
1365 
1366 		entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
1367 						 0, write && !sync);
1368 
1369 		/*
1370 		 * If we are doing synchronous page fault and inode needs fsync,
1371 		 * we can insert PTE into page tables only after that happens.
1372 		 * Skip insertion for now and return the pfn so that caller can
1373 		 * insert it after fsync is done.
1374 		 */
1375 		if (sync) {
1376 			if (WARN_ON_ONCE(!pfnp)) {
1377 				error = -EIO;
1378 				goto error_finish_iomap;
1379 			}
1380 			*pfnp = pfn;
1381 			ret = VM_FAULT_NEEDDSYNC | major;
1382 			goto finish_iomap;
1383 		}
1384 		trace_dax_insert_mapping(inode, vmf, entry);
1385 		if (write)
1386 			ret = vmf_insert_mixed_mkwrite(vma, vaddr, pfn);
1387 		else
1388 			ret = vmf_insert_mixed(vma, vaddr, pfn);
1389 
1390 		goto finish_iomap;
1391 	case IOMAP_UNWRITTEN:
1392 	case IOMAP_HOLE:
1393 		if (!write) {
1394 			ret = dax_load_hole(&xas, mapping, &entry, vmf);
1395 			goto finish_iomap;
1396 		}
1397 		fallthrough;
1398 	default:
1399 		WARN_ON_ONCE(1);
1400 		error = -EIO;
1401 		break;
1402 	}
1403 
1404  error_finish_iomap:
1405 	ret = dax_fault_return(error);
1406  finish_iomap:
1407 	if (ops->iomap_end) {
1408 		int copied = PAGE_SIZE;
1409 
1410 		if (ret & VM_FAULT_ERROR)
1411 			copied = 0;
1412 		/*
1413 		 * The fault is done by now and there's no way back (other
1414 		 * thread may be already happily using PTE we have installed).
1415 		 * Just ignore error from ->iomap_end since we cannot do much
1416 		 * with it.
1417 		 */
1418 		ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1419 	}
1420  unlock_entry:
1421 	dax_unlock_entry(&xas, entry);
1422  out:
1423 	trace_dax_pte_fault_done(inode, vmf, ret);
1424 	return ret | major;
1425 }
1426 
1427 #ifdef CONFIG_FS_DAX_PMD
dax_pmd_load_hole(struct xa_state * xas,struct vm_fault * vmf,struct iomap * iomap,void ** entry)1428 static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1429 		struct iomap *iomap, void **entry)
1430 {
1431 	struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1432 	unsigned long pmd_addr = vmf->address & PMD_MASK;
1433 	struct vm_area_struct *vma = vmf->vma;
1434 	struct inode *inode = mapping->host;
1435 	pgtable_t pgtable = NULL;
1436 	struct page *zero_page;
1437 	spinlock_t *ptl;
1438 	pmd_t pmd_entry;
1439 	pfn_t pfn;
1440 
1441 	zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1442 
1443 	if (unlikely(!zero_page))
1444 		goto fallback;
1445 
1446 	pfn = page_to_pfn_t(zero_page);
1447 	*entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1448 			DAX_PMD | DAX_ZERO_PAGE, false);
1449 
1450 	if (arch_needs_pgtable_deposit()) {
1451 		pgtable = pte_alloc_one(vma->vm_mm);
1452 		if (!pgtable)
1453 			return VM_FAULT_OOM;
1454 	}
1455 
1456 	ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1457 	if (!pmd_none(*(vmf->pmd))) {
1458 		spin_unlock(ptl);
1459 		goto fallback;
1460 	}
1461 
1462 	if (pgtable) {
1463 		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1464 		mm_inc_nr_ptes(vma->vm_mm);
1465 	}
1466 	pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1467 	pmd_entry = pmd_mkhuge(pmd_entry);
1468 	set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1469 	spin_unlock(ptl);
1470 	trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry);
1471 	return VM_FAULT_NOPAGE;
1472 
1473 fallback:
1474 	if (pgtable)
1475 		pte_free(vma->vm_mm, pgtable);
1476 	trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry);
1477 	return VM_FAULT_FALLBACK;
1478 }
1479 
dax_iomap_pmd_fault(struct vm_fault * vmf,pfn_t * pfnp,const struct iomap_ops * ops)1480 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1481 			       const struct iomap_ops *ops)
1482 {
1483 	struct vm_area_struct *vma = vmf->vma;
1484 	struct address_space *mapping = vma->vm_file->f_mapping;
1485 	XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
1486 	unsigned long pmd_addr = vmf->address & PMD_MASK;
1487 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1488 	bool sync;
1489 	unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1490 	struct inode *inode = mapping->host;
1491 	vm_fault_t result = VM_FAULT_FALLBACK;
1492 	struct iomap iomap = { .type = IOMAP_HOLE };
1493 	struct iomap srcmap = { .type = IOMAP_HOLE };
1494 	pgoff_t max_pgoff;
1495 	void *entry;
1496 	loff_t pos;
1497 	int error;
1498 	pfn_t pfn;
1499 
1500 	/*
1501 	 * Check whether offset isn't beyond end of file now. Caller is
1502 	 * supposed to hold locks serializing us with truncate / punch hole so
1503 	 * this is a reliable test.
1504 	 */
1505 	max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
1506 
1507 	trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1508 
1509 	/*
1510 	 * Make sure that the faulting address's PMD offset (color) matches
1511 	 * the PMD offset from the start of the file.  This is necessary so
1512 	 * that a PMD range in the page table overlaps exactly with a PMD
1513 	 * range in the page cache.
1514 	 */
1515 	if ((vmf->pgoff & PG_PMD_COLOUR) !=
1516 	    ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1517 		goto fallback;
1518 
1519 	/* Fall back to PTEs if we're going to COW */
1520 	if (write && !(vma->vm_flags & VM_SHARED))
1521 		goto fallback;
1522 
1523 	/* If the PMD would extend outside the VMA */
1524 	if (pmd_addr < vma->vm_start)
1525 		goto fallback;
1526 	if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1527 		goto fallback;
1528 
1529 	if (xas.xa_index >= max_pgoff) {
1530 		result = VM_FAULT_SIGBUS;
1531 		goto out;
1532 	}
1533 
1534 	/* If the PMD would extend beyond the file size */
1535 	if ((xas.xa_index | PG_PMD_COLOUR) >= max_pgoff)
1536 		goto fallback;
1537 
1538 	/*
1539 	 * grab_mapping_entry() will make sure we get an empty PMD entry,
1540 	 * a zero PMD entry or a DAX PMD.  If it can't (because a PTE
1541 	 * entry is already in the array, for instance), it will return
1542 	 * VM_FAULT_FALLBACK.
1543 	 */
1544 	entry = grab_mapping_entry(&xas, mapping, PMD_ORDER);
1545 	if (xa_is_internal(entry)) {
1546 		result = xa_to_internal(entry);
1547 		goto fallback;
1548 	}
1549 
1550 	/*
1551 	 * It is possible, particularly with mixed reads & writes to private
1552 	 * mappings, that we have raced with a PTE fault that overlaps with
1553 	 * the PMD we need to set up.  If so just return and the fault will be
1554 	 * retried.
1555 	 */
1556 	if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1557 			!pmd_devmap(*vmf->pmd)) {
1558 		result = 0;
1559 		goto unlock_entry;
1560 	}
1561 
1562 	/*
1563 	 * Note that we don't use iomap_apply here.  We aren't doing I/O, only
1564 	 * setting up a mapping, so really we're using iomap_begin() as a way
1565 	 * to look up our filesystem block.
1566 	 */
1567 	pos = (loff_t)xas.xa_index << PAGE_SHIFT;
1568 	error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap,
1569 			&srcmap);
1570 	if (error)
1571 		goto unlock_entry;
1572 
1573 	if (iomap.offset + iomap.length < pos + PMD_SIZE)
1574 		goto finish_iomap;
1575 
1576 	sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap);
1577 
1578 	switch (iomap.type) {
1579 	case IOMAP_MAPPED:
1580 		error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn);
1581 		if (error < 0)
1582 			goto finish_iomap;
1583 
1584 		entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
1585 						DAX_PMD, write && !sync);
1586 
1587 		/*
1588 		 * If we are doing synchronous page fault and inode needs fsync,
1589 		 * we can insert PMD into page tables only after that happens.
1590 		 * Skip insertion for now and return the pfn so that caller can
1591 		 * insert it after fsync is done.
1592 		 */
1593 		if (sync) {
1594 			if (WARN_ON_ONCE(!pfnp))
1595 				goto finish_iomap;
1596 			*pfnp = pfn;
1597 			result = VM_FAULT_NEEDDSYNC;
1598 			goto finish_iomap;
1599 		}
1600 
1601 		trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry);
1602 		result = vmf_insert_pfn_pmd(vmf, pfn, write);
1603 		break;
1604 	case IOMAP_UNWRITTEN:
1605 	case IOMAP_HOLE:
1606 		if (WARN_ON_ONCE(write))
1607 			break;
1608 		result = dax_pmd_load_hole(&xas, vmf, &iomap, &entry);
1609 		break;
1610 	default:
1611 		WARN_ON_ONCE(1);
1612 		break;
1613 	}
1614 
1615  finish_iomap:
1616 	if (ops->iomap_end) {
1617 		int copied = PMD_SIZE;
1618 
1619 		if (result == VM_FAULT_FALLBACK)
1620 			copied = 0;
1621 		/*
1622 		 * The fault is done by now and there's no way back (other
1623 		 * thread may be already happily using PMD we have installed).
1624 		 * Just ignore error from ->iomap_end since we cannot do much
1625 		 * with it.
1626 		 */
1627 		ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1628 				&iomap);
1629 	}
1630  unlock_entry:
1631 	dax_unlock_entry(&xas, entry);
1632  fallback:
1633 	if (result == VM_FAULT_FALLBACK) {
1634 		split_huge_pmd(vma, vmf->pmd, vmf->address);
1635 		count_vm_event(THP_FAULT_FALLBACK);
1636 	}
1637 out:
1638 	trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1639 	return result;
1640 }
1641 #else
dax_iomap_pmd_fault(struct vm_fault * vmf,pfn_t * pfnp,const struct iomap_ops * ops)1642 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1643 			       const struct iomap_ops *ops)
1644 {
1645 	return VM_FAULT_FALLBACK;
1646 }
1647 #endif /* CONFIG_FS_DAX_PMD */
1648 
1649 /**
1650  * dax_iomap_fault - handle a page fault on a DAX file
1651  * @vmf: The description of the fault
1652  * @pe_size: Size of the page to fault in
1653  * @pfnp: PFN to insert for synchronous faults if fsync is required
1654  * @iomap_errp: Storage for detailed error code in case of error
1655  * @ops: Iomap ops passed from the file system
1656  *
1657  * When a page fault occurs, filesystems may call this helper in
1658  * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1659  * has done all the necessary locking for page fault to proceed
1660  * successfully.
1661  */
dax_iomap_fault(struct vm_fault * vmf,enum page_entry_size pe_size,pfn_t * pfnp,int * iomap_errp,const struct iomap_ops * ops)1662 vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1663 		    pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1664 {
1665 	switch (pe_size) {
1666 	case PE_SIZE_PTE:
1667 		return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1668 	case PE_SIZE_PMD:
1669 		return dax_iomap_pmd_fault(vmf, pfnp, ops);
1670 	default:
1671 		return VM_FAULT_FALLBACK;
1672 	}
1673 }
1674 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1675 
1676 /*
1677  * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1678  * @vmf: The description of the fault
1679  * @pfn: PFN to insert
1680  * @order: Order of entry to insert.
1681  *
1682  * This function inserts a writeable PTE or PMD entry into the page tables
1683  * for an mmaped DAX file.  It also marks the page cache entry as dirty.
1684  */
1685 static vm_fault_t
dax_insert_pfn_mkwrite(struct vm_fault * vmf,pfn_t pfn,unsigned int order)1686 dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
1687 {
1688 	struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1689 	XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1690 	void *entry;
1691 	vm_fault_t ret;
1692 
1693 	xas_lock_irq(&xas);
1694 	entry = get_unlocked_entry(&xas, order);
1695 	/* Did we race with someone splitting entry or so? */
1696 	if (!entry || dax_is_conflict(entry) ||
1697 	    (order == 0 && !dax_is_pte_entry(entry))) {
1698 		put_unlocked_entry(&xas, entry, WAKE_NEXT);
1699 		xas_unlock_irq(&xas);
1700 		trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1701 						      VM_FAULT_NOPAGE);
1702 		return VM_FAULT_NOPAGE;
1703 	}
1704 	xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1705 	dax_lock_entry(&xas, entry);
1706 	xas_unlock_irq(&xas);
1707 	if (order == 0)
1708 		ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1709 #ifdef CONFIG_FS_DAX_PMD
1710 	else if (order == PMD_ORDER)
1711 		ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
1712 #endif
1713 	else
1714 		ret = VM_FAULT_FALLBACK;
1715 	dax_unlock_entry(&xas, entry);
1716 	trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1717 	return ret;
1718 }
1719 
1720 /**
1721  * dax_finish_sync_fault - finish synchronous page fault
1722  * @vmf: The description of the fault
1723  * @pe_size: Size of entry to be inserted
1724  * @pfn: PFN to insert
1725  *
1726  * This function ensures that the file range touched by the page fault is
1727  * stored persistently on the media and handles inserting of appropriate page
1728  * table entry.
1729  */
dax_finish_sync_fault(struct vm_fault * vmf,enum page_entry_size pe_size,pfn_t pfn)1730 vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf,
1731 		enum page_entry_size pe_size, pfn_t pfn)
1732 {
1733 	int err;
1734 	loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1735 	unsigned int order = pe_order(pe_size);
1736 	size_t len = PAGE_SIZE << order;
1737 
1738 	err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1739 	if (err)
1740 		return VM_FAULT_SIGBUS;
1741 	return dax_insert_pfn_mkwrite(vmf, pfn, order);
1742 }
1743 EXPORT_SYMBOL_GPL(dax_finish_sync_fault);
1744