blk-mq: don't count completed flush data request as inflight in case of quiesce
[linux-block.git] / fs / dax.c
CommitLineData
2025cf9e 1// SPDX-License-Identifier: GPL-2.0-only
d475c634
MW
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>
d475c634
MW
7 */
8
9#include <linux/atomic.h>
10#include <linux/blkdev.h>
11#include <linux/buffer_head.h>
d77e92e2 12#include <linux/dax.h>
d475c634 13#include <linux/fs.h>
f7ca90b1
MW
14#include <linux/highmem.h>
15#include <linux/memcontrol.h>
16#include <linux/mm.h>
d475c634 17#include <linux/mutex.h>
9973c98e 18#include <linux/pagevec.h>
289c6aed 19#include <linux/sched.h>
f361bf4a 20#include <linux/sched/signal.h>
d475c634 21#include <linux/uio.h>
f7ca90b1 22#include <linux/vmstat.h>
34c0fd54 23#include <linux/pfn_t.h>
0e749e54 24#include <linux/sizes.h>
4b4bb46d 25#include <linux/mmu_notifier.h>
a254e568 26#include <linux/iomap.h>
06083a09 27#include <linux/rmap.h>
11cf9d86 28#include <asm/pgalloc.h>
d475c634 29
282a8e03
RZ
30#define CREATE_TRACE_POINTS
31#include <trace/events/fs_dax.h>
32
ac401cc7
JK
33/* We choose 4096 entries - same as per-zone page wait tables */
34#define DAX_WAIT_TABLE_BITS 12
35#define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
36
917f3452
RZ
37/* The 'colour' (ie low bits) within a PMD of a page offset. */
38#define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
977fbdcd 39#define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT)
917f3452 40
ce95ab0f 41static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
ac401cc7
JK
42
43static int __init init_dax_wait_table(void)
44{
45 int i;
46
47 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
48 init_waitqueue_head(wait_table + i);
49 return 0;
50}
51fs_initcall(init_dax_wait_table);
52
527b19d0 53/*
3159f943
MW
54 * DAX pagecache entries use XArray value entries so they can't be mistaken
55 * for pages. We use one bit for locking, one bit for the entry size (PMD)
56 * and two more to tell us if the entry is a zero page or an empty entry that
57 * is just used for locking. In total four special bits.
527b19d0
RZ
58 *
59 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
60 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
61 * block allocation.
62 */
3159f943
MW
63#define DAX_SHIFT (4)
64#define DAX_LOCKED (1UL << 0)
65#define DAX_PMD (1UL << 1)
66#define DAX_ZERO_PAGE (1UL << 2)
67#define DAX_EMPTY (1UL << 3)
527b19d0 68
a77d19f4 69static unsigned long dax_to_pfn(void *entry)
527b19d0 70{
3159f943 71 return xa_to_value(entry) >> DAX_SHIFT;
527b19d0
RZ
72}
73
9f32d221
MW
74static void *dax_make_entry(pfn_t pfn, unsigned long flags)
75{
76 return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
77}
78
cfc93c6c
MW
79static bool dax_is_locked(void *entry)
80{
81 return xa_to_value(entry) & DAX_LOCKED;
82}
83
a77d19f4 84static unsigned int dax_entry_order(void *entry)
527b19d0 85{
3159f943 86 if (xa_to_value(entry) & DAX_PMD)
cfc93c6c 87 return PMD_ORDER;
527b19d0
RZ
88 return 0;
89}
90
fda490d3 91static unsigned long dax_is_pmd_entry(void *entry)
d1a5f2b4 92{
3159f943 93 return xa_to_value(entry) & DAX_PMD;
d1a5f2b4
DW
94}
95
fda490d3 96static bool dax_is_pte_entry(void *entry)
d475c634 97{
3159f943 98 return !(xa_to_value(entry) & DAX_PMD);
d475c634
MW
99}
100
642261ac 101static int dax_is_zero_entry(void *entry)
d475c634 102{
3159f943 103 return xa_to_value(entry) & DAX_ZERO_PAGE;
d475c634
MW
104}
105
642261ac 106static int dax_is_empty_entry(void *entry)
b2e0d162 107{
3159f943 108 return xa_to_value(entry) & DAX_EMPTY;
b2e0d162
DW
109}
110
23c84eb7
MWO
111/*
112 * true if the entry that was found is of a smaller order than the entry
113 * we were looking for
114 */
115static bool dax_is_conflict(void *entry)
116{
117 return entry == XA_RETRY_ENTRY;
118}
119
ac401cc7 120/*
a77d19f4 121 * DAX page cache entry locking
ac401cc7
JK
122 */
123struct exceptional_entry_key {
ec4907ff 124 struct xarray *xa;
63e95b5c 125 pgoff_t entry_start;
ac401cc7
JK
126};
127
128struct wait_exceptional_entry_queue {
ac6424b9 129 wait_queue_entry_t wait;
ac401cc7
JK
130 struct exceptional_entry_key key;
131};
132
698ab77a
VG
133/**
134 * enum dax_wake_mode: waitqueue wakeup behaviour
135 * @WAKE_ALL: wake all waiters in the waitqueue
136 * @WAKE_NEXT: wake only the first waiter in the waitqueue
137 */
138enum dax_wake_mode {
139 WAKE_ALL,
140 WAKE_NEXT,
141};
142
b15cd800
MW
143static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas,
144 void *entry, struct exceptional_entry_key *key)
63e95b5c
RZ
145{
146 unsigned long hash;
b15cd800 147 unsigned long index = xas->xa_index;
63e95b5c
RZ
148
149 /*
150 * If 'entry' is a PMD, align the 'index' that we use for the wait
151 * queue to the start of that PMD. This ensures that all offsets in
152 * the range covered by the PMD map to the same bit lock.
153 */
642261ac 154 if (dax_is_pmd_entry(entry))
917f3452 155 index &= ~PG_PMD_COLOUR;
b15cd800 156 key->xa = xas->xa;
63e95b5c
RZ
157 key->entry_start = index;
158
b15cd800 159 hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS);
63e95b5c
RZ
160 return wait_table + hash;
161}
162
ec4907ff
MW
163static int wake_exceptional_entry_func(wait_queue_entry_t *wait,
164 unsigned int mode, int sync, void *keyp)
ac401cc7
JK
165{
166 struct exceptional_entry_key *key = keyp;
167 struct wait_exceptional_entry_queue *ewait =
168 container_of(wait, struct wait_exceptional_entry_queue, wait);
169
ec4907ff 170 if (key->xa != ewait->key.xa ||
63e95b5c 171 key->entry_start != ewait->key.entry_start)
ac401cc7
JK
172 return 0;
173 return autoremove_wake_function(wait, mode, sync, NULL);
174}
175
e30331ff 176/*
b93b0163
MW
177 * @entry may no longer be the entry at the index in the mapping.
178 * The important information it's conveying is whether the entry at
179 * this index used to be a PMD entry.
e30331ff 180 */
698ab77a
VG
181static void dax_wake_entry(struct xa_state *xas, void *entry,
182 enum dax_wake_mode mode)
e30331ff
RZ
183{
184 struct exceptional_entry_key key;
185 wait_queue_head_t *wq;
186
b15cd800 187 wq = dax_entry_waitqueue(xas, entry, &key);
e30331ff
RZ
188
189 /*
190 * Checking for locked entry and prepare_to_wait_exclusive() happens
b93b0163 191 * under the i_pages lock, ditto for entry handling in our callers.
e30331ff
RZ
192 * So at this point all tasks that could have seen our entry locked
193 * must be in the waitqueue and the following check will see them.
194 */
195 if (waitqueue_active(wq))
698ab77a 196 __wake_up(wq, TASK_NORMAL, mode == WAKE_ALL ? 0 : 1, &key);
e30331ff
RZ
197}
198
cfc93c6c
MW
199/*
200 * Look up entry in page cache, wait for it to become unlocked if it
201 * is a DAX entry and return it. The caller must subsequently call
202 * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
23c84eb7
MWO
203 * if it did. The entry returned may have a larger order than @order.
204 * If @order is larger than the order of the entry found in i_pages, this
205 * function returns a dax_is_conflict entry.
cfc93c6c
MW
206 *
207 * Must be called with the i_pages lock held.
208 */
23c84eb7 209static void *get_unlocked_entry(struct xa_state *xas, unsigned int order)
cfc93c6c
MW
210{
211 void *entry;
212 struct wait_exceptional_entry_queue ewait;
213 wait_queue_head_t *wq;
214
215 init_wait(&ewait.wait);
216 ewait.wait.func = wake_exceptional_entry_func;
217
218 for (;;) {
0e40de03 219 entry = xas_find_conflict(xas);
6370740e
DW
220 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
221 return entry;
23c84eb7
MWO
222 if (dax_entry_order(entry) < order)
223 return XA_RETRY_ENTRY;
6370740e 224 if (!dax_is_locked(entry))
cfc93c6c
MW
225 return entry;
226
b15cd800 227 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
cfc93c6c
MW
228 prepare_to_wait_exclusive(wq, &ewait.wait,
229 TASK_UNINTERRUPTIBLE);
230 xas_unlock_irq(xas);
231 xas_reset(xas);
232 schedule();
233 finish_wait(wq, &ewait.wait);
234 xas_lock_irq(xas);
235 }
236}
237
55e56f06
MW
238/*
239 * The only thing keeping the address space around is the i_pages lock
240 * (it's cycled in clear_inode() after removing the entries from i_pages)
241 * After we call xas_unlock_irq(), we cannot touch xas->xa.
242 */
243static void wait_entry_unlocked(struct xa_state *xas, void *entry)
244{
245 struct wait_exceptional_entry_queue ewait;
246 wait_queue_head_t *wq;
247
248 init_wait(&ewait.wait);
249 ewait.wait.func = wake_exceptional_entry_func;
250
251 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
d8a70641
DW
252 /*
253 * Unlike get_unlocked_entry() there is no guarantee that this
254 * path ever successfully retrieves an unlocked entry before an
255 * inode dies. Perform a non-exclusive wait in case this path
256 * never successfully performs its own wake up.
257 */
258 prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
55e56f06
MW
259 xas_unlock_irq(xas);
260 schedule();
261 finish_wait(wq, &ewait.wait);
55e56f06
MW
262}
263
4c3d043d
VG
264static void put_unlocked_entry(struct xa_state *xas, void *entry,
265 enum dax_wake_mode mode)
cfc93c6c 266{
61c30c98 267 if (entry && !dax_is_conflict(entry))
4c3d043d 268 dax_wake_entry(xas, entry, mode);
cfc93c6c
MW
269}
270
271/*
272 * We used the xa_state to get the entry, but then we locked the entry and
273 * dropped the xa_lock, so we know the xa_state is stale and must be reset
274 * before use.
275 */
276static void dax_unlock_entry(struct xa_state *xas, void *entry)
277{
278 void *old;
279
7ae2ea7d 280 BUG_ON(dax_is_locked(entry));
cfc93c6c
MW
281 xas_reset(xas);
282 xas_lock_irq(xas);
283 old = xas_store(xas, entry);
284 xas_unlock_irq(xas);
285 BUG_ON(!dax_is_locked(old));
698ab77a 286 dax_wake_entry(xas, entry, WAKE_NEXT);
cfc93c6c
MW
287}
288
289/*
290 * Return: The entry stored at this location before it was locked.
291 */
292static void *dax_lock_entry(struct xa_state *xas, void *entry)
293{
294 unsigned long v = xa_to_value(entry);
295 return xas_store(xas, xa_mk_value(v | DAX_LOCKED));
296}
297
d2c997c0
DW
298static unsigned long dax_entry_size(void *entry)
299{
300 if (dax_is_zero_entry(entry))
301 return 0;
302 else if (dax_is_empty_entry(entry))
303 return 0;
304 else if (dax_is_pmd_entry(entry))
305 return PMD_SIZE;
306 else
307 return PAGE_SIZE;
308}
309
a77d19f4 310static unsigned long dax_end_pfn(void *entry)
d2c997c0 311{
a77d19f4 312 return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
d2c997c0
DW
313}
314
315/*
316 * Iterate through all mapped pfns represented by an entry, i.e. skip
317 * 'empty' and 'zero' entries.
318 */
319#define for_each_mapped_pfn(entry, pfn) \
a77d19f4
MW
320 for (pfn = dax_to_pfn(entry); \
321 pfn < dax_end_pfn(entry); pfn++)
d2c997c0 322
16900426 323static inline bool dax_page_is_shared(struct page *page)
6061b69b 324{
16900426 325 return page->mapping == PAGE_MAPPING_DAX_SHARED;
6061b69b
SR
326}
327
73449daf 328/*
16900426
SR
329 * Set the page->mapping with PAGE_MAPPING_DAX_SHARED flag, increase the
330 * refcount.
6061b69b 331 */
16900426 332static inline void dax_page_share_get(struct page *page)
6061b69b 333{
16900426 334 if (page->mapping != PAGE_MAPPING_DAX_SHARED) {
6061b69b
SR
335 /*
336 * Reset the index if the page was already mapped
337 * regularly before.
338 */
339 if (page->mapping)
16900426
SR
340 page->share = 1;
341 page->mapping = PAGE_MAPPING_DAX_SHARED;
6061b69b 342 }
16900426
SR
343 page->share++;
344}
345
346static inline unsigned long dax_page_share_put(struct page *page)
347{
348 return --page->share;
6061b69b
SR
349}
350
351/*
16900426 352 * When it is called in dax_insert_entry(), the shared flag will indicate that
6061b69b 353 * whether this entry is shared by multiple files. If so, set the page->mapping
16900426 354 * PAGE_MAPPING_DAX_SHARED, and use page->share as refcount.
73449daf
DW
355 */
356static void dax_associate_entry(void *entry, struct address_space *mapping,
16900426 357 struct vm_area_struct *vma, unsigned long address, bool shared)
d2c997c0 358{
73449daf
DW
359 unsigned long size = dax_entry_size(entry), pfn, index;
360 int i = 0;
d2c997c0
DW
361
362 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
363 return;
364
73449daf 365 index = linear_page_index(vma, address & ~(size - 1));
d2c997c0
DW
366 for_each_mapped_pfn(entry, pfn) {
367 struct page *page = pfn_to_page(pfn);
368
16900426
SR
369 if (shared) {
370 dax_page_share_get(page);
6061b69b
SR
371 } else {
372 WARN_ON_ONCE(page->mapping);
373 page->mapping = mapping;
374 page->index = index + i++;
375 }
d2c997c0
DW
376 }
377}
378
379static void dax_disassociate_entry(void *entry, struct address_space *mapping,
380 bool trunc)
381{
382 unsigned long pfn;
383
384 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
385 return;
386
387 for_each_mapped_pfn(entry, pfn) {
388 struct page *page = pfn_to_page(pfn);
389
390 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
16900426
SR
391 if (dax_page_is_shared(page)) {
392 /* keep the shared flag if this page is still shared */
393 if (dax_page_share_put(page) > 0)
6061b69b
SR
394 continue;
395 } else
396 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
d2c997c0 397 page->mapping = NULL;
73449daf 398 page->index = 0;
d2c997c0
DW
399 }
400}
401
5fac7408
DW
402static struct page *dax_busy_page(void *entry)
403{
404 unsigned long pfn;
405
406 for_each_mapped_pfn(entry, pfn) {
407 struct page *page = pfn_to_page(pfn);
408
409 if (page_ref_count(page) > 1)
410 return page;
411 }
412 return NULL;
413}
414
91e79d22
MWO
415/**
416 * dax_lock_folio - Lock the DAX entry corresponding to a folio
417 * @folio: The folio whose entry we want to lock
c5bbd451
MW
418 *
419 * Context: Process context.
91e79d22 420 * Return: A cookie to pass to dax_unlock_folio() or 0 if the entry could
27359fd6 421 * not be locked.
c5bbd451 422 */
91e79d22 423dax_entry_t dax_lock_folio(struct folio *folio)
c2a7d2a1 424{
9f32d221
MW
425 XA_STATE(xas, NULL, 0);
426 void *entry;
c2a7d2a1 427
91e79d22 428 /* Ensure folio->mapping isn't freed while we look at it */
c5bbd451 429 rcu_read_lock();
c2a7d2a1 430 for (;;) {
91e79d22 431 struct address_space *mapping = READ_ONCE(folio->mapping);
c2a7d2a1 432
27359fd6 433 entry = NULL;
c93db7bb 434 if (!mapping || !dax_mapping(mapping))
c5bbd451 435 break;
c2a7d2a1
DW
436
437 /*
438 * In the device-dax case there's no need to lock, a
439 * struct dev_pagemap pin is sufficient to keep the
440 * inode alive, and we assume we have dev_pagemap pin
441 * otherwise we would not have a valid pfn_to_page()
442 * translation.
443 */
27359fd6 444 entry = (void *)~0UL;
9f32d221 445 if (S_ISCHR(mapping->host->i_mode))
c5bbd451 446 break;
c2a7d2a1 447
9f32d221
MW
448 xas.xa = &mapping->i_pages;
449 xas_lock_irq(&xas);
91e79d22 450 if (mapping != folio->mapping) {
9f32d221 451 xas_unlock_irq(&xas);
c2a7d2a1
DW
452 continue;
453 }
91e79d22 454 xas_set(&xas, folio->index);
9f32d221
MW
455 entry = xas_load(&xas);
456 if (dax_is_locked(entry)) {
c5bbd451 457 rcu_read_unlock();
55e56f06 458 wait_entry_unlocked(&xas, entry);
c5bbd451 459 rcu_read_lock();
6d7cd8c1 460 continue;
c2a7d2a1 461 }
9f32d221
MW
462 dax_lock_entry(&xas, entry);
463 xas_unlock_irq(&xas);
c5bbd451 464 break;
c2a7d2a1 465 }
c5bbd451 466 rcu_read_unlock();
27359fd6 467 return (dax_entry_t)entry;
c2a7d2a1
DW
468}
469
91e79d22 470void dax_unlock_folio(struct folio *folio, dax_entry_t cookie)
c2a7d2a1 471{
91e79d22
MWO
472 struct address_space *mapping = folio->mapping;
473 XA_STATE(xas, &mapping->i_pages, folio->index);
c2a7d2a1 474
9f32d221 475 if (S_ISCHR(mapping->host->i_mode))
c2a7d2a1
DW
476 return;
477
27359fd6 478 dax_unlock_entry(&xas, (void *)cookie);
c2a7d2a1
DW
479}
480
2f437eff
SR
481/*
482 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a mapping
483 * @mapping: the file's mapping whose entry we want to lock
484 * @index: the offset within this file
485 * @page: output the dax page corresponding to this dax entry
486 *
487 * Return: A cookie to pass to dax_unlock_mapping_entry() or 0 if the entry
488 * could not be locked.
489 */
490dax_entry_t dax_lock_mapping_entry(struct address_space *mapping, pgoff_t index,
491 struct page **page)
492{
493 XA_STATE(xas, NULL, 0);
494 void *entry;
495
496 rcu_read_lock();
497 for (;;) {
498 entry = NULL;
499 if (!dax_mapping(mapping))
500 break;
501
502 xas.xa = &mapping->i_pages;
503 xas_lock_irq(&xas);
504 xas_set(&xas, index);
505 entry = xas_load(&xas);
506 if (dax_is_locked(entry)) {
507 rcu_read_unlock();
508 wait_entry_unlocked(&xas, entry);
509 rcu_read_lock();
510 continue;
511 }
512 if (!entry ||
513 dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
514 /*
515 * Because we are looking for entry from file's mapping
516 * and index, so the entry may not be inserted for now,
517 * or even a zero/empty entry. We don't think this is
518 * an error case. So, return a special value and do
519 * not output @page.
520 */
521 entry = (void *)~0UL;
522 } else {
523 *page = pfn_to_page(dax_to_pfn(entry));
524 dax_lock_entry(&xas, entry);
525 }
526 xas_unlock_irq(&xas);
527 break;
528 }
529 rcu_read_unlock();
530 return (dax_entry_t)entry;
531}
532
533void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index,
534 dax_entry_t cookie)
535{
536 XA_STATE(xas, &mapping->i_pages, index);
537
538 if (cookie == ~0UL)
539 return;
540
541 dax_unlock_entry(&xas, (void *)cookie);
542}
543
ac401cc7 544/*
a77d19f4
MW
545 * Find page cache entry at given index. If it is a DAX entry, return it
546 * with the entry locked. If the page cache doesn't contain an entry at
547 * that index, add a locked empty entry.
ac401cc7 548 *
3159f943 549 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
b15cd800
MW
550 * either return that locked entry or will return VM_FAULT_FALLBACK.
551 * This will happen if there are any PTE entries within the PMD range
552 * that we are requesting.
642261ac 553 *
b15cd800
MW
554 * We always favor PTE entries over PMD entries. There isn't a flow where we
555 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD
556 * insertion will fail if it finds any PTE entries already in the tree, and a
557 * PTE insertion will cause an existing PMD entry to be unmapped and
558 * downgraded to PTE entries. This happens for both PMD zero pages as
559 * well as PMD empty entries.
642261ac 560 *
b15cd800
MW
561 * The exception to this downgrade path is for PMD entries that have
562 * real storage backing them. We will leave these real PMD entries in
563 * the tree, and PTE writes will simply dirty the entire PMD entry.
642261ac 564 *
ac401cc7
JK
565 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
566 * persistent memory the benefit is doubtful. We can add that later if we can
567 * show it helps.
b15cd800
MW
568 *
569 * On error, this function does not return an ERR_PTR. Instead it returns
570 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values
571 * overlap with xarray value entries.
ac401cc7 572 */
b15cd800 573static void *grab_mapping_entry(struct xa_state *xas,
23c84eb7 574 struct address_space *mapping, unsigned int order)
ac401cc7 575{
b15cd800 576 unsigned long index = xas->xa_index;
1a14e377 577 bool pmd_downgrade; /* splitting PMD entry into PTE entries? */
b15cd800 578 void *entry;
642261ac 579
b15cd800 580retry:
1a14e377 581 pmd_downgrade = false;
b15cd800 582 xas_lock_irq(xas);
23c84eb7 583 entry = get_unlocked_entry(xas, order);
91d25ba8 584
642261ac 585 if (entry) {
23c84eb7
MWO
586 if (dax_is_conflict(entry))
587 goto fallback;
0e40de03 588 if (!xa_is_value(entry)) {
49688e65 589 xas_set_err(xas, -EIO);
b15cd800
MW
590 goto out_unlock;
591 }
592
23c84eb7 593 if (order == 0) {
91d25ba8 594 if (dax_is_pmd_entry(entry) &&
642261ac
RZ
595 (dax_is_zero_entry(entry) ||
596 dax_is_empty_entry(entry))) {
597 pmd_downgrade = true;
598 }
599 }
600 }
601
b15cd800
MW
602 if (pmd_downgrade) {
603 /*
604 * Make sure 'entry' remains valid while we drop
605 * the i_pages lock.
606 */
607 dax_lock_entry(xas, entry);
642261ac 608
642261ac
RZ
609 /*
610 * Besides huge zero pages the only other thing that gets
611 * downgraded are empty entries which don't need to be
612 * unmapped.
613 */
b15cd800
MW
614 if (dax_is_zero_entry(entry)) {
615 xas_unlock_irq(xas);
616 unmap_mapping_pages(mapping,
617 xas->xa_index & ~PG_PMD_COLOUR,
618 PG_PMD_NR, false);
619 xas_reset(xas);
620 xas_lock_irq(xas);
e11f8b7b
RZ
621 }
622
b15cd800
MW
623 dax_disassociate_entry(entry, mapping, false);
624 xas_store(xas, NULL); /* undo the PMD join */
698ab77a 625 dax_wake_entry(xas, entry, WAKE_ALL);
7f0e07fb 626 mapping->nrpages -= PG_PMD_NR;
b15cd800
MW
627 entry = NULL;
628 xas_set(xas, index);
629 }
642261ac 630
b15cd800
MW
631 if (entry) {
632 dax_lock_entry(xas, entry);
633 } else {
23c84eb7
MWO
634 unsigned long flags = DAX_EMPTY;
635
636 if (order > 0)
637 flags |= DAX_PMD;
638 entry = dax_make_entry(pfn_to_pfn_t(0), flags);
b15cd800
MW
639 dax_lock_entry(xas, entry);
640 if (xas_error(xas))
641 goto out_unlock;
7f0e07fb 642 mapping->nrpages += 1UL << order;
ac401cc7 643 }
b15cd800
MW
644
645out_unlock:
646 xas_unlock_irq(xas);
647 if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM))
648 goto retry;
649 if (xas->xa_node == XA_ERROR(-ENOMEM))
650 return xa_mk_internal(VM_FAULT_OOM);
651 if (xas_error(xas))
652 return xa_mk_internal(VM_FAULT_SIGBUS);
e3ad61c6 653 return entry;
b15cd800
MW
654fallback:
655 xas_unlock_irq(xas);
656 return xa_mk_internal(VM_FAULT_FALLBACK);
ac401cc7
JK
657}
658
5fac7408 659/**
6bbdd563 660 * dax_layout_busy_page_range - find first pinned page in @mapping
5fac7408 661 * @mapping: address space to scan for a page with ref count > 1
6bbdd563
VG
662 * @start: Starting offset. Page containing 'start' is included.
663 * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX,
664 * pages from 'start' till the end of file are included.
5fac7408
DW
665 *
666 * DAX requires ZONE_DEVICE mapped pages. These pages are never
667 * 'onlined' to the page allocator so they are considered idle when
668 * page->count == 1. A filesystem uses this interface to determine if
669 * any page in the mapping is busy, i.e. for DMA, or other
670 * get_user_pages() usages.
671 *
672 * It is expected that the filesystem is holding locks to block the
673 * establishment of new mappings in this address_space. I.e. it expects
674 * to be able to run unmap_mapping_range() and subsequently not race
675 * mapping_mapped() becoming true.
676 */
6bbdd563
VG
677struct page *dax_layout_busy_page_range(struct address_space *mapping,
678 loff_t start, loff_t end)
5fac7408 679{
084a8990
MW
680 void *entry;
681 unsigned int scanned = 0;
5fac7408 682 struct page *page = NULL;
6bbdd563
VG
683 pgoff_t start_idx = start >> PAGE_SHIFT;
684 pgoff_t end_idx;
685 XA_STATE(xas, &mapping->i_pages, start_idx);
5fac7408
DW
686
687 /*
688 * In the 'limited' case get_user_pages() for dax is disabled.
689 */
690 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
691 return NULL;
692
693 if (!dax_mapping(mapping) || !mapping_mapped(mapping))
694 return NULL;
695
6bbdd563
VG
696 /* If end == LLONG_MAX, all pages from start to till end of file */
697 if (end == LLONG_MAX)
698 end_idx = ULONG_MAX;
699 else
700 end_idx = end >> PAGE_SHIFT;
5fac7408
DW
701 /*
702 * If we race get_user_pages_fast() here either we'll see the
084a8990 703 * elevated page count in the iteration and wait, or
5fac7408
DW
704 * get_user_pages_fast() will see that the page it took a reference
705 * against is no longer mapped in the page tables and bail to the
706 * get_user_pages() slow path. The slow path is protected by
707 * pte_lock() and pmd_lock(). New references are not taken without
6bbdd563 708 * holding those locks, and unmap_mapping_pages() will not zero the
5fac7408
DW
709 * pte or pmd without holding the respective lock, so we are
710 * guaranteed to either see new references or prevent new
711 * references from being established.
712 */
6bbdd563 713 unmap_mapping_pages(mapping, start_idx, end_idx - start_idx + 1, 0);
5fac7408 714
084a8990 715 xas_lock_irq(&xas);
6bbdd563 716 xas_for_each(&xas, entry, end_idx) {
084a8990
MW
717 if (WARN_ON_ONCE(!xa_is_value(entry)))
718 continue;
719 if (unlikely(dax_is_locked(entry)))
23c84eb7 720 entry = get_unlocked_entry(&xas, 0);
084a8990
MW
721 if (entry)
722 page = dax_busy_page(entry);
4c3d043d 723 put_unlocked_entry(&xas, entry, WAKE_NEXT);
5fac7408
DW
724 if (page)
725 break;
084a8990
MW
726 if (++scanned % XA_CHECK_SCHED)
727 continue;
728
729 xas_pause(&xas);
730 xas_unlock_irq(&xas);
731 cond_resched();
732 xas_lock_irq(&xas);
5fac7408 733 }
084a8990 734 xas_unlock_irq(&xas);
5fac7408
DW
735 return page;
736}
6bbdd563
VG
737EXPORT_SYMBOL_GPL(dax_layout_busy_page_range);
738
739struct page *dax_layout_busy_page(struct address_space *mapping)
740{
741 return dax_layout_busy_page_range(mapping, 0, LLONG_MAX);
742}
5fac7408
DW
743EXPORT_SYMBOL_GPL(dax_layout_busy_page);
744
a77d19f4 745static int __dax_invalidate_entry(struct address_space *mapping,
c6dcf52c
JK
746 pgoff_t index, bool trunc)
747{
07f2d89c 748 XA_STATE(xas, &mapping->i_pages, index);
c6dcf52c
JK
749 int ret = 0;
750 void *entry;
c6dcf52c 751
07f2d89c 752 xas_lock_irq(&xas);
23c84eb7 753 entry = get_unlocked_entry(&xas, 0);
3159f943 754 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
c6dcf52c
JK
755 goto out;
756 if (!trunc &&
07f2d89c
MW
757 (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) ||
758 xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE)))
c6dcf52c 759 goto out;
d2c997c0 760 dax_disassociate_entry(entry, mapping, trunc);
07f2d89c 761 xas_store(&xas, NULL);
7f0e07fb 762 mapping->nrpages -= 1UL << dax_entry_order(entry);
c6dcf52c
JK
763 ret = 1;
764out:
23738832 765 put_unlocked_entry(&xas, entry, WAKE_ALL);
07f2d89c 766 xas_unlock_irq(&xas);
c6dcf52c
JK
767 return ret;
768}
07f2d89c 769
f76b3a32
SR
770static int __dax_clear_dirty_range(struct address_space *mapping,
771 pgoff_t start, pgoff_t end)
772{
773 XA_STATE(xas, &mapping->i_pages, start);
774 unsigned int scanned = 0;
775 void *entry;
776
777 xas_lock_irq(&xas);
778 xas_for_each(&xas, entry, end) {
779 entry = get_unlocked_entry(&xas, 0);
780 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
781 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
782 put_unlocked_entry(&xas, entry, WAKE_NEXT);
783
784 if (++scanned % XA_CHECK_SCHED)
785 continue;
786
787 xas_pause(&xas);
788 xas_unlock_irq(&xas);
789 cond_resched();
790 xas_lock_irq(&xas);
791 }
792 xas_unlock_irq(&xas);
793
794 return 0;
795}
796
ac401cc7 797/*
3159f943
MW
798 * Delete DAX entry at @index from @mapping. Wait for it
799 * to be unlocked before deleting it.
ac401cc7
JK
800 */
801int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
802{
a77d19f4 803 int ret = __dax_invalidate_entry(mapping, index, true);
ac401cc7 804
ac401cc7
JK
805 /*
806 * This gets called from truncate / punch_hole path. As such, the caller
807 * must hold locks protecting against concurrent modifications of the
a77d19f4 808 * page cache (usually fs-private i_mmap_sem for writing). Since the
3159f943 809 * caller has seen a DAX entry for this index, we better find it
ac401cc7
JK
810 * at that index as well...
811 */
c6dcf52c
JK
812 WARN_ON_ONCE(!ret);
813 return ret;
814}
815
c6dcf52c 816/*
3159f943 817 * Invalidate DAX entry if it is clean.
c6dcf52c
JK
818 */
819int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
820 pgoff_t index)
821{
a77d19f4 822 return __dax_invalidate_entry(mapping, index, false);
ac401cc7
JK
823}
824
60696eb2 825static pgoff_t dax_iomap_pgoff(const struct iomap *iomap, loff_t pos)
f7ca90b1 826{
de205114 827 return PHYS_PFN(iomap->addr + (pos & PAGE_MASK) - iomap->offset);
429f8de7
CH
828}
829
830static int copy_cow_page_dax(struct vm_fault *vmf, const struct iomap_iter *iter)
831{
60696eb2 832 pgoff_t pgoff = dax_iomap_pgoff(&iter->iomap, iter->pos);
cccbce67 833 void *vto, *kaddr;
cccbce67
DW
834 long rc;
835 int id;
836
cccbce67 837 id = dax_read_lock();
e511c4a3
JC
838 rc = dax_direct_access(iter->iomap.dax_dev, pgoff, 1, DAX_ACCESS,
839 &kaddr, NULL);
cccbce67
DW
840 if (rc < 0) {
841 dax_read_unlock(id);
842 return rc;
843 }
429f8de7
CH
844 vto = kmap_atomic(vmf->cow_page);
845 copy_user_page(vto, kaddr, vmf->address, vmf->cow_page);
f7ca90b1 846 kunmap_atomic(vto);
cccbce67 847 dax_read_unlock(id);
f7ca90b1
MW
848 return 0;
849}
850
e5d6df73
SR
851/*
852 * MAP_SYNC on a dax mapping guarantees dirty metadata is
853 * flushed on write-faults (non-cow), but not read-faults.
854 */
855static bool dax_fault_is_synchronous(const struct iomap_iter *iter,
856 struct vm_area_struct *vma)
857{
858 return (iter->flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) &&
859 (iter->iomap.flags & IOMAP_F_DIRTY);
860}
861
642261ac
RZ
862/*
863 * By this point grab_mapping_entry() has ensured that we have a locked entry
864 * of the appropriate size so we don't have to worry about downgrading PMDs to
865 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
866 * already in the tree, we will skip the insertion and just dirty the PMD as
867 * appropriate.
868 */
e5d6df73
SR
869static void *dax_insert_entry(struct xa_state *xas, struct vm_fault *vmf,
870 const struct iomap_iter *iter, void *entry, pfn_t pfn,
871 unsigned long flags)
9973c98e 872{
e5d6df73 873 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
b15cd800 874 void *new_entry = dax_make_entry(pfn, flags);
c6f0b395
SR
875 bool write = iter->flags & IOMAP_WRITE;
876 bool dirty = write && !dax_fault_is_synchronous(iter, vmf->vma);
877 bool shared = iter->iomap.flags & IOMAP_F_SHARED;
9973c98e 878
f5b7b748 879 if (dirty)
d2b2a28e 880 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
9973c98e 881
c6f0b395 882 if (shared || (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE))) {
b15cd800 883 unsigned long index = xas->xa_index;
91d25ba8
RZ
884 /* we are replacing a zero page with block mapping */
885 if (dax_is_pmd_entry(entry))
977fbdcd 886 unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
b15cd800 887 PG_PMD_NR, false);
91d25ba8 888 else /* pte entry */
b15cd800 889 unmap_mapping_pages(mapping, index, 1, false);
9973c98e
RZ
890 }
891
b15cd800
MW
892 xas_reset(xas);
893 xas_lock_irq(xas);
c6f0b395 894 if (shared || dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
1571c029
JK
895 void *old;
896
d2c997c0 897 dax_disassociate_entry(entry, mapping, false);
6061b69b 898 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address,
c6f0b395 899 shared);
642261ac 900 /*
a77d19f4 901 * Only swap our new entry into the page cache if the current
642261ac 902 * entry is a zero page or an empty entry. If a normal PTE or
a77d19f4 903 * PMD entry is already in the cache, we leave it alone. This
642261ac
RZ
904 * means that if we are trying to insert a PTE and the
905 * existing entry is a PMD, we will just leave the PMD in the
906 * tree and dirty it if necessary.
907 */
1571c029 908 old = dax_lock_entry(xas, new_entry);
b15cd800
MW
909 WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) |
910 DAX_LOCKED));
91d25ba8 911 entry = new_entry;
b15cd800
MW
912 } else {
913 xas_load(xas); /* Walk the xa_state */
9973c98e 914 }
91d25ba8 915
f5b7b748 916 if (dirty)
b15cd800 917 xas_set_mark(xas, PAGECACHE_TAG_DIRTY);
91d25ba8 918
c6f0b395 919 if (write && shared)
e5d6df73
SR
920 xas_set_mark(xas, PAGECACHE_TAG_TOWRITE);
921
b15cd800 922 xas_unlock_irq(xas);
91d25ba8 923 return entry;
9973c98e
RZ
924}
925
9fc747f6
MW
926static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev,
927 struct address_space *mapping, void *entry)
9973c98e 928{
06083a09 929 unsigned long pfn, index, count, end;
3fe0791c 930 long ret = 0;
06083a09 931 struct vm_area_struct *vma;
9973c98e 932
9973c98e 933 /*
a6abc2c0
JK
934 * A page got tagged dirty in DAX mapping? Something is seriously
935 * wrong.
9973c98e 936 */
3159f943 937 if (WARN_ON(!xa_is_value(entry)))
a6abc2c0 938 return -EIO;
9973c98e 939
9fc747f6
MW
940 if (unlikely(dax_is_locked(entry))) {
941 void *old_entry = entry;
942
23c84eb7 943 entry = get_unlocked_entry(xas, 0);
9fc747f6
MW
944
945 /* Entry got punched out / reallocated? */
946 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
947 goto put_unlocked;
948 /*
949 * Entry got reallocated elsewhere? No need to writeback.
950 * We have to compare pfns as we must not bail out due to
951 * difference in lockbit or entry type.
952 */
953 if (dax_to_pfn(old_entry) != dax_to_pfn(entry))
954 goto put_unlocked;
955 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
956 dax_is_zero_entry(entry))) {
957 ret = -EIO;
958 goto put_unlocked;
959 }
960
961 /* Another fsync thread may have already done this entry */
962 if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE))
963 goto put_unlocked;
9973c98e
RZ
964 }
965
a6abc2c0 966 /* Lock the entry to serialize with page faults */
9fc747f6
MW
967 dax_lock_entry(xas, entry);
968
a6abc2c0
JK
969 /*
970 * We can clear the tag now but we have to be careful so that concurrent
971 * dax_writeback_one() calls for the same index cannot finish before we
972 * actually flush the caches. This is achieved as the calls will look
b93b0163
MW
973 * at the entry only under the i_pages lock and once they do that
974 * they will see the entry locked and wait for it to unlock.
a6abc2c0 975 */
9fc747f6
MW
976 xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE);
977 xas_unlock_irq(xas);
a6abc2c0 978
642261ac 979 /*
e4b3448b
MW
980 * If dax_writeback_mapping_range() was given a wbc->range_start
981 * in the middle of a PMD, the 'index' we use needs to be
982 * aligned to the start of the PMD.
3fe0791c
DW
983 * This allows us to flush for PMD_SIZE and not have to worry about
984 * partial PMD writebacks.
642261ac 985 */
a77d19f4 986 pfn = dax_to_pfn(entry);
e4b3448b
MW
987 count = 1UL << dax_entry_order(entry);
988 index = xas->xa_index & ~(count - 1);
06083a09
MS
989 end = index + count - 1;
990
991 /* Walk all mappings of a given index of a file and writeprotect them */
992 i_mmap_lock_read(mapping);
993 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, end) {
994 pfn_mkclean_range(pfn, count, index, vma);
995 cond_resched();
996 }
997 i_mmap_unlock_read(mapping);
cccbce67 998
e4b3448b 999 dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE);
4b4bb46d
JK
1000 /*
1001 * After we have flushed the cache, we can clear the dirty tag. There
1002 * cannot be new dirty data in the pfn after the flush has completed as
1003 * the pfn mappings are writeprotected and fault waits for mapping
1004 * entry lock.
1005 */
9fc747f6
MW
1006 xas_reset(xas);
1007 xas_lock_irq(xas);
1008 xas_store(xas, entry);
1009 xas_clear_mark(xas, PAGECACHE_TAG_DIRTY);
698ab77a 1010 dax_wake_entry(xas, entry, WAKE_NEXT);
9fc747f6 1011
e4b3448b 1012 trace_dax_writeback_one(mapping->host, index, count);
9973c98e
RZ
1013 return ret;
1014
a6abc2c0 1015 put_unlocked:
4c3d043d 1016 put_unlocked_entry(xas, entry, WAKE_NEXT);
9973c98e
RZ
1017 return ret;
1018}
1019
1020/*
1021 * Flush the mapping to the persistent domain within the byte range of [start,
1022 * end]. This is required by data integrity operations to ensure file data is
1023 * on persistent storage prior to completion of the operation.
1024 */
7f6d5b52 1025int dax_writeback_mapping_range(struct address_space *mapping,
3f666c56 1026 struct dax_device *dax_dev, struct writeback_control *wbc)
9973c98e 1027{
9fc747f6 1028 XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT);
9973c98e 1029 struct inode *inode = mapping->host;
9fc747f6 1030 pgoff_t end_index = wbc->range_end >> PAGE_SHIFT;
9fc747f6
MW
1031 void *entry;
1032 int ret = 0;
1033 unsigned int scanned = 0;
9973c98e
RZ
1034
1035 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
1036 return -EIO;
1037
7716506a 1038 if (mapping_empty(mapping) || wbc->sync_mode != WB_SYNC_ALL)
7f6d5b52
RZ
1039 return 0;
1040
9fc747f6 1041 trace_dax_writeback_range(inode, xas.xa_index, end_index);
9973c98e 1042
9fc747f6 1043 tag_pages_for_writeback(mapping, xas.xa_index, end_index);
9973c98e 1044
9fc747f6
MW
1045 xas_lock_irq(&xas);
1046 xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) {
1047 ret = dax_writeback_one(&xas, dax_dev, mapping, entry);
1048 if (ret < 0) {
1049 mapping_set_error(mapping, ret);
9973c98e 1050 break;
9973c98e 1051 }
9fc747f6
MW
1052 if (++scanned % XA_CHECK_SCHED)
1053 continue;
1054
1055 xas_pause(&xas);
1056 xas_unlock_irq(&xas);
1057 cond_resched();
1058 xas_lock_irq(&xas);
9973c98e 1059 }
9fc747f6 1060 xas_unlock_irq(&xas);
9fc747f6
MW
1061 trace_dax_writeback_range_done(inode, xas.xa_index, end_index);
1062 return ret;
9973c98e
RZ
1063}
1064EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
1065
e28cd3e5
SR
1066static int dax_iomap_direct_access(const struct iomap *iomap, loff_t pos,
1067 size_t size, void **kaddr, pfn_t *pfnp)
f7ca90b1 1068{
60696eb2 1069 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
e28cd3e5 1070 int id, rc = 0;
5e161e40 1071 long length;
f7ca90b1 1072
cccbce67 1073 id = dax_read_lock();
5e161e40 1074 length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
e28cd3e5 1075 DAX_ACCESS, kaddr, pfnp);
5e161e40
JK
1076 if (length < 0) {
1077 rc = length;
1078 goto out;
cccbce67 1079 }
e28cd3e5
SR
1080 if (!pfnp)
1081 goto out_check_addr;
5e161e40
JK
1082 rc = -EINVAL;
1083 if (PFN_PHYS(length) < size)
1084 goto out;
1085 if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1086 goto out;
1087 /* For larger pages we need devmap */
1088 if (length > 1 && !pfn_t_devmap(*pfnp))
1089 goto out;
1090 rc = 0;
e28cd3e5
SR
1091
1092out_check_addr:
1093 if (!kaddr)
1094 goto out;
1095 if (!*kaddr)
1096 rc = -EFAULT;
5e161e40 1097out:
cccbce67 1098 dax_read_unlock(id);
5e161e40 1099 return rc;
0e3b210c 1100}
0e3b210c 1101
ff17b8df 1102/**
708dfad2
SR
1103 * dax_iomap_copy_around - Prepare for an unaligned write to a shared/cow page
1104 * by copying the data before and after the range to be written.
ff17b8df
SR
1105 * @pos: address to do copy from.
1106 * @length: size of copy operation.
1107 * @align_size: aligned w.r.t align_size (either PMD_SIZE or PAGE_SIZE)
1108 * @srcmap: iomap srcmap
1109 * @daddr: destination address to copy to.
1110 *
1111 * This can be called from two places. Either during DAX write fault (page
1112 * aligned), to copy the length size data to daddr. Or, while doing normal DAX
708dfad2 1113 * write operation, dax_iomap_iter() might call this to do the copy of either
ff17b8df 1114 * start or end unaligned address. In the latter case the rest of the copy of
708dfad2
SR
1115 * aligned ranges is taken care by dax_iomap_iter() itself.
1116 * If the srcmap contains invalid data, such as HOLE and UNWRITTEN, zero the
1117 * area to make sure no old data remains.
ff17b8df 1118 */
708dfad2 1119static int dax_iomap_copy_around(loff_t pos, uint64_t length, size_t align_size,
ff17b8df
SR
1120 const struct iomap *srcmap, void *daddr)
1121{
1122 loff_t head_off = pos & (align_size - 1);
1123 size_t size = ALIGN(head_off + length, align_size);
1124 loff_t end = pos + length;
1125 loff_t pg_end = round_up(end, align_size);
708dfad2 1126 /* copy_all is usually in page fault case */
ff17b8df 1127 bool copy_all = head_off == 0 && end == pg_end;
708dfad2
SR
1128 /* zero the edges if srcmap is a HOLE or IOMAP_UNWRITTEN */
1129 bool zero_edge = srcmap->flags & IOMAP_F_SHARED ||
1130 srcmap->type == IOMAP_UNWRITTEN;
ff17b8df
SR
1131 void *saddr = 0;
1132 int ret = 0;
1133
708dfad2
SR
1134 if (!zero_edge) {
1135 ret = dax_iomap_direct_access(srcmap, pos, size, &saddr, NULL);
1136 if (ret)
1ea7ca1b 1137 return dax_mem2blk_err(ret);
708dfad2 1138 }
ff17b8df
SR
1139
1140 if (copy_all) {
708dfad2
SR
1141 if (zero_edge)
1142 memset(daddr, 0, size);
1143 else
1144 ret = copy_mc_to_kernel(daddr, saddr, length);
1145 goto out;
ff17b8df
SR
1146 }
1147
1148 /* Copy the head part of the range */
1149 if (head_off) {
708dfad2
SR
1150 if (zero_edge)
1151 memset(daddr, 0, head_off);
1152 else {
1153 ret = copy_mc_to_kernel(daddr, saddr, head_off);
1154 if (ret)
1155 return -EIO;
1156 }
ff17b8df
SR
1157 }
1158
1159 /* Copy the tail part of the range */
1160 if (end < pg_end) {
1161 loff_t tail_off = head_off + length;
1162 loff_t tail_len = pg_end - end;
1163
708dfad2
SR
1164 if (zero_edge)
1165 memset(daddr + tail_off, 0, tail_len);
1166 else {
1167 ret = copy_mc_to_kernel(daddr + tail_off,
1168 saddr + tail_off, tail_len);
1169 if (ret)
1170 return -EIO;
1171 }
ff17b8df 1172 }
708dfad2
SR
1173out:
1174 if (zero_edge)
1175 dax_flush(srcmap->dax_dev, daddr, size);
1176 return ret ? -EIO : 0;
ff17b8df
SR
1177}
1178
e30331ff 1179/*
91d25ba8
RZ
1180 * The user has performed a load from a hole in the file. Allocating a new
1181 * page in the file would cause excessive storage usage for workloads with
1182 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1183 * If this page is ever written to we will re-fault and change the mapping to
1184 * point to real DAX storage instead.
e30331ff 1185 */
e5d6df73
SR
1186static vm_fault_t dax_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1187 const struct iomap_iter *iter, void **entry)
e30331ff 1188{
e5d6df73 1189 struct inode *inode = iter->inode;
91d25ba8 1190 unsigned long vaddr = vmf->address;
b90ca5cc
MW
1191 pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1192 vm_fault_t ret;
e30331ff 1193
e5d6df73 1194 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, DAX_ZERO_PAGE);
3159f943 1195
ab77dab4 1196 ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
e30331ff
RZ
1197 trace_dax_load_hole(inode, vmf, ret);
1198 return ret;
1199}
1200
c2436190
SR
1201#ifdef CONFIG_FS_DAX_PMD
1202static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
e5d6df73 1203 const struct iomap_iter *iter, void **entry)
c2436190
SR
1204{
1205 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1206 unsigned long pmd_addr = vmf->address & PMD_MASK;
1207 struct vm_area_struct *vma = vmf->vma;
1208 struct inode *inode = mapping->host;
1209 pgtable_t pgtable = NULL;
1210 struct page *zero_page;
1211 spinlock_t *ptl;
1212 pmd_t pmd_entry;
1213 pfn_t pfn;
1214
1215 zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1216
1217 if (unlikely(!zero_page))
1218 goto fallback;
1219
1220 pfn = page_to_pfn_t(zero_page);
e5d6df73
SR
1221 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn,
1222 DAX_PMD | DAX_ZERO_PAGE);
c2436190
SR
1223
1224 if (arch_needs_pgtable_deposit()) {
1225 pgtable = pte_alloc_one(vma->vm_mm);
1226 if (!pgtable)
1227 return VM_FAULT_OOM;
1228 }
1229
1230 ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1231 if (!pmd_none(*(vmf->pmd))) {
1232 spin_unlock(ptl);
1233 goto fallback;
1234 }
1235
1236 if (pgtable) {
1237 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1238 mm_inc_nr_ptes(vma->vm_mm);
1239 }
1240 pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1241 pmd_entry = pmd_mkhuge(pmd_entry);
1242 set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1243 spin_unlock(ptl);
1244 trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry);
1245 return VM_FAULT_NOPAGE;
1246
1247fallback:
1248 if (pgtable)
1249 pte_free(vma->vm_mm, pgtable);
1250 trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry);
1251 return VM_FAULT_FALLBACK;
1252}
1253#else
1254static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
e5d6df73 1255 const struct iomap_iter *iter, void **entry)
c2436190
SR
1256{
1257 return VM_FAULT_FALLBACK;
1258}
1259#endif /* CONFIG_FS_DAX_PMD */
1260
d984648e
SR
1261static s64 dax_unshare_iter(struct iomap_iter *iter)
1262{
1263 struct iomap *iomap = &iter->iomap;
1264 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1265 loff_t pos = iter->pos;
1266 loff_t length = iomap_length(iter);
1267 int id = 0;
1268 s64 ret = 0;
1269 void *daddr = NULL, *saddr = NULL;
1270
1271 /* don't bother with blocks that are not shared to start with */
1272 if (!(iomap->flags & IOMAP_F_SHARED))
1273 return length;
d984648e
SR
1274
1275 id = dax_read_lock();
1276 ret = dax_iomap_direct_access(iomap, pos, length, &daddr, NULL);
1277 if (ret < 0)
1278 goto out_unlock;
1279
13dd4e04
SR
1280 /* zero the distance if srcmap is HOLE or UNWRITTEN */
1281 if (srcmap->flags & IOMAP_F_SHARED || srcmap->type == IOMAP_UNWRITTEN) {
1282 memset(daddr, 0, length);
1283 dax_flush(iomap->dax_dev, daddr, length);
1284 ret = length;
1285 goto out_unlock;
1286 }
1287
d984648e
SR
1288 ret = dax_iomap_direct_access(srcmap, pos, length, &saddr, NULL);
1289 if (ret < 0)
1290 goto out_unlock;
1291
388bc034
SR
1292 if (copy_mc_to_kernel(daddr, saddr, length) == 0)
1293 ret = length;
1294 else
d984648e
SR
1295 ret = -EIO;
1296
1297out_unlock:
1298 dax_read_unlock(id);
1ea7ca1b 1299 return dax_mem2blk_err(ret);
d984648e
SR
1300}
1301
1302int dax_file_unshare(struct inode *inode, loff_t pos, loff_t len,
1303 const struct iomap_ops *ops)
1304{
1305 struct iomap_iter iter = {
1306 .inode = inode,
1307 .pos = pos,
1308 .len = len,
1309 .flags = IOMAP_WRITE | IOMAP_UNSHARE | IOMAP_DAX,
1310 };
1311 int ret;
1312
1313 while ((ret = iomap_iter(&iter, ops)) > 0)
1314 iter.processed = dax_unshare_iter(&iter);
1315 return ret;
1316}
1317EXPORT_SYMBOL_GPL(dax_file_unshare);
1318
8dbfc76d 1319static int dax_memzero(struct iomap_iter *iter, loff_t pos, size_t size)
e5c71954 1320{
8dbfc76d
SR
1321 const struct iomap *iomap = &iter->iomap;
1322 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1323 unsigned offset = offset_in_page(pos);
1324 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
e5c71954
CH
1325 void *kaddr;
1326 long ret;
1327
8dbfc76d
SR
1328 ret = dax_direct_access(iomap->dax_dev, pgoff, 1, DAX_ACCESS, &kaddr,
1329 NULL);
1330 if (ret < 0)
1ea7ca1b
JC
1331 return dax_mem2blk_err(ret);
1332
8dbfc76d 1333 memset(kaddr + offset, 0, size);
708dfad2
SR
1334 if (iomap->flags & IOMAP_F_SHARED)
1335 ret = dax_iomap_copy_around(pos, size, PAGE_SIZE, srcmap,
1336 kaddr);
1337 else
8dbfc76d 1338 dax_flush(iomap->dax_dev, kaddr + offset, size);
e5c71954
CH
1339 return ret;
1340}
1341
c6f40468 1342static s64 dax_zero_iter(struct iomap_iter *iter, bool *did_zero)
679c8bd3 1343{
c6f40468
CH
1344 const struct iomap *iomap = &iter->iomap;
1345 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1346 loff_t pos = iter->pos;
1347 u64 length = iomap_length(iter);
1348 s64 written = 0;
1349
1350 /* already zeroed? we're done. */
1351 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
1352 return length;
1353
f80e1668
SR
1354 /*
1355 * invalidate the pages whose sharing state is to be changed
1356 * because of CoW.
1357 */
1358 if (iomap->flags & IOMAP_F_SHARED)
1359 invalidate_inode_pages2_range(iter->inode->i_mapping,
1360 pos >> PAGE_SHIFT,
1361 (pos + length - 1) >> PAGE_SHIFT);
1362
c6f40468
CH
1363 do {
1364 unsigned offset = offset_in_page(pos);
1365 unsigned size = min_t(u64, PAGE_SIZE - offset, length);
1366 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1367 long rc;
1368 int id;
1369
1370 id = dax_read_lock();
1371 if (IS_ALIGNED(pos, PAGE_SIZE) && size == PAGE_SIZE)
1372 rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1);
1373 else
8dbfc76d 1374 rc = dax_memzero(iter, pos, size);
c6f40468 1375 dax_read_unlock(id);
cccbce67 1376
c6f40468
CH
1377 if (rc < 0)
1378 return rc;
1379 pos += size;
1380 length -= size;
1381 written += size;
c6f40468 1382 } while (length > 0);
e5c71954 1383
f8189d5d
KX
1384 if (did_zero)
1385 *did_zero = true;
c6f40468
CH
1386 return written;
1387}
1388
1389int dax_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1390 const struct iomap_ops *ops)
1391{
1392 struct iomap_iter iter = {
1393 .inode = inode,
1394 .pos = pos,
1395 .len = len,
952da063 1396 .flags = IOMAP_DAX | IOMAP_ZERO,
c6f40468
CH
1397 };
1398 int ret;
1399
1400 while ((ret = iomap_iter(&iter, ops)) > 0)
1401 iter.processed = dax_zero_iter(&iter, did_zero);
1402 return ret;
1403}
1404EXPORT_SYMBOL_GPL(dax_zero_range);
1405
1406int dax_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1407 const struct iomap_ops *ops)
1408{
1409 unsigned int blocksize = i_blocksize(inode);
1410 unsigned int off = pos & (blocksize - 1);
1411
1412 /* Block boundary? Nothing to do */
1413 if (!off)
1414 return 0;
1415 return dax_zero_range(inode, pos, blocksize - off, did_zero, ops);
679c8bd3 1416}
c6f40468 1417EXPORT_SYMBOL_GPL(dax_truncate_page);
679c8bd3 1418
ca289e0b
CH
1419static loff_t dax_iomap_iter(const struct iomap_iter *iomi,
1420 struct iov_iter *iter)
a254e568 1421{
ca289e0b 1422 const struct iomap *iomap = &iomi->iomap;
f80e1668 1423 const struct iomap *srcmap = iomap_iter_srcmap(iomi);
ca289e0b
CH
1424 loff_t length = iomap_length(iomi);
1425 loff_t pos = iomi->pos;
cccbce67 1426 struct dax_device *dax_dev = iomap->dax_dev;
a254e568 1427 loff_t end = pos + length, done = 0;
ff17b8df 1428 bool write = iov_iter_rw(iter) == WRITE;
f80e1668 1429 bool cow = write && iomap->flags & IOMAP_F_SHARED;
a254e568 1430 ssize_t ret = 0;
a77d4786 1431 size_t xfer;
cccbce67 1432 int id;
a254e568 1433
ff17b8df 1434 if (!write) {
ca289e0b 1435 end = min(end, i_size_read(iomi->inode));
a254e568
CH
1436 if (pos >= end)
1437 return 0;
1438
1439 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1440 return iov_iter_zero(min(length, end - pos), iter);
1441 }
1442
ff17b8df
SR
1443 /*
1444 * In DAX mode, enforce either pure overwrites of written extents, or
1445 * writes to unwritten extents as part of a copy-on-write operation.
1446 */
1447 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED &&
1448 !(iomap->flags & IOMAP_F_SHARED)))
a254e568
CH
1449 return -EIO;
1450
e3fce68c
JK
1451 /*
1452 * Write can allocate block for an area which has a hole page mapped
1453 * into page tables. We have to tear down these mappings so that data
1454 * written by write(2) is visible in mmap.
1455 */
f80e1668 1456 if (iomap->flags & IOMAP_F_NEW || cow) {
f76b3a32
SR
1457 /*
1458 * Filesystem allows CoW on non-shared extents. The src extents
1459 * may have been mmapped with dirty mark before. To be able to
1460 * invalidate its dax entries, we need to clear the dirty mark
1461 * in advance.
1462 */
1463 if (cow)
1464 __dax_clear_dirty_range(iomi->inode->i_mapping,
1465 pos >> PAGE_SHIFT,
1466 (end - 1) >> PAGE_SHIFT);
ca289e0b 1467 invalidate_inode_pages2_range(iomi->inode->i_mapping,
e3fce68c
JK
1468 pos >> PAGE_SHIFT,
1469 (end - 1) >> PAGE_SHIFT);
1470 }
1471
cccbce67 1472 id = dax_read_lock();
a254e568
CH
1473 while (pos < end) {
1474 unsigned offset = pos & (PAGE_SIZE - 1);
cccbce67 1475 const size_t size = ALIGN(length + offset, PAGE_SIZE);
60696eb2 1476 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
a254e568 1477 ssize_t map_len;
047218ec 1478 bool recovery = false;
cccbce67 1479 void *kaddr;
a254e568 1480
d1908f52
MH
1481 if (fatal_signal_pending(current)) {
1482 ret = -EINTR;
1483 break;
1484 }
1485
cccbce67 1486 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
e511c4a3 1487 DAX_ACCESS, &kaddr, NULL);
1ea7ca1b 1488 if (map_len == -EHWPOISON && iov_iter_rw(iter) == WRITE) {
047218ec
JC
1489 map_len = dax_direct_access(dax_dev, pgoff,
1490 PHYS_PFN(size), DAX_RECOVERY_WRITE,
1491 &kaddr, NULL);
1492 if (map_len > 0)
1493 recovery = true;
1494 }
a254e568 1495 if (map_len < 0) {
1ea7ca1b 1496 ret = dax_mem2blk_err(map_len);
a254e568
CH
1497 break;
1498 }
1499
f80e1668 1500 if (cow) {
708dfad2
SR
1501 ret = dax_iomap_copy_around(pos, length, PAGE_SIZE,
1502 srcmap, kaddr);
ff17b8df
SR
1503 if (ret)
1504 break;
1505 }
1506
cccbce67
DW
1507 map_len = PFN_PHYS(map_len);
1508 kaddr += offset;
a254e568
CH
1509 map_len -= offset;
1510 if (map_len > end - pos)
1511 map_len = end - pos;
1512
047218ec
JC
1513 if (recovery)
1514 xfer = dax_recovery_write(dax_dev, pgoff, kaddr,
1515 map_len, iter);
ff17b8df 1516 else if (write)
a77d4786 1517 xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
fec53774 1518 map_len, iter);
a254e568 1519 else
a77d4786 1520 xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
b3a9a0c3 1521 map_len, iter);
a254e568 1522
a77d4786
DW
1523 pos += xfer;
1524 length -= xfer;
1525 done += xfer;
1526
1527 if (xfer == 0)
1528 ret = -EFAULT;
1529 if (xfer < map_len)
1530 break;
a254e568 1531 }
cccbce67 1532 dax_read_unlock(id);
a254e568
CH
1533
1534 return done ? done : ret;
1535}
1536
1537/**
11c59c92 1538 * dax_iomap_rw - Perform I/O to a DAX file
a254e568
CH
1539 * @iocb: The control block for this I/O
1540 * @iter: The addresses to do I/O from or to
1541 * @ops: iomap ops passed from the file system
1542 *
1543 * This function performs read and write operations to directly mapped
1544 * persistent memory. The callers needs to take care of read/write exclusion
1545 * and evicting any page cache pages in the region under I/O.
1546 */
1547ssize_t
11c59c92 1548dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
8ff6daa1 1549 const struct iomap_ops *ops)
a254e568 1550{
ca289e0b
CH
1551 struct iomap_iter iomi = {
1552 .inode = iocb->ki_filp->f_mapping->host,
1553 .pos = iocb->ki_pos,
1554 .len = iov_iter_count(iter),
952da063 1555 .flags = IOMAP_DAX,
ca289e0b
CH
1556 };
1557 loff_t done = 0;
1558 int ret;
a254e568 1559
17d9c15c
LJ
1560 if (!iomi.len)
1561 return 0;
1562
168316db 1563 if (iov_iter_rw(iter) == WRITE) {
ca289e0b
CH
1564 lockdep_assert_held_write(&iomi.inode->i_rwsem);
1565 iomi.flags |= IOMAP_WRITE;
168316db 1566 } else {
ca289e0b 1567 lockdep_assert_held(&iomi.inode->i_rwsem);
168316db 1568 }
a254e568 1569
96222d53 1570 if (iocb->ki_flags & IOCB_NOWAIT)
ca289e0b 1571 iomi.flags |= IOMAP_NOWAIT;
96222d53 1572
ca289e0b
CH
1573 while ((ret = iomap_iter(&iomi, ops)) > 0)
1574 iomi.processed = dax_iomap_iter(&iomi, iter);
a254e568 1575
ca289e0b
CH
1576 done = iomi.pos - iocb->ki_pos;
1577 iocb->ki_pos = iomi.pos;
a254e568
CH
1578 return done ? done : ret;
1579}
11c59c92 1580EXPORT_SYMBOL_GPL(dax_iomap_rw);
a7d73fe6 1581
ab77dab4 1582static vm_fault_t dax_fault_return(int error)
9f141d6e
JK
1583{
1584 if (error == 0)
1585 return VM_FAULT_NOPAGE;
c9aed74e 1586 return vmf_error(error);
9f141d6e
JK
1587}
1588
55f81639
SR
1589/*
1590 * When handling a synchronous page fault and the inode need a fsync, we can
1591 * insert the PTE/PMD into page tables only after that fsync happened. Skip
1592 * insertion for now and return the pfn so that caller can insert it after the
1593 * fsync is done.
1594 */
1595static vm_fault_t dax_fault_synchronous_pfnp(pfn_t *pfnp, pfn_t pfn)
1596{
1597 if (WARN_ON_ONCE(!pfnp))
1598 return VM_FAULT_SIGBUS;
1599 *pfnp = pfn;
1600 return VM_FAULT_NEEDDSYNC;
1601}
1602
65dd814a
CH
1603static vm_fault_t dax_fault_cow_page(struct vm_fault *vmf,
1604 const struct iomap_iter *iter)
55f81639 1605{
55f81639
SR
1606 vm_fault_t ret;
1607 int error = 0;
1608
65dd814a 1609 switch (iter->iomap.type) {
55f81639
SR
1610 case IOMAP_HOLE:
1611 case IOMAP_UNWRITTEN:
429f8de7 1612 clear_user_highpage(vmf->cow_page, vmf->address);
55f81639
SR
1613 break;
1614 case IOMAP_MAPPED:
429f8de7 1615 error = copy_cow_page_dax(vmf, iter);
55f81639
SR
1616 break;
1617 default:
1618 WARN_ON_ONCE(1);
1619 error = -EIO;
1620 break;
1621 }
1622
1623 if (error)
1624 return dax_fault_return(error);
1625
1626 __SetPageUptodate(vmf->cow_page);
1627 ret = finish_fault(vmf);
1628 if (!ret)
1629 return VM_FAULT_DONE_COW;
1630 return ret;
1631}
1632
c2436190 1633/**
65dd814a 1634 * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault.
c2436190 1635 * @vmf: vm fault instance
65dd814a 1636 * @iter: iomap iter
c2436190
SR
1637 * @pfnp: pfn to be returned
1638 * @xas: the dax mapping tree of a file
1639 * @entry: an unlocked dax entry to be inserted
1640 * @pmd: distinguish whether it is a pmd fault
c2436190 1641 */
65dd814a
CH
1642static vm_fault_t dax_fault_iter(struct vm_fault *vmf,
1643 const struct iomap_iter *iter, pfn_t *pfnp,
1644 struct xa_state *xas, void **entry, bool pmd)
c2436190 1645{
65dd814a 1646 const struct iomap *iomap = &iter->iomap;
708dfad2 1647 const struct iomap *srcmap = iomap_iter_srcmap(iter);
c2436190
SR
1648 size_t size = pmd ? PMD_SIZE : PAGE_SIZE;
1649 loff_t pos = (loff_t)xas->xa_index << PAGE_SHIFT;
e5d6df73 1650 bool write = iter->flags & IOMAP_WRITE;
c2436190
SR
1651 unsigned long entry_flags = pmd ? DAX_PMD : 0;
1652 int err = 0;
1653 pfn_t pfn;
ff17b8df 1654 void *kaddr;
c2436190 1655
65dd814a
CH
1656 if (!pmd && vmf->cow_page)
1657 return dax_fault_cow_page(vmf, iter);
1658
c2436190
SR
1659 /* if we are reading UNWRITTEN and HOLE, return a hole. */
1660 if (!write &&
1661 (iomap->type == IOMAP_UNWRITTEN || iomap->type == IOMAP_HOLE)) {
1662 if (!pmd)
e5d6df73
SR
1663 return dax_load_hole(xas, vmf, iter, entry);
1664 return dax_pmd_load_hole(xas, vmf, iter, entry);
c2436190
SR
1665 }
1666
ff17b8df 1667 if (iomap->type != IOMAP_MAPPED && !(iomap->flags & IOMAP_F_SHARED)) {
c2436190
SR
1668 WARN_ON_ONCE(1);
1669 return pmd ? VM_FAULT_FALLBACK : VM_FAULT_SIGBUS;
1670 }
1671
ff17b8df 1672 err = dax_iomap_direct_access(iomap, pos, size, &kaddr, &pfn);
c2436190
SR
1673 if (err)
1674 return pmd ? VM_FAULT_FALLBACK : dax_fault_return(err);
1675
e5d6df73 1676 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, entry_flags);
c2436190 1677
708dfad2
SR
1678 if (write && iomap->flags & IOMAP_F_SHARED) {
1679 err = dax_iomap_copy_around(pos, size, size, srcmap, kaddr);
ff17b8df
SR
1680 if (err)
1681 return dax_fault_return(err);
1682 }
c2436190 1683
e5d6df73 1684 if (dax_fault_is_synchronous(iter, vmf->vma))
c2436190
SR
1685 return dax_fault_synchronous_pfnp(pfnp, pfn);
1686
1687 /* insert PMD pfn */
1688 if (pmd)
1689 return vmf_insert_pfn_pmd(vmf, pfn, write);
1690
1691 /* insert PTE pfn */
1692 if (write)
1693 return vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1694 return vmf_insert_mixed(vmf->vma, vmf->address, pfn);
1695}
1696
ab77dab4 1697static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
c0b24625 1698 int *iomap_errp, const struct iomap_ops *ops)
a7d73fe6 1699{
65dd814a 1700 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
b15cd800 1701 XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
65dd814a
CH
1702 struct iomap_iter iter = {
1703 .inode = mapping->host,
1704 .pos = (loff_t)vmf->pgoff << PAGE_SHIFT,
1705 .len = PAGE_SIZE,
952da063 1706 .flags = IOMAP_DAX | IOMAP_FAULT,
65dd814a 1707 };
ab77dab4 1708 vm_fault_t ret = 0;
a7d73fe6 1709 void *entry;
65dd814a 1710 int error;
a7d73fe6 1711
65dd814a 1712 trace_dax_pte_fault(iter.inode, vmf, ret);
a7d73fe6
CH
1713 /*
1714 * Check whether offset isn't beyond end of file now. Caller is supposed
1715 * to hold locks serializing us with truncate / punch hole so this is
1716 * a reliable test.
1717 */
65dd814a 1718 if (iter.pos >= i_size_read(iter.inode)) {
ab77dab4 1719 ret = VM_FAULT_SIGBUS;
a9c42b33
RZ
1720 goto out;
1721 }
a7d73fe6 1722
65dd814a
CH
1723 if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1724 iter.flags |= IOMAP_WRITE;
a7d73fe6 1725
b15cd800
MW
1726 entry = grab_mapping_entry(&xas, mapping, 0);
1727 if (xa_is_internal(entry)) {
1728 ret = xa_to_internal(entry);
13e451fd
JK
1729 goto out;
1730 }
1731
e2093926
RZ
1732 /*
1733 * It is possible, particularly with mixed reads & writes to private
1734 * mappings, that we have raced with a PMD fault that overlaps with
1735 * the PTE we need to set up. If so just return and the fault will be
1736 * retried.
1737 */
1738 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
ab77dab4 1739 ret = VM_FAULT_NOPAGE;
e2093926
RZ
1740 goto unlock_entry;
1741 }
1742
65dd814a
CH
1743 while ((error = iomap_iter(&iter, ops)) > 0) {
1744 if (WARN_ON_ONCE(iomap_length(&iter) < PAGE_SIZE)) {
1745 iter.processed = -EIO; /* fs corruption? */
1746 continue;
a7d73fe6
CH
1747 }
1748
65dd814a
CH
1749 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, false);
1750 if (ret != VM_FAULT_SIGBUS &&
1751 (iter.iomap.flags & IOMAP_F_NEW)) {
a7d73fe6 1752 count_vm_event(PGMAJFAULT);
65dd814a
CH
1753 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
1754 ret |= VM_FAULT_MAJOR;
a7d73fe6 1755 }
1b5a1cb2 1756
65dd814a
CH
1757 if (!(ret & VM_FAULT_ERROR))
1758 iter.processed = PAGE_SIZE;
a7d73fe6
CH
1759 }
1760
65dd814a
CH
1761 if (iomap_errp)
1762 *iomap_errp = error;
1763 if (!ret && error)
1764 ret = dax_fault_return(error);
9f141d6e 1765
c2436190 1766unlock_entry:
b15cd800 1767 dax_unlock_entry(&xas, entry);
c2436190 1768out:
65dd814a
CH
1769 trace_dax_pte_fault_done(iter.inode, vmf, ret);
1770 return ret;
a7d73fe6 1771}
642261ac
RZ
1772
1773#ifdef CONFIG_FS_DAX_PMD
55f81639
SR
1774static bool dax_fault_check_fallback(struct vm_fault *vmf, struct xa_state *xas,
1775 pgoff_t max_pgoff)
642261ac 1776{
f4200391 1777 unsigned long pmd_addr = vmf->address & PMD_MASK;
55f81639 1778 bool write = vmf->flags & FAULT_FLAG_WRITE;
642261ac 1779
55f81639
SR
1780 /*
1781 * Make sure that the faulting address's PMD offset (color) matches
1782 * the PMD offset from the start of the file. This is necessary so
1783 * that a PMD range in the page table overlaps exactly with a PMD
1784 * range in the page cache.
1785 */
1786 if ((vmf->pgoff & PG_PMD_COLOUR) !=
1787 ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1788 return true;
642261ac 1789
55f81639
SR
1790 /* Fall back to PTEs if we're going to COW */
1791 if (write && !(vmf->vma->vm_flags & VM_SHARED))
1792 return true;
11cf9d86 1793
55f81639
SR
1794 /* If the PMD would extend outside the VMA */
1795 if (pmd_addr < vmf->vma->vm_start)
1796 return true;
1797 if ((pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
1798 return true;
642261ac 1799
55f81639
SR
1800 /* If the PMD would extend beyond the file size */
1801 if ((xas->xa_index | PG_PMD_COLOUR) >= max_pgoff)
1802 return true;
653b2ea3 1803
55f81639 1804 return false;
642261ac
RZ
1805}
1806
ab77dab4 1807static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
a2d58167 1808 const struct iomap_ops *ops)
642261ac 1809{
65dd814a 1810 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
b15cd800 1811 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
65dd814a
CH
1812 struct iomap_iter iter = {
1813 .inode = mapping->host,
1814 .len = PMD_SIZE,
952da063 1815 .flags = IOMAP_DAX | IOMAP_FAULT,
65dd814a 1816 };
c2436190 1817 vm_fault_t ret = VM_FAULT_FALLBACK;
b15cd800 1818 pgoff_t max_pgoff;
642261ac 1819 void *entry;
642261ac 1820
65dd814a
CH
1821 if (vmf->flags & FAULT_FLAG_WRITE)
1822 iter.flags |= IOMAP_WRITE;
642261ac 1823
282a8e03
RZ
1824 /*
1825 * Check whether offset isn't beyond end of file now. Caller is
1826 * supposed to hold locks serializing us with truncate / punch hole so
1827 * this is a reliable test.
1828 */
65dd814a 1829 max_pgoff = DIV_ROUND_UP(i_size_read(iter.inode), PAGE_SIZE);
fffa281b 1830
65dd814a 1831 trace_dax_pmd_fault(iter.inode, vmf, max_pgoff, 0);
642261ac 1832
b15cd800 1833 if (xas.xa_index >= max_pgoff) {
c2436190 1834 ret = VM_FAULT_SIGBUS;
282a8e03
RZ
1835 goto out;
1836 }
642261ac 1837
55f81639 1838 if (dax_fault_check_fallback(vmf, &xas, max_pgoff))
642261ac
RZ
1839 goto fallback;
1840
876f2946 1841 /*
b15cd800
MW
1842 * grab_mapping_entry() will make sure we get an empty PMD entry,
1843 * a zero PMD entry or a DAX PMD. If it can't (because a PTE
1844 * entry is already in the array, for instance), it will return
1845 * VM_FAULT_FALLBACK.
876f2946 1846 */
23c84eb7 1847 entry = grab_mapping_entry(&xas, mapping, PMD_ORDER);
b15cd800 1848 if (xa_is_internal(entry)) {
c2436190 1849 ret = xa_to_internal(entry);
876f2946 1850 goto fallback;
b15cd800 1851 }
876f2946 1852
e2093926
RZ
1853 /*
1854 * It is possible, particularly with mixed reads & writes to private
1855 * mappings, that we have raced with a PTE fault that overlaps with
1856 * the PMD we need to set up. If so just return and the fault will be
1857 * retried.
1858 */
1859 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1860 !pmd_devmap(*vmf->pmd)) {
c2436190 1861 ret = 0;
e2093926
RZ
1862 goto unlock_entry;
1863 }
1864
65dd814a 1865 iter.pos = (loff_t)xas.xa_index << PAGE_SHIFT;
dd0c6425 1866 while (iomap_iter(&iter, ops) > 0) {
65dd814a
CH
1867 if (iomap_length(&iter) < PMD_SIZE)
1868 continue; /* actually breaks out of the loop */
caa51d26 1869
65dd814a
CH
1870 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, true);
1871 if (ret != VM_FAULT_FALLBACK)
1872 iter.processed = PMD_SIZE;
642261ac
RZ
1873 }
1874
c2436190 1875unlock_entry:
b15cd800 1876 dax_unlock_entry(&xas, entry);
c2436190
SR
1877fallback:
1878 if (ret == VM_FAULT_FALLBACK) {
65dd814a 1879 split_huge_pmd(vmf->vma, vmf->pmd, vmf->address);
642261ac
RZ
1880 count_vm_event(THP_FAULT_FALLBACK);
1881 }
282a8e03 1882out:
65dd814a 1883 trace_dax_pmd_fault_done(iter.inode, vmf, max_pgoff, ret);
c2436190 1884 return ret;
642261ac 1885}
a2d58167 1886#else
ab77dab4 1887static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
01cddfe9 1888 const struct iomap_ops *ops)
a2d58167
DJ
1889{
1890 return VM_FAULT_FALLBACK;
1891}
642261ac 1892#endif /* CONFIG_FS_DAX_PMD */
a2d58167
DJ
1893
1894/**
1895 * dax_iomap_fault - handle a page fault on a DAX file
1896 * @vmf: The description of the fault
1d024e7a 1897 * @order: Order of the page to fault in
9a0dd422 1898 * @pfnp: PFN to insert for synchronous faults if fsync is required
c0b24625 1899 * @iomap_errp: Storage for detailed error code in case of error
cec04e8c 1900 * @ops: Iomap ops passed from the file system
a2d58167
DJ
1901 *
1902 * When a page fault occurs, filesystems may call this helper in
1903 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1904 * has done all the necessary locking for page fault to proceed
1905 * successfully.
1906 */
1d024e7a 1907vm_fault_t dax_iomap_fault(struct vm_fault *vmf, unsigned int order,
c0b24625 1908 pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
a2d58167 1909{
1d024e7a 1910 if (order == 0)
c0b24625 1911 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1d024e7a 1912 else if (order == PMD_ORDER)
9a0dd422 1913 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1d024e7a 1914 else
a2d58167 1915 return VM_FAULT_FALLBACK;
a2d58167
DJ
1916}
1917EXPORT_SYMBOL_GPL(dax_iomap_fault);
71eab6df 1918
a77d19f4 1919/*
71eab6df
JK
1920 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1921 * @vmf: The description of the fault
71eab6df 1922 * @pfn: PFN to insert
cfc93c6c 1923 * @order: Order of entry to insert.
71eab6df 1924 *
a77d19f4
MW
1925 * This function inserts a writeable PTE or PMD entry into the page tables
1926 * for an mmaped DAX file. It also marks the page cache entry as dirty.
71eab6df 1927 */
cfc93c6c
MW
1928static vm_fault_t
1929dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
71eab6df
JK
1930{
1931 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
cfc93c6c
MW
1932 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1933 void *entry;
ab77dab4 1934 vm_fault_t ret;
71eab6df 1935
cfc93c6c 1936 xas_lock_irq(&xas);
23c84eb7 1937 entry = get_unlocked_entry(&xas, order);
71eab6df 1938 /* Did we race with someone splitting entry or so? */
23c84eb7
MWO
1939 if (!entry || dax_is_conflict(entry) ||
1940 (order == 0 && !dax_is_pte_entry(entry))) {
4c3d043d 1941 put_unlocked_entry(&xas, entry, WAKE_NEXT);
cfc93c6c 1942 xas_unlock_irq(&xas);
71eab6df
JK
1943 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1944 VM_FAULT_NOPAGE);
1945 return VM_FAULT_NOPAGE;
1946 }
cfc93c6c
MW
1947 xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1948 dax_lock_entry(&xas, entry);
1949 xas_unlock_irq(&xas);
1950 if (order == 0)
ab77dab4 1951 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
71eab6df 1952#ifdef CONFIG_FS_DAX_PMD
cfc93c6c 1953 else if (order == PMD_ORDER)
fce86ff5 1954 ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
71eab6df 1955#endif
cfc93c6c 1956 else
ab77dab4 1957 ret = VM_FAULT_FALLBACK;
cfc93c6c 1958 dax_unlock_entry(&xas, entry);
ab77dab4
SJ
1959 trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1960 return ret;
71eab6df
JK
1961}
1962
1963/**
1964 * dax_finish_sync_fault - finish synchronous page fault
1965 * @vmf: The description of the fault
1d024e7a 1966 * @order: Order of entry to be inserted
71eab6df
JK
1967 * @pfn: PFN to insert
1968 *
1969 * This function ensures that the file range touched by the page fault is
1970 * stored persistently on the media and handles inserting of appropriate page
1971 * table entry.
1972 */
1d024e7a
MWO
1973vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf, unsigned int order,
1974 pfn_t pfn)
71eab6df
JK
1975{
1976 int err;
1977 loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
cfc93c6c 1978 size_t len = PAGE_SIZE << order;
71eab6df 1979
71eab6df
JK
1980 err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1981 if (err)
1982 return VM_FAULT_SIGBUS;
cfc93c6c 1983 return dax_insert_pfn_mkwrite(vmf, pfn, order);
71eab6df
JK
1984}
1985EXPORT_SYMBOL_GPL(dax_finish_sync_fault);
6f7db389
SR
1986
1987static loff_t dax_range_compare_iter(struct iomap_iter *it_src,
1988 struct iomap_iter *it_dest, u64 len, bool *same)
1989{
1990 const struct iomap *smap = &it_src->iomap;
1991 const struct iomap *dmap = &it_dest->iomap;
1992 loff_t pos1 = it_src->pos, pos2 = it_dest->pos;
1993 void *saddr, *daddr;
1994 int id, ret;
1995
1996 len = min(len, min(smap->length, dmap->length));
1997
1998 if (smap->type == IOMAP_HOLE && dmap->type == IOMAP_HOLE) {
1999 *same = true;
2000 return len;
2001 }
2002
2003 if (smap->type == IOMAP_HOLE || dmap->type == IOMAP_HOLE) {
2004 *same = false;
2005 return 0;
2006 }
2007
2008 id = dax_read_lock();
2009 ret = dax_iomap_direct_access(smap, pos1, ALIGN(pos1 + len, PAGE_SIZE),
2010 &saddr, NULL);
2011 if (ret < 0)
2012 goto out_unlock;
2013
2014 ret = dax_iomap_direct_access(dmap, pos2, ALIGN(pos2 + len, PAGE_SIZE),
2015 &daddr, NULL);
2016 if (ret < 0)
2017 goto out_unlock;
2018
2019 *same = !memcmp(saddr, daddr, len);
2020 if (!*same)
2021 len = 0;
2022 dax_read_unlock(id);
2023 return len;
2024
2025out_unlock:
2026 dax_read_unlock(id);
2027 return -EIO;
2028}
2029
2030int dax_dedupe_file_range_compare(struct inode *src, loff_t srcoff,
2031 struct inode *dst, loff_t dstoff, loff_t len, bool *same,
2032 const struct iomap_ops *ops)
2033{
2034 struct iomap_iter src_iter = {
2035 .inode = src,
2036 .pos = srcoff,
2037 .len = len,
2038 .flags = IOMAP_DAX,
2039 };
2040 struct iomap_iter dst_iter = {
2041 .inode = dst,
2042 .pos = dstoff,
2043 .len = len,
2044 .flags = IOMAP_DAX,
2045 };
0e79e373 2046 int ret, compared = 0;
6f7db389 2047
0e79e373
SR
2048 while ((ret = iomap_iter(&src_iter, ops)) > 0 &&
2049 (ret = iomap_iter(&dst_iter, ops)) > 0) {
e900ba10
SR
2050 compared = dax_range_compare_iter(&src_iter, &dst_iter,
2051 min(src_iter.len, dst_iter.len), same);
0e79e373
SR
2052 if (compared < 0)
2053 return ret;
2054 src_iter.processed = dst_iter.processed = compared;
6f7db389
SR
2055 }
2056 return ret;
2057}
2058
2059int dax_remap_file_range_prep(struct file *file_in, loff_t pos_in,
2060 struct file *file_out, loff_t pos_out,
2061 loff_t *len, unsigned int remap_flags,
2062 const struct iomap_ops *ops)
2063{
2064 return __generic_remap_file_range_prep(file_in, pos_in, file_out,
2065 pos_out, len, remap_flags, ops);
2066}
2067EXPORT_SYMBOL_GPL(dax_remap_file_range_prep);