blob: ab925dc6647ac764c5a2620a9ef21c53305f90dc [file] [log] [blame]
Matthew Wilcoxd475c632015-02-16 15:58:56 -08001/*
2 * fs/dax.c - Direct Access filesystem code
3 * Copyright (c) 2013-2014 Intel Corporation
4 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
5 * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
6 *
7 * This program is free software; you can redistribute it and/or modify it
8 * under the terms and conditions of the GNU General Public License,
9 * version 2, as published by the Free Software Foundation.
10 *
11 * This program is distributed in the hope it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 * more details.
15 */
16
17#include <linux/atomic.h>
18#include <linux/blkdev.h>
19#include <linux/buffer_head.h>
Ross Zwislerd77e92e2015-09-09 10:29:40 -060020#include <linux/dax.h>
Matthew Wilcoxd475c632015-02-16 15:58:56 -080021#include <linux/fs.h>
22#include <linux/genhd.h>
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -080023#include <linux/highmem.h>
24#include <linux/memcontrol.h>
25#include <linux/mm.h>
Matthew Wilcoxd475c632015-02-16 15:58:56 -080026#include <linux/mutex.h>
Ross Zwisler9973c982016-01-22 15:10:47 -080027#include <linux/pagevec.h>
Matthew Wilcox289c6ae2015-02-16 15:58:59 -080028#include <linux/sched.h>
Ingo Molnarf361bf42017-02-03 23:47:37 +010029#include <linux/sched/signal.h>
Matthew Wilcoxd475c632015-02-16 15:58:56 -080030#include <linux/uio.h>
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -080031#include <linux/vmstat.h>
Dan Williams34c0fd52016-01-15 16:56:14 -080032#include <linux/pfn_t.h>
Dan Williams0e749e52016-01-15 16:55:53 -080033#include <linux/sizes.h>
Jan Kara4b4bb462016-12-14 15:07:53 -080034#include <linux/mmu_notifier.h>
Christoph Hellwiga254e562016-09-19 11:24:49 +100035#include <linux/iomap.h>
36#include "internal.h"
Matthew Wilcoxd475c632015-02-16 15:58:56 -080037
Ross Zwisler282a8e02017-02-22 15:39:50 -080038#define CREATE_TRACE_POINTS
39#include <trace/events/fs_dax.h>
40
Jan Karaac401cc2016-05-12 18:29:18 +020041/* We choose 4096 entries - same as per-zone page wait tables */
42#define DAX_WAIT_TABLE_BITS 12
43#define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
44
Ross Zwislerce95ab0f2016-11-08 11:31:44 +110045static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
Jan Karaac401cc2016-05-12 18:29:18 +020046
47static int __init init_dax_wait_table(void)
48{
49 int i;
50
51 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
52 init_waitqueue_head(wait_table + i);
53 return 0;
54}
55fs_initcall(init_dax_wait_table);
56
Ross Zwisler642261a2016-11-08 11:34:45 +110057static int dax_is_pmd_entry(void *entry)
58{
59 return (unsigned long)entry & RADIX_DAX_PMD;
60}
61
62static int dax_is_pte_entry(void *entry)
63{
64 return !((unsigned long)entry & RADIX_DAX_PMD);
65}
66
67static int dax_is_zero_entry(void *entry)
68{
69 return (unsigned long)entry & RADIX_DAX_HZP;
70}
71
72static int dax_is_empty_entry(void *entry)
73{
74 return (unsigned long)entry & RADIX_DAX_EMPTY;
75}
76
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -080077/*
Jan Karaac401cc2016-05-12 18:29:18 +020078 * DAX radix tree locking
79 */
80struct exceptional_entry_key {
81 struct address_space *mapping;
Ross Zwisler63e95b52016-11-08 11:32:20 +110082 pgoff_t entry_start;
Jan Karaac401cc2016-05-12 18:29:18 +020083};
84
85struct wait_exceptional_entry_queue {
Ingo Molnarac6424b2017-06-20 12:06:13 +020086 wait_queue_entry_t wait;
Jan Karaac401cc2016-05-12 18:29:18 +020087 struct exceptional_entry_key key;
88};
89
Ross Zwisler63e95b52016-11-08 11:32:20 +110090static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
91 pgoff_t index, void *entry, struct exceptional_entry_key *key)
92{
93 unsigned long hash;
94
95 /*
96 * If 'entry' is a PMD, align the 'index' that we use for the wait
97 * queue to the start of that PMD. This ensures that all offsets in
98 * the range covered by the PMD map to the same bit lock.
99 */
Ross Zwisler642261a2016-11-08 11:34:45 +1100100 if (dax_is_pmd_entry(entry))
Ross Zwisler63e95b52016-11-08 11:32:20 +1100101 index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1);
102
103 key->mapping = mapping;
104 key->entry_start = index;
105
106 hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
107 return wait_table + hash;
108}
109
Ingo Molnarac6424b2017-06-20 12:06:13 +0200110static int wake_exceptional_entry_func(wait_queue_entry_t *wait, unsigned int mode,
Jan Karaac401cc2016-05-12 18:29:18 +0200111 int sync, void *keyp)
112{
113 struct exceptional_entry_key *key = keyp;
114 struct wait_exceptional_entry_queue *ewait =
115 container_of(wait, struct wait_exceptional_entry_queue, wait);
116
117 if (key->mapping != ewait->key.mapping ||
Ross Zwisler63e95b52016-11-08 11:32:20 +1100118 key->entry_start != ewait->key.entry_start)
Jan Karaac401cc2016-05-12 18:29:18 +0200119 return 0;
120 return autoremove_wake_function(wait, mode, sync, NULL);
121}
122
123/*
124 * Check whether the given slot is locked. The function must be called with
125 * mapping->tree_lock held
126 */
127static inline int slot_locked(struct address_space *mapping, void **slot)
128{
129 unsigned long entry = (unsigned long)
130 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
131 return entry & RADIX_DAX_ENTRY_LOCK;
132}
133
134/*
135 * Mark the given slot is locked. The function must be called with
136 * mapping->tree_lock held
137 */
138static inline void *lock_slot(struct address_space *mapping, void **slot)
139{
140 unsigned long entry = (unsigned long)
141 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
142
143 entry |= RADIX_DAX_ENTRY_LOCK;
Johannes Weiner6d75f362016-12-12 16:43:43 -0800144 radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
Jan Karaac401cc2016-05-12 18:29:18 +0200145 return (void *)entry;
146}
147
148/*
149 * Mark the given slot is unlocked. The function must be called with
150 * mapping->tree_lock held
151 */
152static inline void *unlock_slot(struct address_space *mapping, void **slot)
153{
154 unsigned long entry = (unsigned long)
155 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
156
157 entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
Johannes Weiner6d75f362016-12-12 16:43:43 -0800158 radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
Jan Karaac401cc2016-05-12 18:29:18 +0200159 return (void *)entry;
160}
161
162/*
163 * Lookup entry in radix tree, wait for it to become unlocked if it is
164 * exceptional entry and return it. The caller must call
165 * put_unlocked_mapping_entry() when he decided not to lock the entry or
166 * put_locked_mapping_entry() when he locked the entry and now wants to
167 * unlock it.
168 *
169 * The function must be called with mapping->tree_lock held.
170 */
171static void *get_unlocked_mapping_entry(struct address_space *mapping,
172 pgoff_t index, void ***slotp)
173{
Ross Zwislere3ad61c2016-11-08 11:32:12 +1100174 void *entry, **slot;
Jan Karaac401cc2016-05-12 18:29:18 +0200175 struct wait_exceptional_entry_queue ewait;
Ross Zwisler63e95b52016-11-08 11:32:20 +1100176 wait_queue_head_t *wq;
Jan Karaac401cc2016-05-12 18:29:18 +0200177
178 init_wait(&ewait.wait);
179 ewait.wait.func = wake_exceptional_entry_func;
Jan Karaac401cc2016-05-12 18:29:18 +0200180
181 for (;;) {
Ross Zwislere3ad61c2016-11-08 11:32:12 +1100182 entry = __radix_tree_lookup(&mapping->page_tree, index, NULL,
Jan Karaac401cc2016-05-12 18:29:18 +0200183 &slot);
Ross Zwislere3ad61c2016-11-08 11:32:12 +1100184 if (!entry || !radix_tree_exceptional_entry(entry) ||
Jan Karaac401cc2016-05-12 18:29:18 +0200185 !slot_locked(mapping, slot)) {
186 if (slotp)
187 *slotp = slot;
Ross Zwislere3ad61c2016-11-08 11:32:12 +1100188 return entry;
Jan Karaac401cc2016-05-12 18:29:18 +0200189 }
Ross Zwisler63e95b52016-11-08 11:32:20 +1100190
191 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
Jan Karaac401cc2016-05-12 18:29:18 +0200192 prepare_to_wait_exclusive(wq, &ewait.wait,
193 TASK_UNINTERRUPTIBLE);
194 spin_unlock_irq(&mapping->tree_lock);
195 schedule();
196 finish_wait(wq, &ewait.wait);
197 spin_lock_irq(&mapping->tree_lock);
198 }
199}
200
Jan Karab1aa8122016-12-14 15:07:24 -0800201static void dax_unlock_mapping_entry(struct address_space *mapping,
202 pgoff_t index)
203{
204 void *entry, **slot;
205
206 spin_lock_irq(&mapping->tree_lock);
207 entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
208 if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
209 !slot_locked(mapping, slot))) {
210 spin_unlock_irq(&mapping->tree_lock);
211 return;
212 }
213 unlock_slot(mapping, slot);
214 spin_unlock_irq(&mapping->tree_lock);
215 dax_wake_mapping_entry_waiter(mapping, index, entry, false);
216}
217
Jan Karaac401cc2016-05-12 18:29:18 +0200218static void put_locked_mapping_entry(struct address_space *mapping,
219 pgoff_t index, void *entry)
220{
221 if (!radix_tree_exceptional_entry(entry)) {
222 unlock_page(entry);
223 put_page(entry);
224 } else {
Jan Karabc2466e2016-05-12 18:29:19 +0200225 dax_unlock_mapping_entry(mapping, index);
Jan Karaac401cc2016-05-12 18:29:18 +0200226 }
227}
228
229/*
230 * Called when we are done with radix tree entry we looked up via
231 * get_unlocked_mapping_entry() and which we didn't lock in the end.
232 */
233static void put_unlocked_mapping_entry(struct address_space *mapping,
234 pgoff_t index, void *entry)
235{
236 if (!radix_tree_exceptional_entry(entry))
237 return;
238
239 /* We have to wake up next waiter for the radix tree entry lock */
Ross Zwisler422476c2016-11-08 11:33:44 +1100240 dax_wake_mapping_entry_waiter(mapping, index, entry, false);
241}
242
Jan Karaac401cc2016-05-12 18:29:18 +0200243/*
244 * Find radix tree entry at given index. If it points to a page, return with
245 * the page locked. If it points to the exceptional entry, return with the
246 * radix tree entry locked. If the radix tree doesn't contain given index,
247 * create empty exceptional entry for the index and return with it locked.
248 *
Ross Zwisler642261a2016-11-08 11:34:45 +1100249 * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
250 * either return that locked entry or will return an error. This error will
251 * happen if there are any 4k entries (either zero pages or DAX entries)
252 * within the 2MiB range that we are requesting.
253 *
254 * We always favor 4k entries over 2MiB entries. There isn't a flow where we
255 * evict 4k entries in order to 'upgrade' them to a 2MiB entry. A 2MiB
256 * insertion will fail if it finds any 4k entries already in the tree, and a
257 * 4k insertion will cause an existing 2MiB entry to be unmapped and
258 * downgraded to 4k entries. This happens for both 2MiB huge zero pages as
259 * well as 2MiB empty entries.
260 *
261 * The exception to this downgrade path is for 2MiB DAX PMD entries that have
262 * real storage backing them. We will leave these real 2MiB DAX entries in
263 * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
264 *
Jan Karaac401cc2016-05-12 18:29:18 +0200265 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
266 * persistent memory the benefit is doubtful. We can add that later if we can
267 * show it helps.
268 */
Ross Zwisler642261a2016-11-08 11:34:45 +1100269static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
270 unsigned long size_flag)
Jan Karaac401cc2016-05-12 18:29:18 +0200271{
Ross Zwisler642261a2016-11-08 11:34:45 +1100272 bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
Ross Zwislere3ad61c2016-11-08 11:32:12 +1100273 void *entry, **slot;
Jan Karaac401cc2016-05-12 18:29:18 +0200274
275restart:
276 spin_lock_irq(&mapping->tree_lock);
Ross Zwislere3ad61c2016-11-08 11:32:12 +1100277 entry = get_unlocked_mapping_entry(mapping, index, &slot);
Ross Zwisler642261a2016-11-08 11:34:45 +1100278
279 if (entry) {
280 if (size_flag & RADIX_DAX_PMD) {
281 if (!radix_tree_exceptional_entry(entry) ||
282 dax_is_pte_entry(entry)) {
283 put_unlocked_mapping_entry(mapping, index,
284 entry);
285 entry = ERR_PTR(-EEXIST);
286 goto out_unlock;
287 }
288 } else { /* trying to grab a PTE entry */
289 if (radix_tree_exceptional_entry(entry) &&
290 dax_is_pmd_entry(entry) &&
291 (dax_is_zero_entry(entry) ||
292 dax_is_empty_entry(entry))) {
293 pmd_downgrade = true;
294 }
295 }
296 }
297
Jan Karaac401cc2016-05-12 18:29:18 +0200298 /* No entry for given index? Make sure radix tree is big enough. */
Ross Zwisler642261a2016-11-08 11:34:45 +1100299 if (!entry || pmd_downgrade) {
Jan Karaac401cc2016-05-12 18:29:18 +0200300 int err;
301
Ross Zwisler642261a2016-11-08 11:34:45 +1100302 if (pmd_downgrade) {
303 /*
304 * Make sure 'entry' remains valid while we drop
305 * mapping->tree_lock.
306 */
307 entry = lock_slot(mapping, slot);
308 }
309
Jan Karaac401cc2016-05-12 18:29:18 +0200310 spin_unlock_irq(&mapping->tree_lock);
Ross Zwisler642261a2016-11-08 11:34:45 +1100311 /*
312 * Besides huge zero pages the only other thing that gets
313 * downgraded are empty entries which don't need to be
314 * unmapped.
315 */
316 if (pmd_downgrade && dax_is_zero_entry(entry))
317 unmap_mapping_range(mapping,
318 (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
319
Jan Kara0cb80b42016-12-12 21:34:12 -0500320 err = radix_tree_preload(
321 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
322 if (err) {
323 if (pmd_downgrade)
324 put_locked_mapping_entry(mapping, index, entry);
325 return ERR_PTR(err);
326 }
Jan Karaac401cc2016-05-12 18:29:18 +0200327 spin_lock_irq(&mapping->tree_lock);
Ross Zwisler642261a2016-11-08 11:34:45 +1100328
Ross Zwislere11f8b72017-04-07 16:04:57 -0700329 if (!entry) {
330 /*
331 * We needed to drop the page_tree lock while calling
332 * radix_tree_preload() and we didn't have an entry to
333 * lock. See if another thread inserted an entry at
334 * our index during this time.
335 */
336 entry = __radix_tree_lookup(&mapping->page_tree, index,
337 NULL, &slot);
338 if (entry) {
339 radix_tree_preload_end();
340 spin_unlock_irq(&mapping->tree_lock);
341 goto restart;
342 }
343 }
344
Ross Zwisler642261a2016-11-08 11:34:45 +1100345 if (pmd_downgrade) {
346 radix_tree_delete(&mapping->page_tree, index);
347 mapping->nrexceptional--;
348 dax_wake_mapping_entry_waiter(mapping, index, entry,
349 true);
350 }
351
352 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
353
354 err = __radix_tree_insert(&mapping->page_tree, index,
355 dax_radix_order(entry), entry);
Jan Karaac401cc2016-05-12 18:29:18 +0200356 radix_tree_preload_end();
357 if (err) {
358 spin_unlock_irq(&mapping->tree_lock);
Ross Zwisler642261a2016-11-08 11:34:45 +1100359 /*
Ross Zwislere11f8b72017-04-07 16:04:57 -0700360 * Our insertion of a DAX entry failed, most likely
361 * because we were inserting a PMD entry and it
362 * collided with a PTE sized entry at a different
363 * index in the PMD range. We haven't inserted
364 * anything into the radix tree and have no waiters to
365 * wake.
Ross Zwisler642261a2016-11-08 11:34:45 +1100366 */
Jan Karaac401cc2016-05-12 18:29:18 +0200367 return ERR_PTR(err);
368 }
369 /* Good, we have inserted empty locked entry into the tree. */
370 mapping->nrexceptional++;
371 spin_unlock_irq(&mapping->tree_lock);
Ross Zwislere3ad61c2016-11-08 11:32:12 +1100372 return entry;
Jan Karaac401cc2016-05-12 18:29:18 +0200373 }
374 /* Normal page in radix tree? */
Ross Zwislere3ad61c2016-11-08 11:32:12 +1100375 if (!radix_tree_exceptional_entry(entry)) {
376 struct page *page = entry;
Jan Karaac401cc2016-05-12 18:29:18 +0200377
378 get_page(page);
379 spin_unlock_irq(&mapping->tree_lock);
380 lock_page(page);
381 /* Page got truncated? Retry... */
382 if (unlikely(page->mapping != mapping)) {
383 unlock_page(page);
384 put_page(page);
385 goto restart;
386 }
387 return page;
388 }
Ross Zwislere3ad61c2016-11-08 11:32:12 +1100389 entry = lock_slot(mapping, slot);
Ross Zwisler642261a2016-11-08 11:34:45 +1100390 out_unlock:
Jan Karaac401cc2016-05-12 18:29:18 +0200391 spin_unlock_irq(&mapping->tree_lock);
Ross Zwislere3ad61c2016-11-08 11:32:12 +1100392 return entry;
Jan Karaac401cc2016-05-12 18:29:18 +0200393}
394
Ross Zwisler63e95b52016-11-08 11:32:20 +1100395/*
396 * We do not necessarily hold the mapping->tree_lock when we call this
397 * function so it is possible that 'entry' is no longer a valid item in the
Ross Zwisler642261a2016-11-08 11:34:45 +1100398 * radix tree. This is okay because all we really need to do is to find the
399 * correct waitqueue where tasks might be waiting for that old 'entry' and
400 * wake them.
Ross Zwisler63e95b52016-11-08 11:32:20 +1100401 */
Jan Karaac401cc2016-05-12 18:29:18 +0200402void dax_wake_mapping_entry_waiter(struct address_space *mapping,
Ross Zwisler63e95b52016-11-08 11:32:20 +1100403 pgoff_t index, void *entry, bool wake_all)
Jan Karaac401cc2016-05-12 18:29:18 +0200404{
Ross Zwisler63e95b52016-11-08 11:32:20 +1100405 struct exceptional_entry_key key;
406 wait_queue_head_t *wq;
407
408 wq = dax_entry_waitqueue(mapping, index, entry, &key);
Jan Karaac401cc2016-05-12 18:29:18 +0200409
410 /*
411 * Checking for locked entry and prepare_to_wait_exclusive() happens
412 * under mapping->tree_lock, ditto for entry handling in our callers.
413 * So at this point all tasks that could have seen our entry locked
414 * must be in the waitqueue and the following check will see them.
415 */
Ross Zwisler63e95b52016-11-08 11:32:20 +1100416 if (waitqueue_active(wq))
Jan Karaac401cc2016-05-12 18:29:18 +0200417 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
Jan Karaac401cc2016-05-12 18:29:18 +0200418}
419
Jan Karac6dcf522016-08-10 17:22:44 +0200420static int __dax_invalidate_mapping_entry(struct address_space *mapping,
421 pgoff_t index, bool trunc)
422{
423 int ret = 0;
424 void *entry;
425 struct radix_tree_root *page_tree = &mapping->page_tree;
426
427 spin_lock_irq(&mapping->tree_lock);
428 entry = get_unlocked_mapping_entry(mapping, index, NULL);
429 if (!entry || !radix_tree_exceptional_entry(entry))
430 goto out;
431 if (!trunc &&
432 (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
433 radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)))
434 goto out;
435 radix_tree_delete(page_tree, index);
436 mapping->nrexceptional--;
437 ret = 1;
438out:
439 put_unlocked_mapping_entry(mapping, index, entry);
440 spin_unlock_irq(&mapping->tree_lock);
441 return ret;
442}
Jan Karaac401cc2016-05-12 18:29:18 +0200443/*
444 * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
445 * entry to get unlocked before deleting it.
446 */
447int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
448{
Jan Karac6dcf522016-08-10 17:22:44 +0200449 int ret = __dax_invalidate_mapping_entry(mapping, index, true);
Jan Karaac401cc2016-05-12 18:29:18 +0200450
Jan Karaac401cc2016-05-12 18:29:18 +0200451 /*
452 * This gets called from truncate / punch_hole path. As such, the caller
453 * must hold locks protecting against concurrent modifications of the
454 * radix tree (usually fs-private i_mmap_sem for writing). Since the
455 * caller has seen exceptional entry for this index, we better find it
456 * at that index as well...
457 */
Jan Karac6dcf522016-08-10 17:22:44 +0200458 WARN_ON_ONCE(!ret);
459 return ret;
460}
Jan Karaac401cc2016-05-12 18:29:18 +0200461
Jan Karac6dcf522016-08-10 17:22:44 +0200462/*
Jan Karac6dcf522016-08-10 17:22:44 +0200463 * Invalidate exceptional DAX entry if it is clean.
464 */
465int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
466 pgoff_t index)
467{
468 return __dax_invalidate_mapping_entry(mapping, index, false);
Jan Karaac401cc2016-05-12 18:29:18 +0200469}
470
471/*
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800472 * The user has performed a load from a hole in the file. Allocating
473 * a new page in the file would cause excessive storage usage for
474 * workloads with sparse files. We allocate a page cache page instead.
475 * We'll kick it out of the page cache if it's ever written to,
476 * otherwise it will simply fall out of the page cache under memory
477 * pressure without ever having been dirtied.
478 */
Jan Karaf449b932016-10-19 14:48:38 +0200479static int dax_load_hole(struct address_space *mapping, void **entry,
Jan Karaac401cc2016-05-12 18:29:18 +0200480 struct vm_fault *vmf)
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800481{
Ross Zwisler678c9fd2017-05-08 16:00:07 -0700482 struct inode *inode = mapping->host;
Jan Karaac401cc2016-05-12 18:29:18 +0200483 struct page *page;
Jan Karaf449b932016-10-19 14:48:38 +0200484 int ret;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800485
Jan Karaac401cc2016-05-12 18:29:18 +0200486 /* Hole page already exists? Return it... */
Jan Karaf449b932016-10-19 14:48:38 +0200487 if (!radix_tree_exceptional_entry(*entry)) {
488 page = *entry;
Ross Zwisler678c9fd2017-05-08 16:00:07 -0700489 goto finish_fault;
Jan Karaac401cc2016-05-12 18:29:18 +0200490 }
491
492 /* This will replace locked radix tree entry with a hole page */
493 page = find_or_create_page(mapping, vmf->pgoff,
494 vmf->gfp_mask | __GFP_ZERO);
Ross Zwisler678c9fd2017-05-08 16:00:07 -0700495 if (!page) {
496 ret = VM_FAULT_OOM;
497 goto out;
498 }
499
500finish_fault:
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800501 vmf->page = page;
Jan Karaf449b932016-10-19 14:48:38 +0200502 ret = finish_fault(vmf);
503 vmf->page = NULL;
504 *entry = page;
505 if (!ret) {
506 /* Grab reference for PTE that is now referencing the page */
507 get_page(page);
Ross Zwisler678c9fd2017-05-08 16:00:07 -0700508 ret = VM_FAULT_NOPAGE;
Jan Karaf449b932016-10-19 14:48:38 +0200509 }
Ross Zwisler678c9fd2017-05-08 16:00:07 -0700510out:
511 trace_dax_load_hole(inode, vmf, ret);
Jan Karaf449b932016-10-19 14:48:38 +0200512 return ret;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800513}
514
Dan Williamscccbce62017-01-27 13:31:42 -0800515static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
516 sector_t sector, size_t size, struct page *to,
517 unsigned long vaddr)
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800518{
Dan Williamscccbce62017-01-27 13:31:42 -0800519 void *vto, *kaddr;
520 pgoff_t pgoff;
521 pfn_t pfn;
522 long rc;
523 int id;
Ross Zwislere2e05392015-08-18 13:55:41 -0600524
Dan Williamscccbce62017-01-27 13:31:42 -0800525 rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
526 if (rc)
527 return rc;
528
529 id = dax_read_lock();
530 rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
531 if (rc < 0) {
532 dax_read_unlock(id);
533 return rc;
534 }
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800535 vto = kmap_atomic(to);
Dan Williamscccbce62017-01-27 13:31:42 -0800536 copy_user_page(vto, (void __force *)kaddr, vaddr, to);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800537 kunmap_atomic(vto);
Dan Williamscccbce62017-01-27 13:31:42 -0800538 dax_read_unlock(id);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800539 return 0;
540}
541
Ross Zwisler642261a2016-11-08 11:34:45 +1100542/*
543 * By this point grab_mapping_entry() has ensured that we have a locked entry
544 * of the appropriate size so we don't have to worry about downgrading PMDs to
545 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
546 * already in the tree, we will skip the insertion and just dirty the PMD as
547 * appropriate.
548 */
Jan Karaac401cc2016-05-12 18:29:18 +0200549static void *dax_insert_mapping_entry(struct address_space *mapping,
550 struct vm_fault *vmf,
Ross Zwisler642261a2016-11-08 11:34:45 +1100551 void *entry, sector_t sector,
552 unsigned long flags)
Ross Zwisler9973c982016-01-22 15:10:47 -0800553{
554 struct radix_tree_root *page_tree = &mapping->page_tree;
Jan Karaac401cc2016-05-12 18:29:18 +0200555 int error = 0;
556 bool hole_fill = false;
557 void *new_entry;
558 pgoff_t index = vmf->pgoff;
Ross Zwisler9973c982016-01-22 15:10:47 -0800559
Jan Karaac401cc2016-05-12 18:29:18 +0200560 if (vmf->flags & FAULT_FLAG_WRITE)
Dmitry Monakhovd2b2a282016-02-05 15:36:55 -0800561 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Ross Zwisler9973c982016-01-22 15:10:47 -0800562
Jan Karaac401cc2016-05-12 18:29:18 +0200563 /* Replacing hole page with block mapping? */
564 if (!radix_tree_exceptional_entry(entry)) {
565 hole_fill = true;
566 /*
567 * Unmap the page now before we remove it from page cache below.
568 * The page is locked so it cannot be faulted in again.
569 */
570 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
571 PAGE_SIZE, 0);
572 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
573 if (error)
574 return ERR_PTR(error);
Ross Zwisler642261a2016-11-08 11:34:45 +1100575 } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) {
576 /* replacing huge zero page with PMD block mapping */
577 unmap_mapping_range(mapping,
578 (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
Ross Zwisler9973c982016-01-22 15:10:47 -0800579 }
580
Jan Karaac401cc2016-05-12 18:29:18 +0200581 spin_lock_irq(&mapping->tree_lock);
Ross Zwisler642261a2016-11-08 11:34:45 +1100582 new_entry = dax_radix_locked_entry(sector, flags);
583
Jan Karaac401cc2016-05-12 18:29:18 +0200584 if (hole_fill) {
585 __delete_from_page_cache(entry, NULL);
586 /* Drop pagecache reference */
587 put_page(entry);
Ross Zwisler642261a2016-11-08 11:34:45 +1100588 error = __radix_tree_insert(page_tree, index,
589 dax_radix_order(new_entry), new_entry);
Jan Karaac401cc2016-05-12 18:29:18 +0200590 if (error) {
591 new_entry = ERR_PTR(error);
Ross Zwisler9973c982016-01-22 15:10:47 -0800592 goto unlock;
593 }
Jan Karaac401cc2016-05-12 18:29:18 +0200594 mapping->nrexceptional++;
Ross Zwisler642261a2016-11-08 11:34:45 +1100595 } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
596 /*
597 * Only swap our new entry into the radix tree if the current
598 * entry is a zero page or an empty entry. If a normal PTE or
599 * PMD entry is already in the tree, we leave it alone. This
600 * means that if we are trying to insert a PTE and the
601 * existing entry is a PMD, we will just leave the PMD in the
602 * tree and dirty it if necessary.
603 */
Johannes Weinerf7942432016-12-12 16:43:41 -0800604 struct radix_tree_node *node;
Jan Karaac401cc2016-05-12 18:29:18 +0200605 void **slot;
606 void *ret;
Ross Zwisler9973c982016-01-22 15:10:47 -0800607
Johannes Weinerf7942432016-12-12 16:43:41 -0800608 ret = __radix_tree_lookup(page_tree, index, &node, &slot);
Jan Karaac401cc2016-05-12 18:29:18 +0200609 WARN_ON_ONCE(ret != entry);
Johannes Weiner4d693d02016-12-12 16:43:49 -0800610 __radix_tree_replace(page_tree, node, slot,
611 new_entry, NULL, NULL);
Ross Zwisler9973c982016-01-22 15:10:47 -0800612 }
Jan Karaac401cc2016-05-12 18:29:18 +0200613 if (vmf->flags & FAULT_FLAG_WRITE)
Ross Zwisler9973c982016-01-22 15:10:47 -0800614 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
615 unlock:
616 spin_unlock_irq(&mapping->tree_lock);
Jan Karaac401cc2016-05-12 18:29:18 +0200617 if (hole_fill) {
618 radix_tree_preload_end();
619 /*
620 * We don't need hole page anymore, it has been replaced with
621 * locked radix tree entry now.
622 */
623 if (mapping->a_ops->freepage)
624 mapping->a_ops->freepage(entry);
625 unlock_page(entry);
626 put_page(entry);
627 }
628 return new_entry;
Ross Zwisler9973c982016-01-22 15:10:47 -0800629}
630
Jan Kara4b4bb462016-12-14 15:07:53 -0800631static inline unsigned long
632pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
633{
634 unsigned long address;
635
636 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
637 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
638 return address;
639}
640
641/* Walk all mappings of a given index of a file and writeprotect them */
642static void dax_mapping_entry_mkclean(struct address_space *mapping,
643 pgoff_t index, unsigned long pfn)
644{
645 struct vm_area_struct *vma;
Ross Zwislerf729c8c2017-01-10 16:57:24 -0800646 pte_t pte, *ptep = NULL;
647 pmd_t *pmdp = NULL;
Jan Kara4b4bb462016-12-14 15:07:53 -0800648 spinlock_t *ptl;
Jan Kara4b4bb462016-12-14 15:07:53 -0800649
650 i_mmap_lock_read(mapping);
651 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
Jérôme Glissea4d1a882017-08-31 17:17:26 -0400652 unsigned long address, start, end;
Jan Kara4b4bb462016-12-14 15:07:53 -0800653
654 cond_resched();
655
656 if (!(vma->vm_flags & VM_SHARED))
657 continue;
658
659 address = pgoff_address(index, vma);
Jérôme Glissea4d1a882017-08-31 17:17:26 -0400660
661 /*
662 * Note because we provide start/end to follow_pte_pmd it will
663 * call mmu_notifier_invalidate_range_start() on our behalf
664 * before taking any lock.
665 */
666 if (follow_pte_pmd(vma->vm_mm, address, &start, &end, &ptep, &pmdp, &ptl))
Jan Kara4b4bb462016-12-14 15:07:53 -0800667 continue;
Jan Kara4b4bb462016-12-14 15:07:53 -0800668
Ross Zwislerf729c8c2017-01-10 16:57:24 -0800669 if (pmdp) {
670#ifdef CONFIG_FS_DAX_PMD
671 pmd_t pmd;
672
673 if (pfn != pmd_pfn(*pmdp))
674 goto unlock_pmd;
675 if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
676 goto unlock_pmd;
677
678 flush_cache_page(vma, address, pfn);
679 pmd = pmdp_huge_clear_flush(vma, address, pmdp);
680 pmd = pmd_wrprotect(pmd);
681 pmd = pmd_mkclean(pmd);
682 set_pmd_at(vma->vm_mm, address, pmdp, pmd);
Jérôme Glissea4d1a882017-08-31 17:17:26 -0400683 mmu_notifier_invalidate_range(vma->vm_mm, start, end);
Ross Zwislerf729c8c2017-01-10 16:57:24 -0800684unlock_pmd:
685 spin_unlock(ptl);
686#endif
687 } else {
688 if (pfn != pte_pfn(*ptep))
689 goto unlock_pte;
690 if (!pte_dirty(*ptep) && !pte_write(*ptep))
691 goto unlock_pte;
692
693 flush_cache_page(vma, address, pfn);
694 pte = ptep_clear_flush(vma, address, ptep);
695 pte = pte_wrprotect(pte);
696 pte = pte_mkclean(pte);
697 set_pte_at(vma->vm_mm, address, ptep, pte);
Jérôme Glissea4d1a882017-08-31 17:17:26 -0400698 mmu_notifier_invalidate_range(vma->vm_mm, start, end);
Ross Zwislerf729c8c2017-01-10 16:57:24 -0800699unlock_pte:
700 pte_unmap_unlock(ptep, ptl);
701 }
Jan Kara4b4bb462016-12-14 15:07:53 -0800702
Jérôme Glissea4d1a882017-08-31 17:17:26 -0400703 mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
Jan Kara4b4bb462016-12-14 15:07:53 -0800704 }
705 i_mmap_unlock_read(mapping);
706}
707
Ross Zwisler9973c982016-01-22 15:10:47 -0800708static int dax_writeback_one(struct block_device *bdev,
Dan Williamscccbce62017-01-27 13:31:42 -0800709 struct dax_device *dax_dev, struct address_space *mapping,
710 pgoff_t index, void *entry)
Ross Zwisler9973c982016-01-22 15:10:47 -0800711{
712 struct radix_tree_root *page_tree = &mapping->page_tree;
Dan Williamscccbce62017-01-27 13:31:42 -0800713 void *entry2, **slot, *kaddr;
714 long ret = 0, id;
715 sector_t sector;
716 pgoff_t pgoff;
717 size_t size;
718 pfn_t pfn;
Ross Zwisler9973c982016-01-22 15:10:47 -0800719
Ross Zwisler9973c982016-01-22 15:10:47 -0800720 /*
Jan Karaa6abc2c2016-12-14 15:07:47 -0800721 * A page got tagged dirty in DAX mapping? Something is seriously
722 * wrong.
Ross Zwisler9973c982016-01-22 15:10:47 -0800723 */
Jan Karaa6abc2c2016-12-14 15:07:47 -0800724 if (WARN_ON(!radix_tree_exceptional_entry(entry)))
725 return -EIO;
Ross Zwisler9973c982016-01-22 15:10:47 -0800726
Jan Karaa6abc2c2016-12-14 15:07:47 -0800727 spin_lock_irq(&mapping->tree_lock);
728 entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
729 /* Entry got punched out / reallocated? */
730 if (!entry2 || !radix_tree_exceptional_entry(entry2))
731 goto put_unlocked;
732 /*
733 * Entry got reallocated elsewhere? No need to writeback. We have to
734 * compare sectors as we must not bail out due to difference in lockbit
735 * or entry type.
736 */
737 if (dax_radix_sector(entry2) != dax_radix_sector(entry))
738 goto put_unlocked;
Ross Zwisler642261a2016-11-08 11:34:45 +1100739 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
740 dax_is_zero_entry(entry))) {
Ross Zwisler9973c982016-01-22 15:10:47 -0800741 ret = -EIO;
Jan Karaa6abc2c2016-12-14 15:07:47 -0800742 goto put_unlocked;
Ross Zwisler9973c982016-01-22 15:10:47 -0800743 }
744
Jan Karaa6abc2c2016-12-14 15:07:47 -0800745 /* Another fsync thread may have already written back this entry */
746 if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
747 goto put_unlocked;
748 /* Lock the entry to serialize with page faults */
749 entry = lock_slot(mapping, slot);
750 /*
751 * We can clear the tag now but we have to be careful so that concurrent
752 * dax_writeback_one() calls for the same index cannot finish before we
753 * actually flush the caches. This is achieved as the calls will look
754 * at the entry only under tree_lock and once they do that they will
755 * see the entry locked and wait for it to unlock.
756 */
757 radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
758 spin_unlock_irq(&mapping->tree_lock);
759
Ross Zwisler642261a2016-11-08 11:34:45 +1100760 /*
761 * Even if dax_writeback_mapping_range() was given a wbc->range_start
762 * in the middle of a PMD, the 'index' we are given will be aligned to
763 * the start index of the PMD, as will the sector we pull from
764 * 'entry'. This allows us to flush for PMD_SIZE and not have to
765 * worry about partial PMD writebacks.
766 */
Dan Williamscccbce62017-01-27 13:31:42 -0800767 sector = dax_radix_sector(entry);
768 size = PAGE_SIZE << dax_radix_order(entry);
769
770 id = dax_read_lock();
771 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
772 if (ret)
773 goto dax_unlock;
Ross Zwisler9973c982016-01-22 15:10:47 -0800774
775 /*
Dan Williamscccbce62017-01-27 13:31:42 -0800776 * dax_direct_access() may sleep, so cannot hold tree_lock over
777 * its invocation.
Ross Zwisler9973c982016-01-22 15:10:47 -0800778 */
Dan Williamscccbce62017-01-27 13:31:42 -0800779 ret = dax_direct_access(dax_dev, pgoff, size / PAGE_SIZE, &kaddr, &pfn);
780 if (ret < 0)
781 goto dax_unlock;
Ross Zwisler9973c982016-01-22 15:10:47 -0800782
Dan Williamscccbce62017-01-27 13:31:42 -0800783 if (WARN_ON_ONCE(ret < size / PAGE_SIZE)) {
Ross Zwisler9973c982016-01-22 15:10:47 -0800784 ret = -EIO;
Dan Williamscccbce62017-01-27 13:31:42 -0800785 goto dax_unlock;
Ross Zwisler9973c982016-01-22 15:10:47 -0800786 }
787
Dan Williamscccbce62017-01-27 13:31:42 -0800788 dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(pfn));
Dan Williams63187702017-05-29 13:07:46 -0700789 dax_flush(dax_dev, pgoff, kaddr, size);
Jan Kara4b4bb462016-12-14 15:07:53 -0800790 /*
791 * After we have flushed the cache, we can clear the dirty tag. There
792 * cannot be new dirty data in the pfn after the flush has completed as
793 * the pfn mappings are writeprotected and fault waits for mapping
794 * entry lock.
795 */
796 spin_lock_irq(&mapping->tree_lock);
797 radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY);
798 spin_unlock_irq(&mapping->tree_lock);
Ross Zwislerf9bc3a02017-05-08 16:00:13 -0700799 trace_dax_writeback_one(mapping->host, index, size >> PAGE_SHIFT);
Dan Williamscccbce62017-01-27 13:31:42 -0800800 dax_unlock:
801 dax_read_unlock(id);
Jan Karaa6abc2c2016-12-14 15:07:47 -0800802 put_locked_mapping_entry(mapping, index, entry);
Ross Zwisler9973c982016-01-22 15:10:47 -0800803 return ret;
804
Jan Karaa6abc2c2016-12-14 15:07:47 -0800805 put_unlocked:
806 put_unlocked_mapping_entry(mapping, index, entry2);
Ross Zwisler9973c982016-01-22 15:10:47 -0800807 spin_unlock_irq(&mapping->tree_lock);
808 return ret;
809}
810
811/*
812 * Flush the mapping to the persistent domain within the byte range of [start,
813 * end]. This is required by data integrity operations to ensure file data is
814 * on persistent storage prior to completion of the operation.
815 */
Ross Zwisler7f6d5b52016-02-26 15:19:55 -0800816int dax_writeback_mapping_range(struct address_space *mapping,
817 struct block_device *bdev, struct writeback_control *wbc)
Ross Zwisler9973c982016-01-22 15:10:47 -0800818{
819 struct inode *inode = mapping->host;
Ross Zwisler642261a2016-11-08 11:34:45 +1100820 pgoff_t start_index, end_index;
Ross Zwisler9973c982016-01-22 15:10:47 -0800821 pgoff_t indices[PAGEVEC_SIZE];
Dan Williamscccbce62017-01-27 13:31:42 -0800822 struct dax_device *dax_dev;
Ross Zwisler9973c982016-01-22 15:10:47 -0800823 struct pagevec pvec;
824 bool done = false;
825 int i, ret = 0;
Ross Zwisler9973c982016-01-22 15:10:47 -0800826
827 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
828 return -EIO;
829
Ross Zwisler7f6d5b52016-02-26 15:19:55 -0800830 if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
831 return 0;
832
Dan Williamscccbce62017-01-27 13:31:42 -0800833 dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
834 if (!dax_dev)
835 return -EIO;
836
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300837 start_index = wbc->range_start >> PAGE_SHIFT;
838 end_index = wbc->range_end >> PAGE_SHIFT;
Ross Zwisler9973c982016-01-22 15:10:47 -0800839
Ross Zwislerd14a3f42017-05-08 16:00:10 -0700840 trace_dax_writeback_range(inode, start_index, end_index);
841
Ross Zwisler9973c982016-01-22 15:10:47 -0800842 tag_pages_for_writeback(mapping, start_index, end_index);
843
844 pagevec_init(&pvec, 0);
845 while (!done) {
846 pvec.nr = find_get_entries_tag(mapping, start_index,
847 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
848 pvec.pages, indices);
849
850 if (pvec.nr == 0)
851 break;
852
853 for (i = 0; i < pvec.nr; i++) {
854 if (indices[i] > end_index) {
855 done = true;
856 break;
857 }
858
Dan Williamscccbce62017-01-27 13:31:42 -0800859 ret = dax_writeback_one(bdev, dax_dev, mapping,
860 indices[i], pvec.pages[i]);
Jeff Layton819ec6b2017-07-06 07:02:27 -0400861 if (ret < 0) {
862 mapping_set_error(mapping, ret);
Ross Zwislerd14a3f42017-05-08 16:00:10 -0700863 goto out;
Jeff Layton819ec6b2017-07-06 07:02:27 -0400864 }
Ross Zwisler9973c982016-01-22 15:10:47 -0800865 }
Jan Kara1eb643d2017-06-23 15:08:46 -0700866 start_index = indices[pvec.nr - 1] + 1;
Ross Zwisler9973c982016-01-22 15:10:47 -0800867 }
Ross Zwislerd14a3f42017-05-08 16:00:10 -0700868out:
Dan Williamscccbce62017-01-27 13:31:42 -0800869 put_dax(dax_dev);
Ross Zwislerd14a3f42017-05-08 16:00:10 -0700870 trace_dax_writeback_range_done(inode, start_index, end_index);
871 return (ret < 0 ? ret : 0);
Ross Zwisler9973c982016-01-22 15:10:47 -0800872}
873EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
874
Jan Karaac401cc2016-05-12 18:29:18 +0200875static int dax_insert_mapping(struct address_space *mapping,
Dan Williamscccbce62017-01-27 13:31:42 -0800876 struct block_device *bdev, struct dax_device *dax_dev,
877 sector_t sector, size_t size, void **entryp,
878 struct vm_area_struct *vma, struct vm_fault *vmf)
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800879{
Jan Kara1a29d852016-12-14 15:07:01 -0800880 unsigned long vaddr = vmf->address;
Jan Karaac401cc2016-05-12 18:29:18 +0200881 void *entry = *entryp;
Dan Williamscccbce62017-01-27 13:31:42 -0800882 void *ret, *kaddr;
883 pgoff_t pgoff;
884 int id, rc;
885 pfn_t pfn;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800886
Dan Williamscccbce62017-01-27 13:31:42 -0800887 rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
888 if (rc)
889 return rc;
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800890
Dan Williamscccbce62017-01-27 13:31:42 -0800891 id = dax_read_lock();
892 rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
893 if (rc < 0) {
894 dax_read_unlock(id);
895 return rc;
896 }
897 dax_read_unlock(id);
898
899 ret = dax_insert_mapping_entry(mapping, vmf, entry, sector, 0);
Jan Kara4d9a2c82016-05-12 18:29:20 +0200900 if (IS_ERR(ret))
901 return PTR_ERR(ret);
Jan Karaac401cc2016-05-12 18:29:18 +0200902 *entryp = ret;
Ross Zwisler9973c982016-01-22 15:10:47 -0800903
Ross Zwislerb4440732017-05-08 16:00:16 -0700904 trace_dax_insert_mapping(mapping->host, vmf, ret);
Dan Williamscccbce62017-01-27 13:31:42 -0800905 return vm_insert_mixed(vma, vaddr, pfn);
Matthew Wilcoxf7ca90b2015-02-16 15:59:02 -0800906}
907
Dave Chinnerce5c5d52015-06-04 09:18:18 +1000908/**
Boaz Harrosh0e3b2102015-04-15 16:15:14 -0700909 * dax_pfn_mkwrite - handle first write to DAX page
Boaz Harrosh0e3b2102015-04-15 16:15:14 -0700910 * @vmf: The description of the fault
Boaz Harrosh0e3b2102015-04-15 16:15:14 -0700911 */
Dave Jiang11bac802017-02-24 14:56:41 -0800912int dax_pfn_mkwrite(struct vm_fault *vmf)
Boaz Harrosh0e3b2102015-04-15 16:15:14 -0700913{
Dave Jiang11bac802017-02-24 14:56:41 -0800914 struct file *file = vmf->vma->vm_file;
Jan Karaac401cc2016-05-12 18:29:18 +0200915 struct address_space *mapping = file->f_mapping;
Ross Zwislerc3ff68d2017-05-08 16:00:03 -0700916 struct inode *inode = mapping->host;
Jan Kara2f89dc12016-12-14 15:07:50 -0800917 void *entry, **slot;
Jan Karaac401cc2016-05-12 18:29:18 +0200918 pgoff_t index = vmf->pgoff;
Boaz Harrosh0e3b2102015-04-15 16:15:14 -0700919
Jan Karaac401cc2016-05-12 18:29:18 +0200920 spin_lock_irq(&mapping->tree_lock);
Jan Kara2f89dc12016-12-14 15:07:50 -0800921 entry = get_unlocked_mapping_entry(mapping, index, &slot);
922 if (!entry || !radix_tree_exceptional_entry(entry)) {
923 if (entry)
924 put_unlocked_mapping_entry(mapping, index, entry);
925 spin_unlock_irq(&mapping->tree_lock);
Ross Zwislerc3ff68d2017-05-08 16:00:03 -0700926 trace_dax_pfn_mkwrite_no_entry(inode, vmf, VM_FAULT_NOPAGE);
Jan Kara2f89dc12016-12-14 15:07:50 -0800927 return VM_FAULT_NOPAGE;
928 }
Jan Karaac401cc2016-05-12 18:29:18 +0200929 radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
Jan Kara2f89dc12016-12-14 15:07:50 -0800930 entry = lock_slot(mapping, slot);
Jan Karaac401cc2016-05-12 18:29:18 +0200931 spin_unlock_irq(&mapping->tree_lock);
Jan Kara2f89dc12016-12-14 15:07:50 -0800932 /*
933 * If we race with somebody updating the PTE and finish_mkwrite_fault()
934 * fails, we don't care. We need to return VM_FAULT_NOPAGE and retry
935 * the fault in either case.
936 */
937 finish_mkwrite_fault(vmf);
938 put_locked_mapping_entry(mapping, index, entry);
Ross Zwislerc3ff68d2017-05-08 16:00:03 -0700939 trace_dax_pfn_mkwrite(inode, vmf, VM_FAULT_NOPAGE);
Boaz Harrosh0e3b2102015-04-15 16:15:14 -0700940 return VM_FAULT_NOPAGE;
941}
942EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
943
Vishal Verma4b0228f2016-04-21 15:13:46 -0400944static bool dax_range_is_aligned(struct block_device *bdev,
945 unsigned int offset, unsigned int length)
946{
947 unsigned short sector_size = bdev_logical_block_size(bdev);
948
949 if (!IS_ALIGNED(offset, sector_size))
950 return false;
951 if (!IS_ALIGNED(length, sector_size))
952 return false;
953
954 return true;
955}
956
Dan Williamscccbce62017-01-27 13:31:42 -0800957int __dax_zero_page_range(struct block_device *bdev,
958 struct dax_device *dax_dev, sector_t sector,
959 unsigned int offset, unsigned int size)
Christoph Hellwig679c8bd2016-05-09 10:47:04 +0200960{
Dan Williamscccbce62017-01-27 13:31:42 -0800961 if (dax_range_is_aligned(bdev, offset, size)) {
962 sector_t start_sector = sector + (offset >> 9);
Vishal Verma4b0228f2016-04-21 15:13:46 -0400963
964 return blkdev_issue_zeroout(bdev, start_sector,
Linus Torvalds53ef7d02017-05-05 18:49:20 -0700965 size >> 9, GFP_NOFS, 0);
Vishal Verma4b0228f2016-04-21 15:13:46 -0400966 } else {
Dan Williamscccbce62017-01-27 13:31:42 -0800967 pgoff_t pgoff;
968 long rc, id;
969 void *kaddr;
970 pfn_t pfn;
971
Dan Williamse84b83b2017-05-10 19:38:13 -0700972 rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
Dan Williamscccbce62017-01-27 13:31:42 -0800973 if (rc)
974 return rc;
975
976 id = dax_read_lock();
Dan Williamse84b83b2017-05-10 19:38:13 -0700977 rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr,
Dan Williamscccbce62017-01-27 13:31:42 -0800978 &pfn);
979 if (rc < 0) {
980 dax_read_unlock(id);
981 return rc;
982 }
Dan Williams81f55872017-05-29 13:12:20 -0700983 memset(kaddr + offset, 0, size);
984 dax_flush(dax_dev, pgoff, kaddr + offset, size);
Dan Williamscccbce62017-01-27 13:31:42 -0800985 dax_read_unlock(id);
Vishal Verma4b0228f2016-04-21 15:13:46 -0400986 }
Christoph Hellwig679c8bd2016-05-09 10:47:04 +0200987 return 0;
988}
989EXPORT_SYMBOL_GPL(__dax_zero_page_range);
990
Ross Zwisler333ccc92016-11-08 11:33:09 +1100991static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
992{
993 return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9);
994}
995
Christoph Hellwiga254e562016-09-19 11:24:49 +1000996static loff_t
Ross Zwisler11c59c92016-11-08 11:32:46 +1100997dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
Christoph Hellwiga254e562016-09-19 11:24:49 +1000998 struct iomap *iomap)
999{
Dan Williamscccbce62017-01-27 13:31:42 -08001000 struct block_device *bdev = iomap->bdev;
1001 struct dax_device *dax_dev = iomap->dax_dev;
Christoph Hellwiga254e562016-09-19 11:24:49 +10001002 struct iov_iter *iter = data;
1003 loff_t end = pos + length, done = 0;
1004 ssize_t ret = 0;
Dan Williamscccbce62017-01-27 13:31:42 -08001005 int id;
Christoph Hellwiga254e562016-09-19 11:24:49 +10001006
1007 if (iov_iter_rw(iter) == READ) {
1008 end = min(end, i_size_read(inode));
1009 if (pos >= end)
1010 return 0;
1011
1012 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1013 return iov_iter_zero(min(length, end - pos), iter);
1014 }
1015
1016 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1017 return -EIO;
1018
Jan Karae3fce682016-08-10 17:10:28 +02001019 /*
1020 * Write can allocate block for an area which has a hole page mapped
1021 * into page tables. We have to tear down these mappings so that data
1022 * written by write(2) is visible in mmap.
1023 */
Jan Karacd656372017-05-12 15:46:50 -07001024 if (iomap->flags & IOMAP_F_NEW) {
Jan Karae3fce682016-08-10 17:10:28 +02001025 invalidate_inode_pages2_range(inode->i_mapping,
1026 pos >> PAGE_SHIFT,
1027 (end - 1) >> PAGE_SHIFT);
1028 }
1029
Dan Williamscccbce62017-01-27 13:31:42 -08001030 id = dax_read_lock();
Christoph Hellwiga254e562016-09-19 11:24:49 +10001031 while (pos < end) {
1032 unsigned offset = pos & (PAGE_SIZE - 1);
Dan Williamscccbce62017-01-27 13:31:42 -08001033 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1034 const sector_t sector = dax_iomap_sector(iomap, pos);
Christoph Hellwiga254e562016-09-19 11:24:49 +10001035 ssize_t map_len;
Dan Williamscccbce62017-01-27 13:31:42 -08001036 pgoff_t pgoff;
1037 void *kaddr;
1038 pfn_t pfn;
Christoph Hellwiga254e562016-09-19 11:24:49 +10001039
Michal Hockod1908f52017-02-03 13:13:26 -08001040 if (fatal_signal_pending(current)) {
1041 ret = -EINTR;
1042 break;
1043 }
1044
Dan Williamscccbce62017-01-27 13:31:42 -08001045 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1046 if (ret)
1047 break;
1048
1049 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1050 &kaddr, &pfn);
Christoph Hellwiga254e562016-09-19 11:24:49 +10001051 if (map_len < 0) {
1052 ret = map_len;
1053 break;
1054 }
1055
Dan Williamscccbce62017-01-27 13:31:42 -08001056 map_len = PFN_PHYS(map_len);
1057 kaddr += offset;
Christoph Hellwiga254e562016-09-19 11:24:49 +10001058 map_len -= offset;
1059 if (map_len > end - pos)
1060 map_len = end - pos;
1061
1062 if (iov_iter_rw(iter) == WRITE)
Dan Williamsfec53772017-05-29 21:56:49 -07001063 map_len = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1064 map_len, iter);
Christoph Hellwiga254e562016-09-19 11:24:49 +10001065 else
Dan Williamscccbce62017-01-27 13:31:42 -08001066 map_len = copy_to_iter(kaddr, map_len, iter);
Christoph Hellwiga254e562016-09-19 11:24:49 +10001067 if (map_len <= 0) {
1068 ret = map_len ? map_len : -EFAULT;
1069 break;
1070 }
1071
1072 pos += map_len;
1073 length -= map_len;
1074 done += map_len;
1075 }
Dan Williamscccbce62017-01-27 13:31:42 -08001076 dax_read_unlock(id);
Christoph Hellwiga254e562016-09-19 11:24:49 +10001077
1078 return done ? done : ret;
1079}
1080
1081/**
Ross Zwisler11c59c92016-11-08 11:32:46 +11001082 * dax_iomap_rw - Perform I/O to a DAX file
Christoph Hellwiga254e562016-09-19 11:24:49 +10001083 * @iocb: The control block for this I/O
1084 * @iter: The addresses to do I/O from or to
1085 * @ops: iomap ops passed from the file system
1086 *
1087 * This function performs read and write operations to directly mapped
1088 * persistent memory. The callers needs to take care of read/write exclusion
1089 * and evicting any page cache pages in the region under I/O.
1090 */
1091ssize_t
Ross Zwisler11c59c92016-11-08 11:32:46 +11001092dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
Christoph Hellwig8ff6daa2017-01-27 23:20:26 -08001093 const struct iomap_ops *ops)
Christoph Hellwiga254e562016-09-19 11:24:49 +10001094{
1095 struct address_space *mapping = iocb->ki_filp->f_mapping;
1096 struct inode *inode = mapping->host;
1097 loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1098 unsigned flags = 0;
1099
Christoph Hellwig168316d2017-02-08 14:43:13 -05001100 if (iov_iter_rw(iter) == WRITE) {
1101 lockdep_assert_held_exclusive(&inode->i_rwsem);
Christoph Hellwiga254e562016-09-19 11:24:49 +10001102 flags |= IOMAP_WRITE;
Christoph Hellwig168316d2017-02-08 14:43:13 -05001103 } else {
1104 lockdep_assert_held(&inode->i_rwsem);
1105 }
Christoph Hellwiga254e562016-09-19 11:24:49 +10001106
Christoph Hellwiga254e562016-09-19 11:24:49 +10001107 while (iov_iter_count(iter)) {
1108 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
Ross Zwisler11c59c92016-11-08 11:32:46 +11001109 iter, dax_iomap_actor);
Christoph Hellwiga254e562016-09-19 11:24:49 +10001110 if (ret <= 0)
1111 break;
1112 pos += ret;
1113 done += ret;
1114 }
1115
1116 iocb->ki_pos += done;
1117 return done ? done : ret;
1118}
Ross Zwisler11c59c92016-11-08 11:32:46 +11001119EXPORT_SYMBOL_GPL(dax_iomap_rw);
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001120
Jan Kara9f141d62016-10-19 14:34:31 +02001121static int dax_fault_return(int error)
1122{
1123 if (error == 0)
1124 return VM_FAULT_NOPAGE;
1125 if (error == -ENOMEM)
1126 return VM_FAULT_OOM;
1127 return VM_FAULT_SIGBUS;
1128}
1129
Dave Jianga2d58162017-02-24 14:56:59 -08001130static int dax_iomap_pte_fault(struct vm_fault *vmf,
1131 const struct iomap_ops *ops)
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001132{
Dave Jiang11bac802017-02-24 14:56:41 -08001133 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001134 struct inode *inode = mapping->host;
Jan Kara1a29d852016-12-14 15:07:01 -08001135 unsigned long vaddr = vmf->address;
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001136 loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1137 sector_t sector;
1138 struct iomap iomap = { 0 };
Jan Kara9484ab12016-11-10 10:26:50 +11001139 unsigned flags = IOMAP_FAULT;
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001140 int error, major = 0;
Jan Karab1aa8122016-12-14 15:07:24 -08001141 int vmf_ret = 0;
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001142 void *entry;
1143
Ross Zwislera9c42b32017-05-08 16:00:00 -07001144 trace_dax_pte_fault(inode, vmf, vmf_ret);
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001145 /*
1146 * Check whether offset isn't beyond end of file now. Caller is supposed
1147 * to hold locks serializing us with truncate / punch hole so this is
1148 * a reliable test.
1149 */
Ross Zwislera9c42b32017-05-08 16:00:00 -07001150 if (pos >= i_size_read(inode)) {
1151 vmf_ret = VM_FAULT_SIGBUS;
1152 goto out;
1153 }
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001154
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001155 if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1156 flags |= IOMAP_WRITE;
1157
Jan Kara13e451f2017-05-12 15:46:57 -07001158 entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1159 if (IS_ERR(entry)) {
1160 vmf_ret = dax_fault_return(PTR_ERR(entry));
1161 goto out;
1162 }
1163
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001164 /*
Ross Zwislere2093922017-06-02 14:46:37 -07001165 * It is possible, particularly with mixed reads & writes to private
1166 * mappings, that we have raced with a PMD fault that overlaps with
1167 * the PTE we need to set up. If so just return and the fault will be
1168 * retried.
1169 */
1170 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1171 vmf_ret = VM_FAULT_NOPAGE;
1172 goto unlock_entry;
1173 }
1174
1175 /*
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001176 * Note that we don't bother to use iomap_apply here: DAX required
1177 * the file system block size to be equal the page size, which means
1178 * that we never have to deal with more than a single extent here.
1179 */
1180 error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
Ross Zwislera9c42b32017-05-08 16:00:00 -07001181 if (error) {
1182 vmf_ret = dax_fault_return(error);
Jan Kara13e451f2017-05-12 15:46:57 -07001183 goto unlock_entry;
Ross Zwislera9c42b32017-05-08 16:00:00 -07001184 }
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001185 if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
Jan Kara13e451f2017-05-12 15:46:57 -07001186 error = -EIO; /* fs corruption? */
1187 goto error_finish_iomap;
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001188 }
1189
Ross Zwisler333ccc92016-11-08 11:33:09 +11001190 sector = dax_iomap_sector(&iomap, pos);
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001191
1192 if (vmf->cow_page) {
1193 switch (iomap.type) {
1194 case IOMAP_HOLE:
1195 case IOMAP_UNWRITTEN:
1196 clear_user_highpage(vmf->cow_page, vaddr);
1197 break;
1198 case IOMAP_MAPPED:
Dan Williamscccbce62017-01-27 13:31:42 -08001199 error = copy_user_dax(iomap.bdev, iomap.dax_dev,
1200 sector, PAGE_SIZE, vmf->cow_page, vaddr);
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001201 break;
1202 default:
1203 WARN_ON_ONCE(1);
1204 error = -EIO;
1205 break;
1206 }
1207
1208 if (error)
Jan Kara13e451f2017-05-12 15:46:57 -07001209 goto error_finish_iomap;
Jan Karab1aa8122016-12-14 15:07:24 -08001210
1211 __SetPageUptodate(vmf->cow_page);
1212 vmf_ret = finish_fault(vmf);
1213 if (!vmf_ret)
1214 vmf_ret = VM_FAULT_DONE_COW;
Jan Kara13e451f2017-05-12 15:46:57 -07001215 goto finish_iomap;
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001216 }
1217
1218 switch (iomap.type) {
1219 case IOMAP_MAPPED:
1220 if (iomap.flags & IOMAP_F_NEW) {
1221 count_vm_event(PGMAJFAULT);
Roman Gushchin22621852017-07-06 15:40:25 -07001222 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001223 major = VM_FAULT_MAJOR;
1224 }
Dan Williamscccbce62017-01-27 13:31:42 -08001225 error = dax_insert_mapping(mapping, iomap.bdev, iomap.dax_dev,
1226 sector, PAGE_SIZE, &entry, vmf->vma, vmf);
Jan Kara9f141d62016-10-19 14:34:31 +02001227 /* -EBUSY is fine, somebody else faulted on the same PTE */
1228 if (error == -EBUSY)
1229 error = 0;
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001230 break;
1231 case IOMAP_UNWRITTEN:
1232 case IOMAP_HOLE:
Ross Zwisler15502902016-11-08 11:33:26 +11001233 if (!(vmf->flags & FAULT_FLAG_WRITE)) {
Jan Karaf449b932016-10-19 14:48:38 +02001234 vmf_ret = dax_load_hole(mapping, &entry, vmf);
Jan Kara13e451f2017-05-12 15:46:57 -07001235 goto finish_iomap;
Ross Zwisler15502902016-11-08 11:33:26 +11001236 }
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001237 /*FALLTHRU*/
1238 default:
1239 WARN_ON_ONCE(1);
1240 error = -EIO;
1241 break;
1242 }
1243
Jan Kara13e451f2017-05-12 15:46:57 -07001244 error_finish_iomap:
Jan Kara9f141d62016-10-19 14:34:31 +02001245 vmf_ret = dax_fault_return(error) | major;
Jan Kara9f141d62016-10-19 14:34:31 +02001246 finish_iomap:
1247 if (ops->iomap_end) {
1248 int copied = PAGE_SIZE;
1249
1250 if (vmf_ret & VM_FAULT_ERROR)
1251 copied = 0;
1252 /*
1253 * The fault is done by now and there's no way back (other
1254 * thread may be already happily using PTE we have installed).
1255 * Just ignore error from ->iomap_end since we cannot do much
1256 * with it.
1257 */
1258 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
Ross Zwisler15502902016-11-08 11:33:26 +11001259 }
Jan Kara13e451f2017-05-12 15:46:57 -07001260 unlock_entry:
1261 put_locked_mapping_entry(mapping, vmf->pgoff, entry);
1262 out:
Ross Zwislera9c42b32017-05-08 16:00:00 -07001263 trace_dax_pte_fault_done(inode, vmf, vmf_ret);
Jan Kara9f141d62016-10-19 14:34:31 +02001264 return vmf_ret;
Christoph Hellwiga7d73fe2016-09-19 11:24:50 +10001265}
Ross Zwisler642261a2016-11-08 11:34:45 +11001266
1267#ifdef CONFIG_FS_DAX_PMD
1268/*
1269 * The 'colour' (ie low bits) within a PMD of a page offset. This comes up
1270 * more often than one might expect in the below functions.
1271 */
1272#define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
1273
Dave Jiangf4200392017-02-22 15:40:06 -08001274static int dax_pmd_insert_mapping(struct vm_fault *vmf, struct iomap *iomap,
1275 loff_t pos, void **entryp)
Ross Zwisler642261a2016-11-08 11:34:45 +11001276{
Dave Jiangf4200392017-02-22 15:40:06 -08001277 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
Dan Williamscccbce62017-01-27 13:31:42 -08001278 const sector_t sector = dax_iomap_sector(iomap, pos);
1279 struct dax_device *dax_dev = iomap->dax_dev;
Ross Zwisler642261a2016-11-08 11:34:45 +11001280 struct block_device *bdev = iomap->bdev;
Ross Zwisler27a7ffa2017-02-22 15:40:00 -08001281 struct inode *inode = mapping->host;
Dan Williamscccbce62017-01-27 13:31:42 -08001282 const size_t size = PMD_SIZE;
1283 void *ret = NULL, *kaddr;
1284 long length = 0;
1285 pgoff_t pgoff;
1286 pfn_t pfn;
1287 int id;
Ross Zwisler642261a2016-11-08 11:34:45 +11001288
Dan Williamscccbce62017-01-27 13:31:42 -08001289 if (bdev_dax_pgoff(bdev, sector, size, &pgoff) != 0)
Ross Zwisler27a7ffa2017-02-22 15:40:00 -08001290 goto fallback;
Ross Zwisler642261a2016-11-08 11:34:45 +11001291
Dan Williamscccbce62017-01-27 13:31:42 -08001292 id = dax_read_lock();
1293 length = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
1294 if (length < 0)
1295 goto unlock_fallback;
1296 length = PFN_PHYS(length);
Ross Zwisler642261a2016-11-08 11:34:45 +11001297
Dan Williamscccbce62017-01-27 13:31:42 -08001298 if (length < size)
1299 goto unlock_fallback;
1300 if (pfn_t_to_pfn(pfn) & PG_PMD_COLOUR)
1301 goto unlock_fallback;
1302 if (!pfn_t_devmap(pfn))
1303 goto unlock_fallback;
1304 dax_read_unlock(id);
1305
1306 ret = dax_insert_mapping_entry(mapping, vmf, *entryp, sector,
Ross Zwisler642261a2016-11-08 11:34:45 +11001307 RADIX_DAX_PMD);
1308 if (IS_ERR(ret))
Ross Zwisler27a7ffa2017-02-22 15:40:00 -08001309 goto fallback;
Ross Zwisler642261a2016-11-08 11:34:45 +11001310 *entryp = ret;
1311
Dan Williamscccbce62017-01-27 13:31:42 -08001312 trace_dax_pmd_insert_mapping(inode, vmf, length, pfn, ret);
Dave Jiangf4200392017-02-22 15:40:06 -08001313 return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd,
Dan Williamscccbce62017-01-27 13:31:42 -08001314 pfn, vmf->flags & FAULT_FLAG_WRITE);
Ross Zwisler642261a2016-11-08 11:34:45 +11001315
Dan Williamscccbce62017-01-27 13:31:42 -08001316unlock_fallback:
1317 dax_read_unlock(id);
Ross Zwisler27a7ffa2017-02-22 15:40:00 -08001318fallback:
Dan Williamscccbce62017-01-27 13:31:42 -08001319 trace_dax_pmd_insert_mapping_fallback(inode, vmf, length, pfn, ret);
Ross Zwisler642261a2016-11-08 11:34:45 +11001320 return VM_FAULT_FALLBACK;
1321}
1322
Dave Jiangf4200392017-02-22 15:40:06 -08001323static int dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap,
1324 void **entryp)
Ross Zwisler642261a2016-11-08 11:34:45 +11001325{
Dave Jiangf4200392017-02-22 15:40:06 -08001326 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1327 unsigned long pmd_addr = vmf->address & PMD_MASK;
Ross Zwisler653b2ea2017-02-22 15:39:57 -08001328 struct inode *inode = mapping->host;
Ross Zwisler642261a2016-11-08 11:34:45 +11001329 struct page *zero_page;
Ross Zwisler653b2ea2017-02-22 15:39:57 -08001330 void *ret = NULL;
Ross Zwisler642261a2016-11-08 11:34:45 +11001331 spinlock_t *ptl;
1332 pmd_t pmd_entry;
Ross Zwisler642261a2016-11-08 11:34:45 +11001333
Dave Jiangf4200392017-02-22 15:40:06 -08001334 zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
Ross Zwisler642261a2016-11-08 11:34:45 +11001335
1336 if (unlikely(!zero_page))
Ross Zwisler653b2ea2017-02-22 15:39:57 -08001337 goto fallback;
Ross Zwisler642261a2016-11-08 11:34:45 +11001338
1339 ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0,
1340 RADIX_DAX_PMD | RADIX_DAX_HZP);
1341 if (IS_ERR(ret))
Ross Zwisler653b2ea2017-02-22 15:39:57 -08001342 goto fallback;
Ross Zwisler642261a2016-11-08 11:34:45 +11001343 *entryp = ret;
1344
Dave Jiangf4200392017-02-22 15:40:06 -08001345 ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1346 if (!pmd_none(*(vmf->pmd))) {
Ross Zwisler642261a2016-11-08 11:34:45 +11001347 spin_unlock(ptl);
Ross Zwisler653b2ea2017-02-22 15:39:57 -08001348 goto fallback;
Ross Zwisler642261a2016-11-08 11:34:45 +11001349 }
1350
Dave Jiangf4200392017-02-22 15:40:06 -08001351 pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
Ross Zwisler642261a2016-11-08 11:34:45 +11001352 pmd_entry = pmd_mkhuge(pmd_entry);
Dave Jiangf4200392017-02-22 15:40:06 -08001353 set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
Ross Zwisler642261a2016-11-08 11:34:45 +11001354 spin_unlock(ptl);
Dave Jiangf4200392017-02-22 15:40:06 -08001355 trace_dax_pmd_load_hole(inode, vmf, zero_page, ret);
Ross Zwisler642261a2016-11-08 11:34:45 +11001356 return VM_FAULT_NOPAGE;
Ross Zwisler653b2ea2017-02-22 15:39:57 -08001357
1358fallback:
Dave Jiangf4200392017-02-22 15:40:06 -08001359 trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret);
Ross Zwisler653b2ea2017-02-22 15:39:57 -08001360 return VM_FAULT_FALLBACK;
Ross Zwisler642261a2016-11-08 11:34:45 +11001361}
1362
Dave Jianga2d58162017-02-24 14:56:59 -08001363static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1364 const struct iomap_ops *ops)
Ross Zwisler642261a2016-11-08 11:34:45 +11001365{
Dave Jiangf4200392017-02-22 15:40:06 -08001366 struct vm_area_struct *vma = vmf->vma;
Ross Zwisler642261a2016-11-08 11:34:45 +11001367 struct address_space *mapping = vma->vm_file->f_mapping;
Dave Jiangd8a849e2017-02-22 15:40:03 -08001368 unsigned long pmd_addr = vmf->address & PMD_MASK;
1369 bool write = vmf->flags & FAULT_FLAG_WRITE;
Jan Kara9484ab12016-11-10 10:26:50 +11001370 unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
Ross Zwisler642261a2016-11-08 11:34:45 +11001371 struct inode *inode = mapping->host;
1372 int result = VM_FAULT_FALLBACK;
1373 struct iomap iomap = { 0 };
1374 pgoff_t max_pgoff, pgoff;
Ross Zwisler642261a2016-11-08 11:34:45 +11001375 void *entry;
1376 loff_t pos;
1377 int error;
1378
Ross Zwisler282a8e02017-02-22 15:39:50 -08001379 /*
1380 * Check whether offset isn't beyond end of file now. Caller is
1381 * supposed to hold locks serializing us with truncate / punch hole so
1382 * this is a reliable test.
1383 */
1384 pgoff = linear_page_index(vma, pmd_addr);
1385 max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT;
1386
Dave Jiangf4200392017-02-22 15:40:06 -08001387 trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
Ross Zwisler282a8e02017-02-22 15:39:50 -08001388
Ross Zwislerfffa2812017-08-25 15:55:36 -07001389 /*
1390 * Make sure that the faulting address's PMD offset (color) matches
1391 * the PMD offset from the start of the file. This is necessary so
1392 * that a PMD range in the page table overlaps exactly with a PMD
1393 * range in the radix tree.
1394 */
1395 if ((vmf->pgoff & PG_PMD_COLOUR) !=
1396 ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1397 goto fallback;
1398
Ross Zwisler642261a2016-11-08 11:34:45 +11001399 /* Fall back to PTEs if we're going to COW */
1400 if (write && !(vma->vm_flags & VM_SHARED))
1401 goto fallback;
1402
1403 /* If the PMD would extend outside the VMA */
1404 if (pmd_addr < vma->vm_start)
1405 goto fallback;
1406 if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1407 goto fallback;
1408
Ross Zwisler282a8e02017-02-22 15:39:50 -08001409 if (pgoff > max_pgoff) {
1410 result = VM_FAULT_SIGBUS;
1411 goto out;
1412 }
Ross Zwisler642261a2016-11-08 11:34:45 +11001413
1414 /* If the PMD would extend beyond the file size */
1415 if ((pgoff | PG_PMD_COLOUR) > max_pgoff)
1416 goto fallback;
1417
1418 /*
Jan Kara9f141d62016-10-19 14:34:31 +02001419 * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX
1420 * PMD or a HZP entry. If it can't (because a 4k page is already in
1421 * the tree, for instance), it will return -EEXIST and we just fall
1422 * back to 4k entries.
1423 */
1424 entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1425 if (IS_ERR(entry))
Ross Zwisler876f2942017-05-12 15:47:00 -07001426 goto fallback;
1427
1428 /*
Ross Zwislere2093922017-06-02 14:46:37 -07001429 * It is possible, particularly with mixed reads & writes to private
1430 * mappings, that we have raced with a PTE fault that overlaps with
1431 * the PMD we need to set up. If so just return and the fault will be
1432 * retried.
1433 */
1434 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1435 !pmd_devmap(*vmf->pmd)) {
1436 result = 0;
1437 goto unlock_entry;
1438 }
1439
1440 /*
Ross Zwisler876f2942017-05-12 15:47:00 -07001441 * Note that we don't use iomap_apply here. We aren't doing I/O, only
1442 * setting up a mapping, so really we're using iomap_begin() as a way
1443 * to look up our filesystem block.
1444 */
1445 pos = (loff_t)pgoff << PAGE_SHIFT;
1446 error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1447 if (error)
1448 goto unlock_entry;
1449
1450 if (iomap.offset + iomap.length < pos + PMD_SIZE)
Jan Kara9f141d62016-10-19 14:34:31 +02001451 goto finish_iomap;
1452
Ross Zwisler642261a2016-11-08 11:34:45 +11001453 switch (iomap.type) {
1454 case IOMAP_MAPPED:
Dave Jiangf4200392017-02-22 15:40:06 -08001455 result = dax_pmd_insert_mapping(vmf, &iomap, pos, &entry);
Ross Zwisler642261a2016-11-08 11:34:45 +11001456 break;
1457 case IOMAP_UNWRITTEN:
1458 case IOMAP_HOLE:
1459 if (WARN_ON_ONCE(write))
Ross Zwisler876f2942017-05-12 15:47:00 -07001460 break;
Dave Jiangf4200392017-02-22 15:40:06 -08001461 result = dax_pmd_load_hole(vmf, &iomap, &entry);
Ross Zwisler642261a2016-11-08 11:34:45 +11001462 break;
1463 default:
1464 WARN_ON_ONCE(1);
1465 break;
1466 }
1467
Jan Kara9f141d62016-10-19 14:34:31 +02001468 finish_iomap:
1469 if (ops->iomap_end) {
1470 int copied = PMD_SIZE;
1471
1472 if (result == VM_FAULT_FALLBACK)
1473 copied = 0;
1474 /*
1475 * The fault is done by now and there's no way back (other
1476 * thread may be already happily using PMD we have installed).
1477 * Just ignore error from ->iomap_end since we cannot do much
1478 * with it.
1479 */
1480 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1481 &iomap);
1482 }
Ross Zwisler876f2942017-05-12 15:47:00 -07001483 unlock_entry:
1484 put_locked_mapping_entry(mapping, pgoff, entry);
Ross Zwisler642261a2016-11-08 11:34:45 +11001485 fallback:
1486 if (result == VM_FAULT_FALLBACK) {
Dave Jiangd8a849e2017-02-22 15:40:03 -08001487 split_huge_pmd(vma, vmf->pmd, vmf->address);
Ross Zwisler642261a2016-11-08 11:34:45 +11001488 count_vm_event(THP_FAULT_FALLBACK);
1489 }
Ross Zwisler282a8e02017-02-22 15:39:50 -08001490out:
Dave Jiangf4200392017-02-22 15:40:06 -08001491 trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
Ross Zwisler642261a2016-11-08 11:34:45 +11001492 return result;
1493}
Dave Jianga2d58162017-02-24 14:56:59 -08001494#else
Arnd Bergmann01cddfe2017-02-27 14:26:44 -08001495static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1496 const struct iomap_ops *ops)
Dave Jianga2d58162017-02-24 14:56:59 -08001497{
1498 return VM_FAULT_FALLBACK;
1499}
Ross Zwisler642261a2016-11-08 11:34:45 +11001500#endif /* CONFIG_FS_DAX_PMD */
Dave Jianga2d58162017-02-24 14:56:59 -08001501
1502/**
1503 * dax_iomap_fault - handle a page fault on a DAX file
1504 * @vmf: The description of the fault
1505 * @ops: iomap ops passed from the file system
1506 *
1507 * When a page fault occurs, filesystems may call this helper in
1508 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1509 * has done all the necessary locking for page fault to proceed
1510 * successfully.
1511 */
Dave Jiangc791ace2017-02-24 14:57:08 -08001512int dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1513 const struct iomap_ops *ops)
Dave Jianga2d58162017-02-24 14:56:59 -08001514{
Dave Jiangc791ace2017-02-24 14:57:08 -08001515 switch (pe_size) {
1516 case PE_SIZE_PTE:
Dave Jianga2d58162017-02-24 14:56:59 -08001517 return dax_iomap_pte_fault(vmf, ops);
Dave Jiangc791ace2017-02-24 14:57:08 -08001518 case PE_SIZE_PMD:
Dave Jianga2d58162017-02-24 14:56:59 -08001519 return dax_iomap_pmd_fault(vmf, ops);
1520 default:
1521 return VM_FAULT_FALLBACK;
1522 }
1523}
1524EXPORT_SYMBOL_GPL(dax_iomap_fault);