blob: d5d7b9393bcaa119e6262b2deeb260f004e723ea [file] [log] [blame]
David Brazdil0f672f62019-12-10 10:32:29 +00001// SPDX-License-Identifier: GPL-2.0-only
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00002/*
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>
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00007 */
8
9#include <linux/atomic.h>
10#include <linux/blkdev.h>
11#include <linux/buffer_head.h>
12#include <linux/dax.h>
13#include <linux/fs.h>
14#include <linux/genhd.h>
15#include <linux/highmem.h>
16#include <linux/memcontrol.h>
17#include <linux/mm.h>
18#include <linux/mutex.h>
19#include <linux/pagevec.h>
20#include <linux/sched.h>
21#include <linux/sched/signal.h>
22#include <linux/uio.h>
23#include <linux/vmstat.h>
24#include <linux/pfn_t.h>
25#include <linux/sizes.h>
26#include <linux/mmu_notifier.h>
27#include <linux/iomap.h>
David Brazdil0f672f62019-12-10 10:32:29 +000028#include <asm/pgalloc.h>
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000029
30#define CREATE_TRACE_POINTS
31#include <trace/events/fs_dax.h>
32
David Brazdil0f672f62019-12-10 10:32:29 +000033static inline unsigned int pe_order(enum page_entry_size pe_size)
34{
35 if (pe_size == PE_SIZE_PTE)
36 return PAGE_SHIFT - PAGE_SHIFT;
37 if (pe_size == PE_SIZE_PMD)
38 return PMD_SHIFT - PAGE_SHIFT;
39 if (pe_size == PE_SIZE_PUD)
40 return PUD_SHIFT - PAGE_SHIFT;
41 return ~0;
42}
43
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000044/* We choose 4096 entries - same as per-zone page wait tables */
45#define DAX_WAIT_TABLE_BITS 12
46#define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
47
48/* The 'colour' (ie low bits) within a PMD of a page offset. */
49#define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
50#define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT)
51
David Brazdil0f672f62019-12-10 10:32:29 +000052/* The order of a PMD entry */
53#define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT)
54
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000055static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
56
57static int __init init_dax_wait_table(void)
58{
59 int i;
60
61 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
62 init_waitqueue_head(wait_table + i);
63 return 0;
64}
65fs_initcall(init_dax_wait_table);
66
67/*
David Brazdil0f672f62019-12-10 10:32:29 +000068 * DAX pagecache entries use XArray value entries so they can't be mistaken
69 * for pages. We use one bit for locking, one bit for the entry size (PMD)
70 * and two more to tell us if the entry is a zero page or an empty entry that
71 * is just used for locking. In total four special bits.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000072 *
73 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
74 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
75 * block allocation.
76 */
David Brazdil0f672f62019-12-10 10:32:29 +000077#define DAX_SHIFT (4)
78#define DAX_LOCKED (1UL << 0)
79#define DAX_PMD (1UL << 1)
80#define DAX_ZERO_PAGE (1UL << 2)
81#define DAX_EMPTY (1UL << 3)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000082
David Brazdil0f672f62019-12-10 10:32:29 +000083static unsigned long dax_to_pfn(void *entry)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000084{
David Brazdil0f672f62019-12-10 10:32:29 +000085 return xa_to_value(entry) >> DAX_SHIFT;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000086}
87
David Brazdil0f672f62019-12-10 10:32:29 +000088static void *dax_make_entry(pfn_t pfn, unsigned long flags)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000089{
David Brazdil0f672f62019-12-10 10:32:29 +000090 return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000091}
92
David Brazdil0f672f62019-12-10 10:32:29 +000093static bool dax_is_locked(void *entry)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000094{
David Brazdil0f672f62019-12-10 10:32:29 +000095 return xa_to_value(entry) & DAX_LOCKED;
96}
97
98static unsigned int dax_entry_order(void *entry)
99{
100 if (xa_to_value(entry) & DAX_PMD)
101 return PMD_ORDER;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000102 return 0;
103}
104
David Brazdil0f672f62019-12-10 10:32:29 +0000105static unsigned long dax_is_pmd_entry(void *entry)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000106{
David Brazdil0f672f62019-12-10 10:32:29 +0000107 return xa_to_value(entry) & DAX_PMD;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000108}
109
David Brazdil0f672f62019-12-10 10:32:29 +0000110static bool dax_is_pte_entry(void *entry)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000111{
David Brazdil0f672f62019-12-10 10:32:29 +0000112 return !(xa_to_value(entry) & DAX_PMD);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000113}
114
115static int dax_is_zero_entry(void *entry)
116{
David Brazdil0f672f62019-12-10 10:32:29 +0000117 return xa_to_value(entry) & DAX_ZERO_PAGE;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000118}
119
120static int dax_is_empty_entry(void *entry)
121{
David Brazdil0f672f62019-12-10 10:32:29 +0000122 return xa_to_value(entry) & DAX_EMPTY;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000123}
124
125/*
David Brazdil0f672f62019-12-10 10:32:29 +0000126 * true if the entry that was found is of a smaller order than the entry
127 * we were looking for
128 */
129static bool dax_is_conflict(void *entry)
130{
131 return entry == XA_RETRY_ENTRY;
132}
133
134/*
135 * DAX page cache entry locking
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000136 */
137struct exceptional_entry_key {
David Brazdil0f672f62019-12-10 10:32:29 +0000138 struct xarray *xa;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000139 pgoff_t entry_start;
140};
141
142struct wait_exceptional_entry_queue {
143 wait_queue_entry_t wait;
144 struct exceptional_entry_key key;
145};
146
Olivier Deprez157378f2022-04-04 15:47:50 +0200147/**
148 * enum dax_wake_mode: waitqueue wakeup behaviour
149 * @WAKE_ALL: wake all waiters in the waitqueue
150 * @WAKE_NEXT: wake only the first waiter in the waitqueue
151 */
152enum dax_wake_mode {
153 WAKE_ALL,
154 WAKE_NEXT,
155};
156
David Brazdil0f672f62019-12-10 10:32:29 +0000157static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas,
158 void *entry, struct exceptional_entry_key *key)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000159{
160 unsigned long hash;
David Brazdil0f672f62019-12-10 10:32:29 +0000161 unsigned long index = xas->xa_index;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000162
163 /*
164 * If 'entry' is a PMD, align the 'index' that we use for the wait
165 * queue to the start of that PMD. This ensures that all offsets in
166 * the range covered by the PMD map to the same bit lock.
167 */
168 if (dax_is_pmd_entry(entry))
169 index &= ~PG_PMD_COLOUR;
David Brazdil0f672f62019-12-10 10:32:29 +0000170 key->xa = xas->xa;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000171 key->entry_start = index;
172
David Brazdil0f672f62019-12-10 10:32:29 +0000173 hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000174 return wait_table + hash;
175}
176
David Brazdil0f672f62019-12-10 10:32:29 +0000177static int wake_exceptional_entry_func(wait_queue_entry_t *wait,
178 unsigned int mode, int sync, void *keyp)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000179{
180 struct exceptional_entry_key *key = keyp;
181 struct wait_exceptional_entry_queue *ewait =
182 container_of(wait, struct wait_exceptional_entry_queue, wait);
183
David Brazdil0f672f62019-12-10 10:32:29 +0000184 if (key->xa != ewait->key.xa ||
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000185 key->entry_start != ewait->key.entry_start)
186 return 0;
187 return autoremove_wake_function(wait, mode, sync, NULL);
188}
189
190/*
191 * @entry may no longer be the entry at the index in the mapping.
192 * The important information it's conveying is whether the entry at
193 * this index used to be a PMD entry.
194 */
Olivier Deprez157378f2022-04-04 15:47:50 +0200195static void dax_wake_entry(struct xa_state *xas, void *entry,
196 enum dax_wake_mode mode)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000197{
198 struct exceptional_entry_key key;
199 wait_queue_head_t *wq;
200
David Brazdil0f672f62019-12-10 10:32:29 +0000201 wq = dax_entry_waitqueue(xas, entry, &key);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000202
203 /*
204 * Checking for locked entry and prepare_to_wait_exclusive() happens
205 * under the i_pages lock, ditto for entry handling in our callers.
206 * So at this point all tasks that could have seen our entry locked
207 * must be in the waitqueue and the following check will see them.
208 */
209 if (waitqueue_active(wq))
Olivier Deprez157378f2022-04-04 15:47:50 +0200210 __wake_up(wq, TASK_NORMAL, mode == WAKE_ALL ? 0 : 1, &key);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000211}
212
213/*
David Brazdil0f672f62019-12-10 10:32:29 +0000214 * Look up entry in page cache, wait for it to become unlocked if it
215 * is a DAX entry and return it. The caller must subsequently call
216 * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
217 * if it did. The entry returned may have a larger order than @order.
218 * If @order is larger than the order of the entry found in i_pages, this
219 * function returns a dax_is_conflict entry.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000220 *
221 * Must be called with the i_pages lock held.
222 */
David Brazdil0f672f62019-12-10 10:32:29 +0000223static void *get_unlocked_entry(struct xa_state *xas, unsigned int order)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000224{
David Brazdil0f672f62019-12-10 10:32:29 +0000225 void *entry;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000226 struct wait_exceptional_entry_queue ewait;
227 wait_queue_head_t *wq;
228
229 init_wait(&ewait.wait);
230 ewait.wait.func = wake_exceptional_entry_func;
231
232 for (;;) {
David Brazdil0f672f62019-12-10 10:32:29 +0000233 entry = xas_find_conflict(xas);
234 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000235 return entry;
David Brazdil0f672f62019-12-10 10:32:29 +0000236 if (dax_entry_order(entry) < order)
237 return XA_RETRY_ENTRY;
238 if (!dax_is_locked(entry))
239 return entry;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000240
David Brazdil0f672f62019-12-10 10:32:29 +0000241 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000242 prepare_to_wait_exclusive(wq, &ewait.wait,
243 TASK_UNINTERRUPTIBLE);
David Brazdil0f672f62019-12-10 10:32:29 +0000244 xas_unlock_irq(xas);
245 xas_reset(xas);
246 schedule();
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000247 finish_wait(wq, &ewait.wait);
David Brazdil0f672f62019-12-10 10:32:29 +0000248 xas_lock_irq(xas);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000249 }
250}
251
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000252/*
David Brazdil0f672f62019-12-10 10:32:29 +0000253 * The only thing keeping the address space around is the i_pages lock
254 * (it's cycled in clear_inode() after removing the entries from i_pages)
255 * After we call xas_unlock_irq(), we cannot touch xas->xa.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000256 */
David Brazdil0f672f62019-12-10 10:32:29 +0000257static void wait_entry_unlocked(struct xa_state *xas, void *entry)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000258{
David Brazdil0f672f62019-12-10 10:32:29 +0000259 struct wait_exceptional_entry_queue ewait;
260 wait_queue_head_t *wq;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000261
David Brazdil0f672f62019-12-10 10:32:29 +0000262 init_wait(&ewait.wait);
263 ewait.wait.func = wake_exceptional_entry_func;
264
265 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
266 /*
267 * Unlike get_unlocked_entry() there is no guarantee that this
268 * path ever successfully retrieves an unlocked entry before an
269 * inode dies. Perform a non-exclusive wait in case this path
270 * never successfully performs its own wake up.
271 */
272 prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
273 xas_unlock_irq(xas);
274 schedule();
275 finish_wait(wq, &ewait.wait);
276}
277
Olivier Deprez157378f2022-04-04 15:47:50 +0200278static void put_unlocked_entry(struct xa_state *xas, void *entry,
279 enum dax_wake_mode mode)
David Brazdil0f672f62019-12-10 10:32:29 +0000280{
David Brazdil0f672f62019-12-10 10:32:29 +0000281 if (entry && !dax_is_conflict(entry))
Olivier Deprez157378f2022-04-04 15:47:50 +0200282 dax_wake_entry(xas, entry, mode);
David Brazdil0f672f62019-12-10 10:32:29 +0000283}
284
285/*
286 * We used the xa_state to get the entry, but then we locked the entry and
287 * dropped the xa_lock, so we know the xa_state is stale and must be reset
288 * before use.
289 */
290static void dax_unlock_entry(struct xa_state *xas, void *entry)
291{
292 void *old;
293
294 BUG_ON(dax_is_locked(entry));
295 xas_reset(xas);
296 xas_lock_irq(xas);
297 old = xas_store(xas, entry);
298 xas_unlock_irq(xas);
299 BUG_ON(!dax_is_locked(old));
Olivier Deprez157378f2022-04-04 15:47:50 +0200300 dax_wake_entry(xas, entry, WAKE_NEXT);
David Brazdil0f672f62019-12-10 10:32:29 +0000301}
302
303/*
304 * Return: The entry stored at this location before it was locked.
305 */
306static void *dax_lock_entry(struct xa_state *xas, void *entry)
307{
308 unsigned long v = xa_to_value(entry);
309 return xas_store(xas, xa_mk_value(v | DAX_LOCKED));
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000310}
311
312static unsigned long dax_entry_size(void *entry)
313{
314 if (dax_is_zero_entry(entry))
315 return 0;
316 else if (dax_is_empty_entry(entry))
317 return 0;
318 else if (dax_is_pmd_entry(entry))
319 return PMD_SIZE;
320 else
321 return PAGE_SIZE;
322}
323
David Brazdil0f672f62019-12-10 10:32:29 +0000324static unsigned long dax_end_pfn(void *entry)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000325{
David Brazdil0f672f62019-12-10 10:32:29 +0000326 return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000327}
328
329/*
330 * Iterate through all mapped pfns represented by an entry, i.e. skip
331 * 'empty' and 'zero' entries.
332 */
333#define for_each_mapped_pfn(entry, pfn) \
David Brazdil0f672f62019-12-10 10:32:29 +0000334 for (pfn = dax_to_pfn(entry); \
335 pfn < dax_end_pfn(entry); pfn++)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000336
337/*
338 * TODO: for reflink+dax we need a way to associate a single page with
339 * multiple address_space instances at different linear_page_index()
340 * offsets.
341 */
342static void dax_associate_entry(void *entry, struct address_space *mapping,
343 struct vm_area_struct *vma, unsigned long address)
344{
345 unsigned long size = dax_entry_size(entry), pfn, index;
346 int i = 0;
347
348 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
349 return;
350
351 index = linear_page_index(vma, address & ~(size - 1));
352 for_each_mapped_pfn(entry, pfn) {
353 struct page *page = pfn_to_page(pfn);
354
355 WARN_ON_ONCE(page->mapping);
356 page->mapping = mapping;
357 page->index = index + i++;
358 }
359}
360
361static void dax_disassociate_entry(void *entry, struct address_space *mapping,
362 bool trunc)
363{
364 unsigned long pfn;
365
366 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
367 return;
368
369 for_each_mapped_pfn(entry, pfn) {
370 struct page *page = pfn_to_page(pfn);
371
372 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
373 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
374 page->mapping = NULL;
375 page->index = 0;
376 }
377}
378
379static struct page *dax_busy_page(void *entry)
380{
381 unsigned long pfn;
382
383 for_each_mapped_pfn(entry, pfn) {
384 struct page *page = pfn_to_page(pfn);
385
386 if (page_ref_count(page) > 1)
387 return page;
388 }
389 return NULL;
390}
391
David Brazdil0f672f62019-12-10 10:32:29 +0000392/*
393 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a page
394 * @page: The page whose entry we want to lock
395 *
396 * Context: Process context.
397 * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could
398 * not be locked.
399 */
400dax_entry_t dax_lock_page(struct page *page)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000401{
David Brazdil0f672f62019-12-10 10:32:29 +0000402 XA_STATE(xas, NULL, 0);
403 void *entry;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000404
David Brazdil0f672f62019-12-10 10:32:29 +0000405 /* Ensure page->mapping isn't freed while we look at it */
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000406 rcu_read_lock();
407 for (;;) {
David Brazdil0f672f62019-12-10 10:32:29 +0000408 struct address_space *mapping = READ_ONCE(page->mapping);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000409
David Brazdil0f672f62019-12-10 10:32:29 +0000410 entry = NULL;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000411 if (!mapping || !dax_mapping(mapping))
412 break;
413
414 /*
415 * In the device-dax case there's no need to lock, a
416 * struct dev_pagemap pin is sufficient to keep the
417 * inode alive, and we assume we have dev_pagemap pin
418 * otherwise we would not have a valid pfn_to_page()
419 * translation.
420 */
David Brazdil0f672f62019-12-10 10:32:29 +0000421 entry = (void *)~0UL;
422 if (S_ISCHR(mapping->host->i_mode))
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000423 break;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000424
David Brazdil0f672f62019-12-10 10:32:29 +0000425 xas.xa = &mapping->i_pages;
426 xas_lock_irq(&xas);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000427 if (mapping != page->mapping) {
David Brazdil0f672f62019-12-10 10:32:29 +0000428 xas_unlock_irq(&xas);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000429 continue;
430 }
David Brazdil0f672f62019-12-10 10:32:29 +0000431 xas_set(&xas, page->index);
432 entry = xas_load(&xas);
433 if (dax_is_locked(entry)) {
434 rcu_read_unlock();
435 wait_entry_unlocked(&xas, entry);
436 rcu_read_lock();
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000437 continue;
438 }
David Brazdil0f672f62019-12-10 10:32:29 +0000439 dax_lock_entry(&xas, entry);
440 xas_unlock_irq(&xas);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000441 break;
442 }
443 rcu_read_unlock();
David Brazdil0f672f62019-12-10 10:32:29 +0000444 return (dax_entry_t)entry;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000445}
446
David Brazdil0f672f62019-12-10 10:32:29 +0000447void dax_unlock_page(struct page *page, dax_entry_t cookie)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000448{
449 struct address_space *mapping = page->mapping;
David Brazdil0f672f62019-12-10 10:32:29 +0000450 XA_STATE(xas, &mapping->i_pages, page->index);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000451
David Brazdil0f672f62019-12-10 10:32:29 +0000452 if (S_ISCHR(mapping->host->i_mode))
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000453 return;
454
David Brazdil0f672f62019-12-10 10:32:29 +0000455 dax_unlock_entry(&xas, (void *)cookie);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000456}
457
458/*
David Brazdil0f672f62019-12-10 10:32:29 +0000459 * Find page cache entry at given index. If it is a DAX entry, return it
460 * with the entry locked. If the page cache doesn't contain an entry at
461 * that index, add a locked empty entry.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000462 *
David Brazdil0f672f62019-12-10 10:32:29 +0000463 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
464 * either return that locked entry or will return VM_FAULT_FALLBACK.
465 * This will happen if there are any PTE entries within the PMD range
466 * that we are requesting.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000467 *
David Brazdil0f672f62019-12-10 10:32:29 +0000468 * We always favor PTE entries over PMD entries. There isn't a flow where we
469 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD
470 * insertion will fail if it finds any PTE entries already in the tree, and a
471 * PTE insertion will cause an existing PMD entry to be unmapped and
472 * downgraded to PTE entries. This happens for both PMD zero pages as
473 * well as PMD empty entries.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000474 *
David Brazdil0f672f62019-12-10 10:32:29 +0000475 * The exception to this downgrade path is for PMD entries that have
476 * real storage backing them. We will leave these real PMD entries in
477 * the tree, and PTE writes will simply dirty the entire PMD entry.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000478 *
479 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
480 * persistent memory the benefit is doubtful. We can add that later if we can
481 * show it helps.
David Brazdil0f672f62019-12-10 10:32:29 +0000482 *
483 * On error, this function does not return an ERR_PTR. Instead it returns
484 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values
485 * overlap with xarray value entries.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000486 */
David Brazdil0f672f62019-12-10 10:32:29 +0000487static void *grab_mapping_entry(struct xa_state *xas,
488 struct address_space *mapping, unsigned int order)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000489{
David Brazdil0f672f62019-12-10 10:32:29 +0000490 unsigned long index = xas->xa_index;
Olivier Deprez0e641232021-09-23 10:07:05 +0200491 bool pmd_downgrade; /* splitting PMD entry into PTE entries? */
David Brazdil0f672f62019-12-10 10:32:29 +0000492 void *entry;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000493
David Brazdil0f672f62019-12-10 10:32:29 +0000494retry:
Olivier Deprez0e641232021-09-23 10:07:05 +0200495 pmd_downgrade = false;
David Brazdil0f672f62019-12-10 10:32:29 +0000496 xas_lock_irq(xas);
497 entry = get_unlocked_entry(xas, order);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000498
499 if (entry) {
David Brazdil0f672f62019-12-10 10:32:29 +0000500 if (dax_is_conflict(entry))
501 goto fallback;
502 if (!xa_is_value(entry)) {
Olivier Deprez157378f2022-04-04 15:47:50 +0200503 xas_set_err(xas, -EIO);
David Brazdil0f672f62019-12-10 10:32:29 +0000504 goto out_unlock;
505 }
506
507 if (order == 0) {
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000508 if (dax_is_pmd_entry(entry) &&
509 (dax_is_zero_entry(entry) ||
510 dax_is_empty_entry(entry))) {
511 pmd_downgrade = true;
512 }
513 }
514 }
515
David Brazdil0f672f62019-12-10 10:32:29 +0000516 if (pmd_downgrade) {
517 /*
518 * Make sure 'entry' remains valid while we drop
519 * the i_pages lock.
520 */
521 dax_lock_entry(xas, entry);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000522
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000523 /*
524 * Besides huge zero pages the only other thing that gets
525 * downgraded are empty entries which don't need to be
526 * unmapped.
527 */
David Brazdil0f672f62019-12-10 10:32:29 +0000528 if (dax_is_zero_entry(entry)) {
529 xas_unlock_irq(xas);
530 unmap_mapping_pages(mapping,
531 xas->xa_index & ~PG_PMD_COLOUR,
532 PG_PMD_NR, false);
533 xas_reset(xas);
534 xas_lock_irq(xas);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000535 }
536
David Brazdil0f672f62019-12-10 10:32:29 +0000537 dax_disassociate_entry(entry, mapping, false);
538 xas_store(xas, NULL); /* undo the PMD join */
Olivier Deprez157378f2022-04-04 15:47:50 +0200539 dax_wake_entry(xas, entry, WAKE_ALL);
David Brazdil0f672f62019-12-10 10:32:29 +0000540 mapping->nrexceptional--;
541 entry = NULL;
542 xas_set(xas, index);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000543 }
David Brazdil0f672f62019-12-10 10:32:29 +0000544
545 if (entry) {
546 dax_lock_entry(xas, entry);
547 } else {
548 unsigned long flags = DAX_EMPTY;
549
550 if (order > 0)
551 flags |= DAX_PMD;
552 entry = dax_make_entry(pfn_to_pfn_t(0), flags);
553 dax_lock_entry(xas, entry);
554 if (xas_error(xas))
555 goto out_unlock;
556 mapping->nrexceptional++;
557 }
558
559out_unlock:
560 xas_unlock_irq(xas);
561 if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM))
562 goto retry;
563 if (xas->xa_node == XA_ERROR(-ENOMEM))
564 return xa_mk_internal(VM_FAULT_OOM);
565 if (xas_error(xas))
566 return xa_mk_internal(VM_FAULT_SIGBUS);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000567 return entry;
David Brazdil0f672f62019-12-10 10:32:29 +0000568fallback:
569 xas_unlock_irq(xas);
570 return xa_mk_internal(VM_FAULT_FALLBACK);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000571}
572
573/**
Olivier Deprez157378f2022-04-04 15:47:50 +0200574 * dax_layout_busy_page_range - find first pinned page in @mapping
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000575 * @mapping: address space to scan for a page with ref count > 1
Olivier Deprez157378f2022-04-04 15:47:50 +0200576 * @start: Starting offset. Page containing 'start' is included.
577 * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX,
578 * pages from 'start' till the end of file are included.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000579 *
580 * DAX requires ZONE_DEVICE mapped pages. These pages are never
581 * 'onlined' to the page allocator so they are considered idle when
582 * page->count == 1. A filesystem uses this interface to determine if
583 * any page in the mapping is busy, i.e. for DMA, or other
584 * get_user_pages() usages.
585 *
586 * It is expected that the filesystem is holding locks to block the
587 * establishment of new mappings in this address_space. I.e. it expects
588 * to be able to run unmap_mapping_range() and subsequently not race
589 * mapping_mapped() becoming true.
590 */
Olivier Deprez157378f2022-04-04 15:47:50 +0200591struct page *dax_layout_busy_page_range(struct address_space *mapping,
592 loff_t start, loff_t end)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000593{
David Brazdil0f672f62019-12-10 10:32:29 +0000594 void *entry;
595 unsigned int scanned = 0;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000596 struct page *page = NULL;
Olivier Deprez157378f2022-04-04 15:47:50 +0200597 pgoff_t start_idx = start >> PAGE_SHIFT;
598 pgoff_t end_idx;
599 XA_STATE(xas, &mapping->i_pages, start_idx);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000600
601 /*
602 * In the 'limited' case get_user_pages() for dax is disabled.
603 */
604 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
605 return NULL;
606
607 if (!dax_mapping(mapping) || !mapping_mapped(mapping))
608 return NULL;
609
Olivier Deprez157378f2022-04-04 15:47:50 +0200610 /* If end == LLONG_MAX, all pages from start to till end of file */
611 if (end == LLONG_MAX)
612 end_idx = ULONG_MAX;
613 else
614 end_idx = end >> PAGE_SHIFT;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000615 /*
616 * If we race get_user_pages_fast() here either we'll see the
David Brazdil0f672f62019-12-10 10:32:29 +0000617 * elevated page count in the iteration and wait, or
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000618 * get_user_pages_fast() will see that the page it took a reference
619 * against is no longer mapped in the page tables and bail to the
620 * get_user_pages() slow path. The slow path is protected by
621 * pte_lock() and pmd_lock(). New references are not taken without
Olivier Deprez157378f2022-04-04 15:47:50 +0200622 * holding those locks, and unmap_mapping_pages() will not zero the
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000623 * pte or pmd without holding the respective lock, so we are
624 * guaranteed to either see new references or prevent new
625 * references from being established.
626 */
Olivier Deprez157378f2022-04-04 15:47:50 +0200627 unmap_mapping_pages(mapping, start_idx, end_idx - start_idx + 1, 0);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000628
David Brazdil0f672f62019-12-10 10:32:29 +0000629 xas_lock_irq(&xas);
Olivier Deprez157378f2022-04-04 15:47:50 +0200630 xas_for_each(&xas, entry, end_idx) {
David Brazdil0f672f62019-12-10 10:32:29 +0000631 if (WARN_ON_ONCE(!xa_is_value(entry)))
632 continue;
633 if (unlikely(dax_is_locked(entry)))
634 entry = get_unlocked_entry(&xas, 0);
635 if (entry)
636 page = dax_busy_page(entry);
Olivier Deprez157378f2022-04-04 15:47:50 +0200637 put_unlocked_entry(&xas, entry, WAKE_NEXT);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000638 if (page)
639 break;
David Brazdil0f672f62019-12-10 10:32:29 +0000640 if (++scanned % XA_CHECK_SCHED)
641 continue;
642
643 xas_pause(&xas);
644 xas_unlock_irq(&xas);
645 cond_resched();
646 xas_lock_irq(&xas);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000647 }
David Brazdil0f672f62019-12-10 10:32:29 +0000648 xas_unlock_irq(&xas);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000649 return page;
650}
Olivier Deprez157378f2022-04-04 15:47:50 +0200651EXPORT_SYMBOL_GPL(dax_layout_busy_page_range);
652
653struct page *dax_layout_busy_page(struct address_space *mapping)
654{
655 return dax_layout_busy_page_range(mapping, 0, LLONG_MAX);
656}
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000657EXPORT_SYMBOL_GPL(dax_layout_busy_page);
658
David Brazdil0f672f62019-12-10 10:32:29 +0000659static int __dax_invalidate_entry(struct address_space *mapping,
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000660 pgoff_t index, bool trunc)
661{
David Brazdil0f672f62019-12-10 10:32:29 +0000662 XA_STATE(xas, &mapping->i_pages, index);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000663 int ret = 0;
664 void *entry;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000665
David Brazdil0f672f62019-12-10 10:32:29 +0000666 xas_lock_irq(&xas);
667 entry = get_unlocked_entry(&xas, 0);
668 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000669 goto out;
670 if (!trunc &&
David Brazdil0f672f62019-12-10 10:32:29 +0000671 (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) ||
672 xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE)))
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000673 goto out;
674 dax_disassociate_entry(entry, mapping, trunc);
David Brazdil0f672f62019-12-10 10:32:29 +0000675 xas_store(&xas, NULL);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000676 mapping->nrexceptional--;
677 ret = 1;
678out:
Olivier Deprez157378f2022-04-04 15:47:50 +0200679 put_unlocked_entry(&xas, entry, WAKE_ALL);
David Brazdil0f672f62019-12-10 10:32:29 +0000680 xas_unlock_irq(&xas);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000681 return ret;
682}
David Brazdil0f672f62019-12-10 10:32:29 +0000683
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000684/*
David Brazdil0f672f62019-12-10 10:32:29 +0000685 * Delete DAX entry at @index from @mapping. Wait for it
686 * to be unlocked before deleting it.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000687 */
688int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
689{
David Brazdil0f672f62019-12-10 10:32:29 +0000690 int ret = __dax_invalidate_entry(mapping, index, true);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000691
692 /*
693 * This gets called from truncate / punch_hole path. As such, the caller
694 * must hold locks protecting against concurrent modifications of the
David Brazdil0f672f62019-12-10 10:32:29 +0000695 * page cache (usually fs-private i_mmap_sem for writing). Since the
696 * caller has seen a DAX entry for this index, we better find it
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000697 * at that index as well...
698 */
699 WARN_ON_ONCE(!ret);
700 return ret;
701}
702
703/*
David Brazdil0f672f62019-12-10 10:32:29 +0000704 * Invalidate DAX entry if it is clean.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000705 */
706int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
707 pgoff_t index)
708{
David Brazdil0f672f62019-12-10 10:32:29 +0000709 return __dax_invalidate_entry(mapping, index, false);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000710}
711
Olivier Deprez157378f2022-04-04 15:47:50 +0200712static int copy_cow_page_dax(struct block_device *bdev, struct dax_device *dax_dev,
713 sector_t sector, struct page *to, unsigned long vaddr)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000714{
715 void *vto, *kaddr;
716 pgoff_t pgoff;
717 long rc;
718 int id;
719
Olivier Deprez157378f2022-04-04 15:47:50 +0200720 rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000721 if (rc)
722 return rc;
723
724 id = dax_read_lock();
Olivier Deprez157378f2022-04-04 15:47:50 +0200725 rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(PAGE_SIZE), &kaddr, NULL);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000726 if (rc < 0) {
727 dax_read_unlock(id);
728 return rc;
729 }
730 vto = kmap_atomic(to);
731 copy_user_page(vto, (void __force *)kaddr, vaddr, to);
732 kunmap_atomic(vto);
733 dax_read_unlock(id);
734 return 0;
735}
736
737/*
738 * By this point grab_mapping_entry() has ensured that we have a locked entry
739 * of the appropriate size so we don't have to worry about downgrading PMDs to
740 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
741 * already in the tree, we will skip the insertion and just dirty the PMD as
742 * appropriate.
743 */
David Brazdil0f672f62019-12-10 10:32:29 +0000744static void *dax_insert_entry(struct xa_state *xas,
745 struct address_space *mapping, struct vm_fault *vmf,
746 void *entry, pfn_t pfn, unsigned long flags, bool dirty)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000747{
David Brazdil0f672f62019-12-10 10:32:29 +0000748 void *new_entry = dax_make_entry(pfn, flags);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000749
750 if (dirty)
751 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
752
David Brazdil0f672f62019-12-10 10:32:29 +0000753 if (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE)) {
754 unsigned long index = xas->xa_index;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000755 /* we are replacing a zero page with block mapping */
756 if (dax_is_pmd_entry(entry))
757 unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
David Brazdil0f672f62019-12-10 10:32:29 +0000758 PG_PMD_NR, false);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000759 else /* pte entry */
David Brazdil0f672f62019-12-10 10:32:29 +0000760 unmap_mapping_pages(mapping, index, 1, false);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000761 }
762
David Brazdil0f672f62019-12-10 10:32:29 +0000763 xas_reset(xas);
764 xas_lock_irq(xas);
765 if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
766 void *old;
767
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000768 dax_disassociate_entry(entry, mapping, false);
769 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000770 /*
David Brazdil0f672f62019-12-10 10:32:29 +0000771 * Only swap our new entry into the page cache if the current
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000772 * entry is a zero page or an empty entry. If a normal PTE or
David Brazdil0f672f62019-12-10 10:32:29 +0000773 * PMD entry is already in the cache, we leave it alone. This
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000774 * means that if we are trying to insert a PTE and the
775 * existing entry is a PMD, we will just leave the PMD in the
776 * tree and dirty it if necessary.
777 */
David Brazdil0f672f62019-12-10 10:32:29 +0000778 old = dax_lock_entry(xas, new_entry);
779 WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) |
780 DAX_LOCKED));
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000781 entry = new_entry;
David Brazdil0f672f62019-12-10 10:32:29 +0000782 } else {
783 xas_load(xas); /* Walk the xa_state */
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000784 }
785
786 if (dirty)
David Brazdil0f672f62019-12-10 10:32:29 +0000787 xas_set_mark(xas, PAGECACHE_TAG_DIRTY);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000788
David Brazdil0f672f62019-12-10 10:32:29 +0000789 xas_unlock_irq(xas);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000790 return entry;
791}
792
David Brazdil0f672f62019-12-10 10:32:29 +0000793static inline
794unsigned long pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000795{
796 unsigned long address;
797
798 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
799 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
800 return address;
801}
802
803/* Walk all mappings of a given index of a file and writeprotect them */
David Brazdil0f672f62019-12-10 10:32:29 +0000804static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index,
805 unsigned long pfn)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000806{
807 struct vm_area_struct *vma;
808 pte_t pte, *ptep = NULL;
809 pmd_t *pmdp = NULL;
810 spinlock_t *ptl;
811
812 i_mmap_lock_read(mapping);
813 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
David Brazdil0f672f62019-12-10 10:32:29 +0000814 struct mmu_notifier_range range;
815 unsigned long address;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000816
817 cond_resched();
818
819 if (!(vma->vm_flags & VM_SHARED))
820 continue;
821
822 address = pgoff_address(index, vma);
823
824 /*
Olivier Deprez0e641232021-09-23 10:07:05 +0200825 * follow_invalidate_pte() will use the range to call
826 * mmu_notifier_invalidate_range_start() on our behalf before
827 * taking any lock.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000828 */
Olivier Deprez0e641232021-09-23 10:07:05 +0200829 if (follow_invalidate_pte(vma->vm_mm, address, &range, &ptep,
830 &pmdp, &ptl))
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000831 continue;
832
833 /*
834 * No need to call mmu_notifier_invalidate_range() as we are
835 * downgrading page table protection not changing it to point
836 * to a new page.
837 *
838 * See Documentation/vm/mmu_notifier.rst
839 */
840 if (pmdp) {
841#ifdef CONFIG_FS_DAX_PMD
842 pmd_t pmd;
843
844 if (pfn != pmd_pfn(*pmdp))
845 goto unlock_pmd;
846 if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
847 goto unlock_pmd;
848
849 flush_cache_page(vma, address, pfn);
David Brazdil0f672f62019-12-10 10:32:29 +0000850 pmd = pmdp_invalidate(vma, address, pmdp);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000851 pmd = pmd_wrprotect(pmd);
852 pmd = pmd_mkclean(pmd);
853 set_pmd_at(vma->vm_mm, address, pmdp, pmd);
854unlock_pmd:
855#endif
856 spin_unlock(ptl);
857 } else {
858 if (pfn != pte_pfn(*ptep))
859 goto unlock_pte;
860 if (!pte_dirty(*ptep) && !pte_write(*ptep))
861 goto unlock_pte;
862
863 flush_cache_page(vma, address, pfn);
864 pte = ptep_clear_flush(vma, address, ptep);
865 pte = pte_wrprotect(pte);
866 pte = pte_mkclean(pte);
867 set_pte_at(vma->vm_mm, address, ptep, pte);
868unlock_pte:
869 pte_unmap_unlock(ptep, ptl);
870 }
871
David Brazdil0f672f62019-12-10 10:32:29 +0000872 mmu_notifier_invalidate_range_end(&range);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000873 }
874 i_mmap_unlock_read(mapping);
875}
876
David Brazdil0f672f62019-12-10 10:32:29 +0000877static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev,
878 struct address_space *mapping, void *entry)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000879{
David Brazdil0f672f62019-12-10 10:32:29 +0000880 unsigned long pfn, index, count;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000881 long ret = 0;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000882
883 /*
884 * A page got tagged dirty in DAX mapping? Something is seriously
885 * wrong.
886 */
David Brazdil0f672f62019-12-10 10:32:29 +0000887 if (WARN_ON(!xa_is_value(entry)))
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000888 return -EIO;
889
David Brazdil0f672f62019-12-10 10:32:29 +0000890 if (unlikely(dax_is_locked(entry))) {
891 void *old_entry = entry;
892
893 entry = get_unlocked_entry(xas, 0);
894
895 /* Entry got punched out / reallocated? */
896 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
897 goto put_unlocked;
898 /*
899 * Entry got reallocated elsewhere? No need to writeback.
900 * We have to compare pfns as we must not bail out due to
901 * difference in lockbit or entry type.
902 */
903 if (dax_to_pfn(old_entry) != dax_to_pfn(entry))
904 goto put_unlocked;
905 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
906 dax_is_zero_entry(entry))) {
907 ret = -EIO;
908 goto put_unlocked;
909 }
910
911 /* Another fsync thread may have already done this entry */
912 if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE))
913 goto put_unlocked;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000914 }
915
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000916 /* Lock the entry to serialize with page faults */
David Brazdil0f672f62019-12-10 10:32:29 +0000917 dax_lock_entry(xas, entry);
918
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000919 /*
920 * We can clear the tag now but we have to be careful so that concurrent
921 * dax_writeback_one() calls for the same index cannot finish before we
922 * actually flush the caches. This is achieved as the calls will look
923 * at the entry only under the i_pages lock and once they do that
924 * they will see the entry locked and wait for it to unlock.
925 */
David Brazdil0f672f62019-12-10 10:32:29 +0000926 xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE);
927 xas_unlock_irq(xas);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000928
929 /*
David Brazdil0f672f62019-12-10 10:32:29 +0000930 * If dax_writeback_mapping_range() was given a wbc->range_start
931 * in the middle of a PMD, the 'index' we use needs to be
932 * aligned to the start of the PMD.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000933 * This allows us to flush for PMD_SIZE and not have to worry about
934 * partial PMD writebacks.
935 */
David Brazdil0f672f62019-12-10 10:32:29 +0000936 pfn = dax_to_pfn(entry);
937 count = 1UL << dax_entry_order(entry);
938 index = xas->xa_index & ~(count - 1);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000939
David Brazdil0f672f62019-12-10 10:32:29 +0000940 dax_entry_mkclean(mapping, index, pfn);
941 dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000942 /*
943 * After we have flushed the cache, we can clear the dirty tag. There
944 * cannot be new dirty data in the pfn after the flush has completed as
945 * the pfn mappings are writeprotected and fault waits for mapping
946 * entry lock.
947 */
David Brazdil0f672f62019-12-10 10:32:29 +0000948 xas_reset(xas);
949 xas_lock_irq(xas);
950 xas_store(xas, entry);
951 xas_clear_mark(xas, PAGECACHE_TAG_DIRTY);
Olivier Deprez157378f2022-04-04 15:47:50 +0200952 dax_wake_entry(xas, entry, WAKE_NEXT);
David Brazdil0f672f62019-12-10 10:32:29 +0000953
954 trace_dax_writeback_one(mapping->host, index, count);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000955 return ret;
956
957 put_unlocked:
Olivier Deprez157378f2022-04-04 15:47:50 +0200958 put_unlocked_entry(xas, entry, WAKE_NEXT);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000959 return ret;
960}
961
962/*
963 * Flush the mapping to the persistent domain within the byte range of [start,
964 * end]. This is required by data integrity operations to ensure file data is
965 * on persistent storage prior to completion of the operation.
966 */
967int dax_writeback_mapping_range(struct address_space *mapping,
Olivier Deprez157378f2022-04-04 15:47:50 +0200968 struct dax_device *dax_dev, struct writeback_control *wbc)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000969{
David Brazdil0f672f62019-12-10 10:32:29 +0000970 XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000971 struct inode *inode = mapping->host;
David Brazdil0f672f62019-12-10 10:32:29 +0000972 pgoff_t end_index = wbc->range_end >> PAGE_SHIFT;
David Brazdil0f672f62019-12-10 10:32:29 +0000973 void *entry;
974 int ret = 0;
975 unsigned int scanned = 0;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000976
977 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
978 return -EIO;
979
980 if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
981 return 0;
982
David Brazdil0f672f62019-12-10 10:32:29 +0000983 trace_dax_writeback_range(inode, xas.xa_index, end_index);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000984
David Brazdil0f672f62019-12-10 10:32:29 +0000985 tag_pages_for_writeback(mapping, xas.xa_index, end_index);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000986
David Brazdil0f672f62019-12-10 10:32:29 +0000987 xas_lock_irq(&xas);
988 xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) {
989 ret = dax_writeback_one(&xas, dax_dev, mapping, entry);
990 if (ret < 0) {
991 mapping_set_error(mapping, ret);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000992 break;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000993 }
David Brazdil0f672f62019-12-10 10:32:29 +0000994 if (++scanned % XA_CHECK_SCHED)
995 continue;
996
997 xas_pause(&xas);
998 xas_unlock_irq(&xas);
999 cond_resched();
1000 xas_lock_irq(&xas);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001001 }
David Brazdil0f672f62019-12-10 10:32:29 +00001002 xas_unlock_irq(&xas);
David Brazdil0f672f62019-12-10 10:32:29 +00001003 trace_dax_writeback_range_done(inode, xas.xa_index, end_index);
1004 return ret;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001005}
1006EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
1007
1008static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
1009{
1010 return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9;
1011}
1012
1013static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size,
1014 pfn_t *pfnp)
1015{
1016 const sector_t sector = dax_iomap_sector(iomap, pos);
1017 pgoff_t pgoff;
1018 int id, rc;
1019 long length;
1020
1021 rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff);
1022 if (rc)
1023 return rc;
1024 id = dax_read_lock();
1025 length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
1026 NULL, pfnp);
1027 if (length < 0) {
1028 rc = length;
1029 goto out;
1030 }
1031 rc = -EINVAL;
1032 if (PFN_PHYS(length) < size)
1033 goto out;
1034 if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1035 goto out;
1036 /* For larger pages we need devmap */
1037 if (length > 1 && !pfn_t_devmap(*pfnp))
1038 goto out;
1039 rc = 0;
1040out:
1041 dax_read_unlock(id);
1042 return rc;
1043}
1044
1045/*
1046 * The user has performed a load from a hole in the file. Allocating a new
1047 * page in the file would cause excessive storage usage for workloads with
1048 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1049 * If this page is ever written to we will re-fault and change the mapping to
1050 * point to real DAX storage instead.
1051 */
David Brazdil0f672f62019-12-10 10:32:29 +00001052static vm_fault_t dax_load_hole(struct xa_state *xas,
1053 struct address_space *mapping, void **entry,
1054 struct vm_fault *vmf)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001055{
1056 struct inode *inode = mapping->host;
1057 unsigned long vaddr = vmf->address;
1058 pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1059 vm_fault_t ret;
1060
David Brazdil0f672f62019-12-10 10:32:29 +00001061 *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1062 DAX_ZERO_PAGE, false);
1063
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001064 ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1065 trace_dax_load_hole(inode, vmf, ret);
1066 return ret;
1067}
1068
Olivier Deprez157378f2022-04-04 15:47:50 +02001069s64 dax_iomap_zero(loff_t pos, u64 length, struct iomap *iomap)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001070{
Olivier Deprez157378f2022-04-04 15:47:50 +02001071 sector_t sector = iomap_sector(iomap, pos & PAGE_MASK);
1072 pgoff_t pgoff;
1073 long rc, id;
1074 void *kaddr;
1075 bool page_aligned = false;
1076 unsigned offset = offset_in_page(pos);
1077 unsigned size = min_t(u64, PAGE_SIZE - offset, length);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001078
Olivier Deprez157378f2022-04-04 15:47:50 +02001079 if (IS_ALIGNED(sector << SECTOR_SHIFT, PAGE_SIZE) &&
1080 (size == PAGE_SIZE))
1081 page_aligned = true;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001082
Olivier Deprez157378f2022-04-04 15:47:50 +02001083 rc = bdev_dax_pgoff(iomap->bdev, sector, PAGE_SIZE, &pgoff);
1084 if (rc)
1085 return rc;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001086
Olivier Deprez157378f2022-04-04 15:47:50 +02001087 id = dax_read_lock();
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001088
Olivier Deprez157378f2022-04-04 15:47:50 +02001089 if (page_aligned)
1090 rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1);
1091 else
1092 rc = dax_direct_access(iomap->dax_dev, pgoff, 1, &kaddr, NULL);
1093 if (rc < 0) {
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001094 dax_read_unlock(id);
Olivier Deprez157378f2022-04-04 15:47:50 +02001095 return rc;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001096 }
Olivier Deprez157378f2022-04-04 15:47:50 +02001097
1098 if (!page_aligned) {
1099 memset(kaddr + offset, 0, size);
1100 dax_flush(iomap->dax_dev, kaddr + offset, size);
1101 }
1102 dax_read_unlock(id);
1103 return size;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001104}
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001105
1106static loff_t
1107dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
Olivier Deprez157378f2022-04-04 15:47:50 +02001108 struct iomap *iomap, struct iomap *srcmap)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001109{
1110 struct block_device *bdev = iomap->bdev;
1111 struct dax_device *dax_dev = iomap->dax_dev;
1112 struct iov_iter *iter = data;
1113 loff_t end = pos + length, done = 0;
1114 ssize_t ret = 0;
1115 size_t xfer;
1116 int id;
1117
1118 if (iov_iter_rw(iter) == READ) {
1119 end = min(end, i_size_read(inode));
1120 if (pos >= end)
1121 return 0;
1122
1123 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1124 return iov_iter_zero(min(length, end - pos), iter);
1125 }
1126
1127 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1128 return -EIO;
1129
1130 /*
1131 * Write can allocate block for an area which has a hole page mapped
1132 * into page tables. We have to tear down these mappings so that data
1133 * written by write(2) is visible in mmap.
1134 */
1135 if (iomap->flags & IOMAP_F_NEW) {
1136 invalidate_inode_pages2_range(inode->i_mapping,
1137 pos >> PAGE_SHIFT,
1138 (end - 1) >> PAGE_SHIFT);
1139 }
1140
1141 id = dax_read_lock();
1142 while (pos < end) {
1143 unsigned offset = pos & (PAGE_SIZE - 1);
1144 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1145 const sector_t sector = dax_iomap_sector(iomap, pos);
1146 ssize_t map_len;
1147 pgoff_t pgoff;
1148 void *kaddr;
1149
1150 if (fatal_signal_pending(current)) {
1151 ret = -EINTR;
1152 break;
1153 }
1154
1155 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1156 if (ret)
1157 break;
1158
1159 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1160 &kaddr, NULL);
1161 if (map_len < 0) {
1162 ret = map_len;
1163 break;
1164 }
1165
1166 map_len = PFN_PHYS(map_len);
1167 kaddr += offset;
1168 map_len -= offset;
1169 if (map_len > end - pos)
1170 map_len = end - pos;
1171
1172 /*
1173 * The userspace address for the memory copy has already been
1174 * validated via access_ok() in either vfs_read() or
1175 * vfs_write(), depending on which operation we are doing.
1176 */
1177 if (iov_iter_rw(iter) == WRITE)
1178 xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1179 map_len, iter);
1180 else
1181 xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1182 map_len, iter);
1183
1184 pos += xfer;
1185 length -= xfer;
1186 done += xfer;
1187
1188 if (xfer == 0)
1189 ret = -EFAULT;
1190 if (xfer < map_len)
1191 break;
1192 }
1193 dax_read_unlock(id);
1194
1195 return done ? done : ret;
1196}
1197
1198/**
1199 * dax_iomap_rw - Perform I/O to a DAX file
1200 * @iocb: The control block for this I/O
1201 * @iter: The addresses to do I/O from or to
1202 * @ops: iomap ops passed from the file system
1203 *
1204 * This function performs read and write operations to directly mapped
1205 * persistent memory. The callers needs to take care of read/write exclusion
1206 * and evicting any page cache pages in the region under I/O.
1207 */
1208ssize_t
1209dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1210 const struct iomap_ops *ops)
1211{
1212 struct address_space *mapping = iocb->ki_filp->f_mapping;
1213 struct inode *inode = mapping->host;
1214 loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1215 unsigned flags = 0;
1216
1217 if (iov_iter_rw(iter) == WRITE) {
David Brazdil0f672f62019-12-10 10:32:29 +00001218 lockdep_assert_held_write(&inode->i_rwsem);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001219 flags |= IOMAP_WRITE;
1220 } else {
1221 lockdep_assert_held(&inode->i_rwsem);
1222 }
1223
Olivier Deprez0e641232021-09-23 10:07:05 +02001224 if (iocb->ki_flags & IOCB_NOWAIT)
1225 flags |= IOMAP_NOWAIT;
1226
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001227 while (iov_iter_count(iter)) {
1228 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1229 iter, dax_iomap_actor);
1230 if (ret <= 0)
1231 break;
1232 pos += ret;
1233 done += ret;
1234 }
1235
1236 iocb->ki_pos += done;
1237 return done ? done : ret;
1238}
1239EXPORT_SYMBOL_GPL(dax_iomap_rw);
1240
1241static vm_fault_t dax_fault_return(int error)
1242{
1243 if (error == 0)
1244 return VM_FAULT_NOPAGE;
David Brazdil0f672f62019-12-10 10:32:29 +00001245 return vmf_error(error);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001246}
1247
1248/*
1249 * MAP_SYNC on a dax mapping guarantees dirty metadata is
1250 * flushed on write-faults (non-cow), but not read-faults.
1251 */
1252static bool dax_fault_is_synchronous(unsigned long flags,
1253 struct vm_area_struct *vma, struct iomap *iomap)
1254{
1255 return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC)
1256 && (iomap->flags & IOMAP_F_DIRTY);
1257}
1258
1259static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1260 int *iomap_errp, const struct iomap_ops *ops)
1261{
1262 struct vm_area_struct *vma = vmf->vma;
1263 struct address_space *mapping = vma->vm_file->f_mapping;
David Brazdil0f672f62019-12-10 10:32:29 +00001264 XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001265 struct inode *inode = mapping->host;
1266 unsigned long vaddr = vmf->address;
1267 loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
Olivier Deprez157378f2022-04-04 15:47:50 +02001268 struct iomap iomap = { .type = IOMAP_HOLE };
1269 struct iomap srcmap = { .type = IOMAP_HOLE };
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001270 unsigned flags = IOMAP_FAULT;
1271 int error, major = 0;
1272 bool write = vmf->flags & FAULT_FLAG_WRITE;
1273 bool sync;
1274 vm_fault_t ret = 0;
1275 void *entry;
1276 pfn_t pfn;
1277
1278 trace_dax_pte_fault(inode, vmf, ret);
1279 /*
1280 * Check whether offset isn't beyond end of file now. Caller is supposed
1281 * to hold locks serializing us with truncate / punch hole so this is
1282 * a reliable test.
1283 */
1284 if (pos >= i_size_read(inode)) {
1285 ret = VM_FAULT_SIGBUS;
1286 goto out;
1287 }
1288
1289 if (write && !vmf->cow_page)
1290 flags |= IOMAP_WRITE;
1291
David Brazdil0f672f62019-12-10 10:32:29 +00001292 entry = grab_mapping_entry(&xas, mapping, 0);
1293 if (xa_is_internal(entry)) {
1294 ret = xa_to_internal(entry);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001295 goto out;
1296 }
1297
1298 /*
1299 * It is possible, particularly with mixed reads & writes to private
1300 * mappings, that we have raced with a PMD fault that overlaps with
1301 * the PTE we need to set up. If so just return and the fault will be
1302 * retried.
1303 */
1304 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1305 ret = VM_FAULT_NOPAGE;
1306 goto unlock_entry;
1307 }
1308
1309 /*
1310 * Note that we don't bother to use iomap_apply here: DAX required
1311 * the file system block size to be equal the page size, which means
1312 * that we never have to deal with more than a single extent here.
1313 */
Olivier Deprez157378f2022-04-04 15:47:50 +02001314 error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap, &srcmap);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001315 if (iomap_errp)
1316 *iomap_errp = error;
1317 if (error) {
1318 ret = dax_fault_return(error);
1319 goto unlock_entry;
1320 }
1321 if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1322 error = -EIO; /* fs corruption? */
1323 goto error_finish_iomap;
1324 }
1325
1326 if (vmf->cow_page) {
1327 sector_t sector = dax_iomap_sector(&iomap, pos);
1328
1329 switch (iomap.type) {
1330 case IOMAP_HOLE:
1331 case IOMAP_UNWRITTEN:
1332 clear_user_highpage(vmf->cow_page, vaddr);
1333 break;
1334 case IOMAP_MAPPED:
Olivier Deprez157378f2022-04-04 15:47:50 +02001335 error = copy_cow_page_dax(iomap.bdev, iomap.dax_dev,
1336 sector, vmf->cow_page, vaddr);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001337 break;
1338 default:
1339 WARN_ON_ONCE(1);
1340 error = -EIO;
1341 break;
1342 }
1343
1344 if (error)
1345 goto error_finish_iomap;
1346
1347 __SetPageUptodate(vmf->cow_page);
1348 ret = finish_fault(vmf);
1349 if (!ret)
1350 ret = VM_FAULT_DONE_COW;
1351 goto finish_iomap;
1352 }
1353
1354 sync = dax_fault_is_synchronous(flags, vma, &iomap);
1355
1356 switch (iomap.type) {
1357 case IOMAP_MAPPED:
1358 if (iomap.flags & IOMAP_F_NEW) {
1359 count_vm_event(PGMAJFAULT);
1360 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
1361 major = VM_FAULT_MAJOR;
1362 }
1363 error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn);
1364 if (error < 0)
1365 goto error_finish_iomap;
1366
David Brazdil0f672f62019-12-10 10:32:29 +00001367 entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001368 0, write && !sync);
1369
1370 /*
1371 * If we are doing synchronous page fault and inode needs fsync,
1372 * we can insert PTE into page tables only after that happens.
1373 * Skip insertion for now and return the pfn so that caller can
1374 * insert it after fsync is done.
1375 */
1376 if (sync) {
1377 if (WARN_ON_ONCE(!pfnp)) {
1378 error = -EIO;
1379 goto error_finish_iomap;
1380 }
1381 *pfnp = pfn;
1382 ret = VM_FAULT_NEEDDSYNC | major;
1383 goto finish_iomap;
1384 }
1385 trace_dax_insert_mapping(inode, vmf, entry);
1386 if (write)
1387 ret = vmf_insert_mixed_mkwrite(vma, vaddr, pfn);
1388 else
1389 ret = vmf_insert_mixed(vma, vaddr, pfn);
1390
1391 goto finish_iomap;
1392 case IOMAP_UNWRITTEN:
1393 case IOMAP_HOLE:
1394 if (!write) {
David Brazdil0f672f62019-12-10 10:32:29 +00001395 ret = dax_load_hole(&xas, mapping, &entry, vmf);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001396 goto finish_iomap;
1397 }
Olivier Deprez157378f2022-04-04 15:47:50 +02001398 fallthrough;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001399 default:
1400 WARN_ON_ONCE(1);
1401 error = -EIO;
1402 break;
1403 }
1404
1405 error_finish_iomap:
1406 ret = dax_fault_return(error);
1407 finish_iomap:
1408 if (ops->iomap_end) {
1409 int copied = PAGE_SIZE;
1410
1411 if (ret & VM_FAULT_ERROR)
1412 copied = 0;
1413 /*
1414 * The fault is done by now and there's no way back (other
1415 * thread may be already happily using PTE we have installed).
1416 * Just ignore error from ->iomap_end since we cannot do much
1417 * with it.
1418 */
1419 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1420 }
1421 unlock_entry:
David Brazdil0f672f62019-12-10 10:32:29 +00001422 dax_unlock_entry(&xas, entry);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001423 out:
1424 trace_dax_pte_fault_done(inode, vmf, ret);
1425 return ret | major;
1426}
1427
1428#ifdef CONFIG_FS_DAX_PMD
David Brazdil0f672f62019-12-10 10:32:29 +00001429static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1430 struct iomap *iomap, void **entry)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001431{
1432 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1433 unsigned long pmd_addr = vmf->address & PMD_MASK;
David Brazdil0f672f62019-12-10 10:32:29 +00001434 struct vm_area_struct *vma = vmf->vma;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001435 struct inode *inode = mapping->host;
David Brazdil0f672f62019-12-10 10:32:29 +00001436 pgtable_t pgtable = NULL;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001437 struct page *zero_page;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001438 spinlock_t *ptl;
1439 pmd_t pmd_entry;
1440 pfn_t pfn;
1441
1442 zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1443
1444 if (unlikely(!zero_page))
1445 goto fallback;
1446
1447 pfn = page_to_pfn_t(zero_page);
David Brazdil0f672f62019-12-10 10:32:29 +00001448 *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1449 DAX_PMD | DAX_ZERO_PAGE, false);
1450
1451 if (arch_needs_pgtable_deposit()) {
1452 pgtable = pte_alloc_one(vma->vm_mm);
1453 if (!pgtable)
1454 return VM_FAULT_OOM;
1455 }
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001456
1457 ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1458 if (!pmd_none(*(vmf->pmd))) {
1459 spin_unlock(ptl);
1460 goto fallback;
1461 }
1462
David Brazdil0f672f62019-12-10 10:32:29 +00001463 if (pgtable) {
1464 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1465 mm_inc_nr_ptes(vma->vm_mm);
1466 }
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001467 pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1468 pmd_entry = pmd_mkhuge(pmd_entry);
1469 set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1470 spin_unlock(ptl);
David Brazdil0f672f62019-12-10 10:32:29 +00001471 trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001472 return VM_FAULT_NOPAGE;
1473
1474fallback:
David Brazdil0f672f62019-12-10 10:32:29 +00001475 if (pgtable)
1476 pte_free(vma->vm_mm, pgtable);
1477 trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001478 return VM_FAULT_FALLBACK;
1479}
1480
1481static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1482 const struct iomap_ops *ops)
1483{
1484 struct vm_area_struct *vma = vmf->vma;
1485 struct address_space *mapping = vma->vm_file->f_mapping;
David Brazdil0f672f62019-12-10 10:32:29 +00001486 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001487 unsigned long pmd_addr = vmf->address & PMD_MASK;
1488 bool write = vmf->flags & FAULT_FLAG_WRITE;
1489 bool sync;
1490 unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1491 struct inode *inode = mapping->host;
1492 vm_fault_t result = VM_FAULT_FALLBACK;
Olivier Deprez157378f2022-04-04 15:47:50 +02001493 struct iomap iomap = { .type = IOMAP_HOLE };
1494 struct iomap srcmap = { .type = IOMAP_HOLE };
David Brazdil0f672f62019-12-10 10:32:29 +00001495 pgoff_t max_pgoff;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001496 void *entry;
1497 loff_t pos;
1498 int error;
1499 pfn_t pfn;
1500
1501 /*
1502 * Check whether offset isn't beyond end of file now. Caller is
1503 * supposed to hold locks serializing us with truncate / punch hole so
1504 * this is a reliable test.
1505 */
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001506 max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
1507
1508 trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1509
1510 /*
1511 * Make sure that the faulting address's PMD offset (color) matches
1512 * the PMD offset from the start of the file. This is necessary so
1513 * that a PMD range in the page table overlaps exactly with a PMD
David Brazdil0f672f62019-12-10 10:32:29 +00001514 * range in the page cache.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001515 */
1516 if ((vmf->pgoff & PG_PMD_COLOUR) !=
1517 ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1518 goto fallback;
1519
1520 /* Fall back to PTEs if we're going to COW */
1521 if (write && !(vma->vm_flags & VM_SHARED))
1522 goto fallback;
1523
1524 /* If the PMD would extend outside the VMA */
1525 if (pmd_addr < vma->vm_start)
1526 goto fallback;
1527 if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1528 goto fallback;
1529
David Brazdil0f672f62019-12-10 10:32:29 +00001530 if (xas.xa_index >= max_pgoff) {
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001531 result = VM_FAULT_SIGBUS;
1532 goto out;
1533 }
1534
1535 /* If the PMD would extend beyond the file size */
David Brazdil0f672f62019-12-10 10:32:29 +00001536 if ((xas.xa_index | PG_PMD_COLOUR) >= max_pgoff)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001537 goto fallback;
1538
1539 /*
David Brazdil0f672f62019-12-10 10:32:29 +00001540 * grab_mapping_entry() will make sure we get an empty PMD entry,
1541 * a zero PMD entry or a DAX PMD. If it can't (because a PTE
1542 * entry is already in the array, for instance), it will return
1543 * VM_FAULT_FALLBACK.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001544 */
David Brazdil0f672f62019-12-10 10:32:29 +00001545 entry = grab_mapping_entry(&xas, mapping, PMD_ORDER);
1546 if (xa_is_internal(entry)) {
1547 result = xa_to_internal(entry);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001548 goto fallback;
David Brazdil0f672f62019-12-10 10:32:29 +00001549 }
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001550
1551 /*
1552 * It is possible, particularly with mixed reads & writes to private
1553 * mappings, that we have raced with a PTE fault that overlaps with
1554 * the PMD we need to set up. If so just return and the fault will be
1555 * retried.
1556 */
1557 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1558 !pmd_devmap(*vmf->pmd)) {
1559 result = 0;
1560 goto unlock_entry;
1561 }
1562
1563 /*
1564 * Note that we don't use iomap_apply here. We aren't doing I/O, only
1565 * setting up a mapping, so really we're using iomap_begin() as a way
1566 * to look up our filesystem block.
1567 */
David Brazdil0f672f62019-12-10 10:32:29 +00001568 pos = (loff_t)xas.xa_index << PAGE_SHIFT;
Olivier Deprez157378f2022-04-04 15:47:50 +02001569 error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap,
1570 &srcmap);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001571 if (error)
1572 goto unlock_entry;
1573
1574 if (iomap.offset + iomap.length < pos + PMD_SIZE)
1575 goto finish_iomap;
1576
1577 sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap);
1578
1579 switch (iomap.type) {
1580 case IOMAP_MAPPED:
1581 error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn);
1582 if (error < 0)
1583 goto finish_iomap;
1584
David Brazdil0f672f62019-12-10 10:32:29 +00001585 entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
1586 DAX_PMD, write && !sync);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001587
1588 /*
1589 * If we are doing synchronous page fault and inode needs fsync,
1590 * we can insert PMD into page tables only after that happens.
1591 * Skip insertion for now and return the pfn so that caller can
1592 * insert it after fsync is done.
1593 */
1594 if (sync) {
1595 if (WARN_ON_ONCE(!pfnp))
1596 goto finish_iomap;
1597 *pfnp = pfn;
1598 result = VM_FAULT_NEEDDSYNC;
1599 goto finish_iomap;
1600 }
1601
1602 trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry);
David Brazdil0f672f62019-12-10 10:32:29 +00001603 result = vmf_insert_pfn_pmd(vmf, pfn, write);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001604 break;
1605 case IOMAP_UNWRITTEN:
1606 case IOMAP_HOLE:
1607 if (WARN_ON_ONCE(write))
1608 break;
David Brazdil0f672f62019-12-10 10:32:29 +00001609 result = dax_pmd_load_hole(&xas, vmf, &iomap, &entry);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001610 break;
1611 default:
1612 WARN_ON_ONCE(1);
1613 break;
1614 }
1615
1616 finish_iomap:
1617 if (ops->iomap_end) {
1618 int copied = PMD_SIZE;
1619
1620 if (result == VM_FAULT_FALLBACK)
1621 copied = 0;
1622 /*
1623 * The fault is done by now and there's no way back (other
1624 * thread may be already happily using PMD we have installed).
1625 * Just ignore error from ->iomap_end since we cannot do much
1626 * with it.
1627 */
1628 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1629 &iomap);
1630 }
1631 unlock_entry:
David Brazdil0f672f62019-12-10 10:32:29 +00001632 dax_unlock_entry(&xas, entry);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001633 fallback:
1634 if (result == VM_FAULT_FALLBACK) {
1635 split_huge_pmd(vma, vmf->pmd, vmf->address);
1636 count_vm_event(THP_FAULT_FALLBACK);
1637 }
1638out:
1639 trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1640 return result;
1641}
1642#else
1643static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1644 const struct iomap_ops *ops)
1645{
1646 return VM_FAULT_FALLBACK;
1647}
1648#endif /* CONFIG_FS_DAX_PMD */
1649
1650/**
1651 * dax_iomap_fault - handle a page fault on a DAX file
1652 * @vmf: The description of the fault
1653 * @pe_size: Size of the page to fault in
1654 * @pfnp: PFN to insert for synchronous faults if fsync is required
1655 * @iomap_errp: Storage for detailed error code in case of error
1656 * @ops: Iomap ops passed from the file system
1657 *
1658 * When a page fault occurs, filesystems may call this helper in
1659 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1660 * has done all the necessary locking for page fault to proceed
1661 * successfully.
1662 */
1663vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1664 pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1665{
1666 switch (pe_size) {
1667 case PE_SIZE_PTE:
1668 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1669 case PE_SIZE_PMD:
1670 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1671 default:
1672 return VM_FAULT_FALLBACK;
1673 }
1674}
1675EXPORT_SYMBOL_GPL(dax_iomap_fault);
1676
David Brazdil0f672f62019-12-10 10:32:29 +00001677/*
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001678 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1679 * @vmf: The description of the fault
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001680 * @pfn: PFN to insert
David Brazdil0f672f62019-12-10 10:32:29 +00001681 * @order: Order of entry to insert.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001682 *
David Brazdil0f672f62019-12-10 10:32:29 +00001683 * This function inserts a writeable PTE or PMD entry into the page tables
1684 * for an mmaped DAX file. It also marks the page cache entry as dirty.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001685 */
David Brazdil0f672f62019-12-10 10:32:29 +00001686static vm_fault_t
1687dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001688{
1689 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
David Brazdil0f672f62019-12-10 10:32:29 +00001690 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1691 void *entry;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001692 vm_fault_t ret;
1693
David Brazdil0f672f62019-12-10 10:32:29 +00001694 xas_lock_irq(&xas);
1695 entry = get_unlocked_entry(&xas, order);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001696 /* Did we race with someone splitting entry or so? */
David Brazdil0f672f62019-12-10 10:32:29 +00001697 if (!entry || dax_is_conflict(entry) ||
1698 (order == 0 && !dax_is_pte_entry(entry))) {
Olivier Deprez157378f2022-04-04 15:47:50 +02001699 put_unlocked_entry(&xas, entry, WAKE_NEXT);
David Brazdil0f672f62019-12-10 10:32:29 +00001700 xas_unlock_irq(&xas);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001701 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1702 VM_FAULT_NOPAGE);
1703 return VM_FAULT_NOPAGE;
1704 }
David Brazdil0f672f62019-12-10 10:32:29 +00001705 xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1706 dax_lock_entry(&xas, entry);
1707 xas_unlock_irq(&xas);
1708 if (order == 0)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001709 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001710#ifdef CONFIG_FS_DAX_PMD
David Brazdil0f672f62019-12-10 10:32:29 +00001711 else if (order == PMD_ORDER)
1712 ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001713#endif
David Brazdil0f672f62019-12-10 10:32:29 +00001714 else
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001715 ret = VM_FAULT_FALLBACK;
David Brazdil0f672f62019-12-10 10:32:29 +00001716 dax_unlock_entry(&xas, entry);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001717 trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1718 return ret;
1719}
1720
1721/**
1722 * dax_finish_sync_fault - finish synchronous page fault
1723 * @vmf: The description of the fault
1724 * @pe_size: Size of entry to be inserted
1725 * @pfn: PFN to insert
1726 *
1727 * This function ensures that the file range touched by the page fault is
1728 * stored persistently on the media and handles inserting of appropriate page
1729 * table entry.
1730 */
1731vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf,
1732 enum page_entry_size pe_size, pfn_t pfn)
1733{
1734 int err;
1735 loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
David Brazdil0f672f62019-12-10 10:32:29 +00001736 unsigned int order = pe_order(pe_size);
1737 size_t len = PAGE_SIZE << order;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001738
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001739 err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1740 if (err)
1741 return VM_FAULT_SIGBUS;
David Brazdil0f672f62019-12-10 10:32:29 +00001742 return dax_insert_pfn_mkwrite(vmf, pfn, order);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001743}
1744EXPORT_SYMBOL_GPL(dax_finish_sync_fault);