1 // SPDX-License-Identifier: GPL-2.0-only 1 // SPDX-License-Identifier: GPL-2.0-only
2 /* 2 /*
3 * fs/dax.c - Direct Access filesystem code 3 * fs/dax.c - Direct Access filesystem code
4 * Copyright (c) 2013-2014 Intel Corporation 4 * Copyright (c) 2013-2014 Intel Corporation
5 * Author: Matthew Wilcox <matthew.r.wilcox@in 5 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
6 * Author: Ross Zwisler <ross.zwisler@linux.in 6 * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
7 */ 7 */
8 8
9 #include <linux/atomic.h> 9 #include <linux/atomic.h>
10 #include <linux/blkdev.h> 10 #include <linux/blkdev.h>
11 #include <linux/buffer_head.h> 11 #include <linux/buffer_head.h>
12 #include <linux/dax.h> 12 #include <linux/dax.h>
13 #include <linux/fs.h> 13 #include <linux/fs.h>
14 #include <linux/highmem.h> 14 #include <linux/highmem.h>
15 #include <linux/memcontrol.h> 15 #include <linux/memcontrol.h>
16 #include <linux/mm.h> 16 #include <linux/mm.h>
17 #include <linux/mutex.h> 17 #include <linux/mutex.h>
18 #include <linux/pagevec.h> 18 #include <linux/pagevec.h>
19 #include <linux/sched.h> 19 #include <linux/sched.h>
20 #include <linux/sched/signal.h> 20 #include <linux/sched/signal.h>
21 #include <linux/uio.h> 21 #include <linux/uio.h>
22 #include <linux/vmstat.h> 22 #include <linux/vmstat.h>
23 #include <linux/pfn_t.h> 23 #include <linux/pfn_t.h>
24 #include <linux/sizes.h> 24 #include <linux/sizes.h>
25 #include <linux/mmu_notifier.h> 25 #include <linux/mmu_notifier.h>
26 #include <linux/iomap.h> 26 #include <linux/iomap.h>
27 #include <linux/rmap.h> 27 #include <linux/rmap.h>
28 #include <asm/pgalloc.h> 28 #include <asm/pgalloc.h>
29 29
30 #define CREATE_TRACE_POINTS 30 #define CREATE_TRACE_POINTS
31 #include <trace/events/fs_dax.h> 31 #include <trace/events/fs_dax.h>
32 32
>> 33 static inline unsigned int pe_order(enum page_entry_size pe_size)
>> 34 {
>> 35 if (pe_size == PE_SIZE_PTE)
>> 36 return PAGE_SHIFT - PAGE_SHIFT;
>> 37 if (pe_size == PE_SIZE_PMD)
>> 38 return PMD_SHIFT - PAGE_SHIFT;
>> 39 if (pe_size == PE_SIZE_PUD)
>> 40 return PUD_SHIFT - PAGE_SHIFT;
>> 41 return ~0;
>> 42 }
>> 43
33 /* We choose 4096 entries - same as per-zone p 44 /* We choose 4096 entries - same as per-zone page wait tables */
34 #define DAX_WAIT_TABLE_BITS 12 45 #define DAX_WAIT_TABLE_BITS 12
35 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_ 46 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
36 47
37 /* The 'colour' (ie low bits) within a PMD of 48 /* The 'colour' (ie low bits) within a PMD of a page offset. */
38 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHI 49 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
39 #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIF 50 #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT)
40 51
>> 52 /* The order of a PMD entry */
>> 53 #define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT)
>> 54
41 static wait_queue_head_t wait_table[DAX_WAIT_T 55 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
42 56
43 static int __init init_dax_wait_table(void) 57 static int __init init_dax_wait_table(void)
44 { 58 {
45 int i; 59 int i;
46 60
47 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES 61 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
48 init_waitqueue_head(wait_table 62 init_waitqueue_head(wait_table + i);
49 return 0; 63 return 0;
50 } 64 }
51 fs_initcall(init_dax_wait_table); 65 fs_initcall(init_dax_wait_table);
52 66
53 /* 67 /*
54 * DAX pagecache entries use XArray value entr 68 * DAX pagecache entries use XArray value entries so they can't be mistaken
55 * for pages. We use one bit for locking, one 69 * for pages. We use one bit for locking, one bit for the entry size (PMD)
56 * and two more to tell us if the entry is a z 70 * and two more to tell us if the entry is a zero page or an empty entry that
57 * is just used for locking. In total four sp 71 * is just used for locking. In total four special bits.
58 * 72 *
59 * If the PMD bit isn't set the entry has size 73 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
60 * and EMPTY bits aren't set the entry is a no 74 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
61 * block allocation. 75 * block allocation.
62 */ 76 */
63 #define DAX_SHIFT (4) 77 #define DAX_SHIFT (4)
64 #define DAX_LOCKED (1UL << 0) 78 #define DAX_LOCKED (1UL << 0)
65 #define DAX_PMD (1UL << 1) 79 #define DAX_PMD (1UL << 1)
66 #define DAX_ZERO_PAGE (1UL << 2) 80 #define DAX_ZERO_PAGE (1UL << 2)
67 #define DAX_EMPTY (1UL << 3) 81 #define DAX_EMPTY (1UL << 3)
68 82
69 static unsigned long dax_to_pfn(void *entry) 83 static unsigned long dax_to_pfn(void *entry)
70 { 84 {
71 return xa_to_value(entry) >> DAX_SHIFT 85 return xa_to_value(entry) >> DAX_SHIFT;
72 } 86 }
73 87
74 static void *dax_make_entry(pfn_t pfn, unsigne 88 static void *dax_make_entry(pfn_t pfn, unsigned long flags)
75 { 89 {
76 return xa_mk_value(flags | (pfn_t_to_p 90 return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
77 } 91 }
78 92
79 static bool dax_is_locked(void *entry) 93 static bool dax_is_locked(void *entry)
80 { 94 {
81 return xa_to_value(entry) & DAX_LOCKED 95 return xa_to_value(entry) & DAX_LOCKED;
82 } 96 }
83 97
84 static unsigned int dax_entry_order(void *entr 98 static unsigned int dax_entry_order(void *entry)
85 { 99 {
86 if (xa_to_value(entry) & DAX_PMD) 100 if (xa_to_value(entry) & DAX_PMD)
87 return PMD_ORDER; 101 return PMD_ORDER;
88 return 0; 102 return 0;
89 } 103 }
90 104
91 static unsigned long dax_is_pmd_entry(void *en 105 static unsigned long dax_is_pmd_entry(void *entry)
92 { 106 {
93 return xa_to_value(entry) & DAX_PMD; 107 return xa_to_value(entry) & DAX_PMD;
94 } 108 }
95 109
96 static bool dax_is_pte_entry(void *entry) 110 static bool dax_is_pte_entry(void *entry)
97 { 111 {
98 return !(xa_to_value(entry) & DAX_PMD) 112 return !(xa_to_value(entry) & DAX_PMD);
99 } 113 }
100 114
101 static int dax_is_zero_entry(void *entry) 115 static int dax_is_zero_entry(void *entry)
102 { 116 {
103 return xa_to_value(entry) & DAX_ZERO_P 117 return xa_to_value(entry) & DAX_ZERO_PAGE;
104 } 118 }
105 119
106 static int dax_is_empty_entry(void *entry) 120 static int dax_is_empty_entry(void *entry)
107 { 121 {
108 return xa_to_value(entry) & DAX_EMPTY; 122 return xa_to_value(entry) & DAX_EMPTY;
109 } 123 }
110 124
111 /* 125 /*
112 * true if the entry that was found is of a sm 126 * true if the entry that was found is of a smaller order than the entry
113 * we were looking for 127 * we were looking for
114 */ 128 */
115 static bool dax_is_conflict(void *entry) 129 static bool dax_is_conflict(void *entry)
116 { 130 {
117 return entry == XA_RETRY_ENTRY; 131 return entry == XA_RETRY_ENTRY;
118 } 132 }
119 133
120 /* 134 /*
121 * DAX page cache entry locking 135 * DAX page cache entry locking
122 */ 136 */
123 struct exceptional_entry_key { 137 struct exceptional_entry_key {
124 struct xarray *xa; 138 struct xarray *xa;
125 pgoff_t entry_start; 139 pgoff_t entry_start;
126 }; 140 };
127 141
128 struct wait_exceptional_entry_queue { 142 struct wait_exceptional_entry_queue {
129 wait_queue_entry_t wait; 143 wait_queue_entry_t wait;
130 struct exceptional_entry_key key; 144 struct exceptional_entry_key key;
131 }; 145 };
132 146
133 /** 147 /**
134 * enum dax_wake_mode: waitqueue wakeup behavi 148 * enum dax_wake_mode: waitqueue wakeup behaviour
135 * @WAKE_ALL: wake all waiters in the waitqueu 149 * @WAKE_ALL: wake all waiters in the waitqueue
136 * @WAKE_NEXT: wake only the first waiter in t 150 * @WAKE_NEXT: wake only the first waiter in the waitqueue
137 */ 151 */
138 enum dax_wake_mode { 152 enum dax_wake_mode {
139 WAKE_ALL, 153 WAKE_ALL,
140 WAKE_NEXT, 154 WAKE_NEXT,
141 }; 155 };
142 156
143 static wait_queue_head_t *dax_entry_waitqueue( 157 static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas,
144 void *entry, struct exceptiona 158 void *entry, struct exceptional_entry_key *key)
145 { 159 {
146 unsigned long hash; 160 unsigned long hash;
147 unsigned long index = xas->xa_index; 161 unsigned long index = xas->xa_index;
148 162
149 /* 163 /*
150 * If 'entry' is a PMD, align the 'ind 164 * If 'entry' is a PMD, align the 'index' that we use for the wait
151 * queue to the start of that PMD. Th 165 * queue to the start of that PMD. This ensures that all offsets in
152 * the range covered by the PMD map to 166 * the range covered by the PMD map to the same bit lock.
153 */ 167 */
154 if (dax_is_pmd_entry(entry)) 168 if (dax_is_pmd_entry(entry))
155 index &= ~PG_PMD_COLOUR; 169 index &= ~PG_PMD_COLOUR;
156 key->xa = xas->xa; 170 key->xa = xas->xa;
157 key->entry_start = index; 171 key->entry_start = index;
158 172
159 hash = hash_long((unsigned long)xas->x 173 hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS);
160 return wait_table + hash; 174 return wait_table + hash;
161 } 175 }
162 176
163 static int wake_exceptional_entry_func(wait_qu 177 static int wake_exceptional_entry_func(wait_queue_entry_t *wait,
164 unsigned int mode, int sync, v 178 unsigned int mode, int sync, void *keyp)
165 { 179 {
166 struct exceptional_entry_key *key = ke 180 struct exceptional_entry_key *key = keyp;
167 struct wait_exceptional_entry_queue *e 181 struct wait_exceptional_entry_queue *ewait =
168 container_of(wait, struct wait 182 container_of(wait, struct wait_exceptional_entry_queue, wait);
169 183
170 if (key->xa != ewait->key.xa || 184 if (key->xa != ewait->key.xa ||
171 key->entry_start != ewait->key.ent 185 key->entry_start != ewait->key.entry_start)
172 return 0; 186 return 0;
173 return autoremove_wake_function(wait, 187 return autoremove_wake_function(wait, mode, sync, NULL);
174 } 188 }
175 189
176 /* 190 /*
177 * @entry may no longer be the entry at the in 191 * @entry may no longer be the entry at the index in the mapping.
178 * The important information it's conveying is 192 * The important information it's conveying is whether the entry at
179 * this index used to be a PMD entry. 193 * this index used to be a PMD entry.
180 */ 194 */
181 static void dax_wake_entry(struct xa_state *xa 195 static void dax_wake_entry(struct xa_state *xas, void *entry,
182 enum dax_wake_mode 196 enum dax_wake_mode mode)
183 { 197 {
184 struct exceptional_entry_key key; 198 struct exceptional_entry_key key;
185 wait_queue_head_t *wq; 199 wait_queue_head_t *wq;
186 200
187 wq = dax_entry_waitqueue(xas, entry, & 201 wq = dax_entry_waitqueue(xas, entry, &key);
188 202
189 /* 203 /*
190 * Checking for locked entry and prepa 204 * Checking for locked entry and prepare_to_wait_exclusive() happens
191 * under the i_pages lock, ditto for e 205 * under the i_pages lock, ditto for entry handling in our callers.
192 * So at this point all tasks that cou 206 * So at this point all tasks that could have seen our entry locked
193 * must be in the waitqueue and the fo 207 * must be in the waitqueue and the following check will see them.
194 */ 208 */
195 if (waitqueue_active(wq)) 209 if (waitqueue_active(wq))
196 __wake_up(wq, TASK_NORMAL, mod 210 __wake_up(wq, TASK_NORMAL, mode == WAKE_ALL ? 0 : 1, &key);
197 } 211 }
198 212
199 /* 213 /*
200 * Look up entry in page cache, wait for it to 214 * Look up entry in page cache, wait for it to become unlocked if it
201 * is a DAX entry and return it. The caller m 215 * is a DAX entry and return it. The caller must subsequently call
202 * put_unlocked_entry() if it did not lock the 216 * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
203 * if it did. The entry returned may have a l 217 * if it did. The entry returned may have a larger order than @order.
204 * If @order is larger than the order of the e 218 * If @order is larger than the order of the entry found in i_pages, this
205 * function returns a dax_is_conflict entry. 219 * function returns a dax_is_conflict entry.
206 * 220 *
207 * Must be called with the i_pages lock held. 221 * Must be called with the i_pages lock held.
208 */ 222 */
209 static void *get_unlocked_entry(struct xa_stat 223 static void *get_unlocked_entry(struct xa_state *xas, unsigned int order)
210 { 224 {
211 void *entry; 225 void *entry;
212 struct wait_exceptional_entry_queue ew 226 struct wait_exceptional_entry_queue ewait;
213 wait_queue_head_t *wq; 227 wait_queue_head_t *wq;
214 228
215 init_wait(&ewait.wait); 229 init_wait(&ewait.wait);
216 ewait.wait.func = wake_exceptional_ent 230 ewait.wait.func = wake_exceptional_entry_func;
217 231
218 for (;;) { 232 for (;;) {
219 entry = xas_find_conflict(xas) 233 entry = xas_find_conflict(xas);
220 if (!entry || WARN_ON_ONCE(!xa 234 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
221 return entry; 235 return entry;
222 if (dax_entry_order(entry) < o 236 if (dax_entry_order(entry) < order)
223 return XA_RETRY_ENTRY; 237 return XA_RETRY_ENTRY;
224 if (!dax_is_locked(entry)) 238 if (!dax_is_locked(entry))
225 return entry; 239 return entry;
226 240
227 wq = dax_entry_waitqueue(xas, 241 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
228 prepare_to_wait_exclusive(wq, 242 prepare_to_wait_exclusive(wq, &ewait.wait,
229 TASK 243 TASK_UNINTERRUPTIBLE);
230 xas_unlock_irq(xas); 244 xas_unlock_irq(xas);
231 xas_reset(xas); 245 xas_reset(xas);
232 schedule(); 246 schedule();
233 finish_wait(wq, &ewait.wait); 247 finish_wait(wq, &ewait.wait);
234 xas_lock_irq(xas); 248 xas_lock_irq(xas);
235 } 249 }
236 } 250 }
237 251
238 /* 252 /*
239 * The only thing keeping the address space ar 253 * The only thing keeping the address space around is the i_pages lock
240 * (it's cycled in clear_inode() after removin 254 * (it's cycled in clear_inode() after removing the entries from i_pages)
241 * After we call xas_unlock_irq(), we cannot t 255 * After we call xas_unlock_irq(), we cannot touch xas->xa.
242 */ 256 */
243 static void wait_entry_unlocked(struct xa_stat 257 static void wait_entry_unlocked(struct xa_state *xas, void *entry)
244 { 258 {
245 struct wait_exceptional_entry_queue ew 259 struct wait_exceptional_entry_queue ewait;
246 wait_queue_head_t *wq; 260 wait_queue_head_t *wq;
247 261
248 init_wait(&ewait.wait); 262 init_wait(&ewait.wait);
249 ewait.wait.func = wake_exceptional_ent 263 ewait.wait.func = wake_exceptional_entry_func;
250 264
251 wq = dax_entry_waitqueue(xas, entry, & 265 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
252 /* 266 /*
253 * Unlike get_unlocked_entry() there i 267 * Unlike get_unlocked_entry() there is no guarantee that this
254 * path ever successfully retrieves an 268 * path ever successfully retrieves an unlocked entry before an
255 * inode dies. Perform a non-exclusive 269 * inode dies. Perform a non-exclusive wait in case this path
256 * never successfully performs its own 270 * never successfully performs its own wake up.
257 */ 271 */
258 prepare_to_wait(wq, &ewait.wait, TASK_ 272 prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
259 xas_unlock_irq(xas); 273 xas_unlock_irq(xas);
260 schedule(); 274 schedule();
261 finish_wait(wq, &ewait.wait); 275 finish_wait(wq, &ewait.wait);
262 } 276 }
263 277
264 static void put_unlocked_entry(struct xa_state 278 static void put_unlocked_entry(struct xa_state *xas, void *entry,
265 enum dax_wake_m 279 enum dax_wake_mode mode)
266 { 280 {
267 if (entry && !dax_is_conflict(entry)) 281 if (entry && !dax_is_conflict(entry))
268 dax_wake_entry(xas, entry, mod 282 dax_wake_entry(xas, entry, mode);
269 } 283 }
270 284
271 /* 285 /*
272 * We used the xa_state to get the entry, but 286 * We used the xa_state to get the entry, but then we locked the entry and
273 * dropped the xa_lock, so we know the xa_stat 287 * dropped the xa_lock, so we know the xa_state is stale and must be reset
274 * before use. 288 * before use.
275 */ 289 */
276 static void dax_unlock_entry(struct xa_state * 290 static void dax_unlock_entry(struct xa_state *xas, void *entry)
277 { 291 {
278 void *old; 292 void *old;
279 293
280 BUG_ON(dax_is_locked(entry)); 294 BUG_ON(dax_is_locked(entry));
281 xas_reset(xas); 295 xas_reset(xas);
282 xas_lock_irq(xas); 296 xas_lock_irq(xas);
283 old = xas_store(xas, entry); 297 old = xas_store(xas, entry);
284 xas_unlock_irq(xas); 298 xas_unlock_irq(xas);
285 BUG_ON(!dax_is_locked(old)); 299 BUG_ON(!dax_is_locked(old));
286 dax_wake_entry(xas, entry, WAKE_NEXT); 300 dax_wake_entry(xas, entry, WAKE_NEXT);
287 } 301 }
288 302
289 /* 303 /*
290 * Return: The entry stored at this location b 304 * Return: The entry stored at this location before it was locked.
291 */ 305 */
292 static void *dax_lock_entry(struct xa_state *x 306 static void *dax_lock_entry(struct xa_state *xas, void *entry)
293 { 307 {
294 unsigned long v = xa_to_value(entry); 308 unsigned long v = xa_to_value(entry);
295 return xas_store(xas, xa_mk_value(v | 309 return xas_store(xas, xa_mk_value(v | DAX_LOCKED));
296 } 310 }
297 311
298 static unsigned long dax_entry_size(void *entr 312 static unsigned long dax_entry_size(void *entry)
299 { 313 {
300 if (dax_is_zero_entry(entry)) 314 if (dax_is_zero_entry(entry))
301 return 0; 315 return 0;
302 else if (dax_is_empty_entry(entry)) 316 else if (dax_is_empty_entry(entry))
303 return 0; 317 return 0;
304 else if (dax_is_pmd_entry(entry)) 318 else if (dax_is_pmd_entry(entry))
305 return PMD_SIZE; 319 return PMD_SIZE;
306 else 320 else
307 return PAGE_SIZE; 321 return PAGE_SIZE;
308 } 322 }
309 323
310 static unsigned long dax_end_pfn(void *entry) 324 static unsigned long dax_end_pfn(void *entry)
311 { 325 {
312 return dax_to_pfn(entry) + dax_entry_s 326 return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
313 } 327 }
314 328
315 /* 329 /*
316 * Iterate through all mapped pfns represented 330 * Iterate through all mapped pfns represented by an entry, i.e. skip
317 * 'empty' and 'zero' entries. 331 * 'empty' and 'zero' entries.
318 */ 332 */
319 #define for_each_mapped_pfn(entry, pfn) \ 333 #define for_each_mapped_pfn(entry, pfn) \
320 for (pfn = dax_to_pfn(entry); \ 334 for (pfn = dax_to_pfn(entry); \
321 pfn < dax_end_pfn(entr 335 pfn < dax_end_pfn(entry); pfn++)
322 336
323 static inline bool dax_page_is_shared(struct p 337 static inline bool dax_page_is_shared(struct page *page)
324 { 338 {
325 return page->mapping == PAGE_MAPPING_D 339 return page->mapping == PAGE_MAPPING_DAX_SHARED;
326 } 340 }
327 341
328 /* 342 /*
329 * Set the page->mapping with PAGE_MAPPING_DAX 343 * Set the page->mapping with PAGE_MAPPING_DAX_SHARED flag, increase the
330 * refcount. 344 * refcount.
331 */ 345 */
332 static inline void dax_page_share_get(struct p 346 static inline void dax_page_share_get(struct page *page)
333 { 347 {
334 if (page->mapping != PAGE_MAPPING_DAX_ 348 if (page->mapping != PAGE_MAPPING_DAX_SHARED) {
335 /* 349 /*
336 * Reset the index if the page 350 * Reset the index if the page was already mapped
337 * regularly before. 351 * regularly before.
338 */ 352 */
339 if (page->mapping) 353 if (page->mapping)
340 page->share = 1; 354 page->share = 1;
341 page->mapping = PAGE_MAPPING_D 355 page->mapping = PAGE_MAPPING_DAX_SHARED;
342 } 356 }
343 page->share++; 357 page->share++;
344 } 358 }
345 359
346 static inline unsigned long dax_page_share_put 360 static inline unsigned long dax_page_share_put(struct page *page)
347 { 361 {
348 return --page->share; 362 return --page->share;
349 } 363 }
350 364
351 /* 365 /*
352 * When it is called in dax_insert_entry(), th 366 * When it is called in dax_insert_entry(), the shared flag will indicate that
353 * whether this entry is shared by multiple fi 367 * whether this entry is shared by multiple files. If so, set the page->mapping
354 * PAGE_MAPPING_DAX_SHARED, and use page->shar 368 * PAGE_MAPPING_DAX_SHARED, and use page->share as refcount.
355 */ 369 */
356 static void dax_associate_entry(void *entry, s 370 static void dax_associate_entry(void *entry, struct address_space *mapping,
357 struct vm_area_struct *vma, un 371 struct vm_area_struct *vma, unsigned long address, bool shared)
358 { 372 {
359 unsigned long size = dax_entry_size(en 373 unsigned long size = dax_entry_size(entry), pfn, index;
360 int i = 0; 374 int i = 0;
361 375
362 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 376 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
363 return; 377 return;
364 378
365 index = linear_page_index(vma, address 379 index = linear_page_index(vma, address & ~(size - 1));
366 for_each_mapped_pfn(entry, pfn) { 380 for_each_mapped_pfn(entry, pfn) {
367 struct page *page = pfn_to_pag 381 struct page *page = pfn_to_page(pfn);
368 382
369 if (shared) { 383 if (shared) {
370 dax_page_share_get(pag 384 dax_page_share_get(page);
371 } else { 385 } else {
372 WARN_ON_ONCE(page->map 386 WARN_ON_ONCE(page->mapping);
373 page->mapping = mappin 387 page->mapping = mapping;
374 page->index = index + 388 page->index = index + i++;
375 } 389 }
376 } 390 }
377 } 391 }
378 392
379 static void dax_disassociate_entry(void *entry 393 static void dax_disassociate_entry(void *entry, struct address_space *mapping,
380 bool trunc) 394 bool trunc)
381 { 395 {
382 unsigned long pfn; 396 unsigned long pfn;
383 397
384 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 398 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
385 return; 399 return;
386 400
387 for_each_mapped_pfn(entry, pfn) { 401 for_each_mapped_pfn(entry, pfn) {
388 struct page *page = pfn_to_pag 402 struct page *page = pfn_to_page(pfn);
389 403
390 WARN_ON_ONCE(trunc && page_ref 404 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
391 if (dax_page_is_shared(page)) 405 if (dax_page_is_shared(page)) {
392 /* keep the shared fla 406 /* keep the shared flag if this page is still shared */
393 if (dax_page_share_put 407 if (dax_page_share_put(page) > 0)
394 continue; 408 continue;
395 } else 409 } else
396 WARN_ON_ONCE(page->map 410 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
397 page->mapping = NULL; 411 page->mapping = NULL;
398 page->index = 0; 412 page->index = 0;
399 } 413 }
400 } 414 }
401 415
402 static struct page *dax_busy_page(void *entry) 416 static struct page *dax_busy_page(void *entry)
403 { 417 {
404 unsigned long pfn; 418 unsigned long pfn;
405 419
406 for_each_mapped_pfn(entry, pfn) { 420 for_each_mapped_pfn(entry, pfn) {
407 struct page *page = pfn_to_pag 421 struct page *page = pfn_to_page(pfn);
408 422
409 if (page_ref_count(page) > 1) 423 if (page_ref_count(page) > 1)
410 return page; 424 return page;
411 } 425 }
412 return NULL; 426 return NULL;
413 } 427 }
414 428
415 /** !! 429 /*
416 * dax_lock_folio - Lock the DAX entry corresp !! 430 * dax_lock_page - Lock the DAX entry corresponding to a page
417 * @folio: The folio whose entry we want to lo !! 431 * @page: The page whose entry we want to lock
418 * 432 *
419 * Context: Process context. 433 * Context: Process context.
420 * Return: A cookie to pass to dax_unlock_foli !! 434 * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could
421 * not be locked. 435 * not be locked.
422 */ 436 */
423 dax_entry_t dax_lock_folio(struct folio *folio !! 437 dax_entry_t dax_lock_page(struct page *page)
424 { 438 {
425 XA_STATE(xas, NULL, 0); 439 XA_STATE(xas, NULL, 0);
426 void *entry; 440 void *entry;
427 441
428 /* Ensure folio->mapping isn't freed w !! 442 /* Ensure page->mapping isn't freed while we look at it */
429 rcu_read_lock(); 443 rcu_read_lock();
430 for (;;) { 444 for (;;) {
431 struct address_space *mapping !! 445 struct address_space *mapping = READ_ONCE(page->mapping);
432 446
433 entry = NULL; 447 entry = NULL;
434 if (!mapping || !dax_mapping(m 448 if (!mapping || !dax_mapping(mapping))
435 break; 449 break;
436 450
437 /* 451 /*
438 * In the device-dax case ther 452 * In the device-dax case there's no need to lock, a
439 * struct dev_pagemap pin is s 453 * struct dev_pagemap pin is sufficient to keep the
440 * inode alive, and we assume 454 * inode alive, and we assume we have dev_pagemap pin
441 * otherwise we would not have 455 * otherwise we would not have a valid pfn_to_page()
442 * translation. 456 * translation.
443 */ 457 */
444 entry = (void *)~0UL; 458 entry = (void *)~0UL;
445 if (S_ISCHR(mapping->host->i_m 459 if (S_ISCHR(mapping->host->i_mode))
446 break; 460 break;
447 461
448 xas.xa = &mapping->i_pages; 462 xas.xa = &mapping->i_pages;
449 xas_lock_irq(&xas); 463 xas_lock_irq(&xas);
450 if (mapping != folio->mapping) !! 464 if (mapping != page->mapping) {
451 xas_unlock_irq(&xas); 465 xas_unlock_irq(&xas);
452 continue; 466 continue;
453 } 467 }
454 xas_set(&xas, folio->index); !! 468 xas_set(&xas, page->index);
455 entry = xas_load(&xas); 469 entry = xas_load(&xas);
456 if (dax_is_locked(entry)) { 470 if (dax_is_locked(entry)) {
457 rcu_read_unlock(); 471 rcu_read_unlock();
458 wait_entry_unlocked(&x 472 wait_entry_unlocked(&xas, entry);
459 rcu_read_lock(); 473 rcu_read_lock();
460 continue; 474 continue;
461 } 475 }
462 dax_lock_entry(&xas, entry); 476 dax_lock_entry(&xas, entry);
463 xas_unlock_irq(&xas); 477 xas_unlock_irq(&xas);
464 break; 478 break;
465 } 479 }
466 rcu_read_unlock(); 480 rcu_read_unlock();
467 return (dax_entry_t)entry; 481 return (dax_entry_t)entry;
468 } 482 }
469 483
470 void dax_unlock_folio(struct folio *folio, dax !! 484 void dax_unlock_page(struct page *page, dax_entry_t cookie)
471 { 485 {
472 struct address_space *mapping = folio- !! 486 struct address_space *mapping = page->mapping;
473 XA_STATE(xas, &mapping->i_pages, folio !! 487 XA_STATE(xas, &mapping->i_pages, page->index);
474 488
475 if (S_ISCHR(mapping->host->i_mode)) 489 if (S_ISCHR(mapping->host->i_mode))
476 return; 490 return;
477 491
478 dax_unlock_entry(&xas, (void *)cookie) 492 dax_unlock_entry(&xas, (void *)cookie);
479 } 493 }
480 494
481 /* 495 /*
482 * dax_lock_mapping_entry - Lock the DAX entry 496 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a mapping
483 * @mapping: the file's mapping whose entry we 497 * @mapping: the file's mapping whose entry we want to lock
484 * @index: the offset within this file 498 * @index: the offset within this file
485 * @page: output the dax page corresponding to 499 * @page: output the dax page corresponding to this dax entry
486 * 500 *
487 * Return: A cookie to pass to dax_unlock_mapp 501 * Return: A cookie to pass to dax_unlock_mapping_entry() or 0 if the entry
488 * could not be locked. 502 * could not be locked.
489 */ 503 */
490 dax_entry_t dax_lock_mapping_entry(struct addr 504 dax_entry_t dax_lock_mapping_entry(struct address_space *mapping, pgoff_t index,
491 struct page **page) 505 struct page **page)
492 { 506 {
493 XA_STATE(xas, NULL, 0); 507 XA_STATE(xas, NULL, 0);
494 void *entry; 508 void *entry;
495 509
496 rcu_read_lock(); 510 rcu_read_lock();
497 for (;;) { 511 for (;;) {
498 entry = NULL; 512 entry = NULL;
499 if (!dax_mapping(mapping)) 513 if (!dax_mapping(mapping))
500 break; 514 break;
501 515
502 xas.xa = &mapping->i_pages; 516 xas.xa = &mapping->i_pages;
503 xas_lock_irq(&xas); 517 xas_lock_irq(&xas);
504 xas_set(&xas, index); 518 xas_set(&xas, index);
505 entry = xas_load(&xas); 519 entry = xas_load(&xas);
506 if (dax_is_locked(entry)) { 520 if (dax_is_locked(entry)) {
507 rcu_read_unlock(); 521 rcu_read_unlock();
508 wait_entry_unlocked(&x 522 wait_entry_unlocked(&xas, entry);
509 rcu_read_lock(); 523 rcu_read_lock();
510 continue; 524 continue;
511 } 525 }
512 if (!entry || 526 if (!entry ||
513 dax_is_zero_entry(entry) | 527 dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
514 /* 528 /*
515 * Because we are look 529 * Because we are looking for entry from file's mapping
516 * and index, so the e 530 * and index, so the entry may not be inserted for now,
517 * or even a zero/empt 531 * or even a zero/empty entry. We don't think this is
518 * an error case. So, 532 * an error case. So, return a special value and do
519 * not output @page. 533 * not output @page.
520 */ 534 */
521 entry = (void *)~0UL; 535 entry = (void *)~0UL;
522 } else { 536 } else {
523 *page = pfn_to_page(da 537 *page = pfn_to_page(dax_to_pfn(entry));
524 dax_lock_entry(&xas, e 538 dax_lock_entry(&xas, entry);
525 } 539 }
526 xas_unlock_irq(&xas); 540 xas_unlock_irq(&xas);
527 break; 541 break;
528 } 542 }
529 rcu_read_unlock(); 543 rcu_read_unlock();
530 return (dax_entry_t)entry; 544 return (dax_entry_t)entry;
531 } 545 }
532 546
533 void dax_unlock_mapping_entry(struct address_s 547 void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index,
534 dax_entry_t cookie) 548 dax_entry_t cookie)
535 { 549 {
536 XA_STATE(xas, &mapping->i_pages, index 550 XA_STATE(xas, &mapping->i_pages, index);
537 551
538 if (cookie == ~0UL) 552 if (cookie == ~0UL)
539 return; 553 return;
540 554
541 dax_unlock_entry(&xas, (void *)cookie) 555 dax_unlock_entry(&xas, (void *)cookie);
542 } 556 }
543 557
544 /* 558 /*
545 * Find page cache entry at given index. If it 559 * Find page cache entry at given index. If it is a DAX entry, return it
546 * with the entry locked. If the page cache do 560 * with the entry locked. If the page cache doesn't contain an entry at
547 * that index, add a locked empty entry. 561 * that index, add a locked empty entry.
548 * 562 *
549 * When requesting an entry with size DAX_PMD, 563 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
550 * either return that locked entry or will ret 564 * either return that locked entry or will return VM_FAULT_FALLBACK.
551 * This will happen if there are any PTE entri 565 * This will happen if there are any PTE entries within the PMD range
552 * that we are requesting. 566 * that we are requesting.
553 * 567 *
554 * We always favor PTE entries over PMD entrie 568 * We always favor PTE entries over PMD entries. There isn't a flow where we
555 * evict PTE entries in order to 'upgrade' the 569 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD
556 * insertion will fail if it finds any PTE ent 570 * insertion will fail if it finds any PTE entries already in the tree, and a
557 * PTE insertion will cause an existing PMD en 571 * PTE insertion will cause an existing PMD entry to be unmapped and
558 * downgraded to PTE entries. This happens fo 572 * downgraded to PTE entries. This happens for both PMD zero pages as
559 * well as PMD empty entries. 573 * well as PMD empty entries.
560 * 574 *
561 * The exception to this downgrade path is for 575 * The exception to this downgrade path is for PMD entries that have
562 * real storage backing them. We will leave t 576 * real storage backing them. We will leave these real PMD entries in
563 * the tree, and PTE writes will simply dirty 577 * the tree, and PTE writes will simply dirty the entire PMD entry.
564 * 578 *
565 * Note: Unlike filemap_fault() we don't honor 579 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
566 * persistent memory the benefit is doubtful. 580 * persistent memory the benefit is doubtful. We can add that later if we can
567 * show it helps. 581 * show it helps.
568 * 582 *
569 * On error, this function does not return an 583 * On error, this function does not return an ERR_PTR. Instead it returns
570 * a VM_FAULT code, encoded as an xarray inter 584 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values
571 * overlap with xarray value entries. 585 * overlap with xarray value entries.
572 */ 586 */
573 static void *grab_mapping_entry(struct xa_stat 587 static void *grab_mapping_entry(struct xa_state *xas,
574 struct address_space *mapping, 588 struct address_space *mapping, unsigned int order)
575 { 589 {
576 unsigned long index = xas->xa_index; 590 unsigned long index = xas->xa_index;
577 bool pmd_downgrade; /* splitting P 591 bool pmd_downgrade; /* splitting PMD entry into PTE entries? */
578 void *entry; 592 void *entry;
579 593
580 retry: 594 retry:
581 pmd_downgrade = false; 595 pmd_downgrade = false;
582 xas_lock_irq(xas); 596 xas_lock_irq(xas);
583 entry = get_unlocked_entry(xas, order) 597 entry = get_unlocked_entry(xas, order);
584 598
585 if (entry) { 599 if (entry) {
586 if (dax_is_conflict(entry)) 600 if (dax_is_conflict(entry))
587 goto fallback; 601 goto fallback;
588 if (!xa_is_value(entry)) { 602 if (!xa_is_value(entry)) {
589 xas_set_err(xas, -EIO) 603 xas_set_err(xas, -EIO);
590 goto out_unlock; 604 goto out_unlock;
591 } 605 }
592 606
593 if (order == 0) { 607 if (order == 0) {
594 if (dax_is_pmd_entry(e 608 if (dax_is_pmd_entry(entry) &&
595 (dax_is_zero_entry 609 (dax_is_zero_entry(entry) ||
596 dax_is_empty_entr 610 dax_is_empty_entry(entry))) {
597 pmd_downgrade 611 pmd_downgrade = true;
598 } 612 }
599 } 613 }
600 } 614 }
601 615
602 if (pmd_downgrade) { 616 if (pmd_downgrade) {
603 /* 617 /*
604 * Make sure 'entry' remains v 618 * Make sure 'entry' remains valid while we drop
605 * the i_pages lock. 619 * the i_pages lock.
606 */ 620 */
607 dax_lock_entry(xas, entry); 621 dax_lock_entry(xas, entry);
608 622
609 /* 623 /*
610 * Besides huge zero pages the 624 * Besides huge zero pages the only other thing that gets
611 * downgraded are empty entrie 625 * downgraded are empty entries which don't need to be
612 * unmapped. 626 * unmapped.
613 */ 627 */
614 if (dax_is_zero_entry(entry)) 628 if (dax_is_zero_entry(entry)) {
615 xas_unlock_irq(xas); 629 xas_unlock_irq(xas);
616 unmap_mapping_pages(ma 630 unmap_mapping_pages(mapping,
617 xas->x 631 xas->xa_index & ~PG_PMD_COLOUR,
618 PG_PMD 632 PG_PMD_NR, false);
619 xas_reset(xas); 633 xas_reset(xas);
620 xas_lock_irq(xas); 634 xas_lock_irq(xas);
621 } 635 }
622 636
623 dax_disassociate_entry(entry, 637 dax_disassociate_entry(entry, mapping, false);
624 xas_store(xas, NULL); /* und 638 xas_store(xas, NULL); /* undo the PMD join */
625 dax_wake_entry(xas, entry, WAK 639 dax_wake_entry(xas, entry, WAKE_ALL);
626 mapping->nrpages -= PG_PMD_NR; 640 mapping->nrpages -= PG_PMD_NR;
627 entry = NULL; 641 entry = NULL;
628 xas_set(xas, index); 642 xas_set(xas, index);
629 } 643 }
630 644
631 if (entry) { 645 if (entry) {
632 dax_lock_entry(xas, entry); 646 dax_lock_entry(xas, entry);
633 } else { 647 } else {
634 unsigned long flags = DAX_EMPT 648 unsigned long flags = DAX_EMPTY;
635 649
636 if (order > 0) 650 if (order > 0)
637 flags |= DAX_PMD; 651 flags |= DAX_PMD;
638 entry = dax_make_entry(pfn_to_ 652 entry = dax_make_entry(pfn_to_pfn_t(0), flags);
639 dax_lock_entry(xas, entry); 653 dax_lock_entry(xas, entry);
640 if (xas_error(xas)) 654 if (xas_error(xas))
641 goto out_unlock; 655 goto out_unlock;
642 mapping->nrpages += 1UL << ord 656 mapping->nrpages += 1UL << order;
643 } 657 }
644 658
645 out_unlock: 659 out_unlock:
646 xas_unlock_irq(xas); 660 xas_unlock_irq(xas);
647 if (xas_nomem(xas, mapping_gfp_mask(ma 661 if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM))
648 goto retry; 662 goto retry;
649 if (xas->xa_node == XA_ERROR(-ENOMEM)) 663 if (xas->xa_node == XA_ERROR(-ENOMEM))
650 return xa_mk_internal(VM_FAULT 664 return xa_mk_internal(VM_FAULT_OOM);
651 if (xas_error(xas)) 665 if (xas_error(xas))
652 return xa_mk_internal(VM_FAULT 666 return xa_mk_internal(VM_FAULT_SIGBUS);
653 return entry; 667 return entry;
654 fallback: 668 fallback:
655 xas_unlock_irq(xas); 669 xas_unlock_irq(xas);
656 return xa_mk_internal(VM_FAULT_FALLBAC 670 return xa_mk_internal(VM_FAULT_FALLBACK);
657 } 671 }
658 672
659 /** 673 /**
660 * dax_layout_busy_page_range - find first pin 674 * dax_layout_busy_page_range - find first pinned page in @mapping
661 * @mapping: address space to scan for a page 675 * @mapping: address space to scan for a page with ref count > 1
662 * @start: Starting offset. Page containing 's 676 * @start: Starting offset. Page containing 'start' is included.
663 * @end: End offset. Page containing 'end' is 677 * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX,
664 * pages from 'start' till the end of fi 678 * pages from 'start' till the end of file are included.
665 * 679 *
666 * DAX requires ZONE_DEVICE mapped pages. Thes 680 * DAX requires ZONE_DEVICE mapped pages. These pages are never
667 * 'onlined' to the page allocator so they are 681 * 'onlined' to the page allocator so they are considered idle when
668 * page->count == 1. A filesystem uses this in 682 * page->count == 1. A filesystem uses this interface to determine if
669 * any page in the mapping is busy, i.e. for D 683 * any page in the mapping is busy, i.e. for DMA, or other
670 * get_user_pages() usages. 684 * get_user_pages() usages.
671 * 685 *
672 * It is expected that the filesystem is holdi 686 * It is expected that the filesystem is holding locks to block the
673 * establishment of new mappings in this addre 687 * establishment of new mappings in this address_space. I.e. it expects
674 * to be able to run unmap_mapping_range() and 688 * to be able to run unmap_mapping_range() and subsequently not race
675 * mapping_mapped() becoming true. 689 * mapping_mapped() becoming true.
676 */ 690 */
677 struct page *dax_layout_busy_page_range(struct 691 struct page *dax_layout_busy_page_range(struct address_space *mapping,
678 loff_t 692 loff_t start, loff_t end)
679 { 693 {
680 void *entry; 694 void *entry;
681 unsigned int scanned = 0; 695 unsigned int scanned = 0;
682 struct page *page = NULL; 696 struct page *page = NULL;
683 pgoff_t start_idx = start >> PAGE_SHIF 697 pgoff_t start_idx = start >> PAGE_SHIFT;
684 pgoff_t end_idx; 698 pgoff_t end_idx;
685 XA_STATE(xas, &mapping->i_pages, start 699 XA_STATE(xas, &mapping->i_pages, start_idx);
686 700
687 /* 701 /*
688 * In the 'limited' case get_user_page 702 * In the 'limited' case get_user_pages() for dax is disabled.
689 */ 703 */
690 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 704 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
691 return NULL; 705 return NULL;
692 706
693 if (!dax_mapping(mapping) || !mapping_ 707 if (!dax_mapping(mapping) || !mapping_mapped(mapping))
694 return NULL; 708 return NULL;
695 709
696 /* If end == LLONG_MAX, all pages from 710 /* If end == LLONG_MAX, all pages from start to till end of file */
697 if (end == LLONG_MAX) 711 if (end == LLONG_MAX)
698 end_idx = ULONG_MAX; 712 end_idx = ULONG_MAX;
699 else 713 else
700 end_idx = end >> PAGE_SHIFT; 714 end_idx = end >> PAGE_SHIFT;
701 /* 715 /*
702 * If we race get_user_pages_fast() he 716 * If we race get_user_pages_fast() here either we'll see the
703 * elevated page count in the iteratio 717 * elevated page count in the iteration and wait, or
704 * get_user_pages_fast() will see that 718 * get_user_pages_fast() will see that the page it took a reference
705 * against is no longer mapped in the 719 * against is no longer mapped in the page tables and bail to the
706 * get_user_pages() slow path. The sl 720 * get_user_pages() slow path. The slow path is protected by
707 * pte_lock() and pmd_lock(). New refe 721 * pte_lock() and pmd_lock(). New references are not taken without
708 * holding those locks, and unmap_mapp 722 * holding those locks, and unmap_mapping_pages() will not zero the
709 * pte or pmd without holding the resp 723 * pte or pmd without holding the respective lock, so we are
710 * guaranteed to either see new refere 724 * guaranteed to either see new references or prevent new
711 * references from being established. 725 * references from being established.
712 */ 726 */
713 unmap_mapping_pages(mapping, start_idx 727 unmap_mapping_pages(mapping, start_idx, end_idx - start_idx + 1, 0);
714 728
715 xas_lock_irq(&xas); 729 xas_lock_irq(&xas);
716 xas_for_each(&xas, entry, end_idx) { 730 xas_for_each(&xas, entry, end_idx) {
717 if (WARN_ON_ONCE(!xa_is_value( 731 if (WARN_ON_ONCE(!xa_is_value(entry)))
718 continue; 732 continue;
719 if (unlikely(dax_is_locked(ent 733 if (unlikely(dax_is_locked(entry)))
720 entry = get_unlocked_e 734 entry = get_unlocked_entry(&xas, 0);
721 if (entry) 735 if (entry)
722 page = dax_busy_page(e 736 page = dax_busy_page(entry);
723 put_unlocked_entry(&xas, entry 737 put_unlocked_entry(&xas, entry, WAKE_NEXT);
724 if (page) 738 if (page)
725 break; 739 break;
726 if (++scanned % XA_CHECK_SCHED 740 if (++scanned % XA_CHECK_SCHED)
727 continue; 741 continue;
728 742
729 xas_pause(&xas); 743 xas_pause(&xas);
730 xas_unlock_irq(&xas); 744 xas_unlock_irq(&xas);
731 cond_resched(); 745 cond_resched();
732 xas_lock_irq(&xas); 746 xas_lock_irq(&xas);
733 } 747 }
734 xas_unlock_irq(&xas); 748 xas_unlock_irq(&xas);
735 return page; 749 return page;
736 } 750 }
737 EXPORT_SYMBOL_GPL(dax_layout_busy_page_range); 751 EXPORT_SYMBOL_GPL(dax_layout_busy_page_range);
738 752
739 struct page *dax_layout_busy_page(struct addre 753 struct page *dax_layout_busy_page(struct address_space *mapping)
740 { 754 {
741 return dax_layout_busy_page_range(mapp 755 return dax_layout_busy_page_range(mapping, 0, LLONG_MAX);
742 } 756 }
743 EXPORT_SYMBOL_GPL(dax_layout_busy_page); 757 EXPORT_SYMBOL_GPL(dax_layout_busy_page);
744 758
745 static int __dax_invalidate_entry(struct addre 759 static int __dax_invalidate_entry(struct address_space *mapping,
746 pgof 760 pgoff_t index, bool trunc)
747 { 761 {
748 XA_STATE(xas, &mapping->i_pages, index 762 XA_STATE(xas, &mapping->i_pages, index);
749 int ret = 0; 763 int ret = 0;
750 void *entry; 764 void *entry;
751 765
752 xas_lock_irq(&xas); 766 xas_lock_irq(&xas);
753 entry = get_unlocked_entry(&xas, 0); 767 entry = get_unlocked_entry(&xas, 0);
754 if (!entry || WARN_ON_ONCE(!xa_is_valu 768 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
755 goto out; 769 goto out;
756 if (!trunc && 770 if (!trunc &&
757 (xas_get_mark(&xas, PAGECACHE_TAG_ 771 (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) ||
758 xas_get_mark(&xas, PAGECACHE_TAG_ 772 xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE)))
759 goto out; 773 goto out;
760 dax_disassociate_entry(entry, mapping, 774 dax_disassociate_entry(entry, mapping, trunc);
761 xas_store(&xas, NULL); 775 xas_store(&xas, NULL);
762 mapping->nrpages -= 1UL << dax_entry_o 776 mapping->nrpages -= 1UL << dax_entry_order(entry);
763 ret = 1; 777 ret = 1;
764 out: 778 out:
765 put_unlocked_entry(&xas, entry, WAKE_A 779 put_unlocked_entry(&xas, entry, WAKE_ALL);
766 xas_unlock_irq(&xas); 780 xas_unlock_irq(&xas);
767 return ret; 781 return ret;
768 } 782 }
769 783
770 static int __dax_clear_dirty_range(struct addr 784 static int __dax_clear_dirty_range(struct address_space *mapping,
771 pgoff_t start, pgoff_t end) 785 pgoff_t start, pgoff_t end)
772 { 786 {
773 XA_STATE(xas, &mapping->i_pages, start 787 XA_STATE(xas, &mapping->i_pages, start);
774 unsigned int scanned = 0; 788 unsigned int scanned = 0;
775 void *entry; 789 void *entry;
776 790
777 xas_lock_irq(&xas); 791 xas_lock_irq(&xas);
778 xas_for_each(&xas, entry, end) { 792 xas_for_each(&xas, entry, end) {
779 entry = get_unlocked_entry(&xa 793 entry = get_unlocked_entry(&xas, 0);
780 xas_clear_mark(&xas, PAGECACHE 794 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
781 xas_clear_mark(&xas, PAGECACHE 795 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
782 put_unlocked_entry(&xas, entry 796 put_unlocked_entry(&xas, entry, WAKE_NEXT);
783 797
784 if (++scanned % XA_CHECK_SCHED 798 if (++scanned % XA_CHECK_SCHED)
785 continue; 799 continue;
786 800
787 xas_pause(&xas); 801 xas_pause(&xas);
788 xas_unlock_irq(&xas); 802 xas_unlock_irq(&xas);
789 cond_resched(); 803 cond_resched();
790 xas_lock_irq(&xas); 804 xas_lock_irq(&xas);
791 } 805 }
792 xas_unlock_irq(&xas); 806 xas_unlock_irq(&xas);
793 807
794 return 0; 808 return 0;
795 } 809 }
796 810
797 /* 811 /*
798 * Delete DAX entry at @index from @mapping. 812 * Delete DAX entry at @index from @mapping. Wait for it
799 * to be unlocked before deleting it. 813 * to be unlocked before deleting it.
800 */ 814 */
801 int dax_delete_mapping_entry(struct address_sp 815 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
802 { 816 {
803 int ret = __dax_invalidate_entry(mappi 817 int ret = __dax_invalidate_entry(mapping, index, true);
804 818
805 /* 819 /*
806 * This gets called from truncate / pu 820 * This gets called from truncate / punch_hole path. As such, the caller
807 * must hold locks protecting against 821 * must hold locks protecting against concurrent modifications of the
808 * page cache (usually fs-private i_mm 822 * page cache (usually fs-private i_mmap_sem for writing). Since the
809 * caller has seen a DAX entry for thi 823 * caller has seen a DAX entry for this index, we better find it
810 * at that index as well... 824 * at that index as well...
811 */ 825 */
812 WARN_ON_ONCE(!ret); 826 WARN_ON_ONCE(!ret);
813 return ret; 827 return ret;
814 } 828 }
815 829
816 /* 830 /*
817 * Invalidate DAX entry if it is clean. 831 * Invalidate DAX entry if it is clean.
818 */ 832 */
819 int dax_invalidate_mapping_entry_sync(struct a 833 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
820 pgoff_t 834 pgoff_t index)
821 { 835 {
822 return __dax_invalidate_entry(mapping, 836 return __dax_invalidate_entry(mapping, index, false);
823 } 837 }
824 838
825 static pgoff_t dax_iomap_pgoff(const struct io 839 static pgoff_t dax_iomap_pgoff(const struct iomap *iomap, loff_t pos)
826 { 840 {
827 return PHYS_PFN(iomap->addr + (pos & P 841 return PHYS_PFN(iomap->addr + (pos & PAGE_MASK) - iomap->offset);
828 } 842 }
829 843
830 static int copy_cow_page_dax(struct vm_fault * 844 static int copy_cow_page_dax(struct vm_fault *vmf, const struct iomap_iter *iter)
831 { 845 {
832 pgoff_t pgoff = dax_iomap_pgoff(&iter- 846 pgoff_t pgoff = dax_iomap_pgoff(&iter->iomap, iter->pos);
833 void *vto, *kaddr; 847 void *vto, *kaddr;
834 long rc; 848 long rc;
835 int id; 849 int id;
836 850
837 id = dax_read_lock(); 851 id = dax_read_lock();
838 rc = dax_direct_access(iter->iomap.dax 852 rc = dax_direct_access(iter->iomap.dax_dev, pgoff, 1, DAX_ACCESS,
839 &kaddr, NULL); 853 &kaddr, NULL);
840 if (rc < 0) { 854 if (rc < 0) {
841 dax_read_unlock(id); 855 dax_read_unlock(id);
842 return rc; 856 return rc;
843 } 857 }
844 vto = kmap_atomic(vmf->cow_page); 858 vto = kmap_atomic(vmf->cow_page);
845 copy_user_page(vto, kaddr, vmf->addres 859 copy_user_page(vto, kaddr, vmf->address, vmf->cow_page);
846 kunmap_atomic(vto); 860 kunmap_atomic(vto);
847 dax_read_unlock(id); 861 dax_read_unlock(id);
848 return 0; 862 return 0;
849 } 863 }
850 864
851 /* 865 /*
852 * MAP_SYNC on a dax mapping guarantees dirty 866 * MAP_SYNC on a dax mapping guarantees dirty metadata is
853 * flushed on write-faults (non-cow), but not 867 * flushed on write-faults (non-cow), but not read-faults.
854 */ 868 */
855 static bool dax_fault_is_synchronous(const str 869 static bool dax_fault_is_synchronous(const struct iomap_iter *iter,
856 struct vm_area_struct *vma) 870 struct vm_area_struct *vma)
857 { 871 {
858 return (iter->flags & IOMAP_WRITE) && 872 return (iter->flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) &&
859 (iter->iomap.flags & IOMAP_F_D 873 (iter->iomap.flags & IOMAP_F_DIRTY);
860 } 874 }
861 875
862 /* 876 /*
863 * By this point grab_mapping_entry() has ensu 877 * By this point grab_mapping_entry() has ensured that we have a locked entry
864 * of the appropriate size so we don't have to 878 * of the appropriate size so we don't have to worry about downgrading PMDs to
865 * PTEs. If we happen to be trying to insert 879 * PTEs. If we happen to be trying to insert a PTE and there is a PMD
866 * already in the tree, we will skip the inser 880 * already in the tree, we will skip the insertion and just dirty the PMD as
867 * appropriate. 881 * appropriate.
868 */ 882 */
869 static void *dax_insert_entry(struct xa_state 883 static void *dax_insert_entry(struct xa_state *xas, struct vm_fault *vmf,
870 const struct iomap_iter *iter, 884 const struct iomap_iter *iter, void *entry, pfn_t pfn,
871 unsigned long flags) 885 unsigned long flags)
872 { 886 {
873 struct address_space *mapping = vmf->v 887 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
874 void *new_entry = dax_make_entry(pfn, 888 void *new_entry = dax_make_entry(pfn, flags);
875 bool write = iter->flags & IOMAP_WRITE 889 bool write = iter->flags & IOMAP_WRITE;
876 bool dirty = write && !dax_fault_is_sy 890 bool dirty = write && !dax_fault_is_synchronous(iter, vmf->vma);
877 bool shared = iter->iomap.flags & IOMA 891 bool shared = iter->iomap.flags & IOMAP_F_SHARED;
878 892
879 if (dirty) 893 if (dirty)
880 __mark_inode_dirty(mapping->ho 894 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
881 895
882 if (shared || (dax_is_zero_entry(entry 896 if (shared || (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE))) {
883 unsigned long index = xas->xa_ 897 unsigned long index = xas->xa_index;
884 /* we are replacing a zero pag 898 /* we are replacing a zero page with block mapping */
885 if (dax_is_pmd_entry(entry)) 899 if (dax_is_pmd_entry(entry))
886 unmap_mapping_pages(ma 900 unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
887 PG_PMD 901 PG_PMD_NR, false);
888 else /* pte entry */ 902 else /* pte entry */
889 unmap_mapping_pages(ma 903 unmap_mapping_pages(mapping, index, 1, false);
890 } 904 }
891 905
892 xas_reset(xas); 906 xas_reset(xas);
893 xas_lock_irq(xas); 907 xas_lock_irq(xas);
894 if (shared || dax_is_zero_entry(entry) 908 if (shared || dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
895 void *old; 909 void *old;
896 910
897 dax_disassociate_entry(entry, 911 dax_disassociate_entry(entry, mapping, false);
898 dax_associate_entry(new_entry, 912 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address,
899 shared); 913 shared);
900 /* 914 /*
901 * Only swap our new entry int 915 * Only swap our new entry into the page cache if the current
902 * entry is a zero page or an 916 * entry is a zero page or an empty entry. If a normal PTE or
903 * PMD entry is already in the 917 * PMD entry is already in the cache, we leave it alone. This
904 * means that if we are trying 918 * means that if we are trying to insert a PTE and the
905 * existing entry is a PMD, we 919 * existing entry is a PMD, we will just leave the PMD in the
906 * tree and dirty it if necess 920 * tree and dirty it if necessary.
907 */ 921 */
908 old = dax_lock_entry(xas, new_ 922 old = dax_lock_entry(xas, new_entry);
909 WARN_ON_ONCE(old != xa_mk_valu 923 WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) |
910 DAX_LO 924 DAX_LOCKED));
911 entry = new_entry; 925 entry = new_entry;
912 } else { 926 } else {
913 xas_load(xas); /* Walk the xa 927 xas_load(xas); /* Walk the xa_state */
914 } 928 }
915 929
916 if (dirty) 930 if (dirty)
917 xas_set_mark(xas, PAGECACHE_TA 931 xas_set_mark(xas, PAGECACHE_TAG_DIRTY);
918 932
919 if (write && shared) 933 if (write && shared)
920 xas_set_mark(xas, PAGECACHE_TA 934 xas_set_mark(xas, PAGECACHE_TAG_TOWRITE);
921 935
922 xas_unlock_irq(xas); 936 xas_unlock_irq(xas);
923 return entry; 937 return entry;
924 } 938 }
925 939
926 static int dax_writeback_one(struct xa_state * 940 static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev,
927 struct address_space *mapping, 941 struct address_space *mapping, void *entry)
928 { 942 {
929 unsigned long pfn, index, count, end; 943 unsigned long pfn, index, count, end;
930 long ret = 0; 944 long ret = 0;
931 struct vm_area_struct *vma; 945 struct vm_area_struct *vma;
932 946
933 /* 947 /*
934 * A page got tagged dirty in DAX mapp 948 * A page got tagged dirty in DAX mapping? Something is seriously
935 * wrong. 949 * wrong.
936 */ 950 */
937 if (WARN_ON(!xa_is_value(entry))) 951 if (WARN_ON(!xa_is_value(entry)))
938 return -EIO; 952 return -EIO;
939 953
940 if (unlikely(dax_is_locked(entry))) { 954 if (unlikely(dax_is_locked(entry))) {
941 void *old_entry = entry; 955 void *old_entry = entry;
942 956
943 entry = get_unlocked_entry(xas 957 entry = get_unlocked_entry(xas, 0);
944 958
945 /* Entry got punched out / rea 959 /* Entry got punched out / reallocated? */
946 if (!entry || WARN_ON_ONCE(!xa 960 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
947 goto put_unlocked; 961 goto put_unlocked;
948 /* 962 /*
949 * Entry got reallocated elsew 963 * Entry got reallocated elsewhere? No need to writeback.
950 * We have to compare pfns as 964 * We have to compare pfns as we must not bail out due to
951 * difference in lockbit or en 965 * difference in lockbit or entry type.
952 */ 966 */
953 if (dax_to_pfn(old_entry) != d 967 if (dax_to_pfn(old_entry) != dax_to_pfn(entry))
954 goto put_unlocked; 968 goto put_unlocked;
955 if (WARN_ON_ONCE(dax_is_empty_ 969 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
956 dax_is 970 dax_is_zero_entry(entry))) {
957 ret = -EIO; 971 ret = -EIO;
958 goto put_unlocked; 972 goto put_unlocked;
959 } 973 }
960 974
961 /* Another fsync thread may ha 975 /* Another fsync thread may have already done this entry */
962 if (!xas_get_mark(xas, PAGECAC 976 if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE))
963 goto put_unlocked; 977 goto put_unlocked;
964 } 978 }
965 979
966 /* Lock the entry to serialize with pa 980 /* Lock the entry to serialize with page faults */
967 dax_lock_entry(xas, entry); 981 dax_lock_entry(xas, entry);
968 982
969 /* 983 /*
970 * We can clear the tag now but we hav 984 * We can clear the tag now but we have to be careful so that concurrent
971 * dax_writeback_one() calls for the s 985 * dax_writeback_one() calls for the same index cannot finish before we
972 * actually flush the caches. This is 986 * actually flush the caches. This is achieved as the calls will look
973 * at the entry only under the i_pages 987 * at the entry only under the i_pages lock and once they do that
974 * they will see the entry locked and 988 * they will see the entry locked and wait for it to unlock.
975 */ 989 */
976 xas_clear_mark(xas, PAGECACHE_TAG_TOWR 990 xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE);
977 xas_unlock_irq(xas); 991 xas_unlock_irq(xas);
978 992
979 /* 993 /*
980 * If dax_writeback_mapping_range() wa 994 * If dax_writeback_mapping_range() was given a wbc->range_start
981 * in the middle of a PMD, the 'index' 995 * in the middle of a PMD, the 'index' we use needs to be
982 * aligned to the start of the PMD. 996 * aligned to the start of the PMD.
983 * This allows us to flush for PMD_SIZ 997 * This allows us to flush for PMD_SIZE and not have to worry about
984 * partial PMD writebacks. 998 * partial PMD writebacks.
985 */ 999 */
986 pfn = dax_to_pfn(entry); 1000 pfn = dax_to_pfn(entry);
987 count = 1UL << dax_entry_order(entry); 1001 count = 1UL << dax_entry_order(entry);
988 index = xas->xa_index & ~(count - 1); 1002 index = xas->xa_index & ~(count - 1);
989 end = index + count - 1; 1003 end = index + count - 1;
990 1004
991 /* Walk all mappings of a given index 1005 /* Walk all mappings of a given index of a file and writeprotect them */
992 i_mmap_lock_read(mapping); 1006 i_mmap_lock_read(mapping);
993 vma_interval_tree_foreach(vma, &mappin 1007 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, end) {
994 pfn_mkclean_range(pfn, count, 1008 pfn_mkclean_range(pfn, count, index, vma);
995 cond_resched(); 1009 cond_resched();
996 } 1010 }
997 i_mmap_unlock_read(mapping); 1011 i_mmap_unlock_read(mapping);
998 1012
999 dax_flush(dax_dev, page_address(pfn_to 1013 dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE);
1000 /* 1014 /*
1001 * After we have flushed the cache, w 1015 * After we have flushed the cache, we can clear the dirty tag. There
1002 * cannot be new dirty data in the pf 1016 * cannot be new dirty data in the pfn after the flush has completed as
1003 * the pfn mappings are writeprotecte 1017 * the pfn mappings are writeprotected and fault waits for mapping
1004 * entry lock. 1018 * entry lock.
1005 */ 1019 */
1006 xas_reset(xas); 1020 xas_reset(xas);
1007 xas_lock_irq(xas); 1021 xas_lock_irq(xas);
1008 xas_store(xas, entry); 1022 xas_store(xas, entry);
1009 xas_clear_mark(xas, PAGECACHE_TAG_DIR 1023 xas_clear_mark(xas, PAGECACHE_TAG_DIRTY);
1010 dax_wake_entry(xas, entry, WAKE_NEXT) 1024 dax_wake_entry(xas, entry, WAKE_NEXT);
1011 1025
1012 trace_dax_writeback_one(mapping->host 1026 trace_dax_writeback_one(mapping->host, index, count);
1013 return ret; 1027 return ret;
1014 1028
1015 put_unlocked: 1029 put_unlocked:
1016 put_unlocked_entry(xas, entry, WAKE_N 1030 put_unlocked_entry(xas, entry, WAKE_NEXT);
1017 return ret; 1031 return ret;
1018 } 1032 }
1019 1033
1020 /* 1034 /*
1021 * Flush the mapping to the persistent domain 1035 * Flush the mapping to the persistent domain within the byte range of [start,
1022 * end]. This is required by data integrity o 1036 * end]. This is required by data integrity operations to ensure file data is
1023 * on persistent storage prior to completion 1037 * on persistent storage prior to completion of the operation.
1024 */ 1038 */
1025 int dax_writeback_mapping_range(struct addres 1039 int dax_writeback_mapping_range(struct address_space *mapping,
1026 struct dax_device *dax_dev, s 1040 struct dax_device *dax_dev, struct writeback_control *wbc)
1027 { 1041 {
1028 XA_STATE(xas, &mapping->i_pages, wbc- 1042 XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT);
1029 struct inode *inode = mapping->host; 1043 struct inode *inode = mapping->host;
1030 pgoff_t end_index = wbc->range_end >> 1044 pgoff_t end_index = wbc->range_end >> PAGE_SHIFT;
1031 void *entry; 1045 void *entry;
1032 int ret = 0; 1046 int ret = 0;
1033 unsigned int scanned = 0; 1047 unsigned int scanned = 0;
1034 1048
1035 if (WARN_ON_ONCE(inode->i_blkbits != 1049 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
1036 return -EIO; 1050 return -EIO;
1037 1051
1038 if (mapping_empty(mapping) || wbc->sy 1052 if (mapping_empty(mapping) || wbc->sync_mode != WB_SYNC_ALL)
1039 return 0; 1053 return 0;
1040 1054
1041 trace_dax_writeback_range(inode, xas. 1055 trace_dax_writeback_range(inode, xas.xa_index, end_index);
1042 1056
1043 tag_pages_for_writeback(mapping, xas. 1057 tag_pages_for_writeback(mapping, xas.xa_index, end_index);
1044 1058
1045 xas_lock_irq(&xas); 1059 xas_lock_irq(&xas);
1046 xas_for_each_marked(&xas, entry, end_ 1060 xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) {
1047 ret = dax_writeback_one(&xas, 1061 ret = dax_writeback_one(&xas, dax_dev, mapping, entry);
1048 if (ret < 0) { 1062 if (ret < 0) {
1049 mapping_set_error(map 1063 mapping_set_error(mapping, ret);
1050 break; 1064 break;
1051 } 1065 }
1052 if (++scanned % XA_CHECK_SCHE 1066 if (++scanned % XA_CHECK_SCHED)
1053 continue; 1067 continue;
1054 1068
1055 xas_pause(&xas); 1069 xas_pause(&xas);
1056 xas_unlock_irq(&xas); 1070 xas_unlock_irq(&xas);
1057 cond_resched(); 1071 cond_resched();
1058 xas_lock_irq(&xas); 1072 xas_lock_irq(&xas);
1059 } 1073 }
1060 xas_unlock_irq(&xas); 1074 xas_unlock_irq(&xas);
1061 trace_dax_writeback_range_done(inode, 1075 trace_dax_writeback_range_done(inode, xas.xa_index, end_index);
1062 return ret; 1076 return ret;
1063 } 1077 }
1064 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range 1078 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
1065 1079
1066 static int dax_iomap_direct_access(const stru 1080 static int dax_iomap_direct_access(const struct iomap *iomap, loff_t pos,
1067 size_t size, void **kaddr, pf 1081 size_t size, void **kaddr, pfn_t *pfnp)
1068 { 1082 {
1069 pgoff_t pgoff = dax_iomap_pgoff(iomap 1083 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1070 int id, rc = 0; 1084 int id, rc = 0;
1071 long length; 1085 long length;
1072 1086
1073 id = dax_read_lock(); 1087 id = dax_read_lock();
1074 length = dax_direct_access(iomap->dax 1088 length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
1075 DAX_ACCESS 1089 DAX_ACCESS, kaddr, pfnp);
1076 if (length < 0) { 1090 if (length < 0) {
1077 rc = length; 1091 rc = length;
1078 goto out; 1092 goto out;
1079 } 1093 }
1080 if (!pfnp) 1094 if (!pfnp)
1081 goto out_check_addr; 1095 goto out_check_addr;
1082 rc = -EINVAL; 1096 rc = -EINVAL;
1083 if (PFN_PHYS(length) < size) 1097 if (PFN_PHYS(length) < size)
1084 goto out; 1098 goto out;
1085 if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(s 1099 if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1086 goto out; 1100 goto out;
1087 /* For larger pages we need devmap */ 1101 /* For larger pages we need devmap */
1088 if (length > 1 && !pfn_t_devmap(*pfnp 1102 if (length > 1 && !pfn_t_devmap(*pfnp))
1089 goto out; 1103 goto out;
1090 rc = 0; 1104 rc = 0;
1091 1105
1092 out_check_addr: 1106 out_check_addr:
1093 if (!kaddr) 1107 if (!kaddr)
1094 goto out; 1108 goto out;
1095 if (!*kaddr) 1109 if (!*kaddr)
1096 rc = -EFAULT; 1110 rc = -EFAULT;
1097 out: 1111 out:
1098 dax_read_unlock(id); 1112 dax_read_unlock(id);
1099 return rc; 1113 return rc;
1100 } 1114 }
1101 1115
1102 /** 1116 /**
1103 * dax_iomap_copy_around - Prepare for an una 1117 * dax_iomap_copy_around - Prepare for an unaligned write to a shared/cow page
1104 * by copying the data before and after the r 1118 * by copying the data before and after the range to be written.
1105 * @pos: address to do copy from. 1119 * @pos: address to do copy from.
1106 * @length: size of copy operation. 1120 * @length: size of copy operation.
1107 * @align_size: aligned w.r.t align_size (eit 1121 * @align_size: aligned w.r.t align_size (either PMD_SIZE or PAGE_SIZE)
1108 * @srcmap: iomap srcmap 1122 * @srcmap: iomap srcmap
1109 * @daddr: destination address to copy t 1123 * @daddr: destination address to copy to.
1110 * 1124 *
1111 * This can be called from two places. Either 1125 * This can be called from two places. Either during DAX write fault (page
1112 * aligned), to copy the length size data to 1126 * aligned), to copy the length size data to daddr. Or, while doing normal DAX
1113 * write operation, dax_iomap_iter() might ca 1127 * write operation, dax_iomap_iter() might call this to do the copy of either
1114 * start or end unaligned address. In the lat 1128 * start or end unaligned address. In the latter case the rest of the copy of
1115 * aligned ranges is taken care by dax_iomap_ 1129 * aligned ranges is taken care by dax_iomap_iter() itself.
1116 * If the srcmap contains invalid data, such 1130 * If the srcmap contains invalid data, such as HOLE and UNWRITTEN, zero the
1117 * area to make sure no old data remains. 1131 * area to make sure no old data remains.
1118 */ 1132 */
1119 static int dax_iomap_copy_around(loff_t pos, 1133 static int dax_iomap_copy_around(loff_t pos, uint64_t length, size_t align_size,
1120 const struct iomap *srcmap, v 1134 const struct iomap *srcmap, void *daddr)
1121 { 1135 {
1122 loff_t head_off = pos & (align_size - 1136 loff_t head_off = pos & (align_size - 1);
1123 size_t size = ALIGN(head_off + length 1137 size_t size = ALIGN(head_off + length, align_size);
1124 loff_t end = pos + length; 1138 loff_t end = pos + length;
1125 loff_t pg_end = round_up(end, align_s 1139 loff_t pg_end = round_up(end, align_size);
1126 /* copy_all is usually in page fault 1140 /* copy_all is usually in page fault case */
1127 bool copy_all = head_off == 0 && end 1141 bool copy_all = head_off == 0 && end == pg_end;
1128 /* zero the edges if srcmap is a HOLE 1142 /* zero the edges if srcmap is a HOLE or IOMAP_UNWRITTEN */
1129 bool zero_edge = srcmap->flags & IOMA 1143 bool zero_edge = srcmap->flags & IOMAP_F_SHARED ||
1130 srcmap->type == IOMA 1144 srcmap->type == IOMAP_UNWRITTEN;
1131 void *saddr = NULL; !! 1145 void *saddr = 0;
1132 int ret = 0; 1146 int ret = 0;
1133 1147
1134 if (!zero_edge) { 1148 if (!zero_edge) {
1135 ret = dax_iomap_direct_access 1149 ret = dax_iomap_direct_access(srcmap, pos, size, &saddr, NULL);
1136 if (ret) 1150 if (ret)
1137 return dax_mem2blk_er !! 1151 return ret;
1138 } 1152 }
1139 1153
1140 if (copy_all) { 1154 if (copy_all) {
1141 if (zero_edge) 1155 if (zero_edge)
1142 memset(daddr, 0, size 1156 memset(daddr, 0, size);
1143 else 1157 else
1144 ret = copy_mc_to_kern 1158 ret = copy_mc_to_kernel(daddr, saddr, length);
1145 goto out; 1159 goto out;
1146 } 1160 }
1147 1161
1148 /* Copy the head part of the range */ 1162 /* Copy the head part of the range */
1149 if (head_off) { 1163 if (head_off) {
1150 if (zero_edge) 1164 if (zero_edge)
1151 memset(daddr, 0, head 1165 memset(daddr, 0, head_off);
1152 else { 1166 else {
1153 ret = copy_mc_to_kern 1167 ret = copy_mc_to_kernel(daddr, saddr, head_off);
1154 if (ret) 1168 if (ret)
1155 return -EIO; 1169 return -EIO;
1156 } 1170 }
1157 } 1171 }
1158 1172
1159 /* Copy the tail part of the range */ 1173 /* Copy the tail part of the range */
1160 if (end < pg_end) { 1174 if (end < pg_end) {
1161 loff_t tail_off = head_off + 1175 loff_t tail_off = head_off + length;
1162 loff_t tail_len = pg_end - en 1176 loff_t tail_len = pg_end - end;
1163 1177
1164 if (zero_edge) 1178 if (zero_edge)
1165 memset(daddr + tail_o 1179 memset(daddr + tail_off, 0, tail_len);
1166 else { 1180 else {
1167 ret = copy_mc_to_kern 1181 ret = copy_mc_to_kernel(daddr + tail_off,
1168 1182 saddr + tail_off, tail_len);
1169 if (ret) 1183 if (ret)
1170 return -EIO; 1184 return -EIO;
1171 } 1185 }
1172 } 1186 }
1173 out: 1187 out:
1174 if (zero_edge) 1188 if (zero_edge)
1175 dax_flush(srcmap->dax_dev, da 1189 dax_flush(srcmap->dax_dev, daddr, size);
1176 return ret ? -EIO : 0; 1190 return ret ? -EIO : 0;
1177 } 1191 }
1178 1192
1179 /* 1193 /*
1180 * The user has performed a load from a hole 1194 * The user has performed a load from a hole in the file. Allocating a new
1181 * page in the file would cause excessive sto 1195 * page in the file would cause excessive storage usage for workloads with
1182 * sparse files. Instead we insert a read-on 1196 * sparse files. Instead we insert a read-only mapping of the 4k zero page.
1183 * If this page is ever written to we will re 1197 * If this page is ever written to we will re-fault and change the mapping to
1184 * point to real DAX storage instead. 1198 * point to real DAX storage instead.
1185 */ 1199 */
1186 static vm_fault_t dax_load_hole(struct xa_sta 1200 static vm_fault_t dax_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1187 const struct iomap_iter *iter 1201 const struct iomap_iter *iter, void **entry)
1188 { 1202 {
1189 struct inode *inode = iter->inode; 1203 struct inode *inode = iter->inode;
1190 unsigned long vaddr = vmf->address; 1204 unsigned long vaddr = vmf->address;
1191 pfn_t pfn = pfn_to_pfn_t(my_zero_pfn( 1205 pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1192 vm_fault_t ret; 1206 vm_fault_t ret;
1193 1207
1194 *entry = dax_insert_entry(xas, vmf, i 1208 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, DAX_ZERO_PAGE);
1195 1209
1196 ret = vmf_insert_mixed(vmf->vma, vadd 1210 ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1197 trace_dax_load_hole(inode, vmf, ret); 1211 trace_dax_load_hole(inode, vmf, ret);
1198 return ret; 1212 return ret;
1199 } 1213 }
1200 1214
1201 #ifdef CONFIG_FS_DAX_PMD 1215 #ifdef CONFIG_FS_DAX_PMD
1202 static vm_fault_t dax_pmd_load_hole(struct xa 1216 static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1203 const struct iomap_iter *iter 1217 const struct iomap_iter *iter, void **entry)
1204 { 1218 {
1205 struct address_space *mapping = vmf-> 1219 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1206 unsigned long pmd_addr = vmf->address 1220 unsigned long pmd_addr = vmf->address & PMD_MASK;
1207 struct vm_area_struct *vma = vmf->vma 1221 struct vm_area_struct *vma = vmf->vma;
1208 struct inode *inode = mapping->host; 1222 struct inode *inode = mapping->host;
1209 pgtable_t pgtable = NULL; 1223 pgtable_t pgtable = NULL;
1210 struct folio *zero_folio; !! 1224 struct page *zero_page;
1211 spinlock_t *ptl; 1225 spinlock_t *ptl;
1212 pmd_t pmd_entry; 1226 pmd_t pmd_entry;
1213 pfn_t pfn; 1227 pfn_t pfn;
1214 1228
1215 zero_folio = mm_get_huge_zero_folio(v !! 1229 zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1216 1230
1217 if (unlikely(!zero_folio)) !! 1231 if (unlikely(!zero_page))
1218 goto fallback; 1232 goto fallback;
1219 1233
1220 pfn = page_to_pfn_t(&zero_folio->page !! 1234 pfn = page_to_pfn_t(zero_page);
1221 *entry = dax_insert_entry(xas, vmf, i 1235 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn,
1222 DAX_PMD | D 1236 DAX_PMD | DAX_ZERO_PAGE);
1223 1237
1224 if (arch_needs_pgtable_deposit()) { 1238 if (arch_needs_pgtable_deposit()) {
1225 pgtable = pte_alloc_one(vma-> 1239 pgtable = pte_alloc_one(vma->vm_mm);
1226 if (!pgtable) 1240 if (!pgtable)
1227 return VM_FAULT_OOM; 1241 return VM_FAULT_OOM;
1228 } 1242 }
1229 1243
1230 ptl = pmd_lock(vmf->vma->vm_mm, vmf-> 1244 ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1231 if (!pmd_none(*(vmf->pmd))) { 1245 if (!pmd_none(*(vmf->pmd))) {
1232 spin_unlock(ptl); 1246 spin_unlock(ptl);
1233 goto fallback; 1247 goto fallback;
1234 } 1248 }
1235 1249
1236 if (pgtable) { 1250 if (pgtable) {
1237 pgtable_trans_huge_deposit(vm 1251 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1238 mm_inc_nr_ptes(vma->vm_mm); 1252 mm_inc_nr_ptes(vma->vm_mm);
1239 } 1253 }
1240 pmd_entry = mk_pmd(&zero_folio->page, !! 1254 pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1241 pmd_entry = pmd_mkhuge(pmd_entry); 1255 pmd_entry = pmd_mkhuge(pmd_entry);
1242 set_pmd_at(vmf->vma->vm_mm, pmd_addr, 1256 set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1243 spin_unlock(ptl); 1257 spin_unlock(ptl);
1244 trace_dax_pmd_load_hole(inode, vmf, z !! 1258 trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry);
1245 return VM_FAULT_NOPAGE; 1259 return VM_FAULT_NOPAGE;
1246 1260
1247 fallback: 1261 fallback:
1248 if (pgtable) 1262 if (pgtable)
1249 pte_free(vma->vm_mm, pgtable) 1263 pte_free(vma->vm_mm, pgtable);
1250 trace_dax_pmd_load_hole_fallback(inod !! 1264 trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry);
1251 return VM_FAULT_FALLBACK; 1265 return VM_FAULT_FALLBACK;
1252 } 1266 }
1253 #else 1267 #else
1254 static vm_fault_t dax_pmd_load_hole(struct xa 1268 static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1255 const struct iomap_iter *iter 1269 const struct iomap_iter *iter, void **entry)
1256 { 1270 {
1257 return VM_FAULT_FALLBACK; 1271 return VM_FAULT_FALLBACK;
1258 } 1272 }
1259 #endif /* CONFIG_FS_DAX_PMD */ 1273 #endif /* CONFIG_FS_DAX_PMD */
1260 1274
1261 static s64 dax_unshare_iter(struct iomap_iter 1275 static s64 dax_unshare_iter(struct iomap_iter *iter)
1262 { 1276 {
1263 struct iomap *iomap = &iter->iomap; 1277 struct iomap *iomap = &iter->iomap;
1264 const struct iomap *srcmap = iomap_it 1278 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1265 loff_t copy_pos = iter->pos; !! 1279 loff_t pos = iter->pos;
1266 u64 copy_len = iomap_length(iter); !! 1280 loff_t length = iomap_length(iter);
1267 u32 mod; <<
1268 int id = 0; 1281 int id = 0;
1269 s64 ret = 0; 1282 s64 ret = 0;
1270 void *daddr = NULL, *saddr = NULL; 1283 void *daddr = NULL, *saddr = NULL;
1271 1284
1272 if (!iomap_want_unshare_iter(iter)) !! 1285 /* don't bother with blocks that are not shared to start with */
1273 return iomap_length(iter); !! 1286 if (!(iomap->flags & IOMAP_F_SHARED))
1274 !! 1287 return length;
1275 /* <<
1276 * Extend the file range to be aligne <<
1277 * we need to copy entire blocks, not <<
1278 * Invalidate the mapping because we' <<
1279 */ <<
1280 mod = offset_in_page(copy_pos); <<
1281 if (mod) { <<
1282 copy_len += mod; <<
1283 copy_pos -= mod; <<
1284 } <<
1285 <<
1286 mod = offset_in_page(copy_pos + copy_ <<
1287 if (mod) <<
1288 copy_len += PAGE_SIZE - mod; <<
1289 <<
1290 invalidate_inode_pages2_range(iter->i <<
1291 copy_po <<
1292 (copy_p <<
1293 1288
1294 id = dax_read_lock(); 1289 id = dax_read_lock();
1295 ret = dax_iomap_direct_access(iomap, !! 1290 ret = dax_iomap_direct_access(iomap, pos, length, &daddr, NULL);
1296 if (ret < 0) 1291 if (ret < 0)
1297 goto out_unlock; 1292 goto out_unlock;
1298 1293
1299 ret = dax_iomap_direct_access(srcmap, !! 1294 /* zero the distance if srcmap is HOLE or UNWRITTEN */
>> 1295 if (srcmap->flags & IOMAP_F_SHARED || srcmap->type == IOMAP_UNWRITTEN) {
>> 1296 memset(daddr, 0, length);
>> 1297 dax_flush(iomap->dax_dev, daddr, length);
>> 1298 ret = length;
>> 1299 goto out_unlock;
>> 1300 }
>> 1301
>> 1302 ret = dax_iomap_direct_access(srcmap, pos, length, &saddr, NULL);
1300 if (ret < 0) 1303 if (ret < 0)
1301 goto out_unlock; 1304 goto out_unlock;
1302 1305
1303 if (copy_mc_to_kernel(daddr, saddr, c !! 1306 if (copy_mc_to_kernel(daddr, saddr, length) == 0)
1304 ret = iomap_length(iter); !! 1307 ret = length;
1305 else 1308 else
1306 ret = -EIO; 1309 ret = -EIO;
1307 1310
1308 out_unlock: 1311 out_unlock:
1309 dax_read_unlock(id); 1312 dax_read_unlock(id);
1310 return dax_mem2blk_err(ret); !! 1313 return ret;
1311 } 1314 }
1312 1315
1313 int dax_file_unshare(struct inode *inode, lof 1316 int dax_file_unshare(struct inode *inode, loff_t pos, loff_t len,
1314 const struct iomap_ops *ops) 1317 const struct iomap_ops *ops)
1315 { 1318 {
1316 struct iomap_iter iter = { 1319 struct iomap_iter iter = {
1317 .inode = inode, 1320 .inode = inode,
1318 .pos = pos, 1321 .pos = pos,
>> 1322 .len = len,
1319 .flags = IOMAP_WRITE 1323 .flags = IOMAP_WRITE | IOMAP_UNSHARE | IOMAP_DAX,
1320 }; 1324 };
1321 loff_t size = i_size_read(inode); <<
1322 int ret; 1325 int ret;
1323 1326
1324 if (pos < 0 || pos >= size) <<
1325 return 0; <<
1326 <<
1327 iter.len = min(len, size - pos); <<
1328 while ((ret = iomap_iter(&iter, ops)) 1327 while ((ret = iomap_iter(&iter, ops)) > 0)
1329 iter.processed = dax_unshare_ 1328 iter.processed = dax_unshare_iter(&iter);
1330 return ret; 1329 return ret;
1331 } 1330 }
1332 EXPORT_SYMBOL_GPL(dax_file_unshare); 1331 EXPORT_SYMBOL_GPL(dax_file_unshare);
1333 1332
1334 static int dax_memzero(struct iomap_iter *ite 1333 static int dax_memzero(struct iomap_iter *iter, loff_t pos, size_t size)
1335 { 1334 {
1336 const struct iomap *iomap = &iter->io 1335 const struct iomap *iomap = &iter->iomap;
1337 const struct iomap *srcmap = iomap_it 1336 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1338 unsigned offset = offset_in_page(pos) 1337 unsigned offset = offset_in_page(pos);
1339 pgoff_t pgoff = dax_iomap_pgoff(iomap 1338 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1340 void *kaddr; 1339 void *kaddr;
1341 long ret; 1340 long ret;
1342 1341
1343 ret = dax_direct_access(iomap->dax_de 1342 ret = dax_direct_access(iomap->dax_dev, pgoff, 1, DAX_ACCESS, &kaddr,
1344 NULL); 1343 NULL);
1345 if (ret < 0) 1344 if (ret < 0)
1346 return dax_mem2blk_err(ret); !! 1345 return ret;
1347 <<
1348 memset(kaddr + offset, 0, size); 1346 memset(kaddr + offset, 0, size);
1349 if (iomap->flags & IOMAP_F_SHARED) 1347 if (iomap->flags & IOMAP_F_SHARED)
1350 ret = dax_iomap_copy_around(p 1348 ret = dax_iomap_copy_around(pos, size, PAGE_SIZE, srcmap,
1351 k 1349 kaddr);
1352 else 1350 else
1353 dax_flush(iomap->dax_dev, kad 1351 dax_flush(iomap->dax_dev, kaddr + offset, size);
1354 return ret; 1352 return ret;
1355 } 1353 }
1356 1354
1357 static s64 dax_zero_iter(struct iomap_iter *i 1355 static s64 dax_zero_iter(struct iomap_iter *iter, bool *did_zero)
1358 { 1356 {
1359 const struct iomap *iomap = &iter->io 1357 const struct iomap *iomap = &iter->iomap;
1360 const struct iomap *srcmap = iomap_it 1358 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1361 loff_t pos = iter->pos; 1359 loff_t pos = iter->pos;
1362 u64 length = iomap_length(iter); 1360 u64 length = iomap_length(iter);
1363 s64 written = 0; 1361 s64 written = 0;
1364 1362
1365 /* already zeroed? we're done. */ 1363 /* already zeroed? we're done. */
1366 if (srcmap->type == IOMAP_HOLE || src 1364 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
1367 return length; 1365 return length;
1368 1366
1369 /* 1367 /*
1370 * invalidate the pages whose sharing 1368 * invalidate the pages whose sharing state is to be changed
1371 * because of CoW. 1369 * because of CoW.
1372 */ 1370 */
1373 if (iomap->flags & IOMAP_F_SHARED) 1371 if (iomap->flags & IOMAP_F_SHARED)
1374 invalidate_inode_pages2_range 1372 invalidate_inode_pages2_range(iter->inode->i_mapping,
1375 1373 pos >> PAGE_SHIFT,
1376 1374 (pos + length - 1) >> PAGE_SHIFT);
1377 1375
1378 do { 1376 do {
1379 unsigned offset = offset_in_p 1377 unsigned offset = offset_in_page(pos);
1380 unsigned size = min_t(u64, PA 1378 unsigned size = min_t(u64, PAGE_SIZE - offset, length);
1381 pgoff_t pgoff = dax_iomap_pgo 1379 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1382 long rc; 1380 long rc;
1383 int id; 1381 int id;
1384 1382
1385 id = dax_read_lock(); 1383 id = dax_read_lock();
1386 if (IS_ALIGNED(pos, PAGE_SIZE 1384 if (IS_ALIGNED(pos, PAGE_SIZE) && size == PAGE_SIZE)
1387 rc = dax_zero_page_ra 1385 rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1);
1388 else 1386 else
1389 rc = dax_memzero(iter 1387 rc = dax_memzero(iter, pos, size);
1390 dax_read_unlock(id); 1388 dax_read_unlock(id);
1391 1389
1392 if (rc < 0) 1390 if (rc < 0)
1393 return rc; 1391 return rc;
1394 pos += size; 1392 pos += size;
1395 length -= size; 1393 length -= size;
1396 written += size; 1394 written += size;
1397 } while (length > 0); 1395 } while (length > 0);
1398 1396
1399 if (did_zero) 1397 if (did_zero)
1400 *did_zero = true; 1398 *did_zero = true;
1401 return written; 1399 return written;
1402 } 1400 }
1403 1401
1404 int dax_zero_range(struct inode *inode, loff_ 1402 int dax_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1405 const struct iomap_ops *ops) 1403 const struct iomap_ops *ops)
1406 { 1404 {
1407 struct iomap_iter iter = { 1405 struct iomap_iter iter = {
1408 .inode = inode, 1406 .inode = inode,
1409 .pos = pos, 1407 .pos = pos,
1410 .len = len, 1408 .len = len,
1411 .flags = IOMAP_DAX | 1409 .flags = IOMAP_DAX | IOMAP_ZERO,
1412 }; 1410 };
1413 int ret; 1411 int ret;
1414 1412
1415 while ((ret = iomap_iter(&iter, ops)) 1413 while ((ret = iomap_iter(&iter, ops)) > 0)
1416 iter.processed = dax_zero_ite 1414 iter.processed = dax_zero_iter(&iter, did_zero);
1417 return ret; 1415 return ret;
1418 } 1416 }
1419 EXPORT_SYMBOL_GPL(dax_zero_range); 1417 EXPORT_SYMBOL_GPL(dax_zero_range);
1420 1418
1421 int dax_truncate_page(struct inode *inode, lo 1419 int dax_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1422 const struct iomap_ops *ops) 1420 const struct iomap_ops *ops)
1423 { 1421 {
1424 unsigned int blocksize = i_blocksize( 1422 unsigned int blocksize = i_blocksize(inode);
1425 unsigned int off = pos & (blocksize - 1423 unsigned int off = pos & (blocksize - 1);
1426 1424
1427 /* Block boundary? Nothing to do */ 1425 /* Block boundary? Nothing to do */
1428 if (!off) 1426 if (!off)
1429 return 0; 1427 return 0;
1430 return dax_zero_range(inode, pos, blo 1428 return dax_zero_range(inode, pos, blocksize - off, did_zero, ops);
1431 } 1429 }
1432 EXPORT_SYMBOL_GPL(dax_truncate_page); 1430 EXPORT_SYMBOL_GPL(dax_truncate_page);
1433 1431
1434 static loff_t dax_iomap_iter(const struct iom 1432 static loff_t dax_iomap_iter(const struct iomap_iter *iomi,
1435 struct iov_iter *iter) 1433 struct iov_iter *iter)
1436 { 1434 {
1437 const struct iomap *iomap = &iomi->io 1435 const struct iomap *iomap = &iomi->iomap;
1438 const struct iomap *srcmap = iomap_it 1436 const struct iomap *srcmap = iomap_iter_srcmap(iomi);
1439 loff_t length = iomap_length(iomi); 1437 loff_t length = iomap_length(iomi);
1440 loff_t pos = iomi->pos; 1438 loff_t pos = iomi->pos;
1441 struct dax_device *dax_dev = iomap->d 1439 struct dax_device *dax_dev = iomap->dax_dev;
1442 loff_t end = pos + length, done = 0; 1440 loff_t end = pos + length, done = 0;
1443 bool write = iov_iter_rw(iter) == WRI 1441 bool write = iov_iter_rw(iter) == WRITE;
1444 bool cow = write && iomap->flags & IO 1442 bool cow = write && iomap->flags & IOMAP_F_SHARED;
1445 ssize_t ret = 0; 1443 ssize_t ret = 0;
1446 size_t xfer; 1444 size_t xfer;
1447 int id; 1445 int id;
1448 1446
1449 if (!write) { 1447 if (!write) {
1450 end = min(end, i_size_read(io 1448 end = min(end, i_size_read(iomi->inode));
1451 if (pos >= end) 1449 if (pos >= end)
1452 return 0; 1450 return 0;
1453 1451
1454 if (iomap->type == IOMAP_HOLE 1452 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1455 return iov_iter_zero( 1453 return iov_iter_zero(min(length, end - pos), iter);
1456 } 1454 }
1457 1455
1458 /* 1456 /*
1459 * In DAX mode, enforce either pure o 1457 * In DAX mode, enforce either pure overwrites of written extents, or
1460 * writes to unwritten extents as par 1458 * writes to unwritten extents as part of a copy-on-write operation.
1461 */ 1459 */
1462 if (WARN_ON_ONCE(iomap->type != IOMAP 1460 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED &&
1463 !(iomap->flags & IOMA 1461 !(iomap->flags & IOMAP_F_SHARED)))
1464 return -EIO; 1462 return -EIO;
1465 1463
1466 /* 1464 /*
1467 * Write can allocate block for an ar 1465 * Write can allocate block for an area which has a hole page mapped
1468 * into page tables. We have to tear 1466 * into page tables. We have to tear down these mappings so that data
1469 * written by write(2) is visible in 1467 * written by write(2) is visible in mmap.
1470 */ 1468 */
1471 if (iomap->flags & IOMAP_F_NEW || cow 1469 if (iomap->flags & IOMAP_F_NEW || cow) {
1472 /* 1470 /*
1473 * Filesystem allows CoW on n 1471 * Filesystem allows CoW on non-shared extents. The src extents
1474 * may have been mmapped with 1472 * may have been mmapped with dirty mark before. To be able to
1475 * invalidate its dax entries 1473 * invalidate its dax entries, we need to clear the dirty mark
1476 * in advance. 1474 * in advance.
1477 */ 1475 */
1478 if (cow) 1476 if (cow)
1479 __dax_clear_dirty_ran 1477 __dax_clear_dirty_range(iomi->inode->i_mapping,
1480 1478 pos >> PAGE_SHIFT,
1481 1479 (end - 1) >> PAGE_SHIFT);
1482 invalidate_inode_pages2_range 1480 invalidate_inode_pages2_range(iomi->inode->i_mapping,
1483 1481 pos >> PAGE_SHIFT,
1484 1482 (end - 1) >> PAGE_SHIFT);
1485 } 1483 }
1486 1484
1487 id = dax_read_lock(); 1485 id = dax_read_lock();
1488 while (pos < end) { 1486 while (pos < end) {
1489 unsigned offset = pos & (PAGE 1487 unsigned offset = pos & (PAGE_SIZE - 1);
1490 const size_t size = ALIGN(len 1488 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1491 pgoff_t pgoff = dax_iomap_pgo 1489 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1492 ssize_t map_len; 1490 ssize_t map_len;
1493 bool recovery = false; 1491 bool recovery = false;
1494 void *kaddr; 1492 void *kaddr;
1495 1493
1496 if (fatal_signal_pending(curr 1494 if (fatal_signal_pending(current)) {
1497 ret = -EINTR; 1495 ret = -EINTR;
1498 break; 1496 break;
1499 } 1497 }
1500 1498
1501 map_len = dax_direct_access(d 1499 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1502 DAX_ACCESS, & 1500 DAX_ACCESS, &kaddr, NULL);
1503 if (map_len == -EHWPOISON && !! 1501 if (map_len == -EIO && iov_iter_rw(iter) == WRITE) {
1504 map_len = dax_direct_ 1502 map_len = dax_direct_access(dax_dev, pgoff,
1505 PHYS_ 1503 PHYS_PFN(size), DAX_RECOVERY_WRITE,
1506 &kadd 1504 &kaddr, NULL);
1507 if (map_len > 0) 1505 if (map_len > 0)
1508 recovery = tr 1506 recovery = true;
1509 } 1507 }
1510 if (map_len < 0) { 1508 if (map_len < 0) {
1511 ret = dax_mem2blk_err !! 1509 ret = map_len;
1512 break; 1510 break;
1513 } 1511 }
1514 1512
1515 if (cow) { 1513 if (cow) {
1516 ret = dax_iomap_copy_ 1514 ret = dax_iomap_copy_around(pos, length, PAGE_SIZE,
1517 1515 srcmap, kaddr);
1518 if (ret) 1516 if (ret)
1519 break; 1517 break;
1520 } 1518 }
1521 1519
1522 map_len = PFN_PHYS(map_len); 1520 map_len = PFN_PHYS(map_len);
1523 kaddr += offset; 1521 kaddr += offset;
1524 map_len -= offset; 1522 map_len -= offset;
1525 if (map_len > end - pos) 1523 if (map_len > end - pos)
1526 map_len = end - pos; 1524 map_len = end - pos;
1527 1525
1528 if (recovery) 1526 if (recovery)
1529 xfer = dax_recovery_w 1527 xfer = dax_recovery_write(dax_dev, pgoff, kaddr,
1530 map_l 1528 map_len, iter);
1531 else if (write) 1529 else if (write)
1532 xfer = dax_copy_from_ 1530 xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1533 map_l 1531 map_len, iter);
1534 else 1532 else
1535 xfer = dax_copy_to_it 1533 xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1536 map_l 1534 map_len, iter);
1537 1535
1538 pos += xfer; 1536 pos += xfer;
1539 length -= xfer; 1537 length -= xfer;
1540 done += xfer; 1538 done += xfer;
1541 1539
1542 if (xfer == 0) 1540 if (xfer == 0)
1543 ret = -EFAULT; 1541 ret = -EFAULT;
1544 if (xfer < map_len) 1542 if (xfer < map_len)
1545 break; 1543 break;
1546 } 1544 }
1547 dax_read_unlock(id); 1545 dax_read_unlock(id);
1548 1546
1549 return done ? done : ret; 1547 return done ? done : ret;
1550 } 1548 }
1551 1549
1552 /** 1550 /**
1553 * dax_iomap_rw - Perform I/O to a DAX file 1551 * dax_iomap_rw - Perform I/O to a DAX file
1554 * @iocb: The control block for this I/ 1552 * @iocb: The control block for this I/O
1555 * @iter: The addresses to do I/O from 1553 * @iter: The addresses to do I/O from or to
1556 * @ops: iomap ops passed from the fil 1554 * @ops: iomap ops passed from the file system
1557 * 1555 *
1558 * This function performs read and write oper 1556 * This function performs read and write operations to directly mapped
1559 * persistent memory. The callers needs to t 1557 * persistent memory. The callers needs to take care of read/write exclusion
1560 * and evicting any page cache pages in the r 1558 * and evicting any page cache pages in the region under I/O.
1561 */ 1559 */
1562 ssize_t 1560 ssize_t
1563 dax_iomap_rw(struct kiocb *iocb, struct iov_i 1561 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1564 const struct iomap_ops *ops) 1562 const struct iomap_ops *ops)
1565 { 1563 {
1566 struct iomap_iter iomi = { 1564 struct iomap_iter iomi = {
1567 .inode = iocb->ki_fi 1565 .inode = iocb->ki_filp->f_mapping->host,
1568 .pos = iocb->ki_po 1566 .pos = iocb->ki_pos,
1569 .len = iov_iter_co 1567 .len = iov_iter_count(iter),
1570 .flags = IOMAP_DAX, 1568 .flags = IOMAP_DAX,
1571 }; 1569 };
1572 loff_t done = 0; 1570 loff_t done = 0;
1573 int ret; 1571 int ret;
1574 1572
1575 if (!iomi.len) 1573 if (!iomi.len)
1576 return 0; 1574 return 0;
1577 1575
1578 if (iov_iter_rw(iter) == WRITE) { 1576 if (iov_iter_rw(iter) == WRITE) {
1579 lockdep_assert_held_write(&io 1577 lockdep_assert_held_write(&iomi.inode->i_rwsem);
1580 iomi.flags |= IOMAP_WRITE; 1578 iomi.flags |= IOMAP_WRITE;
1581 } else { 1579 } else {
1582 lockdep_assert_held(&iomi.ino 1580 lockdep_assert_held(&iomi.inode->i_rwsem);
1583 } 1581 }
1584 1582
1585 if (iocb->ki_flags & IOCB_NOWAIT) 1583 if (iocb->ki_flags & IOCB_NOWAIT)
1586 iomi.flags |= IOMAP_NOWAIT; 1584 iomi.flags |= IOMAP_NOWAIT;
1587 1585
1588 while ((ret = iomap_iter(&iomi, ops)) 1586 while ((ret = iomap_iter(&iomi, ops)) > 0)
1589 iomi.processed = dax_iomap_it 1587 iomi.processed = dax_iomap_iter(&iomi, iter);
1590 1588
1591 done = iomi.pos - iocb->ki_pos; 1589 done = iomi.pos - iocb->ki_pos;
1592 iocb->ki_pos = iomi.pos; 1590 iocb->ki_pos = iomi.pos;
1593 return done ? done : ret; 1591 return done ? done : ret;
1594 } 1592 }
1595 EXPORT_SYMBOL_GPL(dax_iomap_rw); 1593 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1596 1594
1597 static vm_fault_t dax_fault_return(int error) 1595 static vm_fault_t dax_fault_return(int error)
1598 { 1596 {
1599 if (error == 0) 1597 if (error == 0)
1600 return VM_FAULT_NOPAGE; 1598 return VM_FAULT_NOPAGE;
1601 return vmf_error(error); 1599 return vmf_error(error);
1602 } 1600 }
1603 1601
1604 /* 1602 /*
1605 * When handling a synchronous page fault and 1603 * When handling a synchronous page fault and the inode need a fsync, we can
1606 * insert the PTE/PMD into page tables only a 1604 * insert the PTE/PMD into page tables only after that fsync happened. Skip
1607 * insertion for now and return the pfn so th 1605 * insertion for now and return the pfn so that caller can insert it after the
1608 * fsync is done. 1606 * fsync is done.
1609 */ 1607 */
1610 static vm_fault_t dax_fault_synchronous_pfnp( 1608 static vm_fault_t dax_fault_synchronous_pfnp(pfn_t *pfnp, pfn_t pfn)
1611 { 1609 {
1612 if (WARN_ON_ONCE(!pfnp)) 1610 if (WARN_ON_ONCE(!pfnp))
1613 return VM_FAULT_SIGBUS; 1611 return VM_FAULT_SIGBUS;
1614 *pfnp = pfn; 1612 *pfnp = pfn;
1615 return VM_FAULT_NEEDDSYNC; 1613 return VM_FAULT_NEEDDSYNC;
1616 } 1614 }
1617 1615
1618 static vm_fault_t dax_fault_cow_page(struct v 1616 static vm_fault_t dax_fault_cow_page(struct vm_fault *vmf,
1619 const struct iomap_iter *iter 1617 const struct iomap_iter *iter)
1620 { 1618 {
1621 vm_fault_t ret; 1619 vm_fault_t ret;
1622 int error = 0; 1620 int error = 0;
1623 1621
1624 switch (iter->iomap.type) { 1622 switch (iter->iomap.type) {
1625 case IOMAP_HOLE: 1623 case IOMAP_HOLE:
1626 case IOMAP_UNWRITTEN: 1624 case IOMAP_UNWRITTEN:
1627 clear_user_highpage(vmf->cow_ 1625 clear_user_highpage(vmf->cow_page, vmf->address);
1628 break; 1626 break;
1629 case IOMAP_MAPPED: 1627 case IOMAP_MAPPED:
1630 error = copy_cow_page_dax(vmf 1628 error = copy_cow_page_dax(vmf, iter);
1631 break; 1629 break;
1632 default: 1630 default:
1633 WARN_ON_ONCE(1); 1631 WARN_ON_ONCE(1);
1634 error = -EIO; 1632 error = -EIO;
1635 break; 1633 break;
1636 } 1634 }
1637 1635
1638 if (error) 1636 if (error)
1639 return dax_fault_return(error 1637 return dax_fault_return(error);
1640 1638
1641 __SetPageUptodate(vmf->cow_page); 1639 __SetPageUptodate(vmf->cow_page);
1642 ret = finish_fault(vmf); 1640 ret = finish_fault(vmf);
1643 if (!ret) 1641 if (!ret)
1644 return VM_FAULT_DONE_COW; 1642 return VM_FAULT_DONE_COW;
1645 return ret; 1643 return ret;
1646 } 1644 }
1647 1645
1648 /** 1646 /**
1649 * dax_fault_iter - Common actor to handle pf 1647 * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault.
1650 * @vmf: vm fault instance 1648 * @vmf: vm fault instance
1651 * @iter: iomap iter 1649 * @iter: iomap iter
1652 * @pfnp: pfn to be returned 1650 * @pfnp: pfn to be returned
1653 * @xas: the dax mapping tree of a fil 1651 * @xas: the dax mapping tree of a file
1654 * @entry: an unlocked dax entry to be i 1652 * @entry: an unlocked dax entry to be inserted
1655 * @pmd: distinguish whether it is a p 1653 * @pmd: distinguish whether it is a pmd fault
1656 */ 1654 */
1657 static vm_fault_t dax_fault_iter(struct vm_fa 1655 static vm_fault_t dax_fault_iter(struct vm_fault *vmf,
1658 const struct iomap_iter *iter 1656 const struct iomap_iter *iter, pfn_t *pfnp,
1659 struct xa_state *xas, void ** 1657 struct xa_state *xas, void **entry, bool pmd)
1660 { 1658 {
1661 const struct iomap *iomap = &iter->io 1659 const struct iomap *iomap = &iter->iomap;
1662 const struct iomap *srcmap = iomap_it 1660 const struct iomap *srcmap = iomap_iter_srcmap(iter);
1663 size_t size = pmd ? PMD_SIZE : PAGE_S 1661 size_t size = pmd ? PMD_SIZE : PAGE_SIZE;
1664 loff_t pos = (loff_t)xas->xa_index << 1662 loff_t pos = (loff_t)xas->xa_index << PAGE_SHIFT;
1665 bool write = iter->flags & IOMAP_WRIT 1663 bool write = iter->flags & IOMAP_WRITE;
1666 unsigned long entry_flags = pmd ? DAX 1664 unsigned long entry_flags = pmd ? DAX_PMD : 0;
1667 int err = 0; 1665 int err = 0;
1668 pfn_t pfn; 1666 pfn_t pfn;
1669 void *kaddr; 1667 void *kaddr;
1670 1668
1671 if (!pmd && vmf->cow_page) 1669 if (!pmd && vmf->cow_page)
1672 return dax_fault_cow_page(vmf 1670 return dax_fault_cow_page(vmf, iter);
1673 1671
1674 /* if we are reading UNWRITTEN and HO 1672 /* if we are reading UNWRITTEN and HOLE, return a hole. */
1675 if (!write && 1673 if (!write &&
1676 (iomap->type == IOMAP_UNWRITTEN | 1674 (iomap->type == IOMAP_UNWRITTEN || iomap->type == IOMAP_HOLE)) {
1677 if (!pmd) 1675 if (!pmd)
1678 return dax_load_hole( 1676 return dax_load_hole(xas, vmf, iter, entry);
1679 return dax_pmd_load_hole(xas, 1677 return dax_pmd_load_hole(xas, vmf, iter, entry);
1680 } 1678 }
1681 1679
1682 if (iomap->type != IOMAP_MAPPED && !( 1680 if (iomap->type != IOMAP_MAPPED && !(iomap->flags & IOMAP_F_SHARED)) {
1683 WARN_ON_ONCE(1); 1681 WARN_ON_ONCE(1);
1684 return pmd ? VM_FAULT_FALLBAC 1682 return pmd ? VM_FAULT_FALLBACK : VM_FAULT_SIGBUS;
1685 } 1683 }
1686 1684
1687 err = dax_iomap_direct_access(iomap, 1685 err = dax_iomap_direct_access(iomap, pos, size, &kaddr, &pfn);
1688 if (err) 1686 if (err)
1689 return pmd ? VM_FAULT_FALLBAC 1687 return pmd ? VM_FAULT_FALLBACK : dax_fault_return(err);
1690 1688
1691 *entry = dax_insert_entry(xas, vmf, i 1689 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, entry_flags);
1692 1690
1693 if (write && iomap->flags & IOMAP_F_S 1691 if (write && iomap->flags & IOMAP_F_SHARED) {
1694 err = dax_iomap_copy_around(p 1692 err = dax_iomap_copy_around(pos, size, size, srcmap, kaddr);
1695 if (err) 1693 if (err)
1696 return dax_fault_retu 1694 return dax_fault_return(err);
1697 } 1695 }
1698 1696
1699 if (dax_fault_is_synchronous(iter, vm 1697 if (dax_fault_is_synchronous(iter, vmf->vma))
1700 return dax_fault_synchronous_ 1698 return dax_fault_synchronous_pfnp(pfnp, pfn);
1701 1699
1702 /* insert PMD pfn */ 1700 /* insert PMD pfn */
1703 if (pmd) 1701 if (pmd)
1704 return vmf_insert_pfn_pmd(vmf 1702 return vmf_insert_pfn_pmd(vmf, pfn, write);
1705 1703
1706 /* insert PTE pfn */ 1704 /* insert PTE pfn */
1707 if (write) 1705 if (write)
1708 return vmf_insert_mixed_mkwri 1706 return vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1709 return vmf_insert_mixed(vmf->vma, vmf 1707 return vmf_insert_mixed(vmf->vma, vmf->address, pfn);
1710 } 1708 }
1711 1709
1712 static vm_fault_t dax_iomap_pte_fault(struct 1710 static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1713 int *iomap_err 1711 int *iomap_errp, const struct iomap_ops *ops)
1714 { 1712 {
1715 struct address_space *mapping = vmf-> 1713 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1716 XA_STATE(xas, &mapping->i_pages, vmf- 1714 XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
1717 struct iomap_iter iter = { 1715 struct iomap_iter iter = {
1718 .inode = mapping->ho 1716 .inode = mapping->host,
1719 .pos = (loff_t)vmf 1717 .pos = (loff_t)vmf->pgoff << PAGE_SHIFT,
1720 .len = PAGE_SIZE, 1718 .len = PAGE_SIZE,
1721 .flags = IOMAP_DAX | 1719 .flags = IOMAP_DAX | IOMAP_FAULT,
1722 }; 1720 };
1723 vm_fault_t ret = 0; 1721 vm_fault_t ret = 0;
1724 void *entry; 1722 void *entry;
1725 int error; 1723 int error;
1726 1724
1727 trace_dax_pte_fault(iter.inode, vmf, 1725 trace_dax_pte_fault(iter.inode, vmf, ret);
1728 /* 1726 /*
1729 * Check whether offset isn't beyond 1727 * Check whether offset isn't beyond end of file now. Caller is supposed
1730 * to hold locks serializing us with 1728 * to hold locks serializing us with truncate / punch hole so this is
1731 * a reliable test. 1729 * a reliable test.
1732 */ 1730 */
1733 if (iter.pos >= i_size_read(iter.inod 1731 if (iter.pos >= i_size_read(iter.inode)) {
1734 ret = VM_FAULT_SIGBUS; 1732 ret = VM_FAULT_SIGBUS;
1735 goto out; 1733 goto out;
1736 } 1734 }
1737 1735
1738 if ((vmf->flags & FAULT_FLAG_WRITE) & 1736 if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1739 iter.flags |= IOMAP_WRITE; 1737 iter.flags |= IOMAP_WRITE;
1740 1738
1741 entry = grab_mapping_entry(&xas, mapp 1739 entry = grab_mapping_entry(&xas, mapping, 0);
1742 if (xa_is_internal(entry)) { 1740 if (xa_is_internal(entry)) {
1743 ret = xa_to_internal(entry); 1741 ret = xa_to_internal(entry);
1744 goto out; 1742 goto out;
1745 } 1743 }
1746 1744
1747 /* 1745 /*
1748 * It is possible, particularly with 1746 * It is possible, particularly with mixed reads & writes to private
1749 * mappings, that we have raced with 1747 * mappings, that we have raced with a PMD fault that overlaps with
1750 * the PTE we need to set up. If so 1748 * the PTE we need to set up. If so just return and the fault will be
1751 * retried. 1749 * retried.
1752 */ 1750 */
1753 if (pmd_trans_huge(*vmf->pmd) || pmd_ 1751 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1754 ret = VM_FAULT_NOPAGE; 1752 ret = VM_FAULT_NOPAGE;
1755 goto unlock_entry; 1753 goto unlock_entry;
1756 } 1754 }
1757 1755
1758 while ((error = iomap_iter(&iter, ops 1756 while ((error = iomap_iter(&iter, ops)) > 0) {
1759 if (WARN_ON_ONCE(iomap_length 1757 if (WARN_ON_ONCE(iomap_length(&iter) < PAGE_SIZE)) {
1760 iter.processed = -EIO 1758 iter.processed = -EIO; /* fs corruption? */
1761 continue; 1759 continue;
1762 } 1760 }
1763 1761
1764 ret = dax_fault_iter(vmf, &it 1762 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, false);
1765 if (ret != VM_FAULT_SIGBUS && 1763 if (ret != VM_FAULT_SIGBUS &&
1766 (iter.iomap.flags & IOMAP 1764 (iter.iomap.flags & IOMAP_F_NEW)) {
1767 count_vm_event(PGMAJF 1765 count_vm_event(PGMAJFAULT);
1768 count_memcg_event_mm( 1766 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
1769 ret |= VM_FAULT_MAJOR 1767 ret |= VM_FAULT_MAJOR;
1770 } 1768 }
1771 1769
1772 if (!(ret & VM_FAULT_ERROR)) 1770 if (!(ret & VM_FAULT_ERROR))
1773 iter.processed = PAGE 1771 iter.processed = PAGE_SIZE;
1774 } 1772 }
1775 1773
1776 if (iomap_errp) 1774 if (iomap_errp)
1777 *iomap_errp = error; 1775 *iomap_errp = error;
1778 if (!ret && error) 1776 if (!ret && error)
1779 ret = dax_fault_return(error) 1777 ret = dax_fault_return(error);
1780 1778
1781 unlock_entry: 1779 unlock_entry:
1782 dax_unlock_entry(&xas, entry); 1780 dax_unlock_entry(&xas, entry);
1783 out: 1781 out:
1784 trace_dax_pte_fault_done(iter.inode, 1782 trace_dax_pte_fault_done(iter.inode, vmf, ret);
1785 return ret; 1783 return ret;
1786 } 1784 }
1787 1785
1788 #ifdef CONFIG_FS_DAX_PMD 1786 #ifdef CONFIG_FS_DAX_PMD
1789 static bool dax_fault_check_fallback(struct v 1787 static bool dax_fault_check_fallback(struct vm_fault *vmf, struct xa_state *xas,
1790 pgoff_t max_pgoff) 1788 pgoff_t max_pgoff)
1791 { 1789 {
1792 unsigned long pmd_addr = vmf->address 1790 unsigned long pmd_addr = vmf->address & PMD_MASK;
1793 bool write = vmf->flags & FAULT_FLAG_ 1791 bool write = vmf->flags & FAULT_FLAG_WRITE;
1794 1792
1795 /* 1793 /*
1796 * Make sure that the faulting addres 1794 * Make sure that the faulting address's PMD offset (color) matches
1797 * the PMD offset from the start of t 1795 * the PMD offset from the start of the file. This is necessary so
1798 * that a PMD range in the page table 1796 * that a PMD range in the page table overlaps exactly with a PMD
1799 * range in the page cache. 1797 * range in the page cache.
1800 */ 1798 */
1801 if ((vmf->pgoff & PG_PMD_COLOUR) != 1799 if ((vmf->pgoff & PG_PMD_COLOUR) !=
1802 ((vmf->address >> PAGE_SHIFT) & P 1800 ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1803 return true; 1801 return true;
1804 1802
1805 /* Fall back to PTEs if we're going t 1803 /* Fall back to PTEs if we're going to COW */
1806 if (write && !(vmf->vma->vm_flags & V 1804 if (write && !(vmf->vma->vm_flags & VM_SHARED))
1807 return true; 1805 return true;
1808 1806
1809 /* If the PMD would extend outside th 1807 /* If the PMD would extend outside the VMA */
1810 if (pmd_addr < vmf->vma->vm_start) 1808 if (pmd_addr < vmf->vma->vm_start)
1811 return true; 1809 return true;
1812 if ((pmd_addr + PMD_SIZE) > vmf->vma- 1810 if ((pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
1813 return true; 1811 return true;
1814 1812
1815 /* If the PMD would extend beyond the 1813 /* If the PMD would extend beyond the file size */
1816 if ((xas->xa_index | PG_PMD_COLOUR) > 1814 if ((xas->xa_index | PG_PMD_COLOUR) >= max_pgoff)
1817 return true; 1815 return true;
1818 1816
1819 return false; 1817 return false;
1820 } 1818 }
1821 1819
1822 static vm_fault_t dax_iomap_pmd_fault(struct 1820 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1823 const struct i 1821 const struct iomap_ops *ops)
1824 { 1822 {
1825 struct address_space *mapping = vmf-> 1823 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1826 XA_STATE_ORDER(xas, &mapping->i_pages 1824 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
1827 struct iomap_iter iter = { 1825 struct iomap_iter iter = {
1828 .inode = mapping->ho 1826 .inode = mapping->host,
1829 .len = PMD_SIZE, 1827 .len = PMD_SIZE,
1830 .flags = IOMAP_DAX | 1828 .flags = IOMAP_DAX | IOMAP_FAULT,
1831 }; 1829 };
1832 vm_fault_t ret = VM_FAULT_FALLBACK; 1830 vm_fault_t ret = VM_FAULT_FALLBACK;
1833 pgoff_t max_pgoff; 1831 pgoff_t max_pgoff;
1834 void *entry; 1832 void *entry;
>> 1833 int error;
1835 1834
1836 if (vmf->flags & FAULT_FLAG_WRITE) 1835 if (vmf->flags & FAULT_FLAG_WRITE)
1837 iter.flags |= IOMAP_WRITE; 1836 iter.flags |= IOMAP_WRITE;
1838 1837
1839 /* 1838 /*
1840 * Check whether offset isn't beyond 1839 * Check whether offset isn't beyond end of file now. Caller is
1841 * supposed to hold locks serializing 1840 * supposed to hold locks serializing us with truncate / punch hole so
1842 * this is a reliable test. 1841 * this is a reliable test.
1843 */ 1842 */
1844 max_pgoff = DIV_ROUND_UP(i_size_read( 1843 max_pgoff = DIV_ROUND_UP(i_size_read(iter.inode), PAGE_SIZE);
1845 1844
1846 trace_dax_pmd_fault(iter.inode, vmf, 1845 trace_dax_pmd_fault(iter.inode, vmf, max_pgoff, 0);
1847 1846
1848 if (xas.xa_index >= max_pgoff) { 1847 if (xas.xa_index >= max_pgoff) {
1849 ret = VM_FAULT_SIGBUS; 1848 ret = VM_FAULT_SIGBUS;
1850 goto out; 1849 goto out;
1851 } 1850 }
1852 1851
1853 if (dax_fault_check_fallback(vmf, &xa 1852 if (dax_fault_check_fallback(vmf, &xas, max_pgoff))
1854 goto fallback; 1853 goto fallback;
1855 1854
1856 /* 1855 /*
1857 * grab_mapping_entry() will make sur 1856 * grab_mapping_entry() will make sure we get an empty PMD entry,
1858 * a zero PMD entry or a DAX PMD. If 1857 * a zero PMD entry or a DAX PMD. If it can't (because a PTE
1859 * entry is already in the array, for 1858 * entry is already in the array, for instance), it will return
1860 * VM_FAULT_FALLBACK. 1859 * VM_FAULT_FALLBACK.
1861 */ 1860 */
1862 entry = grab_mapping_entry(&xas, mapp 1861 entry = grab_mapping_entry(&xas, mapping, PMD_ORDER);
1863 if (xa_is_internal(entry)) { 1862 if (xa_is_internal(entry)) {
1864 ret = xa_to_internal(entry); 1863 ret = xa_to_internal(entry);
1865 goto fallback; 1864 goto fallback;
1866 } 1865 }
1867 1866
1868 /* 1867 /*
1869 * It is possible, particularly with 1868 * It is possible, particularly with mixed reads & writes to private
1870 * mappings, that we have raced with 1869 * mappings, that we have raced with a PTE fault that overlaps with
1871 * the PMD we need to set up. If so 1870 * the PMD we need to set up. If so just return and the fault will be
1872 * retried. 1871 * retried.
1873 */ 1872 */
1874 if (!pmd_none(*vmf->pmd) && !pmd_tran 1873 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1875 !pmd_devmap(*vmf->pmd 1874 !pmd_devmap(*vmf->pmd)) {
1876 ret = 0; 1875 ret = 0;
1877 goto unlock_entry; 1876 goto unlock_entry;
1878 } 1877 }
1879 1878
1880 iter.pos = (loff_t)xas.xa_index << PA 1879 iter.pos = (loff_t)xas.xa_index << PAGE_SHIFT;
1881 while (iomap_iter(&iter, ops) > 0) { !! 1880 while ((error = iomap_iter(&iter, ops)) > 0) {
1882 if (iomap_length(&iter) < PMD 1881 if (iomap_length(&iter) < PMD_SIZE)
1883 continue; /* actually 1882 continue; /* actually breaks out of the loop */
1884 1883
1885 ret = dax_fault_iter(vmf, &it 1884 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, true);
1886 if (ret != VM_FAULT_FALLBACK) 1885 if (ret != VM_FAULT_FALLBACK)
1887 iter.processed = PMD_ 1886 iter.processed = PMD_SIZE;
1888 } 1887 }
1889 1888
1890 unlock_entry: 1889 unlock_entry:
1891 dax_unlock_entry(&xas, entry); 1890 dax_unlock_entry(&xas, entry);
1892 fallback: 1891 fallback:
1893 if (ret == VM_FAULT_FALLBACK) { 1892 if (ret == VM_FAULT_FALLBACK) {
1894 split_huge_pmd(vmf->vma, vmf- 1893 split_huge_pmd(vmf->vma, vmf->pmd, vmf->address);
1895 count_vm_event(THP_FAULT_FALL 1894 count_vm_event(THP_FAULT_FALLBACK);
1896 } 1895 }
1897 out: 1896 out:
1898 trace_dax_pmd_fault_done(iter.inode, 1897 trace_dax_pmd_fault_done(iter.inode, vmf, max_pgoff, ret);
1899 return ret; 1898 return ret;
1900 } 1899 }
1901 #else 1900 #else
1902 static vm_fault_t dax_iomap_pmd_fault(struct 1901 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1903 const struct i 1902 const struct iomap_ops *ops)
1904 { 1903 {
1905 return VM_FAULT_FALLBACK; 1904 return VM_FAULT_FALLBACK;
1906 } 1905 }
1907 #endif /* CONFIG_FS_DAX_PMD */ 1906 #endif /* CONFIG_FS_DAX_PMD */
1908 1907
1909 /** 1908 /**
1910 * dax_iomap_fault - handle a page fault on a 1909 * dax_iomap_fault - handle a page fault on a DAX file
1911 * @vmf: The description of the fault 1910 * @vmf: The description of the fault
1912 * @order: Order of the page to fault in !! 1911 * @pe_size: Size of the page to fault in
1913 * @pfnp: PFN to insert for synchronous fault 1912 * @pfnp: PFN to insert for synchronous faults if fsync is required
1914 * @iomap_errp: Storage for detailed error co 1913 * @iomap_errp: Storage for detailed error code in case of error
1915 * @ops: Iomap ops passed from the file syste 1914 * @ops: Iomap ops passed from the file system
1916 * 1915 *
1917 * When a page fault occurs, filesystems may 1916 * When a page fault occurs, filesystems may call this helper in
1918 * their fault handler for DAX files. dax_iom 1917 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1919 * has done all the necessary locking for pag 1918 * has done all the necessary locking for page fault to proceed
1920 * successfully. 1919 * successfully.
1921 */ 1920 */
1922 vm_fault_t dax_iomap_fault(struct vm_fault *v !! 1921 vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1923 pfn_t *pfnp, int *iomap_e 1922 pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1924 { 1923 {
1925 if (order == 0) !! 1924 switch (pe_size) {
>> 1925 case PE_SIZE_PTE:
1926 return dax_iomap_pte_fault(vm 1926 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1927 else if (order == PMD_ORDER) !! 1927 case PE_SIZE_PMD:
1928 return dax_iomap_pmd_fault(vm 1928 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1929 else !! 1929 default:
1930 return VM_FAULT_FALLBACK; 1930 return VM_FAULT_FALLBACK;
>> 1931 }
1931 } 1932 }
1932 EXPORT_SYMBOL_GPL(dax_iomap_fault); 1933 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1933 1934
1934 /* 1935 /*
1935 * dax_insert_pfn_mkwrite - insert PTE or PMD 1936 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1936 * @vmf: The description of the fault 1937 * @vmf: The description of the fault
1937 * @pfn: PFN to insert 1938 * @pfn: PFN to insert
1938 * @order: Order of entry to insert. 1939 * @order: Order of entry to insert.
1939 * 1940 *
1940 * This function inserts a writeable PTE or P 1941 * This function inserts a writeable PTE or PMD entry into the page tables
1941 * for an mmaped DAX file. It also marks the 1942 * for an mmaped DAX file. It also marks the page cache entry as dirty.
1942 */ 1943 */
1943 static vm_fault_t 1944 static vm_fault_t
1944 dax_insert_pfn_mkwrite(struct vm_fault *vmf, 1945 dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
1945 { 1946 {
1946 struct address_space *mapping = vmf-> 1947 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1947 XA_STATE_ORDER(xas, &mapping->i_pages 1948 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1948 void *entry; 1949 void *entry;
1949 vm_fault_t ret; 1950 vm_fault_t ret;
1950 1951
1951 xas_lock_irq(&xas); 1952 xas_lock_irq(&xas);
1952 entry = get_unlocked_entry(&xas, orde 1953 entry = get_unlocked_entry(&xas, order);
1953 /* Did we race with someone splitting 1954 /* Did we race with someone splitting entry or so? */
1954 if (!entry || dax_is_conflict(entry) 1955 if (!entry || dax_is_conflict(entry) ||
1955 (order == 0 && !dax_is_pte_entry( 1956 (order == 0 && !dax_is_pte_entry(entry))) {
1956 put_unlocked_entry(&xas, entr 1957 put_unlocked_entry(&xas, entry, WAKE_NEXT);
1957 xas_unlock_irq(&xas); 1958 xas_unlock_irq(&xas);
1958 trace_dax_insert_pfn_mkwrite_ 1959 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1959 1960 VM_FAULT_NOPAGE);
1960 return VM_FAULT_NOPAGE; 1961 return VM_FAULT_NOPAGE;
1961 } 1962 }
1962 xas_set_mark(&xas, PAGECACHE_TAG_DIRT 1963 xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1963 dax_lock_entry(&xas, entry); 1964 dax_lock_entry(&xas, entry);
1964 xas_unlock_irq(&xas); 1965 xas_unlock_irq(&xas);
1965 if (order == 0) 1966 if (order == 0)
1966 ret = vmf_insert_mixed_mkwrit 1967 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1967 #ifdef CONFIG_FS_DAX_PMD 1968 #ifdef CONFIG_FS_DAX_PMD
1968 else if (order == PMD_ORDER) 1969 else if (order == PMD_ORDER)
1969 ret = vmf_insert_pfn_pmd(vmf, 1970 ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
1970 #endif 1971 #endif
1971 else 1972 else
1972 ret = VM_FAULT_FALLBACK; 1973 ret = VM_FAULT_FALLBACK;
1973 dax_unlock_entry(&xas, entry); 1974 dax_unlock_entry(&xas, entry);
1974 trace_dax_insert_pfn_mkwrite(mapping- 1975 trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1975 return ret; 1976 return ret;
1976 } 1977 }
1977 1978
1978 /** 1979 /**
1979 * dax_finish_sync_fault - finish synchronous 1980 * dax_finish_sync_fault - finish synchronous page fault
1980 * @vmf: The description of the fault 1981 * @vmf: The description of the fault
1981 * @order: Order of entry to be inserted !! 1982 * @pe_size: Size of entry to be inserted
1982 * @pfn: PFN to insert 1983 * @pfn: PFN to insert
1983 * 1984 *
1984 * This function ensures that the file range 1985 * This function ensures that the file range touched by the page fault is
1985 * stored persistently on the media and handl 1986 * stored persistently on the media and handles inserting of appropriate page
1986 * table entry. 1987 * table entry.
1987 */ 1988 */
1988 vm_fault_t dax_finish_sync_fault(struct vm_fa !! 1989 vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf,
1989 pfn_t pfn) !! 1990 enum page_entry_size pe_size, pfn_t pfn)
1990 { 1991 {
1991 int err; 1992 int err;
1992 loff_t start = ((loff_t)vmf->pgoff) < 1993 loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
>> 1994 unsigned int order = pe_order(pe_size);
1993 size_t len = PAGE_SIZE << order; 1995 size_t len = PAGE_SIZE << order;
1994 1996
1995 err = vfs_fsync_range(vmf->vma->vm_fi 1997 err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1996 if (err) 1998 if (err)
1997 return VM_FAULT_SIGBUS; 1999 return VM_FAULT_SIGBUS;
1998 return dax_insert_pfn_mkwrite(vmf, pf 2000 return dax_insert_pfn_mkwrite(vmf, pfn, order);
1999 } 2001 }
2000 EXPORT_SYMBOL_GPL(dax_finish_sync_fault); 2002 EXPORT_SYMBOL_GPL(dax_finish_sync_fault);
2001 2003
2002 static loff_t dax_range_compare_iter(struct i 2004 static loff_t dax_range_compare_iter(struct iomap_iter *it_src,
2003 struct iomap_iter *it_dest, u 2005 struct iomap_iter *it_dest, u64 len, bool *same)
2004 { 2006 {
2005 const struct iomap *smap = &it_src->i 2007 const struct iomap *smap = &it_src->iomap;
2006 const struct iomap *dmap = &it_dest-> 2008 const struct iomap *dmap = &it_dest->iomap;
2007 loff_t pos1 = it_src->pos, pos2 = it_ 2009 loff_t pos1 = it_src->pos, pos2 = it_dest->pos;
2008 void *saddr, *daddr; 2010 void *saddr, *daddr;
2009 int id, ret; 2011 int id, ret;
2010 2012
2011 len = min(len, min(smap->length, dmap 2013 len = min(len, min(smap->length, dmap->length));
2012 2014
2013 if (smap->type == IOMAP_HOLE && dmap- 2015 if (smap->type == IOMAP_HOLE && dmap->type == IOMAP_HOLE) {
2014 *same = true; 2016 *same = true;
2015 return len; 2017 return len;
2016 } 2018 }
2017 2019
2018 if (smap->type == IOMAP_HOLE || dmap- 2020 if (smap->type == IOMAP_HOLE || dmap->type == IOMAP_HOLE) {
2019 *same = false; 2021 *same = false;
2020 return 0; 2022 return 0;
2021 } 2023 }
2022 2024
2023 id = dax_read_lock(); 2025 id = dax_read_lock();
2024 ret = dax_iomap_direct_access(smap, p 2026 ret = dax_iomap_direct_access(smap, pos1, ALIGN(pos1 + len, PAGE_SIZE),
2025 &saddr, 2027 &saddr, NULL);
2026 if (ret < 0) 2028 if (ret < 0)
2027 goto out_unlock; 2029 goto out_unlock;
2028 2030
2029 ret = dax_iomap_direct_access(dmap, p 2031 ret = dax_iomap_direct_access(dmap, pos2, ALIGN(pos2 + len, PAGE_SIZE),
2030 &daddr, 2032 &daddr, NULL);
2031 if (ret < 0) 2033 if (ret < 0)
2032 goto out_unlock; 2034 goto out_unlock;
2033 2035
2034 *same = !memcmp(saddr, daddr, len); 2036 *same = !memcmp(saddr, daddr, len);
2035 if (!*same) 2037 if (!*same)
2036 len = 0; 2038 len = 0;
2037 dax_read_unlock(id); 2039 dax_read_unlock(id);
2038 return len; 2040 return len;
2039 2041
2040 out_unlock: 2042 out_unlock:
2041 dax_read_unlock(id); 2043 dax_read_unlock(id);
2042 return -EIO; 2044 return -EIO;
2043 } 2045 }
2044 2046
2045 int dax_dedupe_file_range_compare(struct inod 2047 int dax_dedupe_file_range_compare(struct inode *src, loff_t srcoff,
2046 struct inode *dst, loff_t dst 2048 struct inode *dst, loff_t dstoff, loff_t len, bool *same,
2047 const struct iomap_ops *ops) 2049 const struct iomap_ops *ops)
2048 { 2050 {
2049 struct iomap_iter src_iter = { 2051 struct iomap_iter src_iter = {
2050 .inode = src, 2052 .inode = src,
2051 .pos = srcoff, 2053 .pos = srcoff,
2052 .len = len, 2054 .len = len,
2053 .flags = IOMAP_DAX, 2055 .flags = IOMAP_DAX,
2054 }; 2056 };
2055 struct iomap_iter dst_iter = { 2057 struct iomap_iter dst_iter = {
2056 .inode = dst, 2058 .inode = dst,
2057 .pos = dstoff, 2059 .pos = dstoff,
2058 .len = len, 2060 .len = len,
2059 .flags = IOMAP_DAX, 2061 .flags = IOMAP_DAX,
2060 }; 2062 };
2061 int ret, compared = 0; 2063 int ret, compared = 0;
2062 2064
2063 while ((ret = iomap_iter(&src_iter, o 2065 while ((ret = iomap_iter(&src_iter, ops)) > 0 &&
2064 (ret = iomap_iter(&dst_iter, o 2066 (ret = iomap_iter(&dst_iter, ops)) > 0) {
2065 compared = dax_range_compare_ 2067 compared = dax_range_compare_iter(&src_iter, &dst_iter,
2066 min(src_iter. 2068 min(src_iter.len, dst_iter.len), same);
2067 if (compared < 0) 2069 if (compared < 0)
2068 return ret; 2070 return ret;
2069 src_iter.processed = dst_iter 2071 src_iter.processed = dst_iter.processed = compared;
2070 } 2072 }
2071 return ret; 2073 return ret;
2072 } 2074 }
2073 2075
2074 int dax_remap_file_range_prep(struct file *fi 2076 int dax_remap_file_range_prep(struct file *file_in, loff_t pos_in,
2075 struct file *fi 2077 struct file *file_out, loff_t pos_out,
2076 loff_t *len, un 2078 loff_t *len, unsigned int remap_flags,
2077 const struct io 2079 const struct iomap_ops *ops)
2078 { 2080 {
2079 return __generic_remap_file_range_pre 2081 return __generic_remap_file_range_prep(file_in, pos_in, file_out,
2080 2082 pos_out, len, remap_flags, ops);
2081 } 2083 }
2082 EXPORT_SYMBOL_GPL(dax_remap_file_range_prep); 2084 EXPORT_SYMBOL_GPL(dax_remap_file_range_prep);
2083 2085
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.