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