1 // SPDX-License-Identifier: GPL-2.0-only <<
2 /* 1 /*
3 * fs/dax.c - Direct Access filesystem code 2 * fs/dax.c - Direct Access filesystem code
4 * Copyright (c) 2013-2014 Intel Corporation 3 * Copyright (c) 2013-2014 Intel Corporation
5 * Author: Matthew Wilcox <matthew.r.wilcox@in 4 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
6 * Author: Ross Zwisler <ross.zwisler@linux.in 5 * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
>> 6 *
>> 7 * This program is free software; you can redistribute it and/or modify it
>> 8 * under the terms and conditions of the GNU General Public License,
>> 9 * version 2, as published by the Free Software Foundation.
>> 10 *
>> 11 * This program is distributed in the hope it will be useful, but WITHOUT
>> 12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
>> 13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
>> 14 * more details.
7 */ 15 */
8 16
9 #include <linux/atomic.h> 17 #include <linux/atomic.h>
10 #include <linux/blkdev.h> 18 #include <linux/blkdev.h>
11 #include <linux/buffer_head.h> 19 #include <linux/buffer_head.h>
12 #include <linux/dax.h> 20 #include <linux/dax.h>
13 #include <linux/fs.h> 21 #include <linux/fs.h>
>> 22 #include <linux/genhd.h>
14 #include <linux/highmem.h> 23 #include <linux/highmem.h>
15 #include <linux/memcontrol.h> 24 #include <linux/memcontrol.h>
16 #include <linux/mm.h> 25 #include <linux/mm.h>
17 #include <linux/mutex.h> 26 #include <linux/mutex.h>
18 #include <linux/pagevec.h> 27 #include <linux/pagevec.h>
>> 28 #include <linux/pmem.h>
19 #include <linux/sched.h> 29 #include <linux/sched.h>
20 #include <linux/sched/signal.h> 30 #include <linux/sched/signal.h>
21 #include <linux/uio.h> 31 #include <linux/uio.h>
22 #include <linux/vmstat.h> 32 #include <linux/vmstat.h>
23 #include <linux/pfn_t.h> 33 #include <linux/pfn_t.h>
24 #include <linux/sizes.h> 34 #include <linux/sizes.h>
25 #include <linux/mmu_notifier.h> 35 #include <linux/mmu_notifier.h>
26 #include <linux/iomap.h> 36 #include <linux/iomap.h>
27 #include <linux/rmap.h> !! 37 #include "internal.h"
28 #include <asm/pgalloc.h> <<
29 38
30 #define CREATE_TRACE_POINTS 39 #define CREATE_TRACE_POINTS
31 #include <trace/events/fs_dax.h> 40 #include <trace/events/fs_dax.h>
32 41
33 /* We choose 4096 entries - same as per-zone p 42 /* We choose 4096 entries - same as per-zone page wait tables */
34 #define DAX_WAIT_TABLE_BITS 12 43 #define DAX_WAIT_TABLE_BITS 12
35 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_ 44 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
36 45
37 /* The 'colour' (ie low bits) within a PMD of <<
38 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHI <<
39 #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIF <<
40 <<
41 static wait_queue_head_t wait_table[DAX_WAIT_T 46 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
42 47
43 static int __init init_dax_wait_table(void) 48 static int __init init_dax_wait_table(void)
44 { 49 {
45 int i; 50 int i;
46 51
47 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES 52 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
48 init_waitqueue_head(wait_table 53 init_waitqueue_head(wait_table + i);
49 return 0; 54 return 0;
50 } 55 }
51 fs_initcall(init_dax_wait_table); 56 fs_initcall(init_dax_wait_table);
52 57
53 /* !! 58 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
54 * DAX pagecache entries use XArray value entr <<
55 * for pages. We use one bit for locking, one <<
56 * and two more to tell us if the entry is a z <<
57 * is just used for locking. In total four sp <<
58 * <<
59 * If the PMD bit isn't set the entry has size <<
60 * and EMPTY bits aren't set the entry is a no <<
61 * block allocation. <<
62 */ <<
63 #define DAX_SHIFT (4) <<
64 #define DAX_LOCKED (1UL << 0) <<
65 #define DAX_PMD (1UL << 1) <<
66 #define DAX_ZERO_PAGE (1UL << 2) <<
67 #define DAX_EMPTY (1UL << 3) <<
68 <<
69 static unsigned long dax_to_pfn(void *entry) <<
70 { 59 {
71 return xa_to_value(entry) >> DAX_SHIFT !! 60 struct request_queue *q = bdev->bd_queue;
72 } !! 61 long rc = -EIO;
73 62
74 static void *dax_make_entry(pfn_t pfn, unsigne !! 63 dax->addr = ERR_PTR(-EIO);
75 { !! 64 if (blk_queue_enter(q, true) != 0)
76 return xa_mk_value(flags | (pfn_t_to_p !! 65 return rc;
77 } <<
78 66
79 static bool dax_is_locked(void *entry) !! 67 rc = bdev_direct_access(bdev, dax);
80 { !! 68 if (rc < 0) {
81 return xa_to_value(entry) & DAX_LOCKED !! 69 dax->addr = ERR_PTR(rc);
>> 70 blk_queue_exit(q);
>> 71 return rc;
>> 72 }
>> 73 return rc;
82 } 74 }
83 75
84 static unsigned int dax_entry_order(void *entr !! 76 static void dax_unmap_atomic(struct block_device *bdev,
>> 77 const struct blk_dax_ctl *dax)
85 { 78 {
86 if (xa_to_value(entry) & DAX_PMD) !! 79 if (IS_ERR(dax->addr))
87 return PMD_ORDER; !! 80 return;
88 return 0; !! 81 blk_queue_exit(bdev->bd_queue);
89 } 82 }
90 83
91 static unsigned long dax_is_pmd_entry(void *en !! 84 static int dax_is_pmd_entry(void *entry)
92 { 85 {
93 return xa_to_value(entry) & DAX_PMD; !! 86 return (unsigned long)entry & RADIX_DAX_PMD;
94 } 87 }
95 88
96 static bool dax_is_pte_entry(void *entry) !! 89 static int dax_is_pte_entry(void *entry)
97 { 90 {
98 return !(xa_to_value(entry) & DAX_PMD) !! 91 return !((unsigned long)entry & RADIX_DAX_PMD);
99 } 92 }
100 93
101 static int dax_is_zero_entry(void *entry) 94 static int dax_is_zero_entry(void *entry)
102 { 95 {
103 return xa_to_value(entry) & DAX_ZERO_P !! 96 return (unsigned long)entry & RADIX_DAX_HZP;
104 } 97 }
105 98
106 static int dax_is_empty_entry(void *entry) 99 static int dax_is_empty_entry(void *entry)
107 { 100 {
108 return xa_to_value(entry) & DAX_EMPTY; !! 101 return (unsigned long)entry & RADIX_DAX_EMPTY;
109 } 102 }
110 103
111 /* !! 104 struct page *read_dax_sector(struct block_device *bdev, sector_t n)
112 * true if the entry that was found is of a sm <<
113 * we were looking for <<
114 */ <<
115 static bool dax_is_conflict(void *entry) <<
116 { 105 {
117 return entry == XA_RETRY_ENTRY; !! 106 struct page *page = alloc_pages(GFP_KERNEL, 0);
>> 107 struct blk_dax_ctl dax = {
>> 108 .size = PAGE_SIZE,
>> 109 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
>> 110 };
>> 111 long rc;
>> 112
>> 113 if (!page)
>> 114 return ERR_PTR(-ENOMEM);
>> 115
>> 116 rc = dax_map_atomic(bdev, &dax);
>> 117 if (rc < 0)
>> 118 return ERR_PTR(rc);
>> 119 memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
>> 120 dax_unmap_atomic(bdev, &dax);
>> 121 return page;
118 } 122 }
119 123
120 /* 124 /*
121 * DAX page cache entry locking !! 125 * DAX radix tree locking
122 */ 126 */
123 struct exceptional_entry_key { 127 struct exceptional_entry_key {
124 struct xarray *xa; !! 128 struct address_space *mapping;
125 pgoff_t entry_start; 129 pgoff_t entry_start;
126 }; 130 };
127 131
128 struct wait_exceptional_entry_queue { 132 struct wait_exceptional_entry_queue {
129 wait_queue_entry_t wait; !! 133 wait_queue_t wait;
130 struct exceptional_entry_key key; 134 struct exceptional_entry_key key;
131 }; 135 };
132 136
133 /** !! 137 static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
134 * enum dax_wake_mode: waitqueue wakeup behavi !! 138 pgoff_t index, void *entry, struct exceptional_entry_key *key)
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( <<
144 void *entry, struct exceptiona <<
145 { 139 {
146 unsigned long hash; 140 unsigned long hash;
147 unsigned long index = xas->xa_index; <<
148 141
149 /* 142 /*
150 * If 'entry' is a PMD, align the 'ind 143 * If 'entry' is a PMD, align the 'index' that we use for the wait
151 * queue to the start of that PMD. Th 144 * queue to the start of that PMD. This ensures that all offsets in
152 * the range covered by the PMD map to 145 * the range covered by the PMD map to the same bit lock.
153 */ 146 */
154 if (dax_is_pmd_entry(entry)) 147 if (dax_is_pmd_entry(entry))
155 index &= ~PG_PMD_COLOUR; !! 148 index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1);
156 key->xa = xas->xa; !! 149
>> 150 key->mapping = mapping;
157 key->entry_start = index; 151 key->entry_start = index;
158 152
159 hash = hash_long((unsigned long)xas->x !! 153 hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
160 return wait_table + hash; 154 return wait_table + hash;
161 } 155 }
162 156
163 static int wake_exceptional_entry_func(wait_qu !! 157 static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode,
164 unsigned int mode, int sync, v !! 158 int sync, void *keyp)
165 { 159 {
166 struct exceptional_entry_key *key = ke 160 struct exceptional_entry_key *key = keyp;
167 struct wait_exceptional_entry_queue *e 161 struct wait_exceptional_entry_queue *ewait =
168 container_of(wait, struct wait 162 container_of(wait, struct wait_exceptional_entry_queue, wait);
169 163
170 if (key->xa != ewait->key.xa || !! 164 if (key->mapping != ewait->key.mapping ||
171 key->entry_start != ewait->key.ent 165 key->entry_start != ewait->key.entry_start)
172 return 0; 166 return 0;
173 return autoremove_wake_function(wait, 167 return autoremove_wake_function(wait, mode, sync, NULL);
174 } 168 }
175 169
176 /* 170 /*
177 * @entry may no longer be the entry at the in !! 171 * Check whether the given slot is locked. The function must be called with
178 * The important information it's conveying is !! 172 * mapping->tree_lock held
179 * this index used to be a PMD entry. <<
180 */ 173 */
181 static void dax_wake_entry(struct xa_state *xa !! 174 static inline int slot_locked(struct address_space *mapping, void **slot)
182 enum dax_wake_mode <<
183 { 175 {
184 struct exceptional_entry_key key; !! 176 unsigned long entry = (unsigned long)
185 wait_queue_head_t *wq; !! 177 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
>> 178 return entry & RADIX_DAX_ENTRY_LOCK;
>> 179 }
186 180
187 wq = dax_entry_waitqueue(xas, entry, & !! 181 /*
>> 182 * Mark the given slot is locked. The function must be called with
>> 183 * mapping->tree_lock held
>> 184 */
>> 185 static inline void *lock_slot(struct address_space *mapping, void **slot)
>> 186 {
>> 187 unsigned long entry = (unsigned long)
>> 188 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
188 189
189 /* !! 190 entry |= RADIX_DAX_ENTRY_LOCK;
190 * Checking for locked entry and prepa !! 191 radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
191 * under the i_pages lock, ditto for e !! 192 return (void *)entry;
192 * So at this point all tasks that cou <<
193 * must be in the waitqueue and the fo <<
194 */ <<
195 if (waitqueue_active(wq)) <<
196 __wake_up(wq, TASK_NORMAL, mod <<
197 } 193 }
198 194
199 /* 195 /*
200 * Look up entry in page cache, wait for it to !! 196 * Mark the given slot is unlocked. The function must be called with
201 * is a DAX entry and return it. The caller m !! 197 * mapping->tree_lock held
202 * put_unlocked_entry() if it did not lock the !! 198 */
203 * if it did. The entry returned may have a l !! 199 static inline void *unlock_slot(struct address_space *mapping, void **slot)
204 * If @order is larger than the order of the e !! 200 {
205 * function returns a dax_is_conflict entry. !! 201 unsigned long entry = (unsigned long)
>> 202 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
>> 203
>> 204 entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
>> 205 radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
>> 206 return (void *)entry;
>> 207 }
>> 208
>> 209 /*
>> 210 * Lookup entry in radix tree, wait for it to become unlocked if it is
>> 211 * exceptional entry and return it. The caller must call
>> 212 * put_unlocked_mapping_entry() when he decided not to lock the entry or
>> 213 * put_locked_mapping_entry() when he locked the entry and now wants to
>> 214 * unlock it.
206 * 215 *
207 * Must be called with the i_pages lock held. !! 216 * The function must be called with mapping->tree_lock held.
208 */ 217 */
209 static void *get_unlocked_entry(struct xa_stat !! 218 static void *get_unlocked_mapping_entry(struct address_space *mapping,
>> 219 pgoff_t index, void ***slotp)
210 { 220 {
211 void *entry; !! 221 void *entry, **slot;
212 struct wait_exceptional_entry_queue ew 222 struct wait_exceptional_entry_queue ewait;
213 wait_queue_head_t *wq; 223 wait_queue_head_t *wq;
214 224
215 init_wait(&ewait.wait); 225 init_wait(&ewait.wait);
216 ewait.wait.func = wake_exceptional_ent 226 ewait.wait.func = wake_exceptional_entry_func;
217 227
218 for (;;) { 228 for (;;) {
219 entry = xas_find_conflict(xas) !! 229 entry = __radix_tree_lookup(&mapping->page_tree, index, NULL,
220 if (!entry || WARN_ON_ONCE(!xa !! 230 &slot);
221 return entry; !! 231 if (!entry || !radix_tree_exceptional_entry(entry) ||
222 if (dax_entry_order(entry) < o !! 232 !slot_locked(mapping, slot)) {
223 return XA_RETRY_ENTRY; !! 233 if (slotp)
224 if (!dax_is_locked(entry)) !! 234 *slotp = slot;
225 return entry; 235 return entry;
>> 236 }
226 237
227 wq = dax_entry_waitqueue(xas, !! 238 wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
228 prepare_to_wait_exclusive(wq, 239 prepare_to_wait_exclusive(wq, &ewait.wait,
229 TASK 240 TASK_UNINTERRUPTIBLE);
230 xas_unlock_irq(xas); !! 241 spin_unlock_irq(&mapping->tree_lock);
231 xas_reset(xas); <<
232 schedule(); 242 schedule();
233 finish_wait(wq, &ewait.wait); 243 finish_wait(wq, &ewait.wait);
234 xas_lock_irq(xas); !! 244 spin_lock_irq(&mapping->tree_lock);
235 } <<
236 } <<
237 <<
238 /* <<
239 * The only thing keeping the address space ar <<
240 * (it's cycled in clear_inode() after removin <<
241 * After we call xas_unlock_irq(), we cannot t <<
242 */ <<
243 static void wait_entry_unlocked(struct xa_stat <<
244 { <<
245 struct wait_exceptional_entry_queue ew <<
246 wait_queue_head_t *wq; <<
247 <<
248 init_wait(&ewait.wait); <<
249 ewait.wait.func = wake_exceptional_ent <<
250 <<
251 wq = dax_entry_waitqueue(xas, entry, & <<
252 /* <<
253 * Unlike get_unlocked_entry() there i <<
254 * path ever successfully retrieves an <<
255 * inode dies. Perform a non-exclusive <<
256 * never successfully performs its own <<
257 */ <<
258 prepare_to_wait(wq, &ewait.wait, TASK_ <<
259 xas_unlock_irq(xas); <<
260 schedule(); <<
261 finish_wait(wq, &ewait.wait); <<
262 } <<
263 <<
264 static void put_unlocked_entry(struct xa_state <<
265 enum dax_wake_m <<
266 { <<
267 if (entry && !dax_is_conflict(entry)) <<
268 dax_wake_entry(xas, entry, mod <<
269 } <<
270 <<
271 /* <<
272 * We used the xa_state to get the entry, but <<
273 * dropped the xa_lock, so we know the xa_stat <<
274 * before use. <<
275 */ <<
276 static void dax_unlock_entry(struct xa_state * <<
277 { <<
278 void *old; <<
279 <<
280 BUG_ON(dax_is_locked(entry)); <<
281 xas_reset(xas); <<
282 xas_lock_irq(xas); <<
283 old = xas_store(xas, entry); <<
284 xas_unlock_irq(xas); <<
285 BUG_ON(!dax_is_locked(old)); <<
286 dax_wake_entry(xas, entry, WAKE_NEXT); <<
287 } <<
288 <<
289 /* <<
290 * Return: The entry stored at this location b <<
291 */ <<
292 static void *dax_lock_entry(struct xa_state *x <<
293 { <<
294 unsigned long v = xa_to_value(entry); <<
295 return xas_store(xas, xa_mk_value(v | <<
296 } <<
297 <<
298 static unsigned long dax_entry_size(void *entr <<
299 { <<
300 if (dax_is_zero_entry(entry)) <<
301 return 0; <<
302 else if (dax_is_empty_entry(entry)) <<
303 return 0; <<
304 else if (dax_is_pmd_entry(entry)) <<
305 return PMD_SIZE; <<
306 else <<
307 return PAGE_SIZE; <<
308 } <<
309 <<
310 static unsigned long dax_end_pfn(void *entry) <<
311 { <<
312 return dax_to_pfn(entry) + dax_entry_s <<
313 } <<
314 <<
315 /* <<
316 * Iterate through all mapped pfns represented <<
317 * 'empty' and 'zero' entries. <<
318 */ <<
319 #define for_each_mapped_pfn(entry, pfn) \ <<
320 for (pfn = dax_to_pfn(entry); \ <<
321 pfn < dax_end_pfn(entr <<
322 <<
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 } 245 }
343 page->share++; <<
344 } 246 }
345 247
346 static inline unsigned long dax_page_share_put !! 248 static void dax_unlock_mapping_entry(struct address_space *mapping,
>> 249 pgoff_t index)
347 { 250 {
348 return --page->share; !! 251 void *entry, **slot;
349 } <<
350 252
351 /* !! 253 spin_lock_irq(&mapping->tree_lock);
352 * When it is called in dax_insert_entry(), th !! 254 entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
353 * whether this entry is shared by multiple fi !! 255 if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
354 * PAGE_MAPPING_DAX_SHARED, and use page->shar !! 256 !slot_locked(mapping, slot))) {
355 */ !! 257 spin_unlock_irq(&mapping->tree_lock);
356 static void dax_associate_entry(void *entry, s <<
357 struct vm_area_struct *vma, un <<
358 { <<
359 unsigned long size = dax_entry_size(en <<
360 int i = 0; <<
361 <<
362 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) <<
363 return; <<
364 <<
365 index = linear_page_index(vma, address <<
366 for_each_mapped_pfn(entry, pfn) { <<
367 struct page *page = pfn_to_pag <<
368 <<
369 if (shared) { <<
370 dax_page_share_get(pag <<
371 } else { <<
372 WARN_ON_ONCE(page->map <<
373 page->mapping = mappin <<
374 page->index = index + <<
375 } <<
376 } <<
377 } <<
378 <<
379 static void dax_disassociate_entry(void *entry <<
380 bool trunc) <<
381 { <<
382 unsigned long pfn; <<
383 <<
384 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) <<
385 return; 258 return;
386 <<
387 for_each_mapped_pfn(entry, pfn) { <<
388 struct page *page = pfn_to_pag <<
389 <<
390 WARN_ON_ONCE(trunc && page_ref <<
391 if (dax_page_is_shared(page)) <<
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; <<
398 page->index = 0; <<
399 } <<
400 } <<
401 <<
402 static struct page *dax_busy_page(void *entry) <<
403 { <<
404 unsigned long pfn; <<
405 <<
406 for_each_mapped_pfn(entry, pfn) { <<
407 struct page *page = pfn_to_pag <<
408 <<
409 if (page_ref_count(page) > 1) <<
410 return page; <<
411 } 259 }
412 return NULL; !! 260 unlock_slot(mapping, slot);
>> 261 spin_unlock_irq(&mapping->tree_lock);
>> 262 dax_wake_mapping_entry_waiter(mapping, index, entry, false);
413 } 263 }
414 264
415 /** !! 265 static void put_locked_mapping_entry(struct address_space *mapping,
416 * dax_lock_folio - Lock the DAX entry corresp !! 266 pgoff_t index, void *entry)
417 * @folio: The folio whose entry we want to lo <<
418 * <<
419 * Context: Process context. <<
420 * Return: A cookie to pass to dax_unlock_foli <<
421 * not be locked. <<
422 */ <<
423 dax_entry_t dax_lock_folio(struct folio *folio <<
424 { 267 {
425 XA_STATE(xas, NULL, 0); !! 268 if (!radix_tree_exceptional_entry(entry)) {
426 void *entry; !! 269 unlock_page(entry);
427 !! 270 put_page(entry);
428 /* Ensure folio->mapping isn't freed w !! 271 } else {
429 rcu_read_lock(); !! 272 dax_unlock_mapping_entry(mapping, index);
430 for (;;) { <<
431 struct address_space *mapping <<
432 <<
433 entry = NULL; <<
434 if (!mapping || !dax_mapping(m <<
435 break; <<
436 <<
437 /* <<
438 * In the device-dax case ther <<
439 * struct dev_pagemap pin is s <<
440 * inode alive, and we assume <<
441 * otherwise we would not have <<
442 * translation. <<
443 */ <<
444 entry = (void *)~0UL; <<
445 if (S_ISCHR(mapping->host->i_m <<
446 break; <<
447 <<
448 xas.xa = &mapping->i_pages; <<
449 xas_lock_irq(&xas); <<
450 if (mapping != folio->mapping) <<
451 xas_unlock_irq(&xas); <<
452 continue; <<
453 } <<
454 xas_set(&xas, folio->index); <<
455 entry = xas_load(&xas); <<
456 if (dax_is_locked(entry)) { <<
457 rcu_read_unlock(); <<
458 wait_entry_unlocked(&x <<
459 rcu_read_lock(); <<
460 continue; <<
461 } <<
462 dax_lock_entry(&xas, entry); <<
463 xas_unlock_irq(&xas); <<
464 break; <<
465 } 273 }
466 rcu_read_unlock(); <<
467 return (dax_entry_t)entry; <<
468 } <<
469 <<
470 void dax_unlock_folio(struct folio *folio, dax <<
471 { <<
472 struct address_space *mapping = folio- <<
473 XA_STATE(xas, &mapping->i_pages, folio <<
474 <<
475 if (S_ISCHR(mapping->host->i_mode)) <<
476 return; <<
477 <<
478 dax_unlock_entry(&xas, (void *)cookie) <<
479 } 274 }
480 275
481 /* 276 /*
482 * dax_lock_mapping_entry - Lock the DAX entry !! 277 * Called when we are done with radix tree entry we looked up via
483 * @mapping: the file's mapping whose entry we !! 278 * get_unlocked_mapping_entry() and which we didn't lock in the end.
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 */ 279 */
490 dax_entry_t dax_lock_mapping_entry(struct addr !! 280 static void put_unlocked_mapping_entry(struct address_space *mapping,
491 struct page **page) !! 281 pgoff_t index, void *entry)
492 { 282 {
493 XA_STATE(xas, NULL, 0); !! 283 if (!radix_tree_exceptional_entry(entry))
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; 284 return;
540 285
541 dax_unlock_entry(&xas, (void *)cookie) !! 286 /* We have to wake up next waiter for the radix tree entry lock */
>> 287 dax_wake_mapping_entry_waiter(mapping, index, entry, false);
542 } 288 }
543 289
544 /* 290 /*
545 * Find page cache entry at given index. If it !! 291 * Find radix tree entry at given index. If it points to a page, return with
546 * with the entry locked. If the page cache do !! 292 * the page locked. If it points to the exceptional entry, return with the
547 * that index, add a locked empty entry. !! 293 * radix tree entry locked. If the radix tree doesn't contain given index,
>> 294 * create empty exceptional entry for the index and return with it locked.
548 * 295 *
549 * When requesting an entry with size DAX_PMD, !! 296 * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
550 * either return that locked entry or will ret !! 297 * either return that locked entry or will return an error. This error will
551 * This will happen if there are any PTE entri !! 298 * happen if there are any 4k entries (either zero pages or DAX entries)
552 * that we are requesting. !! 299 * within the 2MiB range that we are requesting.
553 * 300 *
554 * We always favor PTE entries over PMD entrie !! 301 * We always favor 4k entries over 2MiB entries. There isn't a flow where we
555 * evict PTE entries in order to 'upgrade' the !! 302 * evict 4k entries in order to 'upgrade' them to a 2MiB entry. A 2MiB
556 * insertion will fail if it finds any PTE ent !! 303 * insertion will fail if it finds any 4k entries already in the tree, and a
557 * PTE insertion will cause an existing PMD en !! 304 * 4k insertion will cause an existing 2MiB entry to be unmapped and
558 * downgraded to PTE entries. This happens fo !! 305 * downgraded to 4k entries. This happens for both 2MiB huge zero pages as
559 * well as PMD empty entries. !! 306 * well as 2MiB empty entries.
560 * 307 *
561 * The exception to this downgrade path is for !! 308 * The exception to this downgrade path is for 2MiB DAX PMD entries that have
562 * real storage backing them. We will leave t !! 309 * real storage backing them. We will leave these real 2MiB DAX entries in
563 * the tree, and PTE writes will simply dirty !! 310 * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
564 * 311 *
565 * Note: Unlike filemap_fault() we don't honor 312 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
566 * persistent memory the benefit is doubtful. 313 * persistent memory the benefit is doubtful. We can add that later if we can
567 * show it helps. 314 * show it helps.
568 * <<
569 * On error, this function does not return an <<
570 * a VM_FAULT code, encoded as an xarray inter <<
571 * overlap with xarray value entries. <<
572 */ 315 */
573 static void *grab_mapping_entry(struct xa_stat !! 316 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
574 struct address_space *mapping, !! 317 unsigned long size_flag)
575 { 318 {
576 unsigned long index = xas->xa_index; !! 319 bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
577 bool pmd_downgrade; /* splitting P !! 320 void *entry, **slot;
578 void *entry; <<
579 321
580 retry: !! 322 restart:
581 pmd_downgrade = false; !! 323 spin_lock_irq(&mapping->tree_lock);
582 xas_lock_irq(xas); !! 324 entry = get_unlocked_mapping_entry(mapping, index, &slot);
583 entry = get_unlocked_entry(xas, order) <<
584 325
585 if (entry) { 326 if (entry) {
586 if (dax_is_conflict(entry)) !! 327 if (size_flag & RADIX_DAX_PMD) {
587 goto fallback; !! 328 if (!radix_tree_exceptional_entry(entry) ||
588 if (!xa_is_value(entry)) { !! 329 dax_is_pte_entry(entry)) {
589 xas_set_err(xas, -EIO) !! 330 put_unlocked_mapping_entry(mapping, index,
590 goto out_unlock; !! 331 entry);
591 } !! 332 entry = ERR_PTR(-EEXIST);
592 !! 333 goto out_unlock;
593 if (order == 0) { !! 334 }
594 if (dax_is_pmd_entry(e !! 335 } else { /* trying to grab a PTE entry */
>> 336 if (radix_tree_exceptional_entry(entry) &&
>> 337 dax_is_pmd_entry(entry) &&
595 (dax_is_zero_entry 338 (dax_is_zero_entry(entry) ||
596 dax_is_empty_entr 339 dax_is_empty_entry(entry))) {
597 pmd_downgrade 340 pmd_downgrade = true;
598 } 341 }
599 } 342 }
600 } 343 }
601 344
602 if (pmd_downgrade) { !! 345 /* No entry for given index? Make sure radix tree is big enough. */
603 /* !! 346 if (!entry || pmd_downgrade) {
604 * Make sure 'entry' remains v !! 347 int err;
605 * the i_pages lock. <<
606 */ <<
607 dax_lock_entry(xas, entry); <<
608 348
>> 349 if (pmd_downgrade) {
>> 350 /*
>> 351 * Make sure 'entry' remains valid while we drop
>> 352 * mapping->tree_lock.
>> 353 */
>> 354 entry = lock_slot(mapping, slot);
>> 355 }
>> 356
>> 357 spin_unlock_irq(&mapping->tree_lock);
609 /* 358 /*
610 * Besides huge zero pages the 359 * Besides huge zero pages the only other thing that gets
611 * downgraded are empty entrie 360 * downgraded are empty entries which don't need to be
612 * unmapped. 361 * unmapped.
613 */ 362 */
614 if (dax_is_zero_entry(entry)) !! 363 if (pmd_downgrade && dax_is_zero_entry(entry))
615 xas_unlock_irq(xas); !! 364 unmap_mapping_range(mapping,
616 unmap_mapping_pages(ma !! 365 (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
617 xas->x !! 366
618 PG_PMD !! 367 err = radix_tree_preload(
619 xas_reset(xas); !! 368 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
620 xas_lock_irq(xas); !! 369 if (err) {
>> 370 if (pmd_downgrade)
>> 371 put_locked_mapping_entry(mapping, index, entry);
>> 372 return ERR_PTR(err);
621 } 373 }
>> 374 spin_lock_irq(&mapping->tree_lock);
622 375
623 dax_disassociate_entry(entry, !! 376 if (!entry) {
624 xas_store(xas, NULL); /* und !! 377 /*
625 dax_wake_entry(xas, entry, WAK !! 378 * We needed to drop the page_tree lock while calling
626 mapping->nrpages -= PG_PMD_NR; !! 379 * radix_tree_preload() and we didn't have an entry to
627 entry = NULL; !! 380 * lock. See if another thread inserted an entry at
628 xas_set(xas, index); !! 381 * our index during this time.
629 } !! 382 */
>> 383 entry = __radix_tree_lookup(&mapping->page_tree, index,
>> 384 NULL, &slot);
>> 385 if (entry) {
>> 386 radix_tree_preload_end();
>> 387 spin_unlock_irq(&mapping->tree_lock);
>> 388 goto restart;
>> 389 }
>> 390 }
630 391
631 if (entry) { !! 392 if (pmd_downgrade) {
632 dax_lock_entry(xas, entry); !! 393 radix_tree_delete(&mapping->page_tree, index);
633 } else { !! 394 mapping->nrexceptional--;
634 unsigned long flags = DAX_EMPT !! 395 dax_wake_mapping_entry_waiter(mapping, index, entry,
>> 396 true);
>> 397 }
>> 398
>> 399 entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
635 400
636 if (order > 0) !! 401 err = __radix_tree_insert(&mapping->page_tree, index,
637 flags |= DAX_PMD; !! 402 dax_radix_order(entry), entry);
638 entry = dax_make_entry(pfn_to_ !! 403 radix_tree_preload_end();
639 dax_lock_entry(xas, entry); !! 404 if (err) {
640 if (xas_error(xas)) !! 405 spin_unlock_irq(&mapping->tree_lock);
641 goto out_unlock; !! 406 /*
642 mapping->nrpages += 1UL << ord !! 407 * Our insertion of a DAX entry failed, most likely
643 } !! 408 * because we were inserting a PMD entry and it
644 !! 409 * collided with a PTE sized entry at a different
645 out_unlock: !! 410 * index in the PMD range. We haven't inserted
646 xas_unlock_irq(xas); !! 411 * anything into the radix tree and have no waiters to
647 if (xas_nomem(xas, mapping_gfp_mask(ma !! 412 * wake.
648 goto retry; !! 413 */
649 if (xas->xa_node == XA_ERROR(-ENOMEM)) !! 414 return ERR_PTR(err);
650 return xa_mk_internal(VM_FAULT !! 415 }
651 if (xas_error(xas)) !! 416 /* Good, we have inserted empty locked entry into the tree. */
652 return xa_mk_internal(VM_FAULT !! 417 mapping->nrexceptional++;
>> 418 spin_unlock_irq(&mapping->tree_lock);
>> 419 return entry;
>> 420 }
>> 421 /* Normal page in radix tree? */
>> 422 if (!radix_tree_exceptional_entry(entry)) {
>> 423 struct page *page = entry;
>> 424
>> 425 get_page(page);
>> 426 spin_unlock_irq(&mapping->tree_lock);
>> 427 lock_page(page);
>> 428 /* Page got truncated? Retry... */
>> 429 if (unlikely(page->mapping != mapping)) {
>> 430 unlock_page(page);
>> 431 put_page(page);
>> 432 goto restart;
>> 433 }
>> 434 return page;
>> 435 }
>> 436 entry = lock_slot(mapping, slot);
>> 437 out_unlock:
>> 438 spin_unlock_irq(&mapping->tree_lock);
653 return entry; 439 return entry;
654 fallback: <<
655 xas_unlock_irq(xas); <<
656 return xa_mk_internal(VM_FAULT_FALLBAC <<
657 } 440 }
658 441
659 /** !! 442 /*
660 * dax_layout_busy_page_range - find first pin !! 443 * We do not necessarily hold the mapping->tree_lock when we call this
661 * @mapping: address space to scan for a page !! 444 * function so it is possible that 'entry' is no longer a valid item in the
662 * @start: Starting offset. Page containing 's !! 445 * radix tree. This is okay because all we really need to do is to find the
663 * @end: End offset. Page containing 'end' is !! 446 * correct waitqueue where tasks might be waiting for that old 'entry' and
664 * pages from 'start' till the end of fi !! 447 * wake them.
665 * <<
666 * DAX requires ZONE_DEVICE mapped pages. Thes <<
667 * 'onlined' to the page allocator so they are <<
668 * page->count == 1. A filesystem uses this in <<
669 * any page in the mapping is busy, i.e. for D <<
670 * get_user_pages() usages. <<
671 * <<
672 * It is expected that the filesystem is holdi <<
673 * establishment of new mappings in this addre <<
674 * to be able to run unmap_mapping_range() and <<
675 * mapping_mapped() becoming true. <<
676 */ 448 */
677 struct page *dax_layout_busy_page_range(struct !! 449 void dax_wake_mapping_entry_waiter(struct address_space *mapping,
678 loff_t !! 450 pgoff_t index, void *entry, bool wake_all)
679 { 451 {
680 void *entry; !! 452 struct exceptional_entry_key key;
681 unsigned int scanned = 0; !! 453 wait_queue_head_t *wq;
682 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 <<
687 /* <<
688 * In the 'limited' case get_user_page <<
689 */ <<
690 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) <<
691 return NULL; <<
692 454
693 if (!dax_mapping(mapping) || !mapping_ !! 455 wq = dax_entry_waitqueue(mapping, index, entry, &key);
694 return NULL; <<
695 456
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 /* 457 /*
702 * If we race get_user_pages_fast() he !! 458 * Checking for locked entry and prepare_to_wait_exclusive() happens
703 * elevated page count in the iteratio !! 459 * under mapping->tree_lock, ditto for entry handling in our callers.
704 * get_user_pages_fast() will see that !! 460 * So at this point all tasks that could have seen our entry locked
705 * against is no longer mapped in the !! 461 * must be in the waitqueue and the following check will see them.
706 * get_user_pages() slow path. The sl <<
707 * pte_lock() and pmd_lock(). New refe <<
708 * holding those locks, and unmap_mapp <<
709 * pte or pmd without holding the resp <<
710 * guaranteed to either see new refere <<
711 * references from being established. <<
712 */ 462 */
713 unmap_mapping_pages(mapping, start_idx !! 463 if (waitqueue_active(wq))
714 !! 464 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
715 xas_lock_irq(&xas); <<
716 xas_for_each(&xas, entry, end_idx) { <<
717 if (WARN_ON_ONCE(!xa_is_value( <<
718 continue; <<
719 if (unlikely(dax_is_locked(ent <<
720 entry = get_unlocked_e <<
721 if (entry) <<
722 page = dax_busy_page(e <<
723 put_unlocked_entry(&xas, entry <<
724 if (page) <<
725 break; <<
726 if (++scanned % XA_CHECK_SCHED <<
727 continue; <<
728 <<
729 xas_pause(&xas); <<
730 xas_unlock_irq(&xas); <<
731 cond_resched(); <<
732 xas_lock_irq(&xas); <<
733 } <<
734 xas_unlock_irq(&xas); <<
735 return page; <<
736 } <<
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 } 465 }
743 EXPORT_SYMBOL_GPL(dax_layout_busy_page); <<
744 466
745 static int __dax_invalidate_entry(struct addre !! 467 static int __dax_invalidate_mapping_entry(struct address_space *mapping,
746 pgof 468 pgoff_t index, bool trunc)
747 { 469 {
748 XA_STATE(xas, &mapping->i_pages, index <<
749 int ret = 0; 470 int ret = 0;
750 void *entry; 471 void *entry;
>> 472 struct radix_tree_root *page_tree = &mapping->page_tree;
751 473
752 xas_lock_irq(&xas); !! 474 spin_lock_irq(&mapping->tree_lock);
753 entry = get_unlocked_entry(&xas, 0); !! 475 entry = get_unlocked_mapping_entry(mapping, index, NULL);
754 if (!entry || WARN_ON_ONCE(!xa_is_valu !! 476 if (!entry || !radix_tree_exceptional_entry(entry))
755 goto out; 477 goto out;
756 if (!trunc && 478 if (!trunc &&
757 (xas_get_mark(&xas, PAGECACHE_TAG_ !! 479 (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
758 xas_get_mark(&xas, PAGECACHE_TAG_ !! 480 radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)))
759 goto out; 481 goto out;
760 dax_disassociate_entry(entry, mapping, !! 482 radix_tree_delete(page_tree, index);
761 xas_store(&xas, NULL); !! 483 mapping->nrexceptional--;
762 mapping->nrpages -= 1UL << dax_entry_o <<
763 ret = 1; 484 ret = 1;
764 out: 485 out:
765 put_unlocked_entry(&xas, entry, WAKE_A !! 486 put_unlocked_mapping_entry(mapping, index, entry);
766 xas_unlock_irq(&xas); !! 487 spin_unlock_irq(&mapping->tree_lock);
767 return ret; 488 return ret;
768 } 489 }
769 <<
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 /* 490 /*
798 * Delete DAX entry at @index from @mapping. !! 491 * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
799 * to be unlocked before deleting it. !! 492 * entry to get unlocked before deleting it.
800 */ 493 */
801 int dax_delete_mapping_entry(struct address_sp 494 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
802 { 495 {
803 int ret = __dax_invalidate_entry(mappi !! 496 int ret = __dax_invalidate_mapping_entry(mapping, index, true);
804 497
805 /* 498 /*
806 * This gets called from truncate / pu 499 * This gets called from truncate / punch_hole path. As such, the caller
807 * must hold locks protecting against 500 * must hold locks protecting against concurrent modifications of the
808 * page cache (usually fs-private i_mm !! 501 * radix tree (usually fs-private i_mmap_sem for writing). Since the
809 * caller has seen a DAX entry for thi !! 502 * caller has seen exceptional entry for this index, we better find it
810 * at that index as well... 503 * at that index as well...
811 */ 504 */
812 WARN_ON_ONCE(!ret); 505 WARN_ON_ONCE(!ret);
813 return ret; 506 return ret;
814 } 507 }
815 508
816 /* 509 /*
817 * Invalidate DAX entry if it is clean. !! 510 * Invalidate exceptional DAX entry if it is clean.
818 */ 511 */
819 int dax_invalidate_mapping_entry_sync(struct a 512 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
820 pgoff_t 513 pgoff_t index)
821 { 514 {
822 return __dax_invalidate_entry(mapping, !! 515 return __dax_invalidate_mapping_entry(mapping, index, false);
823 } 516 }
824 517
825 static pgoff_t dax_iomap_pgoff(const struct io !! 518 /*
>> 519 * The user has performed a load from a hole in the file. Allocating
>> 520 * a new page in the file would cause excessive storage usage for
>> 521 * workloads with sparse files. We allocate a page cache page instead.
>> 522 * We'll kick it out of the page cache if it's ever written to,
>> 523 * otherwise it will simply fall out of the page cache under memory
>> 524 * pressure without ever having been dirtied.
>> 525 */
>> 526 static int dax_load_hole(struct address_space *mapping, void **entry,
>> 527 struct vm_fault *vmf)
826 { 528 {
827 return PHYS_PFN(iomap->addr + (pos & P !! 529 struct page *page;
828 } !! 530 int ret;
829 531
830 static int copy_cow_page_dax(struct vm_fault * !! 532 /* Hole page already exists? Return it... */
831 { !! 533 if (!radix_tree_exceptional_entry(*entry)) {
832 pgoff_t pgoff = dax_iomap_pgoff(&iter- !! 534 page = *entry;
833 void *vto, *kaddr; !! 535 goto out;
834 long rc; !! 536 }
835 int id; <<
836 537
837 id = dax_read_lock(); !! 538 /* This will replace locked radix tree entry with a hole page */
838 rc = dax_direct_access(iter->iomap.dax !! 539 page = find_or_create_page(mapping, vmf->pgoff,
839 &kaddr, NULL); !! 540 vmf->gfp_mask | __GFP_ZERO);
840 if (rc < 0) { !! 541 if (!page)
841 dax_read_unlock(id); !! 542 return VM_FAULT_OOM;
842 return rc; !! 543 out:
>> 544 vmf->page = page;
>> 545 ret = finish_fault(vmf);
>> 546 vmf->page = NULL;
>> 547 *entry = page;
>> 548 if (!ret) {
>> 549 /* Grab reference for PTE that is now referencing the page */
>> 550 get_page(page);
>> 551 return VM_FAULT_NOPAGE;
843 } 552 }
844 vto = kmap_atomic(vmf->cow_page); !! 553 return ret;
845 copy_user_page(vto, kaddr, vmf->addres <<
846 kunmap_atomic(vto); <<
847 dax_read_unlock(id); <<
848 return 0; <<
849 } 554 }
850 555
851 /* !! 556 static int copy_user_dax(struct block_device *bdev, sector_t sector, size_t size,
852 * MAP_SYNC on a dax mapping guarantees dirty !! 557 struct page *to, unsigned long vaddr)
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 { 558 {
858 return (iter->flags & IOMAP_WRITE) && !! 559 struct blk_dax_ctl dax = {
859 (iter->iomap.flags & IOMAP_F_D !! 560 .sector = sector,
>> 561 .size = size,
>> 562 };
>> 563 void *vto;
>> 564
>> 565 if (dax_map_atomic(bdev, &dax) < 0)
>> 566 return PTR_ERR(dax.addr);
>> 567 vto = kmap_atomic(to);
>> 568 copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
>> 569 kunmap_atomic(vto);
>> 570 dax_unmap_atomic(bdev, &dax);
>> 571 return 0;
860 } 572 }
861 573
862 /* 574 /*
863 * By this point grab_mapping_entry() has ensu 575 * 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 576 * 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 577 * 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 578 * already in the tree, we will skip the insertion and just dirty the PMD as
867 * appropriate. 579 * appropriate.
868 */ 580 */
869 static void *dax_insert_entry(struct xa_state !! 581 static void *dax_insert_mapping_entry(struct address_space *mapping,
870 const struct iomap_iter *iter, !! 582 struct vm_fault *vmf,
871 unsigned long flags) !! 583 void *entry, sector_t sector,
>> 584 unsigned long flags)
872 { 585 {
873 struct address_space *mapping = vmf->v !! 586 struct radix_tree_root *page_tree = &mapping->page_tree;
874 void *new_entry = dax_make_entry(pfn, !! 587 int error = 0;
875 bool write = iter->flags & IOMAP_WRITE !! 588 bool hole_fill = false;
876 bool dirty = write && !dax_fault_is_sy !! 589 void *new_entry;
877 bool shared = iter->iomap.flags & IOMA !! 590 pgoff_t index = vmf->pgoff;
878 591
879 if (dirty) !! 592 if (vmf->flags & FAULT_FLAG_WRITE)
880 __mark_inode_dirty(mapping->ho 593 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
881 594
882 if (shared || (dax_is_zero_entry(entry !! 595 /* Replacing hole page with block mapping? */
883 unsigned long index = xas->xa_ !! 596 if (!radix_tree_exceptional_entry(entry)) {
884 /* we are replacing a zero pag !! 597 hole_fill = true;
885 if (dax_is_pmd_entry(entry)) <<
886 unmap_mapping_pages(ma <<
887 PG_PMD <<
888 else /* pte entry */ <<
889 unmap_mapping_pages(ma <<
890 } <<
891 <<
892 xas_reset(xas); <<
893 xas_lock_irq(xas); <<
894 if (shared || dax_is_zero_entry(entry) <<
895 void *old; <<
896 <<
897 dax_disassociate_entry(entry, <<
898 dax_associate_entry(new_entry, <<
899 shared); <<
900 /* 598 /*
901 * Only swap our new entry int !! 599 * Unmap the page now before we remove it from page cache below.
>> 600 * The page is locked so it cannot be faulted in again.
>> 601 */
>> 602 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
>> 603 PAGE_SIZE, 0);
>> 604 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
>> 605 if (error)
>> 606 return ERR_PTR(error);
>> 607 } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) {
>> 608 /* replacing huge zero page with PMD block mapping */
>> 609 unmap_mapping_range(mapping,
>> 610 (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
>> 611 }
>> 612
>> 613 spin_lock_irq(&mapping->tree_lock);
>> 614 new_entry = dax_radix_locked_entry(sector, flags);
>> 615
>> 616 if (hole_fill) {
>> 617 __delete_from_page_cache(entry, NULL);
>> 618 /* Drop pagecache reference */
>> 619 put_page(entry);
>> 620 error = __radix_tree_insert(page_tree, index,
>> 621 dax_radix_order(new_entry), new_entry);
>> 622 if (error) {
>> 623 new_entry = ERR_PTR(error);
>> 624 goto unlock;
>> 625 }
>> 626 mapping->nrexceptional++;
>> 627 } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
>> 628 /*
>> 629 * Only swap our new entry into the radix tree if the current
902 * entry is a zero page or an 630 * entry is a zero page or an empty entry. If a normal PTE or
903 * PMD entry is already in the !! 631 * PMD entry is already in the tree, we leave it alone. This
904 * means that if we are trying 632 * means that if we are trying to insert a PTE and the
905 * existing entry is a PMD, we 633 * existing entry is a PMD, we will just leave the PMD in the
906 * tree and dirty it if necess 634 * tree and dirty it if necessary.
907 */ 635 */
908 old = dax_lock_entry(xas, new_ !! 636 struct radix_tree_node *node;
909 WARN_ON_ONCE(old != xa_mk_valu !! 637 void **slot;
910 DAX_LO !! 638 void *ret;
911 entry = new_entry; !! 639
912 } else { !! 640 ret = __radix_tree_lookup(page_tree, index, &node, &slot);
913 xas_load(xas); /* Walk the xa !! 641 WARN_ON_ONCE(ret != entry);
>> 642 __radix_tree_replace(page_tree, node, slot,
>> 643 new_entry, NULL, NULL);
914 } 644 }
>> 645 if (vmf->flags & FAULT_FLAG_WRITE)
>> 646 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
>> 647 unlock:
>> 648 spin_unlock_irq(&mapping->tree_lock);
>> 649 if (hole_fill) {
>> 650 radix_tree_preload_end();
>> 651 /*
>> 652 * We don't need hole page anymore, it has been replaced with
>> 653 * locked radix tree entry now.
>> 654 */
>> 655 if (mapping->a_ops->freepage)
>> 656 mapping->a_ops->freepage(entry);
>> 657 unlock_page(entry);
>> 658 put_page(entry);
>> 659 }
>> 660 return new_entry;
>> 661 }
915 662
916 if (dirty) !! 663 static inline unsigned long
917 xas_set_mark(xas, PAGECACHE_TA !! 664 pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
918 !! 665 {
919 if (write && shared) !! 666 unsigned long address;
920 xas_set_mark(xas, PAGECACHE_TA <<
921 667
922 xas_unlock_irq(xas); !! 668 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
923 return entry; !! 669 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
>> 670 return address;
924 } 671 }
925 672
926 static int dax_writeback_one(struct xa_state * !! 673 /* Walk all mappings of a given index of a file and writeprotect them */
927 struct address_space *mapping, !! 674 static void dax_mapping_entry_mkclean(struct address_space *mapping,
>> 675 pgoff_t index, unsigned long pfn)
928 { 676 {
929 unsigned long pfn, index, count, end; <<
930 long ret = 0; <<
931 struct vm_area_struct *vma; 677 struct vm_area_struct *vma;
>> 678 pte_t pte, *ptep = NULL;
>> 679 pmd_t *pmdp = NULL;
>> 680 spinlock_t *ptl;
>> 681 bool changed;
>> 682
>> 683 i_mmap_lock_read(mapping);
>> 684 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
>> 685 unsigned long address;
>> 686
>> 687 cond_resched();
>> 688
>> 689 if (!(vma->vm_flags & VM_SHARED))
>> 690 continue;
>> 691
>> 692 address = pgoff_address(index, vma);
>> 693 changed = false;
>> 694 if (follow_pte_pmd(vma->vm_mm, address, &ptep, &pmdp, &ptl))
>> 695 continue;
>> 696
>> 697 if (pmdp) {
>> 698 #ifdef CONFIG_FS_DAX_PMD
>> 699 pmd_t pmd;
>> 700
>> 701 if (pfn != pmd_pfn(*pmdp))
>> 702 goto unlock_pmd;
>> 703 if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
>> 704 goto unlock_pmd;
>> 705
>> 706 flush_cache_page(vma, address, pfn);
>> 707 pmd = pmdp_huge_clear_flush(vma, address, pmdp);
>> 708 pmd = pmd_wrprotect(pmd);
>> 709 pmd = pmd_mkclean(pmd);
>> 710 set_pmd_at(vma->vm_mm, address, pmdp, pmd);
>> 711 changed = true;
>> 712 unlock_pmd:
>> 713 spin_unlock(ptl);
>> 714 #endif
>> 715 } else {
>> 716 if (pfn != pte_pfn(*ptep))
>> 717 goto unlock_pte;
>> 718 if (!pte_dirty(*ptep) && !pte_write(*ptep))
>> 719 goto unlock_pte;
>> 720
>> 721 flush_cache_page(vma, address, pfn);
>> 722 pte = ptep_clear_flush(vma, address, ptep);
>> 723 pte = pte_wrprotect(pte);
>> 724 pte = pte_mkclean(pte);
>> 725 set_pte_at(vma->vm_mm, address, ptep, pte);
>> 726 changed = true;
>> 727 unlock_pte:
>> 728 pte_unmap_unlock(ptep, ptl);
>> 729 }
>> 730
>> 731 if (changed)
>> 732 mmu_notifier_invalidate_page(vma->vm_mm, address);
>> 733 }
>> 734 i_mmap_unlock_read(mapping);
>> 735 }
>> 736
>> 737 static int dax_writeback_one(struct block_device *bdev,
>> 738 struct address_space *mapping, pgoff_t index, void *entry)
>> 739 {
>> 740 struct radix_tree_root *page_tree = &mapping->page_tree;
>> 741 struct blk_dax_ctl dax;
>> 742 void *entry2, **slot;
>> 743 int ret = 0;
932 744
933 /* 745 /*
934 * A page got tagged dirty in DAX mapp 746 * A page got tagged dirty in DAX mapping? Something is seriously
935 * wrong. 747 * wrong.
936 */ 748 */
937 if (WARN_ON(!xa_is_value(entry))) !! 749 if (WARN_ON(!radix_tree_exceptional_entry(entry)))
938 return -EIO; 750 return -EIO;
939 751
940 if (unlikely(dax_is_locked(entry))) { !! 752 spin_lock_irq(&mapping->tree_lock);
941 void *old_entry = entry; !! 753 entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
942 !! 754 /* Entry got punched out / reallocated? */
943 entry = get_unlocked_entry(xas !! 755 if (!entry2 || !radix_tree_exceptional_entry(entry2))
944 !! 756 goto put_unlocked;
945 /* Entry got punched out / rea !! 757 /*
946 if (!entry || WARN_ON_ONCE(!xa !! 758 * Entry got reallocated elsewhere? No need to writeback. We have to
947 goto put_unlocked; !! 759 * compare sectors as we must not bail out due to difference in lockbit
948 /* !! 760 * or entry type.
949 * Entry got reallocated elsew !! 761 */
950 * We have to compare pfns as !! 762 if (dax_radix_sector(entry2) != dax_radix_sector(entry))
951 * difference in lockbit or en !! 763 goto put_unlocked;
952 */ !! 764 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
953 if (dax_to_pfn(old_entry) != d !! 765 dax_is_zero_entry(entry))) {
954 goto put_unlocked; !! 766 ret = -EIO;
955 if (WARN_ON_ONCE(dax_is_empty_ !! 767 goto put_unlocked;
956 dax_is <<
957 ret = -EIO; <<
958 goto put_unlocked; <<
959 } <<
960 <<
961 /* Another fsync thread may ha <<
962 if (!xas_get_mark(xas, PAGECAC <<
963 goto put_unlocked; <<
964 } 768 }
965 769
>> 770 /* Another fsync thread may have already written back this entry */
>> 771 if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
>> 772 goto put_unlocked;
966 /* Lock the entry to serialize with pa 773 /* Lock the entry to serialize with page faults */
967 dax_lock_entry(xas, entry); !! 774 entry = lock_slot(mapping, slot);
968 <<
969 /* 775 /*
970 * We can clear the tag now but we hav 776 * We can clear the tag now but we have to be careful so that concurrent
971 * dax_writeback_one() calls for the s 777 * dax_writeback_one() calls for the same index cannot finish before we
972 * actually flush the caches. This is 778 * actually flush the caches. This is achieved as the calls will look
973 * at the entry only under the i_pages !! 779 * at the entry only under tree_lock and once they do that they will
974 * they will see the entry locked and !! 780 * see the entry locked and wait for it to unlock.
975 */ 781 */
976 xas_clear_mark(xas, PAGECACHE_TAG_TOWR !! 782 radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
977 xas_unlock_irq(xas); !! 783 spin_unlock_irq(&mapping->tree_lock);
978 784
979 /* 785 /*
980 * If dax_writeback_mapping_range() wa !! 786 * Even if dax_writeback_mapping_range() was given a wbc->range_start
981 * in the middle of a PMD, the 'index' !! 787 * in the middle of a PMD, the 'index' we are given will be aligned to
982 * aligned to the start of the PMD. !! 788 * the start index of the PMD, as will the sector we pull from
983 * This allows us to flush for PMD_SIZ !! 789 * 'entry'. This allows us to flush for PMD_SIZE and not have to
984 * partial PMD writebacks. !! 790 * worry about partial PMD writebacks.
985 */ 791 */
986 pfn = dax_to_pfn(entry); !! 792 dax.sector = dax_radix_sector(entry);
987 count = 1UL << dax_entry_order(entry); !! 793 dax.size = PAGE_SIZE << dax_radix_order(entry);
988 index = xas->xa_index & ~(count - 1); <<
989 end = index + count - 1; <<
990 794
991 /* Walk all mappings of a given index !! 795 /*
992 i_mmap_lock_read(mapping); !! 796 * We cannot hold tree_lock while calling dax_map_atomic() because it
993 vma_interval_tree_foreach(vma, &mappin !! 797 * eventually calls cond_resched().
994 pfn_mkclean_range(pfn, count, !! 798 */
995 cond_resched(); !! 799 ret = dax_map_atomic(bdev, &dax);
>> 800 if (ret < 0) {
>> 801 put_locked_mapping_entry(mapping, index, entry);
>> 802 return ret;
>> 803 }
>> 804
>> 805 if (WARN_ON_ONCE(ret < dax.size)) {
>> 806 ret = -EIO;
>> 807 goto unmap;
996 } 808 }
997 i_mmap_unlock_read(mapping); <<
998 809
999 dax_flush(dax_dev, page_address(pfn_to !! 810 dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(dax.pfn));
>> 811 wb_cache_pmem(dax.addr, dax.size);
1000 /* 812 /*
1001 * After we have flushed the cache, w 813 * After we have flushed the cache, we can clear the dirty tag. There
1002 * cannot be new dirty data in the pf 814 * cannot be new dirty data in the pfn after the flush has completed as
1003 * the pfn mappings are writeprotecte 815 * the pfn mappings are writeprotected and fault waits for mapping
1004 * entry lock. 816 * entry lock.
1005 */ 817 */
1006 xas_reset(xas); !! 818 spin_lock_irq(&mapping->tree_lock);
1007 xas_lock_irq(xas); !! 819 radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY);
1008 xas_store(xas, entry); !! 820 spin_unlock_irq(&mapping->tree_lock);
1009 xas_clear_mark(xas, PAGECACHE_TAG_DIR !! 821 unmap:
1010 dax_wake_entry(xas, entry, WAKE_NEXT) !! 822 dax_unmap_atomic(bdev, &dax);
1011 !! 823 put_locked_mapping_entry(mapping, index, entry);
1012 trace_dax_writeback_one(mapping->host <<
1013 return ret; 824 return ret;
1014 825
1015 put_unlocked: 826 put_unlocked:
1016 put_unlocked_entry(xas, entry, WAKE_N !! 827 put_unlocked_mapping_entry(mapping, index, entry2);
>> 828 spin_unlock_irq(&mapping->tree_lock);
1017 return ret; 829 return ret;
1018 } 830 }
1019 831
1020 /* 832 /*
1021 * Flush the mapping to the persistent domain 833 * Flush the mapping to the persistent domain within the byte range of [start,
1022 * end]. This is required by data integrity o 834 * end]. This is required by data integrity operations to ensure file data is
1023 * on persistent storage prior to completion 835 * on persistent storage prior to completion of the operation.
1024 */ 836 */
1025 int dax_writeback_mapping_range(struct addres 837 int dax_writeback_mapping_range(struct address_space *mapping,
1026 struct dax_device *dax_dev, s !! 838 struct block_device *bdev, struct writeback_control *wbc)
1027 { 839 {
1028 XA_STATE(xas, &mapping->i_pages, wbc- <<
1029 struct inode *inode = mapping->host; 840 struct inode *inode = mapping->host;
1030 pgoff_t end_index = wbc->range_end >> !! 841 pgoff_t start_index, end_index;
1031 void *entry; !! 842 pgoff_t indices[PAGEVEC_SIZE];
1032 int ret = 0; !! 843 struct pagevec pvec;
1033 unsigned int scanned = 0; !! 844 bool done = false;
>> 845 int i, ret = 0;
1034 846
1035 if (WARN_ON_ONCE(inode->i_blkbits != 847 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
1036 return -EIO; 848 return -EIO;
1037 849
1038 if (mapping_empty(mapping) || wbc->sy !! 850 if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
1039 return 0; 851 return 0;
1040 852
1041 trace_dax_writeback_range(inode, xas. !! 853 start_index = wbc->range_start >> PAGE_SHIFT;
1042 !! 854 end_index = wbc->range_end >> PAGE_SHIFT;
1043 tag_pages_for_writeback(mapping, xas. <<
1044 <<
1045 xas_lock_irq(&xas); <<
1046 xas_for_each_marked(&xas, entry, end_ <<
1047 ret = dax_writeback_one(&xas, <<
1048 if (ret < 0) { <<
1049 mapping_set_error(map <<
1050 break; <<
1051 } <<
1052 if (++scanned % XA_CHECK_SCHE <<
1053 continue; <<
1054 <<
1055 xas_pause(&xas); <<
1056 xas_unlock_irq(&xas); <<
1057 cond_resched(); <<
1058 xas_lock_irq(&xas); <<
1059 } <<
1060 xas_unlock_irq(&xas); <<
1061 trace_dax_writeback_range_done(inode, <<
1062 return ret; <<
1063 } <<
1064 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range <<
1065 <<
1066 static int dax_iomap_direct_access(const stru <<
1067 size_t size, void **kaddr, pf <<
1068 { <<
1069 pgoff_t pgoff = dax_iomap_pgoff(iomap <<
1070 int id, rc = 0; <<
1071 long length; <<
1072 <<
1073 id = dax_read_lock(); <<
1074 length = dax_direct_access(iomap->dax <<
1075 DAX_ACCESS <<
1076 if (length < 0) { <<
1077 rc = length; <<
1078 goto out; <<
1079 } <<
1080 if (!pfnp) <<
1081 goto out_check_addr; <<
1082 rc = -EINVAL; <<
1083 if (PFN_PHYS(length) < size) <<
1084 goto out; <<
1085 if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(s <<
1086 goto out; <<
1087 /* For larger pages we need devmap */ <<
1088 if (length > 1 && !pfn_t_devmap(*pfnp <<
1089 goto out; <<
1090 rc = 0; <<
1091 <<
1092 out_check_addr: <<
1093 if (!kaddr) <<
1094 goto out; <<
1095 if (!*kaddr) <<
1096 rc = -EFAULT; <<
1097 out: <<
1098 dax_read_unlock(id); <<
1099 return rc; <<
1100 } <<
1101 855
1102 /** !! 856 tag_pages_for_writeback(mapping, start_index, end_index);
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 857
1134 if (!zero_edge) { !! 858 pagevec_init(&pvec, 0);
1135 ret = dax_iomap_direct_access !! 859 while (!done) {
1136 if (ret) !! 860 pvec.nr = find_get_entries_tag(mapping, start_index,
1137 return dax_mem2blk_er !! 861 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
1138 } !! 862 pvec.pages, indices);
1139 863
1140 if (copy_all) { !! 864 if (pvec.nr == 0)
1141 if (zero_edge) !! 865 break;
1142 memset(daddr, 0, size <<
1143 else <<
1144 ret = copy_mc_to_kern <<
1145 goto out; <<
1146 } <<
1147 866
1148 /* Copy the head part of the range */ !! 867 for (i = 0; i < pvec.nr; i++) {
1149 if (head_off) { !! 868 if (indices[i] > end_index) {
1150 if (zero_edge) !! 869 done = true;
1151 memset(daddr, 0, head !! 870 break;
1152 else { !! 871 }
1153 ret = copy_mc_to_kern <<
1154 if (ret) <<
1155 return -EIO; <<
1156 } <<
1157 } <<
1158 872
1159 /* Copy the tail part of the range */ !! 873 ret = dax_writeback_one(bdev, mapping, indices[i],
1160 if (end < pg_end) { !! 874 pvec.pages[i]);
1161 loff_t tail_off = head_off + !! 875 if (ret < 0)
1162 loff_t tail_len = pg_end - en !! 876 return ret;
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 } 877 }
>> 878 start_index = indices[pvec.nr - 1] + 1;
1172 } 879 }
1173 out: !! 880 return 0;
1174 if (zero_edge) <<
1175 dax_flush(srcmap->dax_dev, da <<
1176 return ret ? -EIO : 0; <<
1177 } 881 }
>> 882 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
1178 883
1179 /* !! 884 static int dax_insert_mapping(struct address_space *mapping,
1180 * The user has performed a load from a hole !! 885 struct block_device *bdev, sector_t sector, size_t size,
1181 * page in the file would cause excessive sto !! 886 void **entryp, struct vm_area_struct *vma, struct vm_fault *vmf)
1182 * sparse files. Instead we insert a read-on <<
1183 * If this page is ever written to we will re <<
1184 * point to real DAX storage instead. <<
1185 */ <<
1186 static vm_fault_t dax_load_hole(struct xa_sta <<
1187 const struct iomap_iter *iter <<
1188 { 887 {
1189 struct inode *inode = iter->inode; <<
1190 unsigned long vaddr = vmf->address; 888 unsigned long vaddr = vmf->address;
1191 pfn_t pfn = pfn_to_pfn_t(my_zero_pfn( !! 889 struct blk_dax_ctl dax = {
1192 vm_fault_t ret; !! 890 .sector = sector,
>> 891 .size = size,
>> 892 };
>> 893 void *ret;
>> 894 void *entry = *entryp;
1193 895
1194 *entry = dax_insert_entry(xas, vmf, i !! 896 if (dax_map_atomic(bdev, &dax) < 0)
>> 897 return PTR_ERR(dax.addr);
>> 898 dax_unmap_atomic(bdev, &dax);
>> 899
>> 900 ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector, 0);
>> 901 if (IS_ERR(ret))
>> 902 return PTR_ERR(ret);
>> 903 *entryp = ret;
1195 904
1196 ret = vmf_insert_mixed(vmf->vma, vadd !! 905 return vm_insert_mixed(vma, vaddr, dax.pfn);
1197 trace_dax_load_hole(inode, vmf, ret); <<
1198 return ret; <<
1199 } 906 }
1200 907
1201 #ifdef CONFIG_FS_DAX_PMD !! 908 /**
1202 static vm_fault_t dax_pmd_load_hole(struct xa !! 909 * dax_pfn_mkwrite - handle first write to DAX page
1203 const struct iomap_iter *iter !! 910 * @vmf: The description of the fault
>> 911 */
>> 912 int dax_pfn_mkwrite(struct vm_fault *vmf)
1204 { 913 {
1205 struct address_space *mapping = vmf-> !! 914 struct file *file = vmf->vma->vm_file;
1206 unsigned long pmd_addr = vmf->address !! 915 struct address_space *mapping = file->f_mapping;
1207 struct vm_area_struct *vma = vmf->vma !! 916 void *entry, **slot;
1208 struct inode *inode = mapping->host; !! 917 pgoff_t index = vmf->pgoff;
1209 pgtable_t pgtable = NULL; !! 918
1210 struct folio *zero_folio; !! 919 spin_lock_irq(&mapping->tree_lock);
1211 spinlock_t *ptl; !! 920 entry = get_unlocked_mapping_entry(mapping, index, &slot);
1212 pmd_t pmd_entry; !! 921 if (!entry || !radix_tree_exceptional_entry(entry)) {
1213 pfn_t pfn; !! 922 if (entry)
1214 !! 923 put_unlocked_mapping_entry(mapping, index, entry);
1215 zero_folio = mm_get_huge_zero_folio(v !! 924 spin_unlock_irq(&mapping->tree_lock);
1216 !! 925 return VM_FAULT_NOPAGE;
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 } 926 }
1240 pmd_entry = mk_pmd(&zero_folio->page, !! 927 radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
1241 pmd_entry = pmd_mkhuge(pmd_entry); !! 928 entry = lock_slot(mapping, slot);
1242 set_pmd_at(vmf->vma->vm_mm, pmd_addr, !! 929 spin_unlock_irq(&mapping->tree_lock);
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 { <<
1263 struct iomap *iomap = &iter->iomap; <<
1264 const struct iomap *srcmap = iomap_it <<
1265 loff_t copy_pos = iter->pos; <<
1266 u64 copy_len = iomap_length(iter); <<
1267 u32 mod; <<
1268 int id = 0; <<
1269 s64 ret = 0; <<
1270 void *daddr = NULL, *saddr = NULL; <<
1271 <<
1272 if (!iomap_want_unshare_iter(iter)) <<
1273 return iomap_length(iter); <<
1274 <<
1275 /* 930 /*
1276 * Extend the file range to be aligne !! 931 * If we race with somebody updating the PTE and finish_mkwrite_fault()
1277 * we need to copy entire blocks, not !! 932 * fails, we don't care. We need to return VM_FAULT_NOPAGE and retry
1278 * Invalidate the mapping because we' !! 933 * the fault in either case.
1279 */ 934 */
1280 mod = offset_in_page(copy_pos); !! 935 finish_mkwrite_fault(vmf);
1281 if (mod) { !! 936 put_locked_mapping_entry(mapping, index, entry);
1282 copy_len += mod; !! 937 return VM_FAULT_NOPAGE;
1283 copy_pos -= mod; <<
1284 } <<
1285 <<
1286 mod = offset_in_page(copy_pos + copy_ <<
1287 if (mod) <<
1288 copy_len += PAGE_SIZE - mod; <<
1289 <<
1290 invalidate_inode_pages2_range(iter->i <<
1291 copy_po <<
1292 (copy_p <<
1293 <<
1294 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 <<
1303 if (copy_mc_to_kernel(daddr, saddr, c <<
1304 ret = iomap_length(iter); <<
1305 else <<
1306 ret = -EIO; <<
1307 <<
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 } 938 }
>> 939 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1356 940
1357 static s64 dax_zero_iter(struct iomap_iter *i !! 941 static bool dax_range_is_aligned(struct block_device *bdev,
>> 942 unsigned int offset, unsigned int length)
1358 { 943 {
1359 const struct iomap *iomap = &iter->io !! 944 unsigned short sector_size = bdev_logical_block_size(bdev);
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 945
1369 /* !! 946 if (!IS_ALIGNED(offset, sector_size))
1370 * invalidate the pages whose sharing !! 947 return false;
1371 * because of CoW. !! 948 if (!IS_ALIGNED(length, sector_size))
1372 */ !! 949 return false;
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); <<
1391 950
1392 if (rc < 0) !! 951 return true;
1393 return rc; <<
1394 pos += size; <<
1395 length -= size; <<
1396 written += size; <<
1397 } while (length > 0); <<
1398 <<
1399 if (did_zero) <<
1400 *did_zero = true; <<
1401 return written; <<
1402 } 952 }
1403 953
1404 int dax_zero_range(struct inode *inode, loff_ !! 954 int __dax_zero_page_range(struct block_device *bdev, sector_t sector,
1405 const struct iomap_ops *ops) !! 955 unsigned int offset, unsigned int length)
1406 { 956 {
1407 struct iomap_iter iter = { !! 957 struct blk_dax_ctl dax = {
1408 .inode = inode, !! 958 .sector = sector,
1409 .pos = pos, !! 959 .size = PAGE_SIZE,
1410 .len = len, <<
1411 .flags = IOMAP_DAX | <<
1412 }; 960 };
1413 int ret; <<
1414 961
1415 while ((ret = iomap_iter(&iter, ops)) !! 962 if (dax_range_is_aligned(bdev, offset, length)) {
1416 iter.processed = dax_zero_ite !! 963 sector_t start_sector = dax.sector + (offset >> 9);
1417 return ret; !! 964
>> 965 return blkdev_issue_zeroout(bdev, start_sector,
>> 966 length >> 9, GFP_NOFS, true);
>> 967 } else {
>> 968 if (dax_map_atomic(bdev, &dax) < 0)
>> 969 return PTR_ERR(dax.addr);
>> 970 clear_pmem(dax.addr + offset, length);
>> 971 dax_unmap_atomic(bdev, &dax);
>> 972 }
>> 973 return 0;
1418 } 974 }
1419 EXPORT_SYMBOL_GPL(dax_zero_range); !! 975 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1420 976
1421 int dax_truncate_page(struct inode *inode, lo !! 977 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
1422 const struct iomap_ops *ops) <<
1423 { 978 {
1424 unsigned int blocksize = i_blocksize( !! 979 return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9);
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 } 980 }
1432 EXPORT_SYMBOL_GPL(dax_truncate_page); <<
1433 981
1434 static loff_t dax_iomap_iter(const struct iom !! 982 static loff_t
1435 struct iov_iter *iter) !! 983 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
>> 984 struct iomap *iomap)
1436 { 985 {
1437 const struct iomap *iomap = &iomi->io !! 986 struct iov_iter *iter = data;
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 <<
1442 loff_t end = pos + length, done = 0; 987 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; 988 ssize_t ret = 0;
1446 size_t xfer; <<
1447 int id; <<
1448 989
1449 if (!write) { !! 990 if (iov_iter_rw(iter) == READ) {
1450 end = min(end, i_size_read(io !! 991 end = min(end, i_size_read(inode));
1451 if (pos >= end) 992 if (pos >= end)
1452 return 0; 993 return 0;
1453 994
1454 if (iomap->type == IOMAP_HOLE 995 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1455 return iov_iter_zero( 996 return iov_iter_zero(min(length, end - pos), iter);
1456 } 997 }
1457 998
1458 /* !! 999 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; 1000 return -EIO;
1465 1001
1466 /* 1002 /*
1467 * Write can allocate block for an ar 1003 * Write can allocate block for an area which has a hole page mapped
1468 * into page tables. We have to tear 1004 * into page tables. We have to tear down these mappings so that data
1469 * written by write(2) is visible in 1005 * written by write(2) is visible in mmap.
1470 */ 1006 */
1471 if (iomap->flags & IOMAP_F_NEW || cow !! 1007 if (iomap->flags & IOMAP_F_NEW) {
1472 /* !! 1008 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 1009 pos >> PAGE_SHIFT,
1484 1010 (end - 1) >> PAGE_SHIFT);
1485 } 1011 }
1486 1012
1487 id = dax_read_lock(); <<
1488 while (pos < end) { 1013 while (pos < end) {
1489 unsigned offset = pos & (PAGE 1014 unsigned offset = pos & (PAGE_SIZE - 1);
1490 const size_t size = ALIGN(len !! 1015 struct blk_dax_ctl dax = { 0 };
1491 pgoff_t pgoff = dax_iomap_pgo <<
1492 ssize_t map_len; 1016 ssize_t map_len;
1493 bool recovery = false; <<
1494 void *kaddr; <<
1495 1017
1496 if (fatal_signal_pending(curr 1018 if (fatal_signal_pending(current)) {
1497 ret = -EINTR; 1019 ret = -EINTR;
1498 break; 1020 break;
1499 } 1021 }
1500 1022
1501 map_len = dax_direct_access(d !! 1023 dax.sector = dax_iomap_sector(iomap, pos);
1502 DAX_ACCESS, & !! 1024 dax.size = (length + offset + PAGE_SIZE - 1) & PAGE_MASK;
1503 if (map_len == -EHWPOISON && !! 1025 map_len = dax_map_atomic(iomap->bdev, &dax);
1504 map_len = dax_direct_ <<
1505 PHYS_ <<
1506 &kadd <<
1507 if (map_len > 0) <<
1508 recovery = tr <<
1509 } <<
1510 if (map_len < 0) { 1026 if (map_len < 0) {
1511 ret = dax_mem2blk_err !! 1027 ret = map_len;
1512 break; 1028 break;
1513 } 1029 }
1514 1030
1515 if (cow) { !! 1031 dax.addr += offset;
1516 ret = dax_iomap_copy_ <<
1517 <<
1518 if (ret) <<
1519 break; <<
1520 } <<
1521 <<
1522 map_len = PFN_PHYS(map_len); <<
1523 kaddr += offset; <<
1524 map_len -= offset; 1032 map_len -= offset;
1525 if (map_len > end - pos) 1033 if (map_len > end - pos)
1526 map_len = end - pos; 1034 map_len = end - pos;
1527 1035
1528 if (recovery) !! 1036 if (iov_iter_rw(iter) == WRITE)
1529 xfer = dax_recovery_w !! 1037 map_len = copy_from_iter_pmem(dax.addr, map_len, iter);
1530 map_l <<
1531 else if (write) <<
1532 xfer = dax_copy_from_ <<
1533 map_l <<
1534 else 1038 else
1535 xfer = dax_copy_to_it !! 1039 map_len = copy_to_iter(dax.addr, map_len, iter);
1536 map_l !! 1040 dax_unmap_atomic(iomap->bdev, &dax);
1537 !! 1041 if (map_len <= 0) {
1538 pos += xfer; !! 1042 ret = map_len ? map_len : -EFAULT;
1539 length -= xfer; <<
1540 done += xfer; <<
1541 <<
1542 if (xfer == 0) <<
1543 ret = -EFAULT; <<
1544 if (xfer < map_len) <<
1545 break; 1043 break;
>> 1044 }
>> 1045
>> 1046 pos += map_len;
>> 1047 length -= map_len;
>> 1048 done += map_len;
1546 } 1049 }
1547 dax_read_unlock(id); <<
1548 1050
1549 return done ? done : ret; 1051 return done ? done : ret;
1550 } 1052 }
1551 1053
1552 /** 1054 /**
1553 * dax_iomap_rw - Perform I/O to a DAX file 1055 * dax_iomap_rw - Perform I/O to a DAX file
1554 * @iocb: The control block for this I/ 1056 * @iocb: The control block for this I/O
1555 * @iter: The addresses to do I/O from 1057 * @iter: The addresses to do I/O from or to
1556 * @ops: iomap ops passed from the fil 1058 * @ops: iomap ops passed from the file system
1557 * 1059 *
1558 * This function performs read and write oper 1060 * This function performs read and write operations to directly mapped
1559 * persistent memory. The callers needs to t 1061 * persistent memory. The callers needs to take care of read/write exclusion
1560 * and evicting any page cache pages in the r 1062 * and evicting any page cache pages in the region under I/O.
1561 */ 1063 */
1562 ssize_t 1064 ssize_t
1563 dax_iomap_rw(struct kiocb *iocb, struct iov_i 1065 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1564 const struct iomap_ops *ops) 1066 const struct iomap_ops *ops)
1565 { 1067 {
1566 struct iomap_iter iomi = { !! 1068 struct address_space *mapping = iocb->ki_filp->f_mapping;
1567 .inode = iocb->ki_fi !! 1069 struct inode *inode = mapping->host;
1568 .pos = iocb->ki_po !! 1070 loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1569 .len = iov_iter_co !! 1071 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 1072
1578 if (iov_iter_rw(iter) == WRITE) { 1073 if (iov_iter_rw(iter) == WRITE) {
1579 lockdep_assert_held_write(&io !! 1074 lockdep_assert_held_exclusive(&inode->i_rwsem);
1580 iomi.flags |= IOMAP_WRITE; !! 1075 flags |= IOMAP_WRITE;
1581 } else { 1076 } else {
1582 lockdep_assert_held(&iomi.ino !! 1077 lockdep_assert_held(&inode->i_rwsem);
1583 } 1078 }
1584 1079
1585 if (iocb->ki_flags & IOCB_NOWAIT) !! 1080 while (iov_iter_count(iter)) {
1586 iomi.flags |= IOMAP_NOWAIT; !! 1081 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1587 !! 1082 iter, dax_iomap_actor);
1588 while ((ret = iomap_iter(&iomi, ops)) !! 1083 if (ret <= 0)
1589 iomi.processed = dax_iomap_it !! 1084 break;
>> 1085 pos += ret;
>> 1086 done += ret;
>> 1087 }
1590 1088
1591 done = iomi.pos - iocb->ki_pos; !! 1089 iocb->ki_pos += done;
1592 iocb->ki_pos = iomi.pos; <<
1593 return done ? done : ret; 1090 return done ? done : ret;
1594 } 1091 }
1595 EXPORT_SYMBOL_GPL(dax_iomap_rw); 1092 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1596 1093
1597 static vm_fault_t dax_fault_return(int error) !! 1094 static int dax_fault_return(int error)
1598 { 1095 {
1599 if (error == 0) 1096 if (error == 0)
1600 return VM_FAULT_NOPAGE; 1097 return VM_FAULT_NOPAGE;
1601 return vmf_error(error); !! 1098 if (error == -ENOMEM)
1602 } !! 1099 return VM_FAULT_OOM;
1603 !! 1100 return VM_FAULT_SIGBUS;
1604 /* <<
1605 * When handling a synchronous page fault and <<
1606 * insert the PTE/PMD into page tables only a <<
1607 * insertion for now and return the pfn so th <<
1608 * fsync is done. <<
1609 */ <<
1610 static vm_fault_t dax_fault_synchronous_pfnp( <<
1611 { <<
1612 if (WARN_ON_ONCE(!pfnp)) <<
1613 return VM_FAULT_SIGBUS; <<
1614 *pfnp = pfn; <<
1615 return VM_FAULT_NEEDDSYNC; <<
1616 } 1101 }
1617 1102
1618 static vm_fault_t dax_fault_cow_page(struct v !! 1103 static int dax_iomap_pte_fault(struct vm_fault *vmf,
1619 const struct iomap_iter *iter !! 1104 const struct iomap_ops *ops)
1620 { <<
1621 vm_fault_t ret; <<
1622 int error = 0; <<
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 } <<
1711 <<
1712 static vm_fault_t dax_iomap_pte_fault(struct <<
1713 int *iomap_err <<
1714 { 1105 {
1715 struct address_space *mapping = vmf-> 1106 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1716 XA_STATE(xas, &mapping->i_pages, vmf- !! 1107 struct inode *inode = mapping->host;
1717 struct iomap_iter iter = { !! 1108 unsigned long vaddr = vmf->address;
1718 .inode = mapping->ho !! 1109 loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1719 .pos = (loff_t)vmf !! 1110 sector_t sector;
1720 .len = PAGE_SIZE, !! 1111 struct iomap iomap = { 0 };
1721 .flags = IOMAP_DAX | !! 1112 unsigned flags = IOMAP_FAULT;
1722 }; !! 1113 int error, major = 0;
1723 vm_fault_t ret = 0; !! 1114 int vmf_ret = 0;
1724 void *entry; 1115 void *entry;
1725 int error; <<
1726 1116
1727 trace_dax_pte_fault(iter.inode, vmf, <<
1728 /* 1117 /*
1729 * Check whether offset isn't beyond 1118 * Check whether offset isn't beyond end of file now. Caller is supposed
1730 * to hold locks serializing us with 1119 * to hold locks serializing us with truncate / punch hole so this is
1731 * a reliable test. 1120 * a reliable test.
1732 */ 1121 */
1733 if (iter.pos >= i_size_read(iter.inod !! 1122 if (pos >= i_size_read(inode))
1734 ret = VM_FAULT_SIGBUS; !! 1123 return VM_FAULT_SIGBUS;
1735 goto out; <<
1736 } <<
1737 1124
1738 if ((vmf->flags & FAULT_FLAG_WRITE) & 1125 if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1739 iter.flags |= IOMAP_WRITE; !! 1126 flags |= IOMAP_WRITE;
1740 1127
1741 entry = grab_mapping_entry(&xas, mapp !! 1128 entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
1742 if (xa_is_internal(entry)) { !! 1129 if (IS_ERR(entry))
1743 ret = xa_to_internal(entry); !! 1130 return dax_fault_return(PTR_ERR(entry));
1744 goto out; <<
1745 } <<
1746 1131
1747 /* 1132 /*
1748 * It is possible, particularly with 1133 * It is possible, particularly with mixed reads & writes to private
1749 * mappings, that we have raced with 1134 * mappings, that we have raced with a PMD fault that overlaps with
1750 * the PTE we need to set up. If so 1135 * the PTE we need to set up. If so just return and the fault will be
1751 * retried. 1136 * retried.
1752 */ 1137 */
1753 if (pmd_trans_huge(*vmf->pmd) || pmd_ 1138 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1754 ret = VM_FAULT_NOPAGE; !! 1139 vmf_ret = VM_FAULT_NOPAGE;
1755 goto unlock_entry; 1140 goto unlock_entry;
1756 } 1141 }
1757 1142
1758 while ((error = iomap_iter(&iter, ops !! 1143 /*
1759 if (WARN_ON_ONCE(iomap_length !! 1144 * Note that we don't bother to use iomap_apply here: DAX required
1760 iter.processed = -EIO !! 1145 * the file system block size to be equal the page size, which means
1761 continue; !! 1146 * that we never have to deal with more than a single extent here.
>> 1147 */
>> 1148 error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
>> 1149 if (error) {
>> 1150 vmf_ret = dax_fault_return(error);
>> 1151 goto unlock_entry;
>> 1152 }
>> 1153 if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
>> 1154 error = -EIO; /* fs corruption? */
>> 1155 goto error_finish_iomap;
>> 1156 }
>> 1157
>> 1158 sector = dax_iomap_sector(&iomap, pos);
>> 1159
>> 1160 if (vmf->cow_page) {
>> 1161 switch (iomap.type) {
>> 1162 case IOMAP_HOLE:
>> 1163 case IOMAP_UNWRITTEN:
>> 1164 clear_user_highpage(vmf->cow_page, vaddr);
>> 1165 break;
>> 1166 case IOMAP_MAPPED:
>> 1167 error = copy_user_dax(iomap.bdev, sector, PAGE_SIZE,
>> 1168 vmf->cow_page, vaddr);
>> 1169 break;
>> 1170 default:
>> 1171 WARN_ON_ONCE(1);
>> 1172 error = -EIO;
>> 1173 break;
1762 } 1174 }
1763 1175
1764 ret = dax_fault_iter(vmf, &it !! 1176 if (error)
1765 if (ret != VM_FAULT_SIGBUS && !! 1177 goto error_finish_iomap;
1766 (iter.iomap.flags & IOMAP !! 1178
>> 1179 __SetPageUptodate(vmf->cow_page);
>> 1180 vmf_ret = finish_fault(vmf);
>> 1181 if (!vmf_ret)
>> 1182 vmf_ret = VM_FAULT_DONE_COW;
>> 1183 goto finish_iomap;
>> 1184 }
>> 1185
>> 1186 switch (iomap.type) {
>> 1187 case IOMAP_MAPPED:
>> 1188 if (iomap.flags & IOMAP_F_NEW) {
1767 count_vm_event(PGMAJF 1189 count_vm_event(PGMAJFAULT);
1768 count_memcg_event_mm( !! 1190 mem_cgroup_count_vm_event(vmf->vma->vm_mm, PGMAJFAULT);
1769 ret |= VM_FAULT_MAJOR !! 1191 major = VM_FAULT_MAJOR;
1770 } 1192 }
1771 !! 1193 error = dax_insert_mapping(mapping, iomap.bdev, sector,
1772 if (!(ret & VM_FAULT_ERROR)) !! 1194 PAGE_SIZE, &entry, vmf->vma, vmf);
1773 iter.processed = PAGE !! 1195 /* -EBUSY is fine, somebody else faulted on the same PTE */
>> 1196 if (error == -EBUSY)
>> 1197 error = 0;
>> 1198 break;
>> 1199 case IOMAP_UNWRITTEN:
>> 1200 case IOMAP_HOLE:
>> 1201 if (!(vmf->flags & FAULT_FLAG_WRITE)) {
>> 1202 vmf_ret = dax_load_hole(mapping, &entry, vmf);
>> 1203 goto finish_iomap;
>> 1204 }
>> 1205 /*FALLTHRU*/
>> 1206 default:
>> 1207 WARN_ON_ONCE(1);
>> 1208 error = -EIO;
>> 1209 break;
1774 } 1210 }
1775 1211
1776 if (iomap_errp) !! 1212 error_finish_iomap:
1777 *iomap_errp = error; !! 1213 vmf_ret = dax_fault_return(error) | major;
1778 if (!ret && error) !! 1214 finish_iomap:
1779 ret = dax_fault_return(error) !! 1215 if (ops->iomap_end) {
>> 1216 int copied = PAGE_SIZE;
1780 1217
1781 unlock_entry: !! 1218 if (vmf_ret & VM_FAULT_ERROR)
1782 dax_unlock_entry(&xas, entry); !! 1219 copied = 0;
1783 out: !! 1220 /*
1784 trace_dax_pte_fault_done(iter.inode, !! 1221 * The fault is done by now and there's no way back (other
1785 return ret; !! 1222 * thread may be already happily using PTE we have installed).
>> 1223 * Just ignore error from ->iomap_end since we cannot do much
>> 1224 * with it.
>> 1225 */
>> 1226 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
>> 1227 }
>> 1228 unlock_entry:
>> 1229 put_locked_mapping_entry(mapping, vmf->pgoff, entry);
>> 1230 return vmf_ret;
1786 } 1231 }
1787 1232
1788 #ifdef CONFIG_FS_DAX_PMD 1233 #ifdef CONFIG_FS_DAX_PMD
1789 static bool dax_fault_check_fallback(struct v !! 1234 /*
1790 pgoff_t max_pgoff) !! 1235 * The 'colour' (ie low bits) within a PMD of a page offset. This comes up
>> 1236 * more often than one might expect in the below functions.
>> 1237 */
>> 1238 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
>> 1239
>> 1240 static int dax_pmd_insert_mapping(struct vm_fault *vmf, struct iomap *iomap,
>> 1241 loff_t pos, void **entryp)
>> 1242 {
>> 1243 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
>> 1244 struct block_device *bdev = iomap->bdev;
>> 1245 struct inode *inode = mapping->host;
>> 1246 struct blk_dax_ctl dax = {
>> 1247 .sector = dax_iomap_sector(iomap, pos),
>> 1248 .size = PMD_SIZE,
>> 1249 };
>> 1250 long length = dax_map_atomic(bdev, &dax);
>> 1251 void *ret = NULL;
>> 1252
>> 1253 if (length < 0) /* dax_map_atomic() failed */
>> 1254 goto fallback;
>> 1255 if (length < PMD_SIZE)
>> 1256 goto unmap_fallback;
>> 1257 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR)
>> 1258 goto unmap_fallback;
>> 1259 if (!pfn_t_devmap(dax.pfn))
>> 1260 goto unmap_fallback;
>> 1261
>> 1262 dax_unmap_atomic(bdev, &dax);
>> 1263
>> 1264 ret = dax_insert_mapping_entry(mapping, vmf, *entryp, dax.sector,
>> 1265 RADIX_DAX_PMD);
>> 1266 if (IS_ERR(ret))
>> 1267 goto fallback;
>> 1268 *entryp = ret;
>> 1269
>> 1270 trace_dax_pmd_insert_mapping(inode, vmf, length, dax.pfn, ret);
>> 1271 return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd,
>> 1272 dax.pfn, vmf->flags & FAULT_FLAG_WRITE);
>> 1273
>> 1274 unmap_fallback:
>> 1275 dax_unmap_atomic(bdev, &dax);
>> 1276 fallback:
>> 1277 trace_dax_pmd_insert_mapping_fallback(inode, vmf, length,
>> 1278 dax.pfn, ret);
>> 1279 return VM_FAULT_FALLBACK;
>> 1280 }
>> 1281
>> 1282 static int dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap,
>> 1283 void **entryp)
1791 { 1284 {
>> 1285 struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1792 unsigned long pmd_addr = vmf->address 1286 unsigned long pmd_addr = vmf->address & PMD_MASK;
1793 bool write = vmf->flags & FAULT_FLAG_ !! 1287 struct inode *inode = mapping->host;
>> 1288 struct page *zero_page;
>> 1289 void *ret = NULL;
>> 1290 spinlock_t *ptl;
>> 1291 pmd_t pmd_entry;
1794 1292
1795 /* !! 1293 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 1294
1805 /* Fall back to PTEs if we're going t !! 1295 if (unlikely(!zero_page))
1806 if (write && !(vmf->vma->vm_flags & V !! 1296 goto fallback;
1807 return true; <<
1808 1297
1809 /* If the PMD would extend outside th !! 1298 ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0,
1810 if (pmd_addr < vmf->vma->vm_start) !! 1299 RADIX_DAX_PMD | RADIX_DAX_HZP);
1811 return true; !! 1300 if (IS_ERR(ret))
1812 if ((pmd_addr + PMD_SIZE) > vmf->vma- !! 1301 goto fallback;
1813 return true; !! 1302 *entryp = ret;
1814 1303
1815 /* If the PMD would extend beyond the !! 1304 ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1816 if ((xas->xa_index | PG_PMD_COLOUR) > !! 1305 if (!pmd_none(*(vmf->pmd))) {
1817 return true; !! 1306 spin_unlock(ptl);
>> 1307 goto fallback;
>> 1308 }
>> 1309
>> 1310 pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
>> 1311 pmd_entry = pmd_mkhuge(pmd_entry);
>> 1312 set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
>> 1313 spin_unlock(ptl);
>> 1314 trace_dax_pmd_load_hole(inode, vmf, zero_page, ret);
>> 1315 return VM_FAULT_NOPAGE;
1818 1316
1819 return false; !! 1317 fallback:
>> 1318 trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret);
>> 1319 return VM_FAULT_FALLBACK;
1820 } 1320 }
1821 1321
1822 static vm_fault_t dax_iomap_pmd_fault(struct !! 1322 static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1823 const struct i 1323 const struct iomap_ops *ops)
1824 { 1324 {
1825 struct address_space *mapping = vmf-> !! 1325 struct vm_area_struct *vma = vmf->vma;
1826 XA_STATE_ORDER(xas, &mapping->i_pages !! 1326 struct address_space *mapping = vma->vm_file->f_mapping;
1827 struct iomap_iter iter = { !! 1327 unsigned long pmd_addr = vmf->address & PMD_MASK;
1828 .inode = mapping->ho !! 1328 bool write = vmf->flags & FAULT_FLAG_WRITE;
1829 .len = PMD_SIZE, !! 1329 unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1830 .flags = IOMAP_DAX | !! 1330 struct inode *inode = mapping->host;
1831 }; !! 1331 int result = VM_FAULT_FALLBACK;
1832 vm_fault_t ret = VM_FAULT_FALLBACK; !! 1332 struct iomap iomap = { 0 };
1833 pgoff_t max_pgoff; !! 1333 pgoff_t max_pgoff, pgoff;
1834 void *entry; 1334 void *entry;
1835 !! 1335 loff_t pos;
1836 if (vmf->flags & FAULT_FLAG_WRITE) !! 1336 int error;
1837 iter.flags |= IOMAP_WRITE; <<
1838 1337
1839 /* 1338 /*
1840 * Check whether offset isn't beyond 1339 * Check whether offset isn't beyond end of file now. Caller is
1841 * supposed to hold locks serializing 1340 * supposed to hold locks serializing us with truncate / punch hole so
1842 * this is a reliable test. 1341 * this is a reliable test.
1843 */ 1342 */
1844 max_pgoff = DIV_ROUND_UP(i_size_read( !! 1343 pgoff = linear_page_index(vma, pmd_addr);
>> 1344 max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT;
1845 1345
1846 trace_dax_pmd_fault(iter.inode, vmf, !! 1346 trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
>> 1347
>> 1348 /* Fall back to PTEs if we're going to COW */
>> 1349 if (write && !(vma->vm_flags & VM_SHARED))
>> 1350 goto fallback;
1847 1351
1848 if (xas.xa_index >= max_pgoff) { !! 1352 /* If the PMD would extend outside the VMA */
1849 ret = VM_FAULT_SIGBUS; !! 1353 if (pmd_addr < vma->vm_start)
>> 1354 goto fallback;
>> 1355 if ((pmd_addr + PMD_SIZE) > vma->vm_end)
>> 1356 goto fallback;
>> 1357
>> 1358 if (pgoff > max_pgoff) {
>> 1359 result = VM_FAULT_SIGBUS;
1850 goto out; 1360 goto out;
1851 } 1361 }
1852 1362
1853 if (dax_fault_check_fallback(vmf, &xa !! 1363 /* If the PMD would extend beyond the file size */
>> 1364 if ((pgoff | PG_PMD_COLOUR) > max_pgoff)
1854 goto fallback; 1365 goto fallback;
1855 1366
1856 /* 1367 /*
1857 * grab_mapping_entry() will make sur !! 1368 * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX
1858 * a zero PMD entry or a DAX PMD. If !! 1369 * PMD or a HZP entry. If it can't (because a 4k page is already in
1859 * entry is already in the array, for !! 1370 * the tree, for instance), it will return -EEXIST and we just fall
1860 * VM_FAULT_FALLBACK. !! 1371 * back to 4k entries.
1861 */ 1372 */
1862 entry = grab_mapping_entry(&xas, mapp !! 1373 entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
1863 if (xa_is_internal(entry)) { !! 1374 if (IS_ERR(entry))
1864 ret = xa_to_internal(entry); <<
1865 goto fallback; 1375 goto fallback;
1866 } <<
1867 1376
1868 /* 1377 /*
1869 * It is possible, particularly with 1378 * It is possible, particularly with mixed reads & writes to private
1870 * mappings, that we have raced with 1379 * mappings, that we have raced with a PTE fault that overlaps with
1871 * the PMD we need to set up. If so 1380 * the PMD we need to set up. If so just return and the fault will be
1872 * retried. 1381 * retried.
1873 */ 1382 */
1874 if (!pmd_none(*vmf->pmd) && !pmd_tran 1383 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1875 !pmd_devmap(*vmf->pmd 1384 !pmd_devmap(*vmf->pmd)) {
1876 ret = 0; !! 1385 result = 0;
1877 goto unlock_entry; 1386 goto unlock_entry;
1878 } 1387 }
1879 1388
1880 iter.pos = (loff_t)xas.xa_index << PA !! 1389 /*
1881 while (iomap_iter(&iter, ops) > 0) { !! 1390 * Note that we don't use iomap_apply here. We aren't doing I/O, only
1882 if (iomap_length(&iter) < PMD !! 1391 * setting up a mapping, so really we're using iomap_begin() as a way
1883 continue; /* actually !! 1392 * to look up our filesystem block.
1884 !! 1393 */
1885 ret = dax_fault_iter(vmf, &it !! 1394 pos = (loff_t)pgoff << PAGE_SHIFT;
1886 if (ret != VM_FAULT_FALLBACK) !! 1395 error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1887 iter.processed = PMD_ !! 1396 if (error)
>> 1397 goto unlock_entry;
>> 1398
>> 1399 if (iomap.offset + iomap.length < pos + PMD_SIZE)
>> 1400 goto finish_iomap;
>> 1401
>> 1402 switch (iomap.type) {
>> 1403 case IOMAP_MAPPED:
>> 1404 result = dax_pmd_insert_mapping(vmf, &iomap, pos, &entry);
>> 1405 break;
>> 1406 case IOMAP_UNWRITTEN:
>> 1407 case IOMAP_HOLE:
>> 1408 if (WARN_ON_ONCE(write))
>> 1409 break;
>> 1410 result = dax_pmd_load_hole(vmf, &iomap, &entry);
>> 1411 break;
>> 1412 default:
>> 1413 WARN_ON_ONCE(1);
>> 1414 break;
1888 } 1415 }
1889 1416
1890 unlock_entry: !! 1417 finish_iomap:
1891 dax_unlock_entry(&xas, entry); !! 1418 if (ops->iomap_end) {
1892 fallback: !! 1419 int copied = PMD_SIZE;
1893 if (ret == VM_FAULT_FALLBACK) { !! 1420
1894 split_huge_pmd(vmf->vma, vmf- !! 1421 if (result == VM_FAULT_FALLBACK)
>> 1422 copied = 0;
>> 1423 /*
>> 1424 * The fault is done by now and there's no way back (other
>> 1425 * thread may be already happily using PMD we have installed).
>> 1426 * Just ignore error from ->iomap_end since we cannot do much
>> 1427 * with it.
>> 1428 */
>> 1429 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
>> 1430 &iomap);
>> 1431 }
>> 1432 unlock_entry:
>> 1433 put_locked_mapping_entry(mapping, pgoff, entry);
>> 1434 fallback:
>> 1435 if (result == VM_FAULT_FALLBACK) {
>> 1436 split_huge_pmd(vma, vmf->pmd, vmf->address);
1895 count_vm_event(THP_FAULT_FALL 1437 count_vm_event(THP_FAULT_FALLBACK);
1896 } 1438 }
1897 out: 1439 out:
1898 trace_dax_pmd_fault_done(iter.inode, !! 1440 trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1899 return ret; !! 1441 return result;
1900 } 1442 }
1901 #else 1443 #else
1902 static vm_fault_t dax_iomap_pmd_fault(struct !! 1444 static int dax_iomap_pmd_fault(struct vm_fault *vmf,
1903 const struct i 1445 const struct iomap_ops *ops)
1904 { 1446 {
1905 return VM_FAULT_FALLBACK; 1447 return VM_FAULT_FALLBACK;
1906 } 1448 }
1907 #endif /* CONFIG_FS_DAX_PMD */ 1449 #endif /* CONFIG_FS_DAX_PMD */
1908 1450
1909 /** 1451 /**
1910 * dax_iomap_fault - handle a page fault on a 1452 * dax_iomap_fault - handle a page fault on a DAX file
1911 * @vmf: The description of the fault 1453 * @vmf: The description of the fault
1912 * @order: Order of the page to fault in !! 1454 * @ops: iomap ops passed from the file system
1913 * @pfnp: PFN to insert for synchronous fault <<
1914 * @iomap_errp: Storage for detailed error co <<
1915 * @ops: Iomap ops passed from the file syste <<
1916 * 1455 *
1917 * When a page fault occurs, filesystems may 1456 * When a page fault occurs, filesystems may call this helper in
1918 * their fault handler for DAX files. dax_iom 1457 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1919 * has done all the necessary locking for pag 1458 * has done all the necessary locking for page fault to proceed
1920 * successfully. 1459 * successfully.
1921 */ 1460 */
1922 vm_fault_t dax_iomap_fault(struct vm_fault *v !! 1461 int dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1923 pfn_t *pfnp, int *iomap_e !! 1462 const struct iomap_ops *ops)
1924 { 1463 {
1925 if (order == 0) !! 1464 switch (pe_size) {
1926 return dax_iomap_pte_fault(vm !! 1465 case PE_SIZE_PTE:
1927 else if (order == PMD_ORDER) !! 1466 return dax_iomap_pte_fault(vmf, ops);
1928 return dax_iomap_pmd_fault(vm !! 1467 case PE_SIZE_PMD:
1929 else !! 1468 return dax_iomap_pmd_fault(vmf, ops);
>> 1469 default:
1930 return VM_FAULT_FALLBACK; 1470 return VM_FAULT_FALLBACK;
1931 } <<
1932 EXPORT_SYMBOL_GPL(dax_iomap_fault); <<
1933 <<
1934 /* <<
1935 * dax_insert_pfn_mkwrite - insert PTE or PMD <<
1936 * @vmf: The description of the fault <<
1937 * @pfn: PFN to insert <<
1938 * @order: Order of entry to insert. <<
1939 * <<
1940 * This function inserts a writeable PTE or P <<
1941 * for an mmaped DAX file. It also marks the <<
1942 */ <<
1943 static vm_fault_t <<
1944 dax_insert_pfn_mkwrite(struct vm_fault *vmf, <<
1945 { <<
1946 struct address_space *mapping = vmf-> <<
1947 XA_STATE_ORDER(xas, &mapping->i_pages <<
1948 void *entry; <<
1949 vm_fault_t ret; <<
1950 <<
1951 xas_lock_irq(&xas); <<
1952 entry = get_unlocked_entry(&xas, orde <<
1953 /* Did we race with someone splitting <<
1954 if (!entry || dax_is_conflict(entry) <<
1955 (order == 0 && !dax_is_pte_entry( <<
1956 put_unlocked_entry(&xas, entr <<
1957 xas_unlock_irq(&xas); <<
1958 trace_dax_insert_pfn_mkwrite_ <<
1959 <<
1960 return VM_FAULT_NOPAGE; <<
1961 } <<
1962 xas_set_mark(&xas, PAGECACHE_TAG_DIRT <<
1963 dax_lock_entry(&xas, entry); <<
1964 xas_unlock_irq(&xas); <<
1965 if (order == 0) <<
1966 ret = vmf_insert_mixed_mkwrit <<
1967 #ifdef CONFIG_FS_DAX_PMD <<
1968 else if (order == PMD_ORDER) <<
1969 ret = vmf_insert_pfn_pmd(vmf, <<
1970 #endif <<
1971 else <<
1972 ret = VM_FAULT_FALLBACK; <<
1973 dax_unlock_entry(&xas, entry); <<
1974 trace_dax_insert_pfn_mkwrite(mapping- <<
1975 return ret; <<
1976 } <<
1977 <<
1978 /** <<
1979 * dax_finish_sync_fault - finish synchronous <<
1980 * @vmf: The description of the fault <<
1981 * @order: Order of entry to be inserted <<
1982 * @pfn: PFN to insert <<
1983 * <<
1984 * This function ensures that the file range <<
1985 * stored persistently on the media and handl <<
1986 * table entry. <<
1987 */ <<
1988 vm_fault_t dax_finish_sync_fault(struct vm_fa <<
1989 pfn_t pfn) <<
1990 { <<
1991 int err; <<
1992 loff_t start = ((loff_t)vmf->pgoff) < <<
1993 size_t len = PAGE_SIZE << order; <<
1994 <<
1995 err = vfs_fsync_range(vmf->vma->vm_fi <<
1996 if (err) <<
1997 return VM_FAULT_SIGBUS; <<
1998 return dax_insert_pfn_mkwrite(vmf, pf <<
1999 } <<
2000 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 } 1471 }
2071 return ret; <<
2072 } 1472 }
2073 !! 1473 EXPORT_SYMBOL_GPL(dax_iomap_fault);
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 1474
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.