~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/dax.c

Version: ~ [ linux-4.20-rc7 ] ~ [ linux-4.19.10 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.89 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.146 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.167 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.130 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.62 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.39.4 ] ~ [ linux-2.6.38.8 ] ~ [ linux-2.6.37.6 ] ~ [ linux-2.6.36.4 ] ~ [ linux-2.6.35.14 ] ~ [ linux-2.6.34.15 ] ~ [ linux-2.6.33.20 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.31.14 ] ~ [ linux-2.6.30.10 ] ~ [ linux-2.6.29.6 ] ~ [ linux-2.6.28.10 ] ~ [ linux-2.6.27.62 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * fs/dax.c - Direct Access filesystem code
  3  * Copyright (c) 2013-2014 Intel Corporation
  4  * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
  5  * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
  6  *
  7  * This program is free software; you can redistribute it and/or modify it
  8  * under the terms and conditions of the GNU General Public License,
  9  * version 2, as published by the Free Software Foundation.
 10  *
 11  * This program is distributed in the hope it will be useful, but WITHOUT
 12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
 14  * more details.
 15  */
 16 
 17 #include <linux/atomic.h>
 18 #include <linux/blkdev.h>
 19 #include <linux/buffer_head.h>
 20 #include <linux/dax.h>
 21 #include <linux/fs.h>
 22 #include <linux/genhd.h>
 23 #include <linux/highmem.h>
 24 #include <linux/memcontrol.h>
 25 #include <linux/mm.h>
 26 #include <linux/mutex.h>
 27 #include <linux/pagevec.h>
 28 #include <linux/sched.h>
 29 #include <linux/sched/signal.h>
 30 #include <linux/uio.h>
 31 #include <linux/vmstat.h>
 32 #include <linux/pfn_t.h>
 33 #include <linux/sizes.h>
 34 #include <linux/mmu_notifier.h>
 35 #include <linux/iomap.h>
 36 #include "internal.h"
 37 
 38 #define CREATE_TRACE_POINTS
 39 #include <trace/events/fs_dax.h>
 40 
 41 static inline unsigned int pe_order(enum page_entry_size pe_size)
 42 {
 43         if (pe_size == PE_SIZE_PTE)
 44                 return PAGE_SHIFT - PAGE_SHIFT;
 45         if (pe_size == PE_SIZE_PMD)
 46                 return PMD_SHIFT - PAGE_SHIFT;
 47         if (pe_size == PE_SIZE_PUD)
 48                 return PUD_SHIFT - PAGE_SHIFT;
 49         return ~0;
 50 }
 51 
 52 /* We choose 4096 entries - same as per-zone page wait tables */
 53 #define DAX_WAIT_TABLE_BITS 12
 54 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
 55 
 56 /* The 'colour' (ie low bits) within a PMD of a page offset.  */
 57 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
 58 #define PG_PMD_NR       (PMD_SIZE >> PAGE_SHIFT)
 59 
 60 /* The order of a PMD entry */
 61 #define PMD_ORDER       (PMD_SHIFT - PAGE_SHIFT)
 62 
 63 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
 64 
 65 static int __init init_dax_wait_table(void)
 66 {
 67         int i;
 68 
 69         for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
 70                 init_waitqueue_head(wait_table + i);
 71         return 0;
 72 }
 73 fs_initcall(init_dax_wait_table);
 74 
 75 /*
 76  * DAX pagecache entries use XArray value entries so they can't be mistaken
 77  * for pages.  We use one bit for locking, one bit for the entry size (PMD)
 78  * and two more to tell us if the entry is a zero page or an empty entry that
 79  * is just used for locking.  In total four special bits.
 80  *
 81  * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
 82  * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
 83  * block allocation.
 84  */
 85 #define DAX_SHIFT       (4)
 86 #define DAX_LOCKED      (1UL << 0)
 87 #define DAX_PMD         (1UL << 1)
 88 #define DAX_ZERO_PAGE   (1UL << 2)
 89 #define DAX_EMPTY       (1UL << 3)
 90 
 91 static unsigned long dax_to_pfn(void *entry)
 92 {
 93         return xa_to_value(entry) >> DAX_SHIFT;
 94 }
 95 
 96 static void *dax_make_entry(pfn_t pfn, unsigned long flags)
 97 {
 98         return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
 99 }
100 
101 static bool dax_is_locked(void *entry)
102 {
103         return xa_to_value(entry) & DAX_LOCKED;
104 }
105 
106 static unsigned int dax_entry_order(void *entry)
107 {
108         if (xa_to_value(entry) & DAX_PMD)
109                 return PMD_ORDER;
110         return 0;
111 }
112 
113 static unsigned long dax_is_pmd_entry(void *entry)
114 {
115         return xa_to_value(entry) & DAX_PMD;
116 }
117 
118 static bool dax_is_pte_entry(void *entry)
119 {
120         return !(xa_to_value(entry) & DAX_PMD);
121 }
122 
123 static int dax_is_zero_entry(void *entry)
124 {
125         return xa_to_value(entry) & DAX_ZERO_PAGE;
126 }
127 
128 static int dax_is_empty_entry(void *entry)
129 {
130         return xa_to_value(entry) & DAX_EMPTY;
131 }
132 
133 /*
134  * DAX page cache entry locking
135  */
136 struct exceptional_entry_key {
137         struct xarray *xa;
138         pgoff_t entry_start;
139 };
140 
141 struct wait_exceptional_entry_queue {
142         wait_queue_entry_t wait;
143         struct exceptional_entry_key key;
144 };
145 
146 static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas,
147                 void *entry, struct exceptional_entry_key *key)
148 {
149         unsigned long hash;
150         unsigned long index = xas->xa_index;
151 
152         /*
153          * If 'entry' is a PMD, align the 'index' that we use for the wait
154          * queue to the start of that PMD.  This ensures that all offsets in
155          * the range covered by the PMD map to the same bit lock.
156          */
157         if (dax_is_pmd_entry(entry))
158                 index &= ~PG_PMD_COLOUR;
159         key->xa = xas->xa;
160         key->entry_start = index;
161 
162         hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS);
163         return wait_table + hash;
164 }
165 
166 static int wake_exceptional_entry_func(wait_queue_entry_t *wait,
167                 unsigned int mode, int sync, void *keyp)
168 {
169         struct exceptional_entry_key *key = keyp;
170         struct wait_exceptional_entry_queue *ewait =
171                 container_of(wait, struct wait_exceptional_entry_queue, wait);
172 
173         if (key->xa != ewait->key.xa ||
174             key->entry_start != ewait->key.entry_start)
175                 return 0;
176         return autoremove_wake_function(wait, mode, sync, NULL);
177 }
178 
179 /*
180  * @entry may no longer be the entry at the index in the mapping.
181  * The important information it's conveying is whether the entry at
182  * this index used to be a PMD entry.
183  */
184 static void dax_wake_entry(struct xa_state *xas, void *entry, bool wake_all)
185 {
186         struct exceptional_entry_key key;
187         wait_queue_head_t *wq;
188 
189         wq = dax_entry_waitqueue(xas, entry, &key);
190 
191         /*
192          * Checking for locked entry and prepare_to_wait_exclusive() happens
193          * under the i_pages lock, ditto for entry handling in our callers.
194          * So at this point all tasks that could have seen our entry locked
195          * must be in the waitqueue and the following check will see them.
196          */
197         if (waitqueue_active(wq))
198                 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
199 }
200 
201 /*
202  * Look up entry in page cache, wait for it to become unlocked if it
203  * is a DAX entry and return it.  The caller must subsequently call
204  * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
205  * if it did.
206  *
207  * Must be called with the i_pages lock held.
208  */
209 static void *get_unlocked_entry(struct xa_state *xas)
210 {
211         void *entry;
212         struct wait_exceptional_entry_queue ewait;
213         wait_queue_head_t *wq;
214 
215         init_wait(&ewait.wait);
216         ewait.wait.func = wake_exceptional_entry_func;
217 
218         for (;;) {
219                 entry = xas_find_conflict(xas);
220                 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)) ||
221                                 !dax_is_locked(entry))
222                         return entry;
223 
224                 wq = dax_entry_waitqueue(xas, entry, &ewait.key);
225                 prepare_to_wait_exclusive(wq, &ewait.wait,
226                                           TASK_UNINTERRUPTIBLE);
227                 xas_unlock_irq(xas);
228                 xas_reset(xas);
229                 schedule();
230                 finish_wait(wq, &ewait.wait);
231                 xas_lock_irq(xas);
232         }
233 }
234 
235 /*
236  * The only thing keeping the address space around is the i_pages lock
237  * (it's cycled in clear_inode() after removing the entries from i_pages)
238  * After we call xas_unlock_irq(), we cannot touch xas->xa.
239  */
240 static void wait_entry_unlocked(struct xa_state *xas, void *entry)
241 {
242         struct wait_exceptional_entry_queue ewait;
243         wait_queue_head_t *wq;
244 
245         init_wait(&ewait.wait);
246         ewait.wait.func = wake_exceptional_entry_func;
247 
248         wq = dax_entry_waitqueue(xas, entry, &ewait.key);
249         prepare_to_wait_exclusive(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
250         xas_unlock_irq(xas);
251         schedule();
252         finish_wait(wq, &ewait.wait);
253 
254         /*
255          * Entry lock waits are exclusive. Wake up the next waiter since
256          * we aren't sure we will acquire the entry lock and thus wake
257          * the next waiter up on unlock.
258          */
259         if (waitqueue_active(wq))
260                 __wake_up(wq, TASK_NORMAL, 1, &ewait.key);
261 }
262 
263 static void put_unlocked_entry(struct xa_state *xas, void *entry)
264 {
265         /* If we were the only waiter woken, wake the next one */
266         if (entry)
267                 dax_wake_entry(xas, entry, false);
268 }
269 
270 /*
271  * We used the xa_state to get the entry, but then we locked the entry and
272  * dropped the xa_lock, so we know the xa_state is stale and must be reset
273  * before use.
274  */
275 static void dax_unlock_entry(struct xa_state *xas, void *entry)
276 {
277         void *old;
278 
279         BUG_ON(dax_is_locked(entry));
280         xas_reset(xas);
281         xas_lock_irq(xas);
282         old = xas_store(xas, entry);
283         xas_unlock_irq(xas);
284         BUG_ON(!dax_is_locked(old));
285         dax_wake_entry(xas, entry, false);
286 }
287 
288 /*
289  * Return: The entry stored at this location before it was locked.
290  */
291 static void *dax_lock_entry(struct xa_state *xas, void *entry)
292 {
293         unsigned long v = xa_to_value(entry);
294         return xas_store(xas, xa_mk_value(v | DAX_LOCKED));
295 }
296 
297 static unsigned long dax_entry_size(void *entry)
298 {
299         if (dax_is_zero_entry(entry))
300                 return 0;
301         else if (dax_is_empty_entry(entry))
302                 return 0;
303         else if (dax_is_pmd_entry(entry))
304                 return PMD_SIZE;
305         else
306                 return PAGE_SIZE;
307 }
308 
309 static unsigned long dax_end_pfn(void *entry)
310 {
311         return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
312 }
313 
314 /*
315  * Iterate through all mapped pfns represented by an entry, i.e. skip
316  * 'empty' and 'zero' entries.
317  */
318 #define for_each_mapped_pfn(entry, pfn) \
319         for (pfn = dax_to_pfn(entry); \
320                         pfn < dax_end_pfn(entry); pfn++)
321 
322 /*
323  * TODO: for reflink+dax we need a way to associate a single page with
324  * multiple address_space instances at different linear_page_index()
325  * offsets.
326  */
327 static void dax_associate_entry(void *entry, struct address_space *mapping,
328                 struct vm_area_struct *vma, unsigned long address)
329 {
330         unsigned long size = dax_entry_size(entry), pfn, index;
331         int i = 0;
332 
333         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
334                 return;
335 
336         index = linear_page_index(vma, address & ~(size - 1));
337         for_each_mapped_pfn(entry, pfn) {
338                 struct page *page = pfn_to_page(pfn);
339 
340                 WARN_ON_ONCE(page->mapping);
341                 page->mapping = mapping;
342                 page->index = index + i++;
343         }
344 }
345 
346 static void dax_disassociate_entry(void *entry, struct address_space *mapping,
347                 bool trunc)
348 {
349         unsigned long pfn;
350 
351         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
352                 return;
353 
354         for_each_mapped_pfn(entry, pfn) {
355                 struct page *page = pfn_to_page(pfn);
356 
357                 WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
358                 WARN_ON_ONCE(page->mapping && page->mapping != mapping);
359                 page->mapping = NULL;
360                 page->index = 0;
361         }
362 }
363 
364 static struct page *dax_busy_page(void *entry)
365 {
366         unsigned long pfn;
367 
368         for_each_mapped_pfn(entry, pfn) {
369                 struct page *page = pfn_to_page(pfn);
370 
371                 if (page_ref_count(page) > 1)
372                         return page;
373         }
374         return NULL;
375 }
376 
377 /*
378  * dax_lock_mapping_entry - Lock the DAX entry corresponding to a page
379  * @page: The page whose entry we want to lock
380  *
381  * Context: Process context.
382  * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could
383  * not be locked.
384  */
385 dax_entry_t dax_lock_page(struct page *page)
386 {
387         XA_STATE(xas, NULL, 0);
388         void *entry;
389 
390         /* Ensure page->mapping isn't freed while we look at it */
391         rcu_read_lock();
392         for (;;) {
393                 struct address_space *mapping = READ_ONCE(page->mapping);
394 
395                 entry = NULL;
396                 if (!mapping || !dax_mapping(mapping))
397                         break;
398 
399                 /*
400                  * In the device-dax case there's no need to lock, a
401                  * struct dev_pagemap pin is sufficient to keep the
402                  * inode alive, and we assume we have dev_pagemap pin
403                  * otherwise we would not have a valid pfn_to_page()
404                  * translation.
405                  */
406                 entry = (void *)~0UL;
407                 if (S_ISCHR(mapping->host->i_mode))
408                         break;
409 
410                 xas.xa = &mapping->i_pages;
411                 xas_lock_irq(&xas);
412                 if (mapping != page->mapping) {
413                         xas_unlock_irq(&xas);
414                         continue;
415                 }
416                 xas_set(&xas, page->index);
417                 entry = xas_load(&xas);
418                 if (dax_is_locked(entry)) {
419                         rcu_read_unlock();
420                         wait_entry_unlocked(&xas, entry);
421                         rcu_read_lock();
422                         continue;
423                 }
424                 dax_lock_entry(&xas, entry);
425                 xas_unlock_irq(&xas);
426                 break;
427         }
428         rcu_read_unlock();
429         return (dax_entry_t)entry;
430 }
431 
432 void dax_unlock_page(struct page *page, dax_entry_t cookie)
433 {
434         struct address_space *mapping = page->mapping;
435         XA_STATE(xas, &mapping->i_pages, page->index);
436 
437         if (S_ISCHR(mapping->host->i_mode))
438                 return;
439 
440         dax_unlock_entry(&xas, (void *)cookie);
441 }
442 
443 /*
444  * Find page cache entry at given index. If it is a DAX entry, return it
445  * with the entry locked. If the page cache doesn't contain an entry at
446  * that index, add a locked empty entry.
447  *
448  * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
449  * either return that locked entry or will return VM_FAULT_FALLBACK.
450  * This will happen if there are any PTE entries within the PMD range
451  * that we are requesting.
452  *
453  * We always favor PTE entries over PMD entries. There isn't a flow where we
454  * evict PTE entries in order to 'upgrade' them to a PMD entry.  A PMD
455  * insertion will fail if it finds any PTE entries already in the tree, and a
456  * PTE insertion will cause an existing PMD entry to be unmapped and
457  * downgraded to PTE entries.  This happens for both PMD zero pages as
458  * well as PMD empty entries.
459  *
460  * The exception to this downgrade path is for PMD entries that have
461  * real storage backing them.  We will leave these real PMD entries in
462  * the tree, and PTE writes will simply dirty the entire PMD entry.
463  *
464  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
465  * persistent memory the benefit is doubtful. We can add that later if we can
466  * show it helps.
467  *
468  * On error, this function does not return an ERR_PTR.  Instead it returns
469  * a VM_FAULT code, encoded as an xarray internal entry.  The ERR_PTR values
470  * overlap with xarray value entries.
471  */
472 static void *grab_mapping_entry(struct xa_state *xas,
473                 struct address_space *mapping, unsigned long size_flag)
474 {
475         unsigned long index = xas->xa_index;
476         bool pmd_downgrade = false; /* splitting PMD entry into PTE entries? */
477         void *entry;
478 
479 retry:
480         xas_lock_irq(xas);
481         entry = get_unlocked_entry(xas);
482 
483         if (entry) {
484                 if (!xa_is_value(entry)) {
485                         xas_set_err(xas, EIO);
486                         goto out_unlock;
487                 }
488 
489                 if (size_flag & DAX_PMD) {
490                         if (dax_is_pte_entry(entry)) {
491                                 put_unlocked_entry(xas, entry);
492                                 goto fallback;
493                         }
494                 } else { /* trying to grab a PTE entry */
495                         if (dax_is_pmd_entry(entry) &&
496                             (dax_is_zero_entry(entry) ||
497                              dax_is_empty_entry(entry))) {
498                                 pmd_downgrade = true;
499                         }
500                 }
501         }
502 
503         if (pmd_downgrade) {
504                 /*
505                  * Make sure 'entry' remains valid while we drop
506                  * the i_pages lock.
507                  */
508                 dax_lock_entry(xas, entry);
509 
510                 /*
511                  * Besides huge zero pages the only other thing that gets
512                  * downgraded are empty entries which don't need to be
513                  * unmapped.
514                  */
515                 if (dax_is_zero_entry(entry)) {
516                         xas_unlock_irq(xas);
517                         unmap_mapping_pages(mapping,
518                                         xas->xa_index & ~PG_PMD_COLOUR,
519                                         PG_PMD_NR, false);
520                         xas_reset(xas);
521                         xas_lock_irq(xas);
522                 }
523 
524                 dax_disassociate_entry(entry, mapping, false);
525                 xas_store(xas, NULL);   /* undo the PMD join */
526                 dax_wake_entry(xas, entry, true);
527                 mapping->nrexceptional--;
528                 entry = NULL;
529                 xas_set(xas, index);
530         }
531 
532         if (entry) {
533                 dax_lock_entry(xas, entry);
534         } else {
535                 entry = dax_make_entry(pfn_to_pfn_t(0), size_flag | DAX_EMPTY);
536                 dax_lock_entry(xas, entry);
537                 if (xas_error(xas))
538                         goto out_unlock;
539                 mapping->nrexceptional++;
540         }
541 
542 out_unlock:
543         xas_unlock_irq(xas);
544         if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM))
545                 goto retry;
546         if (xas->xa_node == XA_ERROR(-ENOMEM))
547                 return xa_mk_internal(VM_FAULT_OOM);
548         if (xas_error(xas))
549                 return xa_mk_internal(VM_FAULT_SIGBUS);
550         return entry;
551 fallback:
552         xas_unlock_irq(xas);
553         return xa_mk_internal(VM_FAULT_FALLBACK);
554 }
555 
556 /**
557  * dax_layout_busy_page - find first pinned page in @mapping
558  * @mapping: address space to scan for a page with ref count > 1
559  *
560  * DAX requires ZONE_DEVICE mapped pages. These pages are never
561  * 'onlined' to the page allocator so they are considered idle when
562  * page->count == 1. A filesystem uses this interface to determine if
563  * any page in the mapping is busy, i.e. for DMA, or other
564  * get_user_pages() usages.
565  *
566  * It is expected that the filesystem is holding locks to block the
567  * establishment of new mappings in this address_space. I.e. it expects
568  * to be able to run unmap_mapping_range() and subsequently not race
569  * mapping_mapped() becoming true.
570  */
571 struct page *dax_layout_busy_page(struct address_space *mapping)
572 {
573         XA_STATE(xas, &mapping->i_pages, 0);
574         void *entry;
575         unsigned int scanned = 0;
576         struct page *page = NULL;
577 
578         /*
579          * In the 'limited' case get_user_pages() for dax is disabled.
580          */
581         if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
582                 return NULL;
583 
584         if (!dax_mapping(mapping) || !mapping_mapped(mapping))
585                 return NULL;
586 
587         /*
588          * If we race get_user_pages_fast() here either we'll see the
589          * elevated page count in the iteration and wait, or
590          * get_user_pages_fast() will see that the page it took a reference
591          * against is no longer mapped in the page tables and bail to the
592          * get_user_pages() slow path.  The slow path is protected by
593          * pte_lock() and pmd_lock(). New references are not taken without
594          * holding those locks, and unmap_mapping_range() will not zero the
595          * pte or pmd without holding the respective lock, so we are
596          * guaranteed to either see new references or prevent new
597          * references from being established.
598          */
599         unmap_mapping_range(mapping, 0, 0, 1);
600 
601         xas_lock_irq(&xas);
602         xas_for_each(&xas, entry, ULONG_MAX) {
603                 if (WARN_ON_ONCE(!xa_is_value(entry)))
604                         continue;
605                 if (unlikely(dax_is_locked(entry)))
606                         entry = get_unlocked_entry(&xas);
607                 if (entry)
608                         page = dax_busy_page(entry);
609                 put_unlocked_entry(&xas, entry);
610                 if (page)
611                         break;
612                 if (++scanned % XA_CHECK_SCHED)
613                         continue;
614 
615                 xas_pause(&xas);
616                 xas_unlock_irq(&xas);
617                 cond_resched();
618                 xas_lock_irq(&xas);
619         }
620         xas_unlock_irq(&xas);
621         return page;
622 }
623 EXPORT_SYMBOL_GPL(dax_layout_busy_page);
624 
625 static int __dax_invalidate_entry(struct address_space *mapping,
626                                           pgoff_t index, bool trunc)
627 {
628         XA_STATE(xas, &mapping->i_pages, index);
629         int ret = 0;
630         void *entry;
631 
632         xas_lock_irq(&xas);
633         entry = get_unlocked_entry(&xas);
634         if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
635                 goto out;
636         if (!trunc &&
637             (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) ||
638              xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE)))
639                 goto out;
640         dax_disassociate_entry(entry, mapping, trunc);
641         xas_store(&xas, NULL);
642         mapping->nrexceptional--;
643         ret = 1;
644 out:
645         put_unlocked_entry(&xas, entry);
646         xas_unlock_irq(&xas);
647         return ret;
648 }
649 
650 /*
651  * Delete DAX entry at @index from @mapping.  Wait for it
652  * to be unlocked before deleting it.
653  */
654 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
655 {
656         int ret = __dax_invalidate_entry(mapping, index, true);
657 
658         /*
659          * This gets called from truncate / punch_hole path. As such, the caller
660          * must hold locks protecting against concurrent modifications of the
661          * page cache (usually fs-private i_mmap_sem for writing). Since the
662          * caller has seen a DAX entry for this index, we better find it
663          * at that index as well...
664          */
665         WARN_ON_ONCE(!ret);
666         return ret;
667 }
668 
669 /*
670  * Invalidate DAX entry if it is clean.
671  */
672 int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
673                                       pgoff_t index)
674 {
675         return __dax_invalidate_entry(mapping, index, false);
676 }
677 
678 static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
679                 sector_t sector, size_t size, struct page *to,
680                 unsigned long vaddr)
681 {
682         void *vto, *kaddr;
683         pgoff_t pgoff;
684         long rc;
685         int id;
686 
687         rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
688         if (rc)
689                 return rc;
690 
691         id = dax_read_lock();
692         rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, NULL);
693         if (rc < 0) {
694                 dax_read_unlock(id);
695                 return rc;
696         }
697         vto = kmap_atomic(to);
698         copy_user_page(vto, (void __force *)kaddr, vaddr, to);
699         kunmap_atomic(vto);
700         dax_read_unlock(id);
701         return 0;
702 }
703 
704 /*
705  * By this point grab_mapping_entry() has ensured that we have a locked entry
706  * of the appropriate size so we don't have to worry about downgrading PMDs to
707  * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
708  * already in the tree, we will skip the insertion and just dirty the PMD as
709  * appropriate.
710  */
711 static void *dax_insert_entry(struct xa_state *xas,
712                 struct address_space *mapping, struct vm_fault *vmf,
713                 void *entry, pfn_t pfn, unsigned long flags, bool dirty)
714 {
715         void *new_entry = dax_make_entry(pfn, flags);
716 
717         if (dirty)
718                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
719 
720         if (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE)) {
721                 unsigned long index = xas->xa_index;
722                 /* we are replacing a zero page with block mapping */
723                 if (dax_is_pmd_entry(entry))
724                         unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
725                                         PG_PMD_NR, false);
726                 else /* pte entry */
727                         unmap_mapping_pages(mapping, index, 1, false);
728         }
729 
730         xas_reset(xas);
731         xas_lock_irq(xas);
732         if (dax_entry_size(entry) != dax_entry_size(new_entry)) {
733                 dax_disassociate_entry(entry, mapping, false);
734                 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address);
735         }
736 
737         if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
738                 /*
739                  * Only swap our new entry into the page cache if the current
740                  * entry is a zero page or an empty entry.  If a normal PTE or
741                  * PMD entry is already in the cache, we leave it alone.  This
742                  * means that if we are trying to insert a PTE and the
743                  * existing entry is a PMD, we will just leave the PMD in the
744                  * tree and dirty it if necessary.
745                  */
746                 void *old = dax_lock_entry(xas, new_entry);
747                 WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) |
748                                         DAX_LOCKED));
749                 entry = new_entry;
750         } else {
751                 xas_load(xas);  /* Walk the xa_state */
752         }
753 
754         if (dirty)
755                 xas_set_mark(xas, PAGECACHE_TAG_DIRTY);
756 
757         xas_unlock_irq(xas);
758         return entry;
759 }
760 
761 static inline
762 unsigned long pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
763 {
764         unsigned long address;
765 
766         address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
767         VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
768         return address;
769 }
770 
771 /* Walk all mappings of a given index of a file and writeprotect them */
772 static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index,
773                 unsigned long pfn)
774 {
775         struct vm_area_struct *vma;
776         pte_t pte, *ptep = NULL;
777         pmd_t *pmdp = NULL;
778         spinlock_t *ptl;
779 
780         i_mmap_lock_read(mapping);
781         vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
782                 unsigned long address, start, end;
783 
784                 cond_resched();
785 
786                 if (!(vma->vm_flags & VM_SHARED))
787                         continue;
788 
789                 address = pgoff_address(index, vma);
790 
791                 /*
792                  * Note because we provide start/end to follow_pte_pmd it will
793                  * call mmu_notifier_invalidate_range_start() on our behalf
794                  * before taking any lock.
795                  */
796                 if (follow_pte_pmd(vma->vm_mm, address, &start, &end, &ptep, &pmdp, &ptl))
797                         continue;
798 
799                 /*
800                  * No need to call mmu_notifier_invalidate_range() as we are
801                  * downgrading page table protection not changing it to point
802                  * to a new page.
803                  *
804                  * See Documentation/vm/mmu_notifier.rst
805                  */
806                 if (pmdp) {
807 #ifdef CONFIG_FS_DAX_PMD
808                         pmd_t pmd;
809 
810                         if (pfn != pmd_pfn(*pmdp))
811                                 goto unlock_pmd;
812                         if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
813                                 goto unlock_pmd;
814 
815                         flush_cache_page(vma, address, pfn);
816                         pmd = pmdp_huge_clear_flush(vma, address, pmdp);
817                         pmd = pmd_wrprotect(pmd);
818                         pmd = pmd_mkclean(pmd);
819                         set_pmd_at(vma->vm_mm, address, pmdp, pmd);
820 unlock_pmd:
821 #endif
822                         spin_unlock(ptl);
823                 } else {
824                         if (pfn != pte_pfn(*ptep))
825                                 goto unlock_pte;
826                         if (!pte_dirty(*ptep) && !pte_write(*ptep))
827                                 goto unlock_pte;
828 
829                         flush_cache_page(vma, address, pfn);
830                         pte = ptep_clear_flush(vma, address, ptep);
831                         pte = pte_wrprotect(pte);
832                         pte = pte_mkclean(pte);
833                         set_pte_at(vma->vm_mm, address, ptep, pte);
834 unlock_pte:
835                         pte_unmap_unlock(ptep, ptl);
836                 }
837 
838                 mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
839         }
840         i_mmap_unlock_read(mapping);
841 }
842 
843 static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev,
844                 struct address_space *mapping, void *entry)
845 {
846         unsigned long pfn;
847         long ret = 0;
848         size_t size;
849 
850         /*
851          * A page got tagged dirty in DAX mapping? Something is seriously
852          * wrong.
853          */
854         if (WARN_ON(!xa_is_value(entry)))
855                 return -EIO;
856 
857         if (unlikely(dax_is_locked(entry))) {
858                 void *old_entry = entry;
859 
860                 entry = get_unlocked_entry(xas);
861 
862                 /* Entry got punched out / reallocated? */
863                 if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
864                         goto put_unlocked;
865                 /*
866                  * Entry got reallocated elsewhere? No need to writeback.
867                  * We have to compare pfns as we must not bail out due to
868                  * difference in lockbit or entry type.
869                  */
870                 if (dax_to_pfn(old_entry) != dax_to_pfn(entry))
871                         goto put_unlocked;
872                 if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
873                                         dax_is_zero_entry(entry))) {
874                         ret = -EIO;
875                         goto put_unlocked;
876                 }
877 
878                 /* Another fsync thread may have already done this entry */
879                 if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE))
880                         goto put_unlocked;
881         }
882 
883         /* Lock the entry to serialize with page faults */
884         dax_lock_entry(xas, entry);
885 
886         /*
887          * We can clear the tag now but we have to be careful so that concurrent
888          * dax_writeback_one() calls for the same index cannot finish before we
889          * actually flush the caches. This is achieved as the calls will look
890          * at the entry only under the i_pages lock and once they do that
891          * they will see the entry locked and wait for it to unlock.
892          */
893         xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE);
894         xas_unlock_irq(xas);
895 
896         /*
897          * Even if dax_writeback_mapping_range() was given a wbc->range_start
898          * in the middle of a PMD, the 'index' we are given will be aligned to
899          * the start index of the PMD, as will the pfn we pull from 'entry'.
900          * This allows us to flush for PMD_SIZE and not have to worry about
901          * partial PMD writebacks.
902          */
903         pfn = dax_to_pfn(entry);
904         size = PAGE_SIZE << dax_entry_order(entry);
905 
906         dax_entry_mkclean(mapping, xas->xa_index, pfn);
907         dax_flush(dax_dev, page_address(pfn_to_page(pfn)), size);
908         /*
909          * After we have flushed the cache, we can clear the dirty tag. There
910          * cannot be new dirty data in the pfn after the flush has completed as
911          * the pfn mappings are writeprotected and fault waits for mapping
912          * entry lock.
913          */
914         xas_reset(xas);
915         xas_lock_irq(xas);
916         xas_store(xas, entry);
917         xas_clear_mark(xas, PAGECACHE_TAG_DIRTY);
918         dax_wake_entry(xas, entry, false);
919 
920         trace_dax_writeback_one(mapping->host, xas->xa_index,
921                         size >> PAGE_SHIFT);
922         return ret;
923 
924  put_unlocked:
925         put_unlocked_entry(xas, entry);
926         return ret;
927 }
928 
929 /*
930  * Flush the mapping to the persistent domain within the byte range of [start,
931  * end]. This is required by data integrity operations to ensure file data is
932  * on persistent storage prior to completion of the operation.
933  */
934 int dax_writeback_mapping_range(struct address_space *mapping,
935                 struct block_device *bdev, struct writeback_control *wbc)
936 {
937         XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT);
938         struct inode *inode = mapping->host;
939         pgoff_t end_index = wbc->range_end >> PAGE_SHIFT;
940         struct dax_device *dax_dev;
941         void *entry;
942         int ret = 0;
943         unsigned int scanned = 0;
944 
945         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
946                 return -EIO;
947 
948         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
949                 return 0;
950 
951         dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
952         if (!dax_dev)
953                 return -EIO;
954 
955         trace_dax_writeback_range(inode, xas.xa_index, end_index);
956 
957         tag_pages_for_writeback(mapping, xas.xa_index, end_index);
958 
959         xas_lock_irq(&xas);
960         xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) {
961                 ret = dax_writeback_one(&xas, dax_dev, mapping, entry);
962                 if (ret < 0) {
963                         mapping_set_error(mapping, ret);
964                         break;
965                 }
966                 if (++scanned % XA_CHECK_SCHED)
967                         continue;
968 
969                 xas_pause(&xas);
970                 xas_unlock_irq(&xas);
971                 cond_resched();
972                 xas_lock_irq(&xas);
973         }
974         xas_unlock_irq(&xas);
975         put_dax(dax_dev);
976         trace_dax_writeback_range_done(inode, xas.xa_index, end_index);
977         return ret;
978 }
979 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
980 
981 static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
982 {
983         return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9;
984 }
985 
986 static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size,
987                          pfn_t *pfnp)
988 {
989         const sector_t sector = dax_iomap_sector(iomap, pos);
990         pgoff_t pgoff;
991         int id, rc;
992         long length;
993 
994         rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff);
995         if (rc)
996                 return rc;
997         id = dax_read_lock();
998         length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
999                                    NULL, pfnp);
1000         if (length < 0) {
1001                 rc = length;
1002                 goto out;
1003         }
1004         rc = -EINVAL;
1005         if (PFN_PHYS(length) < size)
1006                 goto out;
1007         if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1008                 goto out;
1009         /* For larger pages we need devmap */
1010         if (length > 1 && !pfn_t_devmap(*pfnp))
1011                 goto out;
1012         rc = 0;
1013 out:
1014         dax_read_unlock(id);
1015         return rc;
1016 }
1017 
1018 /*
1019  * The user has performed a load from a hole in the file.  Allocating a new
1020  * page in the file would cause excessive storage usage for workloads with
1021  * sparse files.  Instead we insert a read-only mapping of the 4k zero page.
1022  * If this page is ever written to we will re-fault and change the mapping to
1023  * point to real DAX storage instead.
1024  */
1025 static vm_fault_t dax_load_hole(struct xa_state *xas,
1026                 struct address_space *mapping, void **entry,
1027                 struct vm_fault *vmf)
1028 {
1029         struct inode *inode = mapping->host;
1030         unsigned long vaddr = vmf->address;
1031         pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1032         vm_fault_t ret;
1033 
1034         *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1035                         DAX_ZERO_PAGE, false);
1036 
1037         ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1038         trace_dax_load_hole(inode, vmf, ret);
1039         return ret;
1040 }
1041 
1042 static bool dax_range_is_aligned(struct block_device *bdev,
1043                                  unsigned int offset, unsigned int length)
1044 {
1045         unsigned short sector_size = bdev_logical_block_size(bdev);
1046 
1047         if (!IS_ALIGNED(offset, sector_size))
1048                 return false;
1049         if (!IS_ALIGNED(length, sector_size))
1050                 return false;
1051 
1052         return true;
1053 }
1054 
1055 int __dax_zero_page_range(struct block_device *bdev,
1056                 struct dax_device *dax_dev, sector_t sector,
1057                 unsigned int offset, unsigned int size)
1058 {
1059         if (dax_range_is_aligned(bdev, offset, size)) {
1060                 sector_t start_sector = sector + (offset >> 9);
1061 
1062                 return blkdev_issue_zeroout(bdev, start_sector,
1063                                 size >> 9, GFP_NOFS, 0);
1064         } else {
1065                 pgoff_t pgoff;
1066                 long rc, id;
1067                 void *kaddr;
1068 
1069                 rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
1070                 if (rc)
1071                         return rc;
1072 
1073                 id = dax_read_lock();
1074                 rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr, NULL);
1075                 if (rc < 0) {
1076                         dax_read_unlock(id);
1077                         return rc;
1078                 }
1079                 memset(kaddr + offset, 0, size);
1080                 dax_flush(dax_dev, kaddr + offset, size);
1081                 dax_read_unlock(id);
1082         }
1083         return 0;
1084 }
1085 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1086 
1087 static loff_t
1088 dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1089                 struct iomap *iomap)
1090 {
1091         struct block_device *bdev = iomap->bdev;
1092         struct dax_device *dax_dev = iomap->dax_dev;
1093         struct iov_iter *iter = data;
1094         loff_t end = pos + length, done = 0;
1095         ssize_t ret = 0;
1096         size_t xfer;
1097         int id;
1098 
1099         if (iov_iter_rw(iter) == READ) {
1100                 end = min(end, i_size_read(inode));
1101                 if (pos >= end)
1102                         return 0;
1103 
1104                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1105                         return iov_iter_zero(min(length, end - pos), iter);
1106         }
1107 
1108         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1109                 return -EIO;
1110 
1111         /*
1112          * Write can allocate block for an area which has a hole page mapped
1113          * into page tables. We have to tear down these mappings so that data
1114          * written by write(2) is visible in mmap.
1115          */
1116         if (iomap->flags & IOMAP_F_NEW) {
1117                 invalidate_inode_pages2_range(inode->i_mapping,
1118                                               pos >> PAGE_SHIFT,
1119                                               (end - 1) >> PAGE_SHIFT);
1120         }
1121 
1122         id = dax_read_lock();
1123         while (pos < end) {
1124                 unsigned offset = pos & (PAGE_SIZE - 1);
1125                 const size_t size = ALIGN(length + offset, PAGE_SIZE);
1126                 const sector_t sector = dax_iomap_sector(iomap, pos);
1127                 ssize_t map_len;
1128                 pgoff_t pgoff;
1129                 void *kaddr;
1130 
1131                 if (fatal_signal_pending(current)) {
1132                         ret = -EINTR;
1133                         break;
1134                 }
1135 
1136                 ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
1137                 if (ret)
1138                         break;
1139 
1140                 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1141                                 &kaddr, NULL);
1142                 if (map_len < 0) {
1143                         ret = map_len;
1144                         break;
1145                 }
1146 
1147                 map_len = PFN_PHYS(map_len);
1148                 kaddr += offset;
1149                 map_len -= offset;
1150                 if (map_len > end - pos)
1151                         map_len = end - pos;
1152 
1153                 /*
1154                  * The userspace address for the memory copy has already been
1155                  * validated via access_ok() in either vfs_read() or
1156                  * vfs_write(), depending on which operation we are doing.
1157                  */
1158                 if (iov_iter_rw(iter) == WRITE)
1159                         xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1160                                         map_len, iter);
1161                 else
1162                         xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1163                                         map_len, iter);
1164 
1165                 pos += xfer;
1166                 length -= xfer;
1167                 done += xfer;
1168 
1169                 if (xfer == 0)
1170                         ret = -EFAULT;
1171                 if (xfer < map_len)
1172                         break;
1173         }
1174         dax_read_unlock(id);
1175 
1176         return done ? done : ret;
1177 }
1178 
1179 /**
1180  * dax_iomap_rw - Perform I/O to a DAX file
1181  * @iocb:       The control block for this I/O
1182  * @iter:       The addresses to do I/O from or to
1183  * @ops:        iomap ops passed from the file system
1184  *
1185  * This function performs read and write operations to directly mapped
1186  * persistent memory.  The callers needs to take care of read/write exclusion
1187  * and evicting any page cache pages in the region under I/O.
1188  */
1189 ssize_t
1190 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1191                 const struct iomap_ops *ops)
1192 {
1193         struct address_space *mapping = iocb->ki_filp->f_mapping;
1194         struct inode *inode = mapping->host;
1195         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1196         unsigned flags = 0;
1197 
1198         if (iov_iter_rw(iter) == WRITE) {
1199                 lockdep_assert_held_exclusive(&inode->i_rwsem);
1200                 flags |= IOMAP_WRITE;
1201         } else {
1202                 lockdep_assert_held(&inode->i_rwsem);
1203         }
1204 
1205         while (iov_iter_count(iter)) {
1206                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1207                                 iter, dax_iomap_actor);
1208                 if (ret <= 0)
1209                         break;
1210                 pos += ret;
1211                 done += ret;
1212         }
1213 
1214         iocb->ki_pos += done;
1215         return done ? done : ret;
1216 }
1217 EXPORT_SYMBOL_GPL(dax_iomap_rw);
1218 
1219 static vm_fault_t dax_fault_return(int error)
1220 {
1221         if (error == 0)
1222                 return VM_FAULT_NOPAGE;
1223         if (error == -ENOMEM)
1224                 return VM_FAULT_OOM;
1225         return VM_FAULT_SIGBUS;
1226 }
1227 
1228 /*
1229  * MAP_SYNC on a dax mapping guarantees dirty metadata is
1230  * flushed on write-faults (non-cow), but not read-faults.
1231  */
1232 static bool dax_fault_is_synchronous(unsigned long flags,
1233                 struct vm_area_struct *vma, struct iomap *iomap)
1234 {
1235         return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC)
1236                 && (iomap->flags & IOMAP_F_DIRTY);
1237 }
1238 
1239 static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1240                                int *iomap_errp, const struct iomap_ops *ops)
1241 {
1242         struct vm_area_struct *vma = vmf->vma;
1243         struct address_space *mapping = vma->vm_file->f_mapping;
1244         XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
1245         struct inode *inode = mapping->host;
1246         unsigned long vaddr = vmf->address;
1247         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1248         struct iomap iomap = { 0 };
1249         unsigned flags = IOMAP_FAULT;
1250         int error, major = 0;
1251         bool write = vmf->flags & FAULT_FLAG_WRITE;
1252         bool sync;
1253         vm_fault_t ret = 0;
1254         void *entry;
1255         pfn_t pfn;
1256 
1257         trace_dax_pte_fault(inode, vmf, ret);
1258         /*
1259          * Check whether offset isn't beyond end of file now. Caller is supposed
1260          * to hold locks serializing us with truncate / punch hole so this is
1261          * a reliable test.
1262          */
1263         if (pos >= i_size_read(inode)) {
1264                 ret = VM_FAULT_SIGBUS;
1265                 goto out;
1266         }
1267 
1268         if (write && !vmf->cow_page)
1269                 flags |= IOMAP_WRITE;
1270 
1271         entry = grab_mapping_entry(&xas, mapping, 0);
1272         if (xa_is_internal(entry)) {
1273                 ret = xa_to_internal(entry);
1274                 goto out;
1275         }
1276 
1277         /*
1278          * It is possible, particularly with mixed reads & writes to private
1279          * mappings, that we have raced with a PMD fault that overlaps with
1280          * the PTE we need to set up.  If so just return and the fault will be
1281          * retried.
1282          */
1283         if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1284                 ret = VM_FAULT_NOPAGE;
1285                 goto unlock_entry;
1286         }
1287 
1288         /*
1289          * Note that we don't bother to use iomap_apply here: DAX required
1290          * the file system block size to be equal the page size, which means
1291          * that we never have to deal with more than a single extent here.
1292          */
1293         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1294         if (iomap_errp)
1295                 *iomap_errp = error;
1296         if (error) {
1297                 ret = dax_fault_return(error);
1298                 goto unlock_entry;
1299         }
1300         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1301                 error = -EIO;   /* fs corruption? */
1302                 goto error_finish_iomap;
1303         }
1304 
1305         if (vmf->cow_page) {
1306                 sector_t sector = dax_iomap_sector(&iomap, pos);
1307 
1308                 switch (iomap.type) {
1309                 case IOMAP_HOLE:
1310                 case IOMAP_UNWRITTEN:
1311                         clear_user_highpage(vmf->cow_page, vaddr);
1312                         break;
1313                 case IOMAP_MAPPED:
1314                         error = copy_user_dax(iomap.bdev, iomap.dax_dev,
1315                                         sector, PAGE_SIZE, vmf->cow_page, vaddr);
1316                         break;
1317                 default:
1318                         WARN_ON_ONCE(1);
1319                         error = -EIO;
1320                         break;
1321                 }
1322 
1323                 if (error)
1324                         goto error_finish_iomap;
1325 
1326                 __SetPageUptodate(vmf->cow_page);
1327                 ret = finish_fault(vmf);
1328                 if (!ret)
1329                         ret = VM_FAULT_DONE_COW;
1330                 goto finish_iomap;
1331         }
1332 
1333         sync = dax_fault_is_synchronous(flags, vma, &iomap);
1334 
1335         switch (iomap.type) {
1336         case IOMAP_MAPPED:
1337                 if (iomap.flags & IOMAP_F_NEW) {
1338                         count_vm_event(PGMAJFAULT);
1339                         count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
1340                         major = VM_FAULT_MAJOR;
1341                 }
1342                 error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn);
1343                 if (error < 0)
1344                         goto error_finish_iomap;
1345 
1346                 entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
1347                                                  0, write && !sync);
1348 
1349                 /*
1350                  * If we are doing synchronous page fault and inode needs fsync,
1351                  * we can insert PTE into page tables only after that happens.
1352                  * Skip insertion for now and return the pfn so that caller can
1353                  * insert it after fsync is done.
1354                  */
1355                 if (sync) {
1356                         if (WARN_ON_ONCE(!pfnp)) {
1357                                 error = -EIO;
1358                                 goto error_finish_iomap;
1359                         }
1360                         *pfnp = pfn;
1361                         ret = VM_FAULT_NEEDDSYNC | major;
1362                         goto finish_iomap;
1363                 }
1364                 trace_dax_insert_mapping(inode, vmf, entry);
1365                 if (write)
1366                         ret = vmf_insert_mixed_mkwrite(vma, vaddr, pfn);
1367                 else
1368                         ret = vmf_insert_mixed(vma, vaddr, pfn);
1369 
1370                 goto finish_iomap;
1371         case IOMAP_UNWRITTEN:
1372         case IOMAP_HOLE:
1373                 if (!write) {
1374                         ret = dax_load_hole(&xas, mapping, &entry, vmf);
1375                         goto finish_iomap;
1376                 }
1377                 /*FALLTHRU*/
1378         default:
1379                 WARN_ON_ONCE(1);
1380                 error = -EIO;
1381                 break;
1382         }
1383 
1384  error_finish_iomap:
1385         ret = dax_fault_return(error);
1386  finish_iomap:
1387         if (ops->iomap_end) {
1388                 int copied = PAGE_SIZE;
1389 
1390                 if (ret & VM_FAULT_ERROR)
1391                         copied = 0;
1392                 /*
1393                  * The fault is done by now and there's no way back (other
1394                  * thread may be already happily using PTE we have installed).
1395                  * Just ignore error from ->iomap_end since we cannot do much
1396                  * with it.
1397                  */
1398                 ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
1399         }
1400  unlock_entry:
1401         dax_unlock_entry(&xas, entry);
1402  out:
1403         trace_dax_pte_fault_done(inode, vmf, ret);
1404         return ret | major;
1405 }
1406 
1407 #ifdef CONFIG_FS_DAX_PMD
1408 static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1409                 struct iomap *iomap, void **entry)
1410 {
1411         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1412         unsigned long pmd_addr = vmf->address & PMD_MASK;
1413         struct inode *inode = mapping->host;
1414         struct page *zero_page;
1415         spinlock_t *ptl;
1416         pmd_t pmd_entry;
1417         pfn_t pfn;
1418 
1419         zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1420 
1421         if (unlikely(!zero_page))
1422                 goto fallback;
1423 
1424         pfn = page_to_pfn_t(zero_page);
1425         *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn,
1426                         DAX_PMD | DAX_ZERO_PAGE, false);
1427 
1428         ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1429         if (!pmd_none(*(vmf->pmd))) {
1430                 spin_unlock(ptl);
1431                 goto fallback;
1432         }
1433 
1434         pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1435         pmd_entry = pmd_mkhuge(pmd_entry);
1436         set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1437         spin_unlock(ptl);
1438         trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry);
1439         return VM_FAULT_NOPAGE;
1440 
1441 fallback:
1442         trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry);
1443         return VM_FAULT_FALLBACK;
1444 }
1445 
1446 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1447                                const struct iomap_ops *ops)
1448 {
1449         struct vm_area_struct *vma = vmf->vma;
1450         struct address_space *mapping = vma->vm_file->f_mapping;
1451         XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
1452         unsigned long pmd_addr = vmf->address & PMD_MASK;
1453         bool write = vmf->flags & FAULT_FLAG_WRITE;
1454         bool sync;
1455         unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
1456         struct inode *inode = mapping->host;
1457         vm_fault_t result = VM_FAULT_FALLBACK;
1458         struct iomap iomap = { 0 };
1459         pgoff_t max_pgoff;
1460         void *entry;
1461         loff_t pos;
1462         int error;
1463         pfn_t pfn;
1464 
1465         /*
1466          * Check whether offset isn't beyond end of file now. Caller is
1467          * supposed to hold locks serializing us with truncate / punch hole so
1468          * this is a reliable test.
1469          */
1470         max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
1471 
1472         trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
1473 
1474         /*
1475          * Make sure that the faulting address's PMD offset (color) matches
1476          * the PMD offset from the start of the file.  This is necessary so
1477          * that a PMD range in the page table overlaps exactly with a PMD
1478          * range in the page cache.
1479          */
1480         if ((vmf->pgoff & PG_PMD_COLOUR) !=
1481             ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1482                 goto fallback;
1483 
1484         /* Fall back to PTEs if we're going to COW */
1485         if (write && !(vma->vm_flags & VM_SHARED))
1486                 goto fallback;
1487 
1488         /* If the PMD would extend outside the VMA */
1489         if (pmd_addr < vma->vm_start)
1490                 goto fallback;
1491         if ((pmd_addr + PMD_SIZE) > vma->vm_end)
1492                 goto fallback;
1493 
1494         if (xas.xa_index >= max_pgoff) {
1495                 result = VM_FAULT_SIGBUS;
1496                 goto out;
1497         }
1498 
1499         /* If the PMD would extend beyond the file size */
1500         if ((xas.xa_index | PG_PMD_COLOUR) >= max_pgoff)
1501                 goto fallback;
1502 
1503         /*
1504          * grab_mapping_entry() will make sure we get an empty PMD entry,
1505          * a zero PMD entry or a DAX PMD.  If it can't (because a PTE
1506          * entry is already in the array, for instance), it will return
1507          * VM_FAULT_FALLBACK.
1508          */
1509         entry = grab_mapping_entry(&xas, mapping, DAX_PMD);
1510         if (xa_is_internal(entry)) {
1511                 result = xa_to_internal(entry);
1512                 goto fallback;
1513         }
1514 
1515         /*
1516          * It is possible, particularly with mixed reads & writes to private
1517          * mappings, that we have raced with a PTE fault that overlaps with
1518          * the PMD we need to set up.  If so just return and the fault will be
1519          * retried.
1520          */
1521         if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1522                         !pmd_devmap(*vmf->pmd)) {
1523                 result = 0;
1524                 goto unlock_entry;
1525         }
1526 
1527         /*
1528          * Note that we don't use iomap_apply here.  We aren't doing I/O, only
1529          * setting up a mapping, so really we're using iomap_begin() as a way
1530          * to look up our filesystem block.
1531          */
1532         pos = (loff_t)xas.xa_index << PAGE_SHIFT;
1533         error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
1534         if (error)
1535                 goto unlock_entry;
1536 
1537         if (iomap.offset + iomap.length < pos + PMD_SIZE)
1538                 goto finish_iomap;
1539 
1540         sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap);
1541 
1542         switch (iomap.type) {
1543         case IOMAP_MAPPED:
1544                 error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn);
1545                 if (error < 0)
1546                         goto finish_iomap;
1547 
1548                 entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn,
1549                                                 DAX_PMD, write && !sync);
1550 
1551                 /*
1552                  * If we are doing synchronous page fault and inode needs fsync,
1553                  * we can insert PMD into page tables only after that happens.
1554                  * Skip insertion for now and return the pfn so that caller can
1555                  * insert it after fsync is done.
1556                  */
1557                 if (sync) {
1558                         if (WARN_ON_ONCE(!pfnp))
1559                                 goto finish_iomap;
1560                         *pfnp = pfn;
1561                         result = VM_FAULT_NEEDDSYNC;
1562                         goto finish_iomap;
1563                 }
1564 
1565                 trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry);
1566                 result = vmf_insert_pfn_pmd(vma, vmf->address, vmf->pmd, pfn,
1567                                             write);
1568                 break;
1569         case IOMAP_UNWRITTEN:
1570         case IOMAP_HOLE:
1571                 if (WARN_ON_ONCE(write))
1572                         break;
1573                 result = dax_pmd_load_hole(&xas, vmf, &iomap, &entry);
1574                 break;
1575         default:
1576                 WARN_ON_ONCE(1);
1577                 break;
1578         }
1579 
1580  finish_iomap:
1581         if (ops->iomap_end) {
1582                 int copied = PMD_SIZE;
1583 
1584                 if (result == VM_FAULT_FALLBACK)
1585                         copied = 0;
1586                 /*
1587                  * The fault is done by now and there's no way back (other
1588                  * thread may be already happily using PMD we have installed).
1589                  * Just ignore error from ->iomap_end since we cannot do much
1590                  * with it.
1591                  */
1592                 ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
1593                                 &iomap);
1594         }
1595  unlock_entry:
1596         dax_unlock_entry(&xas, entry);
1597  fallback:
1598         if (result == VM_FAULT_FALLBACK) {
1599                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1600                 count_vm_event(THP_FAULT_FALLBACK);
1601         }
1602 out:
1603         trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
1604         return result;
1605 }
1606 #else
1607 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1608                                const struct iomap_ops *ops)
1609 {
1610         return VM_FAULT_FALLBACK;
1611 }
1612 #endif /* CONFIG_FS_DAX_PMD */
1613 
1614 /**
1615  * dax_iomap_fault - handle a page fault on a DAX file
1616  * @vmf: The description of the fault
1617  * @pe_size: Size of the page to fault in
1618  * @pfnp: PFN to insert for synchronous faults if fsync is required
1619  * @iomap_errp: Storage for detailed error code in case of error
1620  * @ops: Iomap ops passed from the file system
1621  *
1622  * When a page fault occurs, filesystems may call this helper in
1623  * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1624  * has done all the necessary locking for page fault to proceed
1625  * successfully.
1626  */
1627 vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1628                     pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1629 {
1630         switch (pe_size) {
1631         case PE_SIZE_PTE:
1632                 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1633         case PE_SIZE_PMD:
1634                 return dax_iomap_pmd_fault(vmf, pfnp, ops);
1635         default:
1636                 return VM_FAULT_FALLBACK;
1637         }
1638 }
1639 EXPORT_SYMBOL_GPL(dax_iomap_fault);
1640 
1641 /*
1642  * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1643  * @vmf: The description of the fault
1644  * @pfn: PFN to insert
1645  * @order: Order of entry to insert.
1646  *
1647  * This function inserts a writeable PTE or PMD entry into the page tables
1648  * for an mmaped DAX file.  It also marks the page cache entry as dirty.
1649  */
1650 static vm_fault_t
1651 dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
1652 {
1653         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1654         XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1655         void *entry;
1656         vm_fault_t ret;
1657 
1658         xas_lock_irq(&xas);
1659         entry = get_unlocked_entry(&xas);
1660         /* Did we race with someone splitting entry or so? */
1661         if (!entry ||
1662             (order == 0 && !dax_is_pte_entry(entry)) ||
1663             (order == PMD_ORDER && !dax_is_pmd_entry(entry))) {
1664                 put_unlocked_entry(&xas, entry);
1665                 xas_unlock_irq(&xas);
1666                 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1667                                                       VM_FAULT_NOPAGE);
1668                 return VM_FAULT_NOPAGE;
1669         }
1670         xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1671         dax_lock_entry(&xas, entry);
1672         xas_unlock_irq(&xas);
1673         if (order == 0)
1674                 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1675 #ifdef CONFIG_FS_DAX_PMD
1676         else if (order == PMD_ORDER)
1677                 ret = vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd,
1678                         pfn, true);
1679 #endif
1680         else
1681                 ret = VM_FAULT_FALLBACK;
1682         dax_unlock_entry(&xas, entry);
1683         trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1684         return ret;
1685 }
1686 
1687 /**
1688  * dax_finish_sync_fault - finish synchronous page fault
1689  * @vmf: The description of the fault
1690  * @pe_size: Size of entry to be inserted
1691  * @pfn: PFN to insert
1692  *
1693  * This function ensures that the file range touched by the page fault is
1694  * stored persistently on the media and handles inserting of appropriate page
1695  * table entry.
1696  */
1697 vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf,
1698                 enum page_entry_size pe_size, pfn_t pfn)
1699 {
1700         int err;
1701         loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1702         unsigned int order = pe_order(pe_size);
1703         size_t len = PAGE_SIZE << order;
1704 
1705         err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1706         if (err)
1707                 return VM_FAULT_SIGBUS;
1708         return dax_insert_pfn_mkwrite(vmf, pfn, order);
1709 }
1710 EXPORT_SYMBOL_GPL(dax_finish_sync_fault);
1711 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp