25c7d7509cf49efe4bc29ed811f05df8de9b6a48
[muen/linux.git] / mm / huge_memory.c
1 /*
2  *  Copyright (C) 2009  Red Hat, Inc.
3  *
4  *  This work is licensed under the terms of the GNU GPL, version 2. See
5  *  the COPYING file in the top-level directory.
6  */
7
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10 #include <linux/mm.h>
11 #include <linux/sched.h>
12 #include <linux/sched/coredump.h>
13 #include <linux/sched/numa_balancing.h>
14 #include <linux/highmem.h>
15 #include <linux/hugetlb.h>
16 #include <linux/mmu_notifier.h>
17 #include <linux/rmap.h>
18 #include <linux/swap.h>
19 #include <linux/shrinker.h>
20 #include <linux/mm_inline.h>
21 #include <linux/swapops.h>
22 #include <linux/dax.h>
23 #include <linux/khugepaged.h>
24 #include <linux/freezer.h>
25 #include <linux/pfn_t.h>
26 #include <linux/mman.h>
27 #include <linux/memremap.h>
28 #include <linux/pagemap.h>
29 #include <linux/debugfs.h>
30 #include <linux/migrate.h>
31 #include <linux/hashtable.h>
32 #include <linux/userfaultfd_k.h>
33 #include <linux/page_idle.h>
34 #include <linux/shmem_fs.h>
35 #include <linux/oom.h>
36
37 #include <asm/tlb.h>
38 #include <asm/pgalloc.h>
39 #include "internal.h"
40
41 /*
42  * By default, transparent hugepage support is disabled in order to avoid
43  * risking an increased memory footprint for applications that are not
44  * guaranteed to benefit from it. When transparent hugepage support is
45  * enabled, it is for all mappings, and khugepaged scans all mappings.
46  * Defrag is invoked by khugepaged hugepage allocations and by page faults
47  * for all hugepage allocations.
48  */
49 unsigned long transparent_hugepage_flags __read_mostly =
50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
51         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
52 #endif
53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
54         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
55 #endif
56         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
57         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
58         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
59
60 static struct shrinker deferred_split_shrinker;
61
62 static atomic_t huge_zero_refcount;
63 struct page *huge_zero_page __read_mostly;
64
65 static struct page *get_huge_zero_page(void)
66 {
67         struct page *zero_page;
68 retry:
69         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
70                 return READ_ONCE(huge_zero_page);
71
72         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
73                         HPAGE_PMD_ORDER);
74         if (!zero_page) {
75                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
76                 return NULL;
77         }
78         count_vm_event(THP_ZERO_PAGE_ALLOC);
79         preempt_disable();
80         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
81                 preempt_enable();
82                 __free_pages(zero_page, compound_order(zero_page));
83                 goto retry;
84         }
85
86         /* We take additional reference here. It will be put back by shrinker */
87         atomic_set(&huge_zero_refcount, 2);
88         preempt_enable();
89         return READ_ONCE(huge_zero_page);
90 }
91
92 static void put_huge_zero_page(void)
93 {
94         /*
95          * Counter should never go to zero here. Only shrinker can put
96          * last reference.
97          */
98         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
99 }
100
101 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
102 {
103         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
104                 return READ_ONCE(huge_zero_page);
105
106         if (!get_huge_zero_page())
107                 return NULL;
108
109         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
110                 put_huge_zero_page();
111
112         return READ_ONCE(huge_zero_page);
113 }
114
115 void mm_put_huge_zero_page(struct mm_struct *mm)
116 {
117         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
118                 put_huge_zero_page();
119 }
120
121 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
122                                         struct shrink_control *sc)
123 {
124         /* we can free zero page only if last reference remains */
125         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
126 }
127
128 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
129                                        struct shrink_control *sc)
130 {
131         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
132                 struct page *zero_page = xchg(&huge_zero_page, NULL);
133                 BUG_ON(zero_page == NULL);
134                 __free_pages(zero_page, compound_order(zero_page));
135                 return HPAGE_PMD_NR;
136         }
137
138         return 0;
139 }
140
141 static struct shrinker huge_zero_page_shrinker = {
142         .count_objects = shrink_huge_zero_page_count,
143         .scan_objects = shrink_huge_zero_page_scan,
144         .seeks = DEFAULT_SEEKS,
145 };
146
147 #ifdef CONFIG_SYSFS
148 static ssize_t enabled_show(struct kobject *kobj,
149                             struct kobj_attribute *attr, char *buf)
150 {
151         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
152                 return sprintf(buf, "[always] madvise never\n");
153         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
154                 return sprintf(buf, "always [madvise] never\n");
155         else
156                 return sprintf(buf, "always madvise [never]\n");
157 }
158
159 static ssize_t enabled_store(struct kobject *kobj,
160                              struct kobj_attribute *attr,
161                              const char *buf, size_t count)
162 {
163         ssize_t ret = count;
164
165         if (!memcmp("always", buf,
166                     min(sizeof("always")-1, count))) {
167                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
168                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
169         } else if (!memcmp("madvise", buf,
170                            min(sizeof("madvise")-1, count))) {
171                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
172                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
173         } else if (!memcmp("never", buf,
174                            min(sizeof("never")-1, count))) {
175                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
176                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
177         } else
178                 ret = -EINVAL;
179
180         if (ret > 0) {
181                 int err = start_stop_khugepaged();
182                 if (err)
183                         ret = err;
184         }
185         return ret;
186 }
187 static struct kobj_attribute enabled_attr =
188         __ATTR(enabled, 0644, enabled_show, enabled_store);
189
190 ssize_t single_hugepage_flag_show(struct kobject *kobj,
191                                 struct kobj_attribute *attr, char *buf,
192                                 enum transparent_hugepage_flag flag)
193 {
194         return sprintf(buf, "%d\n",
195                        !!test_bit(flag, &transparent_hugepage_flags));
196 }
197
198 ssize_t single_hugepage_flag_store(struct kobject *kobj,
199                                  struct kobj_attribute *attr,
200                                  const char *buf, size_t count,
201                                  enum transparent_hugepage_flag flag)
202 {
203         unsigned long value;
204         int ret;
205
206         ret = kstrtoul(buf, 10, &value);
207         if (ret < 0)
208                 return ret;
209         if (value > 1)
210                 return -EINVAL;
211
212         if (value)
213                 set_bit(flag, &transparent_hugepage_flags);
214         else
215                 clear_bit(flag, &transparent_hugepage_flags);
216
217         return count;
218 }
219
220 static ssize_t defrag_show(struct kobject *kobj,
221                            struct kobj_attribute *attr, char *buf)
222 {
223         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
224                 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
225         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
226                 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
227         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
228                 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
229         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
230                 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
231         return sprintf(buf, "always defer defer+madvise madvise [never]\n");
232 }
233
234 static ssize_t defrag_store(struct kobject *kobj,
235                             struct kobj_attribute *attr,
236                             const char *buf, size_t count)
237 {
238         if (!memcmp("always", buf,
239                     min(sizeof("always")-1, count))) {
240                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
241                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
242                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
243                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
244         } else if (!memcmp("defer+madvise", buf,
245                     min(sizeof("defer+madvise")-1, count))) {
246                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
247                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
248                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
249                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
250         } else if (!memcmp("defer", buf,
251                     min(sizeof("defer")-1, count))) {
252                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
253                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
254                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
255                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
256         } else if (!memcmp("madvise", buf,
257                            min(sizeof("madvise")-1, count))) {
258                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
259                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
260                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
261                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
262         } else if (!memcmp("never", buf,
263                            min(sizeof("never")-1, count))) {
264                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
265                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
266                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
267                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
268         } else
269                 return -EINVAL;
270
271         return count;
272 }
273 static struct kobj_attribute defrag_attr =
274         __ATTR(defrag, 0644, defrag_show, defrag_store);
275
276 static ssize_t use_zero_page_show(struct kobject *kobj,
277                 struct kobj_attribute *attr, char *buf)
278 {
279         return single_hugepage_flag_show(kobj, attr, buf,
280                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
281 }
282 static ssize_t use_zero_page_store(struct kobject *kobj,
283                 struct kobj_attribute *attr, const char *buf, size_t count)
284 {
285         return single_hugepage_flag_store(kobj, attr, buf, count,
286                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
287 }
288 static struct kobj_attribute use_zero_page_attr =
289         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
290
291 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
292                 struct kobj_attribute *attr, char *buf)
293 {
294         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
295 }
296 static struct kobj_attribute hpage_pmd_size_attr =
297         __ATTR_RO(hpage_pmd_size);
298
299 #ifdef CONFIG_DEBUG_VM
300 static ssize_t debug_cow_show(struct kobject *kobj,
301                                 struct kobj_attribute *attr, char *buf)
302 {
303         return single_hugepage_flag_show(kobj, attr, buf,
304                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
305 }
306 static ssize_t debug_cow_store(struct kobject *kobj,
307                                struct kobj_attribute *attr,
308                                const char *buf, size_t count)
309 {
310         return single_hugepage_flag_store(kobj, attr, buf, count,
311                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
312 }
313 static struct kobj_attribute debug_cow_attr =
314         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
315 #endif /* CONFIG_DEBUG_VM */
316
317 static struct attribute *hugepage_attr[] = {
318         &enabled_attr.attr,
319         &defrag_attr.attr,
320         &use_zero_page_attr.attr,
321         &hpage_pmd_size_attr.attr,
322 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
323         &shmem_enabled_attr.attr,
324 #endif
325 #ifdef CONFIG_DEBUG_VM
326         &debug_cow_attr.attr,
327 #endif
328         NULL,
329 };
330
331 static const struct attribute_group hugepage_attr_group = {
332         .attrs = hugepage_attr,
333 };
334
335 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
336 {
337         int err;
338
339         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
340         if (unlikely(!*hugepage_kobj)) {
341                 pr_err("failed to create transparent hugepage kobject\n");
342                 return -ENOMEM;
343         }
344
345         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
346         if (err) {
347                 pr_err("failed to register transparent hugepage group\n");
348                 goto delete_obj;
349         }
350
351         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
352         if (err) {
353                 pr_err("failed to register transparent hugepage group\n");
354                 goto remove_hp_group;
355         }
356
357         return 0;
358
359 remove_hp_group:
360         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
361 delete_obj:
362         kobject_put(*hugepage_kobj);
363         return err;
364 }
365
366 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
367 {
368         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
369         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
370         kobject_put(hugepage_kobj);
371 }
372 #else
373 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
374 {
375         return 0;
376 }
377
378 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
379 {
380 }
381 #endif /* CONFIG_SYSFS */
382
383 static int __init hugepage_init(void)
384 {
385         int err;
386         struct kobject *hugepage_kobj;
387
388         if (!has_transparent_hugepage()) {
389                 transparent_hugepage_flags = 0;
390                 return -EINVAL;
391         }
392
393         /*
394          * hugepages can't be allocated by the buddy allocator
395          */
396         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
397         /*
398          * we use page->mapping and page->index in second tail page
399          * as list_head: assuming THP order >= 2
400          */
401         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
402
403         err = hugepage_init_sysfs(&hugepage_kobj);
404         if (err)
405                 goto err_sysfs;
406
407         err = khugepaged_init();
408         if (err)
409                 goto err_slab;
410
411         err = register_shrinker(&huge_zero_page_shrinker);
412         if (err)
413                 goto err_hzp_shrinker;
414         err = register_shrinker(&deferred_split_shrinker);
415         if (err)
416                 goto err_split_shrinker;
417
418         /*
419          * By default disable transparent hugepages on smaller systems,
420          * where the extra memory used could hurt more than TLB overhead
421          * is likely to save.  The admin can still enable it through /sys.
422          */
423         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
424                 transparent_hugepage_flags = 0;
425                 return 0;
426         }
427
428         err = start_stop_khugepaged();
429         if (err)
430                 goto err_khugepaged;
431
432         return 0;
433 err_khugepaged:
434         unregister_shrinker(&deferred_split_shrinker);
435 err_split_shrinker:
436         unregister_shrinker(&huge_zero_page_shrinker);
437 err_hzp_shrinker:
438         khugepaged_destroy();
439 err_slab:
440         hugepage_exit_sysfs(hugepage_kobj);
441 err_sysfs:
442         return err;
443 }
444 subsys_initcall(hugepage_init);
445
446 static int __init setup_transparent_hugepage(char *str)
447 {
448         int ret = 0;
449         if (!str)
450                 goto out;
451         if (!strcmp(str, "always")) {
452                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
453                         &transparent_hugepage_flags);
454                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
455                           &transparent_hugepage_flags);
456                 ret = 1;
457         } else if (!strcmp(str, "madvise")) {
458                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
459                           &transparent_hugepage_flags);
460                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
461                         &transparent_hugepage_flags);
462                 ret = 1;
463         } else if (!strcmp(str, "never")) {
464                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
465                           &transparent_hugepage_flags);
466                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
467                           &transparent_hugepage_flags);
468                 ret = 1;
469         }
470 out:
471         if (!ret)
472                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
473         return ret;
474 }
475 __setup("transparent_hugepage=", setup_transparent_hugepage);
476
477 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
478 {
479         if (likely(vma->vm_flags & VM_WRITE))
480                 pmd = pmd_mkwrite(pmd);
481         return pmd;
482 }
483
484 static inline struct list_head *page_deferred_list(struct page *page)
485 {
486         /* ->lru in the tail pages is occupied by compound_head. */
487         return &page[2].deferred_list;
488 }
489
490 void prep_transhuge_page(struct page *page)
491 {
492         /*
493          * we use page->mapping and page->indexlru in second tail page
494          * as list_head: assuming THP order >= 2
495          */
496
497         INIT_LIST_HEAD(page_deferred_list(page));
498         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
499 }
500
501 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
502                 loff_t off, unsigned long flags, unsigned long size)
503 {
504         unsigned long addr;
505         loff_t off_end = off + len;
506         loff_t off_align = round_up(off, size);
507         unsigned long len_pad;
508
509         if (off_end <= off_align || (off_end - off_align) < size)
510                 return 0;
511
512         len_pad = len + size;
513         if (len_pad < len || (off + len_pad) < off)
514                 return 0;
515
516         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
517                                               off >> PAGE_SHIFT, flags);
518         if (IS_ERR_VALUE(addr))
519                 return 0;
520
521         addr += (off - addr) & (size - 1);
522         return addr;
523 }
524
525 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
526                 unsigned long len, unsigned long pgoff, unsigned long flags)
527 {
528         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
529
530         if (addr)
531                 goto out;
532         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
533                 goto out;
534
535         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
536         if (addr)
537                 return addr;
538
539  out:
540         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
541 }
542 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
543
544 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
545                         struct page *page, gfp_t gfp)
546 {
547         struct vm_area_struct *vma = vmf->vma;
548         struct mem_cgroup *memcg;
549         pgtable_t pgtable;
550         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
551         vm_fault_t ret = 0;
552
553         VM_BUG_ON_PAGE(!PageCompound(page), page);
554
555         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
556                 put_page(page);
557                 count_vm_event(THP_FAULT_FALLBACK);
558                 return VM_FAULT_FALLBACK;
559         }
560
561         pgtable = pte_alloc_one(vma->vm_mm, haddr);
562         if (unlikely(!pgtable)) {
563                 ret = VM_FAULT_OOM;
564                 goto release;
565         }
566
567         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
568         /*
569          * The memory barrier inside __SetPageUptodate makes sure that
570          * clear_huge_page writes become visible before the set_pmd_at()
571          * write.
572          */
573         __SetPageUptodate(page);
574
575         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
576         if (unlikely(!pmd_none(*vmf->pmd))) {
577                 goto unlock_release;
578         } else {
579                 pmd_t entry;
580
581                 ret = check_stable_address_space(vma->vm_mm);
582                 if (ret)
583                         goto unlock_release;
584
585                 /* Deliver the page fault to userland */
586                 if (userfaultfd_missing(vma)) {
587                         vm_fault_t ret2;
588
589                         spin_unlock(vmf->ptl);
590                         mem_cgroup_cancel_charge(page, memcg, true);
591                         put_page(page);
592                         pte_free(vma->vm_mm, pgtable);
593                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
594                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
595                         return ret2;
596                 }
597
598                 entry = mk_huge_pmd(page, vma->vm_page_prot);
599                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
600                 page_add_new_anon_rmap(page, vma, haddr, true);
601                 mem_cgroup_commit_charge(page, memcg, false, true);
602                 lru_cache_add_active_or_unevictable(page, vma);
603                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
604                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
605                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
606                 mm_inc_nr_ptes(vma->vm_mm);
607                 spin_unlock(vmf->ptl);
608                 count_vm_event(THP_FAULT_ALLOC);
609         }
610
611         return 0;
612 unlock_release:
613         spin_unlock(vmf->ptl);
614 release:
615         if (pgtable)
616                 pte_free(vma->vm_mm, pgtable);
617         mem_cgroup_cancel_charge(page, memcg, true);
618         put_page(page);
619         return ret;
620
621 }
622
623 /*
624  * always: directly stall for all thp allocations
625  * defer: wake kswapd and fail if not immediately available
626  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
627  *                fail if not immediately available
628  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
629  *          available
630  * never: never stall for any thp allocation
631  */
632 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
633 {
634         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
635
636         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
637                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
638         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
639                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
640         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
641                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
642                                                              __GFP_KSWAPD_RECLAIM);
643         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
644                 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
645                                                              0);
646         return GFP_TRANSHUGE_LIGHT;
647 }
648
649 /* Caller must hold page table lock. */
650 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
651                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
652                 struct page *zero_page)
653 {
654         pmd_t entry;
655         if (!pmd_none(*pmd))
656                 return false;
657         entry = mk_pmd(zero_page, vma->vm_page_prot);
658         entry = pmd_mkhuge(entry);
659         if (pgtable)
660                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
661         set_pmd_at(mm, haddr, pmd, entry);
662         mm_inc_nr_ptes(mm);
663         return true;
664 }
665
666 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
667 {
668         struct vm_area_struct *vma = vmf->vma;
669         gfp_t gfp;
670         struct page *page;
671         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
672
673         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
674                 return VM_FAULT_FALLBACK;
675         if (unlikely(anon_vma_prepare(vma)))
676                 return VM_FAULT_OOM;
677         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
678                 return VM_FAULT_OOM;
679         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
680                         !mm_forbids_zeropage(vma->vm_mm) &&
681                         transparent_hugepage_use_zero_page()) {
682                 pgtable_t pgtable;
683                 struct page *zero_page;
684                 bool set;
685                 vm_fault_t ret;
686                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
687                 if (unlikely(!pgtable))
688                         return VM_FAULT_OOM;
689                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
690                 if (unlikely(!zero_page)) {
691                         pte_free(vma->vm_mm, pgtable);
692                         count_vm_event(THP_FAULT_FALLBACK);
693                         return VM_FAULT_FALLBACK;
694                 }
695                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
696                 ret = 0;
697                 set = false;
698                 if (pmd_none(*vmf->pmd)) {
699                         ret = check_stable_address_space(vma->vm_mm);
700                         if (ret) {
701                                 spin_unlock(vmf->ptl);
702                         } else if (userfaultfd_missing(vma)) {
703                                 spin_unlock(vmf->ptl);
704                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
705                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
706                         } else {
707                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
708                                                    haddr, vmf->pmd, zero_page);
709                                 spin_unlock(vmf->ptl);
710                                 set = true;
711                         }
712                 } else
713                         spin_unlock(vmf->ptl);
714                 if (!set)
715                         pte_free(vma->vm_mm, pgtable);
716                 return ret;
717         }
718         gfp = alloc_hugepage_direct_gfpmask(vma);
719         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
720         if (unlikely(!page)) {
721                 count_vm_event(THP_FAULT_FALLBACK);
722                 return VM_FAULT_FALLBACK;
723         }
724         prep_transhuge_page(page);
725         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
726 }
727
728 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
729                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
730                 pgtable_t pgtable)
731 {
732         struct mm_struct *mm = vma->vm_mm;
733         pmd_t entry;
734         spinlock_t *ptl;
735
736         ptl = pmd_lock(mm, pmd);
737         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
738         if (pfn_t_devmap(pfn))
739                 entry = pmd_mkdevmap(entry);
740         if (write) {
741                 entry = pmd_mkyoung(pmd_mkdirty(entry));
742                 entry = maybe_pmd_mkwrite(entry, vma);
743         }
744
745         if (pgtable) {
746                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
747                 mm_inc_nr_ptes(mm);
748         }
749
750         set_pmd_at(mm, addr, pmd, entry);
751         update_mmu_cache_pmd(vma, addr, pmd);
752         spin_unlock(ptl);
753 }
754
755 vm_fault_t vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
756                         pmd_t *pmd, pfn_t pfn, bool write)
757 {
758         pgprot_t pgprot = vma->vm_page_prot;
759         pgtable_t pgtable = NULL;
760         /*
761          * If we had pmd_special, we could avoid all these restrictions,
762          * but we need to be consistent with PTEs and architectures that
763          * can't support a 'special' bit.
764          */
765         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
766                         !pfn_t_devmap(pfn));
767         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
768                                                 (VM_PFNMAP|VM_MIXEDMAP));
769         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
770
771         if (addr < vma->vm_start || addr >= vma->vm_end)
772                 return VM_FAULT_SIGBUS;
773
774         if (arch_needs_pgtable_deposit()) {
775                 pgtable = pte_alloc_one(vma->vm_mm, addr);
776                 if (!pgtable)
777                         return VM_FAULT_OOM;
778         }
779
780         track_pfn_insert(vma, &pgprot, pfn);
781
782         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
783         return VM_FAULT_NOPAGE;
784 }
785 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
786
787 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
788 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
789 {
790         if (likely(vma->vm_flags & VM_WRITE))
791                 pud = pud_mkwrite(pud);
792         return pud;
793 }
794
795 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
796                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
797 {
798         struct mm_struct *mm = vma->vm_mm;
799         pud_t entry;
800         spinlock_t *ptl;
801
802         ptl = pud_lock(mm, pud);
803         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
804         if (pfn_t_devmap(pfn))
805                 entry = pud_mkdevmap(entry);
806         if (write) {
807                 entry = pud_mkyoung(pud_mkdirty(entry));
808                 entry = maybe_pud_mkwrite(entry, vma);
809         }
810         set_pud_at(mm, addr, pud, entry);
811         update_mmu_cache_pud(vma, addr, pud);
812         spin_unlock(ptl);
813 }
814
815 vm_fault_t vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
816                         pud_t *pud, pfn_t pfn, bool write)
817 {
818         pgprot_t pgprot = vma->vm_page_prot;
819         /*
820          * If we had pud_special, we could avoid all these restrictions,
821          * but we need to be consistent with PTEs and architectures that
822          * can't support a 'special' bit.
823          */
824         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
825                         !pfn_t_devmap(pfn));
826         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
827                                                 (VM_PFNMAP|VM_MIXEDMAP));
828         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
829
830         if (addr < vma->vm_start || addr >= vma->vm_end)
831                 return VM_FAULT_SIGBUS;
832
833         track_pfn_insert(vma, &pgprot, pfn);
834
835         insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
836         return VM_FAULT_NOPAGE;
837 }
838 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
839 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
840
841 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
842                 pmd_t *pmd, int flags)
843 {
844         pmd_t _pmd;
845
846         _pmd = pmd_mkyoung(*pmd);
847         if (flags & FOLL_WRITE)
848                 _pmd = pmd_mkdirty(_pmd);
849         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
850                                 pmd, _pmd, flags & FOLL_WRITE))
851                 update_mmu_cache_pmd(vma, addr, pmd);
852 }
853
854 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
855                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
856 {
857         unsigned long pfn = pmd_pfn(*pmd);
858         struct mm_struct *mm = vma->vm_mm;
859         struct page *page;
860
861         assert_spin_locked(pmd_lockptr(mm, pmd));
862
863         /*
864          * When we COW a devmap PMD entry, we split it into PTEs, so we should
865          * not be in this function with `flags & FOLL_COW` set.
866          */
867         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
868
869         if (flags & FOLL_WRITE && !pmd_write(*pmd))
870                 return NULL;
871
872         if (pmd_present(*pmd) && pmd_devmap(*pmd))
873                 /* pass */;
874         else
875                 return NULL;
876
877         if (flags & FOLL_TOUCH)
878                 touch_pmd(vma, addr, pmd, flags);
879
880         /*
881          * device mapped pages can only be returned if the
882          * caller will manage the page reference count.
883          */
884         if (!(flags & FOLL_GET))
885                 return ERR_PTR(-EEXIST);
886
887         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
888         *pgmap = get_dev_pagemap(pfn, *pgmap);
889         if (!*pgmap)
890                 return ERR_PTR(-EFAULT);
891         page = pfn_to_page(pfn);
892         get_page(page);
893
894         return page;
895 }
896
897 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
898                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
899                   struct vm_area_struct *vma)
900 {
901         spinlock_t *dst_ptl, *src_ptl;
902         struct page *src_page;
903         pmd_t pmd;
904         pgtable_t pgtable = NULL;
905         int ret = -ENOMEM;
906
907         /* Skip if can be re-fill on fault */
908         if (!vma_is_anonymous(vma))
909                 return 0;
910
911         pgtable = pte_alloc_one(dst_mm, addr);
912         if (unlikely(!pgtable))
913                 goto out;
914
915         dst_ptl = pmd_lock(dst_mm, dst_pmd);
916         src_ptl = pmd_lockptr(src_mm, src_pmd);
917         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
918
919         ret = -EAGAIN;
920         pmd = *src_pmd;
921
922 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
923         if (unlikely(is_swap_pmd(pmd))) {
924                 swp_entry_t entry = pmd_to_swp_entry(pmd);
925
926                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
927                 if (is_write_migration_entry(entry)) {
928                         make_migration_entry_read(&entry);
929                         pmd = swp_entry_to_pmd(entry);
930                         if (pmd_swp_soft_dirty(*src_pmd))
931                                 pmd = pmd_swp_mksoft_dirty(pmd);
932                         set_pmd_at(src_mm, addr, src_pmd, pmd);
933                 }
934                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
935                 mm_inc_nr_ptes(dst_mm);
936                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
937                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
938                 ret = 0;
939                 goto out_unlock;
940         }
941 #endif
942
943         if (unlikely(!pmd_trans_huge(pmd))) {
944                 pte_free(dst_mm, pgtable);
945                 goto out_unlock;
946         }
947         /*
948          * When page table lock is held, the huge zero pmd should not be
949          * under splitting since we don't split the page itself, only pmd to
950          * a page table.
951          */
952         if (is_huge_zero_pmd(pmd)) {
953                 struct page *zero_page;
954                 /*
955                  * get_huge_zero_page() will never allocate a new page here,
956                  * since we already have a zero page to copy. It just takes a
957                  * reference.
958                  */
959                 zero_page = mm_get_huge_zero_page(dst_mm);
960                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
961                                 zero_page);
962                 ret = 0;
963                 goto out_unlock;
964         }
965
966         src_page = pmd_page(pmd);
967         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
968         get_page(src_page);
969         page_dup_rmap(src_page, true);
970         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
971         mm_inc_nr_ptes(dst_mm);
972         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
973
974         pmdp_set_wrprotect(src_mm, addr, src_pmd);
975         pmd = pmd_mkold(pmd_wrprotect(pmd));
976         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
977
978         ret = 0;
979 out_unlock:
980         spin_unlock(src_ptl);
981         spin_unlock(dst_ptl);
982 out:
983         return ret;
984 }
985
986 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
987 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
988                 pud_t *pud, int flags)
989 {
990         pud_t _pud;
991
992         _pud = pud_mkyoung(*pud);
993         if (flags & FOLL_WRITE)
994                 _pud = pud_mkdirty(_pud);
995         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
996                                 pud, _pud, flags & FOLL_WRITE))
997                 update_mmu_cache_pud(vma, addr, pud);
998 }
999
1000 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1001                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1002 {
1003         unsigned long pfn = pud_pfn(*pud);
1004         struct mm_struct *mm = vma->vm_mm;
1005         struct page *page;
1006
1007         assert_spin_locked(pud_lockptr(mm, pud));
1008
1009         if (flags & FOLL_WRITE && !pud_write(*pud))
1010                 return NULL;
1011
1012         if (pud_present(*pud) && pud_devmap(*pud))
1013                 /* pass */;
1014         else
1015                 return NULL;
1016
1017         if (flags & FOLL_TOUCH)
1018                 touch_pud(vma, addr, pud, flags);
1019
1020         /*
1021          * device mapped pages can only be returned if the
1022          * caller will manage the page reference count.
1023          */
1024         if (!(flags & FOLL_GET))
1025                 return ERR_PTR(-EEXIST);
1026
1027         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1028         *pgmap = get_dev_pagemap(pfn, *pgmap);
1029         if (!*pgmap)
1030                 return ERR_PTR(-EFAULT);
1031         page = pfn_to_page(pfn);
1032         get_page(page);
1033
1034         return page;
1035 }
1036
1037 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1038                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1039                   struct vm_area_struct *vma)
1040 {
1041         spinlock_t *dst_ptl, *src_ptl;
1042         pud_t pud;
1043         int ret;
1044
1045         dst_ptl = pud_lock(dst_mm, dst_pud);
1046         src_ptl = pud_lockptr(src_mm, src_pud);
1047         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1048
1049         ret = -EAGAIN;
1050         pud = *src_pud;
1051         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1052                 goto out_unlock;
1053
1054         /*
1055          * When page table lock is held, the huge zero pud should not be
1056          * under splitting since we don't split the page itself, only pud to
1057          * a page table.
1058          */
1059         if (is_huge_zero_pud(pud)) {
1060                 /* No huge zero pud yet */
1061         }
1062
1063         pudp_set_wrprotect(src_mm, addr, src_pud);
1064         pud = pud_mkold(pud_wrprotect(pud));
1065         set_pud_at(dst_mm, addr, dst_pud, pud);
1066
1067         ret = 0;
1068 out_unlock:
1069         spin_unlock(src_ptl);
1070         spin_unlock(dst_ptl);
1071         return ret;
1072 }
1073
1074 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1075 {
1076         pud_t entry;
1077         unsigned long haddr;
1078         bool write = vmf->flags & FAULT_FLAG_WRITE;
1079
1080         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1081         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1082                 goto unlock;
1083
1084         entry = pud_mkyoung(orig_pud);
1085         if (write)
1086                 entry = pud_mkdirty(entry);
1087         haddr = vmf->address & HPAGE_PUD_MASK;
1088         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1089                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1090
1091 unlock:
1092         spin_unlock(vmf->ptl);
1093 }
1094 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1095
1096 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1097 {
1098         pmd_t entry;
1099         unsigned long haddr;
1100         bool write = vmf->flags & FAULT_FLAG_WRITE;
1101
1102         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1103         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1104                 goto unlock;
1105
1106         entry = pmd_mkyoung(orig_pmd);
1107         if (write)
1108                 entry = pmd_mkdirty(entry);
1109         haddr = vmf->address & HPAGE_PMD_MASK;
1110         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1111                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1112
1113 unlock:
1114         spin_unlock(vmf->ptl);
1115 }
1116
1117 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1118                         pmd_t orig_pmd, struct page *page)
1119 {
1120         struct vm_area_struct *vma = vmf->vma;
1121         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1122         struct mem_cgroup *memcg;
1123         pgtable_t pgtable;
1124         pmd_t _pmd;
1125         int i;
1126         vm_fault_t ret = 0;
1127         struct page **pages;
1128         unsigned long mmun_start;       /* For mmu_notifiers */
1129         unsigned long mmun_end;         /* For mmu_notifiers */
1130
1131         pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1132                               GFP_KERNEL);
1133         if (unlikely(!pages)) {
1134                 ret |= VM_FAULT_OOM;
1135                 goto out;
1136         }
1137
1138         for (i = 0; i < HPAGE_PMD_NR; i++) {
1139                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1140                                                vmf->address, page_to_nid(page));
1141                 if (unlikely(!pages[i] ||
1142                              mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1143                                      GFP_KERNEL, &memcg, false))) {
1144                         if (pages[i])
1145                                 put_page(pages[i]);
1146                         while (--i >= 0) {
1147                                 memcg = (void *)page_private(pages[i]);
1148                                 set_page_private(pages[i], 0);
1149                                 mem_cgroup_cancel_charge(pages[i], memcg,
1150                                                 false);
1151                                 put_page(pages[i]);
1152                         }
1153                         kfree(pages);
1154                         ret |= VM_FAULT_OOM;
1155                         goto out;
1156                 }
1157                 set_page_private(pages[i], (unsigned long)memcg);
1158         }
1159
1160         for (i = 0; i < HPAGE_PMD_NR; i++) {
1161                 copy_user_highpage(pages[i], page + i,
1162                                    haddr + PAGE_SIZE * i, vma);
1163                 __SetPageUptodate(pages[i]);
1164                 cond_resched();
1165         }
1166
1167         mmun_start = haddr;
1168         mmun_end   = haddr + HPAGE_PMD_SIZE;
1169         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1170
1171         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1172         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1173                 goto out_free_pages;
1174         VM_BUG_ON_PAGE(!PageHead(page), page);
1175
1176         /*
1177          * Leave pmd empty until pte is filled note we must notify here as
1178          * concurrent CPU thread might write to new page before the call to
1179          * mmu_notifier_invalidate_range_end() happens which can lead to a
1180          * device seeing memory write in different order than CPU.
1181          *
1182          * See Documentation/vm/mmu_notifier.rst
1183          */
1184         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1185
1186         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1187         pmd_populate(vma->vm_mm, &_pmd, pgtable);
1188
1189         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1190                 pte_t entry;
1191                 entry = mk_pte(pages[i], vma->vm_page_prot);
1192                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1193                 memcg = (void *)page_private(pages[i]);
1194                 set_page_private(pages[i], 0);
1195                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1196                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1197                 lru_cache_add_active_or_unevictable(pages[i], vma);
1198                 vmf->pte = pte_offset_map(&_pmd, haddr);
1199                 VM_BUG_ON(!pte_none(*vmf->pte));
1200                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1201                 pte_unmap(vmf->pte);
1202         }
1203         kfree(pages);
1204
1205         smp_wmb(); /* make pte visible before pmd */
1206         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1207         page_remove_rmap(page, true);
1208         spin_unlock(vmf->ptl);
1209
1210         /*
1211          * No need to double call mmu_notifier->invalidate_range() callback as
1212          * the above pmdp_huge_clear_flush_notify() did already call it.
1213          */
1214         mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1215                                                 mmun_end);
1216
1217         ret |= VM_FAULT_WRITE;
1218         put_page(page);
1219
1220 out:
1221         return ret;
1222
1223 out_free_pages:
1224         spin_unlock(vmf->ptl);
1225         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1226         for (i = 0; i < HPAGE_PMD_NR; i++) {
1227                 memcg = (void *)page_private(pages[i]);
1228                 set_page_private(pages[i], 0);
1229                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1230                 put_page(pages[i]);
1231         }
1232         kfree(pages);
1233         goto out;
1234 }
1235
1236 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1237 {
1238         struct vm_area_struct *vma = vmf->vma;
1239         struct page *page = NULL, *new_page;
1240         struct mem_cgroup *memcg;
1241         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1242         unsigned long mmun_start;       /* For mmu_notifiers */
1243         unsigned long mmun_end;         /* For mmu_notifiers */
1244         gfp_t huge_gfp;                 /* for allocation and charge */
1245         vm_fault_t ret = 0;
1246
1247         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1248         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1249         if (is_huge_zero_pmd(orig_pmd))
1250                 goto alloc;
1251         spin_lock(vmf->ptl);
1252         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1253                 goto out_unlock;
1254
1255         page = pmd_page(orig_pmd);
1256         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1257         /*
1258          * We can only reuse the page if nobody else maps the huge page or it's
1259          * part.
1260          */
1261         if (!trylock_page(page)) {
1262                 get_page(page);
1263                 spin_unlock(vmf->ptl);
1264                 lock_page(page);
1265                 spin_lock(vmf->ptl);
1266                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1267                         unlock_page(page);
1268                         put_page(page);
1269                         goto out_unlock;
1270                 }
1271                 put_page(page);
1272         }
1273         if (reuse_swap_page(page, NULL)) {
1274                 pmd_t entry;
1275                 entry = pmd_mkyoung(orig_pmd);
1276                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1277                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1278                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1279                 ret |= VM_FAULT_WRITE;
1280                 unlock_page(page);
1281                 goto out_unlock;
1282         }
1283         unlock_page(page);
1284         get_page(page);
1285         spin_unlock(vmf->ptl);
1286 alloc:
1287         if (transparent_hugepage_enabled(vma) &&
1288             !transparent_hugepage_debug_cow()) {
1289                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1290                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1291         } else
1292                 new_page = NULL;
1293
1294         if (likely(new_page)) {
1295                 prep_transhuge_page(new_page);
1296         } else {
1297                 if (!page) {
1298                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1299                         ret |= VM_FAULT_FALLBACK;
1300                 } else {
1301                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1302                         if (ret & VM_FAULT_OOM) {
1303                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1304                                 ret |= VM_FAULT_FALLBACK;
1305                         }
1306                         put_page(page);
1307                 }
1308                 count_vm_event(THP_FAULT_FALLBACK);
1309                 goto out;
1310         }
1311
1312         if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1313                                         huge_gfp, &memcg, true))) {
1314                 put_page(new_page);
1315                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1316                 if (page)
1317                         put_page(page);
1318                 ret |= VM_FAULT_FALLBACK;
1319                 count_vm_event(THP_FAULT_FALLBACK);
1320                 goto out;
1321         }
1322
1323         count_vm_event(THP_FAULT_ALLOC);
1324
1325         if (!page)
1326                 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1327         else
1328                 copy_user_huge_page(new_page, page, vmf->address,
1329                                     vma, HPAGE_PMD_NR);
1330         __SetPageUptodate(new_page);
1331
1332         mmun_start = haddr;
1333         mmun_end   = haddr + HPAGE_PMD_SIZE;
1334         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1335
1336         spin_lock(vmf->ptl);
1337         if (page)
1338                 put_page(page);
1339         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1340                 spin_unlock(vmf->ptl);
1341                 mem_cgroup_cancel_charge(new_page, memcg, true);
1342                 put_page(new_page);
1343                 goto out_mn;
1344         } else {
1345                 pmd_t entry;
1346                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1347                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1348                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1349                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1350                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1351                 lru_cache_add_active_or_unevictable(new_page, vma);
1352                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1353                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1354                 if (!page) {
1355                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1356                 } else {
1357                         VM_BUG_ON_PAGE(!PageHead(page), page);
1358                         page_remove_rmap(page, true);
1359                         put_page(page);
1360                 }
1361                 ret |= VM_FAULT_WRITE;
1362         }
1363         spin_unlock(vmf->ptl);
1364 out_mn:
1365         /*
1366          * No need to double call mmu_notifier->invalidate_range() callback as
1367          * the above pmdp_huge_clear_flush_notify() did already call it.
1368          */
1369         mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1370                                                mmun_end);
1371 out:
1372         return ret;
1373 out_unlock:
1374         spin_unlock(vmf->ptl);
1375         return ret;
1376 }
1377
1378 /*
1379  * FOLL_FORCE can write to even unwritable pmd's, but only
1380  * after we've gone through a COW cycle and they are dirty.
1381  */
1382 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1383 {
1384         return pmd_write(pmd) ||
1385                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1386 }
1387
1388 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1389                                    unsigned long addr,
1390                                    pmd_t *pmd,
1391                                    unsigned int flags)
1392 {
1393         struct mm_struct *mm = vma->vm_mm;
1394         struct page *page = NULL;
1395
1396         assert_spin_locked(pmd_lockptr(mm, pmd));
1397
1398         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1399                 goto out;
1400
1401         /* Avoid dumping huge zero page */
1402         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1403                 return ERR_PTR(-EFAULT);
1404
1405         /* Full NUMA hinting faults to serialise migration in fault paths */
1406         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1407                 goto out;
1408
1409         page = pmd_page(*pmd);
1410         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1411         if (flags & FOLL_TOUCH)
1412                 touch_pmd(vma, addr, pmd, flags);
1413         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1414                 /*
1415                  * We don't mlock() pte-mapped THPs. This way we can avoid
1416                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1417                  *
1418                  * For anon THP:
1419                  *
1420                  * In most cases the pmd is the only mapping of the page as we
1421                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1422                  * writable private mappings in populate_vma_page_range().
1423                  *
1424                  * The only scenario when we have the page shared here is if we
1425                  * mlocking read-only mapping shared over fork(). We skip
1426                  * mlocking such pages.
1427                  *
1428                  * For file THP:
1429                  *
1430                  * We can expect PageDoubleMap() to be stable under page lock:
1431                  * for file pages we set it in page_add_file_rmap(), which
1432                  * requires page to be locked.
1433                  */
1434
1435                 if (PageAnon(page) && compound_mapcount(page) != 1)
1436                         goto skip_mlock;
1437                 if (PageDoubleMap(page) || !page->mapping)
1438                         goto skip_mlock;
1439                 if (!trylock_page(page))
1440                         goto skip_mlock;
1441                 lru_add_drain();
1442                 if (page->mapping && !PageDoubleMap(page))
1443                         mlock_vma_page(page);
1444                 unlock_page(page);
1445         }
1446 skip_mlock:
1447         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1448         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1449         if (flags & FOLL_GET)
1450                 get_page(page);
1451
1452 out:
1453         return page;
1454 }
1455
1456 /* NUMA hinting page fault entry point for trans huge pmds */
1457 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1458 {
1459         struct vm_area_struct *vma = vmf->vma;
1460         struct anon_vma *anon_vma = NULL;
1461         struct page *page;
1462         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1463         int page_nid = -1, this_nid = numa_node_id();
1464         int target_nid, last_cpupid = -1;
1465         bool page_locked;
1466         bool migrated = false;
1467         bool was_writable;
1468         int flags = 0;
1469
1470         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1471         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1472                 goto out_unlock;
1473
1474         /*
1475          * If there are potential migrations, wait for completion and retry
1476          * without disrupting NUMA hinting information. Do not relock and
1477          * check_same as the page may no longer be mapped.
1478          */
1479         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1480                 page = pmd_page(*vmf->pmd);
1481                 if (!get_page_unless_zero(page))
1482                         goto out_unlock;
1483                 spin_unlock(vmf->ptl);
1484                 wait_on_page_locked(page);
1485                 put_page(page);
1486                 goto out;
1487         }
1488
1489         page = pmd_page(pmd);
1490         BUG_ON(is_huge_zero_page(page));
1491         page_nid = page_to_nid(page);
1492         last_cpupid = page_cpupid_last(page);
1493         count_vm_numa_event(NUMA_HINT_FAULTS);
1494         if (page_nid == this_nid) {
1495                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1496                 flags |= TNF_FAULT_LOCAL;
1497         }
1498
1499         /* See similar comment in do_numa_page for explanation */
1500         if (!pmd_savedwrite(pmd))
1501                 flags |= TNF_NO_GROUP;
1502
1503         /*
1504          * Acquire the page lock to serialise THP migrations but avoid dropping
1505          * page_table_lock if at all possible
1506          */
1507         page_locked = trylock_page(page);
1508         target_nid = mpol_misplaced(page, vma, haddr);
1509         if (target_nid == -1) {
1510                 /* If the page was locked, there are no parallel migrations */
1511                 if (page_locked)
1512                         goto clear_pmdnuma;
1513         }
1514
1515         /* Migration could have started since the pmd_trans_migrating check */
1516         if (!page_locked) {
1517                 page_nid = -1;
1518                 if (!get_page_unless_zero(page))
1519                         goto out_unlock;
1520                 spin_unlock(vmf->ptl);
1521                 wait_on_page_locked(page);
1522                 put_page(page);
1523                 goto out;
1524         }
1525
1526         /*
1527          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1528          * to serialises splits
1529          */
1530         get_page(page);
1531         spin_unlock(vmf->ptl);
1532         anon_vma = page_lock_anon_vma_read(page);
1533
1534         /* Confirm the PMD did not change while page_table_lock was released */
1535         spin_lock(vmf->ptl);
1536         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1537                 unlock_page(page);
1538                 put_page(page);
1539                 page_nid = -1;
1540                 goto out_unlock;
1541         }
1542
1543         /* Bail if we fail to protect against THP splits for any reason */
1544         if (unlikely(!anon_vma)) {
1545                 put_page(page);
1546                 page_nid = -1;
1547                 goto clear_pmdnuma;
1548         }
1549
1550         /*
1551          * Since we took the NUMA fault, we must have observed the !accessible
1552          * bit. Make sure all other CPUs agree with that, to avoid them
1553          * modifying the page we're about to migrate.
1554          *
1555          * Must be done under PTL such that we'll observe the relevant
1556          * inc_tlb_flush_pending().
1557          *
1558          * We are not sure a pending tlb flush here is for a huge page
1559          * mapping or not. Hence use the tlb range variant
1560          */
1561         if (mm_tlb_flush_pending(vma->vm_mm))
1562                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1563
1564         /*
1565          * Migrate the THP to the requested node, returns with page unlocked
1566          * and access rights restored.
1567          */
1568         spin_unlock(vmf->ptl);
1569
1570         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1571                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1572         if (migrated) {
1573                 flags |= TNF_MIGRATED;
1574                 page_nid = target_nid;
1575         } else
1576                 flags |= TNF_MIGRATE_FAIL;
1577
1578         goto out;
1579 clear_pmdnuma:
1580         BUG_ON(!PageLocked(page));
1581         was_writable = pmd_savedwrite(pmd);
1582         pmd = pmd_modify(pmd, vma->vm_page_prot);
1583         pmd = pmd_mkyoung(pmd);
1584         if (was_writable)
1585                 pmd = pmd_mkwrite(pmd);
1586         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1587         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1588         unlock_page(page);
1589 out_unlock:
1590         spin_unlock(vmf->ptl);
1591
1592 out:
1593         if (anon_vma)
1594                 page_unlock_anon_vma_read(anon_vma);
1595
1596         if (page_nid != -1)
1597                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1598                                 flags);
1599
1600         return 0;
1601 }
1602
1603 /*
1604  * Return true if we do MADV_FREE successfully on entire pmd page.
1605  * Otherwise, return false.
1606  */
1607 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1608                 pmd_t *pmd, unsigned long addr, unsigned long next)
1609 {
1610         spinlock_t *ptl;
1611         pmd_t orig_pmd;
1612         struct page *page;
1613         struct mm_struct *mm = tlb->mm;
1614         bool ret = false;
1615
1616         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1617
1618         ptl = pmd_trans_huge_lock(pmd, vma);
1619         if (!ptl)
1620                 goto out_unlocked;
1621
1622         orig_pmd = *pmd;
1623         if (is_huge_zero_pmd(orig_pmd))
1624                 goto out;
1625
1626         if (unlikely(!pmd_present(orig_pmd))) {
1627                 VM_BUG_ON(thp_migration_supported() &&
1628                                   !is_pmd_migration_entry(orig_pmd));
1629                 goto out;
1630         }
1631
1632         page = pmd_page(orig_pmd);
1633         /*
1634          * If other processes are mapping this page, we couldn't discard
1635          * the page unless they all do MADV_FREE so let's skip the page.
1636          */
1637         if (page_mapcount(page) != 1)
1638                 goto out;
1639
1640         if (!trylock_page(page))
1641                 goto out;
1642
1643         /*
1644          * If user want to discard part-pages of THP, split it so MADV_FREE
1645          * will deactivate only them.
1646          */
1647         if (next - addr != HPAGE_PMD_SIZE) {
1648                 get_page(page);
1649                 spin_unlock(ptl);
1650                 split_huge_page(page);
1651                 unlock_page(page);
1652                 put_page(page);
1653                 goto out_unlocked;
1654         }
1655
1656         if (PageDirty(page))
1657                 ClearPageDirty(page);
1658         unlock_page(page);
1659
1660         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1661                 pmdp_invalidate(vma, addr, pmd);
1662                 orig_pmd = pmd_mkold(orig_pmd);
1663                 orig_pmd = pmd_mkclean(orig_pmd);
1664
1665                 set_pmd_at(mm, addr, pmd, orig_pmd);
1666                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1667         }
1668
1669         mark_page_lazyfree(page);
1670         ret = true;
1671 out:
1672         spin_unlock(ptl);
1673 out_unlocked:
1674         return ret;
1675 }
1676
1677 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1678 {
1679         pgtable_t pgtable;
1680
1681         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1682         pte_free(mm, pgtable);
1683         mm_dec_nr_ptes(mm);
1684 }
1685
1686 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1687                  pmd_t *pmd, unsigned long addr)
1688 {
1689         pmd_t orig_pmd;
1690         spinlock_t *ptl;
1691
1692         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1693
1694         ptl = __pmd_trans_huge_lock(pmd, vma);
1695         if (!ptl)
1696                 return 0;
1697         /*
1698          * For architectures like ppc64 we look at deposited pgtable
1699          * when calling pmdp_huge_get_and_clear. So do the
1700          * pgtable_trans_huge_withdraw after finishing pmdp related
1701          * operations.
1702          */
1703         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1704                         tlb->fullmm);
1705         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1706         if (vma_is_dax(vma)) {
1707                 if (arch_needs_pgtable_deposit())
1708                         zap_deposited_table(tlb->mm, pmd);
1709                 spin_unlock(ptl);
1710                 if (is_huge_zero_pmd(orig_pmd))
1711                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1712         } else if (is_huge_zero_pmd(orig_pmd)) {
1713                 zap_deposited_table(tlb->mm, pmd);
1714                 spin_unlock(ptl);
1715                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1716         } else {
1717                 struct page *page = NULL;
1718                 int flush_needed = 1;
1719
1720                 if (pmd_present(orig_pmd)) {
1721                         page = pmd_page(orig_pmd);
1722                         page_remove_rmap(page, true);
1723                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1724                         VM_BUG_ON_PAGE(!PageHead(page), page);
1725                 } else if (thp_migration_supported()) {
1726                         swp_entry_t entry;
1727
1728                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1729                         entry = pmd_to_swp_entry(orig_pmd);
1730                         page = pfn_to_page(swp_offset(entry));
1731                         flush_needed = 0;
1732                 } else
1733                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1734
1735                 if (PageAnon(page)) {
1736                         zap_deposited_table(tlb->mm, pmd);
1737                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1738                 } else {
1739                         if (arch_needs_pgtable_deposit())
1740                                 zap_deposited_table(tlb->mm, pmd);
1741                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1742                 }
1743
1744                 spin_unlock(ptl);
1745                 if (flush_needed)
1746                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1747         }
1748         return 1;
1749 }
1750
1751 #ifndef pmd_move_must_withdraw
1752 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1753                                          spinlock_t *old_pmd_ptl,
1754                                          struct vm_area_struct *vma)
1755 {
1756         /*
1757          * With split pmd lock we also need to move preallocated
1758          * PTE page table if new_pmd is on different PMD page table.
1759          *
1760          * We also don't deposit and withdraw tables for file pages.
1761          */
1762         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1763 }
1764 #endif
1765
1766 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1767 {
1768 #ifdef CONFIG_MEM_SOFT_DIRTY
1769         if (unlikely(is_pmd_migration_entry(pmd)))
1770                 pmd = pmd_swp_mksoft_dirty(pmd);
1771         else if (pmd_present(pmd))
1772                 pmd = pmd_mksoft_dirty(pmd);
1773 #endif
1774         return pmd;
1775 }
1776
1777 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1778                   unsigned long new_addr, unsigned long old_end,
1779                   pmd_t *old_pmd, pmd_t *new_pmd)
1780 {
1781         spinlock_t *old_ptl, *new_ptl;
1782         pmd_t pmd;
1783         struct mm_struct *mm = vma->vm_mm;
1784         bool force_flush = false;
1785
1786         if ((old_addr & ~HPAGE_PMD_MASK) ||
1787             (new_addr & ~HPAGE_PMD_MASK) ||
1788             old_end - old_addr < HPAGE_PMD_SIZE)
1789                 return false;
1790
1791         /*
1792          * The destination pmd shouldn't be established, free_pgtables()
1793          * should have release it.
1794          */
1795         if (WARN_ON(!pmd_none(*new_pmd))) {
1796                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1797                 return false;
1798         }
1799
1800         /*
1801          * We don't have to worry about the ordering of src and dst
1802          * ptlocks because exclusive mmap_sem prevents deadlock.
1803          */
1804         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1805         if (old_ptl) {
1806                 new_ptl = pmd_lockptr(mm, new_pmd);
1807                 if (new_ptl != old_ptl)
1808                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1809                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1810                 if (pmd_present(pmd))
1811                         force_flush = true;
1812                 VM_BUG_ON(!pmd_none(*new_pmd));
1813
1814                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1815                         pgtable_t pgtable;
1816                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1817                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1818                 }
1819                 pmd = move_soft_dirty_pmd(pmd);
1820                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1821                 if (force_flush)
1822                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1823                 if (new_ptl != old_ptl)
1824                         spin_unlock(new_ptl);
1825                 spin_unlock(old_ptl);
1826                 return true;
1827         }
1828         return false;
1829 }
1830
1831 /*
1832  * Returns
1833  *  - 0 if PMD could not be locked
1834  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1835  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1836  */
1837 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1838                 unsigned long addr, pgprot_t newprot, int prot_numa)
1839 {
1840         struct mm_struct *mm = vma->vm_mm;
1841         spinlock_t *ptl;
1842         pmd_t entry;
1843         bool preserve_write;
1844         int ret;
1845
1846         ptl = __pmd_trans_huge_lock(pmd, vma);
1847         if (!ptl)
1848                 return 0;
1849
1850         preserve_write = prot_numa && pmd_write(*pmd);
1851         ret = 1;
1852
1853 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1854         if (is_swap_pmd(*pmd)) {
1855                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1856
1857                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1858                 if (is_write_migration_entry(entry)) {
1859                         pmd_t newpmd;
1860                         /*
1861                          * A protection check is difficult so
1862                          * just be safe and disable write
1863                          */
1864                         make_migration_entry_read(&entry);
1865                         newpmd = swp_entry_to_pmd(entry);
1866                         if (pmd_swp_soft_dirty(*pmd))
1867                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1868                         set_pmd_at(mm, addr, pmd, newpmd);
1869                 }
1870                 goto unlock;
1871         }
1872 #endif
1873
1874         /*
1875          * Avoid trapping faults against the zero page. The read-only
1876          * data is likely to be read-cached on the local CPU and
1877          * local/remote hits to the zero page are not interesting.
1878          */
1879         if (prot_numa && is_huge_zero_pmd(*pmd))
1880                 goto unlock;
1881
1882         if (prot_numa && pmd_protnone(*pmd))
1883                 goto unlock;
1884
1885         /*
1886          * In case prot_numa, we are under down_read(mmap_sem). It's critical
1887          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1888          * which is also under down_read(mmap_sem):
1889          *
1890          *      CPU0:                           CPU1:
1891          *                              change_huge_pmd(prot_numa=1)
1892          *                               pmdp_huge_get_and_clear_notify()
1893          * madvise_dontneed()
1894          *  zap_pmd_range()
1895          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1896          *   // skip the pmd
1897          *                               set_pmd_at();
1898          *                               // pmd is re-established
1899          *
1900          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1901          * which may break userspace.
1902          *
1903          * pmdp_invalidate() is required to make sure we don't miss
1904          * dirty/young flags set by hardware.
1905          */
1906         entry = pmdp_invalidate(vma, addr, pmd);
1907
1908         entry = pmd_modify(entry, newprot);
1909         if (preserve_write)
1910                 entry = pmd_mk_savedwrite(entry);
1911         ret = HPAGE_PMD_NR;
1912         set_pmd_at(mm, addr, pmd, entry);
1913         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1914 unlock:
1915         spin_unlock(ptl);
1916         return ret;
1917 }
1918
1919 /*
1920  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1921  *
1922  * Note that if it returns page table lock pointer, this routine returns without
1923  * unlocking page table lock. So callers must unlock it.
1924  */
1925 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1926 {
1927         spinlock_t *ptl;
1928         ptl = pmd_lock(vma->vm_mm, pmd);
1929         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1930                         pmd_devmap(*pmd)))
1931                 return ptl;
1932         spin_unlock(ptl);
1933         return NULL;
1934 }
1935
1936 /*
1937  * Returns true if a given pud maps a thp, false otherwise.
1938  *
1939  * Note that if it returns true, this routine returns without unlocking page
1940  * table lock. So callers must unlock it.
1941  */
1942 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1943 {
1944         spinlock_t *ptl;
1945
1946         ptl = pud_lock(vma->vm_mm, pud);
1947         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1948                 return ptl;
1949         spin_unlock(ptl);
1950         return NULL;
1951 }
1952
1953 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1954 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1955                  pud_t *pud, unsigned long addr)
1956 {
1957         pud_t orig_pud;
1958         spinlock_t *ptl;
1959
1960         ptl = __pud_trans_huge_lock(pud, vma);
1961         if (!ptl)
1962                 return 0;
1963         /*
1964          * For architectures like ppc64 we look at deposited pgtable
1965          * when calling pudp_huge_get_and_clear. So do the
1966          * pgtable_trans_huge_withdraw after finishing pudp related
1967          * operations.
1968          */
1969         orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1970                         tlb->fullmm);
1971         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1972         if (vma_is_dax(vma)) {
1973                 spin_unlock(ptl);
1974                 /* No zero page support yet */
1975         } else {
1976                 /* No support for anonymous PUD pages yet */
1977                 BUG();
1978         }
1979         return 1;
1980 }
1981
1982 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1983                 unsigned long haddr)
1984 {
1985         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1986         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1987         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1988         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1989
1990         count_vm_event(THP_SPLIT_PUD);
1991
1992         pudp_huge_clear_flush_notify(vma, haddr, pud);
1993 }
1994
1995 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1996                 unsigned long address)
1997 {
1998         spinlock_t *ptl;
1999         struct mm_struct *mm = vma->vm_mm;
2000         unsigned long haddr = address & HPAGE_PUD_MASK;
2001
2002         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
2003         ptl = pud_lock(mm, pud);
2004         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2005                 goto out;
2006         __split_huge_pud_locked(vma, pud, haddr);
2007
2008 out:
2009         spin_unlock(ptl);
2010         /*
2011          * No need to double call mmu_notifier->invalidate_range() callback as
2012          * the above pudp_huge_clear_flush_notify() did already call it.
2013          */
2014         mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2015                                                HPAGE_PUD_SIZE);
2016 }
2017 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2018
2019 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2020                 unsigned long haddr, pmd_t *pmd)
2021 {
2022         struct mm_struct *mm = vma->vm_mm;
2023         pgtable_t pgtable;
2024         pmd_t _pmd;
2025         int i;
2026
2027         /*
2028          * Leave pmd empty until pte is filled note that it is fine to delay
2029          * notification until mmu_notifier_invalidate_range_end() as we are
2030          * replacing a zero pmd write protected page with a zero pte write
2031          * protected page.
2032          *
2033          * See Documentation/vm/mmu_notifier.rst
2034          */
2035         pmdp_huge_clear_flush(vma, haddr, pmd);
2036
2037         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2038         pmd_populate(mm, &_pmd, pgtable);
2039
2040         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2041                 pte_t *pte, entry;
2042                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2043                 entry = pte_mkspecial(entry);
2044                 pte = pte_offset_map(&_pmd, haddr);
2045                 VM_BUG_ON(!pte_none(*pte));
2046                 set_pte_at(mm, haddr, pte, entry);
2047                 pte_unmap(pte);
2048         }
2049         smp_wmb(); /* make pte visible before pmd */
2050         pmd_populate(mm, pmd, pgtable);
2051 }
2052
2053 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2054                 unsigned long haddr, bool freeze)
2055 {
2056         struct mm_struct *mm = vma->vm_mm;
2057         struct page *page;
2058         pgtable_t pgtable;
2059         pmd_t old_pmd, _pmd;
2060         bool young, write, soft_dirty, pmd_migration = false;
2061         unsigned long addr;
2062         int i;
2063
2064         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2065         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2066         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2067         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2068                                 && !pmd_devmap(*pmd));
2069
2070         count_vm_event(THP_SPLIT_PMD);
2071
2072         if (!vma_is_anonymous(vma)) {
2073                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2074                 /*
2075                  * We are going to unmap this huge page. So
2076                  * just go ahead and zap it
2077                  */
2078                 if (arch_needs_pgtable_deposit())
2079                         zap_deposited_table(mm, pmd);
2080                 if (vma_is_dax(vma))
2081                         return;
2082                 page = pmd_page(_pmd);
2083                 if (!PageDirty(page) && pmd_dirty(_pmd))
2084                         set_page_dirty(page);
2085                 if (!PageReferenced(page) && pmd_young(_pmd))
2086                         SetPageReferenced(page);
2087                 page_remove_rmap(page, true);
2088                 put_page(page);
2089                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2090                 return;
2091         } else if (is_huge_zero_pmd(*pmd)) {
2092                 /*
2093                  * FIXME: Do we want to invalidate secondary mmu by calling
2094                  * mmu_notifier_invalidate_range() see comments below inside
2095                  * __split_huge_pmd() ?
2096                  *
2097                  * We are going from a zero huge page write protected to zero
2098                  * small page also write protected so it does not seems useful
2099                  * to invalidate secondary mmu at this time.
2100                  */
2101                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2102         }
2103
2104         /*
2105          * Up to this point the pmd is present and huge and userland has the
2106          * whole access to the hugepage during the split (which happens in
2107          * place). If we overwrite the pmd with the not-huge version pointing
2108          * to the pte here (which of course we could if all CPUs were bug
2109          * free), userland could trigger a small page size TLB miss on the
2110          * small sized TLB while the hugepage TLB entry is still established in
2111          * the huge TLB. Some CPU doesn't like that.
2112          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2113          * 383 on page 93. Intel should be safe but is also warns that it's
2114          * only safe if the permission and cache attributes of the two entries
2115          * loaded in the two TLB is identical (which should be the case here).
2116          * But it is generally safer to never allow small and huge TLB entries
2117          * for the same virtual address to be loaded simultaneously. So instead
2118          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2119          * current pmd notpresent (atomically because here the pmd_trans_huge
2120          * must remain set at all times on the pmd until the split is complete
2121          * for this pmd), then we flush the SMP TLB and finally we write the
2122          * non-huge version of the pmd entry with pmd_populate.
2123          */
2124         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2125
2126 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2127         pmd_migration = is_pmd_migration_entry(old_pmd);
2128         if (pmd_migration) {
2129                 swp_entry_t entry;
2130
2131                 entry = pmd_to_swp_entry(old_pmd);
2132                 page = pfn_to_page(swp_offset(entry));
2133         } else
2134 #endif
2135                 page = pmd_page(old_pmd);
2136         VM_BUG_ON_PAGE(!page_count(page), page);
2137         page_ref_add(page, HPAGE_PMD_NR - 1);
2138         if (pmd_dirty(old_pmd))
2139                 SetPageDirty(page);
2140         write = pmd_write(old_pmd);
2141         young = pmd_young(old_pmd);
2142         soft_dirty = pmd_soft_dirty(old_pmd);
2143
2144         /*
2145          * Withdraw the table only after we mark the pmd entry invalid.
2146          * This's critical for some architectures (Power).
2147          */
2148         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2149         pmd_populate(mm, &_pmd, pgtable);
2150
2151         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2152                 pte_t entry, *pte;
2153                 /*
2154                  * Note that NUMA hinting access restrictions are not
2155                  * transferred to avoid any possibility of altering
2156                  * permissions across VMAs.
2157                  */
2158                 if (freeze || pmd_migration) {
2159                         swp_entry_t swp_entry;
2160                         swp_entry = make_migration_entry(page + i, write);
2161                         entry = swp_entry_to_pte(swp_entry);
2162                         if (soft_dirty)
2163                                 entry = pte_swp_mksoft_dirty(entry);
2164                 } else {
2165                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2166                         entry = maybe_mkwrite(entry, vma);
2167                         if (!write)
2168                                 entry = pte_wrprotect(entry);
2169                         if (!young)
2170                                 entry = pte_mkold(entry);
2171                         if (soft_dirty)
2172                                 entry = pte_mksoft_dirty(entry);
2173                 }
2174                 pte = pte_offset_map(&_pmd, addr);
2175                 BUG_ON(!pte_none(*pte));
2176                 set_pte_at(mm, addr, pte, entry);
2177                 atomic_inc(&page[i]._mapcount);
2178                 pte_unmap(pte);
2179         }
2180
2181         /*
2182          * Set PG_double_map before dropping compound_mapcount to avoid
2183          * false-negative page_mapped().
2184          */
2185         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2186                 for (i = 0; i < HPAGE_PMD_NR; i++)
2187                         atomic_inc(&page[i]._mapcount);
2188         }
2189
2190         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2191                 /* Last compound_mapcount is gone. */
2192                 __dec_node_page_state(page, NR_ANON_THPS);
2193                 if (TestClearPageDoubleMap(page)) {
2194                         /* No need in mapcount reference anymore */
2195                         for (i = 0; i < HPAGE_PMD_NR; i++)
2196                                 atomic_dec(&page[i]._mapcount);
2197                 }
2198         }
2199
2200         smp_wmb(); /* make pte visible before pmd */
2201         pmd_populate(mm, pmd, pgtable);
2202
2203         if (freeze) {
2204                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2205                         page_remove_rmap(page + i, false);
2206                         put_page(page + i);
2207                 }
2208         }
2209 }
2210
2211 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2212                 unsigned long address, bool freeze, struct page *page)
2213 {
2214         spinlock_t *ptl;
2215         struct mm_struct *mm = vma->vm_mm;
2216         unsigned long haddr = address & HPAGE_PMD_MASK;
2217
2218         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2219         ptl = pmd_lock(mm, pmd);
2220
2221         /*
2222          * If caller asks to setup a migration entries, we need a page to check
2223          * pmd against. Otherwise we can end up replacing wrong page.
2224          */
2225         VM_BUG_ON(freeze && !page);
2226         if (page && page != pmd_page(*pmd))
2227                 goto out;
2228
2229         if (pmd_trans_huge(*pmd)) {
2230                 page = pmd_page(*pmd);
2231                 if (PageMlocked(page))
2232                         clear_page_mlock(page);
2233         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2234                 goto out;
2235         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
2236 out:
2237         spin_unlock(ptl);
2238         /*
2239          * No need to double call mmu_notifier->invalidate_range() callback.
2240          * They are 3 cases to consider inside __split_huge_pmd_locked():
2241          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2242          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2243          *    fault will trigger a flush_notify before pointing to a new page
2244          *    (it is fine if the secondary mmu keeps pointing to the old zero
2245          *    page in the meantime)
2246          *  3) Split a huge pmd into pte pointing to the same page. No need
2247          *     to invalidate secondary tlb entry they are all still valid.
2248          *     any further changes to individual pte will notify. So no need
2249          *     to call mmu_notifier->invalidate_range()
2250          */
2251         mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2252                                                HPAGE_PMD_SIZE);
2253 }
2254
2255 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2256                 bool freeze, struct page *page)
2257 {
2258         pgd_t *pgd;
2259         p4d_t *p4d;
2260         pud_t *pud;
2261         pmd_t *pmd;
2262
2263         pgd = pgd_offset(vma->vm_mm, address);
2264         if (!pgd_present(*pgd))
2265                 return;
2266
2267         p4d = p4d_offset(pgd, address);
2268         if (!p4d_present(*p4d))
2269                 return;
2270
2271         pud = pud_offset(p4d, address);
2272         if (!pud_present(*pud))
2273                 return;
2274
2275         pmd = pmd_offset(pud, address);
2276
2277         __split_huge_pmd(vma, pmd, address, freeze, page);
2278 }
2279
2280 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2281                              unsigned long start,
2282                              unsigned long end,
2283                              long adjust_next)
2284 {
2285         /*
2286          * If the new start address isn't hpage aligned and it could
2287          * previously contain an hugepage: check if we need to split
2288          * an huge pmd.
2289          */
2290         if (start & ~HPAGE_PMD_MASK &&
2291             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2292             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2293                 split_huge_pmd_address(vma, start, false, NULL);
2294
2295         /*
2296          * If the new end address isn't hpage aligned and it could
2297          * previously contain an hugepage: check if we need to split
2298          * an huge pmd.
2299          */
2300         if (end & ~HPAGE_PMD_MASK &&
2301             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2302             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2303                 split_huge_pmd_address(vma, end, false, NULL);
2304
2305         /*
2306          * If we're also updating the vma->vm_next->vm_start, if the new
2307          * vm_next->vm_start isn't page aligned and it could previously
2308          * contain an hugepage: check if we need to split an huge pmd.
2309          */
2310         if (adjust_next > 0) {
2311                 struct vm_area_struct *next = vma->vm_next;
2312                 unsigned long nstart = next->vm_start;
2313                 nstart += adjust_next << PAGE_SHIFT;
2314                 if (nstart & ~HPAGE_PMD_MASK &&
2315                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2316                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2317                         split_huge_pmd_address(next, nstart, false, NULL);
2318         }
2319 }
2320
2321 static void freeze_page(struct page *page)
2322 {
2323         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2324                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2325         bool unmap_success;
2326
2327         VM_BUG_ON_PAGE(!PageHead(page), page);
2328
2329         if (PageAnon(page))
2330                 ttu_flags |= TTU_SPLIT_FREEZE;
2331
2332         unmap_success = try_to_unmap(page, ttu_flags);
2333         VM_BUG_ON_PAGE(!unmap_success, page);
2334 }
2335
2336 static void unfreeze_page(struct page *page)
2337 {
2338         int i;
2339         if (PageTransHuge(page)) {
2340                 remove_migration_ptes(page, page, true);
2341         } else {
2342                 for (i = 0; i < HPAGE_PMD_NR; i++)
2343                         remove_migration_ptes(page + i, page + i, true);
2344         }
2345 }
2346
2347 static void __split_huge_page_tail(struct page *head, int tail,
2348                 struct lruvec *lruvec, struct list_head *list)
2349 {
2350         struct page *page_tail = head + tail;
2351
2352         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2353
2354         /*
2355          * Clone page flags before unfreezing refcount.
2356          *
2357          * After successful get_page_unless_zero() might follow flags change,
2358          * for exmaple lock_page() which set PG_waiters.
2359          */
2360         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2361         page_tail->flags |= (head->flags &
2362                         ((1L << PG_referenced) |
2363                          (1L << PG_swapbacked) |
2364                          (1L << PG_swapcache) |
2365                          (1L << PG_mlocked) |
2366                          (1L << PG_uptodate) |
2367                          (1L << PG_active) |
2368                          (1L << PG_workingset) |
2369                          (1L << PG_locked) |
2370                          (1L << PG_unevictable) |
2371                          (1L << PG_dirty)));
2372
2373         /* Page flags must be visible before we make the page non-compound. */
2374         smp_wmb();
2375
2376         /*
2377          * Clear PageTail before unfreezing page refcount.
2378          *
2379          * After successful get_page_unless_zero() might follow put_page()
2380          * which needs correct compound_head().
2381          */
2382         clear_compound_head(page_tail);
2383
2384         /* Finally unfreeze refcount. Additional reference from page cache. */
2385         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2386                                           PageSwapCache(head)));
2387
2388         if (page_is_young(head))
2389                 set_page_young(page_tail);
2390         if (page_is_idle(head))
2391                 set_page_idle(page_tail);
2392
2393         /* ->mapping in first tail page is compound_mapcount */
2394         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2395                         page_tail);
2396         page_tail->mapping = head->mapping;
2397
2398         page_tail->index = head->index + tail;
2399         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2400
2401         /*
2402          * always add to the tail because some iterators expect new
2403          * pages to show after the currently processed elements - e.g.
2404          * migrate_pages
2405          */
2406         lru_add_page_tail(head, page_tail, lruvec, list);
2407 }
2408
2409 static void __split_huge_page(struct page *page, struct list_head *list,
2410                 unsigned long flags)
2411 {
2412         struct page *head = compound_head(page);
2413         struct zone *zone = page_zone(head);
2414         struct lruvec *lruvec;
2415         pgoff_t end = -1;
2416         int i;
2417
2418         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
2419
2420         /* complete memcg works before add pages to LRU */
2421         mem_cgroup_split_huge_fixup(head);
2422
2423         if (!PageAnon(page))
2424                 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2425
2426         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2427                 __split_huge_page_tail(head, i, lruvec, list);
2428                 /* Some pages can be beyond i_size: drop them from page cache */
2429                 if (head[i].index >= end) {
2430                         ClearPageDirty(head + i);
2431                         __delete_from_page_cache(head + i, NULL);
2432                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2433                                 shmem_uncharge(head->mapping->host, 1);
2434                         put_page(head + i);
2435                 }
2436         }
2437
2438         ClearPageCompound(head);
2439         /* See comment in __split_huge_page_tail() */
2440         if (PageAnon(head)) {
2441                 /* Additional pin to radix tree of swap cache */
2442                 if (PageSwapCache(head))
2443                         page_ref_add(head, 2);
2444                 else
2445                         page_ref_inc(head);
2446         } else {
2447                 /* Additional pin to radix tree */
2448                 page_ref_add(head, 2);
2449                 xa_unlock(&head->mapping->i_pages);
2450         }
2451
2452         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2453
2454         unfreeze_page(head);
2455
2456         for (i = 0; i < HPAGE_PMD_NR; i++) {
2457                 struct page *subpage = head + i;
2458                 if (subpage == page)
2459                         continue;
2460                 unlock_page(subpage);
2461
2462                 /*
2463                  * Subpages may be freed if there wasn't any mapping
2464                  * like if add_to_swap() is running on a lru page that
2465                  * had its mapping zapped. And freeing these pages
2466                  * requires taking the lru_lock so we do the put_page
2467                  * of the tail pages after the split is complete.
2468                  */
2469                 put_page(subpage);
2470         }
2471 }
2472
2473 int total_mapcount(struct page *page)
2474 {
2475         int i, compound, ret;
2476
2477         VM_BUG_ON_PAGE(PageTail(page), page);
2478
2479         if (likely(!PageCompound(page)))
2480                 return atomic_read(&page->_mapcount) + 1;
2481
2482         compound = compound_mapcount(page);
2483         if (PageHuge(page))
2484                 return compound;
2485         ret = compound;
2486         for (i = 0; i < HPAGE_PMD_NR; i++)
2487                 ret += atomic_read(&page[i]._mapcount) + 1;
2488         /* File pages has compound_mapcount included in _mapcount */
2489         if (!PageAnon(page))
2490                 return ret - compound * HPAGE_PMD_NR;
2491         if (PageDoubleMap(page))
2492                 ret -= HPAGE_PMD_NR;
2493         return ret;
2494 }
2495
2496 /*
2497  * This calculates accurately how many mappings a transparent hugepage
2498  * has (unlike page_mapcount() which isn't fully accurate). This full
2499  * accuracy is primarily needed to know if copy-on-write faults can
2500  * reuse the page and change the mapping to read-write instead of
2501  * copying them. At the same time this returns the total_mapcount too.
2502  *
2503  * The function returns the highest mapcount any one of the subpages
2504  * has. If the return value is one, even if different processes are
2505  * mapping different subpages of the transparent hugepage, they can
2506  * all reuse it, because each process is reusing a different subpage.
2507  *
2508  * The total_mapcount is instead counting all virtual mappings of the
2509  * subpages. If the total_mapcount is equal to "one", it tells the
2510  * caller all mappings belong to the same "mm" and in turn the
2511  * anon_vma of the transparent hugepage can become the vma->anon_vma
2512  * local one as no other process may be mapping any of the subpages.
2513  *
2514  * It would be more accurate to replace page_mapcount() with
2515  * page_trans_huge_mapcount(), however we only use
2516  * page_trans_huge_mapcount() in the copy-on-write faults where we
2517  * need full accuracy to avoid breaking page pinning, because
2518  * page_trans_huge_mapcount() is slower than page_mapcount().
2519  */
2520 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2521 {
2522         int i, ret, _total_mapcount, mapcount;
2523
2524         /* hugetlbfs shouldn't call it */
2525         VM_BUG_ON_PAGE(PageHuge(page), page);
2526
2527         if (likely(!PageTransCompound(page))) {
2528                 mapcount = atomic_read(&page->_mapcount) + 1;
2529                 if (total_mapcount)
2530                         *total_mapcount = mapcount;
2531                 return mapcount;
2532         }
2533
2534         page = compound_head(page);
2535
2536         _total_mapcount = ret = 0;
2537         for (i = 0; i < HPAGE_PMD_NR; i++) {
2538                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2539                 ret = max(ret, mapcount);
2540                 _total_mapcount += mapcount;
2541         }
2542         if (PageDoubleMap(page)) {
2543                 ret -= 1;
2544                 _total_mapcount -= HPAGE_PMD_NR;
2545         }
2546         mapcount = compound_mapcount(page);
2547         ret += mapcount;
2548         _total_mapcount += mapcount;
2549         if (total_mapcount)
2550                 *total_mapcount = _total_mapcount;
2551         return ret;
2552 }
2553
2554 /* Racy check whether the huge page can be split */
2555 bool can_split_huge_page(struct page *page, int *pextra_pins)
2556 {
2557         int extra_pins;
2558
2559         /* Additional pins from radix tree */
2560         if (PageAnon(page))
2561                 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2562         else
2563                 extra_pins = HPAGE_PMD_NR;
2564         if (pextra_pins)
2565                 *pextra_pins = extra_pins;
2566         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2567 }
2568
2569 /*
2570  * This function splits huge page into normal pages. @page can point to any
2571  * subpage of huge page to split. Split doesn't change the position of @page.
2572  *
2573  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2574  * The huge page must be locked.
2575  *
2576  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2577  *
2578  * Both head page and tail pages will inherit mapping, flags, and so on from
2579  * the hugepage.
2580  *
2581  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2582  * they are not mapped.
2583  *
2584  * Returns 0 if the hugepage is split successfully.
2585  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2586  * us.
2587  */
2588 int split_huge_page_to_list(struct page *page, struct list_head *list)
2589 {
2590         struct page *head = compound_head(page);
2591         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2592         struct anon_vma *anon_vma = NULL;
2593         struct address_space *mapping = NULL;
2594         int count, mapcount, extra_pins, ret;
2595         bool mlocked;
2596         unsigned long flags;
2597
2598         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2599         VM_BUG_ON_PAGE(!PageLocked(page), page);
2600         VM_BUG_ON_PAGE(!PageCompound(page), page);
2601
2602         if (PageWriteback(page))
2603                 return -EBUSY;
2604
2605         if (PageAnon(head)) {
2606                 /*
2607                  * The caller does not necessarily hold an mmap_sem that would
2608                  * prevent the anon_vma disappearing so we first we take a
2609                  * reference to it and then lock the anon_vma for write. This
2610                  * is similar to page_lock_anon_vma_read except the write lock
2611                  * is taken to serialise against parallel split or collapse
2612                  * operations.
2613                  */
2614                 anon_vma = page_get_anon_vma(head);
2615                 if (!anon_vma) {
2616                         ret = -EBUSY;
2617                         goto out;
2618                 }
2619                 mapping = NULL;
2620                 anon_vma_lock_write(anon_vma);
2621         } else {
2622                 mapping = head->mapping;
2623
2624                 /* Truncated ? */
2625                 if (!mapping) {
2626                         ret = -EBUSY;
2627                         goto out;
2628                 }
2629
2630                 anon_vma = NULL;
2631                 i_mmap_lock_read(mapping);
2632         }
2633
2634         /*
2635          * Racy check if we can split the page, before freeze_page() will
2636          * split PMDs
2637          */
2638         if (!can_split_huge_page(head, &extra_pins)) {
2639                 ret = -EBUSY;
2640                 goto out_unlock;
2641         }
2642
2643         mlocked = PageMlocked(page);
2644         freeze_page(head);
2645         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2646
2647         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2648         if (mlocked)
2649                 lru_add_drain();
2650
2651         /* prevent PageLRU to go away from under us, and freeze lru stats */
2652         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2653
2654         if (mapping) {
2655                 void **pslot;
2656
2657                 xa_lock(&mapping->i_pages);
2658                 pslot = radix_tree_lookup_slot(&mapping->i_pages,
2659                                 page_index(head));
2660                 /*
2661                  * Check if the head page is present in radix tree.
2662                  * We assume all tail are present too, if head is there.
2663                  */
2664                 if (radix_tree_deref_slot_protected(pslot,
2665                                         &mapping->i_pages.xa_lock) != head)
2666                         goto fail;
2667         }
2668
2669         /* Prevent deferred_split_scan() touching ->_refcount */
2670         spin_lock(&pgdata->split_queue_lock);
2671         count = page_count(head);
2672         mapcount = total_mapcount(head);
2673         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2674                 if (!list_empty(page_deferred_list(head))) {
2675                         pgdata->split_queue_len--;
2676                         list_del(page_deferred_list(head));
2677                 }
2678                 if (mapping)
2679                         __dec_node_page_state(page, NR_SHMEM_THPS);
2680                 spin_unlock(&pgdata->split_queue_lock);
2681                 __split_huge_page(page, list, flags);
2682                 if (PageSwapCache(head)) {
2683                         swp_entry_t entry = { .val = page_private(head) };
2684
2685                         ret = split_swap_cluster(entry);
2686                 } else
2687                         ret = 0;
2688         } else {
2689                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2690                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2691                                         mapcount, count);
2692                         if (PageTail(page))
2693                                 dump_page(head, NULL);
2694                         dump_page(page, "total_mapcount(head) > 0");
2695                         BUG();
2696                 }
2697                 spin_unlock(&pgdata->split_queue_lock);
2698 fail:           if (mapping)
2699                         xa_unlock(&mapping->i_pages);
2700                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2701                 unfreeze_page(head);
2702                 ret = -EBUSY;
2703         }
2704
2705 out_unlock:
2706         if (anon_vma) {
2707                 anon_vma_unlock_write(anon_vma);
2708                 put_anon_vma(anon_vma);
2709         }
2710         if (mapping)
2711                 i_mmap_unlock_read(mapping);
2712 out:
2713         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2714         return ret;
2715 }
2716
2717 void free_transhuge_page(struct page *page)
2718 {
2719         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2720         unsigned long flags;
2721
2722         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2723         if (!list_empty(page_deferred_list(page))) {
2724                 pgdata->split_queue_len--;
2725                 list_del(page_deferred_list(page));
2726         }
2727         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2728         free_compound_page(page);
2729 }
2730
2731 void deferred_split_huge_page(struct page *page)
2732 {
2733         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2734         unsigned long flags;
2735
2736         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2737
2738         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2739         if (list_empty(page_deferred_list(page))) {
2740                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2741                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2742                 pgdata->split_queue_len++;
2743         }
2744         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2745 }
2746
2747 static unsigned long deferred_split_count(struct shrinker *shrink,
2748                 struct shrink_control *sc)
2749 {
2750         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2751         return READ_ONCE(pgdata->split_queue_len);
2752 }
2753
2754 static unsigned long deferred_split_scan(struct shrinker *shrink,
2755                 struct shrink_control *sc)
2756 {
2757         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2758         unsigned long flags;
2759         LIST_HEAD(list), *pos, *next;
2760         struct page *page;
2761         int split = 0;
2762
2763         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2764         /* Take pin on all head pages to avoid freeing them under us */
2765         list_for_each_safe(pos, next, &pgdata->split_queue) {
2766                 page = list_entry((void *)pos, struct page, mapping);
2767                 page = compound_head(page);
2768                 if (get_page_unless_zero(page)) {
2769                         list_move(page_deferred_list(page), &list);
2770                 } else {
2771                         /* We lost race with put_compound_page() */
2772                         list_del_init(page_deferred_list(page));
2773                         pgdata->split_queue_len--;
2774                 }
2775                 if (!--sc->nr_to_scan)
2776                         break;
2777         }
2778         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2779
2780         list_for_each_safe(pos, next, &list) {
2781                 page = list_entry((void *)pos, struct page, mapping);
2782                 if (!trylock_page(page))
2783                         goto next;
2784                 /* split_huge_page() removes page from list on success */
2785                 if (!split_huge_page(page))
2786                         split++;
2787                 unlock_page(page);
2788 next:
2789                 put_page(page);
2790         }
2791
2792         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2793         list_splice_tail(&list, &pgdata->split_queue);
2794         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2795
2796         /*
2797          * Stop shrinker if we didn't split any page, but the queue is empty.
2798          * This can happen if pages were freed under us.
2799          */
2800         if (!split && list_empty(&pgdata->split_queue))
2801                 return SHRINK_STOP;
2802         return split;
2803 }
2804
2805 static struct shrinker deferred_split_shrinker = {
2806         .count_objects = deferred_split_count,
2807         .scan_objects = deferred_split_scan,
2808         .seeks = DEFAULT_SEEKS,
2809         .flags = SHRINKER_NUMA_AWARE,
2810 };
2811
2812 #ifdef CONFIG_DEBUG_FS
2813 static int split_huge_pages_set(void *data, u64 val)
2814 {
2815         struct zone *zone;
2816         struct page *page;
2817         unsigned long pfn, max_zone_pfn;
2818         unsigned long total = 0, split = 0;
2819
2820         if (val != 1)
2821                 return -EINVAL;
2822
2823         for_each_populated_zone(zone) {
2824                 max_zone_pfn = zone_end_pfn(zone);
2825                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2826                         if (!pfn_valid(pfn))
2827                                 continue;
2828
2829                         page = pfn_to_page(pfn);
2830                         if (!get_page_unless_zero(page))
2831                                 continue;
2832
2833                         if (zone != page_zone(page))
2834                                 goto next;
2835
2836                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2837                                 goto next;
2838
2839                         total++;
2840                         lock_page(page);
2841                         if (!split_huge_page(page))
2842                                 split++;
2843                         unlock_page(page);
2844 next:
2845                         put_page(page);
2846                 }
2847         }
2848
2849         pr_info("%lu of %lu THP split\n", split, total);
2850
2851         return 0;
2852 }
2853 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2854                 "%llu\n");
2855
2856 static int __init split_huge_pages_debugfs(void)
2857 {
2858         void *ret;
2859
2860         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2861                         &split_huge_pages_fops);
2862         if (!ret)
2863                 pr_warn("Failed to create split_huge_pages in debugfs");
2864         return 0;
2865 }
2866 late_initcall(split_huge_pages_debugfs);
2867 #endif
2868
2869 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2870 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2871                 struct page *page)
2872 {
2873         struct vm_area_struct *vma = pvmw->vma;
2874         struct mm_struct *mm = vma->vm_mm;
2875         unsigned long address = pvmw->address;
2876         pmd_t pmdval;
2877         swp_entry_t entry;
2878         pmd_t pmdswp;
2879
2880         if (!(pvmw->pmd && !pvmw->pte))
2881                 return;
2882
2883         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2884         pmdval = *pvmw->pmd;
2885         pmdp_invalidate(vma, address, pvmw->pmd);
2886         if (pmd_dirty(pmdval))
2887                 set_page_dirty(page);
2888         entry = make_migration_entry(page, pmd_write(pmdval));
2889         pmdswp = swp_entry_to_pmd(entry);
2890         if (pmd_soft_dirty(pmdval))
2891                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2892         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2893         page_remove_rmap(page, true);
2894         put_page(page);
2895 }
2896
2897 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2898 {
2899         struct vm_area_struct *vma = pvmw->vma;
2900         struct mm_struct *mm = vma->vm_mm;
2901         unsigned long address = pvmw->address;
2902         unsigned long mmun_start = address & HPAGE_PMD_MASK;
2903         pmd_t pmde;
2904         swp_entry_t entry;
2905
2906         if (!(pvmw->pmd && !pvmw->pte))
2907                 return;
2908
2909         entry = pmd_to_swp_entry(*pvmw->pmd);
2910         get_page(new);
2911         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2912         if (pmd_swp_soft_dirty(*pvmw->pmd))
2913                 pmde = pmd_mksoft_dirty(pmde);
2914         if (is_write_migration_entry(entry))
2915                 pmde = maybe_pmd_mkwrite(pmde, vma);
2916
2917         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2918         if (PageAnon(new))
2919                 page_add_anon_rmap(new, vma, mmun_start, true);
2920         else
2921                 page_add_file_rmap(new, true);
2922         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2923         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
2924                 mlock_vma_page(new);
2925         update_mmu_cache_pmd(vma, address, pvmw->pmd);
2926 }
2927 #endif