Merge tag 'sched-urgent-2020-12-27' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-2.6-microblaze.git] / mm / huge_memory.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  Copyright (C) 2009  Red Hat, Inc.
4  */
5
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7
8 #include <linux/mm.h>
9 #include <linux/sched.h>
10 #include <linux/sched/coredump.h>
11 #include <linux/sched/numa_balancing.h>
12 #include <linux/highmem.h>
13 #include <linux/hugetlb.h>
14 #include <linux/mmu_notifier.h>
15 #include <linux/rmap.h>
16 #include <linux/swap.h>
17 #include <linux/shrinker.h>
18 #include <linux/mm_inline.h>
19 #include <linux/swapops.h>
20 #include <linux/dax.h>
21 #include <linux/khugepaged.h>
22 #include <linux/freezer.h>
23 #include <linux/pfn_t.h>
24 #include <linux/mman.h>
25 #include <linux/memremap.h>
26 #include <linux/pagemap.h>
27 #include <linux/debugfs.h>
28 #include <linux/migrate.h>
29 #include <linux/hashtable.h>
30 #include <linux/userfaultfd_k.h>
31 #include <linux/page_idle.h>
32 #include <linux/shmem_fs.h>
33 #include <linux/oom.h>
34 #include <linux/numa.h>
35 #include <linux/page_owner.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 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
66 {
67         /* The addr is used to check if the vma size fits */
68         unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
69
70         if (!transhuge_vma_suitable(vma, addr))
71                 return false;
72         if (vma_is_anonymous(vma))
73                 return __transparent_hugepage_enabled(vma);
74         if (vma_is_shmem(vma))
75                 return shmem_huge_enabled(vma);
76
77         return false;
78 }
79
80 static struct page *get_huge_zero_page(void)
81 {
82         struct page *zero_page;
83 retry:
84         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
85                 return READ_ONCE(huge_zero_page);
86
87         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
88                         HPAGE_PMD_ORDER);
89         if (!zero_page) {
90                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
91                 return NULL;
92         }
93         count_vm_event(THP_ZERO_PAGE_ALLOC);
94         preempt_disable();
95         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
96                 preempt_enable();
97                 __free_pages(zero_page, compound_order(zero_page));
98                 goto retry;
99         }
100
101         /* We take additional reference here. It will be put back by shrinker */
102         atomic_set(&huge_zero_refcount, 2);
103         preempt_enable();
104         return READ_ONCE(huge_zero_page);
105 }
106
107 static void put_huge_zero_page(void)
108 {
109         /*
110          * Counter should never go to zero here. Only shrinker can put
111          * last reference.
112          */
113         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
114 }
115
116 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
117 {
118         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
119                 return READ_ONCE(huge_zero_page);
120
121         if (!get_huge_zero_page())
122                 return NULL;
123
124         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
125                 put_huge_zero_page();
126
127         return READ_ONCE(huge_zero_page);
128 }
129
130 void mm_put_huge_zero_page(struct mm_struct *mm)
131 {
132         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
133                 put_huge_zero_page();
134 }
135
136 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
137                                         struct shrink_control *sc)
138 {
139         /* we can free zero page only if last reference remains */
140         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
141 }
142
143 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
144                                        struct shrink_control *sc)
145 {
146         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
147                 struct page *zero_page = xchg(&huge_zero_page, NULL);
148                 BUG_ON(zero_page == NULL);
149                 __free_pages(zero_page, compound_order(zero_page));
150                 return HPAGE_PMD_NR;
151         }
152
153         return 0;
154 }
155
156 static struct shrinker huge_zero_page_shrinker = {
157         .count_objects = shrink_huge_zero_page_count,
158         .scan_objects = shrink_huge_zero_page_scan,
159         .seeks = DEFAULT_SEEKS,
160 };
161
162 #ifdef CONFIG_SYSFS
163 static ssize_t enabled_show(struct kobject *kobj,
164                             struct kobj_attribute *attr, char *buf)
165 {
166         const char *output;
167
168         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
169                 output = "[always] madvise never";
170         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
171                           &transparent_hugepage_flags))
172                 output = "always [madvise] never";
173         else
174                 output = "always madvise [never]";
175
176         return sysfs_emit(buf, "%s\n", output);
177 }
178
179 static ssize_t enabled_store(struct kobject *kobj,
180                              struct kobj_attribute *attr,
181                              const char *buf, size_t count)
182 {
183         ssize_t ret = count;
184
185         if (sysfs_streq(buf, "always")) {
186                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
187                 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
188         } else if (sysfs_streq(buf, "madvise")) {
189                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
190                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
191         } else if (sysfs_streq(buf, "never")) {
192                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
193                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
194         } else
195                 ret = -EINVAL;
196
197         if (ret > 0) {
198                 int err = start_stop_khugepaged();
199                 if (err)
200                         ret = err;
201         }
202         return ret;
203 }
204 static struct kobj_attribute enabled_attr =
205         __ATTR(enabled, 0644, enabled_show, enabled_store);
206
207 ssize_t single_hugepage_flag_show(struct kobject *kobj,
208                                   struct kobj_attribute *attr, char *buf,
209                                   enum transparent_hugepage_flag flag)
210 {
211         return sysfs_emit(buf, "%d\n",
212                           !!test_bit(flag, &transparent_hugepage_flags));
213 }
214
215 ssize_t single_hugepage_flag_store(struct kobject *kobj,
216                                  struct kobj_attribute *attr,
217                                  const char *buf, size_t count,
218                                  enum transparent_hugepage_flag flag)
219 {
220         unsigned long value;
221         int ret;
222
223         ret = kstrtoul(buf, 10, &value);
224         if (ret < 0)
225                 return ret;
226         if (value > 1)
227                 return -EINVAL;
228
229         if (value)
230                 set_bit(flag, &transparent_hugepage_flags);
231         else
232                 clear_bit(flag, &transparent_hugepage_flags);
233
234         return count;
235 }
236
237 static ssize_t defrag_show(struct kobject *kobj,
238                            struct kobj_attribute *attr, char *buf)
239 {
240         const char *output;
241
242         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
243                      &transparent_hugepage_flags))
244                 output = "[always] defer defer+madvise madvise never";
245         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
246                           &transparent_hugepage_flags))
247                 output = "always [defer] defer+madvise madvise never";
248         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
249                           &transparent_hugepage_flags))
250                 output = "always defer [defer+madvise] madvise never";
251         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
252                           &transparent_hugepage_flags))
253                 output = "always defer defer+madvise [madvise] never";
254         else
255                 output = "always defer defer+madvise madvise [never]";
256
257         return sysfs_emit(buf, "%s\n", output);
258 }
259
260 static ssize_t defrag_store(struct kobject *kobj,
261                             struct kobj_attribute *attr,
262                             const char *buf, size_t count)
263 {
264         if (sysfs_streq(buf, "always")) {
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                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
269         } else if (sysfs_streq(buf, "defer+madvise")) {
270                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
271                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
272                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
273                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
274         } else if (sysfs_streq(buf, "defer")) {
275                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
276                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
277                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
278                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
279         } else if (sysfs_streq(buf, "madvise")) {
280                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
281                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
282                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
283                 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
284         } else if (sysfs_streq(buf, "never")) {
285                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
286                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
287                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
288                 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
289         } else
290                 return -EINVAL;
291
292         return count;
293 }
294 static struct kobj_attribute defrag_attr =
295         __ATTR(defrag, 0644, defrag_show, defrag_store);
296
297 static ssize_t use_zero_page_show(struct kobject *kobj,
298                                   struct kobj_attribute *attr, char *buf)
299 {
300         return single_hugepage_flag_show(kobj, attr, buf,
301                                          TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
302 }
303 static ssize_t use_zero_page_store(struct kobject *kobj,
304                 struct kobj_attribute *attr, const char *buf, size_t count)
305 {
306         return single_hugepage_flag_store(kobj, attr, buf, count,
307                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
308 }
309 static struct kobj_attribute use_zero_page_attr =
310         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
311
312 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
313                                    struct kobj_attribute *attr, char *buf)
314 {
315         return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
316 }
317 static struct kobj_attribute hpage_pmd_size_attr =
318         __ATTR_RO(hpage_pmd_size);
319
320 static struct attribute *hugepage_attr[] = {
321         &enabled_attr.attr,
322         &defrag_attr.attr,
323         &use_zero_page_attr.attr,
324         &hpage_pmd_size_attr.attr,
325 #ifdef CONFIG_SHMEM
326         &shmem_enabled_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 #ifdef CONFIG_MEMCG
485 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
486 {
487         struct mem_cgroup *memcg = page_memcg(compound_head(page));
488         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
489
490         if (memcg)
491                 return &memcg->deferred_split_queue;
492         else
493                 return &pgdat->deferred_split_queue;
494 }
495 #else
496 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
497 {
498         struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
499
500         return &pgdat->deferred_split_queue;
501 }
502 #endif
503
504 void prep_transhuge_page(struct page *page)
505 {
506         /*
507          * we use page->mapping and page->indexlru in second tail page
508          * as list_head: assuming THP order >= 2
509          */
510
511         INIT_LIST_HEAD(page_deferred_list(page));
512         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
513 }
514
515 bool is_transparent_hugepage(struct page *page)
516 {
517         if (!PageCompound(page))
518                 return false;
519
520         page = compound_head(page);
521         return is_huge_zero_page(page) ||
522                page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
523 }
524 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
525
526 static unsigned long __thp_get_unmapped_area(struct file *filp,
527                 unsigned long addr, unsigned long len,
528                 loff_t off, unsigned long flags, unsigned long size)
529 {
530         loff_t off_end = off + len;
531         loff_t off_align = round_up(off, size);
532         unsigned long len_pad, ret;
533
534         if (off_end <= off_align || (off_end - off_align) < size)
535                 return 0;
536
537         len_pad = len + size;
538         if (len_pad < len || (off + len_pad) < off)
539                 return 0;
540
541         ret = current->mm->get_unmapped_area(filp, addr, len_pad,
542                                               off >> PAGE_SHIFT, flags);
543
544         /*
545          * The failure might be due to length padding. The caller will retry
546          * without the padding.
547          */
548         if (IS_ERR_VALUE(ret))
549                 return 0;
550
551         /*
552          * Do not try to align to THP boundary if allocation at the address
553          * hint succeeds.
554          */
555         if (ret == addr)
556                 return addr;
557
558         ret += (off - ret) & (size - 1);
559         return ret;
560 }
561
562 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
563                 unsigned long len, unsigned long pgoff, unsigned long flags)
564 {
565         unsigned long ret;
566         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
567
568         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
569                 goto out;
570
571         ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
572         if (ret)
573                 return ret;
574 out:
575         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
576 }
577 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
578
579 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
580                         struct page *page, gfp_t gfp)
581 {
582         struct vm_area_struct *vma = vmf->vma;
583         pgtable_t pgtable;
584         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
585         vm_fault_t ret = 0;
586
587         VM_BUG_ON_PAGE(!PageCompound(page), page);
588
589         if (mem_cgroup_charge(page, vma->vm_mm, gfp)) {
590                 put_page(page);
591                 count_vm_event(THP_FAULT_FALLBACK);
592                 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
593                 return VM_FAULT_FALLBACK;
594         }
595         cgroup_throttle_swaprate(page, gfp);
596
597         pgtable = pte_alloc_one(vma->vm_mm);
598         if (unlikely(!pgtable)) {
599                 ret = VM_FAULT_OOM;
600                 goto release;
601         }
602
603         clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
604         /*
605          * The memory barrier inside __SetPageUptodate makes sure that
606          * clear_huge_page writes become visible before the set_pmd_at()
607          * write.
608          */
609         __SetPageUptodate(page);
610
611         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
612         if (unlikely(!pmd_none(*vmf->pmd))) {
613                 goto unlock_release;
614         } else {
615                 pmd_t entry;
616
617                 ret = check_stable_address_space(vma->vm_mm);
618                 if (ret)
619                         goto unlock_release;
620
621                 /* Deliver the page fault to userland */
622                 if (userfaultfd_missing(vma)) {
623                         vm_fault_t ret2;
624
625                         spin_unlock(vmf->ptl);
626                         put_page(page);
627                         pte_free(vma->vm_mm, pgtable);
628                         ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
629                         VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
630                         return ret2;
631                 }
632
633                 entry = mk_huge_pmd(page, vma->vm_page_prot);
634                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
635                 page_add_new_anon_rmap(page, vma, haddr, true);
636                 lru_cache_add_inactive_or_unevictable(page, vma);
637                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
638                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
639                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
640                 mm_inc_nr_ptes(vma->vm_mm);
641                 spin_unlock(vmf->ptl);
642                 count_vm_event(THP_FAULT_ALLOC);
643                 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
644         }
645
646         return 0;
647 unlock_release:
648         spin_unlock(vmf->ptl);
649 release:
650         if (pgtable)
651                 pte_free(vma->vm_mm, pgtable);
652         put_page(page);
653         return ret;
654
655 }
656
657 /*
658  * always: directly stall for all thp allocations
659  * defer: wake kswapd and fail if not immediately available
660  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
661  *                fail if not immediately available
662  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
663  *          available
664  * never: never stall for any thp allocation
665  */
666 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
667 {
668         const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
669
670         /* Always do synchronous compaction */
671         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
672                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
673
674         /* Kick kcompactd and fail quickly */
675         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
676                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
677
678         /* Synchronous compaction if madvised, otherwise kick kcompactd */
679         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
680                 return GFP_TRANSHUGE_LIGHT |
681                         (vma_madvised ? __GFP_DIRECT_RECLAIM :
682                                         __GFP_KSWAPD_RECLAIM);
683
684         /* Only do synchronous compaction if madvised */
685         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
686                 return GFP_TRANSHUGE_LIGHT |
687                        (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
688
689         return GFP_TRANSHUGE_LIGHT;
690 }
691
692 /* Caller must hold page table lock. */
693 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
694                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
695                 struct page *zero_page)
696 {
697         pmd_t entry;
698         if (!pmd_none(*pmd))
699                 return false;
700         entry = mk_pmd(zero_page, vma->vm_page_prot);
701         entry = pmd_mkhuge(entry);
702         if (pgtable)
703                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
704         set_pmd_at(mm, haddr, pmd, entry);
705         mm_inc_nr_ptes(mm);
706         return true;
707 }
708
709 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
710 {
711         struct vm_area_struct *vma = vmf->vma;
712         gfp_t gfp;
713         struct page *page;
714         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
715
716         if (!transhuge_vma_suitable(vma, haddr))
717                 return VM_FAULT_FALLBACK;
718         if (unlikely(anon_vma_prepare(vma)))
719                 return VM_FAULT_OOM;
720         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
721                 return VM_FAULT_OOM;
722         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
723                         !mm_forbids_zeropage(vma->vm_mm) &&
724                         transparent_hugepage_use_zero_page()) {
725                 pgtable_t pgtable;
726                 struct page *zero_page;
727                 vm_fault_t ret;
728                 pgtable = pte_alloc_one(vma->vm_mm);
729                 if (unlikely(!pgtable))
730                         return VM_FAULT_OOM;
731                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
732                 if (unlikely(!zero_page)) {
733                         pte_free(vma->vm_mm, pgtable);
734                         count_vm_event(THP_FAULT_FALLBACK);
735                         return VM_FAULT_FALLBACK;
736                 }
737                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
738                 ret = 0;
739                 if (pmd_none(*vmf->pmd)) {
740                         ret = check_stable_address_space(vma->vm_mm);
741                         if (ret) {
742                                 spin_unlock(vmf->ptl);
743                                 pte_free(vma->vm_mm, pgtable);
744                         } else if (userfaultfd_missing(vma)) {
745                                 spin_unlock(vmf->ptl);
746                                 pte_free(vma->vm_mm, pgtable);
747                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
748                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
749                         } else {
750                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
751                                                    haddr, vmf->pmd, zero_page);
752                                 spin_unlock(vmf->ptl);
753                         }
754                 } else {
755                         spin_unlock(vmf->ptl);
756                         pte_free(vma->vm_mm, pgtable);
757                 }
758                 return ret;
759         }
760         gfp = alloc_hugepage_direct_gfpmask(vma);
761         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
762         if (unlikely(!page)) {
763                 count_vm_event(THP_FAULT_FALLBACK);
764                 return VM_FAULT_FALLBACK;
765         }
766         prep_transhuge_page(page);
767         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
768 }
769
770 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
771                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
772                 pgtable_t pgtable)
773 {
774         struct mm_struct *mm = vma->vm_mm;
775         pmd_t entry;
776         spinlock_t *ptl;
777
778         ptl = pmd_lock(mm, pmd);
779         if (!pmd_none(*pmd)) {
780                 if (write) {
781                         if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
782                                 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
783                                 goto out_unlock;
784                         }
785                         entry = pmd_mkyoung(*pmd);
786                         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
787                         if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
788                                 update_mmu_cache_pmd(vma, addr, pmd);
789                 }
790
791                 goto out_unlock;
792         }
793
794         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
795         if (pfn_t_devmap(pfn))
796                 entry = pmd_mkdevmap(entry);
797         if (write) {
798                 entry = pmd_mkyoung(pmd_mkdirty(entry));
799                 entry = maybe_pmd_mkwrite(entry, vma);
800         }
801
802         if (pgtable) {
803                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
804                 mm_inc_nr_ptes(mm);
805                 pgtable = NULL;
806         }
807
808         set_pmd_at(mm, addr, pmd, entry);
809         update_mmu_cache_pmd(vma, addr, pmd);
810
811 out_unlock:
812         spin_unlock(ptl);
813         if (pgtable)
814                 pte_free(mm, pgtable);
815 }
816
817 /**
818  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
819  * @vmf: Structure describing the fault
820  * @pfn: pfn to insert
821  * @pgprot: page protection to use
822  * @write: whether it's a write fault
823  *
824  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
825  * also consult the vmf_insert_mixed_prot() documentation when
826  * @pgprot != @vmf->vma->vm_page_prot.
827  *
828  * Return: vm_fault_t value.
829  */
830 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
831                                    pgprot_t pgprot, bool write)
832 {
833         unsigned long addr = vmf->address & PMD_MASK;
834         struct vm_area_struct *vma = vmf->vma;
835         pgtable_t pgtable = NULL;
836
837         /*
838          * If we had pmd_special, we could avoid all these restrictions,
839          * but we need to be consistent with PTEs and architectures that
840          * can't support a 'special' bit.
841          */
842         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
843                         !pfn_t_devmap(pfn));
844         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
845                                                 (VM_PFNMAP|VM_MIXEDMAP));
846         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
847
848         if (addr < vma->vm_start || addr >= vma->vm_end)
849                 return VM_FAULT_SIGBUS;
850
851         if (arch_needs_pgtable_deposit()) {
852                 pgtable = pte_alloc_one(vma->vm_mm);
853                 if (!pgtable)
854                         return VM_FAULT_OOM;
855         }
856
857         track_pfn_insert(vma, &pgprot, pfn);
858
859         insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
860         return VM_FAULT_NOPAGE;
861 }
862 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
863
864 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
865 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
866 {
867         if (likely(vma->vm_flags & VM_WRITE))
868                 pud = pud_mkwrite(pud);
869         return pud;
870 }
871
872 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
873                 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
874 {
875         struct mm_struct *mm = vma->vm_mm;
876         pud_t entry;
877         spinlock_t *ptl;
878
879         ptl = pud_lock(mm, pud);
880         if (!pud_none(*pud)) {
881                 if (write) {
882                         if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
883                                 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
884                                 goto out_unlock;
885                         }
886                         entry = pud_mkyoung(*pud);
887                         entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
888                         if (pudp_set_access_flags(vma, addr, pud, entry, 1))
889                                 update_mmu_cache_pud(vma, addr, pud);
890                 }
891                 goto out_unlock;
892         }
893
894         entry = pud_mkhuge(pfn_t_pud(pfn, prot));
895         if (pfn_t_devmap(pfn))
896                 entry = pud_mkdevmap(entry);
897         if (write) {
898                 entry = pud_mkyoung(pud_mkdirty(entry));
899                 entry = maybe_pud_mkwrite(entry, vma);
900         }
901         set_pud_at(mm, addr, pud, entry);
902         update_mmu_cache_pud(vma, addr, pud);
903
904 out_unlock:
905         spin_unlock(ptl);
906 }
907
908 /**
909  * vmf_insert_pfn_pud_prot - insert a pud size pfn
910  * @vmf: Structure describing the fault
911  * @pfn: pfn to insert
912  * @pgprot: page protection to use
913  * @write: whether it's a write fault
914  *
915  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
916  * also consult the vmf_insert_mixed_prot() documentation when
917  * @pgprot != @vmf->vma->vm_page_prot.
918  *
919  * Return: vm_fault_t value.
920  */
921 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
922                                    pgprot_t pgprot, bool write)
923 {
924         unsigned long addr = vmf->address & PUD_MASK;
925         struct vm_area_struct *vma = vmf->vma;
926
927         /*
928          * If we had pud_special, we could avoid all these restrictions,
929          * but we need to be consistent with PTEs and architectures that
930          * can't support a 'special' bit.
931          */
932         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
933                         !pfn_t_devmap(pfn));
934         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
935                                                 (VM_PFNMAP|VM_MIXEDMAP));
936         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
937
938         if (addr < vma->vm_start || addr >= vma->vm_end)
939                 return VM_FAULT_SIGBUS;
940
941         track_pfn_insert(vma, &pgprot, pfn);
942
943         insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
944         return VM_FAULT_NOPAGE;
945 }
946 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
947 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
948
949 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
950                 pmd_t *pmd, int flags)
951 {
952         pmd_t _pmd;
953
954         _pmd = pmd_mkyoung(*pmd);
955         if (flags & FOLL_WRITE)
956                 _pmd = pmd_mkdirty(_pmd);
957         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
958                                 pmd, _pmd, flags & FOLL_WRITE))
959                 update_mmu_cache_pmd(vma, addr, pmd);
960 }
961
962 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
963                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
964 {
965         unsigned long pfn = pmd_pfn(*pmd);
966         struct mm_struct *mm = vma->vm_mm;
967         struct page *page;
968
969         assert_spin_locked(pmd_lockptr(mm, pmd));
970
971         /*
972          * When we COW a devmap PMD entry, we split it into PTEs, so we should
973          * not be in this function with `flags & FOLL_COW` set.
974          */
975         WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
976
977         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
978         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
979                          (FOLL_PIN | FOLL_GET)))
980                 return NULL;
981
982         if (flags & FOLL_WRITE && !pmd_write(*pmd))
983                 return NULL;
984
985         if (pmd_present(*pmd) && pmd_devmap(*pmd))
986                 /* pass */;
987         else
988                 return NULL;
989
990         if (flags & FOLL_TOUCH)
991                 touch_pmd(vma, addr, pmd, flags);
992
993         /*
994          * device mapped pages can only be returned if the
995          * caller will manage the page reference count.
996          */
997         if (!(flags & (FOLL_GET | FOLL_PIN)))
998                 return ERR_PTR(-EEXIST);
999
1000         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1001         *pgmap = get_dev_pagemap(pfn, *pgmap);
1002         if (!*pgmap)
1003                 return ERR_PTR(-EFAULT);
1004         page = pfn_to_page(pfn);
1005         if (!try_grab_page(page, flags))
1006                 page = ERR_PTR(-ENOMEM);
1007
1008         return page;
1009 }
1010
1011 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1012                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1013                   struct vm_area_struct *vma)
1014 {
1015         spinlock_t *dst_ptl, *src_ptl;
1016         struct page *src_page;
1017         pmd_t pmd;
1018         pgtable_t pgtable = NULL;
1019         int ret = -ENOMEM;
1020
1021         /* Skip if can be re-fill on fault */
1022         if (!vma_is_anonymous(vma))
1023                 return 0;
1024
1025         pgtable = pte_alloc_one(dst_mm);
1026         if (unlikely(!pgtable))
1027                 goto out;
1028
1029         dst_ptl = pmd_lock(dst_mm, dst_pmd);
1030         src_ptl = pmd_lockptr(src_mm, src_pmd);
1031         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1032
1033         ret = -EAGAIN;
1034         pmd = *src_pmd;
1035
1036         /*
1037          * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
1038          * does not have the VM_UFFD_WP, which means that the uffd
1039          * fork event is not enabled.
1040          */
1041         if (!(vma->vm_flags & VM_UFFD_WP))
1042                 pmd = pmd_clear_uffd_wp(pmd);
1043
1044 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1045         if (unlikely(is_swap_pmd(pmd))) {
1046                 swp_entry_t entry = pmd_to_swp_entry(pmd);
1047
1048                 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1049                 if (is_write_migration_entry(entry)) {
1050                         make_migration_entry_read(&entry);
1051                         pmd = swp_entry_to_pmd(entry);
1052                         if (pmd_swp_soft_dirty(*src_pmd))
1053                                 pmd = pmd_swp_mksoft_dirty(pmd);
1054                         set_pmd_at(src_mm, addr, src_pmd, pmd);
1055                 }
1056                 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1057                 mm_inc_nr_ptes(dst_mm);
1058                 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1059                 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1060                 ret = 0;
1061                 goto out_unlock;
1062         }
1063 #endif
1064
1065         if (unlikely(!pmd_trans_huge(pmd))) {
1066                 pte_free(dst_mm, pgtable);
1067                 goto out_unlock;
1068         }
1069         /*
1070          * When page table lock is held, the huge zero pmd should not be
1071          * under splitting since we don't split the page itself, only pmd to
1072          * a page table.
1073          */
1074         if (is_huge_zero_pmd(pmd)) {
1075                 struct page *zero_page;
1076                 /*
1077                  * get_huge_zero_page() will never allocate a new page here,
1078                  * since we already have a zero page to copy. It just takes a
1079                  * reference.
1080                  */
1081                 zero_page = mm_get_huge_zero_page(dst_mm);
1082                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1083                                 zero_page);
1084                 ret = 0;
1085                 goto out_unlock;
1086         }
1087
1088         src_page = pmd_page(pmd);
1089         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1090
1091         /*
1092          * If this page is a potentially pinned page, split and retry the fault
1093          * with smaller page size.  Normally this should not happen because the
1094          * userspace should use MADV_DONTFORK upon pinned regions.  This is a
1095          * best effort that the pinned pages won't be replaced by another
1096          * random page during the coming copy-on-write.
1097          */
1098         if (unlikely(is_cow_mapping(vma->vm_flags) &&
1099                      atomic_read(&src_mm->has_pinned) &&
1100                      page_maybe_dma_pinned(src_page))) {
1101                 pte_free(dst_mm, pgtable);
1102                 spin_unlock(src_ptl);
1103                 spin_unlock(dst_ptl);
1104                 __split_huge_pmd(vma, src_pmd, addr, false, NULL);
1105                 return -EAGAIN;
1106         }
1107
1108         get_page(src_page);
1109         page_dup_rmap(src_page, true);
1110         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1111         mm_inc_nr_ptes(dst_mm);
1112         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1113
1114         pmdp_set_wrprotect(src_mm, addr, src_pmd);
1115         pmd = pmd_mkold(pmd_wrprotect(pmd));
1116         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1117
1118         ret = 0;
1119 out_unlock:
1120         spin_unlock(src_ptl);
1121         spin_unlock(dst_ptl);
1122 out:
1123         return ret;
1124 }
1125
1126 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1127 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1128                 pud_t *pud, int flags)
1129 {
1130         pud_t _pud;
1131
1132         _pud = pud_mkyoung(*pud);
1133         if (flags & FOLL_WRITE)
1134                 _pud = pud_mkdirty(_pud);
1135         if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1136                                 pud, _pud, flags & FOLL_WRITE))
1137                 update_mmu_cache_pud(vma, addr, pud);
1138 }
1139
1140 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1141                 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1142 {
1143         unsigned long pfn = pud_pfn(*pud);
1144         struct mm_struct *mm = vma->vm_mm;
1145         struct page *page;
1146
1147         assert_spin_locked(pud_lockptr(mm, pud));
1148
1149         if (flags & FOLL_WRITE && !pud_write(*pud))
1150                 return NULL;
1151
1152         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1153         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1154                          (FOLL_PIN | FOLL_GET)))
1155                 return NULL;
1156
1157         if (pud_present(*pud) && pud_devmap(*pud))
1158                 /* pass */;
1159         else
1160                 return NULL;
1161
1162         if (flags & FOLL_TOUCH)
1163                 touch_pud(vma, addr, pud, flags);
1164
1165         /*
1166          * device mapped pages can only be returned if the
1167          * caller will manage the page reference count.
1168          *
1169          * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1170          */
1171         if (!(flags & (FOLL_GET | FOLL_PIN)))
1172                 return ERR_PTR(-EEXIST);
1173
1174         pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1175         *pgmap = get_dev_pagemap(pfn, *pgmap);
1176         if (!*pgmap)
1177                 return ERR_PTR(-EFAULT);
1178         page = pfn_to_page(pfn);
1179         if (!try_grab_page(page, flags))
1180                 page = ERR_PTR(-ENOMEM);
1181
1182         return page;
1183 }
1184
1185 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1186                   pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1187                   struct vm_area_struct *vma)
1188 {
1189         spinlock_t *dst_ptl, *src_ptl;
1190         pud_t pud;
1191         int ret;
1192
1193         dst_ptl = pud_lock(dst_mm, dst_pud);
1194         src_ptl = pud_lockptr(src_mm, src_pud);
1195         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1196
1197         ret = -EAGAIN;
1198         pud = *src_pud;
1199         if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1200                 goto out_unlock;
1201
1202         /*
1203          * When page table lock is held, the huge zero pud should not be
1204          * under splitting since we don't split the page itself, only pud to
1205          * a page table.
1206          */
1207         if (is_huge_zero_pud(pud)) {
1208                 /* No huge zero pud yet */
1209         }
1210
1211         /* Please refer to comments in copy_huge_pmd() */
1212         if (unlikely(is_cow_mapping(vma->vm_flags) &&
1213                      atomic_read(&src_mm->has_pinned) &&
1214                      page_maybe_dma_pinned(pud_page(pud)))) {
1215                 spin_unlock(src_ptl);
1216                 spin_unlock(dst_ptl);
1217                 __split_huge_pud(vma, src_pud, addr);
1218                 return -EAGAIN;
1219         }
1220
1221         pudp_set_wrprotect(src_mm, addr, src_pud);
1222         pud = pud_mkold(pud_wrprotect(pud));
1223         set_pud_at(dst_mm, addr, dst_pud, pud);
1224
1225         ret = 0;
1226 out_unlock:
1227         spin_unlock(src_ptl);
1228         spin_unlock(dst_ptl);
1229         return ret;
1230 }
1231
1232 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1233 {
1234         pud_t entry;
1235         unsigned long haddr;
1236         bool write = vmf->flags & FAULT_FLAG_WRITE;
1237
1238         vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1239         if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1240                 goto unlock;
1241
1242         entry = pud_mkyoung(orig_pud);
1243         if (write)
1244                 entry = pud_mkdirty(entry);
1245         haddr = vmf->address & HPAGE_PUD_MASK;
1246         if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1247                 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1248
1249 unlock:
1250         spin_unlock(vmf->ptl);
1251 }
1252 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1253
1254 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1255 {
1256         pmd_t entry;
1257         unsigned long haddr;
1258         bool write = vmf->flags & FAULT_FLAG_WRITE;
1259
1260         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1261         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1262                 goto unlock;
1263
1264         entry = pmd_mkyoung(orig_pmd);
1265         if (write)
1266                 entry = pmd_mkdirty(entry);
1267         haddr = vmf->address & HPAGE_PMD_MASK;
1268         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1269                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1270
1271 unlock:
1272         spin_unlock(vmf->ptl);
1273 }
1274
1275 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1276 {
1277         struct vm_area_struct *vma = vmf->vma;
1278         struct page *page;
1279         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1280
1281         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1282         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1283
1284         if (is_huge_zero_pmd(orig_pmd))
1285                 goto fallback;
1286
1287         spin_lock(vmf->ptl);
1288
1289         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1290                 spin_unlock(vmf->ptl);
1291                 return 0;
1292         }
1293
1294         page = pmd_page(orig_pmd);
1295         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1296
1297         /* Lock page for reuse_swap_page() */
1298         if (!trylock_page(page)) {
1299                 get_page(page);
1300                 spin_unlock(vmf->ptl);
1301                 lock_page(page);
1302                 spin_lock(vmf->ptl);
1303                 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1304                         spin_unlock(vmf->ptl);
1305                         unlock_page(page);
1306                         put_page(page);
1307                         return 0;
1308                 }
1309                 put_page(page);
1310         }
1311
1312         /*
1313          * We can only reuse the page if nobody else maps the huge page or it's
1314          * part.
1315          */
1316         if (reuse_swap_page(page, NULL)) {
1317                 pmd_t entry;
1318                 entry = pmd_mkyoung(orig_pmd);
1319                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1320                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1321                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1322                 unlock_page(page);
1323                 spin_unlock(vmf->ptl);
1324                 return VM_FAULT_WRITE;
1325         }
1326
1327         unlock_page(page);
1328         spin_unlock(vmf->ptl);
1329 fallback:
1330         __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1331         return VM_FAULT_FALLBACK;
1332 }
1333
1334 /*
1335  * FOLL_FORCE can write to even unwritable pmd's, but only
1336  * after we've gone through a COW cycle and they are dirty.
1337  */
1338 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1339 {
1340         return pmd_write(pmd) ||
1341                ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1342 }
1343
1344 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1345                                    unsigned long addr,
1346                                    pmd_t *pmd,
1347                                    unsigned int flags)
1348 {
1349         struct mm_struct *mm = vma->vm_mm;
1350         struct page *page = NULL;
1351
1352         assert_spin_locked(pmd_lockptr(mm, pmd));
1353
1354         if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1355                 goto out;
1356
1357         /* Avoid dumping huge zero page */
1358         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1359                 return ERR_PTR(-EFAULT);
1360
1361         /* Full NUMA hinting faults to serialise migration in fault paths */
1362         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1363                 goto out;
1364
1365         page = pmd_page(*pmd);
1366         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1367
1368         if (!try_grab_page(page, flags))
1369                 return ERR_PTR(-ENOMEM);
1370
1371         if (flags & FOLL_TOUCH)
1372                 touch_pmd(vma, addr, pmd, flags);
1373
1374         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1375                 /*
1376                  * We don't mlock() pte-mapped THPs. This way we can avoid
1377                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1378                  *
1379                  * For anon THP:
1380                  *
1381                  * In most cases the pmd is the only mapping of the page as we
1382                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1383                  * writable private mappings in populate_vma_page_range().
1384                  *
1385                  * The only scenario when we have the page shared here is if we
1386                  * mlocking read-only mapping shared over fork(). We skip
1387                  * mlocking such pages.
1388                  *
1389                  * For file THP:
1390                  *
1391                  * We can expect PageDoubleMap() to be stable under page lock:
1392                  * for file pages we set it in page_add_file_rmap(), which
1393                  * requires page to be locked.
1394                  */
1395
1396                 if (PageAnon(page) && compound_mapcount(page) != 1)
1397                         goto skip_mlock;
1398                 if (PageDoubleMap(page) || !page->mapping)
1399                         goto skip_mlock;
1400                 if (!trylock_page(page))
1401                         goto skip_mlock;
1402                 if (page->mapping && !PageDoubleMap(page))
1403                         mlock_vma_page(page);
1404                 unlock_page(page);
1405         }
1406 skip_mlock:
1407         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1408         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1409
1410 out:
1411         return page;
1412 }
1413
1414 /* NUMA hinting page fault entry point for trans huge pmds */
1415 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1416 {
1417         struct vm_area_struct *vma = vmf->vma;
1418         struct anon_vma *anon_vma = NULL;
1419         struct page *page;
1420         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1421         int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1422         int target_nid, last_cpupid = -1;
1423         bool page_locked;
1424         bool migrated = false;
1425         bool was_writable;
1426         int flags = 0;
1427
1428         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1429         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1430                 goto out_unlock;
1431
1432         /*
1433          * If there are potential migrations, wait for completion and retry
1434          * without disrupting NUMA hinting information. Do not relock and
1435          * check_same as the page may no longer be mapped.
1436          */
1437         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1438                 page = pmd_page(*vmf->pmd);
1439                 if (!get_page_unless_zero(page))
1440                         goto out_unlock;
1441                 spin_unlock(vmf->ptl);
1442                 put_and_wait_on_page_locked(page);
1443                 goto out;
1444         }
1445
1446         page = pmd_page(pmd);
1447         BUG_ON(is_huge_zero_page(page));
1448         page_nid = page_to_nid(page);
1449         last_cpupid = page_cpupid_last(page);
1450         count_vm_numa_event(NUMA_HINT_FAULTS);
1451         if (page_nid == this_nid) {
1452                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1453                 flags |= TNF_FAULT_LOCAL;
1454         }
1455
1456         /* See similar comment in do_numa_page for explanation */
1457         if (!pmd_savedwrite(pmd))
1458                 flags |= TNF_NO_GROUP;
1459
1460         /*
1461          * Acquire the page lock to serialise THP migrations but avoid dropping
1462          * page_table_lock if at all possible
1463          */
1464         page_locked = trylock_page(page);
1465         target_nid = mpol_misplaced(page, vma, haddr);
1466         if (target_nid == NUMA_NO_NODE) {
1467                 /* If the page was locked, there are no parallel migrations */
1468                 if (page_locked)
1469                         goto clear_pmdnuma;
1470         }
1471
1472         /* Migration could have started since the pmd_trans_migrating check */
1473         if (!page_locked) {
1474                 page_nid = NUMA_NO_NODE;
1475                 if (!get_page_unless_zero(page))
1476                         goto out_unlock;
1477                 spin_unlock(vmf->ptl);
1478                 put_and_wait_on_page_locked(page);
1479                 goto out;
1480         }
1481
1482         /*
1483          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1484          * to serialises splits
1485          */
1486         get_page(page);
1487         spin_unlock(vmf->ptl);
1488         anon_vma = page_lock_anon_vma_read(page);
1489
1490         /* Confirm the PMD did not change while page_table_lock was released */
1491         spin_lock(vmf->ptl);
1492         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1493                 unlock_page(page);
1494                 put_page(page);
1495                 page_nid = NUMA_NO_NODE;
1496                 goto out_unlock;
1497         }
1498
1499         /* Bail if we fail to protect against THP splits for any reason */
1500         if (unlikely(!anon_vma)) {
1501                 put_page(page);
1502                 page_nid = NUMA_NO_NODE;
1503                 goto clear_pmdnuma;
1504         }
1505
1506         /*
1507          * Since we took the NUMA fault, we must have observed the !accessible
1508          * bit. Make sure all other CPUs agree with that, to avoid them
1509          * modifying the page we're about to migrate.
1510          *
1511          * Must be done under PTL such that we'll observe the relevant
1512          * inc_tlb_flush_pending().
1513          *
1514          * We are not sure a pending tlb flush here is for a huge page
1515          * mapping or not. Hence use the tlb range variant
1516          */
1517         if (mm_tlb_flush_pending(vma->vm_mm)) {
1518                 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1519                 /*
1520                  * change_huge_pmd() released the pmd lock before
1521                  * invalidating the secondary MMUs sharing the primary
1522                  * MMU pagetables (with ->invalidate_range()). The
1523                  * mmu_notifier_invalidate_range_end() (which
1524                  * internally calls ->invalidate_range()) in
1525                  * change_pmd_range() will run after us, so we can't
1526                  * rely on it here and we need an explicit invalidate.
1527                  */
1528                 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1529                                               haddr + HPAGE_PMD_SIZE);
1530         }
1531
1532         /*
1533          * Migrate the THP to the requested node, returns with page unlocked
1534          * and access rights restored.
1535          */
1536         spin_unlock(vmf->ptl);
1537
1538         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1539                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1540         if (migrated) {
1541                 flags |= TNF_MIGRATED;
1542                 page_nid = target_nid;
1543         } else
1544                 flags |= TNF_MIGRATE_FAIL;
1545
1546         goto out;
1547 clear_pmdnuma:
1548         BUG_ON(!PageLocked(page));
1549         was_writable = pmd_savedwrite(pmd);
1550         pmd = pmd_modify(pmd, vma->vm_page_prot);
1551         pmd = pmd_mkyoung(pmd);
1552         if (was_writable)
1553                 pmd = pmd_mkwrite(pmd);
1554         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1555         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1556         unlock_page(page);
1557 out_unlock:
1558         spin_unlock(vmf->ptl);
1559
1560 out:
1561         if (anon_vma)
1562                 page_unlock_anon_vma_read(anon_vma);
1563
1564         if (page_nid != NUMA_NO_NODE)
1565                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1566                                 flags);
1567
1568         return 0;
1569 }
1570
1571 /*
1572  * Return true if we do MADV_FREE successfully on entire pmd page.
1573  * Otherwise, return false.
1574  */
1575 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1576                 pmd_t *pmd, unsigned long addr, unsigned long next)
1577 {
1578         spinlock_t *ptl;
1579         pmd_t orig_pmd;
1580         struct page *page;
1581         struct mm_struct *mm = tlb->mm;
1582         bool ret = false;
1583
1584         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1585
1586         ptl = pmd_trans_huge_lock(pmd, vma);
1587         if (!ptl)
1588                 goto out_unlocked;
1589
1590         orig_pmd = *pmd;
1591         if (is_huge_zero_pmd(orig_pmd))
1592                 goto out;
1593
1594         if (unlikely(!pmd_present(orig_pmd))) {
1595                 VM_BUG_ON(thp_migration_supported() &&
1596                                   !is_pmd_migration_entry(orig_pmd));
1597                 goto out;
1598         }
1599
1600         page = pmd_page(orig_pmd);
1601         /*
1602          * If other processes are mapping this page, we couldn't discard
1603          * the page unless they all do MADV_FREE so let's skip the page.
1604          */
1605         if (page_mapcount(page) != 1)
1606                 goto out;
1607
1608         if (!trylock_page(page))
1609                 goto out;
1610
1611         /*
1612          * If user want to discard part-pages of THP, split it so MADV_FREE
1613          * will deactivate only them.
1614          */
1615         if (next - addr != HPAGE_PMD_SIZE) {
1616                 get_page(page);
1617                 spin_unlock(ptl);
1618                 split_huge_page(page);
1619                 unlock_page(page);
1620                 put_page(page);
1621                 goto out_unlocked;
1622         }
1623
1624         if (PageDirty(page))
1625                 ClearPageDirty(page);
1626         unlock_page(page);
1627
1628         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1629                 pmdp_invalidate(vma, addr, pmd);
1630                 orig_pmd = pmd_mkold(orig_pmd);
1631                 orig_pmd = pmd_mkclean(orig_pmd);
1632
1633                 set_pmd_at(mm, addr, pmd, orig_pmd);
1634                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1635         }
1636
1637         mark_page_lazyfree(page);
1638         ret = true;
1639 out:
1640         spin_unlock(ptl);
1641 out_unlocked:
1642         return ret;
1643 }
1644
1645 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1646 {
1647         pgtable_t pgtable;
1648
1649         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1650         pte_free(mm, pgtable);
1651         mm_dec_nr_ptes(mm);
1652 }
1653
1654 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1655                  pmd_t *pmd, unsigned long addr)
1656 {
1657         pmd_t orig_pmd;
1658         spinlock_t *ptl;
1659
1660         tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1661
1662         ptl = __pmd_trans_huge_lock(pmd, vma);
1663         if (!ptl)
1664                 return 0;
1665         /*
1666          * For architectures like ppc64 we look at deposited pgtable
1667          * when calling pmdp_huge_get_and_clear. So do the
1668          * pgtable_trans_huge_withdraw after finishing pmdp related
1669          * operations.
1670          */
1671         orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1672                                                 tlb->fullmm);
1673         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1674         if (vma_is_special_huge(vma)) {
1675                 if (arch_needs_pgtable_deposit())
1676                         zap_deposited_table(tlb->mm, pmd);
1677                 spin_unlock(ptl);
1678                 if (is_huge_zero_pmd(orig_pmd))
1679                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1680         } else if (is_huge_zero_pmd(orig_pmd)) {
1681                 zap_deposited_table(tlb->mm, pmd);
1682                 spin_unlock(ptl);
1683                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1684         } else {
1685                 struct page *page = NULL;
1686                 int flush_needed = 1;
1687
1688                 if (pmd_present(orig_pmd)) {
1689                         page = pmd_page(orig_pmd);
1690                         page_remove_rmap(page, true);
1691                         VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1692                         VM_BUG_ON_PAGE(!PageHead(page), page);
1693                 } else if (thp_migration_supported()) {
1694                         swp_entry_t entry;
1695
1696                         VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1697                         entry = pmd_to_swp_entry(orig_pmd);
1698                         page = pfn_to_page(swp_offset(entry));
1699                         flush_needed = 0;
1700                 } else
1701                         WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1702
1703                 if (PageAnon(page)) {
1704                         zap_deposited_table(tlb->mm, pmd);
1705                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1706                 } else {
1707                         if (arch_needs_pgtable_deposit())
1708                                 zap_deposited_table(tlb->mm, pmd);
1709                         add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1710                 }
1711
1712                 spin_unlock(ptl);
1713                 if (flush_needed)
1714                         tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1715         }
1716         return 1;
1717 }
1718
1719 #ifndef pmd_move_must_withdraw
1720 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1721                                          spinlock_t *old_pmd_ptl,
1722                                          struct vm_area_struct *vma)
1723 {
1724         /*
1725          * With split pmd lock we also need to move preallocated
1726          * PTE page table if new_pmd is on different PMD page table.
1727          *
1728          * We also don't deposit and withdraw tables for file pages.
1729          */
1730         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1731 }
1732 #endif
1733
1734 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1735 {
1736 #ifdef CONFIG_MEM_SOFT_DIRTY
1737         if (unlikely(is_pmd_migration_entry(pmd)))
1738                 pmd = pmd_swp_mksoft_dirty(pmd);
1739         else if (pmd_present(pmd))
1740                 pmd = pmd_mksoft_dirty(pmd);
1741 #endif
1742         return pmd;
1743 }
1744
1745 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1746                   unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1747 {
1748         spinlock_t *old_ptl, *new_ptl;
1749         pmd_t pmd;
1750         struct mm_struct *mm = vma->vm_mm;
1751         bool force_flush = false;
1752
1753         /*
1754          * The destination pmd shouldn't be established, free_pgtables()
1755          * should have release it.
1756          */
1757         if (WARN_ON(!pmd_none(*new_pmd))) {
1758                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1759                 return false;
1760         }
1761
1762         /*
1763          * We don't have to worry about the ordering of src and dst
1764          * ptlocks because exclusive mmap_lock prevents deadlock.
1765          */
1766         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1767         if (old_ptl) {
1768                 new_ptl = pmd_lockptr(mm, new_pmd);
1769                 if (new_ptl != old_ptl)
1770                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1771                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1772                 if (pmd_present(pmd))
1773                         force_flush = true;
1774                 VM_BUG_ON(!pmd_none(*new_pmd));
1775
1776                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1777                         pgtable_t pgtable;
1778                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1779                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1780                 }
1781                 pmd = move_soft_dirty_pmd(pmd);
1782                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1783                 if (force_flush)
1784                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1785                 if (new_ptl != old_ptl)
1786                         spin_unlock(new_ptl);
1787                 spin_unlock(old_ptl);
1788                 return true;
1789         }
1790         return false;
1791 }
1792
1793 /*
1794  * Returns
1795  *  - 0 if PMD could not be locked
1796  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1797  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1798  */
1799 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1800                 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1801 {
1802         struct mm_struct *mm = vma->vm_mm;
1803         spinlock_t *ptl;
1804         pmd_t entry;
1805         bool preserve_write;
1806         int ret;
1807         bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1808         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1809         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1810
1811         ptl = __pmd_trans_huge_lock(pmd, vma);
1812         if (!ptl)
1813                 return 0;
1814
1815         preserve_write = prot_numa && pmd_write(*pmd);
1816         ret = 1;
1817
1818 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1819         if (is_swap_pmd(*pmd)) {
1820                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1821
1822                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1823                 if (is_write_migration_entry(entry)) {
1824                         pmd_t newpmd;
1825                         /*
1826                          * A protection check is difficult so
1827                          * just be safe and disable write
1828                          */
1829                         make_migration_entry_read(&entry);
1830                         newpmd = swp_entry_to_pmd(entry);
1831                         if (pmd_swp_soft_dirty(*pmd))
1832                                 newpmd = pmd_swp_mksoft_dirty(newpmd);
1833                         set_pmd_at(mm, addr, pmd, newpmd);
1834                 }
1835                 goto unlock;
1836         }
1837 #endif
1838
1839         /*
1840          * Avoid trapping faults against the zero page. The read-only
1841          * data is likely to be read-cached on the local CPU and
1842          * local/remote hits to the zero page are not interesting.
1843          */
1844         if (prot_numa && is_huge_zero_pmd(*pmd))
1845                 goto unlock;
1846
1847         if (prot_numa && pmd_protnone(*pmd))
1848                 goto unlock;
1849
1850         /*
1851          * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1852          * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1853          * which is also under mmap_read_lock(mm):
1854          *
1855          *      CPU0:                           CPU1:
1856          *                              change_huge_pmd(prot_numa=1)
1857          *                               pmdp_huge_get_and_clear_notify()
1858          * madvise_dontneed()
1859          *  zap_pmd_range()
1860          *   pmd_trans_huge(*pmd) == 0 (without ptl)
1861          *   // skip the pmd
1862          *                               set_pmd_at();
1863          *                               // pmd is re-established
1864          *
1865          * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1866          * which may break userspace.
1867          *
1868          * pmdp_invalidate() is required to make sure we don't miss
1869          * dirty/young flags set by hardware.
1870          */
1871         entry = pmdp_invalidate(vma, addr, pmd);
1872
1873         entry = pmd_modify(entry, newprot);
1874         if (preserve_write)
1875                 entry = pmd_mk_savedwrite(entry);
1876         if (uffd_wp) {
1877                 entry = pmd_wrprotect(entry);
1878                 entry = pmd_mkuffd_wp(entry);
1879         } else if (uffd_wp_resolve) {
1880                 /*
1881                  * Leave the write bit to be handled by PF interrupt
1882                  * handler, then things like COW could be properly
1883                  * handled.
1884                  */
1885                 entry = pmd_clear_uffd_wp(entry);
1886         }
1887         ret = HPAGE_PMD_NR;
1888         set_pmd_at(mm, addr, pmd, entry);
1889         BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1890 unlock:
1891         spin_unlock(ptl);
1892         return ret;
1893 }
1894
1895 /*
1896  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1897  *
1898  * Note that if it returns page table lock pointer, this routine returns without
1899  * unlocking page table lock. So callers must unlock it.
1900  */
1901 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1902 {
1903         spinlock_t *ptl;
1904         ptl = pmd_lock(vma->vm_mm, pmd);
1905         if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1906                         pmd_devmap(*pmd)))
1907                 return ptl;
1908         spin_unlock(ptl);
1909         return NULL;
1910 }
1911
1912 /*
1913  * Returns true if a given pud maps a thp, false otherwise.
1914  *
1915  * Note that if it returns true, this routine returns without unlocking page
1916  * table lock. So callers must unlock it.
1917  */
1918 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1919 {
1920         spinlock_t *ptl;
1921
1922         ptl = pud_lock(vma->vm_mm, pud);
1923         if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1924                 return ptl;
1925         spin_unlock(ptl);
1926         return NULL;
1927 }
1928
1929 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1930 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1931                  pud_t *pud, unsigned long addr)
1932 {
1933         spinlock_t *ptl;
1934
1935         ptl = __pud_trans_huge_lock(pud, vma);
1936         if (!ptl)
1937                 return 0;
1938         /*
1939          * For architectures like ppc64 we look at deposited pgtable
1940          * when calling pudp_huge_get_and_clear. So do the
1941          * pgtable_trans_huge_withdraw after finishing pudp related
1942          * operations.
1943          */
1944         pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1945         tlb_remove_pud_tlb_entry(tlb, pud, addr);
1946         if (vma_is_special_huge(vma)) {
1947                 spin_unlock(ptl);
1948                 /* No zero page support yet */
1949         } else {
1950                 /* No support for anonymous PUD pages yet */
1951                 BUG();
1952         }
1953         return 1;
1954 }
1955
1956 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1957                 unsigned long haddr)
1958 {
1959         VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1960         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1961         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1962         VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1963
1964         count_vm_event(THP_SPLIT_PUD);
1965
1966         pudp_huge_clear_flush_notify(vma, haddr, pud);
1967 }
1968
1969 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1970                 unsigned long address)
1971 {
1972         spinlock_t *ptl;
1973         struct mmu_notifier_range range;
1974
1975         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1976                                 address & HPAGE_PUD_MASK,
1977                                 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1978         mmu_notifier_invalidate_range_start(&range);
1979         ptl = pud_lock(vma->vm_mm, pud);
1980         if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1981                 goto out;
1982         __split_huge_pud_locked(vma, pud, range.start);
1983
1984 out:
1985         spin_unlock(ptl);
1986         /*
1987          * No need to double call mmu_notifier->invalidate_range() callback as
1988          * the above pudp_huge_clear_flush_notify() did already call it.
1989          */
1990         mmu_notifier_invalidate_range_only_end(&range);
1991 }
1992 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1993
1994 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1995                 unsigned long haddr, pmd_t *pmd)
1996 {
1997         struct mm_struct *mm = vma->vm_mm;
1998         pgtable_t pgtable;
1999         pmd_t _pmd;
2000         int i;
2001
2002         /*
2003          * Leave pmd empty until pte is filled note that it is fine to delay
2004          * notification until mmu_notifier_invalidate_range_end() as we are
2005          * replacing a zero pmd write protected page with a zero pte write
2006          * protected page.
2007          *
2008          * See Documentation/vm/mmu_notifier.rst
2009          */
2010         pmdp_huge_clear_flush(vma, haddr, pmd);
2011
2012         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2013         pmd_populate(mm, &_pmd, pgtable);
2014
2015         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2016                 pte_t *pte, entry;
2017                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2018                 entry = pte_mkspecial(entry);
2019                 pte = pte_offset_map(&_pmd, haddr);
2020                 VM_BUG_ON(!pte_none(*pte));
2021                 set_pte_at(mm, haddr, pte, entry);
2022                 pte_unmap(pte);
2023         }
2024         smp_wmb(); /* make pte visible before pmd */
2025         pmd_populate(mm, pmd, pgtable);
2026 }
2027
2028 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2029                 unsigned long haddr, bool freeze)
2030 {
2031         struct mm_struct *mm = vma->vm_mm;
2032         struct page *page;
2033         pgtable_t pgtable;
2034         pmd_t old_pmd, _pmd;
2035         bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2036         unsigned long addr;
2037         int i;
2038
2039         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2040         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2041         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2042         VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2043                                 && !pmd_devmap(*pmd));
2044
2045         count_vm_event(THP_SPLIT_PMD);
2046
2047         if (!vma_is_anonymous(vma)) {
2048                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2049                 /*
2050                  * We are going to unmap this huge page. So
2051                  * just go ahead and zap it
2052                  */
2053                 if (arch_needs_pgtable_deposit())
2054                         zap_deposited_table(mm, pmd);
2055                 if (vma_is_special_huge(vma))
2056                         return;
2057                 page = pmd_page(_pmd);
2058                 if (!PageDirty(page) && pmd_dirty(_pmd))
2059                         set_page_dirty(page);
2060                 if (!PageReferenced(page) && pmd_young(_pmd))
2061                         SetPageReferenced(page);
2062                 page_remove_rmap(page, true);
2063                 put_page(page);
2064                 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2065                 return;
2066         } else if (pmd_trans_huge(*pmd) && is_huge_zero_pmd(*pmd)) {
2067                 /*
2068                  * FIXME: Do we want to invalidate secondary mmu by calling
2069                  * mmu_notifier_invalidate_range() see comments below inside
2070                  * __split_huge_pmd() ?
2071                  *
2072                  * We are going from a zero huge page write protected to zero
2073                  * small page also write protected so it does not seems useful
2074                  * to invalidate secondary mmu at this time.
2075                  */
2076                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2077         }
2078
2079         /*
2080          * Up to this point the pmd is present and huge and userland has the
2081          * whole access to the hugepage during the split (which happens in
2082          * place). If we overwrite the pmd with the not-huge version pointing
2083          * to the pte here (which of course we could if all CPUs were bug
2084          * free), userland could trigger a small page size TLB miss on the
2085          * small sized TLB while the hugepage TLB entry is still established in
2086          * the huge TLB. Some CPU doesn't like that.
2087          * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2088          * 383 on page 105. Intel should be safe but is also warns that it's
2089          * only safe if the permission and cache attributes of the two entries
2090          * loaded in the two TLB is identical (which should be the case here).
2091          * But it is generally safer to never allow small and huge TLB entries
2092          * for the same virtual address to be loaded simultaneously. So instead
2093          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2094          * current pmd notpresent (atomically because here the pmd_trans_huge
2095          * must remain set at all times on the pmd until the split is complete
2096          * for this pmd), then we flush the SMP TLB and finally we write the
2097          * non-huge version of the pmd entry with pmd_populate.
2098          */
2099         old_pmd = pmdp_invalidate(vma, haddr, pmd);
2100
2101         pmd_migration = is_pmd_migration_entry(old_pmd);
2102         if (unlikely(pmd_migration)) {
2103                 swp_entry_t entry;
2104
2105                 entry = pmd_to_swp_entry(old_pmd);
2106                 page = pfn_to_page(swp_offset(entry));
2107                 write = is_write_migration_entry(entry);
2108                 young = false;
2109                 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2110                 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2111         } else {
2112                 page = pmd_page(old_pmd);
2113                 if (pmd_dirty(old_pmd))
2114                         SetPageDirty(page);
2115                 write = pmd_write(old_pmd);
2116                 young = pmd_young(old_pmd);
2117                 soft_dirty = pmd_soft_dirty(old_pmd);
2118                 uffd_wp = pmd_uffd_wp(old_pmd);
2119         }
2120         VM_BUG_ON_PAGE(!page_count(page), page);
2121         page_ref_add(page, HPAGE_PMD_NR - 1);
2122
2123         /*
2124          * Withdraw the table only after we mark the pmd entry invalid.
2125          * This's critical for some architectures (Power).
2126          */
2127         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2128         pmd_populate(mm, &_pmd, pgtable);
2129
2130         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2131                 pte_t entry, *pte;
2132                 /*
2133                  * Note that NUMA hinting access restrictions are not
2134                  * transferred to avoid any possibility of altering
2135                  * permissions across VMAs.
2136                  */
2137                 if (freeze || pmd_migration) {
2138                         swp_entry_t swp_entry;
2139                         swp_entry = make_migration_entry(page + i, write);
2140                         entry = swp_entry_to_pte(swp_entry);
2141                         if (soft_dirty)
2142                                 entry = pte_swp_mksoft_dirty(entry);
2143                         if (uffd_wp)
2144                                 entry = pte_swp_mkuffd_wp(entry);
2145                 } else {
2146                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2147                         entry = maybe_mkwrite(entry, vma);
2148                         if (!write)
2149                                 entry = pte_wrprotect(entry);
2150                         if (!young)
2151                                 entry = pte_mkold(entry);
2152                         if (soft_dirty)
2153                                 entry = pte_mksoft_dirty(entry);
2154                         if (uffd_wp)
2155                                 entry = pte_mkuffd_wp(entry);
2156                 }
2157                 pte = pte_offset_map(&_pmd, addr);
2158                 BUG_ON(!pte_none(*pte));
2159                 set_pte_at(mm, addr, pte, entry);
2160                 if (!pmd_migration)
2161                         atomic_inc(&page[i]._mapcount);
2162                 pte_unmap(pte);
2163         }
2164
2165         if (!pmd_migration) {
2166                 /*
2167                  * Set PG_double_map before dropping compound_mapcount to avoid
2168                  * false-negative page_mapped().
2169                  */
2170                 if (compound_mapcount(page) > 1 &&
2171                     !TestSetPageDoubleMap(page)) {
2172                         for (i = 0; i < HPAGE_PMD_NR; i++)
2173                                 atomic_inc(&page[i]._mapcount);
2174                 }
2175
2176                 lock_page_memcg(page);
2177                 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2178                         /* Last compound_mapcount is gone. */
2179                         __dec_lruvec_page_state(page, NR_ANON_THPS);
2180                         if (TestClearPageDoubleMap(page)) {
2181                                 /* No need in mapcount reference anymore */
2182                                 for (i = 0; i < HPAGE_PMD_NR; i++)
2183                                         atomic_dec(&page[i]._mapcount);
2184                         }
2185                 }
2186                 unlock_page_memcg(page);
2187         }
2188
2189         smp_wmb(); /* make pte visible before pmd */
2190         pmd_populate(mm, pmd, pgtable);
2191
2192         if (freeze) {
2193                 for (i = 0; i < HPAGE_PMD_NR; i++) {
2194                         page_remove_rmap(page + i, false);
2195                         put_page(page + i);
2196                 }
2197         }
2198 }
2199
2200 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2201                 unsigned long address, bool freeze, struct page *page)
2202 {
2203         spinlock_t *ptl;
2204         struct mmu_notifier_range range;
2205         bool was_locked = false;
2206         pmd_t _pmd;
2207
2208         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2209                                 address & HPAGE_PMD_MASK,
2210                                 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2211         mmu_notifier_invalidate_range_start(&range);
2212         ptl = pmd_lock(vma->vm_mm, pmd);
2213
2214         /*
2215          * If caller asks to setup a migration entries, we need a page to check
2216          * pmd against. Otherwise we can end up replacing wrong page.
2217          */
2218         VM_BUG_ON(freeze && !page);
2219         if (page) {
2220                 VM_WARN_ON_ONCE(!PageLocked(page));
2221                 was_locked = true;
2222                 if (page != pmd_page(*pmd))
2223                         goto out;
2224         }
2225
2226 repeat:
2227         if (pmd_trans_huge(*pmd)) {
2228                 if (!page) {
2229                         page = pmd_page(*pmd);
2230                         if (unlikely(!trylock_page(page))) {
2231                                 get_page(page);
2232                                 _pmd = *pmd;
2233                                 spin_unlock(ptl);
2234                                 lock_page(page);
2235                                 spin_lock(ptl);
2236                                 if (unlikely(!pmd_same(*pmd, _pmd))) {
2237                                         unlock_page(page);
2238                                         put_page(page);
2239                                         page = NULL;
2240                                         goto repeat;
2241                                 }
2242                                 put_page(page);
2243                         }
2244                 }
2245                 if (PageMlocked(page))
2246                         clear_page_mlock(page);
2247         } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2248                 goto out;
2249         __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2250 out:
2251         spin_unlock(ptl);
2252         if (!was_locked && page)
2253                 unlock_page(page);
2254         /*
2255          * No need to double call mmu_notifier->invalidate_range() callback.
2256          * They are 3 cases to consider inside __split_huge_pmd_locked():
2257          *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2258          *  2) __split_huge_zero_page_pmd() read only zero page and any write
2259          *    fault will trigger a flush_notify before pointing to a new page
2260          *    (it is fine if the secondary mmu keeps pointing to the old zero
2261          *    page in the meantime)
2262          *  3) Split a huge pmd into pte pointing to the same page. No need
2263          *     to invalidate secondary tlb entry they are all still valid.
2264          *     any further changes to individual pte will notify. So no need
2265          *     to call mmu_notifier->invalidate_range()
2266          */
2267         mmu_notifier_invalidate_range_only_end(&range);
2268 }
2269
2270 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2271                 bool freeze, struct page *page)
2272 {
2273         pgd_t *pgd;
2274         p4d_t *p4d;
2275         pud_t *pud;
2276         pmd_t *pmd;
2277
2278         pgd = pgd_offset(vma->vm_mm, address);
2279         if (!pgd_present(*pgd))
2280                 return;
2281
2282         p4d = p4d_offset(pgd, address);
2283         if (!p4d_present(*p4d))
2284                 return;
2285
2286         pud = pud_offset(p4d, address);
2287         if (!pud_present(*pud))
2288                 return;
2289
2290         pmd = pmd_offset(pud, address);
2291
2292         __split_huge_pmd(vma, pmd, address, freeze, page);
2293 }
2294
2295 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2296                              unsigned long start,
2297                              unsigned long end,
2298                              long adjust_next)
2299 {
2300         /*
2301          * If the new start address isn't hpage aligned and it could
2302          * previously contain an hugepage: check if we need to split
2303          * an huge pmd.
2304          */
2305         if (start & ~HPAGE_PMD_MASK &&
2306             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2307             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2308                 split_huge_pmd_address(vma, start, false, NULL);
2309
2310         /*
2311          * If the new end address isn't hpage aligned and it could
2312          * previously contain an hugepage: check if we need to split
2313          * an huge pmd.
2314          */
2315         if (end & ~HPAGE_PMD_MASK &&
2316             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2317             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2318                 split_huge_pmd_address(vma, end, false, NULL);
2319
2320         /*
2321          * If we're also updating the vma->vm_next->vm_start, if the new
2322          * vm_next->vm_start isn't hpage aligned and it could previously
2323          * contain an hugepage: check if we need to split an huge pmd.
2324          */
2325         if (adjust_next > 0) {
2326                 struct vm_area_struct *next = vma->vm_next;
2327                 unsigned long nstart = next->vm_start;
2328                 nstart += adjust_next;
2329                 if (nstart & ~HPAGE_PMD_MASK &&
2330                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2331                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2332                         split_huge_pmd_address(next, nstart, false, NULL);
2333         }
2334 }
2335
2336 static void unmap_page(struct page *page)
2337 {
2338         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK |
2339                 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2340         bool unmap_success;
2341
2342         VM_BUG_ON_PAGE(!PageHead(page), page);
2343
2344         if (PageAnon(page))
2345                 ttu_flags |= TTU_SPLIT_FREEZE;
2346
2347         unmap_success = try_to_unmap(page, ttu_flags);
2348         VM_BUG_ON_PAGE(!unmap_success, page);
2349 }
2350
2351 static void remap_page(struct page *page, unsigned int nr)
2352 {
2353         int i;
2354         if (PageTransHuge(page)) {
2355                 remove_migration_ptes(page, page, true);
2356         } else {
2357                 for (i = 0; i < nr; i++)
2358                         remove_migration_ptes(page + i, page + i, true);
2359         }
2360 }
2361
2362 static void lru_add_page_tail(struct page *head, struct page *tail,
2363                 struct lruvec *lruvec, struct list_head *list)
2364 {
2365         VM_BUG_ON_PAGE(!PageHead(head), head);
2366         VM_BUG_ON_PAGE(PageCompound(tail), head);
2367         VM_BUG_ON_PAGE(PageLRU(tail), head);
2368         lockdep_assert_held(&lruvec->lru_lock);
2369
2370         if (list) {
2371                 /* page reclaim is reclaiming a huge page */
2372                 VM_WARN_ON(PageLRU(head));
2373                 get_page(tail);
2374                 list_add_tail(&tail->lru, list);
2375         } else {
2376                 /* head is still on lru (and we have it frozen) */
2377                 VM_WARN_ON(!PageLRU(head));
2378                 SetPageLRU(tail);
2379                 list_add_tail(&tail->lru, &head->lru);
2380         }
2381 }
2382
2383 static void __split_huge_page_tail(struct page *head, int tail,
2384                 struct lruvec *lruvec, struct list_head *list)
2385 {
2386         struct page *page_tail = head + tail;
2387
2388         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2389
2390         /*
2391          * Clone page flags before unfreezing refcount.
2392          *
2393          * After successful get_page_unless_zero() might follow flags change,
2394          * for example lock_page() which set PG_waiters.
2395          */
2396         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2397         page_tail->flags |= (head->flags &
2398                         ((1L << PG_referenced) |
2399                          (1L << PG_swapbacked) |
2400                          (1L << PG_swapcache) |
2401                          (1L << PG_mlocked) |
2402                          (1L << PG_uptodate) |
2403                          (1L << PG_active) |
2404                          (1L << PG_workingset) |
2405                          (1L << PG_locked) |
2406                          (1L << PG_unevictable) |
2407 #ifdef CONFIG_64BIT
2408                          (1L << PG_arch_2) |
2409 #endif
2410                          (1L << PG_dirty)));
2411
2412         /* ->mapping in first tail page is compound_mapcount */
2413         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2414                         page_tail);
2415         page_tail->mapping = head->mapping;
2416         page_tail->index = head->index + tail;
2417
2418         /* Page flags must be visible before we make the page non-compound. */
2419         smp_wmb();
2420
2421         /*
2422          * Clear PageTail before unfreezing page refcount.
2423          *
2424          * After successful get_page_unless_zero() might follow put_page()
2425          * which needs correct compound_head().
2426          */
2427         clear_compound_head(page_tail);
2428
2429         /* Finally unfreeze refcount. Additional reference from page cache. */
2430         page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2431                                           PageSwapCache(head)));
2432
2433         if (page_is_young(head))
2434                 set_page_young(page_tail);
2435         if (page_is_idle(head))
2436                 set_page_idle(page_tail);
2437
2438         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2439
2440         /*
2441          * always add to the tail because some iterators expect new
2442          * pages to show after the currently processed elements - e.g.
2443          * migrate_pages
2444          */
2445         lru_add_page_tail(head, page_tail, lruvec, list);
2446 }
2447
2448 static void __split_huge_page(struct page *page, struct list_head *list,
2449                 pgoff_t end)
2450 {
2451         struct page *head = compound_head(page);
2452         struct lruvec *lruvec;
2453         struct address_space *swap_cache = NULL;
2454         unsigned long offset = 0;
2455         unsigned int nr = thp_nr_pages(head);
2456         int i;
2457
2458         /* complete memcg works before add pages to LRU */
2459         mem_cgroup_split_huge_fixup(head);
2460
2461         if (PageAnon(head) && PageSwapCache(head)) {
2462                 swp_entry_t entry = { .val = page_private(head) };
2463
2464                 offset = swp_offset(entry);
2465                 swap_cache = swap_address_space(entry);
2466                 xa_lock(&swap_cache->i_pages);
2467         }
2468
2469         /* lock lru list/PageCompound, ref freezed by page_ref_freeze */
2470         lruvec = lock_page_lruvec(head);
2471
2472         for (i = nr - 1; i >= 1; i--) {
2473                 __split_huge_page_tail(head, i, lruvec, list);
2474                 /* Some pages can be beyond i_size: drop them from page cache */
2475                 if (head[i].index >= end) {
2476                         ClearPageDirty(head + i);
2477                         __delete_from_page_cache(head + i, NULL);
2478                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2479                                 shmem_uncharge(head->mapping->host, 1);
2480                         put_page(head + i);
2481                 } else if (!PageAnon(page)) {
2482                         __xa_store(&head->mapping->i_pages, head[i].index,
2483                                         head + i, 0);
2484                 } else if (swap_cache) {
2485                         __xa_store(&swap_cache->i_pages, offset + i,
2486                                         head + i, 0);
2487                 }
2488         }
2489
2490         ClearPageCompound(head);
2491         unlock_page_lruvec(lruvec);
2492         /* Caller disabled irqs, so they are still disabled here */
2493
2494         split_page_owner(head, nr);
2495
2496         /* See comment in __split_huge_page_tail() */
2497         if (PageAnon(head)) {
2498                 /* Additional pin to swap cache */
2499                 if (PageSwapCache(head)) {
2500                         page_ref_add(head, 2);
2501                         xa_unlock(&swap_cache->i_pages);
2502                 } else {
2503                         page_ref_inc(head);
2504                 }
2505         } else {
2506                 /* Additional pin to page cache */
2507                 page_ref_add(head, 2);
2508                 xa_unlock(&head->mapping->i_pages);
2509         }
2510         local_irq_enable();
2511
2512         remap_page(head, nr);
2513
2514         if (PageSwapCache(head)) {
2515                 swp_entry_t entry = { .val = page_private(head) };
2516
2517                 split_swap_cluster(entry);
2518         }
2519
2520         for (i = 0; i < nr; i++) {
2521                 struct page *subpage = head + i;
2522                 if (subpage == page)
2523                         continue;
2524                 unlock_page(subpage);
2525
2526                 /*
2527                  * Subpages may be freed if there wasn't any mapping
2528                  * like if add_to_swap() is running on a lru page that
2529                  * had its mapping zapped. And freeing these pages
2530                  * requires taking the lru_lock so we do the put_page
2531                  * of the tail pages after the split is complete.
2532                  */
2533                 put_page(subpage);
2534         }
2535 }
2536
2537 int total_mapcount(struct page *page)
2538 {
2539         int i, compound, nr, ret;
2540
2541         VM_BUG_ON_PAGE(PageTail(page), page);
2542
2543         if (likely(!PageCompound(page)))
2544                 return atomic_read(&page->_mapcount) + 1;
2545
2546         compound = compound_mapcount(page);
2547         nr = compound_nr(page);
2548         if (PageHuge(page))
2549                 return compound;
2550         ret = compound;
2551         for (i = 0; i < nr; i++)
2552                 ret += atomic_read(&page[i]._mapcount) + 1;
2553         /* File pages has compound_mapcount included in _mapcount */
2554         if (!PageAnon(page))
2555                 return ret - compound * nr;
2556         if (PageDoubleMap(page))
2557                 ret -= nr;
2558         return ret;
2559 }
2560
2561 /*
2562  * This calculates accurately how many mappings a transparent hugepage
2563  * has (unlike page_mapcount() which isn't fully accurate). This full
2564  * accuracy is primarily needed to know if copy-on-write faults can
2565  * reuse the page and change the mapping to read-write instead of
2566  * copying them. At the same time this returns the total_mapcount too.
2567  *
2568  * The function returns the highest mapcount any one of the subpages
2569  * has. If the return value is one, even if different processes are
2570  * mapping different subpages of the transparent hugepage, they can
2571  * all reuse it, because each process is reusing a different subpage.
2572  *
2573  * The total_mapcount is instead counting all virtual mappings of the
2574  * subpages. If the total_mapcount is equal to "one", it tells the
2575  * caller all mappings belong to the same "mm" and in turn the
2576  * anon_vma of the transparent hugepage can become the vma->anon_vma
2577  * local one as no other process may be mapping any of the subpages.
2578  *
2579  * It would be more accurate to replace page_mapcount() with
2580  * page_trans_huge_mapcount(), however we only use
2581  * page_trans_huge_mapcount() in the copy-on-write faults where we
2582  * need full accuracy to avoid breaking page pinning, because
2583  * page_trans_huge_mapcount() is slower than page_mapcount().
2584  */
2585 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2586 {
2587         int i, ret, _total_mapcount, mapcount;
2588
2589         /* hugetlbfs shouldn't call it */
2590         VM_BUG_ON_PAGE(PageHuge(page), page);
2591
2592         if (likely(!PageTransCompound(page))) {
2593                 mapcount = atomic_read(&page->_mapcount) + 1;
2594                 if (total_mapcount)
2595                         *total_mapcount = mapcount;
2596                 return mapcount;
2597         }
2598
2599         page = compound_head(page);
2600
2601         _total_mapcount = ret = 0;
2602         for (i = 0; i < thp_nr_pages(page); i++) {
2603                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2604                 ret = max(ret, mapcount);
2605                 _total_mapcount += mapcount;
2606         }
2607         if (PageDoubleMap(page)) {
2608                 ret -= 1;
2609                 _total_mapcount -= thp_nr_pages(page);
2610         }
2611         mapcount = compound_mapcount(page);
2612         ret += mapcount;
2613         _total_mapcount += mapcount;
2614         if (total_mapcount)
2615                 *total_mapcount = _total_mapcount;
2616         return ret;
2617 }
2618
2619 /* Racy check whether the huge page can be split */
2620 bool can_split_huge_page(struct page *page, int *pextra_pins)
2621 {
2622         int extra_pins;
2623
2624         /* Additional pins from page cache */
2625         if (PageAnon(page))
2626                 extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0;
2627         else
2628                 extra_pins = thp_nr_pages(page);
2629         if (pextra_pins)
2630                 *pextra_pins = extra_pins;
2631         return total_mapcount(page) == page_count(page) - extra_pins - 1;
2632 }
2633
2634 /*
2635  * This function splits huge page into normal pages. @page can point to any
2636  * subpage of huge page to split. Split doesn't change the position of @page.
2637  *
2638  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2639  * The huge page must be locked.
2640  *
2641  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2642  *
2643  * Both head page and tail pages will inherit mapping, flags, and so on from
2644  * the hugepage.
2645  *
2646  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2647  * they are not mapped.
2648  *
2649  * Returns 0 if the hugepage is split successfully.
2650  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2651  * us.
2652  */
2653 int split_huge_page_to_list(struct page *page, struct list_head *list)
2654 {
2655         struct page *head = compound_head(page);
2656         struct deferred_split *ds_queue = get_deferred_split_queue(head);
2657         struct anon_vma *anon_vma = NULL;
2658         struct address_space *mapping = NULL;
2659         int count, mapcount, extra_pins, ret;
2660         pgoff_t end;
2661
2662         VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2663         VM_BUG_ON_PAGE(!PageLocked(head), head);
2664         VM_BUG_ON_PAGE(!PageCompound(head), head);
2665
2666         if (PageWriteback(head))
2667                 return -EBUSY;
2668
2669         if (PageAnon(head)) {
2670                 /*
2671                  * The caller does not necessarily hold an mmap_lock that would
2672                  * prevent the anon_vma disappearing so we first we take a
2673                  * reference to it and then lock the anon_vma for write. This
2674                  * is similar to page_lock_anon_vma_read except the write lock
2675                  * is taken to serialise against parallel split or collapse
2676                  * operations.
2677                  */
2678                 anon_vma = page_get_anon_vma(head);
2679                 if (!anon_vma) {
2680                         ret = -EBUSY;
2681                         goto out;
2682                 }
2683                 end = -1;
2684                 mapping = NULL;
2685                 anon_vma_lock_write(anon_vma);
2686         } else {
2687                 mapping = head->mapping;
2688
2689                 /* Truncated ? */
2690                 if (!mapping) {
2691                         ret = -EBUSY;
2692                         goto out;
2693                 }
2694
2695                 anon_vma = NULL;
2696                 i_mmap_lock_read(mapping);
2697
2698                 /*
2699                  *__split_huge_page() may need to trim off pages beyond EOF:
2700                  * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2701                  * which cannot be nested inside the page tree lock. So note
2702                  * end now: i_size itself may be changed at any moment, but
2703                  * head page lock is good enough to serialize the trimming.
2704                  */
2705                 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2706         }
2707
2708         /*
2709          * Racy check if we can split the page, before unmap_page() will
2710          * split PMDs
2711          */
2712         if (!can_split_huge_page(head, &extra_pins)) {
2713                 ret = -EBUSY;
2714                 goto out_unlock;
2715         }
2716
2717         unmap_page(head);
2718         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2719
2720         /* block interrupt reentry in xa_lock and spinlock */
2721         local_irq_disable();
2722         if (mapping) {
2723                 XA_STATE(xas, &mapping->i_pages, page_index(head));
2724
2725                 /*
2726                  * Check if the head page is present in page cache.
2727                  * We assume all tail are present too, if head is there.
2728                  */
2729                 xa_lock(&mapping->i_pages);
2730                 if (xas_load(&xas) != head)
2731                         goto fail;
2732         }
2733
2734         /* Prevent deferred_split_scan() touching ->_refcount */
2735         spin_lock(&ds_queue->split_queue_lock);
2736         count = page_count(head);
2737         mapcount = total_mapcount(head);
2738         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2739                 if (!list_empty(page_deferred_list(head))) {
2740                         ds_queue->split_queue_len--;
2741                         list_del(page_deferred_list(head));
2742                 }
2743                 spin_unlock(&ds_queue->split_queue_lock);
2744                 if (mapping) {
2745                         if (PageSwapBacked(head))
2746                                 __dec_lruvec_page_state(head, NR_SHMEM_THPS);
2747                         else
2748                                 __dec_lruvec_page_state(head, NR_FILE_THPS);
2749                 }
2750
2751                 __split_huge_page(page, list, end);
2752                 ret = 0;
2753         } else {
2754                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2755                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2756                                         mapcount, count);
2757                         if (PageTail(page))
2758                                 dump_page(head, NULL);
2759                         dump_page(page, "total_mapcount(head) > 0");
2760                         BUG();
2761                 }
2762                 spin_unlock(&ds_queue->split_queue_lock);
2763 fail:           if (mapping)
2764                         xa_unlock(&mapping->i_pages);
2765                 local_irq_enable();
2766                 remap_page(head, thp_nr_pages(head));
2767                 ret = -EBUSY;
2768         }
2769
2770 out_unlock:
2771         if (anon_vma) {
2772                 anon_vma_unlock_write(anon_vma);
2773                 put_anon_vma(anon_vma);
2774         }
2775         if (mapping)
2776                 i_mmap_unlock_read(mapping);
2777 out:
2778         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2779         return ret;
2780 }
2781
2782 void free_transhuge_page(struct page *page)
2783 {
2784         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2785         unsigned long flags;
2786
2787         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2788         if (!list_empty(page_deferred_list(page))) {
2789                 ds_queue->split_queue_len--;
2790                 list_del(page_deferred_list(page));
2791         }
2792         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2793         free_compound_page(page);
2794 }
2795
2796 void deferred_split_huge_page(struct page *page)
2797 {
2798         struct deferred_split *ds_queue = get_deferred_split_queue(page);
2799 #ifdef CONFIG_MEMCG
2800         struct mem_cgroup *memcg = page_memcg(compound_head(page));
2801 #endif
2802         unsigned long flags;
2803
2804         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2805
2806         /*
2807          * The try_to_unmap() in page reclaim path might reach here too,
2808          * this may cause a race condition to corrupt deferred split queue.
2809          * And, if page reclaim is already handling the same page, it is
2810          * unnecessary to handle it again in shrinker.
2811          *
2812          * Check PageSwapCache to determine if the page is being
2813          * handled by page reclaim since THP swap would add the page into
2814          * swap cache before calling try_to_unmap().
2815          */
2816         if (PageSwapCache(page))
2817                 return;
2818
2819         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2820         if (list_empty(page_deferred_list(page))) {
2821                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2822                 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2823                 ds_queue->split_queue_len++;
2824 #ifdef CONFIG_MEMCG
2825                 if (memcg)
2826                         memcg_set_shrinker_bit(memcg, page_to_nid(page),
2827                                                deferred_split_shrinker.id);
2828 #endif
2829         }
2830         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2831 }
2832
2833 static unsigned long deferred_split_count(struct shrinker *shrink,
2834                 struct shrink_control *sc)
2835 {
2836         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2837         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2838
2839 #ifdef CONFIG_MEMCG
2840         if (sc->memcg)
2841                 ds_queue = &sc->memcg->deferred_split_queue;
2842 #endif
2843         return READ_ONCE(ds_queue->split_queue_len);
2844 }
2845
2846 static unsigned long deferred_split_scan(struct shrinker *shrink,
2847                 struct shrink_control *sc)
2848 {
2849         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2850         struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2851         unsigned long flags;
2852         LIST_HEAD(list), *pos, *next;
2853         struct page *page;
2854         int split = 0;
2855
2856 #ifdef CONFIG_MEMCG
2857         if (sc->memcg)
2858                 ds_queue = &sc->memcg->deferred_split_queue;
2859 #endif
2860
2861         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2862         /* Take pin on all head pages to avoid freeing them under us */
2863         list_for_each_safe(pos, next, &ds_queue->split_queue) {
2864                 page = list_entry((void *)pos, struct page, mapping);
2865                 page = compound_head(page);
2866                 if (get_page_unless_zero(page)) {
2867                         list_move(page_deferred_list(page), &list);
2868                 } else {
2869                         /* We lost race with put_compound_page() */
2870                         list_del_init(page_deferred_list(page));
2871                         ds_queue->split_queue_len--;
2872                 }
2873                 if (!--sc->nr_to_scan)
2874                         break;
2875         }
2876         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2877
2878         list_for_each_safe(pos, next, &list) {
2879                 page = list_entry((void *)pos, struct page, mapping);
2880                 if (!trylock_page(page))
2881                         goto next;
2882                 /* split_huge_page() removes page from list on success */
2883                 if (!split_huge_page(page))
2884                         split++;
2885                 unlock_page(page);
2886 next:
2887                 put_page(page);
2888         }
2889
2890         spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2891         list_splice_tail(&list, &ds_queue->split_queue);
2892         spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2893
2894         /*
2895          * Stop shrinker if we didn't split any page, but the queue is empty.
2896          * This can happen if pages were freed under us.
2897          */
2898         if (!split && list_empty(&ds_queue->split_queue))
2899                 return SHRINK_STOP;
2900         return split;
2901 }
2902
2903 static struct shrinker deferred_split_shrinker = {
2904         .count_objects = deferred_split_count,
2905         .scan_objects = deferred_split_scan,
2906         .seeks = DEFAULT_SEEKS,
2907         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2908                  SHRINKER_NONSLAB,
2909 };
2910
2911 #ifdef CONFIG_DEBUG_FS
2912 static int split_huge_pages_set(void *data, u64 val)
2913 {
2914         struct zone *zone;
2915         struct page *page;
2916         unsigned long pfn, max_zone_pfn;
2917         unsigned long total = 0, split = 0;
2918
2919         if (val != 1)
2920                 return -EINVAL;
2921
2922         for_each_populated_zone(zone) {
2923                 max_zone_pfn = zone_end_pfn(zone);
2924                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2925                         if (!pfn_valid(pfn))
2926                                 continue;
2927
2928                         page = pfn_to_page(pfn);
2929                         if (!get_page_unless_zero(page))
2930                                 continue;
2931
2932                         if (zone != page_zone(page))
2933                                 goto next;
2934
2935                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2936                                 goto next;
2937
2938                         total++;
2939                         lock_page(page);
2940                         if (!split_huge_page(page))
2941                                 split++;
2942                         unlock_page(page);
2943 next:
2944                         put_page(page);
2945                 }
2946         }
2947
2948         pr_info("%lu of %lu THP split\n", split, total);
2949
2950         return 0;
2951 }
2952 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2953                 "%llu\n");
2954
2955 static int __init split_huge_pages_debugfs(void)
2956 {
2957         debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2958                             &split_huge_pages_fops);
2959         return 0;
2960 }
2961 late_initcall(split_huge_pages_debugfs);
2962 #endif
2963
2964 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2965 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2966                 struct page *page)
2967 {
2968         struct vm_area_struct *vma = pvmw->vma;
2969         struct mm_struct *mm = vma->vm_mm;
2970         unsigned long address = pvmw->address;
2971         pmd_t pmdval;
2972         swp_entry_t entry;
2973         pmd_t pmdswp;
2974
2975         if (!(pvmw->pmd && !pvmw->pte))
2976                 return;
2977
2978         flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2979         pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
2980         if (pmd_dirty(pmdval))
2981                 set_page_dirty(page);
2982         entry = make_migration_entry(page, pmd_write(pmdval));
2983         pmdswp = swp_entry_to_pmd(entry);
2984         if (pmd_soft_dirty(pmdval))
2985                 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2986         set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2987         page_remove_rmap(page, true);
2988         put_page(page);
2989 }
2990
2991 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2992 {
2993         struct vm_area_struct *vma = pvmw->vma;
2994         struct mm_struct *mm = vma->vm_mm;
2995         unsigned long address = pvmw->address;
2996         unsigned long mmun_start = address & HPAGE_PMD_MASK;
2997         pmd_t pmde;
2998         swp_entry_t entry;
2999
3000         if (!(pvmw->pmd && !pvmw->pte))
3001                 return;
3002
3003         entry = pmd_to_swp_entry(*pvmw->pmd);
3004         get_page(new);
3005         pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3006         if (pmd_swp_soft_dirty(*pvmw->pmd))
3007                 pmde = pmd_mksoft_dirty(pmde);
3008         if (is_write_migration_entry(entry))
3009                 pmde = maybe_pmd_mkwrite(pmde, vma);
3010
3011         flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3012         if (PageAnon(new))
3013                 page_add_anon_rmap(new, vma, mmun_start, true);
3014         else
3015                 page_add_file_rmap(new, true);
3016         set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3017         if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3018                 mlock_vma_page(new);
3019         update_mmu_cache_pmd(vma, address, pvmw->pmd);
3020 }
3021 #endif