IB/mlx5: Add additional checks before processing MADs
[platform/kernel/linux-rpi.git] / mm / huge_memory.c
1 /*
2  *  Copyright (C) 2009  Red Hat, Inc.
3  *
4  *  This work is licensed under the terms of the GNU GPL, version 2. See
5  *  the COPYING file in the top-level directory.
6  */
7
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10 #include <linux/mm.h>
11 #include <linux/sched.h>
12 #include <linux/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
34 #include <asm/tlb.h>
35 #include <asm/pgalloc.h>
36 #include "internal.h"
37
38 /*
39  * By default transparent hugepage support is disabled in order that avoid
40  * to risk increase the memory footprint of applications without a guaranteed
41  * benefit. When transparent hugepage support is enabled, is for all mappings,
42  * and khugepaged scans all mappings.
43  * Defrag is invoked by khugepaged hugepage allocations and by page faults
44  * for all hugepage allocations.
45  */
46 unsigned long transparent_hugepage_flags __read_mostly =
47 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
48         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
49 #endif
50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
51         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
52 #endif
53         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
54         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
55         (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
56
57 static struct shrinker deferred_split_shrinker;
58
59 static atomic_t huge_zero_refcount;
60 struct page *huge_zero_page __read_mostly;
61
62 static struct page *get_huge_zero_page(void)
63 {
64         struct page *zero_page;
65 retry:
66         if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
67                 return READ_ONCE(huge_zero_page);
68
69         zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
70                         HPAGE_PMD_ORDER);
71         if (!zero_page) {
72                 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
73                 return NULL;
74         }
75         count_vm_event(THP_ZERO_PAGE_ALLOC);
76         preempt_disable();
77         if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
78                 preempt_enable();
79                 __free_pages(zero_page, compound_order(zero_page));
80                 goto retry;
81         }
82
83         /* We take additional reference here. It will be put back by shrinker */
84         atomic_set(&huge_zero_refcount, 2);
85         preempt_enable();
86         return READ_ONCE(huge_zero_page);
87 }
88
89 static void put_huge_zero_page(void)
90 {
91         /*
92          * Counter should never go to zero here. Only shrinker can put
93          * last reference.
94          */
95         BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
96 }
97
98 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
99 {
100         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
101                 return READ_ONCE(huge_zero_page);
102
103         if (!get_huge_zero_page())
104                 return NULL;
105
106         if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
107                 put_huge_zero_page();
108
109         return READ_ONCE(huge_zero_page);
110 }
111
112 void mm_put_huge_zero_page(struct mm_struct *mm)
113 {
114         if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
115                 put_huge_zero_page();
116 }
117
118 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
119                                         struct shrink_control *sc)
120 {
121         /* we can free zero page only if last reference remains */
122         return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
123 }
124
125 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
126                                        struct shrink_control *sc)
127 {
128         if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
129                 struct page *zero_page = xchg(&huge_zero_page, NULL);
130                 BUG_ON(zero_page == NULL);
131                 __free_pages(zero_page, compound_order(zero_page));
132                 return HPAGE_PMD_NR;
133         }
134
135         return 0;
136 }
137
138 static struct shrinker huge_zero_page_shrinker = {
139         .count_objects = shrink_huge_zero_page_count,
140         .scan_objects = shrink_huge_zero_page_scan,
141         .seeks = DEFAULT_SEEKS,
142 };
143
144 #ifdef CONFIG_SYSFS
145
146 static ssize_t triple_flag_store(struct kobject *kobj,
147                                  struct kobj_attribute *attr,
148                                  const char *buf, size_t count,
149                                  enum transparent_hugepage_flag enabled,
150                                  enum transparent_hugepage_flag deferred,
151                                  enum transparent_hugepage_flag req_madv)
152 {
153         if (!memcmp("defer", buf,
154                     min(sizeof("defer")-1, count))) {
155                 if (enabled == deferred)
156                         return -EINVAL;
157                 clear_bit(enabled, &transparent_hugepage_flags);
158                 clear_bit(req_madv, &transparent_hugepage_flags);
159                 set_bit(deferred, &transparent_hugepage_flags);
160         } else if (!memcmp("always", buf,
161                     min(sizeof("always")-1, count))) {
162                 clear_bit(deferred, &transparent_hugepage_flags);
163                 clear_bit(req_madv, &transparent_hugepage_flags);
164                 set_bit(enabled, &transparent_hugepage_flags);
165         } else if (!memcmp("madvise", buf,
166                            min(sizeof("madvise")-1, count))) {
167                 clear_bit(enabled, &transparent_hugepage_flags);
168                 clear_bit(deferred, &transparent_hugepage_flags);
169                 set_bit(req_madv, &transparent_hugepage_flags);
170         } else if (!memcmp("never", buf,
171                            min(sizeof("never")-1, count))) {
172                 clear_bit(enabled, &transparent_hugepage_flags);
173                 clear_bit(req_madv, &transparent_hugepage_flags);
174                 clear_bit(deferred, &transparent_hugepage_flags);
175         } else
176                 return -EINVAL;
177
178         return count;
179 }
180
181 static ssize_t enabled_show(struct kobject *kobj,
182                             struct kobj_attribute *attr, char *buf)
183 {
184         if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
185                 return sprintf(buf, "[always] madvise never\n");
186         else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
187                 return sprintf(buf, "always [madvise] never\n");
188         else
189                 return sprintf(buf, "always madvise [never]\n");
190 }
191
192 static ssize_t enabled_store(struct kobject *kobj,
193                              struct kobj_attribute *attr,
194                              const char *buf, size_t count)
195 {
196         ssize_t ret;
197
198         ret = triple_flag_store(kobj, attr, buf, count,
199                                 TRANSPARENT_HUGEPAGE_FLAG,
200                                 TRANSPARENT_HUGEPAGE_FLAG,
201                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
202
203         if (ret > 0) {
204                 int err = start_stop_khugepaged();
205                 if (err)
206                         ret = err;
207         }
208
209         return ret;
210 }
211 static struct kobj_attribute enabled_attr =
212         __ATTR(enabled, 0644, enabled_show, enabled_store);
213
214 ssize_t single_hugepage_flag_show(struct kobject *kobj,
215                                 struct kobj_attribute *attr, char *buf,
216                                 enum transparent_hugepage_flag flag)
217 {
218         return sprintf(buf, "%d\n",
219                        !!test_bit(flag, &transparent_hugepage_flags));
220 }
221
222 ssize_t single_hugepage_flag_store(struct kobject *kobj,
223                                  struct kobj_attribute *attr,
224                                  const char *buf, size_t count,
225                                  enum transparent_hugepage_flag flag)
226 {
227         unsigned long value;
228         int ret;
229
230         ret = kstrtoul(buf, 10, &value);
231         if (ret < 0)
232                 return ret;
233         if (value > 1)
234                 return -EINVAL;
235
236         if (value)
237                 set_bit(flag, &transparent_hugepage_flags);
238         else
239                 clear_bit(flag, &transparent_hugepage_flags);
240
241         return count;
242 }
243
244 /*
245  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
246  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
247  * memory just to allocate one more hugepage.
248  */
249 static ssize_t defrag_show(struct kobject *kobj,
250                            struct kobj_attribute *attr, char *buf)
251 {
252         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
253                 return sprintf(buf, "[always] defer madvise never\n");
254         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
255                 return sprintf(buf, "always [defer] madvise never\n");
256         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
257                 return sprintf(buf, "always defer [madvise] never\n");
258         else
259                 return sprintf(buf, "always defer madvise [never]\n");
260
261 }
262 static ssize_t defrag_store(struct kobject *kobj,
263                             struct kobj_attribute *attr,
264                             const char *buf, size_t count)
265 {
266         return triple_flag_store(kobj, attr, buf, count,
267                                  TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
268                                  TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
269                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
270 }
271 static struct kobj_attribute defrag_attr =
272         __ATTR(defrag, 0644, defrag_show, defrag_store);
273
274 static ssize_t use_zero_page_show(struct kobject *kobj,
275                 struct kobj_attribute *attr, char *buf)
276 {
277         return single_hugepage_flag_show(kobj, attr, buf,
278                                 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
279 }
280 static ssize_t use_zero_page_store(struct kobject *kobj,
281                 struct kobj_attribute *attr, const char *buf, size_t count)
282 {
283         return single_hugepage_flag_store(kobj, attr, buf, count,
284                                  TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
285 }
286 static struct kobj_attribute use_zero_page_attr =
287         __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
288
289 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
290                 struct kobj_attribute *attr, char *buf)
291 {
292         return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
293 }
294 static struct kobj_attribute hpage_pmd_size_attr =
295         __ATTR_RO(hpage_pmd_size);
296
297 #ifdef CONFIG_DEBUG_VM
298 static ssize_t debug_cow_show(struct kobject *kobj,
299                                 struct kobj_attribute *attr, char *buf)
300 {
301         return single_hugepage_flag_show(kobj, attr, buf,
302                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
303 }
304 static ssize_t debug_cow_store(struct kobject *kobj,
305                                struct kobj_attribute *attr,
306                                const char *buf, size_t count)
307 {
308         return single_hugepage_flag_store(kobj, attr, buf, count,
309                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
310 }
311 static struct kobj_attribute debug_cow_attr =
312         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
313 #endif /* CONFIG_DEBUG_VM */
314
315 static struct attribute *hugepage_attr[] = {
316         &enabled_attr.attr,
317         &defrag_attr.attr,
318         &use_zero_page_attr.attr,
319         &hpage_pmd_size_attr.attr,
320 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
321         &shmem_enabled_attr.attr,
322 #endif
323 #ifdef CONFIG_DEBUG_VM
324         &debug_cow_attr.attr,
325 #endif
326         NULL,
327 };
328
329 static struct attribute_group hugepage_attr_group = {
330         .attrs = hugepage_attr,
331 };
332
333 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
334 {
335         int err;
336
337         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
338         if (unlikely(!*hugepage_kobj)) {
339                 pr_err("failed to create transparent hugepage kobject\n");
340                 return -ENOMEM;
341         }
342
343         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
344         if (err) {
345                 pr_err("failed to register transparent hugepage group\n");
346                 goto delete_obj;
347         }
348
349         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
350         if (err) {
351                 pr_err("failed to register transparent hugepage group\n");
352                 goto remove_hp_group;
353         }
354
355         return 0;
356
357 remove_hp_group:
358         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
359 delete_obj:
360         kobject_put(*hugepage_kobj);
361         return err;
362 }
363
364 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
365 {
366         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
367         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
368         kobject_put(hugepage_kobj);
369 }
370 #else
371 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
372 {
373         return 0;
374 }
375
376 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
377 {
378 }
379 #endif /* CONFIG_SYSFS */
380
381 static int __init hugepage_init(void)
382 {
383         int err;
384         struct kobject *hugepage_kobj;
385
386         if (!has_transparent_hugepage()) {
387                 transparent_hugepage_flags = 0;
388                 return -EINVAL;
389         }
390
391         /*
392          * hugepages can't be allocated by the buddy allocator
393          */
394         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
395         /*
396          * we use page->mapping and page->index in second tail page
397          * as list_head: assuming THP order >= 2
398          */
399         MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
400
401         err = hugepage_init_sysfs(&hugepage_kobj);
402         if (err)
403                 goto err_sysfs;
404
405         err = khugepaged_init();
406         if (err)
407                 goto err_slab;
408
409         err = register_shrinker(&huge_zero_page_shrinker);
410         if (err)
411                 goto err_hzp_shrinker;
412         err = register_shrinker(&deferred_split_shrinker);
413         if (err)
414                 goto err_split_shrinker;
415
416         /*
417          * By default disable transparent hugepages on smaller systems,
418          * where the extra memory used could hurt more than TLB overhead
419          * is likely to save.  The admin can still enable it through /sys.
420          */
421         if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
422                 transparent_hugepage_flags = 0;
423                 return 0;
424         }
425
426         err = start_stop_khugepaged();
427         if (err)
428                 goto err_khugepaged;
429
430         return 0;
431 err_khugepaged:
432         unregister_shrinker(&deferred_split_shrinker);
433 err_split_shrinker:
434         unregister_shrinker(&huge_zero_page_shrinker);
435 err_hzp_shrinker:
436         khugepaged_destroy();
437 err_slab:
438         hugepage_exit_sysfs(hugepage_kobj);
439 err_sysfs:
440         return err;
441 }
442 subsys_initcall(hugepage_init);
443
444 static int __init setup_transparent_hugepage(char *str)
445 {
446         int ret = 0;
447         if (!str)
448                 goto out;
449         if (!strcmp(str, "always")) {
450                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
451                         &transparent_hugepage_flags);
452                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
453                           &transparent_hugepage_flags);
454                 ret = 1;
455         } else if (!strcmp(str, "madvise")) {
456                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
457                           &transparent_hugepage_flags);
458                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
459                         &transparent_hugepage_flags);
460                 ret = 1;
461         } else if (!strcmp(str, "never")) {
462                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
463                           &transparent_hugepage_flags);
464                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
465                           &transparent_hugepage_flags);
466                 ret = 1;
467         }
468 out:
469         if (!ret)
470                 pr_warn("transparent_hugepage= cannot parse, ignored\n");
471         return ret;
472 }
473 __setup("transparent_hugepage=", setup_transparent_hugepage);
474
475 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
476 {
477         if (likely(vma->vm_flags & VM_WRITE))
478                 pmd = pmd_mkwrite(pmd);
479         return pmd;
480 }
481
482 static inline struct list_head *page_deferred_list(struct page *page)
483 {
484         /*
485          * ->lru in the tail pages is occupied by compound_head.
486          * Let's use ->mapping + ->index in the second tail page as list_head.
487          */
488         return (struct list_head *)&page[2].mapping;
489 }
490
491 void prep_transhuge_page(struct page *page)
492 {
493         /*
494          * we use page->mapping and page->indexlru in second tail page
495          * as list_head: assuming THP order >= 2
496          */
497
498         INIT_LIST_HEAD(page_deferred_list(page));
499         set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
500 }
501
502 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
503                 loff_t off, unsigned long flags, unsigned long size)
504 {
505         unsigned long addr;
506         loff_t off_end = off + len;
507         loff_t off_align = round_up(off, size);
508         unsigned long len_pad;
509
510         if (off_end <= off_align || (off_end - off_align) < size)
511                 return 0;
512
513         len_pad = len + size;
514         if (len_pad < len || (off + len_pad) < off)
515                 return 0;
516
517         addr = current->mm->get_unmapped_area(filp, 0, len_pad,
518                                               off >> PAGE_SHIFT, flags);
519         if (IS_ERR_VALUE(addr))
520                 return 0;
521
522         addr += (off - addr) & (size - 1);
523         return addr;
524 }
525
526 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
527                 unsigned long len, unsigned long pgoff, unsigned long flags)
528 {
529         loff_t off = (loff_t)pgoff << PAGE_SHIFT;
530
531         if (addr)
532                 goto out;
533         if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
534                 goto out;
535
536         addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
537         if (addr)
538                 return addr;
539
540  out:
541         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
542 }
543 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
544
545 static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
546                 gfp_t gfp)
547 {
548         struct vm_area_struct *vma = vmf->vma;
549         struct mem_cgroup *memcg;
550         pgtable_t pgtable;
551         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
552
553         VM_BUG_ON_PAGE(!PageCompound(page), page);
554
555         if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
556                 put_page(page);
557                 count_vm_event(THP_FAULT_FALLBACK);
558                 return VM_FAULT_FALLBACK;
559         }
560
561         pgtable = pte_alloc_one(vma->vm_mm, haddr);
562         if (unlikely(!pgtable)) {
563                 mem_cgroup_cancel_charge(page, memcg, true);
564                 put_page(page);
565                 return VM_FAULT_OOM;
566         }
567
568         clear_huge_page(page, haddr, HPAGE_PMD_NR);
569         /*
570          * The memory barrier inside __SetPageUptodate makes sure that
571          * clear_huge_page writes become visible before the set_pmd_at()
572          * write.
573          */
574         __SetPageUptodate(page);
575
576         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
577         if (unlikely(!pmd_none(*vmf->pmd))) {
578                 spin_unlock(vmf->ptl);
579                 mem_cgroup_cancel_charge(page, memcg, true);
580                 put_page(page);
581                 pte_free(vma->vm_mm, pgtable);
582         } else {
583                 pmd_t entry;
584
585                 /* Deliver the page fault to userland */
586                 if (userfaultfd_missing(vma)) {
587                         int ret;
588
589                         spin_unlock(vmf->ptl);
590                         mem_cgroup_cancel_charge(page, memcg, true);
591                         put_page(page);
592                         pte_free(vma->vm_mm, pgtable);
593                         ret = handle_userfault(vmf, VM_UFFD_MISSING);
594                         VM_BUG_ON(ret & VM_FAULT_FALLBACK);
595                         return ret;
596                 }
597
598                 entry = mk_huge_pmd(page, vma->vm_page_prot);
599                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
600                 page_add_new_anon_rmap(page, vma, haddr, true);
601                 mem_cgroup_commit_charge(page, memcg, false, true);
602                 lru_cache_add_active_or_unevictable(page, vma);
603                 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
604                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
605                 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
606                 atomic_long_inc(&vma->vm_mm->nr_ptes);
607                 spin_unlock(vmf->ptl);
608                 count_vm_event(THP_FAULT_ALLOC);
609         }
610
611         return 0;
612 }
613
614 /*
615  * If THP defrag is set to always then directly reclaim/compact as necessary
616  * If set to defer then do only background reclaim/compact and defer to khugepaged
617  * If set to madvise and the VMA is flagged then directly reclaim/compact
618  * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
619  */
620 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
621 {
622         bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
623
624         if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
625                                 &transparent_hugepage_flags) && vma_madvised)
626                 return GFP_TRANSHUGE;
627         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
628                                                 &transparent_hugepage_flags))
629                 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
630         else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
631                                                 &transparent_hugepage_flags))
632                 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
633
634         return GFP_TRANSHUGE_LIGHT;
635 }
636
637 /* Caller must hold page table lock. */
638 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
639                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
640                 struct page *zero_page)
641 {
642         pmd_t entry;
643         if (!pmd_none(*pmd))
644                 return false;
645         entry = mk_pmd(zero_page, vma->vm_page_prot);
646         entry = pmd_mkhuge(entry);
647         if (pgtable)
648                 pgtable_trans_huge_deposit(mm, pmd, pgtable);
649         set_pmd_at(mm, haddr, pmd, entry);
650         atomic_long_inc(&mm->nr_ptes);
651         return true;
652 }
653
654 int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
655 {
656         struct vm_area_struct *vma = vmf->vma;
657         gfp_t gfp;
658         struct page *page;
659         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
660
661         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
662                 return VM_FAULT_FALLBACK;
663         if (unlikely(anon_vma_prepare(vma)))
664                 return VM_FAULT_OOM;
665         if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
666                 return VM_FAULT_OOM;
667         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
668                         !mm_forbids_zeropage(vma->vm_mm) &&
669                         transparent_hugepage_use_zero_page()) {
670                 pgtable_t pgtable;
671                 struct page *zero_page;
672                 bool set;
673                 int ret;
674                 pgtable = pte_alloc_one(vma->vm_mm, haddr);
675                 if (unlikely(!pgtable))
676                         return VM_FAULT_OOM;
677                 zero_page = mm_get_huge_zero_page(vma->vm_mm);
678                 if (unlikely(!zero_page)) {
679                         pte_free(vma->vm_mm, pgtable);
680                         count_vm_event(THP_FAULT_FALLBACK);
681                         return VM_FAULT_FALLBACK;
682                 }
683                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
684                 ret = 0;
685                 set = false;
686                 if (pmd_none(*vmf->pmd)) {
687                         if (userfaultfd_missing(vma)) {
688                                 spin_unlock(vmf->ptl);
689                                 ret = handle_userfault(vmf, VM_UFFD_MISSING);
690                                 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
691                         } else {
692                                 set_huge_zero_page(pgtable, vma->vm_mm, vma,
693                                                    haddr, vmf->pmd, zero_page);
694                                 spin_unlock(vmf->ptl);
695                                 set = true;
696                         }
697                 } else
698                         spin_unlock(vmf->ptl);
699                 if (!set)
700                         pte_free(vma->vm_mm, pgtable);
701                 return ret;
702         }
703         gfp = alloc_hugepage_direct_gfpmask(vma);
704         page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
705         if (unlikely(!page)) {
706                 count_vm_event(THP_FAULT_FALLBACK);
707                 return VM_FAULT_FALLBACK;
708         }
709         prep_transhuge_page(page);
710         return __do_huge_pmd_anonymous_page(vmf, page, gfp);
711 }
712
713 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
714                 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
715 {
716         struct mm_struct *mm = vma->vm_mm;
717         pmd_t entry;
718         spinlock_t *ptl;
719
720         ptl = pmd_lock(mm, pmd);
721         entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
722         if (pfn_t_devmap(pfn))
723                 entry = pmd_mkdevmap(entry);
724         if (write) {
725                 entry = pmd_mkyoung(pmd_mkdirty(entry));
726                 entry = maybe_pmd_mkwrite(entry, vma);
727         }
728         set_pmd_at(mm, addr, pmd, entry);
729         update_mmu_cache_pmd(vma, addr, pmd);
730         spin_unlock(ptl);
731 }
732
733 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
734                         pmd_t *pmd, pfn_t pfn, bool write)
735 {
736         pgprot_t pgprot = vma->vm_page_prot;
737         /*
738          * If we had pmd_special, we could avoid all these restrictions,
739          * but we need to be consistent with PTEs and architectures that
740          * can't support a 'special' bit.
741          */
742         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
743         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
744                                                 (VM_PFNMAP|VM_MIXEDMAP));
745         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
746         BUG_ON(!pfn_t_devmap(pfn));
747
748         if (addr < vma->vm_start || addr >= vma->vm_end)
749                 return VM_FAULT_SIGBUS;
750
751         track_pfn_insert(vma, &pgprot, pfn);
752
753         insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
754         return VM_FAULT_NOPAGE;
755 }
756 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
757
758 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
759                 pmd_t *pmd)
760 {
761         pmd_t _pmd;
762
763         /*
764          * We should set the dirty bit only for FOLL_WRITE but for now
765          * the dirty bit in the pmd is meaningless.  And if the dirty
766          * bit will become meaningful and we'll only set it with
767          * FOLL_WRITE, an atomic set_bit will be required on the pmd to
768          * set the young bit, instead of the current set_pmd_at.
769          */
770         _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
771         if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
772                                 pmd, _pmd,  1))
773                 update_mmu_cache_pmd(vma, addr, pmd);
774 }
775
776 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
777                 pmd_t *pmd, int flags)
778 {
779         unsigned long pfn = pmd_pfn(*pmd);
780         struct mm_struct *mm = vma->vm_mm;
781         struct dev_pagemap *pgmap;
782         struct page *page;
783
784         assert_spin_locked(pmd_lockptr(mm, pmd));
785
786         if (flags & FOLL_WRITE && !pmd_write(*pmd))
787                 return NULL;
788
789         if (pmd_present(*pmd) && pmd_devmap(*pmd))
790                 /* pass */;
791         else
792                 return NULL;
793
794         if (flags & FOLL_TOUCH)
795                 touch_pmd(vma, addr, pmd);
796
797         /*
798          * device mapped pages can only be returned if the
799          * caller will manage the page reference count.
800          */
801         if (!(flags & FOLL_GET))
802                 return ERR_PTR(-EEXIST);
803
804         pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
805         pgmap = get_dev_pagemap(pfn, NULL);
806         if (!pgmap)
807                 return ERR_PTR(-EFAULT);
808         page = pfn_to_page(pfn);
809         get_page(page);
810         put_dev_pagemap(pgmap);
811
812         return page;
813 }
814
815 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
816                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
817                   struct vm_area_struct *vma)
818 {
819         spinlock_t *dst_ptl, *src_ptl;
820         struct page *src_page;
821         pmd_t pmd;
822         pgtable_t pgtable = NULL;
823         int ret = -ENOMEM;
824
825         /* Skip if can be re-fill on fault */
826         if (!vma_is_anonymous(vma))
827                 return 0;
828
829         pgtable = pte_alloc_one(dst_mm, addr);
830         if (unlikely(!pgtable))
831                 goto out;
832
833         dst_ptl = pmd_lock(dst_mm, dst_pmd);
834         src_ptl = pmd_lockptr(src_mm, src_pmd);
835         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
836
837         ret = -EAGAIN;
838         pmd = *src_pmd;
839         if (unlikely(!pmd_trans_huge(pmd))) {
840                 pte_free(dst_mm, pgtable);
841                 goto out_unlock;
842         }
843         /*
844          * When page table lock is held, the huge zero pmd should not be
845          * under splitting since we don't split the page itself, only pmd to
846          * a page table.
847          */
848         if (is_huge_zero_pmd(pmd)) {
849                 struct page *zero_page;
850                 /*
851                  * get_huge_zero_page() will never allocate a new page here,
852                  * since we already have a zero page to copy. It just takes a
853                  * reference.
854                  */
855                 zero_page = mm_get_huge_zero_page(dst_mm);
856                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
857                                 zero_page);
858                 ret = 0;
859                 goto out_unlock;
860         }
861
862         src_page = pmd_page(pmd);
863         VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
864         get_page(src_page);
865         page_dup_rmap(src_page, true);
866         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
867         atomic_long_inc(&dst_mm->nr_ptes);
868         pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
869
870         pmdp_set_wrprotect(src_mm, addr, src_pmd);
871         pmd = pmd_mkold(pmd_wrprotect(pmd));
872         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
873
874         ret = 0;
875 out_unlock:
876         spin_unlock(src_ptl);
877         spin_unlock(dst_ptl);
878 out:
879         return ret;
880 }
881
882 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
883 {
884         pmd_t entry;
885         unsigned long haddr;
886
887         vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
888         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
889                 goto unlock;
890
891         entry = pmd_mkyoung(orig_pmd);
892         haddr = vmf->address & HPAGE_PMD_MASK;
893         if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry,
894                                 vmf->flags & FAULT_FLAG_WRITE))
895                 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
896
897 unlock:
898         spin_unlock(vmf->ptl);
899 }
900
901 static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
902                 struct page *page)
903 {
904         struct vm_area_struct *vma = vmf->vma;
905         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
906         struct mem_cgroup *memcg;
907         pgtable_t pgtable;
908         pmd_t _pmd;
909         int ret = 0, i;
910         struct page **pages;
911         unsigned long mmun_start;       /* For mmu_notifiers */
912         unsigned long mmun_end;         /* For mmu_notifiers */
913
914         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
915                         GFP_KERNEL);
916         if (unlikely(!pages)) {
917                 ret |= VM_FAULT_OOM;
918                 goto out;
919         }
920
921         for (i = 0; i < HPAGE_PMD_NR; i++) {
922                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
923                                                __GFP_OTHER_NODE, vma,
924                                                vmf->address, page_to_nid(page));
925                 if (unlikely(!pages[i] ||
926                              mem_cgroup_try_charge(pages[i], vma->vm_mm,
927                                      GFP_KERNEL, &memcg, false))) {
928                         if (pages[i])
929                                 put_page(pages[i]);
930                         while (--i >= 0) {
931                                 memcg = (void *)page_private(pages[i]);
932                                 set_page_private(pages[i], 0);
933                                 mem_cgroup_cancel_charge(pages[i], memcg,
934                                                 false);
935                                 put_page(pages[i]);
936                         }
937                         kfree(pages);
938                         ret |= VM_FAULT_OOM;
939                         goto out;
940                 }
941                 set_page_private(pages[i], (unsigned long)memcg);
942         }
943
944         for (i = 0; i < HPAGE_PMD_NR; i++) {
945                 copy_user_highpage(pages[i], page + i,
946                                    haddr + PAGE_SIZE * i, vma);
947                 __SetPageUptodate(pages[i]);
948                 cond_resched();
949         }
950
951         mmun_start = haddr;
952         mmun_end   = haddr + HPAGE_PMD_SIZE;
953         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
954
955         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
956         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
957                 goto out_free_pages;
958         VM_BUG_ON_PAGE(!PageHead(page), page);
959
960         pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
961         /* leave pmd empty until pte is filled */
962
963         pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
964         pmd_populate(vma->vm_mm, &_pmd, pgtable);
965
966         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
967                 pte_t entry;
968                 entry = mk_pte(pages[i], vma->vm_page_prot);
969                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
970                 memcg = (void *)page_private(pages[i]);
971                 set_page_private(pages[i], 0);
972                 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
973                 mem_cgroup_commit_charge(pages[i], memcg, false, false);
974                 lru_cache_add_active_or_unevictable(pages[i], vma);
975                 vmf->pte = pte_offset_map(&_pmd, haddr);
976                 VM_BUG_ON(!pte_none(*vmf->pte));
977                 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
978                 pte_unmap(vmf->pte);
979         }
980         kfree(pages);
981
982         smp_wmb(); /* make pte visible before pmd */
983         pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
984         page_remove_rmap(page, true);
985         spin_unlock(vmf->ptl);
986
987         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
988
989         ret |= VM_FAULT_WRITE;
990         put_page(page);
991
992 out:
993         return ret;
994
995 out_free_pages:
996         spin_unlock(vmf->ptl);
997         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
998         for (i = 0; i < HPAGE_PMD_NR; i++) {
999                 memcg = (void *)page_private(pages[i]);
1000                 set_page_private(pages[i], 0);
1001                 mem_cgroup_cancel_charge(pages[i], memcg, false);
1002                 put_page(pages[i]);
1003         }
1004         kfree(pages);
1005         goto out;
1006 }
1007
1008 int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1009 {
1010         struct vm_area_struct *vma = vmf->vma;
1011         struct page *page = NULL, *new_page;
1012         struct mem_cgroup *memcg;
1013         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1014         unsigned long mmun_start;       /* For mmu_notifiers */
1015         unsigned long mmun_end;         /* For mmu_notifiers */
1016         gfp_t huge_gfp;                 /* for allocation and charge */
1017         int ret = 0;
1018
1019         vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1020         VM_BUG_ON_VMA(!vma->anon_vma, vma);
1021         if (is_huge_zero_pmd(orig_pmd))
1022                 goto alloc;
1023         spin_lock(vmf->ptl);
1024         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1025                 goto out_unlock;
1026
1027         page = pmd_page(orig_pmd);
1028         VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1029         /*
1030          * We can only reuse the page if nobody else maps the huge page or it's
1031          * part.
1032          */
1033         if (page_trans_huge_mapcount(page, NULL) == 1) {
1034                 pmd_t entry;
1035                 entry = pmd_mkyoung(orig_pmd);
1036                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1037                 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1038                         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1039                 ret |= VM_FAULT_WRITE;
1040                 goto out_unlock;
1041         }
1042         get_page(page);
1043         spin_unlock(vmf->ptl);
1044 alloc:
1045         if (transparent_hugepage_enabled(vma) &&
1046             !transparent_hugepage_debug_cow()) {
1047                 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1048                 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1049         } else
1050                 new_page = NULL;
1051
1052         if (likely(new_page)) {
1053                 prep_transhuge_page(new_page);
1054         } else {
1055                 if (!page) {
1056                         split_huge_pmd(vma, vmf->pmd, vmf->address);
1057                         ret |= VM_FAULT_FALLBACK;
1058                 } else {
1059                         ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1060                         if (ret & VM_FAULT_OOM) {
1061                                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1062                                 ret |= VM_FAULT_FALLBACK;
1063                         }
1064                         put_page(page);
1065                 }
1066                 count_vm_event(THP_FAULT_FALLBACK);
1067                 goto out;
1068         }
1069
1070         if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1071                                         huge_gfp, &memcg, true))) {
1072                 put_page(new_page);
1073                 split_huge_pmd(vma, vmf->pmd, vmf->address);
1074                 if (page)
1075                         put_page(page);
1076                 ret |= VM_FAULT_FALLBACK;
1077                 count_vm_event(THP_FAULT_FALLBACK);
1078                 goto out;
1079         }
1080
1081         count_vm_event(THP_FAULT_ALLOC);
1082
1083         if (!page)
1084                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1085         else
1086                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1087         __SetPageUptodate(new_page);
1088
1089         mmun_start = haddr;
1090         mmun_end   = haddr + HPAGE_PMD_SIZE;
1091         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1092
1093         spin_lock(vmf->ptl);
1094         if (page)
1095                 put_page(page);
1096         if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1097                 spin_unlock(vmf->ptl);
1098                 mem_cgroup_cancel_charge(new_page, memcg, true);
1099                 put_page(new_page);
1100                 goto out_mn;
1101         } else {
1102                 pmd_t entry;
1103                 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1104                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1105                 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1106                 page_add_new_anon_rmap(new_page, vma, haddr, true);
1107                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1108                 lru_cache_add_active_or_unevictable(new_page, vma);
1109                 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1110                 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1111                 if (!page) {
1112                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1113                 } else {
1114                         VM_BUG_ON_PAGE(!PageHead(page), page);
1115                         page_remove_rmap(page, true);
1116                         put_page(page);
1117                 }
1118                 ret |= VM_FAULT_WRITE;
1119         }
1120         spin_unlock(vmf->ptl);
1121 out_mn:
1122         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1123 out:
1124         return ret;
1125 out_unlock:
1126         spin_unlock(vmf->ptl);
1127         return ret;
1128 }
1129
1130 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1131                                    unsigned long addr,
1132                                    pmd_t *pmd,
1133                                    unsigned int flags)
1134 {
1135         struct mm_struct *mm = vma->vm_mm;
1136         struct page *page = NULL;
1137
1138         assert_spin_locked(pmd_lockptr(mm, pmd));
1139
1140         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1141                 goto out;
1142
1143         /* Avoid dumping huge zero page */
1144         if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1145                 return ERR_PTR(-EFAULT);
1146
1147         /* Full NUMA hinting faults to serialise migration in fault paths */
1148         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1149                 goto out;
1150
1151         page = pmd_page(*pmd);
1152         VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1153         if (flags & FOLL_TOUCH)
1154                 touch_pmd(vma, addr, pmd);
1155         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1156                 /*
1157                  * We don't mlock() pte-mapped THPs. This way we can avoid
1158                  * leaking mlocked pages into non-VM_LOCKED VMAs.
1159                  *
1160                  * For anon THP:
1161                  *
1162                  * In most cases the pmd is the only mapping of the page as we
1163                  * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1164                  * writable private mappings in populate_vma_page_range().
1165                  *
1166                  * The only scenario when we have the page shared here is if we
1167                  * mlocking read-only mapping shared over fork(). We skip
1168                  * mlocking such pages.
1169                  *
1170                  * For file THP:
1171                  *
1172                  * We can expect PageDoubleMap() to be stable under page lock:
1173                  * for file pages we set it in page_add_file_rmap(), which
1174                  * requires page to be locked.
1175                  */
1176
1177                 if (PageAnon(page) && compound_mapcount(page) != 1)
1178                         goto skip_mlock;
1179                 if (PageDoubleMap(page) || !page->mapping)
1180                         goto skip_mlock;
1181                 if (!trylock_page(page))
1182                         goto skip_mlock;
1183                 lru_add_drain();
1184                 if (page->mapping && !PageDoubleMap(page))
1185                         mlock_vma_page(page);
1186                 unlock_page(page);
1187         }
1188 skip_mlock:
1189         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1190         VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1191         if (flags & FOLL_GET)
1192                 get_page(page);
1193
1194 out:
1195         return page;
1196 }
1197
1198 /* NUMA hinting page fault entry point for trans huge pmds */
1199 int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1200 {
1201         struct vm_area_struct *vma = vmf->vma;
1202         struct anon_vma *anon_vma = NULL;
1203         struct page *page;
1204         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1205         int page_nid = -1, this_nid = numa_node_id();
1206         int target_nid, last_cpupid = -1;
1207         bool page_locked;
1208         bool migrated = false;
1209         bool was_writable;
1210         int flags = 0;
1211
1212         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1213         if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1214                 goto out_unlock;
1215
1216         /*
1217          * If there are potential migrations, wait for completion and retry
1218          * without disrupting NUMA hinting information. Do not relock and
1219          * check_same as the page may no longer be mapped.
1220          */
1221         if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1222                 page = pmd_page(*vmf->pmd);
1223                 spin_unlock(vmf->ptl);
1224                 wait_on_page_locked(page);
1225                 goto out;
1226         }
1227
1228         page = pmd_page(pmd);
1229         BUG_ON(is_huge_zero_page(page));
1230         page_nid = page_to_nid(page);
1231         last_cpupid = page_cpupid_last(page);
1232         count_vm_numa_event(NUMA_HINT_FAULTS);
1233         if (page_nid == this_nid) {
1234                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1235                 flags |= TNF_FAULT_LOCAL;
1236         }
1237
1238         /* See similar comment in do_numa_page for explanation */
1239         if (!pmd_write(pmd))
1240                 flags |= TNF_NO_GROUP;
1241
1242         /*
1243          * Acquire the page lock to serialise THP migrations but avoid dropping
1244          * page_table_lock if at all possible
1245          */
1246         page_locked = trylock_page(page);
1247         target_nid = mpol_misplaced(page, vma, haddr);
1248         if (target_nid == -1) {
1249                 /* If the page was locked, there are no parallel migrations */
1250                 if (page_locked)
1251                         goto clear_pmdnuma;
1252         }
1253
1254         /* Migration could have started since the pmd_trans_migrating check */
1255         if (!page_locked) {
1256                 spin_unlock(vmf->ptl);
1257                 wait_on_page_locked(page);
1258                 page_nid = -1;
1259                 goto out;
1260         }
1261
1262         /*
1263          * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1264          * to serialises splits
1265          */
1266         get_page(page);
1267         spin_unlock(vmf->ptl);
1268         anon_vma = page_lock_anon_vma_read(page);
1269
1270         /* Confirm the PMD did not change while page_table_lock was released */
1271         spin_lock(vmf->ptl);
1272         if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1273                 unlock_page(page);
1274                 put_page(page);
1275                 page_nid = -1;
1276                 goto out_unlock;
1277         }
1278
1279         /* Bail if we fail to protect against THP splits for any reason */
1280         if (unlikely(!anon_vma)) {
1281                 put_page(page);
1282                 page_nid = -1;
1283                 goto clear_pmdnuma;
1284         }
1285
1286         /*
1287          * Migrate the THP to the requested node, returns with page unlocked
1288          * and access rights restored.
1289          */
1290         spin_unlock(vmf->ptl);
1291         migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1292                                 vmf->pmd, pmd, vmf->address, page, target_nid);
1293         if (migrated) {
1294                 flags |= TNF_MIGRATED;
1295                 page_nid = target_nid;
1296         } else
1297                 flags |= TNF_MIGRATE_FAIL;
1298
1299         goto out;
1300 clear_pmdnuma:
1301         BUG_ON(!PageLocked(page));
1302         was_writable = pmd_write(pmd);
1303         pmd = pmd_modify(pmd, vma->vm_page_prot);
1304         pmd = pmd_mkyoung(pmd);
1305         if (was_writable)
1306                 pmd = pmd_mkwrite(pmd);
1307         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1308         update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1309         unlock_page(page);
1310 out_unlock:
1311         spin_unlock(vmf->ptl);
1312
1313 out:
1314         if (anon_vma)
1315                 page_unlock_anon_vma_read(anon_vma);
1316
1317         if (page_nid != -1)
1318                 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1319                                 vmf->flags);
1320
1321         return 0;
1322 }
1323
1324 /*
1325  * Return true if we do MADV_FREE successfully on entire pmd page.
1326  * Otherwise, return false.
1327  */
1328 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1329                 pmd_t *pmd, unsigned long addr, unsigned long next)
1330 {
1331         spinlock_t *ptl;
1332         pmd_t orig_pmd;
1333         struct page *page;
1334         struct mm_struct *mm = tlb->mm;
1335         bool ret = false;
1336
1337         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1338
1339         ptl = pmd_trans_huge_lock(pmd, vma);
1340         if (!ptl)
1341                 goto out_unlocked;
1342
1343         orig_pmd = *pmd;
1344         if (is_huge_zero_pmd(orig_pmd))
1345                 goto out;
1346
1347         page = pmd_page(orig_pmd);
1348         /*
1349          * If other processes are mapping this page, we couldn't discard
1350          * the page unless they all do MADV_FREE so let's skip the page.
1351          */
1352         if (page_mapcount(page) != 1)
1353                 goto out;
1354
1355         if (!trylock_page(page))
1356                 goto out;
1357
1358         /*
1359          * If user want to discard part-pages of THP, split it so MADV_FREE
1360          * will deactivate only them.
1361          */
1362         if (next - addr != HPAGE_PMD_SIZE) {
1363                 get_page(page);
1364                 spin_unlock(ptl);
1365                 split_huge_page(page);
1366                 put_page(page);
1367                 unlock_page(page);
1368                 goto out_unlocked;
1369         }
1370
1371         if (PageDirty(page))
1372                 ClearPageDirty(page);
1373         unlock_page(page);
1374
1375         if (PageActive(page))
1376                 deactivate_page(page);
1377
1378         if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1379                 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1380                         tlb->fullmm);
1381                 orig_pmd = pmd_mkold(orig_pmd);
1382                 orig_pmd = pmd_mkclean(orig_pmd);
1383
1384                 set_pmd_at(mm, addr, pmd, orig_pmd);
1385                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1386         }
1387         ret = true;
1388 out:
1389         spin_unlock(ptl);
1390 out_unlocked:
1391         return ret;
1392 }
1393
1394 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1395 {
1396         pgtable_t pgtable;
1397
1398         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1399         pte_free(mm, pgtable);
1400         atomic_long_dec(&mm->nr_ptes);
1401 }
1402
1403 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1404                  pmd_t *pmd, unsigned long addr)
1405 {
1406         pmd_t orig_pmd;
1407         spinlock_t *ptl;
1408
1409         tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1410
1411         ptl = __pmd_trans_huge_lock(pmd, vma);
1412         if (!ptl)
1413                 return 0;
1414         /*
1415          * For architectures like ppc64 we look at deposited pgtable
1416          * when calling pmdp_huge_get_and_clear. So do the
1417          * pgtable_trans_huge_withdraw after finishing pmdp related
1418          * operations.
1419          */
1420         orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1421                         tlb->fullmm);
1422         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1423         if (vma_is_dax(vma)) {
1424                 spin_unlock(ptl);
1425                 if (is_huge_zero_pmd(orig_pmd))
1426                         tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1427         } else if (is_huge_zero_pmd(orig_pmd)) {
1428                 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1429                 atomic_long_dec(&tlb->mm->nr_ptes);
1430                 spin_unlock(ptl);
1431                 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1432         } else {
1433                 struct page *page = pmd_page(orig_pmd);
1434                 page_remove_rmap(page, true);
1435                 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1436                 VM_BUG_ON_PAGE(!PageHead(page), page);
1437                 if (PageAnon(page)) {
1438                         pgtable_t pgtable;
1439                         pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1440                         pte_free(tlb->mm, pgtable);
1441                         atomic_long_dec(&tlb->mm->nr_ptes);
1442                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1443                 } else {
1444                         if (arch_needs_pgtable_deposit())
1445                                 zap_deposited_table(tlb->mm, pmd);
1446                         add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1447                 }
1448                 spin_unlock(ptl);
1449                 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1450         }
1451         return 1;
1452 }
1453
1454 #ifndef pmd_move_must_withdraw
1455 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1456                                          spinlock_t *old_pmd_ptl,
1457                                          struct vm_area_struct *vma)
1458 {
1459         /*
1460          * With split pmd lock we also need to move preallocated
1461          * PTE page table if new_pmd is on different PMD page table.
1462          *
1463          * We also don't deposit and withdraw tables for file pages.
1464          */
1465         return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1466 }
1467 #endif
1468
1469 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1470                   unsigned long new_addr, unsigned long old_end,
1471                   pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1472 {
1473         spinlock_t *old_ptl, *new_ptl;
1474         pmd_t pmd;
1475         struct mm_struct *mm = vma->vm_mm;
1476         bool force_flush = false;
1477
1478         if ((old_addr & ~HPAGE_PMD_MASK) ||
1479             (new_addr & ~HPAGE_PMD_MASK) ||
1480             old_end - old_addr < HPAGE_PMD_SIZE)
1481                 return false;
1482
1483         /*
1484          * The destination pmd shouldn't be established, free_pgtables()
1485          * should have release it.
1486          */
1487         if (WARN_ON(!pmd_none(*new_pmd))) {
1488                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1489                 return false;
1490         }
1491
1492         /*
1493          * We don't have to worry about the ordering of src and dst
1494          * ptlocks because exclusive mmap_sem prevents deadlock.
1495          */
1496         old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1497         if (old_ptl) {
1498                 new_ptl = pmd_lockptr(mm, new_pmd);
1499                 if (new_ptl != old_ptl)
1500                         spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1501                 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1502                 if (pmd_present(pmd) && pmd_dirty(pmd))
1503                         force_flush = true;
1504                 VM_BUG_ON(!pmd_none(*new_pmd));
1505
1506                 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1507                         pgtable_t pgtable;
1508                         pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1509                         pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1510                 }
1511                 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1512                 if (new_ptl != old_ptl)
1513                         spin_unlock(new_ptl);
1514                 if (force_flush)
1515                         flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1516                 else
1517                         *need_flush = true;
1518                 spin_unlock(old_ptl);
1519                 return true;
1520         }
1521         return false;
1522 }
1523
1524 /*
1525  * Returns
1526  *  - 0 if PMD could not be locked
1527  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1528  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1529  */
1530 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1531                 unsigned long addr, pgprot_t newprot, int prot_numa)
1532 {
1533         struct mm_struct *mm = vma->vm_mm;
1534         spinlock_t *ptl;
1535         int ret = 0;
1536
1537         ptl = __pmd_trans_huge_lock(pmd, vma);
1538         if (ptl) {
1539                 pmd_t entry;
1540                 bool preserve_write = prot_numa && pmd_write(*pmd);
1541                 ret = 1;
1542
1543                 /*
1544                  * Avoid trapping faults against the zero page. The read-only
1545                  * data is likely to be read-cached on the local CPU and
1546                  * local/remote hits to the zero page are not interesting.
1547                  */
1548                 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1549                         spin_unlock(ptl);
1550                         return ret;
1551                 }
1552
1553                 if (!prot_numa || !pmd_protnone(*pmd)) {
1554                         entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1555                         entry = pmd_modify(entry, newprot);
1556                         if (preserve_write)
1557                                 entry = pmd_mkwrite(entry);
1558                         ret = HPAGE_PMD_NR;
1559                         set_pmd_at(mm, addr, pmd, entry);
1560                         BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1561                                         pmd_write(entry));
1562                 }
1563                 spin_unlock(ptl);
1564         }
1565
1566         return ret;
1567 }
1568
1569 /*
1570  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1571  *
1572  * Note that if it returns page table lock pointer, this routine returns without
1573  * unlocking page table lock. So callers must unlock it.
1574  */
1575 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1576 {
1577         spinlock_t *ptl;
1578         ptl = pmd_lock(vma->vm_mm, pmd);
1579         if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1580                 return ptl;
1581         spin_unlock(ptl);
1582         return NULL;
1583 }
1584
1585 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1586                 unsigned long haddr, pmd_t *pmd)
1587 {
1588         struct mm_struct *mm = vma->vm_mm;
1589         pgtable_t pgtable;
1590         pmd_t _pmd;
1591         int i;
1592
1593         /* leave pmd empty until pte is filled */
1594         pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1595
1596         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1597         pmd_populate(mm, &_pmd, pgtable);
1598
1599         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1600                 pte_t *pte, entry;
1601                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1602                 entry = pte_mkspecial(entry);
1603                 pte = pte_offset_map(&_pmd, haddr);
1604                 VM_BUG_ON(!pte_none(*pte));
1605                 set_pte_at(mm, haddr, pte, entry);
1606                 pte_unmap(pte);
1607         }
1608         smp_wmb(); /* make pte visible before pmd */
1609         pmd_populate(mm, pmd, pgtable);
1610 }
1611
1612 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1613                 unsigned long haddr, bool freeze)
1614 {
1615         struct mm_struct *mm = vma->vm_mm;
1616         struct page *page;
1617         pgtable_t pgtable;
1618         pmd_t _pmd;
1619         bool young, write, dirty, soft_dirty;
1620         unsigned long addr;
1621         int i;
1622
1623         VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1624         VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1625         VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1626         VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1627
1628         count_vm_event(THP_SPLIT_PMD);
1629
1630         if (!vma_is_anonymous(vma)) {
1631                 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1632                 /*
1633                  * We are going to unmap this huge page. So
1634                  * just go ahead and zap it
1635                  */
1636                 if (arch_needs_pgtable_deposit())
1637                         zap_deposited_table(mm, pmd);
1638                 if (vma_is_dax(vma))
1639                         return;
1640                 page = pmd_page(_pmd);
1641                 if (!PageReferenced(page) && pmd_young(_pmd))
1642                         SetPageReferenced(page);
1643                 page_remove_rmap(page, true);
1644                 put_page(page);
1645                 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1646                 return;
1647         } else if (is_huge_zero_pmd(*pmd)) {
1648                 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1649         }
1650
1651         page = pmd_page(*pmd);
1652         VM_BUG_ON_PAGE(!page_count(page), page);
1653         page_ref_add(page, HPAGE_PMD_NR - 1);
1654         write = pmd_write(*pmd);
1655         young = pmd_young(*pmd);
1656         dirty = pmd_dirty(*pmd);
1657         soft_dirty = pmd_soft_dirty(*pmd);
1658
1659         pmdp_huge_split_prepare(vma, haddr, pmd);
1660         pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1661         pmd_populate(mm, &_pmd, pgtable);
1662
1663         for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1664                 pte_t entry, *pte;
1665                 /*
1666                  * Note that NUMA hinting access restrictions are not
1667                  * transferred to avoid any possibility of altering
1668                  * permissions across VMAs.
1669                  */
1670                 if (freeze) {
1671                         swp_entry_t swp_entry;
1672                         swp_entry = make_migration_entry(page + i, write);
1673                         entry = swp_entry_to_pte(swp_entry);
1674                         if (soft_dirty)
1675                                 entry = pte_swp_mksoft_dirty(entry);
1676                 } else {
1677                         entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1678                         entry = maybe_mkwrite(entry, vma);
1679                         if (!write)
1680                                 entry = pte_wrprotect(entry);
1681                         if (!young)
1682                                 entry = pte_mkold(entry);
1683                         if (soft_dirty)
1684                                 entry = pte_mksoft_dirty(entry);
1685                 }
1686                 if (dirty)
1687                         SetPageDirty(page + i);
1688                 pte = pte_offset_map(&_pmd, addr);
1689                 BUG_ON(!pte_none(*pte));
1690                 set_pte_at(mm, addr, pte, entry);
1691                 atomic_inc(&page[i]._mapcount);
1692                 pte_unmap(pte);
1693         }
1694
1695         /*
1696          * Set PG_double_map before dropping compound_mapcount to avoid
1697          * false-negative page_mapped().
1698          */
1699         if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1700                 for (i = 0; i < HPAGE_PMD_NR; i++)
1701                         atomic_inc(&page[i]._mapcount);
1702         }
1703
1704         if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1705                 /* Last compound_mapcount is gone. */
1706                 __dec_node_page_state(page, NR_ANON_THPS);
1707                 if (TestClearPageDoubleMap(page)) {
1708                         /* No need in mapcount reference anymore */
1709                         for (i = 0; i < HPAGE_PMD_NR; i++)
1710                                 atomic_dec(&page[i]._mapcount);
1711                 }
1712         }
1713
1714         smp_wmb(); /* make pte visible before pmd */
1715         /*
1716          * Up to this point the pmd is present and huge and userland has the
1717          * whole access to the hugepage during the split (which happens in
1718          * place). If we overwrite the pmd with the not-huge version pointing
1719          * to the pte here (which of course we could if all CPUs were bug
1720          * free), userland could trigger a small page size TLB miss on the
1721          * small sized TLB while the hugepage TLB entry is still established in
1722          * the huge TLB. Some CPU doesn't like that.
1723          * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1724          * 383 on page 93. Intel should be safe but is also warns that it's
1725          * only safe if the permission and cache attributes of the two entries
1726          * loaded in the two TLB is identical (which should be the case here).
1727          * But it is generally safer to never allow small and huge TLB entries
1728          * for the same virtual address to be loaded simultaneously. So instead
1729          * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1730          * current pmd notpresent (atomically because here the pmd_trans_huge
1731          * and pmd_trans_splitting must remain set at all times on the pmd
1732          * until the split is complete for this pmd), then we flush the SMP TLB
1733          * and finally we write the non-huge version of the pmd entry with
1734          * pmd_populate.
1735          */
1736         pmdp_invalidate(vma, haddr, pmd);
1737         pmd_populate(mm, pmd, pgtable);
1738
1739         if (freeze) {
1740                 for (i = 0; i < HPAGE_PMD_NR; i++) {
1741                         page_remove_rmap(page + i, false);
1742                         put_page(page + i);
1743                 }
1744         }
1745 }
1746
1747 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1748                 unsigned long address, bool freeze, struct page *page)
1749 {
1750         spinlock_t *ptl;
1751         struct mm_struct *mm = vma->vm_mm;
1752         unsigned long haddr = address & HPAGE_PMD_MASK;
1753
1754         mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1755         ptl = pmd_lock(mm, pmd);
1756
1757         /*
1758          * If caller asks to setup a migration entries, we need a page to check
1759          * pmd against. Otherwise we can end up replacing wrong page.
1760          */
1761         VM_BUG_ON(freeze && !page);
1762         if (page && page != pmd_page(*pmd))
1763                 goto out;
1764
1765         if (pmd_trans_huge(*pmd)) {
1766                 page = pmd_page(*pmd);
1767                 if (PageMlocked(page))
1768                         clear_page_mlock(page);
1769         } else if (!pmd_devmap(*pmd))
1770                 goto out;
1771         __split_huge_pmd_locked(vma, pmd, haddr, freeze);
1772 out:
1773         spin_unlock(ptl);
1774         mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1775 }
1776
1777 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1778                 bool freeze, struct page *page)
1779 {
1780         pgd_t *pgd;
1781         pud_t *pud;
1782         pmd_t *pmd;
1783
1784         pgd = pgd_offset(vma->vm_mm, address);
1785         if (!pgd_present(*pgd))
1786                 return;
1787
1788         pud = pud_offset(pgd, address);
1789         if (!pud_present(*pud))
1790                 return;
1791
1792         pmd = pmd_offset(pud, address);
1793
1794         __split_huge_pmd(vma, pmd, address, freeze, page);
1795 }
1796
1797 void vma_adjust_trans_huge(struct vm_area_struct *vma,
1798                              unsigned long start,
1799                              unsigned long end,
1800                              long adjust_next)
1801 {
1802         /*
1803          * If the new start address isn't hpage aligned and it could
1804          * previously contain an hugepage: check if we need to split
1805          * an huge pmd.
1806          */
1807         if (start & ~HPAGE_PMD_MASK &&
1808             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1809             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1810                 split_huge_pmd_address(vma, start, false, NULL);
1811
1812         /*
1813          * If the new end address isn't hpage aligned and it could
1814          * previously contain an hugepage: check if we need to split
1815          * an huge pmd.
1816          */
1817         if (end & ~HPAGE_PMD_MASK &&
1818             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1819             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1820                 split_huge_pmd_address(vma, end, false, NULL);
1821
1822         /*
1823          * If we're also updating the vma->vm_next->vm_start, if the new
1824          * vm_next->vm_start isn't page aligned and it could previously
1825          * contain an hugepage: check if we need to split an huge pmd.
1826          */
1827         if (adjust_next > 0) {
1828                 struct vm_area_struct *next = vma->vm_next;
1829                 unsigned long nstart = next->vm_start;
1830                 nstart += adjust_next << PAGE_SHIFT;
1831                 if (nstart & ~HPAGE_PMD_MASK &&
1832                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1833                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
1834                         split_huge_pmd_address(next, nstart, false, NULL);
1835         }
1836 }
1837
1838 static void freeze_page(struct page *page)
1839 {
1840         enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1841                 TTU_RMAP_LOCKED;
1842         int i, ret;
1843
1844         VM_BUG_ON_PAGE(!PageHead(page), page);
1845
1846         if (PageAnon(page))
1847                 ttu_flags |= TTU_MIGRATION;
1848
1849         /* We only need TTU_SPLIT_HUGE_PMD once */
1850         ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1851         for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1852                 /* Cut short if the page is unmapped */
1853                 if (page_count(page) == 1)
1854                         return;
1855
1856                 ret = try_to_unmap(page + i, ttu_flags);
1857         }
1858         VM_BUG_ON_PAGE(ret, page + i - 1);
1859 }
1860
1861 static void unfreeze_page(struct page *page)
1862 {
1863         int i;
1864
1865         for (i = 0; i < HPAGE_PMD_NR; i++)
1866                 remove_migration_ptes(page + i, page + i, true);
1867 }
1868
1869 static void __split_huge_page_tail(struct page *head, int tail,
1870                 struct lruvec *lruvec, struct list_head *list)
1871 {
1872         struct page *page_tail = head + tail;
1873
1874         VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
1875         VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
1876
1877         /*
1878          * tail_page->_refcount is zero and not changing from under us. But
1879          * get_page_unless_zero() may be running from under us on the
1880          * tail_page. If we used atomic_set() below instead of atomic_inc() or
1881          * atomic_add(), we would then run atomic_set() concurrently with
1882          * get_page_unless_zero(), and atomic_set() is implemented in C not
1883          * using locked ops. spin_unlock on x86 sometime uses locked ops
1884          * because of PPro errata 66, 92, so unless somebody can guarantee
1885          * atomic_set() here would be safe on all archs (and not only on x86),
1886          * it's safer to use atomic_inc()/atomic_add().
1887          */
1888         if (PageAnon(head)) {
1889                 page_ref_inc(page_tail);
1890         } else {
1891                 /* Additional pin to radix tree */
1892                 page_ref_add(page_tail, 2);
1893         }
1894
1895         page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1896         page_tail->flags |= (head->flags &
1897                         ((1L << PG_referenced) |
1898                          (1L << PG_swapbacked) |
1899                          (1L << PG_mlocked) |
1900                          (1L << PG_uptodate) |
1901                          (1L << PG_active) |
1902                          (1L << PG_locked) |
1903                          (1L << PG_unevictable) |
1904                          (1L << PG_dirty)));
1905
1906         /*
1907          * After clearing PageTail the gup refcount can be released.
1908          * Page flags also must be visible before we make the page non-compound.
1909          */
1910         smp_wmb();
1911
1912         clear_compound_head(page_tail);
1913
1914         if (page_is_young(head))
1915                 set_page_young(page_tail);
1916         if (page_is_idle(head))
1917                 set_page_idle(page_tail);
1918
1919         /* ->mapping in first tail page is compound_mapcount */
1920         VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
1921                         page_tail);
1922         page_tail->mapping = head->mapping;
1923
1924         page_tail->index = head->index + tail;
1925         page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1926         lru_add_page_tail(head, page_tail, lruvec, list);
1927 }
1928
1929 static void __split_huge_page(struct page *page, struct list_head *list,
1930                 unsigned long flags)
1931 {
1932         struct page *head = compound_head(page);
1933         struct zone *zone = page_zone(head);
1934         struct lruvec *lruvec;
1935         pgoff_t end = -1;
1936         int i;
1937
1938         lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
1939
1940         /* complete memcg works before add pages to LRU */
1941         mem_cgroup_split_huge_fixup(head);
1942
1943         if (!PageAnon(page))
1944                 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
1945
1946         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1947                 __split_huge_page_tail(head, i, lruvec, list);
1948                 /* Some pages can be beyond i_size: drop them from page cache */
1949                 if (head[i].index >= end) {
1950                         __ClearPageDirty(head + i);
1951                         __delete_from_page_cache(head + i, NULL);
1952                         if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1953                                 shmem_uncharge(head->mapping->host, 1);
1954                         put_page(head + i);
1955                 }
1956         }
1957
1958         ClearPageCompound(head);
1959         /* See comment in __split_huge_page_tail() */
1960         if (PageAnon(head)) {
1961                 page_ref_inc(head);
1962         } else {
1963                 /* Additional pin to radix tree */
1964                 page_ref_add(head, 2);
1965                 spin_unlock(&head->mapping->tree_lock);
1966         }
1967
1968         spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
1969
1970         unfreeze_page(head);
1971
1972         for (i = 0; i < HPAGE_PMD_NR; i++) {
1973                 struct page *subpage = head + i;
1974                 if (subpage == page)
1975                         continue;
1976                 unlock_page(subpage);
1977
1978                 /*
1979                  * Subpages may be freed if there wasn't any mapping
1980                  * like if add_to_swap() is running on a lru page that
1981                  * had its mapping zapped. And freeing these pages
1982                  * requires taking the lru_lock so we do the put_page
1983                  * of the tail pages after the split is complete.
1984                  */
1985                 put_page(subpage);
1986         }
1987 }
1988
1989 int total_mapcount(struct page *page)
1990 {
1991         int i, compound, ret;
1992
1993         VM_BUG_ON_PAGE(PageTail(page), page);
1994
1995         if (likely(!PageCompound(page)))
1996                 return atomic_read(&page->_mapcount) + 1;
1997
1998         compound = compound_mapcount(page);
1999         if (PageHuge(page))
2000                 return compound;
2001         ret = compound;
2002         for (i = 0; i < HPAGE_PMD_NR; i++)
2003                 ret += atomic_read(&page[i]._mapcount) + 1;
2004         /* File pages has compound_mapcount included in _mapcount */
2005         if (!PageAnon(page))
2006                 return ret - compound * HPAGE_PMD_NR;
2007         if (PageDoubleMap(page))
2008                 ret -= HPAGE_PMD_NR;
2009         return ret;
2010 }
2011
2012 /*
2013  * This calculates accurately how many mappings a transparent hugepage
2014  * has (unlike page_mapcount() which isn't fully accurate). This full
2015  * accuracy is primarily needed to know if copy-on-write faults can
2016  * reuse the page and change the mapping to read-write instead of
2017  * copying them. At the same time this returns the total_mapcount too.
2018  *
2019  * The function returns the highest mapcount any one of the subpages
2020  * has. If the return value is one, even if different processes are
2021  * mapping different subpages of the transparent hugepage, they can
2022  * all reuse it, because each process is reusing a different subpage.
2023  *
2024  * The total_mapcount is instead counting all virtual mappings of the
2025  * subpages. If the total_mapcount is equal to "one", it tells the
2026  * caller all mappings belong to the same "mm" and in turn the
2027  * anon_vma of the transparent hugepage can become the vma->anon_vma
2028  * local one as no other process may be mapping any of the subpages.
2029  *
2030  * It would be more accurate to replace page_mapcount() with
2031  * page_trans_huge_mapcount(), however we only use
2032  * page_trans_huge_mapcount() in the copy-on-write faults where we
2033  * need full accuracy to avoid breaking page pinning, because
2034  * page_trans_huge_mapcount() is slower than page_mapcount().
2035  */
2036 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2037 {
2038         int i, ret, _total_mapcount, mapcount;
2039
2040         /* hugetlbfs shouldn't call it */
2041         VM_BUG_ON_PAGE(PageHuge(page), page);
2042
2043         if (likely(!PageTransCompound(page))) {
2044                 mapcount = atomic_read(&page->_mapcount) + 1;
2045                 if (total_mapcount)
2046                         *total_mapcount = mapcount;
2047                 return mapcount;
2048         }
2049
2050         page = compound_head(page);
2051
2052         _total_mapcount = ret = 0;
2053         for (i = 0; i < HPAGE_PMD_NR; i++) {
2054                 mapcount = atomic_read(&page[i]._mapcount) + 1;
2055                 ret = max(ret, mapcount);
2056                 _total_mapcount += mapcount;
2057         }
2058         if (PageDoubleMap(page)) {
2059                 ret -= 1;
2060                 _total_mapcount -= HPAGE_PMD_NR;
2061         }
2062         mapcount = compound_mapcount(page);
2063         ret += mapcount;
2064         _total_mapcount += mapcount;
2065         if (total_mapcount)
2066                 *total_mapcount = _total_mapcount;
2067         return ret;
2068 }
2069
2070 /*
2071  * This function splits huge page into normal pages. @page can point to any
2072  * subpage of huge page to split. Split doesn't change the position of @page.
2073  *
2074  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2075  * The huge page must be locked.
2076  *
2077  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2078  *
2079  * Both head page and tail pages will inherit mapping, flags, and so on from
2080  * the hugepage.
2081  *
2082  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2083  * they are not mapped.
2084  *
2085  * Returns 0 if the hugepage is split successfully.
2086  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2087  * us.
2088  */
2089 int split_huge_page_to_list(struct page *page, struct list_head *list)
2090 {
2091         struct page *head = compound_head(page);
2092         struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2093         struct anon_vma *anon_vma = NULL;
2094         struct address_space *mapping = NULL;
2095         int count, mapcount, extra_pins, ret;
2096         bool mlocked;
2097         unsigned long flags;
2098
2099         VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2100         VM_BUG_ON_PAGE(!PageLocked(page), page);
2101         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2102         VM_BUG_ON_PAGE(!PageCompound(page), page);
2103
2104         if (PageAnon(head)) {
2105                 /*
2106                  * The caller does not necessarily hold an mmap_sem that would
2107                  * prevent the anon_vma disappearing so we first we take a
2108                  * reference to it and then lock the anon_vma for write. This
2109                  * is similar to page_lock_anon_vma_read except the write lock
2110                  * is taken to serialise against parallel split or collapse
2111                  * operations.
2112                  */
2113                 anon_vma = page_get_anon_vma(head);
2114                 if (!anon_vma) {
2115                         ret = -EBUSY;
2116                         goto out;
2117                 }
2118                 extra_pins = 0;
2119                 mapping = NULL;
2120                 anon_vma_lock_write(anon_vma);
2121         } else {
2122                 mapping = head->mapping;
2123
2124                 /* Truncated ? */
2125                 if (!mapping) {
2126                         ret = -EBUSY;
2127                         goto out;
2128                 }
2129
2130                 /* Addidional pins from radix tree */
2131                 extra_pins = HPAGE_PMD_NR;
2132                 anon_vma = NULL;
2133                 i_mmap_lock_read(mapping);
2134         }
2135
2136         /*
2137          * Racy check if we can split the page, before freeze_page() will
2138          * split PMDs
2139          */
2140         if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2141                 ret = -EBUSY;
2142                 goto out_unlock;
2143         }
2144
2145         mlocked = PageMlocked(page);
2146         freeze_page(head);
2147         VM_BUG_ON_PAGE(compound_mapcount(head), head);
2148
2149         /* Make sure the page is not on per-CPU pagevec as it takes pin */
2150         if (mlocked)
2151                 lru_add_drain();
2152
2153         /* prevent PageLRU to go away from under us, and freeze lru stats */
2154         spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2155
2156         if (mapping) {
2157                 void **pslot;
2158
2159                 spin_lock(&mapping->tree_lock);
2160                 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2161                                 page_index(head));
2162                 /*
2163                  * Check if the head page is present in radix tree.
2164                  * We assume all tail are present too, if head is there.
2165                  */
2166                 if (radix_tree_deref_slot_protected(pslot,
2167                                         &mapping->tree_lock) != head)
2168                         goto fail;
2169         }
2170
2171         /* Prevent deferred_split_scan() touching ->_refcount */
2172         spin_lock(&pgdata->split_queue_lock);
2173         count = page_count(head);
2174         mapcount = total_mapcount(head);
2175         if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2176                 if (!list_empty(page_deferred_list(head))) {
2177                         pgdata->split_queue_len--;
2178                         list_del(page_deferred_list(head));
2179                 }
2180                 if (mapping)
2181                         __dec_node_page_state(page, NR_SHMEM_THPS);
2182                 spin_unlock(&pgdata->split_queue_lock);
2183                 __split_huge_page(page, list, flags);
2184                 ret = 0;
2185         } else {
2186                 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2187                         pr_alert("total_mapcount: %u, page_count(): %u\n",
2188                                         mapcount, count);
2189                         if (PageTail(page))
2190                                 dump_page(head, NULL);
2191                         dump_page(page, "total_mapcount(head) > 0");
2192                         BUG();
2193                 }
2194                 spin_unlock(&pgdata->split_queue_lock);
2195 fail:           if (mapping)
2196                         spin_unlock(&mapping->tree_lock);
2197                 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2198                 unfreeze_page(head);
2199                 ret = -EBUSY;
2200         }
2201
2202 out_unlock:
2203         if (anon_vma) {
2204                 anon_vma_unlock_write(anon_vma);
2205                 put_anon_vma(anon_vma);
2206         }
2207         if (mapping)
2208                 i_mmap_unlock_read(mapping);
2209 out:
2210         count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2211         return ret;
2212 }
2213
2214 void free_transhuge_page(struct page *page)
2215 {
2216         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2217         unsigned long flags;
2218
2219         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2220         if (!list_empty(page_deferred_list(page))) {
2221                 pgdata->split_queue_len--;
2222                 list_del(page_deferred_list(page));
2223         }
2224         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2225         free_compound_page(page);
2226 }
2227
2228 void deferred_split_huge_page(struct page *page)
2229 {
2230         struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2231         unsigned long flags;
2232
2233         VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2234
2235         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2236         if (list_empty(page_deferred_list(page))) {
2237                 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2238                 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2239                 pgdata->split_queue_len++;
2240         }
2241         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2242 }
2243
2244 static unsigned long deferred_split_count(struct shrinker *shrink,
2245                 struct shrink_control *sc)
2246 {
2247         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2248         return ACCESS_ONCE(pgdata->split_queue_len);
2249 }
2250
2251 static unsigned long deferred_split_scan(struct shrinker *shrink,
2252                 struct shrink_control *sc)
2253 {
2254         struct pglist_data *pgdata = NODE_DATA(sc->nid);
2255         unsigned long flags;
2256         LIST_HEAD(list), *pos, *next;
2257         struct page *page;
2258         int split = 0;
2259
2260         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2261         /* Take pin on all head pages to avoid freeing them under us */
2262         list_for_each_safe(pos, next, &pgdata->split_queue) {
2263                 page = list_entry((void *)pos, struct page, mapping);
2264                 page = compound_head(page);
2265                 if (get_page_unless_zero(page)) {
2266                         list_move(page_deferred_list(page), &list);
2267                 } else {
2268                         /* We lost race with put_compound_page() */
2269                         list_del_init(page_deferred_list(page));
2270                         pgdata->split_queue_len--;
2271                 }
2272                 if (!--sc->nr_to_scan)
2273                         break;
2274         }
2275         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2276
2277         list_for_each_safe(pos, next, &list) {
2278                 page = list_entry((void *)pos, struct page, mapping);
2279                 lock_page(page);
2280                 /* split_huge_page() removes page from list on success */
2281                 if (!split_huge_page(page))
2282                         split++;
2283                 unlock_page(page);
2284                 put_page(page);
2285         }
2286
2287         spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2288         list_splice_tail(&list, &pgdata->split_queue);
2289         spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2290
2291         /*
2292          * Stop shrinker if we didn't split any page, but the queue is empty.
2293          * This can happen if pages were freed under us.
2294          */
2295         if (!split && list_empty(&pgdata->split_queue))
2296                 return SHRINK_STOP;
2297         return split;
2298 }
2299
2300 static struct shrinker deferred_split_shrinker = {
2301         .count_objects = deferred_split_count,
2302         .scan_objects = deferred_split_scan,
2303         .seeks = DEFAULT_SEEKS,
2304         .flags = SHRINKER_NUMA_AWARE,
2305 };
2306
2307 #ifdef CONFIG_DEBUG_FS
2308 static int split_huge_pages_set(void *data, u64 val)
2309 {
2310         struct zone *zone;
2311         struct page *page;
2312         unsigned long pfn, max_zone_pfn;
2313         unsigned long total = 0, split = 0;
2314
2315         if (val != 1)
2316                 return -EINVAL;
2317
2318         for_each_populated_zone(zone) {
2319                 max_zone_pfn = zone_end_pfn(zone);
2320                 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2321                         if (!pfn_valid(pfn))
2322                                 continue;
2323
2324                         page = pfn_to_page(pfn);
2325                         if (!get_page_unless_zero(page))
2326                                 continue;
2327
2328                         if (zone != page_zone(page))
2329                                 goto next;
2330
2331                         if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2332                                 goto next;
2333
2334                         total++;
2335                         lock_page(page);
2336                         if (!split_huge_page(page))
2337                                 split++;
2338                         unlock_page(page);
2339 next:
2340                         put_page(page);
2341                 }
2342         }
2343
2344         pr_info("%lu of %lu THP split\n", split, total);
2345
2346         return 0;
2347 }
2348 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2349                 "%llu\n");
2350
2351 static int __init split_huge_pages_debugfs(void)
2352 {
2353         void *ret;
2354
2355         ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2356                         &split_huge_pages_fops);
2357         if (!ret)
2358                 pr_warn("Failed to create split_huge_pages in debugfs");
2359         return 0;
2360 }
2361 late_initcall(split_huge_pages_debugfs);
2362 #endif