net: socket: add check for negative optlen in compat setsockopt
[platform/kernel/linux-starfive.git] / mm / khugepaged.c
1 // SPDX-License-Identifier: GPL-2.0
2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
3
4 #include <linux/mm.h>
5 #include <linux/sched.h>
6 #include <linux/sched/mm.h>
7 #include <linux/sched/coredump.h>
8 #include <linux/mmu_notifier.h>
9 #include <linux/rmap.h>
10 #include <linux/swap.h>
11 #include <linux/mm_inline.h>
12 #include <linux/kthread.h>
13 #include <linux/khugepaged.h>
14 #include <linux/freezer.h>
15 #include <linux/mman.h>
16 #include <linux/hashtable.h>
17 #include <linux/userfaultfd_k.h>
18 #include <linux/page_idle.h>
19 #include <linux/swapops.h>
20 #include <linux/shmem_fs.h>
21
22 #include <asm/tlb.h>
23 #include <asm/pgalloc.h>
24 #include "internal.h"
25
26 enum scan_result {
27         SCAN_FAIL,
28         SCAN_SUCCEED,
29         SCAN_PMD_NULL,
30         SCAN_EXCEED_NONE_PTE,
31         SCAN_PTE_NON_PRESENT,
32         SCAN_PAGE_RO,
33         SCAN_LACK_REFERENCED_PAGE,
34         SCAN_PAGE_NULL,
35         SCAN_SCAN_ABORT,
36         SCAN_PAGE_COUNT,
37         SCAN_PAGE_LRU,
38         SCAN_PAGE_LOCK,
39         SCAN_PAGE_ANON,
40         SCAN_PAGE_COMPOUND,
41         SCAN_ANY_PROCESS,
42         SCAN_VMA_NULL,
43         SCAN_VMA_CHECK,
44         SCAN_ADDRESS_RANGE,
45         SCAN_SWAP_CACHE_PAGE,
46         SCAN_DEL_PAGE_LRU,
47         SCAN_ALLOC_HUGE_PAGE_FAIL,
48         SCAN_CGROUP_CHARGE_FAIL,
49         SCAN_EXCEED_SWAP_PTE,
50         SCAN_TRUNCATED,
51 };
52
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/huge_memory.h>
55
56 /* default scan 8*512 pte (or vmas) every 30 second */
57 static unsigned int khugepaged_pages_to_scan __read_mostly;
58 static unsigned int khugepaged_pages_collapsed;
59 static unsigned int khugepaged_full_scans;
60 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
61 /* during fragmentation poll the hugepage allocator once every minute */
62 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
63 static unsigned long khugepaged_sleep_expire;
64 static DEFINE_SPINLOCK(khugepaged_mm_lock);
65 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
66 /*
67  * default collapse hugepages if there is at least one pte mapped like
68  * it would have happened if the vma was large enough during page
69  * fault.
70  */
71 static unsigned int khugepaged_max_ptes_none __read_mostly;
72 static unsigned int khugepaged_max_ptes_swap __read_mostly;
73
74 #define MM_SLOTS_HASH_BITS 10
75 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
76
77 static struct kmem_cache *mm_slot_cache __read_mostly;
78
79 /**
80  * struct mm_slot - hash lookup from mm to mm_slot
81  * @hash: hash collision list
82  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
83  * @mm: the mm that this information is valid for
84  */
85 struct mm_slot {
86         struct hlist_node hash;
87         struct list_head mm_node;
88         struct mm_struct *mm;
89 };
90
91 /**
92  * struct khugepaged_scan - cursor for scanning
93  * @mm_head: the head of the mm list to scan
94  * @mm_slot: the current mm_slot we are scanning
95  * @address: the next address inside that to be scanned
96  *
97  * There is only the one khugepaged_scan instance of this cursor structure.
98  */
99 struct khugepaged_scan {
100         struct list_head mm_head;
101         struct mm_slot *mm_slot;
102         unsigned long address;
103 };
104
105 static struct khugepaged_scan khugepaged_scan = {
106         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
107 };
108
109 #ifdef CONFIG_SYSFS
110 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
111                                          struct kobj_attribute *attr,
112                                          char *buf)
113 {
114         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
115 }
116
117 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
118                                           struct kobj_attribute *attr,
119                                           const char *buf, size_t count)
120 {
121         unsigned long msecs;
122         int err;
123
124         err = kstrtoul(buf, 10, &msecs);
125         if (err || msecs > UINT_MAX)
126                 return -EINVAL;
127
128         khugepaged_scan_sleep_millisecs = msecs;
129         khugepaged_sleep_expire = 0;
130         wake_up_interruptible(&khugepaged_wait);
131
132         return count;
133 }
134 static struct kobj_attribute scan_sleep_millisecs_attr =
135         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
136                scan_sleep_millisecs_store);
137
138 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
139                                           struct kobj_attribute *attr,
140                                           char *buf)
141 {
142         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
143 }
144
145 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
146                                            struct kobj_attribute *attr,
147                                            const char *buf, size_t count)
148 {
149         unsigned long msecs;
150         int err;
151
152         err = kstrtoul(buf, 10, &msecs);
153         if (err || msecs > UINT_MAX)
154                 return -EINVAL;
155
156         khugepaged_alloc_sleep_millisecs = msecs;
157         khugepaged_sleep_expire = 0;
158         wake_up_interruptible(&khugepaged_wait);
159
160         return count;
161 }
162 static struct kobj_attribute alloc_sleep_millisecs_attr =
163         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
164                alloc_sleep_millisecs_store);
165
166 static ssize_t pages_to_scan_show(struct kobject *kobj,
167                                   struct kobj_attribute *attr,
168                                   char *buf)
169 {
170         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
171 }
172 static ssize_t pages_to_scan_store(struct kobject *kobj,
173                                    struct kobj_attribute *attr,
174                                    const char *buf, size_t count)
175 {
176         int err;
177         unsigned long pages;
178
179         err = kstrtoul(buf, 10, &pages);
180         if (err || !pages || pages > UINT_MAX)
181                 return -EINVAL;
182
183         khugepaged_pages_to_scan = pages;
184
185         return count;
186 }
187 static struct kobj_attribute pages_to_scan_attr =
188         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
189                pages_to_scan_store);
190
191 static ssize_t pages_collapsed_show(struct kobject *kobj,
192                                     struct kobj_attribute *attr,
193                                     char *buf)
194 {
195         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
196 }
197 static struct kobj_attribute pages_collapsed_attr =
198         __ATTR_RO(pages_collapsed);
199
200 static ssize_t full_scans_show(struct kobject *kobj,
201                                struct kobj_attribute *attr,
202                                char *buf)
203 {
204         return sprintf(buf, "%u\n", khugepaged_full_scans);
205 }
206 static struct kobj_attribute full_scans_attr =
207         __ATTR_RO(full_scans);
208
209 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
210                                       struct kobj_attribute *attr, char *buf)
211 {
212         return single_hugepage_flag_show(kobj, attr, buf,
213                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
214 }
215 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
216                                        struct kobj_attribute *attr,
217                                        const char *buf, size_t count)
218 {
219         return single_hugepage_flag_store(kobj, attr, buf, count,
220                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
221 }
222 static struct kobj_attribute khugepaged_defrag_attr =
223         __ATTR(defrag, 0644, khugepaged_defrag_show,
224                khugepaged_defrag_store);
225
226 /*
227  * max_ptes_none controls if khugepaged should collapse hugepages over
228  * any unmapped ptes in turn potentially increasing the memory
229  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
230  * reduce the available free memory in the system as it
231  * runs. Increasing max_ptes_none will instead potentially reduce the
232  * free memory in the system during the khugepaged scan.
233  */
234 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
235                                              struct kobj_attribute *attr,
236                                              char *buf)
237 {
238         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
239 }
240 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
241                                               struct kobj_attribute *attr,
242                                               const char *buf, size_t count)
243 {
244         int err;
245         unsigned long max_ptes_none;
246
247         err = kstrtoul(buf, 10, &max_ptes_none);
248         if (err || max_ptes_none > HPAGE_PMD_NR-1)
249                 return -EINVAL;
250
251         khugepaged_max_ptes_none = max_ptes_none;
252
253         return count;
254 }
255 static struct kobj_attribute khugepaged_max_ptes_none_attr =
256         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
257                khugepaged_max_ptes_none_store);
258
259 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
260                                              struct kobj_attribute *attr,
261                                              char *buf)
262 {
263         return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
264 }
265
266 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
267                                               struct kobj_attribute *attr,
268                                               const char *buf, size_t count)
269 {
270         int err;
271         unsigned long max_ptes_swap;
272
273         err  = kstrtoul(buf, 10, &max_ptes_swap);
274         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
275                 return -EINVAL;
276
277         khugepaged_max_ptes_swap = max_ptes_swap;
278
279         return count;
280 }
281
282 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
283         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
284                khugepaged_max_ptes_swap_store);
285
286 static struct attribute *khugepaged_attr[] = {
287         &khugepaged_defrag_attr.attr,
288         &khugepaged_max_ptes_none_attr.attr,
289         &pages_to_scan_attr.attr,
290         &pages_collapsed_attr.attr,
291         &full_scans_attr.attr,
292         &scan_sleep_millisecs_attr.attr,
293         &alloc_sleep_millisecs_attr.attr,
294         &khugepaged_max_ptes_swap_attr.attr,
295         NULL,
296 };
297
298 struct attribute_group khugepaged_attr_group = {
299         .attrs = khugepaged_attr,
300         .name = "khugepaged",
301 };
302 #endif /* CONFIG_SYSFS */
303
304 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
305
306 int hugepage_madvise(struct vm_area_struct *vma,
307                      unsigned long *vm_flags, int advice)
308 {
309         switch (advice) {
310         case MADV_HUGEPAGE:
311 #ifdef CONFIG_S390
312                 /*
313                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
314                  * can't handle this properly after s390_enable_sie, so we simply
315                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
316                  */
317                 if (mm_has_pgste(vma->vm_mm))
318                         return 0;
319 #endif
320                 *vm_flags &= ~VM_NOHUGEPAGE;
321                 *vm_flags |= VM_HUGEPAGE;
322                 /*
323                  * If the vma become good for khugepaged to scan,
324                  * register it here without waiting a page fault that
325                  * may not happen any time soon.
326                  */
327                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
328                                 khugepaged_enter_vma_merge(vma, *vm_flags))
329                         return -ENOMEM;
330                 break;
331         case MADV_NOHUGEPAGE:
332                 *vm_flags &= ~VM_HUGEPAGE;
333                 *vm_flags |= VM_NOHUGEPAGE;
334                 /*
335                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
336                  * this vma even if we leave the mm registered in khugepaged if
337                  * it got registered before VM_NOHUGEPAGE was set.
338                  */
339                 break;
340         }
341
342         return 0;
343 }
344
345 int __init khugepaged_init(void)
346 {
347         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
348                                           sizeof(struct mm_slot),
349                                           __alignof__(struct mm_slot), 0, NULL);
350         if (!mm_slot_cache)
351                 return -ENOMEM;
352
353         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
354         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
355         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
356
357         return 0;
358 }
359
360 void __init khugepaged_destroy(void)
361 {
362         kmem_cache_destroy(mm_slot_cache);
363 }
364
365 static inline struct mm_slot *alloc_mm_slot(void)
366 {
367         if (!mm_slot_cache)     /* initialization failed */
368                 return NULL;
369         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
370 }
371
372 static inline void free_mm_slot(struct mm_slot *mm_slot)
373 {
374         kmem_cache_free(mm_slot_cache, mm_slot);
375 }
376
377 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
378 {
379         struct mm_slot *mm_slot;
380
381         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
382                 if (mm == mm_slot->mm)
383                         return mm_slot;
384
385         return NULL;
386 }
387
388 static void insert_to_mm_slots_hash(struct mm_struct *mm,
389                                     struct mm_slot *mm_slot)
390 {
391         mm_slot->mm = mm;
392         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
393 }
394
395 static inline int khugepaged_test_exit(struct mm_struct *mm)
396 {
397         return atomic_read(&mm->mm_users) == 0;
398 }
399
400 static bool hugepage_vma_check(struct vm_area_struct *vma,
401                                unsigned long vm_flags)
402 {
403         if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
404             (vm_flags & VM_NOHUGEPAGE) ||
405             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
406                 return false;
407         if (shmem_file(vma->vm_file)) {
408                 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
409                         return false;
410                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
411                                 HPAGE_PMD_NR);
412         }
413         if (!vma->anon_vma || vma->vm_ops)
414                 return false;
415         if (is_vma_temporary_stack(vma))
416                 return false;
417         return !(vm_flags & VM_NO_KHUGEPAGED);
418 }
419
420 int __khugepaged_enter(struct mm_struct *mm)
421 {
422         struct mm_slot *mm_slot;
423         int wakeup;
424
425         mm_slot = alloc_mm_slot();
426         if (!mm_slot)
427                 return -ENOMEM;
428
429         /* __khugepaged_exit() must not run from under us */
430         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
431         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
432                 free_mm_slot(mm_slot);
433                 return 0;
434         }
435
436         spin_lock(&khugepaged_mm_lock);
437         insert_to_mm_slots_hash(mm, mm_slot);
438         /*
439          * Insert just behind the scanning cursor, to let the area settle
440          * down a little.
441          */
442         wakeup = list_empty(&khugepaged_scan.mm_head);
443         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
444         spin_unlock(&khugepaged_mm_lock);
445
446         mmgrab(mm);
447         if (wakeup)
448                 wake_up_interruptible(&khugepaged_wait);
449
450         return 0;
451 }
452
453 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
454                                unsigned long vm_flags)
455 {
456         unsigned long hstart, hend;
457
458         /*
459          * khugepaged does not yet work on non-shmem files or special
460          * mappings. And file-private shmem THP is not supported.
461          */
462         if (!hugepage_vma_check(vma, vm_flags))
463                 return 0;
464
465         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
466         hend = vma->vm_end & HPAGE_PMD_MASK;
467         if (hstart < hend)
468                 return khugepaged_enter(vma, vm_flags);
469         return 0;
470 }
471
472 void __khugepaged_exit(struct mm_struct *mm)
473 {
474         struct mm_slot *mm_slot;
475         int free = 0;
476
477         spin_lock(&khugepaged_mm_lock);
478         mm_slot = get_mm_slot(mm);
479         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
480                 hash_del(&mm_slot->hash);
481                 list_del(&mm_slot->mm_node);
482                 free = 1;
483         }
484         spin_unlock(&khugepaged_mm_lock);
485
486         if (free) {
487                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
488                 free_mm_slot(mm_slot);
489                 mmdrop(mm);
490         } else if (mm_slot) {
491                 /*
492                  * This is required to serialize against
493                  * khugepaged_test_exit() (which is guaranteed to run
494                  * under mmap sem read mode). Stop here (after we
495                  * return all pagetables will be destroyed) until
496                  * khugepaged has finished working on the pagetables
497                  * under the mmap_sem.
498                  */
499                 down_write(&mm->mmap_sem);
500                 up_write(&mm->mmap_sem);
501         }
502 }
503
504 static void release_pte_page(struct page *page)
505 {
506         dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
507         unlock_page(page);
508         putback_lru_page(page);
509 }
510
511 static void release_pte_pages(pte_t *pte, pte_t *_pte)
512 {
513         while (--_pte >= pte) {
514                 pte_t pteval = *_pte;
515                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
516                         release_pte_page(pte_page(pteval));
517         }
518 }
519
520 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
521                                         unsigned long address,
522                                         pte_t *pte)
523 {
524         struct page *page = NULL;
525         pte_t *_pte;
526         int none_or_zero = 0, result = 0, referenced = 0;
527         bool writable = false;
528
529         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
530              _pte++, address += PAGE_SIZE) {
531                 pte_t pteval = *_pte;
532                 if (pte_none(pteval) || (pte_present(pteval) &&
533                                 is_zero_pfn(pte_pfn(pteval)))) {
534                         if (!userfaultfd_armed(vma) &&
535                             ++none_or_zero <= khugepaged_max_ptes_none) {
536                                 continue;
537                         } else {
538                                 result = SCAN_EXCEED_NONE_PTE;
539                                 goto out;
540                         }
541                 }
542                 if (!pte_present(pteval)) {
543                         result = SCAN_PTE_NON_PRESENT;
544                         goto out;
545                 }
546                 page = vm_normal_page(vma, address, pteval);
547                 if (unlikely(!page)) {
548                         result = SCAN_PAGE_NULL;
549                         goto out;
550                 }
551
552                 /* TODO: teach khugepaged to collapse THP mapped with pte */
553                 if (PageCompound(page)) {
554                         result = SCAN_PAGE_COMPOUND;
555                         goto out;
556                 }
557
558                 VM_BUG_ON_PAGE(!PageAnon(page), page);
559
560                 /*
561                  * We can do it before isolate_lru_page because the
562                  * page can't be freed from under us. NOTE: PG_lock
563                  * is needed to serialize against split_huge_page
564                  * when invoked from the VM.
565                  */
566                 if (!trylock_page(page)) {
567                         result = SCAN_PAGE_LOCK;
568                         goto out;
569                 }
570
571                 /*
572                  * cannot use mapcount: can't collapse if there's a gup pin.
573                  * The page must only be referenced by the scanned process
574                  * and page swap cache.
575                  */
576                 if (page_count(page) != 1 + PageSwapCache(page)) {
577                         unlock_page(page);
578                         result = SCAN_PAGE_COUNT;
579                         goto out;
580                 }
581                 if (pte_write(pteval)) {
582                         writable = true;
583                 } else {
584                         if (PageSwapCache(page) &&
585                             !reuse_swap_page(page, NULL)) {
586                                 unlock_page(page);
587                                 result = SCAN_SWAP_CACHE_PAGE;
588                                 goto out;
589                         }
590                         /*
591                          * Page is not in the swap cache. It can be collapsed
592                          * into a THP.
593                          */
594                 }
595
596                 /*
597                  * Isolate the page to avoid collapsing an hugepage
598                  * currently in use by the VM.
599                  */
600                 if (isolate_lru_page(page)) {
601                         unlock_page(page);
602                         result = SCAN_DEL_PAGE_LRU;
603                         goto out;
604                 }
605                 inc_node_page_state(page,
606                                 NR_ISOLATED_ANON + page_is_file_cache(page));
607                 VM_BUG_ON_PAGE(!PageLocked(page), page);
608                 VM_BUG_ON_PAGE(PageLRU(page), page);
609
610                 /* There should be enough young pte to collapse the page */
611                 if (pte_young(pteval) ||
612                     page_is_young(page) || PageReferenced(page) ||
613                     mmu_notifier_test_young(vma->vm_mm, address))
614                         referenced++;
615         }
616         if (likely(writable)) {
617                 if (likely(referenced)) {
618                         result = SCAN_SUCCEED;
619                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
620                                                             referenced, writable, result);
621                         return 1;
622                 }
623         } else {
624                 result = SCAN_PAGE_RO;
625         }
626
627 out:
628         release_pte_pages(pte, _pte);
629         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
630                                             referenced, writable, result);
631         return 0;
632 }
633
634 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
635                                       struct vm_area_struct *vma,
636                                       unsigned long address,
637                                       spinlock_t *ptl)
638 {
639         pte_t *_pte;
640         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
641                                 _pte++, page++, address += PAGE_SIZE) {
642                 pte_t pteval = *_pte;
643                 struct page *src_page;
644
645                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
646                         clear_user_highpage(page, address);
647                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
648                         if (is_zero_pfn(pte_pfn(pteval))) {
649                                 /*
650                                  * ptl mostly unnecessary.
651                                  */
652                                 spin_lock(ptl);
653                                 /*
654                                  * paravirt calls inside pte_clear here are
655                                  * superfluous.
656                                  */
657                                 pte_clear(vma->vm_mm, address, _pte);
658                                 spin_unlock(ptl);
659                         }
660                 } else {
661                         src_page = pte_page(pteval);
662                         copy_user_highpage(page, src_page, address, vma);
663                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
664                         release_pte_page(src_page);
665                         /*
666                          * ptl mostly unnecessary, but preempt has to
667                          * be disabled to update the per-cpu stats
668                          * inside page_remove_rmap().
669                          */
670                         spin_lock(ptl);
671                         /*
672                          * paravirt calls inside pte_clear here are
673                          * superfluous.
674                          */
675                         pte_clear(vma->vm_mm, address, _pte);
676                         page_remove_rmap(src_page, false);
677                         spin_unlock(ptl);
678                         free_page_and_swap_cache(src_page);
679                 }
680         }
681 }
682
683 static void khugepaged_alloc_sleep(void)
684 {
685         DEFINE_WAIT(wait);
686
687         add_wait_queue(&khugepaged_wait, &wait);
688         freezable_schedule_timeout_interruptible(
689                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
690         remove_wait_queue(&khugepaged_wait, &wait);
691 }
692
693 static int khugepaged_node_load[MAX_NUMNODES];
694
695 static bool khugepaged_scan_abort(int nid)
696 {
697         int i;
698
699         /*
700          * If node_reclaim_mode is disabled, then no extra effort is made to
701          * allocate memory locally.
702          */
703         if (!node_reclaim_mode)
704                 return false;
705
706         /* If there is a count for this node already, it must be acceptable */
707         if (khugepaged_node_load[nid])
708                 return false;
709
710         for (i = 0; i < MAX_NUMNODES; i++) {
711                 if (!khugepaged_node_load[i])
712                         continue;
713                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
714                         return true;
715         }
716         return false;
717 }
718
719 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
720 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
721 {
722         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
723 }
724
725 #ifdef CONFIG_NUMA
726 static int khugepaged_find_target_node(void)
727 {
728         static int last_khugepaged_target_node = NUMA_NO_NODE;
729         int nid, target_node = 0, max_value = 0;
730
731         /* find first node with max normal pages hit */
732         for (nid = 0; nid < MAX_NUMNODES; nid++)
733                 if (khugepaged_node_load[nid] > max_value) {
734                         max_value = khugepaged_node_load[nid];
735                         target_node = nid;
736                 }
737
738         /* do some balance if several nodes have the same hit record */
739         if (target_node <= last_khugepaged_target_node)
740                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
741                                 nid++)
742                         if (max_value == khugepaged_node_load[nid]) {
743                                 target_node = nid;
744                                 break;
745                         }
746
747         last_khugepaged_target_node = target_node;
748         return target_node;
749 }
750
751 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
752 {
753         if (IS_ERR(*hpage)) {
754                 if (!*wait)
755                         return false;
756
757                 *wait = false;
758                 *hpage = NULL;
759                 khugepaged_alloc_sleep();
760         } else if (*hpage) {
761                 put_page(*hpage);
762                 *hpage = NULL;
763         }
764
765         return true;
766 }
767
768 static struct page *
769 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
770 {
771         VM_BUG_ON_PAGE(*hpage, *hpage);
772
773         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
774         if (unlikely(!*hpage)) {
775                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
776                 *hpage = ERR_PTR(-ENOMEM);
777                 return NULL;
778         }
779
780         prep_transhuge_page(*hpage);
781         count_vm_event(THP_COLLAPSE_ALLOC);
782         return *hpage;
783 }
784 #else
785 static int khugepaged_find_target_node(void)
786 {
787         return 0;
788 }
789
790 static inline struct page *alloc_khugepaged_hugepage(void)
791 {
792         struct page *page;
793
794         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
795                            HPAGE_PMD_ORDER);
796         if (page)
797                 prep_transhuge_page(page);
798         return page;
799 }
800
801 static struct page *khugepaged_alloc_hugepage(bool *wait)
802 {
803         struct page *hpage;
804
805         do {
806                 hpage = alloc_khugepaged_hugepage();
807                 if (!hpage) {
808                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
809                         if (!*wait)
810                                 return NULL;
811
812                         *wait = false;
813                         khugepaged_alloc_sleep();
814                 } else
815                         count_vm_event(THP_COLLAPSE_ALLOC);
816         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
817
818         return hpage;
819 }
820
821 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
822 {
823         if (!*hpage)
824                 *hpage = khugepaged_alloc_hugepage(wait);
825
826         if (unlikely(!*hpage))
827                 return false;
828
829         return true;
830 }
831
832 static struct page *
833 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
834 {
835         VM_BUG_ON(!*hpage);
836
837         return  *hpage;
838 }
839 #endif
840
841 /*
842  * If mmap_sem temporarily dropped, revalidate vma
843  * before taking mmap_sem.
844  * Return 0 if succeeds, otherwise return none-zero
845  * value (scan code).
846  */
847
848 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
849                 struct vm_area_struct **vmap)
850 {
851         struct vm_area_struct *vma;
852         unsigned long hstart, hend;
853
854         if (unlikely(khugepaged_test_exit(mm)))
855                 return SCAN_ANY_PROCESS;
856
857         *vmap = vma = find_vma(mm, address);
858         if (!vma)
859                 return SCAN_VMA_NULL;
860
861         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
862         hend = vma->vm_end & HPAGE_PMD_MASK;
863         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
864                 return SCAN_ADDRESS_RANGE;
865         if (!hugepage_vma_check(vma, vma->vm_flags))
866                 return SCAN_VMA_CHECK;
867         return 0;
868 }
869
870 /*
871  * Bring missing pages in from swap, to complete THP collapse.
872  * Only done if khugepaged_scan_pmd believes it is worthwhile.
873  *
874  * Called and returns without pte mapped or spinlocks held,
875  * but with mmap_sem held to protect against vma changes.
876  */
877
878 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
879                                         struct vm_area_struct *vma,
880                                         unsigned long address, pmd_t *pmd,
881                                         int referenced)
882 {
883         int swapped_in = 0;
884         vm_fault_t ret = 0;
885         struct vm_fault vmf = {
886                 .vma = vma,
887                 .address = address,
888                 .flags = FAULT_FLAG_ALLOW_RETRY,
889                 .pmd = pmd,
890                 .pgoff = linear_page_index(vma, address),
891         };
892
893         /* we only decide to swapin, if there is enough young ptes */
894         if (referenced < HPAGE_PMD_NR/2) {
895                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
896                 return false;
897         }
898         vmf.pte = pte_offset_map(pmd, address);
899         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
900                         vmf.pte++, vmf.address += PAGE_SIZE) {
901                 vmf.orig_pte = *vmf.pte;
902                 if (!is_swap_pte(vmf.orig_pte))
903                         continue;
904                 swapped_in++;
905                 ret = do_swap_page(&vmf);
906
907                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
908                 if (ret & VM_FAULT_RETRY) {
909                         down_read(&mm->mmap_sem);
910                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
911                                 /* vma is no longer available, don't continue to swapin */
912                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
913                                 return false;
914                         }
915                         /* check if the pmd is still valid */
916                         if (mm_find_pmd(mm, address) != pmd) {
917                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
918                                 return false;
919                         }
920                 }
921                 if (ret & VM_FAULT_ERROR) {
922                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
923                         return false;
924                 }
925                 /* pte is unmapped now, we need to map it */
926                 vmf.pte = pte_offset_map(pmd, vmf.address);
927         }
928         vmf.pte--;
929         pte_unmap(vmf.pte);
930         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
931         return true;
932 }
933
934 static void collapse_huge_page(struct mm_struct *mm,
935                                    unsigned long address,
936                                    struct page **hpage,
937                                    int node, int referenced)
938 {
939         pmd_t *pmd, _pmd;
940         pte_t *pte;
941         pgtable_t pgtable;
942         struct page *new_page;
943         spinlock_t *pmd_ptl, *pte_ptl;
944         int isolated = 0, result = 0;
945         struct mem_cgroup *memcg;
946         struct vm_area_struct *vma;
947         struct mmu_notifier_range range;
948         gfp_t gfp;
949
950         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
951
952         /* Only allocate from the target node */
953         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
954
955         /*
956          * Before allocating the hugepage, release the mmap_sem read lock.
957          * The allocation can take potentially a long time if it involves
958          * sync compaction, and we do not need to hold the mmap_sem during
959          * that. We will recheck the vma after taking it again in write mode.
960          */
961         up_read(&mm->mmap_sem);
962         new_page = khugepaged_alloc_page(hpage, gfp, node);
963         if (!new_page) {
964                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
965                 goto out_nolock;
966         }
967
968         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
969                 result = SCAN_CGROUP_CHARGE_FAIL;
970                 goto out_nolock;
971         }
972
973         down_read(&mm->mmap_sem);
974         result = hugepage_vma_revalidate(mm, address, &vma);
975         if (result) {
976                 mem_cgroup_cancel_charge(new_page, memcg, true);
977                 up_read(&mm->mmap_sem);
978                 goto out_nolock;
979         }
980
981         pmd = mm_find_pmd(mm, address);
982         if (!pmd) {
983                 result = SCAN_PMD_NULL;
984                 mem_cgroup_cancel_charge(new_page, memcg, true);
985                 up_read(&mm->mmap_sem);
986                 goto out_nolock;
987         }
988
989         /*
990          * __collapse_huge_page_swapin always returns with mmap_sem locked.
991          * If it fails, we release mmap_sem and jump out_nolock.
992          * Continuing to collapse causes inconsistency.
993          */
994         if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
995                 mem_cgroup_cancel_charge(new_page, memcg, true);
996                 up_read(&mm->mmap_sem);
997                 goto out_nolock;
998         }
999
1000         up_read(&mm->mmap_sem);
1001         /*
1002          * Prevent all access to pagetables with the exception of
1003          * gup_fast later handled by the ptep_clear_flush and the VM
1004          * handled by the anon_vma lock + PG_lock.
1005          */
1006         down_write(&mm->mmap_sem);
1007         result = hugepage_vma_revalidate(mm, address, &vma);
1008         if (result)
1009                 goto out;
1010         /* check if the pmd is still valid */
1011         if (mm_find_pmd(mm, address) != pmd)
1012                 goto out;
1013
1014         anon_vma_lock_write(vma->anon_vma);
1015
1016         pte = pte_offset_map(pmd, address);
1017         pte_ptl = pte_lockptr(mm, pmd);
1018
1019         mmu_notifier_range_init(&range, mm, address, address + HPAGE_PMD_SIZE);
1020         mmu_notifier_invalidate_range_start(&range);
1021         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1022         /*
1023          * After this gup_fast can't run anymore. This also removes
1024          * any huge TLB entry from the CPU so we won't allow
1025          * huge and small TLB entries for the same virtual address
1026          * to avoid the risk of CPU bugs in that area.
1027          */
1028         _pmd = pmdp_collapse_flush(vma, address, pmd);
1029         spin_unlock(pmd_ptl);
1030         mmu_notifier_invalidate_range_end(&range);
1031
1032         spin_lock(pte_ptl);
1033         isolated = __collapse_huge_page_isolate(vma, address, pte);
1034         spin_unlock(pte_ptl);
1035
1036         if (unlikely(!isolated)) {
1037                 pte_unmap(pte);
1038                 spin_lock(pmd_ptl);
1039                 BUG_ON(!pmd_none(*pmd));
1040                 /*
1041                  * We can only use set_pmd_at when establishing
1042                  * hugepmds and never for establishing regular pmds that
1043                  * points to regular pagetables. Use pmd_populate for that
1044                  */
1045                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1046                 spin_unlock(pmd_ptl);
1047                 anon_vma_unlock_write(vma->anon_vma);
1048                 result = SCAN_FAIL;
1049                 goto out;
1050         }
1051
1052         /*
1053          * All pages are isolated and locked so anon_vma rmap
1054          * can't run anymore.
1055          */
1056         anon_vma_unlock_write(vma->anon_vma);
1057
1058         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1059         pte_unmap(pte);
1060         __SetPageUptodate(new_page);
1061         pgtable = pmd_pgtable(_pmd);
1062
1063         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1064         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1065
1066         /*
1067          * spin_lock() below is not the equivalent of smp_wmb(), so
1068          * this is needed to avoid the copy_huge_page writes to become
1069          * visible after the set_pmd_at() write.
1070          */
1071         smp_wmb();
1072
1073         spin_lock(pmd_ptl);
1074         BUG_ON(!pmd_none(*pmd));
1075         page_add_new_anon_rmap(new_page, vma, address, true);
1076         mem_cgroup_commit_charge(new_page, memcg, false, true);
1077         lru_cache_add_active_or_unevictable(new_page, vma);
1078         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1079         set_pmd_at(mm, address, pmd, _pmd);
1080         update_mmu_cache_pmd(vma, address, pmd);
1081         spin_unlock(pmd_ptl);
1082
1083         *hpage = NULL;
1084
1085         khugepaged_pages_collapsed++;
1086         result = SCAN_SUCCEED;
1087 out_up_write:
1088         up_write(&mm->mmap_sem);
1089 out_nolock:
1090         trace_mm_collapse_huge_page(mm, isolated, result);
1091         return;
1092 out:
1093         mem_cgroup_cancel_charge(new_page, memcg, true);
1094         goto out_up_write;
1095 }
1096
1097 static int khugepaged_scan_pmd(struct mm_struct *mm,
1098                                struct vm_area_struct *vma,
1099                                unsigned long address,
1100                                struct page **hpage)
1101 {
1102         pmd_t *pmd;
1103         pte_t *pte, *_pte;
1104         int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1105         struct page *page = NULL;
1106         unsigned long _address;
1107         spinlock_t *ptl;
1108         int node = NUMA_NO_NODE, unmapped = 0;
1109         bool writable = false;
1110
1111         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1112
1113         pmd = mm_find_pmd(mm, address);
1114         if (!pmd) {
1115                 result = SCAN_PMD_NULL;
1116                 goto out;
1117         }
1118
1119         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1120         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1121         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1122              _pte++, _address += PAGE_SIZE) {
1123                 pte_t pteval = *_pte;
1124                 if (is_swap_pte(pteval)) {
1125                         if (++unmapped <= khugepaged_max_ptes_swap) {
1126                                 continue;
1127                         } else {
1128                                 result = SCAN_EXCEED_SWAP_PTE;
1129                                 goto out_unmap;
1130                         }
1131                 }
1132                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1133                         if (!userfaultfd_armed(vma) &&
1134                             ++none_or_zero <= khugepaged_max_ptes_none) {
1135                                 continue;
1136                         } else {
1137                                 result = SCAN_EXCEED_NONE_PTE;
1138                                 goto out_unmap;
1139                         }
1140                 }
1141                 if (!pte_present(pteval)) {
1142                         result = SCAN_PTE_NON_PRESENT;
1143                         goto out_unmap;
1144                 }
1145                 if (pte_write(pteval))
1146                         writable = true;
1147
1148                 page = vm_normal_page(vma, _address, pteval);
1149                 if (unlikely(!page)) {
1150                         result = SCAN_PAGE_NULL;
1151                         goto out_unmap;
1152                 }
1153
1154                 /* TODO: teach khugepaged to collapse THP mapped with pte */
1155                 if (PageCompound(page)) {
1156                         result = SCAN_PAGE_COMPOUND;
1157                         goto out_unmap;
1158                 }
1159
1160                 /*
1161                  * Record which node the original page is from and save this
1162                  * information to khugepaged_node_load[].
1163                  * Khupaged will allocate hugepage from the node has the max
1164                  * hit record.
1165                  */
1166                 node = page_to_nid(page);
1167                 if (khugepaged_scan_abort(node)) {
1168                         result = SCAN_SCAN_ABORT;
1169                         goto out_unmap;
1170                 }
1171                 khugepaged_node_load[node]++;
1172                 if (!PageLRU(page)) {
1173                         result = SCAN_PAGE_LRU;
1174                         goto out_unmap;
1175                 }
1176                 if (PageLocked(page)) {
1177                         result = SCAN_PAGE_LOCK;
1178                         goto out_unmap;
1179                 }
1180                 if (!PageAnon(page)) {
1181                         result = SCAN_PAGE_ANON;
1182                         goto out_unmap;
1183                 }
1184
1185                 /*
1186                  * cannot use mapcount: can't collapse if there's a gup pin.
1187                  * The page must only be referenced by the scanned process
1188                  * and page swap cache.
1189                  */
1190                 if (page_count(page) != 1 + PageSwapCache(page)) {
1191                         result = SCAN_PAGE_COUNT;
1192                         goto out_unmap;
1193                 }
1194                 if (pte_young(pteval) ||
1195                     page_is_young(page) || PageReferenced(page) ||
1196                     mmu_notifier_test_young(vma->vm_mm, address))
1197                         referenced++;
1198         }
1199         if (writable) {
1200                 if (referenced) {
1201                         result = SCAN_SUCCEED;
1202                         ret = 1;
1203                 } else {
1204                         result = SCAN_LACK_REFERENCED_PAGE;
1205                 }
1206         } else {
1207                 result = SCAN_PAGE_RO;
1208         }
1209 out_unmap:
1210         pte_unmap_unlock(pte, ptl);
1211         if (ret) {
1212                 node = khugepaged_find_target_node();
1213                 /* collapse_huge_page will return with the mmap_sem released */
1214                 collapse_huge_page(mm, address, hpage, node, referenced);
1215         }
1216 out:
1217         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1218                                      none_or_zero, result, unmapped);
1219         return ret;
1220 }
1221
1222 static void collect_mm_slot(struct mm_slot *mm_slot)
1223 {
1224         struct mm_struct *mm = mm_slot->mm;
1225
1226         lockdep_assert_held(&khugepaged_mm_lock);
1227
1228         if (khugepaged_test_exit(mm)) {
1229                 /* free mm_slot */
1230                 hash_del(&mm_slot->hash);
1231                 list_del(&mm_slot->mm_node);
1232
1233                 /*
1234                  * Not strictly needed because the mm exited already.
1235                  *
1236                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1237                  */
1238
1239                 /* khugepaged_mm_lock actually not necessary for the below */
1240                 free_mm_slot(mm_slot);
1241                 mmdrop(mm);
1242         }
1243 }
1244
1245 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1246 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1247 {
1248         struct vm_area_struct *vma;
1249         unsigned long addr;
1250         pmd_t *pmd, _pmd;
1251
1252         i_mmap_lock_write(mapping);
1253         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1254                 /* probably overkill */
1255                 if (vma->anon_vma)
1256                         continue;
1257                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1258                 if (addr & ~HPAGE_PMD_MASK)
1259                         continue;
1260                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1261                         continue;
1262                 pmd = mm_find_pmd(vma->vm_mm, addr);
1263                 if (!pmd)
1264                         continue;
1265                 /*
1266                  * We need exclusive mmap_sem to retract page table.
1267                  * If trylock fails we would end up with pte-mapped THP after
1268                  * re-fault. Not ideal, but it's more important to not disturb
1269                  * the system too much.
1270                  */
1271                 if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1272                         spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1273                         /* assume page table is clear */
1274                         _pmd = pmdp_collapse_flush(vma, addr, pmd);
1275                         spin_unlock(ptl);
1276                         up_write(&vma->vm_mm->mmap_sem);
1277                         mm_dec_nr_ptes(vma->vm_mm);
1278                         pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1279                 }
1280         }
1281         i_mmap_unlock_write(mapping);
1282 }
1283
1284 /**
1285  * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1286  *
1287  * Basic scheme is simple, details are more complex:
1288  *  - allocate and lock a new huge page;
1289  *  - scan page cache replacing old pages with the new one
1290  *    + swap in pages if necessary;
1291  *    + fill in gaps;
1292  *    + keep old pages around in case rollback is required;
1293  *  - if replacing succeeds:
1294  *    + copy data over;
1295  *    + free old pages;
1296  *    + unlock huge page;
1297  *  - if replacing failed;
1298  *    + put all pages back and unfreeze them;
1299  *    + restore gaps in the page cache;
1300  *    + unlock and free huge page;
1301  */
1302 static void collapse_shmem(struct mm_struct *mm,
1303                 struct address_space *mapping, pgoff_t start,
1304                 struct page **hpage, int node)
1305 {
1306         gfp_t gfp;
1307         struct page *new_page;
1308         struct mem_cgroup *memcg;
1309         pgoff_t index, end = start + HPAGE_PMD_NR;
1310         LIST_HEAD(pagelist);
1311         XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1312         int nr_none = 0, result = SCAN_SUCCEED;
1313
1314         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1315
1316         /* Only allocate from the target node */
1317         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1318
1319         new_page = khugepaged_alloc_page(hpage, gfp, node);
1320         if (!new_page) {
1321                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1322                 goto out;
1323         }
1324
1325         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1326                 result = SCAN_CGROUP_CHARGE_FAIL;
1327                 goto out;
1328         }
1329
1330         /* This will be less messy when we use multi-index entries */
1331         do {
1332                 xas_lock_irq(&xas);
1333                 xas_create_range(&xas);
1334                 if (!xas_error(&xas))
1335                         break;
1336                 xas_unlock_irq(&xas);
1337                 if (!xas_nomem(&xas, GFP_KERNEL)) {
1338                         mem_cgroup_cancel_charge(new_page, memcg, true);
1339                         result = SCAN_FAIL;
1340                         goto out;
1341                 }
1342         } while (1);
1343
1344         __SetPageLocked(new_page);
1345         __SetPageSwapBacked(new_page);
1346         new_page->index = start;
1347         new_page->mapping = mapping;
1348
1349         /*
1350          * At this point the new_page is locked and not up-to-date.
1351          * It's safe to insert it into the page cache, because nobody would
1352          * be able to map it or use it in another way until we unlock it.
1353          */
1354
1355         xas_set(&xas, start);
1356         for (index = start; index < end; index++) {
1357                 struct page *page = xas_next(&xas);
1358
1359                 VM_BUG_ON(index != xas.xa_index);
1360                 if (!page) {
1361                         /*
1362                          * Stop if extent has been truncated or hole-punched,
1363                          * and is now completely empty.
1364                          */
1365                         if (index == start) {
1366                                 if (!xas_next_entry(&xas, end - 1)) {
1367                                         result = SCAN_TRUNCATED;
1368                                         goto xa_locked;
1369                                 }
1370                                 xas_set(&xas, index);
1371                         }
1372                         if (!shmem_charge(mapping->host, 1)) {
1373                                 result = SCAN_FAIL;
1374                                 goto xa_locked;
1375                         }
1376                         xas_store(&xas, new_page + (index % HPAGE_PMD_NR));
1377                         nr_none++;
1378                         continue;
1379                 }
1380
1381                 if (xa_is_value(page) || !PageUptodate(page)) {
1382                         xas_unlock_irq(&xas);
1383                         /* swap in or instantiate fallocated page */
1384                         if (shmem_getpage(mapping->host, index, &page,
1385                                                 SGP_NOHUGE)) {
1386                                 result = SCAN_FAIL;
1387                                 goto xa_unlocked;
1388                         }
1389                 } else if (trylock_page(page)) {
1390                         get_page(page);
1391                         xas_unlock_irq(&xas);
1392                 } else {
1393                         result = SCAN_PAGE_LOCK;
1394                         goto xa_locked;
1395                 }
1396
1397                 /*
1398                  * The page must be locked, so we can drop the i_pages lock
1399                  * without racing with truncate.
1400                  */
1401                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1402                 VM_BUG_ON_PAGE(!PageUptodate(page), page);
1403
1404                 /*
1405                  * If file was truncated then extended, or hole-punched, before
1406                  * we locked the first page, then a THP might be there already.
1407                  */
1408                 if (PageTransCompound(page)) {
1409                         result = SCAN_PAGE_COMPOUND;
1410                         goto out_unlock;
1411                 }
1412
1413                 if (page_mapping(page) != mapping) {
1414                         result = SCAN_TRUNCATED;
1415                         goto out_unlock;
1416                 }
1417
1418                 if (isolate_lru_page(page)) {
1419                         result = SCAN_DEL_PAGE_LRU;
1420                         goto out_unlock;
1421                 }
1422
1423                 if (page_mapped(page))
1424                         unmap_mapping_pages(mapping, index, 1, false);
1425
1426                 xas_lock_irq(&xas);
1427                 xas_set(&xas, index);
1428
1429                 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1430                 VM_BUG_ON_PAGE(page_mapped(page), page);
1431
1432                 /*
1433                  * The page is expected to have page_count() == 3:
1434                  *  - we hold a pin on it;
1435                  *  - one reference from page cache;
1436                  *  - one from isolate_lru_page;
1437                  */
1438                 if (!page_ref_freeze(page, 3)) {
1439                         result = SCAN_PAGE_COUNT;
1440                         xas_unlock_irq(&xas);
1441                         putback_lru_page(page);
1442                         goto out_unlock;
1443                 }
1444
1445                 /*
1446                  * Add the page to the list to be able to undo the collapse if
1447                  * something go wrong.
1448                  */
1449                 list_add_tail(&page->lru, &pagelist);
1450
1451                 /* Finally, replace with the new page. */
1452                 xas_store(&xas, new_page + (index % HPAGE_PMD_NR));
1453                 continue;
1454 out_unlock:
1455                 unlock_page(page);
1456                 put_page(page);
1457                 goto xa_unlocked;
1458         }
1459
1460         __inc_node_page_state(new_page, NR_SHMEM_THPS);
1461         if (nr_none) {
1462                 struct zone *zone = page_zone(new_page);
1463
1464                 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1465                 __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
1466         }
1467
1468 xa_locked:
1469         xas_unlock_irq(&xas);
1470 xa_unlocked:
1471
1472         if (result == SCAN_SUCCEED) {
1473                 struct page *page, *tmp;
1474
1475                 /*
1476                  * Replacing old pages with new one has succeeded, now we
1477                  * need to copy the content and free the old pages.
1478                  */
1479                 index = start;
1480                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1481                         while (index < page->index) {
1482                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1483                                 index++;
1484                         }
1485                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1486                                         page);
1487                         list_del(&page->lru);
1488                         page->mapping = NULL;
1489                         page_ref_unfreeze(page, 1);
1490                         ClearPageActive(page);
1491                         ClearPageUnevictable(page);
1492                         unlock_page(page);
1493                         put_page(page);
1494                         index++;
1495                 }
1496                 while (index < end) {
1497                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1498                         index++;
1499                 }
1500
1501                 SetPageUptodate(new_page);
1502                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1503                 set_page_dirty(new_page);
1504                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1505                 lru_cache_add_anon(new_page);
1506
1507                 /*
1508                  * Remove pte page tables, so we can re-fault the page as huge.
1509                  */
1510                 retract_page_tables(mapping, start);
1511                 *hpage = NULL;
1512
1513                 khugepaged_pages_collapsed++;
1514         } else {
1515                 struct page *page;
1516
1517                 /* Something went wrong: roll back page cache changes */
1518                 xas_lock_irq(&xas);
1519                 mapping->nrpages -= nr_none;
1520                 shmem_uncharge(mapping->host, nr_none);
1521
1522                 xas_set(&xas, start);
1523                 xas_for_each(&xas, page, end - 1) {
1524                         page = list_first_entry_or_null(&pagelist,
1525                                         struct page, lru);
1526                         if (!page || xas.xa_index < page->index) {
1527                                 if (!nr_none)
1528                                         break;
1529                                 nr_none--;
1530                                 /* Put holes back where they were */
1531                                 xas_store(&xas, NULL);
1532                                 continue;
1533                         }
1534
1535                         VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1536
1537                         /* Unfreeze the page. */
1538                         list_del(&page->lru);
1539                         page_ref_unfreeze(page, 2);
1540                         xas_store(&xas, page);
1541                         xas_pause(&xas);
1542                         xas_unlock_irq(&xas);
1543                         unlock_page(page);
1544                         putback_lru_page(page);
1545                         xas_lock_irq(&xas);
1546                 }
1547                 VM_BUG_ON(nr_none);
1548                 xas_unlock_irq(&xas);
1549
1550                 mem_cgroup_cancel_charge(new_page, memcg, true);
1551                 new_page->mapping = NULL;
1552         }
1553
1554         unlock_page(new_page);
1555 out:
1556         VM_BUG_ON(!list_empty(&pagelist));
1557         /* TODO: tracepoints */
1558 }
1559
1560 static void khugepaged_scan_shmem(struct mm_struct *mm,
1561                 struct address_space *mapping,
1562                 pgoff_t start, struct page **hpage)
1563 {
1564         struct page *page = NULL;
1565         XA_STATE(xas, &mapping->i_pages, start);
1566         int present, swap;
1567         int node = NUMA_NO_NODE;
1568         int result = SCAN_SUCCEED;
1569
1570         present = 0;
1571         swap = 0;
1572         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1573         rcu_read_lock();
1574         xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1575                 if (xas_retry(&xas, page))
1576                         continue;
1577
1578                 if (xa_is_value(page)) {
1579                         if (++swap > khugepaged_max_ptes_swap) {
1580                                 result = SCAN_EXCEED_SWAP_PTE;
1581                                 break;
1582                         }
1583                         continue;
1584                 }
1585
1586                 if (PageTransCompound(page)) {
1587                         result = SCAN_PAGE_COMPOUND;
1588                         break;
1589                 }
1590
1591                 node = page_to_nid(page);
1592                 if (khugepaged_scan_abort(node)) {
1593                         result = SCAN_SCAN_ABORT;
1594                         break;
1595                 }
1596                 khugepaged_node_load[node]++;
1597
1598                 if (!PageLRU(page)) {
1599                         result = SCAN_PAGE_LRU;
1600                         break;
1601                 }
1602
1603                 if (page_count(page) != 1 + page_mapcount(page)) {
1604                         result = SCAN_PAGE_COUNT;
1605                         break;
1606                 }
1607
1608                 /*
1609                  * We probably should check if the page is referenced here, but
1610                  * nobody would transfer pte_young() to PageReferenced() for us.
1611                  * And rmap walk here is just too costly...
1612                  */
1613
1614                 present++;
1615
1616                 if (need_resched()) {
1617                         xas_pause(&xas);
1618                         cond_resched_rcu();
1619                 }
1620         }
1621         rcu_read_unlock();
1622
1623         if (result == SCAN_SUCCEED) {
1624                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1625                         result = SCAN_EXCEED_NONE_PTE;
1626                 } else {
1627                         node = khugepaged_find_target_node();
1628                         collapse_shmem(mm, mapping, start, hpage, node);
1629                 }
1630         }
1631
1632         /* TODO: tracepoints */
1633 }
1634 #else
1635 static void khugepaged_scan_shmem(struct mm_struct *mm,
1636                 struct address_space *mapping,
1637                 pgoff_t start, struct page **hpage)
1638 {
1639         BUILD_BUG();
1640 }
1641 #endif
1642
1643 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1644                                             struct page **hpage)
1645         __releases(&khugepaged_mm_lock)
1646         __acquires(&khugepaged_mm_lock)
1647 {
1648         struct mm_slot *mm_slot;
1649         struct mm_struct *mm;
1650         struct vm_area_struct *vma;
1651         int progress = 0;
1652
1653         VM_BUG_ON(!pages);
1654         lockdep_assert_held(&khugepaged_mm_lock);
1655
1656         if (khugepaged_scan.mm_slot)
1657                 mm_slot = khugepaged_scan.mm_slot;
1658         else {
1659                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1660                                      struct mm_slot, mm_node);
1661                 khugepaged_scan.address = 0;
1662                 khugepaged_scan.mm_slot = mm_slot;
1663         }
1664         spin_unlock(&khugepaged_mm_lock);
1665
1666         mm = mm_slot->mm;
1667         /*
1668          * Don't wait for semaphore (to avoid long wait times).  Just move to
1669          * the next mm on the list.
1670          */
1671         vma = NULL;
1672         if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1673                 goto breakouterloop_mmap_sem;
1674         if (likely(!khugepaged_test_exit(mm)))
1675                 vma = find_vma(mm, khugepaged_scan.address);
1676
1677         progress++;
1678         for (; vma; vma = vma->vm_next) {
1679                 unsigned long hstart, hend;
1680
1681                 cond_resched();
1682                 if (unlikely(khugepaged_test_exit(mm))) {
1683                         progress++;
1684                         break;
1685                 }
1686                 if (!hugepage_vma_check(vma, vma->vm_flags)) {
1687 skip:
1688                         progress++;
1689                         continue;
1690                 }
1691                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1692                 hend = vma->vm_end & HPAGE_PMD_MASK;
1693                 if (hstart >= hend)
1694                         goto skip;
1695                 if (khugepaged_scan.address > hend)
1696                         goto skip;
1697                 if (khugepaged_scan.address < hstart)
1698                         khugepaged_scan.address = hstart;
1699                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1700
1701                 while (khugepaged_scan.address < hend) {
1702                         int ret;
1703                         cond_resched();
1704                         if (unlikely(khugepaged_test_exit(mm)))
1705                                 goto breakouterloop;
1706
1707                         VM_BUG_ON(khugepaged_scan.address < hstart ||
1708                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
1709                                   hend);
1710                         if (shmem_file(vma->vm_file)) {
1711                                 struct file *file;
1712                                 pgoff_t pgoff = linear_page_index(vma,
1713                                                 khugepaged_scan.address);
1714                                 if (!shmem_huge_enabled(vma))
1715                                         goto skip;
1716                                 file = get_file(vma->vm_file);
1717                                 up_read(&mm->mmap_sem);
1718                                 ret = 1;
1719                                 khugepaged_scan_shmem(mm, file->f_mapping,
1720                                                 pgoff, hpage);
1721                                 fput(file);
1722                         } else {
1723                                 ret = khugepaged_scan_pmd(mm, vma,
1724                                                 khugepaged_scan.address,
1725                                                 hpage);
1726                         }
1727                         /* move to next address */
1728                         khugepaged_scan.address += HPAGE_PMD_SIZE;
1729                         progress += HPAGE_PMD_NR;
1730                         if (ret)
1731                                 /* we released mmap_sem so break loop */
1732                                 goto breakouterloop_mmap_sem;
1733                         if (progress >= pages)
1734                                 goto breakouterloop;
1735                 }
1736         }
1737 breakouterloop:
1738         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1739 breakouterloop_mmap_sem:
1740
1741         spin_lock(&khugepaged_mm_lock);
1742         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1743         /*
1744          * Release the current mm_slot if this mm is about to die, or
1745          * if we scanned all vmas of this mm.
1746          */
1747         if (khugepaged_test_exit(mm) || !vma) {
1748                 /*
1749                  * Make sure that if mm_users is reaching zero while
1750                  * khugepaged runs here, khugepaged_exit will find
1751                  * mm_slot not pointing to the exiting mm.
1752                  */
1753                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1754                         khugepaged_scan.mm_slot = list_entry(
1755                                 mm_slot->mm_node.next,
1756                                 struct mm_slot, mm_node);
1757                         khugepaged_scan.address = 0;
1758                 } else {
1759                         khugepaged_scan.mm_slot = NULL;
1760                         khugepaged_full_scans++;
1761                 }
1762
1763                 collect_mm_slot(mm_slot);
1764         }
1765
1766         return progress;
1767 }
1768
1769 static int khugepaged_has_work(void)
1770 {
1771         return !list_empty(&khugepaged_scan.mm_head) &&
1772                 khugepaged_enabled();
1773 }
1774
1775 static int khugepaged_wait_event(void)
1776 {
1777         return !list_empty(&khugepaged_scan.mm_head) ||
1778                 kthread_should_stop();
1779 }
1780
1781 static void khugepaged_do_scan(void)
1782 {
1783         struct page *hpage = NULL;
1784         unsigned int progress = 0, pass_through_head = 0;
1785         unsigned int pages = khugepaged_pages_to_scan;
1786         bool wait = true;
1787
1788         barrier(); /* write khugepaged_pages_to_scan to local stack */
1789
1790         while (progress < pages) {
1791                 if (!khugepaged_prealloc_page(&hpage, &wait))
1792                         break;
1793
1794                 cond_resched();
1795
1796                 if (unlikely(kthread_should_stop() || try_to_freeze()))
1797                         break;
1798
1799                 spin_lock(&khugepaged_mm_lock);
1800                 if (!khugepaged_scan.mm_slot)
1801                         pass_through_head++;
1802                 if (khugepaged_has_work() &&
1803                     pass_through_head < 2)
1804                         progress += khugepaged_scan_mm_slot(pages - progress,
1805                                                             &hpage);
1806                 else
1807                         progress = pages;
1808                 spin_unlock(&khugepaged_mm_lock);
1809         }
1810
1811         if (!IS_ERR_OR_NULL(hpage))
1812                 put_page(hpage);
1813 }
1814
1815 static bool khugepaged_should_wakeup(void)
1816 {
1817         return kthread_should_stop() ||
1818                time_after_eq(jiffies, khugepaged_sleep_expire);
1819 }
1820
1821 static void khugepaged_wait_work(void)
1822 {
1823         if (khugepaged_has_work()) {
1824                 const unsigned long scan_sleep_jiffies =
1825                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1826
1827                 if (!scan_sleep_jiffies)
1828                         return;
1829
1830                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1831                 wait_event_freezable_timeout(khugepaged_wait,
1832                                              khugepaged_should_wakeup(),
1833                                              scan_sleep_jiffies);
1834                 return;
1835         }
1836
1837         if (khugepaged_enabled())
1838                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1839 }
1840
1841 static int khugepaged(void *none)
1842 {
1843         struct mm_slot *mm_slot;
1844
1845         set_freezable();
1846         set_user_nice(current, MAX_NICE);
1847
1848         while (!kthread_should_stop()) {
1849                 khugepaged_do_scan();
1850                 khugepaged_wait_work();
1851         }
1852
1853         spin_lock(&khugepaged_mm_lock);
1854         mm_slot = khugepaged_scan.mm_slot;
1855         khugepaged_scan.mm_slot = NULL;
1856         if (mm_slot)
1857                 collect_mm_slot(mm_slot);
1858         spin_unlock(&khugepaged_mm_lock);
1859         return 0;
1860 }
1861
1862 static void set_recommended_min_free_kbytes(void)
1863 {
1864         struct zone *zone;
1865         int nr_zones = 0;
1866         unsigned long recommended_min;
1867
1868         for_each_populated_zone(zone) {
1869                 /*
1870                  * We don't need to worry about fragmentation of
1871                  * ZONE_MOVABLE since it only has movable pages.
1872                  */
1873                 if (zone_idx(zone) > gfp_zone(GFP_USER))
1874                         continue;
1875
1876                 nr_zones++;
1877         }
1878
1879         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1880         recommended_min = pageblock_nr_pages * nr_zones * 2;
1881
1882         /*
1883          * Make sure that on average at least two pageblocks are almost free
1884          * of another type, one for a migratetype to fall back to and a
1885          * second to avoid subsequent fallbacks of other types There are 3
1886          * MIGRATE_TYPES we care about.
1887          */
1888         recommended_min += pageblock_nr_pages * nr_zones *
1889                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1890
1891         /* don't ever allow to reserve more than 5% of the lowmem */
1892         recommended_min = min(recommended_min,
1893                               (unsigned long) nr_free_buffer_pages() / 20);
1894         recommended_min <<= (PAGE_SHIFT-10);
1895
1896         if (recommended_min > min_free_kbytes) {
1897                 if (user_min_free_kbytes >= 0)
1898                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1899                                 min_free_kbytes, recommended_min);
1900
1901                 min_free_kbytes = recommended_min;
1902         }
1903         setup_per_zone_wmarks();
1904 }
1905
1906 int start_stop_khugepaged(void)
1907 {
1908         static struct task_struct *khugepaged_thread __read_mostly;
1909         static DEFINE_MUTEX(khugepaged_mutex);
1910         int err = 0;
1911
1912         mutex_lock(&khugepaged_mutex);
1913         if (khugepaged_enabled()) {
1914                 if (!khugepaged_thread)
1915                         khugepaged_thread = kthread_run(khugepaged, NULL,
1916                                                         "khugepaged");
1917                 if (IS_ERR(khugepaged_thread)) {
1918                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1919                         err = PTR_ERR(khugepaged_thread);
1920                         khugepaged_thread = NULL;
1921                         goto fail;
1922                 }
1923
1924                 if (!list_empty(&khugepaged_scan.mm_head))
1925                         wake_up_interruptible(&khugepaged_wait);
1926
1927                 set_recommended_min_free_kbytes();
1928         } else if (khugepaged_thread) {
1929                 kthread_stop(khugepaged_thread);
1930                 khugepaged_thread = NULL;
1931         }
1932 fail:
1933         mutex_unlock(&khugepaged_mutex);
1934         return err;
1935 }