Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma
[linux-2.6-microblaze.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 int __khugepaged_enter(struct mm_struct *mm)
401 {
402         struct mm_slot *mm_slot;
403         int wakeup;
404
405         mm_slot = alloc_mm_slot();
406         if (!mm_slot)
407                 return -ENOMEM;
408
409         /* __khugepaged_exit() must not run from under us */
410         VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
411         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
412                 free_mm_slot(mm_slot);
413                 return 0;
414         }
415
416         spin_lock(&khugepaged_mm_lock);
417         insert_to_mm_slots_hash(mm, mm_slot);
418         /*
419          * Insert just behind the scanning cursor, to let the area settle
420          * down a little.
421          */
422         wakeup = list_empty(&khugepaged_scan.mm_head);
423         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
424         spin_unlock(&khugepaged_mm_lock);
425
426         mmgrab(mm);
427         if (wakeup)
428                 wake_up_interruptible(&khugepaged_wait);
429
430         return 0;
431 }
432
433 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
434                                unsigned long vm_flags)
435 {
436         unsigned long hstart, hend;
437         if (!vma->anon_vma)
438                 /*
439                  * Not yet faulted in so we will register later in the
440                  * page fault if needed.
441                  */
442                 return 0;
443         if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
444                 /* khugepaged not yet working on file or special mappings */
445                 return 0;
446         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
447         hend = vma->vm_end & HPAGE_PMD_MASK;
448         if (hstart < hend)
449                 return khugepaged_enter(vma, vm_flags);
450         return 0;
451 }
452
453 void __khugepaged_exit(struct mm_struct *mm)
454 {
455         struct mm_slot *mm_slot;
456         int free = 0;
457
458         spin_lock(&khugepaged_mm_lock);
459         mm_slot = get_mm_slot(mm);
460         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
461                 hash_del(&mm_slot->hash);
462                 list_del(&mm_slot->mm_node);
463                 free = 1;
464         }
465         spin_unlock(&khugepaged_mm_lock);
466
467         if (free) {
468                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
469                 free_mm_slot(mm_slot);
470                 mmdrop(mm);
471         } else if (mm_slot) {
472                 /*
473                  * This is required to serialize against
474                  * khugepaged_test_exit() (which is guaranteed to run
475                  * under mmap sem read mode). Stop here (after we
476                  * return all pagetables will be destroyed) until
477                  * khugepaged has finished working on the pagetables
478                  * under the mmap_sem.
479                  */
480                 down_write(&mm->mmap_sem);
481                 up_write(&mm->mmap_sem);
482         }
483 }
484
485 static void release_pte_page(struct page *page)
486 {
487         dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
488         unlock_page(page);
489         putback_lru_page(page);
490 }
491
492 static void release_pte_pages(pte_t *pte, pte_t *_pte)
493 {
494         while (--_pte >= pte) {
495                 pte_t pteval = *_pte;
496                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
497                         release_pte_page(pte_page(pteval));
498         }
499 }
500
501 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
502                                         unsigned long address,
503                                         pte_t *pte)
504 {
505         struct page *page = NULL;
506         pte_t *_pte;
507         int none_or_zero = 0, result = 0, referenced = 0;
508         bool writable = false;
509
510         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
511              _pte++, address += PAGE_SIZE) {
512                 pte_t pteval = *_pte;
513                 if (pte_none(pteval) || (pte_present(pteval) &&
514                                 is_zero_pfn(pte_pfn(pteval)))) {
515                         if (!userfaultfd_armed(vma) &&
516                             ++none_or_zero <= khugepaged_max_ptes_none) {
517                                 continue;
518                         } else {
519                                 result = SCAN_EXCEED_NONE_PTE;
520                                 goto out;
521                         }
522                 }
523                 if (!pte_present(pteval)) {
524                         result = SCAN_PTE_NON_PRESENT;
525                         goto out;
526                 }
527                 page = vm_normal_page(vma, address, pteval);
528                 if (unlikely(!page)) {
529                         result = SCAN_PAGE_NULL;
530                         goto out;
531                 }
532
533                 VM_BUG_ON_PAGE(PageCompound(page), page);
534                 VM_BUG_ON_PAGE(!PageAnon(page), page);
535
536                 /*
537                  * We can do it before isolate_lru_page because the
538                  * page can't be freed from under us. NOTE: PG_lock
539                  * is needed to serialize against split_huge_page
540                  * when invoked from the VM.
541                  */
542                 if (!trylock_page(page)) {
543                         result = SCAN_PAGE_LOCK;
544                         goto out;
545                 }
546
547                 /*
548                  * cannot use mapcount: can't collapse if there's a gup pin.
549                  * The page must only be referenced by the scanned process
550                  * and page swap cache.
551                  */
552                 if (page_count(page) != 1 + PageSwapCache(page)) {
553                         unlock_page(page);
554                         result = SCAN_PAGE_COUNT;
555                         goto out;
556                 }
557                 if (pte_write(pteval)) {
558                         writable = true;
559                 } else {
560                         if (PageSwapCache(page) &&
561                             !reuse_swap_page(page, NULL)) {
562                                 unlock_page(page);
563                                 result = SCAN_SWAP_CACHE_PAGE;
564                                 goto out;
565                         }
566                         /*
567                          * Page is not in the swap cache. It can be collapsed
568                          * into a THP.
569                          */
570                 }
571
572                 /*
573                  * Isolate the page to avoid collapsing an hugepage
574                  * currently in use by the VM.
575                  */
576                 if (isolate_lru_page(page)) {
577                         unlock_page(page);
578                         result = SCAN_DEL_PAGE_LRU;
579                         goto out;
580                 }
581                 inc_node_page_state(page,
582                                 NR_ISOLATED_ANON + page_is_file_cache(page));
583                 VM_BUG_ON_PAGE(!PageLocked(page), page);
584                 VM_BUG_ON_PAGE(PageLRU(page), page);
585
586                 /* There should be enough young pte to collapse the page */
587                 if (pte_young(pteval) ||
588                     page_is_young(page) || PageReferenced(page) ||
589                     mmu_notifier_test_young(vma->vm_mm, address))
590                         referenced++;
591         }
592         if (likely(writable)) {
593                 if (likely(referenced)) {
594                         result = SCAN_SUCCEED;
595                         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
596                                                             referenced, writable, result);
597                         return 1;
598                 }
599         } else {
600                 result = SCAN_PAGE_RO;
601         }
602
603 out:
604         release_pte_pages(pte, _pte);
605         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
606                                             referenced, writable, result);
607         return 0;
608 }
609
610 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
611                                       struct vm_area_struct *vma,
612                                       unsigned long address,
613                                       spinlock_t *ptl)
614 {
615         pte_t *_pte;
616         for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
617                                 _pte++, page++, address += PAGE_SIZE) {
618                 pte_t pteval = *_pte;
619                 struct page *src_page;
620
621                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
622                         clear_user_highpage(page, address);
623                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
624                         if (is_zero_pfn(pte_pfn(pteval))) {
625                                 /*
626                                  * ptl mostly unnecessary.
627                                  */
628                                 spin_lock(ptl);
629                                 /*
630                                  * paravirt calls inside pte_clear here are
631                                  * superfluous.
632                                  */
633                                 pte_clear(vma->vm_mm, address, _pte);
634                                 spin_unlock(ptl);
635                         }
636                 } else {
637                         src_page = pte_page(pteval);
638                         copy_user_highpage(page, src_page, address, vma);
639                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
640                         release_pte_page(src_page);
641                         /*
642                          * ptl mostly unnecessary, but preempt has to
643                          * be disabled to update the per-cpu stats
644                          * inside page_remove_rmap().
645                          */
646                         spin_lock(ptl);
647                         /*
648                          * paravirt calls inside pte_clear here are
649                          * superfluous.
650                          */
651                         pte_clear(vma->vm_mm, address, _pte);
652                         page_remove_rmap(src_page, false);
653                         spin_unlock(ptl);
654                         free_page_and_swap_cache(src_page);
655                 }
656         }
657 }
658
659 static void khugepaged_alloc_sleep(void)
660 {
661         DEFINE_WAIT(wait);
662
663         add_wait_queue(&khugepaged_wait, &wait);
664         freezable_schedule_timeout_interruptible(
665                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
666         remove_wait_queue(&khugepaged_wait, &wait);
667 }
668
669 static int khugepaged_node_load[MAX_NUMNODES];
670
671 static bool khugepaged_scan_abort(int nid)
672 {
673         int i;
674
675         /*
676          * If node_reclaim_mode is disabled, then no extra effort is made to
677          * allocate memory locally.
678          */
679         if (!node_reclaim_mode)
680                 return false;
681
682         /* If there is a count for this node already, it must be acceptable */
683         if (khugepaged_node_load[nid])
684                 return false;
685
686         for (i = 0; i < MAX_NUMNODES; i++) {
687                 if (!khugepaged_node_load[i])
688                         continue;
689                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
690                         return true;
691         }
692         return false;
693 }
694
695 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
696 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
697 {
698         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
699 }
700
701 #ifdef CONFIG_NUMA
702 static int khugepaged_find_target_node(void)
703 {
704         static int last_khugepaged_target_node = NUMA_NO_NODE;
705         int nid, target_node = 0, max_value = 0;
706
707         /* find first node with max normal pages hit */
708         for (nid = 0; nid < MAX_NUMNODES; nid++)
709                 if (khugepaged_node_load[nid] > max_value) {
710                         max_value = khugepaged_node_load[nid];
711                         target_node = nid;
712                 }
713
714         /* do some balance if several nodes have the same hit record */
715         if (target_node <= last_khugepaged_target_node)
716                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
717                                 nid++)
718                         if (max_value == khugepaged_node_load[nid]) {
719                                 target_node = nid;
720                                 break;
721                         }
722
723         last_khugepaged_target_node = target_node;
724         return target_node;
725 }
726
727 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
728 {
729         if (IS_ERR(*hpage)) {
730                 if (!*wait)
731                         return false;
732
733                 *wait = false;
734                 *hpage = NULL;
735                 khugepaged_alloc_sleep();
736         } else if (*hpage) {
737                 put_page(*hpage);
738                 *hpage = NULL;
739         }
740
741         return true;
742 }
743
744 static struct page *
745 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
746 {
747         VM_BUG_ON_PAGE(*hpage, *hpage);
748
749         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
750         if (unlikely(!*hpage)) {
751                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
752                 *hpage = ERR_PTR(-ENOMEM);
753                 return NULL;
754         }
755
756         prep_transhuge_page(*hpage);
757         count_vm_event(THP_COLLAPSE_ALLOC);
758         return *hpage;
759 }
760 #else
761 static int khugepaged_find_target_node(void)
762 {
763         return 0;
764 }
765
766 static inline struct page *alloc_khugepaged_hugepage(void)
767 {
768         struct page *page;
769
770         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
771                            HPAGE_PMD_ORDER);
772         if (page)
773                 prep_transhuge_page(page);
774         return page;
775 }
776
777 static struct page *khugepaged_alloc_hugepage(bool *wait)
778 {
779         struct page *hpage;
780
781         do {
782                 hpage = alloc_khugepaged_hugepage();
783                 if (!hpage) {
784                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
785                         if (!*wait)
786                                 return NULL;
787
788                         *wait = false;
789                         khugepaged_alloc_sleep();
790                 } else
791                         count_vm_event(THP_COLLAPSE_ALLOC);
792         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
793
794         return hpage;
795 }
796
797 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
798 {
799         if (!*hpage)
800                 *hpage = khugepaged_alloc_hugepage(wait);
801
802         if (unlikely(!*hpage))
803                 return false;
804
805         return true;
806 }
807
808 static struct page *
809 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
810 {
811         VM_BUG_ON(!*hpage);
812
813         return  *hpage;
814 }
815 #endif
816
817 static bool hugepage_vma_check(struct vm_area_struct *vma)
818 {
819         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
820             (vma->vm_flags & VM_NOHUGEPAGE) ||
821             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
822                 return false;
823         if (shmem_file(vma->vm_file)) {
824                 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
825                         return false;
826                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
827                                 HPAGE_PMD_NR);
828         }
829         if (!vma->anon_vma || vma->vm_ops)
830                 return false;
831         if (is_vma_temporary_stack(vma))
832                 return false;
833         return !(vma->vm_flags & VM_NO_KHUGEPAGED);
834 }
835
836 /*
837  * If mmap_sem temporarily dropped, revalidate vma
838  * before taking mmap_sem.
839  * Return 0 if succeeds, otherwise return none-zero
840  * value (scan code).
841  */
842
843 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
844                 struct vm_area_struct **vmap)
845 {
846         struct vm_area_struct *vma;
847         unsigned long hstart, hend;
848
849         if (unlikely(khugepaged_test_exit(mm)))
850                 return SCAN_ANY_PROCESS;
851
852         *vmap = vma = find_vma(mm, address);
853         if (!vma)
854                 return SCAN_VMA_NULL;
855
856         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
857         hend = vma->vm_end & HPAGE_PMD_MASK;
858         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
859                 return SCAN_ADDRESS_RANGE;
860         if (!hugepage_vma_check(vma))
861                 return SCAN_VMA_CHECK;
862         return 0;
863 }
864
865 /*
866  * Bring missing pages in from swap, to complete THP collapse.
867  * Only done if khugepaged_scan_pmd believes it is worthwhile.
868  *
869  * Called and returns without pte mapped or spinlocks held,
870  * but with mmap_sem held to protect against vma changes.
871  */
872
873 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
874                                         struct vm_area_struct *vma,
875                                         unsigned long address, pmd_t *pmd,
876                                         int referenced)
877 {
878         int swapped_in = 0, ret = 0;
879         struct vm_fault vmf = {
880                 .vma = vma,
881                 .address = address,
882                 .flags = FAULT_FLAG_ALLOW_RETRY,
883                 .pmd = pmd,
884                 .pgoff = linear_page_index(vma, address),
885         };
886
887         /* we only decide to swapin, if there is enough young ptes */
888         if (referenced < HPAGE_PMD_NR/2) {
889                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
890                 return false;
891         }
892         vmf.pte = pte_offset_map(pmd, address);
893         for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
894                         vmf.pte++, vmf.address += PAGE_SIZE) {
895                 vmf.orig_pte = *vmf.pte;
896                 if (!is_swap_pte(vmf.orig_pte))
897                         continue;
898                 swapped_in++;
899                 ret = do_swap_page(&vmf);
900
901                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
902                 if (ret & VM_FAULT_RETRY) {
903                         down_read(&mm->mmap_sem);
904                         if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
905                                 /* vma is no longer available, don't continue to swapin */
906                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
907                                 return false;
908                         }
909                         /* check if the pmd is still valid */
910                         if (mm_find_pmd(mm, address) != pmd) {
911                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
912                                 return false;
913                         }
914                 }
915                 if (ret & VM_FAULT_ERROR) {
916                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
917                         return false;
918                 }
919                 /* pte is unmapped now, we need to map it */
920                 vmf.pte = pte_offset_map(pmd, vmf.address);
921         }
922         vmf.pte--;
923         pte_unmap(vmf.pte);
924         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
925         return true;
926 }
927
928 static void collapse_huge_page(struct mm_struct *mm,
929                                    unsigned long address,
930                                    struct page **hpage,
931                                    int node, int referenced)
932 {
933         pmd_t *pmd, _pmd;
934         pte_t *pte;
935         pgtable_t pgtable;
936         struct page *new_page;
937         spinlock_t *pmd_ptl, *pte_ptl;
938         int isolated = 0, result = 0;
939         struct mem_cgroup *memcg;
940         struct vm_area_struct *vma;
941         unsigned long mmun_start;       /* For mmu_notifiers */
942         unsigned long mmun_end;         /* For mmu_notifiers */
943         gfp_t gfp;
944
945         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
946
947         /* Only allocate from the target node */
948         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
949
950         /*
951          * Before allocating the hugepage, release the mmap_sem read lock.
952          * The allocation can take potentially a long time if it involves
953          * sync compaction, and we do not need to hold the mmap_sem during
954          * that. We will recheck the vma after taking it again in write mode.
955          */
956         up_read(&mm->mmap_sem);
957         new_page = khugepaged_alloc_page(hpage, gfp, node);
958         if (!new_page) {
959                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
960                 goto out_nolock;
961         }
962
963         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
964                 result = SCAN_CGROUP_CHARGE_FAIL;
965                 goto out_nolock;
966         }
967
968         down_read(&mm->mmap_sem);
969         result = hugepage_vma_revalidate(mm, address, &vma);
970         if (result) {
971                 mem_cgroup_cancel_charge(new_page, memcg, true);
972                 up_read(&mm->mmap_sem);
973                 goto out_nolock;
974         }
975
976         pmd = mm_find_pmd(mm, address);
977         if (!pmd) {
978                 result = SCAN_PMD_NULL;
979                 mem_cgroup_cancel_charge(new_page, memcg, true);
980                 up_read(&mm->mmap_sem);
981                 goto out_nolock;
982         }
983
984         /*
985          * __collapse_huge_page_swapin always returns with mmap_sem locked.
986          * If it fails, we release mmap_sem and jump out_nolock.
987          * Continuing to collapse causes inconsistency.
988          */
989         if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
990                 mem_cgroup_cancel_charge(new_page, memcg, true);
991                 up_read(&mm->mmap_sem);
992                 goto out_nolock;
993         }
994
995         up_read(&mm->mmap_sem);
996         /*
997          * Prevent all access to pagetables with the exception of
998          * gup_fast later handled by the ptep_clear_flush and the VM
999          * handled by the anon_vma lock + PG_lock.
1000          */
1001         down_write(&mm->mmap_sem);
1002         result = hugepage_vma_revalidate(mm, address, &vma);
1003         if (result)
1004                 goto out;
1005         /* check if the pmd is still valid */
1006         if (mm_find_pmd(mm, address) != pmd)
1007                 goto out;
1008
1009         anon_vma_lock_write(vma->anon_vma);
1010
1011         pte = pte_offset_map(pmd, address);
1012         pte_ptl = pte_lockptr(mm, pmd);
1013
1014         mmun_start = address;
1015         mmun_end   = address + HPAGE_PMD_SIZE;
1016         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1017         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1018         /*
1019          * After this gup_fast can't run anymore. This also removes
1020          * any huge TLB entry from the CPU so we won't allow
1021          * huge and small TLB entries for the same virtual address
1022          * to avoid the risk of CPU bugs in that area.
1023          */
1024         _pmd = pmdp_collapse_flush(vma, address, pmd);
1025         spin_unlock(pmd_ptl);
1026         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1027
1028         spin_lock(pte_ptl);
1029         isolated = __collapse_huge_page_isolate(vma, address, pte);
1030         spin_unlock(pte_ptl);
1031
1032         if (unlikely(!isolated)) {
1033                 pte_unmap(pte);
1034                 spin_lock(pmd_ptl);
1035                 BUG_ON(!pmd_none(*pmd));
1036                 /*
1037                  * We can only use set_pmd_at when establishing
1038                  * hugepmds and never for establishing regular pmds that
1039                  * points to regular pagetables. Use pmd_populate for that
1040                  */
1041                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1042                 spin_unlock(pmd_ptl);
1043                 anon_vma_unlock_write(vma->anon_vma);
1044                 result = SCAN_FAIL;
1045                 goto out;
1046         }
1047
1048         /*
1049          * All pages are isolated and locked so anon_vma rmap
1050          * can't run anymore.
1051          */
1052         anon_vma_unlock_write(vma->anon_vma);
1053
1054         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1055         pte_unmap(pte);
1056         __SetPageUptodate(new_page);
1057         pgtable = pmd_pgtable(_pmd);
1058
1059         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1060         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1061
1062         /*
1063          * spin_lock() below is not the equivalent of smp_wmb(), so
1064          * this is needed to avoid the copy_huge_page writes to become
1065          * visible after the set_pmd_at() write.
1066          */
1067         smp_wmb();
1068
1069         spin_lock(pmd_ptl);
1070         BUG_ON(!pmd_none(*pmd));
1071         page_add_new_anon_rmap(new_page, vma, address, true);
1072         mem_cgroup_commit_charge(new_page, memcg, false, true);
1073         lru_cache_add_active_or_unevictable(new_page, vma);
1074         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1075         set_pmd_at(mm, address, pmd, _pmd);
1076         update_mmu_cache_pmd(vma, address, pmd);
1077         spin_unlock(pmd_ptl);
1078
1079         *hpage = NULL;
1080
1081         khugepaged_pages_collapsed++;
1082         result = SCAN_SUCCEED;
1083 out_up_write:
1084         up_write(&mm->mmap_sem);
1085 out_nolock:
1086         trace_mm_collapse_huge_page(mm, isolated, result);
1087         return;
1088 out:
1089         mem_cgroup_cancel_charge(new_page, memcg, true);
1090         goto out_up_write;
1091 }
1092
1093 static int khugepaged_scan_pmd(struct mm_struct *mm,
1094                                struct vm_area_struct *vma,
1095                                unsigned long address,
1096                                struct page **hpage)
1097 {
1098         pmd_t *pmd;
1099         pte_t *pte, *_pte;
1100         int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1101         struct page *page = NULL;
1102         unsigned long _address;
1103         spinlock_t *ptl;
1104         int node = NUMA_NO_NODE, unmapped = 0;
1105         bool writable = false;
1106
1107         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1108
1109         pmd = mm_find_pmd(mm, address);
1110         if (!pmd) {
1111                 result = SCAN_PMD_NULL;
1112                 goto out;
1113         }
1114
1115         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1116         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1117         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1118              _pte++, _address += PAGE_SIZE) {
1119                 pte_t pteval = *_pte;
1120                 if (is_swap_pte(pteval)) {
1121                         if (++unmapped <= khugepaged_max_ptes_swap) {
1122                                 continue;
1123                         } else {
1124                                 result = SCAN_EXCEED_SWAP_PTE;
1125                                 goto out_unmap;
1126                         }
1127                 }
1128                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1129                         if (!userfaultfd_armed(vma) &&
1130                             ++none_or_zero <= khugepaged_max_ptes_none) {
1131                                 continue;
1132                         } else {
1133                                 result = SCAN_EXCEED_NONE_PTE;
1134                                 goto out_unmap;
1135                         }
1136                 }
1137                 if (!pte_present(pteval)) {
1138                         result = SCAN_PTE_NON_PRESENT;
1139                         goto out_unmap;
1140                 }
1141                 if (pte_write(pteval))
1142                         writable = true;
1143
1144                 page = vm_normal_page(vma, _address, pteval);
1145                 if (unlikely(!page)) {
1146                         result = SCAN_PAGE_NULL;
1147                         goto out_unmap;
1148                 }
1149
1150                 /* TODO: teach khugepaged to collapse THP mapped with pte */
1151                 if (PageCompound(page)) {
1152                         result = SCAN_PAGE_COMPOUND;
1153                         goto out_unmap;
1154                 }
1155
1156                 /*
1157                  * Record which node the original page is from and save this
1158                  * information to khugepaged_node_load[].
1159                  * Khupaged will allocate hugepage from the node has the max
1160                  * hit record.
1161                  */
1162                 node = page_to_nid(page);
1163                 if (khugepaged_scan_abort(node)) {
1164                         result = SCAN_SCAN_ABORT;
1165                         goto out_unmap;
1166                 }
1167                 khugepaged_node_load[node]++;
1168                 if (!PageLRU(page)) {
1169                         result = SCAN_PAGE_LRU;
1170                         goto out_unmap;
1171                 }
1172                 if (PageLocked(page)) {
1173                         result = SCAN_PAGE_LOCK;
1174                         goto out_unmap;
1175                 }
1176                 if (!PageAnon(page)) {
1177                         result = SCAN_PAGE_ANON;
1178                         goto out_unmap;
1179                 }
1180
1181                 /*
1182                  * cannot use mapcount: can't collapse if there's a gup pin.
1183                  * The page must only be referenced by the scanned process
1184                  * and page swap cache.
1185                  */
1186                 if (page_count(page) != 1 + PageSwapCache(page)) {
1187                         result = SCAN_PAGE_COUNT;
1188                         goto out_unmap;
1189                 }
1190                 if (pte_young(pteval) ||
1191                     page_is_young(page) || PageReferenced(page) ||
1192                     mmu_notifier_test_young(vma->vm_mm, address))
1193                         referenced++;
1194         }
1195         if (writable) {
1196                 if (referenced) {
1197                         result = SCAN_SUCCEED;
1198                         ret = 1;
1199                 } else {
1200                         result = SCAN_LACK_REFERENCED_PAGE;
1201                 }
1202         } else {
1203                 result = SCAN_PAGE_RO;
1204         }
1205 out_unmap:
1206         pte_unmap_unlock(pte, ptl);
1207         if (ret) {
1208                 node = khugepaged_find_target_node();
1209                 /* collapse_huge_page will return with the mmap_sem released */
1210                 collapse_huge_page(mm, address, hpage, node, referenced);
1211         }
1212 out:
1213         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1214                                      none_or_zero, result, unmapped);
1215         return ret;
1216 }
1217
1218 static void collect_mm_slot(struct mm_slot *mm_slot)
1219 {
1220         struct mm_struct *mm = mm_slot->mm;
1221
1222         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1223
1224         if (khugepaged_test_exit(mm)) {
1225                 /* free mm_slot */
1226                 hash_del(&mm_slot->hash);
1227                 list_del(&mm_slot->mm_node);
1228
1229                 /*
1230                  * Not strictly needed because the mm exited already.
1231                  *
1232                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1233                  */
1234
1235                 /* khugepaged_mm_lock actually not necessary for the below */
1236                 free_mm_slot(mm_slot);
1237                 mmdrop(mm);
1238         }
1239 }
1240
1241 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1242 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1243 {
1244         struct vm_area_struct *vma;
1245         unsigned long addr;
1246         pmd_t *pmd, _pmd;
1247
1248         i_mmap_lock_write(mapping);
1249         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1250                 /* probably overkill */
1251                 if (vma->anon_vma)
1252                         continue;
1253                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1254                 if (addr & ~HPAGE_PMD_MASK)
1255                         continue;
1256                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1257                         continue;
1258                 pmd = mm_find_pmd(vma->vm_mm, addr);
1259                 if (!pmd)
1260                         continue;
1261                 /*
1262                  * We need exclusive mmap_sem to retract page table.
1263                  * If trylock fails we would end up with pte-mapped THP after
1264                  * re-fault. Not ideal, but it's more important to not disturb
1265                  * the system too much.
1266                  */
1267                 if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1268                         spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1269                         /* assume page table is clear */
1270                         _pmd = pmdp_collapse_flush(vma, addr, pmd);
1271                         spin_unlock(ptl);
1272                         up_write(&vma->vm_mm->mmap_sem);
1273                         mm_dec_nr_ptes(vma->vm_mm);
1274                         pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1275                 }
1276         }
1277         i_mmap_unlock_write(mapping);
1278 }
1279
1280 /**
1281  * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1282  *
1283  * Basic scheme is simple, details are more complex:
1284  *  - allocate and freeze a new huge page;
1285  *  - scan over radix tree replacing old pages the new one
1286  *    + swap in pages if necessary;
1287  *    + fill in gaps;
1288  *    + keep old pages around in case if rollback is required;
1289  *  - if replacing succeed:
1290  *    + copy data over;
1291  *    + free old pages;
1292  *    + unfreeze huge page;
1293  *  - if replacing failed;
1294  *    + put all pages back and unfreeze them;
1295  *    + restore gaps in the radix-tree;
1296  *    + free huge page;
1297  */
1298 static void collapse_shmem(struct mm_struct *mm,
1299                 struct address_space *mapping, pgoff_t start,
1300                 struct page **hpage, int node)
1301 {
1302         gfp_t gfp;
1303         struct page *page, *new_page, *tmp;
1304         struct mem_cgroup *memcg;
1305         pgoff_t index, end = start + HPAGE_PMD_NR;
1306         LIST_HEAD(pagelist);
1307         struct radix_tree_iter iter;
1308         void **slot;
1309         int nr_none = 0, result = SCAN_SUCCEED;
1310
1311         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1312
1313         /* Only allocate from the target node */
1314         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1315
1316         new_page = khugepaged_alloc_page(hpage, gfp, node);
1317         if (!new_page) {
1318                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1319                 goto out;
1320         }
1321
1322         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1323                 result = SCAN_CGROUP_CHARGE_FAIL;
1324                 goto out;
1325         }
1326
1327         new_page->index = start;
1328         new_page->mapping = mapping;
1329         __SetPageSwapBacked(new_page);
1330         __SetPageLocked(new_page);
1331         BUG_ON(!page_ref_freeze(new_page, 1));
1332
1333
1334         /*
1335          * At this point the new_page is 'frozen' (page_count() is zero), locked
1336          * and not up-to-date. It's safe to insert it into radix tree, because
1337          * nobody would be able to map it or use it in other way until we
1338          * unfreeze it.
1339          */
1340
1341         index = start;
1342         spin_lock_irq(&mapping->tree_lock);
1343         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1344                 int n = min(iter.index, end) - index;
1345
1346                 /*
1347                  * Handle holes in the radix tree: charge it from shmem and
1348                  * insert relevant subpage of new_page into the radix-tree.
1349                  */
1350                 if (n && !shmem_charge(mapping->host, n)) {
1351                         result = SCAN_FAIL;
1352                         break;
1353                 }
1354                 nr_none += n;
1355                 for (; index < min(iter.index, end); index++) {
1356                         radix_tree_insert(&mapping->page_tree, index,
1357                                         new_page + (index % HPAGE_PMD_NR));
1358                 }
1359
1360                 /* We are done. */
1361                 if (index >= end)
1362                         break;
1363
1364                 page = radix_tree_deref_slot_protected(slot,
1365                                 &mapping->tree_lock);
1366                 if (radix_tree_exceptional_entry(page) || !PageUptodate(page)) {
1367                         spin_unlock_irq(&mapping->tree_lock);
1368                         /* swap in or instantiate fallocated page */
1369                         if (shmem_getpage(mapping->host, index, &page,
1370                                                 SGP_NOHUGE)) {
1371                                 result = SCAN_FAIL;
1372                                 goto tree_unlocked;
1373                         }
1374                         spin_lock_irq(&mapping->tree_lock);
1375                 } else if (trylock_page(page)) {
1376                         get_page(page);
1377                 } else {
1378                         result = SCAN_PAGE_LOCK;
1379                         break;
1380                 }
1381
1382                 /*
1383                  * The page must be locked, so we can drop the tree_lock
1384                  * without racing with truncate.
1385                  */
1386                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1387                 VM_BUG_ON_PAGE(!PageUptodate(page), page);
1388                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1389
1390                 if (page_mapping(page) != mapping) {
1391                         result = SCAN_TRUNCATED;
1392                         goto out_unlock;
1393                 }
1394                 spin_unlock_irq(&mapping->tree_lock);
1395
1396                 if (isolate_lru_page(page)) {
1397                         result = SCAN_DEL_PAGE_LRU;
1398                         goto out_isolate_failed;
1399                 }
1400
1401                 if (page_mapped(page))
1402                         unmap_mapping_range(mapping, index << PAGE_SHIFT,
1403                                         PAGE_SIZE, 0);
1404
1405                 spin_lock_irq(&mapping->tree_lock);
1406
1407                 slot = radix_tree_lookup_slot(&mapping->page_tree, index);
1408                 VM_BUG_ON_PAGE(page != radix_tree_deref_slot_protected(slot,
1409                                         &mapping->tree_lock), page);
1410                 VM_BUG_ON_PAGE(page_mapped(page), page);
1411
1412                 /*
1413                  * The page is expected to have page_count() == 3:
1414                  *  - we hold a pin on it;
1415                  *  - one reference from radix tree;
1416                  *  - one from isolate_lru_page;
1417                  */
1418                 if (!page_ref_freeze(page, 3)) {
1419                         result = SCAN_PAGE_COUNT;
1420                         goto out_lru;
1421                 }
1422
1423                 /*
1424                  * Add the page to the list to be able to undo the collapse if
1425                  * something go wrong.
1426                  */
1427                 list_add_tail(&page->lru, &pagelist);
1428
1429                 /* Finally, replace with the new page. */
1430                 radix_tree_replace_slot(&mapping->page_tree, slot,
1431                                 new_page + (index % HPAGE_PMD_NR));
1432
1433                 slot = radix_tree_iter_resume(slot, &iter);
1434                 index++;
1435                 continue;
1436 out_lru:
1437                 spin_unlock_irq(&mapping->tree_lock);
1438                 putback_lru_page(page);
1439 out_isolate_failed:
1440                 unlock_page(page);
1441                 put_page(page);
1442                 goto tree_unlocked;
1443 out_unlock:
1444                 unlock_page(page);
1445                 put_page(page);
1446                 break;
1447         }
1448
1449         /*
1450          * Handle hole in radix tree at the end of the range.
1451          * This code only triggers if there's nothing in radix tree
1452          * beyond 'end'.
1453          */
1454         if (result == SCAN_SUCCEED && index < end) {
1455                 int n = end - index;
1456
1457                 if (!shmem_charge(mapping->host, n)) {
1458                         result = SCAN_FAIL;
1459                         goto tree_locked;
1460                 }
1461
1462                 for (; index < end; index++) {
1463                         radix_tree_insert(&mapping->page_tree, index,
1464                                         new_page + (index % HPAGE_PMD_NR));
1465                 }
1466                 nr_none += n;
1467         }
1468
1469 tree_locked:
1470         spin_unlock_irq(&mapping->tree_lock);
1471 tree_unlocked:
1472
1473         if (result == SCAN_SUCCEED) {
1474                 unsigned long flags;
1475                 struct zone *zone = page_zone(new_page);
1476
1477                 /*
1478                  * Replacing old pages with new one has succeed, now we need to
1479                  * copy the content and free old pages.
1480                  */
1481                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1482                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1483                                         page);
1484                         list_del(&page->lru);
1485                         unlock_page(page);
1486                         page_ref_unfreeze(page, 1);
1487                         page->mapping = NULL;
1488                         ClearPageActive(page);
1489                         ClearPageUnevictable(page);
1490                         put_page(page);
1491                 }
1492
1493                 local_irq_save(flags);
1494                 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1495                 if (nr_none) {
1496                         __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1497                         __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
1498                 }
1499                 local_irq_restore(flags);
1500
1501                 /*
1502                  * Remove pte page tables, so we can re-faulti
1503                  * the page as huge.
1504                  */
1505                 retract_page_tables(mapping, start);
1506
1507                 /* Everything is ready, let's unfreeze the new_page */
1508                 set_page_dirty(new_page);
1509                 SetPageUptodate(new_page);
1510                 page_ref_unfreeze(new_page, HPAGE_PMD_NR);
1511                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1512                 lru_cache_add_anon(new_page);
1513                 unlock_page(new_page);
1514
1515                 *hpage = NULL;
1516         } else {
1517                 /* Something went wrong: rollback changes to the radix-tree */
1518                 shmem_uncharge(mapping->host, nr_none);
1519                 spin_lock_irq(&mapping->tree_lock);
1520                 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
1521                                 start) {
1522                         if (iter.index >= end)
1523                                 break;
1524                         page = list_first_entry_or_null(&pagelist,
1525                                         struct page, lru);
1526                         if (!page || iter.index < page->index) {
1527                                 if (!nr_none)
1528                                         break;
1529                                 nr_none--;
1530                                 /* Put holes back where they were */
1531                                 radix_tree_delete(&mapping->page_tree,
1532                                                   iter.index);
1533                                 continue;
1534                         }
1535
1536                         VM_BUG_ON_PAGE(page->index != iter.index, page);
1537
1538                         /* Unfreeze the page. */
1539                         list_del(&page->lru);
1540                         page_ref_unfreeze(page, 2);
1541                         radix_tree_replace_slot(&mapping->page_tree,
1542                                                 slot, page);
1543                         slot = radix_tree_iter_resume(slot, &iter);
1544                         spin_unlock_irq(&mapping->tree_lock);
1545                         putback_lru_page(page);
1546                         unlock_page(page);
1547                         spin_lock_irq(&mapping->tree_lock);
1548                 }
1549                 VM_BUG_ON(nr_none);
1550                 spin_unlock_irq(&mapping->tree_lock);
1551
1552                 /* Unfreeze new_page, caller would take care about freeing it */
1553                 page_ref_unfreeze(new_page, 1);
1554                 mem_cgroup_cancel_charge(new_page, memcg, true);
1555                 unlock_page(new_page);
1556                 new_page->mapping = NULL;
1557         }
1558 out:
1559         VM_BUG_ON(!list_empty(&pagelist));
1560         /* TODO: tracepoints */
1561 }
1562
1563 static void khugepaged_scan_shmem(struct mm_struct *mm,
1564                 struct address_space *mapping,
1565                 pgoff_t start, struct page **hpage)
1566 {
1567         struct page *page = NULL;
1568         struct radix_tree_iter iter;
1569         void **slot;
1570         int present, swap;
1571         int node = NUMA_NO_NODE;
1572         int result = SCAN_SUCCEED;
1573
1574         present = 0;
1575         swap = 0;
1576         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1577         rcu_read_lock();
1578         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1579                 if (iter.index >= start + HPAGE_PMD_NR)
1580                         break;
1581
1582                 page = radix_tree_deref_slot(slot);
1583                 if (radix_tree_deref_retry(page)) {
1584                         slot = radix_tree_iter_retry(&iter);
1585                         continue;
1586                 }
1587
1588                 if (radix_tree_exception(page)) {
1589                         if (++swap > khugepaged_max_ptes_swap) {
1590                                 result = SCAN_EXCEED_SWAP_PTE;
1591                                 break;
1592                         }
1593                         continue;
1594                 }
1595
1596                 if (PageTransCompound(page)) {
1597                         result = SCAN_PAGE_COMPOUND;
1598                         break;
1599                 }
1600
1601                 node = page_to_nid(page);
1602                 if (khugepaged_scan_abort(node)) {
1603                         result = SCAN_SCAN_ABORT;
1604                         break;
1605                 }
1606                 khugepaged_node_load[node]++;
1607
1608                 if (!PageLRU(page)) {
1609                         result = SCAN_PAGE_LRU;
1610                         break;
1611                 }
1612
1613                 if (page_count(page) != 1 + page_mapcount(page)) {
1614                         result = SCAN_PAGE_COUNT;
1615                         break;
1616                 }
1617
1618                 /*
1619                  * We probably should check if the page is referenced here, but
1620                  * nobody would transfer pte_young() to PageReferenced() for us.
1621                  * And rmap walk here is just too costly...
1622                  */
1623
1624                 present++;
1625
1626                 if (need_resched()) {
1627                         slot = radix_tree_iter_resume(slot, &iter);
1628                         cond_resched_rcu();
1629                 }
1630         }
1631         rcu_read_unlock();
1632
1633         if (result == SCAN_SUCCEED) {
1634                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1635                         result = SCAN_EXCEED_NONE_PTE;
1636                 } else {
1637                         node = khugepaged_find_target_node();
1638                         collapse_shmem(mm, mapping, start, hpage, node);
1639                 }
1640         }
1641
1642         /* TODO: tracepoints */
1643 }
1644 #else
1645 static void khugepaged_scan_shmem(struct mm_struct *mm,
1646                 struct address_space *mapping,
1647                 pgoff_t start, struct page **hpage)
1648 {
1649         BUILD_BUG();
1650 }
1651 #endif
1652
1653 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1654                                             struct page **hpage)
1655         __releases(&khugepaged_mm_lock)
1656         __acquires(&khugepaged_mm_lock)
1657 {
1658         struct mm_slot *mm_slot;
1659         struct mm_struct *mm;
1660         struct vm_area_struct *vma;
1661         int progress = 0;
1662
1663         VM_BUG_ON(!pages);
1664         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1665
1666         if (khugepaged_scan.mm_slot)
1667                 mm_slot = khugepaged_scan.mm_slot;
1668         else {
1669                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1670                                      struct mm_slot, mm_node);
1671                 khugepaged_scan.address = 0;
1672                 khugepaged_scan.mm_slot = mm_slot;
1673         }
1674         spin_unlock(&khugepaged_mm_lock);
1675
1676         mm = mm_slot->mm;
1677         down_read(&mm->mmap_sem);
1678         if (unlikely(khugepaged_test_exit(mm)))
1679                 vma = NULL;
1680         else
1681                 vma = find_vma(mm, khugepaged_scan.address);
1682
1683         progress++;
1684         for (; vma; vma = vma->vm_next) {
1685                 unsigned long hstart, hend;
1686
1687                 cond_resched();
1688                 if (unlikely(khugepaged_test_exit(mm))) {
1689                         progress++;
1690                         break;
1691                 }
1692                 if (!hugepage_vma_check(vma)) {
1693 skip:
1694                         progress++;
1695                         continue;
1696                 }
1697                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1698                 hend = vma->vm_end & HPAGE_PMD_MASK;
1699                 if (hstart >= hend)
1700                         goto skip;
1701                 if (khugepaged_scan.address > hend)
1702                         goto skip;
1703                 if (khugepaged_scan.address < hstart)
1704                         khugepaged_scan.address = hstart;
1705                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1706
1707                 while (khugepaged_scan.address < hend) {
1708                         int ret;
1709                         cond_resched();
1710                         if (unlikely(khugepaged_test_exit(mm)))
1711                                 goto breakouterloop;
1712
1713                         VM_BUG_ON(khugepaged_scan.address < hstart ||
1714                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
1715                                   hend);
1716                         if (shmem_file(vma->vm_file)) {
1717                                 struct file *file;
1718                                 pgoff_t pgoff = linear_page_index(vma,
1719                                                 khugepaged_scan.address);
1720                                 if (!shmem_huge_enabled(vma))
1721                                         goto skip;
1722                                 file = get_file(vma->vm_file);
1723                                 up_read(&mm->mmap_sem);
1724                                 ret = 1;
1725                                 khugepaged_scan_shmem(mm, file->f_mapping,
1726                                                 pgoff, hpage);
1727                                 fput(file);
1728                         } else {
1729                                 ret = khugepaged_scan_pmd(mm, vma,
1730                                                 khugepaged_scan.address,
1731                                                 hpage);
1732                         }
1733                         /* move to next address */
1734                         khugepaged_scan.address += HPAGE_PMD_SIZE;
1735                         progress += HPAGE_PMD_NR;
1736                         if (ret)
1737                                 /* we released mmap_sem so break loop */
1738                                 goto breakouterloop_mmap_sem;
1739                         if (progress >= pages)
1740                                 goto breakouterloop;
1741                 }
1742         }
1743 breakouterloop:
1744         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1745 breakouterloop_mmap_sem:
1746
1747         spin_lock(&khugepaged_mm_lock);
1748         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1749         /*
1750          * Release the current mm_slot if this mm is about to die, or
1751          * if we scanned all vmas of this mm.
1752          */
1753         if (khugepaged_test_exit(mm) || !vma) {
1754                 /*
1755                  * Make sure that if mm_users is reaching zero while
1756                  * khugepaged runs here, khugepaged_exit will find
1757                  * mm_slot not pointing to the exiting mm.
1758                  */
1759                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1760                         khugepaged_scan.mm_slot = list_entry(
1761                                 mm_slot->mm_node.next,
1762                                 struct mm_slot, mm_node);
1763                         khugepaged_scan.address = 0;
1764                 } else {
1765                         khugepaged_scan.mm_slot = NULL;
1766                         khugepaged_full_scans++;
1767                 }
1768
1769                 collect_mm_slot(mm_slot);
1770         }
1771
1772         return progress;
1773 }
1774
1775 static int khugepaged_has_work(void)
1776 {
1777         return !list_empty(&khugepaged_scan.mm_head) &&
1778                 khugepaged_enabled();
1779 }
1780
1781 static int khugepaged_wait_event(void)
1782 {
1783         return !list_empty(&khugepaged_scan.mm_head) ||
1784                 kthread_should_stop();
1785 }
1786
1787 static void khugepaged_do_scan(void)
1788 {
1789         struct page *hpage = NULL;
1790         unsigned int progress = 0, pass_through_head = 0;
1791         unsigned int pages = khugepaged_pages_to_scan;
1792         bool wait = true;
1793
1794         barrier(); /* write khugepaged_pages_to_scan to local stack */
1795
1796         while (progress < pages) {
1797                 if (!khugepaged_prealloc_page(&hpage, &wait))
1798                         break;
1799
1800                 cond_resched();
1801
1802                 if (unlikely(kthread_should_stop() || try_to_freeze()))
1803                         break;
1804
1805                 spin_lock(&khugepaged_mm_lock);
1806                 if (!khugepaged_scan.mm_slot)
1807                         pass_through_head++;
1808                 if (khugepaged_has_work() &&
1809                     pass_through_head < 2)
1810                         progress += khugepaged_scan_mm_slot(pages - progress,
1811                                                             &hpage);
1812                 else
1813                         progress = pages;
1814                 spin_unlock(&khugepaged_mm_lock);
1815         }
1816
1817         if (!IS_ERR_OR_NULL(hpage))
1818                 put_page(hpage);
1819 }
1820
1821 static bool khugepaged_should_wakeup(void)
1822 {
1823         return kthread_should_stop() ||
1824                time_after_eq(jiffies, khugepaged_sleep_expire);
1825 }
1826
1827 static void khugepaged_wait_work(void)
1828 {
1829         if (khugepaged_has_work()) {
1830                 const unsigned long scan_sleep_jiffies =
1831                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1832
1833                 if (!scan_sleep_jiffies)
1834                         return;
1835
1836                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1837                 wait_event_freezable_timeout(khugepaged_wait,
1838                                              khugepaged_should_wakeup(),
1839                                              scan_sleep_jiffies);
1840                 return;
1841         }
1842
1843         if (khugepaged_enabled())
1844                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1845 }
1846
1847 static int khugepaged(void *none)
1848 {
1849         struct mm_slot *mm_slot;
1850
1851         set_freezable();
1852         set_user_nice(current, MAX_NICE);
1853
1854         while (!kthread_should_stop()) {
1855                 khugepaged_do_scan();
1856                 khugepaged_wait_work();
1857         }
1858
1859         spin_lock(&khugepaged_mm_lock);
1860         mm_slot = khugepaged_scan.mm_slot;
1861         khugepaged_scan.mm_slot = NULL;
1862         if (mm_slot)
1863                 collect_mm_slot(mm_slot);
1864         spin_unlock(&khugepaged_mm_lock);
1865         return 0;
1866 }
1867
1868 static void set_recommended_min_free_kbytes(void)
1869 {
1870         struct zone *zone;
1871         int nr_zones = 0;
1872         unsigned long recommended_min;
1873
1874         for_each_populated_zone(zone)
1875                 nr_zones++;
1876
1877         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1878         recommended_min = pageblock_nr_pages * nr_zones * 2;
1879
1880         /*
1881          * Make sure that on average at least two pageblocks are almost free
1882          * of another type, one for a migratetype to fall back to and a
1883          * second to avoid subsequent fallbacks of other types There are 3
1884          * MIGRATE_TYPES we care about.
1885          */
1886         recommended_min += pageblock_nr_pages * nr_zones *
1887                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1888
1889         /* don't ever allow to reserve more than 5% of the lowmem */
1890         recommended_min = min(recommended_min,
1891                               (unsigned long) nr_free_buffer_pages() / 20);
1892         recommended_min <<= (PAGE_SHIFT-10);
1893
1894         if (recommended_min > min_free_kbytes) {
1895                 if (user_min_free_kbytes >= 0)
1896                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1897                                 min_free_kbytes, recommended_min);
1898
1899                 min_free_kbytes = recommended_min;
1900         }
1901         setup_per_zone_wmarks();
1902 }
1903
1904 int start_stop_khugepaged(void)
1905 {
1906         static struct task_struct *khugepaged_thread __read_mostly;
1907         static DEFINE_MUTEX(khugepaged_mutex);
1908         int err = 0;
1909
1910         mutex_lock(&khugepaged_mutex);
1911         if (khugepaged_enabled()) {
1912                 if (!khugepaged_thread)
1913                         khugepaged_thread = kthread_run(khugepaged, NULL,
1914                                                         "khugepaged");
1915                 if (IS_ERR(khugepaged_thread)) {
1916                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1917                         err = PTR_ERR(khugepaged_thread);
1918                         khugepaged_thread = NULL;
1919                         goto fail;
1920                 }
1921
1922                 if (!list_empty(&khugepaged_scan.mm_head))
1923                         wake_up_interruptible(&khugepaged_wait);
1924
1925                 set_recommended_min_free_kbytes();
1926         } else if (khugepaged_thread) {
1927                 kthread_stop(khugepaged_thread);
1928                 khugepaged_thread = NULL;
1929         }
1930 fail:
1931         mutex_unlock(&khugepaged_mutex);
1932         return err;
1933 }