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_EXCEED_SWAP_PTE,
32 SCAN_EXCEED_SHARED_PTE,
33 SCAN_PTE_NON_PRESENT,
34 SCAN_PTE_UFFD_WP,
35 SCAN_PAGE_RO,
36 SCAN_LACK_REFERENCED_PAGE,
37 SCAN_PAGE_NULL,
38 SCAN_SCAN_ABORT,
39 SCAN_PAGE_COUNT,
40 SCAN_PAGE_LRU,
41 SCAN_PAGE_LOCK,
42 SCAN_PAGE_ANON,
43 SCAN_PAGE_COMPOUND,
44 SCAN_ANY_PROCESS,
45 SCAN_VMA_NULL,
46 SCAN_VMA_CHECK,
47 SCAN_ADDRESS_RANGE,
48 SCAN_SWAP_CACHE_PAGE,
49 SCAN_DEL_PAGE_LRU,
50 SCAN_ALLOC_HUGE_PAGE_FAIL,
51 SCAN_CGROUP_CHARGE_FAIL,
52 SCAN_TRUNCATED,
53 SCAN_PAGE_HAS_PRIVATE,
54 };
55
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/huge_memory.h>
58
59 static struct task_struct *khugepaged_thread __read_mostly;
60 static DEFINE_MUTEX(khugepaged_mutex);
61
62 /* default scan 8*512 pte (or vmas) every 30 second */
63 static unsigned int khugepaged_pages_to_scan __read_mostly;
64 static unsigned int khugepaged_pages_collapsed;
65 static unsigned int khugepaged_full_scans;
66 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
67 /* during fragmentation poll the hugepage allocator once every minute */
68 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
69 static unsigned long khugepaged_sleep_expire;
70 static DEFINE_SPINLOCK(khugepaged_mm_lock);
71 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
72 /*
73 * default collapse hugepages if there is at least one pte mapped like
74 * it would have happened if the vma was large enough during page
75 * fault.
76 */
77 static unsigned int khugepaged_max_ptes_none __read_mostly;
78 static unsigned int khugepaged_max_ptes_swap __read_mostly;
79 static unsigned int khugepaged_max_ptes_shared __read_mostly;
80
81 #define MM_SLOTS_HASH_BITS 10
82 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
83
84 static struct kmem_cache *mm_slot_cache __read_mostly;
85
86 #define MAX_PTE_MAPPED_THP 8
87
88 /**
89 * struct mm_slot - hash lookup from mm to mm_slot
90 * @hash: hash collision list
91 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
92 * @mm: the mm that this information is valid for
93 */
94 struct mm_slot {
95 struct hlist_node hash;
96 struct list_head mm_node;
97 struct mm_struct *mm;
98
99 /* pte-mapped THP in this mm */
100 int nr_pte_mapped_thp;
101 unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
102 };
103
104 /**
105 * struct khugepaged_scan - cursor for scanning
106 * @mm_head: the head of the mm list to scan
107 * @mm_slot: the current mm_slot we are scanning
108 * @address: the next address inside that to be scanned
109 *
110 * There is only the one khugepaged_scan instance of this cursor structure.
111 */
112 struct khugepaged_scan {
113 struct list_head mm_head;
114 struct mm_slot *mm_slot;
115 unsigned long address;
116 };
117
118 static struct khugepaged_scan khugepaged_scan = {
119 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
120 };
121
122 #ifdef CONFIG_SYSFS
scan_sleep_millisecs_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)123 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
124 struct kobj_attribute *attr,
125 char *buf)
126 {
127 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
128 }
129
scan_sleep_millisecs_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)130 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
131 struct kobj_attribute *attr,
132 const char *buf, size_t count)
133 {
134 unsigned long msecs;
135 int err;
136
137 err = kstrtoul(buf, 10, &msecs);
138 if (err || msecs > UINT_MAX)
139 return -EINVAL;
140
141 khugepaged_scan_sleep_millisecs = msecs;
142 khugepaged_sleep_expire = 0;
143 wake_up_interruptible(&khugepaged_wait);
144
145 return count;
146 }
147 static struct kobj_attribute scan_sleep_millisecs_attr =
148 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
149 scan_sleep_millisecs_store);
150
alloc_sleep_millisecs_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)151 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
152 struct kobj_attribute *attr,
153 char *buf)
154 {
155 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
156 }
157
alloc_sleep_millisecs_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)158 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
159 struct kobj_attribute *attr,
160 const char *buf, size_t count)
161 {
162 unsigned long msecs;
163 int err;
164
165 err = kstrtoul(buf, 10, &msecs);
166 if (err || msecs > UINT_MAX)
167 return -EINVAL;
168
169 khugepaged_alloc_sleep_millisecs = msecs;
170 khugepaged_sleep_expire = 0;
171 wake_up_interruptible(&khugepaged_wait);
172
173 return count;
174 }
175 static struct kobj_attribute alloc_sleep_millisecs_attr =
176 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
177 alloc_sleep_millisecs_store);
178
pages_to_scan_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)179 static ssize_t pages_to_scan_show(struct kobject *kobj,
180 struct kobj_attribute *attr,
181 char *buf)
182 {
183 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
184 }
pages_to_scan_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)185 static ssize_t pages_to_scan_store(struct kobject *kobj,
186 struct kobj_attribute *attr,
187 const char *buf, size_t count)
188 {
189 int err;
190 unsigned long pages;
191
192 err = kstrtoul(buf, 10, &pages);
193 if (err || !pages || pages > UINT_MAX)
194 return -EINVAL;
195
196 khugepaged_pages_to_scan = pages;
197
198 return count;
199 }
200 static struct kobj_attribute pages_to_scan_attr =
201 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
202 pages_to_scan_store);
203
pages_collapsed_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)204 static ssize_t pages_collapsed_show(struct kobject *kobj,
205 struct kobj_attribute *attr,
206 char *buf)
207 {
208 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
209 }
210 static struct kobj_attribute pages_collapsed_attr =
211 __ATTR_RO(pages_collapsed);
212
full_scans_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)213 static ssize_t full_scans_show(struct kobject *kobj,
214 struct kobj_attribute *attr,
215 char *buf)
216 {
217 return sprintf(buf, "%u\n", khugepaged_full_scans);
218 }
219 static struct kobj_attribute full_scans_attr =
220 __ATTR_RO(full_scans);
221
khugepaged_defrag_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)222 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
223 struct kobj_attribute *attr, char *buf)
224 {
225 return single_hugepage_flag_show(kobj, attr, buf,
226 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
227 }
khugepaged_defrag_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)228 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
229 struct kobj_attribute *attr,
230 const char *buf, size_t count)
231 {
232 return single_hugepage_flag_store(kobj, attr, buf, count,
233 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
234 }
235 static struct kobj_attribute khugepaged_defrag_attr =
236 __ATTR(defrag, 0644, khugepaged_defrag_show,
237 khugepaged_defrag_store);
238
239 /*
240 * max_ptes_none controls if khugepaged should collapse hugepages over
241 * any unmapped ptes in turn potentially increasing the memory
242 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
243 * reduce the available free memory in the system as it
244 * runs. Increasing max_ptes_none will instead potentially reduce the
245 * free memory in the system during the khugepaged scan.
246 */
khugepaged_max_ptes_none_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)247 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
248 struct kobj_attribute *attr,
249 char *buf)
250 {
251 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
252 }
khugepaged_max_ptes_none_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)253 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
254 struct kobj_attribute *attr,
255 const char *buf, size_t count)
256 {
257 int err;
258 unsigned long max_ptes_none;
259
260 err = kstrtoul(buf, 10, &max_ptes_none);
261 if (err || max_ptes_none > HPAGE_PMD_NR-1)
262 return -EINVAL;
263
264 khugepaged_max_ptes_none = max_ptes_none;
265
266 return count;
267 }
268 static struct kobj_attribute khugepaged_max_ptes_none_attr =
269 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
270 khugepaged_max_ptes_none_store);
271
khugepaged_max_ptes_swap_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)272 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
273 struct kobj_attribute *attr,
274 char *buf)
275 {
276 return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
277 }
278
khugepaged_max_ptes_swap_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)279 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
280 struct kobj_attribute *attr,
281 const char *buf, size_t count)
282 {
283 int err;
284 unsigned long max_ptes_swap;
285
286 err = kstrtoul(buf, 10, &max_ptes_swap);
287 if (err || max_ptes_swap > HPAGE_PMD_NR-1)
288 return -EINVAL;
289
290 khugepaged_max_ptes_swap = max_ptes_swap;
291
292 return count;
293 }
294
295 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
296 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
297 khugepaged_max_ptes_swap_store);
298
khugepaged_max_ptes_shared_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)299 static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj,
300 struct kobj_attribute *attr,
301 char *buf)
302 {
303 return sprintf(buf, "%u\n", khugepaged_max_ptes_shared);
304 }
305
khugepaged_max_ptes_shared_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)306 static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj,
307 struct kobj_attribute *attr,
308 const char *buf, size_t count)
309 {
310 int err;
311 unsigned long max_ptes_shared;
312
313 err = kstrtoul(buf, 10, &max_ptes_shared);
314 if (err || max_ptes_shared > HPAGE_PMD_NR-1)
315 return -EINVAL;
316
317 khugepaged_max_ptes_shared = max_ptes_shared;
318
319 return count;
320 }
321
322 static struct kobj_attribute khugepaged_max_ptes_shared_attr =
323 __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show,
324 khugepaged_max_ptes_shared_store);
325
326 static struct attribute *khugepaged_attr[] = {
327 &khugepaged_defrag_attr.attr,
328 &khugepaged_max_ptes_none_attr.attr,
329 &khugepaged_max_ptes_swap_attr.attr,
330 &khugepaged_max_ptes_shared_attr.attr,
331 &pages_to_scan_attr.attr,
332 &pages_collapsed_attr.attr,
333 &full_scans_attr.attr,
334 &scan_sleep_millisecs_attr.attr,
335 &alloc_sleep_millisecs_attr.attr,
336 NULL,
337 };
338
339 struct attribute_group khugepaged_attr_group = {
340 .attrs = khugepaged_attr,
341 .name = "khugepaged",
342 };
343 #endif /* CONFIG_SYSFS */
344
hugepage_madvise(struct vm_area_struct * vma,unsigned long * vm_flags,int advice)345 int hugepage_madvise(struct vm_area_struct *vma,
346 unsigned long *vm_flags, int advice)
347 {
348 switch (advice) {
349 case MADV_HUGEPAGE:
350 #ifdef CONFIG_S390
351 /*
352 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
353 * can't handle this properly after s390_enable_sie, so we simply
354 * ignore the madvise to prevent qemu from causing a SIGSEGV.
355 */
356 if (mm_has_pgste(vma->vm_mm))
357 return 0;
358 #endif
359 *vm_flags &= ~VM_NOHUGEPAGE;
360 *vm_flags |= VM_HUGEPAGE;
361 /*
362 * If the vma become good for khugepaged to scan,
363 * register it here without waiting a page fault that
364 * may not happen any time soon.
365 */
366 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
367 khugepaged_enter_vma_merge(vma, *vm_flags))
368 return -ENOMEM;
369 break;
370 case MADV_NOHUGEPAGE:
371 *vm_flags &= ~VM_HUGEPAGE;
372 *vm_flags |= VM_NOHUGEPAGE;
373 /*
374 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
375 * this vma even if we leave the mm registered in khugepaged if
376 * it got registered before VM_NOHUGEPAGE was set.
377 */
378 break;
379 }
380
381 return 0;
382 }
383
khugepaged_init(void)384 int __init khugepaged_init(void)
385 {
386 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
387 sizeof(struct mm_slot),
388 __alignof__(struct mm_slot), 0, NULL);
389 if (!mm_slot_cache)
390 return -ENOMEM;
391
392 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
393 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
394 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
395 khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
396
397 return 0;
398 }
399
khugepaged_destroy(void)400 void __init khugepaged_destroy(void)
401 {
402 kmem_cache_destroy(mm_slot_cache);
403 }
404
alloc_mm_slot(void)405 static inline struct mm_slot *alloc_mm_slot(void)
406 {
407 if (!mm_slot_cache) /* initialization failed */
408 return NULL;
409 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
410 }
411
free_mm_slot(struct mm_slot * mm_slot)412 static inline void free_mm_slot(struct mm_slot *mm_slot)
413 {
414 kmem_cache_free(mm_slot_cache, mm_slot);
415 }
416
get_mm_slot(struct mm_struct * mm)417 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
418 {
419 struct mm_slot *mm_slot;
420
421 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
422 if (mm == mm_slot->mm)
423 return mm_slot;
424
425 return NULL;
426 }
427
insert_to_mm_slots_hash(struct mm_struct * mm,struct mm_slot * mm_slot)428 static void insert_to_mm_slots_hash(struct mm_struct *mm,
429 struct mm_slot *mm_slot)
430 {
431 mm_slot->mm = mm;
432 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
433 }
434
khugepaged_test_exit(struct mm_struct * mm)435 static inline int khugepaged_test_exit(struct mm_struct *mm)
436 {
437 return atomic_read(&mm->mm_users) == 0;
438 }
439
hugepage_vma_check(struct vm_area_struct * vma,unsigned long vm_flags)440 static bool hugepage_vma_check(struct vm_area_struct *vma,
441 unsigned long vm_flags)
442 {
443 if (!transhuge_vma_enabled(vma, vm_flags))
444 return false;
445
446 if (vma->vm_file && !IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) -
447 vma->vm_pgoff, HPAGE_PMD_NR))
448 return false;
449
450 /* Enabled via shmem mount options or sysfs settings. */
451 if (shmem_file(vma->vm_file))
452 return shmem_huge_enabled(vma);
453
454 /* THP settings require madvise. */
455 if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
456 return false;
457
458 /* Only regular file is valid */
459 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
460 !inode_is_open_for_write(vma->vm_file->f_inode) &&
461 (vm_flags & VM_EXEC)) {
462 struct inode *inode = vma->vm_file->f_inode;
463
464 return S_ISREG(inode->i_mode);
465 }
466
467 if (!vma->anon_vma || vma->vm_ops)
468 return false;
469 if (vma_is_temporary_stack(vma))
470 return false;
471 return !(vm_flags & VM_NO_KHUGEPAGED);
472 }
473
__khugepaged_enter(struct mm_struct * mm)474 int __khugepaged_enter(struct mm_struct *mm)
475 {
476 struct mm_slot *mm_slot;
477 int wakeup;
478
479 mm_slot = alloc_mm_slot();
480 if (!mm_slot)
481 return -ENOMEM;
482
483 /* __khugepaged_exit() must not run from under us */
484 VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
485 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
486 free_mm_slot(mm_slot);
487 return 0;
488 }
489
490 spin_lock(&khugepaged_mm_lock);
491 insert_to_mm_slots_hash(mm, mm_slot);
492 /*
493 * Insert just behind the scanning cursor, to let the area settle
494 * down a little.
495 */
496 wakeup = list_empty(&khugepaged_scan.mm_head);
497 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
498 spin_unlock(&khugepaged_mm_lock);
499
500 mmgrab(mm);
501 if (wakeup)
502 wake_up_interruptible(&khugepaged_wait);
503
504 return 0;
505 }
506
khugepaged_enter_vma_merge(struct vm_area_struct * vma,unsigned long vm_flags)507 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
508 unsigned long vm_flags)
509 {
510 unsigned long hstart, hend;
511
512 /*
513 * khugepaged only supports read-only files for non-shmem files.
514 * khugepaged does not yet work on special mappings. And
515 * file-private shmem THP is not supported.
516 */
517 if (!hugepage_vma_check(vma, vm_flags))
518 return 0;
519
520 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
521 hend = vma->vm_end & HPAGE_PMD_MASK;
522 if (hstart < hend)
523 return khugepaged_enter(vma, vm_flags);
524 return 0;
525 }
526
__khugepaged_exit(struct mm_struct * mm)527 void __khugepaged_exit(struct mm_struct *mm)
528 {
529 struct mm_slot *mm_slot;
530 int free = 0;
531
532 spin_lock(&khugepaged_mm_lock);
533 mm_slot = get_mm_slot(mm);
534 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
535 hash_del(&mm_slot->hash);
536 list_del(&mm_slot->mm_node);
537 free = 1;
538 }
539 spin_unlock(&khugepaged_mm_lock);
540
541 if (free) {
542 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
543 free_mm_slot(mm_slot);
544 mmdrop(mm);
545 } else if (mm_slot) {
546 /*
547 * This is required to serialize against
548 * khugepaged_test_exit() (which is guaranteed to run
549 * under mmap sem read mode). Stop here (after we
550 * return all pagetables will be destroyed) until
551 * khugepaged has finished working on the pagetables
552 * under the mmap_lock.
553 */
554 mmap_write_lock(mm);
555 mmap_write_unlock(mm);
556 }
557 }
558
release_pte_page(struct page * page)559 static void release_pte_page(struct page *page)
560 {
561 mod_node_page_state(page_pgdat(page),
562 NR_ISOLATED_ANON + page_is_file_lru(page),
563 -compound_nr(page));
564 unlock_page(page);
565 putback_lru_page(page);
566 }
567
release_pte_pages(pte_t * pte,pte_t * _pte,struct list_head * compound_pagelist)568 static void release_pte_pages(pte_t *pte, pte_t *_pte,
569 struct list_head *compound_pagelist)
570 {
571 struct page *page, *tmp;
572
573 while (--_pte >= pte) {
574 pte_t pteval = *_pte;
575
576 page = pte_page(pteval);
577 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) &&
578 !PageCompound(page))
579 release_pte_page(page);
580 }
581
582 list_for_each_entry_safe(page, tmp, compound_pagelist, lru) {
583 list_del(&page->lru);
584 release_pte_page(page);
585 }
586 }
587
is_refcount_suitable(struct page * page)588 static bool is_refcount_suitable(struct page *page)
589 {
590 int expected_refcount;
591
592 expected_refcount = total_mapcount(page);
593 if (PageSwapCache(page))
594 expected_refcount += compound_nr(page);
595
596 return page_count(page) == expected_refcount;
597 }
598
__collapse_huge_page_isolate(struct vm_area_struct * vma,unsigned long address,pte_t * pte,struct list_head * compound_pagelist)599 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
600 unsigned long address,
601 pte_t *pte,
602 struct list_head *compound_pagelist)
603 {
604 struct page *page = NULL;
605 pte_t *_pte;
606 int none_or_zero = 0, shared = 0, result = 0, referenced = 0;
607 bool writable = false;
608
609 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
610 _pte++, address += PAGE_SIZE) {
611 pte_t pteval = *_pte;
612 if (pte_none(pteval) || (pte_present(pteval) &&
613 is_zero_pfn(pte_pfn(pteval)))) {
614 if (!userfaultfd_armed(vma) &&
615 ++none_or_zero <= khugepaged_max_ptes_none) {
616 continue;
617 } else {
618 result = SCAN_EXCEED_NONE_PTE;
619 goto out;
620 }
621 }
622 if (!pte_present(pteval)) {
623 result = SCAN_PTE_NON_PRESENT;
624 goto out;
625 }
626 if (pte_uffd_wp(pteval)) {
627 result = SCAN_PTE_UFFD_WP;
628 goto out;
629 }
630 page = vm_normal_page(vma, address, pteval);
631 if (unlikely(!page)) {
632 result = SCAN_PAGE_NULL;
633 goto out;
634 }
635
636 VM_BUG_ON_PAGE(!PageAnon(page), page);
637
638 if (page_mapcount(page) > 1 &&
639 ++shared > khugepaged_max_ptes_shared) {
640 result = SCAN_EXCEED_SHARED_PTE;
641 goto out;
642 }
643
644 if (PageCompound(page)) {
645 struct page *p;
646 page = compound_head(page);
647
648 /*
649 * Check if we have dealt with the compound page
650 * already
651 */
652 list_for_each_entry(p, compound_pagelist, lru) {
653 if (page == p)
654 goto next;
655 }
656 }
657
658 /*
659 * We can do it before isolate_lru_page because the
660 * page can't be freed from under us. NOTE: PG_lock
661 * is needed to serialize against split_huge_page
662 * when invoked from the VM.
663 */
664 if (!trylock_page(page)) {
665 result = SCAN_PAGE_LOCK;
666 goto out;
667 }
668
669 /*
670 * Check if the page has any GUP (or other external) pins.
671 *
672 * The page table that maps the page has been already unlinked
673 * from the page table tree and this process cannot get
674 * an additinal pin on the page.
675 *
676 * New pins can come later if the page is shared across fork,
677 * but not from this process. The other process cannot write to
678 * the page, only trigger CoW.
679 */
680 if (!is_refcount_suitable(page)) {
681 unlock_page(page);
682 result = SCAN_PAGE_COUNT;
683 goto out;
684 }
685 if (!pte_write(pteval) && PageSwapCache(page) &&
686 !reuse_swap_page(page, NULL)) {
687 /*
688 * Page is in the swap cache and cannot be re-used.
689 * It cannot be collapsed into a THP.
690 */
691 unlock_page(page);
692 result = SCAN_SWAP_CACHE_PAGE;
693 goto out;
694 }
695
696 /*
697 * Isolate the page to avoid collapsing an hugepage
698 * currently in use by the VM.
699 */
700 if (isolate_lru_page(page)) {
701 unlock_page(page);
702 result = SCAN_DEL_PAGE_LRU;
703 goto out;
704 }
705 mod_node_page_state(page_pgdat(page),
706 NR_ISOLATED_ANON + page_is_file_lru(page),
707 compound_nr(page));
708 VM_BUG_ON_PAGE(!PageLocked(page), page);
709 VM_BUG_ON_PAGE(PageLRU(page), page);
710
711 if (PageCompound(page))
712 list_add_tail(&page->lru, compound_pagelist);
713 next:
714 /* There should be enough young pte to collapse the page */
715 if (pte_young(pteval) ||
716 page_is_young(page) || PageReferenced(page) ||
717 mmu_notifier_test_young(vma->vm_mm, address))
718 referenced++;
719
720 if (pte_write(pteval))
721 writable = true;
722 }
723
724 if (unlikely(!writable)) {
725 result = SCAN_PAGE_RO;
726 } else if (unlikely(!referenced)) {
727 result = SCAN_LACK_REFERENCED_PAGE;
728 } else {
729 result = SCAN_SUCCEED;
730 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
731 referenced, writable, result);
732 return 1;
733 }
734 out:
735 release_pte_pages(pte, _pte, compound_pagelist);
736 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
737 referenced, writable, result);
738 return 0;
739 }
740
__collapse_huge_page_copy(pte_t * pte,struct page * page,struct vm_area_struct * vma,unsigned long address,spinlock_t * ptl,struct list_head * compound_pagelist)741 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
742 struct vm_area_struct *vma,
743 unsigned long address,
744 spinlock_t *ptl,
745 struct list_head *compound_pagelist)
746 {
747 struct page *src_page, *tmp;
748 pte_t *_pte;
749 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
750 _pte++, page++, address += PAGE_SIZE) {
751 pte_t pteval = *_pte;
752
753 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
754 clear_user_highpage(page, address);
755 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
756 if (is_zero_pfn(pte_pfn(pteval))) {
757 /*
758 * ptl mostly unnecessary.
759 */
760 spin_lock(ptl);
761 /*
762 * paravirt calls inside pte_clear here are
763 * superfluous.
764 */
765 pte_clear(vma->vm_mm, address, _pte);
766 spin_unlock(ptl);
767 }
768 } else {
769 src_page = pte_page(pteval);
770 copy_user_highpage(page, src_page, address, vma);
771 if (!PageCompound(src_page))
772 release_pte_page(src_page);
773 /*
774 * ptl mostly unnecessary, but preempt has to
775 * be disabled to update the per-cpu stats
776 * inside page_remove_rmap().
777 */
778 spin_lock(ptl);
779 /*
780 * paravirt calls inside pte_clear here are
781 * superfluous.
782 */
783 pte_clear(vma->vm_mm, address, _pte);
784 page_remove_rmap(src_page, false);
785 spin_unlock(ptl);
786 free_page_and_swap_cache(src_page);
787 }
788 }
789
790 list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
791 list_del(&src_page->lru);
792 release_pte_page(src_page);
793 }
794 }
795
khugepaged_alloc_sleep(void)796 static void khugepaged_alloc_sleep(void)
797 {
798 DEFINE_WAIT(wait);
799
800 add_wait_queue(&khugepaged_wait, &wait);
801 freezable_schedule_timeout_interruptible(
802 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
803 remove_wait_queue(&khugepaged_wait, &wait);
804 }
805
806 static int khugepaged_node_load[MAX_NUMNODES];
807
khugepaged_scan_abort(int nid)808 static bool khugepaged_scan_abort(int nid)
809 {
810 int i;
811
812 /*
813 * If node_reclaim_mode is disabled, then no extra effort is made to
814 * allocate memory locally.
815 */
816 if (!node_reclaim_mode)
817 return false;
818
819 /* If there is a count for this node already, it must be acceptable */
820 if (khugepaged_node_load[nid])
821 return false;
822
823 for (i = 0; i < MAX_NUMNODES; i++) {
824 if (!khugepaged_node_load[i])
825 continue;
826 if (node_distance(nid, i) > node_reclaim_distance)
827 return true;
828 }
829 return false;
830 }
831
832 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
alloc_hugepage_khugepaged_gfpmask(void)833 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
834 {
835 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
836 }
837
838 #ifdef CONFIG_NUMA
khugepaged_find_target_node(void)839 static int khugepaged_find_target_node(void)
840 {
841 static int last_khugepaged_target_node = NUMA_NO_NODE;
842 int nid, target_node = 0, max_value = 0;
843
844 /* find first node with max normal pages hit */
845 for (nid = 0; nid < MAX_NUMNODES; nid++)
846 if (khugepaged_node_load[nid] > max_value) {
847 max_value = khugepaged_node_load[nid];
848 target_node = nid;
849 }
850
851 /* do some balance if several nodes have the same hit record */
852 if (target_node <= last_khugepaged_target_node)
853 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
854 nid++)
855 if (max_value == khugepaged_node_load[nid]) {
856 target_node = nid;
857 break;
858 }
859
860 last_khugepaged_target_node = target_node;
861 return target_node;
862 }
863
khugepaged_prealloc_page(struct page ** hpage,bool * wait)864 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
865 {
866 if (IS_ERR(*hpage)) {
867 if (!*wait)
868 return false;
869
870 *wait = false;
871 *hpage = NULL;
872 khugepaged_alloc_sleep();
873 } else if (*hpage) {
874 put_page(*hpage);
875 *hpage = NULL;
876 }
877
878 return true;
879 }
880
881 static struct page *
khugepaged_alloc_page(struct page ** hpage,gfp_t gfp,int node)882 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
883 {
884 VM_BUG_ON_PAGE(*hpage, *hpage);
885
886 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
887 if (unlikely(!*hpage)) {
888 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
889 *hpage = ERR_PTR(-ENOMEM);
890 return NULL;
891 }
892
893 prep_transhuge_page(*hpage);
894 count_vm_event(THP_COLLAPSE_ALLOC);
895 return *hpage;
896 }
897 #else
khugepaged_find_target_node(void)898 static int khugepaged_find_target_node(void)
899 {
900 return 0;
901 }
902
alloc_khugepaged_hugepage(void)903 static inline struct page *alloc_khugepaged_hugepage(void)
904 {
905 struct page *page;
906
907 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
908 HPAGE_PMD_ORDER);
909 if (page)
910 prep_transhuge_page(page);
911 return page;
912 }
913
khugepaged_alloc_hugepage(bool * wait)914 static struct page *khugepaged_alloc_hugepage(bool *wait)
915 {
916 struct page *hpage;
917
918 do {
919 hpage = alloc_khugepaged_hugepage();
920 if (!hpage) {
921 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
922 if (!*wait)
923 return NULL;
924
925 *wait = false;
926 khugepaged_alloc_sleep();
927 } else
928 count_vm_event(THP_COLLAPSE_ALLOC);
929 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
930
931 return hpage;
932 }
933
khugepaged_prealloc_page(struct page ** hpage,bool * wait)934 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
935 {
936 /*
937 * If the hpage allocated earlier was briefly exposed in page cache
938 * before collapse_file() failed, it is possible that racing lookups
939 * have not yet completed, and would then be unpleasantly surprised by
940 * finding the hpage reused for the same mapping at a different offset.
941 * Just release the previous allocation if there is any danger of that.
942 */
943 if (*hpage && page_count(*hpage) > 1) {
944 put_page(*hpage);
945 *hpage = NULL;
946 }
947
948 if (!*hpage)
949 *hpage = khugepaged_alloc_hugepage(wait);
950
951 if (unlikely(!*hpage))
952 return false;
953
954 return true;
955 }
956
957 static struct page *
khugepaged_alloc_page(struct page ** hpage,gfp_t gfp,int node)958 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
959 {
960 VM_BUG_ON(!*hpage);
961
962 return *hpage;
963 }
964 #endif
965
966 /*
967 * If mmap_lock temporarily dropped, revalidate vma
968 * before taking mmap_lock.
969 * Return 0 if succeeds, otherwise return none-zero
970 * value (scan code).
971 */
972
hugepage_vma_revalidate(struct mm_struct * mm,unsigned long address,struct vm_area_struct ** vmap)973 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
974 struct vm_area_struct **vmap)
975 {
976 struct vm_area_struct *vma;
977 unsigned long hstart, hend;
978
979 if (unlikely(khugepaged_test_exit(mm)))
980 return SCAN_ANY_PROCESS;
981
982 *vmap = vma = find_vma(mm, address);
983 if (!vma)
984 return SCAN_VMA_NULL;
985
986 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
987 hend = vma->vm_end & HPAGE_PMD_MASK;
988 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
989 return SCAN_ADDRESS_RANGE;
990 if (!hugepage_vma_check(vma, vma->vm_flags))
991 return SCAN_VMA_CHECK;
992 /* Anon VMA expected */
993 if (!vma->anon_vma || vma->vm_ops)
994 return SCAN_VMA_CHECK;
995 return 0;
996 }
997
998 /*
999 * Bring missing pages in from swap, to complete THP collapse.
1000 * Only done if khugepaged_scan_pmd believes it is worthwhile.
1001 *
1002 * Called and returns without pte mapped or spinlocks held,
1003 * but with mmap_lock held to protect against vma changes.
1004 */
1005
__collapse_huge_page_swapin(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long haddr,pmd_t * pmd,int referenced)1006 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
1007 struct vm_area_struct *vma,
1008 unsigned long haddr, pmd_t *pmd,
1009 int referenced)
1010 {
1011 int swapped_in = 0;
1012 vm_fault_t ret = 0;
1013 unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);
1014
1015 for (address = haddr; address < end; address += PAGE_SIZE) {
1016 struct vm_fault vmf = {
1017 .vma = vma,
1018 .address = address,
1019 .pgoff = linear_page_index(vma, haddr),
1020 .flags = FAULT_FLAG_ALLOW_RETRY,
1021 .pmd = pmd,
1022 .vma_flags = vma->vm_flags,
1023 .vma_page_prot = vma->vm_page_prot,
1024 };
1025
1026 vmf.pte = pte_offset_map(pmd, address);
1027 vmf.orig_pte = *vmf.pte;
1028 if (!is_swap_pte(vmf.orig_pte)) {
1029 pte_unmap(vmf.pte);
1030 continue;
1031 }
1032 swapped_in++;
1033 ret = do_swap_page(&vmf);
1034
1035 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1036 if (ret & VM_FAULT_RETRY) {
1037 mmap_read_lock(mm);
1038 if (hugepage_vma_revalidate(mm, haddr, &vma)) {
1039 /* vma is no longer available, don't continue to swapin */
1040 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1041 return false;
1042 }
1043 /* check if the pmd is still valid */
1044 if (mm_find_pmd(mm, haddr) != pmd) {
1045 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1046 return false;
1047 }
1048 }
1049 if (ret & VM_FAULT_ERROR) {
1050 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
1051 return false;
1052 }
1053 }
1054
1055 /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1056 if (swapped_in)
1057 lru_add_drain();
1058
1059 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
1060 return true;
1061 }
1062
collapse_huge_page(struct mm_struct * mm,unsigned long address,struct page ** hpage,int node,int referenced,int unmapped)1063 static void collapse_huge_page(struct mm_struct *mm,
1064 unsigned long address,
1065 struct page **hpage,
1066 int node, int referenced, int unmapped)
1067 {
1068 LIST_HEAD(compound_pagelist);
1069 pmd_t *pmd, _pmd;
1070 pte_t *pte;
1071 pgtable_t pgtable;
1072 struct page *new_page;
1073 spinlock_t *pmd_ptl, *pte_ptl;
1074 int isolated = 0, result = 0;
1075 struct vm_area_struct *vma;
1076 struct mmu_notifier_range range;
1077 gfp_t gfp;
1078
1079 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1080
1081 /* Only allocate from the target node */
1082 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1083
1084 /*
1085 * Before allocating the hugepage, release the mmap_lock read lock.
1086 * The allocation can take potentially a long time if it involves
1087 * sync compaction, and we do not need to hold the mmap_lock during
1088 * that. We will recheck the vma after taking it again in write mode.
1089 */
1090 mmap_read_unlock(mm);
1091 new_page = khugepaged_alloc_page(hpage, gfp, node);
1092 if (!new_page) {
1093 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1094 goto out_nolock;
1095 }
1096
1097 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1098 result = SCAN_CGROUP_CHARGE_FAIL;
1099 goto out_nolock;
1100 }
1101 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1102
1103 mmap_read_lock(mm);
1104 result = hugepage_vma_revalidate(mm, address, &vma);
1105 if (result) {
1106 mmap_read_unlock(mm);
1107 goto out_nolock;
1108 }
1109
1110 pmd = mm_find_pmd(mm, address);
1111 if (!pmd) {
1112 result = SCAN_PMD_NULL;
1113 mmap_read_unlock(mm);
1114 goto out_nolock;
1115 }
1116
1117 /*
1118 * __collapse_huge_page_swapin always returns with mmap_lock locked.
1119 * If it fails, we release mmap_lock and jump out_nolock.
1120 * Continuing to collapse causes inconsistency.
1121 */
1122 if (unmapped && !__collapse_huge_page_swapin(mm, vma, address,
1123 pmd, referenced)) {
1124 mmap_read_unlock(mm);
1125 goto out_nolock;
1126 }
1127
1128 mmap_read_unlock(mm);
1129 /*
1130 * Prevent all access to pagetables with the exception of
1131 * gup_fast later handled by the ptep_clear_flush and the VM
1132 * handled by the anon_vma lock + PG_lock.
1133 */
1134 mmap_write_lock(mm);
1135 result = hugepage_vma_revalidate(mm, address, &vma);
1136 if (result)
1137 goto out;
1138 /* check if the pmd is still valid */
1139 if (mm_find_pmd(mm, address) != pmd)
1140 goto out;
1141
1142 vm_write_begin(vma);
1143 anon_vma_lock_write(vma->anon_vma);
1144
1145 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1146 address, address + HPAGE_PMD_SIZE);
1147 mmu_notifier_invalidate_range_start(&range);
1148
1149 pte = pte_offset_map(pmd, address);
1150 pte_ptl = pte_lockptr(mm, pmd);
1151
1152 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1153 /*
1154 * This removes any huge TLB entry from the CPU so we won't allow
1155 * huge and small TLB entries for the same virtual address to
1156 * avoid the risk of CPU bugs in that area.
1157 *
1158 * Parallel fast GUP is fine since fast GUP will back off when
1159 * it detects PMD is changed.
1160 */
1161 _pmd = pmdp_collapse_flush(vma, address, pmd);
1162 spin_unlock(pmd_ptl);
1163 mmu_notifier_invalidate_range_end(&range);
1164 tlb_remove_table_sync_one();
1165
1166 spin_lock(pte_ptl);
1167 isolated = __collapse_huge_page_isolate(vma, address, pte,
1168 &compound_pagelist);
1169 spin_unlock(pte_ptl);
1170
1171 if (unlikely(!isolated)) {
1172 pte_unmap(pte);
1173 spin_lock(pmd_ptl);
1174 BUG_ON(!pmd_none(*pmd));
1175 /*
1176 * We can only use set_pmd_at when establishing
1177 * hugepmds and never for establishing regular pmds that
1178 * points to regular pagetables. Use pmd_populate for that
1179 */
1180 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1181 spin_unlock(pmd_ptl);
1182 anon_vma_unlock_write(vma->anon_vma);
1183 vm_write_end(vma);
1184 result = SCAN_FAIL;
1185 goto out;
1186 }
1187
1188 /*
1189 * All pages are isolated and locked so anon_vma rmap
1190 * can't run anymore.
1191 */
1192 anon_vma_unlock_write(vma->anon_vma);
1193
1194 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl,
1195 &compound_pagelist);
1196 pte_unmap(pte);
1197 __SetPageUptodate(new_page);
1198 pgtable = pmd_pgtable(_pmd);
1199
1200 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1201 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1202
1203 /*
1204 * spin_lock() below is not the equivalent of smp_wmb(), so
1205 * this is needed to avoid the copy_huge_page writes to become
1206 * visible after the set_pmd_at() write.
1207 */
1208 smp_wmb();
1209
1210 spin_lock(pmd_ptl);
1211 BUG_ON(!pmd_none(*pmd));
1212 page_add_new_anon_rmap(new_page, vma, address, true);
1213 lru_cache_add_inactive_or_unevictable(new_page, vma);
1214 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1215 set_pmd_at(mm, address, pmd, _pmd);
1216 update_mmu_cache_pmd(vma, address, pmd);
1217 spin_unlock(pmd_ptl);
1218 vm_write_end(vma);
1219
1220 *hpage = NULL;
1221
1222 khugepaged_pages_collapsed++;
1223 result = SCAN_SUCCEED;
1224 out_up_write:
1225 mmap_write_unlock(mm);
1226 out_nolock:
1227 if (!IS_ERR_OR_NULL(*hpage))
1228 mem_cgroup_uncharge(*hpage);
1229 trace_mm_collapse_huge_page(mm, isolated, result);
1230 return;
1231 out:
1232 goto out_up_write;
1233 }
1234
khugepaged_scan_pmd(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,struct page ** hpage)1235 static int khugepaged_scan_pmd(struct mm_struct *mm,
1236 struct vm_area_struct *vma,
1237 unsigned long address,
1238 struct page **hpage)
1239 {
1240 pmd_t *pmd;
1241 pte_t *pte, *_pte;
1242 int ret = 0, result = 0, referenced = 0;
1243 int none_or_zero = 0, shared = 0;
1244 struct page *page = NULL;
1245 unsigned long _address;
1246 spinlock_t *ptl;
1247 int node = NUMA_NO_NODE, unmapped = 0;
1248 bool writable = false;
1249
1250 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1251
1252 pmd = mm_find_pmd(mm, address);
1253 if (!pmd) {
1254 result = SCAN_PMD_NULL;
1255 goto out;
1256 }
1257
1258 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1259 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1260 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1261 _pte++, _address += PAGE_SIZE) {
1262 pte_t pteval = *_pte;
1263 if (is_swap_pte(pteval)) {
1264 if (++unmapped <= khugepaged_max_ptes_swap) {
1265 /*
1266 * Always be strict with uffd-wp
1267 * enabled swap entries. Please see
1268 * comment below for pte_uffd_wp().
1269 */
1270 if (pte_swp_uffd_wp(pteval)) {
1271 result = SCAN_PTE_UFFD_WP;
1272 goto out_unmap;
1273 }
1274 continue;
1275 } else {
1276 result = SCAN_EXCEED_SWAP_PTE;
1277 goto out_unmap;
1278 }
1279 }
1280 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1281 if (!userfaultfd_armed(vma) &&
1282 ++none_or_zero <= khugepaged_max_ptes_none) {
1283 continue;
1284 } else {
1285 result = SCAN_EXCEED_NONE_PTE;
1286 goto out_unmap;
1287 }
1288 }
1289 if (!pte_present(pteval)) {
1290 result = SCAN_PTE_NON_PRESENT;
1291 goto out_unmap;
1292 }
1293 if (pte_uffd_wp(pteval)) {
1294 /*
1295 * Don't collapse the page if any of the small
1296 * PTEs are armed with uffd write protection.
1297 * Here we can also mark the new huge pmd as
1298 * write protected if any of the small ones is
1299 * marked but that could bring uknown
1300 * userfault messages that falls outside of
1301 * the registered range. So, just be simple.
1302 */
1303 result = SCAN_PTE_UFFD_WP;
1304 goto out_unmap;
1305 }
1306 if (pte_write(pteval))
1307 writable = true;
1308
1309 page = vm_normal_page(vma, _address, pteval);
1310 if (unlikely(!page)) {
1311 result = SCAN_PAGE_NULL;
1312 goto out_unmap;
1313 }
1314
1315 if (page_mapcount(page) > 1 &&
1316 ++shared > khugepaged_max_ptes_shared) {
1317 result = SCAN_EXCEED_SHARED_PTE;
1318 goto out_unmap;
1319 }
1320
1321 page = compound_head(page);
1322
1323 /*
1324 * Record which node the original page is from and save this
1325 * information to khugepaged_node_load[].
1326 * Khupaged will allocate hugepage from the node has the max
1327 * hit record.
1328 */
1329 node = page_to_nid(page);
1330 if (khugepaged_scan_abort(node)) {
1331 result = SCAN_SCAN_ABORT;
1332 goto out_unmap;
1333 }
1334 khugepaged_node_load[node]++;
1335 if (!PageLRU(page)) {
1336 result = SCAN_PAGE_LRU;
1337 goto out_unmap;
1338 }
1339 if (PageLocked(page)) {
1340 result = SCAN_PAGE_LOCK;
1341 goto out_unmap;
1342 }
1343 if (!PageAnon(page)) {
1344 result = SCAN_PAGE_ANON;
1345 goto out_unmap;
1346 }
1347
1348 /*
1349 * Check if the page has any GUP (or other external) pins.
1350 *
1351 * Here the check is racy it may see totmal_mapcount > refcount
1352 * in some cases.
1353 * For example, one process with one forked child process.
1354 * The parent has the PMD split due to MADV_DONTNEED, then
1355 * the child is trying unmap the whole PMD, but khugepaged
1356 * may be scanning the parent between the child has
1357 * PageDoubleMap flag cleared and dec the mapcount. So
1358 * khugepaged may see total_mapcount > refcount.
1359 *
1360 * But such case is ephemeral we could always retry collapse
1361 * later. However it may report false positive if the page
1362 * has excessive GUP pins (i.e. 512). Anyway the same check
1363 * will be done again later the risk seems low.
1364 */
1365 if (!is_refcount_suitable(page)) {
1366 result = SCAN_PAGE_COUNT;
1367 goto out_unmap;
1368 }
1369 if (pte_young(pteval) ||
1370 page_is_young(page) || PageReferenced(page) ||
1371 mmu_notifier_test_young(vma->vm_mm, address))
1372 referenced++;
1373 }
1374 if (!writable) {
1375 result = SCAN_PAGE_RO;
1376 } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) {
1377 result = SCAN_LACK_REFERENCED_PAGE;
1378 } else {
1379 result = SCAN_SUCCEED;
1380 ret = 1;
1381 }
1382 out_unmap:
1383 pte_unmap_unlock(pte, ptl);
1384 if (ret) {
1385 node = khugepaged_find_target_node();
1386 /* collapse_huge_page will return with the mmap_lock released */
1387 collapse_huge_page(mm, address, hpage, node,
1388 referenced, unmapped);
1389 }
1390 out:
1391 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1392 none_or_zero, result, unmapped);
1393 return ret;
1394 }
1395
collect_mm_slot(struct mm_slot * mm_slot)1396 static void collect_mm_slot(struct mm_slot *mm_slot)
1397 {
1398 struct mm_struct *mm = mm_slot->mm;
1399
1400 lockdep_assert_held(&khugepaged_mm_lock);
1401
1402 if (khugepaged_test_exit(mm)) {
1403 /* free mm_slot */
1404 hash_del(&mm_slot->hash);
1405 list_del(&mm_slot->mm_node);
1406
1407 /*
1408 * Not strictly needed because the mm exited already.
1409 *
1410 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1411 */
1412
1413 /* khugepaged_mm_lock actually not necessary for the below */
1414 free_mm_slot(mm_slot);
1415 mmdrop(mm);
1416 }
1417 }
1418
1419 #ifdef CONFIG_SHMEM
1420 /*
1421 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1422 * khugepaged should try to collapse the page table.
1423 */
khugepaged_add_pte_mapped_thp(struct mm_struct * mm,unsigned long addr)1424 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1425 unsigned long addr)
1426 {
1427 struct mm_slot *mm_slot;
1428
1429 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1430
1431 spin_lock(&khugepaged_mm_lock);
1432 mm_slot = get_mm_slot(mm);
1433 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1434 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1435 spin_unlock(&khugepaged_mm_lock);
1436 return 0;
1437 }
1438
1439 /**
1440 * Try to collapse a pte-mapped THP for mm at address haddr.
1441 *
1442 * This function checks whether all the PTEs in the PMD are pointing to the
1443 * right THP. If so, retract the page table so the THP can refault in with
1444 * as pmd-mapped.
1445 */
collapse_pte_mapped_thp(struct mm_struct * mm,unsigned long addr)1446 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1447 {
1448 unsigned long haddr = addr & HPAGE_PMD_MASK;
1449 struct vm_area_struct *vma = find_vma(mm, haddr);
1450 struct page *hpage;
1451 pte_t *start_pte, *pte;
1452 pmd_t *pmd, _pmd;
1453 spinlock_t *ptl;
1454 int count = 0;
1455 int i;
1456 struct mmu_notifier_range range;
1457
1458 if (!vma || !vma->vm_file ||
1459 vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1460 return;
1461
1462 /*
1463 * This vm_flags may not have VM_HUGEPAGE if the page was not
1464 * collapsed by this mm. But we can still collapse if the page is
1465 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1466 * will not fail the vma for missing VM_HUGEPAGE
1467 */
1468 if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1469 return;
1470
1471 hpage = find_lock_page(vma->vm_file->f_mapping,
1472 linear_page_index(vma, haddr));
1473 if (!hpage)
1474 return;
1475
1476 if (!PageHead(hpage))
1477 goto drop_hpage;
1478
1479 pmd = mm_find_pmd(mm, haddr);
1480 if (!pmd)
1481 goto drop_hpage;
1482
1483 vm_write_begin(vma);
1484
1485 /*
1486 * We need to lock the mapping so that from here on, only GUP-fast and
1487 * hardware page walks can access the parts of the page tables that
1488 * we're operating on.
1489 */
1490 i_mmap_lock_write(vma->vm_file->f_mapping);
1491
1492 /*
1493 * This spinlock should be unnecessary: Nobody else should be accessing
1494 * the page tables under spinlock protection here, only
1495 * lockless_pages_from_mm() and the hardware page walker can access page
1496 * tables while all the high-level locks are held in write mode.
1497 */
1498 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1499
1500 /* step 1: check all mapped PTEs are to the right huge page */
1501 for (i = 0, addr = haddr, pte = start_pte;
1502 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1503 struct page *page;
1504
1505 /* empty pte, skip */
1506 if (pte_none(*pte))
1507 continue;
1508
1509 /* page swapped out, abort */
1510 if (!pte_present(*pte))
1511 goto abort;
1512
1513 page = vm_normal_page(vma, addr, *pte);
1514
1515 /*
1516 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1517 * page table, but the new page will not be a subpage of hpage.
1518 */
1519 if (hpage + i != page)
1520 goto abort;
1521 count++;
1522 }
1523
1524 /* step 2: adjust rmap */
1525 for (i = 0, addr = haddr, pte = start_pte;
1526 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1527 struct page *page;
1528
1529 if (pte_none(*pte))
1530 continue;
1531 page = vm_normal_page(vma, addr, *pte);
1532 page_remove_rmap(page, false);
1533 }
1534
1535 pte_unmap_unlock(start_pte, ptl);
1536
1537 /* step 3: set proper refcount and mm_counters. */
1538 if (count) {
1539 page_ref_sub(hpage, count);
1540 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1541 }
1542
1543 /* step 4: collapse pmd */
1544 /* we make no change to anon, but protect concurrent anon page lookup */
1545 if (vma->anon_vma)
1546 anon_vma_lock_write(vma->anon_vma);
1547
1548 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, haddr,
1549 haddr + HPAGE_PMD_SIZE);
1550 mmu_notifier_invalidate_range_start(&range);
1551 _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1552 vm_write_end(vma);
1553 mm_dec_nr_ptes(mm);
1554 tlb_remove_table_sync_one();
1555 mmu_notifier_invalidate_range_end(&range);
1556 pte_free(mm, pmd_pgtable(_pmd));
1557
1558 if (vma->anon_vma)
1559 anon_vma_unlock_write(vma->anon_vma);
1560 i_mmap_unlock_write(vma->vm_file->f_mapping);
1561
1562 drop_hpage:
1563 unlock_page(hpage);
1564 put_page(hpage);
1565 return;
1566
1567 abort:
1568 pte_unmap_unlock(start_pte, ptl);
1569 vm_write_end(vma);
1570 i_mmap_unlock_write(vma->vm_file->f_mapping);
1571 goto drop_hpage;
1572 }
1573
khugepaged_collapse_pte_mapped_thps(struct mm_slot * mm_slot)1574 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1575 {
1576 struct mm_struct *mm = mm_slot->mm;
1577 int i;
1578
1579 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1580 return 0;
1581
1582 if (!mmap_write_trylock(mm))
1583 return -EBUSY;
1584
1585 if (unlikely(khugepaged_test_exit(mm)))
1586 goto out;
1587
1588 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1589 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1590
1591 out:
1592 mm_slot->nr_pte_mapped_thp = 0;
1593 mmap_write_unlock(mm);
1594 return 0;
1595 }
1596
retract_page_tables(struct address_space * mapping,pgoff_t pgoff)1597 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1598 {
1599 struct vm_area_struct *vma;
1600 struct mm_struct *mm;
1601 unsigned long addr;
1602 pmd_t *pmd, _pmd;
1603
1604 i_mmap_lock_write(mapping);
1605 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1606 /*
1607 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1608 * got written to. These VMAs are likely not worth investing
1609 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1610 * later.
1611 *
1612 * Not that vma->anon_vma check is racy: it can be set up after
1613 * the check but before we took mmap_lock by the fault path.
1614 * But page lock would prevent establishing any new ptes of the
1615 * page, so we are safe.
1616 *
1617 * An alternative would be drop the check, but check that page
1618 * table is clear before calling pmdp_collapse_flush() under
1619 * ptl. It has higher chance to recover THP for the VMA, but
1620 * has higher cost too. It would also probably require locking
1621 * the anon_vma.
1622 */
1623 if (vma->anon_vma)
1624 continue;
1625 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1626 if (addr & ~HPAGE_PMD_MASK)
1627 continue;
1628 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1629 continue;
1630 mm = vma->vm_mm;
1631 pmd = mm_find_pmd(mm, addr);
1632 if (!pmd)
1633 continue;
1634 /*
1635 * We need exclusive mmap_lock to retract page table.
1636 *
1637 * We use trylock due to lock inversion: we need to acquire
1638 * mmap_lock while holding page lock. Fault path does it in
1639 * reverse order. Trylock is a way to avoid deadlock.
1640 */
1641 if (mmap_write_trylock(mm)) {
1642 if (!khugepaged_test_exit(mm)) {
1643 struct mmu_notifier_range range;
1644
1645 vm_write_begin(vma);
1646 mmu_notifier_range_init(&range,
1647 MMU_NOTIFY_CLEAR, 0,
1648 NULL, mm, addr,
1649 addr + HPAGE_PMD_SIZE);
1650 mmu_notifier_invalidate_range_start(&range);
1651 /* assume page table is clear */
1652 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1653 vm_write_end(vma);
1654 mm_dec_nr_ptes(mm);
1655 tlb_remove_table_sync_one();
1656 pte_free(mm, pmd_pgtable(_pmd));
1657 mmu_notifier_invalidate_range_end(&range);
1658 }
1659 mmap_write_unlock(mm);
1660 } else {
1661 /* Try again later */
1662 khugepaged_add_pte_mapped_thp(mm, addr);
1663 }
1664 }
1665 i_mmap_unlock_write(mapping);
1666 }
1667
1668 /**
1669 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1670 *
1671 * Basic scheme is simple, details are more complex:
1672 * - allocate and lock a new huge page;
1673 * - scan page cache replacing old pages with the new one
1674 * + swap/gup in pages if necessary;
1675 * + fill in gaps;
1676 * + keep old pages around in case rollback is required;
1677 * - if replacing succeeds:
1678 * + copy data over;
1679 * + free old pages;
1680 * + unlock huge page;
1681 * - if replacing failed;
1682 * + put all pages back and unfreeze them;
1683 * + restore gaps in the page cache;
1684 * + unlock and free huge page;
1685 */
collapse_file(struct mm_struct * mm,struct file * file,pgoff_t start,struct page ** hpage,int node)1686 static void collapse_file(struct mm_struct *mm,
1687 struct file *file, pgoff_t start,
1688 struct page **hpage, int node)
1689 {
1690 struct address_space *mapping = file->f_mapping;
1691 gfp_t gfp;
1692 struct page *new_page;
1693 pgoff_t index, end = start + HPAGE_PMD_NR;
1694 LIST_HEAD(pagelist);
1695 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1696 int nr_none = 0, result = SCAN_SUCCEED;
1697 bool is_shmem = shmem_file(file);
1698
1699 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1700 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1701
1702 /* Only allocate from the target node */
1703 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1704
1705 new_page = khugepaged_alloc_page(hpage, gfp, node);
1706 if (!new_page) {
1707 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1708 goto out;
1709 }
1710
1711 if (unlikely(mem_cgroup_charge(new_page, mm, gfp))) {
1712 result = SCAN_CGROUP_CHARGE_FAIL;
1713 goto out;
1714 }
1715 count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC);
1716
1717 /* This will be less messy when we use multi-index entries */
1718 do {
1719 xas_lock_irq(&xas);
1720 xas_create_range(&xas);
1721 if (!xas_error(&xas))
1722 break;
1723 xas_unlock_irq(&xas);
1724 if (!xas_nomem(&xas, GFP_KERNEL)) {
1725 result = SCAN_FAIL;
1726 goto out;
1727 }
1728 } while (1);
1729
1730 __SetPageLocked(new_page);
1731 if (is_shmem)
1732 __SetPageSwapBacked(new_page);
1733 new_page->index = start;
1734 new_page->mapping = mapping;
1735
1736 /*
1737 * At this point the new_page is locked and not up-to-date.
1738 * It's safe to insert it into the page cache, because nobody would
1739 * be able to map it or use it in another way until we unlock it.
1740 */
1741
1742 xas_set(&xas, start);
1743 for (index = start; index < end; index++) {
1744 struct page *page = xas_next(&xas);
1745
1746 VM_BUG_ON(index != xas.xa_index);
1747 if (is_shmem) {
1748 if (!page) {
1749 /*
1750 * Stop if extent has been truncated or
1751 * hole-punched, and is now completely
1752 * empty.
1753 */
1754 if (index == start) {
1755 if (!xas_next_entry(&xas, end - 1)) {
1756 result = SCAN_TRUNCATED;
1757 goto xa_locked;
1758 }
1759 xas_set(&xas, index);
1760 }
1761 if (!shmem_charge(mapping->host, 1)) {
1762 result = SCAN_FAIL;
1763 goto xa_locked;
1764 }
1765 xas_store(&xas, new_page);
1766 nr_none++;
1767 continue;
1768 }
1769
1770 if (xa_is_value(page) || !PageUptodate(page)) {
1771 xas_unlock_irq(&xas);
1772 /* swap in or instantiate fallocated page */
1773 if (shmem_getpage(mapping->host, index, &page,
1774 SGP_NOHUGE)) {
1775 result = SCAN_FAIL;
1776 goto xa_unlocked;
1777 }
1778 } else if (trylock_page(page)) {
1779 get_page(page);
1780 xas_unlock_irq(&xas);
1781 } else {
1782 result = SCAN_PAGE_LOCK;
1783 goto xa_locked;
1784 }
1785 } else { /* !is_shmem */
1786 if (!page || xa_is_value(page)) {
1787 xas_unlock_irq(&xas);
1788 page_cache_sync_readahead(mapping, &file->f_ra,
1789 file, index,
1790 end - index);
1791 /* drain pagevecs to help isolate_lru_page() */
1792 lru_add_drain();
1793 page = find_lock_page(mapping, index);
1794 if (unlikely(page == NULL)) {
1795 result = SCAN_FAIL;
1796 goto xa_unlocked;
1797 }
1798 } else if (PageDirty(page)) {
1799 /*
1800 * khugepaged only works on read-only fd,
1801 * so this page is dirty because it hasn't
1802 * been flushed since first write. There
1803 * won't be new dirty pages.
1804 *
1805 * Trigger async flush here and hope the
1806 * writeback is done when khugepaged
1807 * revisits this page.
1808 *
1809 * This is a one-off situation. We are not
1810 * forcing writeback in loop.
1811 */
1812 xas_unlock_irq(&xas);
1813 filemap_flush(mapping);
1814 result = SCAN_FAIL;
1815 goto xa_unlocked;
1816 } else if (PageWriteback(page)) {
1817 xas_unlock_irq(&xas);
1818 result = SCAN_FAIL;
1819 goto xa_unlocked;
1820 } else if (trylock_page(page)) {
1821 get_page(page);
1822 xas_unlock_irq(&xas);
1823 } else {
1824 result = SCAN_PAGE_LOCK;
1825 goto xa_locked;
1826 }
1827 }
1828
1829 /*
1830 * The page must be locked, so we can drop the i_pages lock
1831 * without racing with truncate.
1832 */
1833 VM_BUG_ON_PAGE(!PageLocked(page), page);
1834
1835 /* make sure the page is up to date */
1836 if (unlikely(!PageUptodate(page))) {
1837 result = SCAN_FAIL;
1838 goto out_unlock;
1839 }
1840
1841 /*
1842 * If file was truncated then extended, or hole-punched, before
1843 * we locked the first page, then a THP might be there already.
1844 */
1845 if (PageTransCompound(page)) {
1846 result = SCAN_PAGE_COMPOUND;
1847 goto out_unlock;
1848 }
1849
1850 if (page_mapping(page) != mapping) {
1851 result = SCAN_TRUNCATED;
1852 goto out_unlock;
1853 }
1854
1855 if (!is_shmem && (PageDirty(page) ||
1856 PageWriteback(page))) {
1857 /*
1858 * khugepaged only works on read-only fd, so this
1859 * page is dirty because it hasn't been flushed
1860 * since first write.
1861 */
1862 result = SCAN_FAIL;
1863 goto out_unlock;
1864 }
1865
1866 if (isolate_lru_page(page)) {
1867 result = SCAN_DEL_PAGE_LRU;
1868 goto out_unlock;
1869 }
1870
1871 if (page_has_private(page) &&
1872 !try_to_release_page(page, GFP_KERNEL)) {
1873 result = SCAN_PAGE_HAS_PRIVATE;
1874 putback_lru_page(page);
1875 goto out_unlock;
1876 }
1877
1878 if (page_mapped(page))
1879 unmap_mapping_pages(mapping, index, 1, false);
1880
1881 xas_lock_irq(&xas);
1882 xas_set(&xas, index);
1883
1884 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1885 VM_BUG_ON_PAGE(page_mapped(page), page);
1886
1887 /*
1888 * The page is expected to have page_count() == 3:
1889 * - we hold a pin on it;
1890 * - one reference from page cache;
1891 * - one from isolate_lru_page;
1892 */
1893 if (!page_ref_freeze(page, 3)) {
1894 result = SCAN_PAGE_COUNT;
1895 xas_unlock_irq(&xas);
1896 putback_lru_page(page);
1897 goto out_unlock;
1898 }
1899
1900 /*
1901 * Add the page to the list to be able to undo the collapse if
1902 * something go wrong.
1903 */
1904 list_add_tail(&page->lru, &pagelist);
1905
1906 /* Finally, replace with the new page. */
1907 xas_store(&xas, new_page);
1908 continue;
1909 out_unlock:
1910 unlock_page(page);
1911 put_page(page);
1912 goto xa_unlocked;
1913 }
1914
1915 if (is_shmem)
1916 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1917 else {
1918 __inc_node_page_state(new_page, NR_FILE_THPS);
1919 filemap_nr_thps_inc(mapping);
1920 /*
1921 * Paired with smp_mb() in do_dentry_open() to ensure
1922 * i_writecount is up to date and the update to nr_thps is
1923 * visible. Ensures the page cache will be truncated if the
1924 * file is opened writable.
1925 */
1926 smp_mb();
1927 if (inode_is_open_for_write(mapping->host)) {
1928 result = SCAN_FAIL;
1929 __dec_node_page_state(new_page, NR_FILE_THPS);
1930 filemap_nr_thps_dec(mapping);
1931 goto xa_locked;
1932 }
1933 }
1934
1935 if (nr_none) {
1936 __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none);
1937 if (is_shmem)
1938 __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none);
1939 }
1940
1941 xa_locked:
1942 xas_unlock_irq(&xas);
1943 xa_unlocked:
1944
1945 if (result == SCAN_SUCCEED) {
1946 struct page *page, *tmp;
1947
1948 /*
1949 * Replacing old pages with new one has succeeded, now we
1950 * need to copy the content and free the old pages.
1951 */
1952 index = start;
1953 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1954 while (index < page->index) {
1955 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1956 index++;
1957 }
1958 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1959 page);
1960 list_del(&page->lru);
1961 page->mapping = NULL;
1962 page_ref_unfreeze(page, 1);
1963 ClearPageActive(page);
1964 ClearPageUnevictable(page);
1965 unlock_page(page);
1966 put_page(page);
1967 index++;
1968 }
1969 while (index < end) {
1970 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1971 index++;
1972 }
1973
1974 SetPageUptodate(new_page);
1975 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1976 if (is_shmem)
1977 set_page_dirty(new_page);
1978 lru_cache_add(new_page);
1979
1980 /*
1981 * Remove pte page tables, so we can re-fault the page as huge.
1982 */
1983 retract_page_tables(mapping, start);
1984 *hpage = NULL;
1985
1986 khugepaged_pages_collapsed++;
1987 } else {
1988 struct page *page;
1989
1990 /* Something went wrong: roll back page cache changes */
1991 xas_lock_irq(&xas);
1992 mapping->nrpages -= nr_none;
1993
1994 if (is_shmem)
1995 shmem_uncharge(mapping->host, nr_none);
1996
1997 xas_set(&xas, start);
1998 xas_for_each(&xas, page, end - 1) {
1999 page = list_first_entry_or_null(&pagelist,
2000 struct page, lru);
2001 if (!page || xas.xa_index < page->index) {
2002 if (!nr_none)
2003 break;
2004 nr_none--;
2005 /* Put holes back where they were */
2006 xas_store(&xas, NULL);
2007 continue;
2008 }
2009
2010 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
2011
2012 /* Unfreeze the page. */
2013 list_del(&page->lru);
2014 page_ref_unfreeze(page, 2);
2015 xas_store(&xas, page);
2016 xas_pause(&xas);
2017 xas_unlock_irq(&xas);
2018 unlock_page(page);
2019 putback_lru_page(page);
2020 xas_lock_irq(&xas);
2021 }
2022 VM_BUG_ON(nr_none);
2023 xas_unlock_irq(&xas);
2024
2025 new_page->mapping = NULL;
2026 }
2027
2028 unlock_page(new_page);
2029 out:
2030 VM_BUG_ON(!list_empty(&pagelist));
2031 if (!IS_ERR_OR_NULL(*hpage))
2032 mem_cgroup_uncharge(*hpage);
2033 /* TODO: tracepoints */
2034 }
2035
khugepaged_scan_file(struct mm_struct * mm,struct file * file,pgoff_t start,struct page ** hpage)2036 static void khugepaged_scan_file(struct mm_struct *mm,
2037 struct file *file, pgoff_t start, struct page **hpage)
2038 {
2039 struct page *page = NULL;
2040 struct address_space *mapping = file->f_mapping;
2041 XA_STATE(xas, &mapping->i_pages, start);
2042 int present, swap;
2043 int node = NUMA_NO_NODE;
2044 int result = SCAN_SUCCEED;
2045
2046 present = 0;
2047 swap = 0;
2048 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2049 rcu_read_lock();
2050 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
2051 if (xas_retry(&xas, page))
2052 continue;
2053
2054 if (xa_is_value(page)) {
2055 if (++swap > khugepaged_max_ptes_swap) {
2056 result = SCAN_EXCEED_SWAP_PTE;
2057 break;
2058 }
2059 continue;
2060 }
2061
2062 if (PageTransCompound(page)) {
2063 result = SCAN_PAGE_COMPOUND;
2064 break;
2065 }
2066
2067 node = page_to_nid(page);
2068 if (khugepaged_scan_abort(node)) {
2069 result = SCAN_SCAN_ABORT;
2070 break;
2071 }
2072 khugepaged_node_load[node]++;
2073
2074 if (!PageLRU(page)) {
2075 result = SCAN_PAGE_LRU;
2076 break;
2077 }
2078
2079 if (page_count(page) !=
2080 1 + page_mapcount(page) + page_has_private(page)) {
2081 result = SCAN_PAGE_COUNT;
2082 break;
2083 }
2084
2085 /*
2086 * We probably should check if the page is referenced here, but
2087 * nobody would transfer pte_young() to PageReferenced() for us.
2088 * And rmap walk here is just too costly...
2089 */
2090
2091 present++;
2092
2093 if (need_resched()) {
2094 xas_pause(&xas);
2095 cond_resched_rcu();
2096 }
2097 }
2098 rcu_read_unlock();
2099
2100 if (result == SCAN_SUCCEED) {
2101 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2102 result = SCAN_EXCEED_NONE_PTE;
2103 } else {
2104 node = khugepaged_find_target_node();
2105 collapse_file(mm, file, start, hpage, node);
2106 }
2107 }
2108
2109 /* TODO: tracepoints */
2110 }
2111 #else
khugepaged_scan_file(struct mm_struct * mm,struct file * file,pgoff_t start,struct page ** hpage)2112 static void khugepaged_scan_file(struct mm_struct *mm,
2113 struct file *file, pgoff_t start, struct page **hpage)
2114 {
2115 BUILD_BUG();
2116 }
2117
khugepaged_collapse_pte_mapped_thps(struct mm_slot * mm_slot)2118 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
2119 {
2120 return 0;
2121 }
2122 #endif
2123
khugepaged_scan_mm_slot(unsigned int pages,struct page ** hpage)2124 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2125 struct page **hpage)
2126 __releases(&khugepaged_mm_lock)
2127 __acquires(&khugepaged_mm_lock)
2128 {
2129 struct mm_slot *mm_slot;
2130 struct mm_struct *mm;
2131 struct vm_area_struct *vma;
2132 int progress = 0;
2133
2134 VM_BUG_ON(!pages);
2135 lockdep_assert_held(&khugepaged_mm_lock);
2136
2137 if (khugepaged_scan.mm_slot)
2138 mm_slot = khugepaged_scan.mm_slot;
2139 else {
2140 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2141 struct mm_slot, mm_node);
2142 khugepaged_scan.address = 0;
2143 khugepaged_scan.mm_slot = mm_slot;
2144 }
2145 spin_unlock(&khugepaged_mm_lock);
2146 khugepaged_collapse_pte_mapped_thps(mm_slot);
2147
2148 mm = mm_slot->mm;
2149 /*
2150 * Don't wait for semaphore (to avoid long wait times). Just move to
2151 * the next mm on the list.
2152 */
2153 vma = NULL;
2154 if (unlikely(!mmap_read_trylock(mm)))
2155 goto breakouterloop_mmap_lock;
2156 if (likely(!khugepaged_test_exit(mm)))
2157 vma = find_vma(mm, khugepaged_scan.address);
2158
2159 progress++;
2160 for (; vma; vma = vma->vm_next) {
2161 unsigned long hstart, hend;
2162
2163 cond_resched();
2164 if (unlikely(khugepaged_test_exit(mm))) {
2165 progress++;
2166 break;
2167 }
2168 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2169 skip:
2170 progress++;
2171 continue;
2172 }
2173 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2174 hend = vma->vm_end & HPAGE_PMD_MASK;
2175 if (hstart >= hend)
2176 goto skip;
2177 if (khugepaged_scan.address > hend)
2178 goto skip;
2179 if (khugepaged_scan.address < hstart)
2180 khugepaged_scan.address = hstart;
2181 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2182 if (shmem_file(vma->vm_file) && !shmem_huge_enabled(vma))
2183 goto skip;
2184
2185 while (khugepaged_scan.address < hend) {
2186 int ret;
2187 cond_resched();
2188 if (unlikely(khugepaged_test_exit(mm)))
2189 goto breakouterloop;
2190
2191 VM_BUG_ON(khugepaged_scan.address < hstart ||
2192 khugepaged_scan.address + HPAGE_PMD_SIZE >
2193 hend);
2194 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2195 struct file *file = get_file(vma->vm_file);
2196 pgoff_t pgoff = linear_page_index(vma,
2197 khugepaged_scan.address);
2198
2199 mmap_read_unlock(mm);
2200 ret = 1;
2201 khugepaged_scan_file(mm, file, pgoff, hpage);
2202 fput(file);
2203 } else {
2204 ret = khugepaged_scan_pmd(mm, vma,
2205 khugepaged_scan.address,
2206 hpage);
2207 }
2208 /* move to next address */
2209 khugepaged_scan.address += HPAGE_PMD_SIZE;
2210 progress += HPAGE_PMD_NR;
2211 if (ret)
2212 /* we released mmap_lock so break loop */
2213 goto breakouterloop_mmap_lock;
2214 if (progress >= pages)
2215 goto breakouterloop;
2216 }
2217 }
2218 breakouterloop:
2219 mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2220 breakouterloop_mmap_lock:
2221
2222 spin_lock(&khugepaged_mm_lock);
2223 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2224 /*
2225 * Release the current mm_slot if this mm is about to die, or
2226 * if we scanned all vmas of this mm.
2227 */
2228 if (khugepaged_test_exit(mm) || !vma) {
2229 /*
2230 * Make sure that if mm_users is reaching zero while
2231 * khugepaged runs here, khugepaged_exit will find
2232 * mm_slot not pointing to the exiting mm.
2233 */
2234 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2235 khugepaged_scan.mm_slot = list_entry(
2236 mm_slot->mm_node.next,
2237 struct mm_slot, mm_node);
2238 khugepaged_scan.address = 0;
2239 } else {
2240 khugepaged_scan.mm_slot = NULL;
2241 khugepaged_full_scans++;
2242 }
2243
2244 collect_mm_slot(mm_slot);
2245 }
2246
2247 return progress;
2248 }
2249
khugepaged_has_work(void)2250 static int khugepaged_has_work(void)
2251 {
2252 return !list_empty(&khugepaged_scan.mm_head) &&
2253 khugepaged_enabled();
2254 }
2255
khugepaged_wait_event(void)2256 static int khugepaged_wait_event(void)
2257 {
2258 return !list_empty(&khugepaged_scan.mm_head) ||
2259 kthread_should_stop();
2260 }
2261
khugepaged_do_scan(void)2262 static void khugepaged_do_scan(void)
2263 {
2264 struct page *hpage = NULL;
2265 unsigned int progress = 0, pass_through_head = 0;
2266 unsigned int pages = khugepaged_pages_to_scan;
2267 bool wait = true;
2268
2269 barrier(); /* write khugepaged_pages_to_scan to local stack */
2270
2271 lru_add_drain_all();
2272
2273 while (progress < pages) {
2274 if (!khugepaged_prealloc_page(&hpage, &wait))
2275 break;
2276
2277 cond_resched();
2278
2279 if (unlikely(kthread_should_stop() || try_to_freeze()))
2280 break;
2281
2282 spin_lock(&khugepaged_mm_lock);
2283 if (!khugepaged_scan.mm_slot)
2284 pass_through_head++;
2285 if (khugepaged_has_work() &&
2286 pass_through_head < 2)
2287 progress += khugepaged_scan_mm_slot(pages - progress,
2288 &hpage);
2289 else
2290 progress = pages;
2291 spin_unlock(&khugepaged_mm_lock);
2292 }
2293
2294 if (!IS_ERR_OR_NULL(hpage))
2295 put_page(hpage);
2296 }
2297
khugepaged_should_wakeup(void)2298 static bool khugepaged_should_wakeup(void)
2299 {
2300 return kthread_should_stop() ||
2301 time_after_eq(jiffies, khugepaged_sleep_expire);
2302 }
2303
khugepaged_wait_work(void)2304 static void khugepaged_wait_work(void)
2305 {
2306 if (khugepaged_has_work()) {
2307 const unsigned long scan_sleep_jiffies =
2308 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2309
2310 if (!scan_sleep_jiffies)
2311 return;
2312
2313 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2314 wait_event_freezable_timeout(khugepaged_wait,
2315 khugepaged_should_wakeup(),
2316 scan_sleep_jiffies);
2317 return;
2318 }
2319
2320 if (khugepaged_enabled())
2321 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2322 }
2323
khugepaged(void * none)2324 static int khugepaged(void *none)
2325 {
2326 struct mm_slot *mm_slot;
2327
2328 set_freezable();
2329 set_user_nice(current, MAX_NICE);
2330
2331 while (!kthread_should_stop()) {
2332 khugepaged_do_scan();
2333 khugepaged_wait_work();
2334 }
2335
2336 spin_lock(&khugepaged_mm_lock);
2337 mm_slot = khugepaged_scan.mm_slot;
2338 khugepaged_scan.mm_slot = NULL;
2339 if (mm_slot)
2340 collect_mm_slot(mm_slot);
2341 spin_unlock(&khugepaged_mm_lock);
2342 return 0;
2343 }
2344
set_recommended_min_free_kbytes(void)2345 static void set_recommended_min_free_kbytes(void)
2346 {
2347 struct zone *zone;
2348 int nr_zones = 0;
2349 unsigned long recommended_min;
2350
2351 for_each_populated_zone(zone) {
2352 /*
2353 * We don't need to worry about fragmentation of
2354 * ZONE_MOVABLE since it only has movable pages.
2355 */
2356 if (zone_idx(zone) > gfp_zone(GFP_USER))
2357 continue;
2358
2359 nr_zones++;
2360 }
2361
2362 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2363 recommended_min = pageblock_nr_pages * nr_zones * 2;
2364
2365 /*
2366 * Make sure that on average at least two pageblocks are almost free
2367 * of another type, one for a migratetype to fall back to and a
2368 * second to avoid subsequent fallbacks of other types There are 3
2369 * MIGRATE_TYPES we care about.
2370 */
2371 recommended_min += pageblock_nr_pages * nr_zones *
2372 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2373
2374 /* don't ever allow to reserve more than 5% of the lowmem */
2375 recommended_min = min(recommended_min,
2376 (unsigned long) nr_free_buffer_pages() / 20);
2377 recommended_min <<= (PAGE_SHIFT-10);
2378
2379 if (recommended_min > min_free_kbytes) {
2380 if (user_min_free_kbytes >= 0)
2381 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2382 min_free_kbytes, recommended_min);
2383
2384 min_free_kbytes = recommended_min;
2385 }
2386 setup_per_zone_wmarks();
2387 }
2388
start_stop_khugepaged(void)2389 int start_stop_khugepaged(void)
2390 {
2391 int err = 0;
2392
2393 mutex_lock(&khugepaged_mutex);
2394 if (khugepaged_enabled()) {
2395 if (!khugepaged_thread)
2396 khugepaged_thread = kthread_run(khugepaged, NULL,
2397 "khugepaged");
2398 if (IS_ERR(khugepaged_thread)) {
2399 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2400 err = PTR_ERR(khugepaged_thread);
2401 khugepaged_thread = NULL;
2402 goto fail;
2403 }
2404
2405 if (!list_empty(&khugepaged_scan.mm_head))
2406 wake_up_interruptible(&khugepaged_wait);
2407
2408 set_recommended_min_free_kbytes();
2409 } else if (khugepaged_thread) {
2410 kthread_stop(khugepaged_thread);
2411 khugepaged_thread = NULL;
2412 }
2413 fail:
2414 mutex_unlock(&khugepaged_mutex);
2415 return err;
2416 }
2417
khugepaged_min_free_kbytes_update(void)2418 void khugepaged_min_free_kbytes_update(void)
2419 {
2420 mutex_lock(&khugepaged_mutex);
2421 if (khugepaged_enabled() && khugepaged_thread)
2422 set_recommended_min_free_kbytes();
2423 mutex_unlock(&khugepaged_mutex);
2424 }
2425