xref: /linux/mm/khugepaged.c (revision f3a8b6645dc2e60d11f20c1c23afd964ff4e55ae)
1 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
2 
3 #include <linux/mm.h>
4 #include <linux/sched.h>
5 #include <linux/mmu_notifier.h>
6 #include <linux/rmap.h>
7 #include <linux/swap.h>
8 #include <linux/mm_inline.h>
9 #include <linux/kthread.h>
10 #include <linux/khugepaged.h>
11 #include <linux/freezer.h>
12 #include <linux/mman.h>
13 #include <linux/hashtable.h>
14 #include <linux/userfaultfd_k.h>
15 #include <linux/page_idle.h>
16 #include <linux/swapops.h>
17 #include <linux/shmem_fs.h>
18 
19 #include <asm/tlb.h>
20 #include <asm/pgalloc.h>
21 #include "internal.h"
22 
23 enum scan_result {
24 	SCAN_FAIL,
25 	SCAN_SUCCEED,
26 	SCAN_PMD_NULL,
27 	SCAN_EXCEED_NONE_PTE,
28 	SCAN_PTE_NON_PRESENT,
29 	SCAN_PAGE_RO,
30 	SCAN_LACK_REFERENCED_PAGE,
31 	SCAN_PAGE_NULL,
32 	SCAN_SCAN_ABORT,
33 	SCAN_PAGE_COUNT,
34 	SCAN_PAGE_LRU,
35 	SCAN_PAGE_LOCK,
36 	SCAN_PAGE_ANON,
37 	SCAN_PAGE_COMPOUND,
38 	SCAN_ANY_PROCESS,
39 	SCAN_VMA_NULL,
40 	SCAN_VMA_CHECK,
41 	SCAN_ADDRESS_RANGE,
42 	SCAN_SWAP_CACHE_PAGE,
43 	SCAN_DEL_PAGE_LRU,
44 	SCAN_ALLOC_HUGE_PAGE_FAIL,
45 	SCAN_CGROUP_CHARGE_FAIL,
46 	SCAN_EXCEED_SWAP_PTE,
47 	SCAN_TRUNCATED,
48 };
49 
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/huge_memory.h>
52 
53 /* default scan 8*512 pte (or vmas) every 30 second */
54 static unsigned int khugepaged_pages_to_scan __read_mostly;
55 static unsigned int khugepaged_pages_collapsed;
56 static unsigned int khugepaged_full_scans;
57 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
58 /* during fragmentation poll the hugepage allocator once every minute */
59 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
60 static unsigned long khugepaged_sleep_expire;
61 static DEFINE_SPINLOCK(khugepaged_mm_lock);
62 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
63 /*
64  * default collapse hugepages if there is at least one pte mapped like
65  * it would have happened if the vma was large enough during page
66  * fault.
67  */
68 static unsigned int khugepaged_max_ptes_none __read_mostly;
69 static unsigned int khugepaged_max_ptes_swap __read_mostly;
70 
71 #define MM_SLOTS_HASH_BITS 10
72 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
73 
74 static struct kmem_cache *mm_slot_cache __read_mostly;
75 
76 /**
77  * struct mm_slot - hash lookup from mm to mm_slot
78  * @hash: hash collision list
79  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
80  * @mm: the mm that this information is valid for
81  */
82 struct mm_slot {
83 	struct hlist_node hash;
84 	struct list_head mm_node;
85 	struct mm_struct *mm;
86 };
87 
88 /**
89  * struct khugepaged_scan - cursor for scanning
90  * @mm_head: the head of the mm list to scan
91  * @mm_slot: the current mm_slot we are scanning
92  * @address: the next address inside that to be scanned
93  *
94  * There is only the one khugepaged_scan instance of this cursor structure.
95  */
96 struct khugepaged_scan {
97 	struct list_head mm_head;
98 	struct mm_slot *mm_slot;
99 	unsigned long address;
100 };
101 
102 static struct khugepaged_scan khugepaged_scan = {
103 	.mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
104 };
105 
106 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
107 					 struct kobj_attribute *attr,
108 					 char *buf)
109 {
110 	return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
111 }
112 
113 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
114 					  struct kobj_attribute *attr,
115 					  const char *buf, size_t count)
116 {
117 	unsigned long msecs;
118 	int err;
119 
120 	err = kstrtoul(buf, 10, &msecs);
121 	if (err || msecs > UINT_MAX)
122 		return -EINVAL;
123 
124 	khugepaged_scan_sleep_millisecs = msecs;
125 	khugepaged_sleep_expire = 0;
126 	wake_up_interruptible(&khugepaged_wait);
127 
128 	return count;
129 }
130 static struct kobj_attribute scan_sleep_millisecs_attr =
131 	__ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
132 	       scan_sleep_millisecs_store);
133 
134 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
135 					  struct kobj_attribute *attr,
136 					  char *buf)
137 {
138 	return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
139 }
140 
141 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
142 					   struct kobj_attribute *attr,
143 					   const char *buf, size_t count)
144 {
145 	unsigned long msecs;
146 	int err;
147 
148 	err = kstrtoul(buf, 10, &msecs);
149 	if (err || msecs > UINT_MAX)
150 		return -EINVAL;
151 
152 	khugepaged_alloc_sleep_millisecs = msecs;
153 	khugepaged_sleep_expire = 0;
154 	wake_up_interruptible(&khugepaged_wait);
155 
156 	return count;
157 }
158 static struct kobj_attribute alloc_sleep_millisecs_attr =
159 	__ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
160 	       alloc_sleep_millisecs_store);
161 
162 static ssize_t pages_to_scan_show(struct kobject *kobj,
163 				  struct kobj_attribute *attr,
164 				  char *buf)
165 {
166 	return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
167 }
168 static ssize_t pages_to_scan_store(struct kobject *kobj,
169 				   struct kobj_attribute *attr,
170 				   const char *buf, size_t count)
171 {
172 	int err;
173 	unsigned long pages;
174 
175 	err = kstrtoul(buf, 10, &pages);
176 	if (err || !pages || pages > UINT_MAX)
177 		return -EINVAL;
178 
179 	khugepaged_pages_to_scan = pages;
180 
181 	return count;
182 }
183 static struct kobj_attribute pages_to_scan_attr =
184 	__ATTR(pages_to_scan, 0644, pages_to_scan_show,
185 	       pages_to_scan_store);
186 
187 static ssize_t pages_collapsed_show(struct kobject *kobj,
188 				    struct kobj_attribute *attr,
189 				    char *buf)
190 {
191 	return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
192 }
193 static struct kobj_attribute pages_collapsed_attr =
194 	__ATTR_RO(pages_collapsed);
195 
196 static ssize_t full_scans_show(struct kobject *kobj,
197 			       struct kobj_attribute *attr,
198 			       char *buf)
199 {
200 	return sprintf(buf, "%u\n", khugepaged_full_scans);
201 }
202 static struct kobj_attribute full_scans_attr =
203 	__ATTR_RO(full_scans);
204 
205 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
206 				      struct kobj_attribute *attr, char *buf)
207 {
208 	return single_hugepage_flag_show(kobj, attr, buf,
209 				TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
210 }
211 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
212 				       struct kobj_attribute *attr,
213 				       const char *buf, size_t count)
214 {
215 	return single_hugepage_flag_store(kobj, attr, buf, count,
216 				 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
217 }
218 static struct kobj_attribute khugepaged_defrag_attr =
219 	__ATTR(defrag, 0644, khugepaged_defrag_show,
220 	       khugepaged_defrag_store);
221 
222 /*
223  * max_ptes_none controls if khugepaged should collapse hugepages over
224  * any unmapped ptes in turn potentially increasing the memory
225  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
226  * reduce the available free memory in the system as it
227  * runs. Increasing max_ptes_none will instead potentially reduce the
228  * free memory in the system during the khugepaged scan.
229  */
230 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
231 					     struct kobj_attribute *attr,
232 					     char *buf)
233 {
234 	return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
235 }
236 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
237 					      struct kobj_attribute *attr,
238 					      const char *buf, size_t count)
239 {
240 	int err;
241 	unsigned long max_ptes_none;
242 
243 	err = kstrtoul(buf, 10, &max_ptes_none);
244 	if (err || max_ptes_none > HPAGE_PMD_NR-1)
245 		return -EINVAL;
246 
247 	khugepaged_max_ptes_none = max_ptes_none;
248 
249 	return count;
250 }
251 static struct kobj_attribute khugepaged_max_ptes_none_attr =
252 	__ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
253 	       khugepaged_max_ptes_none_store);
254 
255 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
256 					     struct kobj_attribute *attr,
257 					     char *buf)
258 {
259 	return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
260 }
261 
262 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
263 					      struct kobj_attribute *attr,
264 					      const char *buf, size_t count)
265 {
266 	int err;
267 	unsigned long max_ptes_swap;
268 
269 	err  = kstrtoul(buf, 10, &max_ptes_swap);
270 	if (err || max_ptes_swap > HPAGE_PMD_NR-1)
271 		return -EINVAL;
272 
273 	khugepaged_max_ptes_swap = max_ptes_swap;
274 
275 	return count;
276 }
277 
278 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
279 	__ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
280 	       khugepaged_max_ptes_swap_store);
281 
282 static struct attribute *khugepaged_attr[] = {
283 	&khugepaged_defrag_attr.attr,
284 	&khugepaged_max_ptes_none_attr.attr,
285 	&pages_to_scan_attr.attr,
286 	&pages_collapsed_attr.attr,
287 	&full_scans_attr.attr,
288 	&scan_sleep_millisecs_attr.attr,
289 	&alloc_sleep_millisecs_attr.attr,
290 	&khugepaged_max_ptes_swap_attr.attr,
291 	NULL,
292 };
293 
294 struct attribute_group khugepaged_attr_group = {
295 	.attrs = khugepaged_attr,
296 	.name = "khugepaged",
297 };
298 
299 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
300 
301 int hugepage_madvise(struct vm_area_struct *vma,
302 		     unsigned long *vm_flags, int advice)
303 {
304 	switch (advice) {
305 	case MADV_HUGEPAGE:
306 #ifdef CONFIG_S390
307 		/*
308 		 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
309 		 * can't handle this properly after s390_enable_sie, so we simply
310 		 * ignore the madvise to prevent qemu from causing a SIGSEGV.
311 		 */
312 		if (mm_has_pgste(vma->vm_mm))
313 			return 0;
314 #endif
315 		*vm_flags &= ~VM_NOHUGEPAGE;
316 		*vm_flags |= VM_HUGEPAGE;
317 		/*
318 		 * If the vma become good for khugepaged to scan,
319 		 * register it here without waiting a page fault that
320 		 * may not happen any time soon.
321 		 */
322 		if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
323 				khugepaged_enter_vma_merge(vma, *vm_flags))
324 			return -ENOMEM;
325 		break;
326 	case MADV_NOHUGEPAGE:
327 		*vm_flags &= ~VM_HUGEPAGE;
328 		*vm_flags |= VM_NOHUGEPAGE;
329 		/*
330 		 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
331 		 * this vma even if we leave the mm registered in khugepaged if
332 		 * it got registered before VM_NOHUGEPAGE was set.
333 		 */
334 		break;
335 	}
336 
337 	return 0;
338 }
339 
340 int __init khugepaged_init(void)
341 {
342 	mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
343 					  sizeof(struct mm_slot),
344 					  __alignof__(struct mm_slot), 0, NULL);
345 	if (!mm_slot_cache)
346 		return -ENOMEM;
347 
348 	khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
349 	khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
350 	khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
351 
352 	return 0;
353 }
354 
355 void __init khugepaged_destroy(void)
356 {
357 	kmem_cache_destroy(mm_slot_cache);
358 }
359 
360 static inline struct mm_slot *alloc_mm_slot(void)
361 {
362 	if (!mm_slot_cache)	/* initialization failed */
363 		return NULL;
364 	return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
365 }
366 
367 static inline void free_mm_slot(struct mm_slot *mm_slot)
368 {
369 	kmem_cache_free(mm_slot_cache, mm_slot);
370 }
371 
372 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
373 {
374 	struct mm_slot *mm_slot;
375 
376 	hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
377 		if (mm == mm_slot->mm)
378 			return mm_slot;
379 
380 	return NULL;
381 }
382 
383 static void insert_to_mm_slots_hash(struct mm_struct *mm,
384 				    struct mm_slot *mm_slot)
385 {
386 	mm_slot->mm = mm;
387 	hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
388 }
389 
390 static inline int khugepaged_test_exit(struct mm_struct *mm)
391 {
392 	return atomic_read(&mm->mm_users) == 0;
393 }
394 
395 int __khugepaged_enter(struct mm_struct *mm)
396 {
397 	struct mm_slot *mm_slot;
398 	int wakeup;
399 
400 	mm_slot = alloc_mm_slot();
401 	if (!mm_slot)
402 		return -ENOMEM;
403 
404 	/* __khugepaged_exit() must not run from under us */
405 	VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
406 	if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
407 		free_mm_slot(mm_slot);
408 		return 0;
409 	}
410 
411 	spin_lock(&khugepaged_mm_lock);
412 	insert_to_mm_slots_hash(mm, mm_slot);
413 	/*
414 	 * Insert just behind the scanning cursor, to let the area settle
415 	 * down a little.
416 	 */
417 	wakeup = list_empty(&khugepaged_scan.mm_head);
418 	list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
419 	spin_unlock(&khugepaged_mm_lock);
420 
421 	atomic_inc(&mm->mm_count);
422 	if (wakeup)
423 		wake_up_interruptible(&khugepaged_wait);
424 
425 	return 0;
426 }
427 
428 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
429 			       unsigned long vm_flags)
430 {
431 	unsigned long hstart, hend;
432 	if (!vma->anon_vma)
433 		/*
434 		 * Not yet faulted in so we will register later in the
435 		 * page fault if needed.
436 		 */
437 		return 0;
438 	if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
439 		/* khugepaged not yet working on file or special mappings */
440 		return 0;
441 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
442 	hend = vma->vm_end & HPAGE_PMD_MASK;
443 	if (hstart < hend)
444 		return khugepaged_enter(vma, vm_flags);
445 	return 0;
446 }
447 
448 void __khugepaged_exit(struct mm_struct *mm)
449 {
450 	struct mm_slot *mm_slot;
451 	int free = 0;
452 
453 	spin_lock(&khugepaged_mm_lock);
454 	mm_slot = get_mm_slot(mm);
455 	if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
456 		hash_del(&mm_slot->hash);
457 		list_del(&mm_slot->mm_node);
458 		free = 1;
459 	}
460 	spin_unlock(&khugepaged_mm_lock);
461 
462 	if (free) {
463 		clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
464 		free_mm_slot(mm_slot);
465 		mmdrop(mm);
466 	} else if (mm_slot) {
467 		/*
468 		 * This is required to serialize against
469 		 * khugepaged_test_exit() (which is guaranteed to run
470 		 * under mmap sem read mode). Stop here (after we
471 		 * return all pagetables will be destroyed) until
472 		 * khugepaged has finished working on the pagetables
473 		 * under the mmap_sem.
474 		 */
475 		down_write(&mm->mmap_sem);
476 		up_write(&mm->mmap_sem);
477 	}
478 }
479 
480 static void release_pte_page(struct page *page)
481 {
482 	/* 0 stands for page_is_file_cache(page) == false */
483 	dec_node_page_state(page, NR_ISOLATED_ANON + 0);
484 	unlock_page(page);
485 	putback_lru_page(page);
486 }
487 
488 static void release_pte_pages(pte_t *pte, pte_t *_pte)
489 {
490 	while (--_pte >= pte) {
491 		pte_t pteval = *_pte;
492 		if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
493 			release_pte_page(pte_page(pteval));
494 	}
495 }
496 
497 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
498 					unsigned long address,
499 					pte_t *pte)
500 {
501 	struct page *page = NULL;
502 	pte_t *_pte;
503 	int none_or_zero = 0, result = 0, referenced = 0;
504 	bool writable = false;
505 
506 	for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
507 	     _pte++, address += PAGE_SIZE) {
508 		pte_t pteval = *_pte;
509 		if (pte_none(pteval) || (pte_present(pteval) &&
510 				is_zero_pfn(pte_pfn(pteval)))) {
511 			if (!userfaultfd_armed(vma) &&
512 			    ++none_or_zero <= khugepaged_max_ptes_none) {
513 				continue;
514 			} else {
515 				result = SCAN_EXCEED_NONE_PTE;
516 				goto out;
517 			}
518 		}
519 		if (!pte_present(pteval)) {
520 			result = SCAN_PTE_NON_PRESENT;
521 			goto out;
522 		}
523 		page = vm_normal_page(vma, address, pteval);
524 		if (unlikely(!page)) {
525 			result = SCAN_PAGE_NULL;
526 			goto out;
527 		}
528 
529 		VM_BUG_ON_PAGE(PageCompound(page), page);
530 		VM_BUG_ON_PAGE(!PageAnon(page), page);
531 		VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
532 
533 		/*
534 		 * We can do it before isolate_lru_page because the
535 		 * page can't be freed from under us. NOTE: PG_lock
536 		 * is needed to serialize against split_huge_page
537 		 * when invoked from the VM.
538 		 */
539 		if (!trylock_page(page)) {
540 			result = SCAN_PAGE_LOCK;
541 			goto out;
542 		}
543 
544 		/*
545 		 * cannot use mapcount: can't collapse if there's a gup pin.
546 		 * The page must only be referenced by the scanned process
547 		 * and page swap cache.
548 		 */
549 		if (page_count(page) != 1 + !!PageSwapCache(page)) {
550 			unlock_page(page);
551 			result = SCAN_PAGE_COUNT;
552 			goto out;
553 		}
554 		if (pte_write(pteval)) {
555 			writable = true;
556 		} else {
557 			if (PageSwapCache(page) &&
558 			    !reuse_swap_page(page, NULL)) {
559 				unlock_page(page);
560 				result = SCAN_SWAP_CACHE_PAGE;
561 				goto out;
562 			}
563 			/*
564 			 * Page is not in the swap cache. It can be collapsed
565 			 * into a THP.
566 			 */
567 		}
568 
569 		/*
570 		 * Isolate the page to avoid collapsing an hugepage
571 		 * currently in use by the VM.
572 		 */
573 		if (isolate_lru_page(page)) {
574 			unlock_page(page);
575 			result = SCAN_DEL_PAGE_LRU;
576 			goto out;
577 		}
578 		/* 0 stands for page_is_file_cache(page) == false */
579 		inc_node_page_state(page, NR_ISOLATED_ANON + 0);
580 		VM_BUG_ON_PAGE(!PageLocked(page), page);
581 		VM_BUG_ON_PAGE(PageLRU(page), page);
582 
583 		/* There should be enough young pte to collapse the page */
584 		if (pte_young(pteval) ||
585 		    page_is_young(page) || PageReferenced(page) ||
586 		    mmu_notifier_test_young(vma->vm_mm, address))
587 			referenced++;
588 	}
589 	if (likely(writable)) {
590 		if (likely(referenced)) {
591 			result = SCAN_SUCCEED;
592 			trace_mm_collapse_huge_page_isolate(page, none_or_zero,
593 							    referenced, writable, result);
594 			return 1;
595 		}
596 	} else {
597 		result = SCAN_PAGE_RO;
598 	}
599 
600 out:
601 	release_pte_pages(pte, _pte);
602 	trace_mm_collapse_huge_page_isolate(page, none_or_zero,
603 					    referenced, writable, result);
604 	return 0;
605 }
606 
607 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
608 				      struct vm_area_struct *vma,
609 				      unsigned long address,
610 				      spinlock_t *ptl)
611 {
612 	pte_t *_pte;
613 	for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
614 		pte_t pteval = *_pte;
615 		struct page *src_page;
616 
617 		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
618 			clear_user_highpage(page, address);
619 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
620 			if (is_zero_pfn(pte_pfn(pteval))) {
621 				/*
622 				 * ptl mostly unnecessary.
623 				 */
624 				spin_lock(ptl);
625 				/*
626 				 * paravirt calls inside pte_clear here are
627 				 * superfluous.
628 				 */
629 				pte_clear(vma->vm_mm, address, _pte);
630 				spin_unlock(ptl);
631 			}
632 		} else {
633 			src_page = pte_page(pteval);
634 			copy_user_highpage(page, src_page, address, vma);
635 			VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
636 			release_pte_page(src_page);
637 			/*
638 			 * ptl mostly unnecessary, but preempt has to
639 			 * be disabled to update the per-cpu stats
640 			 * inside page_remove_rmap().
641 			 */
642 			spin_lock(ptl);
643 			/*
644 			 * paravirt calls inside pte_clear here are
645 			 * superfluous.
646 			 */
647 			pte_clear(vma->vm_mm, address, _pte);
648 			page_remove_rmap(src_page, false);
649 			spin_unlock(ptl);
650 			free_page_and_swap_cache(src_page);
651 		}
652 
653 		address += PAGE_SIZE;
654 		page++;
655 	}
656 }
657 
658 static void khugepaged_alloc_sleep(void)
659 {
660 	DEFINE_WAIT(wait);
661 
662 	add_wait_queue(&khugepaged_wait, &wait);
663 	freezable_schedule_timeout_interruptible(
664 		msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
665 	remove_wait_queue(&khugepaged_wait, &wait);
666 }
667 
668 static int khugepaged_node_load[MAX_NUMNODES];
669 
670 static bool khugepaged_scan_abort(int nid)
671 {
672 	int i;
673 
674 	/*
675 	 * If node_reclaim_mode is disabled, then no extra effort is made to
676 	 * allocate memory locally.
677 	 */
678 	if (!node_reclaim_mode)
679 		return false;
680 
681 	/* If there is a count for this node already, it must be acceptable */
682 	if (khugepaged_node_load[nid])
683 		return false;
684 
685 	for (i = 0; i < MAX_NUMNODES; i++) {
686 		if (!khugepaged_node_load[i])
687 			continue;
688 		if (node_distance(nid, i) > RECLAIM_DISTANCE)
689 			return true;
690 	}
691 	return false;
692 }
693 
694 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
695 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
696 {
697 	return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
698 }
699 
700 #ifdef CONFIG_NUMA
701 static int khugepaged_find_target_node(void)
702 {
703 	static int last_khugepaged_target_node = NUMA_NO_NODE;
704 	int nid, target_node = 0, max_value = 0;
705 
706 	/* find first node with max normal pages hit */
707 	for (nid = 0; nid < MAX_NUMNODES; nid++)
708 		if (khugepaged_node_load[nid] > max_value) {
709 			max_value = khugepaged_node_load[nid];
710 			target_node = nid;
711 		}
712 
713 	/* do some balance if several nodes have the same hit record */
714 	if (target_node <= last_khugepaged_target_node)
715 		for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
716 				nid++)
717 			if (max_value == khugepaged_node_load[nid]) {
718 				target_node = nid;
719 				break;
720 			}
721 
722 	last_khugepaged_target_node = target_node;
723 	return target_node;
724 }
725 
726 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
727 {
728 	if (IS_ERR(*hpage)) {
729 		if (!*wait)
730 			return false;
731 
732 		*wait = false;
733 		*hpage = NULL;
734 		khugepaged_alloc_sleep();
735 	} else if (*hpage) {
736 		put_page(*hpage);
737 		*hpage = NULL;
738 	}
739 
740 	return true;
741 }
742 
743 static struct page *
744 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
745 {
746 	VM_BUG_ON_PAGE(*hpage, *hpage);
747 
748 	*hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
749 	if (unlikely(!*hpage)) {
750 		count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
751 		*hpage = ERR_PTR(-ENOMEM);
752 		return NULL;
753 	}
754 
755 	prep_transhuge_page(*hpage);
756 	count_vm_event(THP_COLLAPSE_ALLOC);
757 	return *hpage;
758 }
759 #else
760 static int khugepaged_find_target_node(void)
761 {
762 	return 0;
763 }
764 
765 static inline struct page *alloc_khugepaged_hugepage(void)
766 {
767 	struct page *page;
768 
769 	page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
770 			   HPAGE_PMD_ORDER);
771 	if (page)
772 		prep_transhuge_page(page);
773 	return page;
774 }
775 
776 static struct page *khugepaged_alloc_hugepage(bool *wait)
777 {
778 	struct page *hpage;
779 
780 	do {
781 		hpage = alloc_khugepaged_hugepage();
782 		if (!hpage) {
783 			count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
784 			if (!*wait)
785 				return NULL;
786 
787 			*wait = false;
788 			khugepaged_alloc_sleep();
789 		} else
790 			count_vm_event(THP_COLLAPSE_ALLOC);
791 	} while (unlikely(!hpage) && likely(khugepaged_enabled()));
792 
793 	return hpage;
794 }
795 
796 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
797 {
798 	if (!*hpage)
799 		*hpage = khugepaged_alloc_hugepage(wait);
800 
801 	if (unlikely(!*hpage))
802 		return false;
803 
804 	return true;
805 }
806 
807 static struct page *
808 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
809 {
810 	VM_BUG_ON(!*hpage);
811 
812 	return  *hpage;
813 }
814 #endif
815 
816 static bool hugepage_vma_check(struct vm_area_struct *vma)
817 {
818 	if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
819 	    (vma->vm_flags & VM_NOHUGEPAGE))
820 		return false;
821 	if (shmem_file(vma->vm_file)) {
822 		if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
823 			return false;
824 		return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
825 				HPAGE_PMD_NR);
826 	}
827 	if (!vma->anon_vma || vma->vm_ops)
828 		return false;
829 	if (is_vma_temporary_stack(vma))
830 		return false;
831 	return !(vma->vm_flags & VM_NO_KHUGEPAGED);
832 }
833 
834 /*
835  * If mmap_sem temporarily dropped, revalidate vma
836  * before taking mmap_sem.
837  * Return 0 if succeeds, otherwise return none-zero
838  * value (scan code).
839  */
840 
841 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
842 		struct vm_area_struct **vmap)
843 {
844 	struct vm_area_struct *vma;
845 	unsigned long hstart, hend;
846 
847 	if (unlikely(khugepaged_test_exit(mm)))
848 		return SCAN_ANY_PROCESS;
849 
850 	*vmap = vma = find_vma(mm, address);
851 	if (!vma)
852 		return SCAN_VMA_NULL;
853 
854 	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
855 	hend = vma->vm_end & HPAGE_PMD_MASK;
856 	if (address < hstart || address + HPAGE_PMD_SIZE > hend)
857 		return SCAN_ADDRESS_RANGE;
858 	if (!hugepage_vma_check(vma))
859 		return SCAN_VMA_CHECK;
860 	return 0;
861 }
862 
863 /*
864  * Bring missing pages in from swap, to complete THP collapse.
865  * Only done if khugepaged_scan_pmd believes it is worthwhile.
866  *
867  * Called and returns without pte mapped or spinlocks held,
868  * but with mmap_sem held to protect against vma changes.
869  */
870 
871 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
872 					struct vm_area_struct *vma,
873 					unsigned long address, pmd_t *pmd,
874 					int referenced)
875 {
876 	pte_t pteval;
877 	int swapped_in = 0, ret = 0;
878 	struct fault_env fe = {
879 		.vma = vma,
880 		.address = address,
881 		.flags = FAULT_FLAG_ALLOW_RETRY,
882 		.pmd = pmd,
883 	};
884 
885 	/* we only decide to swapin, if there is enough young ptes */
886 	if (referenced < HPAGE_PMD_NR/2) {
887 		trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
888 		return false;
889 	}
890 	fe.pte = pte_offset_map(pmd, address);
891 	for (; fe.address < address + HPAGE_PMD_NR*PAGE_SIZE;
892 			fe.pte++, fe.address += PAGE_SIZE) {
893 		pteval = *fe.pte;
894 		if (!is_swap_pte(pteval))
895 			continue;
896 		swapped_in++;
897 		ret = do_swap_page(&fe, pteval);
898 
899 		/* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
900 		if (ret & VM_FAULT_RETRY) {
901 			down_read(&mm->mmap_sem);
902 			if (hugepage_vma_revalidate(mm, address, &fe.vma)) {
903 				/* vma is no longer available, don't continue to swapin */
904 				trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
905 				return false;
906 			}
907 			/* check if the pmd is still valid */
908 			if (mm_find_pmd(mm, address) != pmd)
909 				return false;
910 		}
911 		if (ret & VM_FAULT_ERROR) {
912 			trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
913 			return false;
914 		}
915 		/* pte is unmapped now, we need to map it */
916 		fe.pte = pte_offset_map(pmd, fe.address);
917 	}
918 	fe.pte--;
919 	pte_unmap(fe.pte);
920 	trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
921 	return true;
922 }
923 
924 static void collapse_huge_page(struct mm_struct *mm,
925 				   unsigned long address,
926 				   struct page **hpage,
927 				   int node, int referenced)
928 {
929 	pmd_t *pmd, _pmd;
930 	pte_t *pte;
931 	pgtable_t pgtable;
932 	struct page *new_page;
933 	spinlock_t *pmd_ptl, *pte_ptl;
934 	int isolated = 0, result = 0;
935 	struct mem_cgroup *memcg;
936 	struct vm_area_struct *vma;
937 	unsigned long mmun_start;	/* For mmu_notifiers */
938 	unsigned long mmun_end;		/* For mmu_notifiers */
939 	gfp_t gfp;
940 
941 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
942 
943 	/* Only allocate from the target node */
944 	gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
945 
946 	/*
947 	 * Before allocating the hugepage, release the mmap_sem read lock.
948 	 * The allocation can take potentially a long time if it involves
949 	 * sync compaction, and we do not need to hold the mmap_sem during
950 	 * that. We will recheck the vma after taking it again in write mode.
951 	 */
952 	up_read(&mm->mmap_sem);
953 	new_page = khugepaged_alloc_page(hpage, gfp, node);
954 	if (!new_page) {
955 		result = SCAN_ALLOC_HUGE_PAGE_FAIL;
956 		goto out_nolock;
957 	}
958 
959 	if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
960 		result = SCAN_CGROUP_CHARGE_FAIL;
961 		goto out_nolock;
962 	}
963 
964 	down_read(&mm->mmap_sem);
965 	result = hugepage_vma_revalidate(mm, address, &vma);
966 	if (result) {
967 		mem_cgroup_cancel_charge(new_page, memcg, true);
968 		up_read(&mm->mmap_sem);
969 		goto out_nolock;
970 	}
971 
972 	pmd = mm_find_pmd(mm, address);
973 	if (!pmd) {
974 		result = SCAN_PMD_NULL;
975 		mem_cgroup_cancel_charge(new_page, memcg, true);
976 		up_read(&mm->mmap_sem);
977 		goto out_nolock;
978 	}
979 
980 	/*
981 	 * __collapse_huge_page_swapin always returns with mmap_sem locked.
982 	 * If it fails, we release mmap_sem and jump out_nolock.
983 	 * Continuing to collapse causes inconsistency.
984 	 */
985 	if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
986 		mem_cgroup_cancel_charge(new_page, memcg, true);
987 		up_read(&mm->mmap_sem);
988 		goto out_nolock;
989 	}
990 
991 	up_read(&mm->mmap_sem);
992 	/*
993 	 * Prevent all access to pagetables with the exception of
994 	 * gup_fast later handled by the ptep_clear_flush and the VM
995 	 * handled by the anon_vma lock + PG_lock.
996 	 */
997 	down_write(&mm->mmap_sem);
998 	result = hugepage_vma_revalidate(mm, address, &vma);
999 	if (result)
1000 		goto out;
1001 	/* check if the pmd is still valid */
1002 	if (mm_find_pmd(mm, address) != pmd)
1003 		goto out;
1004 
1005 	anon_vma_lock_write(vma->anon_vma);
1006 
1007 	pte = pte_offset_map(pmd, address);
1008 	pte_ptl = pte_lockptr(mm, pmd);
1009 
1010 	mmun_start = address;
1011 	mmun_end   = address + HPAGE_PMD_SIZE;
1012 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1013 	pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1014 	/*
1015 	 * After this gup_fast can't run anymore. This also removes
1016 	 * any huge TLB entry from the CPU so we won't allow
1017 	 * huge and small TLB entries for the same virtual address
1018 	 * to avoid the risk of CPU bugs in that area.
1019 	 */
1020 	_pmd = pmdp_collapse_flush(vma, address, pmd);
1021 	spin_unlock(pmd_ptl);
1022 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1023 
1024 	spin_lock(pte_ptl);
1025 	isolated = __collapse_huge_page_isolate(vma, address, pte);
1026 	spin_unlock(pte_ptl);
1027 
1028 	if (unlikely(!isolated)) {
1029 		pte_unmap(pte);
1030 		spin_lock(pmd_ptl);
1031 		BUG_ON(!pmd_none(*pmd));
1032 		/*
1033 		 * We can only use set_pmd_at when establishing
1034 		 * hugepmds and never for establishing regular pmds that
1035 		 * points to regular pagetables. Use pmd_populate for that
1036 		 */
1037 		pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1038 		spin_unlock(pmd_ptl);
1039 		anon_vma_unlock_write(vma->anon_vma);
1040 		result = SCAN_FAIL;
1041 		goto out;
1042 	}
1043 
1044 	/*
1045 	 * All pages are isolated and locked so anon_vma rmap
1046 	 * can't run anymore.
1047 	 */
1048 	anon_vma_unlock_write(vma->anon_vma);
1049 
1050 	__collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1051 	pte_unmap(pte);
1052 	__SetPageUptodate(new_page);
1053 	pgtable = pmd_pgtable(_pmd);
1054 
1055 	_pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1056 	_pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1057 
1058 	/*
1059 	 * spin_lock() below is not the equivalent of smp_wmb(), so
1060 	 * this is needed to avoid the copy_huge_page writes to become
1061 	 * visible after the set_pmd_at() write.
1062 	 */
1063 	smp_wmb();
1064 
1065 	spin_lock(pmd_ptl);
1066 	BUG_ON(!pmd_none(*pmd));
1067 	page_add_new_anon_rmap(new_page, vma, address, true);
1068 	mem_cgroup_commit_charge(new_page, memcg, false, true);
1069 	lru_cache_add_active_or_unevictable(new_page, vma);
1070 	pgtable_trans_huge_deposit(mm, pmd, pgtable);
1071 	set_pmd_at(mm, address, pmd, _pmd);
1072 	update_mmu_cache_pmd(vma, address, pmd);
1073 	spin_unlock(pmd_ptl);
1074 
1075 	*hpage = NULL;
1076 
1077 	khugepaged_pages_collapsed++;
1078 	result = SCAN_SUCCEED;
1079 out_up_write:
1080 	up_write(&mm->mmap_sem);
1081 out_nolock:
1082 	trace_mm_collapse_huge_page(mm, isolated, result);
1083 	return;
1084 out:
1085 	mem_cgroup_cancel_charge(new_page, memcg, true);
1086 	goto out_up_write;
1087 }
1088 
1089 static int khugepaged_scan_pmd(struct mm_struct *mm,
1090 			       struct vm_area_struct *vma,
1091 			       unsigned long address,
1092 			       struct page **hpage)
1093 {
1094 	pmd_t *pmd;
1095 	pte_t *pte, *_pte;
1096 	int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1097 	struct page *page = NULL;
1098 	unsigned long _address;
1099 	spinlock_t *ptl;
1100 	int node = NUMA_NO_NODE, unmapped = 0;
1101 	bool writable = false;
1102 
1103 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1104 
1105 	pmd = mm_find_pmd(mm, address);
1106 	if (!pmd) {
1107 		result = SCAN_PMD_NULL;
1108 		goto out;
1109 	}
1110 
1111 	memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1112 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1113 	for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1114 	     _pte++, _address += PAGE_SIZE) {
1115 		pte_t pteval = *_pte;
1116 		if (is_swap_pte(pteval)) {
1117 			if (++unmapped <= khugepaged_max_ptes_swap) {
1118 				continue;
1119 			} else {
1120 				result = SCAN_EXCEED_SWAP_PTE;
1121 				goto out_unmap;
1122 			}
1123 		}
1124 		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1125 			if (!userfaultfd_armed(vma) &&
1126 			    ++none_or_zero <= khugepaged_max_ptes_none) {
1127 				continue;
1128 			} else {
1129 				result = SCAN_EXCEED_NONE_PTE;
1130 				goto out_unmap;
1131 			}
1132 		}
1133 		if (!pte_present(pteval)) {
1134 			result = SCAN_PTE_NON_PRESENT;
1135 			goto out_unmap;
1136 		}
1137 		if (pte_write(pteval))
1138 			writable = true;
1139 
1140 		page = vm_normal_page(vma, _address, pteval);
1141 		if (unlikely(!page)) {
1142 			result = SCAN_PAGE_NULL;
1143 			goto out_unmap;
1144 		}
1145 
1146 		/* TODO: teach khugepaged to collapse THP mapped with pte */
1147 		if (PageCompound(page)) {
1148 			result = SCAN_PAGE_COMPOUND;
1149 			goto out_unmap;
1150 		}
1151 
1152 		/*
1153 		 * Record which node the original page is from and save this
1154 		 * information to khugepaged_node_load[].
1155 		 * Khupaged will allocate hugepage from the node has the max
1156 		 * hit record.
1157 		 */
1158 		node = page_to_nid(page);
1159 		if (khugepaged_scan_abort(node)) {
1160 			result = SCAN_SCAN_ABORT;
1161 			goto out_unmap;
1162 		}
1163 		khugepaged_node_load[node]++;
1164 		if (!PageLRU(page)) {
1165 			result = SCAN_PAGE_LRU;
1166 			goto out_unmap;
1167 		}
1168 		if (PageLocked(page)) {
1169 			result = SCAN_PAGE_LOCK;
1170 			goto out_unmap;
1171 		}
1172 		if (!PageAnon(page)) {
1173 			result = SCAN_PAGE_ANON;
1174 			goto out_unmap;
1175 		}
1176 
1177 		/*
1178 		 * cannot use mapcount: can't collapse if there's a gup pin.
1179 		 * The page must only be referenced by the scanned process
1180 		 * and page swap cache.
1181 		 */
1182 		if (page_count(page) != 1 + !!PageSwapCache(page)) {
1183 			result = SCAN_PAGE_COUNT;
1184 			goto out_unmap;
1185 		}
1186 		if (pte_young(pteval) ||
1187 		    page_is_young(page) || PageReferenced(page) ||
1188 		    mmu_notifier_test_young(vma->vm_mm, address))
1189 			referenced++;
1190 	}
1191 	if (writable) {
1192 		if (referenced) {
1193 			result = SCAN_SUCCEED;
1194 			ret = 1;
1195 		} else {
1196 			result = SCAN_LACK_REFERENCED_PAGE;
1197 		}
1198 	} else {
1199 		result = SCAN_PAGE_RO;
1200 	}
1201 out_unmap:
1202 	pte_unmap_unlock(pte, ptl);
1203 	if (ret) {
1204 		node = khugepaged_find_target_node();
1205 		/* collapse_huge_page will return with the mmap_sem released */
1206 		collapse_huge_page(mm, address, hpage, node, referenced);
1207 	}
1208 out:
1209 	trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1210 				     none_or_zero, result, unmapped);
1211 	return ret;
1212 }
1213 
1214 static void collect_mm_slot(struct mm_slot *mm_slot)
1215 {
1216 	struct mm_struct *mm = mm_slot->mm;
1217 
1218 	VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1219 
1220 	if (khugepaged_test_exit(mm)) {
1221 		/* free mm_slot */
1222 		hash_del(&mm_slot->hash);
1223 		list_del(&mm_slot->mm_node);
1224 
1225 		/*
1226 		 * Not strictly needed because the mm exited already.
1227 		 *
1228 		 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1229 		 */
1230 
1231 		/* khugepaged_mm_lock actually not necessary for the below */
1232 		free_mm_slot(mm_slot);
1233 		mmdrop(mm);
1234 	}
1235 }
1236 
1237 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1238 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1239 {
1240 	struct vm_area_struct *vma;
1241 	unsigned long addr;
1242 	pmd_t *pmd, _pmd;
1243 
1244 	i_mmap_lock_write(mapping);
1245 	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1246 		/* probably overkill */
1247 		if (vma->anon_vma)
1248 			continue;
1249 		addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1250 		if (addr & ~HPAGE_PMD_MASK)
1251 			continue;
1252 		if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1253 			continue;
1254 		pmd = mm_find_pmd(vma->vm_mm, addr);
1255 		if (!pmd)
1256 			continue;
1257 		/*
1258 		 * We need exclusive mmap_sem to retract page table.
1259 		 * If trylock fails we would end up with pte-mapped THP after
1260 		 * re-fault. Not ideal, but it's more important to not disturb
1261 		 * the system too much.
1262 		 */
1263 		if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1264 			spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1265 			/* assume page table is clear */
1266 			_pmd = pmdp_collapse_flush(vma, addr, pmd);
1267 			spin_unlock(ptl);
1268 			up_write(&vma->vm_mm->mmap_sem);
1269 			atomic_long_dec(&vma->vm_mm->nr_ptes);
1270 			pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1271 		}
1272 	}
1273 	i_mmap_unlock_write(mapping);
1274 }
1275 
1276 /**
1277  * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1278  *
1279  * Basic scheme is simple, details are more complex:
1280  *  - allocate and freeze a new huge page;
1281  *  - scan over radix tree replacing old pages the new one
1282  *    + swap in pages if necessary;
1283  *    + fill in gaps;
1284  *    + keep old pages around in case if rollback is required;
1285  *  - if replacing succeed:
1286  *    + copy data over;
1287  *    + free old pages;
1288  *    + unfreeze huge page;
1289  *  - if replacing failed;
1290  *    + put all pages back and unfreeze them;
1291  *    + restore gaps in the radix-tree;
1292  *    + free huge page;
1293  */
1294 static void collapse_shmem(struct mm_struct *mm,
1295 		struct address_space *mapping, pgoff_t start,
1296 		struct page **hpage, int node)
1297 {
1298 	gfp_t gfp;
1299 	struct page *page, *new_page, *tmp;
1300 	struct mem_cgroup *memcg;
1301 	pgoff_t index, end = start + HPAGE_PMD_NR;
1302 	LIST_HEAD(pagelist);
1303 	struct radix_tree_iter iter;
1304 	void **slot;
1305 	int nr_none = 0, result = SCAN_SUCCEED;
1306 
1307 	VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1308 
1309 	/* Only allocate from the target node */
1310 	gfp = alloc_hugepage_khugepaged_gfpmask() |
1311 		__GFP_OTHER_NODE | __GFP_THISNODE;
1312 
1313 	new_page = khugepaged_alloc_page(hpage, gfp, node);
1314 	if (!new_page) {
1315 		result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1316 		goto out;
1317 	}
1318 
1319 	if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1320 		result = SCAN_CGROUP_CHARGE_FAIL;
1321 		goto out;
1322 	}
1323 
1324 	new_page->index = start;
1325 	new_page->mapping = mapping;
1326 	__SetPageSwapBacked(new_page);
1327 	__SetPageLocked(new_page);
1328 	BUG_ON(!page_ref_freeze(new_page, 1));
1329 
1330 
1331 	/*
1332 	 * At this point the new_page is 'frozen' (page_count() is zero), locked
1333 	 * and not up-to-date. It's safe to insert it into radix tree, because
1334 	 * nobody would be able to map it or use it in other way until we
1335 	 * unfreeze it.
1336 	 */
1337 
1338 	index = start;
1339 	spin_lock_irq(&mapping->tree_lock);
1340 	radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1341 		int n = min(iter.index, end) - index;
1342 
1343 		/*
1344 		 * Handle holes in the radix tree: charge it from shmem and
1345 		 * insert relevant subpage of new_page into the radix-tree.
1346 		 */
1347 		if (n && !shmem_charge(mapping->host, n)) {
1348 			result = SCAN_FAIL;
1349 			break;
1350 		}
1351 		nr_none += n;
1352 		for (; index < min(iter.index, end); index++) {
1353 			radix_tree_insert(&mapping->page_tree, index,
1354 					new_page + (index % HPAGE_PMD_NR));
1355 		}
1356 
1357 		/* We are done. */
1358 		if (index >= end)
1359 			break;
1360 
1361 		page = radix_tree_deref_slot_protected(slot,
1362 				&mapping->tree_lock);
1363 		if (radix_tree_exceptional_entry(page) || !PageUptodate(page)) {
1364 			spin_unlock_irq(&mapping->tree_lock);
1365 			/* swap in or instantiate fallocated page */
1366 			if (shmem_getpage(mapping->host, index, &page,
1367 						SGP_NOHUGE)) {
1368 				result = SCAN_FAIL;
1369 				goto tree_unlocked;
1370 			}
1371 			spin_lock_irq(&mapping->tree_lock);
1372 		} else if (trylock_page(page)) {
1373 			get_page(page);
1374 		} else {
1375 			result = SCAN_PAGE_LOCK;
1376 			break;
1377 		}
1378 
1379 		/*
1380 		 * The page must be locked, so we can drop the tree_lock
1381 		 * without racing with truncate.
1382 		 */
1383 		VM_BUG_ON_PAGE(!PageLocked(page), page);
1384 		VM_BUG_ON_PAGE(!PageUptodate(page), page);
1385 		VM_BUG_ON_PAGE(PageTransCompound(page), page);
1386 
1387 		if (page_mapping(page) != mapping) {
1388 			result = SCAN_TRUNCATED;
1389 			goto out_unlock;
1390 		}
1391 		spin_unlock_irq(&mapping->tree_lock);
1392 
1393 		if (isolate_lru_page(page)) {
1394 			result = SCAN_DEL_PAGE_LRU;
1395 			goto out_isolate_failed;
1396 		}
1397 
1398 		if (page_mapped(page))
1399 			unmap_mapping_range(mapping, index << PAGE_SHIFT,
1400 					PAGE_SIZE, 0);
1401 
1402 		spin_lock_irq(&mapping->tree_lock);
1403 
1404 		VM_BUG_ON_PAGE(page_mapped(page), page);
1405 
1406 		/*
1407 		 * The page is expected to have page_count() == 3:
1408 		 *  - we hold a pin on it;
1409 		 *  - one reference from radix tree;
1410 		 *  - one from isolate_lru_page;
1411 		 */
1412 		if (!page_ref_freeze(page, 3)) {
1413 			result = SCAN_PAGE_COUNT;
1414 			goto out_lru;
1415 		}
1416 
1417 		/*
1418 		 * Add the page to the list to be able to undo the collapse if
1419 		 * something go wrong.
1420 		 */
1421 		list_add_tail(&page->lru, &pagelist);
1422 
1423 		/* Finally, replace with the new page. */
1424 		radix_tree_replace_slot(slot,
1425 				new_page + (index % HPAGE_PMD_NR));
1426 
1427 		index++;
1428 		continue;
1429 out_lru:
1430 		spin_unlock_irq(&mapping->tree_lock);
1431 		putback_lru_page(page);
1432 out_isolate_failed:
1433 		unlock_page(page);
1434 		put_page(page);
1435 		goto tree_unlocked;
1436 out_unlock:
1437 		unlock_page(page);
1438 		put_page(page);
1439 		break;
1440 	}
1441 
1442 	/*
1443 	 * Handle hole in radix tree at the end of the range.
1444 	 * This code only triggers if there's nothing in radix tree
1445 	 * beyond 'end'.
1446 	 */
1447 	if (result == SCAN_SUCCEED && index < end) {
1448 		int n = end - index;
1449 
1450 		if (!shmem_charge(mapping->host, n)) {
1451 			result = SCAN_FAIL;
1452 			goto tree_locked;
1453 		}
1454 
1455 		for (; index < end; index++) {
1456 			radix_tree_insert(&mapping->page_tree, index,
1457 					new_page + (index % HPAGE_PMD_NR));
1458 		}
1459 		nr_none += n;
1460 	}
1461 
1462 tree_locked:
1463 	spin_unlock_irq(&mapping->tree_lock);
1464 tree_unlocked:
1465 
1466 	if (result == SCAN_SUCCEED) {
1467 		unsigned long flags;
1468 		struct zone *zone = page_zone(new_page);
1469 
1470 		/*
1471 		 * Replacing old pages with new one has succeed, now we need to
1472 		 * copy the content and free old pages.
1473 		 */
1474 		list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1475 			copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1476 					page);
1477 			list_del(&page->lru);
1478 			unlock_page(page);
1479 			page_ref_unfreeze(page, 1);
1480 			page->mapping = NULL;
1481 			ClearPageActive(page);
1482 			ClearPageUnevictable(page);
1483 			put_page(page);
1484 		}
1485 
1486 		local_irq_save(flags);
1487 		__inc_node_page_state(new_page, NR_SHMEM_THPS);
1488 		if (nr_none) {
1489 			__mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1490 			__mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
1491 		}
1492 		local_irq_restore(flags);
1493 
1494 		/*
1495 		 * Remove pte page tables, so we can re-faulti
1496 		 * the page as huge.
1497 		 */
1498 		retract_page_tables(mapping, start);
1499 
1500 		/* Everything is ready, let's unfreeze the new_page */
1501 		set_page_dirty(new_page);
1502 		SetPageUptodate(new_page);
1503 		page_ref_unfreeze(new_page, HPAGE_PMD_NR);
1504 		mem_cgroup_commit_charge(new_page, memcg, false, true);
1505 		lru_cache_add_anon(new_page);
1506 		unlock_page(new_page);
1507 
1508 		*hpage = NULL;
1509 	} else {
1510 		/* Something went wrong: rollback changes to the radix-tree */
1511 		shmem_uncharge(mapping->host, nr_none);
1512 		spin_lock_irq(&mapping->tree_lock);
1513 		radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
1514 				start) {
1515 			if (iter.index >= end)
1516 				break;
1517 			page = list_first_entry_or_null(&pagelist,
1518 					struct page, lru);
1519 			if (!page || iter.index < page->index) {
1520 				if (!nr_none)
1521 					break;
1522 				/* Put holes back where they were */
1523 				radix_tree_replace_slot(slot, NULL);
1524 				nr_none--;
1525 				continue;
1526 			}
1527 
1528 			VM_BUG_ON_PAGE(page->index != iter.index, page);
1529 
1530 			/* Unfreeze the page. */
1531 			list_del(&page->lru);
1532 			page_ref_unfreeze(page, 2);
1533 			radix_tree_replace_slot(slot, page);
1534 			spin_unlock_irq(&mapping->tree_lock);
1535 			putback_lru_page(page);
1536 			unlock_page(page);
1537 			spin_lock_irq(&mapping->tree_lock);
1538 		}
1539 		VM_BUG_ON(nr_none);
1540 		spin_unlock_irq(&mapping->tree_lock);
1541 
1542 		/* Unfreeze new_page, caller would take care about freeing it */
1543 		page_ref_unfreeze(new_page, 1);
1544 		mem_cgroup_cancel_charge(new_page, memcg, true);
1545 		unlock_page(new_page);
1546 		new_page->mapping = NULL;
1547 	}
1548 out:
1549 	VM_BUG_ON(!list_empty(&pagelist));
1550 	/* TODO: tracepoints */
1551 }
1552 
1553 static void khugepaged_scan_shmem(struct mm_struct *mm,
1554 		struct address_space *mapping,
1555 		pgoff_t start, struct page **hpage)
1556 {
1557 	struct page *page = NULL;
1558 	struct radix_tree_iter iter;
1559 	void **slot;
1560 	int present, swap;
1561 	int node = NUMA_NO_NODE;
1562 	int result = SCAN_SUCCEED;
1563 
1564 	present = 0;
1565 	swap = 0;
1566 	memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1567 	rcu_read_lock();
1568 	radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1569 		if (iter.index >= start + HPAGE_PMD_NR)
1570 			break;
1571 
1572 		page = radix_tree_deref_slot(slot);
1573 		if (radix_tree_deref_retry(page)) {
1574 			slot = radix_tree_iter_retry(&iter);
1575 			continue;
1576 		}
1577 
1578 		if (radix_tree_exception(page)) {
1579 			if (++swap > khugepaged_max_ptes_swap) {
1580 				result = SCAN_EXCEED_SWAP_PTE;
1581 				break;
1582 			}
1583 			continue;
1584 		}
1585 
1586 		if (PageTransCompound(page)) {
1587 			result = SCAN_PAGE_COMPOUND;
1588 			break;
1589 		}
1590 
1591 		node = page_to_nid(page);
1592 		if (khugepaged_scan_abort(node)) {
1593 			result = SCAN_SCAN_ABORT;
1594 			break;
1595 		}
1596 		khugepaged_node_load[node]++;
1597 
1598 		if (!PageLRU(page)) {
1599 			result = SCAN_PAGE_LRU;
1600 			break;
1601 		}
1602 
1603 		if (page_count(page) != 1 + page_mapcount(page)) {
1604 			result = SCAN_PAGE_COUNT;
1605 			break;
1606 		}
1607 
1608 		/*
1609 		 * We probably should check if the page is referenced here, but
1610 		 * nobody would transfer pte_young() to PageReferenced() for us.
1611 		 * And rmap walk here is just too costly...
1612 		 */
1613 
1614 		present++;
1615 
1616 		if (need_resched()) {
1617 			cond_resched_rcu();
1618 			slot = radix_tree_iter_next(&iter);
1619 		}
1620 	}
1621 	rcu_read_unlock();
1622 
1623 	if (result == SCAN_SUCCEED) {
1624 		if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1625 			result = SCAN_EXCEED_NONE_PTE;
1626 		} else {
1627 			node = khugepaged_find_target_node();
1628 			collapse_shmem(mm, mapping, start, hpage, node);
1629 		}
1630 	}
1631 
1632 	/* TODO: tracepoints */
1633 }
1634 #else
1635 static void khugepaged_scan_shmem(struct mm_struct *mm,
1636 		struct address_space *mapping,
1637 		pgoff_t start, struct page **hpage)
1638 {
1639 	BUILD_BUG();
1640 }
1641 #endif
1642 
1643 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1644 					    struct page **hpage)
1645 	__releases(&khugepaged_mm_lock)
1646 	__acquires(&khugepaged_mm_lock)
1647 {
1648 	struct mm_slot *mm_slot;
1649 	struct mm_struct *mm;
1650 	struct vm_area_struct *vma;
1651 	int progress = 0;
1652 
1653 	VM_BUG_ON(!pages);
1654 	VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1655 
1656 	if (khugepaged_scan.mm_slot)
1657 		mm_slot = khugepaged_scan.mm_slot;
1658 	else {
1659 		mm_slot = list_entry(khugepaged_scan.mm_head.next,
1660 				     struct mm_slot, mm_node);
1661 		khugepaged_scan.address = 0;
1662 		khugepaged_scan.mm_slot = mm_slot;
1663 	}
1664 	spin_unlock(&khugepaged_mm_lock);
1665 
1666 	mm = mm_slot->mm;
1667 	down_read(&mm->mmap_sem);
1668 	if (unlikely(khugepaged_test_exit(mm)))
1669 		vma = NULL;
1670 	else
1671 		vma = find_vma(mm, khugepaged_scan.address);
1672 
1673 	progress++;
1674 	for (; vma; vma = vma->vm_next) {
1675 		unsigned long hstart, hend;
1676 
1677 		cond_resched();
1678 		if (unlikely(khugepaged_test_exit(mm))) {
1679 			progress++;
1680 			break;
1681 		}
1682 		if (!hugepage_vma_check(vma)) {
1683 skip:
1684 			progress++;
1685 			continue;
1686 		}
1687 		hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1688 		hend = vma->vm_end & HPAGE_PMD_MASK;
1689 		if (hstart >= hend)
1690 			goto skip;
1691 		if (khugepaged_scan.address > hend)
1692 			goto skip;
1693 		if (khugepaged_scan.address < hstart)
1694 			khugepaged_scan.address = hstart;
1695 		VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1696 
1697 		while (khugepaged_scan.address < hend) {
1698 			int ret;
1699 			cond_resched();
1700 			if (unlikely(khugepaged_test_exit(mm)))
1701 				goto breakouterloop;
1702 
1703 			VM_BUG_ON(khugepaged_scan.address < hstart ||
1704 				  khugepaged_scan.address + HPAGE_PMD_SIZE >
1705 				  hend);
1706 			if (shmem_file(vma->vm_file)) {
1707 				struct file *file;
1708 				pgoff_t pgoff = linear_page_index(vma,
1709 						khugepaged_scan.address);
1710 				if (!shmem_huge_enabled(vma))
1711 					goto skip;
1712 				file = get_file(vma->vm_file);
1713 				up_read(&mm->mmap_sem);
1714 				ret = 1;
1715 				khugepaged_scan_shmem(mm, file->f_mapping,
1716 						pgoff, hpage);
1717 				fput(file);
1718 			} else {
1719 				ret = khugepaged_scan_pmd(mm, vma,
1720 						khugepaged_scan.address,
1721 						hpage);
1722 			}
1723 			/* move to next address */
1724 			khugepaged_scan.address += HPAGE_PMD_SIZE;
1725 			progress += HPAGE_PMD_NR;
1726 			if (ret)
1727 				/* we released mmap_sem so break loop */
1728 				goto breakouterloop_mmap_sem;
1729 			if (progress >= pages)
1730 				goto breakouterloop;
1731 		}
1732 	}
1733 breakouterloop:
1734 	up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1735 breakouterloop_mmap_sem:
1736 
1737 	spin_lock(&khugepaged_mm_lock);
1738 	VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1739 	/*
1740 	 * Release the current mm_slot if this mm is about to die, or
1741 	 * if we scanned all vmas of this mm.
1742 	 */
1743 	if (khugepaged_test_exit(mm) || !vma) {
1744 		/*
1745 		 * Make sure that if mm_users is reaching zero while
1746 		 * khugepaged runs here, khugepaged_exit will find
1747 		 * mm_slot not pointing to the exiting mm.
1748 		 */
1749 		if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1750 			khugepaged_scan.mm_slot = list_entry(
1751 				mm_slot->mm_node.next,
1752 				struct mm_slot, mm_node);
1753 			khugepaged_scan.address = 0;
1754 		} else {
1755 			khugepaged_scan.mm_slot = NULL;
1756 			khugepaged_full_scans++;
1757 		}
1758 
1759 		collect_mm_slot(mm_slot);
1760 	}
1761 
1762 	return progress;
1763 }
1764 
1765 static int khugepaged_has_work(void)
1766 {
1767 	return !list_empty(&khugepaged_scan.mm_head) &&
1768 		khugepaged_enabled();
1769 }
1770 
1771 static int khugepaged_wait_event(void)
1772 {
1773 	return !list_empty(&khugepaged_scan.mm_head) ||
1774 		kthread_should_stop();
1775 }
1776 
1777 static void khugepaged_do_scan(void)
1778 {
1779 	struct page *hpage = NULL;
1780 	unsigned int progress = 0, pass_through_head = 0;
1781 	unsigned int pages = khugepaged_pages_to_scan;
1782 	bool wait = true;
1783 
1784 	barrier(); /* write khugepaged_pages_to_scan to local stack */
1785 
1786 	while (progress < pages) {
1787 		if (!khugepaged_prealloc_page(&hpage, &wait))
1788 			break;
1789 
1790 		cond_resched();
1791 
1792 		if (unlikely(kthread_should_stop() || try_to_freeze()))
1793 			break;
1794 
1795 		spin_lock(&khugepaged_mm_lock);
1796 		if (!khugepaged_scan.mm_slot)
1797 			pass_through_head++;
1798 		if (khugepaged_has_work() &&
1799 		    pass_through_head < 2)
1800 			progress += khugepaged_scan_mm_slot(pages - progress,
1801 							    &hpage);
1802 		else
1803 			progress = pages;
1804 		spin_unlock(&khugepaged_mm_lock);
1805 	}
1806 
1807 	if (!IS_ERR_OR_NULL(hpage))
1808 		put_page(hpage);
1809 }
1810 
1811 static bool khugepaged_should_wakeup(void)
1812 {
1813 	return kthread_should_stop() ||
1814 	       time_after_eq(jiffies, khugepaged_sleep_expire);
1815 }
1816 
1817 static void khugepaged_wait_work(void)
1818 {
1819 	if (khugepaged_has_work()) {
1820 		const unsigned long scan_sleep_jiffies =
1821 			msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1822 
1823 		if (!scan_sleep_jiffies)
1824 			return;
1825 
1826 		khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1827 		wait_event_freezable_timeout(khugepaged_wait,
1828 					     khugepaged_should_wakeup(),
1829 					     scan_sleep_jiffies);
1830 		return;
1831 	}
1832 
1833 	if (khugepaged_enabled())
1834 		wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1835 }
1836 
1837 static int khugepaged(void *none)
1838 {
1839 	struct mm_slot *mm_slot;
1840 
1841 	set_freezable();
1842 	set_user_nice(current, MAX_NICE);
1843 
1844 	while (!kthread_should_stop()) {
1845 		khugepaged_do_scan();
1846 		khugepaged_wait_work();
1847 	}
1848 
1849 	spin_lock(&khugepaged_mm_lock);
1850 	mm_slot = khugepaged_scan.mm_slot;
1851 	khugepaged_scan.mm_slot = NULL;
1852 	if (mm_slot)
1853 		collect_mm_slot(mm_slot);
1854 	spin_unlock(&khugepaged_mm_lock);
1855 	return 0;
1856 }
1857 
1858 static void set_recommended_min_free_kbytes(void)
1859 {
1860 	struct zone *zone;
1861 	int nr_zones = 0;
1862 	unsigned long recommended_min;
1863 
1864 	for_each_populated_zone(zone)
1865 		nr_zones++;
1866 
1867 	/* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1868 	recommended_min = pageblock_nr_pages * nr_zones * 2;
1869 
1870 	/*
1871 	 * Make sure that on average at least two pageblocks are almost free
1872 	 * of another type, one for a migratetype to fall back to and a
1873 	 * second to avoid subsequent fallbacks of other types There are 3
1874 	 * MIGRATE_TYPES we care about.
1875 	 */
1876 	recommended_min += pageblock_nr_pages * nr_zones *
1877 			   MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1878 
1879 	/* don't ever allow to reserve more than 5% of the lowmem */
1880 	recommended_min = min(recommended_min,
1881 			      (unsigned long) nr_free_buffer_pages() / 20);
1882 	recommended_min <<= (PAGE_SHIFT-10);
1883 
1884 	if (recommended_min > min_free_kbytes) {
1885 		if (user_min_free_kbytes >= 0)
1886 			pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1887 				min_free_kbytes, recommended_min);
1888 
1889 		min_free_kbytes = recommended_min;
1890 	}
1891 	setup_per_zone_wmarks();
1892 }
1893 
1894 int start_stop_khugepaged(void)
1895 {
1896 	static struct task_struct *khugepaged_thread __read_mostly;
1897 	static DEFINE_MUTEX(khugepaged_mutex);
1898 	int err = 0;
1899 
1900 	mutex_lock(&khugepaged_mutex);
1901 	if (khugepaged_enabled()) {
1902 		if (!khugepaged_thread)
1903 			khugepaged_thread = kthread_run(khugepaged, NULL,
1904 							"khugepaged");
1905 		if (IS_ERR(khugepaged_thread)) {
1906 			pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1907 			err = PTR_ERR(khugepaged_thread);
1908 			khugepaged_thread = NULL;
1909 			goto fail;
1910 		}
1911 
1912 		if (!list_empty(&khugepaged_scan.mm_head))
1913 			wake_up_interruptible(&khugepaged_wait);
1914 
1915 		set_recommended_min_free_kbytes();
1916 	} else if (khugepaged_thread) {
1917 		kthread_stop(khugepaged_thread);
1918 		khugepaged_thread = NULL;
1919 	}
1920 fail:
1921 	mutex_unlock(&khugepaged_mutex);
1922 	return err;
1923 }
1924