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