blob: 898eb26f5dc862ec52f4bd570065db6cfe4635d8 [file] [log] [blame]
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -07001#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>
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -070017#include <linux/shmem_fs.h>
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -070018
19#include <asm/tlb.h>
20#include <asm/pgalloc.h>
21#include "internal.h"
22
23enum 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,
Ebru Akagunduz0db501f2016-07-26 15:26:46 -070030 SCAN_LACK_REFERENCED_PAGE,
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -070031 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,
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -070046 SCAN_EXCEED_SWAP_PTE,
47 SCAN_TRUNCATED,
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -070048};
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 */
54static unsigned int khugepaged_pages_to_scan __read_mostly;
55static unsigned int khugepaged_pages_collapsed;
56static unsigned int khugepaged_full_scans;
57static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
58/* during fragmentation poll the hugepage allocator once every minute */
59static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
60static unsigned long khugepaged_sleep_expire;
61static DEFINE_SPINLOCK(khugepaged_mm_lock);
62static 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 */
68static unsigned int khugepaged_max_ptes_none __read_mostly;
69static unsigned int khugepaged_max_ptes_swap __read_mostly;
70
71#define MM_SLOTS_HASH_BITS 10
72static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
73
74static 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 */
82struct 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 */
96struct khugepaged_scan {
97 struct list_head mm_head;
98 struct mm_slot *mm_slot;
99 unsigned long address;
100};
101
102static struct khugepaged_scan khugepaged_scan = {
103 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
104};
105
Jérémy Lefauree1465d12016-11-30 15:54:02 -0800106#ifdef CONFIG_SYSFS
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700107static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
108 struct kobj_attribute *attr,
109 char *buf)
110{
111 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
112}
113
114static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
115 struct kobj_attribute *attr,
116 const char *buf, size_t count)
117{
118 unsigned long msecs;
119 int err;
120
121 err = kstrtoul(buf, 10, &msecs);
122 if (err || msecs > UINT_MAX)
123 return -EINVAL;
124
125 khugepaged_scan_sleep_millisecs = msecs;
126 khugepaged_sleep_expire = 0;
127 wake_up_interruptible(&khugepaged_wait);
128
129 return count;
130}
131static struct kobj_attribute scan_sleep_millisecs_attr =
132 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
133 scan_sleep_millisecs_store);
134
135static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
136 struct kobj_attribute *attr,
137 char *buf)
138{
139 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
140}
141
142static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
143 struct kobj_attribute *attr,
144 const char *buf, size_t count)
145{
146 unsigned long msecs;
147 int err;
148
149 err = kstrtoul(buf, 10, &msecs);
150 if (err || msecs > UINT_MAX)
151 return -EINVAL;
152
153 khugepaged_alloc_sleep_millisecs = msecs;
154 khugepaged_sleep_expire = 0;
155 wake_up_interruptible(&khugepaged_wait);
156
157 return count;
158}
159static struct kobj_attribute alloc_sleep_millisecs_attr =
160 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
161 alloc_sleep_millisecs_store);
162
163static ssize_t pages_to_scan_show(struct kobject *kobj,
164 struct kobj_attribute *attr,
165 char *buf)
166{
167 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
168}
169static ssize_t pages_to_scan_store(struct kobject *kobj,
170 struct kobj_attribute *attr,
171 const char *buf, size_t count)
172{
173 int err;
174 unsigned long pages;
175
176 err = kstrtoul(buf, 10, &pages);
177 if (err || !pages || pages > UINT_MAX)
178 return -EINVAL;
179
180 khugepaged_pages_to_scan = pages;
181
182 return count;
183}
184static struct kobj_attribute pages_to_scan_attr =
185 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
186 pages_to_scan_store);
187
188static ssize_t pages_collapsed_show(struct kobject *kobj,
189 struct kobj_attribute *attr,
190 char *buf)
191{
192 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
193}
194static struct kobj_attribute pages_collapsed_attr =
195 __ATTR_RO(pages_collapsed);
196
197static ssize_t full_scans_show(struct kobject *kobj,
198 struct kobj_attribute *attr,
199 char *buf)
200{
201 return sprintf(buf, "%u\n", khugepaged_full_scans);
202}
203static struct kobj_attribute full_scans_attr =
204 __ATTR_RO(full_scans);
205
206static ssize_t khugepaged_defrag_show(struct kobject *kobj,
207 struct kobj_attribute *attr, char *buf)
208{
209 return single_hugepage_flag_show(kobj, attr, buf,
210 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
211}
212static ssize_t khugepaged_defrag_store(struct kobject *kobj,
213 struct kobj_attribute *attr,
214 const char *buf, size_t count)
215{
216 return single_hugepage_flag_store(kobj, attr, buf, count,
217 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
218}
219static struct kobj_attribute khugepaged_defrag_attr =
220 __ATTR(defrag, 0644, khugepaged_defrag_show,
221 khugepaged_defrag_store);
222
223/*
224 * max_ptes_none controls if khugepaged should collapse hugepages over
225 * any unmapped ptes in turn potentially increasing the memory
226 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
227 * reduce the available free memory in the system as it
228 * runs. Increasing max_ptes_none will instead potentially reduce the
229 * free memory in the system during the khugepaged scan.
230 */
231static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
232 struct kobj_attribute *attr,
233 char *buf)
234{
235 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
236}
237static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
238 struct kobj_attribute *attr,
239 const char *buf, size_t count)
240{
241 int err;
242 unsigned long max_ptes_none;
243
244 err = kstrtoul(buf, 10, &max_ptes_none);
245 if (err || max_ptes_none > HPAGE_PMD_NR-1)
246 return -EINVAL;
247
248 khugepaged_max_ptes_none = max_ptes_none;
249
250 return count;
251}
252static struct kobj_attribute khugepaged_max_ptes_none_attr =
253 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
254 khugepaged_max_ptes_none_store);
255
256static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
257 struct kobj_attribute *attr,
258 char *buf)
259{
260 return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
261}
262
263static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
264 struct kobj_attribute *attr,
265 const char *buf, size_t count)
266{
267 int err;
268 unsigned long max_ptes_swap;
269
270 err = kstrtoul(buf, 10, &max_ptes_swap);
271 if (err || max_ptes_swap > HPAGE_PMD_NR-1)
272 return -EINVAL;
273
274 khugepaged_max_ptes_swap = max_ptes_swap;
275
276 return count;
277}
278
279static struct kobj_attribute khugepaged_max_ptes_swap_attr =
280 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
281 khugepaged_max_ptes_swap_store);
282
283static struct attribute *khugepaged_attr[] = {
284 &khugepaged_defrag_attr.attr,
285 &khugepaged_max_ptes_none_attr.attr,
286 &pages_to_scan_attr.attr,
287 &pages_collapsed_attr.attr,
288 &full_scans_attr.attr,
289 &scan_sleep_millisecs_attr.attr,
290 &alloc_sleep_millisecs_attr.attr,
291 &khugepaged_max_ptes_swap_attr.attr,
292 NULL,
293};
294
295struct attribute_group khugepaged_attr_group = {
296 .attrs = khugepaged_attr,
297 .name = "khugepaged",
298};
Jérémy Lefauree1465d12016-11-30 15:54:02 -0800299#endif /* CONFIG_SYSFS */
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700300
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -0700301#define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700302
303int hugepage_madvise(struct vm_area_struct *vma,
304 unsigned long *vm_flags, int advice)
305{
306 switch (advice) {
307 case MADV_HUGEPAGE:
308#ifdef CONFIG_S390
309 /*
310 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
311 * can't handle this properly after s390_enable_sie, so we simply
312 * ignore the madvise to prevent qemu from causing a SIGSEGV.
313 */
314 if (mm_has_pgste(vma->vm_mm))
315 return 0;
316#endif
317 *vm_flags &= ~VM_NOHUGEPAGE;
318 *vm_flags |= VM_HUGEPAGE;
319 /*
320 * If the vma become good for khugepaged to scan,
321 * register it here without waiting a page fault that
322 * may not happen any time soon.
323 */
324 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
325 khugepaged_enter_vma_merge(vma, *vm_flags))
326 return -ENOMEM;
327 break;
328 case MADV_NOHUGEPAGE:
329 *vm_flags &= ~VM_HUGEPAGE;
330 *vm_flags |= VM_NOHUGEPAGE;
331 /*
332 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
333 * this vma even if we leave the mm registered in khugepaged if
334 * it got registered before VM_NOHUGEPAGE was set.
335 */
336 break;
337 }
338
339 return 0;
340}
341
342int __init khugepaged_init(void)
343{
344 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
345 sizeof(struct mm_slot),
346 __alignof__(struct mm_slot), 0, NULL);
347 if (!mm_slot_cache)
348 return -ENOMEM;
349
350 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
351 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
352 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
353
354 return 0;
355}
356
357void __init khugepaged_destroy(void)
358{
359 kmem_cache_destroy(mm_slot_cache);
360}
361
362static inline struct mm_slot *alloc_mm_slot(void)
363{
364 if (!mm_slot_cache) /* initialization failed */
365 return NULL;
366 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
367}
368
369static inline void free_mm_slot(struct mm_slot *mm_slot)
370{
371 kmem_cache_free(mm_slot_cache, mm_slot);
372}
373
374static struct mm_slot *get_mm_slot(struct mm_struct *mm)
375{
376 struct mm_slot *mm_slot;
377
378 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
379 if (mm == mm_slot->mm)
380 return mm_slot;
381
382 return NULL;
383}
384
385static void insert_to_mm_slots_hash(struct mm_struct *mm,
386 struct mm_slot *mm_slot)
387{
388 mm_slot->mm = mm;
389 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
390}
391
392static inline int khugepaged_test_exit(struct mm_struct *mm)
393{
394 return atomic_read(&mm->mm_users) == 0;
395}
396
397int __khugepaged_enter(struct mm_struct *mm)
398{
399 struct mm_slot *mm_slot;
400 int wakeup;
401
402 mm_slot = alloc_mm_slot();
403 if (!mm_slot)
404 return -ENOMEM;
405
406 /* __khugepaged_exit() must not run from under us */
407 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
408 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
409 free_mm_slot(mm_slot);
410 return 0;
411 }
412
413 spin_lock(&khugepaged_mm_lock);
414 insert_to_mm_slots_hash(mm, mm_slot);
415 /*
416 * Insert just behind the scanning cursor, to let the area settle
417 * down a little.
418 */
419 wakeup = list_empty(&khugepaged_scan.mm_head);
420 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
421 spin_unlock(&khugepaged_mm_lock);
422
423 atomic_inc(&mm->mm_count);
424 if (wakeup)
425 wake_up_interruptible(&khugepaged_wait);
426
427 return 0;
428}
429
430int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
431 unsigned long vm_flags)
432{
433 unsigned long hstart, hend;
434 if (!vma->anon_vma)
435 /*
436 * Not yet faulted in so we will register later in the
437 * page fault if needed.
438 */
439 return 0;
440 if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
441 /* khugepaged not yet working on file or special mappings */
442 return 0;
443 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
444 hend = vma->vm_end & HPAGE_PMD_MASK;
445 if (hstart < hend)
446 return khugepaged_enter(vma, vm_flags);
447 return 0;
448}
449
450void __khugepaged_exit(struct mm_struct *mm)
451{
452 struct mm_slot *mm_slot;
453 int free = 0;
454
455 spin_lock(&khugepaged_mm_lock);
456 mm_slot = get_mm_slot(mm);
457 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
458 hash_del(&mm_slot->hash);
459 list_del(&mm_slot->mm_node);
460 free = 1;
461 }
462 spin_unlock(&khugepaged_mm_lock);
463
464 if (free) {
465 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
466 free_mm_slot(mm_slot);
467 mmdrop(mm);
468 } else if (mm_slot) {
469 /*
470 * This is required to serialize against
471 * khugepaged_test_exit() (which is guaranteed to run
472 * under mmap sem read mode). Stop here (after we
473 * return all pagetables will be destroyed) until
474 * khugepaged has finished working on the pagetables
475 * under the mmap_sem.
476 */
477 down_write(&mm->mmap_sem);
478 up_write(&mm->mmap_sem);
479 }
480}
481
482static void release_pte_page(struct page *page)
483{
484 /* 0 stands for page_is_file_cache(page) == false */
Mel Gorman599d0c92016-07-28 15:45:31 -0700485 dec_node_page_state(page, NR_ISOLATED_ANON + 0);
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700486 unlock_page(page);
487 putback_lru_page(page);
488}
489
490static void release_pte_pages(pte_t *pte, pte_t *_pte)
491{
492 while (--_pte >= pte) {
493 pte_t pteval = *_pte;
494 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
495 release_pte_page(pte_page(pteval));
496 }
497}
498
499static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
500 unsigned long address,
501 pte_t *pte)
502{
503 struct page *page = NULL;
504 pte_t *_pte;
Ebru Akagunduz0db501f2016-07-26 15:26:46 -0700505 int none_or_zero = 0, result = 0, referenced = 0;
506 bool writable = false;
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700507
508 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
509 _pte++, address += PAGE_SIZE) {
510 pte_t pteval = *_pte;
511 if (pte_none(pteval) || (pte_present(pteval) &&
512 is_zero_pfn(pte_pfn(pteval)))) {
513 if (!userfaultfd_armed(vma) &&
514 ++none_or_zero <= khugepaged_max_ptes_none) {
515 continue;
516 } else {
517 result = SCAN_EXCEED_NONE_PTE;
518 goto out;
519 }
520 }
521 if (!pte_present(pteval)) {
522 result = SCAN_PTE_NON_PRESENT;
523 goto out;
524 }
525 page = vm_normal_page(vma, address, pteval);
526 if (unlikely(!page)) {
527 result = SCAN_PAGE_NULL;
528 goto out;
529 }
530
Kirill A. Shutemov24284d52018-03-22 16:17:28 -0700531 /* TODO: teach khugepaged to collapse THP mapped with pte */
532 if (PageCompound(page)) {
533 result = SCAN_PAGE_COMPOUND;
534 goto out;
535 }
536
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700537 VM_BUG_ON_PAGE(!PageAnon(page), page);
538 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
539
540 /*
541 * We can do it before isolate_lru_page because the
542 * page can't be freed from under us. NOTE: PG_lock
543 * is needed to serialize against split_huge_page
544 * when invoked from the VM.
545 */
546 if (!trylock_page(page)) {
547 result = SCAN_PAGE_LOCK;
548 goto out;
549 }
550
551 /*
552 * cannot use mapcount: can't collapse if there's a gup pin.
553 * The page must only be referenced by the scanned process
554 * and page swap cache.
555 */
556 if (page_count(page) != 1 + !!PageSwapCache(page)) {
557 unlock_page(page);
558 result = SCAN_PAGE_COUNT;
559 goto out;
560 }
561 if (pte_write(pteval)) {
562 writable = true;
563 } else {
564 if (PageSwapCache(page) &&
565 !reuse_swap_page(page, NULL)) {
566 unlock_page(page);
567 result = SCAN_SWAP_CACHE_PAGE;
568 goto out;
569 }
570 /*
571 * Page is not in the swap cache. It can be collapsed
572 * into a THP.
573 */
574 }
575
576 /*
577 * Isolate the page to avoid collapsing an hugepage
578 * currently in use by the VM.
579 */
580 if (isolate_lru_page(page)) {
581 unlock_page(page);
582 result = SCAN_DEL_PAGE_LRU;
583 goto out;
584 }
585 /* 0 stands for page_is_file_cache(page) == false */
Mel Gorman599d0c92016-07-28 15:45:31 -0700586 inc_node_page_state(page, NR_ISOLATED_ANON + 0);
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700587 VM_BUG_ON_PAGE(!PageLocked(page), page);
588 VM_BUG_ON_PAGE(PageLRU(page), page);
589
Ebru Akagunduz0db501f2016-07-26 15:26:46 -0700590 /* There should be enough young pte to collapse the page */
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700591 if (pte_young(pteval) ||
592 page_is_young(page) || PageReferenced(page) ||
593 mmu_notifier_test_young(vma->vm_mm, address))
Ebru Akagunduz0db501f2016-07-26 15:26:46 -0700594 referenced++;
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700595 }
596 if (likely(writable)) {
597 if (likely(referenced)) {
598 result = SCAN_SUCCEED;
599 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
600 referenced, writable, result);
601 return 1;
602 }
603 } else {
604 result = SCAN_PAGE_RO;
605 }
606
607out:
608 release_pte_pages(pte, _pte);
609 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
610 referenced, writable, result);
611 return 0;
612}
613
614static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
615 struct vm_area_struct *vma,
616 unsigned long address,
617 spinlock_t *ptl)
618{
619 pte_t *_pte;
620 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
621 pte_t pteval = *_pte;
622 struct page *src_page;
623
624 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
625 clear_user_highpage(page, address);
626 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
627 if (is_zero_pfn(pte_pfn(pteval))) {
628 /*
629 * ptl mostly unnecessary.
630 */
631 spin_lock(ptl);
632 /*
633 * paravirt calls inside pte_clear here are
634 * superfluous.
635 */
636 pte_clear(vma->vm_mm, address, _pte);
637 spin_unlock(ptl);
638 }
639 } else {
640 src_page = pte_page(pteval);
641 copy_user_highpage(page, src_page, address, vma);
642 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
643 release_pte_page(src_page);
644 /*
645 * ptl mostly unnecessary, but preempt has to
646 * be disabled to update the per-cpu stats
647 * inside page_remove_rmap().
648 */
649 spin_lock(ptl);
650 /*
651 * paravirt calls inside pte_clear here are
652 * superfluous.
653 */
654 pte_clear(vma->vm_mm, address, _pte);
655 page_remove_rmap(src_page, false);
656 spin_unlock(ptl);
657 free_page_and_swap_cache(src_page);
658 }
659
660 address += PAGE_SIZE;
661 page++;
662 }
663}
664
665static void khugepaged_alloc_sleep(void)
666{
667 DEFINE_WAIT(wait);
668
669 add_wait_queue(&khugepaged_wait, &wait);
670 freezable_schedule_timeout_interruptible(
671 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
672 remove_wait_queue(&khugepaged_wait, &wait);
673}
674
675static int khugepaged_node_load[MAX_NUMNODES];
676
677static bool khugepaged_scan_abort(int nid)
678{
679 int i;
680
681 /*
Mel Gormana5f5f912016-07-28 15:46:32 -0700682 * If node_reclaim_mode is disabled, then no extra effort is made to
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700683 * allocate memory locally.
684 */
Mel Gormana5f5f912016-07-28 15:46:32 -0700685 if (!node_reclaim_mode)
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700686 return false;
687
688 /* If there is a count for this node already, it must be acceptable */
689 if (khugepaged_node_load[nid])
690 return false;
691
692 for (i = 0; i < MAX_NUMNODES; i++) {
693 if (!khugepaged_node_load[i])
694 continue;
695 if (node_distance(nid, i) > RECLAIM_DISTANCE)
696 return true;
697 }
698 return false;
699}
700
701/* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
702static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
703{
Vlastimil Babka25160352016-07-28 15:49:25 -0700704 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700705}
706
707#ifdef CONFIG_NUMA
708static int khugepaged_find_target_node(void)
709{
710 static int last_khugepaged_target_node = NUMA_NO_NODE;
711 int nid, target_node = 0, max_value = 0;
712
713 /* find first node with max normal pages hit */
714 for (nid = 0; nid < MAX_NUMNODES; nid++)
715 if (khugepaged_node_load[nid] > max_value) {
716 max_value = khugepaged_node_load[nid];
717 target_node = nid;
718 }
719
720 /* do some balance if several nodes have the same hit record */
721 if (target_node <= last_khugepaged_target_node)
722 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
723 nid++)
724 if (max_value == khugepaged_node_load[nid]) {
725 target_node = nid;
726 break;
727 }
728
729 last_khugepaged_target_node = target_node;
730 return target_node;
731}
732
733static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
734{
735 if (IS_ERR(*hpage)) {
736 if (!*wait)
737 return false;
738
739 *wait = false;
740 *hpage = NULL;
741 khugepaged_alloc_sleep();
742 } else if (*hpage) {
743 put_page(*hpage);
744 *hpage = NULL;
745 }
746
747 return true;
748}
749
750static struct page *
Kirill A. Shutemov988ddb72016-07-26 15:26:26 -0700751khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700752{
753 VM_BUG_ON_PAGE(*hpage, *hpage);
754
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700755 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
756 if (unlikely(!*hpage)) {
757 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
758 *hpage = ERR_PTR(-ENOMEM);
759 return NULL;
760 }
761
762 prep_transhuge_page(*hpage);
763 count_vm_event(THP_COLLAPSE_ALLOC);
764 return *hpage;
765}
766#else
767static int khugepaged_find_target_node(void)
768{
769 return 0;
770}
771
772static inline struct page *alloc_khugepaged_hugepage(void)
773{
774 struct page *page;
775
776 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
777 HPAGE_PMD_ORDER);
778 if (page)
779 prep_transhuge_page(page);
780 return page;
781}
782
783static struct page *khugepaged_alloc_hugepage(bool *wait)
784{
785 struct page *hpage;
786
787 do {
788 hpage = alloc_khugepaged_hugepage();
789 if (!hpage) {
790 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
791 if (!*wait)
792 return NULL;
793
794 *wait = false;
795 khugepaged_alloc_sleep();
796 } else
797 count_vm_event(THP_COLLAPSE_ALLOC);
798 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
799
800 return hpage;
801}
802
803static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
804{
805 if (!*hpage)
806 *hpage = khugepaged_alloc_hugepage(wait);
807
808 if (unlikely(!*hpage))
809 return false;
810
811 return true;
812}
813
814static struct page *
Kirill A. Shutemov988ddb72016-07-26 15:26:26 -0700815khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700816{
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700817 VM_BUG_ON(!*hpage);
818
819 return *hpage;
820}
821#endif
822
823static bool hugepage_vma_check(struct vm_area_struct *vma)
824{
825 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
826 (vma->vm_flags & VM_NOHUGEPAGE))
827 return false;
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -0700828 if (shmem_file(vma->vm_file)) {
Kirill A. Shutemove496cf32016-07-26 15:26:35 -0700829 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
830 return false;
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -0700831 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
832 HPAGE_PMD_NR);
833 }
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700834 if (!vma->anon_vma || vma->vm_ops)
835 return false;
836 if (is_vma_temporary_stack(vma))
837 return false;
838 return !(vma->vm_flags & VM_NO_KHUGEPAGED);
839}
840
841/*
842 * If mmap_sem temporarily dropped, revalidate vma
843 * before taking mmap_sem.
844 * Return 0 if succeeds, otherwise return none-zero
845 * value (scan code).
846 */
847
Kirill A. Shutemovc131f752016-09-19 14:44:01 -0700848static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
849 struct vm_area_struct **vmap)
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700850{
851 struct vm_area_struct *vma;
852 unsigned long hstart, hend;
853
854 if (unlikely(khugepaged_test_exit(mm)))
855 return SCAN_ANY_PROCESS;
856
Kirill A. Shutemovc131f752016-09-19 14:44:01 -0700857 *vmap = vma = find_vma(mm, address);
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700858 if (!vma)
859 return SCAN_VMA_NULL;
860
861 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
862 hend = vma->vm_end & HPAGE_PMD_MASK;
863 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
864 return SCAN_ADDRESS_RANGE;
865 if (!hugepage_vma_check(vma))
866 return SCAN_VMA_CHECK;
867 return 0;
868}
869
870/*
871 * Bring missing pages in from swap, to complete THP collapse.
872 * Only done if khugepaged_scan_pmd believes it is worthwhile.
873 *
874 * Called and returns without pte mapped or spinlocks held,
875 * but with mmap_sem held to protect against vma changes.
876 */
877
878static bool __collapse_huge_page_swapin(struct mm_struct *mm,
879 struct vm_area_struct *vma,
Ebru Akagunduz0db501f2016-07-26 15:26:46 -0700880 unsigned long address, pmd_t *pmd,
881 int referenced)
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700882{
883 pte_t pteval;
884 int swapped_in = 0, ret = 0;
885 struct fault_env fe = {
886 .vma = vma,
887 .address = address,
888 .flags = FAULT_FLAG_ALLOW_RETRY,
889 .pmd = pmd,
890 };
891
Ebru Akagunduz982785c2016-09-19 14:44:04 -0700892 /* we only decide to swapin, if there is enough young ptes */
893 if (referenced < HPAGE_PMD_NR/2) {
894 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
895 return false;
896 }
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700897 fe.pte = pte_offset_map(pmd, address);
898 for (; fe.address < address + HPAGE_PMD_NR*PAGE_SIZE;
899 fe.pte++, fe.address += PAGE_SIZE) {
900 pteval = *fe.pte;
901 if (!is_swap_pte(pteval))
902 continue;
903 swapped_in++;
904 ret = do_swap_page(&fe, pteval);
Ebru Akagunduz0db501f2016-07-26 15:26:46 -0700905
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700906 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
907 if (ret & VM_FAULT_RETRY) {
908 down_read(&mm->mmap_sem);
Kirill A. Shutemovc131f752016-09-19 14:44:01 -0700909 if (hugepage_vma_revalidate(mm, address, &fe.vma)) {
Ebru Akagunduz47f863e2016-07-26 15:26:43 -0700910 /* vma is no longer available, don't continue to swapin */
Ebru Akagunduz0db501f2016-07-26 15:26:46 -0700911 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700912 return false;
Ebru Akagunduz47f863e2016-07-26 15:26:43 -0700913 }
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700914 /* check if the pmd is still valid */
915 if (mm_find_pmd(mm, address) != pmd)
916 return false;
917 }
918 if (ret & VM_FAULT_ERROR) {
Ebru Akagunduz0db501f2016-07-26 15:26:46 -0700919 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700920 return false;
921 }
922 /* pte is unmapped now, we need to map it */
923 fe.pte = pte_offset_map(pmd, fe.address);
924 }
925 fe.pte--;
926 pte_unmap(fe.pte);
Ebru Akagunduz0db501f2016-07-26 15:26:46 -0700927 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700928 return true;
929}
930
931static void collapse_huge_page(struct mm_struct *mm,
932 unsigned long address,
933 struct page **hpage,
Ebru Akagunduz0db501f2016-07-26 15:26:46 -0700934 int node, int referenced)
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700935{
936 pmd_t *pmd, _pmd;
937 pte_t *pte;
938 pgtable_t pgtable;
939 struct page *new_page;
940 spinlock_t *pmd_ptl, *pte_ptl;
941 int isolated = 0, result = 0;
942 struct mem_cgroup *memcg;
Kirill A. Shutemovc131f752016-09-19 14:44:01 -0700943 struct vm_area_struct *vma;
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700944 unsigned long mmun_start; /* For mmu_notifiers */
945 unsigned long mmun_end; /* For mmu_notifiers */
946 gfp_t gfp;
947
948 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
949
950 /* Only allocate from the target node */
951 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
952
Kirill A. Shutemov988ddb72016-07-26 15:26:26 -0700953 /*
954 * Before allocating the hugepage, release the mmap_sem read lock.
955 * The allocation can take potentially a long time if it involves
956 * sync compaction, and we do not need to hold the mmap_sem during
957 * that. We will recheck the vma after taking it again in write mode.
958 */
959 up_read(&mm->mmap_sem);
960 new_page = khugepaged_alloc_page(hpage, gfp, node);
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700961 if (!new_page) {
962 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
963 goto out_nolock;
964 }
965
966 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
967 result = SCAN_CGROUP_CHARGE_FAIL;
968 goto out_nolock;
969 }
970
971 down_read(&mm->mmap_sem);
Kirill A. Shutemovc131f752016-09-19 14:44:01 -0700972 result = hugepage_vma_revalidate(mm, address, &vma);
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700973 if (result) {
974 mem_cgroup_cancel_charge(new_page, memcg, true);
975 up_read(&mm->mmap_sem);
976 goto out_nolock;
977 }
978
979 pmd = mm_find_pmd(mm, address);
980 if (!pmd) {
981 result = SCAN_PMD_NULL;
982 mem_cgroup_cancel_charge(new_page, memcg, true);
983 up_read(&mm->mmap_sem);
984 goto out_nolock;
985 }
986
987 /*
988 * __collapse_huge_page_swapin always returns with mmap_sem locked.
Ebru Akagunduz47f863e2016-07-26 15:26:43 -0700989 * If it fails, we release mmap_sem and jump out_nolock.
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700990 * Continuing to collapse causes inconsistency.
991 */
Ebru Akagunduz0db501f2016-07-26 15:26:46 -0700992 if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -0700993 mem_cgroup_cancel_charge(new_page, memcg, true);
994 up_read(&mm->mmap_sem);
995 goto out_nolock;
996 }
997
998 up_read(&mm->mmap_sem);
999 /*
1000 * Prevent all access to pagetables with the exception of
1001 * gup_fast later handled by the ptep_clear_flush and the VM
1002 * handled by the anon_vma lock + PG_lock.
1003 */
1004 down_write(&mm->mmap_sem);
Kirill A. Shutemovc131f752016-09-19 14:44:01 -07001005 result = hugepage_vma_revalidate(mm, address, &vma);
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -07001006 if (result)
1007 goto out;
1008 /* check if the pmd is still valid */
1009 if (mm_find_pmd(mm, address) != pmd)
1010 goto out;
1011
1012 anon_vma_lock_write(vma->anon_vma);
1013
1014 pte = pte_offset_map(pmd, address);
1015 pte_ptl = pte_lockptr(mm, pmd);
1016
1017 mmun_start = address;
1018 mmun_end = address + HPAGE_PMD_SIZE;
1019 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1020 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1021 /*
1022 * After this gup_fast can't run anymore. This also removes
1023 * any huge TLB entry from the CPU so we won't allow
1024 * huge and small TLB entries for the same virtual address
1025 * to avoid the risk of CPU bugs in that area.
1026 */
1027 _pmd = pmdp_collapse_flush(vma, address, pmd);
1028 spin_unlock(pmd_ptl);
1029 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1030
1031 spin_lock(pte_ptl);
1032 isolated = __collapse_huge_page_isolate(vma, address, pte);
1033 spin_unlock(pte_ptl);
1034
1035 if (unlikely(!isolated)) {
1036 pte_unmap(pte);
1037 spin_lock(pmd_ptl);
1038 BUG_ON(!pmd_none(*pmd));
1039 /*
1040 * We can only use set_pmd_at when establishing
1041 * hugepmds and never for establishing regular pmds that
1042 * points to regular pagetables. Use pmd_populate for that
1043 */
1044 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1045 spin_unlock(pmd_ptl);
1046 anon_vma_unlock_write(vma->anon_vma);
1047 result = SCAN_FAIL;
1048 goto out;
1049 }
1050
1051 /*
1052 * All pages are isolated and locked so anon_vma rmap
1053 * can't run anymore.
1054 */
1055 anon_vma_unlock_write(vma->anon_vma);
1056
1057 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1058 pte_unmap(pte);
1059 __SetPageUptodate(new_page);
1060 pgtable = pmd_pgtable(_pmd);
1061
1062 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1063 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1064
1065 /*
1066 * spin_lock() below is not the equivalent of smp_wmb(), so
1067 * this is needed to avoid the copy_huge_page writes to become
1068 * visible after the set_pmd_at() write.
1069 */
1070 smp_wmb();
1071
1072 spin_lock(pmd_ptl);
1073 BUG_ON(!pmd_none(*pmd));
1074 page_add_new_anon_rmap(new_page, vma, address, true);
1075 mem_cgroup_commit_charge(new_page, memcg, false, true);
1076 lru_cache_add_active_or_unevictable(new_page, vma);
1077 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1078 set_pmd_at(mm, address, pmd, _pmd);
1079 update_mmu_cache_pmd(vma, address, pmd);
1080 spin_unlock(pmd_ptl);
1081
1082 *hpage = NULL;
1083
1084 khugepaged_pages_collapsed++;
1085 result = SCAN_SUCCEED;
1086out_up_write:
1087 up_write(&mm->mmap_sem);
1088out_nolock:
1089 trace_mm_collapse_huge_page(mm, isolated, result);
1090 return;
1091out:
1092 mem_cgroup_cancel_charge(new_page, memcg, true);
1093 goto out_up_write;
1094}
1095
1096static int khugepaged_scan_pmd(struct mm_struct *mm,
1097 struct vm_area_struct *vma,
1098 unsigned long address,
1099 struct page **hpage)
1100{
1101 pmd_t *pmd;
1102 pte_t *pte, *_pte;
Ebru Akagunduz0db501f2016-07-26 15:26:46 -07001103 int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -07001104 struct page *page = NULL;
1105 unsigned long _address;
1106 spinlock_t *ptl;
1107 int node = NUMA_NO_NODE, unmapped = 0;
Ebru Akagunduz0db501f2016-07-26 15:26:46 -07001108 bool writable = false;
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -07001109
1110 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1111
1112 pmd = mm_find_pmd(mm, address);
1113 if (!pmd) {
1114 result = SCAN_PMD_NULL;
1115 goto out;
1116 }
1117
1118 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1119 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1120 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1121 _pte++, _address += PAGE_SIZE) {
1122 pte_t pteval = *_pte;
1123 if (is_swap_pte(pteval)) {
1124 if (++unmapped <= khugepaged_max_ptes_swap) {
1125 continue;
1126 } else {
1127 result = SCAN_EXCEED_SWAP_PTE;
1128 goto out_unmap;
1129 }
1130 }
1131 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1132 if (!userfaultfd_armed(vma) &&
1133 ++none_or_zero <= khugepaged_max_ptes_none) {
1134 continue;
1135 } else {
1136 result = SCAN_EXCEED_NONE_PTE;
1137 goto out_unmap;
1138 }
1139 }
1140 if (!pte_present(pteval)) {
1141 result = SCAN_PTE_NON_PRESENT;
1142 goto out_unmap;
1143 }
1144 if (pte_write(pteval))
1145 writable = true;
1146
1147 page = vm_normal_page(vma, _address, pteval);
1148 if (unlikely(!page)) {
1149 result = SCAN_PAGE_NULL;
1150 goto out_unmap;
1151 }
1152
1153 /* TODO: teach khugepaged to collapse THP mapped with pte */
1154 if (PageCompound(page)) {
1155 result = SCAN_PAGE_COMPOUND;
1156 goto out_unmap;
1157 }
1158
1159 /*
1160 * Record which node the original page is from and save this
1161 * information to khugepaged_node_load[].
1162 * Khupaged will allocate hugepage from the node has the max
1163 * hit record.
1164 */
1165 node = page_to_nid(page);
1166 if (khugepaged_scan_abort(node)) {
1167 result = SCAN_SCAN_ABORT;
1168 goto out_unmap;
1169 }
1170 khugepaged_node_load[node]++;
1171 if (!PageLRU(page)) {
1172 result = SCAN_PAGE_LRU;
1173 goto out_unmap;
1174 }
1175 if (PageLocked(page)) {
1176 result = SCAN_PAGE_LOCK;
1177 goto out_unmap;
1178 }
1179 if (!PageAnon(page)) {
1180 result = SCAN_PAGE_ANON;
1181 goto out_unmap;
1182 }
1183
1184 /*
1185 * cannot use mapcount: can't collapse if there's a gup pin.
1186 * The page must only be referenced by the scanned process
1187 * and page swap cache.
1188 */
1189 if (page_count(page) != 1 + !!PageSwapCache(page)) {
1190 result = SCAN_PAGE_COUNT;
1191 goto out_unmap;
1192 }
1193 if (pte_young(pteval) ||
1194 page_is_young(page) || PageReferenced(page) ||
1195 mmu_notifier_test_young(vma->vm_mm, address))
Ebru Akagunduz0db501f2016-07-26 15:26:46 -07001196 referenced++;
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -07001197 }
1198 if (writable) {
1199 if (referenced) {
1200 result = SCAN_SUCCEED;
1201 ret = 1;
1202 } else {
Ebru Akagunduz0db501f2016-07-26 15:26:46 -07001203 result = SCAN_LACK_REFERENCED_PAGE;
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -07001204 }
1205 } else {
1206 result = SCAN_PAGE_RO;
1207 }
1208out_unmap:
1209 pte_unmap_unlock(pte, ptl);
1210 if (ret) {
1211 node = khugepaged_find_target_node();
1212 /* collapse_huge_page will return with the mmap_sem released */
Kirill A. Shutemovc131f752016-09-19 14:44:01 -07001213 collapse_huge_page(mm, address, hpage, node, referenced);
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -07001214 }
1215out:
1216 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1217 none_or_zero, result, unmapped);
1218 return ret;
1219}
1220
1221static void collect_mm_slot(struct mm_slot *mm_slot)
1222{
1223 struct mm_struct *mm = mm_slot->mm;
1224
1225 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1226
1227 if (khugepaged_test_exit(mm)) {
1228 /* free mm_slot */
1229 hash_del(&mm_slot->hash);
1230 list_del(&mm_slot->mm_node);
1231
1232 /*
1233 * Not strictly needed because the mm exited already.
1234 *
1235 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1236 */
1237
1238 /* khugepaged_mm_lock actually not necessary for the below */
1239 free_mm_slot(mm_slot);
1240 mmdrop(mm);
1241 }
1242}
1243
Kirill A. Shutemove496cf32016-07-26 15:26:35 -07001244#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -07001245static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1246{
1247 struct vm_area_struct *vma;
1248 unsigned long addr;
1249 pmd_t *pmd, _pmd;
1250
1251 i_mmap_lock_write(mapping);
1252 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1253 /* probably overkill */
1254 if (vma->anon_vma)
1255 continue;
1256 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1257 if (addr & ~HPAGE_PMD_MASK)
1258 continue;
1259 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1260 continue;
1261 pmd = mm_find_pmd(vma->vm_mm, addr);
1262 if (!pmd)
1263 continue;
1264 /*
1265 * We need exclusive mmap_sem to retract page table.
1266 * If trylock fails we would end up with pte-mapped THP after
1267 * re-fault. Not ideal, but it's more important to not disturb
1268 * the system too much.
1269 */
1270 if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1271 spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1272 /* assume page table is clear */
1273 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1274 spin_unlock(ptl);
1275 up_write(&vma->vm_mm->mmap_sem);
1276 atomic_long_dec(&vma->vm_mm->nr_ptes);
1277 pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1278 }
1279 }
1280 i_mmap_unlock_write(mapping);
1281}
1282
1283/**
1284 * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1285 *
1286 * Basic scheme is simple, details are more complex:
1287 * - allocate and freeze a new huge page;
1288 * - scan over radix tree replacing old pages the new one
1289 * + swap in pages if necessary;
1290 * + fill in gaps;
1291 * + keep old pages around in case if rollback is required;
1292 * - if replacing succeed:
1293 * + copy data over;
1294 * + free old pages;
1295 * + unfreeze huge page;
1296 * - if replacing failed;
1297 * + put all pages back and unfreeze them;
1298 * + restore gaps in the radix-tree;
1299 * + free huge page;
1300 */
1301static void collapse_shmem(struct mm_struct *mm,
1302 struct address_space *mapping, pgoff_t start,
1303 struct page **hpage, int node)
1304{
1305 gfp_t gfp;
1306 struct page *page, *new_page, *tmp;
1307 struct mem_cgroup *memcg;
1308 pgoff_t index, end = start + HPAGE_PMD_NR;
1309 LIST_HEAD(pagelist);
1310 struct radix_tree_iter iter;
1311 void **slot;
1312 int nr_none = 0, result = SCAN_SUCCEED;
1313
1314 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1315
1316 /* Only allocate from the target node */
1317 gfp = alloc_hugepage_khugepaged_gfpmask() |
1318 __GFP_OTHER_NODE | __GFP_THISNODE;
1319
1320 new_page = khugepaged_alloc_page(hpage, gfp, node);
1321 if (!new_page) {
1322 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1323 goto out;
1324 }
1325
1326 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1327 result = SCAN_CGROUP_CHARGE_FAIL;
1328 goto out;
1329 }
1330
1331 new_page->index = start;
1332 new_page->mapping = mapping;
1333 __SetPageSwapBacked(new_page);
1334 __SetPageLocked(new_page);
1335 BUG_ON(!page_ref_freeze(new_page, 1));
1336
1337
1338 /*
1339 * At this point the new_page is 'frozen' (page_count() is zero), locked
1340 * and not up-to-date. It's safe to insert it into radix tree, because
1341 * nobody would be able to map it or use it in other way until we
1342 * unfreeze it.
1343 */
1344
1345 index = start;
1346 spin_lock_irq(&mapping->tree_lock);
1347 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1348 int n = min(iter.index, end) - index;
1349
1350 /*
1351 * Handle holes in the radix tree: charge it from shmem and
1352 * insert relevant subpage of new_page into the radix-tree.
1353 */
1354 if (n && !shmem_charge(mapping->host, n)) {
1355 result = SCAN_FAIL;
1356 break;
1357 }
1358 nr_none += n;
1359 for (; index < min(iter.index, end); index++) {
1360 radix_tree_insert(&mapping->page_tree, index,
1361 new_page + (index % HPAGE_PMD_NR));
1362 }
1363
1364 /* We are done. */
1365 if (index >= end)
1366 break;
1367
1368 page = radix_tree_deref_slot_protected(slot,
1369 &mapping->tree_lock);
1370 if (radix_tree_exceptional_entry(page) || !PageUptodate(page)) {
1371 spin_unlock_irq(&mapping->tree_lock);
1372 /* swap in or instantiate fallocated page */
1373 if (shmem_getpage(mapping->host, index, &page,
1374 SGP_NOHUGE)) {
1375 result = SCAN_FAIL;
1376 goto tree_unlocked;
1377 }
1378 spin_lock_irq(&mapping->tree_lock);
1379 } else if (trylock_page(page)) {
1380 get_page(page);
1381 } else {
1382 result = SCAN_PAGE_LOCK;
1383 break;
1384 }
1385
1386 /*
1387 * The page must be locked, so we can drop the tree_lock
1388 * without racing with truncate.
1389 */
1390 VM_BUG_ON_PAGE(!PageLocked(page), page);
1391 VM_BUG_ON_PAGE(!PageUptodate(page), page);
1392 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1393
1394 if (page_mapping(page) != mapping) {
1395 result = SCAN_TRUNCATED;
1396 goto out_unlock;
1397 }
1398 spin_unlock_irq(&mapping->tree_lock);
1399
1400 if (isolate_lru_page(page)) {
1401 result = SCAN_DEL_PAGE_LRU;
1402 goto out_isolate_failed;
1403 }
1404
1405 if (page_mapped(page))
1406 unmap_mapping_range(mapping, index << PAGE_SHIFT,
1407 PAGE_SIZE, 0);
1408
1409 spin_lock_irq(&mapping->tree_lock);
1410
Johannes Weiner058a4a52016-12-12 16:43:32 -08001411 slot = radix_tree_lookup_slot(&mapping->page_tree, index);
1412 VM_BUG_ON_PAGE(page != radix_tree_deref_slot_protected(slot,
1413 &mapping->tree_lock), page);
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -07001414 VM_BUG_ON_PAGE(page_mapped(page), page);
1415
1416 /*
1417 * The page is expected to have page_count() == 3:
1418 * - we hold a pin on it;
1419 * - one reference from radix tree;
1420 * - one from isolate_lru_page;
1421 */
1422 if (!page_ref_freeze(page, 3)) {
1423 result = SCAN_PAGE_COUNT;
1424 goto out_lru;
1425 }
1426
1427 /*
1428 * Add the page to the list to be able to undo the collapse if
1429 * something go wrong.
1430 */
1431 list_add_tail(&page->lru, &pagelist);
1432
1433 /* Finally, replace with the new page. */
1434 radix_tree_replace_slot(slot,
1435 new_page + (index % HPAGE_PMD_NR));
1436
Johannes Weiner058a4a52016-12-12 16:43:32 -08001437 slot = radix_tree_iter_next(&iter);
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -07001438 index++;
1439 continue;
1440out_lru:
1441 spin_unlock_irq(&mapping->tree_lock);
1442 putback_lru_page(page);
1443out_isolate_failed:
1444 unlock_page(page);
1445 put_page(page);
1446 goto tree_unlocked;
1447out_unlock:
1448 unlock_page(page);
1449 put_page(page);
1450 break;
1451 }
1452
1453 /*
1454 * Handle hole in radix tree at the end of the range.
1455 * This code only triggers if there's nothing in radix tree
1456 * beyond 'end'.
1457 */
1458 if (result == SCAN_SUCCEED && index < end) {
1459 int n = end - index;
1460
1461 if (!shmem_charge(mapping->host, n)) {
1462 result = SCAN_FAIL;
1463 goto tree_locked;
1464 }
1465
1466 for (; index < end; index++) {
1467 radix_tree_insert(&mapping->page_tree, index,
1468 new_page + (index % HPAGE_PMD_NR));
1469 }
1470 nr_none += n;
1471 }
1472
1473tree_locked:
1474 spin_unlock_irq(&mapping->tree_lock);
1475tree_unlocked:
1476
1477 if (result == SCAN_SUCCEED) {
1478 unsigned long flags;
1479 struct zone *zone = page_zone(new_page);
1480
1481 /*
1482 * Replacing old pages with new one has succeed, now we need to
1483 * copy the content and free old pages.
1484 */
1485 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1486 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1487 page);
1488 list_del(&page->lru);
1489 unlock_page(page);
1490 page_ref_unfreeze(page, 1);
1491 page->mapping = NULL;
1492 ClearPageActive(page);
1493 ClearPageUnevictable(page);
1494 put_page(page);
1495 }
1496
1497 local_irq_save(flags);
Mel Gorman11fb9982016-07-28 15:46:20 -07001498 __inc_node_page_state(new_page, NR_SHMEM_THPS);
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -07001499 if (nr_none) {
Mel Gorman11fb9982016-07-28 15:46:20 -07001500 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1501 __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -07001502 }
1503 local_irq_restore(flags);
1504
1505 /*
1506 * Remove pte page tables, so we can re-faulti
1507 * the page as huge.
1508 */
1509 retract_page_tables(mapping, start);
1510
1511 /* Everything is ready, let's unfreeze the new_page */
1512 set_page_dirty(new_page);
1513 SetPageUptodate(new_page);
1514 page_ref_unfreeze(new_page, HPAGE_PMD_NR);
1515 mem_cgroup_commit_charge(new_page, memcg, false, true);
1516 lru_cache_add_anon(new_page);
1517 unlock_page(new_page);
1518
1519 *hpage = NULL;
1520 } else {
1521 /* Something went wrong: rollback changes to the radix-tree */
1522 shmem_uncharge(mapping->host, nr_none);
1523 spin_lock_irq(&mapping->tree_lock);
1524 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
1525 start) {
1526 if (iter.index >= end)
1527 break;
1528 page = list_first_entry_or_null(&pagelist,
1529 struct page, lru);
1530 if (!page || iter.index < page->index) {
1531 if (!nr_none)
1532 break;
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -07001533 nr_none--;
Johannes Weinerdc1b6d02016-12-12 16:43:35 -08001534 /* Put holes back where they were */
1535 radix_tree_delete(&mapping->page_tree,
1536 iter.index);
1537 slot = radix_tree_iter_next(&iter);
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -07001538 continue;
1539 }
1540
1541 VM_BUG_ON_PAGE(page->index != iter.index, page);
1542
1543 /* Unfreeze the page. */
1544 list_del(&page->lru);
1545 page_ref_unfreeze(page, 2);
1546 radix_tree_replace_slot(slot, page);
1547 spin_unlock_irq(&mapping->tree_lock);
1548 putback_lru_page(page);
1549 unlock_page(page);
1550 spin_lock_irq(&mapping->tree_lock);
Johannes Weiner058a4a52016-12-12 16:43:32 -08001551 slot = radix_tree_iter_next(&iter);
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -07001552 }
1553 VM_BUG_ON(nr_none);
1554 spin_unlock_irq(&mapping->tree_lock);
1555
1556 /* Unfreeze new_page, caller would take care about freeing it */
1557 page_ref_unfreeze(new_page, 1);
1558 mem_cgroup_cancel_charge(new_page, memcg, true);
1559 unlock_page(new_page);
1560 new_page->mapping = NULL;
1561 }
1562out:
1563 VM_BUG_ON(!list_empty(&pagelist));
1564 /* TODO: tracepoints */
1565}
1566
1567static void khugepaged_scan_shmem(struct mm_struct *mm,
1568 struct address_space *mapping,
1569 pgoff_t start, struct page **hpage)
1570{
1571 struct page *page = NULL;
1572 struct radix_tree_iter iter;
1573 void **slot;
1574 int present, swap;
1575 int node = NUMA_NO_NODE;
1576 int result = SCAN_SUCCEED;
1577
1578 present = 0;
1579 swap = 0;
1580 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1581 rcu_read_lock();
1582 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1583 if (iter.index >= start + HPAGE_PMD_NR)
1584 break;
1585
1586 page = radix_tree_deref_slot(slot);
1587 if (radix_tree_deref_retry(page)) {
1588 slot = radix_tree_iter_retry(&iter);
1589 continue;
1590 }
1591
1592 if (radix_tree_exception(page)) {
1593 if (++swap > khugepaged_max_ptes_swap) {
1594 result = SCAN_EXCEED_SWAP_PTE;
1595 break;
1596 }
1597 continue;
1598 }
1599
1600 if (PageTransCompound(page)) {
1601 result = SCAN_PAGE_COMPOUND;
1602 break;
1603 }
1604
1605 node = page_to_nid(page);
1606 if (khugepaged_scan_abort(node)) {
1607 result = SCAN_SCAN_ABORT;
1608 break;
1609 }
1610 khugepaged_node_load[node]++;
1611
1612 if (!PageLRU(page)) {
1613 result = SCAN_PAGE_LRU;
1614 break;
1615 }
1616
1617 if (page_count(page) != 1 + page_mapcount(page)) {
1618 result = SCAN_PAGE_COUNT;
1619 break;
1620 }
1621
1622 /*
1623 * We probably should check if the page is referenced here, but
1624 * nobody would transfer pte_young() to PageReferenced() for us.
1625 * And rmap walk here is just too costly...
1626 */
1627
1628 present++;
1629
1630 if (need_resched()) {
1631 cond_resched_rcu();
1632 slot = radix_tree_iter_next(&iter);
1633 }
1634 }
1635 rcu_read_unlock();
1636
1637 if (result == SCAN_SUCCEED) {
1638 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1639 result = SCAN_EXCEED_NONE_PTE;
1640 } else {
1641 node = khugepaged_find_target_node();
1642 collapse_shmem(mm, mapping, start, hpage, node);
1643 }
1644 }
1645
1646 /* TODO: tracepoints */
1647}
1648#else
1649static void khugepaged_scan_shmem(struct mm_struct *mm,
1650 struct address_space *mapping,
1651 pgoff_t start, struct page **hpage)
1652{
1653 BUILD_BUG();
1654}
1655#endif
1656
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -07001657static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1658 struct page **hpage)
1659 __releases(&khugepaged_mm_lock)
1660 __acquires(&khugepaged_mm_lock)
1661{
1662 struct mm_slot *mm_slot;
1663 struct mm_struct *mm;
1664 struct vm_area_struct *vma;
1665 int progress = 0;
1666
1667 VM_BUG_ON(!pages);
1668 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1669
1670 if (khugepaged_scan.mm_slot)
1671 mm_slot = khugepaged_scan.mm_slot;
1672 else {
1673 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1674 struct mm_slot, mm_node);
1675 khugepaged_scan.address = 0;
1676 khugepaged_scan.mm_slot = mm_slot;
1677 }
1678 spin_unlock(&khugepaged_mm_lock);
1679
1680 mm = mm_slot->mm;
1681 down_read(&mm->mmap_sem);
1682 if (unlikely(khugepaged_test_exit(mm)))
1683 vma = NULL;
1684 else
1685 vma = find_vma(mm, khugepaged_scan.address);
1686
1687 progress++;
1688 for (; vma; vma = vma->vm_next) {
1689 unsigned long hstart, hend;
1690
1691 cond_resched();
1692 if (unlikely(khugepaged_test_exit(mm))) {
1693 progress++;
1694 break;
1695 }
1696 if (!hugepage_vma_check(vma)) {
1697skip:
1698 progress++;
1699 continue;
1700 }
1701 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1702 hend = vma->vm_end & HPAGE_PMD_MASK;
1703 if (hstart >= hend)
1704 goto skip;
1705 if (khugepaged_scan.address > hend)
1706 goto skip;
1707 if (khugepaged_scan.address < hstart)
1708 khugepaged_scan.address = hstart;
1709 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1710
1711 while (khugepaged_scan.address < hend) {
1712 int ret;
1713 cond_resched();
1714 if (unlikely(khugepaged_test_exit(mm)))
1715 goto breakouterloop;
1716
1717 VM_BUG_ON(khugepaged_scan.address < hstart ||
1718 khugepaged_scan.address + HPAGE_PMD_SIZE >
1719 hend);
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -07001720 if (shmem_file(vma->vm_file)) {
Kirill A. Shutemove496cf32016-07-26 15:26:35 -07001721 struct file *file;
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -07001722 pgoff_t pgoff = linear_page_index(vma,
1723 khugepaged_scan.address);
Kirill A. Shutemove496cf32016-07-26 15:26:35 -07001724 if (!shmem_huge_enabled(vma))
1725 goto skip;
1726 file = get_file(vma->vm_file);
Kirill A. Shutemovf3f0e1d2016-07-26 15:26:32 -07001727 up_read(&mm->mmap_sem);
1728 ret = 1;
1729 khugepaged_scan_shmem(mm, file->f_mapping,
1730 pgoff, hpage);
1731 fput(file);
1732 } else {
1733 ret = khugepaged_scan_pmd(mm, vma,
1734 khugepaged_scan.address,
1735 hpage);
1736 }
Kirill A. Shutemovb46e7562016-07-26 15:26:24 -07001737 /* move to next address */
1738 khugepaged_scan.address += HPAGE_PMD_SIZE;
1739 progress += HPAGE_PMD_NR;
1740 if (ret)
1741 /* we released mmap_sem so break loop */
1742 goto breakouterloop_mmap_sem;
1743 if (progress >= pages)
1744 goto breakouterloop;
1745 }
1746 }
1747breakouterloop:
1748 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1749breakouterloop_mmap_sem:
1750
1751 spin_lock(&khugepaged_mm_lock);
1752 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1753 /*
1754 * Release the current mm_slot if this mm is about to die, or
1755 * if we scanned all vmas of this mm.
1756 */
1757 if (khugepaged_test_exit(mm) || !vma) {
1758 /*
1759 * Make sure that if mm_users is reaching zero while
1760 * khugepaged runs here, khugepaged_exit will find
1761 * mm_slot not pointing to the exiting mm.
1762 */
1763 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1764 khugepaged_scan.mm_slot = list_entry(
1765 mm_slot->mm_node.next,
1766 struct mm_slot, mm_node);
1767 khugepaged_scan.address = 0;
1768 } else {
1769 khugepaged_scan.mm_slot = NULL;
1770 khugepaged_full_scans++;
1771 }
1772
1773 collect_mm_slot(mm_slot);
1774 }
1775
1776 return progress;
1777}
1778
1779static int khugepaged_has_work(void)
1780{
1781 return !list_empty(&khugepaged_scan.mm_head) &&
1782 khugepaged_enabled();
1783}
1784
1785static int khugepaged_wait_event(void)
1786{
1787 return !list_empty(&khugepaged_scan.mm_head) ||
1788 kthread_should_stop();
1789}
1790
1791static void khugepaged_do_scan(void)
1792{
1793 struct page *hpage = NULL;
1794 unsigned int progress = 0, pass_through_head = 0;
1795 unsigned int pages = khugepaged_pages_to_scan;
1796 bool wait = true;
1797
1798 barrier(); /* write khugepaged_pages_to_scan to local stack */
1799
1800 while (progress < pages) {
1801 if (!khugepaged_prealloc_page(&hpage, &wait))
1802 break;
1803
1804 cond_resched();
1805
1806 if (unlikely(kthread_should_stop() || try_to_freeze()))
1807 break;
1808
1809 spin_lock(&khugepaged_mm_lock);
1810 if (!khugepaged_scan.mm_slot)
1811 pass_through_head++;
1812 if (khugepaged_has_work() &&
1813 pass_through_head < 2)
1814 progress += khugepaged_scan_mm_slot(pages - progress,
1815 &hpage);
1816 else
1817 progress = pages;
1818 spin_unlock(&khugepaged_mm_lock);
1819 }
1820
1821 if (!IS_ERR_OR_NULL(hpage))
1822 put_page(hpage);
1823}
1824
1825static bool khugepaged_should_wakeup(void)
1826{
1827 return kthread_should_stop() ||
1828 time_after_eq(jiffies, khugepaged_sleep_expire);
1829}
1830
1831static void khugepaged_wait_work(void)
1832{
1833 if (khugepaged_has_work()) {
1834 const unsigned long scan_sleep_jiffies =
1835 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1836
1837 if (!scan_sleep_jiffies)
1838 return;
1839
1840 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1841 wait_event_freezable_timeout(khugepaged_wait,
1842 khugepaged_should_wakeup(),
1843 scan_sleep_jiffies);
1844 return;
1845 }
1846
1847 if (khugepaged_enabled())
1848 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1849}
1850
1851static int khugepaged(void *none)
1852{
1853 struct mm_slot *mm_slot;
1854
1855 set_freezable();
1856 set_user_nice(current, MAX_NICE);
1857
1858 while (!kthread_should_stop()) {
1859 khugepaged_do_scan();
1860 khugepaged_wait_work();
1861 }
1862
1863 spin_lock(&khugepaged_mm_lock);
1864 mm_slot = khugepaged_scan.mm_slot;
1865 khugepaged_scan.mm_slot = NULL;
1866 if (mm_slot)
1867 collect_mm_slot(mm_slot);
1868 spin_unlock(&khugepaged_mm_lock);
1869 return 0;
1870}
1871
1872static void set_recommended_min_free_kbytes(void)
1873{
1874 struct zone *zone;
1875 int nr_zones = 0;
1876 unsigned long recommended_min;
1877
1878 for_each_populated_zone(zone)
1879 nr_zones++;
1880
1881 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1882 recommended_min = pageblock_nr_pages * nr_zones * 2;
1883
1884 /*
1885 * Make sure that on average at least two pageblocks are almost free
1886 * of another type, one for a migratetype to fall back to and a
1887 * second to avoid subsequent fallbacks of other types There are 3
1888 * MIGRATE_TYPES we care about.
1889 */
1890 recommended_min += pageblock_nr_pages * nr_zones *
1891 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1892
1893 /* don't ever allow to reserve more than 5% of the lowmem */
1894 recommended_min = min(recommended_min,
1895 (unsigned long) nr_free_buffer_pages() / 20);
1896 recommended_min <<= (PAGE_SHIFT-10);
1897
1898 if (recommended_min > min_free_kbytes) {
1899 if (user_min_free_kbytes >= 0)
1900 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1901 min_free_kbytes, recommended_min);
1902
1903 min_free_kbytes = recommended_min;
1904 }
1905 setup_per_zone_wmarks();
1906}
1907
1908int start_stop_khugepaged(void)
1909{
1910 static struct task_struct *khugepaged_thread __read_mostly;
1911 static DEFINE_MUTEX(khugepaged_mutex);
1912 int err = 0;
1913
1914 mutex_lock(&khugepaged_mutex);
1915 if (khugepaged_enabled()) {
1916 if (!khugepaged_thread)
1917 khugepaged_thread = kthread_run(khugepaged, NULL,
1918 "khugepaged");
1919 if (IS_ERR(khugepaged_thread)) {
1920 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1921 err = PTR_ERR(khugepaged_thread);
1922 khugepaged_thread = NULL;
1923 goto fail;
1924 }
1925
1926 if (!list_empty(&khugepaged_scan.mm_head))
1927 wake_up_interruptible(&khugepaged_wait);
1928
1929 set_recommended_min_free_kbytes();
1930 } else if (khugepaged_thread) {
1931 kthread_stop(khugepaged_thread);
1932 khugepaged_thread = NULL;
1933 }
1934fail:
1935 mutex_unlock(&khugepaged_mutex);
1936 return err;
1937}