]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - mm/khugepaged.c
Merge tag 'pm-urgent-4.8-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael...
[mirror_ubuntu-artful-kernel.git] / mm / khugepaged.c
1 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
2
3 #include <linux/mm.h>
4 #include <linux/sched.h>
5 #include <linux/mmu_notifier.h>
6 #include <linux/rmap.h>
7 #include <linux/swap.h>
8 #include <linux/mm_inline.h>
9 #include <linux/kthread.h>
10 #include <linux/khugepaged.h>
11 #include <linux/freezer.h>
12 #include <linux/mman.h>
13 #include <linux/hashtable.h>
14 #include <linux/userfaultfd_k.h>
15 #include <linux/page_idle.h>
16 #include <linux/swapops.h>
17 #include <linux/shmem_fs.h>
18
19 #include <asm/tlb.h>
20 #include <asm/pgalloc.h>
21 #include "internal.h"
22
23 enum scan_result {
24 SCAN_FAIL,
25 SCAN_SUCCEED,
26 SCAN_PMD_NULL,
27 SCAN_EXCEED_NONE_PTE,
28 SCAN_PTE_NON_PRESENT,
29 SCAN_PAGE_RO,
30 SCAN_LACK_REFERENCED_PAGE,
31 SCAN_PAGE_NULL,
32 SCAN_SCAN_ABORT,
33 SCAN_PAGE_COUNT,
34 SCAN_PAGE_LRU,
35 SCAN_PAGE_LOCK,
36 SCAN_PAGE_ANON,
37 SCAN_PAGE_COMPOUND,
38 SCAN_ANY_PROCESS,
39 SCAN_VMA_NULL,
40 SCAN_VMA_CHECK,
41 SCAN_ADDRESS_RANGE,
42 SCAN_SWAP_CACHE_PAGE,
43 SCAN_DEL_PAGE_LRU,
44 SCAN_ALLOC_HUGE_PAGE_FAIL,
45 SCAN_CGROUP_CHARGE_FAIL,
46 SCAN_EXCEED_SWAP_PTE,
47 SCAN_TRUNCATED,
48 };
49
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/huge_memory.h>
52
53 /* default scan 8*512 pte (or vmas) every 30 second */
54 static unsigned int khugepaged_pages_to_scan __read_mostly;
55 static unsigned int khugepaged_pages_collapsed;
56 static unsigned int khugepaged_full_scans;
57 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
58 /* during fragmentation poll the hugepage allocator once every minute */
59 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
60 static unsigned long khugepaged_sleep_expire;
61 static DEFINE_SPINLOCK(khugepaged_mm_lock);
62 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
63 /*
64 * default collapse hugepages if there is at least one pte mapped like
65 * it would have happened if the vma was large enough during page
66 * fault.
67 */
68 static unsigned int khugepaged_max_ptes_none __read_mostly;
69 static unsigned int khugepaged_max_ptes_swap __read_mostly;
70
71 #define MM_SLOTS_HASH_BITS 10
72 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
73
74 static struct kmem_cache *mm_slot_cache __read_mostly;
75
76 /**
77 * struct mm_slot - hash lookup from mm to mm_slot
78 * @hash: hash collision list
79 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
80 * @mm: the mm that this information is valid for
81 */
82 struct mm_slot {
83 struct hlist_node hash;
84 struct list_head mm_node;
85 struct mm_struct *mm;
86 };
87
88 /**
89 * struct khugepaged_scan - cursor for scanning
90 * @mm_head: the head of the mm list to scan
91 * @mm_slot: the current mm_slot we are scanning
92 * @address: the next address inside that to be scanned
93 *
94 * There is only the one khugepaged_scan instance of this cursor structure.
95 */
96 struct khugepaged_scan {
97 struct list_head mm_head;
98 struct mm_slot *mm_slot;
99 unsigned long address;
100 };
101
102 static struct khugepaged_scan khugepaged_scan = {
103 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
104 };
105
106 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
107 struct kobj_attribute *attr,
108 char *buf)
109 {
110 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
111 }
112
113 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
114 struct kobj_attribute *attr,
115 const char *buf, size_t count)
116 {
117 unsigned long msecs;
118 int err;
119
120 err = kstrtoul(buf, 10, &msecs);
121 if (err || msecs > UINT_MAX)
122 return -EINVAL;
123
124 khugepaged_scan_sleep_millisecs = msecs;
125 khugepaged_sleep_expire = 0;
126 wake_up_interruptible(&khugepaged_wait);
127
128 return count;
129 }
130 static struct kobj_attribute scan_sleep_millisecs_attr =
131 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
132 scan_sleep_millisecs_store);
133
134 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
135 struct kobj_attribute *attr,
136 char *buf)
137 {
138 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
139 }
140
141 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
142 struct kobj_attribute *attr,
143 const char *buf, size_t count)
144 {
145 unsigned long msecs;
146 int err;
147
148 err = kstrtoul(buf, 10, &msecs);
149 if (err || msecs > UINT_MAX)
150 return -EINVAL;
151
152 khugepaged_alloc_sleep_millisecs = msecs;
153 khugepaged_sleep_expire = 0;
154 wake_up_interruptible(&khugepaged_wait);
155
156 return count;
157 }
158 static struct kobj_attribute alloc_sleep_millisecs_attr =
159 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
160 alloc_sleep_millisecs_store);
161
162 static ssize_t pages_to_scan_show(struct kobject *kobj,
163 struct kobj_attribute *attr,
164 char *buf)
165 {
166 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
167 }
168 static ssize_t pages_to_scan_store(struct kobject *kobj,
169 struct kobj_attribute *attr,
170 const char *buf, size_t count)
171 {
172 int err;
173 unsigned long pages;
174
175 err = kstrtoul(buf, 10, &pages);
176 if (err || !pages || pages > UINT_MAX)
177 return -EINVAL;
178
179 khugepaged_pages_to_scan = pages;
180
181 return count;
182 }
183 static struct kobj_attribute pages_to_scan_attr =
184 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
185 pages_to_scan_store);
186
187 static ssize_t pages_collapsed_show(struct kobject *kobj,
188 struct kobj_attribute *attr,
189 char *buf)
190 {
191 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
192 }
193 static struct kobj_attribute pages_collapsed_attr =
194 __ATTR_RO(pages_collapsed);
195
196 static ssize_t full_scans_show(struct kobject *kobj,
197 struct kobj_attribute *attr,
198 char *buf)
199 {
200 return sprintf(buf, "%u\n", khugepaged_full_scans);
201 }
202 static struct kobj_attribute full_scans_attr =
203 __ATTR_RO(full_scans);
204
205 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
206 struct kobj_attribute *attr, char *buf)
207 {
208 return single_hugepage_flag_show(kobj, attr, buf,
209 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
210 }
211 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
212 struct kobj_attribute *attr,
213 const char *buf, size_t count)
214 {
215 return single_hugepage_flag_store(kobj, attr, buf, count,
216 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
217 }
218 static struct kobj_attribute khugepaged_defrag_attr =
219 __ATTR(defrag, 0644, khugepaged_defrag_show,
220 khugepaged_defrag_store);
221
222 /*
223 * max_ptes_none controls if khugepaged should collapse hugepages over
224 * any unmapped ptes in turn potentially increasing the memory
225 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
226 * reduce the available free memory in the system as it
227 * runs. Increasing max_ptes_none will instead potentially reduce the
228 * free memory in the system during the khugepaged scan.
229 */
230 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
231 struct kobj_attribute *attr,
232 char *buf)
233 {
234 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
235 }
236 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
237 struct kobj_attribute *attr,
238 const char *buf, size_t count)
239 {
240 int err;
241 unsigned long max_ptes_none;
242
243 err = kstrtoul(buf, 10, &max_ptes_none);
244 if (err || max_ptes_none > HPAGE_PMD_NR-1)
245 return -EINVAL;
246
247 khugepaged_max_ptes_none = max_ptes_none;
248
249 return count;
250 }
251 static struct kobj_attribute khugepaged_max_ptes_none_attr =
252 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
253 khugepaged_max_ptes_none_store);
254
255 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
256 struct kobj_attribute *attr,
257 char *buf)
258 {
259 return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
260 }
261
262 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
263 struct kobj_attribute *attr,
264 const char *buf, size_t count)
265 {
266 int err;
267 unsigned long max_ptes_swap;
268
269 err = kstrtoul(buf, 10, &max_ptes_swap);
270 if (err || max_ptes_swap > HPAGE_PMD_NR-1)
271 return -EINVAL;
272
273 khugepaged_max_ptes_swap = max_ptes_swap;
274
275 return count;
276 }
277
278 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
279 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
280 khugepaged_max_ptes_swap_store);
281
282 static struct attribute *khugepaged_attr[] = {
283 &khugepaged_defrag_attr.attr,
284 &khugepaged_max_ptes_none_attr.attr,
285 &pages_to_scan_attr.attr,
286 &pages_collapsed_attr.attr,
287 &full_scans_attr.attr,
288 &scan_sleep_millisecs_attr.attr,
289 &alloc_sleep_millisecs_attr.attr,
290 &khugepaged_max_ptes_swap_attr.attr,
291 NULL,
292 };
293
294 struct attribute_group khugepaged_attr_group = {
295 .attrs = khugepaged_attr,
296 .name = "khugepaged",
297 };
298
299 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
300
301 int hugepage_madvise(struct vm_area_struct *vma,
302 unsigned long *vm_flags, int advice)
303 {
304 switch (advice) {
305 case MADV_HUGEPAGE:
306 #ifdef CONFIG_S390
307 /*
308 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
309 * can't handle this properly after s390_enable_sie, so we simply
310 * ignore the madvise to prevent qemu from causing a SIGSEGV.
311 */
312 if (mm_has_pgste(vma->vm_mm))
313 return 0;
314 #endif
315 *vm_flags &= ~VM_NOHUGEPAGE;
316 *vm_flags |= VM_HUGEPAGE;
317 /*
318 * If the vma become good for khugepaged to scan,
319 * register it here without waiting a page fault that
320 * may not happen any time soon.
321 */
322 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
323 khugepaged_enter_vma_merge(vma, *vm_flags))
324 return -ENOMEM;
325 break;
326 case MADV_NOHUGEPAGE:
327 *vm_flags &= ~VM_HUGEPAGE;
328 *vm_flags |= VM_NOHUGEPAGE;
329 /*
330 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
331 * this vma even if we leave the mm registered in khugepaged if
332 * it got registered before VM_NOHUGEPAGE was set.
333 */
334 break;
335 }
336
337 return 0;
338 }
339
340 int __init khugepaged_init(void)
341 {
342 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
343 sizeof(struct mm_slot),
344 __alignof__(struct mm_slot), 0, NULL);
345 if (!mm_slot_cache)
346 return -ENOMEM;
347
348 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
349 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
350 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
351
352 return 0;
353 }
354
355 void __init khugepaged_destroy(void)
356 {
357 kmem_cache_destroy(mm_slot_cache);
358 }
359
360 static inline struct mm_slot *alloc_mm_slot(void)
361 {
362 if (!mm_slot_cache) /* initialization failed */
363 return NULL;
364 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
365 }
366
367 static inline void free_mm_slot(struct mm_slot *mm_slot)
368 {
369 kmem_cache_free(mm_slot_cache, mm_slot);
370 }
371
372 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
373 {
374 struct mm_slot *mm_slot;
375
376 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
377 if (mm == mm_slot->mm)
378 return mm_slot;
379
380 return NULL;
381 }
382
383 static void insert_to_mm_slots_hash(struct mm_struct *mm,
384 struct mm_slot *mm_slot)
385 {
386 mm_slot->mm = mm;
387 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
388 }
389
390 static inline int khugepaged_test_exit(struct mm_struct *mm)
391 {
392 return atomic_read(&mm->mm_users) == 0;
393 }
394
395 int __khugepaged_enter(struct mm_struct *mm)
396 {
397 struct mm_slot *mm_slot;
398 int wakeup;
399
400 mm_slot = alloc_mm_slot();
401 if (!mm_slot)
402 return -ENOMEM;
403
404 /* __khugepaged_exit() must not run from under us */
405 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
406 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
407 free_mm_slot(mm_slot);
408 return 0;
409 }
410
411 spin_lock(&khugepaged_mm_lock);
412 insert_to_mm_slots_hash(mm, mm_slot);
413 /*
414 * Insert just behind the scanning cursor, to let the area settle
415 * down a little.
416 */
417 wakeup = list_empty(&khugepaged_scan.mm_head);
418 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
419 spin_unlock(&khugepaged_mm_lock);
420
421 atomic_inc(&mm->mm_count);
422 if (wakeup)
423 wake_up_interruptible(&khugepaged_wait);
424
425 return 0;
426 }
427
428 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
429 unsigned long vm_flags)
430 {
431 unsigned long hstart, hend;
432 if (!vma->anon_vma)
433 /*
434 * Not yet faulted in so we will register later in the
435 * page fault if needed.
436 */
437 return 0;
438 if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
439 /* khugepaged not yet working on file or special mappings */
440 return 0;
441 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
442 hend = vma->vm_end & HPAGE_PMD_MASK;
443 if (hstart < hend)
444 return khugepaged_enter(vma, vm_flags);
445 return 0;
446 }
447
448 void __khugepaged_exit(struct mm_struct *mm)
449 {
450 struct mm_slot *mm_slot;
451 int free = 0;
452
453 spin_lock(&khugepaged_mm_lock);
454 mm_slot = get_mm_slot(mm);
455 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
456 hash_del(&mm_slot->hash);
457 list_del(&mm_slot->mm_node);
458 free = 1;
459 }
460 spin_unlock(&khugepaged_mm_lock);
461
462 if (free) {
463 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
464 free_mm_slot(mm_slot);
465 mmdrop(mm);
466 } else if (mm_slot) {
467 /*
468 * This is required to serialize against
469 * khugepaged_test_exit() (which is guaranteed to run
470 * under mmap sem read mode). Stop here (after we
471 * return all pagetables will be destroyed) until
472 * khugepaged has finished working on the pagetables
473 * under the mmap_sem.
474 */
475 down_write(&mm->mmap_sem);
476 up_write(&mm->mmap_sem);
477 }
478 }
479
480 static void release_pte_page(struct page *page)
481 {
482 /* 0 stands for page_is_file_cache(page) == false */
483 dec_node_page_state(page, NR_ISOLATED_ANON + 0);
484 unlock_page(page);
485 putback_lru_page(page);
486 }
487
488 static void release_pte_pages(pte_t *pte, pte_t *_pte)
489 {
490 while (--_pte >= pte) {
491 pte_t pteval = *_pte;
492 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
493 release_pte_page(pte_page(pteval));
494 }
495 }
496
497 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
498 unsigned long address,
499 pte_t *pte)
500 {
501 struct page *page = NULL;
502 pte_t *_pte;
503 int none_or_zero = 0, result = 0, referenced = 0;
504 bool writable = false;
505
506 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
507 _pte++, address += PAGE_SIZE) {
508 pte_t pteval = *_pte;
509 if (pte_none(pteval) || (pte_present(pteval) &&
510 is_zero_pfn(pte_pfn(pteval)))) {
511 if (!userfaultfd_armed(vma) &&
512 ++none_or_zero <= khugepaged_max_ptes_none) {
513 continue;
514 } else {
515 result = SCAN_EXCEED_NONE_PTE;
516 goto out;
517 }
518 }
519 if (!pte_present(pteval)) {
520 result = SCAN_PTE_NON_PRESENT;
521 goto out;
522 }
523 page = vm_normal_page(vma, address, pteval);
524 if (unlikely(!page)) {
525 result = SCAN_PAGE_NULL;
526 goto out;
527 }
528
529 VM_BUG_ON_PAGE(PageCompound(page), page);
530 VM_BUG_ON_PAGE(!PageAnon(page), page);
531 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
532
533 /*
534 * We can do it before isolate_lru_page because the
535 * page can't be freed from under us. NOTE: PG_lock
536 * is needed to serialize against split_huge_page
537 * when invoked from the VM.
538 */
539 if (!trylock_page(page)) {
540 result = SCAN_PAGE_LOCK;
541 goto out;
542 }
543
544 /*
545 * cannot use mapcount: can't collapse if there's a gup pin.
546 * The page must only be referenced by the scanned process
547 * and page swap cache.
548 */
549 if (page_count(page) != 1 + !!PageSwapCache(page)) {
550 unlock_page(page);
551 result = SCAN_PAGE_COUNT;
552 goto out;
553 }
554 if (pte_write(pteval)) {
555 writable = true;
556 } else {
557 if (PageSwapCache(page) &&
558 !reuse_swap_page(page, NULL)) {
559 unlock_page(page);
560 result = SCAN_SWAP_CACHE_PAGE;
561 goto out;
562 }
563 /*
564 * Page is not in the swap cache. It can be collapsed
565 * into a THP.
566 */
567 }
568
569 /*
570 * Isolate the page to avoid collapsing an hugepage
571 * currently in use by the VM.
572 */
573 if (isolate_lru_page(page)) {
574 unlock_page(page);
575 result = SCAN_DEL_PAGE_LRU;
576 goto out;
577 }
578 /* 0 stands for page_is_file_cache(page) == false */
579 inc_node_page_state(page, NR_ISOLATED_ANON + 0);
580 VM_BUG_ON_PAGE(!PageLocked(page), page);
581 VM_BUG_ON_PAGE(PageLRU(page), page);
582
583 /* There should be enough young pte to collapse the page */
584 if (pte_young(pteval) ||
585 page_is_young(page) || PageReferenced(page) ||
586 mmu_notifier_test_young(vma->vm_mm, address))
587 referenced++;
588 }
589 if (likely(writable)) {
590 if (likely(referenced)) {
591 result = SCAN_SUCCEED;
592 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
593 referenced, writable, result);
594 return 1;
595 }
596 } else {
597 result = SCAN_PAGE_RO;
598 }
599
600 out:
601 release_pte_pages(pte, _pte);
602 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
603 referenced, writable, result);
604 return 0;
605 }
606
607 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
608 struct vm_area_struct *vma,
609 unsigned long address,
610 spinlock_t *ptl)
611 {
612 pte_t *_pte;
613 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
614 pte_t pteval = *_pte;
615 struct page *src_page;
616
617 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
618 clear_user_highpage(page, address);
619 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
620 if (is_zero_pfn(pte_pfn(pteval))) {
621 /*
622 * ptl mostly unnecessary.
623 */
624 spin_lock(ptl);
625 /*
626 * paravirt calls inside pte_clear here are
627 * superfluous.
628 */
629 pte_clear(vma->vm_mm, address, _pte);
630 spin_unlock(ptl);
631 }
632 } else {
633 src_page = pte_page(pteval);
634 copy_user_highpage(page, src_page, address, vma);
635 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
636 release_pte_page(src_page);
637 /*
638 * ptl mostly unnecessary, but preempt has to
639 * be disabled to update the per-cpu stats
640 * inside page_remove_rmap().
641 */
642 spin_lock(ptl);
643 /*
644 * paravirt calls inside pte_clear here are
645 * superfluous.
646 */
647 pte_clear(vma->vm_mm, address, _pte);
648 page_remove_rmap(src_page, false);
649 spin_unlock(ptl);
650 free_page_and_swap_cache(src_page);
651 }
652
653 address += PAGE_SIZE;
654 page++;
655 }
656 }
657
658 static void khugepaged_alloc_sleep(void)
659 {
660 DEFINE_WAIT(wait);
661
662 add_wait_queue(&khugepaged_wait, &wait);
663 freezable_schedule_timeout_interruptible(
664 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
665 remove_wait_queue(&khugepaged_wait, &wait);
666 }
667
668 static int khugepaged_node_load[MAX_NUMNODES];
669
670 static bool khugepaged_scan_abort(int nid)
671 {
672 int i;
673
674 /*
675 * If node_reclaim_mode is disabled, then no extra effort is made to
676 * allocate memory locally.
677 */
678 if (!node_reclaim_mode)
679 return false;
680
681 /* If there is a count for this node already, it must be acceptable */
682 if (khugepaged_node_load[nid])
683 return false;
684
685 for (i = 0; i < MAX_NUMNODES; i++) {
686 if (!khugepaged_node_load[i])
687 continue;
688 if (node_distance(nid, i) > RECLAIM_DISTANCE)
689 return true;
690 }
691 return false;
692 }
693
694 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
695 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
696 {
697 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
698 }
699
700 #ifdef CONFIG_NUMA
701 static int khugepaged_find_target_node(void)
702 {
703 static int last_khugepaged_target_node = NUMA_NO_NODE;
704 int nid, target_node = 0, max_value = 0;
705
706 /* find first node with max normal pages hit */
707 for (nid = 0; nid < MAX_NUMNODES; nid++)
708 if (khugepaged_node_load[nid] > max_value) {
709 max_value = khugepaged_node_load[nid];
710 target_node = nid;
711 }
712
713 /* do some balance if several nodes have the same hit record */
714 if (target_node <= last_khugepaged_target_node)
715 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
716 nid++)
717 if (max_value == khugepaged_node_load[nid]) {
718 target_node = nid;
719 break;
720 }
721
722 last_khugepaged_target_node = target_node;
723 return target_node;
724 }
725
726 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
727 {
728 if (IS_ERR(*hpage)) {
729 if (!*wait)
730 return false;
731
732 *wait = false;
733 *hpage = NULL;
734 khugepaged_alloc_sleep();
735 } else if (*hpage) {
736 put_page(*hpage);
737 *hpage = NULL;
738 }
739
740 return true;
741 }
742
743 static struct page *
744 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
745 {
746 VM_BUG_ON_PAGE(*hpage, *hpage);
747
748 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
749 if (unlikely(!*hpage)) {
750 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
751 *hpage = ERR_PTR(-ENOMEM);
752 return NULL;
753 }
754
755 prep_transhuge_page(*hpage);
756 count_vm_event(THP_COLLAPSE_ALLOC);
757 return *hpage;
758 }
759 #else
760 static int khugepaged_find_target_node(void)
761 {
762 return 0;
763 }
764
765 static inline struct page *alloc_khugepaged_hugepage(void)
766 {
767 struct page *page;
768
769 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
770 HPAGE_PMD_ORDER);
771 if (page)
772 prep_transhuge_page(page);
773 return page;
774 }
775
776 static struct page *khugepaged_alloc_hugepage(bool *wait)
777 {
778 struct page *hpage;
779
780 do {
781 hpage = alloc_khugepaged_hugepage();
782 if (!hpage) {
783 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
784 if (!*wait)
785 return NULL;
786
787 *wait = false;
788 khugepaged_alloc_sleep();
789 } else
790 count_vm_event(THP_COLLAPSE_ALLOC);
791 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
792
793 return hpage;
794 }
795
796 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
797 {
798 if (!*hpage)
799 *hpage = khugepaged_alloc_hugepage(wait);
800
801 if (unlikely(!*hpage))
802 return false;
803
804 return true;
805 }
806
807 static struct page *
808 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
809 {
810 VM_BUG_ON(!*hpage);
811
812 return *hpage;
813 }
814 #endif
815
816 static bool hugepage_vma_check(struct vm_area_struct *vma)
817 {
818 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
819 (vma->vm_flags & VM_NOHUGEPAGE))
820 return false;
821 if (shmem_file(vma->vm_file)) {
822 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
823 return false;
824 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
825 HPAGE_PMD_NR);
826 }
827 if (!vma->anon_vma || vma->vm_ops)
828 return false;
829 if (is_vma_temporary_stack(vma))
830 return false;
831 return !(vma->vm_flags & VM_NO_KHUGEPAGED);
832 }
833
834 /*
835 * If mmap_sem temporarily dropped, revalidate vma
836 * before taking mmap_sem.
837 * Return 0 if succeeds, otherwise return none-zero
838 * value (scan code).
839 */
840
841 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address)
842 {
843 struct vm_area_struct *vma;
844 unsigned long hstart, hend;
845
846 if (unlikely(khugepaged_test_exit(mm)))
847 return SCAN_ANY_PROCESS;
848
849 vma = find_vma(mm, address);
850 if (!vma)
851 return SCAN_VMA_NULL;
852
853 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
854 hend = vma->vm_end & HPAGE_PMD_MASK;
855 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
856 return SCAN_ADDRESS_RANGE;
857 if (!hugepage_vma_check(vma))
858 return SCAN_VMA_CHECK;
859 return 0;
860 }
861
862 /*
863 * Bring missing pages in from swap, to complete THP collapse.
864 * Only done if khugepaged_scan_pmd believes it is worthwhile.
865 *
866 * Called and returns without pte mapped or spinlocks held,
867 * but with mmap_sem held to protect against vma changes.
868 */
869
870 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
871 struct vm_area_struct *vma,
872 unsigned long address, pmd_t *pmd,
873 int referenced)
874 {
875 pte_t pteval;
876 int swapped_in = 0, ret = 0;
877 struct fault_env fe = {
878 .vma = vma,
879 .address = address,
880 .flags = FAULT_FLAG_ALLOW_RETRY,
881 .pmd = pmd,
882 };
883
884 fe.pte = pte_offset_map(pmd, address);
885 for (; fe.address < address + HPAGE_PMD_NR*PAGE_SIZE;
886 fe.pte++, fe.address += PAGE_SIZE) {
887 pteval = *fe.pte;
888 if (!is_swap_pte(pteval))
889 continue;
890 swapped_in++;
891 /* we only decide to swapin, if there is enough young ptes */
892 if (referenced < HPAGE_PMD_NR/2) {
893 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
894 return false;
895 }
896 ret = do_swap_page(&fe, pteval);
897
898 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
899 if (ret & VM_FAULT_RETRY) {
900 down_read(&mm->mmap_sem);
901 if (hugepage_vma_revalidate(mm, address)) {
902 /* vma is no longer available, don't continue to swapin */
903 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
904 return false;
905 }
906 /* check if the pmd is still valid */
907 if (mm_find_pmd(mm, address) != pmd)
908 return false;
909 }
910 if (ret & VM_FAULT_ERROR) {
911 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
912 return false;
913 }
914 /* pte is unmapped now, we need to map it */
915 fe.pte = pte_offset_map(pmd, fe.address);
916 }
917 fe.pte--;
918 pte_unmap(fe.pte);
919 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
920 return true;
921 }
922
923 static void collapse_huge_page(struct mm_struct *mm,
924 unsigned long address,
925 struct page **hpage,
926 struct vm_area_struct *vma,
927 int node, int referenced)
928 {
929 pmd_t *pmd, _pmd;
930 pte_t *pte;
931 pgtable_t pgtable;
932 struct page *new_page;
933 spinlock_t *pmd_ptl, *pte_ptl;
934 int isolated = 0, result = 0;
935 struct mem_cgroup *memcg;
936 unsigned long mmun_start; /* For mmu_notifiers */
937 unsigned long mmun_end; /* For mmu_notifiers */
938 gfp_t gfp;
939
940 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
941
942 /* Only allocate from the target node */
943 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
944
945 /*
946 * Before allocating the hugepage, release the mmap_sem read lock.
947 * The allocation can take potentially a long time if it involves
948 * sync compaction, and we do not need to hold the mmap_sem during
949 * that. We will recheck the vma after taking it again in write mode.
950 */
951 up_read(&mm->mmap_sem);
952 new_page = khugepaged_alloc_page(hpage, gfp, node);
953 if (!new_page) {
954 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
955 goto out_nolock;
956 }
957
958 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
959 result = SCAN_CGROUP_CHARGE_FAIL;
960 goto out_nolock;
961 }
962
963 down_read(&mm->mmap_sem);
964 result = hugepage_vma_revalidate(mm, address);
965 if (result) {
966 mem_cgroup_cancel_charge(new_page, memcg, true);
967 up_read(&mm->mmap_sem);
968 goto out_nolock;
969 }
970
971 pmd = mm_find_pmd(mm, address);
972 if (!pmd) {
973 result = SCAN_PMD_NULL;
974 mem_cgroup_cancel_charge(new_page, memcg, true);
975 up_read(&mm->mmap_sem);
976 goto out_nolock;
977 }
978
979 /*
980 * __collapse_huge_page_swapin always returns with mmap_sem locked.
981 * If it fails, we release mmap_sem and jump out_nolock.
982 * Continuing to collapse causes inconsistency.
983 */
984 if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
985 mem_cgroup_cancel_charge(new_page, memcg, true);
986 up_read(&mm->mmap_sem);
987 goto out_nolock;
988 }
989
990 up_read(&mm->mmap_sem);
991 /*
992 * Prevent all access to pagetables with the exception of
993 * gup_fast later handled by the ptep_clear_flush and the VM
994 * handled by the anon_vma lock + PG_lock.
995 */
996 down_write(&mm->mmap_sem);
997 result = hugepage_vma_revalidate(mm, address);
998 if (result)
999 goto out;
1000 /* check if the pmd is still valid */
1001 if (mm_find_pmd(mm, address) != pmd)
1002 goto out;
1003
1004 anon_vma_lock_write(vma->anon_vma);
1005
1006 pte = pte_offset_map(pmd, address);
1007 pte_ptl = pte_lockptr(mm, pmd);
1008
1009 mmun_start = address;
1010 mmun_end = address + HPAGE_PMD_SIZE;
1011 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1012 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1013 /*
1014 * After this gup_fast can't run anymore. This also removes
1015 * any huge TLB entry from the CPU so we won't allow
1016 * huge and small TLB entries for the same virtual address
1017 * to avoid the risk of CPU bugs in that area.
1018 */
1019 _pmd = pmdp_collapse_flush(vma, address, pmd);
1020 spin_unlock(pmd_ptl);
1021 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1022
1023 spin_lock(pte_ptl);
1024 isolated = __collapse_huge_page_isolate(vma, address, pte);
1025 spin_unlock(pte_ptl);
1026
1027 if (unlikely(!isolated)) {
1028 pte_unmap(pte);
1029 spin_lock(pmd_ptl);
1030 BUG_ON(!pmd_none(*pmd));
1031 /*
1032 * We can only use set_pmd_at when establishing
1033 * hugepmds and never for establishing regular pmds that
1034 * points to regular pagetables. Use pmd_populate for that
1035 */
1036 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1037 spin_unlock(pmd_ptl);
1038 anon_vma_unlock_write(vma->anon_vma);
1039 result = SCAN_FAIL;
1040 goto out;
1041 }
1042
1043 /*
1044 * All pages are isolated and locked so anon_vma rmap
1045 * can't run anymore.
1046 */
1047 anon_vma_unlock_write(vma->anon_vma);
1048
1049 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1050 pte_unmap(pte);
1051 __SetPageUptodate(new_page);
1052 pgtable = pmd_pgtable(_pmd);
1053
1054 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1055 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1056
1057 /*
1058 * spin_lock() below is not the equivalent of smp_wmb(), so
1059 * this is needed to avoid the copy_huge_page writes to become
1060 * visible after the set_pmd_at() write.
1061 */
1062 smp_wmb();
1063
1064 spin_lock(pmd_ptl);
1065 BUG_ON(!pmd_none(*pmd));
1066 page_add_new_anon_rmap(new_page, vma, address, true);
1067 mem_cgroup_commit_charge(new_page, memcg, false, true);
1068 lru_cache_add_active_or_unevictable(new_page, vma);
1069 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1070 set_pmd_at(mm, address, pmd, _pmd);
1071 update_mmu_cache_pmd(vma, address, pmd);
1072 spin_unlock(pmd_ptl);
1073
1074 *hpage = NULL;
1075
1076 khugepaged_pages_collapsed++;
1077 result = SCAN_SUCCEED;
1078 out_up_write:
1079 up_write(&mm->mmap_sem);
1080 out_nolock:
1081 trace_mm_collapse_huge_page(mm, isolated, result);
1082 return;
1083 out:
1084 mem_cgroup_cancel_charge(new_page, memcg, true);
1085 goto out_up_write;
1086 }
1087
1088 static int khugepaged_scan_pmd(struct mm_struct *mm,
1089 struct vm_area_struct *vma,
1090 unsigned long address,
1091 struct page **hpage)
1092 {
1093 pmd_t *pmd;
1094 pte_t *pte, *_pte;
1095 int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1096 struct page *page = NULL;
1097 unsigned long _address;
1098 spinlock_t *ptl;
1099 int node = NUMA_NO_NODE, unmapped = 0;
1100 bool writable = false;
1101
1102 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1103
1104 pmd = mm_find_pmd(mm, address);
1105 if (!pmd) {
1106 result = SCAN_PMD_NULL;
1107 goto out;
1108 }
1109
1110 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1111 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1112 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1113 _pte++, _address += PAGE_SIZE) {
1114 pte_t pteval = *_pte;
1115 if (is_swap_pte(pteval)) {
1116 if (++unmapped <= khugepaged_max_ptes_swap) {
1117 continue;
1118 } else {
1119 result = SCAN_EXCEED_SWAP_PTE;
1120 goto out_unmap;
1121 }
1122 }
1123 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1124 if (!userfaultfd_armed(vma) &&
1125 ++none_or_zero <= khugepaged_max_ptes_none) {
1126 continue;
1127 } else {
1128 result = SCAN_EXCEED_NONE_PTE;
1129 goto out_unmap;
1130 }
1131 }
1132 if (!pte_present(pteval)) {
1133 result = SCAN_PTE_NON_PRESENT;
1134 goto out_unmap;
1135 }
1136 if (pte_write(pteval))
1137 writable = true;
1138
1139 page = vm_normal_page(vma, _address, pteval);
1140 if (unlikely(!page)) {
1141 result = SCAN_PAGE_NULL;
1142 goto out_unmap;
1143 }
1144
1145 /* TODO: teach khugepaged to collapse THP mapped with pte */
1146 if (PageCompound(page)) {
1147 result = SCAN_PAGE_COMPOUND;
1148 goto out_unmap;
1149 }
1150
1151 /*
1152 * Record which node the original page is from and save this
1153 * information to khugepaged_node_load[].
1154 * Khupaged will allocate hugepage from the node has the max
1155 * hit record.
1156 */
1157 node = page_to_nid(page);
1158 if (khugepaged_scan_abort(node)) {
1159 result = SCAN_SCAN_ABORT;
1160 goto out_unmap;
1161 }
1162 khugepaged_node_load[node]++;
1163 if (!PageLRU(page)) {
1164 result = SCAN_PAGE_LRU;
1165 goto out_unmap;
1166 }
1167 if (PageLocked(page)) {
1168 result = SCAN_PAGE_LOCK;
1169 goto out_unmap;
1170 }
1171 if (!PageAnon(page)) {
1172 result = SCAN_PAGE_ANON;
1173 goto out_unmap;
1174 }
1175
1176 /*
1177 * cannot use mapcount: can't collapse if there's a gup pin.
1178 * The page must only be referenced by the scanned process
1179 * and page swap cache.
1180 */
1181 if (page_count(page) != 1 + !!PageSwapCache(page)) {
1182 result = SCAN_PAGE_COUNT;
1183 goto out_unmap;
1184 }
1185 if (pte_young(pteval) ||
1186 page_is_young(page) || PageReferenced(page) ||
1187 mmu_notifier_test_young(vma->vm_mm, address))
1188 referenced++;
1189 }
1190 if (writable) {
1191 if (referenced) {
1192 result = SCAN_SUCCEED;
1193 ret = 1;
1194 } else {
1195 result = SCAN_LACK_REFERENCED_PAGE;
1196 }
1197 } else {
1198 result = SCAN_PAGE_RO;
1199 }
1200 out_unmap:
1201 pte_unmap_unlock(pte, ptl);
1202 if (ret) {
1203 node = khugepaged_find_target_node();
1204 /* collapse_huge_page will return with the mmap_sem released */
1205 collapse_huge_page(mm, address, hpage, vma, node, referenced);
1206 }
1207 out:
1208 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1209 none_or_zero, result, unmapped);
1210 return ret;
1211 }
1212
1213 static void collect_mm_slot(struct mm_slot *mm_slot)
1214 {
1215 struct mm_struct *mm = mm_slot->mm;
1216
1217 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1218
1219 if (khugepaged_test_exit(mm)) {
1220 /* free mm_slot */
1221 hash_del(&mm_slot->hash);
1222 list_del(&mm_slot->mm_node);
1223
1224 /*
1225 * Not strictly needed because the mm exited already.
1226 *
1227 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1228 */
1229
1230 /* khugepaged_mm_lock actually not necessary for the below */
1231 free_mm_slot(mm_slot);
1232 mmdrop(mm);
1233 }
1234 }
1235
1236 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1237 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1238 {
1239 struct vm_area_struct *vma;
1240 unsigned long addr;
1241 pmd_t *pmd, _pmd;
1242
1243 i_mmap_lock_write(mapping);
1244 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1245 /* probably overkill */
1246 if (vma->anon_vma)
1247 continue;
1248 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1249 if (addr & ~HPAGE_PMD_MASK)
1250 continue;
1251 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1252 continue;
1253 pmd = mm_find_pmd(vma->vm_mm, addr);
1254 if (!pmd)
1255 continue;
1256 /*
1257 * We need exclusive mmap_sem to retract page table.
1258 * If trylock fails we would end up with pte-mapped THP after
1259 * re-fault. Not ideal, but it's more important to not disturb
1260 * the system too much.
1261 */
1262 if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1263 spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1264 /* assume page table is clear */
1265 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1266 spin_unlock(ptl);
1267 up_write(&vma->vm_mm->mmap_sem);
1268 atomic_long_dec(&vma->vm_mm->nr_ptes);
1269 pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1270 }
1271 }
1272 i_mmap_unlock_write(mapping);
1273 }
1274
1275 /**
1276 * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1277 *
1278 * Basic scheme is simple, details are more complex:
1279 * - allocate and freeze a new huge page;
1280 * - scan over radix tree replacing old pages the new one
1281 * + swap in pages if necessary;
1282 * + fill in gaps;
1283 * + keep old pages around in case if rollback is required;
1284 * - if replacing succeed:
1285 * + copy data over;
1286 * + free old pages;
1287 * + unfreeze huge page;
1288 * - if replacing failed;
1289 * + put all pages back and unfreeze them;
1290 * + restore gaps in the radix-tree;
1291 * + free huge page;
1292 */
1293 static void collapse_shmem(struct mm_struct *mm,
1294 struct address_space *mapping, pgoff_t start,
1295 struct page **hpage, int node)
1296 {
1297 gfp_t gfp;
1298 struct page *page, *new_page, *tmp;
1299 struct mem_cgroup *memcg;
1300 pgoff_t index, end = start + HPAGE_PMD_NR;
1301 LIST_HEAD(pagelist);
1302 struct radix_tree_iter iter;
1303 void **slot;
1304 int nr_none = 0, result = SCAN_SUCCEED;
1305
1306 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1307
1308 /* Only allocate from the target node */
1309 gfp = alloc_hugepage_khugepaged_gfpmask() |
1310 __GFP_OTHER_NODE | __GFP_THISNODE;
1311
1312 new_page = khugepaged_alloc_page(hpage, gfp, node);
1313 if (!new_page) {
1314 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1315 goto out;
1316 }
1317
1318 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1319 result = SCAN_CGROUP_CHARGE_FAIL;
1320 goto out;
1321 }
1322
1323 new_page->index = start;
1324 new_page->mapping = mapping;
1325 __SetPageSwapBacked(new_page);
1326 __SetPageLocked(new_page);
1327 BUG_ON(!page_ref_freeze(new_page, 1));
1328
1329
1330 /*
1331 * At this point the new_page is 'frozen' (page_count() is zero), locked
1332 * and not up-to-date. It's safe to insert it into radix tree, because
1333 * nobody would be able to map it or use it in other way until we
1334 * unfreeze it.
1335 */
1336
1337 index = start;
1338 spin_lock_irq(&mapping->tree_lock);
1339 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1340 int n = min(iter.index, end) - index;
1341
1342 /*
1343 * Handle holes in the radix tree: charge it from shmem and
1344 * insert relevant subpage of new_page into the radix-tree.
1345 */
1346 if (n && !shmem_charge(mapping->host, n)) {
1347 result = SCAN_FAIL;
1348 break;
1349 }
1350 nr_none += n;
1351 for (; index < min(iter.index, end); index++) {
1352 radix_tree_insert(&mapping->page_tree, index,
1353 new_page + (index % HPAGE_PMD_NR));
1354 }
1355
1356 /* We are done. */
1357 if (index >= end)
1358 break;
1359
1360 page = radix_tree_deref_slot_protected(slot,
1361 &mapping->tree_lock);
1362 if (radix_tree_exceptional_entry(page) || !PageUptodate(page)) {
1363 spin_unlock_irq(&mapping->tree_lock);
1364 /* swap in or instantiate fallocated page */
1365 if (shmem_getpage(mapping->host, index, &page,
1366 SGP_NOHUGE)) {
1367 result = SCAN_FAIL;
1368 goto tree_unlocked;
1369 }
1370 spin_lock_irq(&mapping->tree_lock);
1371 } else if (trylock_page(page)) {
1372 get_page(page);
1373 } else {
1374 result = SCAN_PAGE_LOCK;
1375 break;
1376 }
1377
1378 /*
1379 * The page must be locked, so we can drop the tree_lock
1380 * without racing with truncate.
1381 */
1382 VM_BUG_ON_PAGE(!PageLocked(page), page);
1383 VM_BUG_ON_PAGE(!PageUptodate(page), page);
1384 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1385
1386 if (page_mapping(page) != mapping) {
1387 result = SCAN_TRUNCATED;
1388 goto out_unlock;
1389 }
1390 spin_unlock_irq(&mapping->tree_lock);
1391
1392 if (isolate_lru_page(page)) {
1393 result = SCAN_DEL_PAGE_LRU;
1394 goto out_isolate_failed;
1395 }
1396
1397 if (page_mapped(page))
1398 unmap_mapping_range(mapping, index << PAGE_SHIFT,
1399 PAGE_SIZE, 0);
1400
1401 spin_lock_irq(&mapping->tree_lock);
1402
1403 VM_BUG_ON_PAGE(page_mapped(page), page);
1404
1405 /*
1406 * The page is expected to have page_count() == 3:
1407 * - we hold a pin on it;
1408 * - one reference from radix tree;
1409 * - one from isolate_lru_page;
1410 */
1411 if (!page_ref_freeze(page, 3)) {
1412 result = SCAN_PAGE_COUNT;
1413 goto out_lru;
1414 }
1415
1416 /*
1417 * Add the page to the list to be able to undo the collapse if
1418 * something go wrong.
1419 */
1420 list_add_tail(&page->lru, &pagelist);
1421
1422 /* Finally, replace with the new page. */
1423 radix_tree_replace_slot(slot,
1424 new_page + (index % HPAGE_PMD_NR));
1425
1426 index++;
1427 continue;
1428 out_lru:
1429 spin_unlock_irq(&mapping->tree_lock);
1430 putback_lru_page(page);
1431 out_isolate_failed:
1432 unlock_page(page);
1433 put_page(page);
1434 goto tree_unlocked;
1435 out_unlock:
1436 unlock_page(page);
1437 put_page(page);
1438 break;
1439 }
1440
1441 /*
1442 * Handle hole in radix tree at the end of the range.
1443 * This code only triggers if there's nothing in radix tree
1444 * beyond 'end'.
1445 */
1446 if (result == SCAN_SUCCEED && index < end) {
1447 int n = end - index;
1448
1449 if (!shmem_charge(mapping->host, n)) {
1450 result = SCAN_FAIL;
1451 goto tree_locked;
1452 }
1453
1454 for (; index < end; index++) {
1455 radix_tree_insert(&mapping->page_tree, index,
1456 new_page + (index % HPAGE_PMD_NR));
1457 }
1458 nr_none += n;
1459 }
1460
1461 tree_locked:
1462 spin_unlock_irq(&mapping->tree_lock);
1463 tree_unlocked:
1464
1465 if (result == SCAN_SUCCEED) {
1466 unsigned long flags;
1467 struct zone *zone = page_zone(new_page);
1468
1469 /*
1470 * Replacing old pages with new one has succeed, now we need to
1471 * copy the content and free old pages.
1472 */
1473 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1474 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1475 page);
1476 list_del(&page->lru);
1477 unlock_page(page);
1478 page_ref_unfreeze(page, 1);
1479 page->mapping = NULL;
1480 ClearPageActive(page);
1481 ClearPageUnevictable(page);
1482 put_page(page);
1483 }
1484
1485 local_irq_save(flags);
1486 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1487 if (nr_none) {
1488 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1489 __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
1490 }
1491 local_irq_restore(flags);
1492
1493 /*
1494 * Remove pte page tables, so we can re-faulti
1495 * the page as huge.
1496 */
1497 retract_page_tables(mapping, start);
1498
1499 /* Everything is ready, let's unfreeze the new_page */
1500 set_page_dirty(new_page);
1501 SetPageUptodate(new_page);
1502 page_ref_unfreeze(new_page, HPAGE_PMD_NR);
1503 mem_cgroup_commit_charge(new_page, memcg, false, true);
1504 lru_cache_add_anon(new_page);
1505 unlock_page(new_page);
1506
1507 *hpage = NULL;
1508 } else {
1509 /* Something went wrong: rollback changes to the radix-tree */
1510 shmem_uncharge(mapping->host, nr_none);
1511 spin_lock_irq(&mapping->tree_lock);
1512 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
1513 start) {
1514 if (iter.index >= end)
1515 break;
1516 page = list_first_entry_or_null(&pagelist,
1517 struct page, lru);
1518 if (!page || iter.index < page->index) {
1519 if (!nr_none)
1520 break;
1521 /* Put holes back where they were */
1522 radix_tree_replace_slot(slot, NULL);
1523 nr_none--;
1524 continue;
1525 }
1526
1527 VM_BUG_ON_PAGE(page->index != iter.index, page);
1528
1529 /* Unfreeze the page. */
1530 list_del(&page->lru);
1531 page_ref_unfreeze(page, 2);
1532 radix_tree_replace_slot(slot, page);
1533 spin_unlock_irq(&mapping->tree_lock);
1534 putback_lru_page(page);
1535 unlock_page(page);
1536 spin_lock_irq(&mapping->tree_lock);
1537 }
1538 VM_BUG_ON(nr_none);
1539 spin_unlock_irq(&mapping->tree_lock);
1540
1541 /* Unfreeze new_page, caller would take care about freeing it */
1542 page_ref_unfreeze(new_page, 1);
1543 mem_cgroup_cancel_charge(new_page, memcg, true);
1544 unlock_page(new_page);
1545 new_page->mapping = NULL;
1546 }
1547 out:
1548 VM_BUG_ON(!list_empty(&pagelist));
1549 /* TODO: tracepoints */
1550 }
1551
1552 static void khugepaged_scan_shmem(struct mm_struct *mm,
1553 struct address_space *mapping,
1554 pgoff_t start, struct page **hpage)
1555 {
1556 struct page *page = NULL;
1557 struct radix_tree_iter iter;
1558 void **slot;
1559 int present, swap;
1560 int node = NUMA_NO_NODE;
1561 int result = SCAN_SUCCEED;
1562
1563 present = 0;
1564 swap = 0;
1565 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1566 rcu_read_lock();
1567 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1568 if (iter.index >= start + HPAGE_PMD_NR)
1569 break;
1570
1571 page = radix_tree_deref_slot(slot);
1572 if (radix_tree_deref_retry(page)) {
1573 slot = radix_tree_iter_retry(&iter);
1574 continue;
1575 }
1576
1577 if (radix_tree_exception(page)) {
1578 if (++swap > khugepaged_max_ptes_swap) {
1579 result = SCAN_EXCEED_SWAP_PTE;
1580 break;
1581 }
1582 continue;
1583 }
1584
1585 if (PageTransCompound(page)) {
1586 result = SCAN_PAGE_COMPOUND;
1587 break;
1588 }
1589
1590 node = page_to_nid(page);
1591 if (khugepaged_scan_abort(node)) {
1592 result = SCAN_SCAN_ABORT;
1593 break;
1594 }
1595 khugepaged_node_load[node]++;
1596
1597 if (!PageLRU(page)) {
1598 result = SCAN_PAGE_LRU;
1599 break;
1600 }
1601
1602 if (page_count(page) != 1 + page_mapcount(page)) {
1603 result = SCAN_PAGE_COUNT;
1604 break;
1605 }
1606
1607 /*
1608 * We probably should check if the page is referenced here, but
1609 * nobody would transfer pte_young() to PageReferenced() for us.
1610 * And rmap walk here is just too costly...
1611 */
1612
1613 present++;
1614
1615 if (need_resched()) {
1616 cond_resched_rcu();
1617 slot = radix_tree_iter_next(&iter);
1618 }
1619 }
1620 rcu_read_unlock();
1621
1622 if (result == SCAN_SUCCEED) {
1623 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1624 result = SCAN_EXCEED_NONE_PTE;
1625 } else {
1626 node = khugepaged_find_target_node();
1627 collapse_shmem(mm, mapping, start, hpage, node);
1628 }
1629 }
1630
1631 /* TODO: tracepoints */
1632 }
1633 #else
1634 static void khugepaged_scan_shmem(struct mm_struct *mm,
1635 struct address_space *mapping,
1636 pgoff_t start, struct page **hpage)
1637 {
1638 BUILD_BUG();
1639 }
1640 #endif
1641
1642 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1643 struct page **hpage)
1644 __releases(&khugepaged_mm_lock)
1645 __acquires(&khugepaged_mm_lock)
1646 {
1647 struct mm_slot *mm_slot;
1648 struct mm_struct *mm;
1649 struct vm_area_struct *vma;
1650 int progress = 0;
1651
1652 VM_BUG_ON(!pages);
1653 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1654
1655 if (khugepaged_scan.mm_slot)
1656 mm_slot = khugepaged_scan.mm_slot;
1657 else {
1658 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1659 struct mm_slot, mm_node);
1660 khugepaged_scan.address = 0;
1661 khugepaged_scan.mm_slot = mm_slot;
1662 }
1663 spin_unlock(&khugepaged_mm_lock);
1664
1665 mm = mm_slot->mm;
1666 down_read(&mm->mmap_sem);
1667 if (unlikely(khugepaged_test_exit(mm)))
1668 vma = NULL;
1669 else
1670 vma = find_vma(mm, khugepaged_scan.address);
1671
1672 progress++;
1673 for (; vma; vma = vma->vm_next) {
1674 unsigned long hstart, hend;
1675
1676 cond_resched();
1677 if (unlikely(khugepaged_test_exit(mm))) {
1678 progress++;
1679 break;
1680 }
1681 if (!hugepage_vma_check(vma)) {
1682 skip:
1683 progress++;
1684 continue;
1685 }
1686 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1687 hend = vma->vm_end & HPAGE_PMD_MASK;
1688 if (hstart >= hend)
1689 goto skip;
1690 if (khugepaged_scan.address > hend)
1691 goto skip;
1692 if (khugepaged_scan.address < hstart)
1693 khugepaged_scan.address = hstart;
1694 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1695
1696 while (khugepaged_scan.address < hend) {
1697 int ret;
1698 cond_resched();
1699 if (unlikely(khugepaged_test_exit(mm)))
1700 goto breakouterloop;
1701
1702 VM_BUG_ON(khugepaged_scan.address < hstart ||
1703 khugepaged_scan.address + HPAGE_PMD_SIZE >
1704 hend);
1705 if (shmem_file(vma->vm_file)) {
1706 struct file *file;
1707 pgoff_t pgoff = linear_page_index(vma,
1708 khugepaged_scan.address);
1709 if (!shmem_huge_enabled(vma))
1710 goto skip;
1711 file = get_file(vma->vm_file);
1712 up_read(&mm->mmap_sem);
1713 ret = 1;
1714 khugepaged_scan_shmem(mm, file->f_mapping,
1715 pgoff, hpage);
1716 fput(file);
1717 } else {
1718 ret = khugepaged_scan_pmd(mm, vma,
1719 khugepaged_scan.address,
1720 hpage);
1721 }
1722 /* move to next address */
1723 khugepaged_scan.address += HPAGE_PMD_SIZE;
1724 progress += HPAGE_PMD_NR;
1725 if (ret)
1726 /* we released mmap_sem so break loop */
1727 goto breakouterloop_mmap_sem;
1728 if (progress >= pages)
1729 goto breakouterloop;
1730 }
1731 }
1732 breakouterloop:
1733 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1734 breakouterloop_mmap_sem:
1735
1736 spin_lock(&khugepaged_mm_lock);
1737 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1738 /*
1739 * Release the current mm_slot if this mm is about to die, or
1740 * if we scanned all vmas of this mm.
1741 */
1742 if (khugepaged_test_exit(mm) || !vma) {
1743 /*
1744 * Make sure that if mm_users is reaching zero while
1745 * khugepaged runs here, khugepaged_exit will find
1746 * mm_slot not pointing to the exiting mm.
1747 */
1748 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1749 khugepaged_scan.mm_slot = list_entry(
1750 mm_slot->mm_node.next,
1751 struct mm_slot, mm_node);
1752 khugepaged_scan.address = 0;
1753 } else {
1754 khugepaged_scan.mm_slot = NULL;
1755 khugepaged_full_scans++;
1756 }
1757
1758 collect_mm_slot(mm_slot);
1759 }
1760
1761 return progress;
1762 }
1763
1764 static int khugepaged_has_work(void)
1765 {
1766 return !list_empty(&khugepaged_scan.mm_head) &&
1767 khugepaged_enabled();
1768 }
1769
1770 static int khugepaged_wait_event(void)
1771 {
1772 return !list_empty(&khugepaged_scan.mm_head) ||
1773 kthread_should_stop();
1774 }
1775
1776 static void khugepaged_do_scan(void)
1777 {
1778 struct page *hpage = NULL;
1779 unsigned int progress = 0, pass_through_head = 0;
1780 unsigned int pages = khugepaged_pages_to_scan;
1781 bool wait = true;
1782
1783 barrier(); /* write khugepaged_pages_to_scan to local stack */
1784
1785 while (progress < pages) {
1786 if (!khugepaged_prealloc_page(&hpage, &wait))
1787 break;
1788
1789 cond_resched();
1790
1791 if (unlikely(kthread_should_stop() || try_to_freeze()))
1792 break;
1793
1794 spin_lock(&khugepaged_mm_lock);
1795 if (!khugepaged_scan.mm_slot)
1796 pass_through_head++;
1797 if (khugepaged_has_work() &&
1798 pass_through_head < 2)
1799 progress += khugepaged_scan_mm_slot(pages - progress,
1800 &hpage);
1801 else
1802 progress = pages;
1803 spin_unlock(&khugepaged_mm_lock);
1804 }
1805
1806 if (!IS_ERR_OR_NULL(hpage))
1807 put_page(hpage);
1808 }
1809
1810 static bool khugepaged_should_wakeup(void)
1811 {
1812 return kthread_should_stop() ||
1813 time_after_eq(jiffies, khugepaged_sleep_expire);
1814 }
1815
1816 static void khugepaged_wait_work(void)
1817 {
1818 if (khugepaged_has_work()) {
1819 const unsigned long scan_sleep_jiffies =
1820 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1821
1822 if (!scan_sleep_jiffies)
1823 return;
1824
1825 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1826 wait_event_freezable_timeout(khugepaged_wait,
1827 khugepaged_should_wakeup(),
1828 scan_sleep_jiffies);
1829 return;
1830 }
1831
1832 if (khugepaged_enabled())
1833 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1834 }
1835
1836 static int khugepaged(void *none)
1837 {
1838 struct mm_slot *mm_slot;
1839
1840 set_freezable();
1841 set_user_nice(current, MAX_NICE);
1842
1843 while (!kthread_should_stop()) {
1844 khugepaged_do_scan();
1845 khugepaged_wait_work();
1846 }
1847
1848 spin_lock(&khugepaged_mm_lock);
1849 mm_slot = khugepaged_scan.mm_slot;
1850 khugepaged_scan.mm_slot = NULL;
1851 if (mm_slot)
1852 collect_mm_slot(mm_slot);
1853 spin_unlock(&khugepaged_mm_lock);
1854 return 0;
1855 }
1856
1857 static void set_recommended_min_free_kbytes(void)
1858 {
1859 struct zone *zone;
1860 int nr_zones = 0;
1861 unsigned long recommended_min;
1862
1863 for_each_populated_zone(zone)
1864 nr_zones++;
1865
1866 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1867 recommended_min = pageblock_nr_pages * nr_zones * 2;
1868
1869 /*
1870 * Make sure that on average at least two pageblocks are almost free
1871 * of another type, one for a migratetype to fall back to and a
1872 * second to avoid subsequent fallbacks of other types There are 3
1873 * MIGRATE_TYPES we care about.
1874 */
1875 recommended_min += pageblock_nr_pages * nr_zones *
1876 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1877
1878 /* don't ever allow to reserve more than 5% of the lowmem */
1879 recommended_min = min(recommended_min,
1880 (unsigned long) nr_free_buffer_pages() / 20);
1881 recommended_min <<= (PAGE_SHIFT-10);
1882
1883 if (recommended_min > min_free_kbytes) {
1884 if (user_min_free_kbytes >= 0)
1885 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1886 min_free_kbytes, recommended_min);
1887
1888 min_free_kbytes = recommended_min;
1889 }
1890 setup_per_zone_wmarks();
1891 }
1892
1893 int start_stop_khugepaged(void)
1894 {
1895 static struct task_struct *khugepaged_thread __read_mostly;
1896 static DEFINE_MUTEX(khugepaged_mutex);
1897 int err = 0;
1898
1899 mutex_lock(&khugepaged_mutex);
1900 if (khugepaged_enabled()) {
1901 if (!khugepaged_thread)
1902 khugepaged_thread = kthread_run(khugepaged, NULL,
1903 "khugepaged");
1904 if (IS_ERR(khugepaged_thread)) {
1905 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1906 err = PTR_ERR(khugepaged_thread);
1907 khugepaged_thread = NULL;
1908 goto fail;
1909 }
1910
1911 if (!list_empty(&khugepaged_scan.mm_head))
1912 wake_up_interruptible(&khugepaged_wait);
1913
1914 set_recommended_min_free_kbytes();
1915 } else if (khugepaged_thread) {
1916 kthread_stop(khugepaged_thread);
1917 khugepaged_thread = NULL;
1918 }
1919 fail:
1920 mutex_unlock(&khugepaged_mutex);
1921 return err;
1922 }