aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--include/linux/huge_mm.h10
-rw-r--r--include/linux/khugepaged.h5
-rw-r--r--mm/Makefile2
-rw-r--r--mm/huge_memory.c1493
-rw-r--r--mm/khugepaged.c1490
5 files changed, 1515 insertions, 1485 deletions
diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
index 7b7406e9fedf..92ce91c03cd0 100644
--- a/include/linux/huge_mm.h
+++ b/include/linux/huge_mm.h
@@ -41,6 +41,16 @@ enum transparent_hugepage_flag {
41#endif 41#endif
42}; 42};
43 43
44struct kobject;
45struct kobj_attribute;
46
47extern ssize_t single_hugepage_flag_store(struct kobject *kobj,
48 struct kobj_attribute *attr,
49 const char *buf, size_t count,
50 enum transparent_hugepage_flag flag);
51extern ssize_t single_hugepage_flag_show(struct kobject *kobj,
52 struct kobj_attribute *attr, char *buf,
53 enum transparent_hugepage_flag flag);
44extern struct kobj_attribute shmem_enabled_attr; 54extern struct kobj_attribute shmem_enabled_attr;
45 55
46#define HPAGE_PMD_ORDER (HPAGE_PMD_SHIFT-PAGE_SHIFT) 56#define HPAGE_PMD_ORDER (HPAGE_PMD_SHIFT-PAGE_SHIFT)
diff --git a/include/linux/khugepaged.h b/include/linux/khugepaged.h
index eeb307985715..1e032a1ddb3e 100644
--- a/include/linux/khugepaged.h
+++ b/include/linux/khugepaged.h
@@ -4,6 +4,11 @@
4#include <linux/sched.h> /* MMF_VM_HUGEPAGE */ 4#include <linux/sched.h> /* MMF_VM_HUGEPAGE */
5 5
6#ifdef CONFIG_TRANSPARENT_HUGEPAGE 6#ifdef CONFIG_TRANSPARENT_HUGEPAGE
7extern struct attribute_group khugepaged_attr_group;
8
9extern int khugepaged_init(void);
10extern void khugepaged_destroy(void);
11extern int start_stop_khugepaged(void);
7extern int __khugepaged_enter(struct mm_struct *mm); 12extern int __khugepaged_enter(struct mm_struct *mm);
8extern void __khugepaged_exit(struct mm_struct *mm); 13extern void __khugepaged_exit(struct mm_struct *mm);
9extern int khugepaged_enter_vma_merge(struct vm_area_struct *vma, 14extern int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
diff --git a/mm/Makefile b/mm/Makefile
index 78c6f7dedb83..fc059666c760 100644
--- a/mm/Makefile
+++ b/mm/Makefile
@@ -74,7 +74,7 @@ obj-$(CONFIG_MEMORY_HOTPLUG) += memory_hotplug.o
74obj-$(CONFIG_MEMTEST) += memtest.o 74obj-$(CONFIG_MEMTEST) += memtest.o
75obj-$(CONFIG_MIGRATION) += migrate.o 75obj-$(CONFIG_MIGRATION) += migrate.o
76obj-$(CONFIG_QUICKLIST) += quicklist.o 76obj-$(CONFIG_QUICKLIST) += quicklist.o
77obj-$(CONFIG_TRANSPARENT_HUGEPAGE) += huge_memory.o 77obj-$(CONFIG_TRANSPARENT_HUGEPAGE) += huge_memory.o khugepaged.o
78obj-$(CONFIG_PAGE_COUNTER) += page_counter.o 78obj-$(CONFIG_PAGE_COUNTER) += page_counter.o
79obj-$(CONFIG_MEMCG) += memcontrol.o vmpressure.o 79obj-$(CONFIG_MEMCG) += memcontrol.o vmpressure.o
80obj-$(CONFIG_MEMCG_SWAP) += swap_cgroup.o 80obj-$(CONFIG_MEMCG_SWAP) += swap_cgroup.o
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 5eba97874ad5..2706182787d8 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -18,7 +18,6 @@
18#include <linux/mm_inline.h> 18#include <linux/mm_inline.h>
19#include <linux/swapops.h> 19#include <linux/swapops.h>
20#include <linux/dax.h> 20#include <linux/dax.h>
21#include <linux/kthread.h>
22#include <linux/khugepaged.h> 21#include <linux/khugepaged.h>
23#include <linux/freezer.h> 22#include <linux/freezer.h>
24#include <linux/pfn_t.h> 23#include <linux/pfn_t.h>
@@ -36,35 +35,6 @@
36#include <asm/pgalloc.h> 35#include <asm/pgalloc.h>
37#include "internal.h" 36#include "internal.h"
38 37
39enum scan_result {
40 SCAN_FAIL,
41 SCAN_SUCCEED,
42 SCAN_PMD_NULL,
43 SCAN_EXCEED_NONE_PTE,
44 SCAN_PTE_NON_PRESENT,
45 SCAN_PAGE_RO,
46 SCAN_NO_REFERENCED_PAGE,
47 SCAN_PAGE_NULL,
48 SCAN_SCAN_ABORT,
49 SCAN_PAGE_COUNT,
50 SCAN_PAGE_LRU,
51 SCAN_PAGE_LOCK,
52 SCAN_PAGE_ANON,
53 SCAN_PAGE_COMPOUND,
54 SCAN_ANY_PROCESS,
55 SCAN_VMA_NULL,
56 SCAN_VMA_CHECK,
57 SCAN_ADDRESS_RANGE,
58 SCAN_SWAP_CACHE_PAGE,
59 SCAN_DEL_PAGE_LRU,
60 SCAN_ALLOC_HUGE_PAGE_FAIL,
61 SCAN_CGROUP_CHARGE_FAIL,
62 SCAN_EXCEED_SWAP_PTE
63};
64
65#define CREATE_TRACE_POINTS
66#include <trace/events/huge_memory.h>
67
68/* 38/*
69 * By default transparent hugepage support is disabled in order that avoid 39 * By default transparent hugepage support is disabled in order that avoid
70 * to risk increase the memory footprint of applications without a guaranteed 40 * to risk increase the memory footprint of applications without a guaranteed
@@ -84,128 +54,8 @@ unsigned long transparent_hugepage_flags __read_mostly =
84 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| 54 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
85 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 55 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
86 56
87/* default scan 8*512 pte (or vmas) every 30 second */
88static unsigned int khugepaged_pages_to_scan __read_mostly;
89static unsigned int khugepaged_pages_collapsed;
90static unsigned int khugepaged_full_scans;
91static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
92/* during fragmentation poll the hugepage allocator once every minute */
93static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
94static unsigned long khugepaged_sleep_expire;
95static struct task_struct *khugepaged_thread __read_mostly;
96static DEFINE_MUTEX(khugepaged_mutex);
97static DEFINE_SPINLOCK(khugepaged_mm_lock);
98static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
99/*
100 * default collapse hugepages if there is at least one pte mapped like
101 * it would have happened if the vma was large enough during page
102 * fault.
103 */
104static unsigned int khugepaged_max_ptes_none __read_mostly;
105static unsigned int khugepaged_max_ptes_swap __read_mostly;
106
107static int khugepaged(void *none);
108static int khugepaged_slab_init(void);
109static void khugepaged_slab_exit(void);
110
111#define MM_SLOTS_HASH_BITS 10
112static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
113
114static struct kmem_cache *mm_slot_cache __read_mostly;
115
116/**
117 * struct mm_slot - hash lookup from mm to mm_slot
118 * @hash: hash collision list
119 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
120 * @mm: the mm that this information is valid for
121 */
122struct mm_slot {
123 struct hlist_node hash;
124 struct list_head mm_node;
125 struct mm_struct *mm;
126};
127
128/**
129 * struct khugepaged_scan - cursor for scanning
130 * @mm_head: the head of the mm list to scan
131 * @mm_slot: the current mm_slot we are scanning
132 * @address: the next address inside that to be scanned
133 *
134 * There is only the one khugepaged_scan instance of this cursor structure.
135 */
136struct khugepaged_scan {
137 struct list_head mm_head;
138 struct mm_slot *mm_slot;
139 unsigned long address;
140};
141static struct khugepaged_scan khugepaged_scan = {
142 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
143};
144
145static struct shrinker deferred_split_shrinker; 57static struct shrinker deferred_split_shrinker;
146 58
147static void set_recommended_min_free_kbytes(void)
148{
149 struct zone *zone;
150 int nr_zones = 0;
151 unsigned long recommended_min;
152
153 for_each_populated_zone(zone)
154 nr_zones++;
155
156 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
157 recommended_min = pageblock_nr_pages * nr_zones * 2;
158
159 /*
160 * Make sure that on average at least two pageblocks are almost free
161 * of another type, one for a migratetype to fall back to and a
162 * second to avoid subsequent fallbacks of other types There are 3
163 * MIGRATE_TYPES we care about.
164 */
165 recommended_min += pageblock_nr_pages * nr_zones *
166 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
167
168 /* don't ever allow to reserve more than 5% of the lowmem */
169 recommended_min = min(recommended_min,
170 (unsigned long) nr_free_buffer_pages() / 20);
171 recommended_min <<= (PAGE_SHIFT-10);
172
173 if (recommended_min > min_free_kbytes) {
174 if (user_min_free_kbytes >= 0)
175 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
176 min_free_kbytes, recommended_min);
177
178 min_free_kbytes = recommended_min;
179 }
180 setup_per_zone_wmarks();
181}
182
183static int start_stop_khugepaged(void)
184{
185 int err = 0;
186 if (khugepaged_enabled()) {
187 if (!khugepaged_thread)
188 khugepaged_thread = kthread_run(khugepaged, NULL,
189 "khugepaged");
190 if (IS_ERR(khugepaged_thread)) {
191 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
192 err = PTR_ERR(khugepaged_thread);
193 khugepaged_thread = NULL;
194 goto fail;
195 }
196
197 if (!list_empty(&khugepaged_scan.mm_head))
198 wake_up_interruptible(&khugepaged_wait);
199
200 set_recommended_min_free_kbytes();
201 } else if (khugepaged_thread) {
202 kthread_stop(khugepaged_thread);
203 khugepaged_thread = NULL;
204 }
205fail:
206 return err;
207}
208
209static atomic_t huge_zero_refcount; 59static atomic_t huge_zero_refcount;
210struct page *huge_zero_page __read_mostly; 60struct page *huge_zero_page __read_mostly;
211 61
@@ -331,12 +181,7 @@ static ssize_t enabled_store(struct kobject *kobj,
331 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); 181 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
332 182
333 if (ret > 0) { 183 if (ret > 0) {
334 int err; 184 int err = start_stop_khugepaged();
335
336 mutex_lock(&khugepaged_mutex);
337 err = start_stop_khugepaged();
338 mutex_unlock(&khugepaged_mutex);
339
340 if (err) 185 if (err)
341 ret = err; 186 ret = err;
342 } 187 }
@@ -346,7 +191,7 @@ static ssize_t enabled_store(struct kobject *kobj,
346static struct kobj_attribute enabled_attr = 191static struct kobj_attribute enabled_attr =
347 __ATTR(enabled, 0644, enabled_show, enabled_store); 192 __ATTR(enabled, 0644, enabled_show, enabled_store);
348 193
349static ssize_t single_flag_show(struct kobject *kobj, 194ssize_t single_hugepage_flag_show(struct kobject *kobj,
350 struct kobj_attribute *attr, char *buf, 195 struct kobj_attribute *attr, char *buf,
351 enum transparent_hugepage_flag flag) 196 enum transparent_hugepage_flag flag)
352{ 197{
@@ -354,7 +199,7 @@ static ssize_t single_flag_show(struct kobject *kobj,
354 !!test_bit(flag, &transparent_hugepage_flags)); 199 !!test_bit(flag, &transparent_hugepage_flags));
355} 200}
356 201
357static ssize_t single_flag_store(struct kobject *kobj, 202ssize_t single_hugepage_flag_store(struct kobject *kobj,
358 struct kobj_attribute *attr, 203 struct kobj_attribute *attr,
359 const char *buf, size_t count, 204 const char *buf, size_t count,
360 enum transparent_hugepage_flag flag) 205 enum transparent_hugepage_flag flag)
@@ -409,13 +254,13 @@ static struct kobj_attribute defrag_attr =
409static ssize_t use_zero_page_show(struct kobject *kobj, 254static ssize_t use_zero_page_show(struct kobject *kobj,
410 struct kobj_attribute *attr, char *buf) 255 struct kobj_attribute *attr, char *buf)
411{ 256{
412 return single_flag_show(kobj, attr, buf, 257 return single_hugepage_flag_show(kobj, attr, buf,
413 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 258 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
414} 259}
415static ssize_t use_zero_page_store(struct kobject *kobj, 260static ssize_t use_zero_page_store(struct kobject *kobj,
416 struct kobj_attribute *attr, const char *buf, size_t count) 261 struct kobj_attribute *attr, const char *buf, size_t count)
417{ 262{
418 return single_flag_store(kobj, attr, buf, count, 263 return single_hugepage_flag_store(kobj, attr, buf, count,
419 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); 264 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
420} 265}
421static struct kobj_attribute use_zero_page_attr = 266static struct kobj_attribute use_zero_page_attr =
@@ -424,14 +269,14 @@ static struct kobj_attribute use_zero_page_attr =
424static ssize_t debug_cow_show(struct kobject *kobj, 269static ssize_t debug_cow_show(struct kobject *kobj,
425 struct kobj_attribute *attr, char *buf) 270 struct kobj_attribute *attr, char *buf)
426{ 271{
427 return single_flag_show(kobj, attr, buf, 272 return single_hugepage_flag_show(kobj, attr, buf,
428 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 273 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
429} 274}
430static ssize_t debug_cow_store(struct kobject *kobj, 275static ssize_t debug_cow_store(struct kobject *kobj,
431 struct kobj_attribute *attr, 276 struct kobj_attribute *attr,
432 const char *buf, size_t count) 277 const char *buf, size_t count)
433{ 278{
434 return single_flag_store(kobj, attr, buf, count, 279 return single_hugepage_flag_store(kobj, attr, buf, count,
435 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); 280 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
436} 281}
437static struct kobj_attribute debug_cow_attr = 282static struct kobj_attribute debug_cow_attr =
@@ -455,199 +300,6 @@ static struct attribute_group hugepage_attr_group = {
455 .attrs = hugepage_attr, 300 .attrs = hugepage_attr,
456}; 301};
457 302
458static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
459 struct kobj_attribute *attr,
460 char *buf)
461{
462 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
463}
464
465static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
466 struct kobj_attribute *attr,
467 const char *buf, size_t count)
468{
469 unsigned long msecs;
470 int err;
471
472 err = kstrtoul(buf, 10, &msecs);
473 if (err || msecs > UINT_MAX)
474 return -EINVAL;
475
476 khugepaged_scan_sleep_millisecs = msecs;
477 khugepaged_sleep_expire = 0;
478 wake_up_interruptible(&khugepaged_wait);
479
480 return count;
481}
482static struct kobj_attribute scan_sleep_millisecs_attr =
483 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
484 scan_sleep_millisecs_store);
485
486static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
487 struct kobj_attribute *attr,
488 char *buf)
489{
490 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
491}
492
493static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
494 struct kobj_attribute *attr,
495 const char *buf, size_t count)
496{
497 unsigned long msecs;
498 int err;
499
500 err = kstrtoul(buf, 10, &msecs);
501 if (err || msecs > UINT_MAX)
502 return -EINVAL;
503
504 khugepaged_alloc_sleep_millisecs = msecs;
505 khugepaged_sleep_expire = 0;
506 wake_up_interruptible(&khugepaged_wait);
507
508 return count;
509}
510static struct kobj_attribute alloc_sleep_millisecs_attr =
511 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
512 alloc_sleep_millisecs_store);
513
514static ssize_t pages_to_scan_show(struct kobject *kobj,
515 struct kobj_attribute *attr,
516 char *buf)
517{
518 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
519}
520static ssize_t pages_to_scan_store(struct kobject *kobj,
521 struct kobj_attribute *attr,
522 const char *buf, size_t count)
523{
524 int err;
525 unsigned long pages;
526
527 err = kstrtoul(buf, 10, &pages);
528 if (err || !pages || pages > UINT_MAX)
529 return -EINVAL;
530
531 khugepaged_pages_to_scan = pages;
532
533 return count;
534}
535static struct kobj_attribute pages_to_scan_attr =
536 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
537 pages_to_scan_store);
538
539static ssize_t pages_collapsed_show(struct kobject *kobj,
540 struct kobj_attribute *attr,
541 char *buf)
542{
543 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
544}
545static struct kobj_attribute pages_collapsed_attr =
546 __ATTR_RO(pages_collapsed);
547
548static ssize_t full_scans_show(struct kobject *kobj,
549 struct kobj_attribute *attr,
550 char *buf)
551{
552 return sprintf(buf, "%u\n", khugepaged_full_scans);
553}
554static struct kobj_attribute full_scans_attr =
555 __ATTR_RO(full_scans);
556
557static ssize_t khugepaged_defrag_show(struct kobject *kobj,
558 struct kobj_attribute *attr, char *buf)
559{
560 return single_flag_show(kobj, attr, buf,
561 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
562}
563static ssize_t khugepaged_defrag_store(struct kobject *kobj,
564 struct kobj_attribute *attr,
565 const char *buf, size_t count)
566{
567 return single_flag_store(kobj, attr, buf, count,
568 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
569}
570static struct kobj_attribute khugepaged_defrag_attr =
571 __ATTR(defrag, 0644, khugepaged_defrag_show,
572 khugepaged_defrag_store);
573
574/*
575 * max_ptes_none controls if khugepaged should collapse hugepages over
576 * any unmapped ptes in turn potentially increasing the memory
577 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
578 * reduce the available free memory in the system as it
579 * runs. Increasing max_ptes_none will instead potentially reduce the
580 * free memory in the system during the khugepaged scan.
581 */
582static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
583 struct kobj_attribute *attr,
584 char *buf)
585{
586 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
587}
588static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
589 struct kobj_attribute *attr,
590 const char *buf, size_t count)
591{
592 int err;
593 unsigned long max_ptes_none;
594
595 err = kstrtoul(buf, 10, &max_ptes_none);
596 if (err || max_ptes_none > HPAGE_PMD_NR-1)
597 return -EINVAL;
598
599 khugepaged_max_ptes_none = max_ptes_none;
600
601 return count;
602}
603static struct kobj_attribute khugepaged_max_ptes_none_attr =
604 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
605 khugepaged_max_ptes_none_store);
606
607static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
608 struct kobj_attribute *attr,
609 char *buf)
610{
611 return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
612}
613
614static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
615 struct kobj_attribute *attr,
616 const char *buf, size_t count)
617{
618 int err;
619 unsigned long max_ptes_swap;
620
621 err = kstrtoul(buf, 10, &max_ptes_swap);
622 if (err || max_ptes_swap > HPAGE_PMD_NR-1)
623 return -EINVAL;
624
625 khugepaged_max_ptes_swap = max_ptes_swap;
626
627 return count;
628}
629
630static struct kobj_attribute khugepaged_max_ptes_swap_attr =
631 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
632 khugepaged_max_ptes_swap_store);
633
634static struct attribute *khugepaged_attr[] = {
635 &khugepaged_defrag_attr.attr,
636 &khugepaged_max_ptes_none_attr.attr,
637 &pages_to_scan_attr.attr,
638 &pages_collapsed_attr.attr,
639 &full_scans_attr.attr,
640 &scan_sleep_millisecs_attr.attr,
641 &alloc_sleep_millisecs_attr.attr,
642 &khugepaged_max_ptes_swap_attr.attr,
643 NULL,
644};
645
646static struct attribute_group khugepaged_attr_group = {
647 .attrs = khugepaged_attr,
648 .name = "khugepaged",
649};
650
651static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) 303static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
652{ 304{
653 int err; 305 int err;
@@ -706,9 +358,6 @@ static int __init hugepage_init(void)
706 return -EINVAL; 358 return -EINVAL;
707 } 359 }
708 360
709 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
710 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
711 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
712 /* 361 /*
713 * hugepages can't be allocated by the buddy allocator 362 * hugepages can't be allocated by the buddy allocator
714 */ 363 */
@@ -723,7 +372,7 @@ static int __init hugepage_init(void)
723 if (err) 372 if (err)
724 goto err_sysfs; 373 goto err_sysfs;
725 374
726 err = khugepaged_slab_init(); 375 err = khugepaged_init();
727 if (err) 376 if (err)
728 goto err_slab; 377 goto err_slab;
729 378
@@ -754,7 +403,7 @@ err_khugepaged:
754err_split_shrinker: 403err_split_shrinker:
755 unregister_shrinker(&huge_zero_page_shrinker); 404 unregister_shrinker(&huge_zero_page_shrinker);
756err_hzp_shrinker: 405err_hzp_shrinker:
757 khugepaged_slab_exit(); 406 khugepaged_destroy();
758err_slab: 407err_slab:
759 hugepage_exit_sysfs(hugepage_kobj); 408 hugepage_exit_sysfs(hugepage_kobj);
760err_sysfs: 409err_sysfs:
@@ -909,12 +558,6 @@ static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
909 return GFP_TRANSHUGE | reclaim_flags; 558 return GFP_TRANSHUGE | reclaim_flags;
910} 559}
911 560
912/* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
913static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
914{
915 return GFP_TRANSHUGE | (khugepaged_defrag() ? __GFP_DIRECT_RECLAIM : 0);
916}
917
918/* Caller must hold page table lock. */ 561/* Caller must hold page table lock. */
919static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, 562static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
920 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, 563 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
@@ -1830,1124 +1473,6 @@ spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1830 return NULL; 1473 return NULL;
1831} 1474}
1832 1475
1833#define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1834
1835int hugepage_madvise(struct vm_area_struct *vma,
1836 unsigned long *vm_flags, int advice)
1837{
1838 switch (advice) {
1839 case MADV_HUGEPAGE:
1840#ifdef CONFIG_S390
1841 /*
1842 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1843 * can't handle this properly after s390_enable_sie, so we simply
1844 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1845 */
1846 if (mm_has_pgste(vma->vm_mm))
1847 return 0;
1848#endif
1849 *vm_flags &= ~VM_NOHUGEPAGE;
1850 *vm_flags |= VM_HUGEPAGE;
1851 /*
1852 * If the vma become good for khugepaged to scan,
1853 * register it here without waiting a page fault that
1854 * may not happen any time soon.
1855 */
1856 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
1857 khugepaged_enter_vma_merge(vma, *vm_flags))
1858 return -ENOMEM;
1859 break;
1860 case MADV_NOHUGEPAGE:
1861 *vm_flags &= ~VM_HUGEPAGE;
1862 *vm_flags |= VM_NOHUGEPAGE;
1863 /*
1864 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1865 * this vma even if we leave the mm registered in khugepaged if
1866 * it got registered before VM_NOHUGEPAGE was set.
1867 */
1868 break;
1869 }
1870
1871 return 0;
1872}
1873
1874static int __init khugepaged_slab_init(void)
1875{
1876 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1877 sizeof(struct mm_slot),
1878 __alignof__(struct mm_slot), 0, NULL);
1879 if (!mm_slot_cache)
1880 return -ENOMEM;
1881
1882 return 0;
1883}
1884
1885static void __init khugepaged_slab_exit(void)
1886{
1887 kmem_cache_destroy(mm_slot_cache);
1888}
1889
1890static inline struct mm_slot *alloc_mm_slot(void)
1891{
1892 if (!mm_slot_cache) /* initialization failed */
1893 return NULL;
1894 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1895}
1896
1897static inline void free_mm_slot(struct mm_slot *mm_slot)
1898{
1899 kmem_cache_free(mm_slot_cache, mm_slot);
1900}
1901
1902static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1903{
1904 struct mm_slot *mm_slot;
1905
1906 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1907 if (mm == mm_slot->mm)
1908 return mm_slot;
1909
1910 return NULL;
1911}
1912
1913static void insert_to_mm_slots_hash(struct mm_struct *mm,
1914 struct mm_slot *mm_slot)
1915{
1916 mm_slot->mm = mm;
1917 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1918}
1919
1920static inline int khugepaged_test_exit(struct mm_struct *mm)
1921{
1922 return atomic_read(&mm->mm_users) == 0;
1923}
1924
1925int __khugepaged_enter(struct mm_struct *mm)
1926{
1927 struct mm_slot *mm_slot;
1928 int wakeup;
1929
1930 mm_slot = alloc_mm_slot();
1931 if (!mm_slot)
1932 return -ENOMEM;
1933
1934 /* __khugepaged_exit() must not run from under us */
1935 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1936 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1937 free_mm_slot(mm_slot);
1938 return 0;
1939 }
1940
1941 spin_lock(&khugepaged_mm_lock);
1942 insert_to_mm_slots_hash(mm, mm_slot);
1943 /*
1944 * Insert just behind the scanning cursor, to let the area settle
1945 * down a little.
1946 */
1947 wakeup = list_empty(&khugepaged_scan.mm_head);
1948 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1949 spin_unlock(&khugepaged_mm_lock);
1950
1951 atomic_inc(&mm->mm_count);
1952 if (wakeup)
1953 wake_up_interruptible(&khugepaged_wait);
1954
1955 return 0;
1956}
1957
1958int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1959 unsigned long vm_flags)
1960{
1961 unsigned long hstart, hend;
1962 if (!vma->anon_vma)
1963 /*
1964 * Not yet faulted in so we will register later in the
1965 * page fault if needed.
1966 */
1967 return 0;
1968 if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
1969 /* khugepaged not yet working on file or special mappings */
1970 return 0;
1971 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1972 hend = vma->vm_end & HPAGE_PMD_MASK;
1973 if (hstart < hend)
1974 return khugepaged_enter(vma, vm_flags);
1975 return 0;
1976}
1977
1978void __khugepaged_exit(struct mm_struct *mm)
1979{
1980 struct mm_slot *mm_slot;
1981 int free = 0;
1982
1983 spin_lock(&khugepaged_mm_lock);
1984 mm_slot = get_mm_slot(mm);
1985 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1986 hash_del(&mm_slot->hash);
1987 list_del(&mm_slot->mm_node);
1988 free = 1;
1989 }
1990 spin_unlock(&khugepaged_mm_lock);
1991
1992 if (free) {
1993 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1994 free_mm_slot(mm_slot);
1995 mmdrop(mm);
1996 } else if (mm_slot) {
1997 /*
1998 * This is required to serialize against
1999 * khugepaged_test_exit() (which is guaranteed to run
2000 * under mmap sem read mode). Stop here (after we
2001 * return all pagetables will be destroyed) until
2002 * khugepaged has finished working on the pagetables
2003 * under the mmap_sem.
2004 */
2005 down_write(&mm->mmap_sem);
2006 up_write(&mm->mmap_sem);
2007 }
2008}
2009
2010static void release_pte_page(struct page *page)
2011{
2012 /* 0 stands for page_is_file_cache(page) == false */
2013 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2014 unlock_page(page);
2015 putback_lru_page(page);
2016}
2017
2018static void release_pte_pages(pte_t *pte, pte_t *_pte)
2019{
2020 while (--_pte >= pte) {
2021 pte_t pteval = *_pte;
2022 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2023 release_pte_page(pte_page(pteval));
2024 }
2025}
2026
2027static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2028 unsigned long address,
2029 pte_t *pte)
2030{
2031 struct page *page = NULL;
2032 pte_t *_pte;
2033 int none_or_zero = 0, result = 0;
2034 bool referenced = false, writable = false;
2035
2036 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2037 _pte++, address += PAGE_SIZE) {
2038 pte_t pteval = *_pte;
2039 if (pte_none(pteval) || (pte_present(pteval) &&
2040 is_zero_pfn(pte_pfn(pteval)))) {
2041 if (!userfaultfd_armed(vma) &&
2042 ++none_or_zero <= khugepaged_max_ptes_none) {
2043 continue;
2044 } else {
2045 result = SCAN_EXCEED_NONE_PTE;
2046 goto out;
2047 }
2048 }
2049 if (!pte_present(pteval)) {
2050 result = SCAN_PTE_NON_PRESENT;
2051 goto out;
2052 }
2053 page = vm_normal_page(vma, address, pteval);
2054 if (unlikely(!page)) {
2055 result = SCAN_PAGE_NULL;
2056 goto out;
2057 }
2058
2059 VM_BUG_ON_PAGE(PageCompound(page), page);
2060 VM_BUG_ON_PAGE(!PageAnon(page), page);
2061 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2062
2063 /*
2064 * We can do it before isolate_lru_page because the
2065 * page can't be freed from under us. NOTE: PG_lock
2066 * is needed to serialize against split_huge_page
2067 * when invoked from the VM.
2068 */
2069 if (!trylock_page(page)) {
2070 result = SCAN_PAGE_LOCK;
2071 goto out;
2072 }
2073
2074 /*
2075 * cannot use mapcount: can't collapse if there's a gup pin.
2076 * The page must only be referenced by the scanned process
2077 * and page swap cache.
2078 */
2079 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2080 unlock_page(page);
2081 result = SCAN_PAGE_COUNT;
2082 goto out;
2083 }
2084 if (pte_write(pteval)) {
2085 writable = true;
2086 } else {
2087 if (PageSwapCache(page) &&
2088 !reuse_swap_page(page, NULL)) {
2089 unlock_page(page);
2090 result = SCAN_SWAP_CACHE_PAGE;
2091 goto out;
2092 }
2093 /*
2094 * Page is not in the swap cache. It can be collapsed
2095 * into a THP.
2096 */
2097 }
2098
2099 /*
2100 * Isolate the page to avoid collapsing an hugepage
2101 * currently in use by the VM.
2102 */
2103 if (isolate_lru_page(page)) {
2104 unlock_page(page);
2105 result = SCAN_DEL_PAGE_LRU;
2106 goto out;
2107 }
2108 /* 0 stands for page_is_file_cache(page) == false */
2109 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2110 VM_BUG_ON_PAGE(!PageLocked(page), page);
2111 VM_BUG_ON_PAGE(PageLRU(page), page);
2112
2113 /* If there is no mapped pte young don't collapse the page */
2114 if (pte_young(pteval) ||
2115 page_is_young(page) || PageReferenced(page) ||
2116 mmu_notifier_test_young(vma->vm_mm, address))
2117 referenced = true;
2118 }
2119 if (likely(writable)) {
2120 if (likely(referenced)) {
2121 result = SCAN_SUCCEED;
2122 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2123 referenced, writable, result);
2124 return 1;
2125 }
2126 } else {
2127 result = SCAN_PAGE_RO;
2128 }
2129
2130out:
2131 release_pte_pages(pte, _pte);
2132 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2133 referenced, writable, result);
2134 return 0;
2135}
2136
2137static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2138 struct vm_area_struct *vma,
2139 unsigned long address,
2140 spinlock_t *ptl)
2141{
2142 pte_t *_pte;
2143 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2144 pte_t pteval = *_pte;
2145 struct page *src_page;
2146
2147 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2148 clear_user_highpage(page, address);
2149 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2150 if (is_zero_pfn(pte_pfn(pteval))) {
2151 /*
2152 * ptl mostly unnecessary.
2153 */
2154 spin_lock(ptl);
2155 /*
2156 * paravirt calls inside pte_clear here are
2157 * superfluous.
2158 */
2159 pte_clear(vma->vm_mm, address, _pte);
2160 spin_unlock(ptl);
2161 }
2162 } else {
2163 src_page = pte_page(pteval);
2164 copy_user_highpage(page, src_page, address, vma);
2165 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2166 release_pte_page(src_page);
2167 /*
2168 * ptl mostly unnecessary, but preempt has to
2169 * be disabled to update the per-cpu stats
2170 * inside page_remove_rmap().
2171 */
2172 spin_lock(ptl);
2173 /*
2174 * paravirt calls inside pte_clear here are
2175 * superfluous.
2176 */
2177 pte_clear(vma->vm_mm, address, _pte);
2178 page_remove_rmap(src_page, false);
2179 spin_unlock(ptl);
2180 free_page_and_swap_cache(src_page);
2181 }
2182
2183 address += PAGE_SIZE;
2184 page++;
2185 }
2186}
2187
2188static void khugepaged_alloc_sleep(void)
2189{
2190 DEFINE_WAIT(wait);
2191
2192 add_wait_queue(&khugepaged_wait, &wait);
2193 freezable_schedule_timeout_interruptible(
2194 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2195 remove_wait_queue(&khugepaged_wait, &wait);
2196}
2197
2198static int khugepaged_node_load[MAX_NUMNODES];
2199
2200static bool khugepaged_scan_abort(int nid)
2201{
2202 int i;
2203
2204 /*
2205 * If zone_reclaim_mode is disabled, then no extra effort is made to
2206 * allocate memory locally.
2207 */
2208 if (!zone_reclaim_mode)
2209 return false;
2210
2211 /* If there is a count for this node already, it must be acceptable */
2212 if (khugepaged_node_load[nid])
2213 return false;
2214
2215 for (i = 0; i < MAX_NUMNODES; i++) {
2216 if (!khugepaged_node_load[i])
2217 continue;
2218 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2219 return true;
2220 }
2221 return false;
2222}
2223
2224#ifdef CONFIG_NUMA
2225static int khugepaged_find_target_node(void)
2226{
2227 static int last_khugepaged_target_node = NUMA_NO_NODE;
2228 int nid, target_node = 0, max_value = 0;
2229
2230 /* find first node with max normal pages hit */
2231 for (nid = 0; nid < MAX_NUMNODES; nid++)
2232 if (khugepaged_node_load[nid] > max_value) {
2233 max_value = khugepaged_node_load[nid];
2234 target_node = nid;
2235 }
2236
2237 /* do some balance if several nodes have the same hit record */
2238 if (target_node <= last_khugepaged_target_node)
2239 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2240 nid++)
2241 if (max_value == khugepaged_node_load[nid]) {
2242 target_node = nid;
2243 break;
2244 }
2245
2246 last_khugepaged_target_node = target_node;
2247 return target_node;
2248}
2249
2250static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2251{
2252 if (IS_ERR(*hpage)) {
2253 if (!*wait)
2254 return false;
2255
2256 *wait = false;
2257 *hpage = NULL;
2258 khugepaged_alloc_sleep();
2259 } else if (*hpage) {
2260 put_page(*hpage);
2261 *hpage = NULL;
2262 }
2263
2264 return true;
2265}
2266
2267static struct page *
2268khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2269 unsigned long address, int node)
2270{
2271 VM_BUG_ON_PAGE(*hpage, *hpage);
2272
2273 /*
2274 * Before allocating the hugepage, release the mmap_sem read lock.
2275 * The allocation can take potentially a long time if it involves
2276 * sync compaction, and we do not need to hold the mmap_sem during
2277 * that. We will recheck the vma after taking it again in write mode.
2278 */
2279 up_read(&mm->mmap_sem);
2280
2281 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2282 if (unlikely(!*hpage)) {
2283 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2284 *hpage = ERR_PTR(-ENOMEM);
2285 return NULL;
2286 }
2287
2288 prep_transhuge_page(*hpage);
2289 count_vm_event(THP_COLLAPSE_ALLOC);
2290 return *hpage;
2291}
2292#else
2293static int khugepaged_find_target_node(void)
2294{
2295 return 0;
2296}
2297
2298static inline struct page *alloc_khugepaged_hugepage(void)
2299{
2300 struct page *page;
2301
2302 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
2303 HPAGE_PMD_ORDER);
2304 if (page)
2305 prep_transhuge_page(page);
2306 return page;
2307}
2308
2309static struct page *khugepaged_alloc_hugepage(bool *wait)
2310{
2311 struct page *hpage;
2312
2313 do {
2314 hpage = alloc_khugepaged_hugepage();
2315 if (!hpage) {
2316 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2317 if (!*wait)
2318 return NULL;
2319
2320 *wait = false;
2321 khugepaged_alloc_sleep();
2322 } else
2323 count_vm_event(THP_COLLAPSE_ALLOC);
2324 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2325
2326 return hpage;
2327}
2328
2329static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2330{
2331 if (!*hpage)
2332 *hpage = khugepaged_alloc_hugepage(wait);
2333
2334 if (unlikely(!*hpage))
2335 return false;
2336
2337 return true;
2338}
2339
2340static struct page *
2341khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2342 unsigned long address, int node)
2343{
2344 up_read(&mm->mmap_sem);
2345 VM_BUG_ON(!*hpage);
2346
2347 return *hpage;
2348}
2349#endif
2350
2351static bool hugepage_vma_check(struct vm_area_struct *vma)
2352{
2353 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2354 (vma->vm_flags & VM_NOHUGEPAGE))
2355 return false;
2356 if (!vma->anon_vma || vma->vm_ops)
2357 return false;
2358 if (is_vma_temporary_stack(vma))
2359 return false;
2360 return !(vma->vm_flags & VM_NO_KHUGEPAGED);
2361}
2362
2363/*
2364 * If mmap_sem temporarily dropped, revalidate vma
2365 * before taking mmap_sem.
2366 * Return 0 if succeeds, otherwise return none-zero
2367 * value (scan code).
2368 */
2369
2370static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address)
2371{
2372 struct vm_area_struct *vma;
2373 unsigned long hstart, hend;
2374
2375 if (unlikely(khugepaged_test_exit(mm)))
2376 return SCAN_ANY_PROCESS;
2377
2378 vma = find_vma(mm, address);
2379 if (!vma)
2380 return SCAN_VMA_NULL;
2381
2382 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2383 hend = vma->vm_end & HPAGE_PMD_MASK;
2384 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2385 return SCAN_ADDRESS_RANGE;
2386 if (!hugepage_vma_check(vma))
2387 return SCAN_VMA_CHECK;
2388 return 0;
2389}
2390
2391/*
2392 * Bring missing pages in from swap, to complete THP collapse.
2393 * Only done if khugepaged_scan_pmd believes it is worthwhile.
2394 *
2395 * Called and returns without pte mapped or spinlocks held,
2396 * but with mmap_sem held to protect against vma changes.
2397 */
2398
2399static bool __collapse_huge_page_swapin(struct mm_struct *mm,
2400 struct vm_area_struct *vma,
2401 unsigned long address, pmd_t *pmd)
2402{
2403 pte_t pteval;
2404 int swapped_in = 0, ret = 0;
2405 struct fault_env fe = {
2406 .vma = vma,
2407 .address = address,
2408 .flags = FAULT_FLAG_ALLOW_RETRY,
2409 .pmd = pmd,
2410 };
2411
2412 fe.pte = pte_offset_map(pmd, address);
2413 for (; fe.address < address + HPAGE_PMD_NR*PAGE_SIZE;
2414 fe.pte++, fe.address += PAGE_SIZE) {
2415 pteval = *fe.pte;
2416 if (!is_swap_pte(pteval))
2417 continue;
2418 swapped_in++;
2419 ret = do_swap_page(&fe, pteval);
2420 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
2421 if (ret & VM_FAULT_RETRY) {
2422 down_read(&mm->mmap_sem);
2423 /* vma is no longer available, don't continue to swapin */
2424 if (hugepage_vma_revalidate(mm, address))
2425 return false;
2426 /* check if the pmd is still valid */
2427 if (mm_find_pmd(mm, address) != pmd)
2428 return false;
2429 }
2430 if (ret & VM_FAULT_ERROR) {
2431 trace_mm_collapse_huge_page_swapin(mm, swapped_in, 0);
2432 return false;
2433 }
2434 /* pte is unmapped now, we need to map it */
2435 fe.pte = pte_offset_map(pmd, fe.address);
2436 }
2437 fe.pte--;
2438 pte_unmap(fe.pte);
2439 trace_mm_collapse_huge_page_swapin(mm, swapped_in, 1);
2440 return true;
2441}
2442
2443static void collapse_huge_page(struct mm_struct *mm,
2444 unsigned long address,
2445 struct page **hpage,
2446 struct vm_area_struct *vma,
2447 int node)
2448{
2449 pmd_t *pmd, _pmd;
2450 pte_t *pte;
2451 pgtable_t pgtable;
2452 struct page *new_page;
2453 spinlock_t *pmd_ptl, *pte_ptl;
2454 int isolated = 0, result = 0;
2455 struct mem_cgroup *memcg;
2456 unsigned long mmun_start; /* For mmu_notifiers */
2457 unsigned long mmun_end; /* For mmu_notifiers */
2458 gfp_t gfp;
2459
2460 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2461
2462 /* Only allocate from the target node */
2463 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
2464
2465 /* release the mmap_sem read lock. */
2466 new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2467 if (!new_page) {
2468 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2469 goto out_nolock;
2470 }
2471
2472 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2473 result = SCAN_CGROUP_CHARGE_FAIL;
2474 goto out_nolock;
2475 }
2476
2477 down_read(&mm->mmap_sem);
2478 result = hugepage_vma_revalidate(mm, address);
2479 if (result) {
2480 mem_cgroup_cancel_charge(new_page, memcg, true);
2481 up_read(&mm->mmap_sem);
2482 goto out_nolock;
2483 }
2484
2485 pmd = mm_find_pmd(mm, address);
2486 if (!pmd) {
2487 result = SCAN_PMD_NULL;
2488 mem_cgroup_cancel_charge(new_page, memcg, true);
2489 up_read(&mm->mmap_sem);
2490 goto out_nolock;
2491 }
2492
2493 /*
2494 * __collapse_huge_page_swapin always returns with mmap_sem locked.
2495 * If it fails, release mmap_sem and jump directly out.
2496 * Continuing to collapse causes inconsistency.
2497 */
2498 if (!__collapse_huge_page_swapin(mm, vma, address, pmd)) {
2499 mem_cgroup_cancel_charge(new_page, memcg, true);
2500 up_read(&mm->mmap_sem);
2501 goto out_nolock;
2502 }
2503
2504 up_read(&mm->mmap_sem);
2505 /*
2506 * Prevent all access to pagetables with the exception of
2507 * gup_fast later handled by the ptep_clear_flush and the VM
2508 * handled by the anon_vma lock + PG_lock.
2509 */
2510 down_write(&mm->mmap_sem);
2511 result = hugepage_vma_revalidate(mm, address);
2512 if (result)
2513 goto out;
2514 /* check if the pmd is still valid */
2515 if (mm_find_pmd(mm, address) != pmd)
2516 goto out;
2517
2518 anon_vma_lock_write(vma->anon_vma);
2519
2520 pte = pte_offset_map(pmd, address);
2521 pte_ptl = pte_lockptr(mm, pmd);
2522
2523 mmun_start = address;
2524 mmun_end = address + HPAGE_PMD_SIZE;
2525 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2526 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2527 /*
2528 * After this gup_fast can't run anymore. This also removes
2529 * any huge TLB entry from the CPU so we won't allow
2530 * huge and small TLB entries for the same virtual address
2531 * to avoid the risk of CPU bugs in that area.
2532 */
2533 _pmd = pmdp_collapse_flush(vma, address, pmd);
2534 spin_unlock(pmd_ptl);
2535 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2536
2537 spin_lock(pte_ptl);
2538 isolated = __collapse_huge_page_isolate(vma, address, pte);
2539 spin_unlock(pte_ptl);
2540
2541 if (unlikely(!isolated)) {
2542 pte_unmap(pte);
2543 spin_lock(pmd_ptl);
2544 BUG_ON(!pmd_none(*pmd));
2545 /*
2546 * We can only use set_pmd_at when establishing
2547 * hugepmds and never for establishing regular pmds that
2548 * points to regular pagetables. Use pmd_populate for that
2549 */
2550 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2551 spin_unlock(pmd_ptl);
2552 anon_vma_unlock_write(vma->anon_vma);
2553 result = SCAN_FAIL;
2554 goto out;
2555 }
2556
2557 /*
2558 * All pages are isolated and locked so anon_vma rmap
2559 * can't run anymore.
2560 */
2561 anon_vma_unlock_write(vma->anon_vma);
2562
2563 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2564 pte_unmap(pte);
2565 __SetPageUptodate(new_page);
2566 pgtable = pmd_pgtable(_pmd);
2567
2568 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2569 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2570
2571 /*
2572 * spin_lock() below is not the equivalent of smp_wmb(), so
2573 * this is needed to avoid the copy_huge_page writes to become
2574 * visible after the set_pmd_at() write.
2575 */
2576 smp_wmb();
2577
2578 spin_lock(pmd_ptl);
2579 BUG_ON(!pmd_none(*pmd));
2580 page_add_new_anon_rmap(new_page, vma, address, true);
2581 mem_cgroup_commit_charge(new_page, memcg, false, true);
2582 lru_cache_add_active_or_unevictable(new_page, vma);
2583 pgtable_trans_huge_deposit(mm, pmd, pgtable);
2584 set_pmd_at(mm, address, pmd, _pmd);
2585 update_mmu_cache_pmd(vma, address, pmd);
2586 spin_unlock(pmd_ptl);
2587
2588 *hpage = NULL;
2589
2590 khugepaged_pages_collapsed++;
2591 result = SCAN_SUCCEED;
2592out_up_write:
2593 up_write(&mm->mmap_sem);
2594out_nolock:
2595 trace_mm_collapse_huge_page(mm, isolated, result);
2596 return;
2597out:
2598 mem_cgroup_cancel_charge(new_page, memcg, true);
2599 goto out_up_write;
2600}
2601
2602static int khugepaged_scan_pmd(struct mm_struct *mm,
2603 struct vm_area_struct *vma,
2604 unsigned long address,
2605 struct page **hpage)
2606{
2607 pmd_t *pmd;
2608 pte_t *pte, *_pte;
2609 int ret = 0, none_or_zero = 0, result = 0;
2610 struct page *page = NULL;
2611 unsigned long _address;
2612 spinlock_t *ptl;
2613 int node = NUMA_NO_NODE, unmapped = 0;
2614 bool writable = false, referenced = false;
2615
2616 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2617
2618 pmd = mm_find_pmd(mm, address);
2619 if (!pmd) {
2620 result = SCAN_PMD_NULL;
2621 goto out;
2622 }
2623
2624 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2625 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2626 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2627 _pte++, _address += PAGE_SIZE) {
2628 pte_t pteval = *_pte;
2629 if (is_swap_pte(pteval)) {
2630 if (++unmapped <= khugepaged_max_ptes_swap) {
2631 continue;
2632 } else {
2633 result = SCAN_EXCEED_SWAP_PTE;
2634 goto out_unmap;
2635 }
2636 }
2637 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2638 if (!userfaultfd_armed(vma) &&
2639 ++none_or_zero <= khugepaged_max_ptes_none) {
2640 continue;
2641 } else {
2642 result = SCAN_EXCEED_NONE_PTE;
2643 goto out_unmap;
2644 }
2645 }
2646 if (!pte_present(pteval)) {
2647 result = SCAN_PTE_NON_PRESENT;
2648 goto out_unmap;
2649 }
2650 if (pte_write(pteval))
2651 writable = true;
2652
2653 page = vm_normal_page(vma, _address, pteval);
2654 if (unlikely(!page)) {
2655 result = SCAN_PAGE_NULL;
2656 goto out_unmap;
2657 }
2658
2659 /* TODO: teach khugepaged to collapse THP mapped with pte */
2660 if (PageCompound(page)) {
2661 result = SCAN_PAGE_COMPOUND;
2662 goto out_unmap;
2663 }
2664
2665 /*
2666 * Record which node the original page is from and save this
2667 * information to khugepaged_node_load[].
2668 * Khupaged will allocate hugepage from the node has the max
2669 * hit record.
2670 */
2671 node = page_to_nid(page);
2672 if (khugepaged_scan_abort(node)) {
2673 result = SCAN_SCAN_ABORT;
2674 goto out_unmap;
2675 }
2676 khugepaged_node_load[node]++;
2677 if (!PageLRU(page)) {
2678 result = SCAN_PAGE_LRU;
2679 goto out_unmap;
2680 }
2681 if (PageLocked(page)) {
2682 result = SCAN_PAGE_LOCK;
2683 goto out_unmap;
2684 }
2685 if (!PageAnon(page)) {
2686 result = SCAN_PAGE_ANON;
2687 goto out_unmap;
2688 }
2689
2690 /*
2691 * cannot use mapcount: can't collapse if there's a gup pin.
2692 * The page must only be referenced by the scanned process
2693 * and page swap cache.
2694 */
2695 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2696 result = SCAN_PAGE_COUNT;
2697 goto out_unmap;
2698 }
2699 if (pte_young(pteval) ||
2700 page_is_young(page) || PageReferenced(page) ||
2701 mmu_notifier_test_young(vma->vm_mm, address))
2702 referenced = true;
2703 }
2704 if (writable) {
2705 if (referenced) {
2706 result = SCAN_SUCCEED;
2707 ret = 1;
2708 } else {
2709 result = SCAN_NO_REFERENCED_PAGE;
2710 }
2711 } else {
2712 result = SCAN_PAGE_RO;
2713 }
2714out_unmap:
2715 pte_unmap_unlock(pte, ptl);
2716 if (ret) {
2717 node = khugepaged_find_target_node();
2718 /* collapse_huge_page will return with the mmap_sem released */
2719 collapse_huge_page(mm, address, hpage, vma, node);
2720 }
2721out:
2722 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
2723 none_or_zero, result, unmapped);
2724 return ret;
2725}
2726
2727static void collect_mm_slot(struct mm_slot *mm_slot)
2728{
2729 struct mm_struct *mm = mm_slot->mm;
2730
2731 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2732
2733 if (khugepaged_test_exit(mm)) {
2734 /* free mm_slot */
2735 hash_del(&mm_slot->hash);
2736 list_del(&mm_slot->mm_node);
2737
2738 /*
2739 * Not strictly needed because the mm exited already.
2740 *
2741 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2742 */
2743
2744 /* khugepaged_mm_lock actually not necessary for the below */
2745 free_mm_slot(mm_slot);
2746 mmdrop(mm);
2747 }
2748}
2749
2750static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2751 struct page **hpage)
2752 __releases(&khugepaged_mm_lock)
2753 __acquires(&khugepaged_mm_lock)
2754{
2755 struct mm_slot *mm_slot;
2756 struct mm_struct *mm;
2757 struct vm_area_struct *vma;
2758 int progress = 0;
2759
2760 VM_BUG_ON(!pages);
2761 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2762
2763 if (khugepaged_scan.mm_slot)
2764 mm_slot = khugepaged_scan.mm_slot;
2765 else {
2766 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2767 struct mm_slot, mm_node);
2768 khugepaged_scan.address = 0;
2769 khugepaged_scan.mm_slot = mm_slot;
2770 }
2771 spin_unlock(&khugepaged_mm_lock);
2772
2773 mm = mm_slot->mm;
2774 down_read(&mm->mmap_sem);
2775 if (unlikely(khugepaged_test_exit(mm)))
2776 vma = NULL;
2777 else
2778 vma = find_vma(mm, khugepaged_scan.address);
2779
2780 progress++;
2781 for (; vma; vma = vma->vm_next) {
2782 unsigned long hstart, hend;
2783
2784 cond_resched();
2785 if (unlikely(khugepaged_test_exit(mm))) {
2786 progress++;
2787 break;
2788 }
2789 if (!hugepage_vma_check(vma)) {
2790skip:
2791 progress++;
2792 continue;
2793 }
2794 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2795 hend = vma->vm_end & HPAGE_PMD_MASK;
2796 if (hstart >= hend)
2797 goto skip;
2798 if (khugepaged_scan.address > hend)
2799 goto skip;
2800 if (khugepaged_scan.address < hstart)
2801 khugepaged_scan.address = hstart;
2802 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2803
2804 while (khugepaged_scan.address < hend) {
2805 int ret;
2806 cond_resched();
2807 if (unlikely(khugepaged_test_exit(mm)))
2808 goto breakouterloop;
2809
2810 VM_BUG_ON(khugepaged_scan.address < hstart ||
2811 khugepaged_scan.address + HPAGE_PMD_SIZE >
2812 hend);
2813 ret = khugepaged_scan_pmd(mm, vma,
2814 khugepaged_scan.address,
2815 hpage);
2816 /* move to next address */
2817 khugepaged_scan.address += HPAGE_PMD_SIZE;
2818 progress += HPAGE_PMD_NR;
2819 if (ret)
2820 /* we released mmap_sem so break loop */
2821 goto breakouterloop_mmap_sem;
2822 if (progress >= pages)
2823 goto breakouterloop;
2824 }
2825 }
2826breakouterloop:
2827 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2828breakouterloop_mmap_sem:
2829
2830 spin_lock(&khugepaged_mm_lock);
2831 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2832 /*
2833 * Release the current mm_slot if this mm is about to die, or
2834 * if we scanned all vmas of this mm.
2835 */
2836 if (khugepaged_test_exit(mm) || !vma) {
2837 /*
2838 * Make sure that if mm_users is reaching zero while
2839 * khugepaged runs here, khugepaged_exit will find
2840 * mm_slot not pointing to the exiting mm.
2841 */
2842 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2843 khugepaged_scan.mm_slot = list_entry(
2844 mm_slot->mm_node.next,
2845 struct mm_slot, mm_node);
2846 khugepaged_scan.address = 0;
2847 } else {
2848 khugepaged_scan.mm_slot = NULL;
2849 khugepaged_full_scans++;
2850 }
2851
2852 collect_mm_slot(mm_slot);
2853 }
2854
2855 return progress;
2856}
2857
2858static int khugepaged_has_work(void)
2859{
2860 return !list_empty(&khugepaged_scan.mm_head) &&
2861 khugepaged_enabled();
2862}
2863
2864static int khugepaged_wait_event(void)
2865{
2866 return !list_empty(&khugepaged_scan.mm_head) ||
2867 kthread_should_stop();
2868}
2869
2870static void khugepaged_do_scan(void)
2871{
2872 struct page *hpage = NULL;
2873 unsigned int progress = 0, pass_through_head = 0;
2874 unsigned int pages = khugepaged_pages_to_scan;
2875 bool wait = true;
2876
2877 barrier(); /* write khugepaged_pages_to_scan to local stack */
2878
2879 while (progress < pages) {
2880 if (!khugepaged_prealloc_page(&hpage, &wait))
2881 break;
2882
2883 cond_resched();
2884
2885 if (unlikely(kthread_should_stop() || try_to_freeze()))
2886 break;
2887
2888 spin_lock(&khugepaged_mm_lock);
2889 if (!khugepaged_scan.mm_slot)
2890 pass_through_head++;
2891 if (khugepaged_has_work() &&
2892 pass_through_head < 2)
2893 progress += khugepaged_scan_mm_slot(pages - progress,
2894 &hpage);
2895 else
2896 progress = pages;
2897 spin_unlock(&khugepaged_mm_lock);
2898 }
2899
2900 if (!IS_ERR_OR_NULL(hpage))
2901 put_page(hpage);
2902}
2903
2904static bool khugepaged_should_wakeup(void)
2905{
2906 return kthread_should_stop() ||
2907 time_after_eq(jiffies, khugepaged_sleep_expire);
2908}
2909
2910static void khugepaged_wait_work(void)
2911{
2912 if (khugepaged_has_work()) {
2913 const unsigned long scan_sleep_jiffies =
2914 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2915
2916 if (!scan_sleep_jiffies)
2917 return;
2918
2919 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2920 wait_event_freezable_timeout(khugepaged_wait,
2921 khugepaged_should_wakeup(),
2922 scan_sleep_jiffies);
2923 return;
2924 }
2925
2926 if (khugepaged_enabled())
2927 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2928}
2929
2930static int khugepaged(void *none)
2931{
2932 struct mm_slot *mm_slot;
2933
2934 set_freezable();
2935 set_user_nice(current, MAX_NICE);
2936
2937 while (!kthread_should_stop()) {
2938 khugepaged_do_scan();
2939 khugepaged_wait_work();
2940 }
2941
2942 spin_lock(&khugepaged_mm_lock);
2943 mm_slot = khugepaged_scan.mm_slot;
2944 khugepaged_scan.mm_slot = NULL;
2945 if (mm_slot)
2946 collect_mm_slot(mm_slot);
2947 spin_unlock(&khugepaged_mm_lock);
2948 return 0;
2949}
2950
2951static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, 1476static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2952 unsigned long haddr, pmd_t *pmd) 1477 unsigned long haddr, pmd_t *pmd)
2953{ 1478{
diff --git a/mm/khugepaged.c b/mm/khugepaged.c
new file mode 100644
index 000000000000..3e6d1a1b7e2c
--- /dev/null
+++ b/mm/khugepaged.c
@@ -0,0 +1,1490 @@
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
18#include <asm/tlb.h>
19#include <asm/pgalloc.h>
20#include "internal.h"
21
22enum scan_result {
23 SCAN_FAIL,
24 SCAN_SUCCEED,
25 SCAN_PMD_NULL,
26 SCAN_EXCEED_NONE_PTE,
27 SCAN_PTE_NON_PRESENT,
28 SCAN_PAGE_RO,
29 SCAN_NO_REFERENCED_PAGE,
30 SCAN_PAGE_NULL,
31 SCAN_SCAN_ABORT,
32 SCAN_PAGE_COUNT,
33 SCAN_PAGE_LRU,
34 SCAN_PAGE_LOCK,
35 SCAN_PAGE_ANON,
36 SCAN_PAGE_COMPOUND,
37 SCAN_ANY_PROCESS,
38 SCAN_VMA_NULL,
39 SCAN_VMA_CHECK,
40 SCAN_ADDRESS_RANGE,
41 SCAN_SWAP_CACHE_PAGE,
42 SCAN_DEL_PAGE_LRU,
43 SCAN_ALLOC_HUGE_PAGE_FAIL,
44 SCAN_CGROUP_CHARGE_FAIL,
45 SCAN_EXCEED_SWAP_PTE
46};
47
48#define CREATE_TRACE_POINTS
49#include <trace/events/huge_memory.h>
50
51/* default scan 8*512 pte (or vmas) every 30 second */
52static unsigned int khugepaged_pages_to_scan __read_mostly;
53static unsigned int khugepaged_pages_collapsed;
54static unsigned int khugepaged_full_scans;
55static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
56/* during fragmentation poll the hugepage allocator once every minute */
57static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
58static unsigned long khugepaged_sleep_expire;
59static DEFINE_SPINLOCK(khugepaged_mm_lock);
60static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
61/*
62 * default collapse hugepages if there is at least one pte mapped like
63 * it would have happened if the vma was large enough during page
64 * fault.
65 */
66static unsigned int khugepaged_max_ptes_none __read_mostly;
67static unsigned int khugepaged_max_ptes_swap __read_mostly;
68
69#define MM_SLOTS_HASH_BITS 10
70static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
71
72static struct kmem_cache *mm_slot_cache __read_mostly;
73
74/**
75 * struct mm_slot - hash lookup from mm to mm_slot
76 * @hash: hash collision list
77 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
78 * @mm: the mm that this information is valid for
79 */
80struct mm_slot {
81 struct hlist_node hash;
82 struct list_head mm_node;
83 struct mm_struct *mm;
84};
85
86/**
87 * struct khugepaged_scan - cursor for scanning
88 * @mm_head: the head of the mm list to scan
89 * @mm_slot: the current mm_slot we are scanning
90 * @address: the next address inside that to be scanned
91 *
92 * There is only the one khugepaged_scan instance of this cursor structure.
93 */
94struct khugepaged_scan {
95 struct list_head mm_head;
96 struct mm_slot *mm_slot;
97 unsigned long address;
98};
99
100static struct khugepaged_scan khugepaged_scan = {
101 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
102};
103
104static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
105 struct kobj_attribute *attr,
106 char *buf)
107{
108 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
109}
110
111static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
112 struct kobj_attribute *attr,
113 const char *buf, size_t count)
114{
115 unsigned long msecs;
116 int err;
117
118 err = kstrtoul(buf, 10, &msecs);
119 if (err || msecs > UINT_MAX)
120 return -EINVAL;
121
122 khugepaged_scan_sleep_millisecs = msecs;
123 khugepaged_sleep_expire = 0;
124 wake_up_interruptible(&khugepaged_wait);
125
126 return count;
127}
128static struct kobj_attribute scan_sleep_millisecs_attr =
129 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
130 scan_sleep_millisecs_store);
131
132static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
133 struct kobj_attribute *attr,
134 char *buf)
135{
136 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
137}
138
139static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
140 struct kobj_attribute *attr,
141 const char *buf, size_t count)
142{
143 unsigned long msecs;
144 int err;
145
146 err = kstrtoul(buf, 10, &msecs);
147 if (err || msecs > UINT_MAX)
148 return -EINVAL;
149
150 khugepaged_alloc_sleep_millisecs = msecs;
151 khugepaged_sleep_expire = 0;
152 wake_up_interruptible(&khugepaged_wait);
153
154 return count;
155}
156static struct kobj_attribute alloc_sleep_millisecs_attr =
157 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
158 alloc_sleep_millisecs_store);
159
160static ssize_t pages_to_scan_show(struct kobject *kobj,
161 struct kobj_attribute *attr,
162 char *buf)
163{
164 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
165}
166static ssize_t pages_to_scan_store(struct kobject *kobj,
167 struct kobj_attribute *attr,
168 const char *buf, size_t count)
169{
170 int err;
171 unsigned long pages;
172
173 err = kstrtoul(buf, 10, &pages);
174 if (err || !pages || pages > UINT_MAX)
175 return -EINVAL;
176
177 khugepaged_pages_to_scan = pages;
178
179 return count;
180}
181static struct kobj_attribute pages_to_scan_attr =
182 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
183 pages_to_scan_store);
184
185static ssize_t pages_collapsed_show(struct kobject *kobj,
186 struct kobj_attribute *attr,
187 char *buf)
188{
189 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
190}
191static struct kobj_attribute pages_collapsed_attr =
192 __ATTR_RO(pages_collapsed);
193
194static ssize_t full_scans_show(struct kobject *kobj,
195 struct kobj_attribute *attr,
196 char *buf)
197{
198 return sprintf(buf, "%u\n", khugepaged_full_scans);
199}
200static struct kobj_attribute full_scans_attr =
201 __ATTR_RO(full_scans);
202
203static ssize_t khugepaged_defrag_show(struct kobject *kobj,
204 struct kobj_attribute *attr, char *buf)
205{
206 return single_hugepage_flag_show(kobj, attr, buf,
207 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
208}
209static ssize_t khugepaged_defrag_store(struct kobject *kobj,
210 struct kobj_attribute *attr,
211 const char *buf, size_t count)
212{
213 return single_hugepage_flag_store(kobj, attr, buf, count,
214 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
215}
216static struct kobj_attribute khugepaged_defrag_attr =
217 __ATTR(defrag, 0644, khugepaged_defrag_show,
218 khugepaged_defrag_store);
219
220/*
221 * max_ptes_none controls if khugepaged should collapse hugepages over
222 * any unmapped ptes in turn potentially increasing the memory
223 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
224 * reduce the available free memory in the system as it
225 * runs. Increasing max_ptes_none will instead potentially reduce the
226 * free memory in the system during the khugepaged scan.
227 */
228static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
229 struct kobj_attribute *attr,
230 char *buf)
231{
232 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
233}
234static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
235 struct kobj_attribute *attr,
236 const char *buf, size_t count)
237{
238 int err;
239 unsigned long max_ptes_none;
240
241 err = kstrtoul(buf, 10, &max_ptes_none);
242 if (err || max_ptes_none > HPAGE_PMD_NR-1)
243 return -EINVAL;
244
245 khugepaged_max_ptes_none = max_ptes_none;
246
247 return count;
248}
249static struct kobj_attribute khugepaged_max_ptes_none_attr =
250 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
251 khugepaged_max_ptes_none_store);
252
253static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
254 struct kobj_attribute *attr,
255 char *buf)
256{
257 return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
258}
259
260static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
261 struct kobj_attribute *attr,
262 const char *buf, size_t count)
263{
264 int err;
265 unsigned long max_ptes_swap;
266
267 err = kstrtoul(buf, 10, &max_ptes_swap);
268 if (err || max_ptes_swap > HPAGE_PMD_NR-1)
269 return -EINVAL;
270
271 khugepaged_max_ptes_swap = max_ptes_swap;
272
273 return count;
274}
275
276static struct kobj_attribute khugepaged_max_ptes_swap_attr =
277 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
278 khugepaged_max_ptes_swap_store);
279
280static struct attribute *khugepaged_attr[] = {
281 &khugepaged_defrag_attr.attr,
282 &khugepaged_max_ptes_none_attr.attr,
283 &pages_to_scan_attr.attr,
284 &pages_collapsed_attr.attr,
285 &full_scans_attr.attr,
286 &scan_sleep_millisecs_attr.attr,
287 &alloc_sleep_millisecs_attr.attr,
288 &khugepaged_max_ptes_swap_attr.attr,
289 NULL,
290};
291
292struct attribute_group khugepaged_attr_group = {
293 .attrs = khugepaged_attr,
294 .name = "khugepaged",
295};
296
297#define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
298
299int hugepage_madvise(struct vm_area_struct *vma,
300 unsigned long *vm_flags, int advice)
301{
302 switch (advice) {
303 case MADV_HUGEPAGE:
304#ifdef CONFIG_S390
305 /*
306 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
307 * can't handle this properly after s390_enable_sie, so we simply
308 * ignore the madvise to prevent qemu from causing a SIGSEGV.
309 */
310 if (mm_has_pgste(vma->vm_mm))
311 return 0;
312#endif
313 *vm_flags &= ~VM_NOHUGEPAGE;
314 *vm_flags |= VM_HUGEPAGE;
315 /*
316 * If the vma become good for khugepaged to scan,
317 * register it here without waiting a page fault that
318 * may not happen any time soon.
319 */
320 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
321 khugepaged_enter_vma_merge(vma, *vm_flags))
322 return -ENOMEM;
323 break;
324 case MADV_NOHUGEPAGE:
325 *vm_flags &= ~VM_HUGEPAGE;
326 *vm_flags |= VM_NOHUGEPAGE;
327 /*
328 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
329 * this vma even if we leave the mm registered in khugepaged if
330 * it got registered before VM_NOHUGEPAGE was set.
331 */
332 break;
333 }
334
335 return 0;
336}
337
338int __init khugepaged_init(void)
339{
340 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
341 sizeof(struct mm_slot),
342 __alignof__(struct mm_slot), 0, NULL);
343 if (!mm_slot_cache)
344 return -ENOMEM;
345
346 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
347 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
348 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
349
350 return 0;
351}
352
353void __init khugepaged_destroy(void)
354{
355 kmem_cache_destroy(mm_slot_cache);
356}
357
358static inline struct mm_slot *alloc_mm_slot(void)
359{
360 if (!mm_slot_cache) /* initialization failed */
361 return NULL;
362 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
363}
364
365static inline void free_mm_slot(struct mm_slot *mm_slot)
366{
367 kmem_cache_free(mm_slot_cache, mm_slot);
368}
369
370static struct mm_slot *get_mm_slot(struct mm_struct *mm)
371{
372 struct mm_slot *mm_slot;
373
374 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
375 if (mm == mm_slot->mm)
376 return mm_slot;
377
378 return NULL;
379}
380
381static void insert_to_mm_slots_hash(struct mm_struct *mm,
382 struct mm_slot *mm_slot)
383{
384 mm_slot->mm = mm;
385 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
386}
387
388static inline int khugepaged_test_exit(struct mm_struct *mm)
389{
390 return atomic_read(&mm->mm_users) == 0;
391}
392
393int __khugepaged_enter(struct mm_struct *mm)
394{
395 struct mm_slot *mm_slot;
396 int wakeup;
397
398 mm_slot = alloc_mm_slot();
399 if (!mm_slot)
400 return -ENOMEM;
401
402 /* __khugepaged_exit() must not run from under us */
403 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
404 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
405 free_mm_slot(mm_slot);
406 return 0;
407 }
408
409 spin_lock(&khugepaged_mm_lock);
410 insert_to_mm_slots_hash(mm, mm_slot);
411 /*
412 * Insert just behind the scanning cursor, to let the area settle
413 * down a little.
414 */
415 wakeup = list_empty(&khugepaged_scan.mm_head);
416 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
417 spin_unlock(&khugepaged_mm_lock);
418
419 atomic_inc(&mm->mm_count);
420 if (wakeup)
421 wake_up_interruptible(&khugepaged_wait);
422
423 return 0;
424}
425
426int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
427 unsigned long vm_flags)
428{
429 unsigned long hstart, hend;
430 if (!vma->anon_vma)
431 /*
432 * Not yet faulted in so we will register later in the
433 * page fault if needed.
434 */
435 return 0;
436 if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
437 /* khugepaged not yet working on file or special mappings */
438 return 0;
439 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
440 hend = vma->vm_end & HPAGE_PMD_MASK;
441 if (hstart < hend)
442 return khugepaged_enter(vma, vm_flags);
443 return 0;
444}
445
446void __khugepaged_exit(struct mm_struct *mm)
447{
448 struct mm_slot *mm_slot;
449 int free = 0;
450
451 spin_lock(&khugepaged_mm_lock);
452 mm_slot = get_mm_slot(mm);
453 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
454 hash_del(&mm_slot->hash);
455 list_del(&mm_slot->mm_node);
456 free = 1;
457 }
458 spin_unlock(&khugepaged_mm_lock);
459
460 if (free) {
461 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
462 free_mm_slot(mm_slot);
463 mmdrop(mm);
464 } else if (mm_slot) {
465 /*
466 * This is required to serialize against
467 * khugepaged_test_exit() (which is guaranteed to run
468 * under mmap sem read mode). Stop here (after we
469 * return all pagetables will be destroyed) until
470 * khugepaged has finished working on the pagetables
471 * under the mmap_sem.
472 */
473 down_write(&mm->mmap_sem);
474 up_write(&mm->mmap_sem);
475 }
476}
477
478static void release_pte_page(struct page *page)
479{
480 /* 0 stands for page_is_file_cache(page) == false */
481 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
482 unlock_page(page);
483 putback_lru_page(page);
484}
485
486static void release_pte_pages(pte_t *pte, pte_t *_pte)
487{
488 while (--_pte >= pte) {
489 pte_t pteval = *_pte;
490 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
491 release_pte_page(pte_page(pteval));
492 }
493}
494
495static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
496 unsigned long address,
497 pte_t *pte)
498{
499 struct page *page = NULL;
500 pte_t *_pte;
501 int none_or_zero = 0, result = 0;
502 bool referenced = false, writable = false;
503
504 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
505 _pte++, address += PAGE_SIZE) {
506 pte_t pteval = *_pte;
507 if (pte_none(pteval) || (pte_present(pteval) &&
508 is_zero_pfn(pte_pfn(pteval)))) {
509 if (!userfaultfd_armed(vma) &&
510 ++none_or_zero <= khugepaged_max_ptes_none) {
511 continue;
512 } else {
513 result = SCAN_EXCEED_NONE_PTE;
514 goto out;
515 }
516 }
517 if (!pte_present(pteval)) {
518 result = SCAN_PTE_NON_PRESENT;
519 goto out;
520 }
521 page = vm_normal_page(vma, address, pteval);
522 if (unlikely(!page)) {
523 result = SCAN_PAGE_NULL;
524 goto out;
525 }
526
527 VM_BUG_ON_PAGE(PageCompound(page), page);
528 VM_BUG_ON_PAGE(!PageAnon(page), page);
529 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
530
531 /*
532 * We can do it before isolate_lru_page because the
533 * page can't be freed from under us. NOTE: PG_lock
534 * is needed to serialize against split_huge_page
535 * when invoked from the VM.
536 */
537 if (!trylock_page(page)) {
538 result = SCAN_PAGE_LOCK;
539 goto out;
540 }
541
542 /*
543 * cannot use mapcount: can't collapse if there's a gup pin.
544 * The page must only be referenced by the scanned process
545 * and page swap cache.
546 */
547 if (page_count(page) != 1 + !!PageSwapCache(page)) {
548 unlock_page(page);
549 result = SCAN_PAGE_COUNT;
550 goto out;
551 }
552 if (pte_write(pteval)) {
553 writable = true;
554 } else {
555 if (PageSwapCache(page) &&
556 !reuse_swap_page(page, NULL)) {
557 unlock_page(page);
558 result = SCAN_SWAP_CACHE_PAGE;
559 goto out;
560 }
561 /*
562 * Page is not in the swap cache. It can be collapsed
563 * into a THP.
564 */
565 }
566
567 /*
568 * Isolate the page to avoid collapsing an hugepage
569 * currently in use by the VM.
570 */
571 if (isolate_lru_page(page)) {
572 unlock_page(page);
573 result = SCAN_DEL_PAGE_LRU;
574 goto out;
575 }
576 /* 0 stands for page_is_file_cache(page) == false */
577 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
578 VM_BUG_ON_PAGE(!PageLocked(page), page);
579 VM_BUG_ON_PAGE(PageLRU(page), page);
580
581 /* If there is no mapped pte young don't collapse the page */
582 if (pte_young(pteval) ||
583 page_is_young(page) || PageReferenced(page) ||
584 mmu_notifier_test_young(vma->vm_mm, address))
585 referenced = true;
586 }
587 if (likely(writable)) {
588 if (likely(referenced)) {
589 result = SCAN_SUCCEED;
590 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
591 referenced, writable, result);
592 return 1;
593 }
594 } else {
595 result = SCAN_PAGE_RO;
596 }
597
598out:
599 release_pte_pages(pte, _pte);
600 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
601 referenced, writable, result);
602 return 0;
603}
604
605static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
606 struct vm_area_struct *vma,
607 unsigned long address,
608 spinlock_t *ptl)
609{
610 pte_t *_pte;
611 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
612 pte_t pteval = *_pte;
613 struct page *src_page;
614
615 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
616 clear_user_highpage(page, address);
617 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
618 if (is_zero_pfn(pte_pfn(pteval))) {
619 /*
620 * ptl mostly unnecessary.
621 */
622 spin_lock(ptl);
623 /*
624 * paravirt calls inside pte_clear here are
625 * superfluous.
626 */
627 pte_clear(vma->vm_mm, address, _pte);
628 spin_unlock(ptl);
629 }
630 } else {
631 src_page = pte_page(pteval);
632 copy_user_highpage(page, src_page, address, vma);
633 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
634 release_pte_page(src_page);
635 /*
636 * ptl mostly unnecessary, but preempt has to
637 * be disabled to update the per-cpu stats
638 * inside page_remove_rmap().
639 */
640 spin_lock(ptl);
641 /*
642 * paravirt calls inside pte_clear here are
643 * superfluous.
644 */
645 pte_clear(vma->vm_mm, address, _pte);
646 page_remove_rmap(src_page, false);
647 spin_unlock(ptl);
648 free_page_and_swap_cache(src_page);
649 }
650
651 address += PAGE_SIZE;
652 page++;
653 }
654}
655
656static void khugepaged_alloc_sleep(void)
657{
658 DEFINE_WAIT(wait);
659
660 add_wait_queue(&khugepaged_wait, &wait);
661 freezable_schedule_timeout_interruptible(
662 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
663 remove_wait_queue(&khugepaged_wait, &wait);
664}
665
666static int khugepaged_node_load[MAX_NUMNODES];
667
668static bool khugepaged_scan_abort(int nid)
669{
670 int i;
671
672 /*
673 * If zone_reclaim_mode is disabled, then no extra effort is made to
674 * allocate memory locally.
675 */
676 if (!zone_reclaim_mode)
677 return false;
678
679 /* If there is a count for this node already, it must be acceptable */
680 if (khugepaged_node_load[nid])
681 return false;
682
683 for (i = 0; i < MAX_NUMNODES; i++) {
684 if (!khugepaged_node_load[i])
685 continue;
686 if (node_distance(nid, i) > RECLAIM_DISTANCE)
687 return true;
688 }
689 return false;
690}
691
692/* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
693static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
694{
695 return GFP_TRANSHUGE | (khugepaged_defrag() ? __GFP_DIRECT_RECLAIM : 0);
696}
697
698#ifdef CONFIG_NUMA
699static int khugepaged_find_target_node(void)
700{
701 static int last_khugepaged_target_node = NUMA_NO_NODE;
702 int nid, target_node = 0, max_value = 0;
703
704 /* find first node with max normal pages hit */
705 for (nid = 0; nid < MAX_NUMNODES; nid++)
706 if (khugepaged_node_load[nid] > max_value) {
707 max_value = khugepaged_node_load[nid];
708 target_node = nid;
709 }
710
711 /* do some balance if several nodes have the same hit record */
712 if (target_node <= last_khugepaged_target_node)
713 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
714 nid++)
715 if (max_value == khugepaged_node_load[nid]) {
716 target_node = nid;
717 break;
718 }
719
720 last_khugepaged_target_node = target_node;
721 return target_node;
722}
723
724static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
725{
726 if (IS_ERR(*hpage)) {
727 if (!*wait)
728 return false;
729
730 *wait = false;
731 *hpage = NULL;
732 khugepaged_alloc_sleep();
733 } else if (*hpage) {
734 put_page(*hpage);
735 *hpage = NULL;
736 }
737
738 return true;
739}
740
741static struct page *
742khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
743 unsigned long address, int node)
744{
745 VM_BUG_ON_PAGE(*hpage, *hpage);
746
747 /*
748 * Before allocating the hugepage, release the mmap_sem read lock.
749 * The allocation can take potentially a long time if it involves
750 * sync compaction, and we do not need to hold the mmap_sem during
751 * that. We will recheck the vma after taking it again in write mode.
752 */
753 up_read(&mm->mmap_sem);
754
755 *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 *
815khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
816 unsigned long address, int node)
817{
818 up_read(&mm->mmap_sem);
819 VM_BUG_ON(!*hpage);
820
821 return *hpage;
822}
823#endif
824
825static bool hugepage_vma_check(struct vm_area_struct *vma)
826{
827 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
828 (vma->vm_flags & VM_NOHUGEPAGE))
829 return false;
830 if (!vma->anon_vma || vma->vm_ops)
831 return false;
832 if (is_vma_temporary_stack(vma))
833 return false;
834 return !(vma->vm_flags & VM_NO_KHUGEPAGED);
835}
836
837/*
838 * If mmap_sem temporarily dropped, revalidate vma
839 * before taking mmap_sem.
840 * Return 0 if succeeds, otherwise return none-zero
841 * value (scan code).
842 */
843
844static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address)
845{
846 struct vm_area_struct *vma;
847 unsigned long hstart, hend;
848
849 if (unlikely(khugepaged_test_exit(mm)))
850 return SCAN_ANY_PROCESS;
851
852 vma = find_vma(mm, address);
853 if (!vma)
854 return SCAN_VMA_NULL;
855
856 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
857 hend = vma->vm_end & HPAGE_PMD_MASK;
858 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
859 return SCAN_ADDRESS_RANGE;
860 if (!hugepage_vma_check(vma))
861 return SCAN_VMA_CHECK;
862 return 0;
863}
864
865/*
866 * Bring missing pages in from swap, to complete THP collapse.
867 * Only done if khugepaged_scan_pmd believes it is worthwhile.
868 *
869 * Called and returns without pte mapped or spinlocks held,
870 * but with mmap_sem held to protect against vma changes.
871 */
872
873static bool __collapse_huge_page_swapin(struct mm_struct *mm,
874 struct vm_area_struct *vma,
875 unsigned long address, pmd_t *pmd)
876{
877 pte_t pteval;
878 int swapped_in = 0, ret = 0;
879 struct fault_env fe = {
880 .vma = vma,
881 .address = address,
882 .flags = FAULT_FLAG_ALLOW_RETRY,
883 .pmd = pmd,
884 };
885
886 fe.pte = pte_offset_map(pmd, address);
887 for (; fe.address < address + HPAGE_PMD_NR*PAGE_SIZE;
888 fe.pte++, fe.address += PAGE_SIZE) {
889 pteval = *fe.pte;
890 if (!is_swap_pte(pteval))
891 continue;
892 swapped_in++;
893 ret = do_swap_page(&fe, pteval);
894 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
895 if (ret & VM_FAULT_RETRY) {
896 down_read(&mm->mmap_sem);
897 /* vma is no longer available, don't continue to swapin */
898 if (hugepage_vma_revalidate(mm, address))
899 return false;
900 /* check if the pmd is still valid */
901 if (mm_find_pmd(mm, address) != pmd)
902 return false;
903 }
904 if (ret & VM_FAULT_ERROR) {
905 trace_mm_collapse_huge_page_swapin(mm, swapped_in, 0);
906 return false;
907 }
908 /* pte is unmapped now, we need to map it */
909 fe.pte = pte_offset_map(pmd, fe.address);
910 }
911 fe.pte--;
912 pte_unmap(fe.pte);
913 trace_mm_collapse_huge_page_swapin(mm, swapped_in, 1);
914 return true;
915}
916
917static void collapse_huge_page(struct mm_struct *mm,
918 unsigned long address,
919 struct page **hpage,
920 struct vm_area_struct *vma,
921 int node)
922{
923 pmd_t *pmd, _pmd;
924 pte_t *pte;
925 pgtable_t pgtable;
926 struct page *new_page;
927 spinlock_t *pmd_ptl, *pte_ptl;
928 int isolated = 0, result = 0;
929 struct mem_cgroup *memcg;
930 unsigned long mmun_start; /* For mmu_notifiers */
931 unsigned long mmun_end; /* For mmu_notifiers */
932 gfp_t gfp;
933
934 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
935
936 /* Only allocate from the target node */
937 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
938
939 /* release the mmap_sem read lock. */
940 new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
941 if (!new_page) {
942 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
943 goto out_nolock;
944 }
945
946 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
947 result = SCAN_CGROUP_CHARGE_FAIL;
948 goto out_nolock;
949 }
950
951 down_read(&mm->mmap_sem);
952 result = hugepage_vma_revalidate(mm, address);
953 if (result) {
954 mem_cgroup_cancel_charge(new_page, memcg, true);
955 up_read(&mm->mmap_sem);
956 goto out_nolock;
957 }
958
959 pmd = mm_find_pmd(mm, address);
960 if (!pmd) {
961 result = SCAN_PMD_NULL;
962 mem_cgroup_cancel_charge(new_page, memcg, true);
963 up_read(&mm->mmap_sem);
964 goto out_nolock;
965 }
966
967 /*
968 * __collapse_huge_page_swapin always returns with mmap_sem locked.
969 * If it fails, release mmap_sem and jump directly out.
970 * Continuing to collapse causes inconsistency.
971 */
972 if (!__collapse_huge_page_swapin(mm, vma, address, pmd)) {
973 mem_cgroup_cancel_charge(new_page, memcg, true);
974 up_read(&mm->mmap_sem);
975 goto out_nolock;
976 }
977
978 up_read(&mm->mmap_sem);
979 /*
980 * Prevent all access to pagetables with the exception of
981 * gup_fast later handled by the ptep_clear_flush and the VM
982 * handled by the anon_vma lock + PG_lock.
983 */
984 down_write(&mm->mmap_sem);
985 result = hugepage_vma_revalidate(mm, address);
986 if (result)
987 goto out;
988 /* check if the pmd is still valid */
989 if (mm_find_pmd(mm, address) != pmd)
990 goto out;
991
992 anon_vma_lock_write(vma->anon_vma);
993
994 pte = pte_offset_map(pmd, address);
995 pte_ptl = pte_lockptr(mm, pmd);
996
997 mmun_start = address;
998 mmun_end = address + HPAGE_PMD_SIZE;
999 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1000 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1001 /*
1002 * After this gup_fast can't run anymore. This also removes
1003 * any huge TLB entry from the CPU so we won't allow
1004 * huge and small TLB entries for the same virtual address
1005 * to avoid the risk of CPU bugs in that area.
1006 */
1007 _pmd = pmdp_collapse_flush(vma, address, pmd);
1008 spin_unlock(pmd_ptl);
1009 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1010
1011 spin_lock(pte_ptl);
1012 isolated = __collapse_huge_page_isolate(vma, address, pte);
1013 spin_unlock(pte_ptl);
1014
1015 if (unlikely(!isolated)) {
1016 pte_unmap(pte);
1017 spin_lock(pmd_ptl);
1018 BUG_ON(!pmd_none(*pmd));
1019 /*
1020 * We can only use set_pmd_at when establishing
1021 * hugepmds and never for establishing regular pmds that
1022 * points to regular pagetables. Use pmd_populate for that
1023 */
1024 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1025 spin_unlock(pmd_ptl);
1026 anon_vma_unlock_write(vma->anon_vma);
1027 result = SCAN_FAIL;
1028 goto out;
1029 }
1030
1031 /*
1032 * All pages are isolated and locked so anon_vma rmap
1033 * can't run anymore.
1034 */
1035 anon_vma_unlock_write(vma->anon_vma);
1036
1037 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1038 pte_unmap(pte);
1039 __SetPageUptodate(new_page);
1040 pgtable = pmd_pgtable(_pmd);
1041
1042 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1043 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1044
1045 /*
1046 * spin_lock() below is not the equivalent of smp_wmb(), so
1047 * this is needed to avoid the copy_huge_page writes to become
1048 * visible after the set_pmd_at() write.
1049 */
1050 smp_wmb();
1051
1052 spin_lock(pmd_ptl);
1053 BUG_ON(!pmd_none(*pmd));
1054 page_add_new_anon_rmap(new_page, vma, address, true);
1055 mem_cgroup_commit_charge(new_page, memcg, false, true);
1056 lru_cache_add_active_or_unevictable(new_page, vma);
1057 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1058 set_pmd_at(mm, address, pmd, _pmd);
1059 update_mmu_cache_pmd(vma, address, pmd);
1060 spin_unlock(pmd_ptl);
1061
1062 *hpage = NULL;
1063
1064 khugepaged_pages_collapsed++;
1065 result = SCAN_SUCCEED;
1066out_up_write:
1067 up_write(&mm->mmap_sem);
1068out_nolock:
1069 trace_mm_collapse_huge_page(mm, isolated, result);
1070 return;
1071out:
1072 mem_cgroup_cancel_charge(new_page, memcg, true);
1073 goto out_up_write;
1074}
1075
1076static int khugepaged_scan_pmd(struct mm_struct *mm,
1077 struct vm_area_struct *vma,
1078 unsigned long address,
1079 struct page **hpage)
1080{
1081 pmd_t *pmd;
1082 pte_t *pte, *_pte;
1083 int ret = 0, none_or_zero = 0, result = 0;
1084 struct page *page = NULL;
1085 unsigned long _address;
1086 spinlock_t *ptl;
1087 int node = NUMA_NO_NODE, unmapped = 0;
1088 bool writable = false, referenced = false;
1089
1090 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1091
1092 pmd = mm_find_pmd(mm, address);
1093 if (!pmd) {
1094 result = SCAN_PMD_NULL;
1095 goto out;
1096 }
1097
1098 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1099 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1100 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1101 _pte++, _address += PAGE_SIZE) {
1102 pte_t pteval = *_pte;
1103 if (is_swap_pte(pteval)) {
1104 if (++unmapped <= khugepaged_max_ptes_swap) {
1105 continue;
1106 } else {
1107 result = SCAN_EXCEED_SWAP_PTE;
1108 goto out_unmap;
1109 }
1110 }
1111 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1112 if (!userfaultfd_armed(vma) &&
1113 ++none_or_zero <= khugepaged_max_ptes_none) {
1114 continue;
1115 } else {
1116 result = SCAN_EXCEED_NONE_PTE;
1117 goto out_unmap;
1118 }
1119 }
1120 if (!pte_present(pteval)) {
1121 result = SCAN_PTE_NON_PRESENT;
1122 goto out_unmap;
1123 }
1124 if (pte_write(pteval))
1125 writable = true;
1126
1127 page = vm_normal_page(vma, _address, pteval);
1128 if (unlikely(!page)) {
1129 result = SCAN_PAGE_NULL;
1130 goto out_unmap;
1131 }
1132
1133 /* TODO: teach khugepaged to collapse THP mapped with pte */
1134 if (PageCompound(page)) {
1135 result = SCAN_PAGE_COMPOUND;
1136 goto out_unmap;
1137 }
1138
1139 /*
1140 * Record which node the original page is from and save this
1141 * information to khugepaged_node_load[].
1142 * Khupaged will allocate hugepage from the node has the max
1143 * hit record.
1144 */
1145 node = page_to_nid(page);
1146 if (khugepaged_scan_abort(node)) {
1147 result = SCAN_SCAN_ABORT;
1148 goto out_unmap;
1149 }
1150 khugepaged_node_load[node]++;
1151 if (!PageLRU(page)) {
1152 result = SCAN_PAGE_LRU;
1153 goto out_unmap;
1154 }
1155 if (PageLocked(page)) {
1156 result = SCAN_PAGE_LOCK;
1157 goto out_unmap;
1158 }
1159 if (!PageAnon(page)) {
1160 result = SCAN_PAGE_ANON;
1161 goto out_unmap;
1162 }
1163
1164 /*
1165 * cannot use mapcount: can't collapse if there's a gup pin.
1166 * The page must only be referenced by the scanned process
1167 * and page swap cache.
1168 */
1169 if (page_count(page) != 1 + !!PageSwapCache(page)) {
1170 result = SCAN_PAGE_COUNT;
1171 goto out_unmap;
1172 }
1173 if (pte_young(pteval) ||
1174 page_is_young(page) || PageReferenced(page) ||
1175 mmu_notifier_test_young(vma->vm_mm, address))
1176 referenced = true;
1177 }
1178 if (writable) {
1179 if (referenced) {
1180 result = SCAN_SUCCEED;
1181 ret = 1;
1182 } else {
1183 result = SCAN_NO_REFERENCED_PAGE;
1184 }
1185 } else {
1186 result = SCAN_PAGE_RO;
1187 }
1188out_unmap:
1189 pte_unmap_unlock(pte, ptl);
1190 if (ret) {
1191 node = khugepaged_find_target_node();
1192 /* collapse_huge_page will return with the mmap_sem released */
1193 collapse_huge_page(mm, address, hpage, vma, node);
1194 }
1195out:
1196 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1197 none_or_zero, result, unmapped);
1198 return ret;
1199}
1200
1201static void collect_mm_slot(struct mm_slot *mm_slot)
1202{
1203 struct mm_struct *mm = mm_slot->mm;
1204
1205 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1206
1207 if (khugepaged_test_exit(mm)) {
1208 /* free mm_slot */
1209 hash_del(&mm_slot->hash);
1210 list_del(&mm_slot->mm_node);
1211
1212 /*
1213 * Not strictly needed because the mm exited already.
1214 *
1215 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1216 */
1217
1218 /* khugepaged_mm_lock actually not necessary for the below */
1219 free_mm_slot(mm_slot);
1220 mmdrop(mm);
1221 }
1222}
1223
1224static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1225 struct page **hpage)
1226 __releases(&khugepaged_mm_lock)
1227 __acquires(&khugepaged_mm_lock)
1228{
1229 struct mm_slot *mm_slot;
1230 struct mm_struct *mm;
1231 struct vm_area_struct *vma;
1232 int progress = 0;
1233
1234 VM_BUG_ON(!pages);
1235 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1236
1237 if (khugepaged_scan.mm_slot)
1238 mm_slot = khugepaged_scan.mm_slot;
1239 else {
1240 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1241 struct mm_slot, mm_node);
1242 khugepaged_scan.address = 0;
1243 khugepaged_scan.mm_slot = mm_slot;
1244 }
1245 spin_unlock(&khugepaged_mm_lock);
1246
1247 mm = mm_slot->mm;
1248 down_read(&mm->mmap_sem);
1249 if (unlikely(khugepaged_test_exit(mm)))
1250 vma = NULL;
1251 else
1252 vma = find_vma(mm, khugepaged_scan.address);
1253
1254 progress++;
1255 for (; vma; vma = vma->vm_next) {
1256 unsigned long hstart, hend;
1257
1258 cond_resched();
1259 if (unlikely(khugepaged_test_exit(mm))) {
1260 progress++;
1261 break;
1262 }
1263 if (!hugepage_vma_check(vma)) {
1264skip:
1265 progress++;
1266 continue;
1267 }
1268 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1269 hend = vma->vm_end & HPAGE_PMD_MASK;
1270 if (hstart >= hend)
1271 goto skip;
1272 if (khugepaged_scan.address > hend)
1273 goto skip;
1274 if (khugepaged_scan.address < hstart)
1275 khugepaged_scan.address = hstart;
1276 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1277
1278 while (khugepaged_scan.address < hend) {
1279 int ret;
1280 cond_resched();
1281 if (unlikely(khugepaged_test_exit(mm)))
1282 goto breakouterloop;
1283
1284 VM_BUG_ON(khugepaged_scan.address < hstart ||
1285 khugepaged_scan.address + HPAGE_PMD_SIZE >
1286 hend);
1287 ret = khugepaged_scan_pmd(mm, vma,
1288 khugepaged_scan.address,
1289 hpage);
1290 /* move to next address */
1291 khugepaged_scan.address += HPAGE_PMD_SIZE;
1292 progress += HPAGE_PMD_NR;
1293 if (ret)
1294 /* we released mmap_sem so break loop */
1295 goto breakouterloop_mmap_sem;
1296 if (progress >= pages)
1297 goto breakouterloop;
1298 }
1299 }
1300breakouterloop:
1301 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1302breakouterloop_mmap_sem:
1303
1304 spin_lock(&khugepaged_mm_lock);
1305 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1306 /*
1307 * Release the current mm_slot if this mm is about to die, or
1308 * if we scanned all vmas of this mm.
1309 */
1310 if (khugepaged_test_exit(mm) || !vma) {
1311 /*
1312 * Make sure that if mm_users is reaching zero while
1313 * khugepaged runs here, khugepaged_exit will find
1314 * mm_slot not pointing to the exiting mm.
1315 */
1316 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1317 khugepaged_scan.mm_slot = list_entry(
1318 mm_slot->mm_node.next,
1319 struct mm_slot, mm_node);
1320 khugepaged_scan.address = 0;
1321 } else {
1322 khugepaged_scan.mm_slot = NULL;
1323 khugepaged_full_scans++;
1324 }
1325
1326 collect_mm_slot(mm_slot);
1327 }
1328
1329 return progress;
1330}
1331
1332static int khugepaged_has_work(void)
1333{
1334 return !list_empty(&khugepaged_scan.mm_head) &&
1335 khugepaged_enabled();
1336}
1337
1338static int khugepaged_wait_event(void)
1339{
1340 return !list_empty(&khugepaged_scan.mm_head) ||
1341 kthread_should_stop();
1342}
1343
1344static void khugepaged_do_scan(void)
1345{
1346 struct page *hpage = NULL;
1347 unsigned int progress = 0, pass_through_head = 0;
1348 unsigned int pages = khugepaged_pages_to_scan;
1349 bool wait = true;
1350
1351 barrier(); /* write khugepaged_pages_to_scan to local stack */
1352
1353 while (progress < pages) {
1354 if (!khugepaged_prealloc_page(&hpage, &wait))
1355 break;
1356
1357 cond_resched();
1358
1359 if (unlikely(kthread_should_stop() || try_to_freeze()))
1360 break;
1361
1362 spin_lock(&khugepaged_mm_lock);
1363 if (!khugepaged_scan.mm_slot)
1364 pass_through_head++;
1365 if (khugepaged_has_work() &&
1366 pass_through_head < 2)
1367 progress += khugepaged_scan_mm_slot(pages - progress,
1368 &hpage);
1369 else
1370 progress = pages;
1371 spin_unlock(&khugepaged_mm_lock);
1372 }
1373
1374 if (!IS_ERR_OR_NULL(hpage))
1375 put_page(hpage);
1376}
1377
1378static bool khugepaged_should_wakeup(void)
1379{
1380 return kthread_should_stop() ||
1381 time_after_eq(jiffies, khugepaged_sleep_expire);
1382}
1383
1384static void khugepaged_wait_work(void)
1385{
1386 if (khugepaged_has_work()) {
1387 const unsigned long scan_sleep_jiffies =
1388 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1389
1390 if (!scan_sleep_jiffies)
1391 return;
1392
1393 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1394 wait_event_freezable_timeout(khugepaged_wait,
1395 khugepaged_should_wakeup(),
1396 scan_sleep_jiffies);
1397 return;
1398 }
1399
1400 if (khugepaged_enabled())
1401 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1402}
1403
1404static int khugepaged(void *none)
1405{
1406 struct mm_slot *mm_slot;
1407
1408 set_freezable();
1409 set_user_nice(current, MAX_NICE);
1410
1411 while (!kthread_should_stop()) {
1412 khugepaged_do_scan();
1413 khugepaged_wait_work();
1414 }
1415
1416 spin_lock(&khugepaged_mm_lock);
1417 mm_slot = khugepaged_scan.mm_slot;
1418 khugepaged_scan.mm_slot = NULL;
1419 if (mm_slot)
1420 collect_mm_slot(mm_slot);
1421 spin_unlock(&khugepaged_mm_lock);
1422 return 0;
1423}
1424
1425static void set_recommended_min_free_kbytes(void)
1426{
1427 struct zone *zone;
1428 int nr_zones = 0;
1429 unsigned long recommended_min;
1430
1431 for_each_populated_zone(zone)
1432 nr_zones++;
1433
1434 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1435 recommended_min = pageblock_nr_pages * nr_zones * 2;
1436
1437 /*
1438 * Make sure that on average at least two pageblocks are almost free
1439 * of another type, one for a migratetype to fall back to and a
1440 * second to avoid subsequent fallbacks of other types There are 3
1441 * MIGRATE_TYPES we care about.
1442 */
1443 recommended_min += pageblock_nr_pages * nr_zones *
1444 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1445
1446 /* don't ever allow to reserve more than 5% of the lowmem */
1447 recommended_min = min(recommended_min,
1448 (unsigned long) nr_free_buffer_pages() / 20);
1449 recommended_min <<= (PAGE_SHIFT-10);
1450
1451 if (recommended_min > min_free_kbytes) {
1452 if (user_min_free_kbytes >= 0)
1453 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1454 min_free_kbytes, recommended_min);
1455
1456 min_free_kbytes = recommended_min;
1457 }
1458 setup_per_zone_wmarks();
1459}
1460
1461int start_stop_khugepaged(void)
1462{
1463 static struct task_struct *khugepaged_thread __read_mostly;
1464 static DEFINE_MUTEX(khugepaged_mutex);
1465 int err = 0;
1466
1467 mutex_lock(&khugepaged_mutex);
1468 if (khugepaged_enabled()) {
1469 if (!khugepaged_thread)
1470 khugepaged_thread = kthread_run(khugepaged, NULL,
1471 "khugepaged");
1472 if (IS_ERR(khugepaged_thread)) {
1473 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1474 err = PTR_ERR(khugepaged_thread);
1475 khugepaged_thread = NULL;
1476 goto fail;
1477 }
1478
1479 if (!list_empty(&khugepaged_scan.mm_head))
1480 wake_up_interruptible(&khugepaged_wait);
1481
1482 set_recommended_min_free_kbytes();
1483 } else if (khugepaged_thread) {
1484 kthread_stop(khugepaged_thread);
1485 khugepaged_thread = NULL;
1486 }
1487fail:
1488 mutex_unlock(&khugepaged_mutex);
1489 return err;
1490}
generated by cgit v1.2.2 (git 2.25.0) at 2026-04-04 17:21:57 -0400