]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
b46e756f KS |
2 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
3 | ||
4 | #include <linux/mm.h> | |
5 | #include <linux/sched.h> | |
6e84f315 | 6 | #include <linux/sched/mm.h> |
f7ccbae4 | 7 | #include <linux/sched/coredump.h> |
b46e756f KS |
8 | #include <linux/mmu_notifier.h> |
9 | #include <linux/rmap.h> | |
10 | #include <linux/swap.h> | |
11 | #include <linux/mm_inline.h> | |
12 | #include <linux/kthread.h> | |
13 | #include <linux/khugepaged.h> | |
14 | #include <linux/freezer.h> | |
15 | #include <linux/mman.h> | |
16 | #include <linux/hashtable.h> | |
17 | #include <linux/userfaultfd_k.h> | |
18 | #include <linux/page_idle.h> | |
19 | #include <linux/swapops.h> | |
f3f0e1d2 | 20 | #include <linux/shmem_fs.h> |
b46e756f KS |
21 | |
22 | #include <asm/tlb.h> | |
23 | #include <asm/pgalloc.h> | |
24 | #include "internal.h" | |
25 | ||
26 | enum scan_result { | |
27 | SCAN_FAIL, | |
28 | SCAN_SUCCEED, | |
29 | SCAN_PMD_NULL, | |
30 | SCAN_EXCEED_NONE_PTE, | |
31 | SCAN_PTE_NON_PRESENT, | |
32 | SCAN_PAGE_RO, | |
0db501f7 | 33 | SCAN_LACK_REFERENCED_PAGE, |
b46e756f KS |
34 | SCAN_PAGE_NULL, |
35 | SCAN_SCAN_ABORT, | |
36 | SCAN_PAGE_COUNT, | |
37 | SCAN_PAGE_LRU, | |
38 | SCAN_PAGE_LOCK, | |
39 | SCAN_PAGE_ANON, | |
40 | SCAN_PAGE_COMPOUND, | |
41 | SCAN_ANY_PROCESS, | |
42 | SCAN_VMA_NULL, | |
43 | SCAN_VMA_CHECK, | |
44 | SCAN_ADDRESS_RANGE, | |
45 | SCAN_SWAP_CACHE_PAGE, | |
46 | SCAN_DEL_PAGE_LRU, | |
47 | SCAN_ALLOC_HUGE_PAGE_FAIL, | |
48 | SCAN_CGROUP_CHARGE_FAIL, | |
f3f0e1d2 KS |
49 | SCAN_EXCEED_SWAP_PTE, |
50 | SCAN_TRUNCATED, | |
b46e756f KS |
51 | }; |
52 | ||
53 | #define CREATE_TRACE_POINTS | |
54 | #include <trace/events/huge_memory.h> | |
55 | ||
56 | /* default scan 8*512 pte (or vmas) every 30 second */ | |
57 | static unsigned int khugepaged_pages_to_scan __read_mostly; | |
58 | static unsigned int khugepaged_pages_collapsed; | |
59 | static unsigned int khugepaged_full_scans; | |
60 | static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; | |
61 | /* during fragmentation poll the hugepage allocator once every minute */ | |
62 | static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; | |
63 | static unsigned long khugepaged_sleep_expire; | |
64 | static DEFINE_SPINLOCK(khugepaged_mm_lock); | |
65 | static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); | |
66 | /* | |
67 | * default collapse hugepages if there is at least one pte mapped like | |
68 | * it would have happened if the vma was large enough during page | |
69 | * fault. | |
70 | */ | |
71 | static unsigned int khugepaged_max_ptes_none __read_mostly; | |
72 | static unsigned int khugepaged_max_ptes_swap __read_mostly; | |
73 | ||
74 | #define MM_SLOTS_HASH_BITS 10 | |
75 | static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
76 | ||
77 | static struct kmem_cache *mm_slot_cache __read_mostly; | |
78 | ||
79 | /** | |
80 | * struct mm_slot - hash lookup from mm to mm_slot | |
81 | * @hash: hash collision list | |
82 | * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head | |
83 | * @mm: the mm that this information is valid for | |
84 | */ | |
85 | struct mm_slot { | |
86 | struct hlist_node hash; | |
87 | struct list_head mm_node; | |
88 | struct mm_struct *mm; | |
89 | }; | |
90 | ||
91 | /** | |
92 | * struct khugepaged_scan - cursor for scanning | |
93 | * @mm_head: the head of the mm list to scan | |
94 | * @mm_slot: the current mm_slot we are scanning | |
95 | * @address: the next address inside that to be scanned | |
96 | * | |
97 | * There is only the one khugepaged_scan instance of this cursor structure. | |
98 | */ | |
99 | struct khugepaged_scan { | |
100 | struct list_head mm_head; | |
101 | struct mm_slot *mm_slot; | |
102 | unsigned long address; | |
103 | }; | |
104 | ||
105 | static struct khugepaged_scan khugepaged_scan = { | |
106 | .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), | |
107 | }; | |
108 | ||
e1465d12 | 109 | #ifdef CONFIG_SYSFS |
b46e756f KS |
110 | static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, |
111 | struct kobj_attribute *attr, | |
112 | char *buf) | |
113 | { | |
114 | return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); | |
115 | } | |
116 | ||
117 | static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, | |
118 | struct kobj_attribute *attr, | |
119 | const char *buf, size_t count) | |
120 | { | |
121 | unsigned long msecs; | |
122 | int err; | |
123 | ||
124 | err = kstrtoul(buf, 10, &msecs); | |
125 | if (err || msecs > UINT_MAX) | |
126 | return -EINVAL; | |
127 | ||
128 | khugepaged_scan_sleep_millisecs = msecs; | |
129 | khugepaged_sleep_expire = 0; | |
130 | wake_up_interruptible(&khugepaged_wait); | |
131 | ||
132 | return count; | |
133 | } | |
134 | static struct kobj_attribute scan_sleep_millisecs_attr = | |
135 | __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, | |
136 | scan_sleep_millisecs_store); | |
137 | ||
138 | static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, | |
139 | struct kobj_attribute *attr, | |
140 | char *buf) | |
141 | { | |
142 | return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); | |
143 | } | |
144 | ||
145 | static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, | |
146 | struct kobj_attribute *attr, | |
147 | const char *buf, size_t count) | |
148 | { | |
149 | unsigned long msecs; | |
150 | int err; | |
151 | ||
152 | err = kstrtoul(buf, 10, &msecs); | |
153 | if (err || msecs > UINT_MAX) | |
154 | return -EINVAL; | |
155 | ||
156 | khugepaged_alloc_sleep_millisecs = msecs; | |
157 | khugepaged_sleep_expire = 0; | |
158 | wake_up_interruptible(&khugepaged_wait); | |
159 | ||
160 | return count; | |
161 | } | |
162 | static struct kobj_attribute alloc_sleep_millisecs_attr = | |
163 | __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, | |
164 | alloc_sleep_millisecs_store); | |
165 | ||
166 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
167 | struct kobj_attribute *attr, | |
168 | char *buf) | |
169 | { | |
170 | return sprintf(buf, "%u\n", khugepaged_pages_to_scan); | |
171 | } | |
172 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
173 | struct kobj_attribute *attr, | |
174 | const char *buf, size_t count) | |
175 | { | |
176 | int err; | |
177 | unsigned long pages; | |
178 | ||
179 | err = kstrtoul(buf, 10, &pages); | |
180 | if (err || !pages || pages > UINT_MAX) | |
181 | return -EINVAL; | |
182 | ||
183 | khugepaged_pages_to_scan = pages; | |
184 | ||
185 | return count; | |
186 | } | |
187 | static struct kobj_attribute pages_to_scan_attr = | |
188 | __ATTR(pages_to_scan, 0644, pages_to_scan_show, | |
189 | pages_to_scan_store); | |
190 | ||
191 | static ssize_t pages_collapsed_show(struct kobject *kobj, | |
192 | struct kobj_attribute *attr, | |
193 | char *buf) | |
194 | { | |
195 | return sprintf(buf, "%u\n", khugepaged_pages_collapsed); | |
196 | } | |
197 | static struct kobj_attribute pages_collapsed_attr = | |
198 | __ATTR_RO(pages_collapsed); | |
199 | ||
200 | static ssize_t full_scans_show(struct kobject *kobj, | |
201 | struct kobj_attribute *attr, | |
202 | char *buf) | |
203 | { | |
204 | return sprintf(buf, "%u\n", khugepaged_full_scans); | |
205 | } | |
206 | static struct kobj_attribute full_scans_attr = | |
207 | __ATTR_RO(full_scans); | |
208 | ||
209 | static ssize_t khugepaged_defrag_show(struct kobject *kobj, | |
210 | struct kobj_attribute *attr, char *buf) | |
211 | { | |
212 | return single_hugepage_flag_show(kobj, attr, buf, | |
213 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
214 | } | |
215 | static ssize_t khugepaged_defrag_store(struct kobject *kobj, | |
216 | struct kobj_attribute *attr, | |
217 | const char *buf, size_t count) | |
218 | { | |
219 | return single_hugepage_flag_store(kobj, attr, buf, count, | |
220 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
221 | } | |
222 | static struct kobj_attribute khugepaged_defrag_attr = | |
223 | __ATTR(defrag, 0644, khugepaged_defrag_show, | |
224 | khugepaged_defrag_store); | |
225 | ||
226 | /* | |
227 | * max_ptes_none controls if khugepaged should collapse hugepages over | |
228 | * any unmapped ptes in turn potentially increasing the memory | |
229 | * footprint of the vmas. When max_ptes_none is 0 khugepaged will not | |
230 | * reduce the available free memory in the system as it | |
231 | * runs. Increasing max_ptes_none will instead potentially reduce the | |
232 | * free memory in the system during the khugepaged scan. | |
233 | */ | |
234 | static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, | |
235 | struct kobj_attribute *attr, | |
236 | char *buf) | |
237 | { | |
238 | return sprintf(buf, "%u\n", khugepaged_max_ptes_none); | |
239 | } | |
240 | static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, | |
241 | struct kobj_attribute *attr, | |
242 | const char *buf, size_t count) | |
243 | { | |
244 | int err; | |
245 | unsigned long max_ptes_none; | |
246 | ||
247 | err = kstrtoul(buf, 10, &max_ptes_none); | |
248 | if (err || max_ptes_none > HPAGE_PMD_NR-1) | |
249 | return -EINVAL; | |
250 | ||
251 | khugepaged_max_ptes_none = max_ptes_none; | |
252 | ||
253 | return count; | |
254 | } | |
255 | static struct kobj_attribute khugepaged_max_ptes_none_attr = | |
256 | __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, | |
257 | khugepaged_max_ptes_none_store); | |
258 | ||
259 | static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj, | |
260 | struct kobj_attribute *attr, | |
261 | char *buf) | |
262 | { | |
263 | return sprintf(buf, "%u\n", khugepaged_max_ptes_swap); | |
264 | } | |
265 | ||
266 | static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj, | |
267 | struct kobj_attribute *attr, | |
268 | const char *buf, size_t count) | |
269 | { | |
270 | int err; | |
271 | unsigned long max_ptes_swap; | |
272 | ||
273 | err = kstrtoul(buf, 10, &max_ptes_swap); | |
274 | if (err || max_ptes_swap > HPAGE_PMD_NR-1) | |
275 | return -EINVAL; | |
276 | ||
277 | khugepaged_max_ptes_swap = max_ptes_swap; | |
278 | ||
279 | return count; | |
280 | } | |
281 | ||
282 | static struct kobj_attribute khugepaged_max_ptes_swap_attr = | |
283 | __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show, | |
284 | khugepaged_max_ptes_swap_store); | |
285 | ||
286 | static struct attribute *khugepaged_attr[] = { | |
287 | &khugepaged_defrag_attr.attr, | |
288 | &khugepaged_max_ptes_none_attr.attr, | |
289 | &pages_to_scan_attr.attr, | |
290 | &pages_collapsed_attr.attr, | |
291 | &full_scans_attr.attr, | |
292 | &scan_sleep_millisecs_attr.attr, | |
293 | &alloc_sleep_millisecs_attr.attr, | |
294 | &khugepaged_max_ptes_swap_attr.attr, | |
295 | NULL, | |
296 | }; | |
297 | ||
298 | struct attribute_group khugepaged_attr_group = { | |
299 | .attrs = khugepaged_attr, | |
300 | .name = "khugepaged", | |
301 | }; | |
e1465d12 | 302 | #endif /* CONFIG_SYSFS */ |
b46e756f | 303 | |
f3f0e1d2 | 304 | #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB) |
b46e756f KS |
305 | |
306 | int hugepage_madvise(struct vm_area_struct *vma, | |
307 | unsigned long *vm_flags, int advice) | |
308 | { | |
309 | switch (advice) { | |
310 | case MADV_HUGEPAGE: | |
311 | #ifdef CONFIG_S390 | |
312 | /* | |
313 | * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 | |
314 | * can't handle this properly after s390_enable_sie, so we simply | |
315 | * ignore the madvise to prevent qemu from causing a SIGSEGV. | |
316 | */ | |
317 | if (mm_has_pgste(vma->vm_mm)) | |
318 | return 0; | |
319 | #endif | |
320 | *vm_flags &= ~VM_NOHUGEPAGE; | |
321 | *vm_flags |= VM_HUGEPAGE; | |
322 | /* | |
323 | * If the vma become good for khugepaged to scan, | |
324 | * register it here without waiting a page fault that | |
325 | * may not happen any time soon. | |
326 | */ | |
327 | if (!(*vm_flags & VM_NO_KHUGEPAGED) && | |
328 | khugepaged_enter_vma_merge(vma, *vm_flags)) | |
329 | return -ENOMEM; | |
330 | break; | |
331 | case MADV_NOHUGEPAGE: | |
332 | *vm_flags &= ~VM_HUGEPAGE; | |
333 | *vm_flags |= VM_NOHUGEPAGE; | |
334 | /* | |
335 | * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning | |
336 | * this vma even if we leave the mm registered in khugepaged if | |
337 | * it got registered before VM_NOHUGEPAGE was set. | |
338 | */ | |
339 | break; | |
340 | } | |
341 | ||
342 | return 0; | |
343 | } | |
344 | ||
345 | int __init khugepaged_init(void) | |
346 | { | |
347 | mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", | |
348 | sizeof(struct mm_slot), | |
349 | __alignof__(struct mm_slot), 0, NULL); | |
350 | if (!mm_slot_cache) | |
351 | return -ENOMEM; | |
352 | ||
353 | khugepaged_pages_to_scan = HPAGE_PMD_NR * 8; | |
354 | khugepaged_max_ptes_none = HPAGE_PMD_NR - 1; | |
355 | khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8; | |
356 | ||
357 | return 0; | |
358 | } | |
359 | ||
360 | void __init khugepaged_destroy(void) | |
361 | { | |
362 | kmem_cache_destroy(mm_slot_cache); | |
363 | } | |
364 | ||
365 | static inline struct mm_slot *alloc_mm_slot(void) | |
366 | { | |
367 | if (!mm_slot_cache) /* initialization failed */ | |
368 | return NULL; | |
369 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
370 | } | |
371 | ||
372 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
373 | { | |
374 | kmem_cache_free(mm_slot_cache, mm_slot); | |
375 | } | |
376 | ||
377 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) | |
378 | { | |
379 | struct mm_slot *mm_slot; | |
380 | ||
381 | hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) | |
382 | if (mm == mm_slot->mm) | |
383 | return mm_slot; | |
384 | ||
385 | return NULL; | |
386 | } | |
387 | ||
388 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
389 | struct mm_slot *mm_slot) | |
390 | { | |
391 | mm_slot->mm = mm; | |
392 | hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); | |
393 | } | |
394 | ||
395 | static inline int khugepaged_test_exit(struct mm_struct *mm) | |
396 | { | |
397 | return atomic_read(&mm->mm_users) == 0; | |
398 | } | |
399 | ||
50f8b92f SL |
400 | static bool hugepage_vma_check(struct vm_area_struct *vma, |
401 | unsigned long vm_flags) | |
c2231020 | 402 | { |
50f8b92f SL |
403 | if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || |
404 | (vm_flags & VM_NOHUGEPAGE) || | |
c2231020 YS |
405 | test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags)) |
406 | return false; | |
407 | if (shmem_file(vma->vm_file)) { | |
408 | if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) | |
409 | return false; | |
410 | return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff, | |
411 | HPAGE_PMD_NR); | |
412 | } | |
413 | if (!vma->anon_vma || vma->vm_ops) | |
414 | return false; | |
415 | if (is_vma_temporary_stack(vma)) | |
416 | return false; | |
50f8b92f | 417 | return !(vm_flags & VM_NO_KHUGEPAGED); |
c2231020 YS |
418 | } |
419 | ||
b46e756f KS |
420 | int __khugepaged_enter(struct mm_struct *mm) |
421 | { | |
422 | struct mm_slot *mm_slot; | |
423 | int wakeup; | |
424 | ||
425 | mm_slot = alloc_mm_slot(); | |
426 | if (!mm_slot) | |
427 | return -ENOMEM; | |
428 | ||
429 | /* __khugepaged_exit() must not run from under us */ | |
430 | VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); | |
431 | if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { | |
432 | free_mm_slot(mm_slot); | |
433 | return 0; | |
434 | } | |
435 | ||
436 | spin_lock(&khugepaged_mm_lock); | |
437 | insert_to_mm_slots_hash(mm, mm_slot); | |
438 | /* | |
439 | * Insert just behind the scanning cursor, to let the area settle | |
440 | * down a little. | |
441 | */ | |
442 | wakeup = list_empty(&khugepaged_scan.mm_head); | |
443 | list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); | |
444 | spin_unlock(&khugepaged_mm_lock); | |
445 | ||
f1f10076 | 446 | mmgrab(mm); |
b46e756f KS |
447 | if (wakeup) |
448 | wake_up_interruptible(&khugepaged_wait); | |
449 | ||
450 | return 0; | |
451 | } | |
452 | ||
453 | int khugepaged_enter_vma_merge(struct vm_area_struct *vma, | |
454 | unsigned long vm_flags) | |
455 | { | |
456 | unsigned long hstart, hend; | |
c2231020 YS |
457 | |
458 | /* | |
459 | * khugepaged does not yet work on non-shmem files or special | |
460 | * mappings. And file-private shmem THP is not supported. | |
461 | */ | |
50f8b92f | 462 | if (!hugepage_vma_check(vma, vm_flags)) |
b46e756f | 463 | return 0; |
c2231020 | 464 | |
b46e756f KS |
465 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
466 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
467 | if (hstart < hend) | |
468 | return khugepaged_enter(vma, vm_flags); | |
469 | return 0; | |
470 | } | |
471 | ||
472 | void __khugepaged_exit(struct mm_struct *mm) | |
473 | { | |
474 | struct mm_slot *mm_slot; | |
475 | int free = 0; | |
476 | ||
477 | spin_lock(&khugepaged_mm_lock); | |
478 | mm_slot = get_mm_slot(mm); | |
479 | if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { | |
480 | hash_del(&mm_slot->hash); | |
481 | list_del(&mm_slot->mm_node); | |
482 | free = 1; | |
483 | } | |
484 | spin_unlock(&khugepaged_mm_lock); | |
485 | ||
486 | if (free) { | |
487 | clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
488 | free_mm_slot(mm_slot); | |
489 | mmdrop(mm); | |
490 | } else if (mm_slot) { | |
491 | /* | |
492 | * This is required to serialize against | |
493 | * khugepaged_test_exit() (which is guaranteed to run | |
494 | * under mmap sem read mode). Stop here (after we | |
495 | * return all pagetables will be destroyed) until | |
496 | * khugepaged has finished working on the pagetables | |
497 | * under the mmap_sem. | |
498 | */ | |
499 | down_write(&mm->mmap_sem); | |
500 | up_write(&mm->mmap_sem); | |
501 | } | |
502 | } | |
503 | ||
504 | static void release_pte_page(struct page *page) | |
505 | { | |
d44d363f | 506 | dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page)); |
b46e756f KS |
507 | unlock_page(page); |
508 | putback_lru_page(page); | |
509 | } | |
510 | ||
511 | static void release_pte_pages(pte_t *pte, pte_t *_pte) | |
512 | { | |
513 | while (--_pte >= pte) { | |
514 | pte_t pteval = *_pte; | |
515 | if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval))) | |
516 | release_pte_page(pte_page(pteval)); | |
517 | } | |
518 | } | |
519 | ||
520 | static int __collapse_huge_page_isolate(struct vm_area_struct *vma, | |
521 | unsigned long address, | |
522 | pte_t *pte) | |
523 | { | |
524 | struct page *page = NULL; | |
525 | pte_t *_pte; | |
0db501f7 EA |
526 | int none_or_zero = 0, result = 0, referenced = 0; |
527 | bool writable = false; | |
b46e756f KS |
528 | |
529 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; | |
530 | _pte++, address += PAGE_SIZE) { | |
531 | pte_t pteval = *_pte; | |
532 | if (pte_none(pteval) || (pte_present(pteval) && | |
533 | is_zero_pfn(pte_pfn(pteval)))) { | |
534 | if (!userfaultfd_armed(vma) && | |
535 | ++none_or_zero <= khugepaged_max_ptes_none) { | |
536 | continue; | |
537 | } else { | |
538 | result = SCAN_EXCEED_NONE_PTE; | |
539 | goto out; | |
540 | } | |
541 | } | |
542 | if (!pte_present(pteval)) { | |
543 | result = SCAN_PTE_NON_PRESENT; | |
544 | goto out; | |
545 | } | |
546 | page = vm_normal_page(vma, address, pteval); | |
547 | if (unlikely(!page)) { | |
548 | result = SCAN_PAGE_NULL; | |
549 | goto out; | |
550 | } | |
551 | ||
fece2029 KS |
552 | /* TODO: teach khugepaged to collapse THP mapped with pte */ |
553 | if (PageCompound(page)) { | |
554 | result = SCAN_PAGE_COMPOUND; | |
555 | goto out; | |
556 | } | |
557 | ||
b46e756f | 558 | VM_BUG_ON_PAGE(!PageAnon(page), page); |
b46e756f KS |
559 | |
560 | /* | |
561 | * We can do it before isolate_lru_page because the | |
562 | * page can't be freed from under us. NOTE: PG_lock | |
563 | * is needed to serialize against split_huge_page | |
564 | * when invoked from the VM. | |
565 | */ | |
566 | if (!trylock_page(page)) { | |
567 | result = SCAN_PAGE_LOCK; | |
568 | goto out; | |
569 | } | |
570 | ||
571 | /* | |
572 | * cannot use mapcount: can't collapse if there's a gup pin. | |
573 | * The page must only be referenced by the scanned process | |
574 | * and page swap cache. | |
575 | */ | |
2948be5a | 576 | if (page_count(page) != 1 + PageSwapCache(page)) { |
b46e756f KS |
577 | unlock_page(page); |
578 | result = SCAN_PAGE_COUNT; | |
579 | goto out; | |
580 | } | |
581 | if (pte_write(pteval)) { | |
582 | writable = true; | |
583 | } else { | |
584 | if (PageSwapCache(page) && | |
585 | !reuse_swap_page(page, NULL)) { | |
586 | unlock_page(page); | |
587 | result = SCAN_SWAP_CACHE_PAGE; | |
588 | goto out; | |
589 | } | |
590 | /* | |
591 | * Page is not in the swap cache. It can be collapsed | |
592 | * into a THP. | |
593 | */ | |
594 | } | |
595 | ||
596 | /* | |
597 | * Isolate the page to avoid collapsing an hugepage | |
598 | * currently in use by the VM. | |
599 | */ | |
600 | if (isolate_lru_page(page)) { | |
601 | unlock_page(page); | |
602 | result = SCAN_DEL_PAGE_LRU; | |
603 | goto out; | |
604 | } | |
d44d363f SL |
605 | inc_node_page_state(page, |
606 | NR_ISOLATED_ANON + page_is_file_cache(page)); | |
b46e756f KS |
607 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
608 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
609 | ||
0db501f7 | 610 | /* There should be enough young pte to collapse the page */ |
b46e756f KS |
611 | if (pte_young(pteval) || |
612 | page_is_young(page) || PageReferenced(page) || | |
613 | mmu_notifier_test_young(vma->vm_mm, address)) | |
0db501f7 | 614 | referenced++; |
b46e756f KS |
615 | } |
616 | if (likely(writable)) { | |
617 | if (likely(referenced)) { | |
618 | result = SCAN_SUCCEED; | |
619 | trace_mm_collapse_huge_page_isolate(page, none_or_zero, | |
620 | referenced, writable, result); | |
621 | return 1; | |
622 | } | |
623 | } else { | |
624 | result = SCAN_PAGE_RO; | |
625 | } | |
626 | ||
627 | out: | |
628 | release_pte_pages(pte, _pte); | |
629 | trace_mm_collapse_huge_page_isolate(page, none_or_zero, | |
630 | referenced, writable, result); | |
631 | return 0; | |
632 | } | |
633 | ||
634 | static void __collapse_huge_page_copy(pte_t *pte, struct page *page, | |
635 | struct vm_area_struct *vma, | |
636 | unsigned long address, | |
637 | spinlock_t *ptl) | |
638 | { | |
639 | pte_t *_pte; | |
338a16ba DR |
640 | for (_pte = pte; _pte < pte + HPAGE_PMD_NR; |
641 | _pte++, page++, address += PAGE_SIZE) { | |
b46e756f KS |
642 | pte_t pteval = *_pte; |
643 | struct page *src_page; | |
644 | ||
645 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { | |
646 | clear_user_highpage(page, address); | |
647 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); | |
648 | if (is_zero_pfn(pte_pfn(pteval))) { | |
649 | /* | |
650 | * ptl mostly unnecessary. | |
651 | */ | |
652 | spin_lock(ptl); | |
653 | /* | |
654 | * paravirt calls inside pte_clear here are | |
655 | * superfluous. | |
656 | */ | |
657 | pte_clear(vma->vm_mm, address, _pte); | |
658 | spin_unlock(ptl); | |
659 | } | |
660 | } else { | |
661 | src_page = pte_page(pteval); | |
662 | copy_user_highpage(page, src_page, address, vma); | |
663 | VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page); | |
664 | release_pte_page(src_page); | |
665 | /* | |
666 | * ptl mostly unnecessary, but preempt has to | |
667 | * be disabled to update the per-cpu stats | |
668 | * inside page_remove_rmap(). | |
669 | */ | |
670 | spin_lock(ptl); | |
671 | /* | |
672 | * paravirt calls inside pte_clear here are | |
673 | * superfluous. | |
674 | */ | |
675 | pte_clear(vma->vm_mm, address, _pte); | |
676 | page_remove_rmap(src_page, false); | |
677 | spin_unlock(ptl); | |
678 | free_page_and_swap_cache(src_page); | |
679 | } | |
b46e756f KS |
680 | } |
681 | } | |
682 | ||
683 | static void khugepaged_alloc_sleep(void) | |
684 | { | |
685 | DEFINE_WAIT(wait); | |
686 | ||
687 | add_wait_queue(&khugepaged_wait, &wait); | |
688 | freezable_schedule_timeout_interruptible( | |
689 | msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); | |
690 | remove_wait_queue(&khugepaged_wait, &wait); | |
691 | } | |
692 | ||
693 | static int khugepaged_node_load[MAX_NUMNODES]; | |
694 | ||
695 | static bool khugepaged_scan_abort(int nid) | |
696 | { | |
697 | int i; | |
698 | ||
699 | /* | |
a5f5f91d | 700 | * If node_reclaim_mode is disabled, then no extra effort is made to |
b46e756f KS |
701 | * allocate memory locally. |
702 | */ | |
a5f5f91d | 703 | if (!node_reclaim_mode) |
b46e756f KS |
704 | return false; |
705 | ||
706 | /* If there is a count for this node already, it must be acceptable */ | |
707 | if (khugepaged_node_load[nid]) | |
708 | return false; | |
709 | ||
710 | for (i = 0; i < MAX_NUMNODES; i++) { | |
711 | if (!khugepaged_node_load[i]) | |
712 | continue; | |
713 | if (node_distance(nid, i) > RECLAIM_DISTANCE) | |
714 | return true; | |
715 | } | |
716 | return false; | |
717 | } | |
718 | ||
719 | /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */ | |
720 | static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void) | |
721 | { | |
25160354 | 722 | return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT; |
b46e756f KS |
723 | } |
724 | ||
725 | #ifdef CONFIG_NUMA | |
726 | static int khugepaged_find_target_node(void) | |
727 | { | |
728 | static int last_khugepaged_target_node = NUMA_NO_NODE; | |
729 | int nid, target_node = 0, max_value = 0; | |
730 | ||
731 | /* find first node with max normal pages hit */ | |
732 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
733 | if (khugepaged_node_load[nid] > max_value) { | |
734 | max_value = khugepaged_node_load[nid]; | |
735 | target_node = nid; | |
736 | } | |
737 | ||
738 | /* do some balance if several nodes have the same hit record */ | |
739 | if (target_node <= last_khugepaged_target_node) | |
740 | for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; | |
741 | nid++) | |
742 | if (max_value == khugepaged_node_load[nid]) { | |
743 | target_node = nid; | |
744 | break; | |
745 | } | |
746 | ||
747 | last_khugepaged_target_node = target_node; | |
748 | return target_node; | |
749 | } | |
750 | ||
751 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) | |
752 | { | |
753 | if (IS_ERR(*hpage)) { | |
754 | if (!*wait) | |
755 | return false; | |
756 | ||
757 | *wait = false; | |
758 | *hpage = NULL; | |
759 | khugepaged_alloc_sleep(); | |
760 | } else if (*hpage) { | |
761 | put_page(*hpage); | |
762 | *hpage = NULL; | |
763 | } | |
764 | ||
765 | return true; | |
766 | } | |
767 | ||
768 | static struct page * | |
988ddb71 | 769 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node) |
b46e756f KS |
770 | { |
771 | VM_BUG_ON_PAGE(*hpage, *hpage); | |
772 | ||
b46e756f KS |
773 | *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER); |
774 | if (unlikely(!*hpage)) { | |
775 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); | |
776 | *hpage = ERR_PTR(-ENOMEM); | |
777 | return NULL; | |
778 | } | |
779 | ||
780 | prep_transhuge_page(*hpage); | |
781 | count_vm_event(THP_COLLAPSE_ALLOC); | |
782 | return *hpage; | |
783 | } | |
784 | #else | |
785 | static int khugepaged_find_target_node(void) | |
786 | { | |
787 | return 0; | |
788 | } | |
789 | ||
790 | static inline struct page *alloc_khugepaged_hugepage(void) | |
791 | { | |
792 | struct page *page; | |
793 | ||
794 | page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(), | |
795 | HPAGE_PMD_ORDER); | |
796 | if (page) | |
797 | prep_transhuge_page(page); | |
798 | return page; | |
799 | } | |
800 | ||
801 | static struct page *khugepaged_alloc_hugepage(bool *wait) | |
802 | { | |
803 | struct page *hpage; | |
804 | ||
805 | do { | |
806 | hpage = alloc_khugepaged_hugepage(); | |
807 | if (!hpage) { | |
808 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); | |
809 | if (!*wait) | |
810 | return NULL; | |
811 | ||
812 | *wait = false; | |
813 | khugepaged_alloc_sleep(); | |
814 | } else | |
815 | count_vm_event(THP_COLLAPSE_ALLOC); | |
816 | } while (unlikely(!hpage) && likely(khugepaged_enabled())); | |
817 | ||
818 | return hpage; | |
819 | } | |
820 | ||
821 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) | |
822 | { | |
823 | if (!*hpage) | |
824 | *hpage = khugepaged_alloc_hugepage(wait); | |
825 | ||
826 | if (unlikely(!*hpage)) | |
827 | return false; | |
828 | ||
829 | return true; | |
830 | } | |
831 | ||
832 | static struct page * | |
988ddb71 | 833 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node) |
b46e756f | 834 | { |
b46e756f KS |
835 | VM_BUG_ON(!*hpage); |
836 | ||
837 | return *hpage; | |
838 | } | |
839 | #endif | |
840 | ||
b46e756f KS |
841 | /* |
842 | * If mmap_sem temporarily dropped, revalidate vma | |
843 | * before taking mmap_sem. | |
844 | * Return 0 if succeeds, otherwise return none-zero | |
845 | * value (scan code). | |
846 | */ | |
847 | ||
c131f751 KS |
848 | static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, |
849 | struct vm_area_struct **vmap) | |
b46e756f KS |
850 | { |
851 | struct vm_area_struct *vma; | |
852 | unsigned long hstart, hend; | |
853 | ||
854 | if (unlikely(khugepaged_test_exit(mm))) | |
855 | return SCAN_ANY_PROCESS; | |
856 | ||
c131f751 | 857 | *vmap = vma = find_vma(mm, address); |
b46e756f KS |
858 | if (!vma) |
859 | return SCAN_VMA_NULL; | |
860 | ||
861 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; | |
862 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
863 | if (address < hstart || address + HPAGE_PMD_SIZE > hend) | |
864 | return SCAN_ADDRESS_RANGE; | |
50f8b92f | 865 | if (!hugepage_vma_check(vma, vma->vm_flags)) |
b46e756f KS |
866 | return SCAN_VMA_CHECK; |
867 | return 0; | |
868 | } | |
869 | ||
870 | /* | |
871 | * Bring missing pages in from swap, to complete THP collapse. | |
872 | * Only done if khugepaged_scan_pmd believes it is worthwhile. | |
873 | * | |
874 | * Called and returns without pte mapped or spinlocks held, | |
875 | * but with mmap_sem held to protect against vma changes. | |
876 | */ | |
877 | ||
878 | static bool __collapse_huge_page_swapin(struct mm_struct *mm, | |
879 | struct vm_area_struct *vma, | |
0db501f7 EA |
880 | unsigned long address, pmd_t *pmd, |
881 | int referenced) | |
b46e756f | 882 | { |
2b740303 SJ |
883 | int swapped_in = 0; |
884 | vm_fault_t ret = 0; | |
82b0f8c3 | 885 | struct vm_fault vmf = { |
b46e756f KS |
886 | .vma = vma, |
887 | .address = address, | |
888 | .flags = FAULT_FLAG_ALLOW_RETRY, | |
889 | .pmd = pmd, | |
0721ec8b | 890 | .pgoff = linear_page_index(vma, address), |
b46e756f KS |
891 | }; |
892 | ||
982785c6 EA |
893 | /* we only decide to swapin, if there is enough young ptes */ |
894 | if (referenced < HPAGE_PMD_NR/2) { | |
895 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); | |
896 | return false; | |
897 | } | |
82b0f8c3 JK |
898 | vmf.pte = pte_offset_map(pmd, address); |
899 | for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE; | |
900 | vmf.pte++, vmf.address += PAGE_SIZE) { | |
2994302b JK |
901 | vmf.orig_pte = *vmf.pte; |
902 | if (!is_swap_pte(vmf.orig_pte)) | |
b46e756f KS |
903 | continue; |
904 | swapped_in++; | |
2994302b | 905 | ret = do_swap_page(&vmf); |
0db501f7 | 906 | |
b46e756f KS |
907 | /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */ |
908 | if (ret & VM_FAULT_RETRY) { | |
909 | down_read(&mm->mmap_sem); | |
82b0f8c3 | 910 | if (hugepage_vma_revalidate(mm, address, &vmf.vma)) { |
47f863ea | 911 | /* vma is no longer available, don't continue to swapin */ |
0db501f7 | 912 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); |
b46e756f | 913 | return false; |
47f863ea | 914 | } |
b46e756f | 915 | /* check if the pmd is still valid */ |
835152a2 SP |
916 | if (mm_find_pmd(mm, address) != pmd) { |
917 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); | |
b46e756f | 918 | return false; |
835152a2 | 919 | } |
b46e756f KS |
920 | } |
921 | if (ret & VM_FAULT_ERROR) { | |
0db501f7 | 922 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); |
b46e756f KS |
923 | return false; |
924 | } | |
925 | /* pte is unmapped now, we need to map it */ | |
82b0f8c3 | 926 | vmf.pte = pte_offset_map(pmd, vmf.address); |
b46e756f | 927 | } |
82b0f8c3 JK |
928 | vmf.pte--; |
929 | pte_unmap(vmf.pte); | |
0db501f7 | 930 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1); |
b46e756f KS |
931 | return true; |
932 | } | |
933 | ||
934 | static void collapse_huge_page(struct mm_struct *mm, | |
935 | unsigned long address, | |
936 | struct page **hpage, | |
0db501f7 | 937 | int node, int referenced) |
b46e756f KS |
938 | { |
939 | pmd_t *pmd, _pmd; | |
940 | pte_t *pte; | |
941 | pgtable_t pgtable; | |
942 | struct page *new_page; | |
943 | spinlock_t *pmd_ptl, *pte_ptl; | |
944 | int isolated = 0, result = 0; | |
945 | struct mem_cgroup *memcg; | |
c131f751 | 946 | struct vm_area_struct *vma; |
ac46d4f3 | 947 | struct mmu_notifier_range range; |
b46e756f KS |
948 | gfp_t gfp; |
949 | ||
950 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
951 | ||
952 | /* Only allocate from the target node */ | |
41b6167e | 953 | gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE; |
b46e756f | 954 | |
988ddb71 KS |
955 | /* |
956 | * Before allocating the hugepage, release the mmap_sem read lock. | |
957 | * The allocation can take potentially a long time if it involves | |
958 | * sync compaction, and we do not need to hold the mmap_sem during | |
959 | * that. We will recheck the vma after taking it again in write mode. | |
960 | */ | |
961 | up_read(&mm->mmap_sem); | |
962 | new_page = khugepaged_alloc_page(hpage, gfp, node); | |
b46e756f KS |
963 | if (!new_page) { |
964 | result = SCAN_ALLOC_HUGE_PAGE_FAIL; | |
965 | goto out_nolock; | |
966 | } | |
967 | ||
2a70f6a7 | 968 | if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) { |
b46e756f KS |
969 | result = SCAN_CGROUP_CHARGE_FAIL; |
970 | goto out_nolock; | |
971 | } | |
972 | ||
973 | down_read(&mm->mmap_sem); | |
c131f751 | 974 | result = hugepage_vma_revalidate(mm, address, &vma); |
b46e756f KS |
975 | if (result) { |
976 | mem_cgroup_cancel_charge(new_page, memcg, true); | |
977 | up_read(&mm->mmap_sem); | |
978 | goto out_nolock; | |
979 | } | |
980 | ||
981 | pmd = mm_find_pmd(mm, address); | |
982 | if (!pmd) { | |
983 | result = SCAN_PMD_NULL; | |
984 | mem_cgroup_cancel_charge(new_page, memcg, true); | |
985 | up_read(&mm->mmap_sem); | |
986 | goto out_nolock; | |
987 | } | |
988 | ||
989 | /* | |
990 | * __collapse_huge_page_swapin always returns with mmap_sem locked. | |
47f863ea | 991 | * If it fails, we release mmap_sem and jump out_nolock. |
b46e756f KS |
992 | * Continuing to collapse causes inconsistency. |
993 | */ | |
0db501f7 | 994 | if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) { |
b46e756f KS |
995 | mem_cgroup_cancel_charge(new_page, memcg, true); |
996 | up_read(&mm->mmap_sem); | |
997 | goto out_nolock; | |
998 | } | |
999 | ||
1000 | up_read(&mm->mmap_sem); | |
1001 | /* | |
1002 | * Prevent all access to pagetables with the exception of | |
1003 | * gup_fast later handled by the ptep_clear_flush and the VM | |
1004 | * handled by the anon_vma lock + PG_lock. | |
1005 | */ | |
1006 | down_write(&mm->mmap_sem); | |
59ea6d06 AA |
1007 | result = SCAN_ANY_PROCESS; |
1008 | if (!mmget_still_valid(mm)) | |
1009 | goto out; | |
c131f751 | 1010 | result = hugepage_vma_revalidate(mm, address, &vma); |
b46e756f KS |
1011 | if (result) |
1012 | goto out; | |
1013 | /* check if the pmd is still valid */ | |
1014 | if (mm_find_pmd(mm, address) != pmd) | |
1015 | goto out; | |
1016 | ||
1017 | anon_vma_lock_write(vma->anon_vma); | |
1018 | ||
7269f999 | 1019 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, |
6f4f13e8 | 1020 | address, address + HPAGE_PMD_SIZE); |
ac46d4f3 | 1021 | mmu_notifier_invalidate_range_start(&range); |
2d617a03 VS |
1022 | |
1023 | pte = pte_offset_map(pmd, address); | |
1024 | pte_ptl = pte_lockptr(mm, pmd); | |
1025 | ||
b46e756f KS |
1026 | pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ |
1027 | /* | |
1028 | * After this gup_fast can't run anymore. This also removes | |
1029 | * any huge TLB entry from the CPU so we won't allow | |
1030 | * huge and small TLB entries for the same virtual address | |
1031 | * to avoid the risk of CPU bugs in that area. | |
1032 | */ | |
1033 | _pmd = pmdp_collapse_flush(vma, address, pmd); | |
1034 | spin_unlock(pmd_ptl); | |
ac46d4f3 | 1035 | mmu_notifier_invalidate_range_end(&range); |
b46e756f KS |
1036 | |
1037 | spin_lock(pte_ptl); | |
1038 | isolated = __collapse_huge_page_isolate(vma, address, pte); | |
1039 | spin_unlock(pte_ptl); | |
1040 | ||
1041 | if (unlikely(!isolated)) { | |
1042 | pte_unmap(pte); | |
1043 | spin_lock(pmd_ptl); | |
1044 | BUG_ON(!pmd_none(*pmd)); | |
1045 | /* | |
1046 | * We can only use set_pmd_at when establishing | |
1047 | * hugepmds and never for establishing regular pmds that | |
1048 | * points to regular pagetables. Use pmd_populate for that | |
1049 | */ | |
1050 | pmd_populate(mm, pmd, pmd_pgtable(_pmd)); | |
1051 | spin_unlock(pmd_ptl); | |
1052 | anon_vma_unlock_write(vma->anon_vma); | |
1053 | result = SCAN_FAIL; | |
1054 | goto out; | |
1055 | } | |
1056 | ||
1057 | /* | |
1058 | * All pages are isolated and locked so anon_vma rmap | |
1059 | * can't run anymore. | |
1060 | */ | |
1061 | anon_vma_unlock_write(vma->anon_vma); | |
1062 | ||
1063 | __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl); | |
1064 | pte_unmap(pte); | |
1065 | __SetPageUptodate(new_page); | |
1066 | pgtable = pmd_pgtable(_pmd); | |
1067 | ||
1068 | _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); | |
f55e1014 | 1069 | _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); |
b46e756f KS |
1070 | |
1071 | /* | |
1072 | * spin_lock() below is not the equivalent of smp_wmb(), so | |
1073 | * this is needed to avoid the copy_huge_page writes to become | |
1074 | * visible after the set_pmd_at() write. | |
1075 | */ | |
1076 | smp_wmb(); | |
1077 | ||
1078 | spin_lock(pmd_ptl); | |
1079 | BUG_ON(!pmd_none(*pmd)); | |
1080 | page_add_new_anon_rmap(new_page, vma, address, true); | |
1081 | mem_cgroup_commit_charge(new_page, memcg, false, true); | |
1ff9e6e1 | 1082 | count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1); |
b46e756f KS |
1083 | lru_cache_add_active_or_unevictable(new_page, vma); |
1084 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
1085 | set_pmd_at(mm, address, pmd, _pmd); | |
1086 | update_mmu_cache_pmd(vma, address, pmd); | |
1087 | spin_unlock(pmd_ptl); | |
1088 | ||
1089 | *hpage = NULL; | |
1090 | ||
1091 | khugepaged_pages_collapsed++; | |
1092 | result = SCAN_SUCCEED; | |
1093 | out_up_write: | |
1094 | up_write(&mm->mmap_sem); | |
1095 | out_nolock: | |
1096 | trace_mm_collapse_huge_page(mm, isolated, result); | |
1097 | return; | |
1098 | out: | |
1099 | mem_cgroup_cancel_charge(new_page, memcg, true); | |
1100 | goto out_up_write; | |
1101 | } | |
1102 | ||
1103 | static int khugepaged_scan_pmd(struct mm_struct *mm, | |
1104 | struct vm_area_struct *vma, | |
1105 | unsigned long address, | |
1106 | struct page **hpage) | |
1107 | { | |
1108 | pmd_t *pmd; | |
1109 | pte_t *pte, *_pte; | |
0db501f7 | 1110 | int ret = 0, none_or_zero = 0, result = 0, referenced = 0; |
b46e756f KS |
1111 | struct page *page = NULL; |
1112 | unsigned long _address; | |
1113 | spinlock_t *ptl; | |
1114 | int node = NUMA_NO_NODE, unmapped = 0; | |
0db501f7 | 1115 | bool writable = false; |
b46e756f KS |
1116 | |
1117 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
1118 | ||
1119 | pmd = mm_find_pmd(mm, address); | |
1120 | if (!pmd) { | |
1121 | result = SCAN_PMD_NULL; | |
1122 | goto out; | |
1123 | } | |
1124 | ||
1125 | memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); | |
1126 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
1127 | for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; | |
1128 | _pte++, _address += PAGE_SIZE) { | |
1129 | pte_t pteval = *_pte; | |
1130 | if (is_swap_pte(pteval)) { | |
1131 | if (++unmapped <= khugepaged_max_ptes_swap) { | |
1132 | continue; | |
1133 | } else { | |
1134 | result = SCAN_EXCEED_SWAP_PTE; | |
1135 | goto out_unmap; | |
1136 | } | |
1137 | } | |
1138 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { | |
1139 | if (!userfaultfd_armed(vma) && | |
1140 | ++none_or_zero <= khugepaged_max_ptes_none) { | |
1141 | continue; | |
1142 | } else { | |
1143 | result = SCAN_EXCEED_NONE_PTE; | |
1144 | goto out_unmap; | |
1145 | } | |
1146 | } | |
1147 | if (!pte_present(pteval)) { | |
1148 | result = SCAN_PTE_NON_PRESENT; | |
1149 | goto out_unmap; | |
1150 | } | |
1151 | if (pte_write(pteval)) | |
1152 | writable = true; | |
1153 | ||
1154 | page = vm_normal_page(vma, _address, pteval); | |
1155 | if (unlikely(!page)) { | |
1156 | result = SCAN_PAGE_NULL; | |
1157 | goto out_unmap; | |
1158 | } | |
1159 | ||
1160 | /* TODO: teach khugepaged to collapse THP mapped with pte */ | |
1161 | if (PageCompound(page)) { | |
1162 | result = SCAN_PAGE_COMPOUND; | |
1163 | goto out_unmap; | |
1164 | } | |
1165 | ||
1166 | /* | |
1167 | * Record which node the original page is from and save this | |
1168 | * information to khugepaged_node_load[]. | |
1169 | * Khupaged will allocate hugepage from the node has the max | |
1170 | * hit record. | |
1171 | */ | |
1172 | node = page_to_nid(page); | |
1173 | if (khugepaged_scan_abort(node)) { | |
1174 | result = SCAN_SCAN_ABORT; | |
1175 | goto out_unmap; | |
1176 | } | |
1177 | khugepaged_node_load[node]++; | |
1178 | if (!PageLRU(page)) { | |
1179 | result = SCAN_PAGE_LRU; | |
1180 | goto out_unmap; | |
1181 | } | |
1182 | if (PageLocked(page)) { | |
1183 | result = SCAN_PAGE_LOCK; | |
1184 | goto out_unmap; | |
1185 | } | |
1186 | if (!PageAnon(page)) { | |
1187 | result = SCAN_PAGE_ANON; | |
1188 | goto out_unmap; | |
1189 | } | |
1190 | ||
1191 | /* | |
1192 | * cannot use mapcount: can't collapse if there's a gup pin. | |
1193 | * The page must only be referenced by the scanned process | |
1194 | * and page swap cache. | |
1195 | */ | |
2948be5a | 1196 | if (page_count(page) != 1 + PageSwapCache(page)) { |
b46e756f KS |
1197 | result = SCAN_PAGE_COUNT; |
1198 | goto out_unmap; | |
1199 | } | |
1200 | if (pte_young(pteval) || | |
1201 | page_is_young(page) || PageReferenced(page) || | |
1202 | mmu_notifier_test_young(vma->vm_mm, address)) | |
0db501f7 | 1203 | referenced++; |
b46e756f KS |
1204 | } |
1205 | if (writable) { | |
1206 | if (referenced) { | |
1207 | result = SCAN_SUCCEED; | |
1208 | ret = 1; | |
1209 | } else { | |
0db501f7 | 1210 | result = SCAN_LACK_REFERENCED_PAGE; |
b46e756f KS |
1211 | } |
1212 | } else { | |
1213 | result = SCAN_PAGE_RO; | |
1214 | } | |
1215 | out_unmap: | |
1216 | pte_unmap_unlock(pte, ptl); | |
1217 | if (ret) { | |
1218 | node = khugepaged_find_target_node(); | |
1219 | /* collapse_huge_page will return with the mmap_sem released */ | |
c131f751 | 1220 | collapse_huge_page(mm, address, hpage, node, referenced); |
b46e756f KS |
1221 | } |
1222 | out: | |
1223 | trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced, | |
1224 | none_or_zero, result, unmapped); | |
1225 | return ret; | |
1226 | } | |
1227 | ||
1228 | static void collect_mm_slot(struct mm_slot *mm_slot) | |
1229 | { | |
1230 | struct mm_struct *mm = mm_slot->mm; | |
1231 | ||
35f3aa39 | 1232 | lockdep_assert_held(&khugepaged_mm_lock); |
b46e756f KS |
1233 | |
1234 | if (khugepaged_test_exit(mm)) { | |
1235 | /* free mm_slot */ | |
1236 | hash_del(&mm_slot->hash); | |
1237 | list_del(&mm_slot->mm_node); | |
1238 | ||
1239 | /* | |
1240 | * Not strictly needed because the mm exited already. | |
1241 | * | |
1242 | * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
1243 | */ | |
1244 | ||
1245 | /* khugepaged_mm_lock actually not necessary for the below */ | |
1246 | free_mm_slot(mm_slot); | |
1247 | mmdrop(mm); | |
1248 | } | |
1249 | } | |
1250 | ||
e496cf3d | 1251 | #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) |
f3f0e1d2 KS |
1252 | static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff) |
1253 | { | |
1254 | struct vm_area_struct *vma; | |
1255 | unsigned long addr; | |
1256 | pmd_t *pmd, _pmd; | |
1257 | ||
1258 | i_mmap_lock_write(mapping); | |
1259 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { | |
1260 | /* probably overkill */ | |
1261 | if (vma->anon_vma) | |
1262 | continue; | |
1263 | addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
1264 | if (addr & ~HPAGE_PMD_MASK) | |
1265 | continue; | |
1266 | if (vma->vm_end < addr + HPAGE_PMD_SIZE) | |
1267 | continue; | |
1268 | pmd = mm_find_pmd(vma->vm_mm, addr); | |
1269 | if (!pmd) | |
1270 | continue; | |
1271 | /* | |
1272 | * We need exclusive mmap_sem to retract page table. | |
1273 | * If trylock fails we would end up with pte-mapped THP after | |
1274 | * re-fault. Not ideal, but it's more important to not disturb | |
1275 | * the system too much. | |
1276 | */ | |
1277 | if (down_write_trylock(&vma->vm_mm->mmap_sem)) { | |
1278 | spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd); | |
1279 | /* assume page table is clear */ | |
1280 | _pmd = pmdp_collapse_flush(vma, addr, pmd); | |
1281 | spin_unlock(ptl); | |
1282 | up_write(&vma->vm_mm->mmap_sem); | |
c4812909 | 1283 | mm_dec_nr_ptes(vma->vm_mm); |
d670ffd8 | 1284 | pte_free(vma->vm_mm, pmd_pgtable(_pmd)); |
f3f0e1d2 KS |
1285 | } |
1286 | } | |
1287 | i_mmap_unlock_write(mapping); | |
1288 | } | |
1289 | ||
1290 | /** | |
1291 | * collapse_shmem - collapse small tmpfs/shmem pages into huge one. | |
1292 | * | |
1293 | * Basic scheme is simple, details are more complex: | |
87c460a0 | 1294 | * - allocate and lock a new huge page; |
77da9389 | 1295 | * - scan page cache replacing old pages with the new one |
f3f0e1d2 KS |
1296 | * + swap in pages if necessary; |
1297 | * + fill in gaps; | |
77da9389 MW |
1298 | * + keep old pages around in case rollback is required; |
1299 | * - if replacing succeeds: | |
f3f0e1d2 KS |
1300 | * + copy data over; |
1301 | * + free old pages; | |
87c460a0 | 1302 | * + unlock huge page; |
f3f0e1d2 KS |
1303 | * - if replacing failed; |
1304 | * + put all pages back and unfreeze them; | |
77da9389 | 1305 | * + restore gaps in the page cache; |
87c460a0 | 1306 | * + unlock and free huge page; |
f3f0e1d2 KS |
1307 | */ |
1308 | static void collapse_shmem(struct mm_struct *mm, | |
1309 | struct address_space *mapping, pgoff_t start, | |
1310 | struct page **hpage, int node) | |
1311 | { | |
1312 | gfp_t gfp; | |
77da9389 | 1313 | struct page *new_page; |
f3f0e1d2 KS |
1314 | struct mem_cgroup *memcg; |
1315 | pgoff_t index, end = start + HPAGE_PMD_NR; | |
1316 | LIST_HEAD(pagelist); | |
77da9389 | 1317 | XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER); |
f3f0e1d2 KS |
1318 | int nr_none = 0, result = SCAN_SUCCEED; |
1319 | ||
1320 | VM_BUG_ON(start & (HPAGE_PMD_NR - 1)); | |
1321 | ||
1322 | /* Only allocate from the target node */ | |
41b6167e | 1323 | gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE; |
f3f0e1d2 KS |
1324 | |
1325 | new_page = khugepaged_alloc_page(hpage, gfp, node); | |
1326 | if (!new_page) { | |
1327 | result = SCAN_ALLOC_HUGE_PAGE_FAIL; | |
1328 | goto out; | |
1329 | } | |
1330 | ||
2a70f6a7 | 1331 | if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) { |
f3f0e1d2 KS |
1332 | result = SCAN_CGROUP_CHARGE_FAIL; |
1333 | goto out; | |
1334 | } | |
1335 | ||
95feeabb HD |
1336 | /* This will be less messy when we use multi-index entries */ |
1337 | do { | |
1338 | xas_lock_irq(&xas); | |
1339 | xas_create_range(&xas); | |
1340 | if (!xas_error(&xas)) | |
1341 | break; | |
1342 | xas_unlock_irq(&xas); | |
1343 | if (!xas_nomem(&xas, GFP_KERNEL)) { | |
1344 | mem_cgroup_cancel_charge(new_page, memcg, true); | |
1345 | result = SCAN_FAIL; | |
1346 | goto out; | |
1347 | } | |
1348 | } while (1); | |
1349 | ||
042a3082 HD |
1350 | __SetPageLocked(new_page); |
1351 | __SetPageSwapBacked(new_page); | |
f3f0e1d2 KS |
1352 | new_page->index = start; |
1353 | new_page->mapping = mapping; | |
f3f0e1d2 | 1354 | |
f3f0e1d2 | 1355 | /* |
87c460a0 HD |
1356 | * At this point the new_page is locked and not up-to-date. |
1357 | * It's safe to insert it into the page cache, because nobody would | |
1358 | * be able to map it or use it in another way until we unlock it. | |
f3f0e1d2 KS |
1359 | */ |
1360 | ||
77da9389 MW |
1361 | xas_set(&xas, start); |
1362 | for (index = start; index < end; index++) { | |
1363 | struct page *page = xas_next(&xas); | |
1364 | ||
1365 | VM_BUG_ON(index != xas.xa_index); | |
1366 | if (!page) { | |
701270fa HD |
1367 | /* |
1368 | * Stop if extent has been truncated or hole-punched, | |
1369 | * and is now completely empty. | |
1370 | */ | |
1371 | if (index == start) { | |
1372 | if (!xas_next_entry(&xas, end - 1)) { | |
1373 | result = SCAN_TRUNCATED; | |
042a3082 | 1374 | goto xa_locked; |
701270fa HD |
1375 | } |
1376 | xas_set(&xas, index); | |
1377 | } | |
77da9389 MW |
1378 | if (!shmem_charge(mapping->host, 1)) { |
1379 | result = SCAN_FAIL; | |
042a3082 | 1380 | goto xa_locked; |
77da9389 | 1381 | } |
69bf4b6b | 1382 | xas_store(&xas, new_page + (index % HPAGE_PMD_NR)); |
77da9389 MW |
1383 | nr_none++; |
1384 | continue; | |
1385 | } | |
f3f0e1d2 | 1386 | |
3159f943 | 1387 | if (xa_is_value(page) || !PageUptodate(page)) { |
77da9389 | 1388 | xas_unlock_irq(&xas); |
f3f0e1d2 KS |
1389 | /* swap in or instantiate fallocated page */ |
1390 | if (shmem_getpage(mapping->host, index, &page, | |
1391 | SGP_NOHUGE)) { | |
1392 | result = SCAN_FAIL; | |
77da9389 | 1393 | goto xa_unlocked; |
f3f0e1d2 | 1394 | } |
f3f0e1d2 KS |
1395 | } else if (trylock_page(page)) { |
1396 | get_page(page); | |
042a3082 | 1397 | xas_unlock_irq(&xas); |
f3f0e1d2 KS |
1398 | } else { |
1399 | result = SCAN_PAGE_LOCK; | |
042a3082 | 1400 | goto xa_locked; |
f3f0e1d2 KS |
1401 | } |
1402 | ||
1403 | /* | |
b93b0163 | 1404 | * The page must be locked, so we can drop the i_pages lock |
f3f0e1d2 KS |
1405 | * without racing with truncate. |
1406 | */ | |
1407 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
1408 | VM_BUG_ON_PAGE(!PageUptodate(page), page); | |
06a5e126 HD |
1409 | |
1410 | /* | |
1411 | * If file was truncated then extended, or hole-punched, before | |
1412 | * we locked the first page, then a THP might be there already. | |
1413 | */ | |
1414 | if (PageTransCompound(page)) { | |
1415 | result = SCAN_PAGE_COMPOUND; | |
1416 | goto out_unlock; | |
1417 | } | |
f3f0e1d2 KS |
1418 | |
1419 | if (page_mapping(page) != mapping) { | |
1420 | result = SCAN_TRUNCATED; | |
1421 | goto out_unlock; | |
1422 | } | |
f3f0e1d2 KS |
1423 | |
1424 | if (isolate_lru_page(page)) { | |
1425 | result = SCAN_DEL_PAGE_LRU; | |
042a3082 | 1426 | goto out_unlock; |
f3f0e1d2 KS |
1427 | } |
1428 | ||
1429 | if (page_mapped(page)) | |
977fbdcd | 1430 | unmap_mapping_pages(mapping, index, 1, false); |
f3f0e1d2 | 1431 | |
77da9389 MW |
1432 | xas_lock_irq(&xas); |
1433 | xas_set(&xas, index); | |
f3f0e1d2 | 1434 | |
77da9389 | 1435 | VM_BUG_ON_PAGE(page != xas_load(&xas), page); |
f3f0e1d2 KS |
1436 | VM_BUG_ON_PAGE(page_mapped(page), page); |
1437 | ||
1438 | /* | |
1439 | * The page is expected to have page_count() == 3: | |
1440 | * - we hold a pin on it; | |
77da9389 | 1441 | * - one reference from page cache; |
f3f0e1d2 KS |
1442 | * - one from isolate_lru_page; |
1443 | */ | |
1444 | if (!page_ref_freeze(page, 3)) { | |
1445 | result = SCAN_PAGE_COUNT; | |
042a3082 HD |
1446 | xas_unlock_irq(&xas); |
1447 | putback_lru_page(page); | |
1448 | goto out_unlock; | |
f3f0e1d2 KS |
1449 | } |
1450 | ||
1451 | /* | |
1452 | * Add the page to the list to be able to undo the collapse if | |
1453 | * something go wrong. | |
1454 | */ | |
1455 | list_add_tail(&page->lru, &pagelist); | |
1456 | ||
1457 | /* Finally, replace with the new page. */ | |
69bf4b6b | 1458 | xas_store(&xas, new_page + (index % HPAGE_PMD_NR)); |
f3f0e1d2 | 1459 | continue; |
f3f0e1d2 KS |
1460 | out_unlock: |
1461 | unlock_page(page); | |
1462 | put_page(page); | |
042a3082 | 1463 | goto xa_unlocked; |
f3f0e1d2 KS |
1464 | } |
1465 | ||
042a3082 HD |
1466 | __inc_node_page_state(new_page, NR_SHMEM_THPS); |
1467 | if (nr_none) { | |
1468 | struct zone *zone = page_zone(new_page); | |
1469 | ||
1470 | __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none); | |
1471 | __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none); | |
1472 | } | |
1473 | ||
1474 | xa_locked: | |
1475 | xas_unlock_irq(&xas); | |
77da9389 | 1476 | xa_unlocked: |
042a3082 | 1477 | |
f3f0e1d2 | 1478 | if (result == SCAN_SUCCEED) { |
77da9389 | 1479 | struct page *page, *tmp; |
f3f0e1d2 KS |
1480 | |
1481 | /* | |
77da9389 MW |
1482 | * Replacing old pages with new one has succeeded, now we |
1483 | * need to copy the content and free the old pages. | |
f3f0e1d2 | 1484 | */ |
2af8ff29 | 1485 | index = start; |
f3f0e1d2 | 1486 | list_for_each_entry_safe(page, tmp, &pagelist, lru) { |
2af8ff29 HD |
1487 | while (index < page->index) { |
1488 | clear_highpage(new_page + (index % HPAGE_PMD_NR)); | |
1489 | index++; | |
1490 | } | |
f3f0e1d2 KS |
1491 | copy_highpage(new_page + (page->index % HPAGE_PMD_NR), |
1492 | page); | |
1493 | list_del(&page->lru); | |
f3f0e1d2 | 1494 | page->mapping = NULL; |
042a3082 | 1495 | page_ref_unfreeze(page, 1); |
f3f0e1d2 KS |
1496 | ClearPageActive(page); |
1497 | ClearPageUnevictable(page); | |
042a3082 | 1498 | unlock_page(page); |
f3f0e1d2 | 1499 | put_page(page); |
2af8ff29 HD |
1500 | index++; |
1501 | } | |
1502 | while (index < end) { | |
1503 | clear_highpage(new_page + (index % HPAGE_PMD_NR)); | |
1504 | index++; | |
f3f0e1d2 KS |
1505 | } |
1506 | ||
f3f0e1d2 | 1507 | SetPageUptodate(new_page); |
87c460a0 | 1508 | page_ref_add(new_page, HPAGE_PMD_NR - 1); |
042a3082 | 1509 | set_page_dirty(new_page); |
f3f0e1d2 | 1510 | mem_cgroup_commit_charge(new_page, memcg, false, true); |
1ff9e6e1 | 1511 | count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1); |
f3f0e1d2 | 1512 | lru_cache_add_anon(new_page); |
f3f0e1d2 | 1513 | |
042a3082 HD |
1514 | /* |
1515 | * Remove pte page tables, so we can re-fault the page as huge. | |
1516 | */ | |
1517 | retract_page_tables(mapping, start); | |
f3f0e1d2 | 1518 | *hpage = NULL; |
87aa7529 YS |
1519 | |
1520 | khugepaged_pages_collapsed++; | |
f3f0e1d2 | 1521 | } else { |
77da9389 | 1522 | struct page *page; |
aaa52e34 | 1523 | |
77da9389 | 1524 | /* Something went wrong: roll back page cache changes */ |
77da9389 | 1525 | xas_lock_irq(&xas); |
aaa52e34 HD |
1526 | mapping->nrpages -= nr_none; |
1527 | shmem_uncharge(mapping->host, nr_none); | |
1528 | ||
77da9389 MW |
1529 | xas_set(&xas, start); |
1530 | xas_for_each(&xas, page, end - 1) { | |
f3f0e1d2 KS |
1531 | page = list_first_entry_or_null(&pagelist, |
1532 | struct page, lru); | |
77da9389 | 1533 | if (!page || xas.xa_index < page->index) { |
f3f0e1d2 KS |
1534 | if (!nr_none) |
1535 | break; | |
f3f0e1d2 | 1536 | nr_none--; |
59749e6c | 1537 | /* Put holes back where they were */ |
77da9389 | 1538 | xas_store(&xas, NULL); |
f3f0e1d2 KS |
1539 | continue; |
1540 | } | |
1541 | ||
77da9389 | 1542 | VM_BUG_ON_PAGE(page->index != xas.xa_index, page); |
f3f0e1d2 KS |
1543 | |
1544 | /* Unfreeze the page. */ | |
1545 | list_del(&page->lru); | |
1546 | page_ref_unfreeze(page, 2); | |
77da9389 MW |
1547 | xas_store(&xas, page); |
1548 | xas_pause(&xas); | |
1549 | xas_unlock_irq(&xas); | |
f3f0e1d2 | 1550 | unlock_page(page); |
042a3082 | 1551 | putback_lru_page(page); |
77da9389 | 1552 | xas_lock_irq(&xas); |
f3f0e1d2 KS |
1553 | } |
1554 | VM_BUG_ON(nr_none); | |
77da9389 | 1555 | xas_unlock_irq(&xas); |
f3f0e1d2 | 1556 | |
f3f0e1d2 | 1557 | mem_cgroup_cancel_charge(new_page, memcg, true); |
f3f0e1d2 KS |
1558 | new_page->mapping = NULL; |
1559 | } | |
042a3082 HD |
1560 | |
1561 | unlock_page(new_page); | |
f3f0e1d2 KS |
1562 | out: |
1563 | VM_BUG_ON(!list_empty(&pagelist)); | |
1564 | /* TODO: tracepoints */ | |
1565 | } | |
1566 | ||
1567 | static void khugepaged_scan_shmem(struct mm_struct *mm, | |
1568 | struct address_space *mapping, | |
1569 | pgoff_t start, struct page **hpage) | |
1570 | { | |
1571 | struct page *page = NULL; | |
85b392db | 1572 | XA_STATE(xas, &mapping->i_pages, start); |
f3f0e1d2 KS |
1573 | int present, swap; |
1574 | int node = NUMA_NO_NODE; | |
1575 | int result = SCAN_SUCCEED; | |
1576 | ||
1577 | present = 0; | |
1578 | swap = 0; | |
1579 | memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); | |
1580 | rcu_read_lock(); | |
85b392db MW |
1581 | xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) { |
1582 | if (xas_retry(&xas, page)) | |
f3f0e1d2 | 1583 | continue; |
f3f0e1d2 | 1584 | |
85b392db | 1585 | if (xa_is_value(page)) { |
f3f0e1d2 KS |
1586 | if (++swap > khugepaged_max_ptes_swap) { |
1587 | result = SCAN_EXCEED_SWAP_PTE; | |
1588 | break; | |
1589 | } | |
1590 | continue; | |
1591 | } | |
1592 | ||
1593 | if (PageTransCompound(page)) { | |
1594 | result = SCAN_PAGE_COMPOUND; | |
1595 | break; | |
1596 | } | |
1597 | ||
1598 | node = page_to_nid(page); | |
1599 | if (khugepaged_scan_abort(node)) { | |
1600 | result = SCAN_SCAN_ABORT; | |
1601 | break; | |
1602 | } | |
1603 | khugepaged_node_load[node]++; | |
1604 | ||
1605 | if (!PageLRU(page)) { | |
1606 | result = SCAN_PAGE_LRU; | |
1607 | break; | |
1608 | } | |
1609 | ||
1610 | if (page_count(page) != 1 + page_mapcount(page)) { | |
1611 | result = SCAN_PAGE_COUNT; | |
1612 | break; | |
1613 | } | |
1614 | ||
1615 | /* | |
1616 | * We probably should check if the page is referenced here, but | |
1617 | * nobody would transfer pte_young() to PageReferenced() for us. | |
1618 | * And rmap walk here is just too costly... | |
1619 | */ | |
1620 | ||
1621 | present++; | |
1622 | ||
1623 | if (need_resched()) { | |
85b392db | 1624 | xas_pause(&xas); |
f3f0e1d2 | 1625 | cond_resched_rcu(); |
f3f0e1d2 KS |
1626 | } |
1627 | } | |
1628 | rcu_read_unlock(); | |
1629 | ||
1630 | if (result == SCAN_SUCCEED) { | |
1631 | if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) { | |
1632 | result = SCAN_EXCEED_NONE_PTE; | |
1633 | } else { | |
1634 | node = khugepaged_find_target_node(); | |
1635 | collapse_shmem(mm, mapping, start, hpage, node); | |
1636 | } | |
1637 | } | |
1638 | ||
1639 | /* TODO: tracepoints */ | |
1640 | } | |
1641 | #else | |
1642 | static void khugepaged_scan_shmem(struct mm_struct *mm, | |
1643 | struct address_space *mapping, | |
1644 | pgoff_t start, struct page **hpage) | |
1645 | { | |
1646 | BUILD_BUG(); | |
1647 | } | |
1648 | #endif | |
1649 | ||
b46e756f KS |
1650 | static unsigned int khugepaged_scan_mm_slot(unsigned int pages, |
1651 | struct page **hpage) | |
1652 | __releases(&khugepaged_mm_lock) | |
1653 | __acquires(&khugepaged_mm_lock) | |
1654 | { | |
1655 | struct mm_slot *mm_slot; | |
1656 | struct mm_struct *mm; | |
1657 | struct vm_area_struct *vma; | |
1658 | int progress = 0; | |
1659 | ||
1660 | VM_BUG_ON(!pages); | |
35f3aa39 | 1661 | lockdep_assert_held(&khugepaged_mm_lock); |
b46e756f KS |
1662 | |
1663 | if (khugepaged_scan.mm_slot) | |
1664 | mm_slot = khugepaged_scan.mm_slot; | |
1665 | else { | |
1666 | mm_slot = list_entry(khugepaged_scan.mm_head.next, | |
1667 | struct mm_slot, mm_node); | |
1668 | khugepaged_scan.address = 0; | |
1669 | khugepaged_scan.mm_slot = mm_slot; | |
1670 | } | |
1671 | spin_unlock(&khugepaged_mm_lock); | |
1672 | ||
1673 | mm = mm_slot->mm; | |
3b454ad3 YS |
1674 | /* |
1675 | * Don't wait for semaphore (to avoid long wait times). Just move to | |
1676 | * the next mm on the list. | |
1677 | */ | |
1678 | vma = NULL; | |
1679 | if (unlikely(!down_read_trylock(&mm->mmap_sem))) | |
1680 | goto breakouterloop_mmap_sem; | |
1681 | if (likely(!khugepaged_test_exit(mm))) | |
b46e756f KS |
1682 | vma = find_vma(mm, khugepaged_scan.address); |
1683 | ||
1684 | progress++; | |
1685 | for (; vma; vma = vma->vm_next) { | |
1686 | unsigned long hstart, hend; | |
1687 | ||
1688 | cond_resched(); | |
1689 | if (unlikely(khugepaged_test_exit(mm))) { | |
1690 | progress++; | |
1691 | break; | |
1692 | } | |
50f8b92f | 1693 | if (!hugepage_vma_check(vma, vma->vm_flags)) { |
b46e756f KS |
1694 | skip: |
1695 | progress++; | |
1696 | continue; | |
1697 | } | |
1698 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; | |
1699 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
1700 | if (hstart >= hend) | |
1701 | goto skip; | |
1702 | if (khugepaged_scan.address > hend) | |
1703 | goto skip; | |
1704 | if (khugepaged_scan.address < hstart) | |
1705 | khugepaged_scan.address = hstart; | |
1706 | VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); | |
1707 | ||
1708 | while (khugepaged_scan.address < hend) { | |
1709 | int ret; | |
1710 | cond_resched(); | |
1711 | if (unlikely(khugepaged_test_exit(mm))) | |
1712 | goto breakouterloop; | |
1713 | ||
1714 | VM_BUG_ON(khugepaged_scan.address < hstart || | |
1715 | khugepaged_scan.address + HPAGE_PMD_SIZE > | |
1716 | hend); | |
f3f0e1d2 | 1717 | if (shmem_file(vma->vm_file)) { |
e496cf3d | 1718 | struct file *file; |
f3f0e1d2 KS |
1719 | pgoff_t pgoff = linear_page_index(vma, |
1720 | khugepaged_scan.address); | |
e496cf3d KS |
1721 | if (!shmem_huge_enabled(vma)) |
1722 | goto skip; | |
1723 | file = get_file(vma->vm_file); | |
f3f0e1d2 KS |
1724 | up_read(&mm->mmap_sem); |
1725 | ret = 1; | |
1726 | khugepaged_scan_shmem(mm, file->f_mapping, | |
1727 | pgoff, hpage); | |
1728 | fput(file); | |
1729 | } else { | |
1730 | ret = khugepaged_scan_pmd(mm, vma, | |
1731 | khugepaged_scan.address, | |
1732 | hpage); | |
1733 | } | |
b46e756f KS |
1734 | /* move to next address */ |
1735 | khugepaged_scan.address += HPAGE_PMD_SIZE; | |
1736 | progress += HPAGE_PMD_NR; | |
1737 | if (ret) | |
1738 | /* we released mmap_sem so break loop */ | |
1739 | goto breakouterloop_mmap_sem; | |
1740 | if (progress >= pages) | |
1741 | goto breakouterloop; | |
1742 | } | |
1743 | } | |
1744 | breakouterloop: | |
1745 | up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ | |
1746 | breakouterloop_mmap_sem: | |
1747 | ||
1748 | spin_lock(&khugepaged_mm_lock); | |
1749 | VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); | |
1750 | /* | |
1751 | * Release the current mm_slot if this mm is about to die, or | |
1752 | * if we scanned all vmas of this mm. | |
1753 | */ | |
1754 | if (khugepaged_test_exit(mm) || !vma) { | |
1755 | /* | |
1756 | * Make sure that if mm_users is reaching zero while | |
1757 | * khugepaged runs here, khugepaged_exit will find | |
1758 | * mm_slot not pointing to the exiting mm. | |
1759 | */ | |
1760 | if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { | |
1761 | khugepaged_scan.mm_slot = list_entry( | |
1762 | mm_slot->mm_node.next, | |
1763 | struct mm_slot, mm_node); | |
1764 | khugepaged_scan.address = 0; | |
1765 | } else { | |
1766 | khugepaged_scan.mm_slot = NULL; | |
1767 | khugepaged_full_scans++; | |
1768 | } | |
1769 | ||
1770 | collect_mm_slot(mm_slot); | |
1771 | } | |
1772 | ||
1773 | return progress; | |
1774 | } | |
1775 | ||
1776 | static int khugepaged_has_work(void) | |
1777 | { | |
1778 | return !list_empty(&khugepaged_scan.mm_head) && | |
1779 | khugepaged_enabled(); | |
1780 | } | |
1781 | ||
1782 | static int khugepaged_wait_event(void) | |
1783 | { | |
1784 | return !list_empty(&khugepaged_scan.mm_head) || | |
1785 | kthread_should_stop(); | |
1786 | } | |
1787 | ||
1788 | static void khugepaged_do_scan(void) | |
1789 | { | |
1790 | struct page *hpage = NULL; | |
1791 | unsigned int progress = 0, pass_through_head = 0; | |
1792 | unsigned int pages = khugepaged_pages_to_scan; | |
1793 | bool wait = true; | |
1794 | ||
1795 | barrier(); /* write khugepaged_pages_to_scan to local stack */ | |
1796 | ||
1797 | while (progress < pages) { | |
1798 | if (!khugepaged_prealloc_page(&hpage, &wait)) | |
1799 | break; | |
1800 | ||
1801 | cond_resched(); | |
1802 | ||
1803 | if (unlikely(kthread_should_stop() || try_to_freeze())) | |
1804 | break; | |
1805 | ||
1806 | spin_lock(&khugepaged_mm_lock); | |
1807 | if (!khugepaged_scan.mm_slot) | |
1808 | pass_through_head++; | |
1809 | if (khugepaged_has_work() && | |
1810 | pass_through_head < 2) | |
1811 | progress += khugepaged_scan_mm_slot(pages - progress, | |
1812 | &hpage); | |
1813 | else | |
1814 | progress = pages; | |
1815 | spin_unlock(&khugepaged_mm_lock); | |
1816 | } | |
1817 | ||
1818 | if (!IS_ERR_OR_NULL(hpage)) | |
1819 | put_page(hpage); | |
1820 | } | |
1821 | ||
1822 | static bool khugepaged_should_wakeup(void) | |
1823 | { | |
1824 | return kthread_should_stop() || | |
1825 | time_after_eq(jiffies, khugepaged_sleep_expire); | |
1826 | } | |
1827 | ||
1828 | static void khugepaged_wait_work(void) | |
1829 | { | |
1830 | if (khugepaged_has_work()) { | |
1831 | const unsigned long scan_sleep_jiffies = | |
1832 | msecs_to_jiffies(khugepaged_scan_sleep_millisecs); | |
1833 | ||
1834 | if (!scan_sleep_jiffies) | |
1835 | return; | |
1836 | ||
1837 | khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; | |
1838 | wait_event_freezable_timeout(khugepaged_wait, | |
1839 | khugepaged_should_wakeup(), | |
1840 | scan_sleep_jiffies); | |
1841 | return; | |
1842 | } | |
1843 | ||
1844 | if (khugepaged_enabled()) | |
1845 | wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); | |
1846 | } | |
1847 | ||
1848 | static int khugepaged(void *none) | |
1849 | { | |
1850 | struct mm_slot *mm_slot; | |
1851 | ||
1852 | set_freezable(); | |
1853 | set_user_nice(current, MAX_NICE); | |
1854 | ||
1855 | while (!kthread_should_stop()) { | |
1856 | khugepaged_do_scan(); | |
1857 | khugepaged_wait_work(); | |
1858 | } | |
1859 | ||
1860 | spin_lock(&khugepaged_mm_lock); | |
1861 | mm_slot = khugepaged_scan.mm_slot; | |
1862 | khugepaged_scan.mm_slot = NULL; | |
1863 | if (mm_slot) | |
1864 | collect_mm_slot(mm_slot); | |
1865 | spin_unlock(&khugepaged_mm_lock); | |
1866 | return 0; | |
1867 | } | |
1868 | ||
1869 | static void set_recommended_min_free_kbytes(void) | |
1870 | { | |
1871 | struct zone *zone; | |
1872 | int nr_zones = 0; | |
1873 | unsigned long recommended_min; | |
1874 | ||
b7d349c7 JK |
1875 | for_each_populated_zone(zone) { |
1876 | /* | |
1877 | * We don't need to worry about fragmentation of | |
1878 | * ZONE_MOVABLE since it only has movable pages. | |
1879 | */ | |
1880 | if (zone_idx(zone) > gfp_zone(GFP_USER)) | |
1881 | continue; | |
1882 | ||
b46e756f | 1883 | nr_zones++; |
b7d349c7 | 1884 | } |
b46e756f KS |
1885 | |
1886 | /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ | |
1887 | recommended_min = pageblock_nr_pages * nr_zones * 2; | |
1888 | ||
1889 | /* | |
1890 | * Make sure that on average at least two pageblocks are almost free | |
1891 | * of another type, one for a migratetype to fall back to and a | |
1892 | * second to avoid subsequent fallbacks of other types There are 3 | |
1893 | * MIGRATE_TYPES we care about. | |
1894 | */ | |
1895 | recommended_min += pageblock_nr_pages * nr_zones * | |
1896 | MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; | |
1897 | ||
1898 | /* don't ever allow to reserve more than 5% of the lowmem */ | |
1899 | recommended_min = min(recommended_min, | |
1900 | (unsigned long) nr_free_buffer_pages() / 20); | |
1901 | recommended_min <<= (PAGE_SHIFT-10); | |
1902 | ||
1903 | if (recommended_min > min_free_kbytes) { | |
1904 | if (user_min_free_kbytes >= 0) | |
1905 | pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", | |
1906 | min_free_kbytes, recommended_min); | |
1907 | ||
1908 | min_free_kbytes = recommended_min; | |
1909 | } | |
1910 | setup_per_zone_wmarks(); | |
1911 | } | |
1912 | ||
1913 | int start_stop_khugepaged(void) | |
1914 | { | |
1915 | static struct task_struct *khugepaged_thread __read_mostly; | |
1916 | static DEFINE_MUTEX(khugepaged_mutex); | |
1917 | int err = 0; | |
1918 | ||
1919 | mutex_lock(&khugepaged_mutex); | |
1920 | if (khugepaged_enabled()) { | |
1921 | if (!khugepaged_thread) | |
1922 | khugepaged_thread = kthread_run(khugepaged, NULL, | |
1923 | "khugepaged"); | |
1924 | if (IS_ERR(khugepaged_thread)) { | |
1925 | pr_err("khugepaged: kthread_run(khugepaged) failed\n"); | |
1926 | err = PTR_ERR(khugepaged_thread); | |
1927 | khugepaged_thread = NULL; | |
1928 | goto fail; | |
1929 | } | |
1930 | ||
1931 | if (!list_empty(&khugepaged_scan.mm_head)) | |
1932 | wake_up_interruptible(&khugepaged_wait); | |
1933 | ||
1934 | set_recommended_min_free_kbytes(); | |
1935 | } else if (khugepaged_thread) { | |
1936 | kthread_stop(khugepaged_thread); | |
1937 | khugepaged_thread = NULL; | |
1938 | } | |
1939 | fail: | |
1940 | mutex_unlock(&khugepaged_mutex); | |
1941 | return err; | |
1942 | } |